]>
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; | |
57 | int cpus = 0; | |
58 | ||
59 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), | |
60 | "sched RCU must be held"); | |
61 | for_each_cpu_and(i, rd->span, cpu_active_mask) | |
62 | cpus++; | |
63 | ||
64 | return cpus; | |
65 | } | |
66 | #else | |
67 | static inline struct dl_bw *dl_bw_of(int i) | |
68 | { | |
69 | return &cpu_rq(i)->dl.dl_bw; | |
70 | } | |
71 | ||
72 | static inline int dl_bw_cpus(int i) | |
73 | { | |
74 | return 1; | |
75 | } | |
76 | #endif | |
77 | ||
e36d8677 | 78 | static inline |
794a56eb | 79 | void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq) |
e36d8677 LA |
80 | { |
81 | u64 old = dl_rq->running_bw; | |
82 | ||
83 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
84 | dl_rq->running_bw += dl_bw; | |
85 | SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */ | |
8fd27231 | 86 | SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw); |
e0367b12 | 87 | /* kick cpufreq (see the comment in kernel/sched/sched.h). */ |
4042d003 | 88 | cpufreq_update_util(rq_of_dl_rq(dl_rq), 0); |
e36d8677 LA |
89 | } |
90 | ||
91 | static inline | |
794a56eb | 92 | void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq) |
e36d8677 LA |
93 | { |
94 | u64 old = dl_rq->running_bw; | |
95 | ||
96 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
97 | dl_rq->running_bw -= dl_bw; | |
98 | SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */ | |
99 | if (dl_rq->running_bw > old) | |
100 | dl_rq->running_bw = 0; | |
e0367b12 | 101 | /* kick cpufreq (see the comment in kernel/sched/sched.h). */ |
4042d003 | 102 | cpufreq_update_util(rq_of_dl_rq(dl_rq), 0); |
e36d8677 LA |
103 | } |
104 | ||
8fd27231 | 105 | static inline |
794a56eb | 106 | void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq) |
8fd27231 LA |
107 | { |
108 | u64 old = dl_rq->this_bw; | |
109 | ||
110 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
111 | dl_rq->this_bw += dl_bw; | |
112 | SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */ | |
113 | } | |
114 | ||
115 | static inline | |
794a56eb | 116 | void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq) |
8fd27231 LA |
117 | { |
118 | u64 old = dl_rq->this_bw; | |
119 | ||
120 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
121 | dl_rq->this_bw -= dl_bw; | |
122 | SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */ | |
123 | if (dl_rq->this_bw > old) | |
124 | dl_rq->this_bw = 0; | |
125 | SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw); | |
126 | } | |
127 | ||
794a56eb JL |
128 | static inline |
129 | void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
130 | { | |
131 | if (!dl_entity_is_special(dl_se)) | |
132 | __add_rq_bw(dl_se->dl_bw, dl_rq); | |
133 | } | |
134 | ||
135 | static inline | |
136 | void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
137 | { | |
138 | if (!dl_entity_is_special(dl_se)) | |
139 | __sub_rq_bw(dl_se->dl_bw, dl_rq); | |
140 | } | |
141 | ||
142 | static inline | |
143 | void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
144 | { | |
145 | if (!dl_entity_is_special(dl_se)) | |
146 | __add_running_bw(dl_se->dl_bw, dl_rq); | |
147 | } | |
148 | ||
149 | static inline | |
150 | void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
151 | { | |
152 | if (!dl_entity_is_special(dl_se)) | |
153 | __sub_running_bw(dl_se->dl_bw, dl_rq); | |
154 | } | |
155 | ||
209a0cbd LA |
156 | void dl_change_utilization(struct task_struct *p, u64 new_bw) |
157 | { | |
8fd27231 | 158 | struct rq *rq; |
209a0cbd | 159 | |
794a56eb JL |
160 | BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV); |
161 | ||
8fd27231 | 162 | if (task_on_rq_queued(p)) |
209a0cbd LA |
163 | return; |
164 | ||
8fd27231 LA |
165 | rq = task_rq(p); |
166 | if (p->dl.dl_non_contending) { | |
794a56eb | 167 | sub_running_bw(&p->dl, &rq->dl); |
8fd27231 LA |
168 | p->dl.dl_non_contending = 0; |
169 | /* | |
170 | * If the timer handler is currently running and the | |
171 | * timer cannot be cancelled, inactive_task_timer() | |
172 | * will see that dl_not_contending is not set, and | |
173 | * will not touch the rq's active utilization, | |
174 | * so we are still safe. | |
175 | */ | |
176 | if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) | |
177 | put_task_struct(p); | |
178 | } | |
794a56eb JL |
179 | __sub_rq_bw(p->dl.dl_bw, &rq->dl); |
180 | __add_rq_bw(new_bw, &rq->dl); | |
209a0cbd LA |
181 | } |
182 | ||
183 | /* | |
184 | * The utilization of a task cannot be immediately removed from | |
185 | * the rq active utilization (running_bw) when the task blocks. | |
186 | * Instead, we have to wait for the so called "0-lag time". | |
187 | * | |
188 | * If a task blocks before the "0-lag time", a timer (the inactive | |
189 | * timer) is armed, and running_bw is decreased when the timer | |
190 | * fires. | |
191 | * | |
192 | * If the task wakes up again before the inactive timer fires, | |
193 | * the timer is cancelled, whereas if the task wakes up after the | |
194 | * inactive timer fired (and running_bw has been decreased) the | |
195 | * task's utilization has to be added to running_bw again. | |
196 | * A flag in the deadline scheduling entity (dl_non_contending) | |
197 | * is used to avoid race conditions between the inactive timer handler | |
198 | * and task wakeups. | |
199 | * | |
200 | * The following diagram shows how running_bw is updated. A task is | |
201 | * "ACTIVE" when its utilization contributes to running_bw; an | |
202 | * "ACTIVE contending" task is in the TASK_RUNNING state, while an | |
203 | * "ACTIVE non contending" task is a blocked task for which the "0-lag time" | |
204 | * has not passed yet. An "INACTIVE" task is a task for which the "0-lag" | |
205 | * time already passed, which does not contribute to running_bw anymore. | |
206 | * +------------------+ | |
207 | * wakeup | ACTIVE | | |
208 | * +------------------>+ contending | | |
209 | * | add_running_bw | | | |
210 | * | +----+------+------+ | |
211 | * | | ^ | |
212 | * | dequeue | | | |
213 | * +--------+-------+ | | | |
214 | * | | t >= 0-lag | | wakeup | |
215 | * | INACTIVE |<---------------+ | | |
216 | * | | sub_running_bw | | | |
217 | * +--------+-------+ | | | |
218 | * ^ | | | |
219 | * | t < 0-lag | | | |
220 | * | | | | |
221 | * | V | | |
222 | * | +----+------+------+ | |
223 | * | sub_running_bw | ACTIVE | | |
224 | * +-------------------+ | | |
225 | * inactive timer | non contending | | |
226 | * fired +------------------+ | |
227 | * | |
228 | * The task_non_contending() function is invoked when a task | |
229 | * blocks, and checks if the 0-lag time already passed or | |
230 | * not (in the first case, it directly updates running_bw; | |
231 | * in the second case, it arms the inactive timer). | |
232 | * | |
233 | * The task_contending() function is invoked when a task wakes | |
234 | * up, and checks if the task is still in the "ACTIVE non contending" | |
235 | * state or not (in the second case, it updates running_bw). | |
236 | */ | |
237 | static void task_non_contending(struct task_struct *p) | |
238 | { | |
239 | struct sched_dl_entity *dl_se = &p->dl; | |
240 | struct hrtimer *timer = &dl_se->inactive_timer; | |
241 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
242 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
243 | s64 zerolag_time; | |
244 | ||
245 | /* | |
246 | * If this is a non-deadline task that has been boosted, | |
247 | * do nothing | |
248 | */ | |
249 | if (dl_se->dl_runtime == 0) | |
250 | return; | |
251 | ||
794a56eb JL |
252 | if (dl_entity_is_special(dl_se)) |
253 | return; | |
254 | ||
209a0cbd LA |
255 | WARN_ON(dl_se->dl_non_contending); |
256 | ||
257 | zerolag_time = dl_se->deadline - | |
258 | div64_long((dl_se->runtime * dl_se->dl_period), | |
259 | dl_se->dl_runtime); | |
260 | ||
261 | /* | |
262 | * Using relative times instead of the absolute "0-lag time" | |
263 | * allows to simplify the code | |
264 | */ | |
265 | zerolag_time -= rq_clock(rq); | |
266 | ||
267 | /* | |
268 | * If the "0-lag time" already passed, decrease the active | |
269 | * utilization now, instead of starting a timer | |
270 | */ | |
1b02cd6a | 271 | if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) { |
209a0cbd | 272 | if (dl_task(p)) |
794a56eb | 273 | sub_running_bw(dl_se, dl_rq); |
387e3130 LA |
274 | if (!dl_task(p) || p->state == TASK_DEAD) { |
275 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); | |
276 | ||
8fd27231 | 277 | if (p->state == TASK_DEAD) |
794a56eb | 278 | sub_rq_bw(&p->dl, &rq->dl); |
387e3130 | 279 | raw_spin_lock(&dl_b->lock); |
8c0944ce | 280 | __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); |
209a0cbd | 281 | __dl_clear_params(p); |
387e3130 LA |
282 | raw_spin_unlock(&dl_b->lock); |
283 | } | |
209a0cbd LA |
284 | |
285 | return; | |
286 | } | |
287 | ||
288 | dl_se->dl_non_contending = 1; | |
289 | get_task_struct(p); | |
290 | hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL); | |
291 | } | |
292 | ||
8fd27231 | 293 | static void task_contending(struct sched_dl_entity *dl_se, int flags) |
209a0cbd LA |
294 | { |
295 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
296 | ||
297 | /* | |
298 | * If this is a non-deadline task that has been boosted, | |
299 | * do nothing | |
300 | */ | |
301 | if (dl_se->dl_runtime == 0) | |
302 | return; | |
303 | ||
8fd27231 | 304 | if (flags & ENQUEUE_MIGRATED) |
794a56eb | 305 | add_rq_bw(dl_se, dl_rq); |
8fd27231 | 306 | |
209a0cbd LA |
307 | if (dl_se->dl_non_contending) { |
308 | dl_se->dl_non_contending = 0; | |
309 | /* | |
310 | * If the timer handler is currently running and the | |
311 | * timer cannot be cancelled, inactive_task_timer() | |
312 | * will see that dl_not_contending is not set, and | |
313 | * will not touch the rq's active utilization, | |
314 | * so we are still safe. | |
315 | */ | |
316 | if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1) | |
317 | put_task_struct(dl_task_of(dl_se)); | |
318 | } else { | |
319 | /* | |
320 | * Since "dl_non_contending" is not set, the | |
321 | * task's utilization has already been removed from | |
322 | * active utilization (either when the task blocked, | |
323 | * when the "inactive timer" fired). | |
324 | * So, add it back. | |
325 | */ | |
794a56eb | 326 | add_running_bw(dl_se, dl_rq); |
209a0cbd LA |
327 | } |
328 | } | |
329 | ||
aab03e05 DF |
330 | static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq) |
331 | { | |
332 | struct sched_dl_entity *dl_se = &p->dl; | |
333 | ||
2161573e | 334 | return dl_rq->root.rb_leftmost == &dl_se->rb_node; |
aab03e05 DF |
335 | } |
336 | ||
332ac17e DF |
337 | void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime) |
338 | { | |
339 | raw_spin_lock_init(&dl_b->dl_runtime_lock); | |
340 | dl_b->dl_period = period; | |
341 | dl_b->dl_runtime = runtime; | |
342 | } | |
343 | ||
332ac17e DF |
344 | void init_dl_bw(struct dl_bw *dl_b) |
345 | { | |
346 | raw_spin_lock_init(&dl_b->lock); | |
347 | raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock); | |
1724813d | 348 | if (global_rt_runtime() == RUNTIME_INF) |
332ac17e DF |
349 | dl_b->bw = -1; |
350 | else | |
1724813d | 351 | dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime()); |
332ac17e DF |
352 | raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock); |
353 | dl_b->total_bw = 0; | |
354 | } | |
355 | ||
07c54f7a | 356 | void init_dl_rq(struct dl_rq *dl_rq) |
aab03e05 | 357 | { |
2161573e | 358 | dl_rq->root = RB_ROOT_CACHED; |
1baca4ce JL |
359 | |
360 | #ifdef CONFIG_SMP | |
361 | /* zero means no -deadline tasks */ | |
362 | dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0; | |
363 | ||
364 | dl_rq->dl_nr_migratory = 0; | |
365 | dl_rq->overloaded = 0; | |
2161573e | 366 | dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED; |
332ac17e DF |
367 | #else |
368 | init_dl_bw(&dl_rq->dl_bw); | |
1baca4ce | 369 | #endif |
e36d8677 LA |
370 | |
371 | dl_rq->running_bw = 0; | |
8fd27231 | 372 | dl_rq->this_bw = 0; |
4da3abce | 373 | init_dl_rq_bw_ratio(dl_rq); |
1baca4ce JL |
374 | } |
375 | ||
376 | #ifdef CONFIG_SMP | |
377 | ||
378 | static inline int dl_overloaded(struct rq *rq) | |
379 | { | |
380 | return atomic_read(&rq->rd->dlo_count); | |
381 | } | |
382 | ||
383 | static inline void dl_set_overload(struct rq *rq) | |
384 | { | |
385 | if (!rq->online) | |
386 | return; | |
387 | ||
388 | cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask); | |
389 | /* | |
390 | * Must be visible before the overload count is | |
391 | * set (as in sched_rt.c). | |
392 | * | |
393 | * Matched by the barrier in pull_dl_task(). | |
394 | */ | |
395 | smp_wmb(); | |
396 | atomic_inc(&rq->rd->dlo_count); | |
397 | } | |
398 | ||
399 | static inline void dl_clear_overload(struct rq *rq) | |
400 | { | |
401 | if (!rq->online) | |
402 | return; | |
403 | ||
404 | atomic_dec(&rq->rd->dlo_count); | |
405 | cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask); | |
406 | } | |
407 | ||
408 | static void update_dl_migration(struct dl_rq *dl_rq) | |
409 | { | |
995b9ea4 | 410 | if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) { |
1baca4ce JL |
411 | if (!dl_rq->overloaded) { |
412 | dl_set_overload(rq_of_dl_rq(dl_rq)); | |
413 | dl_rq->overloaded = 1; | |
414 | } | |
415 | } else if (dl_rq->overloaded) { | |
416 | dl_clear_overload(rq_of_dl_rq(dl_rq)); | |
417 | dl_rq->overloaded = 0; | |
418 | } | |
419 | } | |
420 | ||
421 | static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
422 | { | |
423 | struct task_struct *p = dl_task_of(dl_se); | |
1baca4ce | 424 | |
4b53a341 | 425 | if (p->nr_cpus_allowed > 1) |
1baca4ce JL |
426 | dl_rq->dl_nr_migratory++; |
427 | ||
428 | update_dl_migration(dl_rq); | |
429 | } | |
430 | ||
431 | static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
432 | { | |
433 | struct task_struct *p = dl_task_of(dl_se); | |
1baca4ce | 434 | |
4b53a341 | 435 | if (p->nr_cpus_allowed > 1) |
1baca4ce JL |
436 | dl_rq->dl_nr_migratory--; |
437 | ||
438 | update_dl_migration(dl_rq); | |
439 | } | |
440 | ||
441 | /* | |
442 | * The list of pushable -deadline task is not a plist, like in | |
443 | * sched_rt.c, it is an rb-tree with tasks ordered by deadline. | |
444 | */ | |
445 | static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
446 | { | |
447 | struct dl_rq *dl_rq = &rq->dl; | |
2161573e | 448 | struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_root.rb_node; |
1baca4ce JL |
449 | struct rb_node *parent = NULL; |
450 | struct task_struct *entry; | |
2161573e | 451 | bool leftmost = true; |
1baca4ce JL |
452 | |
453 | BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks)); | |
454 | ||
455 | while (*link) { | |
456 | parent = *link; | |
457 | entry = rb_entry(parent, struct task_struct, | |
458 | pushable_dl_tasks); | |
459 | if (dl_entity_preempt(&p->dl, &entry->dl)) | |
460 | link = &parent->rb_left; | |
461 | else { | |
462 | link = &parent->rb_right; | |
2161573e | 463 | leftmost = false; |
1baca4ce JL |
464 | } |
465 | } | |
466 | ||
2161573e | 467 | if (leftmost) |
7d92de3a | 468 | dl_rq->earliest_dl.next = p->dl.deadline; |
1baca4ce JL |
469 | |
470 | rb_link_node(&p->pushable_dl_tasks, parent, link); | |
2161573e DB |
471 | rb_insert_color_cached(&p->pushable_dl_tasks, |
472 | &dl_rq->pushable_dl_tasks_root, leftmost); | |
aab03e05 DF |
473 | } |
474 | ||
1baca4ce JL |
475 | static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) |
476 | { | |
477 | struct dl_rq *dl_rq = &rq->dl; | |
478 | ||
479 | if (RB_EMPTY_NODE(&p->pushable_dl_tasks)) | |
480 | return; | |
481 | ||
2161573e | 482 | if (dl_rq->pushable_dl_tasks_root.rb_leftmost == &p->pushable_dl_tasks) { |
1baca4ce JL |
483 | struct rb_node *next_node; |
484 | ||
485 | next_node = rb_next(&p->pushable_dl_tasks); | |
7d92de3a WL |
486 | if (next_node) { |
487 | dl_rq->earliest_dl.next = rb_entry(next_node, | |
488 | struct task_struct, pushable_dl_tasks)->dl.deadline; | |
489 | } | |
1baca4ce JL |
490 | } |
491 | ||
2161573e | 492 | rb_erase_cached(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); |
1baca4ce JL |
493 | RB_CLEAR_NODE(&p->pushable_dl_tasks); |
494 | } | |
495 | ||
496 | static inline int has_pushable_dl_tasks(struct rq *rq) | |
497 | { | |
2161573e | 498 | return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root); |
1baca4ce JL |
499 | } |
500 | ||
501 | static int push_dl_task(struct rq *rq); | |
502 | ||
dc877341 PZ |
503 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
504 | { | |
505 | return dl_task(prev); | |
506 | } | |
507 | ||
9916e214 PZ |
508 | static DEFINE_PER_CPU(struct callback_head, dl_push_head); |
509 | static DEFINE_PER_CPU(struct callback_head, dl_pull_head); | |
e3fca9e7 PZ |
510 | |
511 | static void push_dl_tasks(struct rq *); | |
9916e214 | 512 | static void pull_dl_task(struct rq *); |
e3fca9e7 | 513 | |
02d8ec94 | 514 | static inline void deadline_queue_push_tasks(struct rq *rq) |
dc877341 | 515 | { |
e3fca9e7 PZ |
516 | if (!has_pushable_dl_tasks(rq)) |
517 | return; | |
518 | ||
9916e214 PZ |
519 | queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks); |
520 | } | |
521 | ||
02d8ec94 | 522 | static inline void deadline_queue_pull_task(struct rq *rq) |
9916e214 PZ |
523 | { |
524 | queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task); | |
dc877341 PZ |
525 | } |
526 | ||
fa9c9d10 WL |
527 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq); |
528 | ||
a649f237 | 529 | static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p) |
fa9c9d10 WL |
530 | { |
531 | struct rq *later_rq = NULL; | |
59d06cea | 532 | struct dl_bw *dl_b; |
fa9c9d10 WL |
533 | |
534 | later_rq = find_lock_later_rq(p, rq); | |
fa9c9d10 WL |
535 | if (!later_rq) { |
536 | int cpu; | |
537 | ||
538 | /* | |
539 | * If we cannot preempt any rq, fall back to pick any | |
97fb7a0a | 540 | * online CPU: |
fa9c9d10 | 541 | */ |
3bd37062 | 542 | cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr); |
fa9c9d10 WL |
543 | if (cpu >= nr_cpu_ids) { |
544 | /* | |
97fb7a0a | 545 | * Failed to find any suitable CPU. |
fa9c9d10 WL |
546 | * The task will never come back! |
547 | */ | |
548 | BUG_ON(dl_bandwidth_enabled()); | |
549 | ||
550 | /* | |
551 | * If admission control is disabled we | |
552 | * try a little harder to let the task | |
553 | * run. | |
554 | */ | |
555 | cpu = cpumask_any(cpu_active_mask); | |
556 | } | |
557 | later_rq = cpu_rq(cpu); | |
558 | double_lock_balance(rq, later_rq); | |
559 | } | |
560 | ||
59d06cea JL |
561 | if (p->dl.dl_non_contending || p->dl.dl_throttled) { |
562 | /* | |
563 | * Inactive timer is armed (or callback is running, but | |
564 | * waiting for us to release rq locks). In any case, when it | |
565 | * will fire (or continue), it will see running_bw of this | |
566 | * task migrated to later_rq (and correctly handle it). | |
567 | */ | |
568 | sub_running_bw(&p->dl, &rq->dl); | |
569 | sub_rq_bw(&p->dl, &rq->dl); | |
570 | ||
571 | add_rq_bw(&p->dl, &later_rq->dl); | |
572 | add_running_bw(&p->dl, &later_rq->dl); | |
573 | } else { | |
574 | sub_rq_bw(&p->dl, &rq->dl); | |
575 | add_rq_bw(&p->dl, &later_rq->dl); | |
576 | } | |
577 | ||
578 | /* | |
579 | * And we finally need to fixup root_domain(s) bandwidth accounting, | |
580 | * since p is still hanging out in the old (now moved to default) root | |
581 | * domain. | |
582 | */ | |
583 | dl_b = &rq->rd->dl_bw; | |
584 | raw_spin_lock(&dl_b->lock); | |
585 | __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span)); | |
586 | raw_spin_unlock(&dl_b->lock); | |
587 | ||
588 | dl_b = &later_rq->rd->dl_bw; | |
589 | raw_spin_lock(&dl_b->lock); | |
590 | __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span)); | |
591 | raw_spin_unlock(&dl_b->lock); | |
592 | ||
fa9c9d10 | 593 | set_task_cpu(p, later_rq->cpu); |
a649f237 PZ |
594 | double_unlock_balance(later_rq, rq); |
595 | ||
596 | return later_rq; | |
fa9c9d10 WL |
597 | } |
598 | ||
1baca4ce JL |
599 | #else |
600 | ||
601 | static inline | |
602 | void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
603 | { | |
604 | } | |
605 | ||
606 | static inline | |
607 | void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
608 | { | |
609 | } | |
610 | ||
611 | static inline | |
612 | void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
613 | { | |
614 | } | |
615 | ||
616 | static inline | |
617 | void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
618 | { | |
619 | } | |
620 | ||
dc877341 PZ |
621 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
622 | { | |
623 | return false; | |
624 | } | |
625 | ||
0ea60c20 | 626 | static inline void pull_dl_task(struct rq *rq) |
dc877341 | 627 | { |
dc877341 PZ |
628 | } |
629 | ||
02d8ec94 | 630 | static inline void deadline_queue_push_tasks(struct rq *rq) |
dc877341 | 631 | { |
dc877341 PZ |
632 | } |
633 | ||
02d8ec94 | 634 | static inline void deadline_queue_pull_task(struct rq *rq) |
dc877341 PZ |
635 | { |
636 | } | |
1baca4ce JL |
637 | #endif /* CONFIG_SMP */ |
638 | ||
aab03e05 DF |
639 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags); |
640 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags); | |
97fb7a0a | 641 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags); |
aab03e05 DF |
642 | |
643 | /* | |
644 | * We are being explicitly informed that a new instance is starting, | |
645 | * and this means that: | |
646 | * - the absolute deadline of the entity has to be placed at | |
647 | * current time + relative deadline; | |
648 | * - the runtime of the entity has to be set to the maximum value. | |
649 | * | |
650 | * The capability of specifying such event is useful whenever a -deadline | |
651 | * entity wants to (try to!) synchronize its behaviour with the scheduler's | |
652 | * one, and to (try to!) reconcile itself with its own scheduling | |
653 | * parameters. | |
654 | */ | |
98b0a857 | 655 | static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se) |
aab03e05 DF |
656 | { |
657 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
658 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
659 | ||
98b0a857 | 660 | WARN_ON(dl_se->dl_boosted); |
72f9f3fd LA |
661 | WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline)); |
662 | ||
663 | /* | |
664 | * We are racing with the deadline timer. So, do nothing because | |
665 | * the deadline timer handler will take care of properly recharging | |
666 | * the runtime and postponing the deadline | |
667 | */ | |
668 | if (dl_se->dl_throttled) | |
669 | return; | |
aab03e05 DF |
670 | |
671 | /* | |
672 | * We use the regular wall clock time to set deadlines in the | |
673 | * future; in fact, we must consider execution overheads (time | |
674 | * spent on hardirq context, etc.). | |
675 | */ | |
98b0a857 JL |
676 | dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline; |
677 | dl_se->runtime = dl_se->dl_runtime; | |
aab03e05 DF |
678 | } |
679 | ||
680 | /* | |
681 | * Pure Earliest Deadline First (EDF) scheduling does not deal with the | |
682 | * possibility of a entity lasting more than what it declared, and thus | |
683 | * exhausting its runtime. | |
684 | * | |
685 | * Here we are interested in making runtime overrun possible, but we do | |
686 | * not want a entity which is misbehaving to affect the scheduling of all | |
687 | * other entities. | |
688 | * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS) | |
689 | * is used, in order to confine each entity within its own bandwidth. | |
690 | * | |
691 | * This function deals exactly with that, and ensures that when the runtime | |
692 | * of a entity is replenished, its deadline is also postponed. That ensures | |
693 | * the overrunning entity can't interfere with other entity in the system and | |
694 | * can't make them miss their deadlines. Reasons why this kind of overruns | |
695 | * could happen are, typically, a entity voluntarily trying to overcome its | |
1b09d29b | 696 | * runtime, or it just underestimated it during sched_setattr(). |
aab03e05 | 697 | */ |
2d3d891d DF |
698 | static void replenish_dl_entity(struct sched_dl_entity *dl_se, |
699 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
700 | { |
701 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
702 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
703 | ||
2d3d891d DF |
704 | BUG_ON(pi_se->dl_runtime <= 0); |
705 | ||
706 | /* | |
707 | * This could be the case for a !-dl task that is boosted. | |
708 | * Just go with full inherited parameters. | |
709 | */ | |
710 | if (dl_se->dl_deadline == 0) { | |
711 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | |
712 | dl_se->runtime = pi_se->dl_runtime; | |
713 | } | |
714 | ||
48be3a67 PZ |
715 | if (dl_se->dl_yielded && dl_se->runtime > 0) |
716 | dl_se->runtime = 0; | |
717 | ||
aab03e05 DF |
718 | /* |
719 | * We keep moving the deadline away until we get some | |
720 | * available runtime for the entity. This ensures correct | |
721 | * handling of situations where the runtime overrun is | |
722 | * arbitrary large. | |
723 | */ | |
724 | while (dl_se->runtime <= 0) { | |
2d3d891d DF |
725 | dl_se->deadline += pi_se->dl_period; |
726 | dl_se->runtime += pi_se->dl_runtime; | |
aab03e05 DF |
727 | } |
728 | ||
729 | /* | |
730 | * At this point, the deadline really should be "in | |
731 | * the future" with respect to rq->clock. If it's | |
732 | * not, we are, for some reason, lagging too much! | |
733 | * Anyway, after having warn userspace abut that, | |
734 | * we still try to keep the things running by | |
735 | * resetting the deadline and the budget of the | |
736 | * entity. | |
737 | */ | |
738 | if (dl_time_before(dl_se->deadline, rq_clock(rq))) { | |
c219b7dd | 739 | printk_deferred_once("sched: DL replenish lagged too much\n"); |
2d3d891d DF |
740 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
741 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 | 742 | } |
1019a359 PZ |
743 | |
744 | if (dl_se->dl_yielded) | |
745 | dl_se->dl_yielded = 0; | |
746 | if (dl_se->dl_throttled) | |
747 | dl_se->dl_throttled = 0; | |
aab03e05 DF |
748 | } |
749 | ||
750 | /* | |
751 | * Here we check if --at time t-- an entity (which is probably being | |
752 | * [re]activated or, in general, enqueued) can use its remaining runtime | |
753 | * and its current deadline _without_ exceeding the bandwidth it is | |
754 | * assigned (function returns true if it can't). We are in fact applying | |
755 | * one of the CBS rules: when a task wakes up, if the residual runtime | |
756 | * over residual deadline fits within the allocated bandwidth, then we | |
757 | * can keep the current (absolute) deadline and residual budget without | |
758 | * disrupting the schedulability of the system. Otherwise, we should | |
759 | * refill the runtime and set the deadline a period in the future, | |
760 | * because keeping the current (absolute) deadline of the task would | |
712e5e34 | 761 | * result in breaking guarantees promised to other tasks (refer to |
d6a3b247 | 762 | * Documentation/scheduler/sched-deadline.rst for more information). |
aab03e05 DF |
763 | * |
764 | * This function returns true if: | |
765 | * | |
2317d5f1 | 766 | * runtime / (deadline - t) > dl_runtime / dl_deadline , |
aab03e05 DF |
767 | * |
768 | * IOW we can't recycle current parameters. | |
755378a4 | 769 | * |
2317d5f1 | 770 | * Notice that the bandwidth check is done against the deadline. For |
755378a4 | 771 | * task with deadline equal to period this is the same of using |
2317d5f1 | 772 | * dl_period instead of dl_deadline in the equation above. |
aab03e05 | 773 | */ |
2d3d891d DF |
774 | static bool dl_entity_overflow(struct sched_dl_entity *dl_se, |
775 | struct sched_dl_entity *pi_se, u64 t) | |
aab03e05 DF |
776 | { |
777 | u64 left, right; | |
778 | ||
779 | /* | |
780 | * left and right are the two sides of the equation above, | |
781 | * after a bit of shuffling to use multiplications instead | |
782 | * of divisions. | |
783 | * | |
784 | * Note that none of the time values involved in the two | |
785 | * multiplications are absolute: dl_deadline and dl_runtime | |
786 | * are the relative deadline and the maximum runtime of each | |
787 | * instance, runtime is the runtime left for the last instance | |
788 | * and (deadline - t), since t is rq->clock, is the time left | |
789 | * to the (absolute) deadline. Even if overflowing the u64 type | |
790 | * is very unlikely to occur in both cases, here we scale down | |
791 | * as we want to avoid that risk at all. Scaling down by 10 | |
792 | * means that we reduce granularity to 1us. We are fine with it, | |
793 | * since this is only a true/false check and, anyway, thinking | |
794 | * of anything below microseconds resolution is actually fiction | |
795 | * (but still we want to give the user that illusion >;). | |
796 | */ | |
2317d5f1 | 797 | left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE); |
332ac17e DF |
798 | right = ((dl_se->deadline - t) >> DL_SCALE) * |
799 | (pi_se->dl_runtime >> DL_SCALE); | |
aab03e05 DF |
800 | |
801 | return dl_time_before(right, left); | |
802 | } | |
803 | ||
804 | /* | |
3effcb42 DBO |
805 | * Revised wakeup rule [1]: For self-suspending tasks, rather then |
806 | * re-initializing task's runtime and deadline, the revised wakeup | |
807 | * rule adjusts the task's runtime to avoid the task to overrun its | |
808 | * density. | |
aab03e05 | 809 | * |
3effcb42 DBO |
810 | * Reasoning: a task may overrun the density if: |
811 | * runtime / (deadline - t) > dl_runtime / dl_deadline | |
812 | * | |
813 | * Therefore, runtime can be adjusted to: | |
814 | * runtime = (dl_runtime / dl_deadline) * (deadline - t) | |
815 | * | |
816 | * In such way that runtime will be equal to the maximum density | |
817 | * the task can use without breaking any rule. | |
818 | * | |
819 | * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant | |
820 | * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24. | |
821 | */ | |
822 | static void | |
823 | update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq) | |
824 | { | |
825 | u64 laxity = dl_se->deadline - rq_clock(rq); | |
826 | ||
827 | /* | |
828 | * If the task has deadline < period, and the deadline is in the past, | |
829 | * it should already be throttled before this check. | |
830 | * | |
831 | * See update_dl_entity() comments for further details. | |
832 | */ | |
833 | WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq))); | |
834 | ||
835 | dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT; | |
836 | } | |
837 | ||
838 | /* | |
839 | * Regarding the deadline, a task with implicit deadline has a relative | |
840 | * deadline == relative period. A task with constrained deadline has a | |
841 | * relative deadline <= relative period. | |
842 | * | |
843 | * We support constrained deadline tasks. However, there are some restrictions | |
844 | * applied only for tasks which do not have an implicit deadline. See | |
845 | * update_dl_entity() to know more about such restrictions. | |
846 | * | |
847 | * The dl_is_implicit() returns true if the task has an implicit deadline. | |
848 | */ | |
849 | static inline bool dl_is_implicit(struct sched_dl_entity *dl_se) | |
850 | { | |
851 | return dl_se->dl_deadline == dl_se->dl_period; | |
852 | } | |
853 | ||
854 | /* | |
855 | * When a deadline entity is placed in the runqueue, its runtime and deadline | |
856 | * might need to be updated. This is done by a CBS wake up rule. There are two | |
857 | * different rules: 1) the original CBS; and 2) the Revisited CBS. | |
858 | * | |
859 | * When the task is starting a new period, the Original CBS is used. In this | |
860 | * case, the runtime is replenished and a new absolute deadline is set. | |
861 | * | |
862 | * When a task is queued before the begin of the next period, using the | |
863 | * remaining runtime and deadline could make the entity to overflow, see | |
864 | * dl_entity_overflow() to find more about runtime overflow. When such case | |
865 | * is detected, the runtime and deadline need to be updated. | |
866 | * | |
867 | * If the task has an implicit deadline, i.e., deadline == period, the Original | |
868 | * CBS is applied. the runtime is replenished and a new absolute deadline is | |
869 | * set, as in the previous cases. | |
870 | * | |
871 | * However, the Original CBS does not work properly for tasks with | |
872 | * deadline < period, which are said to have a constrained deadline. By | |
873 | * applying the Original CBS, a constrained deadline task would be able to run | |
874 | * runtime/deadline in a period. With deadline < period, the task would | |
875 | * overrun the runtime/period allowed bandwidth, breaking the admission test. | |
876 | * | |
877 | * In order to prevent this misbehave, the Revisited CBS is used for | |
878 | * constrained deadline tasks when a runtime overflow is detected. In the | |
879 | * Revisited CBS, rather than replenishing & setting a new absolute deadline, | |
880 | * the remaining runtime of the task is reduced to avoid runtime overflow. | |
881 | * Please refer to the comments update_dl_revised_wakeup() function to find | |
882 | * more about the Revised CBS rule. | |
aab03e05 | 883 | */ |
2d3d891d DF |
884 | static void update_dl_entity(struct sched_dl_entity *dl_se, |
885 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
886 | { |
887 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
888 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
889 | ||
aab03e05 | 890 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) || |
2d3d891d | 891 | dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { |
3effcb42 DBO |
892 | |
893 | if (unlikely(!dl_is_implicit(dl_se) && | |
894 | !dl_time_before(dl_se->deadline, rq_clock(rq)) && | |
895 | !dl_se->dl_boosted)){ | |
896 | update_dl_revised_wakeup(dl_se, rq); | |
897 | return; | |
898 | } | |
899 | ||
2d3d891d DF |
900 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
901 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 DF |
902 | } |
903 | } | |
904 | ||
5ac69d37 DBO |
905 | static inline u64 dl_next_period(struct sched_dl_entity *dl_se) |
906 | { | |
907 | return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period; | |
908 | } | |
909 | ||
aab03e05 DF |
910 | /* |
911 | * If the entity depleted all its runtime, and if we want it to sleep | |
912 | * while waiting for some new execution time to become available, we | |
5ac69d37 | 913 | * set the bandwidth replenishment timer to the replenishment instant |
aab03e05 DF |
914 | * and try to activate it. |
915 | * | |
916 | * Notice that it is important for the caller to know if the timer | |
917 | * actually started or not (i.e., the replenishment instant is in | |
918 | * the future or in the past). | |
919 | */ | |
a649f237 | 920 | static int start_dl_timer(struct task_struct *p) |
aab03e05 | 921 | { |
a649f237 PZ |
922 | struct sched_dl_entity *dl_se = &p->dl; |
923 | struct hrtimer *timer = &dl_se->dl_timer; | |
924 | struct rq *rq = task_rq(p); | |
aab03e05 | 925 | ktime_t now, act; |
aab03e05 DF |
926 | s64 delta; |
927 | ||
a649f237 PZ |
928 | lockdep_assert_held(&rq->lock); |
929 | ||
aab03e05 DF |
930 | /* |
931 | * We want the timer to fire at the deadline, but considering | |
932 | * that it is actually coming from rq->clock and not from | |
933 | * hrtimer's time base reading. | |
934 | */ | |
5ac69d37 | 935 | act = ns_to_ktime(dl_next_period(dl_se)); |
a649f237 | 936 | now = hrtimer_cb_get_time(timer); |
aab03e05 DF |
937 | delta = ktime_to_ns(now) - rq_clock(rq); |
938 | act = ktime_add_ns(act, delta); | |
939 | ||
940 | /* | |
941 | * If the expiry time already passed, e.g., because the value | |
942 | * chosen as the deadline is too small, don't even try to | |
943 | * start the timer in the past! | |
944 | */ | |
945 | if (ktime_us_delta(act, now) < 0) | |
946 | return 0; | |
947 | ||
a649f237 PZ |
948 | /* |
949 | * !enqueued will guarantee another callback; even if one is already in | |
950 | * progress. This ensures a balanced {get,put}_task_struct(). | |
951 | * | |
952 | * The race against __run_timer() clearing the enqueued state is | |
953 | * harmless because we're holding task_rq()->lock, therefore the timer | |
954 | * expiring after we've done the check will wait on its task_rq_lock() | |
955 | * and observe our state. | |
956 | */ | |
957 | if (!hrtimer_is_queued(timer)) { | |
958 | get_task_struct(p); | |
959 | hrtimer_start(timer, act, HRTIMER_MODE_ABS); | |
960 | } | |
aab03e05 | 961 | |
cc9684d3 | 962 | return 1; |
aab03e05 DF |
963 | } |
964 | ||
965 | /* | |
966 | * This is the bandwidth enforcement timer callback. If here, we know | |
967 | * a task is not on its dl_rq, since the fact that the timer was running | |
968 | * means the task is throttled and needs a runtime replenishment. | |
969 | * | |
970 | * However, what we actually do depends on the fact the task is active, | |
971 | * (it is on its rq) or has been removed from there by a call to | |
972 | * dequeue_task_dl(). In the former case we must issue the runtime | |
973 | * replenishment and add the task back to the dl_rq; in the latter, we just | |
974 | * do nothing but clearing dl_throttled, so that runtime and deadline | |
975 | * updating (and the queueing back to dl_rq) will be done by the | |
976 | * next call to enqueue_task_dl(). | |
977 | */ | |
978 | static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) | |
979 | { | |
980 | struct sched_dl_entity *dl_se = container_of(timer, | |
981 | struct sched_dl_entity, | |
982 | dl_timer); | |
983 | struct task_struct *p = dl_task_of(dl_se); | |
eb580751 | 984 | struct rq_flags rf; |
0f397f2c | 985 | struct rq *rq; |
3960c8c0 | 986 | |
eb580751 | 987 | rq = task_rq_lock(p, &rf); |
0f397f2c | 988 | |
aab03e05 | 989 | /* |
a649f237 | 990 | * The task might have changed its scheduling policy to something |
9846d50d | 991 | * different than SCHED_DEADLINE (through switched_from_dl()). |
a649f237 | 992 | */ |
209a0cbd | 993 | if (!dl_task(p)) |
a649f237 | 994 | goto unlock; |
a649f237 | 995 | |
a649f237 PZ |
996 | /* |
997 | * The task might have been boosted by someone else and might be in the | |
998 | * boosting/deboosting path, its not throttled. | |
999 | */ | |
1000 | if (dl_se->dl_boosted) | |
1001 | goto unlock; | |
a79ec89f | 1002 | |
fa9c9d10 | 1003 | /* |
a649f237 PZ |
1004 | * Spurious timer due to start_dl_timer() race; or we already received |
1005 | * a replenishment from rt_mutex_setprio(). | |
fa9c9d10 | 1006 | */ |
a649f237 | 1007 | if (!dl_se->dl_throttled) |
fa9c9d10 | 1008 | goto unlock; |
a649f237 PZ |
1009 | |
1010 | sched_clock_tick(); | |
1011 | update_rq_clock(rq); | |
fa9c9d10 | 1012 | |
a79ec89f KT |
1013 | /* |
1014 | * If the throttle happened during sched-out; like: | |
1015 | * | |
1016 | * schedule() | |
1017 | * deactivate_task() | |
1018 | * dequeue_task_dl() | |
1019 | * update_curr_dl() | |
1020 | * start_dl_timer() | |
1021 | * __dequeue_task_dl() | |
1022 | * prev->on_rq = 0; | |
1023 | * | |
1024 | * We can be both throttled and !queued. Replenish the counter | |
1025 | * but do not enqueue -- wait for our wakeup to do that. | |
1026 | */ | |
1027 | if (!task_on_rq_queued(p)) { | |
1028 | replenish_dl_entity(dl_se, dl_se); | |
1029 | goto unlock; | |
1030 | } | |
1031 | ||
1baca4ce | 1032 | #ifdef CONFIG_SMP |
c0c8c9fa | 1033 | if (unlikely(!rq->online)) { |
61c7aca6 WL |
1034 | /* |
1035 | * If the runqueue is no longer available, migrate the | |
1036 | * task elsewhere. This necessarily changes rq. | |
1037 | */ | |
c0c8c9fa | 1038 | lockdep_unpin_lock(&rq->lock, rf.cookie); |
a649f237 | 1039 | rq = dl_task_offline_migration(rq, p); |
c0c8c9fa | 1040 | rf.cookie = lockdep_pin_lock(&rq->lock); |
dcc3b5ff | 1041 | update_rq_clock(rq); |
61c7aca6 WL |
1042 | |
1043 | /* | |
1044 | * Now that the task has been migrated to the new RQ and we | |
1045 | * have that locked, proceed as normal and enqueue the task | |
1046 | * there. | |
1047 | */ | |
c0c8c9fa | 1048 | } |
61c7aca6 | 1049 | #endif |
a649f237 | 1050 | |
61c7aca6 WL |
1051 | enqueue_task_dl(rq, p, ENQUEUE_REPLENISH); |
1052 | if (dl_task(rq->curr)) | |
1053 | check_preempt_curr_dl(rq, p, 0); | |
1054 | else | |
1055 | resched_curr(rq); | |
a649f237 | 1056 | |
61c7aca6 | 1057 | #ifdef CONFIG_SMP |
a649f237 PZ |
1058 | /* |
1059 | * Queueing this task back might have overloaded rq, check if we need | |
1060 | * to kick someone away. | |
1019a359 | 1061 | */ |
0aaafaab PZ |
1062 | if (has_pushable_dl_tasks(rq)) { |
1063 | /* | |
1064 | * Nothing relies on rq->lock after this, so its safe to drop | |
1065 | * rq->lock. | |
1066 | */ | |
d8ac8971 | 1067 | rq_unpin_lock(rq, &rf); |
1019a359 | 1068 | push_dl_task(rq); |
d8ac8971 | 1069 | rq_repin_lock(rq, &rf); |
0aaafaab | 1070 | } |
1baca4ce | 1071 | #endif |
a649f237 | 1072 | |
aab03e05 | 1073 | unlock: |
eb580751 | 1074 | task_rq_unlock(rq, p, &rf); |
aab03e05 | 1075 | |
a649f237 PZ |
1076 | /* |
1077 | * This can free the task_struct, including this hrtimer, do not touch | |
1078 | * anything related to that after this. | |
1079 | */ | |
1080 | put_task_struct(p); | |
1081 | ||
aab03e05 DF |
1082 | return HRTIMER_NORESTART; |
1083 | } | |
1084 | ||
1085 | void init_dl_task_timer(struct sched_dl_entity *dl_se) | |
1086 | { | |
1087 | struct hrtimer *timer = &dl_se->dl_timer; | |
1088 | ||
aab03e05 DF |
1089 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
1090 | timer->function = dl_task_timer; | |
1091 | } | |
1092 | ||
df8eac8c DBO |
1093 | /* |
1094 | * During the activation, CBS checks if it can reuse the current task's | |
1095 | * runtime and period. If the deadline of the task is in the past, CBS | |
1096 | * cannot use the runtime, and so it replenishes the task. This rule | |
1097 | * works fine for implicit deadline tasks (deadline == period), and the | |
1098 | * CBS was designed for implicit deadline tasks. However, a task with | |
1099 | * constrained deadline (deadine < period) might be awakened after the | |
1100 | * deadline, but before the next period. In this case, replenishing the | |
1101 | * task would allow it to run for runtime / deadline. As in this case | |
1102 | * deadline < period, CBS enables a task to run for more than the | |
1103 | * runtime / period. In a very loaded system, this can cause a domino | |
1104 | * effect, making other tasks miss their deadlines. | |
1105 | * | |
1106 | * To avoid this problem, in the activation of a constrained deadline | |
1107 | * task after the deadline but before the next period, throttle the | |
1108 | * task and set the replenishing timer to the begin of the next period, | |
1109 | * unless it is boosted. | |
1110 | */ | |
1111 | static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se) | |
1112 | { | |
1113 | struct task_struct *p = dl_task_of(dl_se); | |
1114 | struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se)); | |
1115 | ||
1116 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) && | |
1117 | dl_time_before(rq_clock(rq), dl_next_period(dl_se))) { | |
1118 | if (unlikely(dl_se->dl_boosted || !start_dl_timer(p))) | |
1119 | return; | |
1120 | dl_se->dl_throttled = 1; | |
ae83b56a XP |
1121 | if (dl_se->runtime > 0) |
1122 | dl_se->runtime = 0; | |
df8eac8c DBO |
1123 | } |
1124 | } | |
1125 | ||
aab03e05 | 1126 | static |
6fab5410 | 1127 | int dl_runtime_exceeded(struct sched_dl_entity *dl_se) |
aab03e05 | 1128 | { |
269ad801 | 1129 | return (dl_se->runtime <= 0); |
aab03e05 DF |
1130 | } |
1131 | ||
faa59937 JL |
1132 | extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq); |
1133 | ||
c52f14d3 LA |
1134 | /* |
1135 | * This function implements the GRUB accounting rule: | |
1136 | * according to the GRUB reclaiming algorithm, the runtime is | |
daec5798 LA |
1137 | * not decreased as "dq = -dt", but as |
1138 | * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt", | |
1139 | * where u is the utilization of the task, Umax is the maximum reclaimable | |
1140 | * utilization, Uinact is the (per-runqueue) inactive utilization, computed | |
1141 | * as the difference between the "total runqueue utilization" and the | |
1142 | * runqueue active utilization, and Uextra is the (per runqueue) extra | |
1143 | * reclaimable utilization. | |
9f0d1a50 | 1144 | * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations |
daec5798 LA |
1145 | * multiplied by 2^BW_SHIFT, the result has to be shifted right by |
1146 | * BW_SHIFT. | |
1147 | * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT, | |
1148 | * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT. | |
1149 | * Since delta is a 64 bit variable, to have an overflow its value | |
1150 | * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds. | |
1151 | * So, overflow is not an issue here. | |
c52f14d3 | 1152 | */ |
3febfc8a | 1153 | static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se) |
c52f14d3 | 1154 | { |
9f0d1a50 LA |
1155 | u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */ |
1156 | u64 u_act; | |
daec5798 | 1157 | u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT; |
c52f14d3 | 1158 | |
9f0d1a50 | 1159 | /* |
daec5798 LA |
1160 | * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)}, |
1161 | * we compare u_inact + rq->dl.extra_bw with | |
1162 | * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because | |
1163 | * u_inact + rq->dl.extra_bw can be larger than | |
1164 | * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative | |
1165 | * leading to wrong results) | |
9f0d1a50 | 1166 | */ |
daec5798 LA |
1167 | if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min) |
1168 | u_act = u_act_min; | |
9f0d1a50 | 1169 | else |
daec5798 | 1170 | u_act = BW_UNIT - u_inact - rq->dl.extra_bw; |
9f0d1a50 LA |
1171 | |
1172 | return (delta * u_act) >> BW_SHIFT; | |
c52f14d3 LA |
1173 | } |
1174 | ||
aab03e05 DF |
1175 | /* |
1176 | * Update the current task's runtime statistics (provided it is still | |
1177 | * a -deadline task and has not been removed from the dl_rq). | |
1178 | */ | |
1179 | static void update_curr_dl(struct rq *rq) | |
1180 | { | |
1181 | struct task_struct *curr = rq->curr; | |
1182 | struct sched_dl_entity *dl_se = &curr->dl; | |
07881166 JL |
1183 | u64 delta_exec, scaled_delta_exec; |
1184 | int cpu = cpu_of(rq); | |
6fe0ce1e | 1185 | u64 now; |
aab03e05 DF |
1186 | |
1187 | if (!dl_task(curr) || !on_dl_rq(dl_se)) | |
1188 | return; | |
1189 | ||
1190 | /* | |
1191 | * Consumed budget is computed considering the time as | |
1192 | * observed by schedulable tasks (excluding time spent | |
1193 | * in hardirq context, etc.). Deadlines are instead | |
1194 | * computed using hard walltime. This seems to be the more | |
1195 | * natural solution, but the full ramifications of this | |
1196 | * approach need further study. | |
1197 | */ | |
6fe0ce1e WY |
1198 | now = rq_clock_task(rq); |
1199 | delta_exec = now - curr->se.exec_start; | |
48be3a67 PZ |
1200 | if (unlikely((s64)delta_exec <= 0)) { |
1201 | if (unlikely(dl_se->dl_yielded)) | |
1202 | goto throttle; | |
734ff2a7 | 1203 | return; |
48be3a67 | 1204 | } |
aab03e05 DF |
1205 | |
1206 | schedstat_set(curr->se.statistics.exec_max, | |
1207 | max(curr->se.statistics.exec_max, delta_exec)); | |
1208 | ||
1209 | curr->se.sum_exec_runtime += delta_exec; | |
1210 | account_group_exec_runtime(curr, delta_exec); | |
1211 | ||
6fe0ce1e | 1212 | curr->se.exec_start = now; |
d2cc5ed6 | 1213 | cgroup_account_cputime(curr, delta_exec); |
aab03e05 | 1214 | |
794a56eb JL |
1215 | if (dl_entity_is_special(dl_se)) |
1216 | return; | |
1217 | ||
07881166 JL |
1218 | /* |
1219 | * For tasks that participate in GRUB, we implement GRUB-PA: the | |
1220 | * spare reclaimed bandwidth is used to clock down frequency. | |
1221 | * | |
1222 | * For the others, we still need to scale reservation parameters | |
1223 | * according to current frequency and CPU maximum capacity. | |
1224 | */ | |
1225 | if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) { | |
1226 | scaled_delta_exec = grub_reclaim(delta_exec, | |
1227 | rq, | |
1228 | &curr->dl); | |
1229 | } else { | |
1230 | unsigned long scale_freq = arch_scale_freq_capacity(cpu); | |
8ec59c0f | 1231 | unsigned long scale_cpu = arch_scale_cpu_capacity(cpu); |
07881166 JL |
1232 | |
1233 | scaled_delta_exec = cap_scale(delta_exec, scale_freq); | |
1234 | scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu); | |
1235 | } | |
1236 | ||
1237 | dl_se->runtime -= scaled_delta_exec; | |
48be3a67 PZ |
1238 | |
1239 | throttle: | |
1240 | if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) { | |
1019a359 | 1241 | dl_se->dl_throttled = 1; |
34be3930 JL |
1242 | |
1243 | /* If requested, inform the user about runtime overruns. */ | |
1244 | if (dl_runtime_exceeded(dl_se) && | |
1245 | (dl_se->flags & SCHED_FLAG_DL_OVERRUN)) | |
1246 | dl_se->dl_overrun = 1; | |
1247 | ||
aab03e05 | 1248 | __dequeue_task_dl(rq, curr, 0); |
a649f237 | 1249 | if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr))) |
aab03e05 DF |
1250 | enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH); |
1251 | ||
1252 | if (!is_leftmost(curr, &rq->dl)) | |
8875125e | 1253 | resched_curr(rq); |
aab03e05 | 1254 | } |
1724813d PZ |
1255 | |
1256 | /* | |
1257 | * Because -- for now -- we share the rt bandwidth, we need to | |
1258 | * account our runtime there too, otherwise actual rt tasks | |
1259 | * would be able to exceed the shared quota. | |
1260 | * | |
1261 | * Account to the root rt group for now. | |
1262 | * | |
1263 | * The solution we're working towards is having the RT groups scheduled | |
1264 | * using deadline servers -- however there's a few nasties to figure | |
1265 | * out before that can happen. | |
1266 | */ | |
1267 | if (rt_bandwidth_enabled()) { | |
1268 | struct rt_rq *rt_rq = &rq->rt; | |
1269 | ||
1270 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
1724813d PZ |
1271 | /* |
1272 | * We'll let actual RT tasks worry about the overflow here, we | |
faa59937 JL |
1273 | * have our own CBS to keep us inline; only account when RT |
1274 | * bandwidth is relevant. | |
1724813d | 1275 | */ |
faa59937 JL |
1276 | if (sched_rt_bandwidth_account(rt_rq)) |
1277 | rt_rq->rt_time += delta_exec; | |
1724813d PZ |
1278 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
1279 | } | |
aab03e05 DF |
1280 | } |
1281 | ||
209a0cbd LA |
1282 | static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer) |
1283 | { | |
1284 | struct sched_dl_entity *dl_se = container_of(timer, | |
1285 | struct sched_dl_entity, | |
1286 | inactive_timer); | |
1287 | struct task_struct *p = dl_task_of(dl_se); | |
1288 | struct rq_flags rf; | |
1289 | struct rq *rq; | |
1290 | ||
1291 | rq = task_rq_lock(p, &rf); | |
1292 | ||
ecda2b66 JL |
1293 | sched_clock_tick(); |
1294 | update_rq_clock(rq); | |
1295 | ||
209a0cbd | 1296 | if (!dl_task(p) || p->state == TASK_DEAD) { |
387e3130 LA |
1297 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); |
1298 | ||
209a0cbd | 1299 | if (p->state == TASK_DEAD && dl_se->dl_non_contending) { |
794a56eb JL |
1300 | sub_running_bw(&p->dl, dl_rq_of_se(&p->dl)); |
1301 | sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl)); | |
209a0cbd LA |
1302 | dl_se->dl_non_contending = 0; |
1303 | } | |
387e3130 LA |
1304 | |
1305 | raw_spin_lock(&dl_b->lock); | |
8c0944ce | 1306 | __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); |
387e3130 | 1307 | raw_spin_unlock(&dl_b->lock); |
209a0cbd LA |
1308 | __dl_clear_params(p); |
1309 | ||
1310 | goto unlock; | |
1311 | } | |
1312 | if (dl_se->dl_non_contending == 0) | |
1313 | goto unlock; | |
1314 | ||
794a56eb | 1315 | sub_running_bw(dl_se, &rq->dl); |
209a0cbd LA |
1316 | dl_se->dl_non_contending = 0; |
1317 | unlock: | |
1318 | task_rq_unlock(rq, p, &rf); | |
1319 | put_task_struct(p); | |
1320 | ||
1321 | return HRTIMER_NORESTART; | |
1322 | } | |
1323 | ||
1324 | void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se) | |
1325 | { | |
1326 | struct hrtimer *timer = &dl_se->inactive_timer; | |
1327 | ||
1328 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1329 | timer->function = inactive_task_timer; | |
1330 | } | |
1331 | ||
1baca4ce JL |
1332 | #ifdef CONFIG_SMP |
1333 | ||
1baca4ce JL |
1334 | static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) |
1335 | { | |
1336 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
1337 | ||
1338 | if (dl_rq->earliest_dl.curr == 0 || | |
1339 | dl_time_before(deadline, dl_rq->earliest_dl.curr)) { | |
1baca4ce | 1340 | dl_rq->earliest_dl.curr = deadline; |
d8206bb3 | 1341 | cpudl_set(&rq->rd->cpudl, rq->cpu, deadline); |
1baca4ce JL |
1342 | } |
1343 | } | |
1344 | ||
1345 | static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) | |
1346 | { | |
1347 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
1348 | ||
1349 | /* | |
1350 | * Since we may have removed our earliest (and/or next earliest) | |
1351 | * task we must recompute them. | |
1352 | */ | |
1353 | if (!dl_rq->dl_nr_running) { | |
1354 | dl_rq->earliest_dl.curr = 0; | |
1355 | dl_rq->earliest_dl.next = 0; | |
d8206bb3 | 1356 | cpudl_clear(&rq->rd->cpudl, rq->cpu); |
1baca4ce | 1357 | } else { |
2161573e | 1358 | struct rb_node *leftmost = dl_rq->root.rb_leftmost; |
1baca4ce JL |
1359 | struct sched_dl_entity *entry; |
1360 | ||
1361 | entry = rb_entry(leftmost, struct sched_dl_entity, rb_node); | |
1362 | dl_rq->earliest_dl.curr = entry->deadline; | |
d8206bb3 | 1363 | cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline); |
1baca4ce JL |
1364 | } |
1365 | } | |
1366 | ||
1367 | #else | |
1368 | ||
1369 | static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
1370 | static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
1371 | ||
1372 | #endif /* CONFIG_SMP */ | |
1373 | ||
1374 | static inline | |
1375 | void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
1376 | { | |
1377 | int prio = dl_task_of(dl_se)->prio; | |
1378 | u64 deadline = dl_se->deadline; | |
1379 | ||
1380 | WARN_ON(!dl_prio(prio)); | |
1381 | dl_rq->dl_nr_running++; | |
72465447 | 1382 | add_nr_running(rq_of_dl_rq(dl_rq), 1); |
1baca4ce JL |
1383 | |
1384 | inc_dl_deadline(dl_rq, deadline); | |
1385 | inc_dl_migration(dl_se, dl_rq); | |
1386 | } | |
1387 | ||
1388 | static inline | |
1389 | void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
1390 | { | |
1391 | int prio = dl_task_of(dl_se)->prio; | |
1392 | ||
1393 | WARN_ON(!dl_prio(prio)); | |
1394 | WARN_ON(!dl_rq->dl_nr_running); | |
1395 | dl_rq->dl_nr_running--; | |
72465447 | 1396 | sub_nr_running(rq_of_dl_rq(dl_rq), 1); |
1baca4ce JL |
1397 | |
1398 | dec_dl_deadline(dl_rq, dl_se->deadline); | |
1399 | dec_dl_migration(dl_se, dl_rq); | |
1400 | } | |
1401 | ||
aab03e05 DF |
1402 | static void __enqueue_dl_entity(struct sched_dl_entity *dl_se) |
1403 | { | |
1404 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
2161573e | 1405 | struct rb_node **link = &dl_rq->root.rb_root.rb_node; |
aab03e05 DF |
1406 | struct rb_node *parent = NULL; |
1407 | struct sched_dl_entity *entry; | |
1408 | int leftmost = 1; | |
1409 | ||
1410 | BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node)); | |
1411 | ||
1412 | while (*link) { | |
1413 | parent = *link; | |
1414 | entry = rb_entry(parent, struct sched_dl_entity, rb_node); | |
1415 | if (dl_time_before(dl_se->deadline, entry->deadline)) | |
1416 | link = &parent->rb_left; | |
1417 | else { | |
1418 | link = &parent->rb_right; | |
1419 | leftmost = 0; | |
1420 | } | |
1421 | } | |
1422 | ||
aab03e05 | 1423 | rb_link_node(&dl_se->rb_node, parent, link); |
2161573e | 1424 | rb_insert_color_cached(&dl_se->rb_node, &dl_rq->root, leftmost); |
aab03e05 | 1425 | |
1baca4ce | 1426 | inc_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
1427 | } |
1428 | ||
1429 | static void __dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
1430 | { | |
1431 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
1432 | ||
1433 | if (RB_EMPTY_NODE(&dl_se->rb_node)) | |
1434 | return; | |
1435 | ||
2161573e | 1436 | rb_erase_cached(&dl_se->rb_node, &dl_rq->root); |
aab03e05 DF |
1437 | RB_CLEAR_NODE(&dl_se->rb_node); |
1438 | ||
1baca4ce | 1439 | dec_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
1440 | } |
1441 | ||
1442 | static void | |
2d3d891d DF |
1443 | enqueue_dl_entity(struct sched_dl_entity *dl_se, |
1444 | struct sched_dl_entity *pi_se, int flags) | |
aab03e05 DF |
1445 | { |
1446 | BUG_ON(on_dl_rq(dl_se)); | |
1447 | ||
1448 | /* | |
1449 | * If this is a wakeup or a new instance, the scheduling | |
1450 | * parameters of the task might need updating. Otherwise, | |
1451 | * we want a replenishment of its runtime. | |
1452 | */ | |
e36d8677 | 1453 | if (flags & ENQUEUE_WAKEUP) { |
8fd27231 | 1454 | task_contending(dl_se, flags); |
2d3d891d | 1455 | update_dl_entity(dl_se, pi_se); |
e36d8677 | 1456 | } else if (flags & ENQUEUE_REPLENISH) { |
6a503c3b | 1457 | replenish_dl_entity(dl_se, pi_se); |
295d6d5e LA |
1458 | } else if ((flags & ENQUEUE_RESTORE) && |
1459 | dl_time_before(dl_se->deadline, | |
1460 | rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) { | |
1461 | setup_new_dl_entity(dl_se); | |
e36d8677 | 1462 | } |
aab03e05 DF |
1463 | |
1464 | __enqueue_dl_entity(dl_se); | |
1465 | } | |
1466 | ||
1467 | static void dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
1468 | { | |
1469 | __dequeue_dl_entity(dl_se); | |
1470 | } | |
1471 | ||
1472 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
1473 | { | |
2d3d891d DF |
1474 | struct task_struct *pi_task = rt_mutex_get_top_task(p); |
1475 | struct sched_dl_entity *pi_se = &p->dl; | |
1476 | ||
1477 | /* | |
193be41e JF |
1478 | * Use the scheduling parameters of the top pi-waiter task if: |
1479 | * - we have a top pi-waiter which is a SCHED_DEADLINE task AND | |
1480 | * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is | |
1481 | * smaller than our deadline OR we are a !SCHED_DEADLINE task getting | |
1482 | * boosted due to a SCHED_DEADLINE pi-waiter). | |
1483 | * Otherwise we keep our runtime and deadline. | |
2d3d891d | 1484 | */ |
193be41e | 1485 | if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) { |
2d3d891d | 1486 | pi_se = &pi_task->dl; |
64be6f1f JL |
1487 | } else if (!dl_prio(p->normal_prio)) { |
1488 | /* | |
1489 | * Special case in which we have a !SCHED_DEADLINE task | |
193be41e | 1490 | * that is going to be deboosted, but exceeds its |
64be6f1f JL |
1491 | * runtime while doing so. No point in replenishing |
1492 | * it, as it's going to return back to its original | |
1493 | * scheduling class after this. | |
1494 | */ | |
1495 | BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH); | |
1496 | return; | |
1497 | } | |
2d3d891d | 1498 | |
df8eac8c DBO |
1499 | /* |
1500 | * Check if a constrained deadline task was activated | |
1501 | * after the deadline but before the next period. | |
1502 | * If that is the case, the task will be throttled and | |
1503 | * the replenishment timer will be set to the next period. | |
1504 | */ | |
3effcb42 | 1505 | if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl)) |
df8eac8c DBO |
1506 | dl_check_constrained_dl(&p->dl); |
1507 | ||
8fd27231 | 1508 | if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) { |
794a56eb JL |
1509 | add_rq_bw(&p->dl, &rq->dl); |
1510 | add_running_bw(&p->dl, &rq->dl); | |
8fd27231 | 1511 | } |
e36d8677 | 1512 | |
aab03e05 | 1513 | /* |
e36d8677 | 1514 | * If p is throttled, we do not enqueue it. In fact, if it exhausted |
aab03e05 DF |
1515 | * its budget it needs a replenishment and, since it now is on |
1516 | * its rq, the bandwidth timer callback (which clearly has not | |
1517 | * run yet) will take care of this. | |
e36d8677 LA |
1518 | * However, the active utilization does not depend on the fact |
1519 | * that the task is on the runqueue or not (but depends on the | |
1520 | * task's state - in GRUB parlance, "inactive" vs "active contending"). | |
1521 | * In other words, even if a task is throttled its utilization must | |
1522 | * be counted in the active utilization; hence, we need to call | |
1523 | * add_running_bw(). | |
aab03e05 | 1524 | */ |
e36d8677 | 1525 | if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) { |
209a0cbd | 1526 | if (flags & ENQUEUE_WAKEUP) |
8fd27231 | 1527 | task_contending(&p->dl, flags); |
209a0cbd | 1528 | |
aab03e05 | 1529 | return; |
e36d8677 | 1530 | } |
aab03e05 | 1531 | |
2d3d891d | 1532 | enqueue_dl_entity(&p->dl, pi_se, flags); |
1baca4ce | 1533 | |
4b53a341 | 1534 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) |
1baca4ce | 1535 | enqueue_pushable_dl_task(rq, p); |
aab03e05 DF |
1536 | } |
1537 | ||
1538 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
1539 | { | |
1540 | dequeue_dl_entity(&p->dl); | |
1baca4ce | 1541 | dequeue_pushable_dl_task(rq, p); |
aab03e05 DF |
1542 | } |
1543 | ||
1544 | static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
1545 | { | |
1546 | update_curr_dl(rq); | |
1547 | __dequeue_task_dl(rq, p, flags); | |
e36d8677 | 1548 | |
8fd27231 | 1549 | if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) { |
794a56eb JL |
1550 | sub_running_bw(&p->dl, &rq->dl); |
1551 | sub_rq_bw(&p->dl, &rq->dl); | |
8fd27231 | 1552 | } |
e36d8677 LA |
1553 | |
1554 | /* | |
209a0cbd LA |
1555 | * This check allows to start the inactive timer (or to immediately |
1556 | * decrease the active utilization, if needed) in two cases: | |
e36d8677 LA |
1557 | * when the task blocks and when it is terminating |
1558 | * (p->state == TASK_DEAD). We can handle the two cases in the same | |
1559 | * way, because from GRUB's point of view the same thing is happening | |
1560 | * (the task moves from "active contending" to "active non contending" | |
1561 | * or "inactive") | |
1562 | */ | |
1563 | if (flags & DEQUEUE_SLEEP) | |
209a0cbd | 1564 | task_non_contending(p); |
aab03e05 DF |
1565 | } |
1566 | ||
1567 | /* | |
1568 | * Yield task semantic for -deadline tasks is: | |
1569 | * | |
1570 | * get off from the CPU until our next instance, with | |
1571 | * a new runtime. This is of little use now, since we | |
1572 | * don't have a bandwidth reclaiming mechanism. Anyway, | |
1573 | * bandwidth reclaiming is planned for the future, and | |
1574 | * yield_task_dl will indicate that some spare budget | |
1575 | * is available for other task instances to use it. | |
1576 | */ | |
1577 | static void yield_task_dl(struct rq *rq) | |
1578 | { | |
aab03e05 DF |
1579 | /* |
1580 | * We make the task go to sleep until its current deadline by | |
1581 | * forcing its runtime to zero. This way, update_curr_dl() stops | |
1582 | * it and the bandwidth timer will wake it up and will give it | |
5bfd126e | 1583 | * new scheduling parameters (thanks to dl_yielded=1). |
aab03e05 | 1584 | */ |
48be3a67 PZ |
1585 | rq->curr->dl.dl_yielded = 1; |
1586 | ||
6f1607f1 | 1587 | update_rq_clock(rq); |
aab03e05 | 1588 | update_curr_dl(rq); |
44fb085b WL |
1589 | /* |
1590 | * Tell update_rq_clock() that we've just updated, | |
1591 | * so we don't do microscopic update in schedule() | |
1592 | * and double the fastpath cost. | |
1593 | */ | |
adcc8da8 | 1594 | rq_clock_skip_update(rq); |
aab03e05 DF |
1595 | } |
1596 | ||
1baca4ce JL |
1597 | #ifdef CONFIG_SMP |
1598 | ||
1599 | static int find_later_rq(struct task_struct *task); | |
1baca4ce JL |
1600 | |
1601 | static int | |
1602 | select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags) | |
1603 | { | |
1604 | struct task_struct *curr; | |
1605 | struct rq *rq; | |
1606 | ||
1d7e974c | 1607 | if (sd_flag != SD_BALANCE_WAKE) |
1baca4ce JL |
1608 | goto out; |
1609 | ||
1610 | rq = cpu_rq(cpu); | |
1611 | ||
1612 | rcu_read_lock(); | |
316c1608 | 1613 | curr = READ_ONCE(rq->curr); /* unlocked access */ |
1baca4ce JL |
1614 | |
1615 | /* | |
1616 | * If we are dealing with a -deadline task, we must | |
1617 | * decide where to wake it up. | |
1618 | * If it has a later deadline and the current task | |
1619 | * on this rq can't move (provided the waking task | |
1620 | * can!) we prefer to send it somewhere else. On the | |
1621 | * other hand, if it has a shorter deadline, we | |
1622 | * try to make it stay here, it might be important. | |
1623 | */ | |
1624 | if (unlikely(dl_task(curr)) && | |
4b53a341 | 1625 | (curr->nr_cpus_allowed < 2 || |
1baca4ce | 1626 | !dl_entity_preempt(&p->dl, &curr->dl)) && |
4b53a341 | 1627 | (p->nr_cpus_allowed > 1)) { |
1baca4ce JL |
1628 | int target = find_later_rq(p); |
1629 | ||
9d514262 | 1630 | if (target != -1 && |
5aa50507 LA |
1631 | (dl_time_before(p->dl.deadline, |
1632 | cpu_rq(target)->dl.earliest_dl.curr) || | |
1633 | (cpu_rq(target)->dl.dl_nr_running == 0))) | |
1baca4ce JL |
1634 | cpu = target; |
1635 | } | |
1636 | rcu_read_unlock(); | |
1637 | ||
1638 | out: | |
1639 | return cpu; | |
1640 | } | |
1641 | ||
1327237a | 1642 | static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused) |
209a0cbd LA |
1643 | { |
1644 | struct rq *rq; | |
1645 | ||
8fd27231 | 1646 | if (p->state != TASK_WAKING) |
209a0cbd LA |
1647 | return; |
1648 | ||
1649 | rq = task_rq(p); | |
1650 | /* | |
1651 | * Since p->state == TASK_WAKING, set_task_cpu() has been called | |
1652 | * from try_to_wake_up(). Hence, p->pi_lock is locked, but | |
1653 | * rq->lock is not... So, lock it | |
1654 | */ | |
1655 | raw_spin_lock(&rq->lock); | |
8fd27231 | 1656 | if (p->dl.dl_non_contending) { |
794a56eb | 1657 | sub_running_bw(&p->dl, &rq->dl); |
8fd27231 LA |
1658 | p->dl.dl_non_contending = 0; |
1659 | /* | |
1660 | * If the timer handler is currently running and the | |
1661 | * timer cannot be cancelled, inactive_task_timer() | |
1662 | * will see that dl_not_contending is not set, and | |
1663 | * will not touch the rq's active utilization, | |
1664 | * so we are still safe. | |
1665 | */ | |
1666 | if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) | |
1667 | put_task_struct(p); | |
1668 | } | |
794a56eb | 1669 | sub_rq_bw(&p->dl, &rq->dl); |
209a0cbd LA |
1670 | raw_spin_unlock(&rq->lock); |
1671 | } | |
1672 | ||
1baca4ce JL |
1673 | static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p) |
1674 | { | |
1675 | /* | |
1676 | * Current can't be migrated, useless to reschedule, | |
1677 | * let's hope p can move out. | |
1678 | */ | |
4b53a341 | 1679 | if (rq->curr->nr_cpus_allowed == 1 || |
3261ed0b | 1680 | !cpudl_find(&rq->rd->cpudl, rq->curr, NULL)) |
1baca4ce JL |
1681 | return; |
1682 | ||
1683 | /* | |
1684 | * p is migratable, so let's not schedule it and | |
1685 | * see if it is pushed or pulled somewhere else. | |
1686 | */ | |
4b53a341 | 1687 | if (p->nr_cpus_allowed != 1 && |
3261ed0b | 1688 | cpudl_find(&rq->rd->cpudl, p, NULL)) |
1baca4ce JL |
1689 | return; |
1690 | ||
8875125e | 1691 | resched_curr(rq); |
1baca4ce JL |
1692 | } |
1693 | ||
1694 | #endif /* CONFIG_SMP */ | |
1695 | ||
aab03e05 DF |
1696 | /* |
1697 | * Only called when both the current and waking task are -deadline | |
1698 | * tasks. | |
1699 | */ | |
1700 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | |
1701 | int flags) | |
1702 | { | |
1baca4ce | 1703 | if (dl_entity_preempt(&p->dl, &rq->curr->dl)) { |
8875125e | 1704 | resched_curr(rq); |
1baca4ce JL |
1705 | return; |
1706 | } | |
1707 | ||
1708 | #ifdef CONFIG_SMP | |
1709 | /* | |
1710 | * In the unlikely case current and p have the same deadline | |
1711 | * let us try to decide what's the best thing to do... | |
1712 | */ | |
332ac17e DF |
1713 | if ((p->dl.deadline == rq->curr->dl.deadline) && |
1714 | !test_tsk_need_resched(rq->curr)) | |
1baca4ce JL |
1715 | check_preempt_equal_dl(rq, p); |
1716 | #endif /* CONFIG_SMP */ | |
aab03e05 DF |
1717 | } |
1718 | ||
1719 | #ifdef CONFIG_SCHED_HRTICK | |
1720 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
1721 | { | |
177ef2a6 | 1722 | hrtick_start(rq, p->dl.runtime); |
aab03e05 | 1723 | } |
36ce9881 WL |
1724 | #else /* !CONFIG_SCHED_HRTICK */ |
1725 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
1726 | { | |
1727 | } | |
aab03e05 DF |
1728 | #endif |
1729 | ||
ff1cdc94 MS |
1730 | static inline void set_next_task(struct rq *rq, struct task_struct *p) |
1731 | { | |
1732 | p->se.exec_start = rq_clock_task(rq); | |
1733 | ||
1734 | /* You can't push away the running task */ | |
1735 | dequeue_pushable_dl_task(rq, p); | |
1736 | } | |
1737 | ||
aab03e05 DF |
1738 | static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq, |
1739 | struct dl_rq *dl_rq) | |
1740 | { | |
2161573e | 1741 | struct rb_node *left = rb_first_cached(&dl_rq->root); |
aab03e05 DF |
1742 | |
1743 | if (!left) | |
1744 | return NULL; | |
1745 | ||
1746 | return rb_entry(left, struct sched_dl_entity, rb_node); | |
1747 | } | |
1748 | ||
181a80d1 | 1749 | static struct task_struct * |
d8ac8971 | 1750 | pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
aab03e05 DF |
1751 | { |
1752 | struct sched_dl_entity *dl_se; | |
1753 | struct task_struct *p; | |
1754 | struct dl_rq *dl_rq; | |
1755 | ||
1756 | dl_rq = &rq->dl; | |
1757 | ||
a1d9a323 | 1758 | if (need_pull_dl_task(rq, prev)) { |
cbce1a68 PZ |
1759 | /* |
1760 | * This is OK, because current is on_cpu, which avoids it being | |
1761 | * picked for load-balance and preemption/IRQs are still | |
1762 | * disabled avoiding further scheduler activity on it and we're | |
1763 | * being very careful to re-start the picking loop. | |
1764 | */ | |
d8ac8971 | 1765 | rq_unpin_lock(rq, rf); |
38033c37 | 1766 | pull_dl_task(rq); |
d8ac8971 | 1767 | rq_repin_lock(rq, rf); |
a1d9a323 | 1768 | /* |
176cedc4 | 1769 | * pull_dl_task() can drop (and re-acquire) rq->lock; this |
a1d9a323 KT |
1770 | * means a stop task can slip in, in which case we need to |
1771 | * re-start task selection. | |
1772 | */ | |
da0c1e65 | 1773 | if (rq->stop && task_on_rq_queued(rq->stop)) |
a1d9a323 KT |
1774 | return RETRY_TASK; |
1775 | } | |
1776 | ||
734ff2a7 KT |
1777 | /* |
1778 | * When prev is DL, we may throttle it in put_prev_task(). | |
1779 | * So, we update time before we check for dl_nr_running. | |
1780 | */ | |
1781 | if (prev->sched_class == &dl_sched_class) | |
1782 | update_curr_dl(rq); | |
38033c37 | 1783 | |
aab03e05 DF |
1784 | if (unlikely(!dl_rq->dl_nr_running)) |
1785 | return NULL; | |
1786 | ||
3f1d2a31 | 1787 | put_prev_task(rq, prev); |
606dba2e | 1788 | |
aab03e05 DF |
1789 | dl_se = pick_next_dl_entity(rq, dl_rq); |
1790 | BUG_ON(!dl_se); | |
1791 | ||
1792 | p = dl_task_of(dl_se); | |
1baca4ce | 1793 | |
ff1cdc94 | 1794 | set_next_task(rq, p); |
1baca4ce | 1795 | |
aab03e05 DF |
1796 | if (hrtick_enabled(rq)) |
1797 | start_hrtick_dl(rq, p); | |
1baca4ce | 1798 | |
02d8ec94 | 1799 | deadline_queue_push_tasks(rq); |
1baca4ce | 1800 | |
3727e0e1 | 1801 | if (rq->curr->sched_class != &dl_sched_class) |
23127296 | 1802 | update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0); |
3727e0e1 | 1803 | |
aab03e05 DF |
1804 | return p; |
1805 | } | |
1806 | ||
1807 | static void put_prev_task_dl(struct rq *rq, struct task_struct *p) | |
1808 | { | |
1809 | update_curr_dl(rq); | |
1baca4ce | 1810 | |
23127296 | 1811 | update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1); |
4b53a341 | 1812 | if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1) |
1baca4ce | 1813 | enqueue_pushable_dl_task(rq, p); |
aab03e05 DF |
1814 | } |
1815 | ||
d84b3131 FW |
1816 | /* |
1817 | * scheduler tick hitting a task of our scheduling class. | |
1818 | * | |
1819 | * NOTE: This function can be called remotely by the tick offload that | |
1820 | * goes along full dynticks. Therefore no local assumption can be made | |
1821 | * and everything must be accessed through the @rq and @curr passed in | |
1822 | * parameters. | |
1823 | */ | |
aab03e05 DF |
1824 | static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued) |
1825 | { | |
1826 | update_curr_dl(rq); | |
1827 | ||
23127296 | 1828 | update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1); |
a7bebf48 WL |
1829 | /* |
1830 | * Even when we have runtime, update_curr_dl() might have resulted in us | |
1831 | * not being the leftmost task anymore. In that case NEED_RESCHED will | |
1832 | * be set and schedule() will start a new hrtick for the next task. | |
1833 | */ | |
1834 | if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 && | |
1835 | is_leftmost(p, &rq->dl)) | |
aab03e05 | 1836 | start_hrtick_dl(rq, p); |
aab03e05 DF |
1837 | } |
1838 | ||
1839 | static void task_fork_dl(struct task_struct *p) | |
1840 | { | |
1841 | /* | |
1842 | * SCHED_DEADLINE tasks cannot fork and this is achieved through | |
1843 | * sched_fork() | |
1844 | */ | |
1845 | } | |
1846 | ||
aab03e05 DF |
1847 | static void set_curr_task_dl(struct rq *rq) |
1848 | { | |
ff1cdc94 | 1849 | set_next_task(rq, rq->curr); |
1baca4ce JL |
1850 | } |
1851 | ||
1852 | #ifdef CONFIG_SMP | |
1853 | ||
1854 | /* Only try algorithms three times */ | |
1855 | #define DL_MAX_TRIES 3 | |
1856 | ||
1857 | static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu) | |
1858 | { | |
1859 | if (!task_running(rq, p) && | |
3bd37062 | 1860 | cpumask_test_cpu(cpu, p->cpus_ptr)) |
1baca4ce | 1861 | return 1; |
1baca4ce JL |
1862 | return 0; |
1863 | } | |
1864 | ||
8b5e770e WL |
1865 | /* |
1866 | * Return the earliest pushable rq's task, which is suitable to be executed | |
1867 | * on the CPU, NULL otherwise: | |
1868 | */ | |
1869 | static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu) | |
1870 | { | |
2161573e | 1871 | struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost; |
8b5e770e WL |
1872 | struct task_struct *p = NULL; |
1873 | ||
1874 | if (!has_pushable_dl_tasks(rq)) | |
1875 | return NULL; | |
1876 | ||
1877 | next_node: | |
1878 | if (next_node) { | |
1879 | p = rb_entry(next_node, struct task_struct, pushable_dl_tasks); | |
1880 | ||
1881 | if (pick_dl_task(rq, p, cpu)) | |
1882 | return p; | |
1883 | ||
1884 | next_node = rb_next(next_node); | |
1885 | goto next_node; | |
1886 | } | |
1887 | ||
1888 | return NULL; | |
1889 | } | |
1890 | ||
1baca4ce JL |
1891 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl); |
1892 | ||
1893 | static int find_later_rq(struct task_struct *task) | |
1894 | { | |
1895 | struct sched_domain *sd; | |
4ba29684 | 1896 | struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl); |
1baca4ce | 1897 | int this_cpu = smp_processor_id(); |
b18c3ca1 | 1898 | int cpu = task_cpu(task); |
1baca4ce JL |
1899 | |
1900 | /* Make sure the mask is initialized first */ | |
1901 | if (unlikely(!later_mask)) | |
1902 | return -1; | |
1903 | ||
4b53a341 | 1904 | if (task->nr_cpus_allowed == 1) |
1baca4ce JL |
1905 | return -1; |
1906 | ||
91ec6778 JL |
1907 | /* |
1908 | * We have to consider system topology and task affinity | |
97fb7a0a | 1909 | * first, then we can look for a suitable CPU. |
91ec6778 | 1910 | */ |
3261ed0b | 1911 | if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask)) |
1baca4ce JL |
1912 | return -1; |
1913 | ||
1914 | /* | |
b18c3ca1 BP |
1915 | * If we are here, some targets have been found, including |
1916 | * the most suitable which is, among the runqueues where the | |
1917 | * current tasks have later deadlines than the task's one, the | |
1918 | * rq with the latest possible one. | |
1baca4ce JL |
1919 | * |
1920 | * Now we check how well this matches with task's | |
1921 | * affinity and system topology. | |
1922 | * | |
97fb7a0a | 1923 | * The last CPU where the task run is our first |
1baca4ce JL |
1924 | * guess, since it is most likely cache-hot there. |
1925 | */ | |
1926 | if (cpumask_test_cpu(cpu, later_mask)) | |
1927 | return cpu; | |
1928 | /* | |
1929 | * Check if this_cpu is to be skipped (i.e., it is | |
1930 | * not in the mask) or not. | |
1931 | */ | |
1932 | if (!cpumask_test_cpu(this_cpu, later_mask)) | |
1933 | this_cpu = -1; | |
1934 | ||
1935 | rcu_read_lock(); | |
1936 | for_each_domain(cpu, sd) { | |
1937 | if (sd->flags & SD_WAKE_AFFINE) { | |
b18c3ca1 | 1938 | int best_cpu; |
1baca4ce JL |
1939 | |
1940 | /* | |
1941 | * If possible, preempting this_cpu is | |
1942 | * cheaper than migrating. | |
1943 | */ | |
1944 | if (this_cpu != -1 && | |
1945 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { | |
1946 | rcu_read_unlock(); | |
1947 | return this_cpu; | |
1948 | } | |
1949 | ||
b18c3ca1 BP |
1950 | best_cpu = cpumask_first_and(later_mask, |
1951 | sched_domain_span(sd)); | |
1baca4ce | 1952 | /* |
97fb7a0a | 1953 | * Last chance: if a CPU being in both later_mask |
b18c3ca1 | 1954 | * and current sd span is valid, that becomes our |
97fb7a0a | 1955 | * choice. Of course, the latest possible CPU is |
b18c3ca1 | 1956 | * already under consideration through later_mask. |
1baca4ce | 1957 | */ |
b18c3ca1 | 1958 | if (best_cpu < nr_cpu_ids) { |
1baca4ce JL |
1959 | rcu_read_unlock(); |
1960 | return best_cpu; | |
1961 | } | |
1962 | } | |
1963 | } | |
1964 | rcu_read_unlock(); | |
1965 | ||
1966 | /* | |
1967 | * At this point, all our guesses failed, we just return | |
1968 | * 'something', and let the caller sort the things out. | |
1969 | */ | |
1970 | if (this_cpu != -1) | |
1971 | return this_cpu; | |
1972 | ||
1973 | cpu = cpumask_any(later_mask); | |
1974 | if (cpu < nr_cpu_ids) | |
1975 | return cpu; | |
1976 | ||
1977 | return -1; | |
1978 | } | |
1979 | ||
1980 | /* Locks the rq it finds */ | |
1981 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq) | |
1982 | { | |
1983 | struct rq *later_rq = NULL; | |
1984 | int tries; | |
1985 | int cpu; | |
1986 | ||
1987 | for (tries = 0; tries < DL_MAX_TRIES; tries++) { | |
1988 | cpu = find_later_rq(task); | |
1989 | ||
1990 | if ((cpu == -1) || (cpu == rq->cpu)) | |
1991 | break; | |
1992 | ||
1993 | later_rq = cpu_rq(cpu); | |
1994 | ||
5aa50507 LA |
1995 | if (later_rq->dl.dl_nr_running && |
1996 | !dl_time_before(task->dl.deadline, | |
9d514262 WL |
1997 | later_rq->dl.earliest_dl.curr)) { |
1998 | /* | |
1999 | * Target rq has tasks of equal or earlier deadline, | |
2000 | * retrying does not release any lock and is unlikely | |
2001 | * to yield a different result. | |
2002 | */ | |
2003 | later_rq = NULL; | |
2004 | break; | |
2005 | } | |
2006 | ||
1baca4ce JL |
2007 | /* Retry if something changed. */ |
2008 | if (double_lock_balance(rq, later_rq)) { | |
2009 | if (unlikely(task_rq(task) != rq || | |
3bd37062 | 2010 | !cpumask_test_cpu(later_rq->cpu, task->cpus_ptr) || |
da0c1e65 | 2011 | task_running(rq, task) || |
13b5ab02 | 2012 | !dl_task(task) || |
da0c1e65 | 2013 | !task_on_rq_queued(task))) { |
1baca4ce JL |
2014 | double_unlock_balance(rq, later_rq); |
2015 | later_rq = NULL; | |
2016 | break; | |
2017 | } | |
2018 | } | |
2019 | ||
2020 | /* | |
2021 | * If the rq we found has no -deadline task, or | |
2022 | * its earliest one has a later deadline than our | |
2023 | * task, the rq is a good one. | |
2024 | */ | |
2025 | if (!later_rq->dl.dl_nr_running || | |
2026 | dl_time_before(task->dl.deadline, | |
2027 | later_rq->dl.earliest_dl.curr)) | |
2028 | break; | |
2029 | ||
2030 | /* Otherwise we try again. */ | |
2031 | double_unlock_balance(rq, later_rq); | |
2032 | later_rq = NULL; | |
2033 | } | |
2034 | ||
2035 | return later_rq; | |
2036 | } | |
2037 | ||
2038 | static struct task_struct *pick_next_pushable_dl_task(struct rq *rq) | |
2039 | { | |
2040 | struct task_struct *p; | |
2041 | ||
2042 | if (!has_pushable_dl_tasks(rq)) | |
2043 | return NULL; | |
2044 | ||
2161573e | 2045 | p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost, |
1baca4ce JL |
2046 | struct task_struct, pushable_dl_tasks); |
2047 | ||
2048 | BUG_ON(rq->cpu != task_cpu(p)); | |
2049 | BUG_ON(task_current(rq, p)); | |
4b53a341 | 2050 | BUG_ON(p->nr_cpus_allowed <= 1); |
1baca4ce | 2051 | |
da0c1e65 | 2052 | BUG_ON(!task_on_rq_queued(p)); |
1baca4ce JL |
2053 | BUG_ON(!dl_task(p)); |
2054 | ||
2055 | return p; | |
2056 | } | |
2057 | ||
2058 | /* | |
2059 | * See if the non running -deadline tasks on this rq | |
2060 | * can be sent to some other CPU where they can preempt | |
2061 | * and start executing. | |
2062 | */ | |
2063 | static int push_dl_task(struct rq *rq) | |
2064 | { | |
2065 | struct task_struct *next_task; | |
2066 | struct rq *later_rq; | |
c51b8ab5 | 2067 | int ret = 0; |
1baca4ce JL |
2068 | |
2069 | if (!rq->dl.overloaded) | |
2070 | return 0; | |
2071 | ||
2072 | next_task = pick_next_pushable_dl_task(rq); | |
2073 | if (!next_task) | |
2074 | return 0; | |
2075 | ||
2076 | retry: | |
9ebc6053 | 2077 | if (WARN_ON(next_task == rq->curr)) |
1baca4ce | 2078 | return 0; |
1baca4ce JL |
2079 | |
2080 | /* | |
2081 | * If next_task preempts rq->curr, and rq->curr | |
2082 | * can move away, it makes sense to just reschedule | |
2083 | * without going further in pushing next_task. | |
2084 | */ | |
2085 | if (dl_task(rq->curr) && | |
2086 | dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) && | |
4b53a341 | 2087 | rq->curr->nr_cpus_allowed > 1) { |
8875125e | 2088 | resched_curr(rq); |
1baca4ce JL |
2089 | return 0; |
2090 | } | |
2091 | ||
2092 | /* We might release rq lock */ | |
2093 | get_task_struct(next_task); | |
2094 | ||
2095 | /* Will lock the rq it'll find */ | |
2096 | later_rq = find_lock_later_rq(next_task, rq); | |
2097 | if (!later_rq) { | |
2098 | struct task_struct *task; | |
2099 | ||
2100 | /* | |
2101 | * We must check all this again, since | |
2102 | * find_lock_later_rq releases rq->lock and it is | |
2103 | * then possible that next_task has migrated. | |
2104 | */ | |
2105 | task = pick_next_pushable_dl_task(rq); | |
a776b968 | 2106 | if (task == next_task) { |
1baca4ce JL |
2107 | /* |
2108 | * The task is still there. We don't try | |
97fb7a0a | 2109 | * again, some other CPU will pull it when ready. |
1baca4ce | 2110 | */ |
1baca4ce JL |
2111 | goto out; |
2112 | } | |
2113 | ||
2114 | if (!task) | |
2115 | /* No more tasks */ | |
2116 | goto out; | |
2117 | ||
2118 | put_task_struct(next_task); | |
2119 | next_task = task; | |
2120 | goto retry; | |
2121 | } | |
2122 | ||
2123 | deactivate_task(rq, next_task, 0); | |
794a56eb JL |
2124 | sub_running_bw(&next_task->dl, &rq->dl); |
2125 | sub_rq_bw(&next_task->dl, &rq->dl); | |
1baca4ce | 2126 | set_task_cpu(next_task, later_rq->cpu); |
794a56eb | 2127 | add_rq_bw(&next_task->dl, &later_rq->dl); |
840d7196 DBO |
2128 | |
2129 | /* | |
2130 | * Update the later_rq clock here, because the clock is used | |
2131 | * by the cpufreq_update_util() inside __add_running_bw(). | |
2132 | */ | |
2133 | update_rq_clock(later_rq); | |
794a56eb | 2134 | add_running_bw(&next_task->dl, &later_rq->dl); |
840d7196 | 2135 | activate_task(later_rq, next_task, ENQUEUE_NOCLOCK); |
c51b8ab5 | 2136 | ret = 1; |
1baca4ce | 2137 | |
8875125e | 2138 | resched_curr(later_rq); |
1baca4ce JL |
2139 | |
2140 | double_unlock_balance(rq, later_rq); | |
2141 | ||
2142 | out: | |
2143 | put_task_struct(next_task); | |
2144 | ||
c51b8ab5 | 2145 | return ret; |
1baca4ce JL |
2146 | } |
2147 | ||
2148 | static void push_dl_tasks(struct rq *rq) | |
2149 | { | |
4ffa08ed | 2150 | /* push_dl_task() will return true if it moved a -deadline task */ |
1baca4ce JL |
2151 | while (push_dl_task(rq)) |
2152 | ; | |
aab03e05 DF |
2153 | } |
2154 | ||
0ea60c20 | 2155 | static void pull_dl_task(struct rq *this_rq) |
1baca4ce | 2156 | { |
0ea60c20 | 2157 | int this_cpu = this_rq->cpu, cpu; |
1baca4ce | 2158 | struct task_struct *p; |
0ea60c20 | 2159 | bool resched = false; |
1baca4ce JL |
2160 | struct rq *src_rq; |
2161 | u64 dmin = LONG_MAX; | |
2162 | ||
2163 | if (likely(!dl_overloaded(this_rq))) | |
0ea60c20 | 2164 | return; |
1baca4ce JL |
2165 | |
2166 | /* | |
2167 | * Match the barrier from dl_set_overloaded; this guarantees that if we | |
2168 | * see overloaded we must also see the dlo_mask bit. | |
2169 | */ | |
2170 | smp_rmb(); | |
2171 | ||
2172 | for_each_cpu(cpu, this_rq->rd->dlo_mask) { | |
2173 | if (this_cpu == cpu) | |
2174 | continue; | |
2175 | ||
2176 | src_rq = cpu_rq(cpu); | |
2177 | ||
2178 | /* | |
2179 | * It looks racy, abd it is! However, as in sched_rt.c, | |
2180 | * we are fine with this. | |
2181 | */ | |
2182 | if (this_rq->dl.dl_nr_running && | |
2183 | dl_time_before(this_rq->dl.earliest_dl.curr, | |
2184 | src_rq->dl.earliest_dl.next)) | |
2185 | continue; | |
2186 | ||
2187 | /* Might drop this_rq->lock */ | |
2188 | double_lock_balance(this_rq, src_rq); | |
2189 | ||
2190 | /* | |
2191 | * If there are no more pullable tasks on the | |
2192 | * rq, we're done with it. | |
2193 | */ | |
2194 | if (src_rq->dl.dl_nr_running <= 1) | |
2195 | goto skip; | |
2196 | ||
8b5e770e | 2197 | p = pick_earliest_pushable_dl_task(src_rq, this_cpu); |
1baca4ce JL |
2198 | |
2199 | /* | |
2200 | * We found a task to be pulled if: | |
2201 | * - it preempts our current (if there's one), | |
2202 | * - it will preempt the last one we pulled (if any). | |
2203 | */ | |
2204 | if (p && dl_time_before(p->dl.deadline, dmin) && | |
2205 | (!this_rq->dl.dl_nr_running || | |
2206 | dl_time_before(p->dl.deadline, | |
2207 | this_rq->dl.earliest_dl.curr))) { | |
2208 | WARN_ON(p == src_rq->curr); | |
da0c1e65 | 2209 | WARN_ON(!task_on_rq_queued(p)); |
1baca4ce JL |
2210 | |
2211 | /* | |
2212 | * Then we pull iff p has actually an earlier | |
2213 | * deadline than the current task of its runqueue. | |
2214 | */ | |
2215 | if (dl_time_before(p->dl.deadline, | |
2216 | src_rq->curr->dl.deadline)) | |
2217 | goto skip; | |
2218 | ||
0ea60c20 | 2219 | resched = true; |
1baca4ce JL |
2220 | |
2221 | deactivate_task(src_rq, p, 0); | |
794a56eb JL |
2222 | sub_running_bw(&p->dl, &src_rq->dl); |
2223 | sub_rq_bw(&p->dl, &src_rq->dl); | |
1baca4ce | 2224 | set_task_cpu(p, this_cpu); |
794a56eb JL |
2225 | add_rq_bw(&p->dl, &this_rq->dl); |
2226 | add_running_bw(&p->dl, &this_rq->dl); | |
1baca4ce JL |
2227 | activate_task(this_rq, p, 0); |
2228 | dmin = p->dl.deadline; | |
2229 | ||
2230 | /* Is there any other task even earlier? */ | |
2231 | } | |
2232 | skip: | |
2233 | double_unlock_balance(this_rq, src_rq); | |
2234 | } | |
2235 | ||
0ea60c20 PZ |
2236 | if (resched) |
2237 | resched_curr(this_rq); | |
1baca4ce JL |
2238 | } |
2239 | ||
2240 | /* | |
2241 | * Since the task is not running and a reschedule is not going to happen | |
2242 | * anytime soon on its runqueue, we try pushing it away now. | |
2243 | */ | |
2244 | static void task_woken_dl(struct rq *rq, struct task_struct *p) | |
2245 | { | |
2246 | if (!task_running(rq, p) && | |
2247 | !test_tsk_need_resched(rq->curr) && | |
4b53a341 | 2248 | p->nr_cpus_allowed > 1 && |
1baca4ce | 2249 | dl_task(rq->curr) && |
4b53a341 | 2250 | (rq->curr->nr_cpus_allowed < 2 || |
6b0a563f | 2251 | !dl_entity_preempt(&p->dl, &rq->curr->dl))) { |
1baca4ce JL |
2252 | push_dl_tasks(rq); |
2253 | } | |
2254 | } | |
2255 | ||
2256 | static void set_cpus_allowed_dl(struct task_struct *p, | |
2257 | const struct cpumask *new_mask) | |
2258 | { | |
7f51412a | 2259 | struct root_domain *src_rd; |
6c37067e | 2260 | struct rq *rq; |
1baca4ce JL |
2261 | |
2262 | BUG_ON(!dl_task(p)); | |
2263 | ||
7f51412a JL |
2264 | rq = task_rq(p); |
2265 | src_rd = rq->rd; | |
2266 | /* | |
2267 | * Migrating a SCHED_DEADLINE task between exclusive | |
2268 | * cpusets (different root_domains) entails a bandwidth | |
2269 | * update. We already made space for us in the destination | |
2270 | * domain (see cpuset_can_attach()). | |
2271 | */ | |
2272 | if (!cpumask_intersects(src_rd->span, new_mask)) { | |
2273 | struct dl_bw *src_dl_b; | |
2274 | ||
2275 | src_dl_b = dl_bw_of(cpu_of(rq)); | |
2276 | /* | |
2277 | * We now free resources of the root_domain we are migrating | |
2278 | * off. In the worst case, sched_setattr() may temporary fail | |
2279 | * until we complete the update. | |
2280 | */ | |
2281 | raw_spin_lock(&src_dl_b->lock); | |
8c0944ce | 2282 | __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); |
7f51412a JL |
2283 | raw_spin_unlock(&src_dl_b->lock); |
2284 | } | |
2285 | ||
6c37067e | 2286 | set_cpus_allowed_common(p, new_mask); |
1baca4ce JL |
2287 | } |
2288 | ||
2289 | /* Assumes rq->lock is held */ | |
2290 | static void rq_online_dl(struct rq *rq) | |
2291 | { | |
2292 | if (rq->dl.overloaded) | |
2293 | dl_set_overload(rq); | |
6bfd6d72 | 2294 | |
16b26943 | 2295 | cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu); |
6bfd6d72 | 2296 | if (rq->dl.dl_nr_running > 0) |
d8206bb3 | 2297 | cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr); |
1baca4ce JL |
2298 | } |
2299 | ||
2300 | /* Assumes rq->lock is held */ | |
2301 | static void rq_offline_dl(struct rq *rq) | |
2302 | { | |
2303 | if (rq->dl.overloaded) | |
2304 | dl_clear_overload(rq); | |
6bfd6d72 | 2305 | |
d8206bb3 | 2306 | cpudl_clear(&rq->rd->cpudl, rq->cpu); |
16b26943 | 2307 | cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu); |
1baca4ce JL |
2308 | } |
2309 | ||
a6c0e746 | 2310 | void __init init_sched_dl_class(void) |
1baca4ce JL |
2311 | { |
2312 | unsigned int i; | |
2313 | ||
2314 | for_each_possible_cpu(i) | |
2315 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i), | |
2316 | GFP_KERNEL, cpu_to_node(i)); | |
2317 | } | |
2318 | ||
f9a25f77 MP |
2319 | void dl_add_task_root_domain(struct task_struct *p) |
2320 | { | |
2321 | struct rq_flags rf; | |
2322 | struct rq *rq; | |
2323 | struct dl_bw *dl_b; | |
2324 | ||
2325 | rq = task_rq_lock(p, &rf); | |
2326 | if (!dl_task(p)) | |
2327 | goto unlock; | |
2328 | ||
2329 | dl_b = &rq->rd->dl_bw; | |
2330 | raw_spin_lock(&dl_b->lock); | |
2331 | ||
2332 | __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span)); | |
2333 | ||
2334 | raw_spin_unlock(&dl_b->lock); | |
2335 | ||
2336 | unlock: | |
2337 | task_rq_unlock(rq, p, &rf); | |
2338 | } | |
2339 | ||
2340 | void dl_clear_root_domain(struct root_domain *rd) | |
2341 | { | |
2342 | unsigned long flags; | |
2343 | ||
2344 | raw_spin_lock_irqsave(&rd->dl_bw.lock, flags); | |
2345 | rd->dl_bw.total_bw = 0; | |
2346 | raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags); | |
2347 | } | |
2348 | ||
1baca4ce JL |
2349 | #endif /* CONFIG_SMP */ |
2350 | ||
aab03e05 DF |
2351 | static void switched_from_dl(struct rq *rq, struct task_struct *p) |
2352 | { | |
a649f237 | 2353 | /* |
209a0cbd LA |
2354 | * task_non_contending() can start the "inactive timer" (if the 0-lag |
2355 | * time is in the future). If the task switches back to dl before | |
2356 | * the "inactive timer" fires, it can continue to consume its current | |
2357 | * runtime using its current deadline. If it stays outside of | |
2358 | * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer() | |
2359 | * will reset the task parameters. | |
a649f237 | 2360 | */ |
209a0cbd LA |
2361 | if (task_on_rq_queued(p) && p->dl.dl_runtime) |
2362 | task_non_contending(p); | |
2363 | ||
e117cb52 JL |
2364 | if (!task_on_rq_queued(p)) { |
2365 | /* | |
2366 | * Inactive timer is armed. However, p is leaving DEADLINE and | |
2367 | * might migrate away from this rq while continuing to run on | |
2368 | * some other class. We need to remove its contribution from | |
2369 | * this rq running_bw now, or sub_rq_bw (below) will complain. | |
2370 | */ | |
2371 | if (p->dl.dl_non_contending) | |
2372 | sub_running_bw(&p->dl, &rq->dl); | |
794a56eb | 2373 | sub_rq_bw(&p->dl, &rq->dl); |
e117cb52 | 2374 | } |
8fd27231 | 2375 | |
209a0cbd LA |
2376 | /* |
2377 | * We cannot use inactive_task_timer() to invoke sub_running_bw() | |
2378 | * at the 0-lag time, because the task could have been migrated | |
2379 | * while SCHED_OTHER in the meanwhile. | |
2380 | */ | |
2381 | if (p->dl.dl_non_contending) | |
2382 | p->dl.dl_non_contending = 0; | |
a5e7be3b | 2383 | |
1baca4ce JL |
2384 | /* |
2385 | * Since this might be the only -deadline task on the rq, | |
2386 | * this is the right place to try to pull some other one | |
97fb7a0a | 2387 | * from an overloaded CPU, if any. |
1baca4ce | 2388 | */ |
cd660911 WL |
2389 | if (!task_on_rq_queued(p) || rq->dl.dl_nr_running) |
2390 | return; | |
2391 | ||
02d8ec94 | 2392 | deadline_queue_pull_task(rq); |
aab03e05 DF |
2393 | } |
2394 | ||
1baca4ce JL |
2395 | /* |
2396 | * When switching to -deadline, we may overload the rq, then | |
2397 | * we try to push someone off, if possible. | |
2398 | */ | |
aab03e05 DF |
2399 | static void switched_to_dl(struct rq *rq, struct task_struct *p) |
2400 | { | |
209a0cbd LA |
2401 | if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) |
2402 | put_task_struct(p); | |
98b0a857 JL |
2403 | |
2404 | /* If p is not queued we will update its parameters at next wakeup. */ | |
8fd27231 | 2405 | if (!task_on_rq_queued(p)) { |
794a56eb | 2406 | add_rq_bw(&p->dl, &rq->dl); |
98b0a857 | 2407 | |
8fd27231 LA |
2408 | return; |
2409 | } | |
72f9f3fd | 2410 | |
98b0a857 | 2411 | if (rq->curr != p) { |
1baca4ce | 2412 | #ifdef CONFIG_SMP |
4b53a341 | 2413 | if (p->nr_cpus_allowed > 1 && rq->dl.overloaded) |
02d8ec94 | 2414 | deadline_queue_push_tasks(rq); |
619bd4a7 | 2415 | #endif |
9916e214 PZ |
2416 | if (dl_task(rq->curr)) |
2417 | check_preempt_curr_dl(rq, p, 0); | |
2418 | else | |
2419 | resched_curr(rq); | |
aab03e05 DF |
2420 | } |
2421 | } | |
2422 | ||
1baca4ce JL |
2423 | /* |
2424 | * If the scheduling parameters of a -deadline task changed, | |
2425 | * a push or pull operation might be needed. | |
2426 | */ | |
aab03e05 DF |
2427 | static void prio_changed_dl(struct rq *rq, struct task_struct *p, |
2428 | int oldprio) | |
2429 | { | |
da0c1e65 | 2430 | if (task_on_rq_queued(p) || rq->curr == p) { |
aab03e05 | 2431 | #ifdef CONFIG_SMP |
1baca4ce JL |
2432 | /* |
2433 | * This might be too much, but unfortunately | |
2434 | * we don't have the old deadline value, and | |
2435 | * we can't argue if the task is increasing | |
2436 | * or lowering its prio, so... | |
2437 | */ | |
2438 | if (!rq->dl.overloaded) | |
02d8ec94 | 2439 | deadline_queue_pull_task(rq); |
1baca4ce JL |
2440 | |
2441 | /* | |
2442 | * If we now have a earlier deadline task than p, | |
2443 | * then reschedule, provided p is still on this | |
2444 | * runqueue. | |
2445 | */ | |
9916e214 | 2446 | if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline)) |
8875125e | 2447 | resched_curr(rq); |
1baca4ce JL |
2448 | #else |
2449 | /* | |
2450 | * Again, we don't know if p has a earlier | |
2451 | * or later deadline, so let's blindly set a | |
2452 | * (maybe not needed) rescheduling point. | |
2453 | */ | |
8875125e | 2454 | resched_curr(rq); |
1baca4ce | 2455 | #endif /* CONFIG_SMP */ |
801ccdbf | 2456 | } |
aab03e05 | 2457 | } |
aab03e05 DF |
2458 | |
2459 | const struct sched_class dl_sched_class = { | |
2460 | .next = &rt_sched_class, | |
2461 | .enqueue_task = enqueue_task_dl, | |
2462 | .dequeue_task = dequeue_task_dl, | |
2463 | .yield_task = yield_task_dl, | |
2464 | ||
2465 | .check_preempt_curr = check_preempt_curr_dl, | |
2466 | ||
2467 | .pick_next_task = pick_next_task_dl, | |
2468 | .put_prev_task = put_prev_task_dl, | |
2469 | ||
2470 | #ifdef CONFIG_SMP | |
2471 | .select_task_rq = select_task_rq_dl, | |
209a0cbd | 2472 | .migrate_task_rq = migrate_task_rq_dl, |
1baca4ce JL |
2473 | .set_cpus_allowed = set_cpus_allowed_dl, |
2474 | .rq_online = rq_online_dl, | |
2475 | .rq_offline = rq_offline_dl, | |
1baca4ce | 2476 | .task_woken = task_woken_dl, |
aab03e05 DF |
2477 | #endif |
2478 | ||
2479 | .set_curr_task = set_curr_task_dl, | |
2480 | .task_tick = task_tick_dl, | |
2481 | .task_fork = task_fork_dl, | |
aab03e05 DF |
2482 | |
2483 | .prio_changed = prio_changed_dl, | |
2484 | .switched_from = switched_from_dl, | |
2485 | .switched_to = switched_to_dl, | |
6e998916 SG |
2486 | |
2487 | .update_curr = update_curr_dl, | |
aab03e05 | 2488 | }; |
acb32132 | 2489 | |
06a76fe0 NP |
2490 | int sched_dl_global_validate(void) |
2491 | { | |
2492 | u64 runtime = global_rt_runtime(); | |
2493 | u64 period = global_rt_period(); | |
2494 | u64 new_bw = to_ratio(period, runtime); | |
2495 | struct dl_bw *dl_b; | |
2496 | int cpu, ret = 0; | |
2497 | unsigned long flags; | |
2498 | ||
2499 | /* | |
2500 | * Here we want to check the bandwidth not being set to some | |
2501 | * value smaller than the currently allocated bandwidth in | |
2502 | * any of the root_domains. | |
2503 | * | |
2504 | * FIXME: Cycling on all the CPUs is overdoing, but simpler than | |
2505 | * cycling on root_domains... Discussion on different/better | |
2506 | * solutions is welcome! | |
2507 | */ | |
2508 | for_each_possible_cpu(cpu) { | |
2509 | rcu_read_lock_sched(); | |
2510 | dl_b = dl_bw_of(cpu); | |
2511 | ||
2512 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
2513 | if (new_bw < dl_b->total_bw) | |
2514 | ret = -EBUSY; | |
2515 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2516 | ||
2517 | rcu_read_unlock_sched(); | |
2518 | ||
2519 | if (ret) | |
2520 | break; | |
2521 | } | |
2522 | ||
2523 | return ret; | |
2524 | } | |
2525 | ||
2526 | void init_dl_rq_bw_ratio(struct dl_rq *dl_rq) | |
2527 | { | |
2528 | if (global_rt_runtime() == RUNTIME_INF) { | |
2529 | dl_rq->bw_ratio = 1 << RATIO_SHIFT; | |
2530 | dl_rq->extra_bw = 1 << BW_SHIFT; | |
2531 | } else { | |
2532 | dl_rq->bw_ratio = to_ratio(global_rt_runtime(), | |
2533 | global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT); | |
2534 | dl_rq->extra_bw = to_ratio(global_rt_period(), | |
2535 | global_rt_runtime()); | |
2536 | } | |
2537 | } | |
2538 | ||
2539 | void sched_dl_do_global(void) | |
2540 | { | |
2541 | u64 new_bw = -1; | |
2542 | struct dl_bw *dl_b; | |
2543 | int cpu; | |
2544 | unsigned long flags; | |
2545 | ||
2546 | def_dl_bandwidth.dl_period = global_rt_period(); | |
2547 | def_dl_bandwidth.dl_runtime = global_rt_runtime(); | |
2548 | ||
2549 | if (global_rt_runtime() != RUNTIME_INF) | |
2550 | new_bw = to_ratio(global_rt_period(), global_rt_runtime()); | |
2551 | ||
2552 | /* | |
2553 | * FIXME: As above... | |
2554 | */ | |
2555 | for_each_possible_cpu(cpu) { | |
2556 | rcu_read_lock_sched(); | |
2557 | dl_b = dl_bw_of(cpu); | |
2558 | ||
2559 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
2560 | dl_b->bw = new_bw; | |
2561 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2562 | ||
2563 | rcu_read_unlock_sched(); | |
2564 | init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl); | |
2565 | } | |
2566 | } | |
2567 | ||
2568 | /* | |
2569 | * We must be sure that accepting a new task (or allowing changing the | |
2570 | * parameters of an existing one) is consistent with the bandwidth | |
2571 | * constraints. If yes, this function also accordingly updates the currently | |
2572 | * allocated bandwidth to reflect the new situation. | |
2573 | * | |
2574 | * This function is called while holding p's rq->lock. | |
2575 | */ | |
2576 | int sched_dl_overflow(struct task_struct *p, int policy, | |
2577 | const struct sched_attr *attr) | |
2578 | { | |
2579 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); | |
2580 | u64 period = attr->sched_period ?: attr->sched_deadline; | |
2581 | u64 runtime = attr->sched_runtime; | |
2582 | u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; | |
2583 | int cpus, err = -1; | |
2584 | ||
794a56eb JL |
2585 | if (attr->sched_flags & SCHED_FLAG_SUGOV) |
2586 | return 0; | |
2587 | ||
06a76fe0 NP |
2588 | /* !deadline task may carry old deadline bandwidth */ |
2589 | if (new_bw == p->dl.dl_bw && task_has_dl_policy(p)) | |
2590 | return 0; | |
2591 | ||
2592 | /* | |
2593 | * Either if a task, enters, leave, or stays -deadline but changes | |
2594 | * its parameters, we may need to update accordingly the total | |
2595 | * allocated bandwidth of the container. | |
2596 | */ | |
2597 | raw_spin_lock(&dl_b->lock); | |
2598 | cpus = dl_bw_cpus(task_cpu(p)); | |
2599 | if (dl_policy(policy) && !task_has_dl_policy(p) && | |
2600 | !__dl_overflow(dl_b, cpus, 0, new_bw)) { | |
2601 | if (hrtimer_active(&p->dl.inactive_timer)) | |
8c0944ce | 2602 | __dl_sub(dl_b, p->dl.dl_bw, cpus); |
06a76fe0 NP |
2603 | __dl_add(dl_b, new_bw, cpus); |
2604 | err = 0; | |
2605 | } else if (dl_policy(policy) && task_has_dl_policy(p) && | |
2606 | !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { | |
2607 | /* | |
2608 | * XXX this is slightly incorrect: when the task | |
2609 | * utilization decreases, we should delay the total | |
2610 | * utilization change until the task's 0-lag point. | |
2611 | * But this would require to set the task's "inactive | |
2612 | * timer" when the task is not inactive. | |
2613 | */ | |
8c0944ce | 2614 | __dl_sub(dl_b, p->dl.dl_bw, cpus); |
06a76fe0 NP |
2615 | __dl_add(dl_b, new_bw, cpus); |
2616 | dl_change_utilization(p, new_bw); | |
2617 | err = 0; | |
2618 | } else if (!dl_policy(policy) && task_has_dl_policy(p)) { | |
2619 | /* | |
2620 | * Do not decrease the total deadline utilization here, | |
2621 | * switched_from_dl() will take care to do it at the correct | |
2622 | * (0-lag) time. | |
2623 | */ | |
2624 | err = 0; | |
2625 | } | |
2626 | raw_spin_unlock(&dl_b->lock); | |
2627 | ||
2628 | return err; | |
2629 | } | |
2630 | ||
2631 | /* | |
2632 | * This function initializes the sched_dl_entity of a newly becoming | |
2633 | * SCHED_DEADLINE task. | |
2634 | * | |
2635 | * Only the static values are considered here, the actual runtime and the | |
2636 | * absolute deadline will be properly calculated when the task is enqueued | |
2637 | * for the first time with its new policy. | |
2638 | */ | |
2639 | void __setparam_dl(struct task_struct *p, const struct sched_attr *attr) | |
2640 | { | |
2641 | struct sched_dl_entity *dl_se = &p->dl; | |
2642 | ||
2643 | dl_se->dl_runtime = attr->sched_runtime; | |
2644 | dl_se->dl_deadline = attr->sched_deadline; | |
2645 | dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; | |
2646 | dl_se->flags = attr->sched_flags; | |
2647 | dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); | |
2648 | dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime); | |
2649 | } | |
2650 | ||
2651 | void __getparam_dl(struct task_struct *p, struct sched_attr *attr) | |
2652 | { | |
2653 | struct sched_dl_entity *dl_se = &p->dl; | |
2654 | ||
2655 | attr->sched_priority = p->rt_priority; | |
2656 | attr->sched_runtime = dl_se->dl_runtime; | |
2657 | attr->sched_deadline = dl_se->dl_deadline; | |
2658 | attr->sched_period = dl_se->dl_period; | |
2659 | attr->sched_flags = dl_se->flags; | |
2660 | } | |
2661 | ||
2662 | /* | |
2663 | * This function validates the new parameters of a -deadline task. | |
2664 | * We ask for the deadline not being zero, and greater or equal | |
2665 | * than the runtime, as well as the period of being zero or | |
2666 | * greater than deadline. Furthermore, we have to be sure that | |
2667 | * user parameters are above the internal resolution of 1us (we | |
2668 | * check sched_runtime only since it is always the smaller one) and | |
2669 | * below 2^63 ns (we have to check both sched_deadline and | |
2670 | * sched_period, as the latter can be zero). | |
2671 | */ | |
2672 | bool __checkparam_dl(const struct sched_attr *attr) | |
2673 | { | |
794a56eb JL |
2674 | /* special dl tasks don't actually use any parameter */ |
2675 | if (attr->sched_flags & SCHED_FLAG_SUGOV) | |
2676 | return true; | |
2677 | ||
06a76fe0 NP |
2678 | /* deadline != 0 */ |
2679 | if (attr->sched_deadline == 0) | |
2680 | return false; | |
2681 | ||
2682 | /* | |
2683 | * Since we truncate DL_SCALE bits, make sure we're at least | |
2684 | * that big. | |
2685 | */ | |
2686 | if (attr->sched_runtime < (1ULL << DL_SCALE)) | |
2687 | return false; | |
2688 | ||
2689 | /* | |
2690 | * Since we use the MSB for wrap-around and sign issues, make | |
2691 | * sure it's not set (mind that period can be equal to zero). | |
2692 | */ | |
2693 | if (attr->sched_deadline & (1ULL << 63) || | |
2694 | attr->sched_period & (1ULL << 63)) | |
2695 | return false; | |
2696 | ||
2697 | /* runtime <= deadline <= period (if period != 0) */ | |
2698 | if ((attr->sched_period != 0 && | |
2699 | attr->sched_period < attr->sched_deadline) || | |
2700 | attr->sched_deadline < attr->sched_runtime) | |
2701 | return false; | |
2702 | ||
2703 | return true; | |
2704 | } | |
2705 | ||
2706 | /* | |
2707 | * This function clears the sched_dl_entity static params. | |
2708 | */ | |
2709 | void __dl_clear_params(struct task_struct *p) | |
2710 | { | |
2711 | struct sched_dl_entity *dl_se = &p->dl; | |
2712 | ||
97fb7a0a IM |
2713 | dl_se->dl_runtime = 0; |
2714 | dl_se->dl_deadline = 0; | |
2715 | dl_se->dl_period = 0; | |
2716 | dl_se->flags = 0; | |
2717 | dl_se->dl_bw = 0; | |
2718 | dl_se->dl_density = 0; | |
06a76fe0 | 2719 | |
97fb7a0a IM |
2720 | dl_se->dl_throttled = 0; |
2721 | dl_se->dl_yielded = 0; | |
2722 | dl_se->dl_non_contending = 0; | |
2723 | dl_se->dl_overrun = 0; | |
06a76fe0 NP |
2724 | } |
2725 | ||
2726 | bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr) | |
2727 | { | |
2728 | struct sched_dl_entity *dl_se = &p->dl; | |
2729 | ||
2730 | if (dl_se->dl_runtime != attr->sched_runtime || | |
2731 | dl_se->dl_deadline != attr->sched_deadline || | |
2732 | dl_se->dl_period != attr->sched_period || | |
2733 | dl_se->flags != attr->sched_flags) | |
2734 | return true; | |
2735 | ||
2736 | return false; | |
2737 | } | |
2738 | ||
2739 | #ifdef CONFIG_SMP | |
2740 | int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed) | |
2741 | { | |
97fb7a0a | 2742 | unsigned int dest_cpu; |
06a76fe0 NP |
2743 | struct dl_bw *dl_b; |
2744 | bool overflow; | |
2745 | int cpus, ret; | |
2746 | unsigned long flags; | |
2747 | ||
97fb7a0a IM |
2748 | dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed); |
2749 | ||
06a76fe0 NP |
2750 | rcu_read_lock_sched(); |
2751 | dl_b = dl_bw_of(dest_cpu); | |
2752 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
2753 | cpus = dl_bw_cpus(dest_cpu); | |
2754 | overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw); | |
97fb7a0a | 2755 | if (overflow) { |
06a76fe0 | 2756 | ret = -EBUSY; |
97fb7a0a | 2757 | } else { |
06a76fe0 NP |
2758 | /* |
2759 | * We reserve space for this task in the destination | |
2760 | * root_domain, as we can't fail after this point. | |
2761 | * We will free resources in the source root_domain | |
2762 | * later on (see set_cpus_allowed_dl()). | |
2763 | */ | |
2764 | __dl_add(dl_b, p->dl.dl_bw, cpus); | |
2765 | ret = 0; | |
2766 | } | |
2767 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2768 | rcu_read_unlock_sched(); | |
97fb7a0a | 2769 | |
06a76fe0 NP |
2770 | return ret; |
2771 | } | |
2772 | ||
2773 | int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, | |
2774 | const struct cpumask *trial) | |
2775 | { | |
2776 | int ret = 1, trial_cpus; | |
2777 | struct dl_bw *cur_dl_b; | |
2778 | unsigned long flags; | |
2779 | ||
2780 | rcu_read_lock_sched(); | |
2781 | cur_dl_b = dl_bw_of(cpumask_any(cur)); | |
2782 | trial_cpus = cpumask_weight(trial); | |
2783 | ||
2784 | raw_spin_lock_irqsave(&cur_dl_b->lock, flags); | |
2785 | if (cur_dl_b->bw != -1 && | |
2786 | cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw) | |
2787 | ret = 0; | |
2788 | raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags); | |
2789 | rcu_read_unlock_sched(); | |
97fb7a0a | 2790 | |
06a76fe0 NP |
2791 | return ret; |
2792 | } | |
2793 | ||
2794 | bool dl_cpu_busy(unsigned int cpu) | |
2795 | { | |
2796 | unsigned long flags; | |
2797 | struct dl_bw *dl_b; | |
2798 | bool overflow; | |
2799 | int cpus; | |
2800 | ||
2801 | rcu_read_lock_sched(); | |
2802 | dl_b = dl_bw_of(cpu); | |
2803 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
2804 | cpus = dl_bw_cpus(cpu); | |
2805 | overflow = __dl_overflow(dl_b, cpus, 0, 0); | |
2806 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2807 | rcu_read_unlock_sched(); | |
97fb7a0a | 2808 | |
06a76fe0 NP |
2809 | return overflow; |
2810 | } | |
2811 | #endif | |
2812 | ||
acb32132 | 2813 | #ifdef CONFIG_SCHED_DEBUG |
acb32132 WL |
2814 | void print_dl_stats(struct seq_file *m, int cpu) |
2815 | { | |
2816 | print_dl_rq(m, cpu, &cpu_rq(cpu)->dl); | |
2817 | } | |
2818 | #endif /* CONFIG_SCHED_DEBUG */ |