]>
Commit | Line | Data |
---|---|---|
bb44e5d1 IM |
1 | /* |
2 | * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR | |
3 | * policies) | |
4 | */ | |
5 | ||
029632fb PZ |
6 | #include "sched.h" |
7 | ||
8 | #include <linux/slab.h> | |
b6366f04 | 9 | #include <linux/irq_work.h> |
029632fb | 10 | |
ce0dbbbb | 11 | int sched_rr_timeslice = RR_TIMESLICE; |
975e155e | 12 | int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE; |
ce0dbbbb | 13 | |
029632fb PZ |
14 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); |
15 | ||
16 | struct rt_bandwidth def_rt_bandwidth; | |
17 | ||
18 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) | |
19 | { | |
20 | struct rt_bandwidth *rt_b = | |
21 | container_of(timer, struct rt_bandwidth, rt_period_timer); | |
029632fb | 22 | int idle = 0; |
77a4d1a1 | 23 | int overrun; |
029632fb | 24 | |
77a4d1a1 | 25 | raw_spin_lock(&rt_b->rt_runtime_lock); |
029632fb | 26 | for (;;) { |
77a4d1a1 | 27 | overrun = hrtimer_forward_now(timer, rt_b->rt_period); |
029632fb PZ |
28 | if (!overrun) |
29 | break; | |
30 | ||
77a4d1a1 | 31 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
029632fb | 32 | idle = do_sched_rt_period_timer(rt_b, overrun); |
77a4d1a1 | 33 | raw_spin_lock(&rt_b->rt_runtime_lock); |
029632fb | 34 | } |
4cfafd30 PZ |
35 | if (idle) |
36 | rt_b->rt_period_active = 0; | |
77a4d1a1 | 37 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
029632fb PZ |
38 | |
39 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
40 | } | |
41 | ||
42 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) | |
43 | { | |
44 | rt_b->rt_period = ns_to_ktime(period); | |
45 | rt_b->rt_runtime = runtime; | |
46 | ||
47 | raw_spin_lock_init(&rt_b->rt_runtime_lock); | |
48 | ||
49 | hrtimer_init(&rt_b->rt_period_timer, | |
50 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
51 | rt_b->rt_period_timer.function = sched_rt_period_timer; | |
52 | } | |
53 | ||
54 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) | |
55 | { | |
56 | if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) | |
57 | return; | |
58 | ||
029632fb | 59 | raw_spin_lock(&rt_b->rt_runtime_lock); |
4cfafd30 PZ |
60 | if (!rt_b->rt_period_active) { |
61 | rt_b->rt_period_active = 1; | |
c3a990dc SR |
62 | /* |
63 | * SCHED_DEADLINE updates the bandwidth, as a run away | |
64 | * RT task with a DL task could hog a CPU. But DL does | |
65 | * not reset the period. If a deadline task was running | |
66 | * without an RT task running, it can cause RT tasks to | |
67 | * throttle when they start up. Kick the timer right away | |
68 | * to update the period. | |
69 | */ | |
70 | hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0)); | |
4cfafd30 PZ |
71 | hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED); |
72 | } | |
029632fb PZ |
73 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
74 | } | |
75 | ||
89b41108 | 76 | #if defined(CONFIG_SMP) && defined(HAVE_RT_PUSH_IPI) |
b6366f04 SR |
77 | static void push_irq_work_func(struct irq_work *work); |
78 | #endif | |
79 | ||
07c54f7a | 80 | void init_rt_rq(struct rt_rq *rt_rq) |
029632fb PZ |
81 | { |
82 | struct rt_prio_array *array; | |
83 | int i; | |
84 | ||
85 | array = &rt_rq->active; | |
86 | for (i = 0; i < MAX_RT_PRIO; i++) { | |
87 | INIT_LIST_HEAD(array->queue + i); | |
88 | __clear_bit(i, array->bitmap); | |
89 | } | |
90 | /* delimiter for bitsearch: */ | |
91 | __set_bit(MAX_RT_PRIO, array->bitmap); | |
92 | ||
93 | #if defined CONFIG_SMP | |
94 | rt_rq->highest_prio.curr = MAX_RT_PRIO; | |
95 | rt_rq->highest_prio.next = MAX_RT_PRIO; | |
96 | rt_rq->rt_nr_migratory = 0; | |
97 | rt_rq->overloaded = 0; | |
98 | plist_head_init(&rt_rq->pushable_tasks); | |
b6366f04 SR |
99 | |
100 | #ifdef HAVE_RT_PUSH_IPI | |
101 | rt_rq->push_flags = 0; | |
102 | rt_rq->push_cpu = nr_cpu_ids; | |
103 | raw_spin_lock_init(&rt_rq->push_lock); | |
104 | init_irq_work(&rt_rq->push_work, push_irq_work_func); | |
029632fb | 105 | #endif |
b6366f04 | 106 | #endif /* CONFIG_SMP */ |
f4ebcbc0 KT |
107 | /* We start is dequeued state, because no RT tasks are queued */ |
108 | rt_rq->rt_queued = 0; | |
029632fb PZ |
109 | |
110 | rt_rq->rt_time = 0; | |
111 | rt_rq->rt_throttled = 0; | |
112 | rt_rq->rt_runtime = 0; | |
113 | raw_spin_lock_init(&rt_rq->rt_runtime_lock); | |
114 | } | |
115 | ||
8f48894f | 116 | #ifdef CONFIG_RT_GROUP_SCHED |
029632fb PZ |
117 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) |
118 | { | |
119 | hrtimer_cancel(&rt_b->rt_period_timer); | |
120 | } | |
8f48894f PZ |
121 | |
122 | #define rt_entity_is_task(rt_se) (!(rt_se)->my_q) | |
123 | ||
398a153b GH |
124 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) |
125 | { | |
8f48894f PZ |
126 | #ifdef CONFIG_SCHED_DEBUG |
127 | WARN_ON_ONCE(!rt_entity_is_task(rt_se)); | |
128 | #endif | |
398a153b GH |
129 | return container_of(rt_se, struct task_struct, rt); |
130 | } | |
131 | ||
398a153b GH |
132 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) |
133 | { | |
134 | return rt_rq->rq; | |
135 | } | |
136 | ||
137 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) | |
138 | { | |
139 | return rt_se->rt_rq; | |
140 | } | |
141 | ||
653d07a6 KT |
142 | static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se) |
143 | { | |
144 | struct rt_rq *rt_rq = rt_se->rt_rq; | |
145 | ||
146 | return rt_rq->rq; | |
147 | } | |
148 | ||
029632fb PZ |
149 | void free_rt_sched_group(struct task_group *tg) |
150 | { | |
151 | int i; | |
152 | ||
153 | if (tg->rt_se) | |
154 | destroy_rt_bandwidth(&tg->rt_bandwidth); | |
155 | ||
156 | for_each_possible_cpu(i) { | |
157 | if (tg->rt_rq) | |
158 | kfree(tg->rt_rq[i]); | |
159 | if (tg->rt_se) | |
160 | kfree(tg->rt_se[i]); | |
161 | } | |
162 | ||
163 | kfree(tg->rt_rq); | |
164 | kfree(tg->rt_se); | |
165 | } | |
166 | ||
167 | void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, | |
168 | struct sched_rt_entity *rt_se, int cpu, | |
169 | struct sched_rt_entity *parent) | |
170 | { | |
171 | struct rq *rq = cpu_rq(cpu); | |
172 | ||
173 | rt_rq->highest_prio.curr = MAX_RT_PRIO; | |
174 | rt_rq->rt_nr_boosted = 0; | |
175 | rt_rq->rq = rq; | |
176 | rt_rq->tg = tg; | |
177 | ||
178 | tg->rt_rq[cpu] = rt_rq; | |
179 | tg->rt_se[cpu] = rt_se; | |
180 | ||
181 | if (!rt_se) | |
182 | return; | |
183 | ||
184 | if (!parent) | |
185 | rt_se->rt_rq = &rq->rt; | |
186 | else | |
187 | rt_se->rt_rq = parent->my_q; | |
188 | ||
189 | rt_se->my_q = rt_rq; | |
190 | rt_se->parent = parent; | |
191 | INIT_LIST_HEAD(&rt_se->run_list); | |
192 | } | |
193 | ||
194 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
195 | { | |
196 | struct rt_rq *rt_rq; | |
197 | struct sched_rt_entity *rt_se; | |
198 | int i; | |
199 | ||
200 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); | |
201 | if (!tg->rt_rq) | |
202 | goto err; | |
203 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); | |
204 | if (!tg->rt_se) | |
205 | goto err; | |
206 | ||
207 | init_rt_bandwidth(&tg->rt_bandwidth, | |
208 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); | |
209 | ||
210 | for_each_possible_cpu(i) { | |
211 | rt_rq = kzalloc_node(sizeof(struct rt_rq), | |
212 | GFP_KERNEL, cpu_to_node(i)); | |
213 | if (!rt_rq) | |
214 | goto err; | |
215 | ||
216 | rt_se = kzalloc_node(sizeof(struct sched_rt_entity), | |
217 | GFP_KERNEL, cpu_to_node(i)); | |
218 | if (!rt_se) | |
219 | goto err_free_rq; | |
220 | ||
07c54f7a | 221 | init_rt_rq(rt_rq); |
029632fb PZ |
222 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; |
223 | init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]); | |
224 | } | |
225 | ||
226 | return 1; | |
227 | ||
228 | err_free_rq: | |
229 | kfree(rt_rq); | |
230 | err: | |
231 | return 0; | |
232 | } | |
233 | ||
398a153b GH |
234 | #else /* CONFIG_RT_GROUP_SCHED */ |
235 | ||
a1ba4d8b PZ |
236 | #define rt_entity_is_task(rt_se) (1) |
237 | ||
8f48894f PZ |
238 | static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) |
239 | { | |
240 | return container_of(rt_se, struct task_struct, rt); | |
241 | } | |
242 | ||
398a153b GH |
243 | static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) |
244 | { | |
245 | return container_of(rt_rq, struct rq, rt); | |
246 | } | |
247 | ||
653d07a6 | 248 | static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se) |
398a153b GH |
249 | { |
250 | struct task_struct *p = rt_task_of(rt_se); | |
653d07a6 KT |
251 | |
252 | return task_rq(p); | |
253 | } | |
254 | ||
255 | static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) | |
256 | { | |
257 | struct rq *rq = rq_of_rt_se(rt_se); | |
398a153b GH |
258 | |
259 | return &rq->rt; | |
260 | } | |
261 | ||
029632fb PZ |
262 | void free_rt_sched_group(struct task_group *tg) { } |
263 | ||
264 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
265 | { | |
266 | return 1; | |
267 | } | |
398a153b GH |
268 | #endif /* CONFIG_RT_GROUP_SCHED */ |
269 | ||
4fd29176 | 270 | #ifdef CONFIG_SMP |
84de4274 | 271 | |
8046d680 | 272 | static void pull_rt_task(struct rq *this_rq); |
38033c37 | 273 | |
dc877341 PZ |
274 | static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev) |
275 | { | |
276 | /* Try to pull RT tasks here if we lower this rq's prio */ | |
277 | return rq->rt.highest_prio.curr > prev->prio; | |
278 | } | |
279 | ||
637f5085 | 280 | static inline int rt_overloaded(struct rq *rq) |
4fd29176 | 281 | { |
637f5085 | 282 | return atomic_read(&rq->rd->rto_count); |
4fd29176 | 283 | } |
84de4274 | 284 | |
4fd29176 SR |
285 | static inline void rt_set_overload(struct rq *rq) |
286 | { | |
1f11eb6a GH |
287 | if (!rq->online) |
288 | return; | |
289 | ||
c6c4927b | 290 | cpumask_set_cpu(rq->cpu, rq->rd->rto_mask); |
4fd29176 SR |
291 | /* |
292 | * Make sure the mask is visible before we set | |
293 | * the overload count. That is checked to determine | |
294 | * if we should look at the mask. It would be a shame | |
295 | * if we looked at the mask, but the mask was not | |
296 | * updated yet. | |
7c3f2ab7 PZ |
297 | * |
298 | * Matched by the barrier in pull_rt_task(). | |
4fd29176 | 299 | */ |
7c3f2ab7 | 300 | smp_wmb(); |
637f5085 | 301 | atomic_inc(&rq->rd->rto_count); |
4fd29176 | 302 | } |
84de4274 | 303 | |
4fd29176 SR |
304 | static inline void rt_clear_overload(struct rq *rq) |
305 | { | |
1f11eb6a GH |
306 | if (!rq->online) |
307 | return; | |
308 | ||
4fd29176 | 309 | /* the order here really doesn't matter */ |
637f5085 | 310 | atomic_dec(&rq->rd->rto_count); |
c6c4927b | 311 | cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask); |
4fd29176 | 312 | } |
73fe6aae | 313 | |
398a153b | 314 | static void update_rt_migration(struct rt_rq *rt_rq) |
73fe6aae | 315 | { |
a1ba4d8b | 316 | if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) { |
398a153b GH |
317 | if (!rt_rq->overloaded) { |
318 | rt_set_overload(rq_of_rt_rq(rt_rq)); | |
319 | rt_rq->overloaded = 1; | |
cdc8eb98 | 320 | } |
398a153b GH |
321 | } else if (rt_rq->overloaded) { |
322 | rt_clear_overload(rq_of_rt_rq(rt_rq)); | |
323 | rt_rq->overloaded = 0; | |
637f5085 | 324 | } |
73fe6aae | 325 | } |
4fd29176 | 326 | |
398a153b GH |
327 | static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
328 | { | |
29baa747 PZ |
329 | struct task_struct *p; |
330 | ||
a1ba4d8b PZ |
331 | if (!rt_entity_is_task(rt_se)) |
332 | return; | |
333 | ||
29baa747 | 334 | p = rt_task_of(rt_se); |
a1ba4d8b PZ |
335 | rt_rq = &rq_of_rt_rq(rt_rq)->rt; |
336 | ||
337 | rt_rq->rt_nr_total++; | |
4b53a341 | 338 | if (p->nr_cpus_allowed > 1) |
398a153b GH |
339 | rt_rq->rt_nr_migratory++; |
340 | ||
341 | update_rt_migration(rt_rq); | |
342 | } | |
343 | ||
344 | static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
345 | { | |
29baa747 PZ |
346 | struct task_struct *p; |
347 | ||
a1ba4d8b PZ |
348 | if (!rt_entity_is_task(rt_se)) |
349 | return; | |
350 | ||
29baa747 | 351 | p = rt_task_of(rt_se); |
a1ba4d8b PZ |
352 | rt_rq = &rq_of_rt_rq(rt_rq)->rt; |
353 | ||
354 | rt_rq->rt_nr_total--; | |
4b53a341 | 355 | if (p->nr_cpus_allowed > 1) |
398a153b GH |
356 | rt_rq->rt_nr_migratory--; |
357 | ||
358 | update_rt_migration(rt_rq); | |
359 | } | |
360 | ||
5181f4a4 SR |
361 | static inline int has_pushable_tasks(struct rq *rq) |
362 | { | |
363 | return !plist_head_empty(&rq->rt.pushable_tasks); | |
364 | } | |
365 | ||
fd7a4bed PZ |
366 | static DEFINE_PER_CPU(struct callback_head, rt_push_head); |
367 | static DEFINE_PER_CPU(struct callback_head, rt_pull_head); | |
e3fca9e7 PZ |
368 | |
369 | static void push_rt_tasks(struct rq *); | |
fd7a4bed | 370 | static void pull_rt_task(struct rq *); |
e3fca9e7 PZ |
371 | |
372 | static inline void queue_push_tasks(struct rq *rq) | |
dc877341 | 373 | { |
e3fca9e7 PZ |
374 | if (!has_pushable_tasks(rq)) |
375 | return; | |
376 | ||
fd7a4bed PZ |
377 | queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks); |
378 | } | |
379 | ||
380 | static inline void queue_pull_task(struct rq *rq) | |
381 | { | |
382 | queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task); | |
dc877341 PZ |
383 | } |
384 | ||
917b627d GH |
385 | static void enqueue_pushable_task(struct rq *rq, struct task_struct *p) |
386 | { | |
387 | plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); | |
388 | plist_node_init(&p->pushable_tasks, p->prio); | |
389 | plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks); | |
5181f4a4 SR |
390 | |
391 | /* Update the highest prio pushable task */ | |
392 | if (p->prio < rq->rt.highest_prio.next) | |
393 | rq->rt.highest_prio.next = p->prio; | |
917b627d GH |
394 | } |
395 | ||
396 | static void dequeue_pushable_task(struct rq *rq, struct task_struct *p) | |
397 | { | |
398 | plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); | |
917b627d | 399 | |
5181f4a4 SR |
400 | /* Update the new highest prio pushable task */ |
401 | if (has_pushable_tasks(rq)) { | |
402 | p = plist_first_entry(&rq->rt.pushable_tasks, | |
403 | struct task_struct, pushable_tasks); | |
404 | rq->rt.highest_prio.next = p->prio; | |
405 | } else | |
406 | rq->rt.highest_prio.next = MAX_RT_PRIO; | |
bcf08df3 IM |
407 | } |
408 | ||
917b627d GH |
409 | #else |
410 | ||
ceacc2c1 | 411 | static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p) |
fa85ae24 | 412 | { |
6f505b16 PZ |
413 | } |
414 | ||
ceacc2c1 PZ |
415 | static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p) |
416 | { | |
417 | } | |
418 | ||
b07430ac | 419 | static inline |
ceacc2c1 PZ |
420 | void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
421 | { | |
422 | } | |
423 | ||
398a153b | 424 | static inline |
ceacc2c1 PZ |
425 | void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) |
426 | { | |
427 | } | |
917b627d | 428 | |
dc877341 PZ |
429 | static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev) |
430 | { | |
431 | return false; | |
432 | } | |
433 | ||
8046d680 | 434 | static inline void pull_rt_task(struct rq *this_rq) |
dc877341 | 435 | { |
dc877341 PZ |
436 | } |
437 | ||
e3fca9e7 | 438 | static inline void queue_push_tasks(struct rq *rq) |
dc877341 PZ |
439 | { |
440 | } | |
4fd29176 SR |
441 | #endif /* CONFIG_SMP */ |
442 | ||
f4ebcbc0 KT |
443 | static void enqueue_top_rt_rq(struct rt_rq *rt_rq); |
444 | static void dequeue_top_rt_rq(struct rt_rq *rt_rq); | |
445 | ||
6f505b16 PZ |
446 | static inline int on_rt_rq(struct sched_rt_entity *rt_se) |
447 | { | |
ff77e468 | 448 | return rt_se->on_rq; |
6f505b16 PZ |
449 | } |
450 | ||
052f1dc7 | 451 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 452 | |
9f0c1e56 | 453 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) |
6f505b16 PZ |
454 | { |
455 | if (!rt_rq->tg) | |
9f0c1e56 | 456 | return RUNTIME_INF; |
6f505b16 | 457 | |
ac086bc2 PZ |
458 | return rt_rq->rt_runtime; |
459 | } | |
460 | ||
461 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) | |
462 | { | |
463 | return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period); | |
6f505b16 PZ |
464 | } |
465 | ||
ec514c48 CX |
466 | typedef struct task_group *rt_rq_iter_t; |
467 | ||
1c09ab0d YZ |
468 | static inline struct task_group *next_task_group(struct task_group *tg) |
469 | { | |
470 | do { | |
471 | tg = list_entry_rcu(tg->list.next, | |
472 | typeof(struct task_group), list); | |
473 | } while (&tg->list != &task_groups && task_group_is_autogroup(tg)); | |
474 | ||
475 | if (&tg->list == &task_groups) | |
476 | tg = NULL; | |
477 | ||
478 | return tg; | |
479 | } | |
480 | ||
481 | #define for_each_rt_rq(rt_rq, iter, rq) \ | |
482 | for (iter = container_of(&task_groups, typeof(*iter), list); \ | |
483 | (iter = next_task_group(iter)) && \ | |
484 | (rt_rq = iter->rt_rq[cpu_of(rq)]);) | |
ec514c48 | 485 | |
6f505b16 PZ |
486 | #define for_each_sched_rt_entity(rt_se) \ |
487 | for (; rt_se; rt_se = rt_se->parent) | |
488 | ||
489 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) | |
490 | { | |
491 | return rt_se->my_q; | |
492 | } | |
493 | ||
ff77e468 PZ |
494 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags); |
495 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags); | |
6f505b16 | 496 | |
9f0c1e56 | 497 | static void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
6f505b16 | 498 | { |
f6121f4f | 499 | struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr; |
8875125e | 500 | struct rq *rq = rq_of_rt_rq(rt_rq); |
74b7eb58 YZ |
501 | struct sched_rt_entity *rt_se; |
502 | ||
8875125e | 503 | int cpu = cpu_of(rq); |
0c3b9168 BS |
504 | |
505 | rt_se = rt_rq->tg->rt_se[cpu]; | |
6f505b16 | 506 | |
f6121f4f | 507 | if (rt_rq->rt_nr_running) { |
f4ebcbc0 KT |
508 | if (!rt_se) |
509 | enqueue_top_rt_rq(rt_rq); | |
510 | else if (!on_rt_rq(rt_se)) | |
ff77e468 | 511 | enqueue_rt_entity(rt_se, 0); |
f4ebcbc0 | 512 | |
e864c499 | 513 | if (rt_rq->highest_prio.curr < curr->prio) |
8875125e | 514 | resched_curr(rq); |
6f505b16 PZ |
515 | } |
516 | } | |
517 | ||
9f0c1e56 | 518 | static void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
6f505b16 | 519 | { |
74b7eb58 | 520 | struct sched_rt_entity *rt_se; |
0c3b9168 | 521 | int cpu = cpu_of(rq_of_rt_rq(rt_rq)); |
74b7eb58 | 522 | |
0c3b9168 | 523 | rt_se = rt_rq->tg->rt_se[cpu]; |
6f505b16 | 524 | |
f4ebcbc0 KT |
525 | if (!rt_se) |
526 | dequeue_top_rt_rq(rt_rq); | |
527 | else if (on_rt_rq(rt_se)) | |
ff77e468 | 528 | dequeue_rt_entity(rt_se, 0); |
6f505b16 PZ |
529 | } |
530 | ||
46383648 KT |
531 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
532 | { | |
533 | return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted; | |
534 | } | |
535 | ||
23b0fdfc PZ |
536 | static int rt_se_boosted(struct sched_rt_entity *rt_se) |
537 | { | |
538 | struct rt_rq *rt_rq = group_rt_rq(rt_se); | |
539 | struct task_struct *p; | |
540 | ||
541 | if (rt_rq) | |
542 | return !!rt_rq->rt_nr_boosted; | |
543 | ||
544 | p = rt_task_of(rt_se); | |
545 | return p->prio != p->normal_prio; | |
546 | } | |
547 | ||
d0b27fa7 | 548 | #ifdef CONFIG_SMP |
c6c4927b | 549 | static inline const struct cpumask *sched_rt_period_mask(void) |
d0b27fa7 | 550 | { |
424c93fe | 551 | return this_rq()->rd->span; |
d0b27fa7 | 552 | } |
6f505b16 | 553 | #else |
c6c4927b | 554 | static inline const struct cpumask *sched_rt_period_mask(void) |
d0b27fa7 | 555 | { |
c6c4927b | 556 | return cpu_online_mask; |
d0b27fa7 PZ |
557 | } |
558 | #endif | |
6f505b16 | 559 | |
d0b27fa7 PZ |
560 | static inline |
561 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) | |
6f505b16 | 562 | { |
d0b27fa7 PZ |
563 | return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu]; |
564 | } | |
9f0c1e56 | 565 | |
ac086bc2 PZ |
566 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
567 | { | |
568 | return &rt_rq->tg->rt_bandwidth; | |
569 | } | |
570 | ||
55e12e5e | 571 | #else /* !CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 PZ |
572 | |
573 | static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) | |
574 | { | |
ac086bc2 PZ |
575 | return rt_rq->rt_runtime; |
576 | } | |
577 | ||
578 | static inline u64 sched_rt_period(struct rt_rq *rt_rq) | |
579 | { | |
580 | return ktime_to_ns(def_rt_bandwidth.rt_period); | |
6f505b16 PZ |
581 | } |
582 | ||
ec514c48 CX |
583 | typedef struct rt_rq *rt_rq_iter_t; |
584 | ||
585 | #define for_each_rt_rq(rt_rq, iter, rq) \ | |
586 | for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL) | |
587 | ||
6f505b16 PZ |
588 | #define for_each_sched_rt_entity(rt_se) \ |
589 | for (; rt_se; rt_se = NULL) | |
590 | ||
591 | static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) | |
592 | { | |
593 | return NULL; | |
594 | } | |
595 | ||
9f0c1e56 | 596 | static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq) |
6f505b16 | 597 | { |
f4ebcbc0 KT |
598 | struct rq *rq = rq_of_rt_rq(rt_rq); |
599 | ||
600 | if (!rt_rq->rt_nr_running) | |
601 | return; | |
602 | ||
603 | enqueue_top_rt_rq(rt_rq); | |
8875125e | 604 | resched_curr(rq); |
6f505b16 PZ |
605 | } |
606 | ||
9f0c1e56 | 607 | static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq) |
6f505b16 | 608 | { |
f4ebcbc0 | 609 | dequeue_top_rt_rq(rt_rq); |
6f505b16 PZ |
610 | } |
611 | ||
46383648 KT |
612 | static inline int rt_rq_throttled(struct rt_rq *rt_rq) |
613 | { | |
614 | return rt_rq->rt_throttled; | |
615 | } | |
616 | ||
c6c4927b | 617 | static inline const struct cpumask *sched_rt_period_mask(void) |
d0b27fa7 | 618 | { |
c6c4927b | 619 | return cpu_online_mask; |
d0b27fa7 PZ |
620 | } |
621 | ||
622 | static inline | |
623 | struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) | |
624 | { | |
625 | return &cpu_rq(cpu)->rt; | |
626 | } | |
627 | ||
ac086bc2 PZ |
628 | static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) |
629 | { | |
630 | return &def_rt_bandwidth; | |
631 | } | |
632 | ||
55e12e5e | 633 | #endif /* CONFIG_RT_GROUP_SCHED */ |
d0b27fa7 | 634 | |
faa59937 JL |
635 | bool sched_rt_bandwidth_account(struct rt_rq *rt_rq) |
636 | { | |
637 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); | |
638 | ||
639 | return (hrtimer_active(&rt_b->rt_period_timer) || | |
640 | rt_rq->rt_time < rt_b->rt_runtime); | |
641 | } | |
642 | ||
ac086bc2 | 643 | #ifdef CONFIG_SMP |
78333cdd PZ |
644 | /* |
645 | * We ran out of runtime, see if we can borrow some from our neighbours. | |
646 | */ | |
269b26a5 | 647 | static void do_balance_runtime(struct rt_rq *rt_rq) |
ac086bc2 PZ |
648 | { |
649 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); | |
aa7f6730 | 650 | struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd; |
269b26a5 | 651 | int i, weight; |
ac086bc2 PZ |
652 | u64 rt_period; |
653 | ||
c6c4927b | 654 | weight = cpumask_weight(rd->span); |
ac086bc2 | 655 | |
0986b11b | 656 | raw_spin_lock(&rt_b->rt_runtime_lock); |
ac086bc2 | 657 | rt_period = ktime_to_ns(rt_b->rt_period); |
c6c4927b | 658 | for_each_cpu(i, rd->span) { |
ac086bc2 PZ |
659 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); |
660 | s64 diff; | |
661 | ||
662 | if (iter == rt_rq) | |
663 | continue; | |
664 | ||
0986b11b | 665 | raw_spin_lock(&iter->rt_runtime_lock); |
78333cdd PZ |
666 | /* |
667 | * Either all rqs have inf runtime and there's nothing to steal | |
668 | * or __disable_runtime() below sets a specific rq to inf to | |
669 | * indicate its been disabled and disalow stealing. | |
670 | */ | |
7def2be1 PZ |
671 | if (iter->rt_runtime == RUNTIME_INF) |
672 | goto next; | |
673 | ||
78333cdd PZ |
674 | /* |
675 | * From runqueues with spare time, take 1/n part of their | |
676 | * spare time, but no more than our period. | |
677 | */ | |
ac086bc2 PZ |
678 | diff = iter->rt_runtime - iter->rt_time; |
679 | if (diff > 0) { | |
58838cf3 | 680 | diff = div_u64((u64)diff, weight); |
ac086bc2 PZ |
681 | if (rt_rq->rt_runtime + diff > rt_period) |
682 | diff = rt_period - rt_rq->rt_runtime; | |
683 | iter->rt_runtime -= diff; | |
684 | rt_rq->rt_runtime += diff; | |
ac086bc2 | 685 | if (rt_rq->rt_runtime == rt_period) { |
0986b11b | 686 | raw_spin_unlock(&iter->rt_runtime_lock); |
ac086bc2 PZ |
687 | break; |
688 | } | |
689 | } | |
7def2be1 | 690 | next: |
0986b11b | 691 | raw_spin_unlock(&iter->rt_runtime_lock); |
ac086bc2 | 692 | } |
0986b11b | 693 | raw_spin_unlock(&rt_b->rt_runtime_lock); |
ac086bc2 | 694 | } |
7def2be1 | 695 | |
78333cdd PZ |
696 | /* |
697 | * Ensure this RQ takes back all the runtime it lend to its neighbours. | |
698 | */ | |
7def2be1 PZ |
699 | static void __disable_runtime(struct rq *rq) |
700 | { | |
701 | struct root_domain *rd = rq->rd; | |
ec514c48 | 702 | rt_rq_iter_t iter; |
7def2be1 PZ |
703 | struct rt_rq *rt_rq; |
704 | ||
705 | if (unlikely(!scheduler_running)) | |
706 | return; | |
707 | ||
ec514c48 | 708 | for_each_rt_rq(rt_rq, iter, rq) { |
7def2be1 PZ |
709 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
710 | s64 want; | |
711 | int i; | |
712 | ||
0986b11b TG |
713 | raw_spin_lock(&rt_b->rt_runtime_lock); |
714 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
78333cdd PZ |
715 | /* |
716 | * Either we're all inf and nobody needs to borrow, or we're | |
717 | * already disabled and thus have nothing to do, or we have | |
718 | * exactly the right amount of runtime to take out. | |
719 | */ | |
7def2be1 PZ |
720 | if (rt_rq->rt_runtime == RUNTIME_INF || |
721 | rt_rq->rt_runtime == rt_b->rt_runtime) | |
722 | goto balanced; | |
0986b11b | 723 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
7def2be1 | 724 | |
78333cdd PZ |
725 | /* |
726 | * Calculate the difference between what we started out with | |
727 | * and what we current have, that's the amount of runtime | |
728 | * we lend and now have to reclaim. | |
729 | */ | |
7def2be1 PZ |
730 | want = rt_b->rt_runtime - rt_rq->rt_runtime; |
731 | ||
78333cdd PZ |
732 | /* |
733 | * Greedy reclaim, take back as much as we can. | |
734 | */ | |
c6c4927b | 735 | for_each_cpu(i, rd->span) { |
7def2be1 PZ |
736 | struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); |
737 | s64 diff; | |
738 | ||
78333cdd PZ |
739 | /* |
740 | * Can't reclaim from ourselves or disabled runqueues. | |
741 | */ | |
f1679d08 | 742 | if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF) |
7def2be1 PZ |
743 | continue; |
744 | ||
0986b11b | 745 | raw_spin_lock(&iter->rt_runtime_lock); |
7def2be1 PZ |
746 | if (want > 0) { |
747 | diff = min_t(s64, iter->rt_runtime, want); | |
748 | iter->rt_runtime -= diff; | |
749 | want -= diff; | |
750 | } else { | |
751 | iter->rt_runtime -= want; | |
752 | want -= want; | |
753 | } | |
0986b11b | 754 | raw_spin_unlock(&iter->rt_runtime_lock); |
7def2be1 PZ |
755 | |
756 | if (!want) | |
757 | break; | |
758 | } | |
759 | ||
0986b11b | 760 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
78333cdd PZ |
761 | /* |
762 | * We cannot be left wanting - that would mean some runtime | |
763 | * leaked out of the system. | |
764 | */ | |
7def2be1 PZ |
765 | BUG_ON(want); |
766 | balanced: | |
78333cdd PZ |
767 | /* |
768 | * Disable all the borrow logic by pretending we have inf | |
769 | * runtime - in which case borrowing doesn't make sense. | |
770 | */ | |
7def2be1 | 771 | rt_rq->rt_runtime = RUNTIME_INF; |
a4c96ae3 | 772 | rt_rq->rt_throttled = 0; |
0986b11b TG |
773 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
774 | raw_spin_unlock(&rt_b->rt_runtime_lock); | |
99b62567 KT |
775 | |
776 | /* Make rt_rq available for pick_next_task() */ | |
777 | sched_rt_rq_enqueue(rt_rq); | |
7def2be1 PZ |
778 | } |
779 | } | |
780 | ||
7def2be1 PZ |
781 | static void __enable_runtime(struct rq *rq) |
782 | { | |
ec514c48 | 783 | rt_rq_iter_t iter; |
7def2be1 PZ |
784 | struct rt_rq *rt_rq; |
785 | ||
786 | if (unlikely(!scheduler_running)) | |
787 | return; | |
788 | ||
78333cdd PZ |
789 | /* |
790 | * Reset each runqueue's bandwidth settings | |
791 | */ | |
ec514c48 | 792 | for_each_rt_rq(rt_rq, iter, rq) { |
7def2be1 PZ |
793 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
794 | ||
0986b11b TG |
795 | raw_spin_lock(&rt_b->rt_runtime_lock); |
796 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
7def2be1 PZ |
797 | rt_rq->rt_runtime = rt_b->rt_runtime; |
798 | rt_rq->rt_time = 0; | |
baf25731 | 799 | rt_rq->rt_throttled = 0; |
0986b11b TG |
800 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
801 | raw_spin_unlock(&rt_b->rt_runtime_lock); | |
7def2be1 PZ |
802 | } |
803 | } | |
804 | ||
269b26a5 | 805 | static void balance_runtime(struct rt_rq *rt_rq) |
eff6549b | 806 | { |
4a6184ce | 807 | if (!sched_feat(RT_RUNTIME_SHARE)) |
269b26a5 | 808 | return; |
4a6184ce | 809 | |
eff6549b | 810 | if (rt_rq->rt_time > rt_rq->rt_runtime) { |
0986b11b | 811 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
269b26a5 | 812 | do_balance_runtime(rt_rq); |
0986b11b | 813 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
eff6549b | 814 | } |
eff6549b | 815 | } |
55e12e5e | 816 | #else /* !CONFIG_SMP */ |
269b26a5 | 817 | static inline void balance_runtime(struct rt_rq *rt_rq) {} |
55e12e5e | 818 | #endif /* CONFIG_SMP */ |
ac086bc2 | 819 | |
eff6549b PZ |
820 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun) |
821 | { | |
42c62a58 | 822 | int i, idle = 1, throttled = 0; |
c6c4927b | 823 | const struct cpumask *span; |
eff6549b | 824 | |
eff6549b | 825 | span = sched_rt_period_mask(); |
e221d028 MG |
826 | #ifdef CONFIG_RT_GROUP_SCHED |
827 | /* | |
828 | * FIXME: isolated CPUs should really leave the root task group, | |
829 | * whether they are isolcpus or were isolated via cpusets, lest | |
830 | * the timer run on a CPU which does not service all runqueues, | |
831 | * potentially leaving other CPUs indefinitely throttled. If | |
832 | * isolation is really required, the user will turn the throttle | |
833 | * off to kill the perturbations it causes anyway. Meanwhile, | |
834 | * this maintains functionality for boot and/or troubleshooting. | |
835 | */ | |
836 | if (rt_b == &root_task_group.rt_bandwidth) | |
837 | span = cpu_online_mask; | |
838 | #endif | |
c6c4927b | 839 | for_each_cpu(i, span) { |
eff6549b PZ |
840 | int enqueue = 0; |
841 | struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i); | |
842 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
c249f255 DK |
843 | int skip; |
844 | ||
845 | /* | |
846 | * When span == cpu_online_mask, taking each rq->lock | |
847 | * can be time-consuming. Try to avoid it when possible. | |
848 | */ | |
849 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
850 | skip = !rt_rq->rt_time && !rt_rq->rt_nr_running; | |
851 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
852 | if (skip) | |
853 | continue; | |
eff6549b | 854 | |
05fa785c | 855 | raw_spin_lock(&rq->lock); |
eff6549b PZ |
856 | if (rt_rq->rt_time) { |
857 | u64 runtime; | |
858 | ||
0986b11b | 859 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
eff6549b PZ |
860 | if (rt_rq->rt_throttled) |
861 | balance_runtime(rt_rq); | |
862 | runtime = rt_rq->rt_runtime; | |
863 | rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime); | |
864 | if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) { | |
865 | rt_rq->rt_throttled = 0; | |
866 | enqueue = 1; | |
61eadef6 MG |
867 | |
868 | /* | |
9edfbfed PZ |
869 | * When we're idle and a woken (rt) task is |
870 | * throttled check_preempt_curr() will set | |
871 | * skip_update and the time between the wakeup | |
872 | * and this unthrottle will get accounted as | |
873 | * 'runtime'. | |
61eadef6 MG |
874 | */ |
875 | if (rt_rq->rt_nr_running && rq->curr == rq->idle) | |
9edfbfed | 876 | rq_clock_skip_update(rq, false); |
eff6549b PZ |
877 | } |
878 | if (rt_rq->rt_time || rt_rq->rt_nr_running) | |
879 | idle = 0; | |
0986b11b | 880 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
0c3b9168 | 881 | } else if (rt_rq->rt_nr_running) { |
6c3df255 | 882 | idle = 0; |
0c3b9168 BS |
883 | if (!rt_rq_throttled(rt_rq)) |
884 | enqueue = 1; | |
885 | } | |
42c62a58 PZ |
886 | if (rt_rq->rt_throttled) |
887 | throttled = 1; | |
eff6549b PZ |
888 | |
889 | if (enqueue) | |
890 | sched_rt_rq_enqueue(rt_rq); | |
05fa785c | 891 | raw_spin_unlock(&rq->lock); |
eff6549b PZ |
892 | } |
893 | ||
42c62a58 PZ |
894 | if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)) |
895 | return 1; | |
896 | ||
eff6549b PZ |
897 | return idle; |
898 | } | |
ac086bc2 | 899 | |
6f505b16 PZ |
900 | static inline int rt_se_prio(struct sched_rt_entity *rt_se) |
901 | { | |
052f1dc7 | 902 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
903 | struct rt_rq *rt_rq = group_rt_rq(rt_se); |
904 | ||
905 | if (rt_rq) | |
e864c499 | 906 | return rt_rq->highest_prio.curr; |
6f505b16 PZ |
907 | #endif |
908 | ||
909 | return rt_task_of(rt_se)->prio; | |
910 | } | |
911 | ||
9f0c1e56 | 912 | static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq) |
6f505b16 | 913 | { |
9f0c1e56 | 914 | u64 runtime = sched_rt_runtime(rt_rq); |
fa85ae24 | 915 | |
fa85ae24 | 916 | if (rt_rq->rt_throttled) |
23b0fdfc | 917 | return rt_rq_throttled(rt_rq); |
fa85ae24 | 918 | |
5b680fd6 | 919 | if (runtime >= sched_rt_period(rt_rq)) |
ac086bc2 PZ |
920 | return 0; |
921 | ||
b79f3833 PZ |
922 | balance_runtime(rt_rq); |
923 | runtime = sched_rt_runtime(rt_rq); | |
924 | if (runtime == RUNTIME_INF) | |
925 | return 0; | |
ac086bc2 | 926 | |
9f0c1e56 | 927 | if (rt_rq->rt_time > runtime) { |
7abc63b1 PZ |
928 | struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); |
929 | ||
930 | /* | |
931 | * Don't actually throttle groups that have no runtime assigned | |
932 | * but accrue some time due to boosting. | |
933 | */ | |
934 | if (likely(rt_b->rt_runtime)) { | |
935 | rt_rq->rt_throttled = 1; | |
c224815d | 936 | printk_deferred_once("sched: RT throttling activated\n"); |
7abc63b1 PZ |
937 | } else { |
938 | /* | |
939 | * In case we did anyway, make it go away, | |
940 | * replenishment is a joke, since it will replenish us | |
941 | * with exactly 0 ns. | |
942 | */ | |
943 | rt_rq->rt_time = 0; | |
944 | } | |
945 | ||
23b0fdfc | 946 | if (rt_rq_throttled(rt_rq)) { |
9f0c1e56 | 947 | sched_rt_rq_dequeue(rt_rq); |
23b0fdfc PZ |
948 | return 1; |
949 | } | |
fa85ae24 PZ |
950 | } |
951 | ||
952 | return 0; | |
953 | } | |
954 | ||
bb44e5d1 IM |
955 | /* |
956 | * Update the current task's runtime statistics. Skip current tasks that | |
957 | * are not in our scheduling class. | |
958 | */ | |
a9957449 | 959 | static void update_curr_rt(struct rq *rq) |
bb44e5d1 IM |
960 | { |
961 | struct task_struct *curr = rq->curr; | |
6f505b16 | 962 | struct sched_rt_entity *rt_se = &curr->rt; |
bb44e5d1 IM |
963 | u64 delta_exec; |
964 | ||
06c3bc65 | 965 | if (curr->sched_class != &rt_sched_class) |
bb44e5d1 IM |
966 | return; |
967 | ||
78becc27 | 968 | delta_exec = rq_clock_task(rq) - curr->se.exec_start; |
fc79e240 KT |
969 | if (unlikely((s64)delta_exec <= 0)) |
970 | return; | |
6cfb0d5d | 971 | |
58919e83 | 972 | /* Kick cpufreq (see the comment in kernel/sched/sched.h). */ |
12bde33d | 973 | cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_RT); |
594dd290 | 974 | |
42c62a58 PZ |
975 | schedstat_set(curr->se.statistics.exec_max, |
976 | max(curr->se.statistics.exec_max, delta_exec)); | |
bb44e5d1 IM |
977 | |
978 | curr->se.sum_exec_runtime += delta_exec; | |
f06febc9 FM |
979 | account_group_exec_runtime(curr, delta_exec); |
980 | ||
78becc27 | 981 | curr->se.exec_start = rq_clock_task(rq); |
d842de87 | 982 | cpuacct_charge(curr, delta_exec); |
fa85ae24 | 983 | |
e9e9250b PZ |
984 | sched_rt_avg_update(rq, delta_exec); |
985 | ||
0b148fa0 PZ |
986 | if (!rt_bandwidth_enabled()) |
987 | return; | |
988 | ||
354d60c2 | 989 | for_each_sched_rt_entity(rt_se) { |
0b07939c | 990 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
354d60c2 | 991 | |
cc2991cf | 992 | if (sched_rt_runtime(rt_rq) != RUNTIME_INF) { |
0986b11b | 993 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
cc2991cf PZ |
994 | rt_rq->rt_time += delta_exec; |
995 | if (sched_rt_runtime_exceeded(rt_rq)) | |
8875125e | 996 | resched_curr(rq); |
0986b11b | 997 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
cc2991cf | 998 | } |
354d60c2 | 999 | } |
bb44e5d1 IM |
1000 | } |
1001 | ||
f4ebcbc0 KT |
1002 | static void |
1003 | dequeue_top_rt_rq(struct rt_rq *rt_rq) | |
1004 | { | |
1005 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
1006 | ||
1007 | BUG_ON(&rq->rt != rt_rq); | |
1008 | ||
1009 | if (!rt_rq->rt_queued) | |
1010 | return; | |
1011 | ||
1012 | BUG_ON(!rq->nr_running); | |
1013 | ||
72465447 | 1014 | sub_nr_running(rq, rt_rq->rt_nr_running); |
f4ebcbc0 KT |
1015 | rt_rq->rt_queued = 0; |
1016 | } | |
1017 | ||
1018 | static void | |
1019 | enqueue_top_rt_rq(struct rt_rq *rt_rq) | |
1020 | { | |
1021 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
1022 | ||
1023 | BUG_ON(&rq->rt != rt_rq); | |
1024 | ||
1025 | if (rt_rq->rt_queued) | |
1026 | return; | |
1027 | if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running) | |
1028 | return; | |
1029 | ||
72465447 | 1030 | add_nr_running(rq, rt_rq->rt_nr_running); |
f4ebcbc0 KT |
1031 | rt_rq->rt_queued = 1; |
1032 | } | |
1033 | ||
398a153b | 1034 | #if defined CONFIG_SMP |
e864c499 | 1035 | |
398a153b GH |
1036 | static void |
1037 | inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) | |
63489e45 | 1038 | { |
4d984277 | 1039 | struct rq *rq = rq_of_rt_rq(rt_rq); |
1f11eb6a | 1040 | |
757dfcaa KT |
1041 | #ifdef CONFIG_RT_GROUP_SCHED |
1042 | /* | |
1043 | * Change rq's cpupri only if rt_rq is the top queue. | |
1044 | */ | |
1045 | if (&rq->rt != rt_rq) | |
1046 | return; | |
1047 | #endif | |
5181f4a4 SR |
1048 | if (rq->online && prio < prev_prio) |
1049 | cpupri_set(&rq->rd->cpupri, rq->cpu, prio); | |
398a153b | 1050 | } |
73fe6aae | 1051 | |
398a153b GH |
1052 | static void |
1053 | dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) | |
1054 | { | |
1055 | struct rq *rq = rq_of_rt_rq(rt_rq); | |
d0b27fa7 | 1056 | |
757dfcaa KT |
1057 | #ifdef CONFIG_RT_GROUP_SCHED |
1058 | /* | |
1059 | * Change rq's cpupri only if rt_rq is the top queue. | |
1060 | */ | |
1061 | if (&rq->rt != rt_rq) | |
1062 | return; | |
1063 | #endif | |
398a153b GH |
1064 | if (rq->online && rt_rq->highest_prio.curr != prev_prio) |
1065 | cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr); | |
63489e45 SR |
1066 | } |
1067 | ||
398a153b GH |
1068 | #else /* CONFIG_SMP */ |
1069 | ||
6f505b16 | 1070 | static inline |
398a153b GH |
1071 | void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} |
1072 | static inline | |
1073 | void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} | |
1074 | ||
1075 | #endif /* CONFIG_SMP */ | |
6e0534f2 | 1076 | |
052f1dc7 | 1077 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
398a153b GH |
1078 | static void |
1079 | inc_rt_prio(struct rt_rq *rt_rq, int prio) | |
1080 | { | |
1081 | int prev_prio = rt_rq->highest_prio.curr; | |
1082 | ||
1083 | if (prio < prev_prio) | |
1084 | rt_rq->highest_prio.curr = prio; | |
1085 | ||
1086 | inc_rt_prio_smp(rt_rq, prio, prev_prio); | |
1087 | } | |
1088 | ||
1089 | static void | |
1090 | dec_rt_prio(struct rt_rq *rt_rq, int prio) | |
1091 | { | |
1092 | int prev_prio = rt_rq->highest_prio.curr; | |
1093 | ||
6f505b16 | 1094 | if (rt_rq->rt_nr_running) { |
764a9d6f | 1095 | |
398a153b | 1096 | WARN_ON(prio < prev_prio); |
764a9d6f | 1097 | |
e864c499 | 1098 | /* |
398a153b GH |
1099 | * This may have been our highest task, and therefore |
1100 | * we may have some recomputation to do | |
e864c499 | 1101 | */ |
398a153b | 1102 | if (prio == prev_prio) { |
e864c499 GH |
1103 | struct rt_prio_array *array = &rt_rq->active; |
1104 | ||
1105 | rt_rq->highest_prio.curr = | |
764a9d6f | 1106 | sched_find_first_bit(array->bitmap); |
e864c499 GH |
1107 | } |
1108 | ||
764a9d6f | 1109 | } else |
e864c499 | 1110 | rt_rq->highest_prio.curr = MAX_RT_PRIO; |
73fe6aae | 1111 | |
398a153b GH |
1112 | dec_rt_prio_smp(rt_rq, prio, prev_prio); |
1113 | } | |
1f11eb6a | 1114 | |
398a153b GH |
1115 | #else |
1116 | ||
1117 | static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {} | |
1118 | static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {} | |
1119 | ||
1120 | #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */ | |
6e0534f2 | 1121 | |
052f1dc7 | 1122 | #ifdef CONFIG_RT_GROUP_SCHED |
398a153b GH |
1123 | |
1124 | static void | |
1125 | inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1126 | { | |
1127 | if (rt_se_boosted(rt_se)) | |
1128 | rt_rq->rt_nr_boosted++; | |
1129 | ||
1130 | if (rt_rq->tg) | |
1131 | start_rt_bandwidth(&rt_rq->tg->rt_bandwidth); | |
1132 | } | |
1133 | ||
1134 | static void | |
1135 | dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1136 | { | |
23b0fdfc PZ |
1137 | if (rt_se_boosted(rt_se)) |
1138 | rt_rq->rt_nr_boosted--; | |
1139 | ||
1140 | WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted); | |
398a153b GH |
1141 | } |
1142 | ||
1143 | #else /* CONFIG_RT_GROUP_SCHED */ | |
1144 | ||
1145 | static void | |
1146 | inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1147 | { | |
1148 | start_rt_bandwidth(&def_rt_bandwidth); | |
1149 | } | |
1150 | ||
1151 | static inline | |
1152 | void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {} | |
1153 | ||
1154 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
1155 | ||
22abdef3 KT |
1156 | static inline |
1157 | unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se) | |
1158 | { | |
1159 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
1160 | ||
1161 | if (group_rq) | |
1162 | return group_rq->rt_nr_running; | |
1163 | else | |
1164 | return 1; | |
1165 | } | |
1166 | ||
01d36d0a FW |
1167 | static inline |
1168 | unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se) | |
1169 | { | |
1170 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
1171 | struct task_struct *tsk; | |
1172 | ||
1173 | if (group_rq) | |
1174 | return group_rq->rr_nr_running; | |
1175 | ||
1176 | tsk = rt_task_of(rt_se); | |
1177 | ||
1178 | return (tsk->policy == SCHED_RR) ? 1 : 0; | |
1179 | } | |
1180 | ||
398a153b GH |
1181 | static inline |
1182 | void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1183 | { | |
1184 | int prio = rt_se_prio(rt_se); | |
1185 | ||
1186 | WARN_ON(!rt_prio(prio)); | |
22abdef3 | 1187 | rt_rq->rt_nr_running += rt_se_nr_running(rt_se); |
01d36d0a | 1188 | rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se); |
398a153b GH |
1189 | |
1190 | inc_rt_prio(rt_rq, prio); | |
1191 | inc_rt_migration(rt_se, rt_rq); | |
1192 | inc_rt_group(rt_se, rt_rq); | |
1193 | } | |
1194 | ||
1195 | static inline | |
1196 | void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) | |
1197 | { | |
1198 | WARN_ON(!rt_prio(rt_se_prio(rt_se))); | |
1199 | WARN_ON(!rt_rq->rt_nr_running); | |
22abdef3 | 1200 | rt_rq->rt_nr_running -= rt_se_nr_running(rt_se); |
01d36d0a | 1201 | rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se); |
398a153b GH |
1202 | |
1203 | dec_rt_prio(rt_rq, rt_se_prio(rt_se)); | |
1204 | dec_rt_migration(rt_se, rt_rq); | |
1205 | dec_rt_group(rt_se, rt_rq); | |
63489e45 SR |
1206 | } |
1207 | ||
ff77e468 PZ |
1208 | /* |
1209 | * Change rt_se->run_list location unless SAVE && !MOVE | |
1210 | * | |
1211 | * assumes ENQUEUE/DEQUEUE flags match | |
1212 | */ | |
1213 | static inline bool move_entity(unsigned int flags) | |
1214 | { | |
1215 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE) | |
1216 | return false; | |
1217 | ||
1218 | return true; | |
1219 | } | |
1220 | ||
1221 | static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array) | |
1222 | { | |
1223 | list_del_init(&rt_se->run_list); | |
1224 | ||
1225 | if (list_empty(array->queue + rt_se_prio(rt_se))) | |
1226 | __clear_bit(rt_se_prio(rt_se), array->bitmap); | |
1227 | ||
1228 | rt_se->on_list = 0; | |
1229 | } | |
1230 | ||
1231 | static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) | |
bb44e5d1 | 1232 | { |
6f505b16 PZ |
1233 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); |
1234 | struct rt_prio_array *array = &rt_rq->active; | |
1235 | struct rt_rq *group_rq = group_rt_rq(rt_se); | |
20b6331b | 1236 | struct list_head *queue = array->queue + rt_se_prio(rt_se); |
bb44e5d1 | 1237 | |
ad2a3f13 PZ |
1238 | /* |
1239 | * Don't enqueue the group if its throttled, or when empty. | |
1240 | * The latter is a consequence of the former when a child group | |
1241 | * get throttled and the current group doesn't have any other | |
1242 | * active members. | |
1243 | */ | |
ff77e468 PZ |
1244 | if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) { |
1245 | if (rt_se->on_list) | |
1246 | __delist_rt_entity(rt_se, array); | |
6f505b16 | 1247 | return; |
ff77e468 | 1248 | } |
63489e45 | 1249 | |
ff77e468 PZ |
1250 | if (move_entity(flags)) { |
1251 | WARN_ON_ONCE(rt_se->on_list); | |
1252 | if (flags & ENQUEUE_HEAD) | |
1253 | list_add(&rt_se->run_list, queue); | |
1254 | else | |
1255 | list_add_tail(&rt_se->run_list, queue); | |
1256 | ||
1257 | __set_bit(rt_se_prio(rt_se), array->bitmap); | |
1258 | rt_se->on_list = 1; | |
1259 | } | |
1260 | rt_se->on_rq = 1; | |
78f2c7db | 1261 | |
6f505b16 PZ |
1262 | inc_rt_tasks(rt_se, rt_rq); |
1263 | } | |
1264 | ||
ff77e468 | 1265 | static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) |
6f505b16 PZ |
1266 | { |
1267 | struct rt_rq *rt_rq = rt_rq_of_se(rt_se); | |
1268 | struct rt_prio_array *array = &rt_rq->active; | |
1269 | ||
ff77e468 PZ |
1270 | if (move_entity(flags)) { |
1271 | WARN_ON_ONCE(!rt_se->on_list); | |
1272 | __delist_rt_entity(rt_se, array); | |
1273 | } | |
1274 | rt_se->on_rq = 0; | |
6f505b16 PZ |
1275 | |
1276 | dec_rt_tasks(rt_se, rt_rq); | |
1277 | } | |
1278 | ||
1279 | /* | |
1280 | * Because the prio of an upper entry depends on the lower | |
1281 | * entries, we must remove entries top - down. | |
6f505b16 | 1282 | */ |
ff77e468 | 1283 | static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags) |
6f505b16 | 1284 | { |
ad2a3f13 | 1285 | struct sched_rt_entity *back = NULL; |
6f505b16 | 1286 | |
58d6c2d7 PZ |
1287 | for_each_sched_rt_entity(rt_se) { |
1288 | rt_se->back = back; | |
1289 | back = rt_se; | |
1290 | } | |
1291 | ||
f4ebcbc0 KT |
1292 | dequeue_top_rt_rq(rt_rq_of_se(back)); |
1293 | ||
58d6c2d7 PZ |
1294 | for (rt_se = back; rt_se; rt_se = rt_se->back) { |
1295 | if (on_rt_rq(rt_se)) | |
ff77e468 | 1296 | __dequeue_rt_entity(rt_se, flags); |
ad2a3f13 PZ |
1297 | } |
1298 | } | |
1299 | ||
ff77e468 | 1300 | static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) |
ad2a3f13 | 1301 | { |
f4ebcbc0 KT |
1302 | struct rq *rq = rq_of_rt_se(rt_se); |
1303 | ||
ff77e468 | 1304 | dequeue_rt_stack(rt_se, flags); |
ad2a3f13 | 1305 | for_each_sched_rt_entity(rt_se) |
ff77e468 | 1306 | __enqueue_rt_entity(rt_se, flags); |
f4ebcbc0 | 1307 | enqueue_top_rt_rq(&rq->rt); |
ad2a3f13 PZ |
1308 | } |
1309 | ||
ff77e468 | 1310 | static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) |
ad2a3f13 | 1311 | { |
f4ebcbc0 KT |
1312 | struct rq *rq = rq_of_rt_se(rt_se); |
1313 | ||
ff77e468 | 1314 | dequeue_rt_stack(rt_se, flags); |
ad2a3f13 PZ |
1315 | |
1316 | for_each_sched_rt_entity(rt_se) { | |
1317 | struct rt_rq *rt_rq = group_rt_rq(rt_se); | |
1318 | ||
1319 | if (rt_rq && rt_rq->rt_nr_running) | |
ff77e468 | 1320 | __enqueue_rt_entity(rt_se, flags); |
58d6c2d7 | 1321 | } |
f4ebcbc0 | 1322 | enqueue_top_rt_rq(&rq->rt); |
bb44e5d1 IM |
1323 | } |
1324 | ||
1325 | /* | |
1326 | * Adding/removing a task to/from a priority array: | |
1327 | */ | |
ea87bb78 | 1328 | static void |
371fd7e7 | 1329 | enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags) |
6f505b16 PZ |
1330 | { |
1331 | struct sched_rt_entity *rt_se = &p->rt; | |
1332 | ||
371fd7e7 | 1333 | if (flags & ENQUEUE_WAKEUP) |
6f505b16 PZ |
1334 | rt_se->timeout = 0; |
1335 | ||
ff77e468 | 1336 | enqueue_rt_entity(rt_se, flags); |
c09595f6 | 1337 | |
4b53a341 | 1338 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) |
917b627d | 1339 | enqueue_pushable_task(rq, p); |
6f505b16 PZ |
1340 | } |
1341 | ||
371fd7e7 | 1342 | static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags) |
bb44e5d1 | 1343 | { |
6f505b16 | 1344 | struct sched_rt_entity *rt_se = &p->rt; |
bb44e5d1 | 1345 | |
f1e14ef6 | 1346 | update_curr_rt(rq); |
ff77e468 | 1347 | dequeue_rt_entity(rt_se, flags); |
c09595f6 | 1348 | |
917b627d | 1349 | dequeue_pushable_task(rq, p); |
bb44e5d1 IM |
1350 | } |
1351 | ||
1352 | /* | |
60686317 RW |
1353 | * Put task to the head or the end of the run list without the overhead of |
1354 | * dequeue followed by enqueue. | |
bb44e5d1 | 1355 | */ |
7ebefa8c DA |
1356 | static void |
1357 | requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head) | |
6f505b16 | 1358 | { |
1cdad715 | 1359 | if (on_rt_rq(rt_se)) { |
7ebefa8c DA |
1360 | struct rt_prio_array *array = &rt_rq->active; |
1361 | struct list_head *queue = array->queue + rt_se_prio(rt_se); | |
1362 | ||
1363 | if (head) | |
1364 | list_move(&rt_se->run_list, queue); | |
1365 | else | |
1366 | list_move_tail(&rt_se->run_list, queue); | |
1cdad715 | 1367 | } |
6f505b16 PZ |
1368 | } |
1369 | ||
7ebefa8c | 1370 | static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head) |
bb44e5d1 | 1371 | { |
6f505b16 PZ |
1372 | struct sched_rt_entity *rt_se = &p->rt; |
1373 | struct rt_rq *rt_rq; | |
bb44e5d1 | 1374 | |
6f505b16 PZ |
1375 | for_each_sched_rt_entity(rt_se) { |
1376 | rt_rq = rt_rq_of_se(rt_se); | |
7ebefa8c | 1377 | requeue_rt_entity(rt_rq, rt_se, head); |
6f505b16 | 1378 | } |
bb44e5d1 IM |
1379 | } |
1380 | ||
6f505b16 | 1381 | static void yield_task_rt(struct rq *rq) |
bb44e5d1 | 1382 | { |
7ebefa8c | 1383 | requeue_task_rt(rq, rq->curr, 0); |
bb44e5d1 IM |
1384 | } |
1385 | ||
e7693a36 | 1386 | #ifdef CONFIG_SMP |
318e0893 GH |
1387 | static int find_lowest_rq(struct task_struct *task); |
1388 | ||
0017d735 | 1389 | static int |
ac66f547 | 1390 | select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags) |
e7693a36 | 1391 | { |
7608dec2 PZ |
1392 | struct task_struct *curr; |
1393 | struct rq *rq; | |
c37495fd SR |
1394 | |
1395 | /* For anything but wake ups, just return the task_cpu */ | |
1396 | if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK) | |
1397 | goto out; | |
1398 | ||
7608dec2 PZ |
1399 | rq = cpu_rq(cpu); |
1400 | ||
1401 | rcu_read_lock(); | |
316c1608 | 1402 | curr = READ_ONCE(rq->curr); /* unlocked access */ |
7608dec2 | 1403 | |
318e0893 | 1404 | /* |
7608dec2 | 1405 | * If the current task on @p's runqueue is an RT task, then |
e1f47d89 SR |
1406 | * try to see if we can wake this RT task up on another |
1407 | * runqueue. Otherwise simply start this RT task | |
1408 | * on its current runqueue. | |
1409 | * | |
43fa5460 SR |
1410 | * We want to avoid overloading runqueues. If the woken |
1411 | * task is a higher priority, then it will stay on this CPU | |
1412 | * and the lower prio task should be moved to another CPU. | |
1413 | * Even though this will probably make the lower prio task | |
1414 | * lose its cache, we do not want to bounce a higher task | |
1415 | * around just because it gave up its CPU, perhaps for a | |
1416 | * lock? | |
1417 | * | |
1418 | * For equal prio tasks, we just let the scheduler sort it out. | |
7608dec2 PZ |
1419 | * |
1420 | * Otherwise, just let it ride on the affined RQ and the | |
1421 | * post-schedule router will push the preempted task away | |
1422 | * | |
1423 | * This test is optimistic, if we get it wrong the load-balancer | |
1424 | * will have to sort it out. | |
318e0893 | 1425 | */ |
7608dec2 | 1426 | if (curr && unlikely(rt_task(curr)) && |
4b53a341 | 1427 | (curr->nr_cpus_allowed < 2 || |
6bfa687c | 1428 | curr->prio <= p->prio)) { |
7608dec2 | 1429 | int target = find_lowest_rq(p); |
318e0893 | 1430 | |
80e3d87b TC |
1431 | /* |
1432 | * Don't bother moving it if the destination CPU is | |
1433 | * not running a lower priority task. | |
1434 | */ | |
1435 | if (target != -1 && | |
1436 | p->prio < cpu_rq(target)->rt.highest_prio.curr) | |
7608dec2 | 1437 | cpu = target; |
318e0893 | 1438 | } |
7608dec2 | 1439 | rcu_read_unlock(); |
318e0893 | 1440 | |
c37495fd | 1441 | out: |
7608dec2 | 1442 | return cpu; |
e7693a36 | 1443 | } |
7ebefa8c DA |
1444 | |
1445 | static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p) | |
1446 | { | |
308a623a WL |
1447 | /* |
1448 | * Current can't be migrated, useless to reschedule, | |
1449 | * let's hope p can move out. | |
1450 | */ | |
4b53a341 | 1451 | if (rq->curr->nr_cpus_allowed == 1 || |
308a623a | 1452 | !cpupri_find(&rq->rd->cpupri, rq->curr, NULL)) |
7ebefa8c DA |
1453 | return; |
1454 | ||
308a623a WL |
1455 | /* |
1456 | * p is migratable, so let's not schedule it and | |
1457 | * see if it is pushed or pulled somewhere else. | |
1458 | */ | |
4b53a341 | 1459 | if (p->nr_cpus_allowed != 1 |
13b8bd0a RR |
1460 | && cpupri_find(&rq->rd->cpupri, p, NULL)) |
1461 | return; | |
24600ce8 | 1462 | |
7ebefa8c DA |
1463 | /* |
1464 | * There appears to be other cpus that can accept | |
1465 | * current and none to run 'p', so lets reschedule | |
1466 | * to try and push current away: | |
1467 | */ | |
1468 | requeue_task_rt(rq, p, 1); | |
8875125e | 1469 | resched_curr(rq); |
7ebefa8c DA |
1470 | } |
1471 | ||
e7693a36 GH |
1472 | #endif /* CONFIG_SMP */ |
1473 | ||
bb44e5d1 IM |
1474 | /* |
1475 | * Preempt the current task with a newly woken task if needed: | |
1476 | */ | |
7d478721 | 1477 | static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags) |
bb44e5d1 | 1478 | { |
45c01e82 | 1479 | if (p->prio < rq->curr->prio) { |
8875125e | 1480 | resched_curr(rq); |
45c01e82 GH |
1481 | return; |
1482 | } | |
1483 | ||
1484 | #ifdef CONFIG_SMP | |
1485 | /* | |
1486 | * If: | |
1487 | * | |
1488 | * - the newly woken task is of equal priority to the current task | |
1489 | * - the newly woken task is non-migratable while current is migratable | |
1490 | * - current will be preempted on the next reschedule | |
1491 | * | |
1492 | * we should check to see if current can readily move to a different | |
1493 | * cpu. If so, we will reschedule to allow the push logic to try | |
1494 | * to move current somewhere else, making room for our non-migratable | |
1495 | * task. | |
1496 | */ | |
8dd0de8b | 1497 | if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr)) |
7ebefa8c | 1498 | check_preempt_equal_prio(rq, p); |
45c01e82 | 1499 | #endif |
bb44e5d1 IM |
1500 | } |
1501 | ||
6f505b16 PZ |
1502 | static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq, |
1503 | struct rt_rq *rt_rq) | |
bb44e5d1 | 1504 | { |
6f505b16 PZ |
1505 | struct rt_prio_array *array = &rt_rq->active; |
1506 | struct sched_rt_entity *next = NULL; | |
bb44e5d1 IM |
1507 | struct list_head *queue; |
1508 | int idx; | |
1509 | ||
1510 | idx = sched_find_first_bit(array->bitmap); | |
6f505b16 | 1511 | BUG_ON(idx >= MAX_RT_PRIO); |
bb44e5d1 IM |
1512 | |
1513 | queue = array->queue + idx; | |
6f505b16 | 1514 | next = list_entry(queue->next, struct sched_rt_entity, run_list); |
326587b8 | 1515 | |
6f505b16 PZ |
1516 | return next; |
1517 | } | |
bb44e5d1 | 1518 | |
917b627d | 1519 | static struct task_struct *_pick_next_task_rt(struct rq *rq) |
6f505b16 PZ |
1520 | { |
1521 | struct sched_rt_entity *rt_se; | |
1522 | struct task_struct *p; | |
606dba2e | 1523 | struct rt_rq *rt_rq = &rq->rt; |
6f505b16 PZ |
1524 | |
1525 | do { | |
1526 | rt_se = pick_next_rt_entity(rq, rt_rq); | |
326587b8 | 1527 | BUG_ON(!rt_se); |
6f505b16 PZ |
1528 | rt_rq = group_rt_rq(rt_se); |
1529 | } while (rt_rq); | |
1530 | ||
1531 | p = rt_task_of(rt_se); | |
78becc27 | 1532 | p->se.exec_start = rq_clock_task(rq); |
917b627d GH |
1533 | |
1534 | return p; | |
1535 | } | |
1536 | ||
606dba2e | 1537 | static struct task_struct * |
d8ac8971 | 1538 | pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
917b627d | 1539 | { |
606dba2e PZ |
1540 | struct task_struct *p; |
1541 | struct rt_rq *rt_rq = &rq->rt; | |
1542 | ||
37e117c0 | 1543 | if (need_pull_rt_task(rq, prev)) { |
cbce1a68 PZ |
1544 | /* |
1545 | * This is OK, because current is on_cpu, which avoids it being | |
1546 | * picked for load-balance and preemption/IRQs are still | |
1547 | * disabled avoiding further scheduler activity on it and we're | |
1548 | * being very careful to re-start the picking loop. | |
1549 | */ | |
d8ac8971 | 1550 | rq_unpin_lock(rq, rf); |
38033c37 | 1551 | pull_rt_task(rq); |
d8ac8971 | 1552 | rq_repin_lock(rq, rf); |
37e117c0 PZ |
1553 | /* |
1554 | * pull_rt_task() can drop (and re-acquire) rq->lock; this | |
a1d9a323 KT |
1555 | * means a dl or stop task can slip in, in which case we need |
1556 | * to re-start task selection. | |
37e117c0 | 1557 | */ |
da0c1e65 | 1558 | if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) || |
a1d9a323 | 1559 | rq->dl.dl_nr_running)) |
37e117c0 PZ |
1560 | return RETRY_TASK; |
1561 | } | |
38033c37 | 1562 | |
734ff2a7 KT |
1563 | /* |
1564 | * We may dequeue prev's rt_rq in put_prev_task(). | |
1565 | * So, we update time before rt_nr_running check. | |
1566 | */ | |
1567 | if (prev->sched_class == &rt_sched_class) | |
1568 | update_curr_rt(rq); | |
1569 | ||
f4ebcbc0 | 1570 | if (!rt_rq->rt_queued) |
606dba2e PZ |
1571 | return NULL; |
1572 | ||
3f1d2a31 | 1573 | put_prev_task(rq, prev); |
606dba2e PZ |
1574 | |
1575 | p = _pick_next_task_rt(rq); | |
917b627d GH |
1576 | |
1577 | /* The running task is never eligible for pushing */ | |
f3f1768f | 1578 | dequeue_pushable_task(rq, p); |
917b627d | 1579 | |
e3fca9e7 | 1580 | queue_push_tasks(rq); |
3f029d3c | 1581 | |
6f505b16 | 1582 | return p; |
bb44e5d1 IM |
1583 | } |
1584 | ||
31ee529c | 1585 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p) |
bb44e5d1 | 1586 | { |
f1e14ef6 | 1587 | update_curr_rt(rq); |
917b627d GH |
1588 | |
1589 | /* | |
1590 | * The previous task needs to be made eligible for pushing | |
1591 | * if it is still active | |
1592 | */ | |
4b53a341 | 1593 | if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1) |
917b627d | 1594 | enqueue_pushable_task(rq, p); |
bb44e5d1 IM |
1595 | } |
1596 | ||
681f3e68 | 1597 | #ifdef CONFIG_SMP |
6f505b16 | 1598 | |
e8fa1362 SR |
1599 | /* Only try algorithms three times */ |
1600 | #define RT_MAX_TRIES 3 | |
1601 | ||
f65eda4f SR |
1602 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) |
1603 | { | |
1604 | if (!task_running(rq, p) && | |
0c98d344 | 1605 | cpumask_test_cpu(cpu, &p->cpus_allowed)) |
f65eda4f SR |
1606 | return 1; |
1607 | return 0; | |
1608 | } | |
1609 | ||
e23ee747 KT |
1610 | /* |
1611 | * Return the highest pushable rq's task, which is suitable to be executed | |
1612 | * on the cpu, NULL otherwise | |
1613 | */ | |
1614 | static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu) | |
e8fa1362 | 1615 | { |
e23ee747 KT |
1616 | struct plist_head *head = &rq->rt.pushable_tasks; |
1617 | struct task_struct *p; | |
3d07467b | 1618 | |
e23ee747 KT |
1619 | if (!has_pushable_tasks(rq)) |
1620 | return NULL; | |
3d07467b | 1621 | |
e23ee747 KT |
1622 | plist_for_each_entry(p, head, pushable_tasks) { |
1623 | if (pick_rt_task(rq, p, cpu)) | |
1624 | return p; | |
f65eda4f SR |
1625 | } |
1626 | ||
e23ee747 | 1627 | return NULL; |
e8fa1362 SR |
1628 | } |
1629 | ||
0e3900e6 | 1630 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask); |
e8fa1362 | 1631 | |
6e1254d2 GH |
1632 | static int find_lowest_rq(struct task_struct *task) |
1633 | { | |
1634 | struct sched_domain *sd; | |
4ba29684 | 1635 | struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask); |
6e1254d2 GH |
1636 | int this_cpu = smp_processor_id(); |
1637 | int cpu = task_cpu(task); | |
06f90dbd | 1638 | |
0da938c4 SR |
1639 | /* Make sure the mask is initialized first */ |
1640 | if (unlikely(!lowest_mask)) | |
1641 | return -1; | |
1642 | ||
4b53a341 | 1643 | if (task->nr_cpus_allowed == 1) |
6e0534f2 | 1644 | return -1; /* No other targets possible */ |
6e1254d2 | 1645 | |
6e0534f2 GH |
1646 | if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask)) |
1647 | return -1; /* No targets found */ | |
6e1254d2 GH |
1648 | |
1649 | /* | |
1650 | * At this point we have built a mask of cpus representing the | |
1651 | * lowest priority tasks in the system. Now we want to elect | |
1652 | * the best one based on our affinity and topology. | |
1653 | * | |
1654 | * We prioritize the last cpu that the task executed on since | |
1655 | * it is most likely cache-hot in that location. | |
1656 | */ | |
96f874e2 | 1657 | if (cpumask_test_cpu(cpu, lowest_mask)) |
6e1254d2 GH |
1658 | return cpu; |
1659 | ||
1660 | /* | |
1661 | * Otherwise, we consult the sched_domains span maps to figure | |
1662 | * out which cpu is logically closest to our hot cache data. | |
1663 | */ | |
e2c88063 RR |
1664 | if (!cpumask_test_cpu(this_cpu, lowest_mask)) |
1665 | this_cpu = -1; /* Skip this_cpu opt if not among lowest */ | |
6e1254d2 | 1666 | |
cd4ae6ad | 1667 | rcu_read_lock(); |
e2c88063 RR |
1668 | for_each_domain(cpu, sd) { |
1669 | if (sd->flags & SD_WAKE_AFFINE) { | |
1670 | int best_cpu; | |
6e1254d2 | 1671 | |
e2c88063 RR |
1672 | /* |
1673 | * "this_cpu" is cheaper to preempt than a | |
1674 | * remote processor. | |
1675 | */ | |
1676 | if (this_cpu != -1 && | |
cd4ae6ad XF |
1677 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { |
1678 | rcu_read_unlock(); | |
e2c88063 | 1679 | return this_cpu; |
cd4ae6ad | 1680 | } |
e2c88063 RR |
1681 | |
1682 | best_cpu = cpumask_first_and(lowest_mask, | |
1683 | sched_domain_span(sd)); | |
cd4ae6ad XF |
1684 | if (best_cpu < nr_cpu_ids) { |
1685 | rcu_read_unlock(); | |
e2c88063 | 1686 | return best_cpu; |
cd4ae6ad | 1687 | } |
6e1254d2 GH |
1688 | } |
1689 | } | |
cd4ae6ad | 1690 | rcu_read_unlock(); |
6e1254d2 GH |
1691 | |
1692 | /* | |
1693 | * And finally, if there were no matches within the domains | |
1694 | * just give the caller *something* to work with from the compatible | |
1695 | * locations. | |
1696 | */ | |
e2c88063 RR |
1697 | if (this_cpu != -1) |
1698 | return this_cpu; | |
1699 | ||
1700 | cpu = cpumask_any(lowest_mask); | |
1701 | if (cpu < nr_cpu_ids) | |
1702 | return cpu; | |
1703 | return -1; | |
07b4032c GH |
1704 | } |
1705 | ||
1706 | /* Will lock the rq it finds */ | |
4df64c0b | 1707 | static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) |
07b4032c GH |
1708 | { |
1709 | struct rq *lowest_rq = NULL; | |
07b4032c | 1710 | int tries; |
4df64c0b | 1711 | int cpu; |
e8fa1362 | 1712 | |
07b4032c GH |
1713 | for (tries = 0; tries < RT_MAX_TRIES; tries++) { |
1714 | cpu = find_lowest_rq(task); | |
1715 | ||
2de0b463 | 1716 | if ((cpu == -1) || (cpu == rq->cpu)) |
e8fa1362 SR |
1717 | break; |
1718 | ||
07b4032c GH |
1719 | lowest_rq = cpu_rq(cpu); |
1720 | ||
80e3d87b TC |
1721 | if (lowest_rq->rt.highest_prio.curr <= task->prio) { |
1722 | /* | |
1723 | * Target rq has tasks of equal or higher priority, | |
1724 | * retrying does not release any lock and is unlikely | |
1725 | * to yield a different result. | |
1726 | */ | |
1727 | lowest_rq = NULL; | |
1728 | break; | |
1729 | } | |
1730 | ||
e8fa1362 | 1731 | /* if the prio of this runqueue changed, try again */ |
07b4032c | 1732 | if (double_lock_balance(rq, lowest_rq)) { |
e8fa1362 SR |
1733 | /* |
1734 | * We had to unlock the run queue. In | |
1735 | * the mean time, task could have | |
1736 | * migrated already or had its affinity changed. | |
1737 | * Also make sure that it wasn't scheduled on its rq. | |
1738 | */ | |
07b4032c | 1739 | if (unlikely(task_rq(task) != rq || |
0c98d344 | 1740 | !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_allowed) || |
07b4032c | 1741 | task_running(rq, task) || |
13b5ab02 | 1742 | !rt_task(task) || |
da0c1e65 | 1743 | !task_on_rq_queued(task))) { |
4df64c0b | 1744 | |
7f1b4393 | 1745 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 SR |
1746 | lowest_rq = NULL; |
1747 | break; | |
1748 | } | |
1749 | } | |
1750 | ||
1751 | /* If this rq is still suitable use it. */ | |
e864c499 | 1752 | if (lowest_rq->rt.highest_prio.curr > task->prio) |
e8fa1362 SR |
1753 | break; |
1754 | ||
1755 | /* try again */ | |
1b12bbc7 | 1756 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 SR |
1757 | lowest_rq = NULL; |
1758 | } | |
1759 | ||
1760 | return lowest_rq; | |
1761 | } | |
1762 | ||
917b627d GH |
1763 | static struct task_struct *pick_next_pushable_task(struct rq *rq) |
1764 | { | |
1765 | struct task_struct *p; | |
1766 | ||
1767 | if (!has_pushable_tasks(rq)) | |
1768 | return NULL; | |
1769 | ||
1770 | p = plist_first_entry(&rq->rt.pushable_tasks, | |
1771 | struct task_struct, pushable_tasks); | |
1772 | ||
1773 | BUG_ON(rq->cpu != task_cpu(p)); | |
1774 | BUG_ON(task_current(rq, p)); | |
4b53a341 | 1775 | BUG_ON(p->nr_cpus_allowed <= 1); |
917b627d | 1776 | |
da0c1e65 | 1777 | BUG_ON(!task_on_rq_queued(p)); |
917b627d GH |
1778 | BUG_ON(!rt_task(p)); |
1779 | ||
1780 | return p; | |
1781 | } | |
1782 | ||
e8fa1362 SR |
1783 | /* |
1784 | * If the current CPU has more than one RT task, see if the non | |
1785 | * running task can migrate over to a CPU that is running a task | |
1786 | * of lesser priority. | |
1787 | */ | |
697f0a48 | 1788 | static int push_rt_task(struct rq *rq) |
e8fa1362 SR |
1789 | { |
1790 | struct task_struct *next_task; | |
1791 | struct rq *lowest_rq; | |
311e800e | 1792 | int ret = 0; |
e8fa1362 | 1793 | |
a22d7fc1 GH |
1794 | if (!rq->rt.overloaded) |
1795 | return 0; | |
1796 | ||
917b627d | 1797 | next_task = pick_next_pushable_task(rq); |
e8fa1362 SR |
1798 | if (!next_task) |
1799 | return 0; | |
1800 | ||
49246274 | 1801 | retry: |
697f0a48 | 1802 | if (unlikely(next_task == rq->curr)) { |
f65eda4f | 1803 | WARN_ON(1); |
e8fa1362 | 1804 | return 0; |
f65eda4f | 1805 | } |
e8fa1362 SR |
1806 | |
1807 | /* | |
1808 | * It's possible that the next_task slipped in of | |
1809 | * higher priority than current. If that's the case | |
1810 | * just reschedule current. | |
1811 | */ | |
697f0a48 | 1812 | if (unlikely(next_task->prio < rq->curr->prio)) { |
8875125e | 1813 | resched_curr(rq); |
e8fa1362 SR |
1814 | return 0; |
1815 | } | |
1816 | ||
697f0a48 | 1817 | /* We might release rq lock */ |
e8fa1362 SR |
1818 | get_task_struct(next_task); |
1819 | ||
1820 | /* find_lock_lowest_rq locks the rq if found */ | |
697f0a48 | 1821 | lowest_rq = find_lock_lowest_rq(next_task, rq); |
e8fa1362 SR |
1822 | if (!lowest_rq) { |
1823 | struct task_struct *task; | |
1824 | /* | |
311e800e | 1825 | * find_lock_lowest_rq releases rq->lock |
1563513d GH |
1826 | * so it is possible that next_task has migrated. |
1827 | * | |
1828 | * We need to make sure that the task is still on the same | |
1829 | * run-queue and is also still the next task eligible for | |
1830 | * pushing. | |
e8fa1362 | 1831 | */ |
917b627d | 1832 | task = pick_next_pushable_task(rq); |
de16b91e | 1833 | if (task == next_task) { |
1563513d | 1834 | /* |
311e800e HD |
1835 | * The task hasn't migrated, and is still the next |
1836 | * eligible task, but we failed to find a run-queue | |
1837 | * to push it to. Do not retry in this case, since | |
1838 | * other cpus will pull from us when ready. | |
1563513d | 1839 | */ |
1563513d | 1840 | goto out; |
e8fa1362 | 1841 | } |
917b627d | 1842 | |
1563513d GH |
1843 | if (!task) |
1844 | /* No more tasks, just exit */ | |
1845 | goto out; | |
1846 | ||
917b627d | 1847 | /* |
1563513d | 1848 | * Something has shifted, try again. |
917b627d | 1849 | */ |
1563513d GH |
1850 | put_task_struct(next_task); |
1851 | next_task = task; | |
1852 | goto retry; | |
e8fa1362 SR |
1853 | } |
1854 | ||
697f0a48 | 1855 | deactivate_task(rq, next_task, 0); |
e8fa1362 SR |
1856 | set_task_cpu(next_task, lowest_rq->cpu); |
1857 | activate_task(lowest_rq, next_task, 0); | |
311e800e | 1858 | ret = 1; |
e8fa1362 | 1859 | |
8875125e | 1860 | resched_curr(lowest_rq); |
e8fa1362 | 1861 | |
1b12bbc7 | 1862 | double_unlock_balance(rq, lowest_rq); |
e8fa1362 | 1863 | |
e8fa1362 SR |
1864 | out: |
1865 | put_task_struct(next_task); | |
1866 | ||
311e800e | 1867 | return ret; |
e8fa1362 SR |
1868 | } |
1869 | ||
e8fa1362 SR |
1870 | static void push_rt_tasks(struct rq *rq) |
1871 | { | |
1872 | /* push_rt_task will return true if it moved an RT */ | |
1873 | while (push_rt_task(rq)) | |
1874 | ; | |
1875 | } | |
1876 | ||
b6366f04 SR |
1877 | #ifdef HAVE_RT_PUSH_IPI |
1878 | /* | |
1879 | * The search for the next cpu always starts at rq->cpu and ends | |
1880 | * when we reach rq->cpu again. It will never return rq->cpu. | |
1881 | * This returns the next cpu to check, or nr_cpu_ids if the loop | |
1882 | * is complete. | |
1883 | * | |
1884 | * rq->rt.push_cpu holds the last cpu returned by this function, | |
1885 | * or if this is the first instance, it must hold rq->cpu. | |
1886 | */ | |
1887 | static int rto_next_cpu(struct rq *rq) | |
1888 | { | |
1889 | int prev_cpu = rq->rt.push_cpu; | |
1890 | int cpu; | |
1891 | ||
1892 | cpu = cpumask_next(prev_cpu, rq->rd->rto_mask); | |
1893 | ||
1894 | /* | |
1895 | * If the previous cpu is less than the rq's CPU, then it already | |
1896 | * passed the end of the mask, and has started from the beginning. | |
1897 | * We end if the next CPU is greater or equal to rq's CPU. | |
1898 | */ | |
1899 | if (prev_cpu < rq->cpu) { | |
1900 | if (cpu >= rq->cpu) | |
1901 | return nr_cpu_ids; | |
1902 | ||
1903 | } else if (cpu >= nr_cpu_ids) { | |
1904 | /* | |
1905 | * We passed the end of the mask, start at the beginning. | |
1906 | * If the result is greater or equal to the rq's CPU, then | |
1907 | * the loop is finished. | |
1908 | */ | |
1909 | cpu = cpumask_first(rq->rd->rto_mask); | |
1910 | if (cpu >= rq->cpu) | |
1911 | return nr_cpu_ids; | |
1912 | } | |
1913 | rq->rt.push_cpu = cpu; | |
1914 | ||
1915 | /* Return cpu to let the caller know if the loop is finished or not */ | |
1916 | return cpu; | |
1917 | } | |
1918 | ||
1919 | static int find_next_push_cpu(struct rq *rq) | |
1920 | { | |
1921 | struct rq *next_rq; | |
1922 | int cpu; | |
1923 | ||
1924 | while (1) { | |
1925 | cpu = rto_next_cpu(rq); | |
1926 | if (cpu >= nr_cpu_ids) | |
1927 | break; | |
1928 | next_rq = cpu_rq(cpu); | |
1929 | ||
1930 | /* Make sure the next rq can push to this rq */ | |
1931 | if (next_rq->rt.highest_prio.next < rq->rt.highest_prio.curr) | |
1932 | break; | |
1933 | } | |
1934 | ||
1935 | return cpu; | |
1936 | } | |
1937 | ||
1938 | #define RT_PUSH_IPI_EXECUTING 1 | |
1939 | #define RT_PUSH_IPI_RESTART 2 | |
1940 | ||
3e777f99 SRV |
1941 | /* |
1942 | * When a high priority task schedules out from a CPU and a lower priority | |
1943 | * task is scheduled in, a check is made to see if there's any RT tasks | |
1944 | * on other CPUs that are waiting to run because a higher priority RT task | |
1945 | * is currently running on its CPU. In this case, the CPU with multiple RT | |
1946 | * tasks queued on it (overloaded) needs to be notified that a CPU has opened | |
1947 | * up that may be able to run one of its non-running queued RT tasks. | |
1948 | * | |
1949 | * On large CPU boxes, there's the case that several CPUs could schedule | |
1950 | * a lower priority task at the same time, in which case it will look for | |
1951 | * any overloaded CPUs that it could pull a task from. To do this, the runqueue | |
1952 | * lock must be taken from that overloaded CPU. Having 10s of CPUs all fighting | |
1953 | * for a single overloaded CPU's runqueue lock can produce a large latency. | |
1954 | * (This has actually been observed on large boxes running cyclictest). | |
1955 | * Instead of taking the runqueue lock of the overloaded CPU, each of the | |
1956 | * CPUs that scheduled a lower priority task simply sends an IPI to the | |
1957 | * overloaded CPU. An IPI is much cheaper than taking an runqueue lock with | |
1958 | * lots of contention. The overloaded CPU will look to push its non-running | |
1959 | * RT task off, and if it does, it can then ignore the other IPIs coming | |
1960 | * in, and just pass those IPIs off to any other overloaded CPU. | |
1961 | * | |
1962 | * When a CPU schedules a lower priority task, it only sends an IPI to | |
1963 | * the "next" CPU that has overloaded RT tasks. This prevents IPI storms, | |
1964 | * as having 10 CPUs scheduling lower priority tasks and 10 CPUs with | |
1965 | * RT overloaded tasks, would cause 100 IPIs to go out at once. | |
1966 | * | |
1967 | * The overloaded RT CPU, when receiving an IPI, will try to push off its | |
1968 | * overloaded RT tasks and then send an IPI to the next CPU that has | |
1969 | * overloaded RT tasks. This stops when all CPUs with overloaded RT tasks | |
1970 | * have completed. Just because a CPU may have pushed off its own overloaded | |
1971 | * RT task does not mean it should stop sending the IPI around to other | |
1972 | * overloaded CPUs. There may be another RT task waiting to run on one of | |
1973 | * those CPUs that are of higher priority than the one that was just | |
1974 | * pushed. | |
1975 | * | |
1976 | * An optimization that could possibly be made is to make a CPU array similar | |
1977 | * to the cpupri array mask of all running RT tasks, but for the overloaded | |
1978 | * case, then the IPI could be sent to only the CPU with the highest priority | |
1979 | * RT task waiting, and that CPU could send off further IPIs to the CPU with | |
1980 | * the next highest waiting task. Since the overloaded case is much less likely | |
1981 | * to happen, the complexity of this implementation may not be worth it. | |
1982 | * Instead, just send an IPI around to all overloaded CPUs. | |
1983 | * | |
1984 | * The rq->rt.push_flags holds the status of the IPI that is going around. | |
1985 | * A run queue can only send out a single IPI at a time. The possible flags | |
1986 | * for rq->rt.push_flags are: | |
1987 | * | |
1988 | * (None or zero): No IPI is going around for the current rq | |
1989 | * RT_PUSH_IPI_EXECUTING: An IPI for the rq is being passed around | |
1990 | * RT_PUSH_IPI_RESTART: The priority of the running task for the rq | |
1991 | * has changed, and the IPI should restart | |
1992 | * circulating the overloaded CPUs again. | |
1993 | * | |
1994 | * rq->rt.push_cpu contains the CPU that is being sent the IPI. It is updated | |
1995 | * before sending to the next CPU. | |
1996 | * | |
1997 | * Instead of having all CPUs that schedule a lower priority task send | |
1998 | * an IPI to the same "first" CPU in the RT overload mask, they send it | |
1999 | * to the next overloaded CPU after their own CPU. This helps distribute | |
2000 | * the work when there's more than one overloaded CPU and multiple CPUs | |
2001 | * scheduling in lower priority tasks. | |
2002 | * | |
2003 | * When a rq schedules a lower priority task than what was currently | |
2004 | * running, the next CPU with overloaded RT tasks is examined first. | |
2005 | * That is, if CPU 1 and 5 are overloaded, and CPU 3 schedules a lower | |
2006 | * priority task, it will send an IPI first to CPU 5, then CPU 5 will | |
2007 | * send to CPU 1 if it is still overloaded. CPU 1 will clear the | |
2008 | * rq->rt.push_flags if RT_PUSH_IPI_RESTART is not set. | |
2009 | * | |
2010 | * The first CPU to notice IPI_RESTART is set, will clear that flag and then | |
2011 | * send an IPI to the next overloaded CPU after the rq->cpu and not the next | |
2012 | * CPU after push_cpu. That is, if CPU 1, 4 and 5 are overloaded when CPU 3 | |
2013 | * schedules a lower priority task, and the IPI_RESTART gets set while the | |
2014 | * handling is being done on CPU 5, it will clear the flag and send it back to | |
2015 | * CPU 4 instead of CPU 1. | |
2016 | * | |
2017 | * Note, the above logic can be disabled by turning off the sched_feature | |
2018 | * RT_PUSH_IPI. Then the rq lock of the overloaded CPU will simply be | |
2019 | * taken by the CPU requesting a pull and the waiting RT task will be pulled | |
2020 | * by that CPU. This may be fine for machines with few CPUs. | |
2021 | */ | |
b6366f04 SR |
2022 | static void tell_cpu_to_push(struct rq *rq) |
2023 | { | |
2024 | int cpu; | |
2025 | ||
2026 | if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) { | |
2027 | raw_spin_lock(&rq->rt.push_lock); | |
2028 | /* Make sure it's still executing */ | |
2029 | if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) { | |
2030 | /* | |
2031 | * Tell the IPI to restart the loop as things have | |
2032 | * changed since it started. | |
2033 | */ | |
2034 | rq->rt.push_flags |= RT_PUSH_IPI_RESTART; | |
2035 | raw_spin_unlock(&rq->rt.push_lock); | |
2036 | return; | |
2037 | } | |
2038 | raw_spin_unlock(&rq->rt.push_lock); | |
2039 | } | |
2040 | ||
2041 | /* When here, there's no IPI going around */ | |
2042 | ||
2043 | rq->rt.push_cpu = rq->cpu; | |
2044 | cpu = find_next_push_cpu(rq); | |
2045 | if (cpu >= nr_cpu_ids) | |
2046 | return; | |
2047 | ||
2048 | rq->rt.push_flags = RT_PUSH_IPI_EXECUTING; | |
2049 | ||
2050 | irq_work_queue_on(&rq->rt.push_work, cpu); | |
2051 | } | |
2052 | ||
2053 | /* Called from hardirq context */ | |
2054 | static void try_to_push_tasks(void *arg) | |
2055 | { | |
2056 | struct rt_rq *rt_rq = arg; | |
2057 | struct rq *rq, *src_rq; | |
2058 | int this_cpu; | |
2059 | int cpu; | |
2060 | ||
2061 | this_cpu = rt_rq->push_cpu; | |
2062 | ||
2063 | /* Paranoid check */ | |
2064 | BUG_ON(this_cpu != smp_processor_id()); | |
2065 | ||
2066 | rq = cpu_rq(this_cpu); | |
2067 | src_rq = rq_of_rt_rq(rt_rq); | |
2068 | ||
2069 | again: | |
2070 | if (has_pushable_tasks(rq)) { | |
2071 | raw_spin_lock(&rq->lock); | |
2072 | push_rt_task(rq); | |
2073 | raw_spin_unlock(&rq->lock); | |
2074 | } | |
2075 | ||
2076 | /* Pass the IPI to the next rt overloaded queue */ | |
2077 | raw_spin_lock(&rt_rq->push_lock); | |
2078 | /* | |
2079 | * If the source queue changed since the IPI went out, | |
2080 | * we need to restart the search from that CPU again. | |
2081 | */ | |
2082 | if (rt_rq->push_flags & RT_PUSH_IPI_RESTART) { | |
2083 | rt_rq->push_flags &= ~RT_PUSH_IPI_RESTART; | |
2084 | rt_rq->push_cpu = src_rq->cpu; | |
2085 | } | |
2086 | ||
2087 | cpu = find_next_push_cpu(src_rq); | |
2088 | ||
2089 | if (cpu >= nr_cpu_ids) | |
2090 | rt_rq->push_flags &= ~RT_PUSH_IPI_EXECUTING; | |
2091 | raw_spin_unlock(&rt_rq->push_lock); | |
2092 | ||
2093 | if (cpu >= nr_cpu_ids) | |
2094 | return; | |
2095 | ||
2096 | /* | |
2097 | * It is possible that a restart caused this CPU to be | |
2098 | * chosen again. Don't bother with an IPI, just see if we | |
2099 | * have more to push. | |
2100 | */ | |
2101 | if (unlikely(cpu == rq->cpu)) | |
2102 | goto again; | |
2103 | ||
2104 | /* Try the next RT overloaded CPU */ | |
2105 | irq_work_queue_on(&rt_rq->push_work, cpu); | |
2106 | } | |
2107 | ||
2108 | static void push_irq_work_func(struct irq_work *work) | |
2109 | { | |
2110 | struct rt_rq *rt_rq = container_of(work, struct rt_rq, push_work); | |
2111 | ||
2112 | try_to_push_tasks(rt_rq); | |
2113 | } | |
2114 | #endif /* HAVE_RT_PUSH_IPI */ | |
2115 | ||
8046d680 | 2116 | static void pull_rt_task(struct rq *this_rq) |
f65eda4f | 2117 | { |
8046d680 PZ |
2118 | int this_cpu = this_rq->cpu, cpu; |
2119 | bool resched = false; | |
a8728944 | 2120 | struct task_struct *p; |
f65eda4f | 2121 | struct rq *src_rq; |
f65eda4f | 2122 | |
637f5085 | 2123 | if (likely(!rt_overloaded(this_rq))) |
8046d680 | 2124 | return; |
f65eda4f | 2125 | |
7c3f2ab7 PZ |
2126 | /* |
2127 | * Match the barrier from rt_set_overloaded; this guarantees that if we | |
2128 | * see overloaded we must also see the rto_mask bit. | |
2129 | */ | |
2130 | smp_rmb(); | |
2131 | ||
b6366f04 SR |
2132 | #ifdef HAVE_RT_PUSH_IPI |
2133 | if (sched_feat(RT_PUSH_IPI)) { | |
2134 | tell_cpu_to_push(this_rq); | |
8046d680 | 2135 | return; |
b6366f04 SR |
2136 | } |
2137 | #endif | |
2138 | ||
c6c4927b | 2139 | for_each_cpu(cpu, this_rq->rd->rto_mask) { |
f65eda4f SR |
2140 | if (this_cpu == cpu) |
2141 | continue; | |
2142 | ||
2143 | src_rq = cpu_rq(cpu); | |
74ab8e4f GH |
2144 | |
2145 | /* | |
2146 | * Don't bother taking the src_rq->lock if the next highest | |
2147 | * task is known to be lower-priority than our current task. | |
2148 | * This may look racy, but if this value is about to go | |
2149 | * logically higher, the src_rq will push this task away. | |
2150 | * And if its going logically lower, we do not care | |
2151 | */ | |
2152 | if (src_rq->rt.highest_prio.next >= | |
2153 | this_rq->rt.highest_prio.curr) | |
2154 | continue; | |
2155 | ||
f65eda4f SR |
2156 | /* |
2157 | * We can potentially drop this_rq's lock in | |
2158 | * double_lock_balance, and another CPU could | |
a8728944 | 2159 | * alter this_rq |
f65eda4f | 2160 | */ |
a8728944 | 2161 | double_lock_balance(this_rq, src_rq); |
f65eda4f SR |
2162 | |
2163 | /* | |
e23ee747 KT |
2164 | * We can pull only a task, which is pushable |
2165 | * on its rq, and no others. | |
f65eda4f | 2166 | */ |
e23ee747 | 2167 | p = pick_highest_pushable_task(src_rq, this_cpu); |
f65eda4f SR |
2168 | |
2169 | /* | |
2170 | * Do we have an RT task that preempts | |
2171 | * the to-be-scheduled task? | |
2172 | */ | |
a8728944 | 2173 | if (p && (p->prio < this_rq->rt.highest_prio.curr)) { |
f65eda4f | 2174 | WARN_ON(p == src_rq->curr); |
da0c1e65 | 2175 | WARN_ON(!task_on_rq_queued(p)); |
f65eda4f SR |
2176 | |
2177 | /* | |
2178 | * There's a chance that p is higher in priority | |
2179 | * than what's currently running on its cpu. | |
2180 | * This is just that p is wakeing up and hasn't | |
2181 | * had a chance to schedule. We only pull | |
2182 | * p if it is lower in priority than the | |
a8728944 | 2183 | * current task on the run queue |
f65eda4f | 2184 | */ |
a8728944 | 2185 | if (p->prio < src_rq->curr->prio) |
614ee1f6 | 2186 | goto skip; |
f65eda4f | 2187 | |
8046d680 | 2188 | resched = true; |
f65eda4f SR |
2189 | |
2190 | deactivate_task(src_rq, p, 0); | |
2191 | set_task_cpu(p, this_cpu); | |
2192 | activate_task(this_rq, p, 0); | |
2193 | /* | |
2194 | * We continue with the search, just in | |
2195 | * case there's an even higher prio task | |
25985edc | 2196 | * in another runqueue. (low likelihood |
f65eda4f | 2197 | * but possible) |
f65eda4f | 2198 | */ |
f65eda4f | 2199 | } |
49246274 | 2200 | skip: |
1b12bbc7 | 2201 | double_unlock_balance(this_rq, src_rq); |
f65eda4f SR |
2202 | } |
2203 | ||
8046d680 PZ |
2204 | if (resched) |
2205 | resched_curr(this_rq); | |
f65eda4f SR |
2206 | } |
2207 | ||
8ae121ac GH |
2208 | /* |
2209 | * If we are not running and we are not going to reschedule soon, we should | |
2210 | * try to push tasks away now | |
2211 | */ | |
efbbd05a | 2212 | static void task_woken_rt(struct rq *rq, struct task_struct *p) |
4642dafd | 2213 | { |
9a897c5a | 2214 | if (!task_running(rq, p) && |
8ae121ac | 2215 | !test_tsk_need_resched(rq->curr) && |
4b53a341 | 2216 | p->nr_cpus_allowed > 1 && |
1baca4ce | 2217 | (dl_task(rq->curr) || rt_task(rq->curr)) && |
4b53a341 | 2218 | (rq->curr->nr_cpus_allowed < 2 || |
3be209a8 | 2219 | rq->curr->prio <= p->prio)) |
4642dafd SR |
2220 | push_rt_tasks(rq); |
2221 | } | |
2222 | ||
bdd7c81b | 2223 | /* Assumes rq->lock is held */ |
1f11eb6a | 2224 | static void rq_online_rt(struct rq *rq) |
bdd7c81b IM |
2225 | { |
2226 | if (rq->rt.overloaded) | |
2227 | rt_set_overload(rq); | |
6e0534f2 | 2228 | |
7def2be1 PZ |
2229 | __enable_runtime(rq); |
2230 | ||
e864c499 | 2231 | cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr); |
bdd7c81b IM |
2232 | } |
2233 | ||
2234 | /* Assumes rq->lock is held */ | |
1f11eb6a | 2235 | static void rq_offline_rt(struct rq *rq) |
bdd7c81b IM |
2236 | { |
2237 | if (rq->rt.overloaded) | |
2238 | rt_clear_overload(rq); | |
6e0534f2 | 2239 | |
7def2be1 PZ |
2240 | __disable_runtime(rq); |
2241 | ||
6e0534f2 | 2242 | cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID); |
bdd7c81b | 2243 | } |
cb469845 SR |
2244 | |
2245 | /* | |
2246 | * When switch from the rt queue, we bring ourselves to a position | |
2247 | * that we might want to pull RT tasks from other runqueues. | |
2248 | */ | |
da7a735e | 2249 | static void switched_from_rt(struct rq *rq, struct task_struct *p) |
cb469845 SR |
2250 | { |
2251 | /* | |
2252 | * If there are other RT tasks then we will reschedule | |
2253 | * and the scheduling of the other RT tasks will handle | |
2254 | * the balancing. But if we are the last RT task | |
2255 | * we may need to handle the pulling of RT tasks | |
2256 | * now. | |
2257 | */ | |
da0c1e65 | 2258 | if (!task_on_rq_queued(p) || rq->rt.rt_nr_running) |
1158ddb5 KT |
2259 | return; |
2260 | ||
fd7a4bed | 2261 | queue_pull_task(rq); |
cb469845 | 2262 | } |
3d8cbdf8 | 2263 | |
11c785b7 | 2264 | void __init init_sched_rt_class(void) |
3d8cbdf8 RR |
2265 | { |
2266 | unsigned int i; | |
2267 | ||
029632fb | 2268 | for_each_possible_cpu(i) { |
eaa95840 | 2269 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i), |
6ca09dfc | 2270 | GFP_KERNEL, cpu_to_node(i)); |
029632fb | 2271 | } |
3d8cbdf8 | 2272 | } |
cb469845 SR |
2273 | #endif /* CONFIG_SMP */ |
2274 | ||
2275 | /* | |
2276 | * When switching a task to RT, we may overload the runqueue | |
2277 | * with RT tasks. In this case we try to push them off to | |
2278 | * other runqueues. | |
2279 | */ | |
da7a735e | 2280 | static void switched_to_rt(struct rq *rq, struct task_struct *p) |
cb469845 | 2281 | { |
cb469845 SR |
2282 | /* |
2283 | * If we are already running, then there's nothing | |
2284 | * that needs to be done. But if we are not running | |
2285 | * we may need to preempt the current running task. | |
2286 | * If that current running task is also an RT task | |
2287 | * then see if we can move to another run queue. | |
2288 | */ | |
da0c1e65 | 2289 | if (task_on_rq_queued(p) && rq->curr != p) { |
cb469845 | 2290 | #ifdef CONFIG_SMP |
4b53a341 | 2291 | if (p->nr_cpus_allowed > 1 && rq->rt.overloaded) |
fd7a4bed | 2292 | queue_push_tasks(rq); |
619bd4a7 | 2293 | #endif /* CONFIG_SMP */ |
fd7a4bed | 2294 | if (p->prio < rq->curr->prio) |
8875125e | 2295 | resched_curr(rq); |
cb469845 SR |
2296 | } |
2297 | } | |
2298 | ||
2299 | /* | |
2300 | * Priority of the task has changed. This may cause | |
2301 | * us to initiate a push or pull. | |
2302 | */ | |
da7a735e PZ |
2303 | static void |
2304 | prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 2305 | { |
da0c1e65 | 2306 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
2307 | return; |
2308 | ||
2309 | if (rq->curr == p) { | |
cb469845 SR |
2310 | #ifdef CONFIG_SMP |
2311 | /* | |
2312 | * If our priority decreases while running, we | |
2313 | * may need to pull tasks to this runqueue. | |
2314 | */ | |
2315 | if (oldprio < p->prio) | |
fd7a4bed PZ |
2316 | queue_pull_task(rq); |
2317 | ||
cb469845 SR |
2318 | /* |
2319 | * If there's a higher priority task waiting to run | |
fd7a4bed | 2320 | * then reschedule. |
cb469845 | 2321 | */ |
fd7a4bed | 2322 | if (p->prio > rq->rt.highest_prio.curr) |
8875125e | 2323 | resched_curr(rq); |
cb469845 SR |
2324 | #else |
2325 | /* For UP simply resched on drop of prio */ | |
2326 | if (oldprio < p->prio) | |
8875125e | 2327 | resched_curr(rq); |
e8fa1362 | 2328 | #endif /* CONFIG_SMP */ |
cb469845 SR |
2329 | } else { |
2330 | /* | |
2331 | * This task is not running, but if it is | |
2332 | * greater than the current running task | |
2333 | * then reschedule. | |
2334 | */ | |
2335 | if (p->prio < rq->curr->prio) | |
8875125e | 2336 | resched_curr(rq); |
cb469845 SR |
2337 | } |
2338 | } | |
2339 | ||
b18b6a9c | 2340 | #ifdef CONFIG_POSIX_TIMERS |
78f2c7db PZ |
2341 | static void watchdog(struct rq *rq, struct task_struct *p) |
2342 | { | |
2343 | unsigned long soft, hard; | |
2344 | ||
78d7d407 JS |
2345 | /* max may change after cur was read, this will be fixed next tick */ |
2346 | soft = task_rlimit(p, RLIMIT_RTTIME); | |
2347 | hard = task_rlimit_max(p, RLIMIT_RTTIME); | |
78f2c7db PZ |
2348 | |
2349 | if (soft != RLIM_INFINITY) { | |
2350 | unsigned long next; | |
2351 | ||
57d2aa00 YX |
2352 | if (p->rt.watchdog_stamp != jiffies) { |
2353 | p->rt.timeout++; | |
2354 | p->rt.watchdog_stamp = jiffies; | |
2355 | } | |
2356 | ||
78f2c7db | 2357 | next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ); |
5a52dd50 | 2358 | if (p->rt.timeout > next) |
f06febc9 | 2359 | p->cputime_expires.sched_exp = p->se.sum_exec_runtime; |
78f2c7db PZ |
2360 | } |
2361 | } | |
b18b6a9c NP |
2362 | #else |
2363 | static inline void watchdog(struct rq *rq, struct task_struct *p) { } | |
2364 | #endif | |
bb44e5d1 | 2365 | |
8f4d37ec | 2366 | static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued) |
bb44e5d1 | 2367 | { |
454c7999 CC |
2368 | struct sched_rt_entity *rt_se = &p->rt; |
2369 | ||
67e2be02 PZ |
2370 | update_curr_rt(rq); |
2371 | ||
78f2c7db PZ |
2372 | watchdog(rq, p); |
2373 | ||
bb44e5d1 IM |
2374 | /* |
2375 | * RR tasks need a special form of timeslice management. | |
2376 | * FIFO tasks have no timeslices. | |
2377 | */ | |
2378 | if (p->policy != SCHED_RR) | |
2379 | return; | |
2380 | ||
fa717060 | 2381 | if (--p->rt.time_slice) |
bb44e5d1 IM |
2382 | return; |
2383 | ||
ce0dbbbb | 2384 | p->rt.time_slice = sched_rr_timeslice; |
bb44e5d1 | 2385 | |
98fbc798 | 2386 | /* |
e9aa39bb LB |
2387 | * Requeue to the end of queue if we (and all of our ancestors) are not |
2388 | * the only element on the queue | |
98fbc798 | 2389 | */ |
454c7999 CC |
2390 | for_each_sched_rt_entity(rt_se) { |
2391 | if (rt_se->run_list.prev != rt_se->run_list.next) { | |
2392 | requeue_task_rt(rq, p, 0); | |
8aa6f0eb | 2393 | resched_curr(rq); |
454c7999 CC |
2394 | return; |
2395 | } | |
98fbc798 | 2396 | } |
bb44e5d1 IM |
2397 | } |
2398 | ||
83b699ed SV |
2399 | static void set_curr_task_rt(struct rq *rq) |
2400 | { | |
2401 | struct task_struct *p = rq->curr; | |
2402 | ||
78becc27 | 2403 | p->se.exec_start = rq_clock_task(rq); |
917b627d GH |
2404 | |
2405 | /* The running task is never eligible for pushing */ | |
2406 | dequeue_pushable_task(rq, p); | |
83b699ed SV |
2407 | } |
2408 | ||
6d686f45 | 2409 | static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task) |
0d721cea PW |
2410 | { |
2411 | /* | |
2412 | * Time slice is 0 for SCHED_FIFO tasks | |
2413 | */ | |
2414 | if (task->policy == SCHED_RR) | |
ce0dbbbb | 2415 | return sched_rr_timeslice; |
0d721cea PW |
2416 | else |
2417 | return 0; | |
2418 | } | |
2419 | ||
029632fb | 2420 | const struct sched_class rt_sched_class = { |
5522d5d5 | 2421 | .next = &fair_sched_class, |
bb44e5d1 IM |
2422 | .enqueue_task = enqueue_task_rt, |
2423 | .dequeue_task = dequeue_task_rt, | |
2424 | .yield_task = yield_task_rt, | |
2425 | ||
2426 | .check_preempt_curr = check_preempt_curr_rt, | |
2427 | ||
2428 | .pick_next_task = pick_next_task_rt, | |
2429 | .put_prev_task = put_prev_task_rt, | |
2430 | ||
681f3e68 | 2431 | #ifdef CONFIG_SMP |
4ce72a2c LZ |
2432 | .select_task_rq = select_task_rq_rt, |
2433 | ||
6c37067e | 2434 | .set_cpus_allowed = set_cpus_allowed_common, |
1f11eb6a GH |
2435 | .rq_online = rq_online_rt, |
2436 | .rq_offline = rq_offline_rt, | |
efbbd05a | 2437 | .task_woken = task_woken_rt, |
cb469845 | 2438 | .switched_from = switched_from_rt, |
681f3e68 | 2439 | #endif |
bb44e5d1 | 2440 | |
83b699ed | 2441 | .set_curr_task = set_curr_task_rt, |
bb44e5d1 | 2442 | .task_tick = task_tick_rt, |
cb469845 | 2443 | |
0d721cea PW |
2444 | .get_rr_interval = get_rr_interval_rt, |
2445 | ||
cb469845 SR |
2446 | .prio_changed = prio_changed_rt, |
2447 | .switched_to = switched_to_rt, | |
6e998916 SG |
2448 | |
2449 | .update_curr = update_curr_rt, | |
bb44e5d1 | 2450 | }; |
ada18de2 | 2451 | |
8887cd99 NP |
2452 | #ifdef CONFIG_RT_GROUP_SCHED |
2453 | /* | |
2454 | * Ensure that the real time constraints are schedulable. | |
2455 | */ | |
2456 | static DEFINE_MUTEX(rt_constraints_mutex); | |
2457 | ||
2458 | /* Must be called with tasklist_lock held */ | |
2459 | static inline int tg_has_rt_tasks(struct task_group *tg) | |
2460 | { | |
2461 | struct task_struct *g, *p; | |
2462 | ||
2463 | /* | |
2464 | * Autogroups do not have RT tasks; see autogroup_create(). | |
2465 | */ | |
2466 | if (task_group_is_autogroup(tg)) | |
2467 | return 0; | |
2468 | ||
2469 | for_each_process_thread(g, p) { | |
2470 | if (rt_task(p) && task_group(p) == tg) | |
2471 | return 1; | |
2472 | } | |
2473 | ||
2474 | return 0; | |
2475 | } | |
2476 | ||
2477 | struct rt_schedulable_data { | |
2478 | struct task_group *tg; | |
2479 | u64 rt_period; | |
2480 | u64 rt_runtime; | |
2481 | }; | |
2482 | ||
2483 | static int tg_rt_schedulable(struct task_group *tg, void *data) | |
2484 | { | |
2485 | struct rt_schedulable_data *d = data; | |
2486 | struct task_group *child; | |
2487 | unsigned long total, sum = 0; | |
2488 | u64 period, runtime; | |
2489 | ||
2490 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
2491 | runtime = tg->rt_bandwidth.rt_runtime; | |
2492 | ||
2493 | if (tg == d->tg) { | |
2494 | period = d->rt_period; | |
2495 | runtime = d->rt_runtime; | |
2496 | } | |
2497 | ||
2498 | /* | |
2499 | * Cannot have more runtime than the period. | |
2500 | */ | |
2501 | if (runtime > period && runtime != RUNTIME_INF) | |
2502 | return -EINVAL; | |
2503 | ||
2504 | /* | |
2505 | * Ensure we don't starve existing RT tasks. | |
2506 | */ | |
2507 | if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) | |
2508 | return -EBUSY; | |
2509 | ||
2510 | total = to_ratio(period, runtime); | |
2511 | ||
2512 | /* | |
2513 | * Nobody can have more than the global setting allows. | |
2514 | */ | |
2515 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) | |
2516 | return -EINVAL; | |
2517 | ||
2518 | /* | |
2519 | * The sum of our children's runtime should not exceed our own. | |
2520 | */ | |
2521 | list_for_each_entry_rcu(child, &tg->children, siblings) { | |
2522 | period = ktime_to_ns(child->rt_bandwidth.rt_period); | |
2523 | runtime = child->rt_bandwidth.rt_runtime; | |
2524 | ||
2525 | if (child == d->tg) { | |
2526 | period = d->rt_period; | |
2527 | runtime = d->rt_runtime; | |
2528 | } | |
2529 | ||
2530 | sum += to_ratio(period, runtime); | |
2531 | } | |
2532 | ||
2533 | if (sum > total) | |
2534 | return -EINVAL; | |
2535 | ||
2536 | return 0; | |
2537 | } | |
2538 | ||
2539 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | |
2540 | { | |
2541 | int ret; | |
2542 | ||
2543 | struct rt_schedulable_data data = { | |
2544 | .tg = tg, | |
2545 | .rt_period = period, | |
2546 | .rt_runtime = runtime, | |
2547 | }; | |
2548 | ||
2549 | rcu_read_lock(); | |
2550 | ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); | |
2551 | rcu_read_unlock(); | |
2552 | ||
2553 | return ret; | |
2554 | } | |
2555 | ||
2556 | static int tg_set_rt_bandwidth(struct task_group *tg, | |
2557 | u64 rt_period, u64 rt_runtime) | |
2558 | { | |
2559 | int i, err = 0; | |
2560 | ||
2561 | /* | |
2562 | * Disallowing the root group RT runtime is BAD, it would disallow the | |
2563 | * kernel creating (and or operating) RT threads. | |
2564 | */ | |
2565 | if (tg == &root_task_group && rt_runtime == 0) | |
2566 | return -EINVAL; | |
2567 | ||
2568 | /* No period doesn't make any sense. */ | |
2569 | if (rt_period == 0) | |
2570 | return -EINVAL; | |
2571 | ||
2572 | mutex_lock(&rt_constraints_mutex); | |
2573 | read_lock(&tasklist_lock); | |
2574 | err = __rt_schedulable(tg, rt_period, rt_runtime); | |
2575 | if (err) | |
2576 | goto unlock; | |
2577 | ||
2578 | raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
2579 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); | |
2580 | tg->rt_bandwidth.rt_runtime = rt_runtime; | |
2581 | ||
2582 | for_each_possible_cpu(i) { | |
2583 | struct rt_rq *rt_rq = tg->rt_rq[i]; | |
2584 | ||
2585 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
2586 | rt_rq->rt_runtime = rt_runtime; | |
2587 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
2588 | } | |
2589 | raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
2590 | unlock: | |
2591 | read_unlock(&tasklist_lock); | |
2592 | mutex_unlock(&rt_constraints_mutex); | |
2593 | ||
2594 | return err; | |
2595 | } | |
2596 | ||
2597 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) | |
2598 | { | |
2599 | u64 rt_runtime, rt_period; | |
2600 | ||
2601 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
2602 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | |
2603 | if (rt_runtime_us < 0) | |
2604 | rt_runtime = RUNTIME_INF; | |
2605 | ||
2606 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); | |
2607 | } | |
2608 | ||
2609 | long sched_group_rt_runtime(struct task_group *tg) | |
2610 | { | |
2611 | u64 rt_runtime_us; | |
2612 | ||
2613 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) | |
2614 | return -1; | |
2615 | ||
2616 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; | |
2617 | do_div(rt_runtime_us, NSEC_PER_USEC); | |
2618 | return rt_runtime_us; | |
2619 | } | |
2620 | ||
2621 | int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us) | |
2622 | { | |
2623 | u64 rt_runtime, rt_period; | |
2624 | ||
2625 | rt_period = rt_period_us * NSEC_PER_USEC; | |
2626 | rt_runtime = tg->rt_bandwidth.rt_runtime; | |
2627 | ||
2628 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); | |
2629 | } | |
2630 | ||
2631 | long sched_group_rt_period(struct task_group *tg) | |
2632 | { | |
2633 | u64 rt_period_us; | |
2634 | ||
2635 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
2636 | do_div(rt_period_us, NSEC_PER_USEC); | |
2637 | return rt_period_us; | |
2638 | } | |
2639 | ||
2640 | static int sched_rt_global_constraints(void) | |
2641 | { | |
2642 | int ret = 0; | |
2643 | ||
2644 | mutex_lock(&rt_constraints_mutex); | |
2645 | read_lock(&tasklist_lock); | |
2646 | ret = __rt_schedulable(NULL, 0, 0); | |
2647 | read_unlock(&tasklist_lock); | |
2648 | mutex_unlock(&rt_constraints_mutex); | |
2649 | ||
2650 | return ret; | |
2651 | } | |
2652 | ||
2653 | int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) | |
2654 | { | |
2655 | /* Don't accept realtime tasks when there is no way for them to run */ | |
2656 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) | |
2657 | return 0; | |
2658 | ||
2659 | return 1; | |
2660 | } | |
2661 | ||
2662 | #else /* !CONFIG_RT_GROUP_SCHED */ | |
2663 | static int sched_rt_global_constraints(void) | |
2664 | { | |
2665 | unsigned long flags; | |
2666 | int i; | |
2667 | ||
2668 | raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); | |
2669 | for_each_possible_cpu(i) { | |
2670 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | |
2671 | ||
2672 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
2673 | rt_rq->rt_runtime = global_rt_runtime(); | |
2674 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | |
2675 | } | |
2676 | raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); | |
2677 | ||
2678 | return 0; | |
2679 | } | |
2680 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
2681 | ||
2682 | static int sched_rt_global_validate(void) | |
2683 | { | |
2684 | if (sysctl_sched_rt_period <= 0) | |
2685 | return -EINVAL; | |
2686 | ||
2687 | if ((sysctl_sched_rt_runtime != RUNTIME_INF) && | |
2688 | (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) | |
2689 | return -EINVAL; | |
2690 | ||
2691 | return 0; | |
2692 | } | |
2693 | ||
2694 | static void sched_rt_do_global(void) | |
2695 | { | |
2696 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | |
2697 | def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); | |
2698 | } | |
2699 | ||
2700 | int sched_rt_handler(struct ctl_table *table, int write, | |
2701 | void __user *buffer, size_t *lenp, | |
2702 | loff_t *ppos) | |
2703 | { | |
2704 | int old_period, old_runtime; | |
2705 | static DEFINE_MUTEX(mutex); | |
2706 | int ret; | |
2707 | ||
2708 | mutex_lock(&mutex); | |
2709 | old_period = sysctl_sched_rt_period; | |
2710 | old_runtime = sysctl_sched_rt_runtime; | |
2711 | ||
2712 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
2713 | ||
2714 | if (!ret && write) { | |
2715 | ret = sched_rt_global_validate(); | |
2716 | if (ret) | |
2717 | goto undo; | |
2718 | ||
2719 | ret = sched_dl_global_validate(); | |
2720 | if (ret) | |
2721 | goto undo; | |
2722 | ||
2723 | ret = sched_rt_global_constraints(); | |
2724 | if (ret) | |
2725 | goto undo; | |
2726 | ||
2727 | sched_rt_do_global(); | |
2728 | sched_dl_do_global(); | |
2729 | } | |
2730 | if (0) { | |
2731 | undo: | |
2732 | sysctl_sched_rt_period = old_period; | |
2733 | sysctl_sched_rt_runtime = old_runtime; | |
2734 | } | |
2735 | mutex_unlock(&mutex); | |
2736 | ||
2737 | return ret; | |
2738 | } | |
2739 | ||
2740 | int sched_rr_handler(struct ctl_table *table, int write, | |
2741 | void __user *buffer, size_t *lenp, | |
2742 | loff_t *ppos) | |
2743 | { | |
2744 | int ret; | |
2745 | static DEFINE_MUTEX(mutex); | |
2746 | ||
2747 | mutex_lock(&mutex); | |
2748 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
2749 | /* | |
2750 | * Make sure that internally we keep jiffies. | |
2751 | * Also, writing zero resets the timeslice to default: | |
2752 | */ | |
2753 | if (!ret && write) { | |
2754 | sched_rr_timeslice = | |
2755 | sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE : | |
2756 | msecs_to_jiffies(sysctl_sched_rr_timeslice); | |
2757 | } | |
2758 | mutex_unlock(&mutex); | |
2759 | return ret; | |
2760 | } | |
2761 | ||
ada18de2 PZ |
2762 | #ifdef CONFIG_SCHED_DEBUG |
2763 | extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq); | |
2764 | ||
029632fb | 2765 | void print_rt_stats(struct seq_file *m, int cpu) |
ada18de2 | 2766 | { |
ec514c48 | 2767 | rt_rq_iter_t iter; |
ada18de2 PZ |
2768 | struct rt_rq *rt_rq; |
2769 | ||
2770 | rcu_read_lock(); | |
ec514c48 | 2771 | for_each_rt_rq(rt_rq, iter, cpu_rq(cpu)) |
ada18de2 PZ |
2772 | print_rt_rq(m, cpu, rt_rq); |
2773 | rcu_read_unlock(); | |
2774 | } | |
55e12e5e | 2775 | #endif /* CONFIG_SCHED_DEBUG */ |