1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Deadline Scheduling Class (SCHED_DEADLINE)
4 *
5 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
6 *
7 * Tasks that periodically executes their instances for less than their
8 * runtime won't miss any of their deadlines.
9 * Tasks that are not periodic or sporadic or that tries to execute more
10 * than their reserved bandwidth will be slowed down (and may potentially
11 * miss some of their deadlines), and won't affect any other task.
12 *
13 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
14 * Juri Lelli <juri.lelli@gmail.com>,
15 * Michael Trimarchi <michael@amarulasolutions.com>,
16 * Fabio Checconi <fchecconi@gmail.com>
17 */
18
19 #include <linux/cpuset.h>
20
21 /*
22 * Default limits for DL period; on the top end we guard against small util
23 * tasks still getting ridiculously long effective runtimes, on the bottom end we
24 * guard against timer DoS.
25 */
26 static unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
27 static unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
28 #ifdef CONFIG_SYSCTL
29 static const struct ctl_table sched_dl_sysctls[] = {
30 {
31 .procname = "sched_deadline_period_max_us",
32 .data = &sysctl_sched_dl_period_max,
33 .maxlen = sizeof(unsigned int),
34 .mode = 0644,
35 .proc_handler = proc_douintvec_minmax,
36 .extra1 = (void *)&sysctl_sched_dl_period_min,
37 },
38 {
39 .procname = "sched_deadline_period_min_us",
40 .data = &sysctl_sched_dl_period_min,
41 .maxlen = sizeof(unsigned int),
42 .mode = 0644,
43 .proc_handler = proc_douintvec_minmax,
44 .extra2 = (void *)&sysctl_sched_dl_period_max,
45 },
46 };
47
sched_dl_sysctl_init(void)48 static int __init sched_dl_sysctl_init(void)
49 {
50 register_sysctl_init("kernel", sched_dl_sysctls);
51 return 0;
52 }
53 late_initcall(sched_dl_sysctl_init);
54 #endif
55
dl_server(struct sched_dl_entity * dl_se)56 static bool dl_server(struct sched_dl_entity *dl_se)
57 {
58 return dl_se->dl_server;
59 }
60
dl_task_of(struct sched_dl_entity * dl_se)61 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
62 {
63 BUG_ON(dl_server(dl_se));
64 return container_of(dl_se, struct task_struct, dl);
65 }
66
rq_of_dl_rq(struct dl_rq * dl_rq)67 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
68 {
69 return container_of(dl_rq, struct rq, dl);
70 }
71
rq_of_dl_se(struct sched_dl_entity * dl_se)72 static inline struct rq *rq_of_dl_se(struct sched_dl_entity *dl_se)
73 {
74 struct rq *rq = dl_se->rq;
75
76 if (!dl_server(dl_se))
77 rq = task_rq(dl_task_of(dl_se));
78
79 return rq;
80 }
81
dl_rq_of_se(struct sched_dl_entity * dl_se)82 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
83 {
84 return &rq_of_dl_se(dl_se)->dl;
85 }
86
on_dl_rq(struct sched_dl_entity * dl_se)87 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
88 {
89 return !RB_EMPTY_NODE(&dl_se->rb_node);
90 }
91
92 #ifdef CONFIG_RT_MUTEXES
pi_of(struct sched_dl_entity * dl_se)93 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
94 {
95 return dl_se->pi_se;
96 }
97
is_dl_boosted(struct sched_dl_entity * dl_se)98 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
99 {
100 return pi_of(dl_se) != dl_se;
101 }
102 #else
pi_of(struct sched_dl_entity * dl_se)103 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
104 {
105 return dl_se;
106 }
107
is_dl_boosted(struct sched_dl_entity * dl_se)108 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
109 {
110 return false;
111 }
112 #endif
113
114 #ifdef CONFIG_SMP
dl_bw_of(int i)115 static inline struct dl_bw *dl_bw_of(int i)
116 {
117 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
118 "sched RCU must be held");
119 return &cpu_rq(i)->rd->dl_bw;
120 }
121
dl_bw_cpus(int i)122 static inline int dl_bw_cpus(int i)
123 {
124 struct root_domain *rd = cpu_rq(i)->rd;
125 int cpus;
126
127 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
128 "sched RCU must be held");
129
130 if (cpumask_subset(rd->span, cpu_active_mask))
131 return cpumask_weight(rd->span);
132
133 cpus = 0;
134
135 for_each_cpu_and(i, rd->span, cpu_active_mask)
136 cpus++;
137
138 return cpus;
139 }
140
__dl_bw_capacity(const struct cpumask * mask)141 static inline unsigned long __dl_bw_capacity(const struct cpumask *mask)
142 {
143 unsigned long cap = 0;
144 int i;
145
146 for_each_cpu_and(i, mask, cpu_active_mask)
147 cap += arch_scale_cpu_capacity(i);
148
149 return cap;
150 }
151
152 /*
153 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
154 * of the CPU the task is running on rather rd's \Sum CPU capacity.
155 */
dl_bw_capacity(int i)156 static inline unsigned long dl_bw_capacity(int i)
157 {
158 if (!sched_asym_cpucap_active() &&
159 arch_scale_cpu_capacity(i) == SCHED_CAPACITY_SCALE) {
160 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
161 } else {
162 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
163 "sched RCU must be held");
164
165 return __dl_bw_capacity(cpu_rq(i)->rd->span);
166 }
167 }
168
dl_bw_visited(int cpu,u64 gen)169 static inline bool dl_bw_visited(int cpu, u64 gen)
170 {
171 struct root_domain *rd = cpu_rq(cpu)->rd;
172
173 if (rd->visit_gen == gen)
174 return true;
175
176 rd->visit_gen = gen;
177 return false;
178 }
179
180 static inline
__dl_update(struct dl_bw * dl_b,s64 bw)181 void __dl_update(struct dl_bw *dl_b, s64 bw)
182 {
183 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
184 int i;
185
186 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
187 "sched RCU must be held");
188 for_each_cpu_and(i, rd->span, cpu_active_mask) {
189 struct rq *rq = cpu_rq(i);
190
191 rq->dl.extra_bw += bw;
192 }
193 }
194 #else
dl_bw_of(int i)195 static inline struct dl_bw *dl_bw_of(int i)
196 {
197 return &cpu_rq(i)->dl.dl_bw;
198 }
199
dl_bw_cpus(int i)200 static inline int dl_bw_cpus(int i)
201 {
202 return 1;
203 }
204
dl_bw_capacity(int i)205 static inline unsigned long dl_bw_capacity(int i)
206 {
207 return SCHED_CAPACITY_SCALE;
208 }
209
dl_bw_visited(int cpu,u64 gen)210 static inline bool dl_bw_visited(int cpu, u64 gen)
211 {
212 return false;
213 }
214
215 static inline
__dl_update(struct dl_bw * dl_b,s64 bw)216 void __dl_update(struct dl_bw *dl_b, s64 bw)
217 {
218 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
219
220 dl->extra_bw += bw;
221 }
222 #endif
223
224 static inline
__dl_sub(struct dl_bw * dl_b,u64 tsk_bw,int cpus)225 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
226 {
227 dl_b->total_bw -= tsk_bw;
228 __dl_update(dl_b, (s32)tsk_bw / cpus);
229 }
230
231 static inline
__dl_add(struct dl_bw * dl_b,u64 tsk_bw,int cpus)232 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
233 {
234 dl_b->total_bw += tsk_bw;
235 __dl_update(dl_b, -((s32)tsk_bw / cpus));
236 }
237
238 static inline bool
__dl_overflow(struct dl_bw * dl_b,unsigned long cap,u64 old_bw,u64 new_bw)239 __dl_overflow(struct dl_bw *dl_b, unsigned long cap, u64 old_bw, u64 new_bw)
240 {
241 return dl_b->bw != -1 &&
242 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
243 }
244
245 static inline
__add_running_bw(u64 dl_bw,struct dl_rq * dl_rq)246 void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
247 {
248 u64 old = dl_rq->running_bw;
249
250 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
251 dl_rq->running_bw += dl_bw;
252 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
253 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
254 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
255 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
256 }
257
258 static inline
__sub_running_bw(u64 dl_bw,struct dl_rq * dl_rq)259 void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
260 {
261 u64 old = dl_rq->running_bw;
262
263 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
264 dl_rq->running_bw -= dl_bw;
265 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
266 if (dl_rq->running_bw > old)
267 dl_rq->running_bw = 0;
268 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
269 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
270 }
271
272 static inline
__add_rq_bw(u64 dl_bw,struct dl_rq * dl_rq)273 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
274 {
275 u64 old = dl_rq->this_bw;
276
277 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
278 dl_rq->this_bw += dl_bw;
279 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
280 }
281
282 static inline
__sub_rq_bw(u64 dl_bw,struct dl_rq * dl_rq)283 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
284 {
285 u64 old = dl_rq->this_bw;
286
287 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
288 dl_rq->this_bw -= dl_bw;
289 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
290 if (dl_rq->this_bw > old)
291 dl_rq->this_bw = 0;
292 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
293 }
294
295 static inline
add_rq_bw(struct sched_dl_entity * dl_se,struct dl_rq * dl_rq)296 void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
297 {
298 if (!dl_entity_is_special(dl_se))
299 __add_rq_bw(dl_se->dl_bw, dl_rq);
300 }
301
302 static inline
sub_rq_bw(struct sched_dl_entity * dl_se,struct dl_rq * dl_rq)303 void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
304 {
305 if (!dl_entity_is_special(dl_se))
306 __sub_rq_bw(dl_se->dl_bw, dl_rq);
307 }
308
309 static inline
add_running_bw(struct sched_dl_entity * dl_se,struct dl_rq * dl_rq)310 void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
311 {
312 if (!dl_entity_is_special(dl_se))
313 __add_running_bw(dl_se->dl_bw, dl_rq);
314 }
315
316 static inline
sub_running_bw(struct sched_dl_entity * dl_se,struct dl_rq * dl_rq)317 void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
318 {
319 if (!dl_entity_is_special(dl_se))
320 __sub_running_bw(dl_se->dl_bw, dl_rq);
321 }
322
dl_rq_change_utilization(struct rq * rq,struct sched_dl_entity * dl_se,u64 new_bw)323 static void dl_rq_change_utilization(struct rq *rq, struct sched_dl_entity *dl_se, u64 new_bw)
324 {
325 if (dl_se->dl_non_contending) {
326 sub_running_bw(dl_se, &rq->dl);
327 dl_se->dl_non_contending = 0;
328
329 /*
330 * If the timer handler is currently running and the
331 * timer cannot be canceled, inactive_task_timer()
332 * will see that dl_not_contending is not set, and
333 * will not touch the rq's active utilization,
334 * so we are still safe.
335 */
336 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1) {
337 if (!dl_server(dl_se))
338 put_task_struct(dl_task_of(dl_se));
339 }
340 }
341 __sub_rq_bw(dl_se->dl_bw, &rq->dl);
342 __add_rq_bw(new_bw, &rq->dl);
343 }
344
345 static __always_inline
cancel_dl_timer(struct sched_dl_entity * dl_se,struct hrtimer * timer)346 void cancel_dl_timer(struct sched_dl_entity *dl_se, struct hrtimer *timer)
347 {
348 /*
349 * If the timer callback was running (hrtimer_try_to_cancel == -1),
350 * it will eventually call put_task_struct().
351 */
352 if (hrtimer_try_to_cancel(timer) == 1 && !dl_server(dl_se))
353 put_task_struct(dl_task_of(dl_se));
354 }
355
356 static __always_inline
cancel_replenish_timer(struct sched_dl_entity * dl_se)357 void cancel_replenish_timer(struct sched_dl_entity *dl_se)
358 {
359 cancel_dl_timer(dl_se, &dl_se->dl_timer);
360 }
361
362 static __always_inline
cancel_inactive_timer(struct sched_dl_entity * dl_se)363 void cancel_inactive_timer(struct sched_dl_entity *dl_se)
364 {
365 cancel_dl_timer(dl_se, &dl_se->inactive_timer);
366 }
367
dl_change_utilization(struct task_struct * p,u64 new_bw)368 static void dl_change_utilization(struct task_struct *p, u64 new_bw)
369 {
370 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_SUGOV);
371
372 if (task_on_rq_queued(p))
373 return;
374
375 dl_rq_change_utilization(task_rq(p), &p->dl, new_bw);
376 }
377
378 static void __dl_clear_params(struct sched_dl_entity *dl_se);
379
380 /*
381 * The utilization of a task cannot be immediately removed from
382 * the rq active utilization (running_bw) when the task blocks.
383 * Instead, we have to wait for the so called "0-lag time".
384 *
385 * If a task blocks before the "0-lag time", a timer (the inactive
386 * timer) is armed, and running_bw is decreased when the timer
387 * fires.
388 *
389 * If the task wakes up again before the inactive timer fires,
390 * the timer is canceled, whereas if the task wakes up after the
391 * inactive timer fired (and running_bw has been decreased) the
392 * task's utilization has to be added to running_bw again.
393 * A flag in the deadline scheduling entity (dl_non_contending)
394 * is used to avoid race conditions between the inactive timer handler
395 * and task wakeups.
396 *
397 * The following diagram shows how running_bw is updated. A task is
398 * "ACTIVE" when its utilization contributes to running_bw; an
399 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
400 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
401 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
402 * time already passed, which does not contribute to running_bw anymore.
403 * +------------------+
404 * wakeup | ACTIVE |
405 * +------------------>+ contending |
406 * | add_running_bw | |
407 * | +----+------+------+
408 * | | ^
409 * | dequeue | |
410 * +--------+-------+ | |
411 * | | t >= 0-lag | | wakeup
412 * | INACTIVE |<---------------+ |
413 * | | sub_running_bw | |
414 * +--------+-------+ | |
415 * ^ | |
416 * | t < 0-lag | |
417 * | | |
418 * | V |
419 * | +----+------+------+
420 * | sub_running_bw | ACTIVE |
421 * +-------------------+ |
422 * inactive timer | non contending |
423 * fired +------------------+
424 *
425 * The task_non_contending() function is invoked when a task
426 * blocks, and checks if the 0-lag time already passed or
427 * not (in the first case, it directly updates running_bw;
428 * in the second case, it arms the inactive timer).
429 *
430 * The task_contending() function is invoked when a task wakes
431 * up, and checks if the task is still in the "ACTIVE non contending"
432 * state or not (in the second case, it updates running_bw).
433 */
task_non_contending(struct sched_dl_entity * dl_se)434 static void task_non_contending(struct sched_dl_entity *dl_se)
435 {
436 struct hrtimer *timer = &dl_se->inactive_timer;
437 struct rq *rq = rq_of_dl_se(dl_se);
438 struct dl_rq *dl_rq = &rq->dl;
439 s64 zerolag_time;
440
441 /*
442 * If this is a non-deadline task that has been boosted,
443 * do nothing
444 */
445 if (dl_se->dl_runtime == 0)
446 return;
447
448 if (dl_entity_is_special(dl_se))
449 return;
450
451 WARN_ON(dl_se->dl_non_contending);
452
453 zerolag_time = dl_se->deadline -
454 div64_long((dl_se->runtime * dl_se->dl_period),
455 dl_se->dl_runtime);
456
457 /*
458 * Using relative times instead of the absolute "0-lag time"
459 * allows to simplify the code
460 */
461 zerolag_time -= rq_clock(rq);
462
463 /*
464 * If the "0-lag time" already passed, decrease the active
465 * utilization now, instead of starting a timer
466 */
467 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
468 if (dl_server(dl_se)) {
469 sub_running_bw(dl_se, dl_rq);
470 } else {
471 struct task_struct *p = dl_task_of(dl_se);
472
473 if (dl_task(p))
474 sub_running_bw(dl_se, dl_rq);
475
476 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
477 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
478
479 if (READ_ONCE(p->__state) == TASK_DEAD)
480 sub_rq_bw(dl_se, &rq->dl);
481 raw_spin_lock(&dl_b->lock);
482 __dl_sub(dl_b, dl_se->dl_bw, dl_bw_cpus(task_cpu(p)));
483 raw_spin_unlock(&dl_b->lock);
484 __dl_clear_params(dl_se);
485 }
486 }
487
488 return;
489 }
490
491 dl_se->dl_non_contending = 1;
492 if (!dl_server(dl_se))
493 get_task_struct(dl_task_of(dl_se));
494
495 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
496 }
497
task_contending(struct sched_dl_entity * dl_se,int flags)498 static void task_contending(struct sched_dl_entity *dl_se, int flags)
499 {
500 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
501
502 /*
503 * If this is a non-deadline task that has been boosted,
504 * do nothing
505 */
506 if (dl_se->dl_runtime == 0)
507 return;
508
509 if (flags & ENQUEUE_MIGRATED)
510 add_rq_bw(dl_se, dl_rq);
511
512 if (dl_se->dl_non_contending) {
513 dl_se->dl_non_contending = 0;
514 /*
515 * If the timer handler is currently running and the
516 * timer cannot be canceled, inactive_task_timer()
517 * will see that dl_not_contending is not set, and
518 * will not touch the rq's active utilization,
519 * so we are still safe.
520 */
521 cancel_inactive_timer(dl_se);
522 } else {
523 /*
524 * Since "dl_non_contending" is not set, the
525 * task's utilization has already been removed from
526 * active utilization (either when the task blocked,
527 * when the "inactive timer" fired).
528 * So, add it back.
529 */
530 add_running_bw(dl_se, dl_rq);
531 }
532 }
533
is_leftmost(struct sched_dl_entity * dl_se,struct dl_rq * dl_rq)534 static inline int is_leftmost(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
535 {
536 return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
537 }
538
539 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
540
init_dl_bw(struct dl_bw * dl_b)541 void init_dl_bw(struct dl_bw *dl_b)
542 {
543 raw_spin_lock_init(&dl_b->lock);
544 if (global_rt_runtime() == RUNTIME_INF)
545 dl_b->bw = -1;
546 else
547 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
548 dl_b->total_bw = 0;
549 }
550
init_dl_rq(struct dl_rq * dl_rq)551 void init_dl_rq(struct dl_rq *dl_rq)
552 {
553 dl_rq->root = RB_ROOT_CACHED;
554
555 #ifdef CONFIG_SMP
556 /* zero means no -deadline tasks */
557 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
558
559 dl_rq->overloaded = 0;
560 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
561 #else
562 init_dl_bw(&dl_rq->dl_bw);
563 #endif
564
565 dl_rq->running_bw = 0;
566 dl_rq->this_bw = 0;
567 init_dl_rq_bw_ratio(dl_rq);
568 }
569
570 #ifdef CONFIG_SMP
571
dl_overloaded(struct rq * rq)572 static inline int dl_overloaded(struct rq *rq)
573 {
574 return atomic_read(&rq->rd->dlo_count);
575 }
576
dl_set_overload(struct rq * rq)577 static inline void dl_set_overload(struct rq *rq)
578 {
579 if (!rq->online)
580 return;
581
582 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
583 /*
584 * Must be visible before the overload count is
585 * set (as in sched_rt.c).
586 *
587 * Matched by the barrier in pull_dl_task().
588 */
589 smp_wmb();
590 atomic_inc(&rq->rd->dlo_count);
591 }
592
dl_clear_overload(struct rq * rq)593 static inline void dl_clear_overload(struct rq *rq)
594 {
595 if (!rq->online)
596 return;
597
598 atomic_dec(&rq->rd->dlo_count);
599 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
600 }
601
602 #define __node_2_pdl(node) \
603 rb_entry((node), struct task_struct, pushable_dl_tasks)
604
__pushable_less(struct rb_node * a,const struct rb_node * b)605 static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
606 {
607 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
608 }
609
has_pushable_dl_tasks(struct rq * rq)610 static inline int has_pushable_dl_tasks(struct rq *rq)
611 {
612 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
613 }
614
615 /*
616 * The list of pushable -deadline task is not a plist, like in
617 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
618 */
enqueue_pushable_dl_task(struct rq * rq,struct task_struct * p)619 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
620 {
621 struct rb_node *leftmost;
622
623 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
624
625 leftmost = rb_add_cached(&p->pushable_dl_tasks,
626 &rq->dl.pushable_dl_tasks_root,
627 __pushable_less);
628 if (leftmost)
629 rq->dl.earliest_dl.next = p->dl.deadline;
630
631 if (!rq->dl.overloaded) {
632 dl_set_overload(rq);
633 rq->dl.overloaded = 1;
634 }
635 }
636
dequeue_pushable_dl_task(struct rq * rq,struct task_struct * p)637 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
638 {
639 struct dl_rq *dl_rq = &rq->dl;
640 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
641 struct rb_node *leftmost;
642
643 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
644 return;
645
646 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
647 if (leftmost)
648 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
649
650 RB_CLEAR_NODE(&p->pushable_dl_tasks);
651
652 if (!has_pushable_dl_tasks(rq) && rq->dl.overloaded) {
653 dl_clear_overload(rq);
654 rq->dl.overloaded = 0;
655 }
656 }
657
658 static int push_dl_task(struct rq *rq);
659
need_pull_dl_task(struct rq * rq,struct task_struct * prev)660 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
661 {
662 return rq->online && dl_task(prev);
663 }
664
665 static DEFINE_PER_CPU(struct balance_callback, dl_push_head);
666 static DEFINE_PER_CPU(struct balance_callback, dl_pull_head);
667
668 static void push_dl_tasks(struct rq *);
669 static void pull_dl_task(struct rq *);
670
deadline_queue_push_tasks(struct rq * rq)671 static inline void deadline_queue_push_tasks(struct rq *rq)
672 {
673 if (!has_pushable_dl_tasks(rq))
674 return;
675
676 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
677 }
678
deadline_queue_pull_task(struct rq * rq)679 static inline void deadline_queue_pull_task(struct rq *rq)
680 {
681 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
682 }
683
684 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
685
dl_task_offline_migration(struct rq * rq,struct task_struct * p)686 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
687 {
688 struct rq *later_rq = NULL;
689 struct dl_bw *dl_b;
690
691 later_rq = find_lock_later_rq(p, rq);
692 if (!later_rq) {
693 int cpu;
694
695 /*
696 * If we cannot preempt any rq, fall back to pick any
697 * online CPU:
698 */
699 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
700 if (cpu >= nr_cpu_ids) {
701 /*
702 * Failed to find any suitable CPU.
703 * The task will never come back!
704 */
705 WARN_ON_ONCE(dl_bandwidth_enabled());
706
707 /*
708 * If admission control is disabled we
709 * try a little harder to let the task
710 * run.
711 */
712 cpu = cpumask_any(cpu_active_mask);
713 }
714 later_rq = cpu_rq(cpu);
715 double_lock_balance(rq, later_rq);
716 }
717
718 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
719 /*
720 * Inactive timer is armed (or callback is running, but
721 * waiting for us to release rq locks). In any case, when it
722 * will fire (or continue), it will see running_bw of this
723 * task migrated to later_rq (and correctly handle it).
724 */
725 sub_running_bw(&p->dl, &rq->dl);
726 sub_rq_bw(&p->dl, &rq->dl);
727
728 add_rq_bw(&p->dl, &later_rq->dl);
729 add_running_bw(&p->dl, &later_rq->dl);
730 } else {
731 sub_rq_bw(&p->dl, &rq->dl);
732 add_rq_bw(&p->dl, &later_rq->dl);
733 }
734
735 /*
736 * And we finally need to fix up root_domain(s) bandwidth accounting,
737 * since p is still hanging out in the old (now moved to default) root
738 * domain.
739 */
740 dl_b = &rq->rd->dl_bw;
741 raw_spin_lock(&dl_b->lock);
742 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
743 raw_spin_unlock(&dl_b->lock);
744
745 dl_b = &later_rq->rd->dl_bw;
746 raw_spin_lock(&dl_b->lock);
747 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
748 raw_spin_unlock(&dl_b->lock);
749
750 set_task_cpu(p, later_rq->cpu);
751 double_unlock_balance(later_rq, rq);
752
753 return later_rq;
754 }
755
756 #else
757
758 static inline
enqueue_pushable_dl_task(struct rq * rq,struct task_struct * p)759 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
760 {
761 }
762
763 static inline
dequeue_pushable_dl_task(struct rq * rq,struct task_struct * p)764 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
765 {
766 }
767
768 static inline
inc_dl_migration(struct sched_dl_entity * dl_se,struct dl_rq * dl_rq)769 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
770 {
771 }
772
773 static inline
dec_dl_migration(struct sched_dl_entity * dl_se,struct dl_rq * dl_rq)774 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
775 {
776 }
777
deadline_queue_push_tasks(struct rq * rq)778 static inline void deadline_queue_push_tasks(struct rq *rq)
779 {
780 }
781
deadline_queue_pull_task(struct rq * rq)782 static inline void deadline_queue_pull_task(struct rq *rq)
783 {
784 }
785 #endif /* CONFIG_SMP */
786
787 static void
788 enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags);
789 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
790 static void dequeue_dl_entity(struct sched_dl_entity *dl_se, int flags);
791 static void wakeup_preempt_dl(struct rq *rq, struct task_struct *p, int flags);
792
replenish_dl_new_period(struct sched_dl_entity * dl_se,struct rq * rq)793 static inline void replenish_dl_new_period(struct sched_dl_entity *dl_se,
794 struct rq *rq)
795 {
796 /* for non-boosted task, pi_of(dl_se) == dl_se */
797 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
798 dl_se->runtime = pi_of(dl_se)->dl_runtime;
799
800 /*
801 * If it is a deferred reservation, and the server
802 * is not handling an starvation case, defer it.
803 */
804 if (dl_se->dl_defer && !dl_se->dl_defer_running) {
805 dl_se->dl_throttled = 1;
806 dl_se->dl_defer_armed = 1;
807 }
808 }
809
810 /*
811 * We are being explicitly informed that a new instance is starting,
812 * and this means that:
813 * - the absolute deadline of the entity has to be placed at
814 * current time + relative deadline;
815 * - the runtime of the entity has to be set to the maximum value.
816 *
817 * The capability of specifying such event is useful whenever a -deadline
818 * entity wants to (try to!) synchronize its behaviour with the scheduler's
819 * one, and to (try to!) reconcile itself with its own scheduling
820 * parameters.
821 */
setup_new_dl_entity(struct sched_dl_entity * dl_se)822 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
823 {
824 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
825 struct rq *rq = rq_of_dl_rq(dl_rq);
826
827 WARN_ON(is_dl_boosted(dl_se));
828 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
829
830 /*
831 * We are racing with the deadline timer. So, do nothing because
832 * the deadline timer handler will take care of properly recharging
833 * the runtime and postponing the deadline
834 */
835 if (dl_se->dl_throttled)
836 return;
837
838 /*
839 * We use the regular wall clock time to set deadlines in the
840 * future; in fact, we must consider execution overheads (time
841 * spent on hardirq context, etc.).
842 */
843 replenish_dl_new_period(dl_se, rq);
844 }
845
846 static int start_dl_timer(struct sched_dl_entity *dl_se);
847 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t);
848
849 /*
850 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
851 * possibility of a entity lasting more than what it declared, and thus
852 * exhausting its runtime.
853 *
854 * Here we are interested in making runtime overrun possible, but we do
855 * not want a entity which is misbehaving to affect the scheduling of all
856 * other entities.
857 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
858 * is used, in order to confine each entity within its own bandwidth.
859 *
860 * This function deals exactly with that, and ensures that when the runtime
861 * of a entity is replenished, its deadline is also postponed. That ensures
862 * the overrunning entity can't interfere with other entity in the system and
863 * can't make them miss their deadlines. Reasons why this kind of overruns
864 * could happen are, typically, a entity voluntarily trying to overcome its
865 * runtime, or it just underestimated it during sched_setattr().
866 */
replenish_dl_entity(struct sched_dl_entity * dl_se)867 static void replenish_dl_entity(struct sched_dl_entity *dl_se)
868 {
869 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
870 struct rq *rq = rq_of_dl_rq(dl_rq);
871
872 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime <= 0);
873
874 /*
875 * This could be the case for a !-dl task that is boosted.
876 * Just go with full inherited parameters.
877 *
878 * Or, it could be the case of a deferred reservation that
879 * was not able to consume its runtime in background and
880 * reached this point with current u > U.
881 *
882 * In both cases, set a new period.
883 */
884 if (dl_se->dl_deadline == 0 ||
885 (dl_se->dl_defer_armed && dl_entity_overflow(dl_se, rq_clock(rq)))) {
886 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
887 dl_se->runtime = pi_of(dl_se)->dl_runtime;
888 }
889
890 if (dl_se->dl_yielded && dl_se->runtime > 0)
891 dl_se->runtime = 0;
892
893 /*
894 * We keep moving the deadline away until we get some
895 * available runtime for the entity. This ensures correct
896 * handling of situations where the runtime overrun is
897 * arbitrary large.
898 */
899 while (dl_se->runtime <= 0) {
900 dl_se->deadline += pi_of(dl_se)->dl_period;
901 dl_se->runtime += pi_of(dl_se)->dl_runtime;
902 }
903
904 /*
905 * At this point, the deadline really should be "in
906 * the future" with respect to rq->clock. If it's
907 * not, we are, for some reason, lagging too much!
908 * Anyway, after having warn userspace abut that,
909 * we still try to keep the things running by
910 * resetting the deadline and the budget of the
911 * entity.
912 */
913 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
914 printk_deferred_once("sched: DL replenish lagged too much\n");
915 replenish_dl_new_period(dl_se, rq);
916 }
917
918 if (dl_se->dl_yielded)
919 dl_se->dl_yielded = 0;
920 if (dl_se->dl_throttled)
921 dl_se->dl_throttled = 0;
922
923 /*
924 * If this is the replenishment of a deferred reservation,
925 * clear the flag and return.
926 */
927 if (dl_se->dl_defer_armed) {
928 dl_se->dl_defer_armed = 0;
929 return;
930 }
931
932 /*
933 * A this point, if the deferred server is not armed, and the deadline
934 * is in the future, if it is not running already, throttle the server
935 * and arm the defer timer.
936 */
937 if (dl_se->dl_defer && !dl_se->dl_defer_running &&
938 dl_time_before(rq_clock(dl_se->rq), dl_se->deadline - dl_se->runtime)) {
939 if (!is_dl_boosted(dl_se) && dl_se->server_has_tasks(dl_se)) {
940
941 /*
942 * Set dl_se->dl_defer_armed and dl_throttled variables to
943 * inform the start_dl_timer() that this is a deferred
944 * activation.
945 */
946 dl_se->dl_defer_armed = 1;
947 dl_se->dl_throttled = 1;
948 if (!start_dl_timer(dl_se)) {
949 /*
950 * If for whatever reason (delays), a previous timer was
951 * queued but not serviced, cancel it and clean the
952 * deferrable server variables intended for start_dl_timer().
953 */
954 hrtimer_try_to_cancel(&dl_se->dl_timer);
955 dl_se->dl_defer_armed = 0;
956 dl_se->dl_throttled = 0;
957 }
958 }
959 }
960 }
961
962 /*
963 * Here we check if --at time t-- an entity (which is probably being
964 * [re]activated or, in general, enqueued) can use its remaining runtime
965 * and its current deadline _without_ exceeding the bandwidth it is
966 * assigned (function returns true if it can't). We are in fact applying
967 * one of the CBS rules: when a task wakes up, if the residual runtime
968 * over residual deadline fits within the allocated bandwidth, then we
969 * can keep the current (absolute) deadline and residual budget without
970 * disrupting the schedulability of the system. Otherwise, we should
971 * refill the runtime and set the deadline a period in the future,
972 * because keeping the current (absolute) deadline of the task would
973 * result in breaking guarantees promised to other tasks (refer to
974 * Documentation/scheduler/sched-deadline.rst for more information).
975 *
976 * This function returns true if:
977 *
978 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
979 *
980 * IOW we can't recycle current parameters.
981 *
982 * Notice that the bandwidth check is done against the deadline. For
983 * task with deadline equal to period this is the same of using
984 * dl_period instead of dl_deadline in the equation above.
985 */
dl_entity_overflow(struct sched_dl_entity * dl_se,u64 t)986 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
987 {
988 u64 left, right;
989
990 /*
991 * left and right are the two sides of the equation above,
992 * after a bit of shuffling to use multiplications instead
993 * of divisions.
994 *
995 * Note that none of the time values involved in the two
996 * multiplications are absolute: dl_deadline and dl_runtime
997 * are the relative deadline and the maximum runtime of each
998 * instance, runtime is the runtime left for the last instance
999 * and (deadline - t), since t is rq->clock, is the time left
1000 * to the (absolute) deadline. Even if overflowing the u64 type
1001 * is very unlikely to occur in both cases, here we scale down
1002 * as we want to avoid that risk at all. Scaling down by 10
1003 * means that we reduce granularity to 1us. We are fine with it,
1004 * since this is only a true/false check and, anyway, thinking
1005 * of anything below microseconds resolution is actually fiction
1006 * (but still we want to give the user that illusion >;).
1007 */
1008 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
1009 right = ((dl_se->deadline - t) >> DL_SCALE) *
1010 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
1011
1012 return dl_time_before(right, left);
1013 }
1014
1015 /*
1016 * Revised wakeup rule [1]: For self-suspending tasks, rather then
1017 * re-initializing task's runtime and deadline, the revised wakeup
1018 * rule adjusts the task's runtime to avoid the task to overrun its
1019 * density.
1020 *
1021 * Reasoning: a task may overrun the density if:
1022 * runtime / (deadline - t) > dl_runtime / dl_deadline
1023 *
1024 * Therefore, runtime can be adjusted to:
1025 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
1026 *
1027 * In such way that runtime will be equal to the maximum density
1028 * the task can use without breaking any rule.
1029 *
1030 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
1031 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
1032 */
1033 static void
update_dl_revised_wakeup(struct sched_dl_entity * dl_se,struct rq * rq)1034 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
1035 {
1036 u64 laxity = dl_se->deadline - rq_clock(rq);
1037
1038 /*
1039 * If the task has deadline < period, and the deadline is in the past,
1040 * it should already be throttled before this check.
1041 *
1042 * See update_dl_entity() comments for further details.
1043 */
1044 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
1045
1046 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
1047 }
1048
1049 /*
1050 * Regarding the deadline, a task with implicit deadline has a relative
1051 * deadline == relative period. A task with constrained deadline has a
1052 * relative deadline <= relative period.
1053 *
1054 * We support constrained deadline tasks. However, there are some restrictions
1055 * applied only for tasks which do not have an implicit deadline. See
1056 * update_dl_entity() to know more about such restrictions.
1057 *
1058 * The dl_is_implicit() returns true if the task has an implicit deadline.
1059 */
dl_is_implicit(struct sched_dl_entity * dl_se)1060 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
1061 {
1062 return dl_se->dl_deadline == dl_se->dl_period;
1063 }
1064
1065 /*
1066 * When a deadline entity is placed in the runqueue, its runtime and deadline
1067 * might need to be updated. This is done by a CBS wake up rule. There are two
1068 * different rules: 1) the original CBS; and 2) the Revisited CBS.
1069 *
1070 * When the task is starting a new period, the Original CBS is used. In this
1071 * case, the runtime is replenished and a new absolute deadline is set.
1072 *
1073 * When a task is queued before the begin of the next period, using the
1074 * remaining runtime and deadline could make the entity to overflow, see
1075 * dl_entity_overflow() to find more about runtime overflow. When such case
1076 * is detected, the runtime and deadline need to be updated.
1077 *
1078 * If the task has an implicit deadline, i.e., deadline == period, the Original
1079 * CBS is applied. The runtime is replenished and a new absolute deadline is
1080 * set, as in the previous cases.
1081 *
1082 * However, the Original CBS does not work properly for tasks with
1083 * deadline < period, which are said to have a constrained deadline. By
1084 * applying the Original CBS, a constrained deadline task would be able to run
1085 * runtime/deadline in a period. With deadline < period, the task would
1086 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1087 *
1088 * In order to prevent this misbehave, the Revisited CBS is used for
1089 * constrained deadline tasks when a runtime overflow is detected. In the
1090 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1091 * the remaining runtime of the task is reduced to avoid runtime overflow.
1092 * Please refer to the comments update_dl_revised_wakeup() function to find
1093 * more about the Revised CBS rule.
1094 */
update_dl_entity(struct sched_dl_entity * dl_se)1095 static void update_dl_entity(struct sched_dl_entity *dl_se)
1096 {
1097 struct rq *rq = rq_of_dl_se(dl_se);
1098
1099 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
1100 dl_entity_overflow(dl_se, rq_clock(rq))) {
1101
1102 if (unlikely(!dl_is_implicit(dl_se) &&
1103 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1104 !is_dl_boosted(dl_se))) {
1105 update_dl_revised_wakeup(dl_se, rq);
1106 return;
1107 }
1108
1109 replenish_dl_new_period(dl_se, rq);
1110 } else if (dl_server(dl_se) && dl_se->dl_defer) {
1111 /*
1112 * The server can still use its previous deadline, so check if
1113 * it left the dl_defer_running state.
1114 */
1115 if (!dl_se->dl_defer_running) {
1116 dl_se->dl_defer_armed = 1;
1117 dl_se->dl_throttled = 1;
1118 }
1119 }
1120 }
1121
dl_next_period(struct sched_dl_entity * dl_se)1122 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
1123 {
1124 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1125 }
1126
1127 /*
1128 * If the entity depleted all its runtime, and if we want it to sleep
1129 * while waiting for some new execution time to become available, we
1130 * set the bandwidth replenishment timer to the replenishment instant
1131 * and try to activate it.
1132 *
1133 * Notice that it is important for the caller to know if the timer
1134 * actually started or not (i.e., the replenishment instant is in
1135 * the future or in the past).
1136 */
start_dl_timer(struct sched_dl_entity * dl_se)1137 static int start_dl_timer(struct sched_dl_entity *dl_se)
1138 {
1139 struct hrtimer *timer = &dl_se->dl_timer;
1140 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1141 struct rq *rq = rq_of_dl_rq(dl_rq);
1142 ktime_t now, act;
1143 s64 delta;
1144
1145 lockdep_assert_rq_held(rq);
1146
1147 /*
1148 * We want the timer to fire at the deadline, but considering
1149 * that it is actually coming from rq->clock and not from
1150 * hrtimer's time base reading.
1151 *
1152 * The deferred reservation will have its timer set to
1153 * (deadline - runtime). At that point, the CBS rule will decide
1154 * if the current deadline can be used, or if a replenishment is
1155 * required to avoid add too much pressure on the system
1156 * (current u > U).
1157 */
1158 if (dl_se->dl_defer_armed) {
1159 WARN_ON_ONCE(!dl_se->dl_throttled);
1160 act = ns_to_ktime(dl_se->deadline - dl_se->runtime);
1161 } else {
1162 /* act = deadline - rel-deadline + period */
1163 act = ns_to_ktime(dl_next_period(dl_se));
1164 }
1165
1166 now = hrtimer_cb_get_time(timer);
1167 delta = ktime_to_ns(now) - rq_clock(rq);
1168 act = ktime_add_ns(act, delta);
1169
1170 /*
1171 * If the expiry time already passed, e.g., because the value
1172 * chosen as the deadline is too small, don't even try to
1173 * start the timer in the past!
1174 */
1175 if (ktime_us_delta(act, now) < 0)
1176 return 0;
1177
1178 /*
1179 * !enqueued will guarantee another callback; even if one is already in
1180 * progress. This ensures a balanced {get,put}_task_struct().
1181 *
1182 * The race against __run_timer() clearing the enqueued state is
1183 * harmless because we're holding task_rq()->lock, therefore the timer
1184 * expiring after we've done the check will wait on its task_rq_lock()
1185 * and observe our state.
1186 */
1187 if (!hrtimer_is_queued(timer)) {
1188 if (!dl_server(dl_se))
1189 get_task_struct(dl_task_of(dl_se));
1190 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1191 }
1192
1193 return 1;
1194 }
1195
__push_dl_task(struct rq * rq,struct rq_flags * rf)1196 static void __push_dl_task(struct rq *rq, struct rq_flags *rf)
1197 {
1198 #ifdef CONFIG_SMP
1199 /*
1200 * Queueing this task back might have overloaded rq, check if we need
1201 * to kick someone away.
1202 */
1203 if (has_pushable_dl_tasks(rq)) {
1204 /*
1205 * Nothing relies on rq->lock after this, so its safe to drop
1206 * rq->lock.
1207 */
1208 rq_unpin_lock(rq, rf);
1209 push_dl_task(rq);
1210 rq_repin_lock(rq, rf);
1211 }
1212 #endif
1213 }
1214
1215 /* a defer timer will not be reset if the runtime consumed was < dl_server_min_res */
1216 static const u64 dl_server_min_res = 1 * NSEC_PER_MSEC;
1217
dl_server_timer(struct hrtimer * timer,struct sched_dl_entity * dl_se)1218 static enum hrtimer_restart dl_server_timer(struct hrtimer *timer, struct sched_dl_entity *dl_se)
1219 {
1220 struct rq *rq = rq_of_dl_se(dl_se);
1221 u64 fw;
1222
1223 scoped_guard (rq_lock, rq) {
1224 struct rq_flags *rf = &scope.rf;
1225
1226 if (!dl_se->dl_throttled || !dl_se->dl_runtime)
1227 return HRTIMER_NORESTART;
1228
1229 sched_clock_tick();
1230 update_rq_clock(rq);
1231
1232 if (!dl_se->dl_runtime)
1233 return HRTIMER_NORESTART;
1234
1235 if (!dl_se->server_has_tasks(dl_se)) {
1236 replenish_dl_entity(dl_se);
1237 return HRTIMER_NORESTART;
1238 }
1239
1240 if (dl_se->dl_defer_armed) {
1241 /*
1242 * First check if the server could consume runtime in background.
1243 * If so, it is possible to push the defer timer for this amount
1244 * of time. The dl_server_min_res serves as a limit to avoid
1245 * forwarding the timer for a too small amount of time.
1246 */
1247 if (dl_time_before(rq_clock(dl_se->rq),
1248 (dl_se->deadline - dl_se->runtime - dl_server_min_res))) {
1249
1250 /* reset the defer timer */
1251 fw = dl_se->deadline - rq_clock(dl_se->rq) - dl_se->runtime;
1252
1253 hrtimer_forward_now(timer, ns_to_ktime(fw));
1254 return HRTIMER_RESTART;
1255 }
1256
1257 dl_se->dl_defer_running = 1;
1258 }
1259
1260 enqueue_dl_entity(dl_se, ENQUEUE_REPLENISH);
1261
1262 if (!dl_task(dl_se->rq->curr) || dl_entity_preempt(dl_se, &dl_se->rq->curr->dl))
1263 resched_curr(rq);
1264
1265 __push_dl_task(rq, rf);
1266 }
1267
1268 return HRTIMER_NORESTART;
1269 }
1270
1271 /*
1272 * This is the bandwidth enforcement timer callback. If here, we know
1273 * a task is not on its dl_rq, since the fact that the timer was running
1274 * means the task is throttled and needs a runtime replenishment.
1275 *
1276 * However, what we actually do depends on the fact the task is active,
1277 * (it is on its rq) or has been removed from there by a call to
1278 * dequeue_task_dl(). In the former case we must issue the runtime
1279 * replenishment and add the task back to the dl_rq; in the latter, we just
1280 * do nothing but clearing dl_throttled, so that runtime and deadline
1281 * updating (and the queueing back to dl_rq) will be done by the
1282 * next call to enqueue_task_dl().
1283 */
dl_task_timer(struct hrtimer * timer)1284 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1285 {
1286 struct sched_dl_entity *dl_se = container_of(timer,
1287 struct sched_dl_entity,
1288 dl_timer);
1289 struct task_struct *p;
1290 struct rq_flags rf;
1291 struct rq *rq;
1292
1293 if (dl_server(dl_se))
1294 return dl_server_timer(timer, dl_se);
1295
1296 p = dl_task_of(dl_se);
1297 rq = task_rq_lock(p, &rf);
1298
1299 /*
1300 * The task might have changed its scheduling policy to something
1301 * different than SCHED_DEADLINE (through switched_from_dl()).
1302 */
1303 if (!dl_task(p))
1304 goto unlock;
1305
1306 /*
1307 * The task might have been boosted by someone else and might be in the
1308 * boosting/deboosting path, its not throttled.
1309 */
1310 if (is_dl_boosted(dl_se))
1311 goto unlock;
1312
1313 /*
1314 * Spurious timer due to start_dl_timer() race; or we already received
1315 * a replenishment from rt_mutex_setprio().
1316 */
1317 if (!dl_se->dl_throttled)
1318 goto unlock;
1319
1320 sched_clock_tick();
1321 update_rq_clock(rq);
1322
1323 /*
1324 * If the throttle happened during sched-out; like:
1325 *
1326 * schedule()
1327 * deactivate_task()
1328 * dequeue_task_dl()
1329 * update_curr_dl()
1330 * start_dl_timer()
1331 * __dequeue_task_dl()
1332 * prev->on_rq = 0;
1333 *
1334 * We can be both throttled and !queued. Replenish the counter
1335 * but do not enqueue -- wait for our wakeup to do that.
1336 */
1337 if (!task_on_rq_queued(p)) {
1338 replenish_dl_entity(dl_se);
1339 goto unlock;
1340 }
1341
1342 #ifdef CONFIG_SMP
1343 if (unlikely(!rq->online)) {
1344 /*
1345 * If the runqueue is no longer available, migrate the
1346 * task elsewhere. This necessarily changes rq.
1347 */
1348 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1349 rq = dl_task_offline_migration(rq, p);
1350 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1351 update_rq_clock(rq);
1352
1353 /*
1354 * Now that the task has been migrated to the new RQ and we
1355 * have that locked, proceed as normal and enqueue the task
1356 * there.
1357 */
1358 }
1359 #endif
1360
1361 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1362 if (dl_task(rq->donor))
1363 wakeup_preempt_dl(rq, p, 0);
1364 else
1365 resched_curr(rq);
1366
1367 __push_dl_task(rq, &rf);
1368
1369 unlock:
1370 task_rq_unlock(rq, p, &rf);
1371
1372 /*
1373 * This can free the task_struct, including this hrtimer, do not touch
1374 * anything related to that after this.
1375 */
1376 put_task_struct(p);
1377
1378 return HRTIMER_NORESTART;
1379 }
1380
init_dl_task_timer(struct sched_dl_entity * dl_se)1381 static void init_dl_task_timer(struct sched_dl_entity *dl_se)
1382 {
1383 struct hrtimer *timer = &dl_se->dl_timer;
1384
1385 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1386 timer->function = dl_task_timer;
1387 }
1388
1389 /*
1390 * During the activation, CBS checks if it can reuse the current task's
1391 * runtime and period. If the deadline of the task is in the past, CBS
1392 * cannot use the runtime, and so it replenishes the task. This rule
1393 * works fine for implicit deadline tasks (deadline == period), and the
1394 * CBS was designed for implicit deadline tasks. However, a task with
1395 * constrained deadline (deadline < period) might be awakened after the
1396 * deadline, but before the next period. In this case, replenishing the
1397 * task would allow it to run for runtime / deadline. As in this case
1398 * deadline < period, CBS enables a task to run for more than the
1399 * runtime / period. In a very loaded system, this can cause a domino
1400 * effect, making other tasks miss their deadlines.
1401 *
1402 * To avoid this problem, in the activation of a constrained deadline
1403 * task after the deadline but before the next period, throttle the
1404 * task and set the replenishing timer to the begin of the next period,
1405 * unless it is boosted.
1406 */
dl_check_constrained_dl(struct sched_dl_entity * dl_se)1407 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1408 {
1409 struct rq *rq = rq_of_dl_se(dl_se);
1410
1411 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1412 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1413 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(dl_se)))
1414 return;
1415 dl_se->dl_throttled = 1;
1416 if (dl_se->runtime > 0)
1417 dl_se->runtime = 0;
1418 }
1419 }
1420
1421 static
dl_runtime_exceeded(struct sched_dl_entity * dl_se)1422 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1423 {
1424 return (dl_se->runtime <= 0);
1425 }
1426
1427 /*
1428 * This function implements the GRUB accounting rule. According to the
1429 * GRUB reclaiming algorithm, the runtime is not decreased as "dq = -dt",
1430 * but as "dq = -(max{u, (Umax - Uinact - Uextra)} / Umax) dt",
1431 * where u is the utilization of the task, Umax is the maximum reclaimable
1432 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1433 * as the difference between the "total runqueue utilization" and the
1434 * "runqueue active utilization", and Uextra is the (per runqueue) extra
1435 * reclaimable utilization.
1436 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations multiplied
1437 * by 2^BW_SHIFT, the result has to be shifted right by BW_SHIFT.
1438 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT, dl_bw
1439 * is multiplied by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1440 * Since delta is a 64 bit variable, to have an overflow its value should be
1441 * larger than 2^(64 - 20 - 8), which is more than 64 seconds. So, overflow is
1442 * not an issue here.
1443 */
grub_reclaim(u64 delta,struct rq * rq,struct sched_dl_entity * dl_se)1444 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1445 {
1446 u64 u_act;
1447 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1448
1449 /*
1450 * Instead of computing max{u, (u_max - u_inact - u_extra)}, we
1451 * compare u_inact + u_extra with u_max - u, because u_inact + u_extra
1452 * can be larger than u_max. So, u_max - u_inact - u_extra would be
1453 * negative leading to wrong results.
1454 */
1455 if (u_inact + rq->dl.extra_bw > rq->dl.max_bw - dl_se->dl_bw)
1456 u_act = dl_se->dl_bw;
1457 else
1458 u_act = rq->dl.max_bw - u_inact - rq->dl.extra_bw;
1459
1460 u_act = (u_act * rq->dl.bw_ratio) >> RATIO_SHIFT;
1461 return (delta * u_act) >> BW_SHIFT;
1462 }
1463
dl_scaled_delta_exec(struct rq * rq,struct sched_dl_entity * dl_se,s64 delta_exec)1464 s64 dl_scaled_delta_exec(struct rq *rq, struct sched_dl_entity *dl_se, s64 delta_exec)
1465 {
1466 s64 scaled_delta_exec;
1467
1468 /*
1469 * For tasks that participate in GRUB, we implement GRUB-PA: the
1470 * spare reclaimed bandwidth is used to clock down frequency.
1471 *
1472 * For the others, we still need to scale reservation parameters
1473 * according to current frequency and CPU maximum capacity.
1474 */
1475 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1476 scaled_delta_exec = grub_reclaim(delta_exec, rq, dl_se);
1477 } else {
1478 int cpu = cpu_of(rq);
1479 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1480 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1481
1482 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1483 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1484 }
1485
1486 return scaled_delta_exec;
1487 }
1488
1489 static inline void
1490 update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1491 int flags);
update_curr_dl_se(struct rq * rq,struct sched_dl_entity * dl_se,s64 delta_exec)1492 static void update_curr_dl_se(struct rq *rq, struct sched_dl_entity *dl_se, s64 delta_exec)
1493 {
1494 s64 scaled_delta_exec;
1495
1496 if (unlikely(delta_exec <= 0)) {
1497 if (unlikely(dl_se->dl_yielded))
1498 goto throttle;
1499 return;
1500 }
1501
1502 if (dl_server(dl_se) && dl_se->dl_throttled && !dl_se->dl_defer)
1503 return;
1504
1505 if (dl_entity_is_special(dl_se))
1506 return;
1507
1508 scaled_delta_exec = dl_scaled_delta_exec(rq, dl_se, delta_exec);
1509
1510 dl_se->runtime -= scaled_delta_exec;
1511
1512 /*
1513 * The fair server can consume its runtime while throttled (not queued/
1514 * running as regular CFS).
1515 *
1516 * If the server consumes its entire runtime in this state. The server
1517 * is not required for the current period. Thus, reset the server by
1518 * starting a new period, pushing the activation.
1519 */
1520 if (dl_se->dl_defer && dl_se->dl_throttled && dl_runtime_exceeded(dl_se)) {
1521 /*
1522 * If the server was previously activated - the starving condition
1523 * took place, it this point it went away because the fair scheduler
1524 * was able to get runtime in background. So return to the initial
1525 * state.
1526 */
1527 dl_se->dl_defer_running = 0;
1528
1529 hrtimer_try_to_cancel(&dl_se->dl_timer);
1530
1531 replenish_dl_new_period(dl_se, dl_se->rq);
1532
1533 /*
1534 * Not being able to start the timer seems problematic. If it could not
1535 * be started for whatever reason, we need to "unthrottle" the DL server
1536 * and queue right away. Otherwise nothing might queue it. That's similar
1537 * to what enqueue_dl_entity() does on start_dl_timer==0. For now, just warn.
1538 */
1539 WARN_ON_ONCE(!start_dl_timer(dl_se));
1540
1541 return;
1542 }
1543
1544 throttle:
1545 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1546 dl_se->dl_throttled = 1;
1547
1548 /* If requested, inform the user about runtime overruns. */
1549 if (dl_runtime_exceeded(dl_se) &&
1550 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1551 dl_se->dl_overrun = 1;
1552
1553 dequeue_dl_entity(dl_se, 0);
1554 if (!dl_server(dl_se)) {
1555 update_stats_dequeue_dl(&rq->dl, dl_se, 0);
1556 dequeue_pushable_dl_task(rq, dl_task_of(dl_se));
1557 }
1558
1559 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(dl_se))) {
1560 if (dl_server(dl_se))
1561 enqueue_dl_entity(dl_se, ENQUEUE_REPLENISH);
1562 else
1563 enqueue_task_dl(rq, dl_task_of(dl_se), ENQUEUE_REPLENISH);
1564 }
1565
1566 if (!is_leftmost(dl_se, &rq->dl))
1567 resched_curr(rq);
1568 }
1569
1570 /*
1571 * The fair server (sole dl_server) does not account for real-time
1572 * workload because it is running fair work.
1573 */
1574 if (dl_se == &rq->fair_server)
1575 return;
1576
1577 #ifdef CONFIG_RT_GROUP_SCHED
1578 /*
1579 * Because -- for now -- we share the rt bandwidth, we need to
1580 * account our runtime there too, otherwise actual rt tasks
1581 * would be able to exceed the shared quota.
1582 *
1583 * Account to the root rt group for now.
1584 *
1585 * The solution we're working towards is having the RT groups scheduled
1586 * using deadline servers -- however there's a few nasties to figure
1587 * out before that can happen.
1588 */
1589 if (rt_bandwidth_enabled()) {
1590 struct rt_rq *rt_rq = &rq->rt;
1591
1592 raw_spin_lock(&rt_rq->rt_runtime_lock);
1593 /*
1594 * We'll let actual RT tasks worry about the overflow here, we
1595 * have our own CBS to keep us inline; only account when RT
1596 * bandwidth is relevant.
1597 */
1598 if (sched_rt_bandwidth_account(rt_rq))
1599 rt_rq->rt_time += delta_exec;
1600 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1601 }
1602 #endif
1603 }
1604
1605 /*
1606 * In the non-defer mode, the idle time is not accounted, as the
1607 * server provides a guarantee.
1608 *
1609 * If the dl_server is in defer mode, the idle time is also considered
1610 * as time available for the fair server, avoiding a penalty for the
1611 * rt scheduler that did not consumed that time.
1612 */
dl_server_update_idle_time(struct rq * rq,struct task_struct * p)1613 void dl_server_update_idle_time(struct rq *rq, struct task_struct *p)
1614 {
1615 s64 delta_exec, scaled_delta_exec;
1616
1617 if (!rq->fair_server.dl_defer)
1618 return;
1619
1620 /* no need to discount more */
1621 if (rq->fair_server.runtime < 0)
1622 return;
1623
1624 delta_exec = rq_clock_task(rq) - p->se.exec_start;
1625 if (delta_exec < 0)
1626 return;
1627
1628 scaled_delta_exec = dl_scaled_delta_exec(rq, &rq->fair_server, delta_exec);
1629
1630 rq->fair_server.runtime -= scaled_delta_exec;
1631
1632 if (rq->fair_server.runtime < 0) {
1633 rq->fair_server.dl_defer_running = 0;
1634 rq->fair_server.runtime = 0;
1635 }
1636
1637 p->se.exec_start = rq_clock_task(rq);
1638 }
1639
dl_server_update(struct sched_dl_entity * dl_se,s64 delta_exec)1640 void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec)
1641 {
1642 /* 0 runtime = fair server disabled */
1643 if (dl_se->dl_runtime)
1644 update_curr_dl_se(dl_se->rq, dl_se, delta_exec);
1645 }
1646
dl_server_start(struct sched_dl_entity * dl_se)1647 void dl_server_start(struct sched_dl_entity *dl_se)
1648 {
1649 struct rq *rq = dl_se->rq;
1650
1651 /*
1652 * XXX: the apply do not work fine at the init phase for the
1653 * fair server because things are not yet set. We need to improve
1654 * this before getting generic.
1655 */
1656 if (!dl_server(dl_se)) {
1657 u64 runtime = 50 * NSEC_PER_MSEC;
1658 u64 period = 1000 * NSEC_PER_MSEC;
1659
1660 dl_server_apply_params(dl_se, runtime, period, 1);
1661
1662 dl_se->dl_server = 1;
1663 dl_se->dl_defer = 1;
1664 setup_new_dl_entity(dl_se);
1665 }
1666
1667 if (!dl_se->dl_runtime)
1668 return;
1669
1670 dl_se->dl_server_active = 1;
1671 enqueue_dl_entity(dl_se, ENQUEUE_WAKEUP);
1672 if (!dl_task(dl_se->rq->curr) || dl_entity_preempt(dl_se, &rq->curr->dl))
1673 resched_curr(dl_se->rq);
1674 }
1675
dl_server_stop(struct sched_dl_entity * dl_se)1676 void dl_server_stop(struct sched_dl_entity *dl_se)
1677 {
1678 if (!dl_se->dl_runtime)
1679 return;
1680
1681 dequeue_dl_entity(dl_se, DEQUEUE_SLEEP);
1682 hrtimer_try_to_cancel(&dl_se->dl_timer);
1683 dl_se->dl_defer_armed = 0;
1684 dl_se->dl_throttled = 0;
1685 dl_se->dl_server_active = 0;
1686 }
1687
dl_server_init(struct sched_dl_entity * dl_se,struct rq * rq,dl_server_has_tasks_f has_tasks,dl_server_pick_f pick_task)1688 void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
1689 dl_server_has_tasks_f has_tasks,
1690 dl_server_pick_f pick_task)
1691 {
1692 dl_se->rq = rq;
1693 dl_se->server_has_tasks = has_tasks;
1694 dl_se->server_pick_task = pick_task;
1695 }
1696
__dl_server_attach_root(struct sched_dl_entity * dl_se,struct rq * rq)1697 void __dl_server_attach_root(struct sched_dl_entity *dl_se, struct rq *rq)
1698 {
1699 u64 new_bw = dl_se->dl_bw;
1700 int cpu = cpu_of(rq);
1701 struct dl_bw *dl_b;
1702
1703 dl_b = dl_bw_of(cpu_of(rq));
1704 guard(raw_spinlock)(&dl_b->lock);
1705
1706 if (!dl_bw_cpus(cpu))
1707 return;
1708
1709 __dl_add(dl_b, new_bw, dl_bw_cpus(cpu));
1710 }
1711
dl_server_apply_params(struct sched_dl_entity * dl_se,u64 runtime,u64 period,bool init)1712 int dl_server_apply_params(struct sched_dl_entity *dl_se, u64 runtime, u64 period, bool init)
1713 {
1714 u64 old_bw = init ? 0 : to_ratio(dl_se->dl_period, dl_se->dl_runtime);
1715 u64 new_bw = to_ratio(period, runtime);
1716 struct rq *rq = dl_se->rq;
1717 int cpu = cpu_of(rq);
1718 struct dl_bw *dl_b;
1719 unsigned long cap;
1720 int retval = 0;
1721 int cpus;
1722
1723 dl_b = dl_bw_of(cpu);
1724 guard(raw_spinlock)(&dl_b->lock);
1725
1726 cpus = dl_bw_cpus(cpu);
1727 cap = dl_bw_capacity(cpu);
1728
1729 if (__dl_overflow(dl_b, cap, old_bw, new_bw))
1730 return -EBUSY;
1731
1732 if (init) {
1733 __add_rq_bw(new_bw, &rq->dl);
1734 __dl_add(dl_b, new_bw, cpus);
1735 } else {
1736 __dl_sub(dl_b, dl_se->dl_bw, cpus);
1737 __dl_add(dl_b, new_bw, cpus);
1738
1739 dl_rq_change_utilization(rq, dl_se, new_bw);
1740 }
1741
1742 dl_se->dl_runtime = runtime;
1743 dl_se->dl_deadline = period;
1744 dl_se->dl_period = period;
1745
1746 dl_se->runtime = 0;
1747 dl_se->deadline = 0;
1748
1749 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
1750 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
1751
1752 return retval;
1753 }
1754
1755 /*
1756 * Update the current task's runtime statistics (provided it is still
1757 * a -deadline task and has not been removed from the dl_rq).
1758 */
update_curr_dl(struct rq * rq)1759 static void update_curr_dl(struct rq *rq)
1760 {
1761 struct task_struct *donor = rq->donor;
1762 struct sched_dl_entity *dl_se = &donor->dl;
1763 s64 delta_exec;
1764
1765 if (!dl_task(donor) || !on_dl_rq(dl_se))
1766 return;
1767
1768 /*
1769 * Consumed budget is computed considering the time as
1770 * observed by schedulable tasks (excluding time spent
1771 * in hardirq context, etc.). Deadlines are instead
1772 * computed using hard walltime. This seems to be the more
1773 * natural solution, but the full ramifications of this
1774 * approach need further study.
1775 */
1776 delta_exec = update_curr_common(rq);
1777 update_curr_dl_se(rq, dl_se, delta_exec);
1778 }
1779
inactive_task_timer(struct hrtimer * timer)1780 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1781 {
1782 struct sched_dl_entity *dl_se = container_of(timer,
1783 struct sched_dl_entity,
1784 inactive_timer);
1785 struct task_struct *p = NULL;
1786 struct rq_flags rf;
1787 struct rq *rq;
1788
1789 if (!dl_server(dl_se)) {
1790 p = dl_task_of(dl_se);
1791 rq = task_rq_lock(p, &rf);
1792 } else {
1793 rq = dl_se->rq;
1794 rq_lock(rq, &rf);
1795 }
1796
1797 sched_clock_tick();
1798 update_rq_clock(rq);
1799
1800 if (dl_server(dl_se))
1801 goto no_task;
1802
1803 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1804 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1805
1806 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1807 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1808 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1809 dl_se->dl_non_contending = 0;
1810 }
1811
1812 raw_spin_lock(&dl_b->lock);
1813 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1814 raw_spin_unlock(&dl_b->lock);
1815 __dl_clear_params(dl_se);
1816
1817 goto unlock;
1818 }
1819
1820 no_task:
1821 if (dl_se->dl_non_contending == 0)
1822 goto unlock;
1823
1824 sub_running_bw(dl_se, &rq->dl);
1825 dl_se->dl_non_contending = 0;
1826 unlock:
1827
1828 if (!dl_server(dl_se)) {
1829 task_rq_unlock(rq, p, &rf);
1830 put_task_struct(p);
1831 } else {
1832 rq_unlock(rq, &rf);
1833 }
1834
1835 return HRTIMER_NORESTART;
1836 }
1837
init_dl_inactive_task_timer(struct sched_dl_entity * dl_se)1838 static void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1839 {
1840 struct hrtimer *timer = &dl_se->inactive_timer;
1841
1842 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1843 timer->function = inactive_task_timer;
1844 }
1845
1846 #define __node_2_dle(node) \
1847 rb_entry((node), struct sched_dl_entity, rb_node)
1848
1849 #ifdef CONFIG_SMP
1850
inc_dl_deadline(struct dl_rq * dl_rq,u64 deadline)1851 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1852 {
1853 struct rq *rq = rq_of_dl_rq(dl_rq);
1854
1855 if (dl_rq->earliest_dl.curr == 0 ||
1856 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1857 if (dl_rq->earliest_dl.curr == 0)
1858 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1859 dl_rq->earliest_dl.curr = deadline;
1860 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1861 }
1862 }
1863
dec_dl_deadline(struct dl_rq * dl_rq,u64 deadline)1864 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1865 {
1866 struct rq *rq = rq_of_dl_rq(dl_rq);
1867
1868 /*
1869 * Since we may have removed our earliest (and/or next earliest)
1870 * task we must recompute them.
1871 */
1872 if (!dl_rq->dl_nr_running) {
1873 dl_rq->earliest_dl.curr = 0;
1874 dl_rq->earliest_dl.next = 0;
1875 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1876 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1877 } else {
1878 struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1879 struct sched_dl_entity *entry = __node_2_dle(leftmost);
1880
1881 dl_rq->earliest_dl.curr = entry->deadline;
1882 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1883 }
1884 }
1885
1886 #else
1887
inc_dl_deadline(struct dl_rq * dl_rq,u64 deadline)1888 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
dec_dl_deadline(struct dl_rq * dl_rq,u64 deadline)1889 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1890
1891 #endif /* CONFIG_SMP */
1892
1893 static inline
inc_dl_tasks(struct sched_dl_entity * dl_se,struct dl_rq * dl_rq)1894 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1895 {
1896 u64 deadline = dl_se->deadline;
1897
1898 dl_rq->dl_nr_running++;
1899 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1900
1901 inc_dl_deadline(dl_rq, deadline);
1902 }
1903
1904 static inline
dec_dl_tasks(struct sched_dl_entity * dl_se,struct dl_rq * dl_rq)1905 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1906 {
1907 WARN_ON(!dl_rq->dl_nr_running);
1908 dl_rq->dl_nr_running--;
1909 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1910
1911 dec_dl_deadline(dl_rq, dl_se->deadline);
1912 }
1913
__dl_less(struct rb_node * a,const struct rb_node * b)1914 static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1915 {
1916 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1917 }
1918
1919 static __always_inline struct sched_statistics *
__schedstats_from_dl_se(struct sched_dl_entity * dl_se)1920 __schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1921 {
1922 if (!schedstat_enabled())
1923 return NULL;
1924
1925 if (dl_server(dl_se))
1926 return NULL;
1927
1928 return &dl_task_of(dl_se)->stats;
1929 }
1930
1931 static inline void
update_stats_wait_start_dl(struct dl_rq * dl_rq,struct sched_dl_entity * dl_se)1932 update_stats_wait_start_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1933 {
1934 struct sched_statistics *stats = __schedstats_from_dl_se(dl_se);
1935 if (stats)
1936 __update_stats_wait_start(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1937 }
1938
1939 static inline void
update_stats_wait_end_dl(struct dl_rq * dl_rq,struct sched_dl_entity * dl_se)1940 update_stats_wait_end_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1941 {
1942 struct sched_statistics *stats = __schedstats_from_dl_se(dl_se);
1943 if (stats)
1944 __update_stats_wait_end(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1945 }
1946
1947 static inline void
update_stats_enqueue_sleeper_dl(struct dl_rq * dl_rq,struct sched_dl_entity * dl_se)1948 update_stats_enqueue_sleeper_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1949 {
1950 struct sched_statistics *stats = __schedstats_from_dl_se(dl_se);
1951 if (stats)
1952 __update_stats_enqueue_sleeper(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1953 }
1954
1955 static inline void
update_stats_enqueue_dl(struct dl_rq * dl_rq,struct sched_dl_entity * dl_se,int flags)1956 update_stats_enqueue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1957 int flags)
1958 {
1959 if (!schedstat_enabled())
1960 return;
1961
1962 if (flags & ENQUEUE_WAKEUP)
1963 update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1964 }
1965
1966 static inline void
update_stats_dequeue_dl(struct dl_rq * dl_rq,struct sched_dl_entity * dl_se,int flags)1967 update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1968 int flags)
1969 {
1970 struct task_struct *p = dl_task_of(dl_se);
1971
1972 if (!schedstat_enabled())
1973 return;
1974
1975 if ((flags & DEQUEUE_SLEEP)) {
1976 unsigned int state;
1977
1978 state = READ_ONCE(p->__state);
1979 if (state & TASK_INTERRUPTIBLE)
1980 __schedstat_set(p->stats.sleep_start,
1981 rq_clock(rq_of_dl_rq(dl_rq)));
1982
1983 if (state & TASK_UNINTERRUPTIBLE)
1984 __schedstat_set(p->stats.block_start,
1985 rq_clock(rq_of_dl_rq(dl_rq)));
1986 }
1987 }
1988
__enqueue_dl_entity(struct sched_dl_entity * dl_se)1989 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1990 {
1991 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1992
1993 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->rb_node));
1994
1995 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1996
1997 inc_dl_tasks(dl_se, dl_rq);
1998 }
1999
__dequeue_dl_entity(struct sched_dl_entity * dl_se)2000 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
2001 {
2002 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
2003
2004 if (RB_EMPTY_NODE(&dl_se->rb_node))
2005 return;
2006
2007 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
2008
2009 RB_CLEAR_NODE(&dl_se->rb_node);
2010
2011 dec_dl_tasks(dl_se, dl_rq);
2012 }
2013
2014 static void
enqueue_dl_entity(struct sched_dl_entity * dl_se,int flags)2015 enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
2016 {
2017 WARN_ON_ONCE(on_dl_rq(dl_se));
2018
2019 update_stats_enqueue_dl(dl_rq_of_se(dl_se), dl_se, flags);
2020
2021 /*
2022 * Check if a constrained deadline task was activated
2023 * after the deadline but before the next period.
2024 * If that is the case, the task will be throttled and
2025 * the replenishment timer will be set to the next period.
2026 */
2027 if (!dl_se->dl_throttled && !dl_is_implicit(dl_se))
2028 dl_check_constrained_dl(dl_se);
2029
2030 if (flags & (ENQUEUE_RESTORE|ENQUEUE_MIGRATING)) {
2031 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
2032
2033 add_rq_bw(dl_se, dl_rq);
2034 add_running_bw(dl_se, dl_rq);
2035 }
2036
2037 /*
2038 * If p is throttled, we do not enqueue it. In fact, if it exhausted
2039 * its budget it needs a replenishment and, since it now is on
2040 * its rq, the bandwidth timer callback (which clearly has not
2041 * run yet) will take care of this.
2042 * However, the active utilization does not depend on the fact
2043 * that the task is on the runqueue or not (but depends on the
2044 * task's state - in GRUB parlance, "inactive" vs "active contending").
2045 * In other words, even if a task is throttled its utilization must
2046 * be counted in the active utilization; hence, we need to call
2047 * add_running_bw().
2048 */
2049 if (!dl_se->dl_defer && dl_se->dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
2050 if (flags & ENQUEUE_WAKEUP)
2051 task_contending(dl_se, flags);
2052
2053 return;
2054 }
2055
2056 /*
2057 * If this is a wakeup or a new instance, the scheduling
2058 * parameters of the task might need updating. Otherwise,
2059 * we want a replenishment of its runtime.
2060 */
2061 if (flags & ENQUEUE_WAKEUP) {
2062 task_contending(dl_se, flags);
2063 update_dl_entity(dl_se);
2064 } else if (flags & ENQUEUE_REPLENISH) {
2065 replenish_dl_entity(dl_se);
2066 } else if ((flags & ENQUEUE_RESTORE) &&
2067 !is_dl_boosted(dl_se) &&
2068 dl_time_before(dl_se->deadline, rq_clock(rq_of_dl_se(dl_se)))) {
2069 setup_new_dl_entity(dl_se);
2070 }
2071
2072 /*
2073 * If the reservation is still throttled, e.g., it got replenished but is a
2074 * deferred task and still got to wait, don't enqueue.
2075 */
2076 if (dl_se->dl_throttled && start_dl_timer(dl_se))
2077 return;
2078
2079 /*
2080 * We're about to enqueue, make sure we're not ->dl_throttled!
2081 * In case the timer was not started, say because the defer time
2082 * has passed, mark as not throttled and mark unarmed.
2083 * Also cancel earlier timers, since letting those run is pointless.
2084 */
2085 if (dl_se->dl_throttled) {
2086 hrtimer_try_to_cancel(&dl_se->dl_timer);
2087 dl_se->dl_defer_armed = 0;
2088 dl_se->dl_throttled = 0;
2089 }
2090
2091 __enqueue_dl_entity(dl_se);
2092 }
2093
dequeue_dl_entity(struct sched_dl_entity * dl_se,int flags)2094 static void dequeue_dl_entity(struct sched_dl_entity *dl_se, int flags)
2095 {
2096 __dequeue_dl_entity(dl_se);
2097
2098 if (flags & (DEQUEUE_SAVE|DEQUEUE_MIGRATING)) {
2099 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
2100
2101 sub_running_bw(dl_se, dl_rq);
2102 sub_rq_bw(dl_se, dl_rq);
2103 }
2104
2105 /*
2106 * This check allows to start the inactive timer (or to immediately
2107 * decrease the active utilization, if needed) in two cases:
2108 * when the task blocks and when it is terminating
2109 * (p->state == TASK_DEAD). We can handle the two cases in the same
2110 * way, because from GRUB's point of view the same thing is happening
2111 * (the task moves from "active contending" to "active non contending"
2112 * or "inactive")
2113 */
2114 if (flags & DEQUEUE_SLEEP)
2115 task_non_contending(dl_se);
2116 }
2117
enqueue_task_dl(struct rq * rq,struct task_struct * p,int flags)2118 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2119 {
2120 if (is_dl_boosted(&p->dl)) {
2121 /*
2122 * Because of delays in the detection of the overrun of a
2123 * thread's runtime, it might be the case that a thread
2124 * goes to sleep in a rt mutex with negative runtime. As
2125 * a consequence, the thread will be throttled.
2126 *
2127 * While waiting for the mutex, this thread can also be
2128 * boosted via PI, resulting in a thread that is throttled
2129 * and boosted at the same time.
2130 *
2131 * In this case, the boost overrides the throttle.
2132 */
2133 if (p->dl.dl_throttled) {
2134 /*
2135 * The replenish timer needs to be canceled. No
2136 * problem if it fires concurrently: boosted threads
2137 * are ignored in dl_task_timer().
2138 */
2139 cancel_replenish_timer(&p->dl);
2140 p->dl.dl_throttled = 0;
2141 }
2142 } else if (!dl_prio(p->normal_prio)) {
2143 /*
2144 * Special case in which we have a !SCHED_DEADLINE task that is going
2145 * to be deboosted, but exceeds its runtime while doing so. No point in
2146 * replenishing it, as it's going to return back to its original
2147 * scheduling class after this. If it has been throttled, we need to
2148 * clear the flag, otherwise the task may wake up as throttled after
2149 * being boosted again with no means to replenish the runtime and clear
2150 * the throttle.
2151 */
2152 p->dl.dl_throttled = 0;
2153 if (!(flags & ENQUEUE_REPLENISH))
2154 printk_deferred_once("sched: DL de-boosted task PID %d: REPLENISH flag missing\n",
2155 task_pid_nr(p));
2156
2157 return;
2158 }
2159
2160 check_schedstat_required();
2161 update_stats_wait_start_dl(dl_rq_of_se(&p->dl), &p->dl);
2162
2163 if (p->on_rq == TASK_ON_RQ_MIGRATING)
2164 flags |= ENQUEUE_MIGRATING;
2165
2166 enqueue_dl_entity(&p->dl, flags);
2167
2168 if (dl_server(&p->dl))
2169 return;
2170
2171 if (!task_current(rq, p) && !p->dl.dl_throttled && p->nr_cpus_allowed > 1)
2172 enqueue_pushable_dl_task(rq, p);
2173 }
2174
dequeue_task_dl(struct rq * rq,struct task_struct * p,int flags)2175 static bool dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2176 {
2177 update_curr_dl(rq);
2178
2179 if (p->on_rq == TASK_ON_RQ_MIGRATING)
2180 flags |= DEQUEUE_MIGRATING;
2181
2182 dequeue_dl_entity(&p->dl, flags);
2183 if (!p->dl.dl_throttled && !dl_server(&p->dl))
2184 dequeue_pushable_dl_task(rq, p);
2185
2186 return true;
2187 }
2188
2189 /*
2190 * Yield task semantic for -deadline tasks is:
2191 *
2192 * get off from the CPU until our next instance, with
2193 * a new runtime. This is of little use now, since we
2194 * don't have a bandwidth reclaiming mechanism. Anyway,
2195 * bandwidth reclaiming is planned for the future, and
2196 * yield_task_dl will indicate that some spare budget
2197 * is available for other task instances to use it.
2198 */
yield_task_dl(struct rq * rq)2199 static void yield_task_dl(struct rq *rq)
2200 {
2201 /*
2202 * We make the task go to sleep until its current deadline by
2203 * forcing its runtime to zero. This way, update_curr_dl() stops
2204 * it and the bandwidth timer will wake it up and will give it
2205 * new scheduling parameters (thanks to dl_yielded=1).
2206 */
2207 rq->curr->dl.dl_yielded = 1;
2208
2209 update_rq_clock(rq);
2210 update_curr_dl(rq);
2211 /*
2212 * Tell update_rq_clock() that we've just updated,
2213 * so we don't do microscopic update in schedule()
2214 * and double the fastpath cost.
2215 */
2216 rq_clock_skip_update(rq);
2217 }
2218
2219 #ifdef CONFIG_SMP
2220
dl_task_is_earliest_deadline(struct task_struct * p,struct rq * rq)2221 static inline bool dl_task_is_earliest_deadline(struct task_struct *p,
2222 struct rq *rq)
2223 {
2224 return (!rq->dl.dl_nr_running ||
2225 dl_time_before(p->dl.deadline,
2226 rq->dl.earliest_dl.curr));
2227 }
2228
2229 static int find_later_rq(struct task_struct *task);
2230
2231 static int
select_task_rq_dl(struct task_struct * p,int cpu,int flags)2232 select_task_rq_dl(struct task_struct *p, int cpu, int flags)
2233 {
2234 struct task_struct *curr, *donor;
2235 bool select_rq;
2236 struct rq *rq;
2237
2238 if (!(flags & WF_TTWU))
2239 goto out;
2240
2241 rq = cpu_rq(cpu);
2242
2243 rcu_read_lock();
2244 curr = READ_ONCE(rq->curr); /* unlocked access */
2245 donor = READ_ONCE(rq->donor);
2246
2247 /*
2248 * If we are dealing with a -deadline task, we must
2249 * decide where to wake it up.
2250 * If it has a later deadline and the current task
2251 * on this rq can't move (provided the waking task
2252 * can!) we prefer to send it somewhere else. On the
2253 * other hand, if it has a shorter deadline, we
2254 * try to make it stay here, it might be important.
2255 */
2256 select_rq = unlikely(dl_task(donor)) &&
2257 (curr->nr_cpus_allowed < 2 ||
2258 !dl_entity_preempt(&p->dl, &donor->dl)) &&
2259 p->nr_cpus_allowed > 1;
2260
2261 /*
2262 * Take the capacity of the CPU into account to
2263 * ensure it fits the requirement of the task.
2264 */
2265 if (sched_asym_cpucap_active())
2266 select_rq |= !dl_task_fits_capacity(p, cpu);
2267
2268 if (select_rq) {
2269 int target = find_later_rq(p);
2270
2271 if (target != -1 &&
2272 dl_task_is_earliest_deadline(p, cpu_rq(target)))
2273 cpu = target;
2274 }
2275 rcu_read_unlock();
2276
2277 out:
2278 return cpu;
2279 }
2280
migrate_task_rq_dl(struct task_struct * p,int new_cpu __maybe_unused)2281 static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
2282 {
2283 struct rq_flags rf;
2284 struct rq *rq;
2285
2286 if (READ_ONCE(p->__state) != TASK_WAKING)
2287 return;
2288
2289 rq = task_rq(p);
2290 /*
2291 * Since p->state == TASK_WAKING, set_task_cpu() has been called
2292 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
2293 * rq->lock is not... So, lock it
2294 */
2295 rq_lock(rq, &rf);
2296 if (p->dl.dl_non_contending) {
2297 update_rq_clock(rq);
2298 sub_running_bw(&p->dl, &rq->dl);
2299 p->dl.dl_non_contending = 0;
2300 /*
2301 * If the timer handler is currently running and the
2302 * timer cannot be canceled, inactive_task_timer()
2303 * will see that dl_not_contending is not set, and
2304 * will not touch the rq's active utilization,
2305 * so we are still safe.
2306 */
2307 cancel_inactive_timer(&p->dl);
2308 }
2309 sub_rq_bw(&p->dl, &rq->dl);
2310 rq_unlock(rq, &rf);
2311 }
2312
check_preempt_equal_dl(struct rq * rq,struct task_struct * p)2313 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
2314 {
2315 /*
2316 * Current can't be migrated, useless to reschedule,
2317 * let's hope p can move out.
2318 */
2319 if (rq->curr->nr_cpus_allowed == 1 ||
2320 !cpudl_find(&rq->rd->cpudl, rq->donor, NULL))
2321 return;
2322
2323 /*
2324 * p is migratable, so let's not schedule it and
2325 * see if it is pushed or pulled somewhere else.
2326 */
2327 if (p->nr_cpus_allowed != 1 &&
2328 cpudl_find(&rq->rd->cpudl, p, NULL))
2329 return;
2330
2331 resched_curr(rq);
2332 }
2333
balance_dl(struct rq * rq,struct task_struct * p,struct rq_flags * rf)2334 static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
2335 {
2336 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
2337 /*
2338 * This is OK, because current is on_cpu, which avoids it being
2339 * picked for load-balance and preemption/IRQs are still
2340 * disabled avoiding further scheduler activity on it and we've
2341 * not yet started the picking loop.
2342 */
2343 rq_unpin_lock(rq, rf);
2344 pull_dl_task(rq);
2345 rq_repin_lock(rq, rf);
2346 }
2347
2348 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
2349 }
2350 #endif /* CONFIG_SMP */
2351
2352 /*
2353 * Only called when both the current and waking task are -deadline
2354 * tasks.
2355 */
wakeup_preempt_dl(struct rq * rq,struct task_struct * p,int flags)2356 static void wakeup_preempt_dl(struct rq *rq, struct task_struct *p,
2357 int flags)
2358 {
2359 if (dl_entity_preempt(&p->dl, &rq->donor->dl)) {
2360 resched_curr(rq);
2361 return;
2362 }
2363
2364 #ifdef CONFIG_SMP
2365 /*
2366 * In the unlikely case current and p have the same deadline
2367 * let us try to decide what's the best thing to do...
2368 */
2369 if ((p->dl.deadline == rq->donor->dl.deadline) &&
2370 !test_tsk_need_resched(rq->curr))
2371 check_preempt_equal_dl(rq, p);
2372 #endif /* CONFIG_SMP */
2373 }
2374
2375 #ifdef CONFIG_SCHED_HRTICK
start_hrtick_dl(struct rq * rq,struct sched_dl_entity * dl_se)2376 static void start_hrtick_dl(struct rq *rq, struct sched_dl_entity *dl_se)
2377 {
2378 hrtick_start(rq, dl_se->runtime);
2379 }
2380 #else /* !CONFIG_SCHED_HRTICK */
start_hrtick_dl(struct rq * rq,struct sched_dl_entity * dl_se)2381 static void start_hrtick_dl(struct rq *rq, struct sched_dl_entity *dl_se)
2382 {
2383 }
2384 #endif
2385
set_next_task_dl(struct rq * rq,struct task_struct * p,bool first)2386 static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
2387 {
2388 struct sched_dl_entity *dl_se = &p->dl;
2389 struct dl_rq *dl_rq = &rq->dl;
2390
2391 p->se.exec_start = rq_clock_task(rq);
2392 if (on_dl_rq(&p->dl))
2393 update_stats_wait_end_dl(dl_rq, dl_se);
2394
2395 /* You can't push away the running task */
2396 dequeue_pushable_dl_task(rq, p);
2397
2398 if (!first)
2399 return;
2400
2401 if (rq->donor->sched_class != &dl_sched_class)
2402 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2403
2404 deadline_queue_push_tasks(rq);
2405
2406 if (hrtick_enabled_dl(rq))
2407 start_hrtick_dl(rq, &p->dl);
2408 }
2409
pick_next_dl_entity(struct dl_rq * dl_rq)2410 static struct sched_dl_entity *pick_next_dl_entity(struct dl_rq *dl_rq)
2411 {
2412 struct rb_node *left = rb_first_cached(&dl_rq->root);
2413
2414 if (!left)
2415 return NULL;
2416
2417 return __node_2_dle(left);
2418 }
2419
2420 /*
2421 * __pick_next_task_dl - Helper to pick the next -deadline task to run.
2422 * @rq: The runqueue to pick the next task from.
2423 */
__pick_task_dl(struct rq * rq)2424 static struct task_struct *__pick_task_dl(struct rq *rq)
2425 {
2426 struct sched_dl_entity *dl_se;
2427 struct dl_rq *dl_rq = &rq->dl;
2428 struct task_struct *p;
2429
2430 again:
2431 if (!sched_dl_runnable(rq))
2432 return NULL;
2433
2434 dl_se = pick_next_dl_entity(dl_rq);
2435 WARN_ON_ONCE(!dl_se);
2436
2437 if (dl_server(dl_se)) {
2438 p = dl_se->server_pick_task(dl_se);
2439 if (!p) {
2440 if (dl_server_active(dl_se)) {
2441 dl_se->dl_yielded = 1;
2442 update_curr_dl_se(rq, dl_se, 0);
2443 }
2444 goto again;
2445 }
2446 rq->dl_server = dl_se;
2447 } else {
2448 p = dl_task_of(dl_se);
2449 }
2450
2451 return p;
2452 }
2453
pick_task_dl(struct rq * rq)2454 static struct task_struct *pick_task_dl(struct rq *rq)
2455 {
2456 return __pick_task_dl(rq);
2457 }
2458
put_prev_task_dl(struct rq * rq,struct task_struct * p,struct task_struct * next)2459 static void put_prev_task_dl(struct rq *rq, struct task_struct *p, struct task_struct *next)
2460 {
2461 struct sched_dl_entity *dl_se = &p->dl;
2462 struct dl_rq *dl_rq = &rq->dl;
2463
2464 if (on_dl_rq(&p->dl))
2465 update_stats_wait_start_dl(dl_rq, dl_se);
2466
2467 update_curr_dl(rq);
2468
2469 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2470 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2471 enqueue_pushable_dl_task(rq, p);
2472 }
2473
2474 /*
2475 * scheduler tick hitting a task of our scheduling class.
2476 *
2477 * NOTE: This function can be called remotely by the tick offload that
2478 * goes along full dynticks. Therefore no local assumption can be made
2479 * and everything must be accessed through the @rq and @curr passed in
2480 * parameters.
2481 */
task_tick_dl(struct rq * rq,struct task_struct * p,int queued)2482 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2483 {
2484 update_curr_dl(rq);
2485
2486 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2487 /*
2488 * Even when we have runtime, update_curr_dl() might have resulted in us
2489 * not being the leftmost task anymore. In that case NEED_RESCHED will
2490 * be set and schedule() will start a new hrtick for the next task.
2491 */
2492 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
2493 is_leftmost(&p->dl, &rq->dl))
2494 start_hrtick_dl(rq, &p->dl);
2495 }
2496
task_fork_dl(struct task_struct * p)2497 static void task_fork_dl(struct task_struct *p)
2498 {
2499 /*
2500 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2501 * sched_fork()
2502 */
2503 }
2504
2505 #ifdef CONFIG_SMP
2506
2507 /* Only try algorithms three times */
2508 #define DL_MAX_TRIES 3
2509
2510 /*
2511 * Return the earliest pushable rq's task, which is suitable to be executed
2512 * on the CPU, NULL otherwise:
2513 */
pick_earliest_pushable_dl_task(struct rq * rq,int cpu)2514 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2515 {
2516 struct task_struct *p = NULL;
2517 struct rb_node *next_node;
2518
2519 if (!has_pushable_dl_tasks(rq))
2520 return NULL;
2521
2522 next_node = rb_first_cached(&rq->dl.pushable_dl_tasks_root);
2523 while (next_node) {
2524 p = __node_2_pdl(next_node);
2525
2526 if (task_is_pushable(rq, p, cpu))
2527 return p;
2528
2529 next_node = rb_next(next_node);
2530 }
2531
2532 return NULL;
2533 }
2534
2535 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2536
find_later_rq(struct task_struct * task)2537 static int find_later_rq(struct task_struct *task)
2538 {
2539 struct sched_domain *sd;
2540 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2541 int this_cpu = smp_processor_id();
2542 int cpu = task_cpu(task);
2543
2544 /* Make sure the mask is initialized first */
2545 if (unlikely(!later_mask))
2546 return -1;
2547
2548 if (task->nr_cpus_allowed == 1)
2549 return -1;
2550
2551 /*
2552 * We have to consider system topology and task affinity
2553 * first, then we can look for a suitable CPU.
2554 */
2555 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
2556 return -1;
2557
2558 /*
2559 * If we are here, some targets have been found, including
2560 * the most suitable which is, among the runqueues where the
2561 * current tasks have later deadlines than the task's one, the
2562 * rq with the latest possible one.
2563 *
2564 * Now we check how well this matches with task's
2565 * affinity and system topology.
2566 *
2567 * The last CPU where the task run is our first
2568 * guess, since it is most likely cache-hot there.
2569 */
2570 if (cpumask_test_cpu(cpu, later_mask))
2571 return cpu;
2572 /*
2573 * Check if this_cpu is to be skipped (i.e., it is
2574 * not in the mask) or not.
2575 */
2576 if (!cpumask_test_cpu(this_cpu, later_mask))
2577 this_cpu = -1;
2578
2579 rcu_read_lock();
2580 for_each_domain(cpu, sd) {
2581 if (sd->flags & SD_WAKE_AFFINE) {
2582 int best_cpu;
2583
2584 /*
2585 * If possible, preempting this_cpu is
2586 * cheaper than migrating.
2587 */
2588 if (this_cpu != -1 &&
2589 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2590 rcu_read_unlock();
2591 return this_cpu;
2592 }
2593
2594 best_cpu = cpumask_any_and_distribute(later_mask,
2595 sched_domain_span(sd));
2596 /*
2597 * Last chance: if a CPU being in both later_mask
2598 * and current sd span is valid, that becomes our
2599 * choice. Of course, the latest possible CPU is
2600 * already under consideration through later_mask.
2601 */
2602 if (best_cpu < nr_cpu_ids) {
2603 rcu_read_unlock();
2604 return best_cpu;
2605 }
2606 }
2607 }
2608 rcu_read_unlock();
2609
2610 /*
2611 * At this point, all our guesses failed, we just return
2612 * 'something', and let the caller sort the things out.
2613 */
2614 if (this_cpu != -1)
2615 return this_cpu;
2616
2617 cpu = cpumask_any_distribute(later_mask);
2618 if (cpu < nr_cpu_ids)
2619 return cpu;
2620
2621 return -1;
2622 }
2623
2624 /* Locks the rq it finds */
find_lock_later_rq(struct task_struct * task,struct rq * rq)2625 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2626 {
2627 struct rq *later_rq = NULL;
2628 int tries;
2629 int cpu;
2630
2631 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2632 cpu = find_later_rq(task);
2633
2634 if ((cpu == -1) || (cpu == rq->cpu))
2635 break;
2636
2637 later_rq = cpu_rq(cpu);
2638
2639 if (!dl_task_is_earliest_deadline(task, later_rq)) {
2640 /*
2641 * Target rq has tasks of equal or earlier deadline,
2642 * retrying does not release any lock and is unlikely
2643 * to yield a different result.
2644 */
2645 later_rq = NULL;
2646 break;
2647 }
2648
2649 /* Retry if something changed. */
2650 if (double_lock_balance(rq, later_rq)) {
2651 if (unlikely(task_rq(task) != rq ||
2652 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2653 task_on_cpu(rq, task) ||
2654 !dl_task(task) ||
2655 is_migration_disabled(task) ||
2656 !task_on_rq_queued(task))) {
2657 double_unlock_balance(rq, later_rq);
2658 later_rq = NULL;
2659 break;
2660 }
2661 }
2662
2663 /*
2664 * If the rq we found has no -deadline task, or
2665 * its earliest one has a later deadline than our
2666 * task, the rq is a good one.
2667 */
2668 if (dl_task_is_earliest_deadline(task, later_rq))
2669 break;
2670
2671 /* Otherwise we try again. */
2672 double_unlock_balance(rq, later_rq);
2673 later_rq = NULL;
2674 }
2675
2676 return later_rq;
2677 }
2678
pick_next_pushable_dl_task(struct rq * rq)2679 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2680 {
2681 struct task_struct *p;
2682
2683 if (!has_pushable_dl_tasks(rq))
2684 return NULL;
2685
2686 p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2687
2688 WARN_ON_ONCE(rq->cpu != task_cpu(p));
2689 WARN_ON_ONCE(task_current(rq, p));
2690 WARN_ON_ONCE(p->nr_cpus_allowed <= 1);
2691
2692 WARN_ON_ONCE(!task_on_rq_queued(p));
2693 WARN_ON_ONCE(!dl_task(p));
2694
2695 return p;
2696 }
2697
2698 /*
2699 * See if the non running -deadline tasks on this rq
2700 * can be sent to some other CPU where they can preempt
2701 * and start executing.
2702 */
push_dl_task(struct rq * rq)2703 static int push_dl_task(struct rq *rq)
2704 {
2705 struct task_struct *next_task;
2706 struct rq *later_rq;
2707 int ret = 0;
2708
2709 next_task = pick_next_pushable_dl_task(rq);
2710 if (!next_task)
2711 return 0;
2712
2713 retry:
2714 /*
2715 * If next_task preempts rq->curr, and rq->curr
2716 * can move away, it makes sense to just reschedule
2717 * without going further in pushing next_task.
2718 */
2719 if (dl_task(rq->donor) &&
2720 dl_time_before(next_task->dl.deadline, rq->donor->dl.deadline) &&
2721 rq->curr->nr_cpus_allowed > 1) {
2722 resched_curr(rq);
2723 return 0;
2724 }
2725
2726 if (is_migration_disabled(next_task))
2727 return 0;
2728
2729 if (WARN_ON(next_task == rq->curr))
2730 return 0;
2731
2732 /* We might release rq lock */
2733 get_task_struct(next_task);
2734
2735 /* Will lock the rq it'll find */
2736 later_rq = find_lock_later_rq(next_task, rq);
2737 if (!later_rq) {
2738 struct task_struct *task;
2739
2740 /*
2741 * We must check all this again, since
2742 * find_lock_later_rq releases rq->lock and it is
2743 * then possible that next_task has migrated.
2744 */
2745 task = pick_next_pushable_dl_task(rq);
2746 if (task == next_task) {
2747 /*
2748 * The task is still there. We don't try
2749 * again, some other CPU will pull it when ready.
2750 */
2751 goto out;
2752 }
2753
2754 if (!task)
2755 /* No more tasks */
2756 goto out;
2757
2758 put_task_struct(next_task);
2759 next_task = task;
2760 goto retry;
2761 }
2762
2763 move_queued_task_locked(rq, later_rq, next_task);
2764 ret = 1;
2765
2766 resched_curr(later_rq);
2767
2768 double_unlock_balance(rq, later_rq);
2769
2770 out:
2771 put_task_struct(next_task);
2772
2773 return ret;
2774 }
2775
push_dl_tasks(struct rq * rq)2776 static void push_dl_tasks(struct rq *rq)
2777 {
2778 /* push_dl_task() will return true if it moved a -deadline task */
2779 while (push_dl_task(rq))
2780 ;
2781 }
2782
pull_dl_task(struct rq * this_rq)2783 static void pull_dl_task(struct rq *this_rq)
2784 {
2785 int this_cpu = this_rq->cpu, cpu;
2786 struct task_struct *p, *push_task;
2787 bool resched = false;
2788 struct rq *src_rq;
2789 u64 dmin = LONG_MAX;
2790
2791 if (likely(!dl_overloaded(this_rq)))
2792 return;
2793
2794 /*
2795 * Match the barrier from dl_set_overloaded; this guarantees that if we
2796 * see overloaded we must also see the dlo_mask bit.
2797 */
2798 smp_rmb();
2799
2800 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2801 if (this_cpu == cpu)
2802 continue;
2803
2804 src_rq = cpu_rq(cpu);
2805
2806 /*
2807 * It looks racy, and it is! However, as in sched_rt.c,
2808 * we are fine with this.
2809 */
2810 if (this_rq->dl.dl_nr_running &&
2811 dl_time_before(this_rq->dl.earliest_dl.curr,
2812 src_rq->dl.earliest_dl.next))
2813 continue;
2814
2815 /* Might drop this_rq->lock */
2816 push_task = NULL;
2817 double_lock_balance(this_rq, src_rq);
2818
2819 /*
2820 * If there are no more pullable tasks on the
2821 * rq, we're done with it.
2822 */
2823 if (src_rq->dl.dl_nr_running <= 1)
2824 goto skip;
2825
2826 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2827
2828 /*
2829 * We found a task to be pulled if:
2830 * - it preempts our current (if there's one),
2831 * - it will preempt the last one we pulled (if any).
2832 */
2833 if (p && dl_time_before(p->dl.deadline, dmin) &&
2834 dl_task_is_earliest_deadline(p, this_rq)) {
2835 WARN_ON(p == src_rq->curr);
2836 WARN_ON(!task_on_rq_queued(p));
2837
2838 /*
2839 * Then we pull iff p has actually an earlier
2840 * deadline than the current task of its runqueue.
2841 */
2842 if (dl_time_before(p->dl.deadline,
2843 src_rq->donor->dl.deadline))
2844 goto skip;
2845
2846 if (is_migration_disabled(p)) {
2847 push_task = get_push_task(src_rq);
2848 } else {
2849 move_queued_task_locked(src_rq, this_rq, p);
2850 dmin = p->dl.deadline;
2851 resched = true;
2852 }
2853
2854 /* Is there any other task even earlier? */
2855 }
2856 skip:
2857 double_unlock_balance(this_rq, src_rq);
2858
2859 if (push_task) {
2860 preempt_disable();
2861 raw_spin_rq_unlock(this_rq);
2862 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2863 push_task, &src_rq->push_work);
2864 preempt_enable();
2865 raw_spin_rq_lock(this_rq);
2866 }
2867 }
2868
2869 if (resched)
2870 resched_curr(this_rq);
2871 }
2872
2873 /*
2874 * Since the task is not running and a reschedule is not going to happen
2875 * anytime soon on its runqueue, we try pushing it away now.
2876 */
task_woken_dl(struct rq * rq,struct task_struct * p)2877 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2878 {
2879 if (!task_on_cpu(rq, p) &&
2880 !test_tsk_need_resched(rq->curr) &&
2881 p->nr_cpus_allowed > 1 &&
2882 dl_task(rq->donor) &&
2883 (rq->curr->nr_cpus_allowed < 2 ||
2884 !dl_entity_preempt(&p->dl, &rq->donor->dl))) {
2885 push_dl_tasks(rq);
2886 }
2887 }
2888
set_cpus_allowed_dl(struct task_struct * p,struct affinity_context * ctx)2889 static void set_cpus_allowed_dl(struct task_struct *p,
2890 struct affinity_context *ctx)
2891 {
2892 struct root_domain *src_rd;
2893 struct rq *rq;
2894
2895 WARN_ON_ONCE(!dl_task(p));
2896
2897 rq = task_rq(p);
2898 src_rd = rq->rd;
2899 /*
2900 * Migrating a SCHED_DEADLINE task between exclusive
2901 * cpusets (different root_domains) entails a bandwidth
2902 * update. We already made space for us in the destination
2903 * domain (see cpuset_can_attach()).
2904 */
2905 if (!cpumask_intersects(src_rd->span, ctx->new_mask)) {
2906 struct dl_bw *src_dl_b;
2907
2908 src_dl_b = dl_bw_of(cpu_of(rq));
2909 /*
2910 * We now free resources of the root_domain we are migrating
2911 * off. In the worst case, sched_setattr() may temporary fail
2912 * until we complete the update.
2913 */
2914 raw_spin_lock(&src_dl_b->lock);
2915 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2916 raw_spin_unlock(&src_dl_b->lock);
2917 }
2918
2919 set_cpus_allowed_common(p, ctx);
2920 }
2921
2922 /* Assumes rq->lock is held */
rq_online_dl(struct rq * rq)2923 static void rq_online_dl(struct rq *rq)
2924 {
2925 if (rq->dl.overloaded)
2926 dl_set_overload(rq);
2927
2928 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2929 if (rq->dl.dl_nr_running > 0)
2930 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2931 }
2932
2933 /* Assumes rq->lock is held */
rq_offline_dl(struct rq * rq)2934 static void rq_offline_dl(struct rq *rq)
2935 {
2936 if (rq->dl.overloaded)
2937 dl_clear_overload(rq);
2938
2939 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2940 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2941 }
2942
init_sched_dl_class(void)2943 void __init init_sched_dl_class(void)
2944 {
2945 unsigned int i;
2946
2947 for_each_possible_cpu(i)
2948 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2949 GFP_KERNEL, cpu_to_node(i));
2950 }
2951
dl_add_task_root_domain(struct task_struct * p)2952 void dl_add_task_root_domain(struct task_struct *p)
2953 {
2954 struct rq_flags rf;
2955 struct rq *rq;
2956 struct dl_bw *dl_b;
2957
2958 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2959 if (!dl_task(p)) {
2960 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2961 return;
2962 }
2963
2964 rq = __task_rq_lock(p, &rf);
2965
2966 dl_b = &rq->rd->dl_bw;
2967 raw_spin_lock(&dl_b->lock);
2968
2969 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2970
2971 raw_spin_unlock(&dl_b->lock);
2972
2973 task_rq_unlock(rq, p, &rf);
2974 }
2975
dl_clear_root_domain(struct root_domain * rd)2976 void dl_clear_root_domain(struct root_domain *rd)
2977 {
2978 int i;
2979
2980 guard(raw_spinlock_irqsave)(&rd->dl_bw.lock);
2981 rd->dl_bw.total_bw = 0;
2982
2983 /*
2984 * dl_server bandwidth is only restored when CPUs are attached to root
2985 * domains (after domains are created or CPUs moved back to the
2986 * default root doamin).
2987 */
2988 for_each_cpu(i, rd->span) {
2989 struct sched_dl_entity *dl_se = &cpu_rq(i)->fair_server;
2990
2991 if (dl_server(dl_se) && cpu_active(i))
2992 rd->dl_bw.total_bw += dl_se->dl_bw;
2993 }
2994 }
2995
2996 #endif /* CONFIG_SMP */
2997
switched_from_dl(struct rq * rq,struct task_struct * p)2998 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2999 {
3000 /*
3001 * task_non_contending() can start the "inactive timer" (if the 0-lag
3002 * time is in the future). If the task switches back to dl before
3003 * the "inactive timer" fires, it can continue to consume its current
3004 * runtime using its current deadline. If it stays outside of
3005 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
3006 * will reset the task parameters.
3007 */
3008 if (task_on_rq_queued(p) && p->dl.dl_runtime)
3009 task_non_contending(&p->dl);
3010
3011 /*
3012 * In case a task is setscheduled out from SCHED_DEADLINE we need to
3013 * keep track of that on its cpuset (for correct bandwidth tracking).
3014 */
3015 dec_dl_tasks_cs(p);
3016
3017 if (!task_on_rq_queued(p)) {
3018 /*
3019 * Inactive timer is armed. However, p is leaving DEADLINE and
3020 * might migrate away from this rq while continuing to run on
3021 * some other class. We need to remove its contribution from
3022 * this rq running_bw now, or sub_rq_bw (below) will complain.
3023 */
3024 if (p->dl.dl_non_contending)
3025 sub_running_bw(&p->dl, &rq->dl);
3026 sub_rq_bw(&p->dl, &rq->dl);
3027 }
3028
3029 /*
3030 * We cannot use inactive_task_timer() to invoke sub_running_bw()
3031 * at the 0-lag time, because the task could have been migrated
3032 * while SCHED_OTHER in the meanwhile.
3033 */
3034 if (p->dl.dl_non_contending)
3035 p->dl.dl_non_contending = 0;
3036
3037 /*
3038 * Since this might be the only -deadline task on the rq,
3039 * this is the right place to try to pull some other one
3040 * from an overloaded CPU, if any.
3041 */
3042 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
3043 return;
3044
3045 deadline_queue_pull_task(rq);
3046 }
3047
3048 /*
3049 * When switching to -deadline, we may overload the rq, then
3050 * we try to push someone off, if possible.
3051 */
switched_to_dl(struct rq * rq,struct task_struct * p)3052 static void switched_to_dl(struct rq *rq, struct task_struct *p)
3053 {
3054 cancel_inactive_timer(&p->dl);
3055
3056 /*
3057 * In case a task is setscheduled to SCHED_DEADLINE we need to keep
3058 * track of that on its cpuset (for correct bandwidth tracking).
3059 */
3060 inc_dl_tasks_cs(p);
3061
3062 /* If p is not queued we will update its parameters at next wakeup. */
3063 if (!task_on_rq_queued(p)) {
3064 add_rq_bw(&p->dl, &rq->dl);
3065
3066 return;
3067 }
3068
3069 if (rq->donor != p) {
3070 #ifdef CONFIG_SMP
3071 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
3072 deadline_queue_push_tasks(rq);
3073 #endif
3074 if (dl_task(rq->donor))
3075 wakeup_preempt_dl(rq, p, 0);
3076 else
3077 resched_curr(rq);
3078 } else {
3079 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
3080 }
3081 }
3082
3083 /*
3084 * If the scheduling parameters of a -deadline task changed,
3085 * a push or pull operation might be needed.
3086 */
prio_changed_dl(struct rq * rq,struct task_struct * p,int oldprio)3087 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
3088 int oldprio)
3089 {
3090 if (!task_on_rq_queued(p))
3091 return;
3092
3093 #ifdef CONFIG_SMP
3094 /*
3095 * This might be too much, but unfortunately
3096 * we don't have the old deadline value, and
3097 * we can't argue if the task is increasing
3098 * or lowering its prio, so...
3099 */
3100 if (!rq->dl.overloaded)
3101 deadline_queue_pull_task(rq);
3102
3103 if (task_current_donor(rq, p)) {
3104 /*
3105 * If we now have a earlier deadline task than p,
3106 * then reschedule, provided p is still on this
3107 * runqueue.
3108 */
3109 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
3110 resched_curr(rq);
3111 } else {
3112 /*
3113 * Current may not be deadline in case p was throttled but we
3114 * have just replenished it (e.g. rt_mutex_setprio()).
3115 *
3116 * Otherwise, if p was given an earlier deadline, reschedule.
3117 */
3118 if (!dl_task(rq->curr) ||
3119 dl_time_before(p->dl.deadline, rq->curr->dl.deadline))
3120 resched_curr(rq);
3121 }
3122 #else
3123 /*
3124 * We don't know if p has a earlier or later deadline, so let's blindly
3125 * set a (maybe not needed) rescheduling point.
3126 */
3127 resched_curr(rq);
3128 #endif
3129 }
3130
3131 #ifdef CONFIG_SCHED_CORE
task_is_throttled_dl(struct task_struct * p,int cpu)3132 static int task_is_throttled_dl(struct task_struct *p, int cpu)
3133 {
3134 return p->dl.dl_throttled;
3135 }
3136 #endif
3137
3138 DEFINE_SCHED_CLASS(dl) = {
3139
3140 .enqueue_task = enqueue_task_dl,
3141 .dequeue_task = dequeue_task_dl,
3142 .yield_task = yield_task_dl,
3143
3144 .wakeup_preempt = wakeup_preempt_dl,
3145
3146 .pick_task = pick_task_dl,
3147 .put_prev_task = put_prev_task_dl,
3148 .set_next_task = set_next_task_dl,
3149
3150 #ifdef CONFIG_SMP
3151 .balance = balance_dl,
3152 .select_task_rq = select_task_rq_dl,
3153 .migrate_task_rq = migrate_task_rq_dl,
3154 .set_cpus_allowed = set_cpus_allowed_dl,
3155 .rq_online = rq_online_dl,
3156 .rq_offline = rq_offline_dl,
3157 .task_woken = task_woken_dl,
3158 .find_lock_rq = find_lock_later_rq,
3159 #endif
3160
3161 .task_tick = task_tick_dl,
3162 .task_fork = task_fork_dl,
3163
3164 .prio_changed = prio_changed_dl,
3165 .switched_from = switched_from_dl,
3166 .switched_to = switched_to_dl,
3167
3168 .update_curr = update_curr_dl,
3169 #ifdef CONFIG_SCHED_CORE
3170 .task_is_throttled = task_is_throttled_dl,
3171 #endif
3172 };
3173
3174 /* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
3175 static u64 dl_generation;
3176
sched_dl_global_validate(void)3177 int sched_dl_global_validate(void)
3178 {
3179 u64 runtime = global_rt_runtime();
3180 u64 period = global_rt_period();
3181 u64 new_bw = to_ratio(period, runtime);
3182 u64 gen = ++dl_generation;
3183 struct dl_bw *dl_b;
3184 int cpu, cpus, ret = 0;
3185 unsigned long flags;
3186
3187 /*
3188 * Here we want to check the bandwidth not being set to some
3189 * value smaller than the currently allocated bandwidth in
3190 * any of the root_domains.
3191 */
3192 for_each_possible_cpu(cpu) {
3193 rcu_read_lock_sched();
3194
3195 if (dl_bw_visited(cpu, gen))
3196 goto next;
3197
3198 dl_b = dl_bw_of(cpu);
3199 cpus = dl_bw_cpus(cpu);
3200
3201 raw_spin_lock_irqsave(&dl_b->lock, flags);
3202 if (new_bw * cpus < dl_b->total_bw)
3203 ret = -EBUSY;
3204 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3205
3206 next:
3207 rcu_read_unlock_sched();
3208
3209 if (ret)
3210 break;
3211 }
3212
3213 return ret;
3214 }
3215
init_dl_rq_bw_ratio(struct dl_rq * dl_rq)3216 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
3217 {
3218 if (global_rt_runtime() == RUNTIME_INF) {
3219 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
3220 dl_rq->max_bw = dl_rq->extra_bw = 1 << BW_SHIFT;
3221 } else {
3222 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
3223 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
3224 dl_rq->max_bw = dl_rq->extra_bw =
3225 to_ratio(global_rt_period(), global_rt_runtime());
3226 }
3227 }
3228
sched_dl_do_global(void)3229 void sched_dl_do_global(void)
3230 {
3231 u64 new_bw = -1;
3232 u64 gen = ++dl_generation;
3233 struct dl_bw *dl_b;
3234 int cpu;
3235 unsigned long flags;
3236
3237 if (global_rt_runtime() != RUNTIME_INF)
3238 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
3239
3240 for_each_possible_cpu(cpu) {
3241 rcu_read_lock_sched();
3242
3243 if (dl_bw_visited(cpu, gen)) {
3244 rcu_read_unlock_sched();
3245 continue;
3246 }
3247
3248 dl_b = dl_bw_of(cpu);
3249
3250 raw_spin_lock_irqsave(&dl_b->lock, flags);
3251 dl_b->bw = new_bw;
3252 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3253
3254 rcu_read_unlock_sched();
3255 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
3256 }
3257 }
3258
3259 /*
3260 * We must be sure that accepting a new task (or allowing changing the
3261 * parameters of an existing one) is consistent with the bandwidth
3262 * constraints. If yes, this function also accordingly updates the currently
3263 * allocated bandwidth to reflect the new situation.
3264 *
3265 * This function is called while holding p's rq->lock.
3266 */
sched_dl_overflow(struct task_struct * p,int policy,const struct sched_attr * attr)3267 int sched_dl_overflow(struct task_struct *p, int policy,
3268 const struct sched_attr *attr)
3269 {
3270 u64 period = attr->sched_period ?: attr->sched_deadline;
3271 u64 runtime = attr->sched_runtime;
3272 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
3273 int cpus, err = -1, cpu = task_cpu(p);
3274 struct dl_bw *dl_b = dl_bw_of(cpu);
3275 unsigned long cap;
3276
3277 if (attr->sched_flags & SCHED_FLAG_SUGOV)
3278 return 0;
3279
3280 /* !deadline task may carry old deadline bandwidth */
3281 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
3282 return 0;
3283
3284 /*
3285 * Either if a task, enters, leave, or stays -deadline but changes
3286 * its parameters, we may need to update accordingly the total
3287 * allocated bandwidth of the container.
3288 */
3289 raw_spin_lock(&dl_b->lock);
3290 cpus = dl_bw_cpus(cpu);
3291 cap = dl_bw_capacity(cpu);
3292
3293 if (dl_policy(policy) && !task_has_dl_policy(p) &&
3294 !__dl_overflow(dl_b, cap, 0, new_bw)) {
3295 if (hrtimer_active(&p->dl.inactive_timer))
3296 __dl_sub(dl_b, p->dl.dl_bw, cpus);
3297 __dl_add(dl_b, new_bw, cpus);
3298 err = 0;
3299 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
3300 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
3301 /*
3302 * XXX this is slightly incorrect: when the task
3303 * utilization decreases, we should delay the total
3304 * utilization change until the task's 0-lag point.
3305 * But this would require to set the task's "inactive
3306 * timer" when the task is not inactive.
3307 */
3308 __dl_sub(dl_b, p->dl.dl_bw, cpus);
3309 __dl_add(dl_b, new_bw, cpus);
3310 dl_change_utilization(p, new_bw);
3311 err = 0;
3312 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
3313 /*
3314 * Do not decrease the total deadline utilization here,
3315 * switched_from_dl() will take care to do it at the correct
3316 * (0-lag) time.
3317 */
3318 err = 0;
3319 }
3320 raw_spin_unlock(&dl_b->lock);
3321
3322 return err;
3323 }
3324
3325 /*
3326 * This function initializes the sched_dl_entity of a newly becoming
3327 * SCHED_DEADLINE task.
3328 *
3329 * Only the static values are considered here, the actual runtime and the
3330 * absolute deadline will be properly calculated when the task is enqueued
3331 * for the first time with its new policy.
3332 */
__setparam_dl(struct task_struct * p,const struct sched_attr * attr)3333 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3334 {
3335 struct sched_dl_entity *dl_se = &p->dl;
3336
3337 dl_se->dl_runtime = attr->sched_runtime;
3338 dl_se->dl_deadline = attr->sched_deadline;
3339 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3340 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
3341 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3342 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
3343 }
3344
__getparam_dl(struct task_struct * p,struct sched_attr * attr)3345 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3346 {
3347 struct sched_dl_entity *dl_se = &p->dl;
3348
3349 attr->sched_priority = p->rt_priority;
3350 attr->sched_runtime = dl_se->dl_runtime;
3351 attr->sched_deadline = dl_se->dl_deadline;
3352 attr->sched_period = dl_se->dl_period;
3353 attr->sched_flags &= ~SCHED_DL_FLAGS;
3354 attr->sched_flags |= dl_se->flags;
3355 }
3356
3357 /*
3358 * This function validates the new parameters of a -deadline task.
3359 * We ask for the deadline not being zero, and greater or equal
3360 * than the runtime, as well as the period of being zero or
3361 * greater than deadline. Furthermore, we have to be sure that
3362 * user parameters are above the internal resolution of 1us (we
3363 * check sched_runtime only since it is always the smaller one) and
3364 * below 2^63 ns (we have to check both sched_deadline and
3365 * sched_period, as the latter can be zero).
3366 */
__checkparam_dl(const struct sched_attr * attr)3367 bool __checkparam_dl(const struct sched_attr *attr)
3368 {
3369 u64 period, max, min;
3370
3371 /* special dl tasks don't actually use any parameter */
3372 if (attr->sched_flags & SCHED_FLAG_SUGOV)
3373 return true;
3374
3375 /* deadline != 0 */
3376 if (attr->sched_deadline == 0)
3377 return false;
3378
3379 /*
3380 * Since we truncate DL_SCALE bits, make sure we're at least
3381 * that big.
3382 */
3383 if (attr->sched_runtime < (1ULL << DL_SCALE))
3384 return false;
3385
3386 /*
3387 * Since we use the MSB for wrap-around and sign issues, make
3388 * sure it's not set (mind that period can be equal to zero).
3389 */
3390 if (attr->sched_deadline & (1ULL << 63) ||
3391 attr->sched_period & (1ULL << 63))
3392 return false;
3393
3394 period = attr->sched_period;
3395 if (!period)
3396 period = attr->sched_deadline;
3397
3398 /* runtime <= deadline <= period (if period != 0) */
3399 if (period < attr->sched_deadline ||
3400 attr->sched_deadline < attr->sched_runtime)
3401 return false;
3402
3403 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
3404 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
3405
3406 if (period < min || period > max)
3407 return false;
3408
3409 return true;
3410 }
3411
3412 /*
3413 * This function clears the sched_dl_entity static params.
3414 */
__dl_clear_params(struct sched_dl_entity * dl_se)3415 static void __dl_clear_params(struct sched_dl_entity *dl_se)
3416 {
3417 dl_se->dl_runtime = 0;
3418 dl_se->dl_deadline = 0;
3419 dl_se->dl_period = 0;
3420 dl_se->flags = 0;
3421 dl_se->dl_bw = 0;
3422 dl_se->dl_density = 0;
3423
3424 dl_se->dl_throttled = 0;
3425 dl_se->dl_yielded = 0;
3426 dl_se->dl_non_contending = 0;
3427 dl_se->dl_overrun = 0;
3428 dl_se->dl_server = 0;
3429
3430 #ifdef CONFIG_RT_MUTEXES
3431 dl_se->pi_se = dl_se;
3432 #endif
3433 }
3434
init_dl_entity(struct sched_dl_entity * dl_se)3435 void init_dl_entity(struct sched_dl_entity *dl_se)
3436 {
3437 RB_CLEAR_NODE(&dl_se->rb_node);
3438 init_dl_task_timer(dl_se);
3439 init_dl_inactive_task_timer(dl_se);
3440 __dl_clear_params(dl_se);
3441 }
3442
dl_param_changed(struct task_struct * p,const struct sched_attr * attr)3443 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
3444 {
3445 struct sched_dl_entity *dl_se = &p->dl;
3446
3447 if (dl_se->dl_runtime != attr->sched_runtime ||
3448 dl_se->dl_deadline != attr->sched_deadline ||
3449 dl_se->dl_period != attr->sched_period ||
3450 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
3451 return true;
3452
3453 return false;
3454 }
3455
3456 #ifdef CONFIG_SMP
dl_cpuset_cpumask_can_shrink(const struct cpumask * cur,const struct cpumask * trial)3457 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3458 const struct cpumask *trial)
3459 {
3460 unsigned long flags, cap;
3461 struct dl_bw *cur_dl_b;
3462 int ret = 1;
3463
3464 rcu_read_lock_sched();
3465 cur_dl_b = dl_bw_of(cpumask_any(cur));
3466 cap = __dl_bw_capacity(trial);
3467 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3468 if (__dl_overflow(cur_dl_b, cap, 0, 0))
3469 ret = 0;
3470 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3471 rcu_read_unlock_sched();
3472
3473 return ret;
3474 }
3475
3476 enum dl_bw_request {
3477 dl_bw_req_deactivate = 0,
3478 dl_bw_req_alloc,
3479 dl_bw_req_free
3480 };
3481
dl_bw_manage(enum dl_bw_request req,int cpu,u64 dl_bw)3482 static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
3483 {
3484 unsigned long flags, cap;
3485 struct dl_bw *dl_b;
3486 bool overflow = 0;
3487 u64 fair_server_bw = 0;
3488
3489 rcu_read_lock_sched();
3490 dl_b = dl_bw_of(cpu);
3491 raw_spin_lock_irqsave(&dl_b->lock, flags);
3492
3493 cap = dl_bw_capacity(cpu);
3494 switch (req) {
3495 case dl_bw_req_free:
3496 __dl_sub(dl_b, dl_bw, dl_bw_cpus(cpu));
3497 break;
3498 case dl_bw_req_alloc:
3499 overflow = __dl_overflow(dl_b, cap, 0, dl_bw);
3500
3501 if (!overflow) {
3502 /*
3503 * We reserve space in the destination
3504 * root_domain, as we can't fail after this point.
3505 * We will free resources in the source root_domain
3506 * later on (see set_cpus_allowed_dl()).
3507 */
3508 __dl_add(dl_b, dl_bw, dl_bw_cpus(cpu));
3509 }
3510 break;
3511 case dl_bw_req_deactivate:
3512 /*
3513 * cpu is not off yet, but we need to do the math by
3514 * considering it off already (i.e., what would happen if we
3515 * turn cpu off?).
3516 */
3517 cap -= arch_scale_cpu_capacity(cpu);
3518
3519 /*
3520 * cpu is going offline and NORMAL tasks will be moved away
3521 * from it. We can thus discount dl_server bandwidth
3522 * contribution as it won't need to be servicing tasks after
3523 * the cpu is off.
3524 */
3525 if (cpu_rq(cpu)->fair_server.dl_server)
3526 fair_server_bw = cpu_rq(cpu)->fair_server.dl_bw;
3527
3528 /*
3529 * Not much to check if no DEADLINE bandwidth is present.
3530 * dl_servers we can discount, as tasks will be moved out the
3531 * offlined CPUs anyway.
3532 */
3533 if (dl_b->total_bw - fair_server_bw > 0) {
3534 /*
3535 * Leaving at least one CPU for DEADLINE tasks seems a
3536 * wise thing to do. As said above, cpu is not offline
3537 * yet, so account for that.
3538 */
3539 if (dl_bw_cpus(cpu) - 1)
3540 overflow = __dl_overflow(dl_b, cap, fair_server_bw, 0);
3541 else
3542 overflow = 1;
3543 }
3544
3545 break;
3546 }
3547
3548 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3549 rcu_read_unlock_sched();
3550
3551 return overflow ? -EBUSY : 0;
3552 }
3553
dl_bw_deactivate(int cpu)3554 int dl_bw_deactivate(int cpu)
3555 {
3556 return dl_bw_manage(dl_bw_req_deactivate, cpu, 0);
3557 }
3558
dl_bw_alloc(int cpu,u64 dl_bw)3559 int dl_bw_alloc(int cpu, u64 dl_bw)
3560 {
3561 return dl_bw_manage(dl_bw_req_alloc, cpu, dl_bw);
3562 }
3563
dl_bw_free(int cpu,u64 dl_bw)3564 void dl_bw_free(int cpu, u64 dl_bw)
3565 {
3566 dl_bw_manage(dl_bw_req_free, cpu, dl_bw);
3567 }
3568 #endif
3569
3570 #ifdef CONFIG_SCHED_DEBUG
print_dl_stats(struct seq_file * m,int cpu)3571 void print_dl_stats(struct seq_file *m, int cpu)
3572 {
3573 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3574 }
3575 #endif /* CONFIG_SCHED_DEBUG */
3576