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