Line data Source code
1 : // SPDX-License-Identifier: GPL-2.0
2 : /*
3 : * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
4 : * policies)
5 : */
6 :
7 : int sched_rr_timeslice = RR_TIMESLICE;
8 : int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE;
9 : /* More than 4 hours if BW_SHIFT equals 20. */
10 : static const u64 max_rt_runtime = MAX_BW;
11 :
12 : static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
13 :
14 : struct rt_bandwidth def_rt_bandwidth;
15 :
16 0 : static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
17 : {
18 0 : struct rt_bandwidth *rt_b =
19 0 : container_of(timer, struct rt_bandwidth, rt_period_timer);
20 0 : int idle = 0;
21 : int overrun;
22 :
23 0 : raw_spin_lock(&rt_b->rt_runtime_lock);
24 : for (;;) {
25 0 : overrun = hrtimer_forward_now(timer, rt_b->rt_period);
26 0 : if (!overrun)
27 : break;
28 :
29 0 : raw_spin_unlock(&rt_b->rt_runtime_lock);
30 0 : idle = do_sched_rt_period_timer(rt_b, overrun);
31 0 : raw_spin_lock(&rt_b->rt_runtime_lock);
32 : }
33 0 : if (idle)
34 0 : rt_b->rt_period_active = 0;
35 0 : raw_spin_unlock(&rt_b->rt_runtime_lock);
36 :
37 0 : return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
38 : }
39 :
40 1 : void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
41 : {
42 1 : rt_b->rt_period = ns_to_ktime(period);
43 1 : rt_b->rt_runtime = runtime;
44 :
45 : raw_spin_lock_init(&rt_b->rt_runtime_lock);
46 :
47 1 : hrtimer_init(&rt_b->rt_period_timer, CLOCK_MONOTONIC,
48 : HRTIMER_MODE_REL_HARD);
49 1 : rt_b->rt_period_timer.function = sched_rt_period_timer;
50 1 : }
51 :
52 0 : static inline void do_start_rt_bandwidth(struct rt_bandwidth *rt_b)
53 : {
54 0 : raw_spin_lock(&rt_b->rt_runtime_lock);
55 0 : if (!rt_b->rt_period_active) {
56 0 : rt_b->rt_period_active = 1;
57 : /*
58 : * SCHED_DEADLINE updates the bandwidth, as a run away
59 : * RT task with a DL task could hog a CPU. But DL does
60 : * not reset the period. If a deadline task was running
61 : * without an RT task running, it can cause RT tasks to
62 : * throttle when they start up. Kick the timer right away
63 : * to update the period.
64 : */
65 0 : hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
66 0 : hrtimer_start_expires(&rt_b->rt_period_timer,
67 : HRTIMER_MODE_ABS_PINNED_HARD);
68 : }
69 0 : raw_spin_unlock(&rt_b->rt_runtime_lock);
70 0 : }
71 :
72 : static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
73 : {
74 0 : if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
75 : return;
76 :
77 0 : do_start_rt_bandwidth(rt_b);
78 : }
79 :
80 1 : void init_rt_rq(struct rt_rq *rt_rq)
81 : {
82 : struct rt_prio_array *array;
83 : int i;
84 :
85 1 : array = &rt_rq->active;
86 101 : for (i = 0; i < MAX_RT_PRIO; i++) {
87 200 : INIT_LIST_HEAD(array->queue + i);
88 200 : __clear_bit(i, array->bitmap);
89 : }
90 : /* delimiter for bitsearch: */
91 2 : __set_bit(MAX_RT_PRIO, array->bitmap);
92 :
93 : #if defined CONFIG_SMP
94 : rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
95 : rt_rq->highest_prio.next = MAX_RT_PRIO-1;
96 : rt_rq->rt_nr_migratory = 0;
97 : rt_rq->overloaded = 0;
98 : plist_head_init(&rt_rq->pushable_tasks);
99 : #endif /* CONFIG_SMP */
100 : /* We start is dequeued state, because no RT tasks are queued */
101 1 : rt_rq->rt_queued = 0;
102 :
103 1 : rt_rq->rt_time = 0;
104 1 : rt_rq->rt_throttled = 0;
105 1 : rt_rq->rt_runtime = 0;
106 : raw_spin_lock_init(&rt_rq->rt_runtime_lock);
107 1 : }
108 :
109 : #ifdef CONFIG_RT_GROUP_SCHED
110 : static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
111 : {
112 : hrtimer_cancel(&rt_b->rt_period_timer);
113 : }
114 :
115 : #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
116 :
117 : static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
118 : {
119 : #ifdef CONFIG_SCHED_DEBUG
120 : WARN_ON_ONCE(!rt_entity_is_task(rt_se));
121 : #endif
122 : return container_of(rt_se, struct task_struct, rt);
123 : }
124 :
125 : static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
126 : {
127 : return rt_rq->rq;
128 : }
129 :
130 : static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
131 : {
132 : return rt_se->rt_rq;
133 : }
134 :
135 : static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
136 : {
137 : struct rt_rq *rt_rq = rt_se->rt_rq;
138 :
139 : return rt_rq->rq;
140 : }
141 :
142 : void unregister_rt_sched_group(struct task_group *tg)
143 : {
144 : if (tg->rt_se)
145 : destroy_rt_bandwidth(&tg->rt_bandwidth);
146 :
147 : }
148 :
149 : void free_rt_sched_group(struct task_group *tg)
150 : {
151 : int i;
152 :
153 : for_each_possible_cpu(i) {
154 : if (tg->rt_rq)
155 : kfree(tg->rt_rq[i]);
156 : if (tg->rt_se)
157 : kfree(tg->rt_se[i]);
158 : }
159 :
160 : kfree(tg->rt_rq);
161 : kfree(tg->rt_se);
162 : }
163 :
164 : void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
165 : struct sched_rt_entity *rt_se, int cpu,
166 : struct sched_rt_entity *parent)
167 : {
168 : struct rq *rq = cpu_rq(cpu);
169 :
170 : rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
171 : rt_rq->rt_nr_boosted = 0;
172 : rt_rq->rq = rq;
173 : rt_rq->tg = tg;
174 :
175 : tg->rt_rq[cpu] = rt_rq;
176 : tg->rt_se[cpu] = rt_se;
177 :
178 : if (!rt_se)
179 : return;
180 :
181 : if (!parent)
182 : rt_se->rt_rq = &rq->rt;
183 : else
184 : rt_se->rt_rq = parent->my_q;
185 :
186 : rt_se->my_q = rt_rq;
187 : rt_se->parent = parent;
188 : INIT_LIST_HEAD(&rt_se->run_list);
189 : }
190 :
191 : int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
192 : {
193 : struct rt_rq *rt_rq;
194 : struct sched_rt_entity *rt_se;
195 : int i;
196 :
197 : tg->rt_rq = kcalloc(nr_cpu_ids, sizeof(rt_rq), GFP_KERNEL);
198 : if (!tg->rt_rq)
199 : goto err;
200 : tg->rt_se = kcalloc(nr_cpu_ids, sizeof(rt_se), GFP_KERNEL);
201 : if (!tg->rt_se)
202 : goto err;
203 :
204 : init_rt_bandwidth(&tg->rt_bandwidth,
205 : ktime_to_ns(def_rt_bandwidth.rt_period), 0);
206 :
207 : for_each_possible_cpu(i) {
208 : rt_rq = kzalloc_node(sizeof(struct rt_rq),
209 : GFP_KERNEL, cpu_to_node(i));
210 : if (!rt_rq)
211 : goto err;
212 :
213 : rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
214 : GFP_KERNEL, cpu_to_node(i));
215 : if (!rt_se)
216 : goto err_free_rq;
217 :
218 : init_rt_rq(rt_rq);
219 : rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
220 : init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
221 : }
222 :
223 : return 1;
224 :
225 : err_free_rq:
226 : kfree(rt_rq);
227 : err:
228 : return 0;
229 : }
230 :
231 : #else /* CONFIG_RT_GROUP_SCHED */
232 :
233 : #define rt_entity_is_task(rt_se) (1)
234 :
235 : static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
236 : {
237 0 : return container_of(rt_se, struct task_struct, rt);
238 : }
239 :
240 : static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
241 : {
242 0 : return container_of(rt_rq, struct rq, rt);
243 : }
244 :
245 : static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
246 : {
247 0 : struct task_struct *p = rt_task_of(rt_se);
248 :
249 0 : return task_rq(p);
250 : }
251 :
252 : static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
253 : {
254 0 : struct rq *rq = rq_of_rt_se(rt_se);
255 :
256 : return &rq->rt;
257 : }
258 :
259 0 : void unregister_rt_sched_group(struct task_group *tg) { }
260 :
261 0 : void free_rt_sched_group(struct task_group *tg) { }
262 :
263 0 : int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
264 : {
265 0 : return 1;
266 : }
267 : #endif /* CONFIG_RT_GROUP_SCHED */
268 :
269 : #ifdef CONFIG_SMP
270 :
271 : static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
272 : {
273 : /* Try to pull RT tasks here if we lower this rq's prio */
274 : return rq->online && rq->rt.highest_prio.curr > prev->prio;
275 : }
276 :
277 : static inline int rt_overloaded(struct rq *rq)
278 : {
279 : return atomic_read(&rq->rd->rto_count);
280 : }
281 :
282 : static inline void rt_set_overload(struct rq *rq)
283 : {
284 : if (!rq->online)
285 : return;
286 :
287 : cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
288 : /*
289 : * Make sure the mask is visible before we set
290 : * the overload count. That is checked to determine
291 : * if we should look at the mask. It would be a shame
292 : * if we looked at the mask, but the mask was not
293 : * updated yet.
294 : *
295 : * Matched by the barrier in pull_rt_task().
296 : */
297 : smp_wmb();
298 : atomic_inc(&rq->rd->rto_count);
299 : }
300 :
301 : static inline void rt_clear_overload(struct rq *rq)
302 : {
303 : if (!rq->online)
304 : return;
305 :
306 : /* the order here really doesn't matter */
307 : atomic_dec(&rq->rd->rto_count);
308 : cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
309 : }
310 :
311 : static void update_rt_migration(struct rt_rq *rt_rq)
312 : {
313 : if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
314 : if (!rt_rq->overloaded) {
315 : rt_set_overload(rq_of_rt_rq(rt_rq));
316 : rt_rq->overloaded = 1;
317 : }
318 : } else if (rt_rq->overloaded) {
319 : rt_clear_overload(rq_of_rt_rq(rt_rq));
320 : rt_rq->overloaded = 0;
321 : }
322 : }
323 :
324 : static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
325 : {
326 : struct task_struct *p;
327 :
328 : if (!rt_entity_is_task(rt_se))
329 : return;
330 :
331 : p = rt_task_of(rt_se);
332 : rt_rq = &rq_of_rt_rq(rt_rq)->rt;
333 :
334 : rt_rq->rt_nr_total++;
335 : if (p->nr_cpus_allowed > 1)
336 : rt_rq->rt_nr_migratory++;
337 :
338 : update_rt_migration(rt_rq);
339 : }
340 :
341 : static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
342 : {
343 : struct task_struct *p;
344 :
345 : if (!rt_entity_is_task(rt_se))
346 : return;
347 :
348 : p = rt_task_of(rt_se);
349 : rt_rq = &rq_of_rt_rq(rt_rq)->rt;
350 :
351 : rt_rq->rt_nr_total--;
352 : if (p->nr_cpus_allowed > 1)
353 : rt_rq->rt_nr_migratory--;
354 :
355 : update_rt_migration(rt_rq);
356 : }
357 :
358 : static inline int has_pushable_tasks(struct rq *rq)
359 : {
360 : return !plist_head_empty(&rq->rt.pushable_tasks);
361 : }
362 :
363 : static DEFINE_PER_CPU(struct callback_head, rt_push_head);
364 : static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
365 :
366 : static void push_rt_tasks(struct rq *);
367 : static void pull_rt_task(struct rq *);
368 :
369 : static inline void rt_queue_push_tasks(struct rq *rq)
370 : {
371 : if (!has_pushable_tasks(rq))
372 : return;
373 :
374 : queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
375 : }
376 :
377 : static inline void rt_queue_pull_task(struct rq *rq)
378 : {
379 : queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
380 : }
381 :
382 : static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
383 : {
384 : plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
385 : plist_node_init(&p->pushable_tasks, p->prio);
386 : plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
387 :
388 : /* Update the highest prio pushable task */
389 : if (p->prio < rq->rt.highest_prio.next)
390 : rq->rt.highest_prio.next = p->prio;
391 : }
392 :
393 : static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
394 : {
395 : plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
396 :
397 : /* Update the new highest prio pushable task */
398 : if (has_pushable_tasks(rq)) {
399 : p = plist_first_entry(&rq->rt.pushable_tasks,
400 : struct task_struct, pushable_tasks);
401 : rq->rt.highest_prio.next = p->prio;
402 : } else {
403 : rq->rt.highest_prio.next = MAX_RT_PRIO-1;
404 : }
405 : }
406 :
407 : #else
408 :
409 : static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
410 : {
411 : }
412 :
413 : static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
414 : {
415 : }
416 :
417 : static inline
418 : void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
419 : {
420 : }
421 :
422 : static inline
423 : void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
424 : {
425 : }
426 :
427 : static inline void rt_queue_push_tasks(struct rq *rq)
428 : {
429 : }
430 : #endif /* CONFIG_SMP */
431 :
432 : static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
433 : static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
434 :
435 : static inline int on_rt_rq(struct sched_rt_entity *rt_se)
436 : {
437 0 : return rt_se->on_rq;
438 : }
439 :
440 : #ifdef CONFIG_UCLAMP_TASK
441 : /*
442 : * Verify the fitness of task @p to run on @cpu taking into account the uclamp
443 : * settings.
444 : *
445 : * This check is only important for heterogeneous systems where uclamp_min value
446 : * is higher than the capacity of a @cpu. For non-heterogeneous system this
447 : * function will always return true.
448 : *
449 : * The function will return true if the capacity of the @cpu is >= the
450 : * uclamp_min and false otherwise.
451 : *
452 : * Note that uclamp_min will be clamped to uclamp_max if uclamp_min
453 : * > uclamp_max.
454 : */
455 : static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
456 : {
457 : unsigned int min_cap;
458 : unsigned int max_cap;
459 : unsigned int cpu_cap;
460 :
461 : /* Only heterogeneous systems can benefit from this check */
462 : if (!static_branch_unlikely(&sched_asym_cpucapacity))
463 : return true;
464 :
465 : min_cap = uclamp_eff_value(p, UCLAMP_MIN);
466 : max_cap = uclamp_eff_value(p, UCLAMP_MAX);
467 :
468 : cpu_cap = capacity_orig_of(cpu);
469 :
470 : return cpu_cap >= min(min_cap, max_cap);
471 : }
472 : #else
473 : static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
474 : {
475 : return true;
476 : }
477 : #endif
478 :
479 : #ifdef CONFIG_RT_GROUP_SCHED
480 :
481 : static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
482 : {
483 : if (!rt_rq->tg)
484 : return RUNTIME_INF;
485 :
486 : return rt_rq->rt_runtime;
487 : }
488 :
489 : static inline u64 sched_rt_period(struct rt_rq *rt_rq)
490 : {
491 : return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
492 : }
493 :
494 : typedef struct task_group *rt_rq_iter_t;
495 :
496 : static inline struct task_group *next_task_group(struct task_group *tg)
497 : {
498 : do {
499 : tg = list_entry_rcu(tg->list.next,
500 : typeof(struct task_group), list);
501 : } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
502 :
503 : if (&tg->list == &task_groups)
504 : tg = NULL;
505 :
506 : return tg;
507 : }
508 :
509 : #define for_each_rt_rq(rt_rq, iter, rq) \
510 : for (iter = container_of(&task_groups, typeof(*iter), list); \
511 : (iter = next_task_group(iter)) && \
512 : (rt_rq = iter->rt_rq[cpu_of(rq)]);)
513 :
514 : #define for_each_sched_rt_entity(rt_se) \
515 : for (; rt_se; rt_se = rt_se->parent)
516 :
517 : static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
518 : {
519 : return rt_se->my_q;
520 : }
521 :
522 : static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
523 : static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
524 :
525 : static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
526 : {
527 : struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
528 : struct rq *rq = rq_of_rt_rq(rt_rq);
529 : struct sched_rt_entity *rt_se;
530 :
531 : int cpu = cpu_of(rq);
532 :
533 : rt_se = rt_rq->tg->rt_se[cpu];
534 :
535 : if (rt_rq->rt_nr_running) {
536 : if (!rt_se)
537 : enqueue_top_rt_rq(rt_rq);
538 : else if (!on_rt_rq(rt_se))
539 : enqueue_rt_entity(rt_se, 0);
540 :
541 : if (rt_rq->highest_prio.curr < curr->prio)
542 : resched_curr(rq);
543 : }
544 : }
545 :
546 : static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
547 : {
548 : struct sched_rt_entity *rt_se;
549 : int cpu = cpu_of(rq_of_rt_rq(rt_rq));
550 :
551 : rt_se = rt_rq->tg->rt_se[cpu];
552 :
553 : if (!rt_se) {
554 : dequeue_top_rt_rq(rt_rq);
555 : /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
556 : cpufreq_update_util(rq_of_rt_rq(rt_rq), 0);
557 : }
558 : else if (on_rt_rq(rt_se))
559 : dequeue_rt_entity(rt_se, 0);
560 : }
561 :
562 : static inline int rt_rq_throttled(struct rt_rq *rt_rq)
563 : {
564 : return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
565 : }
566 :
567 : static int rt_se_boosted(struct sched_rt_entity *rt_se)
568 : {
569 : struct rt_rq *rt_rq = group_rt_rq(rt_se);
570 : struct task_struct *p;
571 :
572 : if (rt_rq)
573 : return !!rt_rq->rt_nr_boosted;
574 :
575 : p = rt_task_of(rt_se);
576 : return p->prio != p->normal_prio;
577 : }
578 :
579 : #ifdef CONFIG_SMP
580 : static inline const struct cpumask *sched_rt_period_mask(void)
581 : {
582 : return this_rq()->rd->span;
583 : }
584 : #else
585 : static inline const struct cpumask *sched_rt_period_mask(void)
586 : {
587 : return cpu_online_mask;
588 : }
589 : #endif
590 :
591 : static inline
592 : struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
593 : {
594 : return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
595 : }
596 :
597 : static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
598 : {
599 : return &rt_rq->tg->rt_bandwidth;
600 : }
601 :
602 : #else /* !CONFIG_RT_GROUP_SCHED */
603 :
604 : static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
605 : {
606 : return rt_rq->rt_runtime;
607 : }
608 :
609 : static inline u64 sched_rt_period(struct rt_rq *rt_rq)
610 : {
611 0 : return ktime_to_ns(def_rt_bandwidth.rt_period);
612 : }
613 :
614 : typedef struct rt_rq *rt_rq_iter_t;
615 :
616 : #define for_each_rt_rq(rt_rq, iter, rq) \
617 : for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
618 :
619 : #define for_each_sched_rt_entity(rt_se) \
620 : for (; rt_se; rt_se = NULL)
621 :
622 : static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
623 : {
624 : return NULL;
625 : }
626 :
627 0 : static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
628 : {
629 0 : struct rq *rq = rq_of_rt_rq(rt_rq);
630 :
631 0 : if (!rt_rq->rt_nr_running)
632 : return;
633 :
634 0 : enqueue_top_rt_rq(rt_rq);
635 0 : resched_curr(rq);
636 : }
637 :
638 : static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
639 : {
640 0 : dequeue_top_rt_rq(rt_rq);
641 : }
642 :
643 : static inline int rt_rq_throttled(struct rt_rq *rt_rq)
644 : {
645 : return rt_rq->rt_throttled;
646 : }
647 :
648 : static inline const struct cpumask *sched_rt_period_mask(void)
649 : {
650 : return cpu_online_mask;
651 : }
652 :
653 : static inline
654 : struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
655 : {
656 0 : return &cpu_rq(cpu)->rt;
657 : }
658 :
659 : static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
660 : {
661 : return &def_rt_bandwidth;
662 : }
663 :
664 : #endif /* CONFIG_RT_GROUP_SCHED */
665 :
666 0 : bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
667 : {
668 0 : struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
669 :
670 0 : return (hrtimer_active(&rt_b->rt_period_timer) ||
671 0 : rt_rq->rt_time < rt_b->rt_runtime);
672 : }
673 :
674 : #ifdef CONFIG_SMP
675 : /*
676 : * We ran out of runtime, see if we can borrow some from our neighbours.
677 : */
678 : static void do_balance_runtime(struct rt_rq *rt_rq)
679 : {
680 : struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
681 : struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
682 : int i, weight;
683 : u64 rt_period;
684 :
685 : weight = cpumask_weight(rd->span);
686 :
687 : raw_spin_lock(&rt_b->rt_runtime_lock);
688 : rt_period = ktime_to_ns(rt_b->rt_period);
689 : for_each_cpu(i, rd->span) {
690 : struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
691 : s64 diff;
692 :
693 : if (iter == rt_rq)
694 : continue;
695 :
696 : raw_spin_lock(&iter->rt_runtime_lock);
697 : /*
698 : * Either all rqs have inf runtime and there's nothing to steal
699 : * or __disable_runtime() below sets a specific rq to inf to
700 : * indicate its been disabled and disallow stealing.
701 : */
702 : if (iter->rt_runtime == RUNTIME_INF)
703 : goto next;
704 :
705 : /*
706 : * From runqueues with spare time, take 1/n part of their
707 : * spare time, but no more than our period.
708 : */
709 : diff = iter->rt_runtime - iter->rt_time;
710 : if (diff > 0) {
711 : diff = div_u64((u64)diff, weight);
712 : if (rt_rq->rt_runtime + diff > rt_period)
713 : diff = rt_period - rt_rq->rt_runtime;
714 : iter->rt_runtime -= diff;
715 : rt_rq->rt_runtime += diff;
716 : if (rt_rq->rt_runtime == rt_period) {
717 : raw_spin_unlock(&iter->rt_runtime_lock);
718 : break;
719 : }
720 : }
721 : next:
722 : raw_spin_unlock(&iter->rt_runtime_lock);
723 : }
724 : raw_spin_unlock(&rt_b->rt_runtime_lock);
725 : }
726 :
727 : /*
728 : * Ensure this RQ takes back all the runtime it lend to its neighbours.
729 : */
730 : static void __disable_runtime(struct rq *rq)
731 : {
732 : struct root_domain *rd = rq->rd;
733 : rt_rq_iter_t iter;
734 : struct rt_rq *rt_rq;
735 :
736 : if (unlikely(!scheduler_running))
737 : return;
738 :
739 : for_each_rt_rq(rt_rq, iter, rq) {
740 : struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
741 : s64 want;
742 : int i;
743 :
744 : raw_spin_lock(&rt_b->rt_runtime_lock);
745 : raw_spin_lock(&rt_rq->rt_runtime_lock);
746 : /*
747 : * Either we're all inf and nobody needs to borrow, or we're
748 : * already disabled and thus have nothing to do, or we have
749 : * exactly the right amount of runtime to take out.
750 : */
751 : if (rt_rq->rt_runtime == RUNTIME_INF ||
752 : rt_rq->rt_runtime == rt_b->rt_runtime)
753 : goto balanced;
754 : raw_spin_unlock(&rt_rq->rt_runtime_lock);
755 :
756 : /*
757 : * Calculate the difference between what we started out with
758 : * and what we current have, that's the amount of runtime
759 : * we lend and now have to reclaim.
760 : */
761 : want = rt_b->rt_runtime - rt_rq->rt_runtime;
762 :
763 : /*
764 : * Greedy reclaim, take back as much as we can.
765 : */
766 : for_each_cpu(i, rd->span) {
767 : struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
768 : s64 diff;
769 :
770 : /*
771 : * Can't reclaim from ourselves or disabled runqueues.
772 : */
773 : if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
774 : continue;
775 :
776 : raw_spin_lock(&iter->rt_runtime_lock);
777 : if (want > 0) {
778 : diff = min_t(s64, iter->rt_runtime, want);
779 : iter->rt_runtime -= diff;
780 : want -= diff;
781 : } else {
782 : iter->rt_runtime -= want;
783 : want -= want;
784 : }
785 : raw_spin_unlock(&iter->rt_runtime_lock);
786 :
787 : if (!want)
788 : break;
789 : }
790 :
791 : raw_spin_lock(&rt_rq->rt_runtime_lock);
792 : /*
793 : * We cannot be left wanting - that would mean some runtime
794 : * leaked out of the system.
795 : */
796 : BUG_ON(want);
797 : balanced:
798 : /*
799 : * Disable all the borrow logic by pretending we have inf
800 : * runtime - in which case borrowing doesn't make sense.
801 : */
802 : rt_rq->rt_runtime = RUNTIME_INF;
803 : rt_rq->rt_throttled = 0;
804 : raw_spin_unlock(&rt_rq->rt_runtime_lock);
805 : raw_spin_unlock(&rt_b->rt_runtime_lock);
806 :
807 : /* Make rt_rq available for pick_next_task() */
808 : sched_rt_rq_enqueue(rt_rq);
809 : }
810 : }
811 :
812 : static void __enable_runtime(struct rq *rq)
813 : {
814 : rt_rq_iter_t iter;
815 : struct rt_rq *rt_rq;
816 :
817 : if (unlikely(!scheduler_running))
818 : return;
819 :
820 : /*
821 : * Reset each runqueue's bandwidth settings
822 : */
823 : for_each_rt_rq(rt_rq, iter, rq) {
824 : struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
825 :
826 : raw_spin_lock(&rt_b->rt_runtime_lock);
827 : raw_spin_lock(&rt_rq->rt_runtime_lock);
828 : rt_rq->rt_runtime = rt_b->rt_runtime;
829 : rt_rq->rt_time = 0;
830 : rt_rq->rt_throttled = 0;
831 : raw_spin_unlock(&rt_rq->rt_runtime_lock);
832 : raw_spin_unlock(&rt_b->rt_runtime_lock);
833 : }
834 : }
835 :
836 : static void balance_runtime(struct rt_rq *rt_rq)
837 : {
838 : if (!sched_feat(RT_RUNTIME_SHARE))
839 : return;
840 :
841 : if (rt_rq->rt_time > rt_rq->rt_runtime) {
842 : raw_spin_unlock(&rt_rq->rt_runtime_lock);
843 : do_balance_runtime(rt_rq);
844 : raw_spin_lock(&rt_rq->rt_runtime_lock);
845 : }
846 : }
847 : #else /* !CONFIG_SMP */
848 : static inline void balance_runtime(struct rt_rq *rt_rq) {}
849 : #endif /* CONFIG_SMP */
850 :
851 0 : static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
852 : {
853 0 : int i, idle = 1, throttled = 0;
854 : const struct cpumask *span;
855 :
856 0 : span = sched_rt_period_mask();
857 : #ifdef CONFIG_RT_GROUP_SCHED
858 : /*
859 : * FIXME: isolated CPUs should really leave the root task group,
860 : * whether they are isolcpus or were isolated via cpusets, lest
861 : * the timer run on a CPU which does not service all runqueues,
862 : * potentially leaving other CPUs indefinitely throttled. If
863 : * isolation is really required, the user will turn the throttle
864 : * off to kill the perturbations it causes anyway. Meanwhile,
865 : * this maintains functionality for boot and/or troubleshooting.
866 : */
867 : if (rt_b == &root_task_group.rt_bandwidth)
868 : span = cpu_online_mask;
869 : #endif
870 0 : for_each_cpu(i, span) {
871 0 : int enqueue = 0;
872 0 : struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
873 0 : struct rq *rq = rq_of_rt_rq(rt_rq);
874 : int skip;
875 :
876 : /*
877 : * When span == cpu_online_mask, taking each rq->lock
878 : * can be time-consuming. Try to avoid it when possible.
879 : */
880 0 : raw_spin_lock(&rt_rq->rt_runtime_lock);
881 0 : if (!sched_feat(RT_RUNTIME_SHARE) && rt_rq->rt_runtime != RUNTIME_INF)
882 0 : rt_rq->rt_runtime = rt_b->rt_runtime;
883 0 : skip = !rt_rq->rt_time && !rt_rq->rt_nr_running;
884 0 : raw_spin_unlock(&rt_rq->rt_runtime_lock);
885 0 : if (skip)
886 0 : continue;
887 :
888 0 : raw_spin_rq_lock(rq);
889 0 : update_rq_clock(rq);
890 :
891 0 : if (rt_rq->rt_time) {
892 : u64 runtime;
893 :
894 0 : raw_spin_lock(&rt_rq->rt_runtime_lock);
895 0 : if (rt_rq->rt_throttled)
896 : balance_runtime(rt_rq);
897 0 : runtime = rt_rq->rt_runtime;
898 0 : rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
899 0 : if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
900 0 : rt_rq->rt_throttled = 0;
901 0 : enqueue = 1;
902 :
903 : /*
904 : * When we're idle and a woken (rt) task is
905 : * throttled check_preempt_curr() will set
906 : * skip_update and the time between the wakeup
907 : * and this unthrottle will get accounted as
908 : * 'runtime'.
909 : */
910 0 : if (rt_rq->rt_nr_running && rq->curr == rq->idle)
911 : rq_clock_cancel_skipupdate(rq);
912 : }
913 0 : if (rt_rq->rt_time || rt_rq->rt_nr_running)
914 0 : idle = 0;
915 0 : raw_spin_unlock(&rt_rq->rt_runtime_lock);
916 0 : } else if (rt_rq->rt_nr_running) {
917 0 : idle = 0;
918 0 : if (!rt_rq_throttled(rt_rq))
919 0 : enqueue = 1;
920 : }
921 0 : if (rt_rq->rt_throttled)
922 0 : throttled = 1;
923 :
924 0 : if (enqueue)
925 0 : sched_rt_rq_enqueue(rt_rq);
926 0 : raw_spin_rq_unlock(rq);
927 : }
928 :
929 0 : if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
930 : return 1;
931 :
932 : return idle;
933 : }
934 :
935 : static inline int rt_se_prio(struct sched_rt_entity *rt_se)
936 : {
937 : #ifdef CONFIG_RT_GROUP_SCHED
938 : struct rt_rq *rt_rq = group_rt_rq(rt_se);
939 :
940 : if (rt_rq)
941 : return rt_rq->highest_prio.curr;
942 : #endif
943 :
944 0 : return rt_task_of(rt_se)->prio;
945 : }
946 :
947 0 : static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
948 : {
949 0 : u64 runtime = sched_rt_runtime(rt_rq);
950 :
951 0 : if (rt_rq->rt_throttled)
952 : return rt_rq_throttled(rt_rq);
953 :
954 0 : if (runtime >= sched_rt_period(rt_rq))
955 : return 0;
956 :
957 0 : balance_runtime(rt_rq);
958 0 : runtime = sched_rt_runtime(rt_rq);
959 0 : if (runtime == RUNTIME_INF)
960 : return 0;
961 :
962 0 : if (rt_rq->rt_time > runtime) {
963 0 : struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
964 :
965 : /*
966 : * Don't actually throttle groups that have no runtime assigned
967 : * but accrue some time due to boosting.
968 : */
969 0 : if (likely(rt_b->rt_runtime)) {
970 0 : rt_rq->rt_throttled = 1;
971 0 : printk_deferred_once("sched: RT throttling activated\n");
972 : } else {
973 : /*
974 : * In case we did anyway, make it go away,
975 : * replenishment is a joke, since it will replenish us
976 : * with exactly 0 ns.
977 : */
978 0 : rt_rq->rt_time = 0;
979 : }
980 :
981 0 : if (rt_rq_throttled(rt_rq)) {
982 0 : sched_rt_rq_dequeue(rt_rq);
983 0 : return 1;
984 : }
985 : }
986 :
987 : return 0;
988 : }
989 :
990 : /*
991 : * Update the current task's runtime statistics. Skip current tasks that
992 : * are not in our scheduling class.
993 : */
994 0 : static void update_curr_rt(struct rq *rq)
995 : {
996 0 : struct task_struct *curr = rq->curr;
997 0 : struct sched_rt_entity *rt_se = &curr->rt;
998 : u64 delta_exec;
999 : u64 now;
1000 :
1001 0 : if (curr->sched_class != &rt_sched_class)
1002 : return;
1003 :
1004 0 : now = rq_clock_task(rq);
1005 0 : delta_exec = now - curr->se.exec_start;
1006 0 : if (unlikely((s64)delta_exec <= 0))
1007 : return;
1008 :
1009 : schedstat_set(curr->stats.exec_max,
1010 : max(curr->stats.exec_max, delta_exec));
1011 :
1012 0 : trace_sched_stat_runtime(curr, delta_exec, 0);
1013 :
1014 0 : curr->se.sum_exec_runtime += delta_exec;
1015 0 : account_group_exec_runtime(curr, delta_exec);
1016 :
1017 0 : curr->se.exec_start = now;
1018 0 : cgroup_account_cputime(curr, delta_exec);
1019 :
1020 0 : if (!rt_bandwidth_enabled())
1021 : return;
1022 :
1023 0 : for_each_sched_rt_entity(rt_se) {
1024 0 : struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1025 : int exceeded;
1026 :
1027 0 : if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
1028 0 : raw_spin_lock(&rt_rq->rt_runtime_lock);
1029 0 : rt_rq->rt_time += delta_exec;
1030 0 : exceeded = sched_rt_runtime_exceeded(rt_rq);
1031 0 : if (exceeded)
1032 0 : resched_curr(rq);
1033 0 : raw_spin_unlock(&rt_rq->rt_runtime_lock);
1034 0 : if (exceeded)
1035 0 : do_start_rt_bandwidth(sched_rt_bandwidth(rt_rq));
1036 : }
1037 : }
1038 : }
1039 :
1040 : static void
1041 0 : dequeue_top_rt_rq(struct rt_rq *rt_rq)
1042 : {
1043 0 : struct rq *rq = rq_of_rt_rq(rt_rq);
1044 :
1045 0 : BUG_ON(&rq->rt != rt_rq);
1046 :
1047 0 : if (!rt_rq->rt_queued)
1048 : return;
1049 :
1050 0 : BUG_ON(!rq->nr_running);
1051 :
1052 0 : sub_nr_running(rq, rt_rq->rt_nr_running);
1053 0 : rt_rq->rt_queued = 0;
1054 :
1055 : }
1056 :
1057 : static void
1058 0 : enqueue_top_rt_rq(struct rt_rq *rt_rq)
1059 : {
1060 0 : struct rq *rq = rq_of_rt_rq(rt_rq);
1061 :
1062 0 : BUG_ON(&rq->rt != rt_rq);
1063 :
1064 0 : if (rt_rq->rt_queued)
1065 : return;
1066 :
1067 0 : if (rt_rq_throttled(rt_rq))
1068 : return;
1069 :
1070 0 : if (rt_rq->rt_nr_running) {
1071 0 : add_nr_running(rq, rt_rq->rt_nr_running);
1072 0 : rt_rq->rt_queued = 1;
1073 : }
1074 :
1075 : /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
1076 : cpufreq_update_util(rq, 0);
1077 : }
1078 :
1079 : #if defined CONFIG_SMP
1080 :
1081 : static void
1082 : inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1083 : {
1084 : struct rq *rq = rq_of_rt_rq(rt_rq);
1085 :
1086 : #ifdef CONFIG_RT_GROUP_SCHED
1087 : /*
1088 : * Change rq's cpupri only if rt_rq is the top queue.
1089 : */
1090 : if (&rq->rt != rt_rq)
1091 : return;
1092 : #endif
1093 : if (rq->online && prio < prev_prio)
1094 : cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1095 : }
1096 :
1097 : static void
1098 : dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1099 : {
1100 : struct rq *rq = rq_of_rt_rq(rt_rq);
1101 :
1102 : #ifdef CONFIG_RT_GROUP_SCHED
1103 : /*
1104 : * Change rq's cpupri only if rt_rq is the top queue.
1105 : */
1106 : if (&rq->rt != rt_rq)
1107 : return;
1108 : #endif
1109 : if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1110 : cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1111 : }
1112 :
1113 : #else /* CONFIG_SMP */
1114 :
1115 : static inline
1116 : void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1117 : static inline
1118 : void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1119 :
1120 : #endif /* CONFIG_SMP */
1121 :
1122 : #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1123 : static void
1124 : inc_rt_prio(struct rt_rq *rt_rq, int prio)
1125 : {
1126 : int prev_prio = rt_rq->highest_prio.curr;
1127 :
1128 : if (prio < prev_prio)
1129 : rt_rq->highest_prio.curr = prio;
1130 :
1131 : inc_rt_prio_smp(rt_rq, prio, prev_prio);
1132 : }
1133 :
1134 : static void
1135 : dec_rt_prio(struct rt_rq *rt_rq, int prio)
1136 : {
1137 : int prev_prio = rt_rq->highest_prio.curr;
1138 :
1139 : if (rt_rq->rt_nr_running) {
1140 :
1141 : WARN_ON(prio < prev_prio);
1142 :
1143 : /*
1144 : * This may have been our highest task, and therefore
1145 : * we may have some recomputation to do
1146 : */
1147 : if (prio == prev_prio) {
1148 : struct rt_prio_array *array = &rt_rq->active;
1149 :
1150 : rt_rq->highest_prio.curr =
1151 : sched_find_first_bit(array->bitmap);
1152 : }
1153 :
1154 : } else {
1155 : rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
1156 : }
1157 :
1158 : dec_rt_prio_smp(rt_rq, prio, prev_prio);
1159 : }
1160 :
1161 : #else
1162 :
1163 : static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1164 : static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1165 :
1166 : #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1167 :
1168 : #ifdef CONFIG_RT_GROUP_SCHED
1169 :
1170 : static void
1171 : inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1172 : {
1173 : if (rt_se_boosted(rt_se))
1174 : rt_rq->rt_nr_boosted++;
1175 :
1176 : if (rt_rq->tg)
1177 : start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1178 : }
1179 :
1180 : static void
1181 : dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1182 : {
1183 : if (rt_se_boosted(rt_se))
1184 : rt_rq->rt_nr_boosted--;
1185 :
1186 : WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1187 : }
1188 :
1189 : #else /* CONFIG_RT_GROUP_SCHED */
1190 :
1191 : static void
1192 : inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1193 : {
1194 0 : start_rt_bandwidth(&def_rt_bandwidth);
1195 : }
1196 :
1197 : static inline
1198 : void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1199 :
1200 : #endif /* CONFIG_RT_GROUP_SCHED */
1201 :
1202 : static inline
1203 : unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1204 : {
1205 0 : struct rt_rq *group_rq = group_rt_rq(rt_se);
1206 :
1207 : if (group_rq)
1208 : return group_rq->rt_nr_running;
1209 : else
1210 : return 1;
1211 : }
1212 :
1213 : static inline
1214 : unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1215 : {
1216 0 : struct rt_rq *group_rq = group_rt_rq(rt_se);
1217 : struct task_struct *tsk;
1218 :
1219 : if (group_rq)
1220 : return group_rq->rr_nr_running;
1221 :
1222 0 : tsk = rt_task_of(rt_se);
1223 :
1224 0 : return (tsk->policy == SCHED_RR) ? 1 : 0;
1225 : }
1226 :
1227 : static inline
1228 0 : void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1229 : {
1230 0 : int prio = rt_se_prio(rt_se);
1231 :
1232 0 : WARN_ON(!rt_prio(prio));
1233 0 : rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1234 0 : rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1235 :
1236 0 : inc_rt_prio(rt_rq, prio);
1237 0 : inc_rt_migration(rt_se, rt_rq);
1238 0 : inc_rt_group(rt_se, rt_rq);
1239 0 : }
1240 :
1241 : static inline
1242 0 : void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1243 : {
1244 0 : WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1245 0 : WARN_ON(!rt_rq->rt_nr_running);
1246 0 : rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1247 0 : rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1248 :
1249 0 : dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1250 0 : dec_rt_migration(rt_se, rt_rq);
1251 0 : dec_rt_group(rt_se, rt_rq);
1252 0 : }
1253 :
1254 : /*
1255 : * Change rt_se->run_list location unless SAVE && !MOVE
1256 : *
1257 : * assumes ENQUEUE/DEQUEUE flags match
1258 : */
1259 : static inline bool move_entity(unsigned int flags)
1260 : {
1261 0 : if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1262 : return false;
1263 :
1264 : return true;
1265 : }
1266 :
1267 : static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1268 : {
1269 0 : list_del_init(&rt_se->run_list);
1270 :
1271 0 : if (list_empty(array->queue + rt_se_prio(rt_se)))
1272 0 : __clear_bit(rt_se_prio(rt_se), array->bitmap);
1273 :
1274 0 : rt_se->on_list = 0;
1275 : }
1276 :
1277 : static inline struct sched_statistics *
1278 : __schedstats_from_rt_se(struct sched_rt_entity *rt_se)
1279 : {
1280 : #ifdef CONFIG_RT_GROUP_SCHED
1281 : /* schedstats is not supported for rt group. */
1282 : if (!rt_entity_is_task(rt_se))
1283 : return NULL;
1284 : #endif
1285 :
1286 : return &rt_task_of(rt_se)->stats;
1287 : }
1288 :
1289 : static inline void
1290 : update_stats_wait_start_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1291 : {
1292 : struct sched_statistics *stats;
1293 0 : struct task_struct *p = NULL;
1294 :
1295 : if (!schedstat_enabled())
1296 : return;
1297 :
1298 : if (rt_entity_is_task(rt_se))
1299 : p = rt_task_of(rt_se);
1300 :
1301 : stats = __schedstats_from_rt_se(rt_se);
1302 : if (!stats)
1303 : return;
1304 :
1305 : __update_stats_wait_start(rq_of_rt_rq(rt_rq), p, stats);
1306 : }
1307 :
1308 : static inline void
1309 : update_stats_enqueue_sleeper_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1310 : {
1311 : struct sched_statistics *stats;
1312 : struct task_struct *p = NULL;
1313 :
1314 : if (!schedstat_enabled())
1315 : return;
1316 :
1317 : if (rt_entity_is_task(rt_se))
1318 : p = rt_task_of(rt_se);
1319 :
1320 : stats = __schedstats_from_rt_se(rt_se);
1321 : if (!stats)
1322 : return;
1323 :
1324 : __update_stats_enqueue_sleeper(rq_of_rt_rq(rt_rq), p, stats);
1325 : }
1326 :
1327 : static inline void
1328 : update_stats_enqueue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
1329 : int flags)
1330 : {
1331 : if (!schedstat_enabled())
1332 : return;
1333 :
1334 : if (flags & ENQUEUE_WAKEUP)
1335 : update_stats_enqueue_sleeper_rt(rt_rq, rt_se);
1336 : }
1337 :
1338 : static inline void
1339 : update_stats_wait_end_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1340 : {
1341 : struct sched_statistics *stats;
1342 : struct task_struct *p = NULL;
1343 :
1344 : if (!schedstat_enabled())
1345 : return;
1346 :
1347 : if (rt_entity_is_task(rt_se))
1348 : p = rt_task_of(rt_se);
1349 :
1350 : stats = __schedstats_from_rt_se(rt_se);
1351 : if (!stats)
1352 : return;
1353 :
1354 : __update_stats_wait_end(rq_of_rt_rq(rt_rq), p, stats);
1355 : }
1356 :
1357 : static inline void
1358 : update_stats_dequeue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
1359 : int flags)
1360 : {
1361 0 : struct task_struct *p = NULL;
1362 :
1363 : if (!schedstat_enabled())
1364 : return;
1365 :
1366 : if (rt_entity_is_task(rt_se))
1367 : p = rt_task_of(rt_se);
1368 :
1369 : if ((flags & DEQUEUE_SLEEP) && p) {
1370 : unsigned int state;
1371 :
1372 : state = READ_ONCE(p->__state);
1373 : if (state & TASK_INTERRUPTIBLE)
1374 : __schedstat_set(p->stats.sleep_start,
1375 : rq_clock(rq_of_rt_rq(rt_rq)));
1376 :
1377 : if (state & TASK_UNINTERRUPTIBLE)
1378 : __schedstat_set(p->stats.block_start,
1379 : rq_clock(rq_of_rt_rq(rt_rq)));
1380 : }
1381 : }
1382 :
1383 0 : static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1384 : {
1385 0 : struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1386 0 : struct rt_prio_array *array = &rt_rq->active;
1387 0 : struct rt_rq *group_rq = group_rt_rq(rt_se);
1388 0 : struct list_head *queue = array->queue + rt_se_prio(rt_se);
1389 :
1390 : /*
1391 : * Don't enqueue the group if its throttled, or when empty.
1392 : * The latter is a consequence of the former when a child group
1393 : * get throttled and the current group doesn't have any other
1394 : * active members.
1395 : */
1396 : if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1397 : if (rt_se->on_list)
1398 : __delist_rt_entity(rt_se, array);
1399 : return;
1400 : }
1401 :
1402 0 : if (move_entity(flags)) {
1403 0 : WARN_ON_ONCE(rt_se->on_list);
1404 0 : if (flags & ENQUEUE_HEAD)
1405 0 : list_add(&rt_se->run_list, queue);
1406 : else
1407 0 : list_add_tail(&rt_se->run_list, queue);
1408 :
1409 0 : __set_bit(rt_se_prio(rt_se), array->bitmap);
1410 0 : rt_se->on_list = 1;
1411 : }
1412 0 : rt_se->on_rq = 1;
1413 :
1414 0 : inc_rt_tasks(rt_se, rt_rq);
1415 : }
1416 :
1417 0 : static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1418 : {
1419 0 : struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1420 0 : struct rt_prio_array *array = &rt_rq->active;
1421 :
1422 0 : if (move_entity(flags)) {
1423 0 : WARN_ON_ONCE(!rt_se->on_list);
1424 : __delist_rt_entity(rt_se, array);
1425 : }
1426 0 : rt_se->on_rq = 0;
1427 :
1428 0 : dec_rt_tasks(rt_se, rt_rq);
1429 0 : }
1430 :
1431 : /*
1432 : * Because the prio of an upper entry depends on the lower
1433 : * entries, we must remove entries top - down.
1434 : */
1435 0 : static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1436 : {
1437 0 : struct sched_rt_entity *back = NULL;
1438 :
1439 0 : for_each_sched_rt_entity(rt_se) {
1440 0 : rt_se->back = back;
1441 0 : back = rt_se;
1442 : }
1443 :
1444 0 : dequeue_top_rt_rq(rt_rq_of_se(back));
1445 :
1446 0 : for (rt_se = back; rt_se; rt_se = rt_se->back) {
1447 0 : if (on_rt_rq(rt_se))
1448 0 : __dequeue_rt_entity(rt_se, flags);
1449 : }
1450 0 : }
1451 :
1452 0 : static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1453 : {
1454 0 : struct rq *rq = rq_of_rt_se(rt_se);
1455 :
1456 0 : update_stats_enqueue_rt(rt_rq_of_se(rt_se), rt_se, flags);
1457 :
1458 0 : dequeue_rt_stack(rt_se, flags);
1459 0 : for_each_sched_rt_entity(rt_se)
1460 0 : __enqueue_rt_entity(rt_se, flags);
1461 0 : enqueue_top_rt_rq(&rq->rt);
1462 0 : }
1463 :
1464 : static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1465 : {
1466 0 : struct rq *rq = rq_of_rt_se(rt_se);
1467 :
1468 0 : update_stats_dequeue_rt(rt_rq_of_se(rt_se), rt_se, flags);
1469 :
1470 0 : dequeue_rt_stack(rt_se, flags);
1471 :
1472 0 : for_each_sched_rt_entity(rt_se) {
1473 : struct rt_rq *rt_rq = group_rt_rq(rt_se);
1474 :
1475 : if (rt_rq && rt_rq->rt_nr_running)
1476 : __enqueue_rt_entity(rt_se, flags);
1477 : }
1478 0 : enqueue_top_rt_rq(&rq->rt);
1479 : }
1480 :
1481 : /*
1482 : * Adding/removing a task to/from a priority array:
1483 : */
1484 : static void
1485 0 : enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1486 : {
1487 0 : struct sched_rt_entity *rt_se = &p->rt;
1488 :
1489 0 : if (flags & ENQUEUE_WAKEUP)
1490 0 : rt_se->timeout = 0;
1491 :
1492 : check_schedstat_required();
1493 0 : update_stats_wait_start_rt(rt_rq_of_se(rt_se), rt_se);
1494 :
1495 0 : enqueue_rt_entity(rt_se, flags);
1496 :
1497 0 : if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1498 : enqueue_pushable_task(rq, p);
1499 0 : }
1500 :
1501 0 : static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1502 : {
1503 0 : struct sched_rt_entity *rt_se = &p->rt;
1504 :
1505 0 : update_curr_rt(rq);
1506 0 : dequeue_rt_entity(rt_se, flags);
1507 :
1508 0 : dequeue_pushable_task(rq, p);
1509 0 : }
1510 :
1511 : /*
1512 : * Put task to the head or the end of the run list without the overhead of
1513 : * dequeue followed by enqueue.
1514 : */
1515 : static void
1516 0 : requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1517 : {
1518 0 : if (on_rt_rq(rt_se)) {
1519 0 : struct rt_prio_array *array = &rt_rq->active;
1520 0 : struct list_head *queue = array->queue + rt_se_prio(rt_se);
1521 :
1522 0 : if (head)
1523 0 : list_move(&rt_se->run_list, queue);
1524 : else
1525 0 : list_move_tail(&rt_se->run_list, queue);
1526 : }
1527 0 : }
1528 :
1529 : static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1530 : {
1531 0 : struct sched_rt_entity *rt_se = &p->rt;
1532 : struct rt_rq *rt_rq;
1533 :
1534 0 : for_each_sched_rt_entity(rt_se) {
1535 0 : rt_rq = rt_rq_of_se(rt_se);
1536 0 : requeue_rt_entity(rt_rq, rt_se, head);
1537 : }
1538 : }
1539 :
1540 0 : static void yield_task_rt(struct rq *rq)
1541 : {
1542 0 : requeue_task_rt(rq, rq->curr, 0);
1543 0 : }
1544 :
1545 : #ifdef CONFIG_SMP
1546 : static int find_lowest_rq(struct task_struct *task);
1547 :
1548 : static int
1549 : select_task_rq_rt(struct task_struct *p, int cpu, int flags)
1550 : {
1551 : struct task_struct *curr;
1552 : struct rq *rq;
1553 : bool test;
1554 :
1555 : /* For anything but wake ups, just return the task_cpu */
1556 : if (!(flags & (WF_TTWU | WF_FORK)))
1557 : goto out;
1558 :
1559 : rq = cpu_rq(cpu);
1560 :
1561 : rcu_read_lock();
1562 : curr = READ_ONCE(rq->curr); /* unlocked access */
1563 :
1564 : /*
1565 : * If the current task on @p's runqueue is an RT task, then
1566 : * try to see if we can wake this RT task up on another
1567 : * runqueue. Otherwise simply start this RT task
1568 : * on its current runqueue.
1569 : *
1570 : * We want to avoid overloading runqueues. If the woken
1571 : * task is a higher priority, then it will stay on this CPU
1572 : * and the lower prio task should be moved to another CPU.
1573 : * Even though this will probably make the lower prio task
1574 : * lose its cache, we do not want to bounce a higher task
1575 : * around just because it gave up its CPU, perhaps for a
1576 : * lock?
1577 : *
1578 : * For equal prio tasks, we just let the scheduler sort it out.
1579 : *
1580 : * Otherwise, just let it ride on the affined RQ and the
1581 : * post-schedule router will push the preempted task away
1582 : *
1583 : * This test is optimistic, if we get it wrong the load-balancer
1584 : * will have to sort it out.
1585 : *
1586 : * We take into account the capacity of the CPU to ensure it fits the
1587 : * requirement of the task - which is only important on heterogeneous
1588 : * systems like big.LITTLE.
1589 : */
1590 : test = curr &&
1591 : unlikely(rt_task(curr)) &&
1592 : (curr->nr_cpus_allowed < 2 || curr->prio <= p->prio);
1593 :
1594 : if (test || !rt_task_fits_capacity(p, cpu)) {
1595 : int target = find_lowest_rq(p);
1596 :
1597 : /*
1598 : * Bail out if we were forcing a migration to find a better
1599 : * fitting CPU but our search failed.
1600 : */
1601 : if (!test && target != -1 && !rt_task_fits_capacity(p, target))
1602 : goto out_unlock;
1603 :
1604 : /*
1605 : * Don't bother moving it if the destination CPU is
1606 : * not running a lower priority task.
1607 : */
1608 : if (target != -1 &&
1609 : p->prio < cpu_rq(target)->rt.highest_prio.curr)
1610 : cpu = target;
1611 : }
1612 :
1613 : out_unlock:
1614 : rcu_read_unlock();
1615 :
1616 : out:
1617 : return cpu;
1618 : }
1619 :
1620 : static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1621 : {
1622 : /*
1623 : * Current can't be migrated, useless to reschedule,
1624 : * let's hope p can move out.
1625 : */
1626 : if (rq->curr->nr_cpus_allowed == 1 ||
1627 : !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1628 : return;
1629 :
1630 : /*
1631 : * p is migratable, so let's not schedule it and
1632 : * see if it is pushed or pulled somewhere else.
1633 : */
1634 : if (p->nr_cpus_allowed != 1 &&
1635 : cpupri_find(&rq->rd->cpupri, p, NULL))
1636 : return;
1637 :
1638 : /*
1639 : * There appear to be other CPUs that can accept
1640 : * the current task but none can run 'p', so lets reschedule
1641 : * to try and push the current task away:
1642 : */
1643 : requeue_task_rt(rq, p, 1);
1644 : resched_curr(rq);
1645 : }
1646 :
1647 : static int balance_rt(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1648 : {
1649 : if (!on_rt_rq(&p->rt) && need_pull_rt_task(rq, p)) {
1650 : /*
1651 : * This is OK, because current is on_cpu, which avoids it being
1652 : * picked for load-balance and preemption/IRQs are still
1653 : * disabled avoiding further scheduler activity on it and we've
1654 : * not yet started the picking loop.
1655 : */
1656 : rq_unpin_lock(rq, rf);
1657 : pull_rt_task(rq);
1658 : rq_repin_lock(rq, rf);
1659 : }
1660 :
1661 : return sched_stop_runnable(rq) || sched_dl_runnable(rq) || sched_rt_runnable(rq);
1662 : }
1663 : #endif /* CONFIG_SMP */
1664 :
1665 : /*
1666 : * Preempt the current task with a newly woken task if needed:
1667 : */
1668 0 : static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1669 : {
1670 0 : if (p->prio < rq->curr->prio) {
1671 0 : resched_curr(rq);
1672 0 : return;
1673 : }
1674 :
1675 : #ifdef CONFIG_SMP
1676 : /*
1677 : * If:
1678 : *
1679 : * - the newly woken task is of equal priority to the current task
1680 : * - the newly woken task is non-migratable while current is migratable
1681 : * - current will be preempted on the next reschedule
1682 : *
1683 : * we should check to see if current can readily move to a different
1684 : * cpu. If so, we will reschedule to allow the push logic to try
1685 : * to move current somewhere else, making room for our non-migratable
1686 : * task.
1687 : */
1688 : if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1689 : check_preempt_equal_prio(rq, p);
1690 : #endif
1691 : }
1692 :
1693 0 : static inline void set_next_task_rt(struct rq *rq, struct task_struct *p, bool first)
1694 : {
1695 0 : struct sched_rt_entity *rt_se = &p->rt;
1696 0 : struct rt_rq *rt_rq = &rq->rt;
1697 :
1698 0 : p->se.exec_start = rq_clock_task(rq);
1699 0 : if (on_rt_rq(&p->rt))
1700 : update_stats_wait_end_rt(rt_rq, rt_se);
1701 :
1702 : /* The running task is never eligible for pushing */
1703 0 : dequeue_pushable_task(rq, p);
1704 :
1705 0 : if (!first)
1706 : return;
1707 :
1708 : /*
1709 : * If prev task was rt, put_prev_task() has already updated the
1710 : * utilization. We only care of the case where we start to schedule a
1711 : * rt task
1712 : */
1713 0 : if (rq->curr->sched_class != &rt_sched_class)
1714 0 : update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1715 :
1716 : rt_queue_push_tasks(rq);
1717 : }
1718 :
1719 0 : static struct sched_rt_entity *pick_next_rt_entity(struct rt_rq *rt_rq)
1720 : {
1721 0 : struct rt_prio_array *array = &rt_rq->active;
1722 0 : struct sched_rt_entity *next = NULL;
1723 : struct list_head *queue;
1724 : int idx;
1725 :
1726 0 : idx = sched_find_first_bit(array->bitmap);
1727 0 : BUG_ON(idx >= MAX_RT_PRIO);
1728 :
1729 0 : queue = array->queue + idx;
1730 0 : next = list_entry(queue->next, struct sched_rt_entity, run_list);
1731 :
1732 0 : return next;
1733 : }
1734 :
1735 0 : static struct task_struct *_pick_next_task_rt(struct rq *rq)
1736 : {
1737 : struct sched_rt_entity *rt_se;
1738 0 : struct rt_rq *rt_rq = &rq->rt;
1739 :
1740 : do {
1741 0 : rt_se = pick_next_rt_entity(rt_rq);
1742 0 : BUG_ON(!rt_se);
1743 0 : rt_rq = group_rt_rq(rt_se);
1744 : } while (rt_rq);
1745 :
1746 0 : return rt_task_of(rt_se);
1747 : }
1748 :
1749 : static struct task_struct *pick_task_rt(struct rq *rq)
1750 : {
1751 : struct task_struct *p;
1752 :
1753 0 : if (!sched_rt_runnable(rq))
1754 : return NULL;
1755 :
1756 0 : p = _pick_next_task_rt(rq);
1757 :
1758 : return p;
1759 : }
1760 :
1761 0 : static struct task_struct *pick_next_task_rt(struct rq *rq)
1762 : {
1763 0 : struct task_struct *p = pick_task_rt(rq);
1764 :
1765 0 : if (p)
1766 0 : set_next_task_rt(rq, p, true);
1767 :
1768 0 : return p;
1769 : }
1770 :
1771 0 : static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1772 : {
1773 0 : struct sched_rt_entity *rt_se = &p->rt;
1774 0 : struct rt_rq *rt_rq = &rq->rt;
1775 :
1776 0 : if (on_rt_rq(&p->rt))
1777 : update_stats_wait_start_rt(rt_rq, rt_se);
1778 :
1779 0 : update_curr_rt(rq);
1780 :
1781 0 : update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1782 :
1783 : /*
1784 : * The previous task needs to be made eligible for pushing
1785 : * if it is still active
1786 : */
1787 0 : if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1788 : enqueue_pushable_task(rq, p);
1789 0 : }
1790 :
1791 : #ifdef CONFIG_SMP
1792 :
1793 : /* Only try algorithms three times */
1794 : #define RT_MAX_TRIES 3
1795 :
1796 : static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1797 : {
1798 : if (!task_running(rq, p) &&
1799 : cpumask_test_cpu(cpu, &p->cpus_mask))
1800 : return 1;
1801 :
1802 : return 0;
1803 : }
1804 :
1805 : /*
1806 : * Return the highest pushable rq's task, which is suitable to be executed
1807 : * on the CPU, NULL otherwise
1808 : */
1809 : static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1810 : {
1811 : struct plist_head *head = &rq->rt.pushable_tasks;
1812 : struct task_struct *p;
1813 :
1814 : if (!has_pushable_tasks(rq))
1815 : return NULL;
1816 :
1817 : plist_for_each_entry(p, head, pushable_tasks) {
1818 : if (pick_rt_task(rq, p, cpu))
1819 : return p;
1820 : }
1821 :
1822 : return NULL;
1823 : }
1824 :
1825 : static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1826 :
1827 : static int find_lowest_rq(struct task_struct *task)
1828 : {
1829 : struct sched_domain *sd;
1830 : struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1831 : int this_cpu = smp_processor_id();
1832 : int cpu = task_cpu(task);
1833 : int ret;
1834 :
1835 : /* Make sure the mask is initialized first */
1836 : if (unlikely(!lowest_mask))
1837 : return -1;
1838 :
1839 : if (task->nr_cpus_allowed == 1)
1840 : return -1; /* No other targets possible */
1841 :
1842 : /*
1843 : * If we're on asym system ensure we consider the different capacities
1844 : * of the CPUs when searching for the lowest_mask.
1845 : */
1846 : if (static_branch_unlikely(&sched_asym_cpucapacity)) {
1847 :
1848 : ret = cpupri_find_fitness(&task_rq(task)->rd->cpupri,
1849 : task, lowest_mask,
1850 : rt_task_fits_capacity);
1851 : } else {
1852 :
1853 : ret = cpupri_find(&task_rq(task)->rd->cpupri,
1854 : task, lowest_mask);
1855 : }
1856 :
1857 : if (!ret)
1858 : return -1; /* No targets found */
1859 :
1860 : /*
1861 : * At this point we have built a mask of CPUs representing the
1862 : * lowest priority tasks in the system. Now we want to elect
1863 : * the best one based on our affinity and topology.
1864 : *
1865 : * We prioritize the last CPU that the task executed on since
1866 : * it is most likely cache-hot in that location.
1867 : */
1868 : if (cpumask_test_cpu(cpu, lowest_mask))
1869 : return cpu;
1870 :
1871 : /*
1872 : * Otherwise, we consult the sched_domains span maps to figure
1873 : * out which CPU is logically closest to our hot cache data.
1874 : */
1875 : if (!cpumask_test_cpu(this_cpu, lowest_mask))
1876 : this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1877 :
1878 : rcu_read_lock();
1879 : for_each_domain(cpu, sd) {
1880 : if (sd->flags & SD_WAKE_AFFINE) {
1881 : int best_cpu;
1882 :
1883 : /*
1884 : * "this_cpu" is cheaper to preempt than a
1885 : * remote processor.
1886 : */
1887 : if (this_cpu != -1 &&
1888 : cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1889 : rcu_read_unlock();
1890 : return this_cpu;
1891 : }
1892 :
1893 : best_cpu = cpumask_any_and_distribute(lowest_mask,
1894 : sched_domain_span(sd));
1895 : if (best_cpu < nr_cpu_ids) {
1896 : rcu_read_unlock();
1897 : return best_cpu;
1898 : }
1899 : }
1900 : }
1901 : rcu_read_unlock();
1902 :
1903 : /*
1904 : * And finally, if there were no matches within the domains
1905 : * just give the caller *something* to work with from the compatible
1906 : * locations.
1907 : */
1908 : if (this_cpu != -1)
1909 : return this_cpu;
1910 :
1911 : cpu = cpumask_any_distribute(lowest_mask);
1912 : if (cpu < nr_cpu_ids)
1913 : return cpu;
1914 :
1915 : return -1;
1916 : }
1917 :
1918 : /* Will lock the rq it finds */
1919 : static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1920 : {
1921 : struct rq *lowest_rq = NULL;
1922 : int tries;
1923 : int cpu;
1924 :
1925 : for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1926 : cpu = find_lowest_rq(task);
1927 :
1928 : if ((cpu == -1) || (cpu == rq->cpu))
1929 : break;
1930 :
1931 : lowest_rq = cpu_rq(cpu);
1932 :
1933 : if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1934 : /*
1935 : * Target rq has tasks of equal or higher priority,
1936 : * retrying does not release any lock and is unlikely
1937 : * to yield a different result.
1938 : */
1939 : lowest_rq = NULL;
1940 : break;
1941 : }
1942 :
1943 : /* if the prio of this runqueue changed, try again */
1944 : if (double_lock_balance(rq, lowest_rq)) {
1945 : /*
1946 : * We had to unlock the run queue. In
1947 : * the mean time, task could have
1948 : * migrated already or had its affinity changed.
1949 : * Also make sure that it wasn't scheduled on its rq.
1950 : */
1951 : if (unlikely(task_rq(task) != rq ||
1952 : !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_mask) ||
1953 : task_running(rq, task) ||
1954 : !rt_task(task) ||
1955 : !task_on_rq_queued(task))) {
1956 :
1957 : double_unlock_balance(rq, lowest_rq);
1958 : lowest_rq = NULL;
1959 : break;
1960 : }
1961 : }
1962 :
1963 : /* If this rq is still suitable use it. */
1964 : if (lowest_rq->rt.highest_prio.curr > task->prio)
1965 : break;
1966 :
1967 : /* try again */
1968 : double_unlock_balance(rq, lowest_rq);
1969 : lowest_rq = NULL;
1970 : }
1971 :
1972 : return lowest_rq;
1973 : }
1974 :
1975 : static struct task_struct *pick_next_pushable_task(struct rq *rq)
1976 : {
1977 : struct task_struct *p;
1978 :
1979 : if (!has_pushable_tasks(rq))
1980 : return NULL;
1981 :
1982 : p = plist_first_entry(&rq->rt.pushable_tasks,
1983 : struct task_struct, pushable_tasks);
1984 :
1985 : BUG_ON(rq->cpu != task_cpu(p));
1986 : BUG_ON(task_current(rq, p));
1987 : BUG_ON(p->nr_cpus_allowed <= 1);
1988 :
1989 : BUG_ON(!task_on_rq_queued(p));
1990 : BUG_ON(!rt_task(p));
1991 :
1992 : return p;
1993 : }
1994 :
1995 : /*
1996 : * If the current CPU has more than one RT task, see if the non
1997 : * running task can migrate over to a CPU that is running a task
1998 : * of lesser priority.
1999 : */
2000 : static int push_rt_task(struct rq *rq, bool pull)
2001 : {
2002 : struct task_struct *next_task;
2003 : struct rq *lowest_rq;
2004 : int ret = 0;
2005 :
2006 : if (!rq->rt.overloaded)
2007 : return 0;
2008 :
2009 : next_task = pick_next_pushable_task(rq);
2010 : if (!next_task)
2011 : return 0;
2012 :
2013 : retry:
2014 : /*
2015 : * It's possible that the next_task slipped in of
2016 : * higher priority than current. If that's the case
2017 : * just reschedule current.
2018 : */
2019 : if (unlikely(next_task->prio < rq->curr->prio)) {
2020 : resched_curr(rq);
2021 : return 0;
2022 : }
2023 :
2024 : if (is_migration_disabled(next_task)) {
2025 : struct task_struct *push_task = NULL;
2026 : int cpu;
2027 :
2028 : if (!pull || rq->push_busy)
2029 : return 0;
2030 :
2031 : /*
2032 : * Invoking find_lowest_rq() on anything but an RT task doesn't
2033 : * make sense. Per the above priority check, curr has to
2034 : * be of higher priority than next_task, so no need to
2035 : * reschedule when bailing out.
2036 : *
2037 : * Note that the stoppers are masqueraded as SCHED_FIFO
2038 : * (cf. sched_set_stop_task()), so we can't rely on rt_task().
2039 : */
2040 : if (rq->curr->sched_class != &rt_sched_class)
2041 : return 0;
2042 :
2043 : cpu = find_lowest_rq(rq->curr);
2044 : if (cpu == -1 || cpu == rq->cpu)
2045 : return 0;
2046 :
2047 : /*
2048 : * Given we found a CPU with lower priority than @next_task,
2049 : * therefore it should be running. However we cannot migrate it
2050 : * to this other CPU, instead attempt to push the current
2051 : * running task on this CPU away.
2052 : */
2053 : push_task = get_push_task(rq);
2054 : if (push_task) {
2055 : raw_spin_rq_unlock(rq);
2056 : stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2057 : push_task, &rq->push_work);
2058 : raw_spin_rq_lock(rq);
2059 : }
2060 :
2061 : return 0;
2062 : }
2063 :
2064 : if (WARN_ON(next_task == rq->curr))
2065 : return 0;
2066 :
2067 : /* We might release rq lock */
2068 : get_task_struct(next_task);
2069 :
2070 : /* find_lock_lowest_rq locks the rq if found */
2071 : lowest_rq = find_lock_lowest_rq(next_task, rq);
2072 : if (!lowest_rq) {
2073 : struct task_struct *task;
2074 : /*
2075 : * find_lock_lowest_rq releases rq->lock
2076 : * so it is possible that next_task has migrated.
2077 : *
2078 : * We need to make sure that the task is still on the same
2079 : * run-queue and is also still the next task eligible for
2080 : * pushing.
2081 : */
2082 : task = pick_next_pushable_task(rq);
2083 : if (task == next_task) {
2084 : /*
2085 : * The task hasn't migrated, and is still the next
2086 : * eligible task, but we failed to find a run-queue
2087 : * to push it to. Do not retry in this case, since
2088 : * other CPUs will pull from us when ready.
2089 : */
2090 : goto out;
2091 : }
2092 :
2093 : if (!task)
2094 : /* No more tasks, just exit */
2095 : goto out;
2096 :
2097 : /*
2098 : * Something has shifted, try again.
2099 : */
2100 : put_task_struct(next_task);
2101 : next_task = task;
2102 : goto retry;
2103 : }
2104 :
2105 : deactivate_task(rq, next_task, 0);
2106 : set_task_cpu(next_task, lowest_rq->cpu);
2107 : activate_task(lowest_rq, next_task, 0);
2108 : resched_curr(lowest_rq);
2109 : ret = 1;
2110 :
2111 : double_unlock_balance(rq, lowest_rq);
2112 : out:
2113 : put_task_struct(next_task);
2114 :
2115 : return ret;
2116 : }
2117 :
2118 : static void push_rt_tasks(struct rq *rq)
2119 : {
2120 : /* push_rt_task will return true if it moved an RT */
2121 : while (push_rt_task(rq, false))
2122 : ;
2123 : }
2124 :
2125 : #ifdef HAVE_RT_PUSH_IPI
2126 :
2127 : /*
2128 : * When a high priority task schedules out from a CPU and a lower priority
2129 : * task is scheduled in, a check is made to see if there's any RT tasks
2130 : * on other CPUs that are waiting to run because a higher priority RT task
2131 : * is currently running on its CPU. In this case, the CPU with multiple RT
2132 : * tasks queued on it (overloaded) needs to be notified that a CPU has opened
2133 : * up that may be able to run one of its non-running queued RT tasks.
2134 : *
2135 : * All CPUs with overloaded RT tasks need to be notified as there is currently
2136 : * no way to know which of these CPUs have the highest priority task waiting
2137 : * to run. Instead of trying to take a spinlock on each of these CPUs,
2138 : * which has shown to cause large latency when done on machines with many
2139 : * CPUs, sending an IPI to the CPUs to have them push off the overloaded
2140 : * RT tasks waiting to run.
2141 : *
2142 : * Just sending an IPI to each of the CPUs is also an issue, as on large
2143 : * count CPU machines, this can cause an IPI storm on a CPU, especially
2144 : * if its the only CPU with multiple RT tasks queued, and a large number
2145 : * of CPUs scheduling a lower priority task at the same time.
2146 : *
2147 : * Each root domain has its own irq work function that can iterate over
2148 : * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
2149 : * task must be checked if there's one or many CPUs that are lowering
2150 : * their priority, there's a single irq work iterator that will try to
2151 : * push off RT tasks that are waiting to run.
2152 : *
2153 : * When a CPU schedules a lower priority task, it will kick off the
2154 : * irq work iterator that will jump to each CPU with overloaded RT tasks.
2155 : * As it only takes the first CPU that schedules a lower priority task
2156 : * to start the process, the rto_start variable is incremented and if
2157 : * the atomic result is one, then that CPU will try to take the rto_lock.
2158 : * This prevents high contention on the lock as the process handles all
2159 : * CPUs scheduling lower priority tasks.
2160 : *
2161 : * All CPUs that are scheduling a lower priority task will increment the
2162 : * rt_loop_next variable. This will make sure that the irq work iterator
2163 : * checks all RT overloaded CPUs whenever a CPU schedules a new lower
2164 : * priority task, even if the iterator is in the middle of a scan. Incrementing
2165 : * the rt_loop_next will cause the iterator to perform another scan.
2166 : *
2167 : */
2168 : static int rto_next_cpu(struct root_domain *rd)
2169 : {
2170 : int next;
2171 : int cpu;
2172 :
2173 : /*
2174 : * When starting the IPI RT pushing, the rto_cpu is set to -1,
2175 : * rt_next_cpu() will simply return the first CPU found in
2176 : * the rto_mask.
2177 : *
2178 : * If rto_next_cpu() is called with rto_cpu is a valid CPU, it
2179 : * will return the next CPU found in the rto_mask.
2180 : *
2181 : * If there are no more CPUs left in the rto_mask, then a check is made
2182 : * against rto_loop and rto_loop_next. rto_loop is only updated with
2183 : * the rto_lock held, but any CPU may increment the rto_loop_next
2184 : * without any locking.
2185 : */
2186 : for (;;) {
2187 :
2188 : /* When rto_cpu is -1 this acts like cpumask_first() */
2189 : cpu = cpumask_next(rd->rto_cpu, rd->rto_mask);
2190 :
2191 : rd->rto_cpu = cpu;
2192 :
2193 : if (cpu < nr_cpu_ids)
2194 : return cpu;
2195 :
2196 : rd->rto_cpu = -1;
2197 :
2198 : /*
2199 : * ACQUIRE ensures we see the @rto_mask changes
2200 : * made prior to the @next value observed.
2201 : *
2202 : * Matches WMB in rt_set_overload().
2203 : */
2204 : next = atomic_read_acquire(&rd->rto_loop_next);
2205 :
2206 : if (rd->rto_loop == next)
2207 : break;
2208 :
2209 : rd->rto_loop = next;
2210 : }
2211 :
2212 : return -1;
2213 : }
2214 :
2215 : static inline bool rto_start_trylock(atomic_t *v)
2216 : {
2217 : return !atomic_cmpxchg_acquire(v, 0, 1);
2218 : }
2219 :
2220 : static inline void rto_start_unlock(atomic_t *v)
2221 : {
2222 : atomic_set_release(v, 0);
2223 : }
2224 :
2225 : static void tell_cpu_to_push(struct rq *rq)
2226 : {
2227 : int cpu = -1;
2228 :
2229 : /* Keep the loop going if the IPI is currently active */
2230 : atomic_inc(&rq->rd->rto_loop_next);
2231 :
2232 : /* Only one CPU can initiate a loop at a time */
2233 : if (!rto_start_trylock(&rq->rd->rto_loop_start))
2234 : return;
2235 :
2236 : raw_spin_lock(&rq->rd->rto_lock);
2237 :
2238 : /*
2239 : * The rto_cpu is updated under the lock, if it has a valid CPU
2240 : * then the IPI is still running and will continue due to the
2241 : * update to loop_next, and nothing needs to be done here.
2242 : * Otherwise it is finishing up and an ipi needs to be sent.
2243 : */
2244 : if (rq->rd->rto_cpu < 0)
2245 : cpu = rto_next_cpu(rq->rd);
2246 :
2247 : raw_spin_unlock(&rq->rd->rto_lock);
2248 :
2249 : rto_start_unlock(&rq->rd->rto_loop_start);
2250 :
2251 : if (cpu >= 0) {
2252 : /* Make sure the rd does not get freed while pushing */
2253 : sched_get_rd(rq->rd);
2254 : irq_work_queue_on(&rq->rd->rto_push_work, cpu);
2255 : }
2256 : }
2257 :
2258 : /* Called from hardirq context */
2259 : void rto_push_irq_work_func(struct irq_work *work)
2260 : {
2261 : struct root_domain *rd =
2262 : container_of(work, struct root_domain, rto_push_work);
2263 : struct rq *rq;
2264 : int cpu;
2265 :
2266 : rq = this_rq();
2267 :
2268 : /*
2269 : * We do not need to grab the lock to check for has_pushable_tasks.
2270 : * When it gets updated, a check is made if a push is possible.
2271 : */
2272 : if (has_pushable_tasks(rq)) {
2273 : raw_spin_rq_lock(rq);
2274 : while (push_rt_task(rq, true))
2275 : ;
2276 : raw_spin_rq_unlock(rq);
2277 : }
2278 :
2279 : raw_spin_lock(&rd->rto_lock);
2280 :
2281 : /* Pass the IPI to the next rt overloaded queue */
2282 : cpu = rto_next_cpu(rd);
2283 :
2284 : raw_spin_unlock(&rd->rto_lock);
2285 :
2286 : if (cpu < 0) {
2287 : sched_put_rd(rd);
2288 : return;
2289 : }
2290 :
2291 : /* Try the next RT overloaded CPU */
2292 : irq_work_queue_on(&rd->rto_push_work, cpu);
2293 : }
2294 : #endif /* HAVE_RT_PUSH_IPI */
2295 :
2296 : static void pull_rt_task(struct rq *this_rq)
2297 : {
2298 : int this_cpu = this_rq->cpu, cpu;
2299 : bool resched = false;
2300 : struct task_struct *p, *push_task;
2301 : struct rq *src_rq;
2302 : int rt_overload_count = rt_overloaded(this_rq);
2303 :
2304 : if (likely(!rt_overload_count))
2305 : return;
2306 :
2307 : /*
2308 : * Match the barrier from rt_set_overloaded; this guarantees that if we
2309 : * see overloaded we must also see the rto_mask bit.
2310 : */
2311 : smp_rmb();
2312 :
2313 : /* If we are the only overloaded CPU do nothing */
2314 : if (rt_overload_count == 1 &&
2315 : cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask))
2316 : return;
2317 :
2318 : #ifdef HAVE_RT_PUSH_IPI
2319 : if (sched_feat(RT_PUSH_IPI)) {
2320 : tell_cpu_to_push(this_rq);
2321 : return;
2322 : }
2323 : #endif
2324 :
2325 : for_each_cpu(cpu, this_rq->rd->rto_mask) {
2326 : if (this_cpu == cpu)
2327 : continue;
2328 :
2329 : src_rq = cpu_rq(cpu);
2330 :
2331 : /*
2332 : * Don't bother taking the src_rq->lock if the next highest
2333 : * task is known to be lower-priority than our current task.
2334 : * This may look racy, but if this value is about to go
2335 : * logically higher, the src_rq will push this task away.
2336 : * And if its going logically lower, we do not care
2337 : */
2338 : if (src_rq->rt.highest_prio.next >=
2339 : this_rq->rt.highest_prio.curr)
2340 : continue;
2341 :
2342 : /*
2343 : * We can potentially drop this_rq's lock in
2344 : * double_lock_balance, and another CPU could
2345 : * alter this_rq
2346 : */
2347 : push_task = NULL;
2348 : double_lock_balance(this_rq, src_rq);
2349 :
2350 : /*
2351 : * We can pull only a task, which is pushable
2352 : * on its rq, and no others.
2353 : */
2354 : p = pick_highest_pushable_task(src_rq, this_cpu);
2355 :
2356 : /*
2357 : * Do we have an RT task that preempts
2358 : * the to-be-scheduled task?
2359 : */
2360 : if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2361 : WARN_ON(p == src_rq->curr);
2362 : WARN_ON(!task_on_rq_queued(p));
2363 :
2364 : /*
2365 : * There's a chance that p is higher in priority
2366 : * than what's currently running on its CPU.
2367 : * This is just that p is waking up and hasn't
2368 : * had a chance to schedule. We only pull
2369 : * p if it is lower in priority than the
2370 : * current task on the run queue
2371 : */
2372 : if (p->prio < src_rq->curr->prio)
2373 : goto skip;
2374 :
2375 : if (is_migration_disabled(p)) {
2376 : push_task = get_push_task(src_rq);
2377 : } else {
2378 : deactivate_task(src_rq, p, 0);
2379 : set_task_cpu(p, this_cpu);
2380 : activate_task(this_rq, p, 0);
2381 : resched = true;
2382 : }
2383 : /*
2384 : * We continue with the search, just in
2385 : * case there's an even higher prio task
2386 : * in another runqueue. (low likelihood
2387 : * but possible)
2388 : */
2389 : }
2390 : skip:
2391 : double_unlock_balance(this_rq, src_rq);
2392 :
2393 : if (push_task) {
2394 : raw_spin_rq_unlock(this_rq);
2395 : stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2396 : push_task, &src_rq->push_work);
2397 : raw_spin_rq_lock(this_rq);
2398 : }
2399 : }
2400 :
2401 : if (resched)
2402 : resched_curr(this_rq);
2403 : }
2404 :
2405 : /*
2406 : * If we are not running and we are not going to reschedule soon, we should
2407 : * try to push tasks away now
2408 : */
2409 : static void task_woken_rt(struct rq *rq, struct task_struct *p)
2410 : {
2411 : bool need_to_push = !task_running(rq, p) &&
2412 : !test_tsk_need_resched(rq->curr) &&
2413 : p->nr_cpus_allowed > 1 &&
2414 : (dl_task(rq->curr) || rt_task(rq->curr)) &&
2415 : (rq->curr->nr_cpus_allowed < 2 ||
2416 : rq->curr->prio <= p->prio);
2417 :
2418 : if (need_to_push)
2419 : push_rt_tasks(rq);
2420 : }
2421 :
2422 : /* Assumes rq->lock is held */
2423 : static void rq_online_rt(struct rq *rq)
2424 : {
2425 : if (rq->rt.overloaded)
2426 : rt_set_overload(rq);
2427 :
2428 : __enable_runtime(rq);
2429 :
2430 : cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2431 : }
2432 :
2433 : /* Assumes rq->lock is held */
2434 : static void rq_offline_rt(struct rq *rq)
2435 : {
2436 : if (rq->rt.overloaded)
2437 : rt_clear_overload(rq);
2438 :
2439 : __disable_runtime(rq);
2440 :
2441 : cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2442 : }
2443 :
2444 : /*
2445 : * When switch from the rt queue, we bring ourselves to a position
2446 : * that we might want to pull RT tasks from other runqueues.
2447 : */
2448 : static void switched_from_rt(struct rq *rq, struct task_struct *p)
2449 : {
2450 : /*
2451 : * If there are other RT tasks then we will reschedule
2452 : * and the scheduling of the other RT tasks will handle
2453 : * the balancing. But if we are the last RT task
2454 : * we may need to handle the pulling of RT tasks
2455 : * now.
2456 : */
2457 : if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2458 : return;
2459 :
2460 : rt_queue_pull_task(rq);
2461 : }
2462 :
2463 : void __init init_sched_rt_class(void)
2464 : {
2465 : unsigned int i;
2466 :
2467 : for_each_possible_cpu(i) {
2468 : zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2469 : GFP_KERNEL, cpu_to_node(i));
2470 : }
2471 : }
2472 : #endif /* CONFIG_SMP */
2473 :
2474 : /*
2475 : * When switching a task to RT, we may overload the runqueue
2476 : * with RT tasks. In this case we try to push them off to
2477 : * other runqueues.
2478 : */
2479 0 : static void switched_to_rt(struct rq *rq, struct task_struct *p)
2480 : {
2481 : /*
2482 : * If we are running, update the avg_rt tracking, as the running time
2483 : * will now on be accounted into the latter.
2484 : */
2485 0 : if (task_current(rq, p)) {
2486 0 : update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2487 : return;
2488 : }
2489 :
2490 : /*
2491 : * If we are not running we may need to preempt the current
2492 : * running task. If that current running task is also an RT task
2493 : * then see if we can move to another run queue.
2494 : */
2495 0 : if (task_on_rq_queued(p)) {
2496 : #ifdef CONFIG_SMP
2497 : if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2498 : rt_queue_push_tasks(rq);
2499 : #endif /* CONFIG_SMP */
2500 0 : if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq)))
2501 0 : resched_curr(rq);
2502 : }
2503 : }
2504 :
2505 : /*
2506 : * Priority of the task has changed. This may cause
2507 : * us to initiate a push or pull.
2508 : */
2509 : static void
2510 0 : prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2511 : {
2512 0 : if (!task_on_rq_queued(p))
2513 : return;
2514 :
2515 0 : if (task_current(rq, p)) {
2516 : #ifdef CONFIG_SMP
2517 : /*
2518 : * If our priority decreases while running, we
2519 : * may need to pull tasks to this runqueue.
2520 : */
2521 : if (oldprio < p->prio)
2522 : rt_queue_pull_task(rq);
2523 :
2524 : /*
2525 : * If there's a higher priority task waiting to run
2526 : * then reschedule.
2527 : */
2528 : if (p->prio > rq->rt.highest_prio.curr)
2529 : resched_curr(rq);
2530 : #else
2531 : /* For UP simply resched on drop of prio */
2532 0 : if (oldprio < p->prio)
2533 0 : resched_curr(rq);
2534 : #endif /* CONFIG_SMP */
2535 : } else {
2536 : /*
2537 : * This task is not running, but if it is
2538 : * greater than the current running task
2539 : * then reschedule.
2540 : */
2541 0 : if (p->prio < rq->curr->prio)
2542 0 : resched_curr(rq);
2543 : }
2544 : }
2545 :
2546 : #ifdef CONFIG_POSIX_TIMERS
2547 0 : static void watchdog(struct rq *rq, struct task_struct *p)
2548 : {
2549 : unsigned long soft, hard;
2550 :
2551 : /* max may change after cur was read, this will be fixed next tick */
2552 0 : soft = task_rlimit(p, RLIMIT_RTTIME);
2553 0 : hard = task_rlimit_max(p, RLIMIT_RTTIME);
2554 :
2555 0 : if (soft != RLIM_INFINITY) {
2556 : unsigned long next;
2557 :
2558 0 : if (p->rt.watchdog_stamp != jiffies) {
2559 0 : p->rt.timeout++;
2560 0 : p->rt.watchdog_stamp = jiffies;
2561 : }
2562 :
2563 0 : next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2564 0 : if (p->rt.timeout > next) {
2565 0 : posix_cputimers_rt_watchdog(&p->posix_cputimers,
2566 : p->se.sum_exec_runtime);
2567 : }
2568 : }
2569 0 : }
2570 : #else
2571 : static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2572 : #endif
2573 :
2574 : /*
2575 : * scheduler tick hitting a task of our scheduling class.
2576 : *
2577 : * NOTE: This function can be called remotely by the tick offload that
2578 : * goes along full dynticks. Therefore no local assumption can be made
2579 : * and everything must be accessed through the @rq and @curr passed in
2580 : * parameters.
2581 : */
2582 0 : static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2583 : {
2584 0 : struct sched_rt_entity *rt_se = &p->rt;
2585 :
2586 0 : update_curr_rt(rq);
2587 0 : update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2588 :
2589 0 : watchdog(rq, p);
2590 :
2591 : /*
2592 : * RR tasks need a special form of timeslice management.
2593 : * FIFO tasks have no timeslices.
2594 : */
2595 0 : if (p->policy != SCHED_RR)
2596 : return;
2597 :
2598 0 : if (--p->rt.time_slice)
2599 : return;
2600 :
2601 0 : p->rt.time_slice = sched_rr_timeslice;
2602 :
2603 : /*
2604 : * Requeue to the end of queue if we (and all of our ancestors) are not
2605 : * the only element on the queue
2606 : */
2607 0 : for_each_sched_rt_entity(rt_se) {
2608 0 : if (rt_se->run_list.prev != rt_se->run_list.next) {
2609 0 : requeue_task_rt(rq, p, 0);
2610 0 : resched_curr(rq);
2611 0 : return;
2612 : }
2613 : }
2614 : }
2615 :
2616 0 : static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2617 : {
2618 : /*
2619 : * Time slice is 0 for SCHED_FIFO tasks
2620 : */
2621 0 : if (task->policy == SCHED_RR)
2622 0 : return sched_rr_timeslice;
2623 : else
2624 : return 0;
2625 : }
2626 :
2627 : DEFINE_SCHED_CLASS(rt) = {
2628 :
2629 : .enqueue_task = enqueue_task_rt,
2630 : .dequeue_task = dequeue_task_rt,
2631 : .yield_task = yield_task_rt,
2632 :
2633 : .check_preempt_curr = check_preempt_curr_rt,
2634 :
2635 : .pick_next_task = pick_next_task_rt,
2636 : .put_prev_task = put_prev_task_rt,
2637 : .set_next_task = set_next_task_rt,
2638 :
2639 : #ifdef CONFIG_SMP
2640 : .balance = balance_rt,
2641 : .pick_task = pick_task_rt,
2642 : .select_task_rq = select_task_rq_rt,
2643 : .set_cpus_allowed = set_cpus_allowed_common,
2644 : .rq_online = rq_online_rt,
2645 : .rq_offline = rq_offline_rt,
2646 : .task_woken = task_woken_rt,
2647 : .switched_from = switched_from_rt,
2648 : .find_lock_rq = find_lock_lowest_rq,
2649 : #endif
2650 :
2651 : .task_tick = task_tick_rt,
2652 :
2653 : .get_rr_interval = get_rr_interval_rt,
2654 :
2655 : .prio_changed = prio_changed_rt,
2656 : .switched_to = switched_to_rt,
2657 :
2658 : .update_curr = update_curr_rt,
2659 :
2660 : #ifdef CONFIG_UCLAMP_TASK
2661 : .uclamp_enabled = 1,
2662 : #endif
2663 : };
2664 :
2665 : #ifdef CONFIG_RT_GROUP_SCHED
2666 : /*
2667 : * Ensure that the real time constraints are schedulable.
2668 : */
2669 : static DEFINE_MUTEX(rt_constraints_mutex);
2670 :
2671 : static inline int tg_has_rt_tasks(struct task_group *tg)
2672 : {
2673 : struct task_struct *task;
2674 : struct css_task_iter it;
2675 : int ret = 0;
2676 :
2677 : /*
2678 : * Autogroups do not have RT tasks; see autogroup_create().
2679 : */
2680 : if (task_group_is_autogroup(tg))
2681 : return 0;
2682 :
2683 : css_task_iter_start(&tg->css, 0, &it);
2684 : while (!ret && (task = css_task_iter_next(&it)))
2685 : ret |= rt_task(task);
2686 : css_task_iter_end(&it);
2687 :
2688 : return ret;
2689 : }
2690 :
2691 : struct rt_schedulable_data {
2692 : struct task_group *tg;
2693 : u64 rt_period;
2694 : u64 rt_runtime;
2695 : };
2696 :
2697 : static int tg_rt_schedulable(struct task_group *tg, void *data)
2698 : {
2699 : struct rt_schedulable_data *d = data;
2700 : struct task_group *child;
2701 : unsigned long total, sum = 0;
2702 : u64 period, runtime;
2703 :
2704 : period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2705 : runtime = tg->rt_bandwidth.rt_runtime;
2706 :
2707 : if (tg == d->tg) {
2708 : period = d->rt_period;
2709 : runtime = d->rt_runtime;
2710 : }
2711 :
2712 : /*
2713 : * Cannot have more runtime than the period.
2714 : */
2715 : if (runtime > period && runtime != RUNTIME_INF)
2716 : return -EINVAL;
2717 :
2718 : /*
2719 : * Ensure we don't starve existing RT tasks if runtime turns zero.
2720 : */
2721 : if (rt_bandwidth_enabled() && !runtime &&
2722 : tg->rt_bandwidth.rt_runtime && tg_has_rt_tasks(tg))
2723 : return -EBUSY;
2724 :
2725 : total = to_ratio(period, runtime);
2726 :
2727 : /*
2728 : * Nobody can have more than the global setting allows.
2729 : */
2730 : if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2731 : return -EINVAL;
2732 :
2733 : /*
2734 : * The sum of our children's runtime should not exceed our own.
2735 : */
2736 : list_for_each_entry_rcu(child, &tg->children, siblings) {
2737 : period = ktime_to_ns(child->rt_bandwidth.rt_period);
2738 : runtime = child->rt_bandwidth.rt_runtime;
2739 :
2740 : if (child == d->tg) {
2741 : period = d->rt_period;
2742 : runtime = d->rt_runtime;
2743 : }
2744 :
2745 : sum += to_ratio(period, runtime);
2746 : }
2747 :
2748 : if (sum > total)
2749 : return -EINVAL;
2750 :
2751 : return 0;
2752 : }
2753 :
2754 : static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2755 : {
2756 : int ret;
2757 :
2758 : struct rt_schedulable_data data = {
2759 : .tg = tg,
2760 : .rt_period = period,
2761 : .rt_runtime = runtime,
2762 : };
2763 :
2764 : rcu_read_lock();
2765 : ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2766 : rcu_read_unlock();
2767 :
2768 : return ret;
2769 : }
2770 :
2771 : static int tg_set_rt_bandwidth(struct task_group *tg,
2772 : u64 rt_period, u64 rt_runtime)
2773 : {
2774 : int i, err = 0;
2775 :
2776 : /*
2777 : * Disallowing the root group RT runtime is BAD, it would disallow the
2778 : * kernel creating (and or operating) RT threads.
2779 : */
2780 : if (tg == &root_task_group && rt_runtime == 0)
2781 : return -EINVAL;
2782 :
2783 : /* No period doesn't make any sense. */
2784 : if (rt_period == 0)
2785 : return -EINVAL;
2786 :
2787 : /*
2788 : * Bound quota to defend quota against overflow during bandwidth shift.
2789 : */
2790 : if (rt_runtime != RUNTIME_INF && rt_runtime > max_rt_runtime)
2791 : return -EINVAL;
2792 :
2793 : mutex_lock(&rt_constraints_mutex);
2794 : err = __rt_schedulable(tg, rt_period, rt_runtime);
2795 : if (err)
2796 : goto unlock;
2797 :
2798 : raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2799 : tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2800 : tg->rt_bandwidth.rt_runtime = rt_runtime;
2801 :
2802 : for_each_possible_cpu(i) {
2803 : struct rt_rq *rt_rq = tg->rt_rq[i];
2804 :
2805 : raw_spin_lock(&rt_rq->rt_runtime_lock);
2806 : rt_rq->rt_runtime = rt_runtime;
2807 : raw_spin_unlock(&rt_rq->rt_runtime_lock);
2808 : }
2809 : raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2810 : unlock:
2811 : mutex_unlock(&rt_constraints_mutex);
2812 :
2813 : return err;
2814 : }
2815 :
2816 : int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2817 : {
2818 : u64 rt_runtime, rt_period;
2819 :
2820 : rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2821 : rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2822 : if (rt_runtime_us < 0)
2823 : rt_runtime = RUNTIME_INF;
2824 : else if ((u64)rt_runtime_us > U64_MAX / NSEC_PER_USEC)
2825 : return -EINVAL;
2826 :
2827 : return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2828 : }
2829 :
2830 : long sched_group_rt_runtime(struct task_group *tg)
2831 : {
2832 : u64 rt_runtime_us;
2833 :
2834 : if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2835 : return -1;
2836 :
2837 : rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2838 : do_div(rt_runtime_us, NSEC_PER_USEC);
2839 : return rt_runtime_us;
2840 : }
2841 :
2842 : int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2843 : {
2844 : u64 rt_runtime, rt_period;
2845 :
2846 : if (rt_period_us > U64_MAX / NSEC_PER_USEC)
2847 : return -EINVAL;
2848 :
2849 : rt_period = rt_period_us * NSEC_PER_USEC;
2850 : rt_runtime = tg->rt_bandwidth.rt_runtime;
2851 :
2852 : return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2853 : }
2854 :
2855 : long sched_group_rt_period(struct task_group *tg)
2856 : {
2857 : u64 rt_period_us;
2858 :
2859 : rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2860 : do_div(rt_period_us, NSEC_PER_USEC);
2861 : return rt_period_us;
2862 : }
2863 :
2864 : static int sched_rt_global_constraints(void)
2865 : {
2866 : int ret = 0;
2867 :
2868 : mutex_lock(&rt_constraints_mutex);
2869 : ret = __rt_schedulable(NULL, 0, 0);
2870 : mutex_unlock(&rt_constraints_mutex);
2871 :
2872 : return ret;
2873 : }
2874 :
2875 : int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2876 : {
2877 : /* Don't accept realtime tasks when there is no way for them to run */
2878 : if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2879 : return 0;
2880 :
2881 : return 1;
2882 : }
2883 :
2884 : #else /* !CONFIG_RT_GROUP_SCHED */
2885 0 : static int sched_rt_global_constraints(void)
2886 : {
2887 : unsigned long flags;
2888 : int i;
2889 :
2890 0 : raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2891 0 : for_each_possible_cpu(i) {
2892 0 : struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2893 :
2894 0 : raw_spin_lock(&rt_rq->rt_runtime_lock);
2895 0 : rt_rq->rt_runtime = global_rt_runtime();
2896 0 : raw_spin_unlock(&rt_rq->rt_runtime_lock);
2897 : }
2898 0 : raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2899 :
2900 0 : return 0;
2901 : }
2902 : #endif /* CONFIG_RT_GROUP_SCHED */
2903 :
2904 : static int sched_rt_global_validate(void)
2905 : {
2906 0 : if (sysctl_sched_rt_period <= 0)
2907 : return -EINVAL;
2908 :
2909 0 : if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2910 0 : ((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||
2911 0 : ((u64)sysctl_sched_rt_runtime *
2912 : NSEC_PER_USEC > max_rt_runtime)))
2913 : return -EINVAL;
2914 :
2915 : return 0;
2916 : }
2917 :
2918 0 : static void sched_rt_do_global(void)
2919 : {
2920 : unsigned long flags;
2921 :
2922 0 : raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2923 0 : def_rt_bandwidth.rt_runtime = global_rt_runtime();
2924 0 : def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2925 0 : raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2926 0 : }
2927 :
2928 0 : int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
2929 : size_t *lenp, loff_t *ppos)
2930 : {
2931 : int old_period, old_runtime;
2932 : static DEFINE_MUTEX(mutex);
2933 : int ret;
2934 :
2935 0 : mutex_lock(&mutex);
2936 0 : old_period = sysctl_sched_rt_period;
2937 0 : old_runtime = sysctl_sched_rt_runtime;
2938 :
2939 0 : ret = proc_dointvec(table, write, buffer, lenp, ppos);
2940 :
2941 0 : if (!ret && write) {
2942 0 : ret = sched_rt_global_validate();
2943 0 : if (ret)
2944 : goto undo;
2945 :
2946 0 : ret = sched_dl_global_validate();
2947 0 : if (ret)
2948 : goto undo;
2949 :
2950 0 : ret = sched_rt_global_constraints();
2951 0 : if (ret)
2952 : goto undo;
2953 :
2954 0 : sched_rt_do_global();
2955 0 : sched_dl_do_global();
2956 : }
2957 : if (0) {
2958 : undo:
2959 0 : sysctl_sched_rt_period = old_period;
2960 0 : sysctl_sched_rt_runtime = old_runtime;
2961 : }
2962 0 : mutex_unlock(&mutex);
2963 :
2964 0 : return ret;
2965 : }
2966 :
2967 0 : int sched_rr_handler(struct ctl_table *table, int write, void *buffer,
2968 : size_t *lenp, loff_t *ppos)
2969 : {
2970 : int ret;
2971 : static DEFINE_MUTEX(mutex);
2972 :
2973 0 : mutex_lock(&mutex);
2974 0 : ret = proc_dointvec(table, write, buffer, lenp, ppos);
2975 : /*
2976 : * Make sure that internally we keep jiffies.
2977 : * Also, writing zero resets the timeslice to default:
2978 : */
2979 0 : if (!ret && write) {
2980 0 : sched_rr_timeslice =
2981 0 : sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
2982 0 : msecs_to_jiffies(sysctl_sched_rr_timeslice);
2983 : }
2984 0 : mutex_unlock(&mutex);
2985 :
2986 0 : return ret;
2987 : }
2988 :
2989 : #ifdef CONFIG_SCHED_DEBUG
2990 0 : void print_rt_stats(struct seq_file *m, int cpu)
2991 : {
2992 : rt_rq_iter_t iter;
2993 : struct rt_rq *rt_rq;
2994 :
2995 : rcu_read_lock();
2996 0 : for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2997 0 : print_rt_rq(m, cpu, rt_rq);
2998 : rcu_read_unlock();
2999 0 : }
3000 : #endif /* CONFIG_SCHED_DEBUG */
|
