sched: remove redundant on_rq status change
[linux-2.6-block.git] / kernel / sched / deadline.c
CommitLineData
b2441318 1// SPDX-License-Identifier: GPL-2.0
aab03e05
DF
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>,
1baca4ce 14 * Juri Lelli <juri.lelli@gmail.com>,
aab03e05
DF
15 * Michael Trimarchi <michael@amarulasolutions.com>,
16 * Fabio Checconi <fchecconi@gmail.com>
17 */
18#include "sched.h"
3727e0e1 19#include "pelt.h"
aab03e05 20
332ac17e
DF
21struct dl_bandwidth def_dl_bandwidth;
22
aab03e05
DF
23static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
24{
25 return container_of(dl_se, struct task_struct, dl);
26}
27
28static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
29{
30 return container_of(dl_rq, struct rq, dl);
31}
32
33static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
34{
35 struct task_struct *p = dl_task_of(dl_se);
36 struct rq *rq = task_rq(p);
37
38 return &rq->dl;
39}
40
41static inline int on_dl_rq(struct sched_dl_entity *dl_se)
42{
43 return !RB_EMPTY_NODE(&dl_se->rb_node);
44}
45
2279f540
JL
46#ifdef CONFIG_RT_MUTEXES
47static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
48{
49 return dl_se->pi_se;
50}
51
52static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
53{
54 return pi_of(dl_se) != dl_se;
55}
56#else
57static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
58{
59 return dl_se;
60}
61
62static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
63{
64 return false;
65}
66#endif
67
06a76fe0
NP
68#ifdef CONFIG_SMP
69static inline struct dl_bw *dl_bw_of(int i)
70{
71 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
72 "sched RCU must be held");
73 return &cpu_rq(i)->rd->dl_bw;
74}
75
76static inline int dl_bw_cpus(int i)
77{
78 struct root_domain *rd = cpu_rq(i)->rd;
c81b8932 79 int cpus;
06a76fe0
NP
80
81 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
82 "sched RCU must be held");
c81b8932
DE
83
84 if (cpumask_subset(rd->span, cpu_active_mask))
85 return cpumask_weight(rd->span);
86
87 cpus = 0;
88
06a76fe0
NP
89 for_each_cpu_and(i, rd->span, cpu_active_mask)
90 cpus++;
91
92 return cpus;
93}
fc9dc698
DE
94
95static inline unsigned long __dl_bw_capacity(int i)
96{
97 struct root_domain *rd = cpu_rq(i)->rd;
98 unsigned long cap = 0;
99
100 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
101 "sched RCU must be held");
102
103 for_each_cpu_and(i, rd->span, cpu_active_mask)
104 cap += capacity_orig_of(i);
105
106 return cap;
107}
108
109/*
110 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
111 * of the CPU the task is running on rather rd's \Sum CPU capacity.
112 */
113static inline unsigned long dl_bw_capacity(int i)
114{
115 if (!static_branch_unlikely(&sched_asym_cpucapacity) &&
116 capacity_orig_of(i) == SCHED_CAPACITY_SCALE) {
117 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
118 } else {
119 return __dl_bw_capacity(i);
120 }
121}
26762423
PL
122
123static inline bool dl_bw_visited(int cpu, u64 gen)
124{
125 struct root_domain *rd = cpu_rq(cpu)->rd;
126
127 if (rd->visit_gen == gen)
128 return true;
129
130 rd->visit_gen = gen;
131 return false;
132}
06a76fe0
NP
133#else
134static inline struct dl_bw *dl_bw_of(int i)
135{
136 return &cpu_rq(i)->dl.dl_bw;
137}
138
139static inline int dl_bw_cpus(int i)
140{
141 return 1;
142}
fc9dc698
DE
143
144static inline unsigned long dl_bw_capacity(int i)
145{
146 return SCHED_CAPACITY_SCALE;
147}
26762423
PL
148
149static inline bool dl_bw_visited(int cpu, u64 gen)
150{
151 return false;
152}
06a76fe0
NP
153#endif
154
e36d8677 155static inline
794a56eb 156void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
e36d8677
LA
157{
158 u64 old = dl_rq->running_bw;
159
5cb9eaa3 160 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
e36d8677
LA
161 dl_rq->running_bw += dl_bw;
162 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
8fd27231 163 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
e0367b12 164 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
4042d003 165 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
e36d8677
LA
166}
167
168static inline
794a56eb 169void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
e36d8677
LA
170{
171 u64 old = dl_rq->running_bw;
172
5cb9eaa3 173 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
e36d8677
LA
174 dl_rq->running_bw -= dl_bw;
175 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
176 if (dl_rq->running_bw > old)
177 dl_rq->running_bw = 0;
e0367b12 178 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
4042d003 179 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
e36d8677
LA
180}
181
8fd27231 182static inline
794a56eb 183void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
8fd27231
LA
184{
185 u64 old = dl_rq->this_bw;
186
5cb9eaa3 187 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
8fd27231
LA
188 dl_rq->this_bw += dl_bw;
189 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
190}
191
192static inline
794a56eb 193void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
8fd27231
LA
194{
195 u64 old = dl_rq->this_bw;
196
5cb9eaa3 197 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
8fd27231
LA
198 dl_rq->this_bw -= dl_bw;
199 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
200 if (dl_rq->this_bw > old)
201 dl_rq->this_bw = 0;
202 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
203}
204
794a56eb
JL
205static inline
206void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
207{
208 if (!dl_entity_is_special(dl_se))
209 __add_rq_bw(dl_se->dl_bw, dl_rq);
210}
211
212static inline
213void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
214{
215 if (!dl_entity_is_special(dl_se))
216 __sub_rq_bw(dl_se->dl_bw, dl_rq);
217}
218
219static inline
220void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
221{
222 if (!dl_entity_is_special(dl_se))
223 __add_running_bw(dl_se->dl_bw, dl_rq);
224}
225
226static inline
227void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
228{
229 if (!dl_entity_is_special(dl_se))
230 __sub_running_bw(dl_se->dl_bw, dl_rq);
231}
232
ba4f7bc1 233static void dl_change_utilization(struct task_struct *p, u64 new_bw)
209a0cbd 234{
8fd27231 235 struct rq *rq;
209a0cbd 236
794a56eb
JL
237 BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV);
238
8fd27231 239 if (task_on_rq_queued(p))
209a0cbd
LA
240 return;
241
8fd27231
LA
242 rq = task_rq(p);
243 if (p->dl.dl_non_contending) {
794a56eb 244 sub_running_bw(&p->dl, &rq->dl);
8fd27231
LA
245 p->dl.dl_non_contending = 0;
246 /*
247 * If the timer handler is currently running and the
3b03706f 248 * timer cannot be canceled, inactive_task_timer()
8fd27231
LA
249 * will see that dl_not_contending is not set, and
250 * will not touch the rq's active utilization,
251 * so we are still safe.
252 */
253 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
254 put_task_struct(p);
255 }
794a56eb
JL
256 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
257 __add_rq_bw(new_bw, &rq->dl);
209a0cbd
LA
258}
259
260/*
261 * The utilization of a task cannot be immediately removed from
262 * the rq active utilization (running_bw) when the task blocks.
263 * Instead, we have to wait for the so called "0-lag time".
264 *
265 * If a task blocks before the "0-lag time", a timer (the inactive
266 * timer) is armed, and running_bw is decreased when the timer
267 * fires.
268 *
269 * If the task wakes up again before the inactive timer fires,
3b03706f 270 * the timer is canceled, whereas if the task wakes up after the
209a0cbd
LA
271 * inactive timer fired (and running_bw has been decreased) the
272 * task's utilization has to be added to running_bw again.
273 * A flag in the deadline scheduling entity (dl_non_contending)
274 * is used to avoid race conditions between the inactive timer handler
275 * and task wakeups.
276 *
277 * The following diagram shows how running_bw is updated. A task is
278 * "ACTIVE" when its utilization contributes to running_bw; an
279 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
280 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
281 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
282 * time already passed, which does not contribute to running_bw anymore.
283 * +------------------+
284 * wakeup | ACTIVE |
285 * +------------------>+ contending |
286 * | add_running_bw | |
287 * | +----+------+------+
288 * | | ^
289 * | dequeue | |
290 * +--------+-------+ | |
291 * | | t >= 0-lag | | wakeup
292 * | INACTIVE |<---------------+ |
293 * | | sub_running_bw | |
294 * +--------+-------+ | |
295 * ^ | |
296 * | t < 0-lag | |
297 * | | |
298 * | V |
299 * | +----+------+------+
300 * | sub_running_bw | ACTIVE |
301 * +-------------------+ |
302 * inactive timer | non contending |
303 * fired +------------------+
304 *
305 * The task_non_contending() function is invoked when a task
306 * blocks, and checks if the 0-lag time already passed or
307 * not (in the first case, it directly updates running_bw;
308 * in the second case, it arms the inactive timer).
309 *
310 * The task_contending() function is invoked when a task wakes
311 * up, and checks if the task is still in the "ACTIVE non contending"
312 * state or not (in the second case, it updates running_bw).
313 */
314static void task_non_contending(struct task_struct *p)
315{
316 struct sched_dl_entity *dl_se = &p->dl;
317 struct hrtimer *timer = &dl_se->inactive_timer;
318 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
319 struct rq *rq = rq_of_dl_rq(dl_rq);
320 s64 zerolag_time;
321
322 /*
323 * If this is a non-deadline task that has been boosted,
324 * do nothing
325 */
326 if (dl_se->dl_runtime == 0)
327 return;
328
794a56eb
JL
329 if (dl_entity_is_special(dl_se))
330 return;
331
209a0cbd
LA
332 WARN_ON(dl_se->dl_non_contending);
333
334 zerolag_time = dl_se->deadline -
335 div64_long((dl_se->runtime * dl_se->dl_period),
336 dl_se->dl_runtime);
337
338 /*
339 * Using relative times instead of the absolute "0-lag time"
340 * allows to simplify the code
341 */
342 zerolag_time -= rq_clock(rq);
343
344 /*
345 * If the "0-lag time" already passed, decrease the active
346 * utilization now, instead of starting a timer
347 */
1b02cd6a 348 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
209a0cbd 349 if (dl_task(p))
794a56eb 350 sub_running_bw(dl_se, dl_rq);
2f064a59 351 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
387e3130
LA
352 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
353
2f064a59 354 if (READ_ONCE(p->__state) == TASK_DEAD)
794a56eb 355 sub_rq_bw(&p->dl, &rq->dl);
387e3130 356 raw_spin_lock(&dl_b->lock);
8c0944ce 357 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
209a0cbd 358 __dl_clear_params(p);
387e3130
LA
359 raw_spin_unlock(&dl_b->lock);
360 }
209a0cbd
LA
361
362 return;
363 }
364
365 dl_se->dl_non_contending = 1;
366 get_task_struct(p);
850377a8 367 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
209a0cbd
LA
368}
369
8fd27231 370static void task_contending(struct sched_dl_entity *dl_se, int flags)
209a0cbd
LA
371{
372 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
373
374 /*
375 * If this is a non-deadline task that has been boosted,
376 * do nothing
377 */
378 if (dl_se->dl_runtime == 0)
379 return;
380
8fd27231 381 if (flags & ENQUEUE_MIGRATED)
794a56eb 382 add_rq_bw(dl_se, dl_rq);
8fd27231 383
209a0cbd
LA
384 if (dl_se->dl_non_contending) {
385 dl_se->dl_non_contending = 0;
386 /*
387 * If the timer handler is currently running and the
3b03706f 388 * timer cannot be canceled, inactive_task_timer()
209a0cbd
LA
389 * will see that dl_not_contending is not set, and
390 * will not touch the rq's active utilization,
391 * so we are still safe.
392 */
393 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
394 put_task_struct(dl_task_of(dl_se));
395 } else {
396 /*
397 * Since "dl_non_contending" is not set, the
398 * task's utilization has already been removed from
399 * active utilization (either when the task blocked,
400 * when the "inactive timer" fired).
401 * So, add it back.
402 */
794a56eb 403 add_running_bw(dl_se, dl_rq);
209a0cbd
LA
404 }
405}
406
aab03e05
DF
407static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
408{
409 struct sched_dl_entity *dl_se = &p->dl;
410
2161573e 411 return dl_rq->root.rb_leftmost == &dl_se->rb_node;
aab03e05
DF
412}
413
ba4f7bc1
YC
414static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
415
332ac17e
DF
416void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
417{
418 raw_spin_lock_init(&dl_b->dl_runtime_lock);
419 dl_b->dl_period = period;
420 dl_b->dl_runtime = runtime;
421}
422
332ac17e
DF
423void init_dl_bw(struct dl_bw *dl_b)
424{
425 raw_spin_lock_init(&dl_b->lock);
426 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
1724813d 427 if (global_rt_runtime() == RUNTIME_INF)
332ac17e
DF
428 dl_b->bw = -1;
429 else
1724813d 430 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
332ac17e
DF
431 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
432 dl_b->total_bw = 0;
433}
434
07c54f7a 435void init_dl_rq(struct dl_rq *dl_rq)
aab03e05 436{
2161573e 437 dl_rq->root = RB_ROOT_CACHED;
1baca4ce
JL
438
439#ifdef CONFIG_SMP
440 /* zero means no -deadline tasks */
441 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
442
443 dl_rq->dl_nr_migratory = 0;
444 dl_rq->overloaded = 0;
2161573e 445 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
332ac17e
DF
446#else
447 init_dl_bw(&dl_rq->dl_bw);
1baca4ce 448#endif
e36d8677
LA
449
450 dl_rq->running_bw = 0;
8fd27231 451 dl_rq->this_bw = 0;
4da3abce 452 init_dl_rq_bw_ratio(dl_rq);
1baca4ce
JL
453}
454
455#ifdef CONFIG_SMP
456
457static inline int dl_overloaded(struct rq *rq)
458{
459 return atomic_read(&rq->rd->dlo_count);
460}
461
462static inline void dl_set_overload(struct rq *rq)
463{
464 if (!rq->online)
465 return;
466
467 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
468 /*
469 * Must be visible before the overload count is
470 * set (as in sched_rt.c).
471 *
472 * Matched by the barrier in pull_dl_task().
473 */
474 smp_wmb();
475 atomic_inc(&rq->rd->dlo_count);
476}
477
478static inline void dl_clear_overload(struct rq *rq)
479{
480 if (!rq->online)
481 return;
482
483 atomic_dec(&rq->rd->dlo_count);
484 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
485}
486
487static void update_dl_migration(struct dl_rq *dl_rq)
488{
995b9ea4 489 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
1baca4ce
JL
490 if (!dl_rq->overloaded) {
491 dl_set_overload(rq_of_dl_rq(dl_rq));
492 dl_rq->overloaded = 1;
493 }
494 } else if (dl_rq->overloaded) {
495 dl_clear_overload(rq_of_dl_rq(dl_rq));
496 dl_rq->overloaded = 0;
497 }
498}
499
500static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
501{
502 struct task_struct *p = dl_task_of(dl_se);
1baca4ce 503
4b53a341 504 if (p->nr_cpus_allowed > 1)
1baca4ce
JL
505 dl_rq->dl_nr_migratory++;
506
507 update_dl_migration(dl_rq);
508}
509
510static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
511{
512 struct task_struct *p = dl_task_of(dl_se);
1baca4ce 513
4b53a341 514 if (p->nr_cpus_allowed > 1)
1baca4ce
JL
515 dl_rq->dl_nr_migratory--;
516
517 update_dl_migration(dl_rq);
518}
519
8ecca394
PZ
520#define __node_2_pdl(node) \
521 rb_entry((node), struct task_struct, pushable_dl_tasks)
522
523static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
524{
525 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
526}
527
1baca4ce
JL
528/*
529 * The list of pushable -deadline task is not a plist, like in
530 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
531 */
532static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
533{
8ecca394 534 struct rb_node *leftmost;
1baca4ce
JL
535
536 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
537
8ecca394
PZ
538 leftmost = rb_add_cached(&p->pushable_dl_tasks,
539 &rq->dl.pushable_dl_tasks_root,
540 __pushable_less);
2161573e 541 if (leftmost)
8ecca394 542 rq->dl.earliest_dl.next = p->dl.deadline;
aab03e05
DF
543}
544
1baca4ce
JL
545static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
546{
547 struct dl_rq *dl_rq = &rq->dl;
8ecca394
PZ
548 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
549 struct rb_node *leftmost;
1baca4ce
JL
550
551 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
552 return;
553
8ecca394
PZ
554 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
555 if (leftmost)
556 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
1baca4ce 557
1baca4ce
JL
558 RB_CLEAR_NODE(&p->pushable_dl_tasks);
559}
560
561static inline int has_pushable_dl_tasks(struct rq *rq)
562{
2161573e 563 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
1baca4ce
JL
564}
565
566static int push_dl_task(struct rq *rq);
567
dc877341
PZ
568static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
569{
120455c5 570 return rq->online && dl_task(prev);
dc877341
PZ
571}
572
9916e214
PZ
573static DEFINE_PER_CPU(struct callback_head, dl_push_head);
574static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
e3fca9e7
PZ
575
576static void push_dl_tasks(struct rq *);
9916e214 577static void pull_dl_task(struct rq *);
e3fca9e7 578
02d8ec94 579static inline void deadline_queue_push_tasks(struct rq *rq)
dc877341 580{
e3fca9e7
PZ
581 if (!has_pushable_dl_tasks(rq))
582 return;
583
9916e214
PZ
584 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
585}
586
02d8ec94 587static inline void deadline_queue_pull_task(struct rq *rq)
9916e214
PZ
588{
589 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
dc877341
PZ
590}
591
fa9c9d10
WL
592static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
593
a649f237 594static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
fa9c9d10
WL
595{
596 struct rq *later_rq = NULL;
59d06cea 597 struct dl_bw *dl_b;
fa9c9d10
WL
598
599 later_rq = find_lock_later_rq(p, rq);
fa9c9d10
WL
600 if (!later_rq) {
601 int cpu;
602
603 /*
604 * If we cannot preempt any rq, fall back to pick any
97fb7a0a 605 * online CPU:
fa9c9d10 606 */
3bd37062 607 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
fa9c9d10
WL
608 if (cpu >= nr_cpu_ids) {
609 /*
97fb7a0a 610 * Failed to find any suitable CPU.
fa9c9d10
WL
611 * The task will never come back!
612 */
613 BUG_ON(dl_bandwidth_enabled());
614
615 /*
616 * If admission control is disabled we
617 * try a little harder to let the task
618 * run.
619 */
620 cpu = cpumask_any(cpu_active_mask);
621 }
622 later_rq = cpu_rq(cpu);
623 double_lock_balance(rq, later_rq);
624 }
625
59d06cea
JL
626 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
627 /*
628 * Inactive timer is armed (or callback is running, but
629 * waiting for us to release rq locks). In any case, when it
630 * will fire (or continue), it will see running_bw of this
631 * task migrated to later_rq (and correctly handle it).
632 */
633 sub_running_bw(&p->dl, &rq->dl);
634 sub_rq_bw(&p->dl, &rq->dl);
635
636 add_rq_bw(&p->dl, &later_rq->dl);
637 add_running_bw(&p->dl, &later_rq->dl);
638 } else {
639 sub_rq_bw(&p->dl, &rq->dl);
640 add_rq_bw(&p->dl, &later_rq->dl);
641 }
642
643 /*
644 * And we finally need to fixup root_domain(s) bandwidth accounting,
645 * since p is still hanging out in the old (now moved to default) root
646 * domain.
647 */
648 dl_b = &rq->rd->dl_bw;
649 raw_spin_lock(&dl_b->lock);
650 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
651 raw_spin_unlock(&dl_b->lock);
652
653 dl_b = &later_rq->rd->dl_bw;
654 raw_spin_lock(&dl_b->lock);
655 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
656 raw_spin_unlock(&dl_b->lock);
657
fa9c9d10 658 set_task_cpu(p, later_rq->cpu);
a649f237
PZ
659 double_unlock_balance(later_rq, rq);
660
661 return later_rq;
fa9c9d10
WL
662}
663
1baca4ce
JL
664#else
665
666static inline
667void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
668{
669}
670
671static inline
672void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
673{
674}
675
676static inline
677void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
678{
679}
680
681static inline
682void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
683{
684}
685
dc877341
PZ
686static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
687{
688 return false;
689}
690
0ea60c20 691static inline void pull_dl_task(struct rq *rq)
dc877341 692{
dc877341
PZ
693}
694
02d8ec94 695static inline void deadline_queue_push_tasks(struct rq *rq)
dc877341 696{
dc877341
PZ
697}
698
02d8ec94 699static inline void deadline_queue_pull_task(struct rq *rq)
dc877341
PZ
700{
701}
1baca4ce
JL
702#endif /* CONFIG_SMP */
703
aab03e05
DF
704static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
705static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
97fb7a0a 706static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
aab03e05
DF
707
708/*
709 * We are being explicitly informed that a new instance is starting,
710 * and this means that:
711 * - the absolute deadline of the entity has to be placed at
712 * current time + relative deadline;
713 * - the runtime of the entity has to be set to the maximum value.
714 *
715 * The capability of specifying such event is useful whenever a -deadline
716 * entity wants to (try to!) synchronize its behaviour with the scheduler's
717 * one, and to (try to!) reconcile itself with its own scheduling
718 * parameters.
719 */
98b0a857 720static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
aab03e05
DF
721{
722 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
723 struct rq *rq = rq_of_dl_rq(dl_rq);
724
2279f540 725 WARN_ON(is_dl_boosted(dl_se));
72f9f3fd
LA
726 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
727
728 /*
729 * We are racing with the deadline timer. So, do nothing because
730 * the deadline timer handler will take care of properly recharging
731 * the runtime and postponing the deadline
732 */
733 if (dl_se->dl_throttled)
734 return;
aab03e05
DF
735
736 /*
737 * We use the regular wall clock time to set deadlines in the
738 * future; in fact, we must consider execution overheads (time
739 * spent on hardirq context, etc.).
740 */
98b0a857
JL
741 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
742 dl_se->runtime = dl_se->dl_runtime;
aab03e05
DF
743}
744
745/*
746 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
747 * possibility of a entity lasting more than what it declared, and thus
748 * exhausting its runtime.
749 *
750 * Here we are interested in making runtime overrun possible, but we do
751 * not want a entity which is misbehaving to affect the scheduling of all
752 * other entities.
753 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
754 * is used, in order to confine each entity within its own bandwidth.
755 *
756 * This function deals exactly with that, and ensures that when the runtime
757 * of a entity is replenished, its deadline is also postponed. That ensures
758 * the overrunning entity can't interfere with other entity in the system and
759 * can't make them miss their deadlines. Reasons why this kind of overruns
760 * could happen are, typically, a entity voluntarily trying to overcome its
1b09d29b 761 * runtime, or it just underestimated it during sched_setattr().
aab03e05 762 */
2279f540 763static void replenish_dl_entity(struct sched_dl_entity *dl_se)
aab03e05
DF
764{
765 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
766 struct rq *rq = rq_of_dl_rq(dl_rq);
767
2279f540 768 BUG_ON(pi_of(dl_se)->dl_runtime <= 0);
2d3d891d
DF
769
770 /*
771 * This could be the case for a !-dl task that is boosted.
772 * Just go with full inherited parameters.
773 */
774 if (dl_se->dl_deadline == 0) {
2279f540
JL
775 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
776 dl_se->runtime = pi_of(dl_se)->dl_runtime;
2d3d891d
DF
777 }
778
48be3a67
PZ
779 if (dl_se->dl_yielded && dl_se->runtime > 0)
780 dl_se->runtime = 0;
781
aab03e05
DF
782 /*
783 * We keep moving the deadline away until we get some
784 * available runtime for the entity. This ensures correct
785 * handling of situations where the runtime overrun is
786 * arbitrary large.
787 */
788 while (dl_se->runtime <= 0) {
2279f540
JL
789 dl_se->deadline += pi_of(dl_se)->dl_period;
790 dl_se->runtime += pi_of(dl_se)->dl_runtime;
aab03e05
DF
791 }
792
793 /*
794 * At this point, the deadline really should be "in
795 * the future" with respect to rq->clock. If it's
796 * not, we are, for some reason, lagging too much!
797 * Anyway, after having warn userspace abut that,
798 * we still try to keep the things running by
799 * resetting the deadline and the budget of the
800 * entity.
801 */
802 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
c219b7dd 803 printk_deferred_once("sched: DL replenish lagged too much\n");
2279f540
JL
804 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
805 dl_se->runtime = pi_of(dl_se)->dl_runtime;
aab03e05 806 }
1019a359
PZ
807
808 if (dl_se->dl_yielded)
809 dl_se->dl_yielded = 0;
810 if (dl_se->dl_throttled)
811 dl_se->dl_throttled = 0;
aab03e05
DF
812}
813
814/*
815 * Here we check if --at time t-- an entity (which is probably being
816 * [re]activated or, in general, enqueued) can use its remaining runtime
817 * and its current deadline _without_ exceeding the bandwidth it is
818 * assigned (function returns true if it can't). We are in fact applying
819 * one of the CBS rules: when a task wakes up, if the residual runtime
820 * over residual deadline fits within the allocated bandwidth, then we
821 * can keep the current (absolute) deadline and residual budget without
822 * disrupting the schedulability of the system. Otherwise, we should
823 * refill the runtime and set the deadline a period in the future,
824 * because keeping the current (absolute) deadline of the task would
712e5e34 825 * result in breaking guarantees promised to other tasks (refer to
d6a3b247 826 * Documentation/scheduler/sched-deadline.rst for more information).
aab03e05
DF
827 *
828 * This function returns true if:
829 *
2317d5f1 830 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
aab03e05
DF
831 *
832 * IOW we can't recycle current parameters.
755378a4 833 *
2317d5f1 834 * Notice that the bandwidth check is done against the deadline. For
755378a4 835 * task with deadline equal to period this is the same of using
2317d5f1 836 * dl_period instead of dl_deadline in the equation above.
aab03e05 837 */
2279f540 838static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
aab03e05
DF
839{
840 u64 left, right;
841
842 /*
843 * left and right are the two sides of the equation above,
844 * after a bit of shuffling to use multiplications instead
845 * of divisions.
846 *
847 * Note that none of the time values involved in the two
848 * multiplications are absolute: dl_deadline and dl_runtime
849 * are the relative deadline and the maximum runtime of each
850 * instance, runtime is the runtime left for the last instance
851 * and (deadline - t), since t is rq->clock, is the time left
852 * to the (absolute) deadline. Even if overflowing the u64 type
853 * is very unlikely to occur in both cases, here we scale down
854 * as we want to avoid that risk at all. Scaling down by 10
855 * means that we reduce granularity to 1us. We are fine with it,
856 * since this is only a true/false check and, anyway, thinking
857 * of anything below microseconds resolution is actually fiction
858 * (but still we want to give the user that illusion >;).
859 */
2279f540 860 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
332ac17e 861 right = ((dl_se->deadline - t) >> DL_SCALE) *
2279f540 862 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
aab03e05
DF
863
864 return dl_time_before(right, left);
865}
866
867/*
3effcb42
DBO
868 * Revised wakeup rule [1]: For self-suspending tasks, rather then
869 * re-initializing task's runtime and deadline, the revised wakeup
870 * rule adjusts the task's runtime to avoid the task to overrun its
871 * density.
aab03e05 872 *
3effcb42
DBO
873 * Reasoning: a task may overrun the density if:
874 * runtime / (deadline - t) > dl_runtime / dl_deadline
875 *
876 * Therefore, runtime can be adjusted to:
877 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
878 *
879 * In such way that runtime will be equal to the maximum density
880 * the task can use without breaking any rule.
881 *
882 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
883 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
884 */
885static void
886update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
887{
888 u64 laxity = dl_se->deadline - rq_clock(rq);
889
890 /*
891 * If the task has deadline < period, and the deadline is in the past,
892 * it should already be throttled before this check.
893 *
894 * See update_dl_entity() comments for further details.
895 */
896 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
897
898 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
899}
900
901/*
902 * Regarding the deadline, a task with implicit deadline has a relative
903 * deadline == relative period. A task with constrained deadline has a
904 * relative deadline <= relative period.
905 *
906 * We support constrained deadline tasks. However, there are some restrictions
907 * applied only for tasks which do not have an implicit deadline. See
908 * update_dl_entity() to know more about such restrictions.
909 *
910 * The dl_is_implicit() returns true if the task has an implicit deadline.
911 */
912static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
913{
914 return dl_se->dl_deadline == dl_se->dl_period;
915}
916
917/*
918 * When a deadline entity is placed in the runqueue, its runtime and deadline
919 * might need to be updated. This is done by a CBS wake up rule. There are two
920 * different rules: 1) the original CBS; and 2) the Revisited CBS.
921 *
922 * When the task is starting a new period, the Original CBS is used. In this
923 * case, the runtime is replenished and a new absolute deadline is set.
924 *
925 * When a task is queued before the begin of the next period, using the
926 * remaining runtime and deadline could make the entity to overflow, see
927 * dl_entity_overflow() to find more about runtime overflow. When such case
928 * is detected, the runtime and deadline need to be updated.
929 *
930 * If the task has an implicit deadline, i.e., deadline == period, the Original
931 * CBS is applied. the runtime is replenished and a new absolute deadline is
932 * set, as in the previous cases.
933 *
934 * However, the Original CBS does not work properly for tasks with
935 * deadline < period, which are said to have a constrained deadline. By
936 * applying the Original CBS, a constrained deadline task would be able to run
937 * runtime/deadline in a period. With deadline < period, the task would
938 * overrun the runtime/period allowed bandwidth, breaking the admission test.
939 *
940 * In order to prevent this misbehave, the Revisited CBS is used for
941 * constrained deadline tasks when a runtime overflow is detected. In the
942 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
943 * the remaining runtime of the task is reduced to avoid runtime overflow.
944 * Please refer to the comments update_dl_revised_wakeup() function to find
945 * more about the Revised CBS rule.
aab03e05 946 */
2279f540 947static void update_dl_entity(struct sched_dl_entity *dl_se)
aab03e05
DF
948{
949 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
950 struct rq *rq = rq_of_dl_rq(dl_rq);
951
aab03e05 952 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
2279f540 953 dl_entity_overflow(dl_se, rq_clock(rq))) {
3effcb42
DBO
954
955 if (unlikely(!dl_is_implicit(dl_se) &&
956 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
2279f540 957 !is_dl_boosted(dl_se))) {
3effcb42
DBO
958 update_dl_revised_wakeup(dl_se, rq);
959 return;
960 }
961
2279f540
JL
962 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
963 dl_se->runtime = pi_of(dl_se)->dl_runtime;
aab03e05
DF
964 }
965}
966
5ac69d37
DBO
967static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
968{
969 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
970}
971
aab03e05
DF
972/*
973 * If the entity depleted all its runtime, and if we want it to sleep
974 * while waiting for some new execution time to become available, we
5ac69d37 975 * set the bandwidth replenishment timer to the replenishment instant
aab03e05
DF
976 * and try to activate it.
977 *
978 * Notice that it is important for the caller to know if the timer
979 * actually started or not (i.e., the replenishment instant is in
980 * the future or in the past).
981 */
a649f237 982static int start_dl_timer(struct task_struct *p)
aab03e05 983{
a649f237
PZ
984 struct sched_dl_entity *dl_se = &p->dl;
985 struct hrtimer *timer = &dl_se->dl_timer;
986 struct rq *rq = task_rq(p);
aab03e05 987 ktime_t now, act;
aab03e05
DF
988 s64 delta;
989
5cb9eaa3 990 lockdep_assert_rq_held(rq);
a649f237 991
aab03e05
DF
992 /*
993 * We want the timer to fire at the deadline, but considering
994 * that it is actually coming from rq->clock and not from
995 * hrtimer's time base reading.
996 */
5ac69d37 997 act = ns_to_ktime(dl_next_period(dl_se));
a649f237 998 now = hrtimer_cb_get_time(timer);
aab03e05
DF
999 delta = ktime_to_ns(now) - rq_clock(rq);
1000 act = ktime_add_ns(act, delta);
1001
1002 /*
1003 * If the expiry time already passed, e.g., because the value
1004 * chosen as the deadline is too small, don't even try to
1005 * start the timer in the past!
1006 */
1007 if (ktime_us_delta(act, now) < 0)
1008 return 0;
1009
a649f237
PZ
1010 /*
1011 * !enqueued will guarantee another callback; even if one is already in
1012 * progress. This ensures a balanced {get,put}_task_struct().
1013 *
1014 * The race against __run_timer() clearing the enqueued state is
1015 * harmless because we're holding task_rq()->lock, therefore the timer
1016 * expiring after we've done the check will wait on its task_rq_lock()
1017 * and observe our state.
1018 */
1019 if (!hrtimer_is_queued(timer)) {
1020 get_task_struct(p);
d5096aa6 1021 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
a649f237 1022 }
aab03e05 1023
cc9684d3 1024 return 1;
aab03e05
DF
1025}
1026
1027/*
1028 * This is the bandwidth enforcement timer callback. If here, we know
1029 * a task is not on its dl_rq, since the fact that the timer was running
1030 * means the task is throttled and needs a runtime replenishment.
1031 *
1032 * However, what we actually do depends on the fact the task is active,
1033 * (it is on its rq) or has been removed from there by a call to
1034 * dequeue_task_dl(). In the former case we must issue the runtime
1035 * replenishment and add the task back to the dl_rq; in the latter, we just
1036 * do nothing but clearing dl_throttled, so that runtime and deadline
1037 * updating (and the queueing back to dl_rq) will be done by the
1038 * next call to enqueue_task_dl().
1039 */
1040static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1041{
1042 struct sched_dl_entity *dl_se = container_of(timer,
1043 struct sched_dl_entity,
1044 dl_timer);
1045 struct task_struct *p = dl_task_of(dl_se);
eb580751 1046 struct rq_flags rf;
0f397f2c 1047 struct rq *rq;
3960c8c0 1048
eb580751 1049 rq = task_rq_lock(p, &rf);
0f397f2c 1050
aab03e05 1051 /*
a649f237 1052 * The task might have changed its scheduling policy to something
9846d50d 1053 * different than SCHED_DEADLINE (through switched_from_dl()).
a649f237 1054 */
209a0cbd 1055 if (!dl_task(p))
a649f237 1056 goto unlock;
a649f237 1057
a649f237
PZ
1058 /*
1059 * The task might have been boosted by someone else and might be in the
1060 * boosting/deboosting path, its not throttled.
1061 */
2279f540 1062 if (is_dl_boosted(dl_se))
a649f237 1063 goto unlock;
a79ec89f 1064
fa9c9d10 1065 /*
a649f237
PZ
1066 * Spurious timer due to start_dl_timer() race; or we already received
1067 * a replenishment from rt_mutex_setprio().
fa9c9d10 1068 */
a649f237 1069 if (!dl_se->dl_throttled)
fa9c9d10 1070 goto unlock;
a649f237
PZ
1071
1072 sched_clock_tick();
1073 update_rq_clock(rq);
fa9c9d10 1074
a79ec89f
KT
1075 /*
1076 * If the throttle happened during sched-out; like:
1077 *
1078 * schedule()
1079 * deactivate_task()
1080 * dequeue_task_dl()
1081 * update_curr_dl()
1082 * start_dl_timer()
1083 * __dequeue_task_dl()
1084 * prev->on_rq = 0;
1085 *
1086 * We can be both throttled and !queued. Replenish the counter
1087 * but do not enqueue -- wait for our wakeup to do that.
1088 */
1089 if (!task_on_rq_queued(p)) {
2279f540 1090 replenish_dl_entity(dl_se);
a79ec89f
KT
1091 goto unlock;
1092 }
1093
1baca4ce 1094#ifdef CONFIG_SMP
c0c8c9fa 1095 if (unlikely(!rq->online)) {
61c7aca6
WL
1096 /*
1097 * If the runqueue is no longer available, migrate the
1098 * task elsewhere. This necessarily changes rq.
1099 */
9ef7e7e3 1100 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
a649f237 1101 rq = dl_task_offline_migration(rq, p);
9ef7e7e3 1102 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
dcc3b5ff 1103 update_rq_clock(rq);
61c7aca6
WL
1104
1105 /*
1106 * Now that the task has been migrated to the new RQ and we
1107 * have that locked, proceed as normal and enqueue the task
1108 * there.
1109 */
c0c8c9fa 1110 }
61c7aca6 1111#endif
a649f237 1112
61c7aca6
WL
1113 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1114 if (dl_task(rq->curr))
1115 check_preempt_curr_dl(rq, p, 0);
1116 else
1117 resched_curr(rq);
a649f237 1118
61c7aca6 1119#ifdef CONFIG_SMP
a649f237
PZ
1120 /*
1121 * Queueing this task back might have overloaded rq, check if we need
1122 * to kick someone away.
1019a359 1123 */
0aaafaab
PZ
1124 if (has_pushable_dl_tasks(rq)) {
1125 /*
1126 * Nothing relies on rq->lock after this, so its safe to drop
1127 * rq->lock.
1128 */
d8ac8971 1129 rq_unpin_lock(rq, &rf);
1019a359 1130 push_dl_task(rq);
d8ac8971 1131 rq_repin_lock(rq, &rf);
0aaafaab 1132 }
1baca4ce 1133#endif
a649f237 1134
aab03e05 1135unlock:
eb580751 1136 task_rq_unlock(rq, p, &rf);
aab03e05 1137
a649f237
PZ
1138 /*
1139 * This can free the task_struct, including this hrtimer, do not touch
1140 * anything related to that after this.
1141 */
1142 put_task_struct(p);
1143
aab03e05
DF
1144 return HRTIMER_NORESTART;
1145}
1146
1147void init_dl_task_timer(struct sched_dl_entity *dl_se)
1148{
1149 struct hrtimer *timer = &dl_se->dl_timer;
1150
d5096aa6 1151 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
aab03e05
DF
1152 timer->function = dl_task_timer;
1153}
1154
df8eac8c
DBO
1155/*
1156 * During the activation, CBS checks if it can reuse the current task's
1157 * runtime and period. If the deadline of the task is in the past, CBS
1158 * cannot use the runtime, and so it replenishes the task. This rule
1159 * works fine for implicit deadline tasks (deadline == period), and the
1160 * CBS was designed for implicit deadline tasks. However, a task with
c4969417 1161 * constrained deadline (deadline < period) might be awakened after the
df8eac8c
DBO
1162 * deadline, but before the next period. In this case, replenishing the
1163 * task would allow it to run for runtime / deadline. As in this case
1164 * deadline < period, CBS enables a task to run for more than the
1165 * runtime / period. In a very loaded system, this can cause a domino
1166 * effect, making other tasks miss their deadlines.
1167 *
1168 * To avoid this problem, in the activation of a constrained deadline
1169 * task after the deadline but before the next period, throttle the
1170 * task and set the replenishing timer to the begin of the next period,
1171 * unless it is boosted.
1172 */
1173static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1174{
1175 struct task_struct *p = dl_task_of(dl_se);
1176 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1177
1178 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1179 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
2279f540 1180 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(p)))
df8eac8c
DBO
1181 return;
1182 dl_se->dl_throttled = 1;
ae83b56a
XP
1183 if (dl_se->runtime > 0)
1184 dl_se->runtime = 0;
df8eac8c
DBO
1185 }
1186}
1187
aab03e05 1188static
6fab5410 1189int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
aab03e05 1190{
269ad801 1191 return (dl_se->runtime <= 0);
aab03e05
DF
1192}
1193
faa59937
JL
1194extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
1195
c52f14d3
LA
1196/*
1197 * This function implements the GRUB accounting rule:
1198 * according to the GRUB reclaiming algorithm, the runtime is
daec5798
LA
1199 * not decreased as "dq = -dt", but as
1200 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1201 * where u is the utilization of the task, Umax is the maximum reclaimable
1202 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1203 * as the difference between the "total runqueue utilization" and the
1204 * runqueue active utilization, and Uextra is the (per runqueue) extra
1205 * reclaimable utilization.
9f0d1a50 1206 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
daec5798
LA
1207 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1208 * BW_SHIFT.
3b03706f 1209 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT,
daec5798
LA
1210 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1211 * Since delta is a 64 bit variable, to have an overflow its value
1212 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1213 * So, overflow is not an issue here.
c52f14d3 1214 */
3febfc8a 1215static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
c52f14d3 1216{
9f0d1a50
LA
1217 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1218 u64 u_act;
daec5798 1219 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
c52f14d3 1220
9f0d1a50 1221 /*
daec5798
LA
1222 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1223 * we compare u_inact + rq->dl.extra_bw with
1224 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1225 * u_inact + rq->dl.extra_bw can be larger than
1226 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1227 * leading to wrong results)
9f0d1a50 1228 */
daec5798
LA
1229 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1230 u_act = u_act_min;
9f0d1a50 1231 else
daec5798 1232 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
9f0d1a50
LA
1233
1234 return (delta * u_act) >> BW_SHIFT;
c52f14d3
LA
1235}
1236
aab03e05
DF
1237/*
1238 * Update the current task's runtime statistics (provided it is still
1239 * a -deadline task and has not been removed from the dl_rq).
1240 */
1241static void update_curr_dl(struct rq *rq)
1242{
1243 struct task_struct *curr = rq->curr;
1244 struct sched_dl_entity *dl_se = &curr->dl;
07881166
JL
1245 u64 delta_exec, scaled_delta_exec;
1246 int cpu = cpu_of(rq);
6fe0ce1e 1247 u64 now;
aab03e05
DF
1248
1249 if (!dl_task(curr) || !on_dl_rq(dl_se))
1250 return;
1251
1252 /*
1253 * Consumed budget is computed considering the time as
1254 * observed by schedulable tasks (excluding time spent
1255 * in hardirq context, etc.). Deadlines are instead
1256 * computed using hard walltime. This seems to be the more
1257 * natural solution, but the full ramifications of this
1258 * approach need further study.
1259 */
6fe0ce1e
WY
1260 now = rq_clock_task(rq);
1261 delta_exec = now - curr->se.exec_start;
48be3a67
PZ
1262 if (unlikely((s64)delta_exec <= 0)) {
1263 if (unlikely(dl_se->dl_yielded))
1264 goto throttle;
734ff2a7 1265 return;
48be3a67 1266 }
aab03e05
DF
1267
1268 schedstat_set(curr->se.statistics.exec_max,
1269 max(curr->se.statistics.exec_max, delta_exec));
1270
1271 curr->se.sum_exec_runtime += delta_exec;
1272 account_group_exec_runtime(curr, delta_exec);
1273
6fe0ce1e 1274 curr->se.exec_start = now;
d2cc5ed6 1275 cgroup_account_cputime(curr, delta_exec);
aab03e05 1276
794a56eb
JL
1277 if (dl_entity_is_special(dl_se))
1278 return;
1279
07881166
JL
1280 /*
1281 * For tasks that participate in GRUB, we implement GRUB-PA: the
1282 * spare reclaimed bandwidth is used to clock down frequency.
1283 *
1284 * For the others, we still need to scale reservation parameters
1285 * according to current frequency and CPU maximum capacity.
1286 */
1287 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1288 scaled_delta_exec = grub_reclaim(delta_exec,
1289 rq,
1290 &curr->dl);
1291 } else {
1292 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
8ec59c0f 1293 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
07881166
JL
1294
1295 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1296 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1297 }
1298
1299 dl_se->runtime -= scaled_delta_exec;
48be3a67
PZ
1300
1301throttle:
1302 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1019a359 1303 dl_se->dl_throttled = 1;
34be3930
JL
1304
1305 /* If requested, inform the user about runtime overruns. */
1306 if (dl_runtime_exceeded(dl_se) &&
1307 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1308 dl_se->dl_overrun = 1;
1309
aab03e05 1310 __dequeue_task_dl(rq, curr, 0);
2279f540 1311 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(curr)))
aab03e05
DF
1312 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1313
1314 if (!is_leftmost(curr, &rq->dl))
8875125e 1315 resched_curr(rq);
aab03e05 1316 }
1724813d
PZ
1317
1318 /*
1319 * Because -- for now -- we share the rt bandwidth, we need to
1320 * account our runtime there too, otherwise actual rt tasks
1321 * would be able to exceed the shared quota.
1322 *
1323 * Account to the root rt group for now.
1324 *
1325 * The solution we're working towards is having the RT groups scheduled
1326 * using deadline servers -- however there's a few nasties to figure
1327 * out before that can happen.
1328 */
1329 if (rt_bandwidth_enabled()) {
1330 struct rt_rq *rt_rq = &rq->rt;
1331
1332 raw_spin_lock(&rt_rq->rt_runtime_lock);
1724813d
PZ
1333 /*
1334 * We'll let actual RT tasks worry about the overflow here, we
faa59937
JL
1335 * have our own CBS to keep us inline; only account when RT
1336 * bandwidth is relevant.
1724813d 1337 */
faa59937
JL
1338 if (sched_rt_bandwidth_account(rt_rq))
1339 rt_rq->rt_time += delta_exec;
1724813d
PZ
1340 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1341 }
aab03e05
DF
1342}
1343
209a0cbd
LA
1344static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1345{
1346 struct sched_dl_entity *dl_se = container_of(timer,
1347 struct sched_dl_entity,
1348 inactive_timer);
1349 struct task_struct *p = dl_task_of(dl_se);
1350 struct rq_flags rf;
1351 struct rq *rq;
1352
1353 rq = task_rq_lock(p, &rf);
1354
ecda2b66
JL
1355 sched_clock_tick();
1356 update_rq_clock(rq);
1357
2f064a59 1358 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
387e3130
LA
1359 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1360
2f064a59 1361 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
794a56eb
JL
1362 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1363 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
209a0cbd
LA
1364 dl_se->dl_non_contending = 0;
1365 }
387e3130
LA
1366
1367 raw_spin_lock(&dl_b->lock);
8c0944ce 1368 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
387e3130 1369 raw_spin_unlock(&dl_b->lock);
209a0cbd
LA
1370 __dl_clear_params(p);
1371
1372 goto unlock;
1373 }
1374 if (dl_se->dl_non_contending == 0)
1375 goto unlock;
1376
794a56eb 1377 sub_running_bw(dl_se, &rq->dl);
209a0cbd
LA
1378 dl_se->dl_non_contending = 0;
1379unlock:
1380 task_rq_unlock(rq, p, &rf);
1381 put_task_struct(p);
1382
1383 return HRTIMER_NORESTART;
1384}
1385
1386void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1387{
1388 struct hrtimer *timer = &dl_se->inactive_timer;
1389
850377a8 1390 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
209a0cbd
LA
1391 timer->function = inactive_task_timer;
1392}
1393
1baca4ce
JL
1394#ifdef CONFIG_SMP
1395
1baca4ce
JL
1396static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1397{
1398 struct rq *rq = rq_of_dl_rq(dl_rq);
1399
1400 if (dl_rq->earliest_dl.curr == 0 ||
1401 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
b13772f8
PZ
1402 if (dl_rq->earliest_dl.curr == 0)
1403 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1baca4ce 1404 dl_rq->earliest_dl.curr = deadline;
d8206bb3 1405 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1baca4ce
JL
1406 }
1407}
1408
1409static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1410{
1411 struct rq *rq = rq_of_dl_rq(dl_rq);
1412
1413 /*
1414 * Since we may have removed our earliest (and/or next earliest)
1415 * task we must recompute them.
1416 */
1417 if (!dl_rq->dl_nr_running) {
1418 dl_rq->earliest_dl.curr = 0;
1419 dl_rq->earliest_dl.next = 0;
d8206bb3 1420 cpudl_clear(&rq->rd->cpudl, rq->cpu);
b13772f8 1421 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1baca4ce 1422 } else {
2161573e 1423 struct rb_node *leftmost = dl_rq->root.rb_leftmost;
1baca4ce
JL
1424 struct sched_dl_entity *entry;
1425
1426 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1427 dl_rq->earliest_dl.curr = entry->deadline;
d8206bb3 1428 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1baca4ce
JL
1429 }
1430}
1431
1432#else
1433
1434static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1435static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1436
1437#endif /* CONFIG_SMP */
1438
1439static inline
1440void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1441{
1442 int prio = dl_task_of(dl_se)->prio;
1443 u64 deadline = dl_se->deadline;
1444
1445 WARN_ON(!dl_prio(prio));
1446 dl_rq->dl_nr_running++;
72465447 1447 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1baca4ce
JL
1448
1449 inc_dl_deadline(dl_rq, deadline);
1450 inc_dl_migration(dl_se, dl_rq);
1451}
1452
1453static inline
1454void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1455{
1456 int prio = dl_task_of(dl_se)->prio;
1457
1458 WARN_ON(!dl_prio(prio));
1459 WARN_ON(!dl_rq->dl_nr_running);
1460 dl_rq->dl_nr_running--;
72465447 1461 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1baca4ce
JL
1462
1463 dec_dl_deadline(dl_rq, dl_se->deadline);
1464 dec_dl_migration(dl_se, dl_rq);
1465}
1466
8ecca394
PZ
1467#define __node_2_dle(node) \
1468 rb_entry((node), struct sched_dl_entity, rb_node)
1469
1470static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1471{
1472 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1473}
1474
aab03e05
DF
1475static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1476{
1477 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
aab03e05
DF
1478
1479 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1480
8ecca394 1481 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
aab03e05 1482
1baca4ce 1483 inc_dl_tasks(dl_se, dl_rq);
aab03e05
DF
1484}
1485
1486static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1487{
1488 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1489
1490 if (RB_EMPTY_NODE(&dl_se->rb_node))
1491 return;
1492
2161573e 1493 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
8ecca394 1494
aab03e05
DF
1495 RB_CLEAR_NODE(&dl_se->rb_node);
1496
1baca4ce 1497 dec_dl_tasks(dl_se, dl_rq);
aab03e05
DF
1498}
1499
1500static void
2279f540 1501enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
aab03e05
DF
1502{
1503 BUG_ON(on_dl_rq(dl_se));
1504
1505 /*
1506 * If this is a wakeup or a new instance, the scheduling
1507 * parameters of the task might need updating. Otherwise,
1508 * we want a replenishment of its runtime.
1509 */
e36d8677 1510 if (flags & ENQUEUE_WAKEUP) {
8fd27231 1511 task_contending(dl_se, flags);
2279f540 1512 update_dl_entity(dl_se);
e36d8677 1513 } else if (flags & ENQUEUE_REPLENISH) {
2279f540 1514 replenish_dl_entity(dl_se);
295d6d5e
LA
1515 } else if ((flags & ENQUEUE_RESTORE) &&
1516 dl_time_before(dl_se->deadline,
1517 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1518 setup_new_dl_entity(dl_se);
e36d8677 1519 }
aab03e05
DF
1520
1521 __enqueue_dl_entity(dl_se);
1522}
1523
1524static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1525{
1526 __dequeue_dl_entity(dl_se);
1527}
1528
1529static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1530{
2279f540 1531 if (is_dl_boosted(&p->dl)) {
feff2e65
DBO
1532 /*
1533 * Because of delays in the detection of the overrun of a
1534 * thread's runtime, it might be the case that a thread
1535 * goes to sleep in a rt mutex with negative runtime. As
1536 * a consequence, the thread will be throttled.
1537 *
1538 * While waiting for the mutex, this thread can also be
1539 * boosted via PI, resulting in a thread that is throttled
1540 * and boosted at the same time.
1541 *
1542 * In this case, the boost overrides the throttle.
1543 */
1544 if (p->dl.dl_throttled) {
1545 /*
1546 * The replenish timer needs to be canceled. No
1547 * problem if it fires concurrently: boosted threads
1548 * are ignored in dl_task_timer().
1549 */
1550 hrtimer_try_to_cancel(&p->dl.dl_timer);
1551 p->dl.dl_throttled = 0;
1552 }
64be6f1f
JL
1553 } else if (!dl_prio(p->normal_prio)) {
1554 /*
46fcc4b0
LS
1555 * Special case in which we have a !SCHED_DEADLINE task that is going
1556 * to be deboosted, but exceeds its runtime while doing so. No point in
1557 * replenishing it, as it's going to return back to its original
1558 * scheduling class after this. If it has been throttled, we need to
1559 * clear the flag, otherwise the task may wake up as throttled after
1560 * being boosted again with no means to replenish the runtime and clear
1561 * the throttle.
64be6f1f 1562 */
46fcc4b0 1563 p->dl.dl_throttled = 0;
2279f540 1564 BUG_ON(!is_dl_boosted(&p->dl) || flags != ENQUEUE_REPLENISH);
64be6f1f
JL
1565 return;
1566 }
2d3d891d 1567
df8eac8c
DBO
1568 /*
1569 * Check if a constrained deadline task was activated
1570 * after the deadline but before the next period.
1571 * If that is the case, the task will be throttled and
1572 * the replenishment timer will be set to the next period.
1573 */
3effcb42 1574 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
df8eac8c
DBO
1575 dl_check_constrained_dl(&p->dl);
1576
8fd27231 1577 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
794a56eb
JL
1578 add_rq_bw(&p->dl, &rq->dl);
1579 add_running_bw(&p->dl, &rq->dl);
8fd27231 1580 }
e36d8677 1581
aab03e05 1582 /*
e36d8677 1583 * If p is throttled, we do not enqueue it. In fact, if it exhausted
aab03e05
DF
1584 * its budget it needs a replenishment and, since it now is on
1585 * its rq, the bandwidth timer callback (which clearly has not
1586 * run yet) will take care of this.
e36d8677
LA
1587 * However, the active utilization does not depend on the fact
1588 * that the task is on the runqueue or not (but depends on the
1589 * task's state - in GRUB parlance, "inactive" vs "active contending").
1590 * In other words, even if a task is throttled its utilization must
1591 * be counted in the active utilization; hence, we need to call
1592 * add_running_bw().
aab03e05 1593 */
e36d8677 1594 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
209a0cbd 1595 if (flags & ENQUEUE_WAKEUP)
8fd27231 1596 task_contending(&p->dl, flags);
209a0cbd 1597
aab03e05 1598 return;
e36d8677 1599 }
aab03e05 1600
2279f540 1601 enqueue_dl_entity(&p->dl, flags);
1baca4ce 1602
4b53a341 1603 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1baca4ce 1604 enqueue_pushable_dl_task(rq, p);
aab03e05
DF
1605}
1606
1607static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1608{
1609 dequeue_dl_entity(&p->dl);
1baca4ce 1610 dequeue_pushable_dl_task(rq, p);
aab03e05
DF
1611}
1612
1613static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1614{
1615 update_curr_dl(rq);
1616 __dequeue_task_dl(rq, p, flags);
e36d8677 1617
8fd27231 1618 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
794a56eb
JL
1619 sub_running_bw(&p->dl, &rq->dl);
1620 sub_rq_bw(&p->dl, &rq->dl);
8fd27231 1621 }
e36d8677
LA
1622
1623 /*
209a0cbd
LA
1624 * This check allows to start the inactive timer (or to immediately
1625 * decrease the active utilization, if needed) in two cases:
e36d8677
LA
1626 * when the task blocks and when it is terminating
1627 * (p->state == TASK_DEAD). We can handle the two cases in the same
1628 * way, because from GRUB's point of view the same thing is happening
1629 * (the task moves from "active contending" to "active non contending"
1630 * or "inactive")
1631 */
1632 if (flags & DEQUEUE_SLEEP)
209a0cbd 1633 task_non_contending(p);
aab03e05
DF
1634}
1635
1636/*
1637 * Yield task semantic for -deadline tasks is:
1638 *
1639 * get off from the CPU until our next instance, with
1640 * a new runtime. This is of little use now, since we
1641 * don't have a bandwidth reclaiming mechanism. Anyway,
1642 * bandwidth reclaiming is planned for the future, and
1643 * yield_task_dl will indicate that some spare budget
1644 * is available for other task instances to use it.
1645 */
1646static void yield_task_dl(struct rq *rq)
1647{
aab03e05
DF
1648 /*
1649 * We make the task go to sleep until its current deadline by
1650 * forcing its runtime to zero. This way, update_curr_dl() stops
1651 * it and the bandwidth timer will wake it up and will give it
5bfd126e 1652 * new scheduling parameters (thanks to dl_yielded=1).
aab03e05 1653 */
48be3a67
PZ
1654 rq->curr->dl.dl_yielded = 1;
1655
6f1607f1 1656 update_rq_clock(rq);
aab03e05 1657 update_curr_dl(rq);
44fb085b
WL
1658 /*
1659 * Tell update_rq_clock() that we've just updated,
1660 * so we don't do microscopic update in schedule()
1661 * and double the fastpath cost.
1662 */
adcc8da8 1663 rq_clock_skip_update(rq);
aab03e05
DF
1664}
1665
1baca4ce
JL
1666#ifdef CONFIG_SMP
1667
1668static int find_later_rq(struct task_struct *task);
1baca4ce
JL
1669
1670static int
3aef1551 1671select_task_rq_dl(struct task_struct *p, int cpu, int flags)
1baca4ce
JL
1672{
1673 struct task_struct *curr;
b4118988 1674 bool select_rq;
1baca4ce
JL
1675 struct rq *rq;
1676
3aef1551 1677 if (!(flags & WF_TTWU))
1baca4ce
JL
1678 goto out;
1679
1680 rq = cpu_rq(cpu);
1681
1682 rcu_read_lock();
316c1608 1683 curr = READ_ONCE(rq->curr); /* unlocked access */
1baca4ce
JL
1684
1685 /*
1686 * If we are dealing with a -deadline task, we must
1687 * decide where to wake it up.
1688 * If it has a later deadline and the current task
1689 * on this rq can't move (provided the waking task
1690 * can!) we prefer to send it somewhere else. On the
1691 * other hand, if it has a shorter deadline, we
1692 * try to make it stay here, it might be important.
1693 */
b4118988
LA
1694 select_rq = unlikely(dl_task(curr)) &&
1695 (curr->nr_cpus_allowed < 2 ||
1696 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1697 p->nr_cpus_allowed > 1;
1698
1699 /*
1700 * Take the capacity of the CPU into account to
1701 * ensure it fits the requirement of the task.
1702 */
1703 if (static_branch_unlikely(&sched_asym_cpucapacity))
1704 select_rq |= !dl_task_fits_capacity(p, cpu);
1705
1706 if (select_rq) {
1baca4ce
JL
1707 int target = find_later_rq(p);
1708
9d514262 1709 if (target != -1 &&
5aa50507
LA
1710 (dl_time_before(p->dl.deadline,
1711 cpu_rq(target)->dl.earliest_dl.curr) ||
1712 (cpu_rq(target)->dl.dl_nr_running == 0)))
1baca4ce
JL
1713 cpu = target;
1714 }
1715 rcu_read_unlock();
1716
1717out:
1718 return cpu;
1719}
1720
1327237a 1721static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
209a0cbd
LA
1722{
1723 struct rq *rq;
1724
2f064a59 1725 if (READ_ONCE(p->__state) != TASK_WAKING)
209a0cbd
LA
1726 return;
1727
1728 rq = task_rq(p);
1729 /*
1730 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1731 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1732 * rq->lock is not... So, lock it
1733 */
5cb9eaa3 1734 raw_spin_rq_lock(rq);
8fd27231 1735 if (p->dl.dl_non_contending) {
794a56eb 1736 sub_running_bw(&p->dl, &rq->dl);
8fd27231
LA
1737 p->dl.dl_non_contending = 0;
1738 /*
1739 * If the timer handler is currently running and the
3b03706f 1740 * timer cannot be canceled, inactive_task_timer()
8fd27231
LA
1741 * will see that dl_not_contending is not set, and
1742 * will not touch the rq's active utilization,
1743 * so we are still safe.
1744 */
1745 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1746 put_task_struct(p);
1747 }
794a56eb 1748 sub_rq_bw(&p->dl, &rq->dl);
5cb9eaa3 1749 raw_spin_rq_unlock(rq);
209a0cbd
LA
1750}
1751
1baca4ce
JL
1752static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1753{
1754 /*
1755 * Current can't be migrated, useless to reschedule,
1756 * let's hope p can move out.
1757 */
4b53a341 1758 if (rq->curr->nr_cpus_allowed == 1 ||
3261ed0b 1759 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1baca4ce
JL
1760 return;
1761
1762 /*
1763 * p is migratable, so let's not schedule it and
1764 * see if it is pushed or pulled somewhere else.
1765 */
4b53a341 1766 if (p->nr_cpus_allowed != 1 &&
3261ed0b 1767 cpudl_find(&rq->rd->cpudl, p, NULL))
1baca4ce
JL
1768 return;
1769
8875125e 1770 resched_curr(rq);
1baca4ce
JL
1771}
1772
6e2df058
PZ
1773static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1774{
1775 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
1776 /*
1777 * This is OK, because current is on_cpu, which avoids it being
1778 * picked for load-balance and preemption/IRQs are still
1779 * disabled avoiding further scheduler activity on it and we've
1780 * not yet started the picking loop.
1781 */
1782 rq_unpin_lock(rq, rf);
1783 pull_dl_task(rq);
1784 rq_repin_lock(rq, rf);
1785 }
1786
1787 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
1788}
1baca4ce
JL
1789#endif /* CONFIG_SMP */
1790
aab03e05
DF
1791/*
1792 * Only called when both the current and waking task are -deadline
1793 * tasks.
1794 */
1795static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1796 int flags)
1797{
1baca4ce 1798 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
8875125e 1799 resched_curr(rq);
1baca4ce
JL
1800 return;
1801 }
1802
1803#ifdef CONFIG_SMP
1804 /*
1805 * In the unlikely case current and p have the same deadline
1806 * let us try to decide what's the best thing to do...
1807 */
332ac17e
DF
1808 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1809 !test_tsk_need_resched(rq->curr))
1baca4ce
JL
1810 check_preempt_equal_dl(rq, p);
1811#endif /* CONFIG_SMP */
aab03e05
DF
1812}
1813
1814#ifdef CONFIG_SCHED_HRTICK
1815static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1816{
177ef2a6 1817 hrtick_start(rq, p->dl.runtime);
aab03e05 1818}
36ce9881
WL
1819#else /* !CONFIG_SCHED_HRTICK */
1820static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1821{
1822}
aab03e05
DF
1823#endif
1824
a0e813f2 1825static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
ff1cdc94
MS
1826{
1827 p->se.exec_start = rq_clock_task(rq);
1828
1829 /* You can't push away the running task */
1830 dequeue_pushable_dl_task(rq, p);
f95d4eae 1831
a0e813f2
PZ
1832 if (!first)
1833 return;
1834
e0ee463c 1835 if (hrtick_enabled_dl(rq))
f95d4eae
PZ
1836 start_hrtick_dl(rq, p);
1837
1838 if (rq->curr->sched_class != &dl_sched_class)
1839 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1840
1841 deadline_queue_push_tasks(rq);
ff1cdc94
MS
1842}
1843
aab03e05
DF
1844static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1845 struct dl_rq *dl_rq)
1846{
2161573e 1847 struct rb_node *left = rb_first_cached(&dl_rq->root);
aab03e05
DF
1848
1849 if (!left)
1850 return NULL;
1851
1852 return rb_entry(left, struct sched_dl_entity, rb_node);
1853}
1854
21f56ffe 1855static struct task_struct *pick_task_dl(struct rq *rq)
aab03e05
DF
1856{
1857 struct sched_dl_entity *dl_se;
6e2df058 1858 struct dl_rq *dl_rq = &rq->dl;
aab03e05 1859 struct task_struct *p;
aab03e05 1860
6e2df058 1861 if (!sched_dl_runnable(rq))
aab03e05
DF
1862 return NULL;
1863
1864 dl_se = pick_next_dl_entity(rq, dl_rq);
1865 BUG_ON(!dl_se);
aab03e05 1866 p = dl_task_of(dl_se);
21f56ffe
PZ
1867
1868 return p;
1869}
1870
1871static struct task_struct *pick_next_task_dl(struct rq *rq)
1872{
1873 struct task_struct *p;
1874
1875 p = pick_task_dl(rq);
1876 if (p)
1877 set_next_task_dl(rq, p, true);
1878
aab03e05
DF
1879 return p;
1880}
1881
6e2df058 1882static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
aab03e05
DF
1883{
1884 update_curr_dl(rq);
1baca4ce 1885
23127296 1886 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
4b53a341 1887 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1baca4ce 1888 enqueue_pushable_dl_task(rq, p);
aab03e05
DF
1889}
1890
d84b3131
FW
1891/*
1892 * scheduler tick hitting a task of our scheduling class.
1893 *
1894 * NOTE: This function can be called remotely by the tick offload that
1895 * goes along full dynticks. Therefore no local assumption can be made
1896 * and everything must be accessed through the @rq and @curr passed in
1897 * parameters.
1898 */
aab03e05
DF
1899static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1900{
1901 update_curr_dl(rq);
1902
23127296 1903 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
a7bebf48
WL
1904 /*
1905 * Even when we have runtime, update_curr_dl() might have resulted in us
1906 * not being the leftmost task anymore. In that case NEED_RESCHED will
1907 * be set and schedule() will start a new hrtick for the next task.
1908 */
e0ee463c 1909 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
a7bebf48 1910 is_leftmost(p, &rq->dl))
aab03e05 1911 start_hrtick_dl(rq, p);
aab03e05
DF
1912}
1913
1914static void task_fork_dl(struct task_struct *p)
1915{
1916 /*
1917 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1918 * sched_fork()
1919 */
1920}
1921
1baca4ce
JL
1922#ifdef CONFIG_SMP
1923
1924/* Only try algorithms three times */
1925#define DL_MAX_TRIES 3
1926
1927static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1928{
1929 if (!task_running(rq, p) &&
95158a89 1930 cpumask_test_cpu(cpu, &p->cpus_mask))
1baca4ce 1931 return 1;
1baca4ce
JL
1932 return 0;
1933}
1934
8b5e770e
WL
1935/*
1936 * Return the earliest pushable rq's task, which is suitable to be executed
1937 * on the CPU, NULL otherwise:
1938 */
1939static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1940{
2161573e 1941 struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost;
8b5e770e
WL
1942 struct task_struct *p = NULL;
1943
1944 if (!has_pushable_dl_tasks(rq))
1945 return NULL;
1946
1947next_node:
1948 if (next_node) {
1949 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1950
1951 if (pick_dl_task(rq, p, cpu))
1952 return p;
1953
1954 next_node = rb_next(next_node);
1955 goto next_node;
1956 }
1957
1958 return NULL;
1959}
1960
1baca4ce
JL
1961static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1962
1963static int find_later_rq(struct task_struct *task)
1964{
1965 struct sched_domain *sd;
4ba29684 1966 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1baca4ce 1967 int this_cpu = smp_processor_id();
b18c3ca1 1968 int cpu = task_cpu(task);
1baca4ce
JL
1969
1970 /* Make sure the mask is initialized first */
1971 if (unlikely(!later_mask))
1972 return -1;
1973
4b53a341 1974 if (task->nr_cpus_allowed == 1)
1baca4ce
JL
1975 return -1;
1976
91ec6778
JL
1977 /*
1978 * We have to consider system topology and task affinity
97fb7a0a 1979 * first, then we can look for a suitable CPU.
91ec6778 1980 */
3261ed0b 1981 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
1baca4ce
JL
1982 return -1;
1983
1984 /*
b18c3ca1
BP
1985 * If we are here, some targets have been found, including
1986 * the most suitable which is, among the runqueues where the
1987 * current tasks have later deadlines than the task's one, the
1988 * rq with the latest possible one.
1baca4ce
JL
1989 *
1990 * Now we check how well this matches with task's
1991 * affinity and system topology.
1992 *
97fb7a0a 1993 * The last CPU where the task run is our first
1baca4ce
JL
1994 * guess, since it is most likely cache-hot there.
1995 */
1996 if (cpumask_test_cpu(cpu, later_mask))
1997 return cpu;
1998 /*
1999 * Check if this_cpu is to be skipped (i.e., it is
2000 * not in the mask) or not.
2001 */
2002 if (!cpumask_test_cpu(this_cpu, later_mask))
2003 this_cpu = -1;
2004
2005 rcu_read_lock();
2006 for_each_domain(cpu, sd) {
2007 if (sd->flags & SD_WAKE_AFFINE) {
b18c3ca1 2008 int best_cpu;
1baca4ce
JL
2009
2010 /*
2011 * If possible, preempting this_cpu is
2012 * cheaper than migrating.
2013 */
2014 if (this_cpu != -1 &&
2015 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2016 rcu_read_unlock();
2017 return this_cpu;
2018 }
2019
14e292f8
PZ
2020 best_cpu = cpumask_any_and_distribute(later_mask,
2021 sched_domain_span(sd));
1baca4ce 2022 /*
97fb7a0a 2023 * Last chance: if a CPU being in both later_mask
b18c3ca1 2024 * and current sd span is valid, that becomes our
97fb7a0a 2025 * choice. Of course, the latest possible CPU is
b18c3ca1 2026 * already under consideration through later_mask.
1baca4ce 2027 */
b18c3ca1 2028 if (best_cpu < nr_cpu_ids) {
1baca4ce
JL
2029 rcu_read_unlock();
2030 return best_cpu;
2031 }
2032 }
2033 }
2034 rcu_read_unlock();
2035
2036 /*
2037 * At this point, all our guesses failed, we just return
2038 * 'something', and let the caller sort the things out.
2039 */
2040 if (this_cpu != -1)
2041 return this_cpu;
2042
14e292f8 2043 cpu = cpumask_any_distribute(later_mask);
1baca4ce
JL
2044 if (cpu < nr_cpu_ids)
2045 return cpu;
2046
2047 return -1;
2048}
2049
2050/* Locks the rq it finds */
2051static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2052{
2053 struct rq *later_rq = NULL;
2054 int tries;
2055 int cpu;
2056
2057 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2058 cpu = find_later_rq(task);
2059
2060 if ((cpu == -1) || (cpu == rq->cpu))
2061 break;
2062
2063 later_rq = cpu_rq(cpu);
2064
5aa50507
LA
2065 if (later_rq->dl.dl_nr_running &&
2066 !dl_time_before(task->dl.deadline,
9d514262
WL
2067 later_rq->dl.earliest_dl.curr)) {
2068 /*
2069 * Target rq has tasks of equal or earlier deadline,
2070 * retrying does not release any lock and is unlikely
2071 * to yield a different result.
2072 */
2073 later_rq = NULL;
2074 break;
2075 }
2076
1baca4ce
JL
2077 /* Retry if something changed. */
2078 if (double_lock_balance(rq, later_rq)) {
2079 if (unlikely(task_rq(task) != rq ||
95158a89 2080 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
da0c1e65 2081 task_running(rq, task) ||
13b5ab02 2082 !dl_task(task) ||
da0c1e65 2083 !task_on_rq_queued(task))) {
1baca4ce
JL
2084 double_unlock_balance(rq, later_rq);
2085 later_rq = NULL;
2086 break;
2087 }
2088 }
2089
2090 /*
2091 * If the rq we found has no -deadline task, or
2092 * its earliest one has a later deadline than our
2093 * task, the rq is a good one.
2094 */
2095 if (!later_rq->dl.dl_nr_running ||
2096 dl_time_before(task->dl.deadline,
2097 later_rq->dl.earliest_dl.curr))
2098 break;
2099
2100 /* Otherwise we try again. */
2101 double_unlock_balance(rq, later_rq);
2102 later_rq = NULL;
2103 }
2104
2105 return later_rq;
2106}
2107
2108static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2109{
2110 struct task_struct *p;
2111
2112 if (!has_pushable_dl_tasks(rq))
2113 return NULL;
2114
2161573e 2115 p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost,
1baca4ce
JL
2116 struct task_struct, pushable_dl_tasks);
2117
2118 BUG_ON(rq->cpu != task_cpu(p));
2119 BUG_ON(task_current(rq, p));
4b53a341 2120 BUG_ON(p->nr_cpus_allowed <= 1);
1baca4ce 2121
da0c1e65 2122 BUG_ON(!task_on_rq_queued(p));
1baca4ce
JL
2123 BUG_ON(!dl_task(p));
2124
2125 return p;
2126}
2127
2128/*
2129 * See if the non running -deadline tasks on this rq
2130 * can be sent to some other CPU where they can preempt
2131 * and start executing.
2132 */
2133static int push_dl_task(struct rq *rq)
2134{
2135 struct task_struct *next_task;
2136 struct rq *later_rq;
c51b8ab5 2137 int ret = 0;
1baca4ce
JL
2138
2139 if (!rq->dl.overloaded)
2140 return 0;
2141
2142 next_task = pick_next_pushable_dl_task(rq);
2143 if (!next_task)
2144 return 0;
2145
2146retry:
a7c81556
PZ
2147 if (is_migration_disabled(next_task))
2148 return 0;
2149
9ebc6053 2150 if (WARN_ON(next_task == rq->curr))
1baca4ce 2151 return 0;
1baca4ce
JL
2152
2153 /*
2154 * If next_task preempts rq->curr, and rq->curr
2155 * can move away, it makes sense to just reschedule
2156 * without going further in pushing next_task.
2157 */
2158 if (dl_task(rq->curr) &&
2159 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
4b53a341 2160 rq->curr->nr_cpus_allowed > 1) {
8875125e 2161 resched_curr(rq);
1baca4ce
JL
2162 return 0;
2163 }
2164
2165 /* We might release rq lock */
2166 get_task_struct(next_task);
2167
2168 /* Will lock the rq it'll find */
2169 later_rq = find_lock_later_rq(next_task, rq);
2170 if (!later_rq) {
2171 struct task_struct *task;
2172
2173 /*
2174 * We must check all this again, since
2175 * find_lock_later_rq releases rq->lock and it is
2176 * then possible that next_task has migrated.
2177 */
2178 task = pick_next_pushable_dl_task(rq);
a776b968 2179 if (task == next_task) {
1baca4ce
JL
2180 /*
2181 * The task is still there. We don't try
97fb7a0a 2182 * again, some other CPU will pull it when ready.
1baca4ce 2183 */
1baca4ce
JL
2184 goto out;
2185 }
2186
2187 if (!task)
2188 /* No more tasks */
2189 goto out;
2190
2191 put_task_struct(next_task);
2192 next_task = task;
2193 goto retry;
2194 }
2195
2196 deactivate_task(rq, next_task, 0);
2197 set_task_cpu(next_task, later_rq->cpu);
840d7196
DBO
2198
2199 /*
2200 * Update the later_rq clock here, because the clock is used
2201 * by the cpufreq_update_util() inside __add_running_bw().
2202 */
2203 update_rq_clock(later_rq);
840d7196 2204 activate_task(later_rq, next_task, ENQUEUE_NOCLOCK);
c51b8ab5 2205 ret = 1;
1baca4ce 2206
8875125e 2207 resched_curr(later_rq);
1baca4ce
JL
2208
2209 double_unlock_balance(rq, later_rq);
2210
2211out:
2212 put_task_struct(next_task);
2213
c51b8ab5 2214 return ret;
1baca4ce
JL
2215}
2216
2217static void push_dl_tasks(struct rq *rq)
2218{
4ffa08ed 2219 /* push_dl_task() will return true if it moved a -deadline task */
1baca4ce
JL
2220 while (push_dl_task(rq))
2221 ;
aab03e05
DF
2222}
2223
0ea60c20 2224static void pull_dl_task(struct rq *this_rq)
1baca4ce 2225{
0ea60c20 2226 int this_cpu = this_rq->cpu, cpu;
a7c81556 2227 struct task_struct *p, *push_task;
0ea60c20 2228 bool resched = false;
1baca4ce
JL
2229 struct rq *src_rq;
2230 u64 dmin = LONG_MAX;
2231
2232 if (likely(!dl_overloaded(this_rq)))
0ea60c20 2233 return;
1baca4ce
JL
2234
2235 /*
2236 * Match the barrier from dl_set_overloaded; this guarantees that if we
2237 * see overloaded we must also see the dlo_mask bit.
2238 */
2239 smp_rmb();
2240
2241 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2242 if (this_cpu == cpu)
2243 continue;
2244
2245 src_rq = cpu_rq(cpu);
2246
2247 /*
2248 * It looks racy, abd it is! However, as in sched_rt.c,
2249 * we are fine with this.
2250 */
2251 if (this_rq->dl.dl_nr_running &&
2252 dl_time_before(this_rq->dl.earliest_dl.curr,
2253 src_rq->dl.earliest_dl.next))
2254 continue;
2255
2256 /* Might drop this_rq->lock */
a7c81556 2257 push_task = NULL;
1baca4ce
JL
2258 double_lock_balance(this_rq, src_rq);
2259
2260 /*
2261 * If there are no more pullable tasks on the
2262 * rq, we're done with it.
2263 */
2264 if (src_rq->dl.dl_nr_running <= 1)
2265 goto skip;
2266
8b5e770e 2267 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
1baca4ce
JL
2268
2269 /*
2270 * We found a task to be pulled if:
2271 * - it preempts our current (if there's one),
2272 * - it will preempt the last one we pulled (if any).
2273 */
2274 if (p && dl_time_before(p->dl.deadline, dmin) &&
2275 (!this_rq->dl.dl_nr_running ||
2276 dl_time_before(p->dl.deadline,
2277 this_rq->dl.earliest_dl.curr))) {
2278 WARN_ON(p == src_rq->curr);
da0c1e65 2279 WARN_ON(!task_on_rq_queued(p));
1baca4ce
JL
2280
2281 /*
2282 * Then we pull iff p has actually an earlier
2283 * deadline than the current task of its runqueue.
2284 */
2285 if (dl_time_before(p->dl.deadline,
2286 src_rq->curr->dl.deadline))
2287 goto skip;
2288
a7c81556
PZ
2289 if (is_migration_disabled(p)) {
2290 push_task = get_push_task(src_rq);
2291 } else {
2292 deactivate_task(src_rq, p, 0);
2293 set_task_cpu(p, this_cpu);
2294 activate_task(this_rq, p, 0);
2295 dmin = p->dl.deadline;
2296 resched = true;
2297 }
1baca4ce
JL
2298
2299 /* Is there any other task even earlier? */
2300 }
2301skip:
2302 double_unlock_balance(this_rq, src_rq);
a7c81556
PZ
2303
2304 if (push_task) {
5cb9eaa3 2305 raw_spin_rq_unlock(this_rq);
a7c81556
PZ
2306 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2307 push_task, &src_rq->push_work);
5cb9eaa3 2308 raw_spin_rq_lock(this_rq);
a7c81556 2309 }
1baca4ce
JL
2310 }
2311
0ea60c20
PZ
2312 if (resched)
2313 resched_curr(this_rq);
1baca4ce
JL
2314}
2315
2316/*
2317 * Since the task is not running and a reschedule is not going to happen
2318 * anytime soon on its runqueue, we try pushing it away now.
2319 */
2320static void task_woken_dl(struct rq *rq, struct task_struct *p)
2321{
2322 if (!task_running(rq, p) &&
2323 !test_tsk_need_resched(rq->curr) &&
4b53a341 2324 p->nr_cpus_allowed > 1 &&
1baca4ce 2325 dl_task(rq->curr) &&
4b53a341 2326 (rq->curr->nr_cpus_allowed < 2 ||
6b0a563f 2327 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
1baca4ce
JL
2328 push_dl_tasks(rq);
2329 }
2330}
2331
2332static void set_cpus_allowed_dl(struct task_struct *p,
9cfc3e18
PZ
2333 const struct cpumask *new_mask,
2334 u32 flags)
1baca4ce 2335{
7f51412a 2336 struct root_domain *src_rd;
6c37067e 2337 struct rq *rq;
1baca4ce
JL
2338
2339 BUG_ON(!dl_task(p));
2340
7f51412a
JL
2341 rq = task_rq(p);
2342 src_rd = rq->rd;
2343 /*
2344 * Migrating a SCHED_DEADLINE task between exclusive
2345 * cpusets (different root_domains) entails a bandwidth
2346 * update. We already made space for us in the destination
2347 * domain (see cpuset_can_attach()).
2348 */
2349 if (!cpumask_intersects(src_rd->span, new_mask)) {
2350 struct dl_bw *src_dl_b;
2351
2352 src_dl_b = dl_bw_of(cpu_of(rq));
2353 /*
2354 * We now free resources of the root_domain we are migrating
2355 * off. In the worst case, sched_setattr() may temporary fail
2356 * until we complete the update.
2357 */
2358 raw_spin_lock(&src_dl_b->lock);
8c0944ce 2359 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
7f51412a
JL
2360 raw_spin_unlock(&src_dl_b->lock);
2361 }
2362
9cfc3e18 2363 set_cpus_allowed_common(p, new_mask, flags);
1baca4ce
JL
2364}
2365
2366/* Assumes rq->lock is held */
2367static void rq_online_dl(struct rq *rq)
2368{
2369 if (rq->dl.overloaded)
2370 dl_set_overload(rq);
6bfd6d72 2371
16b26943 2372 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
6bfd6d72 2373 if (rq->dl.dl_nr_running > 0)
d8206bb3 2374 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
1baca4ce
JL
2375}
2376
2377/* Assumes rq->lock is held */
2378static void rq_offline_dl(struct rq *rq)
2379{
2380 if (rq->dl.overloaded)
2381 dl_clear_overload(rq);
6bfd6d72 2382
d8206bb3 2383 cpudl_clear(&rq->rd->cpudl, rq->cpu);
16b26943 2384 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
1baca4ce
JL
2385}
2386
a6c0e746 2387void __init init_sched_dl_class(void)
1baca4ce
JL
2388{
2389 unsigned int i;
2390
2391 for_each_possible_cpu(i)
2392 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2393 GFP_KERNEL, cpu_to_node(i));
2394}
2395
f9a25f77
MP
2396void dl_add_task_root_domain(struct task_struct *p)
2397{
2398 struct rq_flags rf;
2399 struct rq *rq;
2400 struct dl_bw *dl_b;
2401
de40f33e
DE
2402 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2403 if (!dl_task(p)) {
2404 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2405 return;
2406 }
2407
2408 rq = __task_rq_lock(p, &rf);
f9a25f77
MP
2409
2410 dl_b = &rq->rd->dl_bw;
2411 raw_spin_lock(&dl_b->lock);
2412
2413 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2414
2415 raw_spin_unlock(&dl_b->lock);
2416
f9a25f77
MP
2417 task_rq_unlock(rq, p, &rf);
2418}
2419
2420void dl_clear_root_domain(struct root_domain *rd)
2421{
2422 unsigned long flags;
2423
2424 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2425 rd->dl_bw.total_bw = 0;
2426 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2427}
2428
1baca4ce
JL
2429#endif /* CONFIG_SMP */
2430
aab03e05
DF
2431static void switched_from_dl(struct rq *rq, struct task_struct *p)
2432{
a649f237 2433 /*
209a0cbd
LA
2434 * task_non_contending() can start the "inactive timer" (if the 0-lag
2435 * time is in the future). If the task switches back to dl before
2436 * the "inactive timer" fires, it can continue to consume its current
2437 * runtime using its current deadline. If it stays outside of
2438 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2439 * will reset the task parameters.
a649f237 2440 */
209a0cbd
LA
2441 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2442 task_non_contending(p);
2443
e117cb52
JL
2444 if (!task_on_rq_queued(p)) {
2445 /*
2446 * Inactive timer is armed. However, p is leaving DEADLINE and
2447 * might migrate away from this rq while continuing to run on
2448 * some other class. We need to remove its contribution from
2449 * this rq running_bw now, or sub_rq_bw (below) will complain.
2450 */
2451 if (p->dl.dl_non_contending)
2452 sub_running_bw(&p->dl, &rq->dl);
794a56eb 2453 sub_rq_bw(&p->dl, &rq->dl);
e117cb52 2454 }
8fd27231 2455
209a0cbd
LA
2456 /*
2457 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2458 * at the 0-lag time, because the task could have been migrated
2459 * while SCHED_OTHER in the meanwhile.
2460 */
2461 if (p->dl.dl_non_contending)
2462 p->dl.dl_non_contending = 0;
a5e7be3b 2463
1baca4ce
JL
2464 /*
2465 * Since this might be the only -deadline task on the rq,
2466 * this is the right place to try to pull some other one
97fb7a0a 2467 * from an overloaded CPU, if any.
1baca4ce 2468 */
cd660911
WL
2469 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2470 return;
2471
02d8ec94 2472 deadline_queue_pull_task(rq);
aab03e05
DF
2473}
2474
1baca4ce
JL
2475/*
2476 * When switching to -deadline, we may overload the rq, then
2477 * we try to push someone off, if possible.
2478 */
aab03e05
DF
2479static void switched_to_dl(struct rq *rq, struct task_struct *p)
2480{
209a0cbd
LA
2481 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2482 put_task_struct(p);
98b0a857
JL
2483
2484 /* If p is not queued we will update its parameters at next wakeup. */
8fd27231 2485 if (!task_on_rq_queued(p)) {
794a56eb 2486 add_rq_bw(&p->dl, &rq->dl);
98b0a857 2487
8fd27231
LA
2488 return;
2489 }
72f9f3fd 2490
98b0a857 2491 if (rq->curr != p) {
1baca4ce 2492#ifdef CONFIG_SMP
4b53a341 2493 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
02d8ec94 2494 deadline_queue_push_tasks(rq);
619bd4a7 2495#endif
9916e214
PZ
2496 if (dl_task(rq->curr))
2497 check_preempt_curr_dl(rq, p, 0);
2498 else
2499 resched_curr(rq);
d7d60709
VD
2500 } else {
2501 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
aab03e05
DF
2502 }
2503}
2504
1baca4ce
JL
2505/*
2506 * If the scheduling parameters of a -deadline task changed,
2507 * a push or pull operation might be needed.
2508 */
aab03e05
DF
2509static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2510 int oldprio)
2511{
65bcf072 2512 if (task_on_rq_queued(p) || task_current(rq, p)) {
aab03e05 2513#ifdef CONFIG_SMP
1baca4ce
JL
2514 /*
2515 * This might be too much, but unfortunately
2516 * we don't have the old deadline value, and
2517 * we can't argue if the task is increasing
2518 * or lowering its prio, so...
2519 */
2520 if (!rq->dl.overloaded)
02d8ec94 2521 deadline_queue_pull_task(rq);
1baca4ce
JL
2522
2523 /*
2524 * If we now have a earlier deadline task than p,
2525 * then reschedule, provided p is still on this
2526 * runqueue.
2527 */
9916e214 2528 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
8875125e 2529 resched_curr(rq);
1baca4ce
JL
2530#else
2531 /*
2532 * Again, we don't know if p has a earlier
2533 * or later deadline, so let's blindly set a
2534 * (maybe not needed) rescheduling point.
2535 */
8875125e 2536 resched_curr(rq);
1baca4ce 2537#endif /* CONFIG_SMP */
801ccdbf 2538 }
aab03e05 2539}
aab03e05 2540
43c31ac0
PZ
2541DEFINE_SCHED_CLASS(dl) = {
2542
aab03e05
DF
2543 .enqueue_task = enqueue_task_dl,
2544 .dequeue_task = dequeue_task_dl,
2545 .yield_task = yield_task_dl,
2546
2547 .check_preempt_curr = check_preempt_curr_dl,
2548
2549 .pick_next_task = pick_next_task_dl,
2550 .put_prev_task = put_prev_task_dl,
03b7fad1 2551 .set_next_task = set_next_task_dl,
aab03e05
DF
2552
2553#ifdef CONFIG_SMP
6e2df058 2554 .balance = balance_dl,
21f56ffe 2555 .pick_task = pick_task_dl,
aab03e05 2556 .select_task_rq = select_task_rq_dl,
209a0cbd 2557 .migrate_task_rq = migrate_task_rq_dl,
1baca4ce
JL
2558 .set_cpus_allowed = set_cpus_allowed_dl,
2559 .rq_online = rq_online_dl,
2560 .rq_offline = rq_offline_dl,
1baca4ce 2561 .task_woken = task_woken_dl,
a7c81556 2562 .find_lock_rq = find_lock_later_rq,
aab03e05
DF
2563#endif
2564
aab03e05
DF
2565 .task_tick = task_tick_dl,
2566 .task_fork = task_fork_dl,
aab03e05
DF
2567
2568 .prio_changed = prio_changed_dl,
2569 .switched_from = switched_from_dl,
2570 .switched_to = switched_to_dl,
6e998916
SG
2571
2572 .update_curr = update_curr_dl,
aab03e05 2573};
acb32132 2574
26762423
PL
2575/* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
2576static u64 dl_generation;
2577
06a76fe0
NP
2578int sched_dl_global_validate(void)
2579{
2580 u64 runtime = global_rt_runtime();
2581 u64 period = global_rt_period();
2582 u64 new_bw = to_ratio(period, runtime);
26762423 2583 u64 gen = ++dl_generation;
06a76fe0 2584 struct dl_bw *dl_b;
a57415f5 2585 int cpu, cpus, ret = 0;
06a76fe0
NP
2586 unsigned long flags;
2587
2588 /*
2589 * Here we want to check the bandwidth not being set to some
2590 * value smaller than the currently allocated bandwidth in
2591 * any of the root_domains.
06a76fe0
NP
2592 */
2593 for_each_possible_cpu(cpu) {
2594 rcu_read_lock_sched();
26762423
PL
2595
2596 if (dl_bw_visited(cpu, gen))
2597 goto next;
2598
06a76fe0 2599 dl_b = dl_bw_of(cpu);
a57415f5 2600 cpus = dl_bw_cpus(cpu);
06a76fe0
NP
2601
2602 raw_spin_lock_irqsave(&dl_b->lock, flags);
a57415f5 2603 if (new_bw * cpus < dl_b->total_bw)
06a76fe0
NP
2604 ret = -EBUSY;
2605 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2606
26762423 2607next:
06a76fe0
NP
2608 rcu_read_unlock_sched();
2609
2610 if (ret)
2611 break;
2612 }
2613
2614 return ret;
2615}
2616
ba4f7bc1 2617static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
06a76fe0
NP
2618{
2619 if (global_rt_runtime() == RUNTIME_INF) {
2620 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2621 dl_rq->extra_bw = 1 << BW_SHIFT;
2622 } else {
2623 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2624 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2625 dl_rq->extra_bw = to_ratio(global_rt_period(),
2626 global_rt_runtime());
2627 }
2628}
2629
2630void sched_dl_do_global(void)
2631{
2632 u64 new_bw = -1;
26762423 2633 u64 gen = ++dl_generation;
06a76fe0
NP
2634 struct dl_bw *dl_b;
2635 int cpu;
2636 unsigned long flags;
2637
2638 def_dl_bandwidth.dl_period = global_rt_period();
2639 def_dl_bandwidth.dl_runtime = global_rt_runtime();
2640
2641 if (global_rt_runtime() != RUNTIME_INF)
2642 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2643
06a76fe0
NP
2644 for_each_possible_cpu(cpu) {
2645 rcu_read_lock_sched();
26762423
PL
2646
2647 if (dl_bw_visited(cpu, gen)) {
2648 rcu_read_unlock_sched();
2649 continue;
2650 }
2651
06a76fe0
NP
2652 dl_b = dl_bw_of(cpu);
2653
2654 raw_spin_lock_irqsave(&dl_b->lock, flags);
2655 dl_b->bw = new_bw;
2656 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2657
2658 rcu_read_unlock_sched();
2659 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2660 }
2661}
2662
2663/*
2664 * We must be sure that accepting a new task (or allowing changing the
2665 * parameters of an existing one) is consistent with the bandwidth
2666 * constraints. If yes, this function also accordingly updates the currently
2667 * allocated bandwidth to reflect the new situation.
2668 *
2669 * This function is called while holding p's rq->lock.
2670 */
2671int sched_dl_overflow(struct task_struct *p, int policy,
2672 const struct sched_attr *attr)
2673{
06a76fe0
NP
2674 u64 period = attr->sched_period ?: attr->sched_deadline;
2675 u64 runtime = attr->sched_runtime;
2676 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
60ffd5ed
LA
2677 int cpus, err = -1, cpu = task_cpu(p);
2678 struct dl_bw *dl_b = dl_bw_of(cpu);
2679 unsigned long cap;
06a76fe0 2680
794a56eb
JL
2681 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2682 return 0;
2683
06a76fe0
NP
2684 /* !deadline task may carry old deadline bandwidth */
2685 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2686 return 0;
2687
2688 /*
2689 * Either if a task, enters, leave, or stays -deadline but changes
2690 * its parameters, we may need to update accordingly the total
2691 * allocated bandwidth of the container.
2692 */
2693 raw_spin_lock(&dl_b->lock);
60ffd5ed
LA
2694 cpus = dl_bw_cpus(cpu);
2695 cap = dl_bw_capacity(cpu);
2696
06a76fe0 2697 if (dl_policy(policy) && !task_has_dl_policy(p) &&
60ffd5ed 2698 !__dl_overflow(dl_b, cap, 0, new_bw)) {
06a76fe0 2699 if (hrtimer_active(&p->dl.inactive_timer))
8c0944ce 2700 __dl_sub(dl_b, p->dl.dl_bw, cpus);
06a76fe0
NP
2701 __dl_add(dl_b, new_bw, cpus);
2702 err = 0;
2703 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
60ffd5ed 2704 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
06a76fe0
NP
2705 /*
2706 * XXX this is slightly incorrect: when the task
2707 * utilization decreases, we should delay the total
2708 * utilization change until the task's 0-lag point.
2709 * But this would require to set the task's "inactive
2710 * timer" when the task is not inactive.
2711 */
8c0944ce 2712 __dl_sub(dl_b, p->dl.dl_bw, cpus);
06a76fe0
NP
2713 __dl_add(dl_b, new_bw, cpus);
2714 dl_change_utilization(p, new_bw);
2715 err = 0;
2716 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2717 /*
2718 * Do not decrease the total deadline utilization here,
2719 * switched_from_dl() will take care to do it at the correct
2720 * (0-lag) time.
2721 */
2722 err = 0;
2723 }
2724 raw_spin_unlock(&dl_b->lock);
2725
2726 return err;
2727}
2728
2729/*
2730 * This function initializes the sched_dl_entity of a newly becoming
2731 * SCHED_DEADLINE task.
2732 *
2733 * Only the static values are considered here, the actual runtime and the
2734 * absolute deadline will be properly calculated when the task is enqueued
2735 * for the first time with its new policy.
2736 */
2737void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2738{
2739 struct sched_dl_entity *dl_se = &p->dl;
2740
2741 dl_se->dl_runtime = attr->sched_runtime;
2742 dl_se->dl_deadline = attr->sched_deadline;
2743 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2744 dl_se->flags = attr->sched_flags;
2745 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2746 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2747}
2748
2749void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2750{
2751 struct sched_dl_entity *dl_se = &p->dl;
2752
2753 attr->sched_priority = p->rt_priority;
2754 attr->sched_runtime = dl_se->dl_runtime;
2755 attr->sched_deadline = dl_se->dl_deadline;
2756 attr->sched_period = dl_se->dl_period;
2757 attr->sched_flags = dl_se->flags;
2758}
2759
b4098bfc
PZ
2760/*
2761 * Default limits for DL period; on the top end we guard against small util
3b03706f 2762 * tasks still getting ridiculously long effective runtimes, on the bottom end we
b4098bfc
PZ
2763 * guard against timer DoS.
2764 */
2765unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
2766unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
2767
06a76fe0
NP
2768/*
2769 * This function validates the new parameters of a -deadline task.
2770 * We ask for the deadline not being zero, and greater or equal
2771 * than the runtime, as well as the period of being zero or
2772 * greater than deadline. Furthermore, we have to be sure that
2773 * user parameters are above the internal resolution of 1us (we
2774 * check sched_runtime only since it is always the smaller one) and
2775 * below 2^63 ns (we have to check both sched_deadline and
2776 * sched_period, as the latter can be zero).
2777 */
2778bool __checkparam_dl(const struct sched_attr *attr)
2779{
b4098bfc
PZ
2780 u64 period, max, min;
2781
794a56eb
JL
2782 /* special dl tasks don't actually use any parameter */
2783 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2784 return true;
2785
06a76fe0
NP
2786 /* deadline != 0 */
2787 if (attr->sched_deadline == 0)
2788 return false;
2789
2790 /*
2791 * Since we truncate DL_SCALE bits, make sure we're at least
2792 * that big.
2793 */
2794 if (attr->sched_runtime < (1ULL << DL_SCALE))
2795 return false;
2796
2797 /*
2798 * Since we use the MSB for wrap-around and sign issues, make
2799 * sure it's not set (mind that period can be equal to zero).
2800 */
2801 if (attr->sched_deadline & (1ULL << 63) ||
2802 attr->sched_period & (1ULL << 63))
2803 return false;
2804
b4098bfc
PZ
2805 period = attr->sched_period;
2806 if (!period)
2807 period = attr->sched_deadline;
2808
06a76fe0 2809 /* runtime <= deadline <= period (if period != 0) */
b4098bfc 2810 if (period < attr->sched_deadline ||
06a76fe0
NP
2811 attr->sched_deadline < attr->sched_runtime)
2812 return false;
2813
b4098bfc
PZ
2814 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
2815 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
2816
2817 if (period < min || period > max)
2818 return false;
2819
06a76fe0
NP
2820 return true;
2821}
2822
2823/*
2824 * This function clears the sched_dl_entity static params.
2825 */
2826void __dl_clear_params(struct task_struct *p)
2827{
2828 struct sched_dl_entity *dl_se = &p->dl;
2829
97fb7a0a
IM
2830 dl_se->dl_runtime = 0;
2831 dl_se->dl_deadline = 0;
2832 dl_se->dl_period = 0;
2833 dl_se->flags = 0;
2834 dl_se->dl_bw = 0;
2835 dl_se->dl_density = 0;
06a76fe0 2836
97fb7a0a
IM
2837 dl_se->dl_throttled = 0;
2838 dl_se->dl_yielded = 0;
2839 dl_se->dl_non_contending = 0;
2840 dl_se->dl_overrun = 0;
2279f540
JL
2841
2842#ifdef CONFIG_RT_MUTEXES
2843 dl_se->pi_se = dl_se;
2844#endif
06a76fe0
NP
2845}
2846
2847bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2848{
2849 struct sched_dl_entity *dl_se = &p->dl;
2850
2851 if (dl_se->dl_runtime != attr->sched_runtime ||
2852 dl_se->dl_deadline != attr->sched_deadline ||
2853 dl_se->dl_period != attr->sched_period ||
2854 dl_se->flags != attr->sched_flags)
2855 return true;
2856
2857 return false;
2858}
2859
2860#ifdef CONFIG_SMP
2861int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed)
2862{
60ffd5ed 2863 unsigned long flags, cap;
97fb7a0a 2864 unsigned int dest_cpu;
06a76fe0
NP
2865 struct dl_bw *dl_b;
2866 bool overflow;
60ffd5ed 2867 int ret;
06a76fe0 2868
97fb7a0a
IM
2869 dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
2870
06a76fe0
NP
2871 rcu_read_lock_sched();
2872 dl_b = dl_bw_of(dest_cpu);
2873 raw_spin_lock_irqsave(&dl_b->lock, flags);
60ffd5ed
LA
2874 cap = dl_bw_capacity(dest_cpu);
2875 overflow = __dl_overflow(dl_b, cap, 0, p->dl.dl_bw);
97fb7a0a 2876 if (overflow) {
06a76fe0 2877 ret = -EBUSY;
97fb7a0a 2878 } else {
06a76fe0
NP
2879 /*
2880 * We reserve space for this task in the destination
2881 * root_domain, as we can't fail after this point.
2882 * We will free resources in the source root_domain
2883 * later on (see set_cpus_allowed_dl()).
2884 */
60ffd5ed
LA
2885 int cpus = dl_bw_cpus(dest_cpu);
2886
06a76fe0
NP
2887 __dl_add(dl_b, p->dl.dl_bw, cpus);
2888 ret = 0;
2889 }
2890 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2891 rcu_read_unlock_sched();
97fb7a0a 2892
06a76fe0
NP
2893 return ret;
2894}
2895
2896int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
2897 const struct cpumask *trial)
2898{
2899 int ret = 1, trial_cpus;
2900 struct dl_bw *cur_dl_b;
2901 unsigned long flags;
2902
2903 rcu_read_lock_sched();
2904 cur_dl_b = dl_bw_of(cpumask_any(cur));
2905 trial_cpus = cpumask_weight(trial);
2906
2907 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
2908 if (cur_dl_b->bw != -1 &&
2909 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
2910 ret = 0;
2911 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
2912 rcu_read_unlock_sched();
97fb7a0a 2913
06a76fe0
NP
2914 return ret;
2915}
2916
2917bool dl_cpu_busy(unsigned int cpu)
2918{
60ffd5ed 2919 unsigned long flags, cap;
06a76fe0
NP
2920 struct dl_bw *dl_b;
2921 bool overflow;
06a76fe0
NP
2922
2923 rcu_read_lock_sched();
2924 dl_b = dl_bw_of(cpu);
2925 raw_spin_lock_irqsave(&dl_b->lock, flags);
60ffd5ed
LA
2926 cap = dl_bw_capacity(cpu);
2927 overflow = __dl_overflow(dl_b, cap, 0, 0);
06a76fe0
NP
2928 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2929 rcu_read_unlock_sched();
97fb7a0a 2930
06a76fe0
NP
2931 return overflow;
2932}
2933#endif
2934
acb32132 2935#ifdef CONFIG_SCHED_DEBUG
acb32132
WL
2936void print_dl_stats(struct seq_file *m, int cpu)
2937{
2938 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
2939}
2940#endif /* CONFIG_SCHED_DEBUG */