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