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 | */ | |
029632fb PZ |
6 | #include "sched.h" |
7 | ||
371bf427 VG |
8 | #include "pelt.h" |
9 | ||
ce0dbbbb | 10 | int sched_rr_timeslice = RR_TIMESLICE; |
975e155e | 11 | int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE; |
ce0dbbbb | 12 | |
029632fb PZ |
13 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); |
14 | ||
15 | struct rt_bandwidth def_rt_bandwidth; | |
16 | ||
17 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) | |
18 | { | |
19 | struct rt_bandwidth *rt_b = | |
20 | container_of(timer, struct rt_bandwidth, rt_period_timer); | |
029632fb | 21 | int idle = 0; |
77a4d1a1 | 22 | int overrun; |
029632fb | 23 | |
77a4d1a1 | 24 | raw_spin_lock(&rt_b->rt_runtime_lock); |
029632fb | 25 | for (;;) { |
77a4d1a1 | 26 | overrun = hrtimer_forward_now(timer, rt_b->rt_period); |
029632fb PZ |
27 | if (!overrun) |
28 | break; | |
29 | ||
77a4d1a1 | 30 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
029632fb | 31 | idle = do_sched_rt_period_timer(rt_b, overrun); |
77a4d1a1 | 32 | raw_spin_lock(&rt_b->rt_runtime_lock); |
029632fb | 33 | } |
4cfafd30 PZ |
34 | if (idle) |
35 | rt_b->rt_period_active = 0; | |
77a4d1a1 | 36 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
029632fb PZ |
37 | |
38 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
39 | } | |
40 | ||
41 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) | |
42 | { | |
43 | rt_b->rt_period = ns_to_ktime(period); | |
44 | rt_b->rt_runtime = runtime; | |
45 | ||
46 | raw_spin_lock_init(&rt_b->rt_runtime_lock); | |
47 | ||
d5096aa6 SAS |
48 | hrtimer_init(&rt_b->rt_period_timer, CLOCK_MONOTONIC, |
49 | HRTIMER_MODE_REL_HARD); | |
029632fb PZ |
50 | rt_b->rt_period_timer.function = sched_rt_period_timer; |
51 | } | |
52 | ||
53 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) | |
54 | { | |
55 | if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) | |
56 | return; | |
57 | ||
029632fb | 58 | raw_spin_lock(&rt_b->rt_runtime_lock); |
4cfafd30 PZ |
59 | if (!rt_b->rt_period_active) { |
60 | rt_b->rt_period_active = 1; | |
c3a990dc SR |
61 | /* |
62 | * SCHED_DEADLINE updates the bandwidth, as a run away | |
63 | * RT task with a DL task could hog a CPU. But DL does | |
64 | * not reset the period. If a deadline task was running | |
65 | * without an RT task running, it can cause RT tasks to | |
66 | * throttle when they start up. Kick the timer right away | |
67 | * to update the period. | |
68 | */ | |
69 | hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0)); | |
d5096aa6 SAS |
70 | hrtimer_start_expires(&rt_b->rt_period_timer, |
71 | HRTIMER_MODE_ABS_PINNED_HARD); | |
4cfafd30 | 72 | } |
029632fb PZ |
73 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
74 | } | |
75 | ||
07c54f7a | 76 | void init_rt_rq(struct rt_rq *rt_rq) |
029632fb PZ |
77 | { |
78 | struct rt_prio_array *array; | |
79 | int i; | |
80 | ||
81 | array = &rt_rq->active; | |
82 | for (i = 0; i < MAX_RT_PRIO; i++) { | |
83 | INIT_LIST_HEAD(array->queue + i); | |
84 | __clear_bit(i, array->bitmap); | |
85 | } | |
86 | /* delimiter for bitsearch: */ | |
87 | __set_bit(MAX_RT_PRIO, array->bitmap); | |
88 | ||
89 | #if defined CONFIG_SMP | |
90 | rt_rq->highest_prio.curr = MAX_RT_PRIO; | |
91 | rt_rq->highest_prio.next = MAX_RT_PRIO; | |
92 | rt_rq->rt_nr_migratory = 0; | |
93 | rt_rq->overloaded = 0; | |
94 | plist_head_init(&rt_rq->pushable_tasks); | |
b6366f04 | 95 | #endif /* CONFIG_SMP */ |
f4ebcbc0 KT |
96 | /* We start is dequeued state, because no RT tasks are queued */ |
97 | rt_rq->rt_queued = 0; | |
029632fb PZ |
98 | |
99 | rt_rq->rt_time = 0; | |
100 | rt_rq->rt_throttled = 0; | |
101 | rt_rq->rt_runtime = 0; | |
102 | raw_spin_lock_init(&rt_rq->rt_runtime_lock); | |
103 | } | |
104 | ||
8f48894f | 105 | #ifdef CONFIG_RT_GROUP_SCHED |
029632fb PZ |
106 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) |
107 | { | |
108 | hrtimer_cancel(&rt_b->rt_period_timer); | |
109 | } | |
8f48894f PZ |
110 | |
111 | #define rt_entity_is_task(rt_se) (!(rt_se)->my_q) | |
112 | ||
398a153b GH |
113 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) |
114 | { | |
8f48894f PZ |
115 | #ifdef CONFIG_SCHED_DEBUG |
116 | WARN_ON_ONCE(!rt_entity_is_task(rt_se)); | |
117 | #endif | |
398a153b GH |
118 | return container_of(rt_se, struct task_struct, rt); |
119 | } | |
120 | ||
398a153b GH |
121 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) |
122 | { | |
123 | return rt_rq->rq; | |
124 | } | |
125 | ||
126 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) | |
127 | { | |
128 | return rt_se->rt_rq; | |
129 | } | |
130 | ||
653d07a6 KT |
131 | static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se) |
132 | { | |
133 | struct rt_rq *rt_rq = rt_se->rt_rq; | |
134 | ||
135 | return rt_rq->rq; | |
136 | } | |
137 | ||
029632fb PZ |
138 | void free_rt_sched_group(struct task_group *tg) |
139 | { | |
140 | int i; | |
141 | ||
142 | if (tg->rt_se) | |
143 | destroy_rt_bandwidth(&tg->rt_bandwidth); | |
144 | ||
145 | for_each_possible_cpu(i) { | |
146 | if (tg->rt_rq) | |
147 | kfree(tg->rt_rq[i]); | |
148 | if (tg->rt_se) | |
149 | kfree(tg->rt_se[i]); | |
150 | } | |
151 | ||
152 | kfree(tg->rt_rq); | |
153 | kfree(tg->rt_se); | |
154 | } | |
155 | ||
156 | void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, | |
157 | struct sched_rt_entity *rt_se, int cpu, | |
158 | struct sched_rt_entity *parent) | |
159 | { | |
160 | struct rq *rq = cpu_rq(cpu); | |
161 | ||
162 | rt_rq->highest_prio.curr = MAX_RT_PRIO; | |
163 | rt_rq->rt_nr_boosted = 0; | |
164 | rt_rq->rq = rq; | |
165 | rt_rq->tg = tg; | |
166 | ||
167 | tg->rt_rq[cpu] = rt_rq; | |
168 | tg->rt_se[cpu] = rt_se; | |
169 | ||
170 | if (!rt_se) | |
171 | return; | |
172 | ||
173 | if (!parent) | |
174 | rt_se->rt_rq = &rq->rt; | |
175 | else | |
176 | rt_se->rt_rq = parent->my_q; | |
177 | ||
178 | rt_se->my_q = rt_rq; | |
179 | rt_se->parent = parent; | |
180 | INIT_LIST_HEAD(&rt_se->run_list); | |
181 | } | |
182 | ||
183 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
184 | { | |
185 | struct rt_rq *rt_rq; | |
186 | struct sched_rt_entity *rt_se; | |
187 | int i; | |
188 | ||
6396bb22 | 189 | tg->rt_rq = kcalloc(nr_cpu_ids, sizeof(rt_rq), GFP_KERNEL); |
029632fb PZ |
190 | if (!tg->rt_rq) |
191 | goto err; | |
6396bb22 | 192 | tg->rt_se = kcalloc(nr_cpu_ids, sizeof(rt_se), GFP_KERNEL); |
029632fb PZ |
193 | if (!tg->rt_se) |
194 | goto err; | |
195 | ||
196 | init_rt_bandwidth(&tg->rt_bandwidth, | |
197 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); | |
198 | ||
199 | for_each_possible_cpu(i) { | |
200 | rt_rq = kzalloc_node(sizeof(struct rt_rq), | |
201 | GFP_KERNEL, cpu_to_node(i)); | |
202 | if (!rt_rq) | |
203 | goto err; | |
204 | ||
205 | rt_se = kzalloc_node(sizeof(struct sched_rt_entity), | |
206 | GFP_KERNEL, cpu_to_node(i)); | |
207 | if (!rt_se) | |
208 | goto err_free_rq; | |
209 | ||
07c54f7a | 210 | init_rt_rq(rt_rq); |
029632fb PZ |
211 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; |
212 | init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]); | |
213 | } | |
214 | ||
215 | return 1; | |
216 | ||
217 | err_free_rq: | |
218 | kfree(rt_rq); | |
219 | err: | |
220 | return 0; | |
221 | } | |
222 | ||
398a153b GH |
223 | #else /* CONFIG_RT_GROUP_SCHED */ |
224 | ||
a1ba4d8b PZ |
225 | #define rt_entity_is_task(rt_se) (1) |
226 | ||
8f48894f PZ |
227 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) |
228 | { | |
229 | return container_of(rt_se, struct task_struct, rt); | |
230 | } | |
231 | ||
398a153b GH |
232 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) |
233 | { | |
234 | return container_of(rt_rq, struct rq, rt); | |
235 | } | |
236 | ||
653d07a6 | 237 | static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se) |
398a153b GH |
238 | { |
239 | struct task_struct *p = rt_task_of(rt_se); | |
653d07a6 KT |
240 | |
241 | return task_rq(p); | |
242 | } | |
243 | ||
244 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) | |
245 | { | |
246 | struct rq *rq = rq_of_rt_se(rt_se); | |
398a153b GH |
247 | |
248 | return &rq->rt; | |
249 | } | |
250 | ||
029632fb PZ |
251 | void free_rt_sched_group(struct task_group *tg) { } |
252 | ||
253 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
254 | { | |
255 | return 1; | |
256 | } | |
398a153b GH |
257 | #endif /* CONFIG_RT_GROUP_SCHED */ |
258 | ||
4fd29176 | 259 | #ifdef CONFIG_SMP |
84de4274 | 260 | |
8046d680 | 261 | static void pull_rt_task(struct rq *this_rq); |
38033c37 | 262 | |
dc877341 PZ |
263 | static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev) |
264 | { | |
265 | /* Try to pull RT tasks here if we lower this rq's prio */ | |
266 | return rq->rt.highest_prio.curr > prev->prio; | |
267 | } | |
268 | ||
637f5085 | 269 | static inline int rt_overloaded(struct rq *rq) |
4fd29176 | 270 | { |
637f5085 | 271 | return atomic_read(&rq->rd->rto_count); |
4fd29176 | 272 | } |
84de4274 | 273 | |
4fd29176 SR |
274 | static inline void rt_set_overload(struct rq *rq) |
275 | { | |
1f11eb6a GH |
276 | if (!rq->online) |
277 | return; | |
278 | ||
c6c4927b | 279 | cpumask_set_cpu(rq->cpu, rq->rd->rto_mask); |
4fd29176 SR |
280 | /* |
281 | * Make sure the mask is visible before we set | |
282 | * the overload count. That is checked to determine | |
283 | * if we should look at the mask. It would be a shame | |
284 | * if we looked at the mask, but the mask was not | |
285 | * updated yet. | |
7c3f2ab7 PZ |
286 | * |
287 | * Matched by the barrier in pull_rt_task(). | |
4fd29176 | 288 | */ |
7c3f2ab7 | 289 | smp_wmb(); |
637f5085 | 290 | atomic_inc(&rq->rd->rto_count); |
4fd29176 | 291 | } |
84de4274 | 292 | |
4fd29176 SR |
293 | static inline void rt_clear_overload(struct rq *rq) |
294 | { | |
1f11eb6a GH |
295 | if (!rq->online) |
296 | return; | |
297 | ||
4fd29176 | 298 | /* the order here really doesn't matter */ |
637f5085 | 299 | atomic_dec(&rq->rd->rto_count); |
c6c4927b | 300 | cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask); |
4fd29176 | 301 | } |
73fe6aae | 302 | |
398a153b | 303 | static void update_rt_migration(struct rt_rq *rt_rq) |
73fe6aae | 304 | { |
a1ba4d8b | 305 | if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) { |
398a153b GH |
306 | if (!rt_rq->overloaded) { |
307 | rt_set_overload(rq_of_rt_rq(rt_rq)); | |
308 | rt_rq->overloaded = 1; | |
cdc8eb98 | 309 | } |
398a153b GH |
310 | } else if (rt_rq->overloaded) { |
311 | rt_clear_overload(rq_of_rt_rq(rt_rq)); | |
312 | rt_rq->overloaded = 0; | |
637f5085 | 313 | } |
73fe6aae | 314 | } |
4fd29176 | 315 | |
398a153b GH |
316 | static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
317 | { | |
29baa747 PZ |
318 | struct task_struct *p; |
319 | ||
a1ba4d8b PZ |
320 | if (!rt_entity_is_task(rt_se)) |
321 | return; | |
322 | ||
29baa747 | 323 | p = rt_task_of(rt_se); |
a1ba4d8b PZ |
324 | rt_rq = &rq_of_rt_rq(rt_rq)->rt; |
325 | ||
326 | rt_rq->rt_nr_total++; | |
4b53a341 | 327 | if (p->nr_cpus_allowed > 1) |
398a153b GH |
328 | rt_rq->rt_nr_migratory++; |
329 | ||
330 | update_rt_migration(rt_rq); | |
331 | } | |
332 | ||
333 | static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
334 | { | |
29baa747 PZ |
335 | struct task_struct *p; |
336 | ||
a1ba4d8b PZ |
337 | if (!rt_entity_is_task(rt_se)) |
338 | return; | |
339 | ||
29baa747 | 340 | p = rt_task_of(rt_se); |
a1ba4d8b PZ |
341 | rt_rq = &rq_of_rt_rq(rt_rq)->rt; |
342 | ||
343 | rt_rq->rt_nr_total--; | |
4b53a341 | 344 | if (p->nr_cpus_allowed > 1) |
398a153b GH |
345 | rt_rq->rt_nr_migratory--; |
346 | ||
347 | update_rt_migration(rt_rq); | |
348 | } | |
349 | ||
5181f4a4 SR |
350 | static inline int has_pushable_tasks(struct rq *rq) |
351 | { | |
352 | return !plist_head_empty(&rq->rt.pushable_tasks); | |
353 | } | |
354 | ||
fd7a4bed PZ |
355 | static DEFINE_PER_CPU(struct callback_head, rt_push_head); |
356 | static DEFINE_PER_CPU(struct callback_head, rt_pull_head); | |
e3fca9e7 PZ |
357 | |
358 | static void push_rt_tasks(struct rq *); | |
fd7a4bed | 359 | static void pull_rt_task(struct rq *); |
e3fca9e7 | 360 | |
02d8ec94 | 361 | static inline void rt_queue_push_tasks(struct rq *rq) |
dc877341 | 362 | { |
e3fca9e7 PZ |
363 | if (!has_pushable_tasks(rq)) |
364 | return; | |
365 | ||
fd7a4bed PZ |
366 | queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks); |
367 | } | |
368 | ||
02d8ec94 | 369 | static inline void rt_queue_pull_task(struct rq *rq) |
fd7a4bed PZ |
370 | { |
371 | queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task); | |
dc877341 PZ |
372 | } |
373 | ||
917b627d GH |
374 | static void enqueue_pushable_task(struct rq *rq, struct task_struct *p) |
375 | { | |
376 | plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); | |
377 | plist_node_init(&p->pushable_tasks, p->prio); | |
378 | plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks); | |
5181f4a4 SR |
379 | |
380 | /* Update the highest prio pushable task */ | |
381 | if (p->prio < rq->rt.highest_prio.next) | |
382 | rq->rt.highest_prio.next = p->prio; | |
917b627d GH |
383 | } |
384 | ||
385 | static void dequeue_pushable_task(struct rq *rq, struct task_struct *p) | |
386 | { | |
387 | plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); | |
917b627d | 388 | |
5181f4a4 SR |
389 | /* Update the new highest prio pushable task */ |
390 | if (has_pushable_tasks(rq)) { | |
391 | p = plist_first_entry(&rq->rt.pushable_tasks, | |
392 | struct task_struct, pushable_tasks); | |
393 | rq->rt.highest_prio.next = p->prio; | |
394 | } else | |
395 | rq->rt.highest_prio.next = MAX_RT_PRIO; | |
bcf08df3 IM |
396 | } |
397 | ||
917b627d GH |
398 | #else |
399 | ||
ceacc2c1 | 400 | static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p) |
fa85ae24 | 401 | { |
6f505b16 PZ |
402 | } |
403 | ||
ceacc2c1 PZ |
404 | static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p) |
405 | { | |
406 | } | |
407 | ||
b07430ac | 408 | static inline |
ceacc2c1 PZ |
409 | void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
410 | { | |
411 | } | |
412 | ||
398a153b | 413 | static inline |
ceacc2c1 PZ |
414 | void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
415 | { | |
416 | } | |
917b627d | 417 | |
dc877341 PZ |
418 | static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev) |
419 | { | |
420 | return false; | |
421 | } | |
422 | ||
8046d680 | 423 | static inline void pull_rt_task(struct rq *this_rq) |
dc877341 | 424 | { |
dc877341 PZ |
425 | } |
426 | ||
02d8ec94 | 427 | static inline void rt_queue_push_tasks(struct rq *rq) |
dc877341 PZ |
428 | { |
429 | } | |
4fd29176 SR |
430 | #endif /* CONFIG_SMP */ |
431 | ||
f4ebcbc0 KT |
432 | static void enqueue_top_rt_rq(struct rt_rq *rt_rq); |
433 | static void dequeue_top_rt_rq(struct rt_rq *rt_rq); | |
434 | ||
6f505b16 PZ |
435 | static inline int on_rt_rq(struct sched_rt_entity *rt_se) |
436 | { | |
ff77e468 | 437 | return rt_se->on_rq; |
6f505b16 PZ |
438 | } |
439 | ||
804d402f QY |
440 | #ifdef CONFIG_UCLAMP_TASK |
441 | /* | |
442 | * Verify the fitness of task @p to run on @cpu taking into account the uclamp | |
443 | * settings. | |
444 | * | |
445 | * This check is only important for heterogeneous systems where uclamp_min value | |
446 | * is higher than the capacity of a @cpu. For non-heterogeneous system this | |
447 | * function will always return true. | |
448 | * | |
449 | * The function will return true if the capacity of the @cpu is >= the | |
450 | * uclamp_min and false otherwise. | |
451 | * | |
452 | * Note that uclamp_min will be clamped to uclamp_max if uclamp_min | |
453 | * > uclamp_max. | |
454 | */ | |
455 | static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu) | |
456 | { | |
457 | unsigned int min_cap; | |
458 | unsigned int max_cap; | |
459 | unsigned int cpu_cap; | |
460 | ||
461 | /* Only heterogeneous systems can benefit from this check */ | |
462 | if (!static_branch_unlikely(&sched_asym_cpucapacity)) | |
463 | return true; | |
464 | ||
465 | min_cap = uclamp_eff_value(p, UCLAMP_MIN); | |
466 | max_cap = uclamp_eff_value(p, UCLAMP_MAX); | |
467 | ||
468 | cpu_cap = capacity_orig_of(cpu); | |
469 | ||
470 | return cpu_cap >= min(min_cap, max_cap); | |
471 | } | |
472 | #else | |
473 | static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu) | |
474 | { | |
475 | return true; | |
476 | } | |
477 | #endif | |
478 | ||
052f1dc7 | 479 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 480 | |
9f0c1e56 | 481 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) |
6f505b16 PZ |
482 | { |
483 | if (!rt_rq->tg) | |
9f0c1e56 | 484 | return RUNTIME_INF; |
6f505b16 | 485 | |
ac086bc2 PZ |
486 | return rt_rq->rt_runtime; |
487 | } | |
488 | ||
489 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) | |
490 | { | |
491 | return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period); | |
6f505b16 PZ |
492 | } |
493 | ||
ec514c48 CX |
494 | typedef struct task_group *rt_rq_iter_t; |
495 | ||
1c09ab0d YZ |
496 | static inline struct task_group *next_task_group(struct task_group *tg) |
497 | { | |
498 | do { | |
499 | tg = list_entry_rcu(tg->list.next, | |
500 | typeof(struct task_group), list); | |
501 | } while (&tg->list != &task_groups && task_group_is_autogroup(tg)); | |
502 | ||
503 | if (&tg->list == &task_groups) | |
504 | tg = NULL; | |
505 | ||
506 | return tg; | |
507 | } | |
508 | ||
509 | #define for_each_rt_rq(rt_rq, iter, rq) \ | |
510 | for (iter = container_of(&task_groups, typeof(*iter), list); \ | |
511 | (iter = next_task_group(iter)) && \ | |
512 | (rt_rq = iter->rt_rq[cpu_of(rq)]);) | |
ec514c48 | 513 | |
6f505b16 PZ |
514 | #define for_each_sched_rt_entity(rt_se) \ |
515 | for (; rt_se; rt_se = rt_se->parent) | |
516 | ||
517 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) | |
518 | { | |
519 | return rt_se->my_q; | |
520 | } | |
521 | ||
ff77e468 PZ |
522 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags); |
523 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags); | |
6f505b16 | 524 | |
9f0c1e56 | 525 | static void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
6f505b16 | 526 | { |
f6121f4f | 527 | struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr; |
8875125e | 528 | struct rq *rq = rq_of_rt_rq(rt_rq); |
74b7eb58 YZ |
529 | struct sched_rt_entity *rt_se; |
530 | ||
8875125e | 531 | int cpu = cpu_of(rq); |
0c3b9168 BS |
532 | |
533 | rt_se = rt_rq->tg->rt_se[cpu]; | |
6f505b16 | 534 | |
f6121f4f | 535 | if (rt_rq->rt_nr_running) { |
f4ebcbc0 KT |
536 | if (!rt_se) |
537 | enqueue_top_rt_rq(rt_rq); | |
538 | else if (!on_rt_rq(rt_se)) | |
ff77e468 | 539 | enqueue_rt_entity(rt_se, 0); |
f4ebcbc0 | 540 | |
e864c499 | 541 | if (rt_rq->highest_prio.curr < curr->prio) |
8875125e | 542 | resched_curr(rq); |
6f505b16 PZ |
543 | } |
544 | } | |
545 | ||
9f0c1e56 | 546 | static void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
6f505b16 | 547 | { |
74b7eb58 | 548 | struct sched_rt_entity *rt_se; |
0c3b9168 | 549 | int cpu = cpu_of(rq_of_rt_rq(rt_rq)); |
74b7eb58 | 550 | |
0c3b9168 | 551 | rt_se = rt_rq->tg->rt_se[cpu]; |
6f505b16 | 552 | |
296b2ffe | 553 | if (!rt_se) { |
f4ebcbc0 | 554 | dequeue_top_rt_rq(rt_rq); |
296b2ffe VG |
555 | /* Kick cpufreq (see the comment in kernel/sched/sched.h). */ |
556 | cpufreq_update_util(rq_of_rt_rq(rt_rq), 0); | |
557 | } | |
f4ebcbc0 | 558 | else if (on_rt_rq(rt_se)) |
ff77e468 | 559 | dequeue_rt_entity(rt_se, 0); |
6f505b16 PZ |
560 | } |
561 | ||
46383648 KT |
562 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
563 | { | |
564 | return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted; | |
565 | } | |
566 | ||
23b0fdfc PZ |
567 | static int rt_se_boosted(struct sched_rt_entity *rt_se) |
568 | { | |
569 | struct rt_rq *rt_rq = group_rt_rq(rt_se); | |
570 | struct task_struct *p; | |
571 | ||
572 | if (rt_rq) | |
573 | return !!rt_rq->rt_nr_boosted; | |
574 | ||
575 | p = rt_task_of(rt_se); | |
576 | return p->prio != p->normal_prio; | |
577 | } | |
578 | ||
d0b27fa7 | 579 | #ifdef CONFIG_SMP |
c6c4927b | 580 | static inline const struct cpumask *sched_rt_period_mask(void) |
d0b27fa7 | 581 | { |
424c93fe | 582 | return this_rq()->rd->span; |
d0b27fa7 | 583 | } |
6f505b16 | 584 | #else |
c6c4927b | 585 | static inline const struct cpumask *sched_rt_period_mask(void) |
d0b27fa7 | 586 | { |
c6c4927b | 587 | return cpu_online_mask; |
d0b27fa7 PZ |
588 | } |
589 | #endif | |
6f505b16 | 590 | |
d0b27fa7 PZ |
591 | static inline |
592 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) | |
6f505b16 | 593 | { |
d0b27fa7 PZ |
594 | return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu]; |
595 | } | |
9f0c1e56 | 596 | |
ac086bc2 PZ |
597 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
598 | { | |
599 | return &rt_rq->tg->rt_bandwidth; | |
600 | } | |
601 | ||
55e12e5e | 602 | #else /* !CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 PZ |
603 | |
604 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) | |
605 | { | |
ac086bc2 PZ |
606 | return rt_rq->rt_runtime; |
607 | } | |
608 | ||
609 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) | |
610 | { | |
611 | return ktime_to_ns(def_rt_bandwidth.rt_period); | |
6f505b16 PZ |
612 | } |
613 | ||
ec514c48 CX |
614 | typedef struct rt_rq *rt_rq_iter_t; |
615 | ||
616 | #define for_each_rt_rq(rt_rq, iter, rq) \ | |
617 | for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL) | |
618 | ||
6f505b16 PZ |
619 | #define for_each_sched_rt_entity(rt_se) \ |
620 | for (; rt_se; rt_se = NULL) | |
621 | ||
622 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) | |
623 | { | |
624 | return NULL; | |
625 | } | |
626 | ||
9f0c1e56 | 627 | static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
6f505b16 | 628 | { |
f4ebcbc0 KT |
629 | struct rq *rq = rq_of_rt_rq(rt_rq); |
630 | ||
631 | if (!rt_rq->rt_nr_running) | |
632 | return; | |
633 | ||
634 | enqueue_top_rt_rq(rt_rq); | |
8875125e | 635 | resched_curr(rq); |
6f505b16 PZ |
636 | } |
637 | ||
9f0c1e56 | 638 | static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
6f505b16 | 639 | { |
f4ebcbc0 | 640 | dequeue_top_rt_rq(rt_rq); |
6f505b16 PZ |
641 | } |
642 | ||
46383648 KT |
643 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
644 | { | |
645 | return rt_rq->rt_throttled; | |
646 | } | |
647 | ||
c6c4927b | 648 | static inline const struct cpumask *sched_rt_period_mask(void) |
d0b27fa7 | 649 | { |
c6c4927b | 650 | return cpu_online_mask; |
d0b27fa7 PZ |
651 | } |
652 | ||
653 | static inline | |
654 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) | |
655 | { | |
656 | return &cpu_rq(cpu)->rt; | |
657 | } | |
658 | ||
ac086bc2 PZ |
659 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
660 | { | |
661 | return &def_rt_bandwidth; | |
662 | } | |
663 | ||
55e12e5e | 664 | #endif /* CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 | 665 | |
faa59937 JL |
666 | bool sched_rt_bandwidth_account(struct rt_rq *rt_rq) |
667 | { | |
668 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); | |
669 | ||
670 | return (hrtimer_active(&rt_b->rt_period_timer) || | |
671 | rt_rq->rt_time < rt_b->rt_runtime); | |
672 | } | |
673 | ||
ac086bc2 | 674 | #ifdef CONFIG_SMP |
78333cdd PZ |
675 | /* |
676 | * We ran out of runtime, see if we can borrow some from our neighbours. | |
677 | */ | |
269b26a5 | 678 | static void do_balance_runtime(struct rt_rq *rt_rq) |
ac086bc2 PZ |
679 | { |
680 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); | |
aa7f6730 | 681 | struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd; |
269b26a5 | 682 | int i, weight; |
ac086bc2 PZ |
683 | u64 rt_period; |
684 | ||
c6c4927b | 685 | weight = cpumask_weight(rd->span); |
ac086bc2 | 686 | |
0986b11b | 687 | raw_spin_lock(&rt_b->rt_runtime_lock); |
ac086bc2 | 688 | rt_period = ktime_to_ns(rt_b->rt_period); |
c6c4927b | 689 | for_each_cpu(i, rd->span) { |
ac086bc2 PZ |
690 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); |
691 | s64 diff; | |
692 | ||
693 | if (iter == rt_rq) | |
694 | continue; | |
695 | ||
0986b11b | 696 | raw_spin_lock(&iter->rt_runtime_lock); |
78333cdd PZ |
697 | /* |
698 | * Either all rqs have inf runtime and there's nothing to steal | |
699 | * or __disable_runtime() below sets a specific rq to inf to | |
700 | * indicate its been disabled and disalow stealing. | |
701 | */ | |
7def2be1 PZ |
702 | if (iter->rt_runtime == RUNTIME_INF) |
703 | goto next; | |
704 | ||
78333cdd PZ |
705 | /* |
706 | * From runqueues with spare time, take 1/n part of their | |
707 | * spare time, but no more than our period. | |
708 | */ | |
ac086bc2 PZ |
709 | diff = iter->rt_runtime - iter->rt_time; |
710 | if (diff > 0) { | |
58838cf3 | 711 | diff = div_u64((u64)diff, weight); |
ac086bc2 PZ |
712 | if (rt_rq->rt_runtime + diff > rt_period) |
713 | diff = rt_period - rt_rq->rt_runtime; | |
714 | iter->rt_runtime -= diff; | |
715 | rt_rq->rt_runtime += diff; | |
ac086bc2 | 716 | if (rt_rq->rt_runtime == rt_period) { |
0986b11b | 717 | raw_spin_unlock(&iter->rt_runtime_lock); |
ac086bc2 PZ |
718 | break; |
719 | } | |
720 | } | |
7def2be1 | 721 | next: |
0986b11b | 722 | raw_spin_unlock(&iter->rt_runtime_lock); |
ac086bc2 | 723 | } |
0986b11b | 724 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
ac086bc2 | 725 | } |
7def2be1 | 726 | |
78333cdd PZ |
727 | /* |
728 | * Ensure this RQ takes back all the runtime it lend to its neighbours. | |
729 | */ | |
7def2be1 PZ |
730 | static void __disable_runtime(struct rq *rq) |
731 | { | |
732 | struct root_domain *rd = rq->rd; | |
ec514c48 | 733 | rt_rq_iter_t iter; |
7def2be1 PZ |
734 | struct rt_rq *rt_rq; |
735 | ||
736 | if (unlikely(!scheduler_running)) | |
737 | return; | |
738 | ||
ec514c48 | 739 | for_each_rt_rq(rt_rq, iter, rq) { |
7def2be1 PZ |
740 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
741 | s64 want; | |
742 | int i; | |
743 | ||
0986b11b TG |
744 | raw_spin_lock(&rt_b->rt_runtime_lock); |
745 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
78333cdd PZ |
746 | /* |
747 | * Either we're all inf and nobody needs to borrow, or we're | |
748 | * already disabled and thus have nothing to do, or we have | |
749 | * exactly the right amount of runtime to take out. | |
750 | */ | |
7def2be1 PZ |
751 | if (rt_rq->rt_runtime == RUNTIME_INF || |
752 | rt_rq->rt_runtime == rt_b->rt_runtime) | |
753 | goto balanced; | |
0986b11b | 754 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
7def2be1 | 755 | |
78333cdd PZ |
756 | /* |
757 | * Calculate the difference between what we started out with | |
758 | * and what we current have, that's the amount of runtime | |
759 | * we lend and now have to reclaim. | |
760 | */ | |
7def2be1 PZ |
761 | want = rt_b->rt_runtime - rt_rq->rt_runtime; |
762 | ||
78333cdd PZ |
763 | /* |
764 | * Greedy reclaim, take back as much as we can. | |
765 | */ | |
c6c4927b | 766 | for_each_cpu(i, rd->span) { |
7def2be1 PZ |
767 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); |
768 | s64 diff; | |
769 | ||
78333cdd PZ |
770 | /* |
771 | * Can't reclaim from ourselves or disabled runqueues. | |
772 | */ | |
f1679d08 | 773 | if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF) |
7def2be1 PZ |
774 | continue; |
775 | ||
0986b11b | 776 | raw_spin_lock(&iter->rt_runtime_lock); |
7def2be1 PZ |
777 | if (want > 0) { |
778 | diff = min_t(s64, iter->rt_runtime, want); | |
779 | iter->rt_runtime -= diff; | |
780 | want -= diff; | |
781 | } else { | |
782 | iter->rt_runtime -= want; | |
783 | want -= want; | |
784 | } | |
0986b11b | 785 | raw_spin_unlock(&iter->rt_runtime_lock); |
7def2be1 PZ |
786 | |
787 | if (!want) | |
788 | break; | |
789 | } | |
790 | ||
0986b11b | 791 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
78333cdd PZ |
792 | /* |
793 | * We cannot be left wanting - that would mean some runtime | |
794 | * leaked out of the system. | |
795 | */ | |
7def2be1 PZ |
796 | BUG_ON(want); |
797 | balanced: | |
78333cdd PZ |
798 | /* |
799 | * Disable all the borrow logic by pretending we have inf | |
800 | * runtime - in which case borrowing doesn't make sense. | |
801 | */ | |
7def2be1 | 802 | rt_rq->rt_runtime = RUNTIME_INF; |
a4c96ae3 | 803 | rt_rq->rt_throttled = 0; |
0986b11b TG |
804 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
805 | raw_spin_unlock(&rt_b->rt_runtime_lock); | |
99b62567 KT |
806 | |
807 | /* Make rt_rq available for pick_next_task() */ | |
808 | sched_rt_rq_enqueue(rt_rq); | |
7def2be1 PZ |
809 | } |
810 | } | |
811 | ||
7def2be1 PZ |
812 | static void __enable_runtime(struct rq *rq) |
813 | { | |
ec514c48 | 814 | rt_rq_iter_t iter; |
7def2be1 PZ |
815 | struct rt_rq *rt_rq; |
816 | ||
817 | if (unlikely(!scheduler_running)) | |
818 | return; | |
819 | ||
78333cdd PZ |
820 | /* |
821 | * Reset each runqueue's bandwidth settings | |
822 | */ | |
ec514c48 | 823 | for_each_rt_rq(rt_rq, iter, rq) { |
7def2be1 PZ |
824 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
825 | ||
0986b11b TG |
826 | raw_spin_lock(&rt_b->rt_runtime_lock); |
827 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
7def2be1 PZ |
828 | rt_rq->rt_runtime = rt_b->rt_runtime; |
829 | rt_rq->rt_time = 0; | |
baf25731 | 830 | rt_rq->rt_throttled = 0; |
0986b11b TG |
831 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
832 | raw_spin_unlock(&rt_b->rt_runtime_lock); | |
7def2be1 PZ |
833 | } |
834 | } | |
835 | ||
269b26a5 | 836 | static void balance_runtime(struct rt_rq *rt_rq) |
eff6549b | 837 | { |
4a6184ce | 838 | if (!sched_feat(RT_RUNTIME_SHARE)) |
269b26a5 | 839 | return; |
4a6184ce | 840 | |
eff6549b | 841 | if (rt_rq->rt_time > rt_rq->rt_runtime) { |
0986b11b | 842 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
269b26a5 | 843 | do_balance_runtime(rt_rq); |
0986b11b | 844 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
eff6549b | 845 | } |
eff6549b | 846 | } |
55e12e5e | 847 | #else /* !CONFIG_SMP */ |
269b26a5 | 848 | static inline void balance_runtime(struct rt_rq *rt_rq) {} |
55e12e5e | 849 | #endif /* CONFIG_SMP */ |
ac086bc2 | 850 | |
eff6549b PZ |
851 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun) |
852 | { | |
42c62a58 | 853 | int i, idle = 1, throttled = 0; |
c6c4927b | 854 | const struct cpumask *span; |
eff6549b | 855 | |
eff6549b | 856 | span = sched_rt_period_mask(); |
e221d028 MG |
857 | #ifdef CONFIG_RT_GROUP_SCHED |
858 | /* | |
859 | * FIXME: isolated CPUs should really leave the root task group, | |
860 | * whether they are isolcpus or were isolated via cpusets, lest | |
861 | * the timer run on a CPU which does not service all runqueues, | |
862 | * potentially leaving other CPUs indefinitely throttled. If | |
863 | * isolation is really required, the user will turn the throttle | |
864 | * off to kill the perturbations it causes anyway. Meanwhile, | |
865 | * this maintains functionality for boot and/or troubleshooting. | |
866 | */ | |
867 | if (rt_b == &root_task_group.rt_bandwidth) | |
868 | span = cpu_online_mask; | |
869 | #endif | |
c6c4927b | 870 | for_each_cpu(i, span) { |
eff6549b PZ |
871 | int enqueue = 0; |
872 | struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i); | |
873 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
c249f255 DK |
874 | int skip; |
875 | ||
876 | /* | |
877 | * When span == cpu_online_mask, taking each rq->lock | |
878 | * can be time-consuming. Try to avoid it when possible. | |
879 | */ | |
880 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
f3d133ee HL |
881 | if (!sched_feat(RT_RUNTIME_SHARE) && rt_rq->rt_runtime != RUNTIME_INF) |
882 | rt_rq->rt_runtime = rt_b->rt_runtime; | |
c249f255 DK |
883 | skip = !rt_rq->rt_time && !rt_rq->rt_nr_running; |
884 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
885 | if (skip) | |
886 | continue; | |
eff6549b | 887 | |
05fa785c | 888 | raw_spin_lock(&rq->lock); |
d29a2064 DB |
889 | update_rq_clock(rq); |
890 | ||
eff6549b PZ |
891 | if (rt_rq->rt_time) { |
892 | u64 runtime; | |
893 | ||
0986b11b | 894 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
eff6549b PZ |
895 | if (rt_rq->rt_throttled) |
896 | balance_runtime(rt_rq); | |
897 | runtime = rt_rq->rt_runtime; | |
898 | rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime); | |
899 | if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) { | |
900 | rt_rq->rt_throttled = 0; | |
901 | enqueue = 1; | |
61eadef6 MG |
902 | |
903 | /* | |
9edfbfed PZ |
904 | * When we're idle and a woken (rt) task is |
905 | * throttled check_preempt_curr() will set | |
906 | * skip_update and the time between the wakeup | |
907 | * and this unthrottle will get accounted as | |
908 | * 'runtime'. | |
61eadef6 MG |
909 | */ |
910 | if (rt_rq->rt_nr_running && rq->curr == rq->idle) | |
adcc8da8 | 911 | rq_clock_cancel_skipupdate(rq); |
eff6549b PZ |
912 | } |
913 | if (rt_rq->rt_time || rt_rq->rt_nr_running) | |
914 | idle = 0; | |
0986b11b | 915 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
0c3b9168 | 916 | } else if (rt_rq->rt_nr_running) { |
6c3df255 | 917 | idle = 0; |
0c3b9168 BS |
918 | if (!rt_rq_throttled(rt_rq)) |
919 | enqueue = 1; | |
920 | } | |
42c62a58 PZ |
921 | if (rt_rq->rt_throttled) |
922 | throttled = 1; | |
eff6549b PZ |
923 | |
924 | if (enqueue) | |
925 | sched_rt_rq_enqueue(rt_rq); | |
05fa785c | 926 | raw_spin_unlock(&rq->lock); |
eff6549b PZ |
927 | } |
928 | ||
42c62a58 PZ |
929 | if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)) |
930 | return 1; | |
931 | ||
eff6549b PZ |
932 | return idle; |
933 | } | |
ac086bc2 | 934 | |
6f505b16 PZ |
935 | static inline int rt_se_prio(struct sched_rt_entity *rt_se) |
936 | { | |
052f1dc7 | 937 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
938 | struct rt_rq *rt_rq = group_rt_rq(rt_se); |
939 | ||
940 | if (rt_rq) | |
e864c499 | 941 | return rt_rq->highest_prio.curr; |
6f505b16 PZ |
942 | #endif |
943 | ||
944 | return rt_task_of(rt_se)->prio; | |
945 | } | |
946 | ||
9f0c1e56 | 947 | static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq) |
6f505b16 | 948 | { |
9f0c1e56 | 949 | u64 runtime = sched_rt_runtime(rt_rq); |
fa85ae24 | 950 | |
fa85ae24 | 951 | if (rt_rq->rt_throttled) |
23b0fdfc | 952 | return rt_rq_throttled(rt_rq); |
fa85ae24 | 953 | |
5b680fd6 | 954 | if (runtime >= sched_rt_period(rt_rq)) |
ac086bc2 PZ |
955 | return 0; |
956 | ||
b79f3833 PZ |
957 | balance_runtime(rt_rq); |
958 | runtime = sched_rt_runtime(rt_rq); | |
959 | if (runtime == RUNTIME_INF) | |
960 | return 0; | |
ac086bc2 | 961 | |
9f0c1e56 | 962 | if (rt_rq->rt_time > runtime) { |
7abc63b1 PZ |
963 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
964 | ||
965 | /* | |
966 | * Don't actually throttle groups that have no runtime assigned | |
967 | * but accrue some time due to boosting. | |
968 | */ | |
969 | if (likely(rt_b->rt_runtime)) { | |
970 | rt_rq->rt_throttled = 1; | |
c224815d | 971 | printk_deferred_once("sched: RT throttling activated\n"); |
7abc63b1 PZ |
972 | } else { |
973 | /* | |
974 | * In case we did anyway, make it go away, | |
975 | * replenishment is a joke, since it will replenish us | |
976 | * with exactly 0 ns. | |
977 | */ | |
978 | rt_rq->rt_time = 0; | |
979 | } | |
980 | ||
23b0fdfc | 981 | if (rt_rq_throttled(rt_rq)) { |
9f0c1e56 | 982 | sched_rt_rq_dequeue(rt_rq); |
23b0fdfc PZ |
983 | return 1; |
984 | } | |
fa85ae24 PZ |
985 | } |
986 | ||
987 | return 0; | |
988 | } | |
989 | ||
bb44e5d1 IM |
990 | /* |
991 | * Update the current task's runtime statistics. Skip current tasks that | |
992 | * are not in our scheduling class. | |
993 | */ | |
a9957449 | 994 | static void update_curr_rt(struct rq *rq) |
bb44e5d1 IM |
995 | { |
996 | struct task_struct *curr = rq->curr; | |
6f505b16 | 997 | struct sched_rt_entity *rt_se = &curr->rt; |
bb44e5d1 | 998 | u64 delta_exec; |
a7711602 | 999 | u64 now; |
bb44e5d1 | 1000 | |
06c3bc65 | 1001 | if (curr->sched_class != &rt_sched_class) |
bb44e5d1 IM |
1002 | return; |
1003 | ||
a7711602 | 1004 | now = rq_clock_task(rq); |
e7ad2031 | 1005 | delta_exec = now - curr->se.exec_start; |
fc79e240 KT |
1006 | if (unlikely((s64)delta_exec <= 0)) |
1007 | return; | |
6cfb0d5d | 1008 | |
42c62a58 PZ |
1009 | schedstat_set(curr->se.statistics.exec_max, |
1010 | max(curr->se.statistics.exec_max, delta_exec)); | |
bb44e5d1 IM |
1011 | |
1012 | curr->se.sum_exec_runtime += delta_exec; | |
f06febc9 FM |
1013 | account_group_exec_runtime(curr, delta_exec); |
1014 | ||
e7ad2031 | 1015 | curr->se.exec_start = now; |
d2cc5ed6 | 1016 | cgroup_account_cputime(curr, delta_exec); |
fa85ae24 | 1017 | |
0b148fa0 PZ |
1018 | if (!rt_bandwidth_enabled()) |
1019 | return; | |
1020 | ||
354d60c2 | 1021 | for_each_sched_rt_entity(rt_se) { |
0b07939c | 1022 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
354d60c2 | 1023 | |
cc2991cf | 1024 | if (sched_rt_runtime(rt_rq) != RUNTIME_INF) { |
0986b11b | 1025 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
cc2991cf PZ |
1026 | rt_rq->rt_time += delta_exec; |
1027 | if (sched_rt_runtime_exceeded(rt_rq)) | |
8875125e | 1028 | resched_curr(rq); |
0986b11b | 1029 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
cc2991cf | 1030 | } |
354d60c2 | 1031 | } |
bb44e5d1 IM |
1032 | } |
1033 | ||
f4ebcbc0 KT |
1034 | static void |
1035 | dequeue_top_rt_rq(struct rt_rq *rt_rq) | |
1036 | { | |
1037 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
1038 | ||
1039 | BUG_ON(&rq->rt != rt_rq); | |
1040 | ||
1041 | if (!rt_rq->rt_queued) | |
1042 | return; | |
1043 | ||
1044 | BUG_ON(!rq->nr_running); | |
1045 | ||
72465447 | 1046 | sub_nr_running(rq, rt_rq->rt_nr_running); |
f4ebcbc0 | 1047 | rt_rq->rt_queued = 0; |
8f111bc3 | 1048 | |
f4ebcbc0 KT |
1049 | } |
1050 | ||
1051 | static void | |
1052 | enqueue_top_rt_rq(struct rt_rq *rt_rq) | |
1053 | { | |
1054 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
1055 | ||
1056 | BUG_ON(&rq->rt != rt_rq); | |
1057 | ||
1058 | if (rt_rq->rt_queued) | |
1059 | return; | |
296b2ffe VG |
1060 | |
1061 | if (rt_rq_throttled(rt_rq)) | |
f4ebcbc0 KT |
1062 | return; |
1063 | ||
296b2ffe VG |
1064 | if (rt_rq->rt_nr_running) { |
1065 | add_nr_running(rq, rt_rq->rt_nr_running); | |
1066 | rt_rq->rt_queued = 1; | |
1067 | } | |
8f111bc3 PZ |
1068 | |
1069 | /* Kick cpufreq (see the comment in kernel/sched/sched.h). */ | |
1070 | cpufreq_update_util(rq, 0); | |
f4ebcbc0 KT |
1071 | } |
1072 | ||
398a153b | 1073 | #if defined CONFIG_SMP |
e864c499 | 1074 | |
398a153b GH |
1075 | static void |
1076 | inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) | |
63489e45 | 1077 | { |
4d984277 | 1078 | struct rq *rq = rq_of_rt_rq(rt_rq); |
1f11eb6a | 1079 | |
757dfcaa KT |
1080 | #ifdef CONFIG_RT_GROUP_SCHED |
1081 | /* | |
1082 | * Change rq's cpupri only if rt_rq is the top queue. | |
1083 | */ | |
1084 | if (&rq->rt != rt_rq) | |
1085 | return; | |
1086 | #endif | |
5181f4a4 SR |
1087 | if (rq->online && prio < prev_prio) |
1088 | cpupri_set(&rq->rd->cpupri, rq->cpu, prio); | |
398a153b | 1089 | } |
73fe6aae | 1090 | |
398a153b GH |
1091 | static void |
1092 | dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) | |
1093 | { | |
1094 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
d0b27fa7 | 1095 | |
757dfcaa KT |
1096 | #ifdef CONFIG_RT_GROUP_SCHED |
1097 | /* | |
1098 | * Change rq's cpupri only if rt_rq is the top queue. | |
1099 | */ | |
1100 | if (&rq->rt != rt_rq) | |
1101 | return; | |
1102 | #endif | |
398a153b GH |
1103 | if (rq->online && rt_rq->highest_prio.curr != prev_prio) |
1104 | cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr); | |
63489e45 SR |
1105 | } |
1106 | ||
398a153b GH |
1107 | #else /* CONFIG_SMP */ |
1108 | ||
6f505b16 | 1109 | static inline |
398a153b GH |
1110 | void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} |
1111 | static inline | |
1112 | void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} | |
1113 | ||
1114 | #endif /* CONFIG_SMP */ | |
6e0534f2 | 1115 | |
052f1dc7 | 1116 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
398a153b GH |
1117 | static void |
1118 | inc_rt_prio(struct rt_rq *rt_rq, int prio) | |
1119 | { | |
1120 | int prev_prio = rt_rq->highest_prio.curr; | |
1121 | ||
1122 | if (prio < prev_prio) | |
1123 | rt_rq->highest_prio.curr = prio; | |
1124 | ||
1125 | inc_rt_prio_smp(rt_rq, prio, prev_prio); | |
1126 | } | |
1127 | ||
1128 | static void | |
1129 | dec_rt_prio(struct rt_rq *rt_rq, int prio) | |
1130 | { | |
1131 | int prev_prio = rt_rq->highest_prio.curr; | |
1132 | ||
6f505b16 | 1133 | if (rt_rq->rt_nr_running) { |
764a9d6f | 1134 | |
398a153b | 1135 | WARN_ON(prio < prev_prio); |
764a9d6f | 1136 | |
e864c499 | 1137 | /* |
398a153b GH |
1138 | * This may have been our highest task, and therefore |
1139 | * we may have some recomputation to do | |
e864c499 | 1140 | */ |
398a153b | 1141 | if (prio == prev_prio) { |
e864c499 GH |
1142 | struct rt_prio_array *array = &rt_rq->active; |
1143 | ||
1144 | rt_rq->highest_prio.curr = | |
764a9d6f | 1145 | sched_find_first_bit(array->bitmap); |
e864c499 GH |
1146 | } |
1147 | ||
764a9d6f | 1148 | } else |
e864c499 | 1149 | rt_rq->highest_prio.curr = MAX_RT_PRIO; |
73fe6aae | 1150 | |
398a153b GH |
1151 | dec_rt_prio_smp(rt_rq, prio, prev_prio); |
1152 | } | |
1f11eb6a | 1153 | |
398a153b GH |
1154 | #else |
1155 | ||
1156 | static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {} | |
1157 | static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {} | |
1158 | ||
1159 | #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */ | |
6e0534f2 | 1160 | |
052f1dc7 | 1161 | #ifdef CONFIG_RT_GROUP_SCHED |
398a153b GH |
1162 | |
1163 | static void | |
1164 | inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1165 | { | |
1166 | if (rt_se_boosted(rt_se)) | |
1167 | rt_rq->rt_nr_boosted++; | |
1168 | ||
1169 | if (rt_rq->tg) | |
1170 | start_rt_bandwidth(&rt_rq->tg->rt_bandwidth); | |
1171 | } | |
1172 | ||
1173 | static void | |
1174 | dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1175 | { | |
23b0fdfc PZ |
1176 | if (rt_se_boosted(rt_se)) |
1177 | rt_rq->rt_nr_boosted--; | |
1178 | ||
1179 | WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted); | |
398a153b GH |
1180 | } |
1181 | ||
1182 | #else /* CONFIG_RT_GROUP_SCHED */ | |
1183 | ||
1184 | static void | |
1185 | inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1186 | { | |
1187 | start_rt_bandwidth(&def_rt_bandwidth); | |
1188 | } | |
1189 | ||
1190 | static inline | |
1191 | void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {} | |
1192 | ||
1193 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
1194 | ||
22abdef3 KT |
1195 | static inline |
1196 | unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se) | |
1197 | { | |
1198 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
1199 | ||
1200 | if (group_rq) | |
1201 | return group_rq->rt_nr_running; | |
1202 | else | |
1203 | return 1; | |
1204 | } | |
1205 | ||
01d36d0a FW |
1206 | static inline |
1207 | unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se) | |
1208 | { | |
1209 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
1210 | struct task_struct *tsk; | |
1211 | ||
1212 | if (group_rq) | |
1213 | return group_rq->rr_nr_running; | |
1214 | ||
1215 | tsk = rt_task_of(rt_se); | |
1216 | ||
1217 | return (tsk->policy == SCHED_RR) ? 1 : 0; | |
1218 | } | |
1219 | ||
398a153b GH |
1220 | static inline |
1221 | void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1222 | { | |
1223 | int prio = rt_se_prio(rt_se); | |
1224 | ||
1225 | WARN_ON(!rt_prio(prio)); | |
22abdef3 | 1226 | rt_rq->rt_nr_running += rt_se_nr_running(rt_se); |
01d36d0a | 1227 | rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se); |
398a153b GH |
1228 | |
1229 | inc_rt_prio(rt_rq, prio); | |
1230 | inc_rt_migration(rt_se, rt_rq); | |
1231 | inc_rt_group(rt_se, rt_rq); | |
1232 | } | |
1233 | ||
1234 | static inline | |
1235 | void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1236 | { | |
1237 | WARN_ON(!rt_prio(rt_se_prio(rt_se))); | |
1238 | WARN_ON(!rt_rq->rt_nr_running); | |
22abdef3 | 1239 | rt_rq->rt_nr_running -= rt_se_nr_running(rt_se); |
01d36d0a | 1240 | rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se); |
398a153b GH |
1241 | |
1242 | dec_rt_prio(rt_rq, rt_se_prio(rt_se)); | |
1243 | dec_rt_migration(rt_se, rt_rq); | |
1244 | dec_rt_group(rt_se, rt_rq); | |
63489e45 SR |
1245 | } |
1246 | ||
ff77e468 PZ |
1247 | /* |
1248 | * Change rt_se->run_list location unless SAVE && !MOVE | |
1249 | * | |
1250 | * assumes ENQUEUE/DEQUEUE flags match | |
1251 | */ | |
1252 | static inline bool move_entity(unsigned int flags) | |
1253 | { | |
1254 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE) | |
1255 | return false; | |
1256 | ||
1257 | return true; | |
1258 | } | |
1259 | ||
1260 | static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array) | |
1261 | { | |
1262 | list_del_init(&rt_se->run_list); | |
1263 | ||
1264 | if (list_empty(array->queue + rt_se_prio(rt_se))) | |
1265 | __clear_bit(rt_se_prio(rt_se), array->bitmap); | |
1266 | ||
1267 | rt_se->on_list = 0; | |
1268 | } | |
1269 | ||
1270 | static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) | |
bb44e5d1 | 1271 | { |
6f505b16 PZ |
1272 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
1273 | struct rt_prio_array *array = &rt_rq->active; | |
1274 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
20b6331b | 1275 | struct list_head *queue = array->queue + rt_se_prio(rt_se); |
bb44e5d1 | 1276 | |
ad2a3f13 PZ |
1277 | /* |
1278 | * Don't enqueue the group if its throttled, or when empty. | |
1279 | * The latter is a consequence of the former when a child group | |
1280 | * get throttled and the current group doesn't have any other | |
1281 | * active members. | |
1282 | */ | |
ff77e468 PZ |
1283 | if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) { |
1284 | if (rt_se->on_list) | |
1285 | __delist_rt_entity(rt_se, array); | |
6f505b16 | 1286 | return; |
ff77e468 | 1287 | } |
63489e45 | 1288 | |
ff77e468 PZ |
1289 | if (move_entity(flags)) { |
1290 | WARN_ON_ONCE(rt_se->on_list); | |
1291 | if (flags & ENQUEUE_HEAD) | |
1292 | list_add(&rt_se->run_list, queue); | |
1293 | else | |
1294 | list_add_tail(&rt_se->run_list, queue); | |
1295 | ||
1296 | __set_bit(rt_se_prio(rt_se), array->bitmap); | |
1297 | rt_se->on_list = 1; | |
1298 | } | |
1299 | rt_se->on_rq = 1; | |
78f2c7db | 1300 | |
6f505b16 PZ |
1301 | inc_rt_tasks(rt_se, rt_rq); |
1302 | } | |
1303 | ||
ff77e468 | 1304 | static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) |
6f505b16 PZ |
1305 | { |
1306 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); | |
1307 | struct rt_prio_array *array = &rt_rq->active; | |
1308 | ||
ff77e468 PZ |
1309 | if (move_entity(flags)) { |
1310 | WARN_ON_ONCE(!rt_se->on_list); | |
1311 | __delist_rt_entity(rt_se, array); | |
1312 | } | |
1313 | rt_se->on_rq = 0; | |
6f505b16 PZ |
1314 | |
1315 | dec_rt_tasks(rt_se, rt_rq); | |
1316 | } | |
1317 | ||
1318 | /* | |
1319 | * Because the prio of an upper entry depends on the lower | |
1320 | * entries, we must remove entries top - down. | |
6f505b16 | 1321 | */ |
ff77e468 | 1322 | static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags) |
6f505b16 | 1323 | { |
ad2a3f13 | 1324 | struct sched_rt_entity *back = NULL; |
6f505b16 | 1325 | |
58d6c2d7 PZ |
1326 | for_each_sched_rt_entity(rt_se) { |
1327 | rt_se->back = back; | |
1328 | back = rt_se; | |
1329 | } | |
1330 | ||
f4ebcbc0 KT |
1331 | dequeue_top_rt_rq(rt_rq_of_se(back)); |
1332 | ||
58d6c2d7 PZ |
1333 | for (rt_se = back; rt_se; rt_se = rt_se->back) { |
1334 | if (on_rt_rq(rt_se)) | |
ff77e468 | 1335 | __dequeue_rt_entity(rt_se, flags); |
ad2a3f13 PZ |
1336 | } |
1337 | } | |
1338 | ||
ff77e468 | 1339 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) |
ad2a3f13 | 1340 | { |
f4ebcbc0 KT |
1341 | struct rq *rq = rq_of_rt_se(rt_se); |
1342 | ||
ff77e468 | 1343 | dequeue_rt_stack(rt_se, flags); |
ad2a3f13 | 1344 | for_each_sched_rt_entity(rt_se) |
ff77e468 | 1345 | __enqueue_rt_entity(rt_se, flags); |
f4ebcbc0 | 1346 | enqueue_top_rt_rq(&rq->rt); |
ad2a3f13 PZ |
1347 | } |
1348 | ||
ff77e468 | 1349 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) |
ad2a3f13 | 1350 | { |
f4ebcbc0 KT |
1351 | struct rq *rq = rq_of_rt_se(rt_se); |
1352 | ||
ff77e468 | 1353 | dequeue_rt_stack(rt_se, flags); |
ad2a3f13 PZ |
1354 | |
1355 | for_each_sched_rt_entity(rt_se) { | |
1356 | struct rt_rq *rt_rq = group_rt_rq(rt_se); | |
1357 | ||
1358 | if (rt_rq && rt_rq->rt_nr_running) | |
ff77e468 | 1359 | __enqueue_rt_entity(rt_se, flags); |
58d6c2d7 | 1360 | } |
f4ebcbc0 | 1361 | enqueue_top_rt_rq(&rq->rt); |
bb44e5d1 IM |
1362 | } |
1363 | ||
1364 | /* | |
1365 | * Adding/removing a task to/from a priority array: | |
1366 | */ | |
ea87bb78 | 1367 | static void |
371fd7e7 | 1368 | enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags) |
6f505b16 PZ |
1369 | { |
1370 | struct sched_rt_entity *rt_se = &p->rt; | |
1371 | ||
371fd7e7 | 1372 | if (flags & ENQUEUE_WAKEUP) |
6f505b16 PZ |
1373 | rt_se->timeout = 0; |
1374 | ||
ff77e468 | 1375 | enqueue_rt_entity(rt_se, flags); |
c09595f6 | 1376 | |
4b53a341 | 1377 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) |
917b627d | 1378 | enqueue_pushable_task(rq, p); |
6f505b16 PZ |
1379 | } |
1380 | ||
371fd7e7 | 1381 | static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags) |
bb44e5d1 | 1382 | { |
6f505b16 | 1383 | struct sched_rt_entity *rt_se = &p->rt; |
bb44e5d1 | 1384 | |
f1e14ef6 | 1385 | update_curr_rt(rq); |
ff77e468 | 1386 | dequeue_rt_entity(rt_se, flags); |
c09595f6 | 1387 | |
917b627d | 1388 | dequeue_pushable_task(rq, p); |
bb44e5d1 IM |
1389 | } |
1390 | ||
1391 | /* | |
60686317 RW |
1392 | * Put task to the head or the end of the run list without the overhead of |
1393 | * dequeue followed by enqueue. | |
bb44e5d1 | 1394 | */ |
7ebefa8c DA |
1395 | static void |
1396 | requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head) | |
6f505b16 | 1397 | { |
1cdad715 | 1398 | if (on_rt_rq(rt_se)) { |
7ebefa8c DA |
1399 | struct rt_prio_array *array = &rt_rq->active; |
1400 | struct list_head *queue = array->queue + rt_se_prio(rt_se); | |
1401 | ||
1402 | if (head) | |
1403 | list_move(&rt_se->run_list, queue); | |
1404 | else | |
1405 | list_move_tail(&rt_se->run_list, queue); | |
1cdad715 | 1406 | } |
6f505b16 PZ |
1407 | } |
1408 | ||
7ebefa8c | 1409 | static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head) |
bb44e5d1 | 1410 | { |
6f505b16 PZ |
1411 | struct sched_rt_entity *rt_se = &p->rt; |
1412 | struct rt_rq *rt_rq; | |
bb44e5d1 | 1413 | |
6f505b16 PZ |
1414 | for_each_sched_rt_entity(rt_se) { |
1415 | rt_rq = rt_rq_of_se(rt_se); | |
7ebefa8c | 1416 | requeue_rt_entity(rt_rq, rt_se, head); |
6f505b16 | 1417 | } |
bb44e5d1 IM |
1418 | } |
1419 | ||
6f505b16 | 1420 | static void yield_task_rt(struct rq *rq) |
bb44e5d1 | 1421 | { |
7ebefa8c | 1422 | requeue_task_rt(rq, rq->curr, 0); |
bb44e5d1 IM |
1423 | } |
1424 | ||
e7693a36 | 1425 | #ifdef CONFIG_SMP |
318e0893 GH |
1426 | static int find_lowest_rq(struct task_struct *task); |
1427 | ||
0017d735 | 1428 | static int |
ac66f547 | 1429 | select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags) |
e7693a36 | 1430 | { |
7608dec2 PZ |
1431 | struct task_struct *curr; |
1432 | struct rq *rq; | |
804d402f | 1433 | bool test; |
c37495fd SR |
1434 | |
1435 | /* For anything but wake ups, just return the task_cpu */ | |
1436 | if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK) | |
1437 | goto out; | |
1438 | ||
7608dec2 PZ |
1439 | rq = cpu_rq(cpu); |
1440 | ||
1441 | rcu_read_lock(); | |
316c1608 | 1442 | curr = READ_ONCE(rq->curr); /* unlocked access */ |
7608dec2 | 1443 | |
318e0893 | 1444 | /* |
7608dec2 | 1445 | * If the current task on @p's runqueue is an RT task, then |
e1f47d89 SR |
1446 | * try to see if we can wake this RT task up on another |
1447 | * runqueue. Otherwise simply start this RT task | |
1448 | * on its current runqueue. | |
1449 | * | |
43fa5460 SR |
1450 | * We want to avoid overloading runqueues. If the woken |
1451 | * task is a higher priority, then it will stay on this CPU | |
1452 | * and the lower prio task should be moved to another CPU. | |
1453 | * Even though this will probably make the lower prio task | |
1454 | * lose its cache, we do not want to bounce a higher task | |
1455 | * around just because it gave up its CPU, perhaps for a | |
1456 | * lock? | |
1457 | * | |
1458 | * For equal prio tasks, we just let the scheduler sort it out. | |
7608dec2 PZ |
1459 | * |
1460 | * Otherwise, just let it ride on the affined RQ and the | |
1461 | * post-schedule router will push the preempted task away | |
1462 | * | |
1463 | * This test is optimistic, if we get it wrong the load-balancer | |
1464 | * will have to sort it out. | |
804d402f QY |
1465 | * |
1466 | * We take into account the capacity of the CPU to ensure it fits the | |
1467 | * requirement of the task - which is only important on heterogeneous | |
1468 | * systems like big.LITTLE. | |
318e0893 | 1469 | */ |
804d402f QY |
1470 | test = curr && |
1471 | unlikely(rt_task(curr)) && | |
1472 | (curr->nr_cpus_allowed < 2 || curr->prio <= p->prio); | |
1473 | ||
1474 | if (test || !rt_task_fits_capacity(p, cpu)) { | |
7608dec2 | 1475 | int target = find_lowest_rq(p); |
318e0893 | 1476 | |
b28bc1e0 QY |
1477 | /* |
1478 | * Bail out if we were forcing a migration to find a better | |
1479 | * fitting CPU but our search failed. | |
1480 | */ | |
1481 | if (!test && target != -1 && !rt_task_fits_capacity(p, target)) | |
1482 | goto out_unlock; | |
1483 | ||
80e3d87b TC |
1484 | /* |
1485 | * Don't bother moving it if the destination CPU is | |
1486 | * not running a lower priority task. | |
1487 | */ | |
1488 | if (target != -1 && | |
1489 | p->prio < cpu_rq(target)->rt.highest_prio.curr) | |
7608dec2 | 1490 | cpu = target; |
318e0893 | 1491 | } |
b28bc1e0 QY |
1492 | |
1493 | out_unlock: | |
7608dec2 | 1494 | rcu_read_unlock(); |
318e0893 | 1495 | |
c37495fd | 1496 | out: |
7608dec2 | 1497 | return cpu; |
e7693a36 | 1498 | } |
7ebefa8c DA |
1499 | |
1500 | static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p) | |
1501 | { | |
308a623a WL |
1502 | /* |
1503 | * Current can't be migrated, useless to reschedule, | |
1504 | * let's hope p can move out. | |
1505 | */ | |
4b53a341 | 1506 | if (rq->curr->nr_cpus_allowed == 1 || |
a1bd02e1 | 1507 | !cpupri_find(&rq->rd->cpupri, rq->curr, NULL)) |
7ebefa8c DA |
1508 | return; |
1509 | ||
308a623a WL |
1510 | /* |
1511 | * p is migratable, so let's not schedule it and | |
1512 | * see if it is pushed or pulled somewhere else. | |
1513 | */ | |
804d402f | 1514 | if (p->nr_cpus_allowed != 1 && |
a1bd02e1 | 1515 | cpupri_find(&rq->rd->cpupri, p, NULL)) |
13b8bd0a | 1516 | return; |
24600ce8 | 1517 | |
7ebefa8c | 1518 | /* |
97fb7a0a IM |
1519 | * There appear to be other CPUs that can accept |
1520 | * the current task but none can run 'p', so lets reschedule | |
1521 | * to try and push the current task away: | |
7ebefa8c DA |
1522 | */ |
1523 | requeue_task_rt(rq, p, 1); | |
8875125e | 1524 | resched_curr(rq); |
7ebefa8c DA |
1525 | } |
1526 | ||
6e2df058 PZ |
1527 | static int balance_rt(struct rq *rq, struct task_struct *p, struct rq_flags *rf) |
1528 | { | |
1529 | if (!on_rt_rq(&p->rt) && need_pull_rt_task(rq, p)) { | |
1530 | /* | |
1531 | * This is OK, because current is on_cpu, which avoids it being | |
1532 | * picked for load-balance and preemption/IRQs are still | |
1533 | * disabled avoiding further scheduler activity on it and we've | |
1534 | * not yet started the picking loop. | |
1535 | */ | |
1536 | rq_unpin_lock(rq, rf); | |
1537 | pull_rt_task(rq); | |
1538 | rq_repin_lock(rq, rf); | |
1539 | } | |
1540 | ||
1541 | return sched_stop_runnable(rq) || sched_dl_runnable(rq) || sched_rt_runnable(rq); | |
1542 | } | |
e7693a36 GH |
1543 | #endif /* CONFIG_SMP */ |
1544 | ||
bb44e5d1 IM |
1545 | /* |
1546 | * Preempt the current task with a newly woken task if needed: | |
1547 | */ | |
7d478721 | 1548 | static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags) |
bb44e5d1 | 1549 | { |
45c01e82 | 1550 | if (p->prio < rq->curr->prio) { |
8875125e | 1551 | resched_curr(rq); |
45c01e82 GH |
1552 | return; |
1553 | } | |
1554 | ||
1555 | #ifdef CONFIG_SMP | |
1556 | /* | |
1557 | * If: | |
1558 | * | |
1559 | * - the newly woken task is of equal priority to the current task | |
1560 | * - the newly woken task is non-migratable while current is migratable | |
1561 | * - current will be preempted on the next reschedule | |
1562 | * | |
1563 | * we should check to see if current can readily move to a different | |
1564 | * cpu. If so, we will reschedule to allow the push logic to try | |
1565 | * to move current somewhere else, making room for our non-migratable | |
1566 | * task. | |
1567 | */ | |
8dd0de8b | 1568 | if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr)) |
7ebefa8c | 1569 | check_preempt_equal_prio(rq, p); |
45c01e82 | 1570 | #endif |
bb44e5d1 IM |
1571 | } |
1572 | ||
a0e813f2 | 1573 | static inline void set_next_task_rt(struct rq *rq, struct task_struct *p, bool first) |
ff1cdc94 MS |
1574 | { |
1575 | p->se.exec_start = rq_clock_task(rq); | |
1576 | ||
1577 | /* The running task is never eligible for pushing */ | |
1578 | dequeue_pushable_task(rq, p); | |
f95d4eae | 1579 | |
a0e813f2 PZ |
1580 | if (!first) |
1581 | return; | |
1582 | ||
f95d4eae PZ |
1583 | /* |
1584 | * If prev task was rt, put_prev_task() has already updated the | |
1585 | * utilization. We only care of the case where we start to schedule a | |
1586 | * rt task | |
1587 | */ | |
1588 | if (rq->curr->sched_class != &rt_sched_class) | |
1589 | update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0); | |
1590 | ||
1591 | rt_queue_push_tasks(rq); | |
ff1cdc94 MS |
1592 | } |
1593 | ||
6f505b16 PZ |
1594 | static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq, |
1595 | struct rt_rq *rt_rq) | |
bb44e5d1 | 1596 | { |
6f505b16 PZ |
1597 | struct rt_prio_array *array = &rt_rq->active; |
1598 | struct sched_rt_entity *next = NULL; | |
bb44e5d1 IM |
1599 | struct list_head *queue; |
1600 | int idx; | |
1601 | ||
1602 | idx = sched_find_first_bit(array->bitmap); | |
6f505b16 | 1603 | BUG_ON(idx >= MAX_RT_PRIO); |
bb44e5d1 IM |
1604 | |
1605 | queue = array->queue + idx; | |
6f505b16 | 1606 | next = list_entry(queue->next, struct sched_rt_entity, run_list); |
326587b8 | 1607 | |
6f505b16 PZ |
1608 | return next; |
1609 | } | |
bb44e5d1 | 1610 | |
917b627d | 1611 | static struct task_struct *_pick_next_task_rt(struct rq *rq) |
6f505b16 PZ |
1612 | { |
1613 | struct sched_rt_entity *rt_se; | |
606dba2e | 1614 | struct rt_rq *rt_rq = &rq->rt; |
6f505b16 PZ |
1615 | |
1616 | do { | |
1617 | rt_se = pick_next_rt_entity(rq, rt_rq); | |
326587b8 | 1618 | BUG_ON(!rt_se); |
6f505b16 PZ |
1619 | rt_rq = group_rt_rq(rt_se); |
1620 | } while (rt_rq); | |
1621 | ||
ff1cdc94 | 1622 | return rt_task_of(rt_se); |
917b627d GH |
1623 | } |
1624 | ||
98c2f700 | 1625 | static struct task_struct *pick_next_task_rt(struct rq *rq) |
917b627d | 1626 | { |
606dba2e | 1627 | struct task_struct *p; |
606dba2e | 1628 | |
6e2df058 | 1629 | if (!sched_rt_runnable(rq)) |
606dba2e PZ |
1630 | return NULL; |
1631 | ||
606dba2e | 1632 | p = _pick_next_task_rt(rq); |
a0e813f2 | 1633 | set_next_task_rt(rq, p, true); |
6f505b16 | 1634 | return p; |
bb44e5d1 IM |
1635 | } |
1636 | ||
6e2df058 | 1637 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p) |
bb44e5d1 | 1638 | { |
f1e14ef6 | 1639 | update_curr_rt(rq); |
917b627d | 1640 | |
23127296 | 1641 | update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1); |
371bf427 | 1642 | |
917b627d GH |
1643 | /* |
1644 | * The previous task needs to be made eligible for pushing | |
1645 | * if it is still active | |
1646 | */ | |
4b53a341 | 1647 | if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1) |
917b627d | 1648 | enqueue_pushable_task(rq, p); |
bb44e5d1 IM |
1649 | } |
1650 | ||
681f3e68 | 1651 | #ifdef CONFIG_SMP |
6f505b16 | 1652 | |
e8fa1362 SR |
1653 | /* Only try algorithms three times */ |
1654 | #define RT_MAX_TRIES 3 | |
1655 | ||
f65eda4f SR |
1656 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) |
1657 | { | |
1658 | if (!task_running(rq, p) && | |
98ca645f | 1659 | cpumask_test_cpu(cpu, p->cpus_ptr)) |
f65eda4f | 1660 | return 1; |
97fb7a0a | 1661 | |
f65eda4f SR |
1662 | return 0; |
1663 | } | |
1664 | ||
e23ee747 KT |
1665 | /* |
1666 | * Return the highest pushable rq's task, which is suitable to be executed | |
97fb7a0a | 1667 | * on the CPU, NULL otherwise |
e23ee747 KT |
1668 | */ |
1669 | static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu) | |
e8fa1362 | 1670 | { |
e23ee747 KT |
1671 | struct plist_head *head = &rq->rt.pushable_tasks; |
1672 | struct task_struct *p; | |
3d07467b | 1673 | |
e23ee747 KT |
1674 | if (!has_pushable_tasks(rq)) |
1675 | return NULL; | |
3d07467b | 1676 | |
e23ee747 KT |
1677 | plist_for_each_entry(p, head, pushable_tasks) { |
1678 | if (pick_rt_task(rq, p, cpu)) | |
1679 | return p; | |
f65eda4f SR |
1680 | } |
1681 | ||
e23ee747 | 1682 | return NULL; |
e8fa1362 SR |
1683 | } |
1684 | ||
0e3900e6 | 1685 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask); |
e8fa1362 | 1686 | |
6e1254d2 GH |
1687 | static int find_lowest_rq(struct task_struct *task) |
1688 | { | |
1689 | struct sched_domain *sd; | |
4ba29684 | 1690 | struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask); |
6e1254d2 GH |
1691 | int this_cpu = smp_processor_id(); |
1692 | int cpu = task_cpu(task); | |
a1bd02e1 | 1693 | int ret; |
06f90dbd | 1694 | |
0da938c4 SR |
1695 | /* Make sure the mask is initialized first */ |
1696 | if (unlikely(!lowest_mask)) | |
1697 | return -1; | |
1698 | ||
4b53a341 | 1699 | if (task->nr_cpus_allowed == 1) |
6e0534f2 | 1700 | return -1; /* No other targets possible */ |
6e1254d2 | 1701 | |
a1bd02e1 QY |
1702 | /* |
1703 | * If we're on asym system ensure we consider the different capacities | |
1704 | * of the CPUs when searching for the lowest_mask. | |
1705 | */ | |
1706 | if (static_branch_unlikely(&sched_asym_cpucapacity)) { | |
1707 | ||
1708 | ret = cpupri_find_fitness(&task_rq(task)->rd->cpupri, | |
1709 | task, lowest_mask, | |
1710 | rt_task_fits_capacity); | |
1711 | } else { | |
1712 | ||
1713 | ret = cpupri_find(&task_rq(task)->rd->cpupri, | |
1714 | task, lowest_mask); | |
1715 | } | |
1716 | ||
1717 | if (!ret) | |
6e0534f2 | 1718 | return -1; /* No targets found */ |
6e1254d2 GH |
1719 | |
1720 | /* | |
97fb7a0a | 1721 | * At this point we have built a mask of CPUs representing the |
6e1254d2 GH |
1722 | * lowest priority tasks in the system. Now we want to elect |
1723 | * the best one based on our affinity and topology. | |
1724 | * | |
97fb7a0a | 1725 | * We prioritize the last CPU that the task executed on since |
6e1254d2 GH |
1726 | * it is most likely cache-hot in that location. |
1727 | */ | |
96f874e2 | 1728 | if (cpumask_test_cpu(cpu, lowest_mask)) |
6e1254d2 GH |
1729 | return cpu; |
1730 | ||
1731 | /* | |
1732 | * Otherwise, we consult the sched_domains span maps to figure | |
97fb7a0a | 1733 | * out which CPU is logically closest to our hot cache data. |
6e1254d2 | 1734 | */ |
e2c88063 RR |
1735 | if (!cpumask_test_cpu(this_cpu, lowest_mask)) |
1736 | this_cpu = -1; /* Skip this_cpu opt if not among lowest */ | |
6e1254d2 | 1737 | |
cd4ae6ad | 1738 | rcu_read_lock(); |
e2c88063 RR |
1739 | for_each_domain(cpu, sd) { |
1740 | if (sd->flags & SD_WAKE_AFFINE) { | |
1741 | int best_cpu; | |
6e1254d2 | 1742 | |
e2c88063 RR |
1743 | /* |
1744 | * "this_cpu" is cheaper to preempt than a | |
1745 | * remote processor. | |
1746 | */ | |
1747 | if (this_cpu != -1 && | |
cd4ae6ad XF |
1748 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { |
1749 | rcu_read_unlock(); | |
e2c88063 | 1750 | return this_cpu; |
cd4ae6ad | 1751 | } |
e2c88063 RR |
1752 | |
1753 | best_cpu = cpumask_first_and(lowest_mask, | |
1754 | sched_domain_span(sd)); | |
cd4ae6ad XF |
1755 | if (best_cpu < nr_cpu_ids) { |
1756 | rcu_read_unlock(); | |
e2c88063 | 1757 | return best_cpu; |
cd4ae6ad | 1758 | } |
6e1254d2 GH |
1759 | } |
1760 | } | |
cd4ae6ad | 1761 | rcu_read_unlock(); |
6e1254d2 GH |
1762 | |
1763 | /* | |
1764 | * And finally, if there were no matches within the domains | |
1765 | * just give the caller *something* to work with from the compatible | |
1766 | * locations. | |
1767 | */ | |
e2c88063 RR |
1768 | if (this_cpu != -1) |
1769 | return this_cpu; | |
1770 | ||
1771 | cpu = cpumask_any(lowest_mask); | |
1772 | if (cpu < nr_cpu_ids) | |
1773 | return cpu; | |
97fb7a0a | 1774 | |
e2c88063 | 1775 | return -1; |
07b4032c GH |
1776 | } |
1777 | ||
1778 | /* Will lock the rq it finds */ | |
4df64c0b | 1779 | static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) |
07b4032c GH |
1780 | { |
1781 | struct rq *lowest_rq = NULL; | |
07b4032c | 1782 | int tries; |
4df64c0b | 1783 | int cpu; |
e8fa1362 | 1784 | |
07b4032c GH |
1785 | for (tries = 0; tries < RT_MAX_TRIES; tries++) { |
1786 | cpu = find_lowest_rq(task); | |
1787 | ||
2de0b463 | 1788 | if ((cpu == -1) || (cpu == rq->cpu)) |
e8fa1362 SR |
1789 | break; |
1790 | ||
07b4032c GH |
1791 | lowest_rq = cpu_rq(cpu); |
1792 | ||
80e3d87b TC |
1793 | if (lowest_rq->rt.highest_prio.curr <= task->prio) { |
1794 | /* | |
1795 | * Target rq has tasks of equal or higher priority, | |
1796 | * retrying does not release any lock and is unlikely | |
1797 | * to yield a different result. | |
1798 | */ | |
1799 | lowest_rq = NULL; | |
1800 | break; | |
1801 | } | |
1802 | ||
e8fa1362 | 1803 | /* if the prio of this runqueue changed, try again */ |
07b4032c | 1804 | if (double_lock_balance(rq, lowest_rq)) { |
e8fa1362 SR |
1805 | /* |
1806 | * We had to unlock the run queue. In | |
1807 | * the mean time, task could have | |
1808 | * migrated already or had its affinity changed. | |
1809 | * Also make sure that it wasn't scheduled on its rq. | |
1810 | */ | |
07b4032c | 1811 | if (unlikely(task_rq(task) != rq || |
3bd37062 | 1812 | !cpumask_test_cpu(lowest_rq->cpu, task->cpus_ptr) || |
07b4032c | 1813 | task_running(rq, task) || |
13b5ab02 | 1814 | !rt_task(task) || |
da0c1e65 | 1815 | !task_on_rq_queued(task))) { |
4df64c0b | 1816 | |
7f1b4393 | 1817 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 SR |
1818 | lowest_rq = NULL; |
1819 | break; | |
1820 | } | |
1821 | } | |
1822 | ||
1823 | /* If this rq is still suitable use it. */ | |
e864c499 | 1824 | if (lowest_rq->rt.highest_prio.curr > task->prio) |
e8fa1362 SR |
1825 | break; |
1826 | ||
1827 | /* try again */ | |
1b12bbc7 | 1828 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 SR |
1829 | lowest_rq = NULL; |
1830 | } | |
1831 | ||
1832 | return lowest_rq; | |
1833 | } | |
1834 | ||
917b627d GH |
1835 | static struct task_struct *pick_next_pushable_task(struct rq *rq) |
1836 | { | |
1837 | struct task_struct *p; | |
1838 | ||
1839 | if (!has_pushable_tasks(rq)) | |
1840 | return NULL; | |
1841 | ||
1842 | p = plist_first_entry(&rq->rt.pushable_tasks, | |
1843 | struct task_struct, pushable_tasks); | |
1844 | ||
1845 | BUG_ON(rq->cpu != task_cpu(p)); | |
1846 | BUG_ON(task_current(rq, p)); | |
4b53a341 | 1847 | BUG_ON(p->nr_cpus_allowed <= 1); |
917b627d | 1848 | |
da0c1e65 | 1849 | BUG_ON(!task_on_rq_queued(p)); |
917b627d GH |
1850 | BUG_ON(!rt_task(p)); |
1851 | ||
1852 | return p; | |
1853 | } | |
1854 | ||
e8fa1362 SR |
1855 | /* |
1856 | * If the current CPU has more than one RT task, see if the non | |
1857 | * running task can migrate over to a CPU that is running a task | |
1858 | * of lesser priority. | |
1859 | */ | |
697f0a48 | 1860 | static int push_rt_task(struct rq *rq) |
e8fa1362 SR |
1861 | { |
1862 | struct task_struct *next_task; | |
1863 | struct rq *lowest_rq; | |
311e800e | 1864 | int ret = 0; |
e8fa1362 | 1865 | |
a22d7fc1 GH |
1866 | if (!rq->rt.overloaded) |
1867 | return 0; | |
1868 | ||
917b627d | 1869 | next_task = pick_next_pushable_task(rq); |
e8fa1362 SR |
1870 | if (!next_task) |
1871 | return 0; | |
1872 | ||
49246274 | 1873 | retry: |
9ebc6053 | 1874 | if (WARN_ON(next_task == rq->curr)) |
e8fa1362 SR |
1875 | return 0; |
1876 | ||
1877 | /* | |
1878 | * It's possible that the next_task slipped in of | |
1879 | * higher priority than current. If that's the case | |
1880 | * just reschedule current. | |
1881 | */ | |
697f0a48 | 1882 | if (unlikely(next_task->prio < rq->curr->prio)) { |
8875125e | 1883 | resched_curr(rq); |
e8fa1362 SR |
1884 | return 0; |
1885 | } | |
1886 | ||
697f0a48 | 1887 | /* We might release rq lock */ |
e8fa1362 SR |
1888 | get_task_struct(next_task); |
1889 | ||
1890 | /* find_lock_lowest_rq locks the rq if found */ | |
697f0a48 | 1891 | lowest_rq = find_lock_lowest_rq(next_task, rq); |
e8fa1362 SR |
1892 | if (!lowest_rq) { |
1893 | struct task_struct *task; | |
1894 | /* | |
311e800e | 1895 | * find_lock_lowest_rq releases rq->lock |
1563513d GH |
1896 | * so it is possible that next_task has migrated. |
1897 | * | |
1898 | * We need to make sure that the task is still on the same | |
1899 | * run-queue and is also still the next task eligible for | |
1900 | * pushing. | |
e8fa1362 | 1901 | */ |
917b627d | 1902 | task = pick_next_pushable_task(rq); |
de16b91e | 1903 | if (task == next_task) { |
1563513d | 1904 | /* |
311e800e HD |
1905 | * The task hasn't migrated, and is still the next |
1906 | * eligible task, but we failed to find a run-queue | |
1907 | * to push it to. Do not retry in this case, since | |
97fb7a0a | 1908 | * other CPUs will pull from us when ready. |
1563513d | 1909 | */ |
1563513d | 1910 | goto out; |
e8fa1362 | 1911 | } |
917b627d | 1912 | |
1563513d GH |
1913 | if (!task) |
1914 | /* No more tasks, just exit */ | |
1915 | goto out; | |
1916 | ||
917b627d | 1917 | /* |
1563513d | 1918 | * Something has shifted, try again. |
917b627d | 1919 | */ |
1563513d GH |
1920 | put_task_struct(next_task); |
1921 | next_task = task; | |
1922 | goto retry; | |
e8fa1362 SR |
1923 | } |
1924 | ||
697f0a48 | 1925 | deactivate_task(rq, next_task, 0); |
e8fa1362 SR |
1926 | set_task_cpu(next_task, lowest_rq->cpu); |
1927 | activate_task(lowest_rq, next_task, 0); | |
311e800e | 1928 | ret = 1; |
e8fa1362 | 1929 | |
8875125e | 1930 | resched_curr(lowest_rq); |
e8fa1362 | 1931 | |
1b12bbc7 | 1932 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 | 1933 | |
e8fa1362 SR |
1934 | out: |
1935 | put_task_struct(next_task); | |
1936 | ||
311e800e | 1937 | return ret; |
e8fa1362 SR |
1938 | } |
1939 | ||
e8fa1362 SR |
1940 | static void push_rt_tasks(struct rq *rq) |
1941 | { | |
1942 | /* push_rt_task will return true if it moved an RT */ | |
1943 | while (push_rt_task(rq)) | |
1944 | ; | |
1945 | } | |
1946 | ||
b6366f04 | 1947 | #ifdef HAVE_RT_PUSH_IPI |
4bdced5c | 1948 | |
b6366f04 | 1949 | /* |
4bdced5c SRRH |
1950 | * When a high priority task schedules out from a CPU and a lower priority |
1951 | * task is scheduled in, a check is made to see if there's any RT tasks | |
1952 | * on other CPUs that are waiting to run because a higher priority RT task | |
1953 | * is currently running on its CPU. In this case, the CPU with multiple RT | |
1954 | * tasks queued on it (overloaded) needs to be notified that a CPU has opened | |
1955 | * up that may be able to run one of its non-running queued RT tasks. | |
1956 | * | |
1957 | * All CPUs with overloaded RT tasks need to be notified as there is currently | |
1958 | * no way to know which of these CPUs have the highest priority task waiting | |
1959 | * to run. Instead of trying to take a spinlock on each of these CPUs, | |
1960 | * which has shown to cause large latency when done on machines with many | |
1961 | * CPUs, sending an IPI to the CPUs to have them push off the overloaded | |
1962 | * RT tasks waiting to run. | |
1963 | * | |
1964 | * Just sending an IPI to each of the CPUs is also an issue, as on large | |
1965 | * count CPU machines, this can cause an IPI storm on a CPU, especially | |
1966 | * if its the only CPU with multiple RT tasks queued, and a large number | |
1967 | * of CPUs scheduling a lower priority task at the same time. | |
1968 | * | |
1969 | * Each root domain has its own irq work function that can iterate over | |
1970 | * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT | |
1971 | * tassk must be checked if there's one or many CPUs that are lowering | |
1972 | * their priority, there's a single irq work iterator that will try to | |
1973 | * push off RT tasks that are waiting to run. | |
1974 | * | |
1975 | * When a CPU schedules a lower priority task, it will kick off the | |
1976 | * irq work iterator that will jump to each CPU with overloaded RT tasks. | |
1977 | * As it only takes the first CPU that schedules a lower priority task | |
1978 | * to start the process, the rto_start variable is incremented and if | |
1979 | * the atomic result is one, then that CPU will try to take the rto_lock. | |
1980 | * This prevents high contention on the lock as the process handles all | |
1981 | * CPUs scheduling lower priority tasks. | |
1982 | * | |
1983 | * All CPUs that are scheduling a lower priority task will increment the | |
1984 | * rt_loop_next variable. This will make sure that the irq work iterator | |
1985 | * checks all RT overloaded CPUs whenever a CPU schedules a new lower | |
1986 | * priority task, even if the iterator is in the middle of a scan. Incrementing | |
1987 | * the rt_loop_next will cause the iterator to perform another scan. | |
b6366f04 | 1988 | * |
b6366f04 | 1989 | */ |
ad0f1d9d | 1990 | static int rto_next_cpu(struct root_domain *rd) |
b6366f04 | 1991 | { |
4bdced5c | 1992 | int next; |
b6366f04 SR |
1993 | int cpu; |
1994 | ||
b6366f04 | 1995 | /* |
4bdced5c SRRH |
1996 | * When starting the IPI RT pushing, the rto_cpu is set to -1, |
1997 | * rt_next_cpu() will simply return the first CPU found in | |
1998 | * the rto_mask. | |
1999 | * | |
97fb7a0a | 2000 | * If rto_next_cpu() is called with rto_cpu is a valid CPU, it |
4bdced5c SRRH |
2001 | * will return the next CPU found in the rto_mask. |
2002 | * | |
2003 | * If there are no more CPUs left in the rto_mask, then a check is made | |
2004 | * against rto_loop and rto_loop_next. rto_loop is only updated with | |
2005 | * the rto_lock held, but any CPU may increment the rto_loop_next | |
2006 | * without any locking. | |
b6366f04 | 2007 | */ |
4bdced5c | 2008 | for (;;) { |
b6366f04 | 2009 | |
4bdced5c SRRH |
2010 | /* When rto_cpu is -1 this acts like cpumask_first() */ |
2011 | cpu = cpumask_next(rd->rto_cpu, rd->rto_mask); | |
b6366f04 | 2012 | |
4bdced5c | 2013 | rd->rto_cpu = cpu; |
b6366f04 | 2014 | |
4bdced5c SRRH |
2015 | if (cpu < nr_cpu_ids) |
2016 | return cpu; | |
b6366f04 | 2017 | |
4bdced5c SRRH |
2018 | rd->rto_cpu = -1; |
2019 | ||
2020 | /* | |
2021 | * ACQUIRE ensures we see the @rto_mask changes | |
2022 | * made prior to the @next value observed. | |
2023 | * | |
2024 | * Matches WMB in rt_set_overload(). | |
2025 | */ | |
2026 | next = atomic_read_acquire(&rd->rto_loop_next); | |
b6366f04 | 2027 | |
4bdced5c | 2028 | if (rd->rto_loop == next) |
b6366f04 | 2029 | break; |
4bdced5c SRRH |
2030 | |
2031 | rd->rto_loop = next; | |
b6366f04 SR |
2032 | } |
2033 | ||
4bdced5c | 2034 | return -1; |
b6366f04 SR |
2035 | } |
2036 | ||
4bdced5c SRRH |
2037 | static inline bool rto_start_trylock(atomic_t *v) |
2038 | { | |
2039 | return !atomic_cmpxchg_acquire(v, 0, 1); | |
2040 | } | |
b6366f04 | 2041 | |
4bdced5c | 2042 | static inline void rto_start_unlock(atomic_t *v) |
b6366f04 | 2043 | { |
4bdced5c SRRH |
2044 | atomic_set_release(v, 0); |
2045 | } | |
b6366f04 | 2046 | |
4bdced5c SRRH |
2047 | static void tell_cpu_to_push(struct rq *rq) |
2048 | { | |
2049 | int cpu = -1; | |
b6366f04 | 2050 | |
4bdced5c SRRH |
2051 | /* Keep the loop going if the IPI is currently active */ |
2052 | atomic_inc(&rq->rd->rto_loop_next); | |
b6366f04 | 2053 | |
4bdced5c SRRH |
2054 | /* Only one CPU can initiate a loop at a time */ |
2055 | if (!rto_start_trylock(&rq->rd->rto_loop_start)) | |
b6366f04 SR |
2056 | return; |
2057 | ||
4bdced5c | 2058 | raw_spin_lock(&rq->rd->rto_lock); |
b6366f04 | 2059 | |
4bdced5c | 2060 | /* |
97fb7a0a | 2061 | * The rto_cpu is updated under the lock, if it has a valid CPU |
4bdced5c SRRH |
2062 | * then the IPI is still running and will continue due to the |
2063 | * update to loop_next, and nothing needs to be done here. | |
2064 | * Otherwise it is finishing up and an ipi needs to be sent. | |
2065 | */ | |
2066 | if (rq->rd->rto_cpu < 0) | |
ad0f1d9d | 2067 | cpu = rto_next_cpu(rq->rd); |
4bdced5c SRRH |
2068 | |
2069 | raw_spin_unlock(&rq->rd->rto_lock); | |
2070 | ||
2071 | rto_start_unlock(&rq->rd->rto_loop_start); | |
2072 | ||
364f5665 SRV |
2073 | if (cpu >= 0) { |
2074 | /* Make sure the rd does not get freed while pushing */ | |
2075 | sched_get_rd(rq->rd); | |
4bdced5c | 2076 | irq_work_queue_on(&rq->rd->rto_push_work, cpu); |
364f5665 | 2077 | } |
b6366f04 SR |
2078 | } |
2079 | ||
2080 | /* Called from hardirq context */ | |
4bdced5c | 2081 | void rto_push_irq_work_func(struct irq_work *work) |
b6366f04 | 2082 | { |
ad0f1d9d SRV |
2083 | struct root_domain *rd = |
2084 | container_of(work, struct root_domain, rto_push_work); | |
4bdced5c | 2085 | struct rq *rq; |
b6366f04 SR |
2086 | int cpu; |
2087 | ||
4bdced5c | 2088 | rq = this_rq(); |
b6366f04 | 2089 | |
4bdced5c SRRH |
2090 | /* |
2091 | * We do not need to grab the lock to check for has_pushable_tasks. | |
2092 | * When it gets updated, a check is made if a push is possible. | |
2093 | */ | |
b6366f04 SR |
2094 | if (has_pushable_tasks(rq)) { |
2095 | raw_spin_lock(&rq->lock); | |
4bdced5c | 2096 | push_rt_tasks(rq); |
b6366f04 SR |
2097 | raw_spin_unlock(&rq->lock); |
2098 | } | |
2099 | ||
ad0f1d9d | 2100 | raw_spin_lock(&rd->rto_lock); |
b6366f04 | 2101 | |
4bdced5c | 2102 | /* Pass the IPI to the next rt overloaded queue */ |
ad0f1d9d | 2103 | cpu = rto_next_cpu(rd); |
b6366f04 | 2104 | |
ad0f1d9d | 2105 | raw_spin_unlock(&rd->rto_lock); |
b6366f04 | 2106 | |
364f5665 SRV |
2107 | if (cpu < 0) { |
2108 | sched_put_rd(rd); | |
b6366f04 | 2109 | return; |
364f5665 | 2110 | } |
b6366f04 | 2111 | |
b6366f04 | 2112 | /* Try the next RT overloaded CPU */ |
ad0f1d9d | 2113 | irq_work_queue_on(&rd->rto_push_work, cpu); |
b6366f04 SR |
2114 | } |
2115 | #endif /* HAVE_RT_PUSH_IPI */ | |
2116 | ||
8046d680 | 2117 | static void pull_rt_task(struct rq *this_rq) |
f65eda4f | 2118 | { |
8046d680 PZ |
2119 | int this_cpu = this_rq->cpu, cpu; |
2120 | bool resched = false; | |
a8728944 | 2121 | struct task_struct *p; |
f65eda4f | 2122 | struct rq *src_rq; |
f73c52a5 | 2123 | int rt_overload_count = rt_overloaded(this_rq); |
f65eda4f | 2124 | |
f73c52a5 | 2125 | if (likely(!rt_overload_count)) |
8046d680 | 2126 | return; |
f65eda4f | 2127 | |
7c3f2ab7 PZ |
2128 | /* |
2129 | * Match the barrier from rt_set_overloaded; this guarantees that if we | |
2130 | * see overloaded we must also see the rto_mask bit. | |
2131 | */ | |
2132 | smp_rmb(); | |
2133 | ||
f73c52a5 SR |
2134 | /* If we are the only overloaded CPU do nothing */ |
2135 | if (rt_overload_count == 1 && | |
2136 | cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask)) | |
2137 | return; | |
2138 | ||
b6366f04 SR |
2139 | #ifdef HAVE_RT_PUSH_IPI |
2140 | if (sched_feat(RT_PUSH_IPI)) { | |
2141 | tell_cpu_to_push(this_rq); | |
8046d680 | 2142 | return; |
b6366f04 SR |
2143 | } |
2144 | #endif | |
2145 | ||
c6c4927b | 2146 | for_each_cpu(cpu, this_rq->rd->rto_mask) { |
f65eda4f SR |
2147 | if (this_cpu == cpu) |
2148 | continue; | |
2149 | ||
2150 | src_rq = cpu_rq(cpu); | |
74ab8e4f GH |
2151 | |
2152 | /* | |
2153 | * Don't bother taking the src_rq->lock if the next highest | |
2154 | * task is known to be lower-priority than our current task. | |
2155 | * This may look racy, but if this value is about to go | |
2156 | * logically higher, the src_rq will push this task away. | |
2157 | * And if its going logically lower, we do not care | |
2158 | */ | |
2159 | if (src_rq->rt.highest_prio.next >= | |
2160 | this_rq->rt.highest_prio.curr) | |
2161 | continue; | |
2162 | ||
f65eda4f SR |
2163 | /* |
2164 | * We can potentially drop this_rq's lock in | |
2165 | * double_lock_balance, and another CPU could | |
a8728944 | 2166 | * alter this_rq |
f65eda4f | 2167 | */ |
a8728944 | 2168 | double_lock_balance(this_rq, src_rq); |
f65eda4f SR |
2169 | |
2170 | /* | |
e23ee747 KT |
2171 | * We can pull only a task, which is pushable |
2172 | * on its rq, and no others. | |
f65eda4f | 2173 | */ |
e23ee747 | 2174 | p = pick_highest_pushable_task(src_rq, this_cpu); |
f65eda4f SR |
2175 | |
2176 | /* | |
2177 | * Do we have an RT task that preempts | |
2178 | * the to-be-scheduled task? | |
2179 | */ | |
a8728944 | 2180 | if (p && (p->prio < this_rq->rt.highest_prio.curr)) { |
f65eda4f | 2181 | WARN_ON(p == src_rq->curr); |
da0c1e65 | 2182 | WARN_ON(!task_on_rq_queued(p)); |
f65eda4f SR |
2183 | |
2184 | /* | |
2185 | * There's a chance that p is higher in priority | |
97fb7a0a | 2186 | * than what's currently running on its CPU. |
f65eda4f SR |
2187 | * This is just that p is wakeing up and hasn't |
2188 | * had a chance to schedule. We only pull | |
2189 | * p if it is lower in priority than the | |
a8728944 | 2190 | * current task on the run queue |
f65eda4f | 2191 | */ |
a8728944 | 2192 | if (p->prio < src_rq->curr->prio) |
614ee1f6 | 2193 | goto skip; |
f65eda4f | 2194 | |
8046d680 | 2195 | resched = true; |
f65eda4f SR |
2196 | |
2197 | deactivate_task(src_rq, p, 0); | |
2198 | set_task_cpu(p, this_cpu); | |
2199 | activate_task(this_rq, p, 0); | |
2200 | /* | |
2201 | * We continue with the search, just in | |
2202 | * case there's an even higher prio task | |
25985edc | 2203 | * in another runqueue. (low likelihood |
f65eda4f | 2204 | * but possible) |
f65eda4f | 2205 | */ |
f65eda4f | 2206 | } |
49246274 | 2207 | skip: |
1b12bbc7 | 2208 | double_unlock_balance(this_rq, src_rq); |
f65eda4f SR |
2209 | } |
2210 | ||
8046d680 PZ |
2211 | if (resched) |
2212 | resched_curr(this_rq); | |
f65eda4f SR |
2213 | } |
2214 | ||
8ae121ac GH |
2215 | /* |
2216 | * If we are not running and we are not going to reschedule soon, we should | |
2217 | * try to push tasks away now | |
2218 | */ | |
efbbd05a | 2219 | static void task_woken_rt(struct rq *rq, struct task_struct *p) |
4642dafd | 2220 | { |
804d402f QY |
2221 | bool need_to_push = !task_running(rq, p) && |
2222 | !test_tsk_need_resched(rq->curr) && | |
2223 | p->nr_cpus_allowed > 1 && | |
2224 | (dl_task(rq->curr) || rt_task(rq->curr)) && | |
2225 | (rq->curr->nr_cpus_allowed < 2 || | |
2226 | rq->curr->prio <= p->prio); | |
2227 | ||
d94a9df4 | 2228 | if (need_to_push) |
4642dafd SR |
2229 | push_rt_tasks(rq); |
2230 | } | |
2231 | ||
bdd7c81b | 2232 | /* Assumes rq->lock is held */ |
1f11eb6a | 2233 | static void rq_online_rt(struct rq *rq) |
bdd7c81b IM |
2234 | { |
2235 | if (rq->rt.overloaded) | |
2236 | rt_set_overload(rq); | |
6e0534f2 | 2237 | |
7def2be1 PZ |
2238 | __enable_runtime(rq); |
2239 | ||
e864c499 | 2240 | cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr); |
bdd7c81b IM |
2241 | } |
2242 | ||
2243 | /* Assumes rq->lock is held */ | |
1f11eb6a | 2244 | static void rq_offline_rt(struct rq *rq) |
bdd7c81b IM |
2245 | { |
2246 | if (rq->rt.overloaded) | |
2247 | rt_clear_overload(rq); | |
6e0534f2 | 2248 | |
7def2be1 PZ |
2249 | __disable_runtime(rq); |
2250 | ||
6e0534f2 | 2251 | cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID); |
bdd7c81b | 2252 | } |
cb469845 SR |
2253 | |
2254 | /* | |
2255 | * When switch from the rt queue, we bring ourselves to a position | |
2256 | * that we might want to pull RT tasks from other runqueues. | |
2257 | */ | |
da7a735e | 2258 | static void switched_from_rt(struct rq *rq, struct task_struct *p) |
cb469845 SR |
2259 | { |
2260 | /* | |
2261 | * If there are other RT tasks then we will reschedule | |
2262 | * and the scheduling of the other RT tasks will handle | |
2263 | * the balancing. But if we are the last RT task | |
2264 | * we may need to handle the pulling of RT tasks | |
2265 | * now. | |
2266 | */ | |
da0c1e65 | 2267 | if (!task_on_rq_queued(p) || rq->rt.rt_nr_running) |
1158ddb5 KT |
2268 | return; |
2269 | ||
02d8ec94 | 2270 | rt_queue_pull_task(rq); |
cb469845 | 2271 | } |
3d8cbdf8 | 2272 | |
11c785b7 | 2273 | void __init init_sched_rt_class(void) |
3d8cbdf8 RR |
2274 | { |
2275 | unsigned int i; | |
2276 | ||
029632fb | 2277 | for_each_possible_cpu(i) { |
eaa95840 | 2278 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i), |
6ca09dfc | 2279 | GFP_KERNEL, cpu_to_node(i)); |
029632fb | 2280 | } |
3d8cbdf8 | 2281 | } |
cb469845 SR |
2282 | #endif /* CONFIG_SMP */ |
2283 | ||
2284 | /* | |
2285 | * When switching a task to RT, we may overload the runqueue | |
2286 | * with RT tasks. In this case we try to push them off to | |
2287 | * other runqueues. | |
2288 | */ | |
da7a735e | 2289 | static void switched_to_rt(struct rq *rq, struct task_struct *p) |
cb469845 | 2290 | { |
cb469845 SR |
2291 | /* |
2292 | * If we are already running, then there's nothing | |
2293 | * that needs to be done. But if we are not running | |
2294 | * we may need to preempt the current running task. | |
2295 | * If that current running task is also an RT task | |
2296 | * then see if we can move to another run queue. | |
2297 | */ | |
da0c1e65 | 2298 | if (task_on_rq_queued(p) && rq->curr != p) { |
cb469845 | 2299 | #ifdef CONFIG_SMP |
d94a9df4 | 2300 | if (p->nr_cpus_allowed > 1 && rq->rt.overloaded) |
02d8ec94 | 2301 | rt_queue_push_tasks(rq); |
619bd4a7 | 2302 | #endif /* CONFIG_SMP */ |
2fe25826 | 2303 | if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq))) |
8875125e | 2304 | resched_curr(rq); |
cb469845 SR |
2305 | } |
2306 | } | |
2307 | ||
2308 | /* | |
2309 | * Priority of the task has changed. This may cause | |
2310 | * us to initiate a push or pull. | |
2311 | */ | |
da7a735e PZ |
2312 | static void |
2313 | prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 2314 | { |
da0c1e65 | 2315 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
2316 | return; |
2317 | ||
2318 | if (rq->curr == p) { | |
cb469845 SR |
2319 | #ifdef CONFIG_SMP |
2320 | /* | |
2321 | * If our priority decreases while running, we | |
2322 | * may need to pull tasks to this runqueue. | |
2323 | */ | |
2324 | if (oldprio < p->prio) | |
02d8ec94 | 2325 | rt_queue_pull_task(rq); |
fd7a4bed | 2326 | |
cb469845 SR |
2327 | /* |
2328 | * If there's a higher priority task waiting to run | |
fd7a4bed | 2329 | * then reschedule. |
cb469845 | 2330 | */ |
fd7a4bed | 2331 | if (p->prio > rq->rt.highest_prio.curr) |
8875125e | 2332 | resched_curr(rq); |
cb469845 SR |
2333 | #else |
2334 | /* For UP simply resched on drop of prio */ | |
2335 | if (oldprio < p->prio) | |
8875125e | 2336 | resched_curr(rq); |
e8fa1362 | 2337 | #endif /* CONFIG_SMP */ |
cb469845 SR |
2338 | } else { |
2339 | /* | |
2340 | * This task is not running, but if it is | |
2341 | * greater than the current running task | |
2342 | * then reschedule. | |
2343 | */ | |
2344 | if (p->prio < rq->curr->prio) | |
8875125e | 2345 | resched_curr(rq); |
cb469845 SR |
2346 | } |
2347 | } | |
2348 | ||
b18b6a9c | 2349 | #ifdef CONFIG_POSIX_TIMERS |
78f2c7db PZ |
2350 | static void watchdog(struct rq *rq, struct task_struct *p) |
2351 | { | |
2352 | unsigned long soft, hard; | |
2353 | ||
78d7d407 JS |
2354 | /* max may change after cur was read, this will be fixed next tick */ |
2355 | soft = task_rlimit(p, RLIMIT_RTTIME); | |
2356 | hard = task_rlimit_max(p, RLIMIT_RTTIME); | |
78f2c7db PZ |
2357 | |
2358 | if (soft != RLIM_INFINITY) { | |
2359 | unsigned long next; | |
2360 | ||
57d2aa00 YX |
2361 | if (p->rt.watchdog_stamp != jiffies) { |
2362 | p->rt.timeout++; | |
2363 | p->rt.watchdog_stamp = jiffies; | |
2364 | } | |
2365 | ||
78f2c7db | 2366 | next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ); |
3a245c0f TG |
2367 | if (p->rt.timeout > next) { |
2368 | posix_cputimers_rt_watchdog(&p->posix_cputimers, | |
2369 | p->se.sum_exec_runtime); | |
2370 | } | |
78f2c7db PZ |
2371 | } |
2372 | } | |
b18b6a9c NP |
2373 | #else |
2374 | static inline void watchdog(struct rq *rq, struct task_struct *p) { } | |
2375 | #endif | |
bb44e5d1 | 2376 | |
d84b3131 FW |
2377 | /* |
2378 | * scheduler tick hitting a task of our scheduling class. | |
2379 | * | |
2380 | * NOTE: This function can be called remotely by the tick offload that | |
2381 | * goes along full dynticks. Therefore no local assumption can be made | |
2382 | * and everything must be accessed through the @rq and @curr passed in | |
2383 | * parameters. | |
2384 | */ | |
8f4d37ec | 2385 | static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued) |
bb44e5d1 | 2386 | { |
454c7999 CC |
2387 | struct sched_rt_entity *rt_se = &p->rt; |
2388 | ||
67e2be02 | 2389 | update_curr_rt(rq); |
23127296 | 2390 | update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1); |
67e2be02 | 2391 | |
78f2c7db PZ |
2392 | watchdog(rq, p); |
2393 | ||
bb44e5d1 IM |
2394 | /* |
2395 | * RR tasks need a special form of timeslice management. | |
2396 | * FIFO tasks have no timeslices. | |
2397 | */ | |
2398 | if (p->policy != SCHED_RR) | |
2399 | return; | |
2400 | ||
fa717060 | 2401 | if (--p->rt.time_slice) |
bb44e5d1 IM |
2402 | return; |
2403 | ||
ce0dbbbb | 2404 | p->rt.time_slice = sched_rr_timeslice; |
bb44e5d1 | 2405 | |
98fbc798 | 2406 | /* |
e9aa39bb LB |
2407 | * Requeue to the end of queue if we (and all of our ancestors) are not |
2408 | * the only element on the queue | |
98fbc798 | 2409 | */ |
454c7999 CC |
2410 | for_each_sched_rt_entity(rt_se) { |
2411 | if (rt_se->run_list.prev != rt_se->run_list.next) { | |
2412 | requeue_task_rt(rq, p, 0); | |
8aa6f0eb | 2413 | resched_curr(rq); |
454c7999 CC |
2414 | return; |
2415 | } | |
98fbc798 | 2416 | } |
bb44e5d1 IM |
2417 | } |
2418 | ||
6d686f45 | 2419 | static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task) |
0d721cea PW |
2420 | { |
2421 | /* | |
2422 | * Time slice is 0 for SCHED_FIFO tasks | |
2423 | */ | |
2424 | if (task->policy == SCHED_RR) | |
ce0dbbbb | 2425 | return sched_rr_timeslice; |
0d721cea PW |
2426 | else |
2427 | return 0; | |
2428 | } | |
2429 | ||
029632fb | 2430 | const struct sched_class rt_sched_class = { |
5522d5d5 | 2431 | .next = &fair_sched_class, |
bb44e5d1 IM |
2432 | .enqueue_task = enqueue_task_rt, |
2433 | .dequeue_task = dequeue_task_rt, | |
2434 | .yield_task = yield_task_rt, | |
2435 | ||
2436 | .check_preempt_curr = check_preempt_curr_rt, | |
2437 | ||
2438 | .pick_next_task = pick_next_task_rt, | |
2439 | .put_prev_task = put_prev_task_rt, | |
03b7fad1 | 2440 | .set_next_task = set_next_task_rt, |
bb44e5d1 | 2441 | |
681f3e68 | 2442 | #ifdef CONFIG_SMP |
6e2df058 | 2443 | .balance = balance_rt, |
4ce72a2c | 2444 | .select_task_rq = select_task_rq_rt, |
6c37067e | 2445 | .set_cpus_allowed = set_cpus_allowed_common, |
1f11eb6a GH |
2446 | .rq_online = rq_online_rt, |
2447 | .rq_offline = rq_offline_rt, | |
efbbd05a | 2448 | .task_woken = task_woken_rt, |
cb469845 | 2449 | .switched_from = switched_from_rt, |
681f3e68 | 2450 | #endif |
bb44e5d1 IM |
2451 | |
2452 | .task_tick = task_tick_rt, | |
cb469845 | 2453 | |
0d721cea PW |
2454 | .get_rr_interval = get_rr_interval_rt, |
2455 | ||
cb469845 SR |
2456 | .prio_changed = prio_changed_rt, |
2457 | .switched_to = switched_to_rt, | |
6e998916 SG |
2458 | |
2459 | .update_curr = update_curr_rt, | |
982d9cdc PB |
2460 | |
2461 | #ifdef CONFIG_UCLAMP_TASK | |
2462 | .uclamp_enabled = 1, | |
2463 | #endif | |
bb44e5d1 | 2464 | }; |
ada18de2 | 2465 | |
8887cd99 NP |
2466 | #ifdef CONFIG_RT_GROUP_SCHED |
2467 | /* | |
2468 | * Ensure that the real time constraints are schedulable. | |
2469 | */ | |
2470 | static DEFINE_MUTEX(rt_constraints_mutex); | |
2471 | ||
8887cd99 NP |
2472 | static inline int tg_has_rt_tasks(struct task_group *tg) |
2473 | { | |
b4fb015e KK |
2474 | struct task_struct *task; |
2475 | struct css_task_iter it; | |
2476 | int ret = 0; | |
8887cd99 NP |
2477 | |
2478 | /* | |
2479 | * Autogroups do not have RT tasks; see autogroup_create(). | |
2480 | */ | |
2481 | if (task_group_is_autogroup(tg)) | |
2482 | return 0; | |
2483 | ||
b4fb015e KK |
2484 | css_task_iter_start(&tg->css, 0, &it); |
2485 | while (!ret && (task = css_task_iter_next(&it))) | |
2486 | ret |= rt_task(task); | |
2487 | css_task_iter_end(&it); | |
8887cd99 | 2488 | |
b4fb015e | 2489 | return ret; |
8887cd99 NP |
2490 | } |
2491 | ||
2492 | struct rt_schedulable_data { | |
2493 | struct task_group *tg; | |
2494 | u64 rt_period; | |
2495 | u64 rt_runtime; | |
2496 | }; | |
2497 | ||
2498 | static int tg_rt_schedulable(struct task_group *tg, void *data) | |
2499 | { | |
2500 | struct rt_schedulable_data *d = data; | |
2501 | struct task_group *child; | |
2502 | unsigned long total, sum = 0; | |
2503 | u64 period, runtime; | |
2504 | ||
2505 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
2506 | runtime = tg->rt_bandwidth.rt_runtime; | |
2507 | ||
2508 | if (tg == d->tg) { | |
2509 | period = d->rt_period; | |
2510 | runtime = d->rt_runtime; | |
2511 | } | |
2512 | ||
2513 | /* | |
2514 | * Cannot have more runtime than the period. | |
2515 | */ | |
2516 | if (runtime > period && runtime != RUNTIME_INF) | |
2517 | return -EINVAL; | |
2518 | ||
2519 | /* | |
b4fb015e | 2520 | * Ensure we don't starve existing RT tasks if runtime turns zero. |
8887cd99 | 2521 | */ |
b4fb015e KK |
2522 | if (rt_bandwidth_enabled() && !runtime && |
2523 | tg->rt_bandwidth.rt_runtime && tg_has_rt_tasks(tg)) | |
8887cd99 NP |
2524 | return -EBUSY; |
2525 | ||
2526 | total = to_ratio(period, runtime); | |
2527 | ||
2528 | /* | |
2529 | * Nobody can have more than the global setting allows. | |
2530 | */ | |
2531 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) | |
2532 | return -EINVAL; | |
2533 | ||
2534 | /* | |
2535 | * The sum of our children's runtime should not exceed our own. | |
2536 | */ | |
2537 | list_for_each_entry_rcu(child, &tg->children, siblings) { | |
2538 | period = ktime_to_ns(child->rt_bandwidth.rt_period); | |
2539 | runtime = child->rt_bandwidth.rt_runtime; | |
2540 | ||
2541 | if (child == d->tg) { | |
2542 | period = d->rt_period; | |
2543 | runtime = d->rt_runtime; | |
2544 | } | |
2545 | ||
2546 | sum += to_ratio(period, runtime); | |
2547 | } | |
2548 | ||
2549 | if (sum > total) | |
2550 | return -EINVAL; | |
2551 | ||
2552 | return 0; | |
2553 | } | |
2554 | ||
2555 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | |
2556 | { | |
2557 | int ret; | |
2558 | ||
2559 | struct rt_schedulable_data data = { | |
2560 | .tg = tg, | |
2561 | .rt_period = period, | |
2562 | .rt_runtime = runtime, | |
2563 | }; | |
2564 | ||
2565 | rcu_read_lock(); | |
2566 | ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); | |
2567 | rcu_read_unlock(); | |
2568 | ||
2569 | return ret; | |
2570 | } | |
2571 | ||
2572 | static int tg_set_rt_bandwidth(struct task_group *tg, | |
2573 | u64 rt_period, u64 rt_runtime) | |
2574 | { | |
2575 | int i, err = 0; | |
2576 | ||
2577 | /* | |
2578 | * Disallowing the root group RT runtime is BAD, it would disallow the | |
2579 | * kernel creating (and or operating) RT threads. | |
2580 | */ | |
2581 | if (tg == &root_task_group && rt_runtime == 0) | |
2582 | return -EINVAL; | |
2583 | ||
2584 | /* No period doesn't make any sense. */ | |
2585 | if (rt_period == 0) | |
2586 | return -EINVAL; | |
2587 | ||
2588 | mutex_lock(&rt_constraints_mutex); | |
8887cd99 NP |
2589 | err = __rt_schedulable(tg, rt_period, rt_runtime); |
2590 | if (err) | |
2591 | goto unlock; | |
2592 | ||
2593 | raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
2594 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); | |
2595 | tg->rt_bandwidth.rt_runtime = rt_runtime; | |
2596 | ||
2597 | for_each_possible_cpu(i) { | |
2598 | struct rt_rq *rt_rq = tg->rt_rq[i]; | |
2599 | ||
2600 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
2601 | rt_rq->rt_runtime = rt_runtime; | |
2602 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
2603 | } | |
2604 | raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
2605 | unlock: | |
8887cd99 NP |
2606 | mutex_unlock(&rt_constraints_mutex); |
2607 | ||
2608 | return err; | |
2609 | } | |
2610 | ||
2611 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) | |
2612 | { | |
2613 | u64 rt_runtime, rt_period; | |
2614 | ||
2615 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
2616 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | |
2617 | if (rt_runtime_us < 0) | |
2618 | rt_runtime = RUNTIME_INF; | |
1a010e29 KK |
2619 | else if ((u64)rt_runtime_us > U64_MAX / NSEC_PER_USEC) |
2620 | return -EINVAL; | |
8887cd99 NP |
2621 | |
2622 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); | |
2623 | } | |
2624 | ||
2625 | long sched_group_rt_runtime(struct task_group *tg) | |
2626 | { | |
2627 | u64 rt_runtime_us; | |
2628 | ||
2629 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) | |
2630 | return -1; | |
2631 | ||
2632 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; | |
2633 | do_div(rt_runtime_us, NSEC_PER_USEC); | |
2634 | return rt_runtime_us; | |
2635 | } | |
2636 | ||
2637 | int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us) | |
2638 | { | |
2639 | u64 rt_runtime, rt_period; | |
2640 | ||
1a010e29 KK |
2641 | if (rt_period_us > U64_MAX / NSEC_PER_USEC) |
2642 | return -EINVAL; | |
2643 | ||
8887cd99 NP |
2644 | rt_period = rt_period_us * NSEC_PER_USEC; |
2645 | rt_runtime = tg->rt_bandwidth.rt_runtime; | |
2646 | ||
2647 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); | |
2648 | } | |
2649 | ||
2650 | long sched_group_rt_period(struct task_group *tg) | |
2651 | { | |
2652 | u64 rt_period_us; | |
2653 | ||
2654 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
2655 | do_div(rt_period_us, NSEC_PER_USEC); | |
2656 | return rt_period_us; | |
2657 | } | |
2658 | ||
2659 | static int sched_rt_global_constraints(void) | |
2660 | { | |
2661 | int ret = 0; | |
2662 | ||
2663 | mutex_lock(&rt_constraints_mutex); | |
8887cd99 | 2664 | ret = __rt_schedulable(NULL, 0, 0); |
8887cd99 NP |
2665 | mutex_unlock(&rt_constraints_mutex); |
2666 | ||
2667 | return ret; | |
2668 | } | |
2669 | ||
2670 | int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) | |
2671 | { | |
2672 | /* Don't accept realtime tasks when there is no way for them to run */ | |
2673 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) | |
2674 | return 0; | |
2675 | ||
2676 | return 1; | |
2677 | } | |
2678 | ||
2679 | #else /* !CONFIG_RT_GROUP_SCHED */ | |
2680 | static int sched_rt_global_constraints(void) | |
2681 | { | |
2682 | unsigned long flags; | |
2683 | int i; | |
2684 | ||
2685 | raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); | |
2686 | for_each_possible_cpu(i) { | |
2687 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | |
2688 | ||
2689 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
2690 | rt_rq->rt_runtime = global_rt_runtime(); | |
2691 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
2692 | } | |
2693 | raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); | |
2694 | ||
2695 | return 0; | |
2696 | } | |
2697 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
2698 | ||
2699 | static int sched_rt_global_validate(void) | |
2700 | { | |
2701 | if (sysctl_sched_rt_period <= 0) | |
2702 | return -EINVAL; | |
2703 | ||
2704 | if ((sysctl_sched_rt_runtime != RUNTIME_INF) && | |
2705 | (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) | |
2706 | return -EINVAL; | |
2707 | ||
2708 | return 0; | |
2709 | } | |
2710 | ||
2711 | static void sched_rt_do_global(void) | |
2712 | { | |
2713 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | |
2714 | def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); | |
2715 | } | |
2716 | ||
32927393 CH |
2717 | int sched_rt_handler(struct ctl_table *table, int write, void *buffer, |
2718 | size_t *lenp, loff_t *ppos) | |
8887cd99 NP |
2719 | { |
2720 | int old_period, old_runtime; | |
2721 | static DEFINE_MUTEX(mutex); | |
2722 | int ret; | |
2723 | ||
2724 | mutex_lock(&mutex); | |
2725 | old_period = sysctl_sched_rt_period; | |
2726 | old_runtime = sysctl_sched_rt_runtime; | |
2727 | ||
2728 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
2729 | ||
2730 | if (!ret && write) { | |
2731 | ret = sched_rt_global_validate(); | |
2732 | if (ret) | |
2733 | goto undo; | |
2734 | ||
2735 | ret = sched_dl_global_validate(); | |
2736 | if (ret) | |
2737 | goto undo; | |
2738 | ||
2739 | ret = sched_rt_global_constraints(); | |
2740 | if (ret) | |
2741 | goto undo; | |
2742 | ||
2743 | sched_rt_do_global(); | |
2744 | sched_dl_do_global(); | |
2745 | } | |
2746 | if (0) { | |
2747 | undo: | |
2748 | sysctl_sched_rt_period = old_period; | |
2749 | sysctl_sched_rt_runtime = old_runtime; | |
2750 | } | |
2751 | mutex_unlock(&mutex); | |
2752 | ||
2753 | return ret; | |
2754 | } | |
2755 | ||
32927393 CH |
2756 | int sched_rr_handler(struct ctl_table *table, int write, void *buffer, |
2757 | size_t *lenp, loff_t *ppos) | |
8887cd99 NP |
2758 | { |
2759 | int ret; | |
2760 | static DEFINE_MUTEX(mutex); | |
2761 | ||
2762 | mutex_lock(&mutex); | |
2763 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
2764 | /* | |
2765 | * Make sure that internally we keep jiffies. | |
2766 | * Also, writing zero resets the timeslice to default: | |
2767 | */ | |
2768 | if (!ret && write) { | |
2769 | sched_rr_timeslice = | |
2770 | sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE : | |
2771 | msecs_to_jiffies(sysctl_sched_rr_timeslice); | |
2772 | } | |
2773 | mutex_unlock(&mutex); | |
97fb7a0a | 2774 | |
8887cd99 NP |
2775 | return ret; |
2776 | } | |
2777 | ||
ada18de2 | 2778 | #ifdef CONFIG_SCHED_DEBUG |
029632fb | 2779 | void print_rt_stats(struct seq_file *m, int cpu) |
ada18de2 | 2780 | { |
ec514c48 | 2781 | rt_rq_iter_t iter; |
ada18de2 PZ |
2782 | struct rt_rq *rt_rq; |
2783 | ||
2784 | rcu_read_lock(); | |
ec514c48 | 2785 | for_each_rt_rq(rt_rq, iter, cpu_rq(cpu)) |
ada18de2 PZ |
2786 | print_rt_rq(m, cpu, rt_rq); |
2787 | rcu_read_unlock(); | |
2788 | } | |
55e12e5e | 2789 | #endif /* CONFIG_SCHED_DEBUG */ |