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" | |
19 | ||
6bfd6d72 | 20 | #include <linux/slab.h> |
06a76fe0 | 21 | #include <uapi/linux/sched/types.h> |
6bfd6d72 | 22 | |
332ac17e DF |
23 | struct dl_bandwidth def_dl_bandwidth; |
24 | ||
aab03e05 DF |
25 | static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se) |
26 | { | |
27 | return container_of(dl_se, struct task_struct, dl); | |
28 | } | |
29 | ||
30 | static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq) | |
31 | { | |
32 | return container_of(dl_rq, struct rq, dl); | |
33 | } | |
34 | ||
35 | static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se) | |
36 | { | |
37 | struct task_struct *p = dl_task_of(dl_se); | |
38 | struct rq *rq = task_rq(p); | |
39 | ||
40 | return &rq->dl; | |
41 | } | |
42 | ||
43 | static inline int on_dl_rq(struct sched_dl_entity *dl_se) | |
44 | { | |
45 | return !RB_EMPTY_NODE(&dl_se->rb_node); | |
46 | } | |
47 | ||
06a76fe0 NP |
48 | #ifdef CONFIG_SMP |
49 | static inline struct dl_bw *dl_bw_of(int i) | |
50 | { | |
51 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), | |
52 | "sched RCU must be held"); | |
53 | return &cpu_rq(i)->rd->dl_bw; | |
54 | } | |
55 | ||
56 | static inline int dl_bw_cpus(int i) | |
57 | { | |
58 | struct root_domain *rd = cpu_rq(i)->rd; | |
59 | int cpus = 0; | |
60 | ||
61 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), | |
62 | "sched RCU must be held"); | |
63 | for_each_cpu_and(i, rd->span, cpu_active_mask) | |
64 | cpus++; | |
65 | ||
66 | return cpus; | |
67 | } | |
68 | #else | |
69 | static inline struct dl_bw *dl_bw_of(int i) | |
70 | { | |
71 | return &cpu_rq(i)->dl.dl_bw; | |
72 | } | |
73 | ||
74 | static inline int dl_bw_cpus(int i) | |
75 | { | |
76 | return 1; | |
77 | } | |
78 | #endif | |
79 | ||
e36d8677 | 80 | static inline |
794a56eb | 81 | void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq) |
e36d8677 LA |
82 | { |
83 | u64 old = dl_rq->running_bw; | |
84 | ||
85 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
86 | dl_rq->running_bw += dl_bw; | |
87 | SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */ | |
8fd27231 | 88 | SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw); |
e0367b12 JL |
89 | /* kick cpufreq (see the comment in kernel/sched/sched.h). */ |
90 | cpufreq_update_util(rq_of_dl_rq(dl_rq), SCHED_CPUFREQ_DL); | |
e36d8677 LA |
91 | } |
92 | ||
93 | static inline | |
794a56eb | 94 | void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq) |
e36d8677 LA |
95 | { |
96 | u64 old = dl_rq->running_bw; | |
97 | ||
98 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
99 | dl_rq->running_bw -= dl_bw; | |
100 | SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */ | |
101 | if (dl_rq->running_bw > old) | |
102 | dl_rq->running_bw = 0; | |
e0367b12 JL |
103 | /* kick cpufreq (see the comment in kernel/sched/sched.h). */ |
104 | cpufreq_update_util(rq_of_dl_rq(dl_rq), SCHED_CPUFREQ_DL); | |
e36d8677 LA |
105 | } |
106 | ||
8fd27231 | 107 | static inline |
794a56eb | 108 | void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq) |
8fd27231 LA |
109 | { |
110 | u64 old = dl_rq->this_bw; | |
111 | ||
112 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
113 | dl_rq->this_bw += dl_bw; | |
114 | SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */ | |
115 | } | |
116 | ||
117 | static inline | |
794a56eb | 118 | void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq) |
8fd27231 LA |
119 | { |
120 | u64 old = dl_rq->this_bw; | |
121 | ||
122 | lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); | |
123 | dl_rq->this_bw -= dl_bw; | |
124 | SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */ | |
125 | if (dl_rq->this_bw > old) | |
126 | dl_rq->this_bw = 0; | |
127 | SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw); | |
128 | } | |
129 | ||
794a56eb JL |
130 | static inline |
131 | void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
132 | { | |
133 | if (!dl_entity_is_special(dl_se)) | |
134 | __add_rq_bw(dl_se->dl_bw, dl_rq); | |
135 | } | |
136 | ||
137 | static inline | |
138 | void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
139 | { | |
140 | if (!dl_entity_is_special(dl_se)) | |
141 | __sub_rq_bw(dl_se->dl_bw, dl_rq); | |
142 | } | |
143 | ||
144 | static inline | |
145 | void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
146 | { | |
147 | if (!dl_entity_is_special(dl_se)) | |
148 | __add_running_bw(dl_se->dl_bw, dl_rq); | |
149 | } | |
150 | ||
151 | static inline | |
152 | void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
153 | { | |
154 | if (!dl_entity_is_special(dl_se)) | |
155 | __sub_running_bw(dl_se->dl_bw, dl_rq); | |
156 | } | |
157 | ||
209a0cbd LA |
158 | void dl_change_utilization(struct task_struct *p, u64 new_bw) |
159 | { | |
8fd27231 | 160 | struct rq *rq; |
209a0cbd | 161 | |
794a56eb JL |
162 | BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV); |
163 | ||
8fd27231 | 164 | if (task_on_rq_queued(p)) |
209a0cbd LA |
165 | return; |
166 | ||
8fd27231 LA |
167 | rq = task_rq(p); |
168 | if (p->dl.dl_non_contending) { | |
794a56eb | 169 | sub_running_bw(&p->dl, &rq->dl); |
8fd27231 LA |
170 | p->dl.dl_non_contending = 0; |
171 | /* | |
172 | * If the timer handler is currently running and the | |
173 | * timer cannot be cancelled, inactive_task_timer() | |
174 | * will see that dl_not_contending is not set, and | |
175 | * will not touch the rq's active utilization, | |
176 | * so we are still safe. | |
177 | */ | |
178 | if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) | |
179 | put_task_struct(p); | |
180 | } | |
794a56eb JL |
181 | __sub_rq_bw(p->dl.dl_bw, &rq->dl); |
182 | __add_rq_bw(new_bw, &rq->dl); | |
209a0cbd LA |
183 | } |
184 | ||
185 | /* | |
186 | * The utilization of a task cannot be immediately removed from | |
187 | * the rq active utilization (running_bw) when the task blocks. | |
188 | * Instead, we have to wait for the so called "0-lag time". | |
189 | * | |
190 | * If a task blocks before the "0-lag time", a timer (the inactive | |
191 | * timer) is armed, and running_bw is decreased when the timer | |
192 | * fires. | |
193 | * | |
194 | * If the task wakes up again before the inactive timer fires, | |
195 | * the timer is cancelled, whereas if the task wakes up after the | |
196 | * inactive timer fired (and running_bw has been decreased) the | |
197 | * task's utilization has to be added to running_bw again. | |
198 | * A flag in the deadline scheduling entity (dl_non_contending) | |
199 | * is used to avoid race conditions between the inactive timer handler | |
200 | * and task wakeups. | |
201 | * | |
202 | * The following diagram shows how running_bw is updated. A task is | |
203 | * "ACTIVE" when its utilization contributes to running_bw; an | |
204 | * "ACTIVE contending" task is in the TASK_RUNNING state, while an | |
205 | * "ACTIVE non contending" task is a blocked task for which the "0-lag time" | |
206 | * has not passed yet. An "INACTIVE" task is a task for which the "0-lag" | |
207 | * time already passed, which does not contribute to running_bw anymore. | |
208 | * +------------------+ | |
209 | * wakeup | ACTIVE | | |
210 | * +------------------>+ contending | | |
211 | * | add_running_bw | | | |
212 | * | +----+------+------+ | |
213 | * | | ^ | |
214 | * | dequeue | | | |
215 | * +--------+-------+ | | | |
216 | * | | t >= 0-lag | | wakeup | |
217 | * | INACTIVE |<---------------+ | | |
218 | * | | sub_running_bw | | | |
219 | * +--------+-------+ | | | |
220 | * ^ | | | |
221 | * | t < 0-lag | | | |
222 | * | | | | |
223 | * | V | | |
224 | * | +----+------+------+ | |
225 | * | sub_running_bw | ACTIVE | | |
226 | * +-------------------+ | | |
227 | * inactive timer | non contending | | |
228 | * fired +------------------+ | |
229 | * | |
230 | * The task_non_contending() function is invoked when a task | |
231 | * blocks, and checks if the 0-lag time already passed or | |
232 | * not (in the first case, it directly updates running_bw; | |
233 | * in the second case, it arms the inactive timer). | |
234 | * | |
235 | * The task_contending() function is invoked when a task wakes | |
236 | * up, and checks if the task is still in the "ACTIVE non contending" | |
237 | * state or not (in the second case, it updates running_bw). | |
238 | */ | |
239 | static void task_non_contending(struct task_struct *p) | |
240 | { | |
241 | struct sched_dl_entity *dl_se = &p->dl; | |
242 | struct hrtimer *timer = &dl_se->inactive_timer; | |
243 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
244 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
245 | s64 zerolag_time; | |
246 | ||
247 | /* | |
248 | * If this is a non-deadline task that has been boosted, | |
249 | * do nothing | |
250 | */ | |
251 | if (dl_se->dl_runtime == 0) | |
252 | return; | |
253 | ||
794a56eb JL |
254 | if (dl_entity_is_special(dl_se)) |
255 | return; | |
256 | ||
209a0cbd LA |
257 | WARN_ON(hrtimer_active(&dl_se->inactive_timer)); |
258 | WARN_ON(dl_se->dl_non_contending); | |
259 | ||
260 | zerolag_time = dl_se->deadline - | |
261 | div64_long((dl_se->runtime * dl_se->dl_period), | |
262 | dl_se->dl_runtime); | |
263 | ||
264 | /* | |
265 | * Using relative times instead of the absolute "0-lag time" | |
266 | * allows to simplify the code | |
267 | */ | |
268 | zerolag_time -= rq_clock(rq); | |
269 | ||
270 | /* | |
271 | * If the "0-lag time" already passed, decrease the active | |
272 | * utilization now, instead of starting a timer | |
273 | */ | |
274 | if (zerolag_time < 0) { | |
275 | if (dl_task(p)) | |
794a56eb | 276 | sub_running_bw(dl_se, dl_rq); |
387e3130 LA |
277 | if (!dl_task(p) || p->state == TASK_DEAD) { |
278 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); | |
279 | ||
8fd27231 | 280 | if (p->state == TASK_DEAD) |
794a56eb | 281 | sub_rq_bw(&p->dl, &rq->dl); |
387e3130 | 282 | raw_spin_lock(&dl_b->lock); |
8c0944ce | 283 | __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); |
209a0cbd | 284 | __dl_clear_params(p); |
387e3130 LA |
285 | raw_spin_unlock(&dl_b->lock); |
286 | } | |
209a0cbd LA |
287 | |
288 | return; | |
289 | } | |
290 | ||
291 | dl_se->dl_non_contending = 1; | |
292 | get_task_struct(p); | |
293 | hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL); | |
294 | } | |
295 | ||
8fd27231 | 296 | static void task_contending(struct sched_dl_entity *dl_se, int flags) |
209a0cbd LA |
297 | { |
298 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
299 | ||
300 | /* | |
301 | * If this is a non-deadline task that has been boosted, | |
302 | * do nothing | |
303 | */ | |
304 | if (dl_se->dl_runtime == 0) | |
305 | return; | |
306 | ||
8fd27231 | 307 | if (flags & ENQUEUE_MIGRATED) |
794a56eb | 308 | add_rq_bw(dl_se, dl_rq); |
8fd27231 | 309 | |
209a0cbd LA |
310 | if (dl_se->dl_non_contending) { |
311 | dl_se->dl_non_contending = 0; | |
312 | /* | |
313 | * If the timer handler is currently running and the | |
314 | * timer cannot be cancelled, inactive_task_timer() | |
315 | * will see that dl_not_contending is not set, and | |
316 | * will not touch the rq's active utilization, | |
317 | * so we are still safe. | |
318 | */ | |
319 | if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1) | |
320 | put_task_struct(dl_task_of(dl_se)); | |
321 | } else { | |
322 | /* | |
323 | * Since "dl_non_contending" is not set, the | |
324 | * task's utilization has already been removed from | |
325 | * active utilization (either when the task blocked, | |
326 | * when the "inactive timer" fired). | |
327 | * So, add it back. | |
328 | */ | |
794a56eb | 329 | add_running_bw(dl_se, dl_rq); |
209a0cbd LA |
330 | } |
331 | } | |
332 | ||
aab03e05 DF |
333 | static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq) |
334 | { | |
335 | struct sched_dl_entity *dl_se = &p->dl; | |
336 | ||
2161573e | 337 | return dl_rq->root.rb_leftmost == &dl_se->rb_node; |
aab03e05 DF |
338 | } |
339 | ||
332ac17e DF |
340 | void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime) |
341 | { | |
342 | raw_spin_lock_init(&dl_b->dl_runtime_lock); | |
343 | dl_b->dl_period = period; | |
344 | dl_b->dl_runtime = runtime; | |
345 | } | |
346 | ||
332ac17e DF |
347 | void init_dl_bw(struct dl_bw *dl_b) |
348 | { | |
349 | raw_spin_lock_init(&dl_b->lock); | |
350 | raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock); | |
1724813d | 351 | if (global_rt_runtime() == RUNTIME_INF) |
332ac17e DF |
352 | dl_b->bw = -1; |
353 | else | |
1724813d | 354 | dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime()); |
332ac17e DF |
355 | raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock); |
356 | dl_b->total_bw = 0; | |
357 | } | |
358 | ||
07c54f7a | 359 | void init_dl_rq(struct dl_rq *dl_rq) |
aab03e05 | 360 | { |
2161573e | 361 | dl_rq->root = RB_ROOT_CACHED; |
1baca4ce JL |
362 | |
363 | #ifdef CONFIG_SMP | |
364 | /* zero means no -deadline tasks */ | |
365 | dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0; | |
366 | ||
367 | dl_rq->dl_nr_migratory = 0; | |
368 | dl_rq->overloaded = 0; | |
2161573e | 369 | dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED; |
332ac17e DF |
370 | #else |
371 | init_dl_bw(&dl_rq->dl_bw); | |
1baca4ce | 372 | #endif |
e36d8677 LA |
373 | |
374 | dl_rq->running_bw = 0; | |
8fd27231 | 375 | dl_rq->this_bw = 0; |
4da3abce | 376 | init_dl_rq_bw_ratio(dl_rq); |
1baca4ce JL |
377 | } |
378 | ||
379 | #ifdef CONFIG_SMP | |
380 | ||
381 | static inline int dl_overloaded(struct rq *rq) | |
382 | { | |
383 | return atomic_read(&rq->rd->dlo_count); | |
384 | } | |
385 | ||
386 | static inline void dl_set_overload(struct rq *rq) | |
387 | { | |
388 | if (!rq->online) | |
389 | return; | |
390 | ||
391 | cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask); | |
392 | /* | |
393 | * Must be visible before the overload count is | |
394 | * set (as in sched_rt.c). | |
395 | * | |
396 | * Matched by the barrier in pull_dl_task(). | |
397 | */ | |
398 | smp_wmb(); | |
399 | atomic_inc(&rq->rd->dlo_count); | |
400 | } | |
401 | ||
402 | static inline void dl_clear_overload(struct rq *rq) | |
403 | { | |
404 | if (!rq->online) | |
405 | return; | |
406 | ||
407 | atomic_dec(&rq->rd->dlo_count); | |
408 | cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask); | |
409 | } | |
410 | ||
411 | static void update_dl_migration(struct dl_rq *dl_rq) | |
412 | { | |
995b9ea4 | 413 | if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) { |
1baca4ce JL |
414 | if (!dl_rq->overloaded) { |
415 | dl_set_overload(rq_of_dl_rq(dl_rq)); | |
416 | dl_rq->overloaded = 1; | |
417 | } | |
418 | } else if (dl_rq->overloaded) { | |
419 | dl_clear_overload(rq_of_dl_rq(dl_rq)); | |
420 | dl_rq->overloaded = 0; | |
421 | } | |
422 | } | |
423 | ||
424 | static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
425 | { | |
426 | struct task_struct *p = dl_task_of(dl_se); | |
1baca4ce | 427 | |
4b53a341 | 428 | if (p->nr_cpus_allowed > 1) |
1baca4ce JL |
429 | dl_rq->dl_nr_migratory++; |
430 | ||
431 | update_dl_migration(dl_rq); | |
432 | } | |
433 | ||
434 | static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
435 | { | |
436 | struct task_struct *p = dl_task_of(dl_se); | |
1baca4ce | 437 | |
4b53a341 | 438 | if (p->nr_cpus_allowed > 1) |
1baca4ce JL |
439 | dl_rq->dl_nr_migratory--; |
440 | ||
441 | update_dl_migration(dl_rq); | |
442 | } | |
443 | ||
444 | /* | |
445 | * The list of pushable -deadline task is not a plist, like in | |
446 | * sched_rt.c, it is an rb-tree with tasks ordered by deadline. | |
447 | */ | |
448 | static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
449 | { | |
450 | struct dl_rq *dl_rq = &rq->dl; | |
2161573e | 451 | struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_root.rb_node; |
1baca4ce JL |
452 | struct rb_node *parent = NULL; |
453 | struct task_struct *entry; | |
2161573e | 454 | bool leftmost = true; |
1baca4ce JL |
455 | |
456 | BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks)); | |
457 | ||
458 | while (*link) { | |
459 | parent = *link; | |
460 | entry = rb_entry(parent, struct task_struct, | |
461 | pushable_dl_tasks); | |
462 | if (dl_entity_preempt(&p->dl, &entry->dl)) | |
463 | link = &parent->rb_left; | |
464 | else { | |
465 | link = &parent->rb_right; | |
2161573e | 466 | leftmost = false; |
1baca4ce JL |
467 | } |
468 | } | |
469 | ||
2161573e | 470 | if (leftmost) |
7d92de3a | 471 | dl_rq->earliest_dl.next = p->dl.deadline; |
1baca4ce JL |
472 | |
473 | rb_link_node(&p->pushable_dl_tasks, parent, link); | |
2161573e DB |
474 | rb_insert_color_cached(&p->pushable_dl_tasks, |
475 | &dl_rq->pushable_dl_tasks_root, leftmost); | |
aab03e05 DF |
476 | } |
477 | ||
1baca4ce JL |
478 | static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) |
479 | { | |
480 | struct dl_rq *dl_rq = &rq->dl; | |
481 | ||
482 | if (RB_EMPTY_NODE(&p->pushable_dl_tasks)) | |
483 | return; | |
484 | ||
2161573e | 485 | if (dl_rq->pushable_dl_tasks_root.rb_leftmost == &p->pushable_dl_tasks) { |
1baca4ce JL |
486 | struct rb_node *next_node; |
487 | ||
488 | next_node = rb_next(&p->pushable_dl_tasks); | |
7d92de3a WL |
489 | if (next_node) { |
490 | dl_rq->earliest_dl.next = rb_entry(next_node, | |
491 | struct task_struct, pushable_dl_tasks)->dl.deadline; | |
492 | } | |
1baca4ce JL |
493 | } |
494 | ||
2161573e | 495 | rb_erase_cached(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); |
1baca4ce JL |
496 | RB_CLEAR_NODE(&p->pushable_dl_tasks); |
497 | } | |
498 | ||
499 | static inline int has_pushable_dl_tasks(struct rq *rq) | |
500 | { | |
2161573e | 501 | return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root); |
1baca4ce JL |
502 | } |
503 | ||
504 | static int push_dl_task(struct rq *rq); | |
505 | ||
dc877341 PZ |
506 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
507 | { | |
508 | return dl_task(prev); | |
509 | } | |
510 | ||
9916e214 PZ |
511 | static DEFINE_PER_CPU(struct callback_head, dl_push_head); |
512 | static DEFINE_PER_CPU(struct callback_head, dl_pull_head); | |
e3fca9e7 PZ |
513 | |
514 | static void push_dl_tasks(struct rq *); | |
9916e214 | 515 | static void pull_dl_task(struct rq *); |
e3fca9e7 PZ |
516 | |
517 | static inline void queue_push_tasks(struct rq *rq) | |
dc877341 | 518 | { |
e3fca9e7 PZ |
519 | if (!has_pushable_dl_tasks(rq)) |
520 | return; | |
521 | ||
9916e214 PZ |
522 | queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks); |
523 | } | |
524 | ||
525 | static inline void queue_pull_task(struct rq *rq) | |
526 | { | |
527 | queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task); | |
dc877341 PZ |
528 | } |
529 | ||
fa9c9d10 WL |
530 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq); |
531 | ||
a649f237 | 532 | static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p) |
fa9c9d10 WL |
533 | { |
534 | struct rq *later_rq = NULL; | |
fa9c9d10 WL |
535 | |
536 | later_rq = find_lock_later_rq(p, rq); | |
fa9c9d10 WL |
537 | if (!later_rq) { |
538 | int cpu; | |
539 | ||
540 | /* | |
541 | * If we cannot preempt any rq, fall back to pick any | |
542 | * online cpu. | |
543 | */ | |
0c98d344 | 544 | cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed); |
fa9c9d10 WL |
545 | if (cpu >= nr_cpu_ids) { |
546 | /* | |
547 | * Fail to find any suitable cpu. | |
548 | * The task will never come back! | |
549 | */ | |
550 | BUG_ON(dl_bandwidth_enabled()); | |
551 | ||
552 | /* | |
553 | * If admission control is disabled we | |
554 | * try a little harder to let the task | |
555 | * run. | |
556 | */ | |
557 | cpu = cpumask_any(cpu_active_mask); | |
558 | } | |
559 | later_rq = cpu_rq(cpu); | |
560 | double_lock_balance(rq, later_rq); | |
561 | } | |
562 | ||
fa9c9d10 | 563 | set_task_cpu(p, later_rq->cpu); |
a649f237 PZ |
564 | double_unlock_balance(later_rq, rq); |
565 | ||
566 | return later_rq; | |
fa9c9d10 WL |
567 | } |
568 | ||
1baca4ce JL |
569 | #else |
570 | ||
571 | static inline | |
572 | void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
573 | { | |
574 | } | |
575 | ||
576 | static inline | |
577 | void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) | |
578 | { | |
579 | } | |
580 | ||
581 | static inline | |
582 | void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
583 | { | |
584 | } | |
585 | ||
586 | static inline | |
587 | void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
588 | { | |
589 | } | |
590 | ||
dc877341 PZ |
591 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
592 | { | |
593 | return false; | |
594 | } | |
595 | ||
0ea60c20 | 596 | static inline void pull_dl_task(struct rq *rq) |
dc877341 | 597 | { |
dc877341 PZ |
598 | } |
599 | ||
e3fca9e7 | 600 | static inline void queue_push_tasks(struct rq *rq) |
dc877341 | 601 | { |
dc877341 PZ |
602 | } |
603 | ||
9916e214 | 604 | static inline void queue_pull_task(struct rq *rq) |
dc877341 PZ |
605 | { |
606 | } | |
1baca4ce JL |
607 | #endif /* CONFIG_SMP */ |
608 | ||
aab03e05 DF |
609 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags); |
610 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags); | |
611 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | |
612 | int flags); | |
613 | ||
614 | /* | |
615 | * We are being explicitly informed that a new instance is starting, | |
616 | * and this means that: | |
617 | * - the absolute deadline of the entity has to be placed at | |
618 | * current time + relative deadline; | |
619 | * - the runtime of the entity has to be set to the maximum value. | |
620 | * | |
621 | * The capability of specifying such event is useful whenever a -deadline | |
622 | * entity wants to (try to!) synchronize its behaviour with the scheduler's | |
623 | * one, and to (try to!) reconcile itself with its own scheduling | |
624 | * parameters. | |
625 | */ | |
98b0a857 | 626 | static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se) |
aab03e05 DF |
627 | { |
628 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
629 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
630 | ||
98b0a857 | 631 | WARN_ON(dl_se->dl_boosted); |
72f9f3fd LA |
632 | WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline)); |
633 | ||
634 | /* | |
635 | * We are racing with the deadline timer. So, do nothing because | |
636 | * the deadline timer handler will take care of properly recharging | |
637 | * the runtime and postponing the deadline | |
638 | */ | |
639 | if (dl_se->dl_throttled) | |
640 | return; | |
aab03e05 DF |
641 | |
642 | /* | |
643 | * We use the regular wall clock time to set deadlines in the | |
644 | * future; in fact, we must consider execution overheads (time | |
645 | * spent on hardirq context, etc.). | |
646 | */ | |
98b0a857 JL |
647 | dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline; |
648 | dl_se->runtime = dl_se->dl_runtime; | |
aab03e05 DF |
649 | } |
650 | ||
651 | /* | |
652 | * Pure Earliest Deadline First (EDF) scheduling does not deal with the | |
653 | * possibility of a entity lasting more than what it declared, and thus | |
654 | * exhausting its runtime. | |
655 | * | |
656 | * Here we are interested in making runtime overrun possible, but we do | |
657 | * not want a entity which is misbehaving to affect the scheduling of all | |
658 | * other entities. | |
659 | * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS) | |
660 | * is used, in order to confine each entity within its own bandwidth. | |
661 | * | |
662 | * This function deals exactly with that, and ensures that when the runtime | |
663 | * of a entity is replenished, its deadline is also postponed. That ensures | |
664 | * the overrunning entity can't interfere with other entity in the system and | |
665 | * can't make them miss their deadlines. Reasons why this kind of overruns | |
666 | * could happen are, typically, a entity voluntarily trying to overcome its | |
1b09d29b | 667 | * runtime, or it just underestimated it during sched_setattr(). |
aab03e05 | 668 | */ |
2d3d891d DF |
669 | static void replenish_dl_entity(struct sched_dl_entity *dl_se, |
670 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
671 | { |
672 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
673 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
674 | ||
2d3d891d DF |
675 | BUG_ON(pi_se->dl_runtime <= 0); |
676 | ||
677 | /* | |
678 | * This could be the case for a !-dl task that is boosted. | |
679 | * Just go with full inherited parameters. | |
680 | */ | |
681 | if (dl_se->dl_deadline == 0) { | |
682 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | |
683 | dl_se->runtime = pi_se->dl_runtime; | |
684 | } | |
685 | ||
48be3a67 PZ |
686 | if (dl_se->dl_yielded && dl_se->runtime > 0) |
687 | dl_se->runtime = 0; | |
688 | ||
aab03e05 DF |
689 | /* |
690 | * We keep moving the deadline away until we get some | |
691 | * available runtime for the entity. This ensures correct | |
692 | * handling of situations where the runtime overrun is | |
693 | * arbitrary large. | |
694 | */ | |
695 | while (dl_se->runtime <= 0) { | |
2d3d891d DF |
696 | dl_se->deadline += pi_se->dl_period; |
697 | dl_se->runtime += pi_se->dl_runtime; | |
aab03e05 DF |
698 | } |
699 | ||
700 | /* | |
701 | * At this point, the deadline really should be "in | |
702 | * the future" with respect to rq->clock. If it's | |
703 | * not, we are, for some reason, lagging too much! | |
704 | * Anyway, after having warn userspace abut that, | |
705 | * we still try to keep the things running by | |
706 | * resetting the deadline and the budget of the | |
707 | * entity. | |
708 | */ | |
709 | if (dl_time_before(dl_se->deadline, rq_clock(rq))) { | |
c219b7dd | 710 | printk_deferred_once("sched: DL replenish lagged too much\n"); |
2d3d891d DF |
711 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
712 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 | 713 | } |
1019a359 PZ |
714 | |
715 | if (dl_se->dl_yielded) | |
716 | dl_se->dl_yielded = 0; | |
717 | if (dl_se->dl_throttled) | |
718 | dl_se->dl_throttled = 0; | |
aab03e05 DF |
719 | } |
720 | ||
721 | /* | |
722 | * Here we check if --at time t-- an entity (which is probably being | |
723 | * [re]activated or, in general, enqueued) can use its remaining runtime | |
724 | * and its current deadline _without_ exceeding the bandwidth it is | |
725 | * assigned (function returns true if it can't). We are in fact applying | |
726 | * one of the CBS rules: when a task wakes up, if the residual runtime | |
727 | * over residual deadline fits within the allocated bandwidth, then we | |
728 | * can keep the current (absolute) deadline and residual budget without | |
729 | * disrupting the schedulability of the system. Otherwise, we should | |
730 | * refill the runtime and set the deadline a period in the future, | |
731 | * because keeping the current (absolute) deadline of the task would | |
712e5e34 DF |
732 | * result in breaking guarantees promised to other tasks (refer to |
733 | * Documentation/scheduler/sched-deadline.txt for more informations). | |
aab03e05 DF |
734 | * |
735 | * This function returns true if: | |
736 | * | |
2317d5f1 | 737 | * runtime / (deadline - t) > dl_runtime / dl_deadline , |
aab03e05 DF |
738 | * |
739 | * IOW we can't recycle current parameters. | |
755378a4 | 740 | * |
2317d5f1 | 741 | * Notice that the bandwidth check is done against the deadline. For |
755378a4 | 742 | * task with deadline equal to period this is the same of using |
2317d5f1 | 743 | * dl_period instead of dl_deadline in the equation above. |
aab03e05 | 744 | */ |
2d3d891d DF |
745 | static bool dl_entity_overflow(struct sched_dl_entity *dl_se, |
746 | struct sched_dl_entity *pi_se, u64 t) | |
aab03e05 DF |
747 | { |
748 | u64 left, right; | |
749 | ||
750 | /* | |
751 | * left and right are the two sides of the equation above, | |
752 | * after a bit of shuffling to use multiplications instead | |
753 | * of divisions. | |
754 | * | |
755 | * Note that none of the time values involved in the two | |
756 | * multiplications are absolute: dl_deadline and dl_runtime | |
757 | * are the relative deadline and the maximum runtime of each | |
758 | * instance, runtime is the runtime left for the last instance | |
759 | * and (deadline - t), since t is rq->clock, is the time left | |
760 | * to the (absolute) deadline. Even if overflowing the u64 type | |
761 | * is very unlikely to occur in both cases, here we scale down | |
762 | * as we want to avoid that risk at all. Scaling down by 10 | |
763 | * means that we reduce granularity to 1us. We are fine with it, | |
764 | * since this is only a true/false check and, anyway, thinking | |
765 | * of anything below microseconds resolution is actually fiction | |
766 | * (but still we want to give the user that illusion >;). | |
767 | */ | |
2317d5f1 | 768 | left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE); |
332ac17e DF |
769 | right = ((dl_se->deadline - t) >> DL_SCALE) * |
770 | (pi_se->dl_runtime >> DL_SCALE); | |
aab03e05 DF |
771 | |
772 | return dl_time_before(right, left); | |
773 | } | |
774 | ||
775 | /* | |
3effcb42 DBO |
776 | * Revised wakeup rule [1]: For self-suspending tasks, rather then |
777 | * re-initializing task's runtime and deadline, the revised wakeup | |
778 | * rule adjusts the task's runtime to avoid the task to overrun its | |
779 | * density. | |
aab03e05 | 780 | * |
3effcb42 DBO |
781 | * Reasoning: a task may overrun the density if: |
782 | * runtime / (deadline - t) > dl_runtime / dl_deadline | |
783 | * | |
784 | * Therefore, runtime can be adjusted to: | |
785 | * runtime = (dl_runtime / dl_deadline) * (deadline - t) | |
786 | * | |
787 | * In such way that runtime will be equal to the maximum density | |
788 | * the task can use without breaking any rule. | |
789 | * | |
790 | * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant | |
791 | * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24. | |
792 | */ | |
793 | static void | |
794 | update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq) | |
795 | { | |
796 | u64 laxity = dl_se->deadline - rq_clock(rq); | |
797 | ||
798 | /* | |
799 | * If the task has deadline < period, and the deadline is in the past, | |
800 | * it should already be throttled before this check. | |
801 | * | |
802 | * See update_dl_entity() comments for further details. | |
803 | */ | |
804 | WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq))); | |
805 | ||
806 | dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT; | |
807 | } | |
808 | ||
809 | /* | |
810 | * Regarding the deadline, a task with implicit deadline has a relative | |
811 | * deadline == relative period. A task with constrained deadline has a | |
812 | * relative deadline <= relative period. | |
813 | * | |
814 | * We support constrained deadline tasks. However, there are some restrictions | |
815 | * applied only for tasks which do not have an implicit deadline. See | |
816 | * update_dl_entity() to know more about such restrictions. | |
817 | * | |
818 | * The dl_is_implicit() returns true if the task has an implicit deadline. | |
819 | */ | |
820 | static inline bool dl_is_implicit(struct sched_dl_entity *dl_se) | |
821 | { | |
822 | return dl_se->dl_deadline == dl_se->dl_period; | |
823 | } | |
824 | ||
825 | /* | |
826 | * When a deadline entity is placed in the runqueue, its runtime and deadline | |
827 | * might need to be updated. This is done by a CBS wake up rule. There are two | |
828 | * different rules: 1) the original CBS; and 2) the Revisited CBS. | |
829 | * | |
830 | * When the task is starting a new period, the Original CBS is used. In this | |
831 | * case, the runtime is replenished and a new absolute deadline is set. | |
832 | * | |
833 | * When a task is queued before the begin of the next period, using the | |
834 | * remaining runtime and deadline could make the entity to overflow, see | |
835 | * dl_entity_overflow() to find more about runtime overflow. When such case | |
836 | * is detected, the runtime and deadline need to be updated. | |
837 | * | |
838 | * If the task has an implicit deadline, i.e., deadline == period, the Original | |
839 | * CBS is applied. the runtime is replenished and a new absolute deadline is | |
840 | * set, as in the previous cases. | |
841 | * | |
842 | * However, the Original CBS does not work properly for tasks with | |
843 | * deadline < period, which are said to have a constrained deadline. By | |
844 | * applying the Original CBS, a constrained deadline task would be able to run | |
845 | * runtime/deadline in a period. With deadline < period, the task would | |
846 | * overrun the runtime/period allowed bandwidth, breaking the admission test. | |
847 | * | |
848 | * In order to prevent this misbehave, the Revisited CBS is used for | |
849 | * constrained deadline tasks when a runtime overflow is detected. In the | |
850 | * Revisited CBS, rather than replenishing & setting a new absolute deadline, | |
851 | * the remaining runtime of the task is reduced to avoid runtime overflow. | |
852 | * Please refer to the comments update_dl_revised_wakeup() function to find | |
853 | * more about the Revised CBS rule. | |
aab03e05 | 854 | */ |
2d3d891d DF |
855 | static void update_dl_entity(struct sched_dl_entity *dl_se, |
856 | struct sched_dl_entity *pi_se) | |
aab03e05 DF |
857 | { |
858 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
859 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
860 | ||
aab03e05 | 861 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) || |
2d3d891d | 862 | dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { |
3effcb42 DBO |
863 | |
864 | if (unlikely(!dl_is_implicit(dl_se) && | |
865 | !dl_time_before(dl_se->deadline, rq_clock(rq)) && | |
866 | !dl_se->dl_boosted)){ | |
867 | update_dl_revised_wakeup(dl_se, rq); | |
868 | return; | |
869 | } | |
870 | ||
2d3d891d DF |
871 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
872 | dl_se->runtime = pi_se->dl_runtime; | |
aab03e05 DF |
873 | } |
874 | } | |
875 | ||
5ac69d37 DBO |
876 | static inline u64 dl_next_period(struct sched_dl_entity *dl_se) |
877 | { | |
878 | return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period; | |
879 | } | |
880 | ||
aab03e05 DF |
881 | /* |
882 | * If the entity depleted all its runtime, and if we want it to sleep | |
883 | * while waiting for some new execution time to become available, we | |
5ac69d37 | 884 | * set the bandwidth replenishment timer to the replenishment instant |
aab03e05 DF |
885 | * and try to activate it. |
886 | * | |
887 | * Notice that it is important for the caller to know if the timer | |
888 | * actually started or not (i.e., the replenishment instant is in | |
889 | * the future or in the past). | |
890 | */ | |
a649f237 | 891 | static int start_dl_timer(struct task_struct *p) |
aab03e05 | 892 | { |
a649f237 PZ |
893 | struct sched_dl_entity *dl_se = &p->dl; |
894 | struct hrtimer *timer = &dl_se->dl_timer; | |
895 | struct rq *rq = task_rq(p); | |
aab03e05 | 896 | ktime_t now, act; |
aab03e05 DF |
897 | s64 delta; |
898 | ||
a649f237 PZ |
899 | lockdep_assert_held(&rq->lock); |
900 | ||
aab03e05 DF |
901 | /* |
902 | * We want the timer to fire at the deadline, but considering | |
903 | * that it is actually coming from rq->clock and not from | |
904 | * hrtimer's time base reading. | |
905 | */ | |
5ac69d37 | 906 | act = ns_to_ktime(dl_next_period(dl_se)); |
a649f237 | 907 | now = hrtimer_cb_get_time(timer); |
aab03e05 DF |
908 | delta = ktime_to_ns(now) - rq_clock(rq); |
909 | act = ktime_add_ns(act, delta); | |
910 | ||
911 | /* | |
912 | * If the expiry time already passed, e.g., because the value | |
913 | * chosen as the deadline is too small, don't even try to | |
914 | * start the timer in the past! | |
915 | */ | |
916 | if (ktime_us_delta(act, now) < 0) | |
917 | return 0; | |
918 | ||
a649f237 PZ |
919 | /* |
920 | * !enqueued will guarantee another callback; even if one is already in | |
921 | * progress. This ensures a balanced {get,put}_task_struct(). | |
922 | * | |
923 | * The race against __run_timer() clearing the enqueued state is | |
924 | * harmless because we're holding task_rq()->lock, therefore the timer | |
925 | * expiring after we've done the check will wait on its task_rq_lock() | |
926 | * and observe our state. | |
927 | */ | |
928 | if (!hrtimer_is_queued(timer)) { | |
929 | get_task_struct(p); | |
930 | hrtimer_start(timer, act, HRTIMER_MODE_ABS); | |
931 | } | |
aab03e05 | 932 | |
cc9684d3 | 933 | return 1; |
aab03e05 DF |
934 | } |
935 | ||
936 | /* | |
937 | * This is the bandwidth enforcement timer callback. If here, we know | |
938 | * a task is not on its dl_rq, since the fact that the timer was running | |
939 | * means the task is throttled and needs a runtime replenishment. | |
940 | * | |
941 | * However, what we actually do depends on the fact the task is active, | |
942 | * (it is on its rq) or has been removed from there by a call to | |
943 | * dequeue_task_dl(). In the former case we must issue the runtime | |
944 | * replenishment and add the task back to the dl_rq; in the latter, we just | |
945 | * do nothing but clearing dl_throttled, so that runtime and deadline | |
946 | * updating (and the queueing back to dl_rq) will be done by the | |
947 | * next call to enqueue_task_dl(). | |
948 | */ | |
949 | static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) | |
950 | { | |
951 | struct sched_dl_entity *dl_se = container_of(timer, | |
952 | struct sched_dl_entity, | |
953 | dl_timer); | |
954 | struct task_struct *p = dl_task_of(dl_se); | |
eb580751 | 955 | struct rq_flags rf; |
0f397f2c | 956 | struct rq *rq; |
3960c8c0 | 957 | |
eb580751 | 958 | rq = task_rq_lock(p, &rf); |
0f397f2c | 959 | |
aab03e05 | 960 | /* |
a649f237 | 961 | * The task might have changed its scheduling policy to something |
9846d50d | 962 | * different than SCHED_DEADLINE (through switched_from_dl()). |
a649f237 | 963 | */ |
209a0cbd | 964 | if (!dl_task(p)) |
a649f237 | 965 | goto unlock; |
a649f237 | 966 | |
a649f237 PZ |
967 | /* |
968 | * The task might have been boosted by someone else and might be in the | |
969 | * boosting/deboosting path, its not throttled. | |
970 | */ | |
971 | if (dl_se->dl_boosted) | |
972 | goto unlock; | |
a79ec89f | 973 | |
fa9c9d10 | 974 | /* |
a649f237 PZ |
975 | * Spurious timer due to start_dl_timer() race; or we already received |
976 | * a replenishment from rt_mutex_setprio(). | |
fa9c9d10 | 977 | */ |
a649f237 | 978 | if (!dl_se->dl_throttled) |
fa9c9d10 | 979 | goto unlock; |
a649f237 PZ |
980 | |
981 | sched_clock_tick(); | |
982 | update_rq_clock(rq); | |
fa9c9d10 | 983 | |
a79ec89f KT |
984 | /* |
985 | * If the throttle happened during sched-out; like: | |
986 | * | |
987 | * schedule() | |
988 | * deactivate_task() | |
989 | * dequeue_task_dl() | |
990 | * update_curr_dl() | |
991 | * start_dl_timer() | |
992 | * __dequeue_task_dl() | |
993 | * prev->on_rq = 0; | |
994 | * | |
995 | * We can be both throttled and !queued. Replenish the counter | |
996 | * but do not enqueue -- wait for our wakeup to do that. | |
997 | */ | |
998 | if (!task_on_rq_queued(p)) { | |
999 | replenish_dl_entity(dl_se, dl_se); | |
1000 | goto unlock; | |
1001 | } | |
1002 | ||
1baca4ce | 1003 | #ifdef CONFIG_SMP |
c0c8c9fa | 1004 | if (unlikely(!rq->online)) { |
61c7aca6 WL |
1005 | /* |
1006 | * If the runqueue is no longer available, migrate the | |
1007 | * task elsewhere. This necessarily changes rq. | |
1008 | */ | |
c0c8c9fa | 1009 | lockdep_unpin_lock(&rq->lock, rf.cookie); |
a649f237 | 1010 | rq = dl_task_offline_migration(rq, p); |
c0c8c9fa | 1011 | rf.cookie = lockdep_pin_lock(&rq->lock); |
dcc3b5ff | 1012 | update_rq_clock(rq); |
61c7aca6 WL |
1013 | |
1014 | /* | |
1015 | * Now that the task has been migrated to the new RQ and we | |
1016 | * have that locked, proceed as normal and enqueue the task | |
1017 | * there. | |
1018 | */ | |
c0c8c9fa | 1019 | } |
61c7aca6 | 1020 | #endif |
a649f237 | 1021 | |
61c7aca6 WL |
1022 | enqueue_task_dl(rq, p, ENQUEUE_REPLENISH); |
1023 | if (dl_task(rq->curr)) | |
1024 | check_preempt_curr_dl(rq, p, 0); | |
1025 | else | |
1026 | resched_curr(rq); | |
a649f237 | 1027 | |
61c7aca6 | 1028 | #ifdef CONFIG_SMP |
a649f237 PZ |
1029 | /* |
1030 | * Queueing this task back might have overloaded rq, check if we need | |
1031 | * to kick someone away. | |
1019a359 | 1032 | */ |
0aaafaab PZ |
1033 | if (has_pushable_dl_tasks(rq)) { |
1034 | /* | |
1035 | * Nothing relies on rq->lock after this, so its safe to drop | |
1036 | * rq->lock. | |
1037 | */ | |
d8ac8971 | 1038 | rq_unpin_lock(rq, &rf); |
1019a359 | 1039 | push_dl_task(rq); |
d8ac8971 | 1040 | rq_repin_lock(rq, &rf); |
0aaafaab | 1041 | } |
1baca4ce | 1042 | #endif |
a649f237 | 1043 | |
aab03e05 | 1044 | unlock: |
eb580751 | 1045 | task_rq_unlock(rq, p, &rf); |
aab03e05 | 1046 | |
a649f237 PZ |
1047 | /* |
1048 | * This can free the task_struct, including this hrtimer, do not touch | |
1049 | * anything related to that after this. | |
1050 | */ | |
1051 | put_task_struct(p); | |
1052 | ||
aab03e05 DF |
1053 | return HRTIMER_NORESTART; |
1054 | } | |
1055 | ||
1056 | void init_dl_task_timer(struct sched_dl_entity *dl_se) | |
1057 | { | |
1058 | struct hrtimer *timer = &dl_se->dl_timer; | |
1059 | ||
aab03e05 DF |
1060 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
1061 | timer->function = dl_task_timer; | |
1062 | } | |
1063 | ||
df8eac8c DBO |
1064 | /* |
1065 | * During the activation, CBS checks if it can reuse the current task's | |
1066 | * runtime and period. If the deadline of the task is in the past, CBS | |
1067 | * cannot use the runtime, and so it replenishes the task. This rule | |
1068 | * works fine for implicit deadline tasks (deadline == period), and the | |
1069 | * CBS was designed for implicit deadline tasks. However, a task with | |
1070 | * constrained deadline (deadine < period) might be awakened after the | |
1071 | * deadline, but before the next period. In this case, replenishing the | |
1072 | * task would allow it to run for runtime / deadline. As in this case | |
1073 | * deadline < period, CBS enables a task to run for more than the | |
1074 | * runtime / period. In a very loaded system, this can cause a domino | |
1075 | * effect, making other tasks miss their deadlines. | |
1076 | * | |
1077 | * To avoid this problem, in the activation of a constrained deadline | |
1078 | * task after the deadline but before the next period, throttle the | |
1079 | * task and set the replenishing timer to the begin of the next period, | |
1080 | * unless it is boosted. | |
1081 | */ | |
1082 | static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se) | |
1083 | { | |
1084 | struct task_struct *p = dl_task_of(dl_se); | |
1085 | struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se)); | |
1086 | ||
1087 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) && | |
1088 | dl_time_before(rq_clock(rq), dl_next_period(dl_se))) { | |
1089 | if (unlikely(dl_se->dl_boosted || !start_dl_timer(p))) | |
1090 | return; | |
1091 | dl_se->dl_throttled = 1; | |
ae83b56a XP |
1092 | if (dl_se->runtime > 0) |
1093 | dl_se->runtime = 0; | |
df8eac8c DBO |
1094 | } |
1095 | } | |
1096 | ||
aab03e05 | 1097 | static |
6fab5410 | 1098 | int dl_runtime_exceeded(struct sched_dl_entity *dl_se) |
aab03e05 | 1099 | { |
269ad801 | 1100 | return (dl_se->runtime <= 0); |
aab03e05 DF |
1101 | } |
1102 | ||
faa59937 JL |
1103 | extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq); |
1104 | ||
c52f14d3 LA |
1105 | /* |
1106 | * This function implements the GRUB accounting rule: | |
1107 | * according to the GRUB reclaiming algorithm, the runtime is | |
daec5798 LA |
1108 | * not decreased as "dq = -dt", but as |
1109 | * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt", | |
1110 | * where u is the utilization of the task, Umax is the maximum reclaimable | |
1111 | * utilization, Uinact is the (per-runqueue) inactive utilization, computed | |
1112 | * as the difference between the "total runqueue utilization" and the | |
1113 | * runqueue active utilization, and Uextra is the (per runqueue) extra | |
1114 | * reclaimable utilization. | |
9f0d1a50 | 1115 | * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations |
daec5798 LA |
1116 | * multiplied by 2^BW_SHIFT, the result has to be shifted right by |
1117 | * BW_SHIFT. | |
1118 | * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT, | |
1119 | * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT. | |
1120 | * Since delta is a 64 bit variable, to have an overflow its value | |
1121 | * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds. | |
1122 | * So, overflow is not an issue here. | |
c52f14d3 | 1123 | */ |
9f0d1a50 | 1124 | u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se) |
c52f14d3 | 1125 | { |
9f0d1a50 LA |
1126 | u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */ |
1127 | u64 u_act; | |
daec5798 | 1128 | u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT; |
c52f14d3 | 1129 | |
9f0d1a50 | 1130 | /* |
daec5798 LA |
1131 | * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)}, |
1132 | * we compare u_inact + rq->dl.extra_bw with | |
1133 | * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because | |
1134 | * u_inact + rq->dl.extra_bw can be larger than | |
1135 | * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative | |
1136 | * leading to wrong results) | |
9f0d1a50 | 1137 | */ |
daec5798 LA |
1138 | if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min) |
1139 | u_act = u_act_min; | |
9f0d1a50 | 1140 | else |
daec5798 | 1141 | u_act = BW_UNIT - u_inact - rq->dl.extra_bw; |
9f0d1a50 LA |
1142 | |
1143 | return (delta * u_act) >> BW_SHIFT; | |
c52f14d3 LA |
1144 | } |
1145 | ||
aab03e05 DF |
1146 | /* |
1147 | * Update the current task's runtime statistics (provided it is still | |
1148 | * a -deadline task and has not been removed from the dl_rq). | |
1149 | */ | |
1150 | static void update_curr_dl(struct rq *rq) | |
1151 | { | |
1152 | struct task_struct *curr = rq->curr; | |
1153 | struct sched_dl_entity *dl_se = &curr->dl; | |
07881166 JL |
1154 | u64 delta_exec, scaled_delta_exec; |
1155 | int cpu = cpu_of(rq); | |
6fe0ce1e | 1156 | u64 now; |
aab03e05 DF |
1157 | |
1158 | if (!dl_task(curr) || !on_dl_rq(dl_se)) | |
1159 | return; | |
1160 | ||
1161 | /* | |
1162 | * Consumed budget is computed considering the time as | |
1163 | * observed by schedulable tasks (excluding time spent | |
1164 | * in hardirq context, etc.). Deadlines are instead | |
1165 | * computed using hard walltime. This seems to be the more | |
1166 | * natural solution, but the full ramifications of this | |
1167 | * approach need further study. | |
1168 | */ | |
6fe0ce1e WY |
1169 | now = rq_clock_task(rq); |
1170 | delta_exec = now - curr->se.exec_start; | |
48be3a67 PZ |
1171 | if (unlikely((s64)delta_exec <= 0)) { |
1172 | if (unlikely(dl_se->dl_yielded)) | |
1173 | goto throttle; | |
734ff2a7 | 1174 | return; |
48be3a67 | 1175 | } |
aab03e05 DF |
1176 | |
1177 | schedstat_set(curr->se.statistics.exec_max, | |
1178 | max(curr->se.statistics.exec_max, delta_exec)); | |
1179 | ||
1180 | curr->se.sum_exec_runtime += delta_exec; | |
1181 | account_group_exec_runtime(curr, delta_exec); | |
1182 | ||
6fe0ce1e | 1183 | curr->se.exec_start = now; |
d2cc5ed6 | 1184 | cgroup_account_cputime(curr, delta_exec); |
aab03e05 | 1185 | |
239be4a9 DF |
1186 | sched_rt_avg_update(rq, delta_exec); |
1187 | ||
794a56eb JL |
1188 | if (dl_entity_is_special(dl_se)) |
1189 | return; | |
1190 | ||
07881166 JL |
1191 | /* |
1192 | * For tasks that participate in GRUB, we implement GRUB-PA: the | |
1193 | * spare reclaimed bandwidth is used to clock down frequency. | |
1194 | * | |
1195 | * For the others, we still need to scale reservation parameters | |
1196 | * according to current frequency and CPU maximum capacity. | |
1197 | */ | |
1198 | if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) { | |
1199 | scaled_delta_exec = grub_reclaim(delta_exec, | |
1200 | rq, | |
1201 | &curr->dl); | |
1202 | } else { | |
1203 | unsigned long scale_freq = arch_scale_freq_capacity(cpu); | |
1204 | unsigned long scale_cpu = arch_scale_cpu_capacity(NULL, cpu); | |
1205 | ||
1206 | scaled_delta_exec = cap_scale(delta_exec, scale_freq); | |
1207 | scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu); | |
1208 | } | |
1209 | ||
1210 | dl_se->runtime -= scaled_delta_exec; | |
48be3a67 PZ |
1211 | |
1212 | throttle: | |
1213 | if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) { | |
1019a359 | 1214 | dl_se->dl_throttled = 1; |
34be3930 JL |
1215 | |
1216 | /* If requested, inform the user about runtime overruns. */ | |
1217 | if (dl_runtime_exceeded(dl_se) && | |
1218 | (dl_se->flags & SCHED_FLAG_DL_OVERRUN)) | |
1219 | dl_se->dl_overrun = 1; | |
1220 | ||
aab03e05 | 1221 | __dequeue_task_dl(rq, curr, 0); |
a649f237 | 1222 | if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr))) |
aab03e05 DF |
1223 | enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH); |
1224 | ||
1225 | if (!is_leftmost(curr, &rq->dl)) | |
8875125e | 1226 | resched_curr(rq); |
aab03e05 | 1227 | } |
1724813d PZ |
1228 | |
1229 | /* | |
1230 | * Because -- for now -- we share the rt bandwidth, we need to | |
1231 | * account our runtime there too, otherwise actual rt tasks | |
1232 | * would be able to exceed the shared quota. | |
1233 | * | |
1234 | * Account to the root rt group for now. | |
1235 | * | |
1236 | * The solution we're working towards is having the RT groups scheduled | |
1237 | * using deadline servers -- however there's a few nasties to figure | |
1238 | * out before that can happen. | |
1239 | */ | |
1240 | if (rt_bandwidth_enabled()) { | |
1241 | struct rt_rq *rt_rq = &rq->rt; | |
1242 | ||
1243 | raw_spin_lock(&rt_rq->rt_runtime_lock); | |
1724813d PZ |
1244 | /* |
1245 | * We'll let actual RT tasks worry about the overflow here, we | |
faa59937 JL |
1246 | * have our own CBS to keep us inline; only account when RT |
1247 | * bandwidth is relevant. | |
1724813d | 1248 | */ |
faa59937 JL |
1249 | if (sched_rt_bandwidth_account(rt_rq)) |
1250 | rt_rq->rt_time += delta_exec; | |
1724813d PZ |
1251 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
1252 | } | |
aab03e05 DF |
1253 | } |
1254 | ||
209a0cbd LA |
1255 | static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer) |
1256 | { | |
1257 | struct sched_dl_entity *dl_se = container_of(timer, | |
1258 | struct sched_dl_entity, | |
1259 | inactive_timer); | |
1260 | struct task_struct *p = dl_task_of(dl_se); | |
1261 | struct rq_flags rf; | |
1262 | struct rq *rq; | |
1263 | ||
1264 | rq = task_rq_lock(p, &rf); | |
1265 | ||
1266 | if (!dl_task(p) || p->state == TASK_DEAD) { | |
387e3130 LA |
1267 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); |
1268 | ||
209a0cbd | 1269 | if (p->state == TASK_DEAD && dl_se->dl_non_contending) { |
794a56eb JL |
1270 | sub_running_bw(&p->dl, dl_rq_of_se(&p->dl)); |
1271 | sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl)); | |
209a0cbd LA |
1272 | dl_se->dl_non_contending = 0; |
1273 | } | |
387e3130 LA |
1274 | |
1275 | raw_spin_lock(&dl_b->lock); | |
8c0944ce | 1276 | __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); |
387e3130 | 1277 | raw_spin_unlock(&dl_b->lock); |
209a0cbd LA |
1278 | __dl_clear_params(p); |
1279 | ||
1280 | goto unlock; | |
1281 | } | |
1282 | if (dl_se->dl_non_contending == 0) | |
1283 | goto unlock; | |
1284 | ||
1285 | sched_clock_tick(); | |
1286 | update_rq_clock(rq); | |
1287 | ||
794a56eb | 1288 | sub_running_bw(dl_se, &rq->dl); |
209a0cbd LA |
1289 | dl_se->dl_non_contending = 0; |
1290 | unlock: | |
1291 | task_rq_unlock(rq, p, &rf); | |
1292 | put_task_struct(p); | |
1293 | ||
1294 | return HRTIMER_NORESTART; | |
1295 | } | |
1296 | ||
1297 | void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se) | |
1298 | { | |
1299 | struct hrtimer *timer = &dl_se->inactive_timer; | |
1300 | ||
1301 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1302 | timer->function = inactive_task_timer; | |
1303 | } | |
1304 | ||
1baca4ce JL |
1305 | #ifdef CONFIG_SMP |
1306 | ||
1baca4ce JL |
1307 | static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) |
1308 | { | |
1309 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
1310 | ||
1311 | if (dl_rq->earliest_dl.curr == 0 || | |
1312 | dl_time_before(deadline, dl_rq->earliest_dl.curr)) { | |
1baca4ce | 1313 | dl_rq->earliest_dl.curr = deadline; |
d8206bb3 | 1314 | cpudl_set(&rq->rd->cpudl, rq->cpu, deadline); |
1baca4ce JL |
1315 | } |
1316 | } | |
1317 | ||
1318 | static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) | |
1319 | { | |
1320 | struct rq *rq = rq_of_dl_rq(dl_rq); | |
1321 | ||
1322 | /* | |
1323 | * Since we may have removed our earliest (and/or next earliest) | |
1324 | * task we must recompute them. | |
1325 | */ | |
1326 | if (!dl_rq->dl_nr_running) { | |
1327 | dl_rq->earliest_dl.curr = 0; | |
1328 | dl_rq->earliest_dl.next = 0; | |
d8206bb3 | 1329 | cpudl_clear(&rq->rd->cpudl, rq->cpu); |
1baca4ce | 1330 | } else { |
2161573e | 1331 | struct rb_node *leftmost = dl_rq->root.rb_leftmost; |
1baca4ce JL |
1332 | struct sched_dl_entity *entry; |
1333 | ||
1334 | entry = rb_entry(leftmost, struct sched_dl_entity, rb_node); | |
1335 | dl_rq->earliest_dl.curr = entry->deadline; | |
d8206bb3 | 1336 | cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline); |
1baca4ce JL |
1337 | } |
1338 | } | |
1339 | ||
1340 | #else | |
1341 | ||
1342 | static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
1343 | static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | |
1344 | ||
1345 | #endif /* CONFIG_SMP */ | |
1346 | ||
1347 | static inline | |
1348 | void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
1349 | { | |
1350 | int prio = dl_task_of(dl_se)->prio; | |
1351 | u64 deadline = dl_se->deadline; | |
1352 | ||
1353 | WARN_ON(!dl_prio(prio)); | |
1354 | dl_rq->dl_nr_running++; | |
72465447 | 1355 | add_nr_running(rq_of_dl_rq(dl_rq), 1); |
1baca4ce JL |
1356 | |
1357 | inc_dl_deadline(dl_rq, deadline); | |
1358 | inc_dl_migration(dl_se, dl_rq); | |
1359 | } | |
1360 | ||
1361 | static inline | |
1362 | void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | |
1363 | { | |
1364 | int prio = dl_task_of(dl_se)->prio; | |
1365 | ||
1366 | WARN_ON(!dl_prio(prio)); | |
1367 | WARN_ON(!dl_rq->dl_nr_running); | |
1368 | dl_rq->dl_nr_running--; | |
72465447 | 1369 | sub_nr_running(rq_of_dl_rq(dl_rq), 1); |
1baca4ce JL |
1370 | |
1371 | dec_dl_deadline(dl_rq, dl_se->deadline); | |
1372 | dec_dl_migration(dl_se, dl_rq); | |
1373 | } | |
1374 | ||
aab03e05 DF |
1375 | static void __enqueue_dl_entity(struct sched_dl_entity *dl_se) |
1376 | { | |
1377 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
2161573e | 1378 | struct rb_node **link = &dl_rq->root.rb_root.rb_node; |
aab03e05 DF |
1379 | struct rb_node *parent = NULL; |
1380 | struct sched_dl_entity *entry; | |
1381 | int leftmost = 1; | |
1382 | ||
1383 | BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node)); | |
1384 | ||
1385 | while (*link) { | |
1386 | parent = *link; | |
1387 | entry = rb_entry(parent, struct sched_dl_entity, rb_node); | |
1388 | if (dl_time_before(dl_se->deadline, entry->deadline)) | |
1389 | link = &parent->rb_left; | |
1390 | else { | |
1391 | link = &parent->rb_right; | |
1392 | leftmost = 0; | |
1393 | } | |
1394 | } | |
1395 | ||
aab03e05 | 1396 | rb_link_node(&dl_se->rb_node, parent, link); |
2161573e | 1397 | rb_insert_color_cached(&dl_se->rb_node, &dl_rq->root, leftmost); |
aab03e05 | 1398 | |
1baca4ce | 1399 | inc_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
1400 | } |
1401 | ||
1402 | static void __dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
1403 | { | |
1404 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | |
1405 | ||
1406 | if (RB_EMPTY_NODE(&dl_se->rb_node)) | |
1407 | return; | |
1408 | ||
2161573e | 1409 | rb_erase_cached(&dl_se->rb_node, &dl_rq->root); |
aab03e05 DF |
1410 | RB_CLEAR_NODE(&dl_se->rb_node); |
1411 | ||
1baca4ce | 1412 | dec_dl_tasks(dl_se, dl_rq); |
aab03e05 DF |
1413 | } |
1414 | ||
1415 | static void | |
2d3d891d DF |
1416 | enqueue_dl_entity(struct sched_dl_entity *dl_se, |
1417 | struct sched_dl_entity *pi_se, int flags) | |
aab03e05 DF |
1418 | { |
1419 | BUG_ON(on_dl_rq(dl_se)); | |
1420 | ||
1421 | /* | |
1422 | * If this is a wakeup or a new instance, the scheduling | |
1423 | * parameters of the task might need updating. Otherwise, | |
1424 | * we want a replenishment of its runtime. | |
1425 | */ | |
e36d8677 | 1426 | if (flags & ENQUEUE_WAKEUP) { |
8fd27231 | 1427 | task_contending(dl_se, flags); |
2d3d891d | 1428 | update_dl_entity(dl_se, pi_se); |
e36d8677 | 1429 | } else if (flags & ENQUEUE_REPLENISH) { |
6a503c3b | 1430 | replenish_dl_entity(dl_se, pi_se); |
295d6d5e LA |
1431 | } else if ((flags & ENQUEUE_RESTORE) && |
1432 | dl_time_before(dl_se->deadline, | |
1433 | rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) { | |
1434 | setup_new_dl_entity(dl_se); | |
e36d8677 | 1435 | } |
aab03e05 DF |
1436 | |
1437 | __enqueue_dl_entity(dl_se); | |
1438 | } | |
1439 | ||
1440 | static void dequeue_dl_entity(struct sched_dl_entity *dl_se) | |
1441 | { | |
1442 | __dequeue_dl_entity(dl_se); | |
1443 | } | |
1444 | ||
1445 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
1446 | { | |
2d3d891d DF |
1447 | struct task_struct *pi_task = rt_mutex_get_top_task(p); |
1448 | struct sched_dl_entity *pi_se = &p->dl; | |
1449 | ||
1450 | /* | |
193be41e JF |
1451 | * Use the scheduling parameters of the top pi-waiter task if: |
1452 | * - we have a top pi-waiter which is a SCHED_DEADLINE task AND | |
1453 | * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is | |
1454 | * smaller than our deadline OR we are a !SCHED_DEADLINE task getting | |
1455 | * boosted due to a SCHED_DEADLINE pi-waiter). | |
1456 | * Otherwise we keep our runtime and deadline. | |
2d3d891d | 1457 | */ |
193be41e | 1458 | if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) { |
2d3d891d | 1459 | pi_se = &pi_task->dl; |
64be6f1f JL |
1460 | } else if (!dl_prio(p->normal_prio)) { |
1461 | /* | |
1462 | * Special case in which we have a !SCHED_DEADLINE task | |
193be41e | 1463 | * that is going to be deboosted, but exceeds its |
64be6f1f JL |
1464 | * runtime while doing so. No point in replenishing |
1465 | * it, as it's going to return back to its original | |
1466 | * scheduling class after this. | |
1467 | */ | |
1468 | BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH); | |
1469 | return; | |
1470 | } | |
2d3d891d | 1471 | |
df8eac8c DBO |
1472 | /* |
1473 | * Check if a constrained deadline task was activated | |
1474 | * after the deadline but before the next period. | |
1475 | * If that is the case, the task will be throttled and | |
1476 | * the replenishment timer will be set to the next period. | |
1477 | */ | |
3effcb42 | 1478 | if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl)) |
df8eac8c DBO |
1479 | dl_check_constrained_dl(&p->dl); |
1480 | ||
8fd27231 | 1481 | if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) { |
794a56eb JL |
1482 | add_rq_bw(&p->dl, &rq->dl); |
1483 | add_running_bw(&p->dl, &rq->dl); | |
8fd27231 | 1484 | } |
e36d8677 | 1485 | |
aab03e05 | 1486 | /* |
e36d8677 | 1487 | * If p is throttled, we do not enqueue it. In fact, if it exhausted |
aab03e05 DF |
1488 | * its budget it needs a replenishment and, since it now is on |
1489 | * its rq, the bandwidth timer callback (which clearly has not | |
1490 | * run yet) will take care of this. | |
e36d8677 LA |
1491 | * However, the active utilization does not depend on the fact |
1492 | * that the task is on the runqueue or not (but depends on the | |
1493 | * task's state - in GRUB parlance, "inactive" vs "active contending"). | |
1494 | * In other words, even if a task is throttled its utilization must | |
1495 | * be counted in the active utilization; hence, we need to call | |
1496 | * add_running_bw(). | |
aab03e05 | 1497 | */ |
e36d8677 | 1498 | if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) { |
209a0cbd | 1499 | if (flags & ENQUEUE_WAKEUP) |
8fd27231 | 1500 | task_contending(&p->dl, flags); |
209a0cbd | 1501 | |
aab03e05 | 1502 | return; |
e36d8677 | 1503 | } |
aab03e05 | 1504 | |
2d3d891d | 1505 | enqueue_dl_entity(&p->dl, pi_se, flags); |
1baca4ce | 1506 | |
4b53a341 | 1507 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) |
1baca4ce | 1508 | enqueue_pushable_dl_task(rq, p); |
aab03e05 DF |
1509 | } |
1510 | ||
1511 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
1512 | { | |
1513 | dequeue_dl_entity(&p->dl); | |
1baca4ce | 1514 | dequeue_pushable_dl_task(rq, p); |
aab03e05 DF |
1515 | } |
1516 | ||
1517 | static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | |
1518 | { | |
1519 | update_curr_dl(rq); | |
1520 | __dequeue_task_dl(rq, p, flags); | |
e36d8677 | 1521 | |
8fd27231 | 1522 | if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) { |
794a56eb JL |
1523 | sub_running_bw(&p->dl, &rq->dl); |
1524 | sub_rq_bw(&p->dl, &rq->dl); | |
8fd27231 | 1525 | } |
e36d8677 LA |
1526 | |
1527 | /* | |
209a0cbd LA |
1528 | * This check allows to start the inactive timer (or to immediately |
1529 | * decrease the active utilization, if needed) in two cases: | |
e36d8677 LA |
1530 | * when the task blocks and when it is terminating |
1531 | * (p->state == TASK_DEAD). We can handle the two cases in the same | |
1532 | * way, because from GRUB's point of view the same thing is happening | |
1533 | * (the task moves from "active contending" to "active non contending" | |
1534 | * or "inactive") | |
1535 | */ | |
1536 | if (flags & DEQUEUE_SLEEP) | |
209a0cbd | 1537 | task_non_contending(p); |
aab03e05 DF |
1538 | } |
1539 | ||
1540 | /* | |
1541 | * Yield task semantic for -deadline tasks is: | |
1542 | * | |
1543 | * get off from the CPU until our next instance, with | |
1544 | * a new runtime. This is of little use now, since we | |
1545 | * don't have a bandwidth reclaiming mechanism. Anyway, | |
1546 | * bandwidth reclaiming is planned for the future, and | |
1547 | * yield_task_dl will indicate that some spare budget | |
1548 | * is available for other task instances to use it. | |
1549 | */ | |
1550 | static void yield_task_dl(struct rq *rq) | |
1551 | { | |
aab03e05 DF |
1552 | /* |
1553 | * We make the task go to sleep until its current deadline by | |
1554 | * forcing its runtime to zero. This way, update_curr_dl() stops | |
1555 | * it and the bandwidth timer will wake it up and will give it | |
5bfd126e | 1556 | * new scheduling parameters (thanks to dl_yielded=1). |
aab03e05 | 1557 | */ |
48be3a67 PZ |
1558 | rq->curr->dl.dl_yielded = 1; |
1559 | ||
6f1607f1 | 1560 | update_rq_clock(rq); |
aab03e05 | 1561 | update_curr_dl(rq); |
44fb085b WL |
1562 | /* |
1563 | * Tell update_rq_clock() that we've just updated, | |
1564 | * so we don't do microscopic update in schedule() | |
1565 | * and double the fastpath cost. | |
1566 | */ | |
1567 | rq_clock_skip_update(rq, true); | |
aab03e05 DF |
1568 | } |
1569 | ||
1baca4ce JL |
1570 | #ifdef CONFIG_SMP |
1571 | ||
1572 | static int find_later_rq(struct task_struct *task); | |
1baca4ce JL |
1573 | |
1574 | static int | |
1575 | select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags) | |
1576 | { | |
1577 | struct task_struct *curr; | |
1578 | struct rq *rq; | |
1579 | ||
1d7e974c | 1580 | if (sd_flag != SD_BALANCE_WAKE) |
1baca4ce JL |
1581 | goto out; |
1582 | ||
1583 | rq = cpu_rq(cpu); | |
1584 | ||
1585 | rcu_read_lock(); | |
316c1608 | 1586 | curr = READ_ONCE(rq->curr); /* unlocked access */ |
1baca4ce JL |
1587 | |
1588 | /* | |
1589 | * If we are dealing with a -deadline task, we must | |
1590 | * decide where to wake it up. | |
1591 | * If it has a later deadline and the current task | |
1592 | * on this rq can't move (provided the waking task | |
1593 | * can!) we prefer to send it somewhere else. On the | |
1594 | * other hand, if it has a shorter deadline, we | |
1595 | * try to make it stay here, it might be important. | |
1596 | */ | |
1597 | if (unlikely(dl_task(curr)) && | |
4b53a341 | 1598 | (curr->nr_cpus_allowed < 2 || |
1baca4ce | 1599 | !dl_entity_preempt(&p->dl, &curr->dl)) && |
4b53a341 | 1600 | (p->nr_cpus_allowed > 1)) { |
1baca4ce JL |
1601 | int target = find_later_rq(p); |
1602 | ||
9d514262 | 1603 | if (target != -1 && |
5aa50507 LA |
1604 | (dl_time_before(p->dl.deadline, |
1605 | cpu_rq(target)->dl.earliest_dl.curr) || | |
1606 | (cpu_rq(target)->dl.dl_nr_running == 0))) | |
1baca4ce JL |
1607 | cpu = target; |
1608 | } | |
1609 | rcu_read_unlock(); | |
1610 | ||
1611 | out: | |
1612 | return cpu; | |
1613 | } | |
1614 | ||
209a0cbd LA |
1615 | static void migrate_task_rq_dl(struct task_struct *p) |
1616 | { | |
1617 | struct rq *rq; | |
1618 | ||
8fd27231 | 1619 | if (p->state != TASK_WAKING) |
209a0cbd LA |
1620 | return; |
1621 | ||
1622 | rq = task_rq(p); | |
1623 | /* | |
1624 | * Since p->state == TASK_WAKING, set_task_cpu() has been called | |
1625 | * from try_to_wake_up(). Hence, p->pi_lock is locked, but | |
1626 | * rq->lock is not... So, lock it | |
1627 | */ | |
1628 | raw_spin_lock(&rq->lock); | |
8fd27231 | 1629 | if (p->dl.dl_non_contending) { |
794a56eb | 1630 | sub_running_bw(&p->dl, &rq->dl); |
8fd27231 LA |
1631 | p->dl.dl_non_contending = 0; |
1632 | /* | |
1633 | * If the timer handler is currently running and the | |
1634 | * timer cannot be cancelled, inactive_task_timer() | |
1635 | * will see that dl_not_contending is not set, and | |
1636 | * will not touch the rq's active utilization, | |
1637 | * so we are still safe. | |
1638 | */ | |
1639 | if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) | |
1640 | put_task_struct(p); | |
1641 | } | |
794a56eb | 1642 | sub_rq_bw(&p->dl, &rq->dl); |
209a0cbd LA |
1643 | raw_spin_unlock(&rq->lock); |
1644 | } | |
1645 | ||
1baca4ce JL |
1646 | static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p) |
1647 | { | |
1648 | /* | |
1649 | * Current can't be migrated, useless to reschedule, | |
1650 | * let's hope p can move out. | |
1651 | */ | |
4b53a341 | 1652 | if (rq->curr->nr_cpus_allowed == 1 || |
3261ed0b | 1653 | !cpudl_find(&rq->rd->cpudl, rq->curr, NULL)) |
1baca4ce JL |
1654 | return; |
1655 | ||
1656 | /* | |
1657 | * p is migratable, so let's not schedule it and | |
1658 | * see if it is pushed or pulled somewhere else. | |
1659 | */ | |
4b53a341 | 1660 | if (p->nr_cpus_allowed != 1 && |
3261ed0b | 1661 | cpudl_find(&rq->rd->cpudl, p, NULL)) |
1baca4ce JL |
1662 | return; |
1663 | ||
8875125e | 1664 | resched_curr(rq); |
1baca4ce JL |
1665 | } |
1666 | ||
1667 | #endif /* CONFIG_SMP */ | |
1668 | ||
aab03e05 DF |
1669 | /* |
1670 | * Only called when both the current and waking task are -deadline | |
1671 | * tasks. | |
1672 | */ | |
1673 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | |
1674 | int flags) | |
1675 | { | |
1baca4ce | 1676 | if (dl_entity_preempt(&p->dl, &rq->curr->dl)) { |
8875125e | 1677 | resched_curr(rq); |
1baca4ce JL |
1678 | return; |
1679 | } | |
1680 | ||
1681 | #ifdef CONFIG_SMP | |
1682 | /* | |
1683 | * In the unlikely case current and p have the same deadline | |
1684 | * let us try to decide what's the best thing to do... | |
1685 | */ | |
332ac17e DF |
1686 | if ((p->dl.deadline == rq->curr->dl.deadline) && |
1687 | !test_tsk_need_resched(rq->curr)) | |
1baca4ce JL |
1688 | check_preempt_equal_dl(rq, p); |
1689 | #endif /* CONFIG_SMP */ | |
aab03e05 DF |
1690 | } |
1691 | ||
1692 | #ifdef CONFIG_SCHED_HRTICK | |
1693 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
1694 | { | |
177ef2a6 | 1695 | hrtick_start(rq, p->dl.runtime); |
aab03e05 | 1696 | } |
36ce9881 WL |
1697 | #else /* !CONFIG_SCHED_HRTICK */ |
1698 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | |
1699 | { | |
1700 | } | |
aab03e05 DF |
1701 | #endif |
1702 | ||
1703 | static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq, | |
1704 | struct dl_rq *dl_rq) | |
1705 | { | |
2161573e | 1706 | struct rb_node *left = rb_first_cached(&dl_rq->root); |
aab03e05 DF |
1707 | |
1708 | if (!left) | |
1709 | return NULL; | |
1710 | ||
1711 | return rb_entry(left, struct sched_dl_entity, rb_node); | |
1712 | } | |
1713 | ||
181a80d1 | 1714 | static struct task_struct * |
d8ac8971 | 1715 | pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
aab03e05 DF |
1716 | { |
1717 | struct sched_dl_entity *dl_se; | |
1718 | struct task_struct *p; | |
1719 | struct dl_rq *dl_rq; | |
1720 | ||
1721 | dl_rq = &rq->dl; | |
1722 | ||
a1d9a323 | 1723 | if (need_pull_dl_task(rq, prev)) { |
cbce1a68 PZ |
1724 | /* |
1725 | * This is OK, because current is on_cpu, which avoids it being | |
1726 | * picked for load-balance and preemption/IRQs are still | |
1727 | * disabled avoiding further scheduler activity on it and we're | |
1728 | * being very careful to re-start the picking loop. | |
1729 | */ | |
d8ac8971 | 1730 | rq_unpin_lock(rq, rf); |
38033c37 | 1731 | pull_dl_task(rq); |
d8ac8971 | 1732 | rq_repin_lock(rq, rf); |
a1d9a323 | 1733 | /* |
176cedc4 | 1734 | * pull_dl_task() can drop (and re-acquire) rq->lock; this |
a1d9a323 KT |
1735 | * means a stop task can slip in, in which case we need to |
1736 | * re-start task selection. | |
1737 | */ | |
da0c1e65 | 1738 | if (rq->stop && task_on_rq_queued(rq->stop)) |
a1d9a323 KT |
1739 | return RETRY_TASK; |
1740 | } | |
1741 | ||
734ff2a7 KT |
1742 | /* |
1743 | * When prev is DL, we may throttle it in put_prev_task(). | |
1744 | * So, we update time before we check for dl_nr_running. | |
1745 | */ | |
1746 | if (prev->sched_class == &dl_sched_class) | |
1747 | update_curr_dl(rq); | |
38033c37 | 1748 | |
aab03e05 DF |
1749 | if (unlikely(!dl_rq->dl_nr_running)) |
1750 | return NULL; | |
1751 | ||
3f1d2a31 | 1752 | put_prev_task(rq, prev); |
606dba2e | 1753 | |
aab03e05 DF |
1754 | dl_se = pick_next_dl_entity(rq, dl_rq); |
1755 | BUG_ON(!dl_se); | |
1756 | ||
1757 | p = dl_task_of(dl_se); | |
1758 | p->se.exec_start = rq_clock_task(rq); | |
1baca4ce JL |
1759 | |
1760 | /* Running task will never be pushed. */ | |
71362650 | 1761 | dequeue_pushable_dl_task(rq, p); |
1baca4ce | 1762 | |
aab03e05 DF |
1763 | if (hrtick_enabled(rq)) |
1764 | start_hrtick_dl(rq, p); | |
1baca4ce | 1765 | |
e3fca9e7 | 1766 | queue_push_tasks(rq); |
1baca4ce | 1767 | |
aab03e05 DF |
1768 | return p; |
1769 | } | |
1770 | ||
1771 | static void put_prev_task_dl(struct rq *rq, struct task_struct *p) | |
1772 | { | |
1773 | update_curr_dl(rq); | |
1baca4ce | 1774 | |
4b53a341 | 1775 | if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1) |
1baca4ce | 1776 | enqueue_pushable_dl_task(rq, p); |
aab03e05 DF |
1777 | } |
1778 | ||
1779 | static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued) | |
1780 | { | |
1781 | update_curr_dl(rq); | |
1782 | ||
a7bebf48 WL |
1783 | /* |
1784 | * Even when we have runtime, update_curr_dl() might have resulted in us | |
1785 | * not being the leftmost task anymore. In that case NEED_RESCHED will | |
1786 | * be set and schedule() will start a new hrtick for the next task. | |
1787 | */ | |
1788 | if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 && | |
1789 | is_leftmost(p, &rq->dl)) | |
aab03e05 | 1790 | start_hrtick_dl(rq, p); |
aab03e05 DF |
1791 | } |
1792 | ||
1793 | static void task_fork_dl(struct task_struct *p) | |
1794 | { | |
1795 | /* | |
1796 | * SCHED_DEADLINE tasks cannot fork and this is achieved through | |
1797 | * sched_fork() | |
1798 | */ | |
1799 | } | |
1800 | ||
aab03e05 DF |
1801 | static void set_curr_task_dl(struct rq *rq) |
1802 | { | |
1803 | struct task_struct *p = rq->curr; | |
1804 | ||
1805 | p->se.exec_start = rq_clock_task(rq); | |
1baca4ce JL |
1806 | |
1807 | /* You can't push away the running task */ | |
1808 | dequeue_pushable_dl_task(rq, p); | |
1809 | } | |
1810 | ||
1811 | #ifdef CONFIG_SMP | |
1812 | ||
1813 | /* Only try algorithms three times */ | |
1814 | #define DL_MAX_TRIES 3 | |
1815 | ||
1816 | static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu) | |
1817 | { | |
1818 | if (!task_running(rq, p) && | |
0c98d344 | 1819 | cpumask_test_cpu(cpu, &p->cpus_allowed)) |
1baca4ce | 1820 | return 1; |
1baca4ce JL |
1821 | return 0; |
1822 | } | |
1823 | ||
8b5e770e WL |
1824 | /* |
1825 | * Return the earliest pushable rq's task, which is suitable to be executed | |
1826 | * on the CPU, NULL otherwise: | |
1827 | */ | |
1828 | static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu) | |
1829 | { | |
2161573e | 1830 | struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost; |
8b5e770e WL |
1831 | struct task_struct *p = NULL; |
1832 | ||
1833 | if (!has_pushable_dl_tasks(rq)) | |
1834 | return NULL; | |
1835 | ||
1836 | next_node: | |
1837 | if (next_node) { | |
1838 | p = rb_entry(next_node, struct task_struct, pushable_dl_tasks); | |
1839 | ||
1840 | if (pick_dl_task(rq, p, cpu)) | |
1841 | return p; | |
1842 | ||
1843 | next_node = rb_next(next_node); | |
1844 | goto next_node; | |
1845 | } | |
1846 | ||
1847 | return NULL; | |
1848 | } | |
1849 | ||
1baca4ce JL |
1850 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl); |
1851 | ||
1852 | static int find_later_rq(struct task_struct *task) | |
1853 | { | |
1854 | struct sched_domain *sd; | |
4ba29684 | 1855 | struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl); |
1baca4ce | 1856 | int this_cpu = smp_processor_id(); |
b18c3ca1 | 1857 | int cpu = task_cpu(task); |
1baca4ce JL |
1858 | |
1859 | /* Make sure the mask is initialized first */ | |
1860 | if (unlikely(!later_mask)) | |
1861 | return -1; | |
1862 | ||
4b53a341 | 1863 | if (task->nr_cpus_allowed == 1) |
1baca4ce JL |
1864 | return -1; |
1865 | ||
91ec6778 JL |
1866 | /* |
1867 | * We have to consider system topology and task affinity | |
1868 | * first, then we can look for a suitable cpu. | |
1869 | */ | |
3261ed0b | 1870 | if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask)) |
1baca4ce JL |
1871 | return -1; |
1872 | ||
1873 | /* | |
b18c3ca1 BP |
1874 | * If we are here, some targets have been found, including |
1875 | * the most suitable which is, among the runqueues where the | |
1876 | * current tasks have later deadlines than the task's one, the | |
1877 | * rq with the latest possible one. | |
1baca4ce JL |
1878 | * |
1879 | * Now we check how well this matches with task's | |
1880 | * affinity and system topology. | |
1881 | * | |
1882 | * The last cpu where the task run is our first | |
1883 | * guess, since it is most likely cache-hot there. | |
1884 | */ | |
1885 | if (cpumask_test_cpu(cpu, later_mask)) | |
1886 | return cpu; | |
1887 | /* | |
1888 | * Check if this_cpu is to be skipped (i.e., it is | |
1889 | * not in the mask) or not. | |
1890 | */ | |
1891 | if (!cpumask_test_cpu(this_cpu, later_mask)) | |
1892 | this_cpu = -1; | |
1893 | ||
1894 | rcu_read_lock(); | |
1895 | for_each_domain(cpu, sd) { | |
1896 | if (sd->flags & SD_WAKE_AFFINE) { | |
b18c3ca1 | 1897 | int best_cpu; |
1baca4ce JL |
1898 | |
1899 | /* | |
1900 | * If possible, preempting this_cpu is | |
1901 | * cheaper than migrating. | |
1902 | */ | |
1903 | if (this_cpu != -1 && | |
1904 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { | |
1905 | rcu_read_unlock(); | |
1906 | return this_cpu; | |
1907 | } | |
1908 | ||
b18c3ca1 BP |
1909 | best_cpu = cpumask_first_and(later_mask, |
1910 | sched_domain_span(sd)); | |
1baca4ce | 1911 | /* |
b18c3ca1 BP |
1912 | * Last chance: if a cpu being in both later_mask |
1913 | * and current sd span is valid, that becomes our | |
1914 | * choice. Of course, the latest possible cpu is | |
1915 | * already under consideration through later_mask. | |
1baca4ce | 1916 | */ |
b18c3ca1 | 1917 | if (best_cpu < nr_cpu_ids) { |
1baca4ce JL |
1918 | rcu_read_unlock(); |
1919 | return best_cpu; | |
1920 | } | |
1921 | } | |
1922 | } | |
1923 | rcu_read_unlock(); | |
1924 | ||
1925 | /* | |
1926 | * At this point, all our guesses failed, we just return | |
1927 | * 'something', and let the caller sort the things out. | |
1928 | */ | |
1929 | if (this_cpu != -1) | |
1930 | return this_cpu; | |
1931 | ||
1932 | cpu = cpumask_any(later_mask); | |
1933 | if (cpu < nr_cpu_ids) | |
1934 | return cpu; | |
1935 | ||
1936 | return -1; | |
1937 | } | |
1938 | ||
1939 | /* Locks the rq it finds */ | |
1940 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq) | |
1941 | { | |
1942 | struct rq *later_rq = NULL; | |
1943 | int tries; | |
1944 | int cpu; | |
1945 | ||
1946 | for (tries = 0; tries < DL_MAX_TRIES; tries++) { | |
1947 | cpu = find_later_rq(task); | |
1948 | ||
1949 | if ((cpu == -1) || (cpu == rq->cpu)) | |
1950 | break; | |
1951 | ||
1952 | later_rq = cpu_rq(cpu); | |
1953 | ||
5aa50507 LA |
1954 | if (later_rq->dl.dl_nr_running && |
1955 | !dl_time_before(task->dl.deadline, | |
9d514262 WL |
1956 | later_rq->dl.earliest_dl.curr)) { |
1957 | /* | |
1958 | * Target rq has tasks of equal or earlier deadline, | |
1959 | * retrying does not release any lock and is unlikely | |
1960 | * to yield a different result. | |
1961 | */ | |
1962 | later_rq = NULL; | |
1963 | break; | |
1964 | } | |
1965 | ||
1baca4ce JL |
1966 | /* Retry if something changed. */ |
1967 | if (double_lock_balance(rq, later_rq)) { | |
1968 | if (unlikely(task_rq(task) != rq || | |
0c98d344 | 1969 | !cpumask_test_cpu(later_rq->cpu, &task->cpus_allowed) || |
da0c1e65 | 1970 | task_running(rq, task) || |
13b5ab02 | 1971 | !dl_task(task) || |
da0c1e65 | 1972 | !task_on_rq_queued(task))) { |
1baca4ce JL |
1973 | double_unlock_balance(rq, later_rq); |
1974 | later_rq = NULL; | |
1975 | break; | |
1976 | } | |
1977 | } | |
1978 | ||
1979 | /* | |
1980 | * If the rq we found has no -deadline task, or | |
1981 | * its earliest one has a later deadline than our | |
1982 | * task, the rq is a good one. | |
1983 | */ | |
1984 | if (!later_rq->dl.dl_nr_running || | |
1985 | dl_time_before(task->dl.deadline, | |
1986 | later_rq->dl.earliest_dl.curr)) | |
1987 | break; | |
1988 | ||
1989 | /* Otherwise we try again. */ | |
1990 | double_unlock_balance(rq, later_rq); | |
1991 | later_rq = NULL; | |
1992 | } | |
1993 | ||
1994 | return later_rq; | |
1995 | } | |
1996 | ||
1997 | static struct task_struct *pick_next_pushable_dl_task(struct rq *rq) | |
1998 | { | |
1999 | struct task_struct *p; | |
2000 | ||
2001 | if (!has_pushable_dl_tasks(rq)) | |
2002 | return NULL; | |
2003 | ||
2161573e | 2004 | p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost, |
1baca4ce JL |
2005 | struct task_struct, pushable_dl_tasks); |
2006 | ||
2007 | BUG_ON(rq->cpu != task_cpu(p)); | |
2008 | BUG_ON(task_current(rq, p)); | |
4b53a341 | 2009 | BUG_ON(p->nr_cpus_allowed <= 1); |
1baca4ce | 2010 | |
da0c1e65 | 2011 | BUG_ON(!task_on_rq_queued(p)); |
1baca4ce JL |
2012 | BUG_ON(!dl_task(p)); |
2013 | ||
2014 | return p; | |
2015 | } | |
2016 | ||
2017 | /* | |
2018 | * See if the non running -deadline tasks on this rq | |
2019 | * can be sent to some other CPU where they can preempt | |
2020 | * and start executing. | |
2021 | */ | |
2022 | static int push_dl_task(struct rq *rq) | |
2023 | { | |
2024 | struct task_struct *next_task; | |
2025 | struct rq *later_rq; | |
c51b8ab5 | 2026 | int ret = 0; |
1baca4ce JL |
2027 | |
2028 | if (!rq->dl.overloaded) | |
2029 | return 0; | |
2030 | ||
2031 | next_task = pick_next_pushable_dl_task(rq); | |
2032 | if (!next_task) | |
2033 | return 0; | |
2034 | ||
2035 | retry: | |
2036 | if (unlikely(next_task == rq->curr)) { | |
2037 | WARN_ON(1); | |
2038 | return 0; | |
2039 | } | |
2040 | ||
2041 | /* | |
2042 | * If next_task preempts rq->curr, and rq->curr | |
2043 | * can move away, it makes sense to just reschedule | |
2044 | * without going further in pushing next_task. | |
2045 | */ | |
2046 | if (dl_task(rq->curr) && | |
2047 | dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) && | |
4b53a341 | 2048 | rq->curr->nr_cpus_allowed > 1) { |
8875125e | 2049 | resched_curr(rq); |
1baca4ce JL |
2050 | return 0; |
2051 | } | |
2052 | ||
2053 | /* We might release rq lock */ | |
2054 | get_task_struct(next_task); | |
2055 | ||
2056 | /* Will lock the rq it'll find */ | |
2057 | later_rq = find_lock_later_rq(next_task, rq); | |
2058 | if (!later_rq) { | |
2059 | struct task_struct *task; | |
2060 | ||
2061 | /* | |
2062 | * We must check all this again, since | |
2063 | * find_lock_later_rq releases rq->lock and it is | |
2064 | * then possible that next_task has migrated. | |
2065 | */ | |
2066 | task = pick_next_pushable_dl_task(rq); | |
a776b968 | 2067 | if (task == next_task) { |
1baca4ce JL |
2068 | /* |
2069 | * The task is still there. We don't try | |
2070 | * again, some other cpu will pull it when ready. | |
2071 | */ | |
1baca4ce JL |
2072 | goto out; |
2073 | } | |
2074 | ||
2075 | if (!task) | |
2076 | /* No more tasks */ | |
2077 | goto out; | |
2078 | ||
2079 | put_task_struct(next_task); | |
2080 | next_task = task; | |
2081 | goto retry; | |
2082 | } | |
2083 | ||
2084 | deactivate_task(rq, next_task, 0); | |
794a56eb JL |
2085 | sub_running_bw(&next_task->dl, &rq->dl); |
2086 | sub_rq_bw(&next_task->dl, &rq->dl); | |
1baca4ce | 2087 | set_task_cpu(next_task, later_rq->cpu); |
794a56eb JL |
2088 | add_rq_bw(&next_task->dl, &later_rq->dl); |
2089 | add_running_bw(&next_task->dl, &later_rq->dl); | |
1baca4ce | 2090 | activate_task(later_rq, next_task, 0); |
c51b8ab5 | 2091 | ret = 1; |
1baca4ce | 2092 | |
8875125e | 2093 | resched_curr(later_rq); |
1baca4ce JL |
2094 | |
2095 | double_unlock_balance(rq, later_rq); | |
2096 | ||
2097 | out: | |
2098 | put_task_struct(next_task); | |
2099 | ||
c51b8ab5 | 2100 | return ret; |
1baca4ce JL |
2101 | } |
2102 | ||
2103 | static void push_dl_tasks(struct rq *rq) | |
2104 | { | |
4ffa08ed | 2105 | /* push_dl_task() will return true if it moved a -deadline task */ |
1baca4ce JL |
2106 | while (push_dl_task(rq)) |
2107 | ; | |
aab03e05 DF |
2108 | } |
2109 | ||
0ea60c20 | 2110 | static void pull_dl_task(struct rq *this_rq) |
1baca4ce | 2111 | { |
0ea60c20 | 2112 | int this_cpu = this_rq->cpu, cpu; |
1baca4ce | 2113 | struct task_struct *p; |
0ea60c20 | 2114 | bool resched = false; |
1baca4ce JL |
2115 | struct rq *src_rq; |
2116 | u64 dmin = LONG_MAX; | |
2117 | ||
2118 | if (likely(!dl_overloaded(this_rq))) | |
0ea60c20 | 2119 | return; |
1baca4ce JL |
2120 | |
2121 | /* | |
2122 | * Match the barrier from dl_set_overloaded; this guarantees that if we | |
2123 | * see overloaded we must also see the dlo_mask bit. | |
2124 | */ | |
2125 | smp_rmb(); | |
2126 | ||
2127 | for_each_cpu(cpu, this_rq->rd->dlo_mask) { | |
2128 | if (this_cpu == cpu) | |
2129 | continue; | |
2130 | ||
2131 | src_rq = cpu_rq(cpu); | |
2132 | ||
2133 | /* | |
2134 | * It looks racy, abd it is! However, as in sched_rt.c, | |
2135 | * we are fine with this. | |
2136 | */ | |
2137 | if (this_rq->dl.dl_nr_running && | |
2138 | dl_time_before(this_rq->dl.earliest_dl.curr, | |
2139 | src_rq->dl.earliest_dl.next)) | |
2140 | continue; | |
2141 | ||
2142 | /* Might drop this_rq->lock */ | |
2143 | double_lock_balance(this_rq, src_rq); | |
2144 | ||
2145 | /* | |
2146 | * If there are no more pullable tasks on the | |
2147 | * rq, we're done with it. | |
2148 | */ | |
2149 | if (src_rq->dl.dl_nr_running <= 1) | |
2150 | goto skip; | |
2151 | ||
8b5e770e | 2152 | p = pick_earliest_pushable_dl_task(src_rq, this_cpu); |
1baca4ce JL |
2153 | |
2154 | /* | |
2155 | * We found a task to be pulled if: | |
2156 | * - it preempts our current (if there's one), | |
2157 | * - it will preempt the last one we pulled (if any). | |
2158 | */ | |
2159 | if (p && dl_time_before(p->dl.deadline, dmin) && | |
2160 | (!this_rq->dl.dl_nr_running || | |
2161 | dl_time_before(p->dl.deadline, | |
2162 | this_rq->dl.earliest_dl.curr))) { | |
2163 | WARN_ON(p == src_rq->curr); | |
da0c1e65 | 2164 | WARN_ON(!task_on_rq_queued(p)); |
1baca4ce JL |
2165 | |
2166 | /* | |
2167 | * Then we pull iff p has actually an earlier | |
2168 | * deadline than the current task of its runqueue. | |
2169 | */ | |
2170 | if (dl_time_before(p->dl.deadline, | |
2171 | src_rq->curr->dl.deadline)) | |
2172 | goto skip; | |
2173 | ||
0ea60c20 | 2174 | resched = true; |
1baca4ce JL |
2175 | |
2176 | deactivate_task(src_rq, p, 0); | |
794a56eb JL |
2177 | sub_running_bw(&p->dl, &src_rq->dl); |
2178 | sub_rq_bw(&p->dl, &src_rq->dl); | |
1baca4ce | 2179 | set_task_cpu(p, this_cpu); |
794a56eb JL |
2180 | add_rq_bw(&p->dl, &this_rq->dl); |
2181 | add_running_bw(&p->dl, &this_rq->dl); | |
1baca4ce JL |
2182 | activate_task(this_rq, p, 0); |
2183 | dmin = p->dl.deadline; | |
2184 | ||
2185 | /* Is there any other task even earlier? */ | |
2186 | } | |
2187 | skip: | |
2188 | double_unlock_balance(this_rq, src_rq); | |
2189 | } | |
2190 | ||
0ea60c20 PZ |
2191 | if (resched) |
2192 | resched_curr(this_rq); | |
1baca4ce JL |
2193 | } |
2194 | ||
2195 | /* | |
2196 | * Since the task is not running and a reschedule is not going to happen | |
2197 | * anytime soon on its runqueue, we try pushing it away now. | |
2198 | */ | |
2199 | static void task_woken_dl(struct rq *rq, struct task_struct *p) | |
2200 | { | |
2201 | if (!task_running(rq, p) && | |
2202 | !test_tsk_need_resched(rq->curr) && | |
4b53a341 | 2203 | p->nr_cpus_allowed > 1 && |
1baca4ce | 2204 | dl_task(rq->curr) && |
4b53a341 | 2205 | (rq->curr->nr_cpus_allowed < 2 || |
6b0a563f | 2206 | !dl_entity_preempt(&p->dl, &rq->curr->dl))) { |
1baca4ce JL |
2207 | push_dl_tasks(rq); |
2208 | } | |
2209 | } | |
2210 | ||
2211 | static void set_cpus_allowed_dl(struct task_struct *p, | |
2212 | const struct cpumask *new_mask) | |
2213 | { | |
7f51412a | 2214 | struct root_domain *src_rd; |
6c37067e | 2215 | struct rq *rq; |
1baca4ce JL |
2216 | |
2217 | BUG_ON(!dl_task(p)); | |
2218 | ||
7f51412a JL |
2219 | rq = task_rq(p); |
2220 | src_rd = rq->rd; | |
2221 | /* | |
2222 | * Migrating a SCHED_DEADLINE task between exclusive | |
2223 | * cpusets (different root_domains) entails a bandwidth | |
2224 | * update. We already made space for us in the destination | |
2225 | * domain (see cpuset_can_attach()). | |
2226 | */ | |
2227 | if (!cpumask_intersects(src_rd->span, new_mask)) { | |
2228 | struct dl_bw *src_dl_b; | |
2229 | ||
2230 | src_dl_b = dl_bw_of(cpu_of(rq)); | |
2231 | /* | |
2232 | * We now free resources of the root_domain we are migrating | |
2233 | * off. In the worst case, sched_setattr() may temporary fail | |
2234 | * until we complete the update. | |
2235 | */ | |
2236 | raw_spin_lock(&src_dl_b->lock); | |
8c0944ce | 2237 | __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); |
7f51412a JL |
2238 | raw_spin_unlock(&src_dl_b->lock); |
2239 | } | |
2240 | ||
6c37067e | 2241 | set_cpus_allowed_common(p, new_mask); |
1baca4ce JL |
2242 | } |
2243 | ||
2244 | /* Assumes rq->lock is held */ | |
2245 | static void rq_online_dl(struct rq *rq) | |
2246 | { | |
2247 | if (rq->dl.overloaded) | |
2248 | dl_set_overload(rq); | |
6bfd6d72 | 2249 | |
16b26943 | 2250 | cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu); |
6bfd6d72 | 2251 | if (rq->dl.dl_nr_running > 0) |
d8206bb3 | 2252 | cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr); |
1baca4ce JL |
2253 | } |
2254 | ||
2255 | /* Assumes rq->lock is held */ | |
2256 | static void rq_offline_dl(struct rq *rq) | |
2257 | { | |
2258 | if (rq->dl.overloaded) | |
2259 | dl_clear_overload(rq); | |
6bfd6d72 | 2260 | |
d8206bb3 | 2261 | cpudl_clear(&rq->rd->cpudl, rq->cpu); |
16b26943 | 2262 | cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu); |
1baca4ce JL |
2263 | } |
2264 | ||
a6c0e746 | 2265 | void __init init_sched_dl_class(void) |
1baca4ce JL |
2266 | { |
2267 | unsigned int i; | |
2268 | ||
2269 | for_each_possible_cpu(i) | |
2270 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i), | |
2271 | GFP_KERNEL, cpu_to_node(i)); | |
2272 | } | |
2273 | ||
2274 | #endif /* CONFIG_SMP */ | |
2275 | ||
aab03e05 DF |
2276 | static void switched_from_dl(struct rq *rq, struct task_struct *p) |
2277 | { | |
a649f237 | 2278 | /* |
209a0cbd LA |
2279 | * task_non_contending() can start the "inactive timer" (if the 0-lag |
2280 | * time is in the future). If the task switches back to dl before | |
2281 | * the "inactive timer" fires, it can continue to consume its current | |
2282 | * runtime using its current deadline. If it stays outside of | |
2283 | * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer() | |
2284 | * will reset the task parameters. | |
a649f237 | 2285 | */ |
209a0cbd LA |
2286 | if (task_on_rq_queued(p) && p->dl.dl_runtime) |
2287 | task_non_contending(p); | |
2288 | ||
8fd27231 | 2289 | if (!task_on_rq_queued(p)) |
794a56eb | 2290 | sub_rq_bw(&p->dl, &rq->dl); |
8fd27231 | 2291 | |
209a0cbd LA |
2292 | /* |
2293 | * We cannot use inactive_task_timer() to invoke sub_running_bw() | |
2294 | * at the 0-lag time, because the task could have been migrated | |
2295 | * while SCHED_OTHER in the meanwhile. | |
2296 | */ | |
2297 | if (p->dl.dl_non_contending) | |
2298 | p->dl.dl_non_contending = 0; | |
a5e7be3b | 2299 | |
1baca4ce JL |
2300 | /* |
2301 | * Since this might be the only -deadline task on the rq, | |
2302 | * this is the right place to try to pull some other one | |
2303 | * from an overloaded cpu, if any. | |
2304 | */ | |
cd660911 WL |
2305 | if (!task_on_rq_queued(p) || rq->dl.dl_nr_running) |
2306 | return; | |
2307 | ||
9916e214 | 2308 | queue_pull_task(rq); |
aab03e05 DF |
2309 | } |
2310 | ||
1baca4ce JL |
2311 | /* |
2312 | * When switching to -deadline, we may overload the rq, then | |
2313 | * we try to push someone off, if possible. | |
2314 | */ | |
aab03e05 DF |
2315 | static void switched_to_dl(struct rq *rq, struct task_struct *p) |
2316 | { | |
209a0cbd LA |
2317 | if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) |
2318 | put_task_struct(p); | |
98b0a857 JL |
2319 | |
2320 | /* If p is not queued we will update its parameters at next wakeup. */ | |
8fd27231 | 2321 | if (!task_on_rq_queued(p)) { |
794a56eb | 2322 | add_rq_bw(&p->dl, &rq->dl); |
98b0a857 | 2323 | |
8fd27231 LA |
2324 | return; |
2325 | } | |
72f9f3fd | 2326 | |
98b0a857 | 2327 | if (rq->curr != p) { |
1baca4ce | 2328 | #ifdef CONFIG_SMP |
4b53a341 | 2329 | if (p->nr_cpus_allowed > 1 && rq->dl.overloaded) |
9916e214 | 2330 | queue_push_tasks(rq); |
619bd4a7 | 2331 | #endif |
9916e214 PZ |
2332 | if (dl_task(rq->curr)) |
2333 | check_preempt_curr_dl(rq, p, 0); | |
2334 | else | |
2335 | resched_curr(rq); | |
aab03e05 DF |
2336 | } |
2337 | } | |
2338 | ||
1baca4ce JL |
2339 | /* |
2340 | * If the scheduling parameters of a -deadline task changed, | |
2341 | * a push or pull operation might be needed. | |
2342 | */ | |
aab03e05 DF |
2343 | static void prio_changed_dl(struct rq *rq, struct task_struct *p, |
2344 | int oldprio) | |
2345 | { | |
da0c1e65 | 2346 | if (task_on_rq_queued(p) || rq->curr == p) { |
aab03e05 | 2347 | #ifdef CONFIG_SMP |
1baca4ce JL |
2348 | /* |
2349 | * This might be too much, but unfortunately | |
2350 | * we don't have the old deadline value, and | |
2351 | * we can't argue if the task is increasing | |
2352 | * or lowering its prio, so... | |
2353 | */ | |
2354 | if (!rq->dl.overloaded) | |
9916e214 | 2355 | queue_pull_task(rq); |
1baca4ce JL |
2356 | |
2357 | /* | |
2358 | * If we now have a earlier deadline task than p, | |
2359 | * then reschedule, provided p is still on this | |
2360 | * runqueue. | |
2361 | */ | |
9916e214 | 2362 | if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline)) |
8875125e | 2363 | resched_curr(rq); |
1baca4ce JL |
2364 | #else |
2365 | /* | |
2366 | * Again, we don't know if p has a earlier | |
2367 | * or later deadline, so let's blindly set a | |
2368 | * (maybe not needed) rescheduling point. | |
2369 | */ | |
8875125e | 2370 | resched_curr(rq); |
1baca4ce | 2371 | #endif /* CONFIG_SMP */ |
801ccdbf | 2372 | } |
aab03e05 | 2373 | } |
aab03e05 DF |
2374 | |
2375 | const struct sched_class dl_sched_class = { | |
2376 | .next = &rt_sched_class, | |
2377 | .enqueue_task = enqueue_task_dl, | |
2378 | .dequeue_task = dequeue_task_dl, | |
2379 | .yield_task = yield_task_dl, | |
2380 | ||
2381 | .check_preempt_curr = check_preempt_curr_dl, | |
2382 | ||
2383 | .pick_next_task = pick_next_task_dl, | |
2384 | .put_prev_task = put_prev_task_dl, | |
2385 | ||
2386 | #ifdef CONFIG_SMP | |
2387 | .select_task_rq = select_task_rq_dl, | |
209a0cbd | 2388 | .migrate_task_rq = migrate_task_rq_dl, |
1baca4ce JL |
2389 | .set_cpus_allowed = set_cpus_allowed_dl, |
2390 | .rq_online = rq_online_dl, | |
2391 | .rq_offline = rq_offline_dl, | |
1baca4ce | 2392 | .task_woken = task_woken_dl, |
aab03e05 DF |
2393 | #endif |
2394 | ||
2395 | .set_curr_task = set_curr_task_dl, | |
2396 | .task_tick = task_tick_dl, | |
2397 | .task_fork = task_fork_dl, | |
aab03e05 DF |
2398 | |
2399 | .prio_changed = prio_changed_dl, | |
2400 | .switched_from = switched_from_dl, | |
2401 | .switched_to = switched_to_dl, | |
6e998916 SG |
2402 | |
2403 | .update_curr = update_curr_dl, | |
aab03e05 | 2404 | }; |
acb32132 | 2405 | |
06a76fe0 NP |
2406 | int sched_dl_global_validate(void) |
2407 | { | |
2408 | u64 runtime = global_rt_runtime(); | |
2409 | u64 period = global_rt_period(); | |
2410 | u64 new_bw = to_ratio(period, runtime); | |
2411 | struct dl_bw *dl_b; | |
2412 | int cpu, ret = 0; | |
2413 | unsigned long flags; | |
2414 | ||
2415 | /* | |
2416 | * Here we want to check the bandwidth not being set to some | |
2417 | * value smaller than the currently allocated bandwidth in | |
2418 | * any of the root_domains. | |
2419 | * | |
2420 | * FIXME: Cycling on all the CPUs is overdoing, but simpler than | |
2421 | * cycling on root_domains... Discussion on different/better | |
2422 | * solutions is welcome! | |
2423 | */ | |
2424 | for_each_possible_cpu(cpu) { | |
2425 | rcu_read_lock_sched(); | |
2426 | dl_b = dl_bw_of(cpu); | |
2427 | ||
2428 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
2429 | if (new_bw < dl_b->total_bw) | |
2430 | ret = -EBUSY; | |
2431 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2432 | ||
2433 | rcu_read_unlock_sched(); | |
2434 | ||
2435 | if (ret) | |
2436 | break; | |
2437 | } | |
2438 | ||
2439 | return ret; | |
2440 | } | |
2441 | ||
2442 | void init_dl_rq_bw_ratio(struct dl_rq *dl_rq) | |
2443 | { | |
2444 | if (global_rt_runtime() == RUNTIME_INF) { | |
2445 | dl_rq->bw_ratio = 1 << RATIO_SHIFT; | |
2446 | dl_rq->extra_bw = 1 << BW_SHIFT; | |
2447 | } else { | |
2448 | dl_rq->bw_ratio = to_ratio(global_rt_runtime(), | |
2449 | global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT); | |
2450 | dl_rq->extra_bw = to_ratio(global_rt_period(), | |
2451 | global_rt_runtime()); | |
2452 | } | |
2453 | } | |
2454 | ||
2455 | void sched_dl_do_global(void) | |
2456 | { | |
2457 | u64 new_bw = -1; | |
2458 | struct dl_bw *dl_b; | |
2459 | int cpu; | |
2460 | unsigned long flags; | |
2461 | ||
2462 | def_dl_bandwidth.dl_period = global_rt_period(); | |
2463 | def_dl_bandwidth.dl_runtime = global_rt_runtime(); | |
2464 | ||
2465 | if (global_rt_runtime() != RUNTIME_INF) | |
2466 | new_bw = to_ratio(global_rt_period(), global_rt_runtime()); | |
2467 | ||
2468 | /* | |
2469 | * FIXME: As above... | |
2470 | */ | |
2471 | for_each_possible_cpu(cpu) { | |
2472 | rcu_read_lock_sched(); | |
2473 | dl_b = dl_bw_of(cpu); | |
2474 | ||
2475 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
2476 | dl_b->bw = new_bw; | |
2477 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2478 | ||
2479 | rcu_read_unlock_sched(); | |
2480 | init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl); | |
2481 | } | |
2482 | } | |
2483 | ||
2484 | /* | |
2485 | * We must be sure that accepting a new task (or allowing changing the | |
2486 | * parameters of an existing one) is consistent with the bandwidth | |
2487 | * constraints. If yes, this function also accordingly updates the currently | |
2488 | * allocated bandwidth to reflect the new situation. | |
2489 | * | |
2490 | * This function is called while holding p's rq->lock. | |
2491 | */ | |
2492 | int sched_dl_overflow(struct task_struct *p, int policy, | |
2493 | const struct sched_attr *attr) | |
2494 | { | |
2495 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); | |
2496 | u64 period = attr->sched_period ?: attr->sched_deadline; | |
2497 | u64 runtime = attr->sched_runtime; | |
2498 | u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; | |
2499 | int cpus, err = -1; | |
2500 | ||
794a56eb JL |
2501 | if (attr->sched_flags & SCHED_FLAG_SUGOV) |
2502 | return 0; | |
2503 | ||
06a76fe0 NP |
2504 | /* !deadline task may carry old deadline bandwidth */ |
2505 | if (new_bw == p->dl.dl_bw && task_has_dl_policy(p)) | |
2506 | return 0; | |
2507 | ||
2508 | /* | |
2509 | * Either if a task, enters, leave, or stays -deadline but changes | |
2510 | * its parameters, we may need to update accordingly the total | |
2511 | * allocated bandwidth of the container. | |
2512 | */ | |
2513 | raw_spin_lock(&dl_b->lock); | |
2514 | cpus = dl_bw_cpus(task_cpu(p)); | |
2515 | if (dl_policy(policy) && !task_has_dl_policy(p) && | |
2516 | !__dl_overflow(dl_b, cpus, 0, new_bw)) { | |
2517 | if (hrtimer_active(&p->dl.inactive_timer)) | |
8c0944ce | 2518 | __dl_sub(dl_b, p->dl.dl_bw, cpus); |
06a76fe0 NP |
2519 | __dl_add(dl_b, new_bw, cpus); |
2520 | err = 0; | |
2521 | } else if (dl_policy(policy) && task_has_dl_policy(p) && | |
2522 | !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { | |
2523 | /* | |
2524 | * XXX this is slightly incorrect: when the task | |
2525 | * utilization decreases, we should delay the total | |
2526 | * utilization change until the task's 0-lag point. | |
2527 | * But this would require to set the task's "inactive | |
2528 | * timer" when the task is not inactive. | |
2529 | */ | |
8c0944ce | 2530 | __dl_sub(dl_b, p->dl.dl_bw, cpus); |
06a76fe0 NP |
2531 | __dl_add(dl_b, new_bw, cpus); |
2532 | dl_change_utilization(p, new_bw); | |
2533 | err = 0; | |
2534 | } else if (!dl_policy(policy) && task_has_dl_policy(p)) { | |
2535 | /* | |
2536 | * Do not decrease the total deadline utilization here, | |
2537 | * switched_from_dl() will take care to do it at the correct | |
2538 | * (0-lag) time. | |
2539 | */ | |
2540 | err = 0; | |
2541 | } | |
2542 | raw_spin_unlock(&dl_b->lock); | |
2543 | ||
2544 | return err; | |
2545 | } | |
2546 | ||
2547 | /* | |
2548 | * This function initializes the sched_dl_entity of a newly becoming | |
2549 | * SCHED_DEADLINE task. | |
2550 | * | |
2551 | * Only the static values are considered here, the actual runtime and the | |
2552 | * absolute deadline will be properly calculated when the task is enqueued | |
2553 | * for the first time with its new policy. | |
2554 | */ | |
2555 | void __setparam_dl(struct task_struct *p, const struct sched_attr *attr) | |
2556 | { | |
2557 | struct sched_dl_entity *dl_se = &p->dl; | |
2558 | ||
2559 | dl_se->dl_runtime = attr->sched_runtime; | |
2560 | dl_se->dl_deadline = attr->sched_deadline; | |
2561 | dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; | |
2562 | dl_se->flags = attr->sched_flags; | |
2563 | dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); | |
2564 | dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime); | |
2565 | } | |
2566 | ||
2567 | void __getparam_dl(struct task_struct *p, struct sched_attr *attr) | |
2568 | { | |
2569 | struct sched_dl_entity *dl_se = &p->dl; | |
2570 | ||
2571 | attr->sched_priority = p->rt_priority; | |
2572 | attr->sched_runtime = dl_se->dl_runtime; | |
2573 | attr->sched_deadline = dl_se->dl_deadline; | |
2574 | attr->sched_period = dl_se->dl_period; | |
2575 | attr->sched_flags = dl_se->flags; | |
2576 | } | |
2577 | ||
2578 | /* | |
2579 | * This function validates the new parameters of a -deadline task. | |
2580 | * We ask for the deadline not being zero, and greater or equal | |
2581 | * than the runtime, as well as the period of being zero or | |
2582 | * greater than deadline. Furthermore, we have to be sure that | |
2583 | * user parameters are above the internal resolution of 1us (we | |
2584 | * check sched_runtime only since it is always the smaller one) and | |
2585 | * below 2^63 ns (we have to check both sched_deadline and | |
2586 | * sched_period, as the latter can be zero). | |
2587 | */ | |
2588 | bool __checkparam_dl(const struct sched_attr *attr) | |
2589 | { | |
794a56eb JL |
2590 | /* special dl tasks don't actually use any parameter */ |
2591 | if (attr->sched_flags & SCHED_FLAG_SUGOV) | |
2592 | return true; | |
2593 | ||
06a76fe0 NP |
2594 | /* deadline != 0 */ |
2595 | if (attr->sched_deadline == 0) | |
2596 | return false; | |
2597 | ||
2598 | /* | |
2599 | * Since we truncate DL_SCALE bits, make sure we're at least | |
2600 | * that big. | |
2601 | */ | |
2602 | if (attr->sched_runtime < (1ULL << DL_SCALE)) | |
2603 | return false; | |
2604 | ||
2605 | /* | |
2606 | * Since we use the MSB for wrap-around and sign issues, make | |
2607 | * sure it's not set (mind that period can be equal to zero). | |
2608 | */ | |
2609 | if (attr->sched_deadline & (1ULL << 63) || | |
2610 | attr->sched_period & (1ULL << 63)) | |
2611 | return false; | |
2612 | ||
2613 | /* runtime <= deadline <= period (if period != 0) */ | |
2614 | if ((attr->sched_period != 0 && | |
2615 | attr->sched_period < attr->sched_deadline) || | |
2616 | attr->sched_deadline < attr->sched_runtime) | |
2617 | return false; | |
2618 | ||
2619 | return true; | |
2620 | } | |
2621 | ||
2622 | /* | |
2623 | * This function clears the sched_dl_entity static params. | |
2624 | */ | |
2625 | void __dl_clear_params(struct task_struct *p) | |
2626 | { | |
2627 | struct sched_dl_entity *dl_se = &p->dl; | |
2628 | ||
2629 | dl_se->dl_runtime = 0; | |
2630 | dl_se->dl_deadline = 0; | |
2631 | dl_se->dl_period = 0; | |
2632 | dl_se->flags = 0; | |
2633 | dl_se->dl_bw = 0; | |
2634 | dl_se->dl_density = 0; | |
2635 | ||
2636 | dl_se->dl_throttled = 0; | |
2637 | dl_se->dl_yielded = 0; | |
2638 | dl_se->dl_non_contending = 0; | |
34be3930 | 2639 | dl_se->dl_overrun = 0; |
06a76fe0 NP |
2640 | } |
2641 | ||
2642 | bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr) | |
2643 | { | |
2644 | struct sched_dl_entity *dl_se = &p->dl; | |
2645 | ||
2646 | if (dl_se->dl_runtime != attr->sched_runtime || | |
2647 | dl_se->dl_deadline != attr->sched_deadline || | |
2648 | dl_se->dl_period != attr->sched_period || | |
2649 | dl_se->flags != attr->sched_flags) | |
2650 | return true; | |
2651 | ||
2652 | return false; | |
2653 | } | |
2654 | ||
2655 | #ifdef CONFIG_SMP | |
2656 | int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed) | |
2657 | { | |
2658 | unsigned int dest_cpu = cpumask_any_and(cpu_active_mask, | |
2659 | cs_cpus_allowed); | |
2660 | struct dl_bw *dl_b; | |
2661 | bool overflow; | |
2662 | int cpus, ret; | |
2663 | unsigned long flags; | |
2664 | ||
2665 | rcu_read_lock_sched(); | |
2666 | dl_b = dl_bw_of(dest_cpu); | |
2667 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
2668 | cpus = dl_bw_cpus(dest_cpu); | |
2669 | overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw); | |
2670 | if (overflow) | |
2671 | ret = -EBUSY; | |
2672 | else { | |
2673 | /* | |
2674 | * We reserve space for this task in the destination | |
2675 | * root_domain, as we can't fail after this point. | |
2676 | * We will free resources in the source root_domain | |
2677 | * later on (see set_cpus_allowed_dl()). | |
2678 | */ | |
2679 | __dl_add(dl_b, p->dl.dl_bw, cpus); | |
2680 | ret = 0; | |
2681 | } | |
2682 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2683 | rcu_read_unlock_sched(); | |
2684 | return ret; | |
2685 | } | |
2686 | ||
2687 | int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, | |
2688 | const struct cpumask *trial) | |
2689 | { | |
2690 | int ret = 1, trial_cpus; | |
2691 | struct dl_bw *cur_dl_b; | |
2692 | unsigned long flags; | |
2693 | ||
2694 | rcu_read_lock_sched(); | |
2695 | cur_dl_b = dl_bw_of(cpumask_any(cur)); | |
2696 | trial_cpus = cpumask_weight(trial); | |
2697 | ||
2698 | raw_spin_lock_irqsave(&cur_dl_b->lock, flags); | |
2699 | if (cur_dl_b->bw != -1 && | |
2700 | cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw) | |
2701 | ret = 0; | |
2702 | raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags); | |
2703 | rcu_read_unlock_sched(); | |
2704 | return ret; | |
2705 | } | |
2706 | ||
2707 | bool dl_cpu_busy(unsigned int cpu) | |
2708 | { | |
2709 | unsigned long flags; | |
2710 | struct dl_bw *dl_b; | |
2711 | bool overflow; | |
2712 | int cpus; | |
2713 | ||
2714 | rcu_read_lock_sched(); | |
2715 | dl_b = dl_bw_of(cpu); | |
2716 | raw_spin_lock_irqsave(&dl_b->lock, flags); | |
2717 | cpus = dl_bw_cpus(cpu); | |
2718 | overflow = __dl_overflow(dl_b, cpus, 0, 0); | |
2719 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | |
2720 | rcu_read_unlock_sched(); | |
2721 | return overflow; | |
2722 | } | |
2723 | #endif | |
2724 | ||
acb32132 WL |
2725 | #ifdef CONFIG_SCHED_DEBUG |
2726 | extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq); | |
2727 | ||
2728 | void print_dl_stats(struct seq_file *m, int cpu) | |
2729 | { | |
2730 | print_dl_rq(m, cpu, &cpu_rq(cpu)->dl); | |
2731 | } | |
2732 | #endif /* CONFIG_SCHED_DEBUG */ |