Commit | Line | Data |
---|---|---|
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 |
21 | struct dl_bandwidth def_dl_bandwidth; |
22 | ||
aab03e05 DF |
23 | static 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 | ||
28 | static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq) | |
29 | { | |
30 | return container_of(dl_rq, struct rq, dl); | |
31 | } | |
32 | ||
33 | static 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 | ||
41 | static inline int on_dl_rq(struct sched_dl_entity *dl_se) | |
42 | { | |
43 | return !RB_EMPTY_NODE(&dl_se->rb_node); | |
44 | } | |
45 | ||
06a76fe0 NP |
46 | #ifdef CONFIG_SMP |
47 | static inline struct dl_bw *dl_bw_of(int i) | |
48 | { | |
49 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), | |
50 | "sched RCU must be held"); | |
51 | return &cpu_rq(i)->rd->dl_bw; | |
52 | } | |
53 | ||
54 | static inline int dl_bw_cpus(int i) | |
55 | { | |
56 | struct root_domain *rd = cpu_rq(i)->rd; | |
c81b8932 | 57 | int cpus; |
06a76fe0 NP |
58 | |
59 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), | |
60 | "sched RCU must be held"); | |
c81b8932 DE |
61 | |
62 | if (cpumask_subset(rd->span, cpu_active_mask)) | |
63 | return cpumask_weight(rd->span); | |
64 | ||
65 | cpus = 0; | |
66 | ||
06a76fe0 NP |
67 | for_each_cpu_and(i, rd->span, cpu_active_mask) |
68 | cpus++; | |
69 | ||
70 | return cpus; | |
71 | } | |
fc9dc698 DE |
72 | |
73 | static inline unsigned long __dl_bw_capacity(int i) | |
74 | { | |
75 | struct root_domain *rd = cpu_rq(i)->rd; | |
76 | unsigned long cap = 0; | |
77 | ||
78 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), | |
79 | "sched RCU must be held"); | |
80 | ||
81 | for_each_cpu_and(i, rd->span, cpu_active_mask) | |
82 | cap += capacity_orig_of(i); | |
83 | ||
84 | return cap; | |
85 | } | |
86 | ||
87 | /* | |
88 | * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity | |
89 | * of the CPU the task is running on rather rd's \Sum CPU capacity. | |
90 | */ | |
91 | static inline unsigned long dl_bw_capacity(int i) | |
92 | { | |
93 | if (!static_branch_unlikely(&sched_asym_cpucapacity) && | |
94 | capacity_orig_of(i) == SCHED_CAPACITY_SCALE) { | |
95 | return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT; | |
96 | } else { | |
97 | return __dl_bw_capacity(i); | |
98 | } | |
99 | } | |
06a76fe0 NP |
100 | #else |
101 | static inline struct dl_bw *dl_bw_of(int i) | |
102 | { | |
103 | return &cpu_rq(i)->dl.dl_bw; | |
104 | } | |
105 | ||
106 | static inline int dl_bw_cpus(int i) | |
107 | { | |
108 | return 1; | |
109 | } | |
fc9dc698 DE |
110 | |
111 | static inline unsigned long dl_bw_capacity(int i) | |
112 | { | |
113 | return SCHED_CAPACITY_SCALE; | |
114 | } | |
06a76fe0 NP |
115 | #endif |
116 | ||
e36d8677 | 117 | static inline |
794a56eb | 118 | void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq) |
e36d8677 LA |
119 | { |
120 | u64 old = dl_rq->running_bw; | |
121 | ||
122 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
123 | dl_rq->running_bw += dl_bw; | |
124 | SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */ | |
8fd27231 | 125 | SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw); |
e0367b12 | 126 | /* kick cpufreq (see the comment in kernel/sched/sched.h). */ |
4042d003 | 127 | cpufreq_update_util(rq_of_dl_rq(dl_rq), 0); |
e36d8677 LA |
128 | } |
129 | ||
130 | static inline | |
794a56eb | 131 | void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq) |
e36d8677 LA |
132 | { |
133 | u64 old = dl_rq->running_bw; | |
134 | ||
135 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
136 | dl_rq->running_bw -= dl_bw; | |
137 | SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */ | |
138 | if (dl_rq->running_bw > old) | |
139 | dl_rq->running_bw = 0; | |
e0367b12 | 140 | /* kick cpufreq (see the comment in kernel/sched/sched.h). */ |
4042d003 | 141 | cpufreq_update_util(rq_of_dl_rq(dl_rq), 0); |
e36d8677 LA |
142 | } |
143 | ||
8fd27231 | 144 | static inline |
794a56eb | 145 | void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq) |
8fd27231 LA |
146 | { |
147 | u64 old = dl_rq->this_bw; | |
148 | ||
149 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
150 | dl_rq->this_bw += dl_bw; | |
151 | SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */ | |
152 | } | |
153 | ||
154 | static inline | |
794a56eb | 155 | void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq) |
8fd27231 LA |
156 | { |
157 | u64 old = dl_rq->this_bw; | |
158 | ||
159 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
160 | dl_rq->this_bw -= dl_bw; | |
161 | SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */ | |
162 | if (dl_rq->this_bw > old) | |
163 | dl_rq->this_bw = 0; | |
164 | SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw); | |
165 | } | |
166 | ||
794a56eb JL |
167 | static inline |
168 | void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
169 | { | |
170 | if (!dl_entity_is_special(dl_se)) | |
171 | __add_rq_bw(dl_se->dl_bw, dl_rq); | |
172 | } | |
173 | ||
174 | static inline | |
175 | void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
176 | { | |
177 | if (!dl_entity_is_special(dl_se)) | |
178 | __sub_rq_bw(dl_se->dl_bw, dl_rq); | |
179 | } | |
180 | ||
181 | static inline | |
182 | void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
183 | { | |
184 | if (!dl_entity_is_special(dl_se)) | |
185 | __add_running_bw(dl_se->dl_bw, dl_rq); | |
186 | } | |
187 | ||
188 | static inline | |
189 | void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
190 | { | |
191 | if (!dl_entity_is_special(dl_se)) | |
192 | __sub_running_bw(dl_se->dl_bw, dl_rq); | |
193 | } | |
194 | ||
ba4f7bc1 | 195 | static void dl_change_utilization(struct task_struct *p, u64 new_bw) |
209a0cbd | 196 | { |
8fd27231 | 197 | struct rq *rq; |
209a0cbd | 198 | |
794a56eb JL |
199 | BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV); |
200 | ||
8fd27231 | 201 | if (task_on_rq_queued(p)) |
209a0cbd LA |
202 | return; |
203 | ||
8fd27231 LA |
204 | rq = task_rq(p); |
205 | if (p->dl.dl_non_contending) { | |
794a56eb | 206 | sub_running_bw(&p->dl, &rq->dl); |
8fd27231 LA |
207 | p->dl.dl_non_contending = 0; |
208 | /* | |
209 | * If the timer handler is currently running and the | |
210 | * timer cannot be cancelled, inactive_task_timer() | |
211 | * will see that dl_not_contending is not set, and | |
212 | * will not touch the rq's active utilization, | |
213 | * so we are still safe. | |
214 | */ | |
215 | if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) | |
216 | put_task_struct(p); | |
217 | } | |
794a56eb JL |
218 | __sub_rq_bw(p->dl.dl_bw, &rq->dl); |
219 | __add_rq_bw(new_bw, &rq->dl); | |
209a0cbd LA |
220 | } |
221 | ||
222 | /* | |
223 | * The utilization of a task cannot be immediately removed from | |
224 | * the rq active utilization (running_bw) when the task blocks. | |
225 | * Instead, we have to wait for the so called "0-lag time". | |
226 | * | |
227 | * If a task blocks before the "0-lag time", a timer (the inactive | |
228 | * timer) is armed, and running_bw is decreased when the timer | |
229 | * fires. | |
230 | * | |
231 | * If the task wakes up again before the inactive timer fires, | |
232 | * the timer is cancelled, whereas if the task wakes up after the | |
233 | * inactive timer fired (and running_bw has been decreased) the | |
234 | * task's utilization has to be added to running_bw again. | |
235 | * A flag in the deadline scheduling entity (dl_non_contending) | |
236 | * is used to avoid race conditions between the inactive timer handler | |
237 | * and task wakeups. | |
238 | * | |
239 | * The following diagram shows how running_bw is updated. A task is | |
240 | * "ACTIVE" when its utilization contributes to running_bw; an | |
241 | * "ACTIVE contending" task is in the TASK_RUNNING state, while an | |
242 | * "ACTIVE non contending" task is a blocked task for which the "0-lag time" | |
243 | * has not passed yet. An "INACTIVE" task is a task for which the "0-lag" | |
244 | * time already passed, which does not contribute to running_bw anymore. | |
245 | * +------------------+ | |
246 | * wakeup | ACTIVE | | |
247 | * +------------------>+ contending | | |
248 | * | add_running_bw | | | |
249 | * | +----+------+------+ | |
250 | * | | ^ | |
251 | * | dequeue | | | |
252 | * +--------+-------+ | | | |
253 | * | | t >= 0-lag | | wakeup | |
254 | * | INACTIVE |<---------------+ | | |
255 | * | | sub_running_bw | | | |
256 | * +--------+-------+ | | | |
257 | * ^ | | | |
258 | * | t < 0-lag | | | |
259 | * | | | | |
260 | * | V | | |
261 | * | +----+------+------+ | |
262 | * | sub_running_bw | ACTIVE | | |
263 | * +-------------------+ | | |
264 | * inactive timer | non contending | | |
265 | * fired +------------------+ | |
266 | * | |
267 | * The task_non_contending() function is invoked when a task | |
268 | * blocks, and checks if the 0-lag time already passed or | |
269 | * not (in the first case, it directly updates running_bw; | |
270 | * in the second case, it arms the inactive timer). | |
271 | * | |
272 | * The task_contending() function is invoked when a task wakes | |
273 | * up, and checks if the task is still in the "ACTIVE non contending" | |
274 | * state or not (in the second case, it updates running_bw). | |
275 | */ | |
276 | static void task_non_contending(struct task_struct *p) | |
277 | { | |
278 | struct sched_dl_entity *dl_se = &p->dl; | |
279 | struct hrtimer *timer = &dl_se->inactive_timer; | |
280 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
281 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
282 | s64 zerolag_time; | |
283 | ||
284 | /* | |
285 | * If this is a non-deadline task that has been boosted, | |
286 | * do nothing | |
287 | */ | |
288 | if (dl_se->dl_runtime == 0) | |
289 | return; | |
290 | ||
794a56eb JL |
291 | if (dl_entity_is_special(dl_se)) |
292 | return; | |
293 | ||
209a0cbd LA |
294 | WARN_ON(dl_se->dl_non_contending); |
295 | ||
296 | zerolag_time = dl_se->deadline - | |
297 | div64_long((dl_se->runtime * dl_se->dl_period), | |
298 | dl_se->dl_runtime); | |
299 | ||
300 | /* | |
301 | * Using relative times instead of the absolute "0-lag time" | |
302 | * allows to simplify the code | |
303 | */ | |
304 | zerolag_time -= rq_clock(rq); | |
305 | ||
306 | /* | |
307 | * If the "0-lag time" already passed, decrease the active | |
308 | * utilization now, instead of starting a timer | |
309 | */ | |
1b02cd6a | 310 | if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) { |
209a0cbd | 311 | if (dl_task(p)) |
794a56eb | 312 | sub_running_bw(dl_se, dl_rq); |
387e3130 LA |
313 | if (!dl_task(p) || p->state == TASK_DEAD) { |
314 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); | |
315 | ||
8fd27231 | 316 | if (p->state == TASK_DEAD) |
794a56eb | 317 | sub_rq_bw(&p->dl, &rq->dl); |
387e3130 | 318 | raw_spin_lock(&dl_b->lock); |
8c0944ce | 319 | __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); |
209a0cbd | 320 | __dl_clear_params(p); |
387e3130 LA |
321 | raw_spin_unlock(&dl_b->lock); |
322 | } | |
209a0cbd LA |
323 | |
324 | return; | |
325 | } | |
326 | ||
327 | dl_se->dl_non_contending = 1; | |
328 | get_task_struct(p); | |
850377a8 | 329 | hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD); |
209a0cbd LA |
330 | } |
331 | ||
8fd27231 | 332 | static void task_contending(struct sched_dl_entity *dl_se, int flags) |
209a0cbd LA |
333 | { |
334 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
335 | ||
336 | /* | |
337 | * If this is a non-deadline task that has been boosted, | |
338 | * do nothing | |
339 | */ | |
340 | if (dl_se->dl_runtime == 0) | |
341 | return; | |
342 | ||
8fd27231 | 343 | if (flags & ENQUEUE_MIGRATED) |
794a56eb | 344 | add_rq_bw(dl_se, dl_rq); |
8fd27231 | 345 | |
209a0cbd LA |
346 | if (dl_se->dl_non_contending) { |
347 | dl_se->dl_non_contending = 0; | |
348 | /* | |
349 | * If the timer handler is currently running and the | |
350 | * timer cannot be cancelled, inactive_task_timer() | |
351 | * will see that dl_not_contending is not set, and | |
352 | * will not touch the rq's active utilization, | |
353 | * so we are still safe. | |
354 | */ | |
355 | if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1) | |
356 | put_task_struct(dl_task_of(dl_se)); | |
357 | } else { | |
358 | /* | |
359 | * Since "dl_non_contending" is not set, the | |
360 | * task's utilization has already been removed from | |
361 | * active utilization (either when the task blocked, | |
362 | * when the "inactive timer" fired). | |
363 | * So, add it back. | |
364 | */ | |
794a56eb | 365 | add_running_bw(dl_se, dl_rq); |
209a0cbd LA |
366 | } |
367 | } | |
368 | ||
aab03e05 DF |
369 | static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq) |
370 | { | |
371 | struct sched_dl_entity *dl_se = &p->dl; | |
372 | ||
2161573e | 373 | return dl_rq->root.rb_leftmost == &dl_se->rb_node; |
aab03e05 DF |
374 | } |
375 | ||
ba4f7bc1 YC |
376 | static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq); |
377 | ||
332ac17e DF |
378 | void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime) |
379 | { | |
380 | raw_spin_lock_init(&dl_b->dl_runtime_lock); | |
381 | dl_b->dl_period = period; | |
382 | dl_b->dl_runtime = runtime; | |
383 | } | |
384 | ||
332ac17e DF |
385 | void init_dl_bw(struct dl_bw *dl_b) |
386 | { | |
387 | raw_spin_lock_init(&dl_b->lock); | |
388 | raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock); | |
1724813d | 389 | if (global_rt_runtime() == RUNTIME_INF) |
332ac17e DF |
390 | dl_b->bw = -1; |
391 | else | |
1724813d | 392 | dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime()); |
332ac17e DF |
393 | raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock); |
394 | dl_b->total_bw = 0; | |
395 | } | |
396 | ||
07c54f7a | 397 | void init_dl_rq(struct dl_rq *dl_rq) |
aab03e05 | 398 | { |
2161573e | 399 | dl_rq->root = RB_ROOT_CACHED; |
1baca4ce JL |
400 | |
401 | #ifdef CONFIG_SMP | |
402 | /* zero means no -deadline tasks */ | |
403 | dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0; | |
404 | ||
405 | dl_rq->dl_nr_migratory = 0; | |
406 | dl_rq->overloaded = 0; | |
2161573e | 407 | dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED; |
332ac17e DF |
408 | #else |
409 | init_dl_bw(&dl_rq->dl_bw); | |
1baca4ce | 410 | #endif |
e36d8677 LA |
411 | |
412 | dl_rq->running_bw = 0; | |
8fd27231 | 413 | dl_rq->this_bw = 0; |
4da3abce | 414 | init_dl_rq_bw_ratio(dl_rq); |
1baca4ce JL |
415 | } |
416 | ||
417 | #ifdef CONFIG_SMP | |
418 | ||
419 | static inline int dl_overloaded(struct rq *rq) | |
420 | { | |
421 | return atomic_read(&rq->rd->dlo_count); | |
422 | } | |
423 | ||
424 | static inline void dl_set_overload(struct rq *rq) | |
425 | { | |
426 | if (!rq->online) | |
427 | return; | |
428 | ||
429 | cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask); | |
430 | /* | |
431 | * Must be visible before the overload count is | |
432 | * set (as in sched_rt.c). | |
433 | * | |
434 | * Matched by the barrier in pull_dl_task(). | |
435 | */ | |
436 | smp_wmb(); | |
437 | atomic_inc(&rq->rd->dlo_count); | |
438 | } | |
439 | ||
440 | static inline void dl_clear_overload(struct rq *rq) | |
441 | { | |
442 | if (!rq->online) | |
443 | return; | |
444 | ||
445 | atomic_dec(&rq->rd->dlo_count); | |
446 | cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask); | |
447 | } | |
448 | ||
449 | static void update_dl_migration(struct dl_rq *dl_rq) | |
450 | { | |
995b9ea4 | 451 | if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) { |
1baca4ce JL |
452 | if (!dl_rq->overloaded) { |
453 | dl_set_overload(rq_of_dl_rq(dl_rq)); | |
454 | dl_rq->overloaded = 1; | |
455 | } | |
456 | } else if (dl_rq->overloaded) { | |
457 | dl_clear_overload(rq_of_dl_rq(dl_rq)); | |
458 | dl_rq->overloaded = 0; | |
459 | } | |
460 | } | |
461 | ||
462 | static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
463 | { | |
464 | struct task_struct *p = dl_task_of(dl_se); | |
1baca4ce | 465 | |
4b53a341 | 466 | if (p->nr_cpus_allowed > 1) |
1baca4ce JL |
467 | dl_rq->dl_nr_migratory++; |
468 | ||
469 | update_dl_migration(dl_rq); | |
470 | } | |
471 | ||
472 | static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
473 | { | |
474 | struct task_struct *p = dl_task_of(dl_se); | |
1baca4ce | 475 | |
4b53a341 | 476 | if (p->nr_cpus_allowed > 1) |
1baca4ce JL |
477 | dl_rq->dl_nr_migratory--; |
478 | ||
479 | update_dl_migration(dl_rq); | |
480 | } | |
481 | ||
482 | /* | |
483 | * The list of pushable -deadline task is not a plist, like in | |
484 | * sched_rt.c, it is an rb-tree with tasks ordered by deadline. | |
485 | */ | |
486 | static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
487 | { | |
488 | struct dl_rq *dl_rq = &rq->dl; | |
2161573e | 489 | struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_root.rb_node; |
1baca4ce JL |
490 | struct rb_node *parent = NULL; |
491 | struct task_struct *entry; | |
2161573e | 492 | bool leftmost = true; |
1baca4ce JL |
493 | |
494 | BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks)); | |
495 | ||
496 | while (*link) { | |
497 | parent = *link; | |
498 | entry = rb_entry(parent, struct task_struct, | |
499 | pushable_dl_tasks); | |
500 | if (dl_entity_preempt(&p->dl, &entry->dl)) | |
501 | link = &parent->rb_left; | |
502 | else { | |
503 | link = &parent->rb_right; | |
2161573e | 504 | leftmost = false; |
1baca4ce JL |
505 | } |
506 | } | |
507 | ||
2161573e | 508 | if (leftmost) |
7d92de3a | 509 | dl_rq->earliest_dl.next = p->dl.deadline; |
1baca4ce JL |
510 | |
511 | rb_link_node(&p->pushable_dl_tasks, parent, link); | |
2161573e DB |
512 | rb_insert_color_cached(&p->pushable_dl_tasks, |
513 | &dl_rq->pushable_dl_tasks_root, leftmost); | |
aab03e05 DF |
514 | } |
515 | ||
1baca4ce JL |
516 | static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) |
517 | { | |
518 | struct dl_rq *dl_rq = &rq->dl; | |
519 | ||
520 | if (RB_EMPTY_NODE(&p->pushable_dl_tasks)) | |
521 | return; | |
522 | ||
2161573e | 523 | if (dl_rq->pushable_dl_tasks_root.rb_leftmost == &p->pushable_dl_tasks) { |
1baca4ce JL |
524 | struct rb_node *next_node; |
525 | ||
526 | next_node = rb_next(&p->pushable_dl_tasks); | |
7d92de3a WL |
527 | if (next_node) { |
528 | dl_rq->earliest_dl.next = rb_entry(next_node, | |
529 | struct task_struct, pushable_dl_tasks)->dl.deadline; | |
530 | } | |
1baca4ce JL |
531 | } |
532 | ||
2161573e | 533 | rb_erase_cached(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); |
1baca4ce JL |
534 | RB_CLEAR_NODE(&p->pushable_dl_tasks); |
535 | } | |
536 | ||
537 | static inline int has_pushable_dl_tasks(struct rq *rq) | |
538 | { | |
2161573e | 539 | return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root); |
1baca4ce JL |
540 | } |
541 | ||
542 | static int push_dl_task(struct rq *rq); | |
543 | ||
dc877341 PZ |
544 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
545 | { | |
546 | return dl_task(prev); | |
547 | } | |
548 | ||
9916e214 PZ |
549 | static DEFINE_PER_CPU(struct callback_head, dl_push_head); |
550 | static DEFINE_PER_CPU(struct callback_head, dl_pull_head); | |
e3fca9e7 PZ |
551 | |
552 | static void push_dl_tasks(struct rq *); | |
9916e214 | 553 | static void pull_dl_task(struct rq *); |
e3fca9e7 | 554 | |
02d8ec94 | 555 | static inline void deadline_queue_push_tasks(struct rq *rq) |
dc877341 | 556 | { |
e3fca9e7 PZ |
557 | if (!has_pushable_dl_tasks(rq)) |
558 | return; | |
559 | ||
9916e214 PZ |
560 | queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks); |
561 | } | |
562 | ||
02d8ec94 | 563 | static inline void deadline_queue_pull_task(struct rq *rq) |
9916e214 PZ |
564 | { |
565 | queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task); | |
dc877341 PZ |
566 | } |
567 | ||
fa9c9d10 WL |
568 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq); |
569 | ||
a649f237 | 570 | static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p) |
fa9c9d10 WL |
571 | { |
572 | struct rq *later_rq = NULL; | |
59d06cea | 573 | struct dl_bw *dl_b; |
fa9c9d10 WL |
574 | |
575 | later_rq = find_lock_later_rq(p, rq); | |
fa9c9d10 WL |
576 | if (!later_rq) { |
577 | int cpu; | |
578 | ||
579 | /* | |
580 | * If we cannot preempt any rq, fall back to pick any | |
97fb7a0a | 581 | * online CPU: |
fa9c9d10 | 582 | */ |
3bd37062 | 583 | cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr); |
fa9c9d10 WL |
584 | if (cpu >= nr_cpu_ids) { |
585 | /* | |
97fb7a0a | 586 | * Failed to find any suitable CPU. |
fa9c9d10 WL |
587 | * The task will never come back! |
588 | */ | |
589 | BUG_ON(dl_bandwidth_enabled()); | |
590 | ||
591 | /* | |
592 | * If admission control is disabled we | |
593 | * try a little harder to let the task | |
594 | * run. | |
595 | */ | |
596 | cpu = cpumask_any(cpu_active_mask); | |
597 | } | |
598 | later_rq = cpu_rq(cpu); | |
599 | double_lock_balance(rq, later_rq); | |
600 | } | |
601 | ||
59d06cea JL |
602 | if (p->dl.dl_non_contending || p->dl.dl_throttled) { |
603 | /* | |
604 | * Inactive timer is armed (or callback is running, but | |
605 | * waiting for us to release rq locks). In any case, when it | |
606 | * will fire (or continue), it will see running_bw of this | |
607 | * task migrated to later_rq (and correctly handle it). | |
608 | */ | |
609 | sub_running_bw(&p->dl, &rq->dl); | |
610 | sub_rq_bw(&p->dl, &rq->dl); | |
611 | ||
612 | add_rq_bw(&p->dl, &later_rq->dl); | |
613 | add_running_bw(&p->dl, &later_rq->dl); | |
614 | } else { | |
615 | sub_rq_bw(&p->dl, &rq->dl); | |
616 | add_rq_bw(&p->dl, &later_rq->dl); | |
617 | } | |
618 | ||
619 | /* | |
620 | * And we finally need to fixup root_domain(s) bandwidth accounting, | |
621 | * since p is still hanging out in the old (now moved to default) root | |
622 | * domain. | |
623 | */ | |
624 | dl_b = &rq->rd->dl_bw; | |
625 | raw_spin_lock(&dl_b->lock); | |
626 | __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span)); | |
627 | raw_spin_unlock(&dl_b->lock); | |
628 | ||
629 | dl_b = &later_rq->rd->dl_bw; | |
630 | raw_spin_lock(&dl_b->lock); | |
631 | __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span)); | |
632 | raw_spin_unlock(&dl_b->lock); | |
633 | ||
fa9c9d10 | 634 | set_task_cpu(p, later_rq->cpu); |
a649f237 PZ |
635 | double_unlock_balance(later_rq, rq); |
636 | ||
637 | return later_rq; | |
fa9c9d10 WL |
638 | } |
639 | ||
1baca4ce JL |
640 | #else |
641 | ||
642 | static inline | |
643 | void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
644 | { | |
645 | } | |
646 | ||
647 | static inline | |
648 | void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
649 | { | |
650 | } | |
651 | ||
652 | static inline | |
653 | void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
654 | { | |
655 | } | |
656 | ||
657 | static inline | |
658 | void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
659 | { | |
660 | } | |
661 | ||
dc877341 PZ |
662 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
663 | { | |
664 | return false; | |
665 | } | |
666 | ||
0ea60c20 | 667 | static inline void pull_dl_task(struct rq *rq) |
dc877341 | 668 | { |
dc877341 PZ |
669 | } |
670 | ||
02d8ec94 | 671 | static inline void deadline_queue_push_tasks(struct rq *rq) |
dc877341 | 672 | { |
dc877341 PZ |
673 | } |
674 | ||
02d8ec94 | 675 | static inline void deadline_queue_pull_task(struct rq *rq) |
dc877341 PZ |
676 | { |
677 | } | |
1baca4ce JL |
678 | #endif /* CONFIG_SMP */ |
679 | ||
aab03e05 DF |
680 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags); |
681 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags); | |
97fb7a0a | 682 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags); |
aab03e05 DF |
683 | |
684 | /* | |
685 | * We are being explicitly informed that a new instance is starting, | |
686 | * and this means that: | |
687 | * - the absolute deadline of the entity has to be placed at | |
688 | * current time + relative deadline; | |
689 | * - the runtime of the entity has to be set to the maximum value. | |
690 | * | |
691 | * The capability of specifying such event is useful whenever a -deadline | |
692 | * entity wants to (try to!) synchronize its behaviour with the scheduler's | |
693 | * one, and to (try to!) reconcile itself with its own scheduling | |
694 | * parameters. | |
695 | */ | |
98b0a857 | 696 | static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se) |
aab03e05 DF |
697 | { |
698 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
699 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
700 | ||
98b0a857 | 701 | WARN_ON(dl_se->dl_boosted); |
72f9f3fd LA |
702 | WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline)); |
703 | ||
704 | /* | |
705 | * We are racing with the deadline timer. So, do nothing because | |
706 | * the deadline timer handler will take care of properly recharging | |
707 | * the runtime and postponing the deadline | |
708 | */ | |
709 | if (dl_se->dl_throttled) | |
710 | return; | |
aab03e05 DF |
711 | |
712 | /* | |
713 | * We use the regular wall clock time to set deadlines in the | |
714 | * future; in fact, we must consider execution overheads (time | |
715 | * spent on hardirq context, etc.). | |
716 | */ | |
98b0a857 JL |
717 | dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline; |
718 | dl_se->runtime = dl_se->dl_runtime; | |
aab03e05 DF |
719 | } |
720 | ||
721 | /* | |
722 | * Pure Earliest Deadline First (EDF) scheduling does not deal with the | |
723 | * possibility of a entity lasting more than what it declared, and thus | |
724 | * exhausting its runtime. | |
725 | * | |
726 | * Here we are interested in making runtime overrun possible, but we do | |
727 | * not want a entity which is misbehaving to affect the scheduling of all | |
728 | * other entities. | |
729 | * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS) | |
730 | * is used, in order to confine each entity within its own bandwidth. | |
731 | * | |
732 | * This function deals exactly with that, and ensures that when the runtime | |
733 | * of a entity is replenished, its deadline is also postponed. That ensures | |
734 | * the overrunning entity can't interfere with other entity in the system and | |
735 | * can't make them miss their deadlines. Reasons why this kind of overruns | |
736 | * could happen are, typically, a entity voluntarily trying to overcome its | |
1b09d29b | 737 | * runtime, or it just underestimated it during sched_setattr(). |
aab03e05 | 738 | */ |
2d3d891d DF |
739 | static void replenish_dl_entity(struct sched_dl_entity *dl_se, |
740 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
741 | { |
742 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
743 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
744 | ||
2d3d891d DF |
745 | BUG_ON(pi_se->dl_runtime <= 0); |
746 | ||
747 | /* | |
748 | * This could be the case for a !-dl task that is boosted. | |
749 | * Just go with full inherited parameters. | |
750 | */ | |
751 | if (dl_se->dl_deadline == 0) { | |
752 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | |
753 | dl_se->runtime = pi_se->dl_runtime; | |
754 | } | |
755 | ||
48be3a67 PZ |
756 | if (dl_se->dl_yielded && dl_se->runtime > 0) |
757 | dl_se->runtime = 0; | |
758 | ||
aab03e05 DF |
759 | /* |
760 | * We keep moving the deadline away until we get some | |
761 | * available runtime for the entity. This ensures correct | |
762 | * handling of situations where the runtime overrun is | |
763 | * arbitrary large. | |
764 | */ | |
765 | while (dl_se->runtime <= 0) { | |
2d3d891d DF |
766 | dl_se->deadline += pi_se->dl_period; |
767 | dl_se->runtime += pi_se->dl_runtime; | |
aab03e05 DF |
768 | } |
769 | ||
770 | /* | |
771 | * At this point, the deadline really should be "in | |
772 | * the future" with respect to rq->clock. If it's | |
773 | * not, we are, for some reason, lagging too much! | |
774 | * Anyway, after having warn userspace abut that, | |
775 | * we still try to keep the things running by | |
776 | * resetting the deadline and the budget of the | |
777 | * entity. | |
778 | */ | |
779 | if (dl_time_before(dl_se->deadline, rq_clock(rq))) { | |
c219b7dd | 780 | printk_deferred_once("sched: DL replenish lagged too much\n"); |
2d3d891d DF |
781 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
782 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 | 783 | } |
1019a359 PZ |
784 | |
785 | if (dl_se->dl_yielded) | |
786 | dl_se->dl_yielded = 0; | |
787 | if (dl_se->dl_throttled) | |
788 | dl_se->dl_throttled = 0; | |
aab03e05 DF |
789 | } |
790 | ||
791 | /* | |
792 | * Here we check if --at time t-- an entity (which is probably being | |
793 | * [re]activated or, in general, enqueued) can use its remaining runtime | |
794 | * and its current deadline _without_ exceeding the bandwidth it is | |
795 | * assigned (function returns true if it can't). We are in fact applying | |
796 | * one of the CBS rules: when a task wakes up, if the residual runtime | |
797 | * over residual deadline fits within the allocated bandwidth, then we | |
798 | * can keep the current (absolute) deadline and residual budget without | |
799 | * disrupting the schedulability of the system. Otherwise, we should | |
800 | * refill the runtime and set the deadline a period in the future, | |
801 | * because keeping the current (absolute) deadline of the task would | |
712e5e34 | 802 | * result in breaking guarantees promised to other tasks (refer to |
d6a3b247 | 803 | * Documentation/scheduler/sched-deadline.rst for more information). |
aab03e05 DF |
804 | * |
805 | * This function returns true if: | |
806 | * | |
2317d5f1 | 807 | * runtime / (deadline - t) > dl_runtime / dl_deadline , |
aab03e05 DF |
808 | * |
809 | * IOW we can't recycle current parameters. | |
755378a4 | 810 | * |
2317d5f1 | 811 | * Notice that the bandwidth check is done against the deadline. For |
755378a4 | 812 | * task with deadline equal to period this is the same of using |
2317d5f1 | 813 | * dl_period instead of dl_deadline in the equation above. |
aab03e05 | 814 | */ |
2d3d891d DF |
815 | static bool dl_entity_overflow(struct sched_dl_entity *dl_se, |
816 | struct sched_dl_entity *pi_se, u64 t) | |
aab03e05 DF |
817 | { |
818 | u64 left, right; | |
819 | ||
820 | /* | |
821 | * left and right are the two sides of the equation above, | |
822 | * after a bit of shuffling to use multiplications instead | |
823 | * of divisions. | |
824 | * | |
825 | * Note that none of the time values involved in the two | |
826 | * multiplications are absolute: dl_deadline and dl_runtime | |
827 | * are the relative deadline and the maximum runtime of each | |
828 | * instance, runtime is the runtime left for the last instance | |
829 | * and (deadline - t), since t is rq->clock, is the time left | |
830 | * to the (absolute) deadline. Even if overflowing the u64 type | |
831 | * is very unlikely to occur in both cases, here we scale down | |
832 | * as we want to avoid that risk at all. Scaling down by 10 | |
833 | * means that we reduce granularity to 1us. We are fine with it, | |
834 | * since this is only a true/false check and, anyway, thinking | |
835 | * of anything below microseconds resolution is actually fiction | |
836 | * (but still we want to give the user that illusion >;). | |
837 | */ | |
2317d5f1 | 838 | left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE); |
332ac17e DF |
839 | right = ((dl_se->deadline - t) >> DL_SCALE) * |
840 | (pi_se->dl_runtime >> DL_SCALE); | |
aab03e05 DF |
841 | |
842 | return dl_time_before(right, left); | |
843 | } | |
844 | ||
845 | /* | |
3effcb42 DBO |
846 | * Revised wakeup rule [1]: For self-suspending tasks, rather then |
847 | * re-initializing task's runtime and deadline, the revised wakeup | |
848 | * rule adjusts the task's runtime to avoid the task to overrun its | |
849 | * density. | |
aab03e05 | 850 | * |
3effcb42 DBO |
851 | * Reasoning: a task may overrun the density if: |
852 | * runtime / (deadline - t) > dl_runtime / dl_deadline | |
853 | * | |
854 | * Therefore, runtime can be adjusted to: | |
855 | * runtime = (dl_runtime / dl_deadline) * (deadline - t) | |
856 | * | |
857 | * In such way that runtime will be equal to the maximum density | |
858 | * the task can use without breaking any rule. | |
859 | * | |
860 | * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant | |
861 | * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24. | |
862 | */ | |
863 | static void | |
864 | update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq) | |
865 | { | |
866 | u64 laxity = dl_se->deadline - rq_clock(rq); | |
867 | ||
868 | /* | |
869 | * If the task has deadline < period, and the deadline is in the past, | |
870 | * it should already be throttled before this check. | |
871 | * | |
872 | * See update_dl_entity() comments for further details. | |
873 | */ | |
874 | WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq))); | |
875 | ||
876 | dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT; | |
877 | } | |
878 | ||
879 | /* | |
880 | * Regarding the deadline, a task with implicit deadline has a relative | |
881 | * deadline == relative period. A task with constrained deadline has a | |
882 | * relative deadline <= relative period. | |
883 | * | |
884 | * We support constrained deadline tasks. However, there are some restrictions | |
885 | * applied only for tasks which do not have an implicit deadline. See | |
886 | * update_dl_entity() to know more about such restrictions. | |
887 | * | |
888 | * The dl_is_implicit() returns true if the task has an implicit deadline. | |
889 | */ | |
890 | static inline bool dl_is_implicit(struct sched_dl_entity *dl_se) | |
891 | { | |
892 | return dl_se->dl_deadline == dl_se->dl_period; | |
893 | } | |
894 | ||
895 | /* | |
896 | * When a deadline entity is placed in the runqueue, its runtime and deadline | |
897 | * might need to be updated. This is done by a CBS wake up rule. There are two | |
898 | * different rules: 1) the original CBS; and 2) the Revisited CBS. | |
899 | * | |
900 | * When the task is starting a new period, the Original CBS is used. In this | |
901 | * case, the runtime is replenished and a new absolute deadline is set. | |
902 | * | |
903 | * When a task is queued before the begin of the next period, using the | |
904 | * remaining runtime and deadline could make the entity to overflow, see | |
905 | * dl_entity_overflow() to find more about runtime overflow. When such case | |
906 | * is detected, the runtime and deadline need to be updated. | |
907 | * | |
908 | * If the task has an implicit deadline, i.e., deadline == period, the Original | |
909 | * CBS is applied. the runtime is replenished and a new absolute deadline is | |
910 | * set, as in the previous cases. | |
911 | * | |
912 | * However, the Original CBS does not work properly for tasks with | |
913 | * deadline < period, which are said to have a constrained deadline. By | |
914 | * applying the Original CBS, a constrained deadline task would be able to run | |
915 | * runtime/deadline in a period. With deadline < period, the task would | |
916 | * overrun the runtime/period allowed bandwidth, breaking the admission test. | |
917 | * | |
918 | * In order to prevent this misbehave, the Revisited CBS is used for | |
919 | * constrained deadline tasks when a runtime overflow is detected. In the | |
920 | * Revisited CBS, rather than replenishing & setting a new absolute deadline, | |
921 | * the remaining runtime of the task is reduced to avoid runtime overflow. | |
922 | * Please refer to the comments update_dl_revised_wakeup() function to find | |
923 | * more about the Revised CBS rule. | |
aab03e05 | 924 | */ |
2d3d891d DF |
925 | static void update_dl_entity(struct sched_dl_entity *dl_se, |
926 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
927 | { |
928 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
929 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
930 | ||
aab03e05 | 931 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) || |
2d3d891d | 932 | dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { |
3effcb42 DBO |
933 | |
934 | if (unlikely(!dl_is_implicit(dl_se) && | |
935 | !dl_time_before(dl_se->deadline, rq_clock(rq)) && | |
936 | !dl_se->dl_boosted)){ | |
937 | update_dl_revised_wakeup(dl_se, rq); | |
938 | return; | |
939 | } | |
940 | ||
2d3d891d DF |
941 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
942 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 DF |
943 | } |
944 | } | |
945 | ||
5ac69d37 DBO |
946 | static inline u64 dl_next_period(struct sched_dl_entity *dl_se) |
947 | { | |
948 | return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period; | |
949 | } | |
950 | ||
aab03e05 DF |
951 | /* |
952 | * If the entity depleted all its runtime, and if we want it to sleep | |
953 | * while waiting for some new execution time to become available, we | |
5ac69d37 | 954 | * set the bandwidth replenishment timer to the replenishment instant |
aab03e05 DF |
955 | * and try to activate it. |
956 | * | |
957 | * Notice that it is important for the caller to know if the timer | |
958 | * actually started or not (i.e., the replenishment instant is in | |
959 | * the future or in the past). | |
960 | */ | |
a649f237 | 961 | static int start_dl_timer(struct task_struct *p) |
aab03e05 | 962 | { |
a649f237 PZ |
963 | struct sched_dl_entity *dl_se = &p->dl; |
964 | struct hrtimer *timer = &dl_se->dl_timer; | |
965 | struct rq *rq = task_rq(p); | |
aab03e05 | 966 | ktime_t now, act; |
aab03e05 DF |
967 | s64 delta; |
968 | ||
a649f237 PZ |
969 | lockdep_assert_held(&rq->lock); |
970 | ||
aab03e05 DF |
971 | /* |
972 | * We want the timer to fire at the deadline, but considering | |
973 | * that it is actually coming from rq->clock and not from | |
974 | * hrtimer's time base reading. | |
975 | */ | |
5ac69d37 | 976 | act = ns_to_ktime(dl_next_period(dl_se)); |
a649f237 | 977 | now = hrtimer_cb_get_time(timer); |
aab03e05 DF |
978 | delta = ktime_to_ns(now) - rq_clock(rq); |
979 | act = ktime_add_ns(act, delta); | |
980 | ||
981 | /* | |
982 | * If the expiry time already passed, e.g., because the value | |
983 | * chosen as the deadline is too small, don't even try to | |
984 | * start the timer in the past! | |
985 | */ | |
986 | if (ktime_us_delta(act, now) < 0) | |
987 | return 0; | |
988 | ||
a649f237 PZ |
989 | /* |
990 | * !enqueued will guarantee another callback; even if one is already in | |
991 | * progress. This ensures a balanced {get,put}_task_struct(). | |
992 | * | |
993 | * The race against __run_timer() clearing the enqueued state is | |
994 | * harmless because we're holding task_rq()->lock, therefore the timer | |
995 | * expiring after we've done the check will wait on its task_rq_lock() | |
996 | * and observe our state. | |
997 | */ | |
998 | if (!hrtimer_is_queued(timer)) { | |
999 | get_task_struct(p); | |
d5096aa6 | 1000 | hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD); |
a649f237 | 1001 | } |
aab03e05 | 1002 | |
cc9684d3 | 1003 | return 1; |
aab03e05 DF |
1004 | } |
1005 | ||
1006 | /* | |
1007 | * This is the bandwidth enforcement timer callback. If here, we know | |
1008 | * a task is not on its dl_rq, since the fact that the timer was running | |
1009 | * means the task is throttled and needs a runtime replenishment. | |
1010 | * | |
1011 | * However, what we actually do depends on the fact the task is active, | |
1012 | * (it is on its rq) or has been removed from there by a call to | |
1013 | * dequeue_task_dl(). In the former case we must issue the runtime | |
1014 | * replenishment and add the task back to the dl_rq; in the latter, we just | |
1015 | * do nothing but clearing dl_throttled, so that runtime and deadline | |
1016 | * updating (and the queueing back to dl_rq) will be done by the | |
1017 | * next call to enqueue_task_dl(). | |
1018 | */ | |
1019 | static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) | |
1020 | { | |
1021 | struct sched_dl_entity *dl_se = container_of(timer, | |
1022 | struct sched_dl_entity, | |
1023 | dl_timer); | |
1024 | struct task_struct *p = dl_task_of(dl_se); | |
eb580751 | 1025 | struct rq_flags rf; |
0f397f2c | 1026 | struct rq *rq; |
3960c8c0 | 1027 | |
eb580751 | 1028 | rq = task_rq_lock(p, &rf); |
0f397f2c | 1029 | |
aab03e05 | 1030 | /* |
a649f237 | 1031 | * The task might have changed its scheduling policy to something |
9846d50d | 1032 | * different than SCHED_DEADLINE (through switched_from_dl()). |
a649f237 | 1033 | */ |
209a0cbd | 1034 | if (!dl_task(p)) |
a649f237 | 1035 | goto unlock; |
a649f237 | 1036 | |
a649f237 PZ |
1037 | /* |
1038 | * The task might have been boosted by someone else and might be in the | |
1039 | * boosting/deboosting path, its not throttled. | |
1040 | */ | |
1041 | if (dl_se->dl_boosted) | |
1042 | goto unlock; | |
a79ec89f | 1043 | |
fa9c9d10 | 1044 | /* |
a649f237 PZ |
1045 | * Spurious timer due to start_dl_timer() race; or we already received |
1046 | * a replenishment from rt_mutex_setprio(). | |
fa9c9d10 | 1047 | */ |
a649f237 | 1048 | if (!dl_se->dl_throttled) |
fa9c9d10 | 1049 | goto unlock; |
a649f237 PZ |
1050 | |
1051 | sched_clock_tick(); | |
1052 | update_rq_clock(rq); | |
fa9c9d10 | 1053 | |
a79ec89f KT |
1054 | /* |
1055 | * If the throttle happened during sched-out; like: | |
1056 | * | |
1057 | * schedule() | |
1058 | * deactivate_task() | |
1059 | * dequeue_task_dl() | |
1060 | * update_curr_dl() | |
1061 | * start_dl_timer() | |
1062 | * __dequeue_task_dl() | |
1063 | * prev->on_rq = 0; | |
1064 | * | |
1065 | * We can be both throttled and !queued. Replenish the counter | |
1066 | * but do not enqueue -- wait for our wakeup to do that. | |
1067 | */ | |
1068 | if (!task_on_rq_queued(p)) { | |
1069 | replenish_dl_entity(dl_se, dl_se); | |
1070 | goto unlock; | |
1071 | } | |
1072 | ||
1baca4ce | 1073 | #ifdef CONFIG_SMP |
c0c8c9fa | 1074 | if (unlikely(!rq->online)) { |
61c7aca6 WL |
1075 | /* |
1076 | * If the runqueue is no longer available, migrate the | |
1077 | * task elsewhere. This necessarily changes rq. | |
1078 | */ | |
c0c8c9fa | 1079 | lockdep_unpin_lock(&rq->lock, rf.cookie); |
a649f237 | 1080 | rq = dl_task_offline_migration(rq, p); |
c0c8c9fa | 1081 | rf.cookie = lockdep_pin_lock(&rq->lock); |
dcc3b5ff | 1082 | update_rq_clock(rq); |
61c7aca6 WL |
1083 | |
1084 | /* | |
1085 | * Now that the task has been migrated to the new RQ and we | |
1086 | * have that locked, proceed as normal and enqueue the task | |
1087 | * there. | |
1088 | */ | |
c0c8c9fa | 1089 | } |
61c7aca6 | 1090 | #endif |
a649f237 | 1091 | |
61c7aca6 WL |
1092 | enqueue_task_dl(rq, p, ENQUEUE_REPLENISH); |
1093 | if (dl_task(rq->curr)) | |
1094 | check_preempt_curr_dl(rq, p, 0); | |
1095 | else | |
1096 | resched_curr(rq); | |
a649f237 | 1097 | |
61c7aca6 | 1098 | #ifdef CONFIG_SMP |
a649f237 PZ |
1099 | /* |
1100 | * Queueing this task back might have overloaded rq, check if we need | |
1101 | * to kick someone away. | |
1019a359 | 1102 | */ |
0aaafaab PZ |
1103 | if (has_pushable_dl_tasks(rq)) { |
1104 | /* | |
1105 | * Nothing relies on rq->lock after this, so its safe to drop | |
1106 | * rq->lock. | |
1107 | */ | |
d8ac8971 | 1108 | rq_unpin_lock(rq, &rf); |
1019a359 | 1109 | push_dl_task(rq); |
d8ac8971 | 1110 | rq_repin_lock(rq, &rf); |
0aaafaab | 1111 | } |
1baca4ce | 1112 | #endif |
a649f237 | 1113 | |
aab03e05 | 1114 | unlock: |
eb580751 | 1115 | task_rq_unlock(rq, p, &rf); |
aab03e05 | 1116 | |
a649f237 PZ |
1117 | /* |
1118 | * This can free the task_struct, including this hrtimer, do not touch | |
1119 | * anything related to that after this. | |
1120 | */ | |
1121 | put_task_struct(p); | |
1122 | ||
aab03e05 DF |
1123 | return HRTIMER_NORESTART; |
1124 | } | |
1125 | ||
1126 | void init_dl_task_timer(struct sched_dl_entity *dl_se) | |
1127 | { | |
1128 | struct hrtimer *timer = &dl_se->dl_timer; | |
1129 | ||
d5096aa6 | 1130 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD); |
aab03e05 DF |
1131 | timer->function = dl_task_timer; |
1132 | } | |
1133 | ||
df8eac8c DBO |
1134 | /* |
1135 | * During the activation, CBS checks if it can reuse the current task's | |
1136 | * runtime and period. If the deadline of the task is in the past, CBS | |
1137 | * cannot use the runtime, and so it replenishes the task. This rule | |
1138 | * works fine for implicit deadline tasks (deadline == period), and the | |
1139 | * CBS was designed for implicit deadline tasks. However, a task with | |
1140 | * constrained deadline (deadine < period) might be awakened after the | |
1141 | * deadline, but before the next period. In this case, replenishing the | |
1142 | * task would allow it to run for runtime / deadline. As in this case | |
1143 | * deadline < period, CBS enables a task to run for more than the | |
1144 | * runtime / period. In a very loaded system, this can cause a domino | |
1145 | * effect, making other tasks miss their deadlines. | |
1146 | * | |
1147 | * To avoid this problem, in the activation of a constrained deadline | |
1148 | * task after the deadline but before the next period, throttle the | |
1149 | * task and set the replenishing timer to the begin of the next period, | |
1150 | * unless it is boosted. | |
1151 | */ | |
1152 | static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se) | |
1153 | { | |
1154 | struct task_struct *p = dl_task_of(dl_se); | |
1155 | struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se)); | |
1156 | ||
1157 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) && | |
1158 | dl_time_before(rq_clock(rq), dl_next_period(dl_se))) { | |
1159 | if (unlikely(dl_se->dl_boosted || !start_dl_timer(p))) | |
1160 | return; | |
1161 | dl_se->dl_throttled = 1; | |
ae83b56a XP |
1162 | if (dl_se->runtime > 0) |
1163 | dl_se->runtime = 0; | |
df8eac8c DBO |
1164 | } |
1165 | } | |
1166 | ||
aab03e05 | 1167 | static |
6fab5410 | 1168 | int dl_runtime_exceeded(struct sched_dl_entity *dl_se) |
aab03e05 | 1169 | { |
269ad801 | 1170 | return (dl_se->runtime <= 0); |
aab03e05 DF |
1171 | } |
1172 | ||
faa59937 JL |
1173 | extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq); |
1174 | ||
c52f14d3 LA |
1175 | /* |
1176 | * This function implements the GRUB accounting rule: | |
1177 | * according to the GRUB reclaiming algorithm, the runtime is | |
daec5798 LA |
1178 | * not decreased as "dq = -dt", but as |
1179 | * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt", | |
1180 | * where u is the utilization of the task, Umax is the maximum reclaimable | |
1181 | * utilization, Uinact is the (per-runqueue) inactive utilization, computed | |
1182 | * as the difference between the "total runqueue utilization" and the | |
1183 | * runqueue active utilization, and Uextra is the (per runqueue) extra | |
1184 | * reclaimable utilization. | |
9f0d1a50 | 1185 | * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations |
daec5798 LA |
1186 | * multiplied by 2^BW_SHIFT, the result has to be shifted right by |
1187 | * BW_SHIFT. | |
1188 | * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT, | |
1189 | * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT. | |
1190 | * Since delta is a 64 bit variable, to have an overflow its value | |
1191 | * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds. | |
1192 | * So, overflow is not an issue here. | |
c52f14d3 | 1193 | */ |
3febfc8a | 1194 | static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se) |
c52f14d3 | 1195 | { |
9f0d1a50 LA |
1196 | u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */ |
1197 | u64 u_act; | |
daec5798 | 1198 | u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT; |
c52f14d3 | 1199 | |
9f0d1a50 | 1200 | /* |
daec5798 LA |
1201 | * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)}, |
1202 | * we compare u_inact + rq->dl.extra_bw with | |
1203 | * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because | |
1204 | * u_inact + rq->dl.extra_bw can be larger than | |
1205 | * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative | |
1206 | * leading to wrong results) | |
9f0d1a50 | 1207 | */ |
daec5798 LA |
1208 | if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min) |
1209 | u_act = u_act_min; | |
9f0d1a50 | 1210 | else |
daec5798 | 1211 | u_act = BW_UNIT - u_inact - rq->dl.extra_bw; |
9f0d1a50 LA |
1212 | |
1213 | return (delta * u_act) >> BW_SHIFT; | |
c52f14d3 LA |
1214 | } |
1215 | ||
aab03e05 DF |
1216 | /* |
1217 | * Update the current task's runtime statistics (provided it is still | |
1218 | * a -deadline task and has not been removed from the dl_rq). | |
1219 | */ | |
1220 | static void update_curr_dl(struct rq *rq) | |
1221 | { | |
1222 | struct task_struct *curr = rq->curr; | |
1223 | struct sched_dl_entity *dl_se = &curr->dl; | |
07881166 JL |
1224 | u64 delta_exec, scaled_delta_exec; |
1225 | int cpu = cpu_of(rq); | |
6fe0ce1e | 1226 | u64 now; |
aab03e05 DF |
1227 | |
1228 | if (!dl_task(curr) || !on_dl_rq(dl_se)) | |
1229 | return; | |
1230 | ||
1231 | /* | |
1232 | * Consumed budget is computed considering the time as | |
1233 | * observed by schedulable tasks (excluding time spent | |
1234 | * in hardirq context, etc.). Deadlines are instead | |
1235 | * computed using hard walltime. This seems to be the more | |
1236 | * natural solution, but the full ramifications of this | |
1237 | * approach need further study. | |
1238 | */ | |
6fe0ce1e WY |
1239 | now = rq_clock_task(rq); |
1240 | delta_exec = now - curr->se.exec_start; | |
48be3a67 PZ |
1241 | if (unlikely((s64)delta_exec <= 0)) { |
1242 | if (unlikely(dl_se->dl_yielded)) | |
1243 | goto throttle; | |
734ff2a7 | 1244 | return; |
48be3a67 | 1245 | } |
aab03e05 DF |
1246 | |
1247 | schedstat_set(curr->se.statistics.exec_max, | |
1248 | max(curr->se.statistics.exec_max, delta_exec)); | |
1249 | ||
1250 | curr->se.sum_exec_runtime += delta_exec; | |
1251 | account_group_exec_runtime(curr, delta_exec); | |
1252 | ||
6fe0ce1e | 1253 | curr->se.exec_start = now; |
d2cc5ed6 | 1254 | cgroup_account_cputime(curr, delta_exec); |
aab03e05 | 1255 | |
794a56eb JL |
1256 | if (dl_entity_is_special(dl_se)) |
1257 | return; | |
1258 | ||
07881166 JL |
1259 | /* |
1260 | * For tasks that participate in GRUB, we implement GRUB-PA: the | |
1261 | * spare reclaimed bandwidth is used to clock down frequency. | |
1262 | * | |
1263 | * For the others, we still need to scale reservation parameters | |
1264 | * according to current frequency and CPU maximum capacity. | |
1265 | */ | |
1266 | if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) { | |
1267 | scaled_delta_exec = grub_reclaim(delta_exec, | |
1268 | rq, | |
1269 | &curr->dl); | |
1270 | } else { | |
1271 | unsigned long scale_freq = arch_scale_freq_capacity(cpu); | |
8ec59c0f | 1272 | unsigned long scale_cpu = arch_scale_cpu_capacity(cpu); |
07881166 JL |
1273 | |
1274 | scaled_delta_exec = cap_scale(delta_exec, scale_freq); | |
1275 | scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu); | |
1276 | } | |
1277 | ||
1278 | dl_se->runtime -= scaled_delta_exec; | |
48be3a67 PZ |
1279 | |
1280 | throttle: | |
1281 | if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) { | |
1019a359 | 1282 | dl_se->dl_throttled = 1; |
34be3930 JL |
1283 | |
1284 | /* If requested, inform the user about runtime overruns. */ | |
1285 | if (dl_runtime_exceeded(dl_se) && | |
1286 | (dl_se->flags & SCHED_FLAG_DL_OVERRUN)) | |
1287 | dl_se->dl_overrun = 1; | |
1288 | ||
aab03e05 | 1289 | __dequeue_task_dl(rq, curr, 0); |
a649f237 | 1290 | if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr))) |
aab03e05 DF |
1291 | enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH); |
1292 | ||
1293 | if (!is_leftmost(curr, &rq->dl)) | |
8875125e | 1294 | resched_curr(rq); |
aab03e05 | 1295 | } |
1724813d PZ |
1296 | |
1297 | /* | |
1298 | * Because -- for now -- we share the rt bandwidth, we need to | |
1299 | * account our runtime there too, otherwise actual rt tasks | |
1300 | * would be able to exceed the shared quota. | |
1301 | * | |
1302 | * Account to the root rt group for now. | |
1303 | * | |
1304 | * The solution we're working towards is having the RT groups scheduled | |
1305 | * using deadline servers -- however there's a few nasties to figure | |
1306 | * out before that can happen. | |
1307 | */ | |
1308 | if (rt_bandwidth_enabled()) { | |
1309 | struct rt_rq *rt_rq = &rq->rt; | |
1310 | ||
1311 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
1724813d PZ |
1312 | /* |
1313 | * We'll let actual RT tasks worry about the overflow here, we | |
faa59937 JL |
1314 | * have our own CBS to keep us inline; only account when RT |
1315 | * bandwidth is relevant. | |
1724813d | 1316 | */ |
faa59937 JL |
1317 | if (sched_rt_bandwidth_account(rt_rq)) |
1318 | rt_rq->rt_time += delta_exec; | |
1724813d PZ |
1319 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
1320 | } | |
aab03e05 DF |
1321 | } |
1322 | ||
209a0cbd LA |
1323 | static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer) |
1324 | { | |
1325 | struct sched_dl_entity *dl_se = container_of(timer, | |
1326 | struct sched_dl_entity, | |
1327 | inactive_timer); | |
1328 | struct task_struct *p = dl_task_of(dl_se); | |
1329 | struct rq_flags rf; | |
1330 | struct rq *rq; | |
1331 | ||
1332 | rq = task_rq_lock(p, &rf); | |
1333 | ||
ecda2b66 JL |
1334 | sched_clock_tick(); |
1335 | update_rq_clock(rq); | |
1336 | ||
209a0cbd | 1337 | if (!dl_task(p) || p->state == TASK_DEAD) { |
387e3130 LA |
1338 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); |
1339 | ||
209a0cbd | 1340 | if (p->state == TASK_DEAD && dl_se->dl_non_contending) { |
794a56eb JL |
1341 | sub_running_bw(&p->dl, dl_rq_of_se(&p->dl)); |
1342 | sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl)); | |
209a0cbd LA |
1343 | dl_se->dl_non_contending = 0; |
1344 | } | |
387e3130 LA |
1345 | |
1346 | raw_spin_lock(&dl_b->lock); | |
8c0944ce | 1347 | __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); |
387e3130 | 1348 | raw_spin_unlock(&dl_b->lock); |
209a0cbd LA |
1349 | __dl_clear_params(p); |
1350 | ||
1351 | goto unlock; | |
1352 | } | |
1353 | if (dl_se->dl_non_contending == 0) | |
1354 | goto unlock; | |
1355 | ||
794a56eb | 1356 | sub_running_bw(dl_se, &rq->dl); |
209a0cbd LA |
1357 | dl_se->dl_non_contending = 0; |
1358 | unlock: | |
1359 | task_rq_unlock(rq, p, &rf); | |
1360 | put_task_struct(p); | |
1361 | ||
1362 | return HRTIMER_NORESTART; | |
1363 | } | |
1364 | ||
1365 | void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se) | |
1366 | { | |
1367 | struct hrtimer *timer = &dl_se->inactive_timer; | |
1368 | ||
850377a8 | 1369 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD); |
209a0cbd LA |
1370 | timer->function = inactive_task_timer; |
1371 | } | |
1372 | ||
1baca4ce JL |
1373 | #ifdef CONFIG_SMP |
1374 | ||
1baca4ce JL |
1375 | static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) |
1376 | { | |
1377 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
1378 | ||
1379 | if (dl_rq->earliest_dl.curr == 0 || | |
1380 | dl_time_before(deadline, dl_rq->earliest_dl.curr)) { | |
1baca4ce | 1381 | dl_rq->earliest_dl.curr = deadline; |
d8206bb3 | 1382 | cpudl_set(&rq->rd->cpudl, rq->cpu, deadline); |
1baca4ce JL |
1383 | } |
1384 | } | |
1385 | ||
1386 | static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) | |
1387 | { | |
1388 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
1389 | ||
1390 | /* | |
1391 | * Since we may have removed our earliest (and/or next earliest) | |
1392 | * task we must recompute them. | |
1393 | */ | |
1394 | if (!dl_rq->dl_nr_running) { | |
1395 | dl_rq->earliest_dl.curr = 0; | |
1396 | dl_rq->earliest_dl.next = 0; | |
d8206bb3 | 1397 | cpudl_clear(&rq->rd->cpudl, rq->cpu); |
1baca4ce | 1398 | } else { |
2161573e | 1399 | struct rb_node *leftmost = dl_rq->root.rb_leftmost; |
1baca4ce JL |
1400 | struct sched_dl_entity *entry; |
1401 | ||
1402 | entry = rb_entry(leftmost, struct sched_dl_entity, rb_node); | |
1403 | dl_rq->earliest_dl.curr = entry->deadline; | |
d8206bb3 | 1404 | cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline); |
1baca4ce JL |
1405 | } |
1406 | } | |
1407 | ||
1408 | #else | |
1409 | ||
1410 | static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
1411 | static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
1412 | ||
1413 | #endif /* CONFIG_SMP */ | |
1414 | ||
1415 | static inline | |
1416 | void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
1417 | { | |
1418 | int prio = dl_task_of(dl_se)->prio; | |
1419 | u64 deadline = dl_se->deadline; | |
1420 | ||
1421 | WARN_ON(!dl_prio(prio)); | |
1422 | dl_rq->dl_nr_running++; | |
72465447 | 1423 | add_nr_running(rq_of_dl_rq(dl_rq), 1); |
1baca4ce JL |
1424 | |
1425 | inc_dl_deadline(dl_rq, deadline); | |
1426 | inc_dl_migration(dl_se, dl_rq); | |
1427 | } | |
1428 | ||
1429 | static inline | |
1430 | void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
1431 | { | |
1432 | int prio = dl_task_of(dl_se)->prio; | |
1433 | ||
1434 | WARN_ON(!dl_prio(prio)); | |
1435 | WARN_ON(!dl_rq->dl_nr_running); | |
1436 | dl_rq->dl_nr_running--; | |
72465447 | 1437 | sub_nr_running(rq_of_dl_rq(dl_rq), 1); |
1baca4ce JL |
1438 | |
1439 | dec_dl_deadline(dl_rq, dl_se->deadline); | |
1440 | dec_dl_migration(dl_se, dl_rq); | |
1441 | } | |
1442 | ||
aab03e05 DF |
1443 | static void __enqueue_dl_entity(struct sched_dl_entity *dl_se) |
1444 | { | |
1445 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
2161573e | 1446 | struct rb_node **link = &dl_rq->root.rb_root.rb_node; |
aab03e05 DF |
1447 | struct rb_node *parent = NULL; |
1448 | struct sched_dl_entity *entry; | |
1449 | int leftmost = 1; | |
1450 | ||
1451 | BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node)); | |
1452 | ||
1453 | while (*link) { | |
1454 | parent = *link; | |
1455 | entry = rb_entry(parent, struct sched_dl_entity, rb_node); | |
1456 | if (dl_time_before(dl_se->deadline, entry->deadline)) | |
1457 | link = &parent->rb_left; | |
1458 | else { | |
1459 | link = &parent->rb_right; | |
1460 | leftmost = 0; | |
1461 | } | |
1462 | } | |
1463 | ||
aab03e05 | 1464 | rb_link_node(&dl_se->rb_node, parent, link); |
2161573e | 1465 | rb_insert_color_cached(&dl_se->rb_node, &dl_rq->root, leftmost); |
aab03e05 | 1466 | |
1baca4ce | 1467 | inc_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
1468 | } |
1469 | ||
1470 | static void __dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
1471 | { | |
1472 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
1473 | ||
1474 | if (RB_EMPTY_NODE(&dl_se->rb_node)) | |
1475 | return; | |
1476 | ||
2161573e | 1477 | rb_erase_cached(&dl_se->rb_node, &dl_rq->root); |
aab03e05 DF |
1478 | RB_CLEAR_NODE(&dl_se->rb_node); |
1479 | ||
1baca4ce | 1480 | dec_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
1481 | } |
1482 | ||
1483 | static void | |
2d3d891d DF |
1484 | enqueue_dl_entity(struct sched_dl_entity *dl_se, |
1485 | struct sched_dl_entity *pi_se, int flags) | |
aab03e05 DF |
1486 | { |
1487 | BUG_ON(on_dl_rq(dl_se)); | |
1488 | ||
1489 | /* | |
1490 | * If this is a wakeup or a new instance, the scheduling | |
1491 | * parameters of the task might need updating. Otherwise, | |
1492 | * we want a replenishment of its runtime. | |
1493 | */ | |
e36d8677 | 1494 | if (flags & ENQUEUE_WAKEUP) { |
8fd27231 | 1495 | task_contending(dl_se, flags); |
2d3d891d | 1496 | update_dl_entity(dl_se, pi_se); |
e36d8677 | 1497 | } else if (flags & ENQUEUE_REPLENISH) { |
6a503c3b | 1498 | replenish_dl_entity(dl_se, pi_se); |
295d6d5e LA |
1499 | } else if ((flags & ENQUEUE_RESTORE) && |
1500 | dl_time_before(dl_se->deadline, | |
1501 | rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) { | |
1502 | setup_new_dl_entity(dl_se); | |
e36d8677 | 1503 | } |
aab03e05 DF |
1504 | |
1505 | __enqueue_dl_entity(dl_se); | |
1506 | } | |
1507 | ||
1508 | static void dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
1509 | { | |
1510 | __dequeue_dl_entity(dl_se); | |
1511 | } | |
1512 | ||
1513 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
1514 | { | |
2d3d891d DF |
1515 | struct task_struct *pi_task = rt_mutex_get_top_task(p); |
1516 | struct sched_dl_entity *pi_se = &p->dl; | |
1517 | ||
1518 | /* | |
193be41e JF |
1519 | * Use the scheduling parameters of the top pi-waiter task if: |
1520 | * - we have a top pi-waiter which is a SCHED_DEADLINE task AND | |
1521 | * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is | |
1522 | * smaller than our deadline OR we are a !SCHED_DEADLINE task getting | |
1523 | * boosted due to a SCHED_DEADLINE pi-waiter). | |
1524 | * Otherwise we keep our runtime and deadline. | |
2d3d891d | 1525 | */ |
193be41e | 1526 | if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) { |
2d3d891d | 1527 | pi_se = &pi_task->dl; |
64be6f1f JL |
1528 | } else if (!dl_prio(p->normal_prio)) { |
1529 | /* | |
1530 | * Special case in which we have a !SCHED_DEADLINE task | |
193be41e | 1531 | * that is going to be deboosted, but exceeds its |
64be6f1f JL |
1532 | * runtime while doing so. No point in replenishing |
1533 | * it, as it's going to return back to its original | |
1534 | * scheduling class after this. | |
1535 | */ | |
1536 | BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH); | |
1537 | return; | |
1538 | } | |
2d3d891d | 1539 | |
df8eac8c DBO |
1540 | /* |
1541 | * Check if a constrained deadline task was activated | |
1542 | * after the deadline but before the next period. | |
1543 | * If that is the case, the task will be throttled and | |
1544 | * the replenishment timer will be set to the next period. | |
1545 | */ | |
3effcb42 | 1546 | if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl)) |
df8eac8c DBO |
1547 | dl_check_constrained_dl(&p->dl); |
1548 | ||
8fd27231 | 1549 | if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) { |
794a56eb JL |
1550 | add_rq_bw(&p->dl, &rq->dl); |
1551 | add_running_bw(&p->dl, &rq->dl); | |
8fd27231 | 1552 | } |
e36d8677 | 1553 | |
aab03e05 | 1554 | /* |
e36d8677 | 1555 | * If p is throttled, we do not enqueue it. In fact, if it exhausted |
aab03e05 DF |
1556 | * its budget it needs a replenishment and, since it now is on |
1557 | * its rq, the bandwidth timer callback (which clearly has not | |
1558 | * run yet) will take care of this. | |
e36d8677 LA |
1559 | * However, the active utilization does not depend on the fact |
1560 | * that the task is on the runqueue or not (but depends on the | |
1561 | * task's state - in GRUB parlance, "inactive" vs "active contending"). | |
1562 | * In other words, even if a task is throttled its utilization must | |
1563 | * be counted in the active utilization; hence, we need to call | |
1564 | * add_running_bw(). | |
aab03e05 | 1565 | */ |
e36d8677 | 1566 | if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) { |
209a0cbd | 1567 | if (flags & ENQUEUE_WAKEUP) |
8fd27231 | 1568 | task_contending(&p->dl, flags); |
209a0cbd | 1569 | |
aab03e05 | 1570 | return; |
e36d8677 | 1571 | } |
aab03e05 | 1572 | |
2d3d891d | 1573 | enqueue_dl_entity(&p->dl, pi_se, flags); |
1baca4ce | 1574 | |
4b53a341 | 1575 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) |
1baca4ce | 1576 | enqueue_pushable_dl_task(rq, p); |
aab03e05 DF |
1577 | } |
1578 | ||
1579 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
1580 | { | |
1581 | dequeue_dl_entity(&p->dl); | |
1baca4ce | 1582 | dequeue_pushable_dl_task(rq, p); |
aab03e05 DF |
1583 | } |
1584 | ||
1585 | static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
1586 | { | |
1587 | update_curr_dl(rq); | |
1588 | __dequeue_task_dl(rq, p, flags); | |
e36d8677 | 1589 | |
8fd27231 | 1590 | if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) { |
794a56eb JL |
1591 | sub_running_bw(&p->dl, &rq->dl); |
1592 | sub_rq_bw(&p->dl, &rq->dl); | |
8fd27231 | 1593 | } |
e36d8677 LA |
1594 | |
1595 | /* | |
209a0cbd LA |
1596 | * This check allows to start the inactive timer (or to immediately |
1597 | * decrease the active utilization, if needed) in two cases: | |
e36d8677 LA |
1598 | * when the task blocks and when it is terminating |
1599 | * (p->state == TASK_DEAD). We can handle the two cases in the same | |
1600 | * way, because from GRUB's point of view the same thing is happening | |
1601 | * (the task moves from "active contending" to "active non contending" | |
1602 | * or "inactive") | |
1603 | */ | |
1604 | if (flags & DEQUEUE_SLEEP) | |
209a0cbd | 1605 | task_non_contending(p); |
aab03e05 DF |
1606 | } |
1607 | ||
1608 | /* | |
1609 | * Yield task semantic for -deadline tasks is: | |
1610 | * | |
1611 | * get off from the CPU until our next instance, with | |
1612 | * a new runtime. This is of little use now, since we | |
1613 | * don't have a bandwidth reclaiming mechanism. Anyway, | |
1614 | * bandwidth reclaiming is planned for the future, and | |
1615 | * yield_task_dl will indicate that some spare budget | |
1616 | * is available for other task instances to use it. | |
1617 | */ | |
1618 | static void yield_task_dl(struct rq *rq) | |
1619 | { | |
aab03e05 DF |
1620 | /* |
1621 | * We make the task go to sleep until its current deadline by | |
1622 | * forcing its runtime to zero. This way, update_curr_dl() stops | |
1623 | * it and the bandwidth timer will wake it up and will give it | |
5bfd126e | 1624 | * new scheduling parameters (thanks to dl_yielded=1). |
aab03e05 | 1625 | */ |
48be3a67 PZ |
1626 | rq->curr->dl.dl_yielded = 1; |
1627 | ||
6f1607f1 | 1628 | update_rq_clock(rq); |
aab03e05 | 1629 | update_curr_dl(rq); |
44fb085b WL |
1630 | /* |
1631 | * Tell update_rq_clock() that we've just updated, | |
1632 | * so we don't do microscopic update in schedule() | |
1633 | * and double the fastpath cost. | |
1634 | */ | |
adcc8da8 | 1635 | rq_clock_skip_update(rq); |
aab03e05 DF |
1636 | } |
1637 | ||
1baca4ce JL |
1638 | #ifdef CONFIG_SMP |
1639 | ||
1640 | static int find_later_rq(struct task_struct *task); | |
1baca4ce JL |
1641 | |
1642 | static int | |
1643 | select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags) | |
1644 | { | |
1645 | struct task_struct *curr; | |
b4118988 | 1646 | bool select_rq; |
1baca4ce JL |
1647 | struct rq *rq; |
1648 | ||
1d7e974c | 1649 | if (sd_flag != SD_BALANCE_WAKE) |
1baca4ce JL |
1650 | goto out; |
1651 | ||
1652 | rq = cpu_rq(cpu); | |
1653 | ||
1654 | rcu_read_lock(); | |
316c1608 | 1655 | curr = READ_ONCE(rq->curr); /* unlocked access */ |
1baca4ce JL |
1656 | |
1657 | /* | |
1658 | * If we are dealing with a -deadline task, we must | |
1659 | * decide where to wake it up. | |
1660 | * If it has a later deadline and the current task | |
1661 | * on this rq can't move (provided the waking task | |
1662 | * can!) we prefer to send it somewhere else. On the | |
1663 | * other hand, if it has a shorter deadline, we | |
1664 | * try to make it stay here, it might be important. | |
1665 | */ | |
b4118988 LA |
1666 | select_rq = unlikely(dl_task(curr)) && |
1667 | (curr->nr_cpus_allowed < 2 || | |
1668 | !dl_entity_preempt(&p->dl, &curr->dl)) && | |
1669 | p->nr_cpus_allowed > 1; | |
1670 | ||
1671 | /* | |
1672 | * Take the capacity of the CPU into account to | |
1673 | * ensure it fits the requirement of the task. | |
1674 | */ | |
1675 | if (static_branch_unlikely(&sched_asym_cpucapacity)) | |
1676 | select_rq |= !dl_task_fits_capacity(p, cpu); | |
1677 | ||
1678 | if (select_rq) { | |
1baca4ce JL |
1679 | int target = find_later_rq(p); |
1680 | ||
9d514262 | 1681 | if (target != -1 && |
5aa50507 LA |
1682 | (dl_time_before(p->dl.deadline, |
1683 | cpu_rq(target)->dl.earliest_dl.curr) || | |
1684 | (cpu_rq(target)->dl.dl_nr_running == 0))) | |
1baca4ce JL |
1685 | cpu = target; |
1686 | } | |
1687 | rcu_read_unlock(); | |
1688 | ||
1689 | out: | |
1690 | return cpu; | |
1691 | } | |
1692 | ||
1327237a | 1693 | static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused) |
209a0cbd LA |
1694 | { |
1695 | struct rq *rq; | |
1696 | ||
8fd27231 | 1697 | if (p->state != TASK_WAKING) |
209a0cbd LA |
1698 | return; |
1699 | ||
1700 | rq = task_rq(p); | |
1701 | /* | |
1702 | * Since p->state == TASK_WAKING, set_task_cpu() has been called | |
1703 | * from try_to_wake_up(). Hence, p->pi_lock is locked, but | |
1704 | * rq->lock is not... So, lock it | |
1705 | */ | |
1706 | raw_spin_lock(&rq->lock); | |
8fd27231 | 1707 | if (p->dl.dl_non_contending) { |
794a56eb | 1708 | sub_running_bw(&p->dl, &rq->dl); |
8fd27231 LA |
1709 | p->dl.dl_non_contending = 0; |
1710 | /* | |
1711 | * If the timer handler is currently running and the | |
1712 | * timer cannot be cancelled, inactive_task_timer() | |
1713 | * will see that dl_not_contending is not set, and | |
1714 | * will not touch the rq's active utilization, | |
1715 | * so we are still safe. | |
1716 | */ | |
1717 | if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) | |
1718 | put_task_struct(p); | |
1719 | } | |
794a56eb | 1720 | sub_rq_bw(&p->dl, &rq->dl); |
209a0cbd LA |
1721 | raw_spin_unlock(&rq->lock); |
1722 | } | |
1723 | ||
1baca4ce JL |
1724 | static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p) |
1725 | { | |
1726 | /* | |
1727 | * Current can't be migrated, useless to reschedule, | |
1728 | * let's hope p can move out. | |
1729 | */ | |
4b53a341 | 1730 | if (rq->curr->nr_cpus_allowed == 1 || |
3261ed0b | 1731 | !cpudl_find(&rq->rd->cpudl, rq->curr, NULL)) |
1baca4ce JL |
1732 | return; |
1733 | ||
1734 | /* | |
1735 | * p is migratable, so let's not schedule it and | |
1736 | * see if it is pushed or pulled somewhere else. | |
1737 | */ | |
4b53a341 | 1738 | if (p->nr_cpus_allowed != 1 && |
3261ed0b | 1739 | cpudl_find(&rq->rd->cpudl, p, NULL)) |
1baca4ce JL |
1740 | return; |
1741 | ||
8875125e | 1742 | resched_curr(rq); |
1baca4ce JL |
1743 | } |
1744 | ||
6e2df058 PZ |
1745 | static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf) |
1746 | { | |
1747 | if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) { | |
1748 | /* | |
1749 | * This is OK, because current is on_cpu, which avoids it being | |
1750 | * picked for load-balance and preemption/IRQs are still | |
1751 | * disabled avoiding further scheduler activity on it and we've | |
1752 | * not yet started the picking loop. | |
1753 | */ | |
1754 | rq_unpin_lock(rq, rf); | |
1755 | pull_dl_task(rq); | |
1756 | rq_repin_lock(rq, rf); | |
1757 | } | |
1758 | ||
1759 | return sched_stop_runnable(rq) || sched_dl_runnable(rq); | |
1760 | } | |
1baca4ce JL |
1761 | #endif /* CONFIG_SMP */ |
1762 | ||
aab03e05 DF |
1763 | /* |
1764 | * Only called when both the current and waking task are -deadline | |
1765 | * tasks. | |
1766 | */ | |
1767 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | |
1768 | int flags) | |
1769 | { | |
1baca4ce | 1770 | if (dl_entity_preempt(&p->dl, &rq->curr->dl)) { |
8875125e | 1771 | resched_curr(rq); |
1baca4ce JL |
1772 | return; |
1773 | } | |
1774 | ||
1775 | #ifdef CONFIG_SMP | |
1776 | /* | |
1777 | * In the unlikely case current and p have the same deadline | |
1778 | * let us try to decide what's the best thing to do... | |
1779 | */ | |
332ac17e DF |
1780 | if ((p->dl.deadline == rq->curr->dl.deadline) && |
1781 | !test_tsk_need_resched(rq->curr)) | |
1baca4ce JL |
1782 | check_preempt_equal_dl(rq, p); |
1783 | #endif /* CONFIG_SMP */ | |
aab03e05 DF |
1784 | } |
1785 | ||
1786 | #ifdef CONFIG_SCHED_HRTICK | |
1787 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
1788 | { | |
177ef2a6 | 1789 | hrtick_start(rq, p->dl.runtime); |
aab03e05 | 1790 | } |
36ce9881 WL |
1791 | #else /* !CONFIG_SCHED_HRTICK */ |
1792 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
1793 | { | |
1794 | } | |
aab03e05 DF |
1795 | #endif |
1796 | ||
a0e813f2 | 1797 | static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first) |
ff1cdc94 MS |
1798 | { |
1799 | p->se.exec_start = rq_clock_task(rq); | |
1800 | ||
1801 | /* You can't push away the running task */ | |
1802 | dequeue_pushable_dl_task(rq, p); | |
f95d4eae | 1803 | |
a0e813f2 PZ |
1804 | if (!first) |
1805 | return; | |
1806 | ||
f95d4eae PZ |
1807 | if (hrtick_enabled(rq)) |
1808 | start_hrtick_dl(rq, p); | |
1809 | ||
1810 | if (rq->curr->sched_class != &dl_sched_class) | |
1811 | update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0); | |
1812 | ||
1813 | deadline_queue_push_tasks(rq); | |
ff1cdc94 MS |
1814 | } |
1815 | ||
aab03e05 DF |
1816 | static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq, |
1817 | struct dl_rq *dl_rq) | |
1818 | { | |
2161573e | 1819 | struct rb_node *left = rb_first_cached(&dl_rq->root); |
aab03e05 DF |
1820 | |
1821 | if (!left) | |
1822 | return NULL; | |
1823 | ||
1824 | return rb_entry(left, struct sched_dl_entity, rb_node); | |
1825 | } | |
1826 | ||
98c2f700 | 1827 | static struct task_struct *pick_next_task_dl(struct rq *rq) |
aab03e05 DF |
1828 | { |
1829 | struct sched_dl_entity *dl_se; | |
6e2df058 | 1830 | struct dl_rq *dl_rq = &rq->dl; |
aab03e05 | 1831 | struct task_struct *p; |
aab03e05 | 1832 | |
6e2df058 | 1833 | if (!sched_dl_runnable(rq)) |
aab03e05 DF |
1834 | return NULL; |
1835 | ||
1836 | dl_se = pick_next_dl_entity(rq, dl_rq); | |
1837 | BUG_ON(!dl_se); | |
aab03e05 | 1838 | p = dl_task_of(dl_se); |
a0e813f2 | 1839 | set_next_task_dl(rq, p, true); |
aab03e05 DF |
1840 | return p; |
1841 | } | |
1842 | ||
6e2df058 | 1843 | static void put_prev_task_dl(struct rq *rq, struct task_struct *p) |
aab03e05 DF |
1844 | { |
1845 | update_curr_dl(rq); | |
1baca4ce | 1846 | |
23127296 | 1847 | update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1); |
4b53a341 | 1848 | if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1) |
1baca4ce | 1849 | enqueue_pushable_dl_task(rq, p); |
aab03e05 DF |
1850 | } |
1851 | ||
d84b3131 FW |
1852 | /* |
1853 | * scheduler tick hitting a task of our scheduling class. | |
1854 | * | |
1855 | * NOTE: This function can be called remotely by the tick offload that | |
1856 | * goes along full dynticks. Therefore no local assumption can be made | |
1857 | * and everything must be accessed through the @rq and @curr passed in | |
1858 | * parameters. | |
1859 | */ | |
aab03e05 DF |
1860 | static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued) |
1861 | { | |
1862 | update_curr_dl(rq); | |
1863 | ||
23127296 | 1864 | update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1); |
a7bebf48 WL |
1865 | /* |
1866 | * Even when we have runtime, update_curr_dl() might have resulted in us | |
1867 | * not being the leftmost task anymore. In that case NEED_RESCHED will | |
1868 | * be set and schedule() will start a new hrtick for the next task. | |
1869 | */ | |
1870 | if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 && | |
1871 | is_leftmost(p, &rq->dl)) | |
aab03e05 | 1872 | start_hrtick_dl(rq, p); |
aab03e05 DF |
1873 | } |
1874 | ||
1875 | static void task_fork_dl(struct task_struct *p) | |
1876 | { | |
1877 | /* | |
1878 | * SCHED_DEADLINE tasks cannot fork and this is achieved through | |
1879 | * sched_fork() | |
1880 | */ | |
1881 | } | |
1882 | ||
1baca4ce JL |
1883 | #ifdef CONFIG_SMP |
1884 | ||
1885 | /* Only try algorithms three times */ | |
1886 | #define DL_MAX_TRIES 3 | |
1887 | ||
1888 | static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu) | |
1889 | { | |
1890 | if (!task_running(rq, p) && | |
3bd37062 | 1891 | cpumask_test_cpu(cpu, p->cpus_ptr)) |
1baca4ce | 1892 | return 1; |
1baca4ce JL |
1893 | return 0; |
1894 | } | |
1895 | ||
8b5e770e WL |
1896 | /* |
1897 | * Return the earliest pushable rq's task, which is suitable to be executed | |
1898 | * on the CPU, NULL otherwise: | |
1899 | */ | |
1900 | static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu) | |
1901 | { | |
2161573e | 1902 | struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost; |
8b5e770e WL |
1903 | struct task_struct *p = NULL; |
1904 | ||
1905 | if (!has_pushable_dl_tasks(rq)) | |
1906 | return NULL; | |
1907 | ||
1908 | next_node: | |
1909 | if (next_node) { | |
1910 | p = rb_entry(next_node, struct task_struct, pushable_dl_tasks); | |
1911 | ||
1912 | if (pick_dl_task(rq, p, cpu)) | |
1913 | return p; | |
1914 | ||
1915 | next_node = rb_next(next_node); | |
1916 | goto next_node; | |
1917 | } | |
1918 | ||
1919 | return NULL; | |
1920 | } | |
1921 | ||
1baca4ce JL |
1922 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl); |
1923 | ||
1924 | static int find_later_rq(struct task_struct *task) | |
1925 | { | |
1926 | struct sched_domain *sd; | |
4ba29684 | 1927 | struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl); |
1baca4ce | 1928 | int this_cpu = smp_processor_id(); |
b18c3ca1 | 1929 | int cpu = task_cpu(task); |
1baca4ce JL |
1930 | |
1931 | /* Make sure the mask is initialized first */ | |
1932 | if (unlikely(!later_mask)) | |
1933 | return -1; | |
1934 | ||
4b53a341 | 1935 | if (task->nr_cpus_allowed == 1) |
1baca4ce JL |
1936 | return -1; |
1937 | ||
91ec6778 JL |
1938 | /* |
1939 | * We have to consider system topology and task affinity | |
97fb7a0a | 1940 | * first, then we can look for a suitable CPU. |
91ec6778 | 1941 | */ |
3261ed0b | 1942 | if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask)) |
1baca4ce JL |
1943 | return -1; |
1944 | ||
1945 | /* | |
b18c3ca1 BP |
1946 | * If we are here, some targets have been found, including |
1947 | * the most suitable which is, among the runqueues where the | |
1948 | * current tasks have later deadlines than the task's one, the | |
1949 | * rq with the latest possible one. | |
1baca4ce JL |
1950 | * |
1951 | * Now we check how well this matches with task's | |
1952 | * affinity and system topology. | |
1953 | * | |
97fb7a0a | 1954 | * The last CPU where the task run is our first |
1baca4ce JL |
1955 | * guess, since it is most likely cache-hot there. |
1956 | */ | |
1957 | if (cpumask_test_cpu(cpu, later_mask)) | |
1958 | return cpu; | |
1959 | /* | |
1960 | * Check if this_cpu is to be skipped (i.e., it is | |
1961 | * not in the mask) or not. | |
1962 | */ | |
1963 | if (!cpumask_test_cpu(this_cpu, later_mask)) | |
1964 | this_cpu = -1; | |
1965 | ||
1966 | rcu_read_lock(); | |
1967 | for_each_domain(cpu, sd) { | |
1968 | if (sd->flags & SD_WAKE_AFFINE) { | |
b18c3ca1 | 1969 | int best_cpu; |
1baca4ce JL |
1970 | |
1971 | /* | |
1972 | * If possible, preempting this_cpu is | |
1973 | * cheaper than migrating. | |
1974 | */ | |
1975 | if (this_cpu != -1 && | |
1976 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { | |
1977 | rcu_read_unlock(); | |
1978 | return this_cpu; | |
1979 | } | |
1980 | ||
b18c3ca1 BP |
1981 | best_cpu = cpumask_first_and(later_mask, |
1982 | sched_domain_span(sd)); | |
1baca4ce | 1983 | /* |
97fb7a0a | 1984 | * Last chance: if a CPU being in both later_mask |
b18c3ca1 | 1985 | * and current sd span is valid, that becomes our |
97fb7a0a | 1986 | * choice. Of course, the latest possible CPU is |
b18c3ca1 | 1987 | * already under consideration through later_mask. |
1baca4ce | 1988 | */ |
b18c3ca1 | 1989 | if (best_cpu < nr_cpu_ids) { |
1baca4ce JL |
1990 | rcu_read_unlock(); |
1991 | return best_cpu; | |
1992 | } | |
1993 | } | |
1994 | } | |
1995 | rcu_read_unlock(); | |
1996 | ||
1997 | /* | |
1998 | * At this point, all our guesses failed, we just return | |
1999 | * 'something', and let the caller sort the things out. | |
2000 | */ | |
2001 | if (this_cpu != -1) | |
2002 | return this_cpu; | |
2003 | ||
2004 | cpu = cpumask_any(later_mask); | |
2005 | if (cpu < nr_cpu_ids) | |
2006 | return cpu; | |
2007 | ||
2008 | return -1; | |
2009 | } | |
2010 | ||
2011 | /* Locks the rq it finds */ | |
2012 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq) | |
2013 | { | |
2014 | struct rq *later_rq = NULL; | |
2015 | int tries; | |
2016 | int cpu; | |
2017 | ||
2018 | for (tries = 0; tries < DL_MAX_TRIES; tries++) { | |
2019 | cpu = find_later_rq(task); | |
2020 | ||
2021 | if ((cpu == -1) || (cpu == rq->cpu)) | |
2022 | break; | |
2023 | ||
2024 | later_rq = cpu_rq(cpu); | |
2025 | ||
5aa50507 LA |
2026 | if (later_rq->dl.dl_nr_running && |
2027 | !dl_time_before(task->dl.deadline, | |
9d514262 WL |
2028 | later_rq->dl.earliest_dl.curr)) { |
2029 | /* | |
2030 | * Target rq has tasks of equal or earlier deadline, | |
2031 | * retrying does not release any lock and is unlikely | |
2032 | * to yield a different result. | |
2033 | */ | |
2034 | later_rq = NULL; | |
2035 | break; | |
2036 | } | |
2037 | ||
1baca4ce JL |
2038 | /* Retry if something changed. */ |
2039 | if (double_lock_balance(rq, later_rq)) { | |
2040 | if (unlikely(task_rq(task) != rq || | |
3bd37062 | 2041 | !cpumask_test_cpu(later_rq->cpu, task->cpus_ptr) || |
da0c1e65 | 2042 | task_running(rq, task) || |
13b5ab02 | 2043 | !dl_task(task) || |
da0c1e65 | 2044 | !task_on_rq_queued(task))) { |
1baca4ce JL |
2045 | double_unlock_balance(rq, later_rq); |
2046 | later_rq = NULL; | |
2047 | break; | |
2048 | } | |
2049 | } | |
2050 | ||
2051 | /* | |
2052 | * If the rq we found has no -deadline task, or | |
2053 | * its earliest one has a later deadline than our | |
2054 | * task, the rq is a good one. | |
2055 | */ | |
2056 | if (!later_rq->dl.dl_nr_running || | |
2057 | dl_time_before(task->dl.deadline, | |
2058 | later_rq->dl.earliest_dl.curr)) | |
2059 | break; | |
2060 | ||
2061 | /* Otherwise we try again. */ | |
2062 | double_unlock_balance(rq, later_rq); | |
2063 | later_rq = NULL; | |
2064 | } | |
2065 | ||
2066 | return later_rq; | |
2067 | } | |
2068 | ||
2069 | static struct task_struct *pick_next_pushable_dl_task(struct rq *rq) | |
2070 | { | |
2071 | struct task_struct *p; | |
2072 | ||
2073 | if (!has_pushable_dl_tasks(rq)) | |
2074 | return NULL; | |
2075 | ||
2161573e | 2076 | p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost, |
1baca4ce JL |
2077 | struct task_struct, pushable_dl_tasks); |
2078 | ||
2079 | BUG_ON(rq->cpu != task_cpu(p)); | |
2080 | BUG_ON(task_current(rq, p)); | |
4b53a341 | 2081 | BUG_ON(p->nr_cpus_allowed <= 1); |
1baca4ce | 2082 | |
da0c1e65 | 2083 | BUG_ON(!task_on_rq_queued(p)); |
1baca4ce JL |
2084 | BUG_ON(!dl_task(p)); |
2085 | ||
2086 | return p; | |
2087 | } | |
2088 | ||
2089 | /* | |
2090 | * See if the non running -deadline tasks on this rq | |
2091 | * can be sent to some other CPU where they can preempt | |
2092 | * and start executing. | |
2093 | */ | |
2094 | static int push_dl_task(struct rq *rq) | |
2095 | { | |
2096 | struct task_struct *next_task; | |
2097 | struct rq *later_rq; | |
c51b8ab5 | 2098 | int ret = 0; |
1baca4ce JL |
2099 | |
2100 | if (!rq->dl.overloaded) | |
2101 | return 0; | |
2102 | ||
2103 | next_task = pick_next_pushable_dl_task(rq); | |
2104 | if (!next_task) | |
2105 | return 0; | |
2106 | ||
2107 | retry: | |
9ebc6053 | 2108 | if (WARN_ON(next_task == rq->curr)) |
1baca4ce | 2109 | return 0; |
1baca4ce JL |
2110 | |
2111 | /* | |
2112 | * If next_task preempts rq->curr, and rq->curr | |
2113 | * can move away, it makes sense to just reschedule | |
2114 | * without going further in pushing next_task. | |
2115 | */ | |
2116 | if (dl_task(rq->curr) && | |
2117 | dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) && | |
4b53a341 | 2118 | rq->curr->nr_cpus_allowed > 1) { |
8875125e | 2119 | resched_curr(rq); |
1baca4ce JL |
2120 | return 0; |
2121 | } | |
2122 | ||
2123 | /* We might release rq lock */ | |
2124 | get_task_struct(next_task); | |
2125 | ||
2126 | /* Will lock the rq it'll find */ | |
2127 | later_rq = find_lock_later_rq(next_task, rq); | |
2128 | if (!later_rq) { | |
2129 | struct task_struct *task; | |
2130 | ||
2131 | /* | |
2132 | * We must check all this again, since | |
2133 | * find_lock_later_rq releases rq->lock and it is | |
2134 | * then possible that next_task has migrated. | |
2135 | */ | |
2136 | task = pick_next_pushable_dl_task(rq); | |
a776b968 | 2137 | if (task == next_task) { |
1baca4ce JL |
2138 | /* |
2139 | * The task is still there. We don't try | |
97fb7a0a | 2140 | * again, some other CPU will pull it when ready. |
1baca4ce | 2141 | */ |
1baca4ce JL |
2142 | goto out; |
2143 | } | |
2144 | ||
2145 | if (!task) | |
2146 | /* No more tasks */ | |
2147 | goto out; | |
2148 | ||
2149 | put_task_struct(next_task); | |
2150 | next_task = task; | |
2151 | goto retry; | |
2152 | } | |
2153 | ||
2154 | deactivate_task(rq, next_task, 0); | |
2155 | set_task_cpu(next_task, later_rq->cpu); | |
840d7196 DBO |
2156 | |
2157 | /* | |
2158 | * Update the later_rq clock here, because the clock is used | |
2159 | * by the cpufreq_update_util() inside __add_running_bw(). | |
2160 | */ | |
2161 | update_rq_clock(later_rq); | |
840d7196 | 2162 | activate_task(later_rq, next_task, ENQUEUE_NOCLOCK); |
c51b8ab5 | 2163 | ret = 1; |
1baca4ce | 2164 | |
8875125e | 2165 | resched_curr(later_rq); |
1baca4ce JL |
2166 | |
2167 | double_unlock_balance(rq, later_rq); | |
2168 | ||
2169 | out: | |
2170 | put_task_struct(next_task); | |
2171 | ||
c51b8ab5 | 2172 | return ret; |
1baca4ce JL |
2173 | } |
2174 | ||
2175 | static void push_dl_tasks(struct rq *rq) | |
2176 | { | |
4ffa08ed | 2177 | /* push_dl_task() will return true if it moved a -deadline task */ |
1baca4ce JL |
2178 | while (push_dl_task(rq)) |
2179 | ; | |
aab03e05 DF |
2180 | } |
2181 | ||
0ea60c20 | 2182 | static void pull_dl_task(struct rq *this_rq) |
1baca4ce | 2183 | { |
0ea60c20 | 2184 | int this_cpu = this_rq->cpu, cpu; |
1baca4ce | 2185 | struct task_struct *p; |
0ea60c20 | 2186 | bool resched = false; |
1baca4ce JL |
2187 | struct rq *src_rq; |
2188 | u64 dmin = LONG_MAX; | |
2189 | ||
2190 | if (likely(!dl_overloaded(this_rq))) | |
0ea60c20 | 2191 | return; |
1baca4ce JL |
2192 | |
2193 | /* | |
2194 | * Match the barrier from dl_set_overloaded; this guarantees that if we | |
2195 | * see overloaded we must also see the dlo_mask bit. | |
2196 | */ | |
2197 | smp_rmb(); | |
2198 | ||
2199 | for_each_cpu(cpu, this_rq->rd->dlo_mask) { | |
2200 | if (this_cpu == cpu) | |
2201 | continue; | |
2202 | ||
2203 | src_rq = cpu_rq(cpu); | |
2204 | ||
2205 | /* | |
2206 | * It looks racy, abd it is! However, as in sched_rt.c, | |
2207 | * we are fine with this. | |
2208 | */ | |
2209 | if (this_rq->dl.dl_nr_running && | |
2210 | dl_time_before(this_rq->dl.earliest_dl.curr, | |
2211 | src_rq->dl.earliest_dl.next)) | |
2212 | continue; | |
2213 | ||
2214 | /* Might drop this_rq->lock */ | |
2215 | double_lock_balance(this_rq, src_rq); | |
2216 | ||
2217 | /* | |
2218 | * If there are no more pullable tasks on the | |
2219 | * rq, we're done with it. | |
2220 | */ | |
2221 | if (src_rq->dl.dl_nr_running <= 1) | |
2222 | goto skip; | |
2223 | ||
8b5e770e | 2224 | p = pick_earliest_pushable_dl_task(src_rq, this_cpu); |
1baca4ce JL |
2225 | |
2226 | /* | |
2227 | * We found a task to be pulled if: | |
2228 | * - it preempts our current (if there's one), | |
2229 | * - it will preempt the last one we pulled (if any). | |
2230 | */ | |
2231 | if (p && dl_time_before(p->dl.deadline, dmin) && | |
2232 | (!this_rq->dl.dl_nr_running || | |
2233 | dl_time_before(p->dl.deadline, | |
2234 | this_rq->dl.earliest_dl.curr))) { | |
2235 | WARN_ON(p == src_rq->curr); | |
da0c1e65 | 2236 | WARN_ON(!task_on_rq_queued(p)); |
1baca4ce JL |
2237 | |
2238 | /* | |
2239 | * Then we pull iff p has actually an earlier | |
2240 | * deadline than the current task of its runqueue. | |
2241 | */ | |
2242 | if (dl_time_before(p->dl.deadline, | |
2243 | src_rq->curr->dl.deadline)) | |
2244 | goto skip; | |
2245 | ||
0ea60c20 | 2246 | resched = true; |
1baca4ce JL |
2247 | |
2248 | deactivate_task(src_rq, p, 0); | |
2249 | set_task_cpu(p, this_cpu); | |
2250 | activate_task(this_rq, p, 0); | |
2251 | dmin = p->dl.deadline; | |
2252 | ||
2253 | /* Is there any other task even earlier? */ | |
2254 | } | |
2255 | skip: | |
2256 | double_unlock_balance(this_rq, src_rq); | |
2257 | } | |
2258 | ||
0ea60c20 PZ |
2259 | if (resched) |
2260 | resched_curr(this_rq); | |
1baca4ce JL |
2261 | } |
2262 | ||
2263 | /* | |
2264 | * Since the task is not running and a reschedule is not going to happen | |
2265 | * anytime soon on its runqueue, we try pushing it away now. | |
2266 | */ | |
2267 | static void task_woken_dl(struct rq *rq, struct task_struct *p) | |
2268 | { | |
2269 | if (!task_running(rq, p) && | |
2270 | !test_tsk_need_resched(rq->curr) && | |
4b53a341 | 2271 | p->nr_cpus_allowed > 1 && |
1baca4ce | 2272 | dl_task(rq->curr) && |
4b53a341 | 2273 | (rq->curr->nr_cpus_allowed < 2 || |
6b0a563f | 2274 | !dl_entity_preempt(&p->dl, &rq->curr->dl))) { |
1baca4ce JL |
2275 | push_dl_tasks(rq); |
2276 | } | |
2277 | } | |
2278 | ||
2279 | static void set_cpus_allowed_dl(struct task_struct *p, | |
2280 | const struct cpumask *new_mask) | |
2281 | { | |
7f51412a | 2282 | struct root_domain *src_rd; |
6c37067e | 2283 | struct rq *rq; |
1baca4ce JL |
2284 | |
2285 | BUG_ON(!dl_task(p)); | |
2286 | ||
7f51412a JL |
2287 | rq = task_rq(p); |
2288 | src_rd = rq->rd; | |
2289 | /* | |
2290 | * Migrating a SCHED_DEADLINE task between exclusive | |
2291 | * cpusets (different root_domains) entails a bandwidth | |
2292 | * update. We already made space for us in the destination | |
2293 | * domain (see cpuset_can_attach()). | |
2294 | */ | |
2295 | if (!cpumask_intersects(src_rd->span, new_mask)) { | |
2296 | struct dl_bw *src_dl_b; | |
2297 | ||
2298 | src_dl_b = dl_bw_of(cpu_of(rq)); | |
2299 | /* | |
2300 | * We now free resources of the root_domain we are migrating | |
2301 | * off. In the worst case, sched_setattr() may temporary fail | |
2302 | * until we complete the update. | |
2303 | */ | |
2304 | raw_spin_lock(&src_dl_b->lock); | |
8c0944ce | 2305 | __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); |
7f51412a JL |
2306 | raw_spin_unlock(&src_dl_b->lock); |
2307 | } | |
2308 | ||
6c37067e | 2309 | set_cpus_allowed_common(p, new_mask); |
1baca4ce JL |
2310 | } |
2311 | ||
2312 | /* Assumes rq->lock is held */ | |
2313 | static void rq_online_dl(struct rq *rq) | |
2314 | { | |
2315 | if (rq->dl.overloaded) | |
2316 | dl_set_overload(rq); | |
6bfd6d72 | 2317 | |
16b26943 | 2318 | cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu); |
6bfd6d72 | 2319 | if (rq->dl.dl_nr_running > 0) |
d8206bb3 | 2320 | cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr); |
1baca4ce JL |
2321 | } |
2322 | ||
2323 | /* Assumes rq->lock is held */ | |
2324 | static void rq_offline_dl(struct rq *rq) | |
2325 | { | |
2326 | if (rq->dl.overloaded) | |
2327 | dl_clear_overload(rq); | |
6bfd6d72 | 2328 | |
d8206bb3 | 2329 | cpudl_clear(&rq->rd->cpudl, rq->cpu); |
16b26943 | 2330 | cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu); |
1baca4ce JL |
2331 | } |
2332 | ||
a6c0e746 | 2333 | void __init init_sched_dl_class(void) |
1baca4ce JL |
2334 | { |
2335 | unsigned int i; | |
2336 | ||
2337 | for_each_possible_cpu(i) | |
2338 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i), | |
2339 | GFP_KERNEL, cpu_to_node(i)); | |
2340 | } | |
2341 | ||
f9a25f77 MP |
2342 | void dl_add_task_root_domain(struct task_struct *p) |
2343 | { | |
2344 | struct rq_flags rf; | |
2345 | struct rq *rq; | |
2346 | struct dl_bw *dl_b; | |
2347 | ||
2348 | rq = task_rq_lock(p, &rf); | |
2349 | if (!dl_task(p)) | |
2350 | goto unlock; | |
2351 | ||
2352 | dl_b = &rq->rd->dl_bw; | |
2353 | raw_spin_lock(&dl_b->lock); | |
2354 | ||
2355 | __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span)); | |
2356 | ||
2357 | raw_spin_unlock(&dl_b->lock); | |
2358 | ||
2359 | unlock: | |
2360 | task_rq_unlock(rq, p, &rf); | |
2361 | } | |
2362 | ||
2363 | void dl_clear_root_domain(struct root_domain *rd) | |
2364 | { | |
2365 | unsigned long flags; | |
2366 | ||
2367 | raw_spin_lock_irqsave(&rd->dl_bw.lock, flags); | |
2368 | rd->dl_bw.total_bw = 0; | |
2369 | raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags); | |
2370 | } | |
2371 | ||
1baca4ce JL |
2372 | #endif /* CONFIG_SMP */ |
2373 | ||
aab03e05 DF |
2374 | static void switched_from_dl(struct rq *rq, struct task_struct *p) |
2375 | { | |
a649f237 | 2376 | /* |
209a0cbd LA |
2377 | * task_non_contending() can start the "inactive timer" (if the 0-lag |
2378 | * time is in the future). If the task switches back to dl before | |
2379 | * the "inactive timer" fires, it can continue to consume its current | |
2380 | * runtime using its current deadline. If it stays outside of | |
2381 | * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer() | |
2382 | * will reset the task parameters. | |
a649f237 | 2383 | */ |
209a0cbd LA |
2384 | if (task_on_rq_queued(p) && p->dl.dl_runtime) |
2385 | task_non_contending(p); | |
2386 | ||
e117cb52 JL |
2387 | if (!task_on_rq_queued(p)) { |
2388 | /* | |
2389 | * Inactive timer is armed. However, p is leaving DEADLINE and | |
2390 | * might migrate away from this rq while continuing to run on | |
2391 | * some other class. We need to remove its contribution from | |
2392 | * this rq running_bw now, or sub_rq_bw (below) will complain. | |
2393 | */ | |
2394 | if (p->dl.dl_non_contending) | |
2395 | sub_running_bw(&p->dl, &rq->dl); | |
794a56eb | 2396 | sub_rq_bw(&p->dl, &rq->dl); |
e117cb52 | 2397 | } |
8fd27231 | 2398 | |
209a0cbd LA |
2399 | /* |
2400 | * We cannot use inactive_task_timer() to invoke sub_running_bw() | |
2401 | * at the 0-lag time, because the task could have been migrated | |
2402 | * while SCHED_OTHER in the meanwhile. | |
2403 | */ | |
2404 | if (p->dl.dl_non_contending) | |
2405 | p->dl.dl_non_contending = 0; | |
a5e7be3b | 2406 | |
1baca4ce JL |
2407 | /* |
2408 | * Since this might be the only -deadline task on the rq, | |
2409 | * this is the right place to try to pull some other one | |
97fb7a0a | 2410 | * from an overloaded CPU, if any. |
1baca4ce | 2411 | */ |
cd660911 WL |
2412 | if (!task_on_rq_queued(p) || rq->dl.dl_nr_running) |
2413 | return; | |
2414 | ||
02d8ec94 | 2415 | deadline_queue_pull_task(rq); |
aab03e05 DF |
2416 | } |
2417 | ||
1baca4ce JL |
2418 | /* |
2419 | * When switching to -deadline, we may overload the rq, then | |
2420 | * we try to push someone off, if possible. | |
2421 | */ | |
aab03e05 DF |
2422 | static void switched_to_dl(struct rq *rq, struct task_struct *p) |
2423 | { | |
209a0cbd LA |
2424 | if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) |
2425 | put_task_struct(p); | |
98b0a857 JL |
2426 | |
2427 | /* If p is not queued we will update its parameters at next wakeup. */ | |
8fd27231 | 2428 | if (!task_on_rq_queued(p)) { |
794a56eb | 2429 | add_rq_bw(&p->dl, &rq->dl); |
98b0a857 | 2430 | |
8fd27231 LA |
2431 | return; |
2432 | } | |
72f9f3fd | 2433 | |
98b0a857 | 2434 | if (rq->curr != p) { |
1baca4ce | 2435 | #ifdef CONFIG_SMP |
4b53a341 | 2436 | if (p->nr_cpus_allowed > 1 && rq->dl.overloaded) |
02d8ec94 | 2437 | deadline_queue_push_tasks(rq); |
619bd4a7 | 2438 | #endif |
9916e214 PZ |
2439 | if (dl_task(rq->curr)) |
2440 | check_preempt_curr_dl(rq, p, 0); | |
2441 | else | |
2442 | resched_curr(rq); | |
aab03e05 DF |
2443 | } |
2444 | } | |
2445 | ||
1baca4ce JL |
2446 | /* |
2447 | * If the scheduling parameters of a -deadline task changed, | |
2448 | * a push or pull operation might be needed. | |
2449 | */ | |
aab03e05 DF |
2450 | static void prio_changed_dl(struct rq *rq, struct task_struct *p, |
2451 | int oldprio) | |
2452 | { | |
da0c1e65 | 2453 | if (task_on_rq_queued(p) || rq->curr == p) { |
aab03e05 | 2454 | #ifdef CONFIG_SMP |
1baca4ce JL |
2455 | /* |
2456 | * This might be too much, but unfortunately | |
2457 | * we don't have the old deadline value, and | |
2458 | * we can't argue if the task is increasing | |
2459 | * or lowering its prio, so... | |
2460 | */ | |
2461 | if (!rq->dl.overloaded) | |
02d8ec94 | 2462 | deadline_queue_pull_task(rq); |
1baca4ce JL |
2463 | |
2464 | /* | |
2465 | * If we now have a earlier deadline task than p, | |
2466 | * then reschedule, provided p is still on this | |
2467 | * runqueue. | |
2468 | */ | |
9916e214 | 2469 | if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline)) |
8875125e | 2470 | resched_curr(rq); |
1baca4ce JL |
2471 | #else |
2472 | /* | |
2473 | * Again, we don't know if p has a earlier | |
2474 | * or later deadline, so let's blindly set a | |
2475 | * (maybe not needed) rescheduling point. | |
2476 | */ | |
8875125e | 2477 | resched_curr(rq); |
1baca4ce | 2478 | #endif /* CONFIG_SMP */ |
801ccdbf | 2479 | } |
aab03e05 | 2480 | } |
aab03e05 DF |
2481 | |
2482 | const struct sched_class dl_sched_class = { | |
2483 | .next = &rt_sched_class, | |
2484 | .enqueue_task = enqueue_task_dl, | |
2485 | .dequeue_task = dequeue_task_dl, | |
2486 | .yield_task = yield_task_dl, | |
2487 | ||
2488 | .check_preempt_curr = check_preempt_curr_dl, | |
2489 | ||
2490 | .pick_next_task = pick_next_task_dl, | |
2491 | .put_prev_task = put_prev_task_dl, | |
03b7fad1 | 2492 | .set_next_task = set_next_task_dl, |
aab03e05 DF |
2493 | |
2494 | #ifdef CONFIG_SMP | |
6e2df058 | 2495 | .balance = balance_dl, |
aab03e05 | 2496 | .select_task_rq = select_task_rq_dl, |
209a0cbd | 2497 | .migrate_task_rq = migrate_task_rq_dl, |
1baca4ce JL |
2498 | .set_cpus_allowed = set_cpus_allowed_dl, |
2499 | .rq_online = rq_online_dl, | |
2500 | .rq_offline = rq_offline_dl, | |
1baca4ce | 2501 | .task_woken = task_woken_dl, |
aab03e05 DF |
2502 | #endif |
2503 | ||
aab03e05 DF |
2504 | .task_tick = task_tick_dl, |
2505 | .task_fork = task_fork_dl, | |
aab03e05 DF |
2506 | |
2507 | .prio_changed = prio_changed_dl, | |
2508 | .switched_from = switched_from_dl, | |
2509 | .switched_to = switched_to_dl, | |
6e998916 SG |
2510 | |
2511 | .update_curr = update_curr_dl, | |
aab03e05 | 2512 | }; |
acb32132 | 2513 | |
06a76fe0 NP |
2514 | int sched_dl_global_validate(void) |
2515 | { | |
2516 | u64 runtime = global_rt_runtime(); | |
2517 | u64 period = global_rt_period(); | |
2518 | u64 new_bw = to_ratio(period, runtime); | |
2519 | struct dl_bw *dl_b; | |
2520 | int cpu, ret = 0; | |
2521 | unsigned long flags; | |
2522 | ||
2523 | /* | |
2524 | * Here we want to check the bandwidth not being set to some | |
2525 | * value smaller than the currently allocated bandwidth in | |
2526 | * any of the root_domains. | |
2527 | * | |
2528 | * FIXME: Cycling on all the CPUs is overdoing, but simpler than | |
2529 | * cycling on root_domains... Discussion on different/better | |
2530 | * solutions is welcome! | |
2531 | */ | |
2532 | for_each_possible_cpu(cpu) { | |
2533 | rcu_read_lock_sched(); | |
2534 | dl_b = dl_bw_of(cpu); | |
2535 | ||
2536 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
2537 | if (new_bw < dl_b->total_bw) | |
2538 | ret = -EBUSY; | |
2539 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2540 | ||
2541 | rcu_read_unlock_sched(); | |
2542 | ||
2543 | if (ret) | |
2544 | break; | |
2545 | } | |
2546 | ||
2547 | return ret; | |
2548 | } | |
2549 | ||
ba4f7bc1 | 2550 | static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq) |
06a76fe0 NP |
2551 | { |
2552 | if (global_rt_runtime() == RUNTIME_INF) { | |
2553 | dl_rq->bw_ratio = 1 << RATIO_SHIFT; | |
2554 | dl_rq->extra_bw = 1 << BW_SHIFT; | |
2555 | } else { | |
2556 | dl_rq->bw_ratio = to_ratio(global_rt_runtime(), | |
2557 | global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT); | |
2558 | dl_rq->extra_bw = to_ratio(global_rt_period(), | |
2559 | global_rt_runtime()); | |
2560 | } | |
2561 | } | |
2562 | ||
2563 | void sched_dl_do_global(void) | |
2564 | { | |
2565 | u64 new_bw = -1; | |
2566 | struct dl_bw *dl_b; | |
2567 | int cpu; | |
2568 | unsigned long flags; | |
2569 | ||
2570 | def_dl_bandwidth.dl_period = global_rt_period(); | |
2571 | def_dl_bandwidth.dl_runtime = global_rt_runtime(); | |
2572 | ||
2573 | if (global_rt_runtime() != RUNTIME_INF) | |
2574 | new_bw = to_ratio(global_rt_period(), global_rt_runtime()); | |
2575 | ||
2576 | /* | |
2577 | * FIXME: As above... | |
2578 | */ | |
2579 | for_each_possible_cpu(cpu) { | |
2580 | rcu_read_lock_sched(); | |
2581 | dl_b = dl_bw_of(cpu); | |
2582 | ||
2583 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
2584 | dl_b->bw = new_bw; | |
2585 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2586 | ||
2587 | rcu_read_unlock_sched(); | |
2588 | init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl); | |
2589 | } | |
2590 | } | |
2591 | ||
2592 | /* | |
2593 | * We must be sure that accepting a new task (or allowing changing the | |
2594 | * parameters of an existing one) is consistent with the bandwidth | |
2595 | * constraints. If yes, this function also accordingly updates the currently | |
2596 | * allocated bandwidth to reflect the new situation. | |
2597 | * | |
2598 | * This function is called while holding p's rq->lock. | |
2599 | */ | |
2600 | int sched_dl_overflow(struct task_struct *p, int policy, | |
2601 | const struct sched_attr *attr) | |
2602 | { | |
06a76fe0 NP |
2603 | u64 period = attr->sched_period ?: attr->sched_deadline; |
2604 | u64 runtime = attr->sched_runtime; | |
2605 | u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; | |
60ffd5ed LA |
2606 | int cpus, err = -1, cpu = task_cpu(p); |
2607 | struct dl_bw *dl_b = dl_bw_of(cpu); | |
2608 | unsigned long cap; | |
06a76fe0 | 2609 | |
794a56eb JL |
2610 | if (attr->sched_flags & SCHED_FLAG_SUGOV) |
2611 | return 0; | |
2612 | ||
06a76fe0 NP |
2613 | /* !deadline task may carry old deadline bandwidth */ |
2614 | if (new_bw == p->dl.dl_bw && task_has_dl_policy(p)) | |
2615 | return 0; | |
2616 | ||
2617 | /* | |
2618 | * Either if a task, enters, leave, or stays -deadline but changes | |
2619 | * its parameters, we may need to update accordingly the total | |
2620 | * allocated bandwidth of the container. | |
2621 | */ | |
2622 | raw_spin_lock(&dl_b->lock); | |
60ffd5ed LA |
2623 | cpus = dl_bw_cpus(cpu); |
2624 | cap = dl_bw_capacity(cpu); | |
2625 | ||
06a76fe0 | 2626 | if (dl_policy(policy) && !task_has_dl_policy(p) && |
60ffd5ed | 2627 | !__dl_overflow(dl_b, cap, 0, new_bw)) { |
06a76fe0 | 2628 | if (hrtimer_active(&p->dl.inactive_timer)) |
8c0944ce | 2629 | __dl_sub(dl_b, p->dl.dl_bw, cpus); |
06a76fe0 NP |
2630 | __dl_add(dl_b, new_bw, cpus); |
2631 | err = 0; | |
2632 | } else if (dl_policy(policy) && task_has_dl_policy(p) && | |
60ffd5ed | 2633 | !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) { |
06a76fe0 NP |
2634 | /* |
2635 | * XXX this is slightly incorrect: when the task | |
2636 | * utilization decreases, we should delay the total | |
2637 | * utilization change until the task's 0-lag point. | |
2638 | * But this would require to set the task's "inactive | |
2639 | * timer" when the task is not inactive. | |
2640 | */ | |
8c0944ce | 2641 | __dl_sub(dl_b, p->dl.dl_bw, cpus); |
06a76fe0 NP |
2642 | __dl_add(dl_b, new_bw, cpus); |
2643 | dl_change_utilization(p, new_bw); | |
2644 | err = 0; | |
2645 | } else if (!dl_policy(policy) && task_has_dl_policy(p)) { | |
2646 | /* | |
2647 | * Do not decrease the total deadline utilization here, | |
2648 | * switched_from_dl() will take care to do it at the correct | |
2649 | * (0-lag) time. | |
2650 | */ | |
2651 | err = 0; | |
2652 | } | |
2653 | raw_spin_unlock(&dl_b->lock); | |
2654 | ||
2655 | return err; | |
2656 | } | |
2657 | ||
2658 | /* | |
2659 | * This function initializes the sched_dl_entity of a newly becoming | |
2660 | * SCHED_DEADLINE task. | |
2661 | * | |
2662 | * Only the static values are considered here, the actual runtime and the | |
2663 | * absolute deadline will be properly calculated when the task is enqueued | |
2664 | * for the first time with its new policy. | |
2665 | */ | |
2666 | void __setparam_dl(struct task_struct *p, const struct sched_attr *attr) | |
2667 | { | |
2668 | struct sched_dl_entity *dl_se = &p->dl; | |
2669 | ||
2670 | dl_se->dl_runtime = attr->sched_runtime; | |
2671 | dl_se->dl_deadline = attr->sched_deadline; | |
2672 | dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; | |
2673 | dl_se->flags = attr->sched_flags; | |
2674 | dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); | |
2675 | dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime); | |
2676 | } | |
2677 | ||
2678 | void __getparam_dl(struct task_struct *p, struct sched_attr *attr) | |
2679 | { | |
2680 | struct sched_dl_entity *dl_se = &p->dl; | |
2681 | ||
2682 | attr->sched_priority = p->rt_priority; | |
2683 | attr->sched_runtime = dl_se->dl_runtime; | |
2684 | attr->sched_deadline = dl_se->dl_deadline; | |
2685 | attr->sched_period = dl_se->dl_period; | |
2686 | attr->sched_flags = dl_se->flags; | |
2687 | } | |
2688 | ||
b4098bfc PZ |
2689 | /* |
2690 | * Default limits for DL period; on the top end we guard against small util | |
2691 | * tasks still getting rediculous long effective runtimes, on the bottom end we | |
2692 | * guard against timer DoS. | |
2693 | */ | |
2694 | unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */ | |
2695 | unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */ | |
2696 | ||
06a76fe0 NP |
2697 | /* |
2698 | * This function validates the new parameters of a -deadline task. | |
2699 | * We ask for the deadline not being zero, and greater or equal | |
2700 | * than the runtime, as well as the period of being zero or | |
2701 | * greater than deadline. Furthermore, we have to be sure that | |
2702 | * user parameters are above the internal resolution of 1us (we | |
2703 | * check sched_runtime only since it is always the smaller one) and | |
2704 | * below 2^63 ns (we have to check both sched_deadline and | |
2705 | * sched_period, as the latter can be zero). | |
2706 | */ | |
2707 | bool __checkparam_dl(const struct sched_attr *attr) | |
2708 | { | |
b4098bfc PZ |
2709 | u64 period, max, min; |
2710 | ||
794a56eb JL |
2711 | /* special dl tasks don't actually use any parameter */ |
2712 | if (attr->sched_flags & SCHED_FLAG_SUGOV) | |
2713 | return true; | |
2714 | ||
06a76fe0 NP |
2715 | /* deadline != 0 */ |
2716 | if (attr->sched_deadline == 0) | |
2717 | return false; | |
2718 | ||
2719 | /* | |
2720 | * Since we truncate DL_SCALE bits, make sure we're at least | |
2721 | * that big. | |
2722 | */ | |
2723 | if (attr->sched_runtime < (1ULL << DL_SCALE)) | |
2724 | return false; | |
2725 | ||
2726 | /* | |
2727 | * Since we use the MSB for wrap-around and sign issues, make | |
2728 | * sure it's not set (mind that period can be equal to zero). | |
2729 | */ | |
2730 | if (attr->sched_deadline & (1ULL << 63) || | |
2731 | attr->sched_period & (1ULL << 63)) | |
2732 | return false; | |
2733 | ||
b4098bfc PZ |
2734 | period = attr->sched_period; |
2735 | if (!period) | |
2736 | period = attr->sched_deadline; | |
2737 | ||
06a76fe0 | 2738 | /* runtime <= deadline <= period (if period != 0) */ |
b4098bfc | 2739 | if (period < attr->sched_deadline || |
06a76fe0 NP |
2740 | attr->sched_deadline < attr->sched_runtime) |
2741 | return false; | |
2742 | ||
b4098bfc PZ |
2743 | max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC; |
2744 | min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC; | |
2745 | ||
2746 | if (period < min || period > max) | |
2747 | return false; | |
2748 | ||
06a76fe0 NP |
2749 | return true; |
2750 | } | |
2751 | ||
2752 | /* | |
2753 | * This function clears the sched_dl_entity static params. | |
2754 | */ | |
2755 | void __dl_clear_params(struct task_struct *p) | |
2756 | { | |
2757 | struct sched_dl_entity *dl_se = &p->dl; | |
2758 | ||
97fb7a0a IM |
2759 | dl_se->dl_runtime = 0; |
2760 | dl_se->dl_deadline = 0; | |
2761 | dl_se->dl_period = 0; | |
2762 | dl_se->flags = 0; | |
2763 | dl_se->dl_bw = 0; | |
2764 | dl_se->dl_density = 0; | |
06a76fe0 | 2765 | |
97fb7a0a IM |
2766 | dl_se->dl_throttled = 0; |
2767 | dl_se->dl_yielded = 0; | |
2768 | dl_se->dl_non_contending = 0; | |
2769 | dl_se->dl_overrun = 0; | |
06a76fe0 NP |
2770 | } |
2771 | ||
2772 | bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr) | |
2773 | { | |
2774 | struct sched_dl_entity *dl_se = &p->dl; | |
2775 | ||
2776 | if (dl_se->dl_runtime != attr->sched_runtime || | |
2777 | dl_se->dl_deadline != attr->sched_deadline || | |
2778 | dl_se->dl_period != attr->sched_period || | |
2779 | dl_se->flags != attr->sched_flags) | |
2780 | return true; | |
2781 | ||
2782 | return false; | |
2783 | } | |
2784 | ||
2785 | #ifdef CONFIG_SMP | |
2786 | int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed) | |
2787 | { | |
60ffd5ed | 2788 | unsigned long flags, cap; |
97fb7a0a | 2789 | unsigned int dest_cpu; |
06a76fe0 NP |
2790 | struct dl_bw *dl_b; |
2791 | bool overflow; | |
60ffd5ed | 2792 | int ret; |
06a76fe0 | 2793 | |
97fb7a0a IM |
2794 | dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed); |
2795 | ||
06a76fe0 NP |
2796 | rcu_read_lock_sched(); |
2797 | dl_b = dl_bw_of(dest_cpu); | |
2798 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
60ffd5ed LA |
2799 | cap = dl_bw_capacity(dest_cpu); |
2800 | overflow = __dl_overflow(dl_b, cap, 0, p->dl.dl_bw); | |
97fb7a0a | 2801 | if (overflow) { |
06a76fe0 | 2802 | ret = -EBUSY; |
97fb7a0a | 2803 | } else { |
06a76fe0 NP |
2804 | /* |
2805 | * We reserve space for this task in the destination | |
2806 | * root_domain, as we can't fail after this point. | |
2807 | * We will free resources in the source root_domain | |
2808 | * later on (see set_cpus_allowed_dl()). | |
2809 | */ | |
60ffd5ed LA |
2810 | int cpus = dl_bw_cpus(dest_cpu); |
2811 | ||
06a76fe0 NP |
2812 | __dl_add(dl_b, p->dl.dl_bw, cpus); |
2813 | ret = 0; | |
2814 | } | |
2815 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2816 | rcu_read_unlock_sched(); | |
97fb7a0a | 2817 | |
06a76fe0 NP |
2818 | return ret; |
2819 | } | |
2820 | ||
2821 | int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, | |
2822 | const struct cpumask *trial) | |
2823 | { | |
2824 | int ret = 1, trial_cpus; | |
2825 | struct dl_bw *cur_dl_b; | |
2826 | unsigned long flags; | |
2827 | ||
2828 | rcu_read_lock_sched(); | |
2829 | cur_dl_b = dl_bw_of(cpumask_any(cur)); | |
2830 | trial_cpus = cpumask_weight(trial); | |
2831 | ||
2832 | raw_spin_lock_irqsave(&cur_dl_b->lock, flags); | |
2833 | if (cur_dl_b->bw != -1 && | |
2834 | cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw) | |
2835 | ret = 0; | |
2836 | raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags); | |
2837 | rcu_read_unlock_sched(); | |
97fb7a0a | 2838 | |
06a76fe0 NP |
2839 | return ret; |
2840 | } | |
2841 | ||
2842 | bool dl_cpu_busy(unsigned int cpu) | |
2843 | { | |
60ffd5ed | 2844 | unsigned long flags, cap; |
06a76fe0 NP |
2845 | struct dl_bw *dl_b; |
2846 | bool overflow; | |
06a76fe0 NP |
2847 | |
2848 | rcu_read_lock_sched(); | |
2849 | dl_b = dl_bw_of(cpu); | |
2850 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
60ffd5ed LA |
2851 | cap = dl_bw_capacity(cpu); |
2852 | overflow = __dl_overflow(dl_b, cap, 0, 0); | |
06a76fe0 NP |
2853 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); |
2854 | rcu_read_unlock_sched(); | |
97fb7a0a | 2855 | |
06a76fe0 NP |
2856 | return overflow; |
2857 | } | |
2858 | #endif | |
2859 | ||
acb32132 | 2860 | #ifdef CONFIG_SCHED_DEBUG |
acb32132 WL |
2861 | void print_dl_stats(struct seq_file *m, int cpu) |
2862 | { | |
2863 | print_dl_rq(m, cpu, &cpu_rq(cpu)->dl); | |
2864 | } | |
2865 | #endif /* CONFIG_SCHED_DEBUG */ |