]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * kernel/sched/core.c | |
3 | * | |
4 | * Core kernel scheduler code and related syscalls | |
5 | * | |
6 | * Copyright (C) 1991-2002 Linus Torvalds | |
7 | */ | |
8 | #include <linux/sched.h> | |
9 | #include <linux/sched/clock.h> | |
10 | #include <uapi/linux/sched/types.h> | |
11 | #include <linux/sched/loadavg.h> | |
12 | #include <linux/sched/hotplug.h> | |
13 | #include <linux/wait_bit.h> | |
14 | #include <linux/cpuset.h> | |
15 | #include <linux/delayacct.h> | |
16 | #include <linux/init_task.h> | |
17 | #include <linux/context_tracking.h> | |
18 | #include <linux/rcupdate_wait.h> | |
19 | ||
20 | #include <linux/blkdev.h> | |
21 | #include <linux/kprobes.h> | |
22 | #include <linux/mmu_context.h> | |
23 | #include <linux/module.h> | |
24 | #include <linux/nmi.h> | |
25 | #include <linux/prefetch.h> | |
26 | #include <linux/profile.h> | |
27 | #include <linux/security.h> | |
28 | #include <linux/syscalls.h> | |
29 | ||
30 | #include <asm/switch_to.h> | |
31 | #include <asm/tlb.h> | |
32 | #ifdef CONFIG_PARAVIRT | |
33 | #include <asm/paravirt.h> | |
34 | #endif | |
35 | ||
36 | #include "sched.h" | |
37 | #include "../workqueue_internal.h" | |
38 | #include "../smpboot.h" | |
39 | ||
40 | #define CREATE_TRACE_POINTS | |
41 | #include <trace/events/sched.h> | |
42 | ||
43 | DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); | |
44 | ||
45 | /* | |
46 | * Debugging: various feature bits | |
47 | */ | |
48 | ||
49 | #define SCHED_FEAT(name, enabled) \ | |
50 | (1UL << __SCHED_FEAT_##name) * enabled | | |
51 | ||
52 | const_debug unsigned int sysctl_sched_features = | |
53 | #include "features.h" | |
54 | 0; | |
55 | ||
56 | #undef SCHED_FEAT | |
57 | ||
58 | /* | |
59 | * Number of tasks to iterate in a single balance run. | |
60 | * Limited because this is done with IRQs disabled. | |
61 | */ | |
62 | const_debug unsigned int sysctl_sched_nr_migrate = 32; | |
63 | ||
64 | /* | |
65 | * period over which we average the RT time consumption, measured | |
66 | * in ms. | |
67 | * | |
68 | * default: 1s | |
69 | */ | |
70 | const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; | |
71 | ||
72 | /* | |
73 | * period over which we measure -rt task CPU usage in us. | |
74 | * default: 1s | |
75 | */ | |
76 | unsigned int sysctl_sched_rt_period = 1000000; | |
77 | ||
78 | __read_mostly int scheduler_running; | |
79 | ||
80 | /* | |
81 | * part of the period that we allow rt tasks to run in us. | |
82 | * default: 0.95s | |
83 | */ | |
84 | int sysctl_sched_rt_runtime = 950000; | |
85 | ||
86 | /* CPUs with isolated domains */ | |
87 | cpumask_var_t cpu_isolated_map; | |
88 | ||
89 | /* | |
90 | * __task_rq_lock - lock the rq @p resides on. | |
91 | */ | |
92 | struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf) | |
93 | __acquires(rq->lock) | |
94 | { | |
95 | struct rq *rq; | |
96 | ||
97 | lockdep_assert_held(&p->pi_lock); | |
98 | ||
99 | for (;;) { | |
100 | rq = task_rq(p); | |
101 | raw_spin_lock(&rq->lock); | |
102 | if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { | |
103 | rq_pin_lock(rq, rf); | |
104 | return rq; | |
105 | } | |
106 | raw_spin_unlock(&rq->lock); | |
107 | ||
108 | while (unlikely(task_on_rq_migrating(p))) | |
109 | cpu_relax(); | |
110 | } | |
111 | } | |
112 | ||
113 | /* | |
114 | * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. | |
115 | */ | |
116 | struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf) | |
117 | __acquires(p->pi_lock) | |
118 | __acquires(rq->lock) | |
119 | { | |
120 | struct rq *rq; | |
121 | ||
122 | for (;;) { | |
123 | raw_spin_lock_irqsave(&p->pi_lock, rf->flags); | |
124 | rq = task_rq(p); | |
125 | raw_spin_lock(&rq->lock); | |
126 | /* | |
127 | * move_queued_task() task_rq_lock() | |
128 | * | |
129 | * ACQUIRE (rq->lock) | |
130 | * [S] ->on_rq = MIGRATING [L] rq = task_rq() | |
131 | * WMB (__set_task_cpu()) ACQUIRE (rq->lock); | |
132 | * [S] ->cpu = new_cpu [L] task_rq() | |
133 | * [L] ->on_rq | |
134 | * RELEASE (rq->lock) | |
135 | * | |
136 | * If we observe the old cpu in task_rq_lock, the acquire of | |
137 | * the old rq->lock will fully serialize against the stores. | |
138 | * | |
139 | * If we observe the new CPU in task_rq_lock, the acquire will | |
140 | * pair with the WMB to ensure we must then also see migrating. | |
141 | */ | |
142 | if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { | |
143 | rq_pin_lock(rq, rf); | |
144 | return rq; | |
145 | } | |
146 | raw_spin_unlock(&rq->lock); | |
147 | raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags); | |
148 | ||
149 | while (unlikely(task_on_rq_migrating(p))) | |
150 | cpu_relax(); | |
151 | } | |
152 | } | |
153 | ||
154 | /* | |
155 | * RQ-clock updating methods: | |
156 | */ | |
157 | ||
158 | static void update_rq_clock_task(struct rq *rq, s64 delta) | |
159 | { | |
160 | /* | |
161 | * In theory, the compile should just see 0 here, and optimize out the call | |
162 | * to sched_rt_avg_update. But I don't trust it... | |
163 | */ | |
164 | #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) | |
165 | s64 steal = 0, irq_delta = 0; | |
166 | #endif | |
167 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING | |
168 | irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; | |
169 | ||
170 | /* | |
171 | * Since irq_time is only updated on {soft,}irq_exit, we might run into | |
172 | * this case when a previous update_rq_clock() happened inside a | |
173 | * {soft,}irq region. | |
174 | * | |
175 | * When this happens, we stop ->clock_task and only update the | |
176 | * prev_irq_time stamp to account for the part that fit, so that a next | |
177 | * update will consume the rest. This ensures ->clock_task is | |
178 | * monotonic. | |
179 | * | |
180 | * It does however cause some slight miss-attribution of {soft,}irq | |
181 | * time, a more accurate solution would be to update the irq_time using | |
182 | * the current rq->clock timestamp, except that would require using | |
183 | * atomic ops. | |
184 | */ | |
185 | if (irq_delta > delta) | |
186 | irq_delta = delta; | |
187 | ||
188 | rq->prev_irq_time += irq_delta; | |
189 | delta -= irq_delta; | |
190 | #endif | |
191 | #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING | |
192 | if (static_key_false((¶virt_steal_rq_enabled))) { | |
193 | steal = paravirt_steal_clock(cpu_of(rq)); | |
194 | steal -= rq->prev_steal_time_rq; | |
195 | ||
196 | if (unlikely(steal > delta)) | |
197 | steal = delta; | |
198 | ||
199 | rq->prev_steal_time_rq += steal; | |
200 | delta -= steal; | |
201 | } | |
202 | #endif | |
203 | ||
204 | rq->clock_task += delta; | |
205 | ||
206 | #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) | |
207 | if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) | |
208 | sched_rt_avg_update(rq, irq_delta + steal); | |
209 | #endif | |
210 | } | |
211 | ||
212 | void update_rq_clock(struct rq *rq) | |
213 | { | |
214 | s64 delta; | |
215 | ||
216 | lockdep_assert_held(&rq->lock); | |
217 | ||
218 | if (rq->clock_update_flags & RQCF_ACT_SKIP) | |
219 | return; | |
220 | ||
221 | #ifdef CONFIG_SCHED_DEBUG | |
222 | if (sched_feat(WARN_DOUBLE_CLOCK)) | |
223 | SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED); | |
224 | rq->clock_update_flags |= RQCF_UPDATED; | |
225 | #endif | |
226 | ||
227 | delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; | |
228 | if (delta < 0) | |
229 | return; | |
230 | rq->clock += delta; | |
231 | update_rq_clock_task(rq, delta); | |
232 | } | |
233 | ||
234 | ||
235 | #ifdef CONFIG_SCHED_HRTICK | |
236 | /* | |
237 | * Use HR-timers to deliver accurate preemption points. | |
238 | */ | |
239 | ||
240 | static void hrtick_clear(struct rq *rq) | |
241 | { | |
242 | if (hrtimer_active(&rq->hrtick_timer)) | |
243 | hrtimer_cancel(&rq->hrtick_timer); | |
244 | } | |
245 | ||
246 | /* | |
247 | * High-resolution timer tick. | |
248 | * Runs from hardirq context with interrupts disabled. | |
249 | */ | |
250 | static enum hrtimer_restart hrtick(struct hrtimer *timer) | |
251 | { | |
252 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); | |
253 | struct rq_flags rf; | |
254 | ||
255 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
256 | ||
257 | rq_lock(rq, &rf); | |
258 | update_rq_clock(rq); | |
259 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); | |
260 | rq_unlock(rq, &rf); | |
261 | ||
262 | return HRTIMER_NORESTART; | |
263 | } | |
264 | ||
265 | #ifdef CONFIG_SMP | |
266 | ||
267 | static void __hrtick_restart(struct rq *rq) | |
268 | { | |
269 | struct hrtimer *timer = &rq->hrtick_timer; | |
270 | ||
271 | hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED); | |
272 | } | |
273 | ||
274 | /* | |
275 | * called from hardirq (IPI) context | |
276 | */ | |
277 | static void __hrtick_start(void *arg) | |
278 | { | |
279 | struct rq *rq = arg; | |
280 | struct rq_flags rf; | |
281 | ||
282 | rq_lock(rq, &rf); | |
283 | __hrtick_restart(rq); | |
284 | rq->hrtick_csd_pending = 0; | |
285 | rq_unlock(rq, &rf); | |
286 | } | |
287 | ||
288 | /* | |
289 | * Called to set the hrtick timer state. | |
290 | * | |
291 | * called with rq->lock held and irqs disabled | |
292 | */ | |
293 | void hrtick_start(struct rq *rq, u64 delay) | |
294 | { | |
295 | struct hrtimer *timer = &rq->hrtick_timer; | |
296 | ktime_t time; | |
297 | s64 delta; | |
298 | ||
299 | /* | |
300 | * Don't schedule slices shorter than 10000ns, that just | |
301 | * doesn't make sense and can cause timer DoS. | |
302 | */ | |
303 | delta = max_t(s64, delay, 10000LL); | |
304 | time = ktime_add_ns(timer->base->get_time(), delta); | |
305 | ||
306 | hrtimer_set_expires(timer, time); | |
307 | ||
308 | if (rq == this_rq()) { | |
309 | __hrtick_restart(rq); | |
310 | } else if (!rq->hrtick_csd_pending) { | |
311 | smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); | |
312 | rq->hrtick_csd_pending = 1; | |
313 | } | |
314 | } | |
315 | ||
316 | #else | |
317 | /* | |
318 | * Called to set the hrtick timer state. | |
319 | * | |
320 | * called with rq->lock held and irqs disabled | |
321 | */ | |
322 | void hrtick_start(struct rq *rq, u64 delay) | |
323 | { | |
324 | /* | |
325 | * Don't schedule slices shorter than 10000ns, that just | |
326 | * doesn't make sense. Rely on vruntime for fairness. | |
327 | */ | |
328 | delay = max_t(u64, delay, 10000LL); | |
329 | hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), | |
330 | HRTIMER_MODE_REL_PINNED); | |
331 | } | |
332 | #endif /* CONFIG_SMP */ | |
333 | ||
334 | static void init_rq_hrtick(struct rq *rq) | |
335 | { | |
336 | #ifdef CONFIG_SMP | |
337 | rq->hrtick_csd_pending = 0; | |
338 | ||
339 | rq->hrtick_csd.flags = 0; | |
340 | rq->hrtick_csd.func = __hrtick_start; | |
341 | rq->hrtick_csd.info = rq; | |
342 | #endif | |
343 | ||
344 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
345 | rq->hrtick_timer.function = hrtick; | |
346 | } | |
347 | #else /* CONFIG_SCHED_HRTICK */ | |
348 | static inline void hrtick_clear(struct rq *rq) | |
349 | { | |
350 | } | |
351 | ||
352 | static inline void init_rq_hrtick(struct rq *rq) | |
353 | { | |
354 | } | |
355 | #endif /* CONFIG_SCHED_HRTICK */ | |
356 | ||
357 | /* | |
358 | * cmpxchg based fetch_or, macro so it works for different integer types | |
359 | */ | |
360 | #define fetch_or(ptr, mask) \ | |
361 | ({ \ | |
362 | typeof(ptr) _ptr = (ptr); \ | |
363 | typeof(mask) _mask = (mask); \ | |
364 | typeof(*_ptr) _old, _val = *_ptr; \ | |
365 | \ | |
366 | for (;;) { \ | |
367 | _old = cmpxchg(_ptr, _val, _val | _mask); \ | |
368 | if (_old == _val) \ | |
369 | break; \ | |
370 | _val = _old; \ | |
371 | } \ | |
372 | _old; \ | |
373 | }) | |
374 | ||
375 | #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) | |
376 | /* | |
377 | * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, | |
378 | * this avoids any races wrt polling state changes and thereby avoids | |
379 | * spurious IPIs. | |
380 | */ | |
381 | static bool set_nr_and_not_polling(struct task_struct *p) | |
382 | { | |
383 | struct thread_info *ti = task_thread_info(p); | |
384 | return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); | |
385 | } | |
386 | ||
387 | /* | |
388 | * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. | |
389 | * | |
390 | * If this returns true, then the idle task promises to call | |
391 | * sched_ttwu_pending() and reschedule soon. | |
392 | */ | |
393 | static bool set_nr_if_polling(struct task_struct *p) | |
394 | { | |
395 | struct thread_info *ti = task_thread_info(p); | |
396 | typeof(ti->flags) old, val = READ_ONCE(ti->flags); | |
397 | ||
398 | for (;;) { | |
399 | if (!(val & _TIF_POLLING_NRFLAG)) | |
400 | return false; | |
401 | if (val & _TIF_NEED_RESCHED) | |
402 | return true; | |
403 | old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); | |
404 | if (old == val) | |
405 | break; | |
406 | val = old; | |
407 | } | |
408 | return true; | |
409 | } | |
410 | ||
411 | #else | |
412 | static bool set_nr_and_not_polling(struct task_struct *p) | |
413 | { | |
414 | set_tsk_need_resched(p); | |
415 | return true; | |
416 | } | |
417 | ||
418 | #ifdef CONFIG_SMP | |
419 | static bool set_nr_if_polling(struct task_struct *p) | |
420 | { | |
421 | return false; | |
422 | } | |
423 | #endif | |
424 | #endif | |
425 | ||
426 | void wake_q_add(struct wake_q_head *head, struct task_struct *task) | |
427 | { | |
428 | struct wake_q_node *node = &task->wake_q; | |
429 | ||
430 | /* | |
431 | * Atomically grab the task, if ->wake_q is !nil already it means | |
432 | * its already queued (either by us or someone else) and will get the | |
433 | * wakeup due to that. | |
434 | * | |
435 | * This cmpxchg() implies a full barrier, which pairs with the write | |
436 | * barrier implied by the wakeup in wake_up_q(). | |
437 | */ | |
438 | if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL)) | |
439 | return; | |
440 | ||
441 | get_task_struct(task); | |
442 | ||
443 | /* | |
444 | * The head is context local, there can be no concurrency. | |
445 | */ | |
446 | *head->lastp = node; | |
447 | head->lastp = &node->next; | |
448 | } | |
449 | ||
450 | void wake_up_q(struct wake_q_head *head) | |
451 | { | |
452 | struct wake_q_node *node = head->first; | |
453 | ||
454 | while (node != WAKE_Q_TAIL) { | |
455 | struct task_struct *task; | |
456 | ||
457 | task = container_of(node, struct task_struct, wake_q); | |
458 | BUG_ON(!task); | |
459 | /* Task can safely be re-inserted now: */ | |
460 | node = node->next; | |
461 | task->wake_q.next = NULL; | |
462 | ||
463 | /* | |
464 | * wake_up_process() implies a wmb() to pair with the queueing | |
465 | * in wake_q_add() so as not to miss wakeups. | |
466 | */ | |
467 | wake_up_process(task); | |
468 | put_task_struct(task); | |
469 | } | |
470 | } | |
471 | ||
472 | /* | |
473 | * resched_curr - mark rq's current task 'to be rescheduled now'. | |
474 | * | |
475 | * On UP this means the setting of the need_resched flag, on SMP it | |
476 | * might also involve a cross-CPU call to trigger the scheduler on | |
477 | * the target CPU. | |
478 | */ | |
479 | void resched_curr(struct rq *rq) | |
480 | { | |
481 | struct task_struct *curr = rq->curr; | |
482 | int cpu; | |
483 | ||
484 | lockdep_assert_held(&rq->lock); | |
485 | ||
486 | if (test_tsk_need_resched(curr)) | |
487 | return; | |
488 | ||
489 | cpu = cpu_of(rq); | |
490 | ||
491 | if (cpu == smp_processor_id()) { | |
492 | set_tsk_need_resched(curr); | |
493 | set_preempt_need_resched(); | |
494 | return; | |
495 | } | |
496 | ||
497 | if (set_nr_and_not_polling(curr)) | |
498 | smp_send_reschedule(cpu); | |
499 | else | |
500 | trace_sched_wake_idle_without_ipi(cpu); | |
501 | } | |
502 | ||
503 | void resched_cpu(int cpu) | |
504 | { | |
505 | struct rq *rq = cpu_rq(cpu); | |
506 | unsigned long flags; | |
507 | ||
508 | if (!raw_spin_trylock_irqsave(&rq->lock, flags)) | |
509 | return; | |
510 | resched_curr(rq); | |
511 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
512 | } | |
513 | ||
514 | #ifdef CONFIG_SMP | |
515 | #ifdef CONFIG_NO_HZ_COMMON | |
516 | /* | |
517 | * In the semi idle case, use the nearest busy CPU for migrating timers | |
518 | * from an idle CPU. This is good for power-savings. | |
519 | * | |
520 | * We don't do similar optimization for completely idle system, as | |
521 | * selecting an idle CPU will add more delays to the timers than intended | |
522 | * (as that CPU's timer base may not be uptodate wrt jiffies etc). | |
523 | */ | |
524 | int get_nohz_timer_target(void) | |
525 | { | |
526 | int i, cpu = smp_processor_id(); | |
527 | struct sched_domain *sd; | |
528 | ||
529 | if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu)) | |
530 | return cpu; | |
531 | ||
532 | rcu_read_lock(); | |
533 | for_each_domain(cpu, sd) { | |
534 | for_each_cpu(i, sched_domain_span(sd)) { | |
535 | if (cpu == i) | |
536 | continue; | |
537 | ||
538 | if (!idle_cpu(i) && is_housekeeping_cpu(i)) { | |
539 | cpu = i; | |
540 | goto unlock; | |
541 | } | |
542 | } | |
543 | } | |
544 | ||
545 | if (!is_housekeeping_cpu(cpu)) | |
546 | cpu = housekeeping_any_cpu(); | |
547 | unlock: | |
548 | rcu_read_unlock(); | |
549 | return cpu; | |
550 | } | |
551 | ||
552 | /* | |
553 | * When add_timer_on() enqueues a timer into the timer wheel of an | |
554 | * idle CPU then this timer might expire before the next timer event | |
555 | * which is scheduled to wake up that CPU. In case of a completely | |
556 | * idle system the next event might even be infinite time into the | |
557 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and | |
558 | * leaves the inner idle loop so the newly added timer is taken into | |
559 | * account when the CPU goes back to idle and evaluates the timer | |
560 | * wheel for the next timer event. | |
561 | */ | |
562 | static void wake_up_idle_cpu(int cpu) | |
563 | { | |
564 | struct rq *rq = cpu_rq(cpu); | |
565 | ||
566 | if (cpu == smp_processor_id()) | |
567 | return; | |
568 | ||
569 | if (set_nr_and_not_polling(rq->idle)) | |
570 | smp_send_reschedule(cpu); | |
571 | else | |
572 | trace_sched_wake_idle_without_ipi(cpu); | |
573 | } | |
574 | ||
575 | static bool wake_up_full_nohz_cpu(int cpu) | |
576 | { | |
577 | /* | |
578 | * We just need the target to call irq_exit() and re-evaluate | |
579 | * the next tick. The nohz full kick at least implies that. | |
580 | * If needed we can still optimize that later with an | |
581 | * empty IRQ. | |
582 | */ | |
583 | if (cpu_is_offline(cpu)) | |
584 | return true; /* Don't try to wake offline CPUs. */ | |
585 | if (tick_nohz_full_cpu(cpu)) { | |
586 | if (cpu != smp_processor_id() || | |
587 | tick_nohz_tick_stopped()) | |
588 | tick_nohz_full_kick_cpu(cpu); | |
589 | return true; | |
590 | } | |
591 | ||
592 | return false; | |
593 | } | |
594 | ||
595 | /* | |
596 | * Wake up the specified CPU. If the CPU is going offline, it is the | |
597 | * caller's responsibility to deal with the lost wakeup, for example, | |
598 | * by hooking into the CPU_DEAD notifier like timers and hrtimers do. | |
599 | */ | |
600 | void wake_up_nohz_cpu(int cpu) | |
601 | { | |
602 | if (!wake_up_full_nohz_cpu(cpu)) | |
603 | wake_up_idle_cpu(cpu); | |
604 | } | |
605 | ||
606 | static inline bool got_nohz_idle_kick(void) | |
607 | { | |
608 | int cpu = smp_processor_id(); | |
609 | ||
610 | if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) | |
611 | return false; | |
612 | ||
613 | if (idle_cpu(cpu) && !need_resched()) | |
614 | return true; | |
615 | ||
616 | /* | |
617 | * We can't run Idle Load Balance on this CPU for this time so we | |
618 | * cancel it and clear NOHZ_BALANCE_KICK | |
619 | */ | |
620 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); | |
621 | return false; | |
622 | } | |
623 | ||
624 | #else /* CONFIG_NO_HZ_COMMON */ | |
625 | ||
626 | static inline bool got_nohz_idle_kick(void) | |
627 | { | |
628 | return false; | |
629 | } | |
630 | ||
631 | #endif /* CONFIG_NO_HZ_COMMON */ | |
632 | ||
633 | #ifdef CONFIG_NO_HZ_FULL | |
634 | bool sched_can_stop_tick(struct rq *rq) | |
635 | { | |
636 | int fifo_nr_running; | |
637 | ||
638 | /* Deadline tasks, even if single, need the tick */ | |
639 | if (rq->dl.dl_nr_running) | |
640 | return false; | |
641 | ||
642 | /* | |
643 | * If there are more than one RR tasks, we need the tick to effect the | |
644 | * actual RR behaviour. | |
645 | */ | |
646 | if (rq->rt.rr_nr_running) { | |
647 | if (rq->rt.rr_nr_running == 1) | |
648 | return true; | |
649 | else | |
650 | return false; | |
651 | } | |
652 | ||
653 | /* | |
654 | * If there's no RR tasks, but FIFO tasks, we can skip the tick, no | |
655 | * forced preemption between FIFO tasks. | |
656 | */ | |
657 | fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running; | |
658 | if (fifo_nr_running) | |
659 | return true; | |
660 | ||
661 | /* | |
662 | * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left; | |
663 | * if there's more than one we need the tick for involuntary | |
664 | * preemption. | |
665 | */ | |
666 | if (rq->nr_running > 1) | |
667 | return false; | |
668 | ||
669 | return true; | |
670 | } | |
671 | #endif /* CONFIG_NO_HZ_FULL */ | |
672 | ||
673 | void sched_avg_update(struct rq *rq) | |
674 | { | |
675 | s64 period = sched_avg_period(); | |
676 | ||
677 | while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { | |
678 | /* | |
679 | * Inline assembly required to prevent the compiler | |
680 | * optimising this loop into a divmod call. | |
681 | * See __iter_div_u64_rem() for another example of this. | |
682 | */ | |
683 | asm("" : "+rm" (rq->age_stamp)); | |
684 | rq->age_stamp += period; | |
685 | rq->rt_avg /= 2; | |
686 | } | |
687 | } | |
688 | ||
689 | #endif /* CONFIG_SMP */ | |
690 | ||
691 | #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ | |
692 | (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) | |
693 | /* | |
694 | * Iterate task_group tree rooted at *from, calling @down when first entering a | |
695 | * node and @up when leaving it for the final time. | |
696 | * | |
697 | * Caller must hold rcu_lock or sufficient equivalent. | |
698 | */ | |
699 | int walk_tg_tree_from(struct task_group *from, | |
700 | tg_visitor down, tg_visitor up, void *data) | |
701 | { | |
702 | struct task_group *parent, *child; | |
703 | int ret; | |
704 | ||
705 | parent = from; | |
706 | ||
707 | down: | |
708 | ret = (*down)(parent, data); | |
709 | if (ret) | |
710 | goto out; | |
711 | list_for_each_entry_rcu(child, &parent->children, siblings) { | |
712 | parent = child; | |
713 | goto down; | |
714 | ||
715 | up: | |
716 | continue; | |
717 | } | |
718 | ret = (*up)(parent, data); | |
719 | if (ret || parent == from) | |
720 | goto out; | |
721 | ||
722 | child = parent; | |
723 | parent = parent->parent; | |
724 | if (parent) | |
725 | goto up; | |
726 | out: | |
727 | return ret; | |
728 | } | |
729 | ||
730 | int tg_nop(struct task_group *tg, void *data) | |
731 | { | |
732 | return 0; | |
733 | } | |
734 | #endif | |
735 | ||
736 | static void set_load_weight(struct task_struct *p) | |
737 | { | |
738 | int prio = p->static_prio - MAX_RT_PRIO; | |
739 | struct load_weight *load = &p->se.load; | |
740 | ||
741 | /* | |
742 | * SCHED_IDLE tasks get minimal weight: | |
743 | */ | |
744 | if (idle_policy(p->policy)) { | |
745 | load->weight = scale_load(WEIGHT_IDLEPRIO); | |
746 | load->inv_weight = WMULT_IDLEPRIO; | |
747 | return; | |
748 | } | |
749 | ||
750 | load->weight = scale_load(sched_prio_to_weight[prio]); | |
751 | load->inv_weight = sched_prio_to_wmult[prio]; | |
752 | } | |
753 | ||
754 | static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags) | |
755 | { | |
756 | if (!(flags & ENQUEUE_NOCLOCK)) | |
757 | update_rq_clock(rq); | |
758 | ||
759 | if (!(flags & ENQUEUE_RESTORE)) | |
760 | sched_info_queued(rq, p); | |
761 | ||
762 | p->sched_class->enqueue_task(rq, p, flags); | |
763 | } | |
764 | ||
765 | static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags) | |
766 | { | |
767 | if (!(flags & DEQUEUE_NOCLOCK)) | |
768 | update_rq_clock(rq); | |
769 | ||
770 | if (!(flags & DEQUEUE_SAVE)) | |
771 | sched_info_dequeued(rq, p); | |
772 | ||
773 | p->sched_class->dequeue_task(rq, p, flags); | |
774 | } | |
775 | ||
776 | void activate_task(struct rq *rq, struct task_struct *p, int flags) | |
777 | { | |
778 | if (task_contributes_to_load(p)) | |
779 | rq->nr_uninterruptible--; | |
780 | ||
781 | enqueue_task(rq, p, flags); | |
782 | } | |
783 | ||
784 | void deactivate_task(struct rq *rq, struct task_struct *p, int flags) | |
785 | { | |
786 | if (task_contributes_to_load(p)) | |
787 | rq->nr_uninterruptible++; | |
788 | ||
789 | dequeue_task(rq, p, flags); | |
790 | } | |
791 | ||
792 | /* | |
793 | * __normal_prio - return the priority that is based on the static prio | |
794 | */ | |
795 | static inline int __normal_prio(struct task_struct *p) | |
796 | { | |
797 | return p->static_prio; | |
798 | } | |
799 | ||
800 | /* | |
801 | * Calculate the expected normal priority: i.e. priority | |
802 | * without taking RT-inheritance into account. Might be | |
803 | * boosted by interactivity modifiers. Changes upon fork, | |
804 | * setprio syscalls, and whenever the interactivity | |
805 | * estimator recalculates. | |
806 | */ | |
807 | static inline int normal_prio(struct task_struct *p) | |
808 | { | |
809 | int prio; | |
810 | ||
811 | if (task_has_dl_policy(p)) | |
812 | prio = MAX_DL_PRIO-1; | |
813 | else if (task_has_rt_policy(p)) | |
814 | prio = MAX_RT_PRIO-1 - p->rt_priority; | |
815 | else | |
816 | prio = __normal_prio(p); | |
817 | return prio; | |
818 | } | |
819 | ||
820 | /* | |
821 | * Calculate the current priority, i.e. the priority | |
822 | * taken into account by the scheduler. This value might | |
823 | * be boosted by RT tasks, or might be boosted by | |
824 | * interactivity modifiers. Will be RT if the task got | |
825 | * RT-boosted. If not then it returns p->normal_prio. | |
826 | */ | |
827 | static int effective_prio(struct task_struct *p) | |
828 | { | |
829 | p->normal_prio = normal_prio(p); | |
830 | /* | |
831 | * If we are RT tasks or we were boosted to RT priority, | |
832 | * keep the priority unchanged. Otherwise, update priority | |
833 | * to the normal priority: | |
834 | */ | |
835 | if (!rt_prio(p->prio)) | |
836 | return p->normal_prio; | |
837 | return p->prio; | |
838 | } | |
839 | ||
840 | /** | |
841 | * task_curr - is this task currently executing on a CPU? | |
842 | * @p: the task in question. | |
843 | * | |
844 | * Return: 1 if the task is currently executing. 0 otherwise. | |
845 | */ | |
846 | inline int task_curr(const struct task_struct *p) | |
847 | { | |
848 | return cpu_curr(task_cpu(p)) == p; | |
849 | } | |
850 | ||
851 | /* | |
852 | * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock, | |
853 | * use the balance_callback list if you want balancing. | |
854 | * | |
855 | * this means any call to check_class_changed() must be followed by a call to | |
856 | * balance_callback(). | |
857 | */ | |
858 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, | |
859 | const struct sched_class *prev_class, | |
860 | int oldprio) | |
861 | { | |
862 | if (prev_class != p->sched_class) { | |
863 | if (prev_class->switched_from) | |
864 | prev_class->switched_from(rq, p); | |
865 | ||
866 | p->sched_class->switched_to(rq, p); | |
867 | } else if (oldprio != p->prio || dl_task(p)) | |
868 | p->sched_class->prio_changed(rq, p, oldprio); | |
869 | } | |
870 | ||
871 | void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) | |
872 | { | |
873 | const struct sched_class *class; | |
874 | ||
875 | if (p->sched_class == rq->curr->sched_class) { | |
876 | rq->curr->sched_class->check_preempt_curr(rq, p, flags); | |
877 | } else { | |
878 | for_each_class(class) { | |
879 | if (class == rq->curr->sched_class) | |
880 | break; | |
881 | if (class == p->sched_class) { | |
882 | resched_curr(rq); | |
883 | break; | |
884 | } | |
885 | } | |
886 | } | |
887 | ||
888 | /* | |
889 | * A queue event has occurred, and we're going to schedule. In | |
890 | * this case, we can save a useless back to back clock update. | |
891 | */ | |
892 | if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr)) | |
893 | rq_clock_skip_update(rq, true); | |
894 | } | |
895 | ||
896 | #ifdef CONFIG_SMP | |
897 | /* | |
898 | * This is how migration works: | |
899 | * | |
900 | * 1) we invoke migration_cpu_stop() on the target CPU using | |
901 | * stop_one_cpu(). | |
902 | * 2) stopper starts to run (implicitly forcing the migrated thread | |
903 | * off the CPU) | |
904 | * 3) it checks whether the migrated task is still in the wrong runqueue. | |
905 | * 4) if it's in the wrong runqueue then the migration thread removes | |
906 | * it and puts it into the right queue. | |
907 | * 5) stopper completes and stop_one_cpu() returns and the migration | |
908 | * is done. | |
909 | */ | |
910 | ||
911 | /* | |
912 | * move_queued_task - move a queued task to new rq. | |
913 | * | |
914 | * Returns (locked) new rq. Old rq's lock is released. | |
915 | */ | |
916 | static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf, | |
917 | struct task_struct *p, int new_cpu) | |
918 | { | |
919 | lockdep_assert_held(&rq->lock); | |
920 | ||
921 | p->on_rq = TASK_ON_RQ_MIGRATING; | |
922 | dequeue_task(rq, p, DEQUEUE_NOCLOCK); | |
923 | set_task_cpu(p, new_cpu); | |
924 | rq_unlock(rq, rf); | |
925 | ||
926 | rq = cpu_rq(new_cpu); | |
927 | ||
928 | rq_lock(rq, rf); | |
929 | BUG_ON(task_cpu(p) != new_cpu); | |
930 | enqueue_task(rq, p, 0); | |
931 | p->on_rq = TASK_ON_RQ_QUEUED; | |
932 | check_preempt_curr(rq, p, 0); | |
933 | ||
934 | return rq; | |
935 | } | |
936 | ||
937 | struct migration_arg { | |
938 | struct task_struct *task; | |
939 | int dest_cpu; | |
940 | }; | |
941 | ||
942 | /* | |
943 | * Move (not current) task off this CPU, onto the destination CPU. We're doing | |
944 | * this because either it can't run here any more (set_cpus_allowed() | |
945 | * away from this CPU, or CPU going down), or because we're | |
946 | * attempting to rebalance this task on exec (sched_exec). | |
947 | * | |
948 | * So we race with normal scheduler movements, but that's OK, as long | |
949 | * as the task is no longer on this CPU. | |
950 | */ | |
951 | static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf, | |
952 | struct task_struct *p, int dest_cpu) | |
953 | { | |
954 | if (unlikely(!cpu_active(dest_cpu))) | |
955 | return rq; | |
956 | ||
957 | /* Affinity changed (again). */ | |
958 | if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) | |
959 | return rq; | |
960 | ||
961 | update_rq_clock(rq); | |
962 | rq = move_queued_task(rq, rf, p, dest_cpu); | |
963 | ||
964 | return rq; | |
965 | } | |
966 | ||
967 | /* | |
968 | * migration_cpu_stop - this will be executed by a highprio stopper thread | |
969 | * and performs thread migration by bumping thread off CPU then | |
970 | * 'pushing' onto another runqueue. | |
971 | */ | |
972 | static int migration_cpu_stop(void *data) | |
973 | { | |
974 | struct migration_arg *arg = data; | |
975 | struct task_struct *p = arg->task; | |
976 | struct rq *rq = this_rq(); | |
977 | struct rq_flags rf; | |
978 | ||
979 | /* | |
980 | * The original target CPU might have gone down and we might | |
981 | * be on another CPU but it doesn't matter. | |
982 | */ | |
983 | local_irq_disable(); | |
984 | /* | |
985 | * We need to explicitly wake pending tasks before running | |
986 | * __migrate_task() such that we will not miss enforcing cpus_allowed | |
987 | * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. | |
988 | */ | |
989 | sched_ttwu_pending(); | |
990 | ||
991 | raw_spin_lock(&p->pi_lock); | |
992 | rq_lock(rq, &rf); | |
993 | /* | |
994 | * If task_rq(p) != rq, it cannot be migrated here, because we're | |
995 | * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because | |
996 | * we're holding p->pi_lock. | |
997 | */ | |
998 | if (task_rq(p) == rq) { | |
999 | if (task_on_rq_queued(p)) | |
1000 | rq = __migrate_task(rq, &rf, p, arg->dest_cpu); | |
1001 | else | |
1002 | p->wake_cpu = arg->dest_cpu; | |
1003 | } | |
1004 | rq_unlock(rq, &rf); | |
1005 | raw_spin_unlock(&p->pi_lock); | |
1006 | ||
1007 | local_irq_enable(); | |
1008 | return 0; | |
1009 | } | |
1010 | ||
1011 | /* | |
1012 | * sched_class::set_cpus_allowed must do the below, but is not required to | |
1013 | * actually call this function. | |
1014 | */ | |
1015 | void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask) | |
1016 | { | |
1017 | cpumask_copy(&p->cpus_allowed, new_mask); | |
1018 | p->nr_cpus_allowed = cpumask_weight(new_mask); | |
1019 | } | |
1020 | ||
1021 | void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) | |
1022 | { | |
1023 | struct rq *rq = task_rq(p); | |
1024 | bool queued, running; | |
1025 | ||
1026 | lockdep_assert_held(&p->pi_lock); | |
1027 | ||
1028 | queued = task_on_rq_queued(p); | |
1029 | running = task_current(rq, p); | |
1030 | ||
1031 | if (queued) { | |
1032 | /* | |
1033 | * Because __kthread_bind() calls this on blocked tasks without | |
1034 | * holding rq->lock. | |
1035 | */ | |
1036 | lockdep_assert_held(&rq->lock); | |
1037 | dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK); | |
1038 | } | |
1039 | if (running) | |
1040 | put_prev_task(rq, p); | |
1041 | ||
1042 | p->sched_class->set_cpus_allowed(p, new_mask); | |
1043 | ||
1044 | if (queued) | |
1045 | enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK); | |
1046 | if (running) | |
1047 | set_curr_task(rq, p); | |
1048 | } | |
1049 | ||
1050 | /* | |
1051 | * Change a given task's CPU affinity. Migrate the thread to a | |
1052 | * proper CPU and schedule it away if the CPU it's executing on | |
1053 | * is removed from the allowed bitmask. | |
1054 | * | |
1055 | * NOTE: the caller must have a valid reference to the task, the | |
1056 | * task must not exit() & deallocate itself prematurely. The | |
1057 | * call is not atomic; no spinlocks may be held. | |
1058 | */ | |
1059 | static int __set_cpus_allowed_ptr(struct task_struct *p, | |
1060 | const struct cpumask *new_mask, bool check) | |
1061 | { | |
1062 | const struct cpumask *cpu_valid_mask = cpu_active_mask; | |
1063 | unsigned int dest_cpu; | |
1064 | struct rq_flags rf; | |
1065 | struct rq *rq; | |
1066 | int ret = 0; | |
1067 | ||
1068 | rq = task_rq_lock(p, &rf); | |
1069 | update_rq_clock(rq); | |
1070 | ||
1071 | if (p->flags & PF_KTHREAD) { | |
1072 | /* | |
1073 | * Kernel threads are allowed on online && !active CPUs | |
1074 | */ | |
1075 | cpu_valid_mask = cpu_online_mask; | |
1076 | } | |
1077 | ||
1078 | /* | |
1079 | * Must re-check here, to close a race against __kthread_bind(), | |
1080 | * sched_setaffinity() is not guaranteed to observe the flag. | |
1081 | */ | |
1082 | if (check && (p->flags & PF_NO_SETAFFINITY)) { | |
1083 | ret = -EINVAL; | |
1084 | goto out; | |
1085 | } | |
1086 | ||
1087 | if (cpumask_equal(&p->cpus_allowed, new_mask)) | |
1088 | goto out; | |
1089 | ||
1090 | if (!cpumask_intersects(new_mask, cpu_valid_mask)) { | |
1091 | ret = -EINVAL; | |
1092 | goto out; | |
1093 | } | |
1094 | ||
1095 | do_set_cpus_allowed(p, new_mask); | |
1096 | ||
1097 | if (p->flags & PF_KTHREAD) { | |
1098 | /* | |
1099 | * For kernel threads that do indeed end up on online && | |
1100 | * !active we want to ensure they are strict per-CPU threads. | |
1101 | */ | |
1102 | WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) && | |
1103 | !cpumask_intersects(new_mask, cpu_active_mask) && | |
1104 | p->nr_cpus_allowed != 1); | |
1105 | } | |
1106 | ||
1107 | /* Can the task run on the task's current CPU? If so, we're done */ | |
1108 | if (cpumask_test_cpu(task_cpu(p), new_mask)) | |
1109 | goto out; | |
1110 | ||
1111 | dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask); | |
1112 | if (task_running(rq, p) || p->state == TASK_WAKING) { | |
1113 | struct migration_arg arg = { p, dest_cpu }; | |
1114 | /* Need help from migration thread: drop lock and wait. */ | |
1115 | task_rq_unlock(rq, p, &rf); | |
1116 | stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); | |
1117 | tlb_migrate_finish(p->mm); | |
1118 | return 0; | |
1119 | } else if (task_on_rq_queued(p)) { | |
1120 | /* | |
1121 | * OK, since we're going to drop the lock immediately | |
1122 | * afterwards anyway. | |
1123 | */ | |
1124 | rq = move_queued_task(rq, &rf, p, dest_cpu); | |
1125 | } | |
1126 | out: | |
1127 | task_rq_unlock(rq, p, &rf); | |
1128 | ||
1129 | return ret; | |
1130 | } | |
1131 | ||
1132 | int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) | |
1133 | { | |
1134 | return __set_cpus_allowed_ptr(p, new_mask, false); | |
1135 | } | |
1136 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); | |
1137 | ||
1138 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) | |
1139 | { | |
1140 | #ifdef CONFIG_SCHED_DEBUG | |
1141 | /* | |
1142 | * We should never call set_task_cpu() on a blocked task, | |
1143 | * ttwu() will sort out the placement. | |
1144 | */ | |
1145 | WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && | |
1146 | !p->on_rq); | |
1147 | ||
1148 | /* | |
1149 | * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING, | |
1150 | * because schedstat_wait_{start,end} rebase migrating task's wait_start | |
1151 | * time relying on p->on_rq. | |
1152 | */ | |
1153 | WARN_ON_ONCE(p->state == TASK_RUNNING && | |
1154 | p->sched_class == &fair_sched_class && | |
1155 | (p->on_rq && !task_on_rq_migrating(p))); | |
1156 | ||
1157 | #ifdef CONFIG_LOCKDEP | |
1158 | /* | |
1159 | * The caller should hold either p->pi_lock or rq->lock, when changing | |
1160 | * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. | |
1161 | * | |
1162 | * sched_move_task() holds both and thus holding either pins the cgroup, | |
1163 | * see task_group(). | |
1164 | * | |
1165 | * Furthermore, all task_rq users should acquire both locks, see | |
1166 | * task_rq_lock(). | |
1167 | */ | |
1168 | WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || | |
1169 | lockdep_is_held(&task_rq(p)->lock))); | |
1170 | #endif | |
1171 | #endif | |
1172 | ||
1173 | trace_sched_migrate_task(p, new_cpu); | |
1174 | ||
1175 | if (task_cpu(p) != new_cpu) { | |
1176 | if (p->sched_class->migrate_task_rq) | |
1177 | p->sched_class->migrate_task_rq(p); | |
1178 | p->se.nr_migrations++; | |
1179 | perf_event_task_migrate(p); | |
1180 | } | |
1181 | ||
1182 | __set_task_cpu(p, new_cpu); | |
1183 | } | |
1184 | ||
1185 | static void __migrate_swap_task(struct task_struct *p, int cpu) | |
1186 | { | |
1187 | if (task_on_rq_queued(p)) { | |
1188 | struct rq *src_rq, *dst_rq; | |
1189 | struct rq_flags srf, drf; | |
1190 | ||
1191 | src_rq = task_rq(p); | |
1192 | dst_rq = cpu_rq(cpu); | |
1193 | ||
1194 | rq_pin_lock(src_rq, &srf); | |
1195 | rq_pin_lock(dst_rq, &drf); | |
1196 | ||
1197 | p->on_rq = TASK_ON_RQ_MIGRATING; | |
1198 | deactivate_task(src_rq, p, 0); | |
1199 | set_task_cpu(p, cpu); | |
1200 | activate_task(dst_rq, p, 0); | |
1201 | p->on_rq = TASK_ON_RQ_QUEUED; | |
1202 | check_preempt_curr(dst_rq, p, 0); | |
1203 | ||
1204 | rq_unpin_lock(dst_rq, &drf); | |
1205 | rq_unpin_lock(src_rq, &srf); | |
1206 | ||
1207 | } else { | |
1208 | /* | |
1209 | * Task isn't running anymore; make it appear like we migrated | |
1210 | * it before it went to sleep. This means on wakeup we make the | |
1211 | * previous CPU our target instead of where it really is. | |
1212 | */ | |
1213 | p->wake_cpu = cpu; | |
1214 | } | |
1215 | } | |
1216 | ||
1217 | struct migration_swap_arg { | |
1218 | struct task_struct *src_task, *dst_task; | |
1219 | int src_cpu, dst_cpu; | |
1220 | }; | |
1221 | ||
1222 | static int migrate_swap_stop(void *data) | |
1223 | { | |
1224 | struct migration_swap_arg *arg = data; | |
1225 | struct rq *src_rq, *dst_rq; | |
1226 | int ret = -EAGAIN; | |
1227 | ||
1228 | if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu)) | |
1229 | return -EAGAIN; | |
1230 | ||
1231 | src_rq = cpu_rq(arg->src_cpu); | |
1232 | dst_rq = cpu_rq(arg->dst_cpu); | |
1233 | ||
1234 | double_raw_lock(&arg->src_task->pi_lock, | |
1235 | &arg->dst_task->pi_lock); | |
1236 | double_rq_lock(src_rq, dst_rq); | |
1237 | ||
1238 | if (task_cpu(arg->dst_task) != arg->dst_cpu) | |
1239 | goto unlock; | |
1240 | ||
1241 | if (task_cpu(arg->src_task) != arg->src_cpu) | |
1242 | goto unlock; | |
1243 | ||
1244 | if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed)) | |
1245 | goto unlock; | |
1246 | ||
1247 | if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed)) | |
1248 | goto unlock; | |
1249 | ||
1250 | __migrate_swap_task(arg->src_task, arg->dst_cpu); | |
1251 | __migrate_swap_task(arg->dst_task, arg->src_cpu); | |
1252 | ||
1253 | ret = 0; | |
1254 | ||
1255 | unlock: | |
1256 | double_rq_unlock(src_rq, dst_rq); | |
1257 | raw_spin_unlock(&arg->dst_task->pi_lock); | |
1258 | raw_spin_unlock(&arg->src_task->pi_lock); | |
1259 | ||
1260 | return ret; | |
1261 | } | |
1262 | ||
1263 | /* | |
1264 | * Cross migrate two tasks | |
1265 | */ | |
1266 | int migrate_swap(struct task_struct *cur, struct task_struct *p) | |
1267 | { | |
1268 | struct migration_swap_arg arg; | |
1269 | int ret = -EINVAL; | |
1270 | ||
1271 | arg = (struct migration_swap_arg){ | |
1272 | .src_task = cur, | |
1273 | .src_cpu = task_cpu(cur), | |
1274 | .dst_task = p, | |
1275 | .dst_cpu = task_cpu(p), | |
1276 | }; | |
1277 | ||
1278 | if (arg.src_cpu == arg.dst_cpu) | |
1279 | goto out; | |
1280 | ||
1281 | /* | |
1282 | * These three tests are all lockless; this is OK since all of them | |
1283 | * will be re-checked with proper locks held further down the line. | |
1284 | */ | |
1285 | if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) | |
1286 | goto out; | |
1287 | ||
1288 | if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed)) | |
1289 | goto out; | |
1290 | ||
1291 | if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed)) | |
1292 | goto out; | |
1293 | ||
1294 | trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); | |
1295 | ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); | |
1296 | ||
1297 | out: | |
1298 | return ret; | |
1299 | } | |
1300 | ||
1301 | /* | |
1302 | * wait_task_inactive - wait for a thread to unschedule. | |
1303 | * | |
1304 | * If @match_state is nonzero, it's the @p->state value just checked and | |
1305 | * not expected to change. If it changes, i.e. @p might have woken up, | |
1306 | * then return zero. When we succeed in waiting for @p to be off its CPU, | |
1307 | * we return a positive number (its total switch count). If a second call | |
1308 | * a short while later returns the same number, the caller can be sure that | |
1309 | * @p has remained unscheduled the whole time. | |
1310 | * | |
1311 | * The caller must ensure that the task *will* unschedule sometime soon, | |
1312 | * else this function might spin for a *long* time. This function can't | |
1313 | * be called with interrupts off, or it may introduce deadlock with | |
1314 | * smp_call_function() if an IPI is sent by the same process we are | |
1315 | * waiting to become inactive. | |
1316 | */ | |
1317 | unsigned long wait_task_inactive(struct task_struct *p, long match_state) | |
1318 | { | |
1319 | int running, queued; | |
1320 | struct rq_flags rf; | |
1321 | unsigned long ncsw; | |
1322 | struct rq *rq; | |
1323 | ||
1324 | for (;;) { | |
1325 | /* | |
1326 | * We do the initial early heuristics without holding | |
1327 | * any task-queue locks at all. We'll only try to get | |
1328 | * the runqueue lock when things look like they will | |
1329 | * work out! | |
1330 | */ | |
1331 | rq = task_rq(p); | |
1332 | ||
1333 | /* | |
1334 | * If the task is actively running on another CPU | |
1335 | * still, just relax and busy-wait without holding | |
1336 | * any locks. | |
1337 | * | |
1338 | * NOTE! Since we don't hold any locks, it's not | |
1339 | * even sure that "rq" stays as the right runqueue! | |
1340 | * But we don't care, since "task_running()" will | |
1341 | * return false if the runqueue has changed and p | |
1342 | * is actually now running somewhere else! | |
1343 | */ | |
1344 | while (task_running(rq, p)) { | |
1345 | if (match_state && unlikely(p->state != match_state)) | |
1346 | return 0; | |
1347 | cpu_relax(); | |
1348 | } | |
1349 | ||
1350 | /* | |
1351 | * Ok, time to look more closely! We need the rq | |
1352 | * lock now, to be *sure*. If we're wrong, we'll | |
1353 | * just go back and repeat. | |
1354 | */ | |
1355 | rq = task_rq_lock(p, &rf); | |
1356 | trace_sched_wait_task(p); | |
1357 | running = task_running(rq, p); | |
1358 | queued = task_on_rq_queued(p); | |
1359 | ncsw = 0; | |
1360 | if (!match_state || p->state == match_state) | |
1361 | ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ | |
1362 | task_rq_unlock(rq, p, &rf); | |
1363 | ||
1364 | /* | |
1365 | * If it changed from the expected state, bail out now. | |
1366 | */ | |
1367 | if (unlikely(!ncsw)) | |
1368 | break; | |
1369 | ||
1370 | /* | |
1371 | * Was it really running after all now that we | |
1372 | * checked with the proper locks actually held? | |
1373 | * | |
1374 | * Oops. Go back and try again.. | |
1375 | */ | |
1376 | if (unlikely(running)) { | |
1377 | cpu_relax(); | |
1378 | continue; | |
1379 | } | |
1380 | ||
1381 | /* | |
1382 | * It's not enough that it's not actively running, | |
1383 | * it must be off the runqueue _entirely_, and not | |
1384 | * preempted! | |
1385 | * | |
1386 | * So if it was still runnable (but just not actively | |
1387 | * running right now), it's preempted, and we should | |
1388 | * yield - it could be a while. | |
1389 | */ | |
1390 | if (unlikely(queued)) { | |
1391 | ktime_t to = NSEC_PER_SEC / HZ; | |
1392 | ||
1393 | set_current_state(TASK_UNINTERRUPTIBLE); | |
1394 | schedule_hrtimeout(&to, HRTIMER_MODE_REL); | |
1395 | continue; | |
1396 | } | |
1397 | ||
1398 | /* | |
1399 | * Ahh, all good. It wasn't running, and it wasn't | |
1400 | * runnable, which means that it will never become | |
1401 | * running in the future either. We're all done! | |
1402 | */ | |
1403 | break; | |
1404 | } | |
1405 | ||
1406 | return ncsw; | |
1407 | } | |
1408 | ||
1409 | /*** | |
1410 | * kick_process - kick a running thread to enter/exit the kernel | |
1411 | * @p: the to-be-kicked thread | |
1412 | * | |
1413 | * Cause a process which is running on another CPU to enter | |
1414 | * kernel-mode, without any delay. (to get signals handled.) | |
1415 | * | |
1416 | * NOTE: this function doesn't have to take the runqueue lock, | |
1417 | * because all it wants to ensure is that the remote task enters | |
1418 | * the kernel. If the IPI races and the task has been migrated | |
1419 | * to another CPU then no harm is done and the purpose has been | |
1420 | * achieved as well. | |
1421 | */ | |
1422 | void kick_process(struct task_struct *p) | |
1423 | { | |
1424 | int cpu; | |
1425 | ||
1426 | preempt_disable(); | |
1427 | cpu = task_cpu(p); | |
1428 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
1429 | smp_send_reschedule(cpu); | |
1430 | preempt_enable(); | |
1431 | } | |
1432 | EXPORT_SYMBOL_GPL(kick_process); | |
1433 | ||
1434 | /* | |
1435 | * ->cpus_allowed is protected by both rq->lock and p->pi_lock | |
1436 | * | |
1437 | * A few notes on cpu_active vs cpu_online: | |
1438 | * | |
1439 | * - cpu_active must be a subset of cpu_online | |
1440 | * | |
1441 | * - on cpu-up we allow per-cpu kthreads on the online && !active cpu, | |
1442 | * see __set_cpus_allowed_ptr(). At this point the newly online | |
1443 | * CPU isn't yet part of the sched domains, and balancing will not | |
1444 | * see it. | |
1445 | * | |
1446 | * - on CPU-down we clear cpu_active() to mask the sched domains and | |
1447 | * avoid the load balancer to place new tasks on the to be removed | |
1448 | * CPU. Existing tasks will remain running there and will be taken | |
1449 | * off. | |
1450 | * | |
1451 | * This means that fallback selection must not select !active CPUs. | |
1452 | * And can assume that any active CPU must be online. Conversely | |
1453 | * select_task_rq() below may allow selection of !active CPUs in order | |
1454 | * to satisfy the above rules. | |
1455 | */ | |
1456 | static int select_fallback_rq(int cpu, struct task_struct *p) | |
1457 | { | |
1458 | int nid = cpu_to_node(cpu); | |
1459 | const struct cpumask *nodemask = NULL; | |
1460 | enum { cpuset, possible, fail } state = cpuset; | |
1461 | int dest_cpu; | |
1462 | ||
1463 | /* | |
1464 | * If the node that the CPU is on has been offlined, cpu_to_node() | |
1465 | * will return -1. There is no CPU on the node, and we should | |
1466 | * select the CPU on the other node. | |
1467 | */ | |
1468 | if (nid != -1) { | |
1469 | nodemask = cpumask_of_node(nid); | |
1470 | ||
1471 | /* Look for allowed, online CPU in same node. */ | |
1472 | for_each_cpu(dest_cpu, nodemask) { | |
1473 | if (!cpu_active(dest_cpu)) | |
1474 | continue; | |
1475 | if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed)) | |
1476 | return dest_cpu; | |
1477 | } | |
1478 | } | |
1479 | ||
1480 | for (;;) { | |
1481 | /* Any allowed, online CPU? */ | |
1482 | for_each_cpu(dest_cpu, &p->cpus_allowed) { | |
1483 | if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu)) | |
1484 | continue; | |
1485 | if (!cpu_online(dest_cpu)) | |
1486 | continue; | |
1487 | goto out; | |
1488 | } | |
1489 | ||
1490 | /* No more Mr. Nice Guy. */ | |
1491 | switch (state) { | |
1492 | case cpuset: | |
1493 | if (IS_ENABLED(CONFIG_CPUSETS)) { | |
1494 | cpuset_cpus_allowed_fallback(p); | |
1495 | state = possible; | |
1496 | break; | |
1497 | } | |
1498 | /* Fall-through */ | |
1499 | case possible: | |
1500 | do_set_cpus_allowed(p, cpu_possible_mask); | |
1501 | state = fail; | |
1502 | break; | |
1503 | ||
1504 | case fail: | |
1505 | BUG(); | |
1506 | break; | |
1507 | } | |
1508 | } | |
1509 | ||
1510 | out: | |
1511 | if (state != cpuset) { | |
1512 | /* | |
1513 | * Don't tell them about moving exiting tasks or | |
1514 | * kernel threads (both mm NULL), since they never | |
1515 | * leave kernel. | |
1516 | */ | |
1517 | if (p->mm && printk_ratelimit()) { | |
1518 | printk_deferred("process %d (%s) no longer affine to cpu%d\n", | |
1519 | task_pid_nr(p), p->comm, cpu); | |
1520 | } | |
1521 | } | |
1522 | ||
1523 | return dest_cpu; | |
1524 | } | |
1525 | ||
1526 | /* | |
1527 | * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. | |
1528 | */ | |
1529 | static inline | |
1530 | int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) | |
1531 | { | |
1532 | lockdep_assert_held(&p->pi_lock); | |
1533 | ||
1534 | if (p->nr_cpus_allowed > 1) | |
1535 | cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); | |
1536 | else | |
1537 | cpu = cpumask_any(&p->cpus_allowed); | |
1538 | ||
1539 | /* | |
1540 | * In order not to call set_task_cpu() on a blocking task we need | |
1541 | * to rely on ttwu() to place the task on a valid ->cpus_allowed | |
1542 | * CPU. | |
1543 | * | |
1544 | * Since this is common to all placement strategies, this lives here. | |
1545 | * | |
1546 | * [ this allows ->select_task() to simply return task_cpu(p) and | |
1547 | * not worry about this generic constraint ] | |
1548 | */ | |
1549 | if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) || | |
1550 | !cpu_online(cpu))) | |
1551 | cpu = select_fallback_rq(task_cpu(p), p); | |
1552 | ||
1553 | return cpu; | |
1554 | } | |
1555 | ||
1556 | static void update_avg(u64 *avg, u64 sample) | |
1557 | { | |
1558 | s64 diff = sample - *avg; | |
1559 | *avg += diff >> 3; | |
1560 | } | |
1561 | ||
1562 | void sched_set_stop_task(int cpu, struct task_struct *stop) | |
1563 | { | |
1564 | struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; | |
1565 | struct task_struct *old_stop = cpu_rq(cpu)->stop; | |
1566 | ||
1567 | if (stop) { | |
1568 | /* | |
1569 | * Make it appear like a SCHED_FIFO task, its something | |
1570 | * userspace knows about and won't get confused about. | |
1571 | * | |
1572 | * Also, it will make PI more or less work without too | |
1573 | * much confusion -- but then, stop work should not | |
1574 | * rely on PI working anyway. | |
1575 | */ | |
1576 | sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); | |
1577 | ||
1578 | stop->sched_class = &stop_sched_class; | |
1579 | } | |
1580 | ||
1581 | cpu_rq(cpu)->stop = stop; | |
1582 | ||
1583 | if (old_stop) { | |
1584 | /* | |
1585 | * Reset it back to a normal scheduling class so that | |
1586 | * it can die in pieces. | |
1587 | */ | |
1588 | old_stop->sched_class = &rt_sched_class; | |
1589 | } | |
1590 | } | |
1591 | ||
1592 | #else | |
1593 | ||
1594 | static inline int __set_cpus_allowed_ptr(struct task_struct *p, | |
1595 | const struct cpumask *new_mask, bool check) | |
1596 | { | |
1597 | return set_cpus_allowed_ptr(p, new_mask); | |
1598 | } | |
1599 | ||
1600 | #endif /* CONFIG_SMP */ | |
1601 | ||
1602 | static void | |
1603 | ttwu_stat(struct task_struct *p, int cpu, int wake_flags) | |
1604 | { | |
1605 | struct rq *rq; | |
1606 | ||
1607 | if (!schedstat_enabled()) | |
1608 | return; | |
1609 | ||
1610 | rq = this_rq(); | |
1611 | ||
1612 | #ifdef CONFIG_SMP | |
1613 | if (cpu == rq->cpu) { | |
1614 | schedstat_inc(rq->ttwu_local); | |
1615 | schedstat_inc(p->se.statistics.nr_wakeups_local); | |
1616 | } else { | |
1617 | struct sched_domain *sd; | |
1618 | ||
1619 | schedstat_inc(p->se.statistics.nr_wakeups_remote); | |
1620 | rcu_read_lock(); | |
1621 | for_each_domain(rq->cpu, sd) { | |
1622 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { | |
1623 | schedstat_inc(sd->ttwu_wake_remote); | |
1624 | break; | |
1625 | } | |
1626 | } | |
1627 | rcu_read_unlock(); | |
1628 | } | |
1629 | ||
1630 | if (wake_flags & WF_MIGRATED) | |
1631 | schedstat_inc(p->se.statistics.nr_wakeups_migrate); | |
1632 | #endif /* CONFIG_SMP */ | |
1633 | ||
1634 | schedstat_inc(rq->ttwu_count); | |
1635 | schedstat_inc(p->se.statistics.nr_wakeups); | |
1636 | ||
1637 | if (wake_flags & WF_SYNC) | |
1638 | schedstat_inc(p->se.statistics.nr_wakeups_sync); | |
1639 | } | |
1640 | ||
1641 | static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) | |
1642 | { | |
1643 | activate_task(rq, p, en_flags); | |
1644 | p->on_rq = TASK_ON_RQ_QUEUED; | |
1645 | ||
1646 | /* If a worker is waking up, notify the workqueue: */ | |
1647 | if (p->flags & PF_WQ_WORKER) | |
1648 | wq_worker_waking_up(p, cpu_of(rq)); | |
1649 | } | |
1650 | ||
1651 | /* | |
1652 | * Mark the task runnable and perform wakeup-preemption. | |
1653 | */ | |
1654 | static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags, | |
1655 | struct rq_flags *rf) | |
1656 | { | |
1657 | check_preempt_curr(rq, p, wake_flags); | |
1658 | p->state = TASK_RUNNING; | |
1659 | trace_sched_wakeup(p); | |
1660 | ||
1661 | #ifdef CONFIG_SMP | |
1662 | if (p->sched_class->task_woken) { | |
1663 | /* | |
1664 | * Our task @p is fully woken up and running; so its safe to | |
1665 | * drop the rq->lock, hereafter rq is only used for statistics. | |
1666 | */ | |
1667 | rq_unpin_lock(rq, rf); | |
1668 | p->sched_class->task_woken(rq, p); | |
1669 | rq_repin_lock(rq, rf); | |
1670 | } | |
1671 | ||
1672 | if (rq->idle_stamp) { | |
1673 | u64 delta = rq_clock(rq) - rq->idle_stamp; | |
1674 | u64 max = 2*rq->max_idle_balance_cost; | |
1675 | ||
1676 | update_avg(&rq->avg_idle, delta); | |
1677 | ||
1678 | if (rq->avg_idle > max) | |
1679 | rq->avg_idle = max; | |
1680 | ||
1681 | rq->idle_stamp = 0; | |
1682 | } | |
1683 | #endif | |
1684 | } | |
1685 | ||
1686 | static void | |
1687 | ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags, | |
1688 | struct rq_flags *rf) | |
1689 | { | |
1690 | int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK; | |
1691 | ||
1692 | lockdep_assert_held(&rq->lock); | |
1693 | ||
1694 | #ifdef CONFIG_SMP | |
1695 | if (p->sched_contributes_to_load) | |
1696 | rq->nr_uninterruptible--; | |
1697 | ||
1698 | if (wake_flags & WF_MIGRATED) | |
1699 | en_flags |= ENQUEUE_MIGRATED; | |
1700 | #endif | |
1701 | ||
1702 | ttwu_activate(rq, p, en_flags); | |
1703 | ttwu_do_wakeup(rq, p, wake_flags, rf); | |
1704 | } | |
1705 | ||
1706 | /* | |
1707 | * Called in case the task @p isn't fully descheduled from its runqueue, | |
1708 | * in this case we must do a remote wakeup. Its a 'light' wakeup though, | |
1709 | * since all we need to do is flip p->state to TASK_RUNNING, since | |
1710 | * the task is still ->on_rq. | |
1711 | */ | |
1712 | static int ttwu_remote(struct task_struct *p, int wake_flags) | |
1713 | { | |
1714 | struct rq_flags rf; | |
1715 | struct rq *rq; | |
1716 | int ret = 0; | |
1717 | ||
1718 | rq = __task_rq_lock(p, &rf); | |
1719 | if (task_on_rq_queued(p)) { | |
1720 | /* check_preempt_curr() may use rq clock */ | |
1721 | update_rq_clock(rq); | |
1722 | ttwu_do_wakeup(rq, p, wake_flags, &rf); | |
1723 | ret = 1; | |
1724 | } | |
1725 | __task_rq_unlock(rq, &rf); | |
1726 | ||
1727 | return ret; | |
1728 | } | |
1729 | ||
1730 | #ifdef CONFIG_SMP | |
1731 | void sched_ttwu_pending(void) | |
1732 | { | |
1733 | struct rq *rq = this_rq(); | |
1734 | struct llist_node *llist = llist_del_all(&rq->wake_list); | |
1735 | struct task_struct *p, *t; | |
1736 | struct rq_flags rf; | |
1737 | ||
1738 | if (!llist) | |
1739 | return; | |
1740 | ||
1741 | rq_lock_irqsave(rq, &rf); | |
1742 | update_rq_clock(rq); | |
1743 | ||
1744 | llist_for_each_entry_safe(p, t, llist, wake_entry) | |
1745 | ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf); | |
1746 | ||
1747 | rq_unlock_irqrestore(rq, &rf); | |
1748 | } | |
1749 | ||
1750 | void scheduler_ipi(void) | |
1751 | { | |
1752 | /* | |
1753 | * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting | |
1754 | * TIF_NEED_RESCHED remotely (for the first time) will also send | |
1755 | * this IPI. | |
1756 | */ | |
1757 | preempt_fold_need_resched(); | |
1758 | ||
1759 | if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) | |
1760 | return; | |
1761 | ||
1762 | /* | |
1763 | * Not all reschedule IPI handlers call irq_enter/irq_exit, since | |
1764 | * traditionally all their work was done from the interrupt return | |
1765 | * path. Now that we actually do some work, we need to make sure | |
1766 | * we do call them. | |
1767 | * | |
1768 | * Some archs already do call them, luckily irq_enter/exit nest | |
1769 | * properly. | |
1770 | * | |
1771 | * Arguably we should visit all archs and update all handlers, | |
1772 | * however a fair share of IPIs are still resched only so this would | |
1773 | * somewhat pessimize the simple resched case. | |
1774 | */ | |
1775 | irq_enter(); | |
1776 | sched_ttwu_pending(); | |
1777 | ||
1778 | /* | |
1779 | * Check if someone kicked us for doing the nohz idle load balance. | |
1780 | */ | |
1781 | if (unlikely(got_nohz_idle_kick())) { | |
1782 | this_rq()->idle_balance = 1; | |
1783 | raise_softirq_irqoff(SCHED_SOFTIRQ); | |
1784 | } | |
1785 | irq_exit(); | |
1786 | } | |
1787 | ||
1788 | static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags) | |
1789 | { | |
1790 | struct rq *rq = cpu_rq(cpu); | |
1791 | ||
1792 | p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED); | |
1793 | ||
1794 | if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { | |
1795 | if (!set_nr_if_polling(rq->idle)) | |
1796 | smp_send_reschedule(cpu); | |
1797 | else | |
1798 | trace_sched_wake_idle_without_ipi(cpu); | |
1799 | } | |
1800 | } | |
1801 | ||
1802 | void wake_up_if_idle(int cpu) | |
1803 | { | |
1804 | struct rq *rq = cpu_rq(cpu); | |
1805 | struct rq_flags rf; | |
1806 | ||
1807 | rcu_read_lock(); | |
1808 | ||
1809 | if (!is_idle_task(rcu_dereference(rq->curr))) | |
1810 | goto out; | |
1811 | ||
1812 | if (set_nr_if_polling(rq->idle)) { | |
1813 | trace_sched_wake_idle_without_ipi(cpu); | |
1814 | } else { | |
1815 | rq_lock_irqsave(rq, &rf); | |
1816 | if (is_idle_task(rq->curr)) | |
1817 | smp_send_reschedule(cpu); | |
1818 | /* Else CPU is not idle, do nothing here: */ | |
1819 | rq_unlock_irqrestore(rq, &rf); | |
1820 | } | |
1821 | ||
1822 | out: | |
1823 | rcu_read_unlock(); | |
1824 | } | |
1825 | ||
1826 | bool cpus_share_cache(int this_cpu, int that_cpu) | |
1827 | { | |
1828 | return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); | |
1829 | } | |
1830 | #endif /* CONFIG_SMP */ | |
1831 | ||
1832 | static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags) | |
1833 | { | |
1834 | struct rq *rq = cpu_rq(cpu); | |
1835 | struct rq_flags rf; | |
1836 | ||
1837 | #if defined(CONFIG_SMP) | |
1838 | if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { | |
1839 | sched_clock_cpu(cpu); /* Sync clocks across CPUs */ | |
1840 | ttwu_queue_remote(p, cpu, wake_flags); | |
1841 | return; | |
1842 | } | |
1843 | #endif | |
1844 | ||
1845 | rq_lock(rq, &rf); | |
1846 | update_rq_clock(rq); | |
1847 | ttwu_do_activate(rq, p, wake_flags, &rf); | |
1848 | rq_unlock(rq, &rf); | |
1849 | } | |
1850 | ||
1851 | /* | |
1852 | * Notes on Program-Order guarantees on SMP systems. | |
1853 | * | |
1854 | * MIGRATION | |
1855 | * | |
1856 | * The basic program-order guarantee on SMP systems is that when a task [t] | |
1857 | * migrates, all its activity on its old CPU [c0] happens-before any subsequent | |
1858 | * execution on its new CPU [c1]. | |
1859 | * | |
1860 | * For migration (of runnable tasks) this is provided by the following means: | |
1861 | * | |
1862 | * A) UNLOCK of the rq(c0)->lock scheduling out task t | |
1863 | * B) migration for t is required to synchronize *both* rq(c0)->lock and | |
1864 | * rq(c1)->lock (if not at the same time, then in that order). | |
1865 | * C) LOCK of the rq(c1)->lock scheduling in task | |
1866 | * | |
1867 | * Transitivity guarantees that B happens after A and C after B. | |
1868 | * Note: we only require RCpc transitivity. | |
1869 | * Note: the CPU doing B need not be c0 or c1 | |
1870 | * | |
1871 | * Example: | |
1872 | * | |
1873 | * CPU0 CPU1 CPU2 | |
1874 | * | |
1875 | * LOCK rq(0)->lock | |
1876 | * sched-out X | |
1877 | * sched-in Y | |
1878 | * UNLOCK rq(0)->lock | |
1879 | * | |
1880 | * LOCK rq(0)->lock // orders against CPU0 | |
1881 | * dequeue X | |
1882 | * UNLOCK rq(0)->lock | |
1883 | * | |
1884 | * LOCK rq(1)->lock | |
1885 | * enqueue X | |
1886 | * UNLOCK rq(1)->lock | |
1887 | * | |
1888 | * LOCK rq(1)->lock // orders against CPU2 | |
1889 | * sched-out Z | |
1890 | * sched-in X | |
1891 | * UNLOCK rq(1)->lock | |
1892 | * | |
1893 | * | |
1894 | * BLOCKING -- aka. SLEEP + WAKEUP | |
1895 | * | |
1896 | * For blocking we (obviously) need to provide the same guarantee as for | |
1897 | * migration. However the means are completely different as there is no lock | |
1898 | * chain to provide order. Instead we do: | |
1899 | * | |
1900 | * 1) smp_store_release(X->on_cpu, 0) | |
1901 | * 2) smp_cond_load_acquire(!X->on_cpu) | |
1902 | * | |
1903 | * Example: | |
1904 | * | |
1905 | * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule) | |
1906 | * | |
1907 | * LOCK rq(0)->lock LOCK X->pi_lock | |
1908 | * dequeue X | |
1909 | * sched-out X | |
1910 | * smp_store_release(X->on_cpu, 0); | |
1911 | * | |
1912 | * smp_cond_load_acquire(&X->on_cpu, !VAL); | |
1913 | * X->state = WAKING | |
1914 | * set_task_cpu(X,2) | |
1915 | * | |
1916 | * LOCK rq(2)->lock | |
1917 | * enqueue X | |
1918 | * X->state = RUNNING | |
1919 | * UNLOCK rq(2)->lock | |
1920 | * | |
1921 | * LOCK rq(2)->lock // orders against CPU1 | |
1922 | * sched-out Z | |
1923 | * sched-in X | |
1924 | * UNLOCK rq(2)->lock | |
1925 | * | |
1926 | * UNLOCK X->pi_lock | |
1927 | * UNLOCK rq(0)->lock | |
1928 | * | |
1929 | * | |
1930 | * However; for wakeups there is a second guarantee we must provide, namely we | |
1931 | * must observe the state that lead to our wakeup. That is, not only must our | |
1932 | * task observe its own prior state, it must also observe the stores prior to | |
1933 | * its wakeup. | |
1934 | * | |
1935 | * This means that any means of doing remote wakeups must order the CPU doing | |
1936 | * the wakeup against the CPU the task is going to end up running on. This, | |
1937 | * however, is already required for the regular Program-Order guarantee above, | |
1938 | * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire). | |
1939 | * | |
1940 | */ | |
1941 | ||
1942 | /** | |
1943 | * try_to_wake_up - wake up a thread | |
1944 | * @p: the thread to be awakened | |
1945 | * @state: the mask of task states that can be woken | |
1946 | * @wake_flags: wake modifier flags (WF_*) | |
1947 | * | |
1948 | * If (@state & @p->state) @p->state = TASK_RUNNING. | |
1949 | * | |
1950 | * If the task was not queued/runnable, also place it back on a runqueue. | |
1951 | * | |
1952 | * Atomic against schedule() which would dequeue a task, also see | |
1953 | * set_current_state(). | |
1954 | * | |
1955 | * Return: %true if @p->state changes (an actual wakeup was done), | |
1956 | * %false otherwise. | |
1957 | */ | |
1958 | static int | |
1959 | try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) | |
1960 | { | |
1961 | unsigned long flags; | |
1962 | int cpu, success = 0; | |
1963 | ||
1964 | /* | |
1965 | * If we are going to wake up a thread waiting for CONDITION we | |
1966 | * need to ensure that CONDITION=1 done by the caller can not be | |
1967 | * reordered with p->state check below. This pairs with mb() in | |
1968 | * set_current_state() the waiting thread does. | |
1969 | */ | |
1970 | smp_mb__before_spinlock(); | |
1971 | raw_spin_lock_irqsave(&p->pi_lock, flags); | |
1972 | if (!(p->state & state)) | |
1973 | goto out; | |
1974 | ||
1975 | trace_sched_waking(p); | |
1976 | ||
1977 | /* We're going to change ->state: */ | |
1978 | success = 1; | |
1979 | cpu = task_cpu(p); | |
1980 | ||
1981 | /* | |
1982 | * Ensure we load p->on_rq _after_ p->state, otherwise it would | |
1983 | * be possible to, falsely, observe p->on_rq == 0 and get stuck | |
1984 | * in smp_cond_load_acquire() below. | |
1985 | * | |
1986 | * sched_ttwu_pending() try_to_wake_up() | |
1987 | * [S] p->on_rq = 1; [L] P->state | |
1988 | * UNLOCK rq->lock -----. | |
1989 | * \ | |
1990 | * +--- RMB | |
1991 | * schedule() / | |
1992 | * LOCK rq->lock -----' | |
1993 | * UNLOCK rq->lock | |
1994 | * | |
1995 | * [task p] | |
1996 | * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq | |
1997 | * | |
1998 | * Pairs with the UNLOCK+LOCK on rq->lock from the | |
1999 | * last wakeup of our task and the schedule that got our task | |
2000 | * current. | |
2001 | */ | |
2002 | smp_rmb(); | |
2003 | if (p->on_rq && ttwu_remote(p, wake_flags)) | |
2004 | goto stat; | |
2005 | ||
2006 | #ifdef CONFIG_SMP | |
2007 | /* | |
2008 | * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be | |
2009 | * possible to, falsely, observe p->on_cpu == 0. | |
2010 | * | |
2011 | * One must be running (->on_cpu == 1) in order to remove oneself | |
2012 | * from the runqueue. | |
2013 | * | |
2014 | * [S] ->on_cpu = 1; [L] ->on_rq | |
2015 | * UNLOCK rq->lock | |
2016 | * RMB | |
2017 | * LOCK rq->lock | |
2018 | * [S] ->on_rq = 0; [L] ->on_cpu | |
2019 | * | |
2020 | * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock | |
2021 | * from the consecutive calls to schedule(); the first switching to our | |
2022 | * task, the second putting it to sleep. | |
2023 | */ | |
2024 | smp_rmb(); | |
2025 | ||
2026 | /* | |
2027 | * If the owning (remote) CPU is still in the middle of schedule() with | |
2028 | * this task as prev, wait until its done referencing the task. | |
2029 | * | |
2030 | * Pairs with the smp_store_release() in finish_lock_switch(). | |
2031 | * | |
2032 | * This ensures that tasks getting woken will be fully ordered against | |
2033 | * their previous state and preserve Program Order. | |
2034 | */ | |
2035 | smp_cond_load_acquire(&p->on_cpu, !VAL); | |
2036 | ||
2037 | p->sched_contributes_to_load = !!task_contributes_to_load(p); | |
2038 | p->state = TASK_WAKING; | |
2039 | ||
2040 | if (p->in_iowait) { | |
2041 | delayacct_blkio_end(); | |
2042 | atomic_dec(&task_rq(p)->nr_iowait); | |
2043 | } | |
2044 | ||
2045 | cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); | |
2046 | if (task_cpu(p) != cpu) { | |
2047 | wake_flags |= WF_MIGRATED; | |
2048 | set_task_cpu(p, cpu); | |
2049 | } | |
2050 | ||
2051 | #else /* CONFIG_SMP */ | |
2052 | ||
2053 | if (p->in_iowait) { | |
2054 | delayacct_blkio_end(); | |
2055 | atomic_dec(&task_rq(p)->nr_iowait); | |
2056 | } | |
2057 | ||
2058 | #endif /* CONFIG_SMP */ | |
2059 | ||
2060 | ttwu_queue(p, cpu, wake_flags); | |
2061 | stat: | |
2062 | ttwu_stat(p, cpu, wake_flags); | |
2063 | out: | |
2064 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | |
2065 | ||
2066 | return success; | |
2067 | } | |
2068 | ||
2069 | /** | |
2070 | * try_to_wake_up_local - try to wake up a local task with rq lock held | |
2071 | * @p: the thread to be awakened | |
2072 | * @rf: request-queue flags for pinning | |
2073 | * | |
2074 | * Put @p on the run-queue if it's not already there. The caller must | |
2075 | * ensure that this_rq() is locked, @p is bound to this_rq() and not | |
2076 | * the current task. | |
2077 | */ | |
2078 | static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf) | |
2079 | { | |
2080 | struct rq *rq = task_rq(p); | |
2081 | ||
2082 | if (WARN_ON_ONCE(rq != this_rq()) || | |
2083 | WARN_ON_ONCE(p == current)) | |
2084 | return; | |
2085 | ||
2086 | lockdep_assert_held(&rq->lock); | |
2087 | ||
2088 | if (!raw_spin_trylock(&p->pi_lock)) { | |
2089 | /* | |
2090 | * This is OK, because current is on_cpu, which avoids it being | |
2091 | * picked for load-balance and preemption/IRQs are still | |
2092 | * disabled avoiding further scheduler activity on it and we've | |
2093 | * not yet picked a replacement task. | |
2094 | */ | |
2095 | rq_unlock(rq, rf); | |
2096 | raw_spin_lock(&p->pi_lock); | |
2097 | rq_relock(rq, rf); | |
2098 | } | |
2099 | ||
2100 | if (!(p->state & TASK_NORMAL)) | |
2101 | goto out; | |
2102 | ||
2103 | trace_sched_waking(p); | |
2104 | ||
2105 | if (!task_on_rq_queued(p)) { | |
2106 | if (p->in_iowait) { | |
2107 | delayacct_blkio_end(); | |
2108 | atomic_dec(&rq->nr_iowait); | |
2109 | } | |
2110 | ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK); | |
2111 | } | |
2112 | ||
2113 | ttwu_do_wakeup(rq, p, 0, rf); | |
2114 | ttwu_stat(p, smp_processor_id(), 0); | |
2115 | out: | |
2116 | raw_spin_unlock(&p->pi_lock); | |
2117 | } | |
2118 | ||
2119 | /** | |
2120 | * wake_up_process - Wake up a specific process | |
2121 | * @p: The process to be woken up. | |
2122 | * | |
2123 | * Attempt to wake up the nominated process and move it to the set of runnable | |
2124 | * processes. | |
2125 | * | |
2126 | * Return: 1 if the process was woken up, 0 if it was already running. | |
2127 | * | |
2128 | * It may be assumed that this function implies a write memory barrier before | |
2129 | * changing the task state if and only if any tasks are woken up. | |
2130 | */ | |
2131 | int wake_up_process(struct task_struct *p) | |
2132 | { | |
2133 | return try_to_wake_up(p, TASK_NORMAL, 0); | |
2134 | } | |
2135 | EXPORT_SYMBOL(wake_up_process); | |
2136 | ||
2137 | int wake_up_state(struct task_struct *p, unsigned int state) | |
2138 | { | |
2139 | return try_to_wake_up(p, state, 0); | |
2140 | } | |
2141 | ||
2142 | /* | |
2143 | * Perform scheduler related setup for a newly forked process p. | |
2144 | * p is forked by current. | |
2145 | * | |
2146 | * __sched_fork() is basic setup used by init_idle() too: | |
2147 | */ | |
2148 | static void __sched_fork(unsigned long clone_flags, struct task_struct *p) | |
2149 | { | |
2150 | p->on_rq = 0; | |
2151 | ||
2152 | p->se.on_rq = 0; | |
2153 | p->se.exec_start = 0; | |
2154 | p->se.sum_exec_runtime = 0; | |
2155 | p->se.prev_sum_exec_runtime = 0; | |
2156 | p->se.nr_migrations = 0; | |
2157 | p->se.vruntime = 0; | |
2158 | INIT_LIST_HEAD(&p->se.group_node); | |
2159 | ||
2160 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
2161 | p->se.cfs_rq = NULL; | |
2162 | #endif | |
2163 | ||
2164 | #ifdef CONFIG_SCHEDSTATS | |
2165 | /* Even if schedstat is disabled, there should not be garbage */ | |
2166 | memset(&p->se.statistics, 0, sizeof(p->se.statistics)); | |
2167 | #endif | |
2168 | ||
2169 | RB_CLEAR_NODE(&p->dl.rb_node); | |
2170 | init_dl_task_timer(&p->dl); | |
2171 | init_dl_inactive_task_timer(&p->dl); | |
2172 | __dl_clear_params(p); | |
2173 | ||
2174 | INIT_LIST_HEAD(&p->rt.run_list); | |
2175 | p->rt.timeout = 0; | |
2176 | p->rt.time_slice = sched_rr_timeslice; | |
2177 | p->rt.on_rq = 0; | |
2178 | p->rt.on_list = 0; | |
2179 | ||
2180 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
2181 | INIT_HLIST_HEAD(&p->preempt_notifiers); | |
2182 | #endif | |
2183 | ||
2184 | #ifdef CONFIG_NUMA_BALANCING | |
2185 | if (p->mm && atomic_read(&p->mm->mm_users) == 1) { | |
2186 | p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
2187 | p->mm->numa_scan_seq = 0; | |
2188 | } | |
2189 | ||
2190 | if (clone_flags & CLONE_VM) | |
2191 | p->numa_preferred_nid = current->numa_preferred_nid; | |
2192 | else | |
2193 | p->numa_preferred_nid = -1; | |
2194 | ||
2195 | p->node_stamp = 0ULL; | |
2196 | p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; | |
2197 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; | |
2198 | p->numa_work.next = &p->numa_work; | |
2199 | p->numa_faults = NULL; | |
2200 | p->last_task_numa_placement = 0; | |
2201 | p->last_sum_exec_runtime = 0; | |
2202 | ||
2203 | p->numa_group = NULL; | |
2204 | #endif /* CONFIG_NUMA_BALANCING */ | |
2205 | } | |
2206 | ||
2207 | DEFINE_STATIC_KEY_FALSE(sched_numa_balancing); | |
2208 | ||
2209 | #ifdef CONFIG_NUMA_BALANCING | |
2210 | ||
2211 | void set_numabalancing_state(bool enabled) | |
2212 | { | |
2213 | if (enabled) | |
2214 | static_branch_enable(&sched_numa_balancing); | |
2215 | else | |
2216 | static_branch_disable(&sched_numa_balancing); | |
2217 | } | |
2218 | ||
2219 | #ifdef CONFIG_PROC_SYSCTL | |
2220 | int sysctl_numa_balancing(struct ctl_table *table, int write, | |
2221 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
2222 | { | |
2223 | struct ctl_table t; | |
2224 | int err; | |
2225 | int state = static_branch_likely(&sched_numa_balancing); | |
2226 | ||
2227 | if (write && !capable(CAP_SYS_ADMIN)) | |
2228 | return -EPERM; | |
2229 | ||
2230 | t = *table; | |
2231 | t.data = &state; | |
2232 | err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); | |
2233 | if (err < 0) | |
2234 | return err; | |
2235 | if (write) | |
2236 | set_numabalancing_state(state); | |
2237 | return err; | |
2238 | } | |
2239 | #endif | |
2240 | #endif | |
2241 | ||
2242 | #ifdef CONFIG_SCHEDSTATS | |
2243 | ||
2244 | DEFINE_STATIC_KEY_FALSE(sched_schedstats); | |
2245 | static bool __initdata __sched_schedstats = false; | |
2246 | ||
2247 | static void set_schedstats(bool enabled) | |
2248 | { | |
2249 | if (enabled) | |
2250 | static_branch_enable(&sched_schedstats); | |
2251 | else | |
2252 | static_branch_disable(&sched_schedstats); | |
2253 | } | |
2254 | ||
2255 | void force_schedstat_enabled(void) | |
2256 | { | |
2257 | if (!schedstat_enabled()) { | |
2258 | pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n"); | |
2259 | static_branch_enable(&sched_schedstats); | |
2260 | } | |
2261 | } | |
2262 | ||
2263 | static int __init setup_schedstats(char *str) | |
2264 | { | |
2265 | int ret = 0; | |
2266 | if (!str) | |
2267 | goto out; | |
2268 | ||
2269 | /* | |
2270 | * This code is called before jump labels have been set up, so we can't | |
2271 | * change the static branch directly just yet. Instead set a temporary | |
2272 | * variable so init_schedstats() can do it later. | |
2273 | */ | |
2274 | if (!strcmp(str, "enable")) { | |
2275 | __sched_schedstats = true; | |
2276 | ret = 1; | |
2277 | } else if (!strcmp(str, "disable")) { | |
2278 | __sched_schedstats = false; | |
2279 | ret = 1; | |
2280 | } | |
2281 | out: | |
2282 | if (!ret) | |
2283 | pr_warn("Unable to parse schedstats=\n"); | |
2284 | ||
2285 | return ret; | |
2286 | } | |
2287 | __setup("schedstats=", setup_schedstats); | |
2288 | ||
2289 | static void __init init_schedstats(void) | |
2290 | { | |
2291 | set_schedstats(__sched_schedstats); | |
2292 | } | |
2293 | ||
2294 | #ifdef CONFIG_PROC_SYSCTL | |
2295 | int sysctl_schedstats(struct ctl_table *table, int write, | |
2296 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
2297 | { | |
2298 | struct ctl_table t; | |
2299 | int err; | |
2300 | int state = static_branch_likely(&sched_schedstats); | |
2301 | ||
2302 | if (write && !capable(CAP_SYS_ADMIN)) | |
2303 | return -EPERM; | |
2304 | ||
2305 | t = *table; | |
2306 | t.data = &state; | |
2307 | err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); | |
2308 | if (err < 0) | |
2309 | return err; | |
2310 | if (write) | |
2311 | set_schedstats(state); | |
2312 | return err; | |
2313 | } | |
2314 | #endif /* CONFIG_PROC_SYSCTL */ | |
2315 | #else /* !CONFIG_SCHEDSTATS */ | |
2316 | static inline void init_schedstats(void) {} | |
2317 | #endif /* CONFIG_SCHEDSTATS */ | |
2318 | ||
2319 | /* | |
2320 | * fork()/clone()-time setup: | |
2321 | */ | |
2322 | int sched_fork(unsigned long clone_flags, struct task_struct *p) | |
2323 | { | |
2324 | unsigned long flags; | |
2325 | int cpu = get_cpu(); | |
2326 | ||
2327 | __sched_fork(clone_flags, p); | |
2328 | /* | |
2329 | * We mark the process as NEW here. This guarantees that | |
2330 | * nobody will actually run it, and a signal or other external | |
2331 | * event cannot wake it up and insert it on the runqueue either. | |
2332 | */ | |
2333 | p->state = TASK_NEW; | |
2334 | ||
2335 | /* | |
2336 | * Make sure we do not leak PI boosting priority to the child. | |
2337 | */ | |
2338 | p->prio = current->normal_prio; | |
2339 | ||
2340 | /* | |
2341 | * Revert to default priority/policy on fork if requested. | |
2342 | */ | |
2343 | if (unlikely(p->sched_reset_on_fork)) { | |
2344 | if (task_has_dl_policy(p) || task_has_rt_policy(p)) { | |
2345 | p->policy = SCHED_NORMAL; | |
2346 | p->static_prio = NICE_TO_PRIO(0); | |
2347 | p->rt_priority = 0; | |
2348 | } else if (PRIO_TO_NICE(p->static_prio) < 0) | |
2349 | p->static_prio = NICE_TO_PRIO(0); | |
2350 | ||
2351 | p->prio = p->normal_prio = __normal_prio(p); | |
2352 | set_load_weight(p); | |
2353 | ||
2354 | /* | |
2355 | * We don't need the reset flag anymore after the fork. It has | |
2356 | * fulfilled its duty: | |
2357 | */ | |
2358 | p->sched_reset_on_fork = 0; | |
2359 | } | |
2360 | ||
2361 | if (dl_prio(p->prio)) { | |
2362 | put_cpu(); | |
2363 | return -EAGAIN; | |
2364 | } else if (rt_prio(p->prio)) { | |
2365 | p->sched_class = &rt_sched_class; | |
2366 | } else { | |
2367 | p->sched_class = &fair_sched_class; | |
2368 | } | |
2369 | ||
2370 | init_entity_runnable_average(&p->se); | |
2371 | ||
2372 | /* | |
2373 | * The child is not yet in the pid-hash so no cgroup attach races, | |
2374 | * and the cgroup is pinned to this child due to cgroup_fork() | |
2375 | * is ran before sched_fork(). | |
2376 | * | |
2377 | * Silence PROVE_RCU. | |
2378 | */ | |
2379 | raw_spin_lock_irqsave(&p->pi_lock, flags); | |
2380 | /* | |
2381 | * We're setting the CPU for the first time, we don't migrate, | |
2382 | * so use __set_task_cpu(). | |
2383 | */ | |
2384 | __set_task_cpu(p, cpu); | |
2385 | if (p->sched_class->task_fork) | |
2386 | p->sched_class->task_fork(p); | |
2387 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | |
2388 | ||
2389 | #ifdef CONFIG_SCHED_INFO | |
2390 | if (likely(sched_info_on())) | |
2391 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | |
2392 | #endif | |
2393 | #if defined(CONFIG_SMP) | |
2394 | p->on_cpu = 0; | |
2395 | #endif | |
2396 | init_task_preempt_count(p); | |
2397 | #ifdef CONFIG_SMP | |
2398 | plist_node_init(&p->pushable_tasks, MAX_PRIO); | |
2399 | RB_CLEAR_NODE(&p->pushable_dl_tasks); | |
2400 | #endif | |
2401 | ||
2402 | put_cpu(); | |
2403 | return 0; | |
2404 | } | |
2405 | ||
2406 | unsigned long to_ratio(u64 period, u64 runtime) | |
2407 | { | |
2408 | if (runtime == RUNTIME_INF) | |
2409 | return BW_UNIT; | |
2410 | ||
2411 | /* | |
2412 | * Doing this here saves a lot of checks in all | |
2413 | * the calling paths, and returning zero seems | |
2414 | * safe for them anyway. | |
2415 | */ | |
2416 | if (period == 0) | |
2417 | return 0; | |
2418 | ||
2419 | return div64_u64(runtime << BW_SHIFT, period); | |
2420 | } | |
2421 | ||
2422 | /* | |
2423 | * wake_up_new_task - wake up a newly created task for the first time. | |
2424 | * | |
2425 | * This function will do some initial scheduler statistics housekeeping | |
2426 | * that must be done for every newly created context, then puts the task | |
2427 | * on the runqueue and wakes it. | |
2428 | */ | |
2429 | void wake_up_new_task(struct task_struct *p) | |
2430 | { | |
2431 | struct rq_flags rf; | |
2432 | struct rq *rq; | |
2433 | ||
2434 | raw_spin_lock_irqsave(&p->pi_lock, rf.flags); | |
2435 | p->state = TASK_RUNNING; | |
2436 | #ifdef CONFIG_SMP | |
2437 | /* | |
2438 | * Fork balancing, do it here and not earlier because: | |
2439 | * - cpus_allowed can change in the fork path | |
2440 | * - any previously selected CPU might disappear through hotplug | |
2441 | * | |
2442 | * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq, | |
2443 | * as we're not fully set-up yet. | |
2444 | */ | |
2445 | __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); | |
2446 | #endif | |
2447 | rq = __task_rq_lock(p, &rf); | |
2448 | update_rq_clock(rq); | |
2449 | post_init_entity_util_avg(&p->se); | |
2450 | ||
2451 | activate_task(rq, p, ENQUEUE_NOCLOCK); | |
2452 | p->on_rq = TASK_ON_RQ_QUEUED; | |
2453 | trace_sched_wakeup_new(p); | |
2454 | check_preempt_curr(rq, p, WF_FORK); | |
2455 | #ifdef CONFIG_SMP | |
2456 | if (p->sched_class->task_woken) { | |
2457 | /* | |
2458 | * Nothing relies on rq->lock after this, so its fine to | |
2459 | * drop it. | |
2460 | */ | |
2461 | rq_unpin_lock(rq, &rf); | |
2462 | p->sched_class->task_woken(rq, p); | |
2463 | rq_repin_lock(rq, &rf); | |
2464 | } | |
2465 | #endif | |
2466 | task_rq_unlock(rq, p, &rf); | |
2467 | } | |
2468 | ||
2469 | #ifdef CONFIG_PREEMPT_NOTIFIERS | |
2470 | ||
2471 | static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE; | |
2472 | ||
2473 | void preempt_notifier_inc(void) | |
2474 | { | |
2475 | static_key_slow_inc(&preempt_notifier_key); | |
2476 | } | |
2477 | EXPORT_SYMBOL_GPL(preempt_notifier_inc); | |
2478 | ||
2479 | void preempt_notifier_dec(void) | |
2480 | { | |
2481 | static_key_slow_dec(&preempt_notifier_key); | |
2482 | } | |
2483 | EXPORT_SYMBOL_GPL(preempt_notifier_dec); | |
2484 | ||
2485 | /** | |
2486 | * preempt_notifier_register - tell me when current is being preempted & rescheduled | |
2487 | * @notifier: notifier struct to register | |
2488 | */ | |
2489 | void preempt_notifier_register(struct preempt_notifier *notifier) | |
2490 | { | |
2491 | if (!static_key_false(&preempt_notifier_key)) | |
2492 | WARN(1, "registering preempt_notifier while notifiers disabled\n"); | |
2493 | ||
2494 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); | |
2495 | } | |
2496 | EXPORT_SYMBOL_GPL(preempt_notifier_register); | |
2497 | ||
2498 | /** | |
2499 | * preempt_notifier_unregister - no longer interested in preemption notifications | |
2500 | * @notifier: notifier struct to unregister | |
2501 | * | |
2502 | * This is *not* safe to call from within a preemption notifier. | |
2503 | */ | |
2504 | void preempt_notifier_unregister(struct preempt_notifier *notifier) | |
2505 | { | |
2506 | hlist_del(¬ifier->link); | |
2507 | } | |
2508 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); | |
2509 | ||
2510 | static void __fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2511 | { | |
2512 | struct preempt_notifier *notifier; | |
2513 | ||
2514 | hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) | |
2515 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); | |
2516 | } | |
2517 | ||
2518 | static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2519 | { | |
2520 | if (static_key_false(&preempt_notifier_key)) | |
2521 | __fire_sched_in_preempt_notifiers(curr); | |
2522 | } | |
2523 | ||
2524 | static void | |
2525 | __fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2526 | struct task_struct *next) | |
2527 | { | |
2528 | struct preempt_notifier *notifier; | |
2529 | ||
2530 | hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) | |
2531 | notifier->ops->sched_out(notifier, next); | |
2532 | } | |
2533 | ||
2534 | static __always_inline void | |
2535 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2536 | struct task_struct *next) | |
2537 | { | |
2538 | if (static_key_false(&preempt_notifier_key)) | |
2539 | __fire_sched_out_preempt_notifiers(curr, next); | |
2540 | } | |
2541 | ||
2542 | #else /* !CONFIG_PREEMPT_NOTIFIERS */ | |
2543 | ||
2544 | static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2545 | { | |
2546 | } | |
2547 | ||
2548 | static inline void | |
2549 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2550 | struct task_struct *next) | |
2551 | { | |
2552 | } | |
2553 | ||
2554 | #endif /* CONFIG_PREEMPT_NOTIFIERS */ | |
2555 | ||
2556 | /** | |
2557 | * prepare_task_switch - prepare to switch tasks | |
2558 | * @rq: the runqueue preparing to switch | |
2559 | * @prev: the current task that is being switched out | |
2560 | * @next: the task we are going to switch to. | |
2561 | * | |
2562 | * This is called with the rq lock held and interrupts off. It must | |
2563 | * be paired with a subsequent finish_task_switch after the context | |
2564 | * switch. | |
2565 | * | |
2566 | * prepare_task_switch sets up locking and calls architecture specific | |
2567 | * hooks. | |
2568 | */ | |
2569 | static inline void | |
2570 | prepare_task_switch(struct rq *rq, struct task_struct *prev, | |
2571 | struct task_struct *next) | |
2572 | { | |
2573 | sched_info_switch(rq, prev, next); | |
2574 | perf_event_task_sched_out(prev, next); | |
2575 | fire_sched_out_preempt_notifiers(prev, next); | |
2576 | prepare_lock_switch(rq, next); | |
2577 | prepare_arch_switch(next); | |
2578 | } | |
2579 | ||
2580 | /** | |
2581 | * finish_task_switch - clean up after a task-switch | |
2582 | * @prev: the thread we just switched away from. | |
2583 | * | |
2584 | * finish_task_switch must be called after the context switch, paired | |
2585 | * with a prepare_task_switch call before the context switch. | |
2586 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
2587 | * and do any other architecture-specific cleanup actions. | |
2588 | * | |
2589 | * Note that we may have delayed dropping an mm in context_switch(). If | |
2590 | * so, we finish that here outside of the runqueue lock. (Doing it | |
2591 | * with the lock held can cause deadlocks; see schedule() for | |
2592 | * details.) | |
2593 | * | |
2594 | * The context switch have flipped the stack from under us and restored the | |
2595 | * local variables which were saved when this task called schedule() in the | |
2596 | * past. prev == current is still correct but we need to recalculate this_rq | |
2597 | * because prev may have moved to another CPU. | |
2598 | */ | |
2599 | static struct rq *finish_task_switch(struct task_struct *prev) | |
2600 | __releases(rq->lock) | |
2601 | { | |
2602 | struct rq *rq = this_rq(); | |
2603 | struct mm_struct *mm = rq->prev_mm; | |
2604 | long prev_state; | |
2605 | ||
2606 | /* | |
2607 | * The previous task will have left us with a preempt_count of 2 | |
2608 | * because it left us after: | |
2609 | * | |
2610 | * schedule() | |
2611 | * preempt_disable(); // 1 | |
2612 | * __schedule() | |
2613 | * raw_spin_lock_irq(&rq->lock) // 2 | |
2614 | * | |
2615 | * Also, see FORK_PREEMPT_COUNT. | |
2616 | */ | |
2617 | if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET, | |
2618 | "corrupted preempt_count: %s/%d/0x%x\n", | |
2619 | current->comm, current->pid, preempt_count())) | |
2620 | preempt_count_set(FORK_PREEMPT_COUNT); | |
2621 | ||
2622 | rq->prev_mm = NULL; | |
2623 | ||
2624 | /* | |
2625 | * A task struct has one reference for the use as "current". | |
2626 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls | |
2627 | * schedule one last time. The schedule call will never return, and | |
2628 | * the scheduled task must drop that reference. | |
2629 | * | |
2630 | * We must observe prev->state before clearing prev->on_cpu (in | |
2631 | * finish_lock_switch), otherwise a concurrent wakeup can get prev | |
2632 | * running on another CPU and we could rave with its RUNNING -> DEAD | |
2633 | * transition, resulting in a double drop. | |
2634 | */ | |
2635 | prev_state = prev->state; | |
2636 | vtime_task_switch(prev); | |
2637 | perf_event_task_sched_in(prev, current); | |
2638 | finish_lock_switch(rq, prev); | |
2639 | finish_arch_post_lock_switch(); | |
2640 | ||
2641 | fire_sched_in_preempt_notifiers(current); | |
2642 | if (mm) | |
2643 | mmdrop(mm); | |
2644 | if (unlikely(prev_state == TASK_DEAD)) { | |
2645 | if (prev->sched_class->task_dead) | |
2646 | prev->sched_class->task_dead(prev); | |
2647 | ||
2648 | /* | |
2649 | * Remove function-return probe instances associated with this | |
2650 | * task and put them back on the free list. | |
2651 | */ | |
2652 | kprobe_flush_task(prev); | |
2653 | ||
2654 | /* Task is done with its stack. */ | |
2655 | put_task_stack(prev); | |
2656 | ||
2657 | put_task_struct(prev); | |
2658 | } | |
2659 | ||
2660 | tick_nohz_task_switch(); | |
2661 | return rq; | |
2662 | } | |
2663 | ||
2664 | #ifdef CONFIG_SMP | |
2665 | ||
2666 | /* rq->lock is NOT held, but preemption is disabled */ | |
2667 | static void __balance_callback(struct rq *rq) | |
2668 | { | |
2669 | struct callback_head *head, *next; | |
2670 | void (*func)(struct rq *rq); | |
2671 | unsigned long flags; | |
2672 | ||
2673 | raw_spin_lock_irqsave(&rq->lock, flags); | |
2674 | head = rq->balance_callback; | |
2675 | rq->balance_callback = NULL; | |
2676 | while (head) { | |
2677 | func = (void (*)(struct rq *))head->func; | |
2678 | next = head->next; | |
2679 | head->next = NULL; | |
2680 | head = next; | |
2681 | ||
2682 | func(rq); | |
2683 | } | |
2684 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
2685 | } | |
2686 | ||
2687 | static inline void balance_callback(struct rq *rq) | |
2688 | { | |
2689 | if (unlikely(rq->balance_callback)) | |
2690 | __balance_callback(rq); | |
2691 | } | |
2692 | ||
2693 | #else | |
2694 | ||
2695 | static inline void balance_callback(struct rq *rq) | |
2696 | { | |
2697 | } | |
2698 | ||
2699 | #endif | |
2700 | ||
2701 | /** | |
2702 | * schedule_tail - first thing a freshly forked thread must call. | |
2703 | * @prev: the thread we just switched away from. | |
2704 | */ | |
2705 | asmlinkage __visible void schedule_tail(struct task_struct *prev) | |
2706 | __releases(rq->lock) | |
2707 | { | |
2708 | struct rq *rq; | |
2709 | ||
2710 | /* | |
2711 | * New tasks start with FORK_PREEMPT_COUNT, see there and | |
2712 | * finish_task_switch() for details. | |
2713 | * | |
2714 | * finish_task_switch() will drop rq->lock() and lower preempt_count | |
2715 | * and the preempt_enable() will end up enabling preemption (on | |
2716 | * PREEMPT_COUNT kernels). | |
2717 | */ | |
2718 | ||
2719 | rq = finish_task_switch(prev); | |
2720 | balance_callback(rq); | |
2721 | preempt_enable(); | |
2722 | ||
2723 | if (current->set_child_tid) | |
2724 | put_user(task_pid_vnr(current), current->set_child_tid); | |
2725 | } | |
2726 | ||
2727 | /* | |
2728 | * context_switch - switch to the new MM and the new thread's register state. | |
2729 | */ | |
2730 | static __always_inline struct rq * | |
2731 | context_switch(struct rq *rq, struct task_struct *prev, | |
2732 | struct task_struct *next, struct rq_flags *rf) | |
2733 | { | |
2734 | struct mm_struct *mm, *oldmm; | |
2735 | ||
2736 | prepare_task_switch(rq, prev, next); | |
2737 | ||
2738 | mm = next->mm; | |
2739 | oldmm = prev->active_mm; | |
2740 | /* | |
2741 | * For paravirt, this is coupled with an exit in switch_to to | |
2742 | * combine the page table reload and the switch backend into | |
2743 | * one hypercall. | |
2744 | */ | |
2745 | arch_start_context_switch(prev); | |
2746 | ||
2747 | if (!mm) { | |
2748 | next->active_mm = oldmm; | |
2749 | mmgrab(oldmm); | |
2750 | enter_lazy_tlb(oldmm, next); | |
2751 | } else | |
2752 | switch_mm_irqs_off(oldmm, mm, next); | |
2753 | ||
2754 | if (!prev->mm) { | |
2755 | prev->active_mm = NULL; | |
2756 | rq->prev_mm = oldmm; | |
2757 | } | |
2758 | ||
2759 | rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP); | |
2760 | ||
2761 | /* | |
2762 | * Since the runqueue lock will be released by the next | |
2763 | * task (which is an invalid locking op but in the case | |
2764 | * of the scheduler it's an obvious special-case), so we | |
2765 | * do an early lockdep release here: | |
2766 | */ | |
2767 | rq_unpin_lock(rq, rf); | |
2768 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); | |
2769 | ||
2770 | /* Here we just switch the register state and the stack. */ | |
2771 | switch_to(prev, next, prev); | |
2772 | barrier(); | |
2773 | ||
2774 | return finish_task_switch(prev); | |
2775 | } | |
2776 | ||
2777 | /* | |
2778 | * nr_running and nr_context_switches: | |
2779 | * | |
2780 | * externally visible scheduler statistics: current number of runnable | |
2781 | * threads, total number of context switches performed since bootup. | |
2782 | */ | |
2783 | unsigned long nr_running(void) | |
2784 | { | |
2785 | unsigned long i, sum = 0; | |
2786 | ||
2787 | for_each_online_cpu(i) | |
2788 | sum += cpu_rq(i)->nr_running; | |
2789 | ||
2790 | return sum; | |
2791 | } | |
2792 | ||
2793 | /* | |
2794 | * Check if only the current task is running on the CPU. | |
2795 | * | |
2796 | * Caution: this function does not check that the caller has disabled | |
2797 | * preemption, thus the result might have a time-of-check-to-time-of-use | |
2798 | * race. The caller is responsible to use it correctly, for example: | |
2799 | * | |
2800 | * - from a non-preemptable section (of course) | |
2801 | * | |
2802 | * - from a thread that is bound to a single CPU | |
2803 | * | |
2804 | * - in a loop with very short iterations (e.g. a polling loop) | |
2805 | */ | |
2806 | bool single_task_running(void) | |
2807 | { | |
2808 | return raw_rq()->nr_running == 1; | |
2809 | } | |
2810 | EXPORT_SYMBOL(single_task_running); | |
2811 | ||
2812 | unsigned long long nr_context_switches(void) | |
2813 | { | |
2814 | int i; | |
2815 | unsigned long long sum = 0; | |
2816 | ||
2817 | for_each_possible_cpu(i) | |
2818 | sum += cpu_rq(i)->nr_switches; | |
2819 | ||
2820 | return sum; | |
2821 | } | |
2822 | ||
2823 | /* | |
2824 | * IO-wait accounting, and how its mostly bollocks (on SMP). | |
2825 | * | |
2826 | * The idea behind IO-wait account is to account the idle time that we could | |
2827 | * have spend running if it were not for IO. That is, if we were to improve the | |
2828 | * storage performance, we'd have a proportional reduction in IO-wait time. | |
2829 | * | |
2830 | * This all works nicely on UP, where, when a task blocks on IO, we account | |
2831 | * idle time as IO-wait, because if the storage were faster, it could've been | |
2832 | * running and we'd not be idle. | |
2833 | * | |
2834 | * This has been extended to SMP, by doing the same for each CPU. This however | |
2835 | * is broken. | |
2836 | * | |
2837 | * Imagine for instance the case where two tasks block on one CPU, only the one | |
2838 | * CPU will have IO-wait accounted, while the other has regular idle. Even | |
2839 | * though, if the storage were faster, both could've ran at the same time, | |
2840 | * utilising both CPUs. | |
2841 | * | |
2842 | * This means, that when looking globally, the current IO-wait accounting on | |
2843 | * SMP is a lower bound, by reason of under accounting. | |
2844 | * | |
2845 | * Worse, since the numbers are provided per CPU, they are sometimes | |
2846 | * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly | |
2847 | * associated with any one particular CPU, it can wake to another CPU than it | |
2848 | * blocked on. This means the per CPU IO-wait number is meaningless. | |
2849 | * | |
2850 | * Task CPU affinities can make all that even more 'interesting'. | |
2851 | */ | |
2852 | ||
2853 | unsigned long nr_iowait(void) | |
2854 | { | |
2855 | unsigned long i, sum = 0; | |
2856 | ||
2857 | for_each_possible_cpu(i) | |
2858 | sum += atomic_read(&cpu_rq(i)->nr_iowait); | |
2859 | ||
2860 | return sum; | |
2861 | } | |
2862 | ||
2863 | /* | |
2864 | * Consumers of these two interfaces, like for example the cpufreq menu | |
2865 | * governor are using nonsensical data. Boosting frequency for a CPU that has | |
2866 | * IO-wait which might not even end up running the task when it does become | |
2867 | * runnable. | |
2868 | */ | |
2869 | ||
2870 | unsigned long nr_iowait_cpu(int cpu) | |
2871 | { | |
2872 | struct rq *this = cpu_rq(cpu); | |
2873 | return atomic_read(&this->nr_iowait); | |
2874 | } | |
2875 | ||
2876 | void get_iowait_load(unsigned long *nr_waiters, unsigned long *load) | |
2877 | { | |
2878 | struct rq *rq = this_rq(); | |
2879 | *nr_waiters = atomic_read(&rq->nr_iowait); | |
2880 | *load = rq->load.weight; | |
2881 | } | |
2882 | ||
2883 | #ifdef CONFIG_SMP | |
2884 | ||
2885 | /* | |
2886 | * sched_exec - execve() is a valuable balancing opportunity, because at | |
2887 | * this point the task has the smallest effective memory and cache footprint. | |
2888 | */ | |
2889 | void sched_exec(void) | |
2890 | { | |
2891 | struct task_struct *p = current; | |
2892 | unsigned long flags; | |
2893 | int dest_cpu; | |
2894 | ||
2895 | raw_spin_lock_irqsave(&p->pi_lock, flags); | |
2896 | dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); | |
2897 | if (dest_cpu == smp_processor_id()) | |
2898 | goto unlock; | |
2899 | ||
2900 | if (likely(cpu_active(dest_cpu))) { | |
2901 | struct migration_arg arg = { p, dest_cpu }; | |
2902 | ||
2903 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | |
2904 | stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); | |
2905 | return; | |
2906 | } | |
2907 | unlock: | |
2908 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | |
2909 | } | |
2910 | ||
2911 | #endif | |
2912 | ||
2913 | DEFINE_PER_CPU(struct kernel_stat, kstat); | |
2914 | DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); | |
2915 | ||
2916 | EXPORT_PER_CPU_SYMBOL(kstat); | |
2917 | EXPORT_PER_CPU_SYMBOL(kernel_cpustat); | |
2918 | ||
2919 | /* | |
2920 | * The function fair_sched_class.update_curr accesses the struct curr | |
2921 | * and its field curr->exec_start; when called from task_sched_runtime(), | |
2922 | * we observe a high rate of cache misses in practice. | |
2923 | * Prefetching this data results in improved performance. | |
2924 | */ | |
2925 | static inline void prefetch_curr_exec_start(struct task_struct *p) | |
2926 | { | |
2927 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
2928 | struct sched_entity *curr = (&p->se)->cfs_rq->curr; | |
2929 | #else | |
2930 | struct sched_entity *curr = (&task_rq(p)->cfs)->curr; | |
2931 | #endif | |
2932 | prefetch(curr); | |
2933 | prefetch(&curr->exec_start); | |
2934 | } | |
2935 | ||
2936 | /* | |
2937 | * Return accounted runtime for the task. | |
2938 | * In case the task is currently running, return the runtime plus current's | |
2939 | * pending runtime that have not been accounted yet. | |
2940 | */ | |
2941 | unsigned long long task_sched_runtime(struct task_struct *p) | |
2942 | { | |
2943 | struct rq_flags rf; | |
2944 | struct rq *rq; | |
2945 | u64 ns; | |
2946 | ||
2947 | #if defined(CONFIG_64BIT) && defined(CONFIG_SMP) | |
2948 | /* | |
2949 | * 64-bit doesn't need locks to atomically read a 64bit value. | |
2950 | * So we have a optimization chance when the task's delta_exec is 0. | |
2951 | * Reading ->on_cpu is racy, but this is ok. | |
2952 | * | |
2953 | * If we race with it leaving CPU, we'll take a lock. So we're correct. | |
2954 | * If we race with it entering CPU, unaccounted time is 0. This is | |
2955 | * indistinguishable from the read occurring a few cycles earlier. | |
2956 | * If we see ->on_cpu without ->on_rq, the task is leaving, and has | |
2957 | * been accounted, so we're correct here as well. | |
2958 | */ | |
2959 | if (!p->on_cpu || !task_on_rq_queued(p)) | |
2960 | return p->se.sum_exec_runtime; | |
2961 | #endif | |
2962 | ||
2963 | rq = task_rq_lock(p, &rf); | |
2964 | /* | |
2965 | * Must be ->curr _and_ ->on_rq. If dequeued, we would | |
2966 | * project cycles that may never be accounted to this | |
2967 | * thread, breaking clock_gettime(). | |
2968 | */ | |
2969 | if (task_current(rq, p) && task_on_rq_queued(p)) { | |
2970 | prefetch_curr_exec_start(p); | |
2971 | update_rq_clock(rq); | |
2972 | p->sched_class->update_curr(rq); | |
2973 | } | |
2974 | ns = p->se.sum_exec_runtime; | |
2975 | task_rq_unlock(rq, p, &rf); | |
2976 | ||
2977 | return ns; | |
2978 | } | |
2979 | ||
2980 | /* | |
2981 | * This function gets called by the timer code, with HZ frequency. | |
2982 | * We call it with interrupts disabled. | |
2983 | */ | |
2984 | void scheduler_tick(void) | |
2985 | { | |
2986 | int cpu = smp_processor_id(); | |
2987 | struct rq *rq = cpu_rq(cpu); | |
2988 | struct task_struct *curr = rq->curr; | |
2989 | struct rq_flags rf; | |
2990 | ||
2991 | sched_clock_tick(); | |
2992 | ||
2993 | rq_lock(rq, &rf); | |
2994 | ||
2995 | update_rq_clock(rq); | |
2996 | curr->sched_class->task_tick(rq, curr, 0); | |
2997 | cpu_load_update_active(rq); | |
2998 | calc_global_load_tick(rq); | |
2999 | ||
3000 | rq_unlock(rq, &rf); | |
3001 | ||
3002 | perf_event_task_tick(); | |
3003 | ||
3004 | #ifdef CONFIG_SMP | |
3005 | rq->idle_balance = idle_cpu(cpu); | |
3006 | trigger_load_balance(rq); | |
3007 | #endif | |
3008 | rq_last_tick_reset(rq); | |
3009 | } | |
3010 | ||
3011 | #ifdef CONFIG_NO_HZ_FULL | |
3012 | /** | |
3013 | * scheduler_tick_max_deferment | |
3014 | * | |
3015 | * Keep at least one tick per second when a single | |
3016 | * active task is running because the scheduler doesn't | |
3017 | * yet completely support full dynticks environment. | |
3018 | * | |
3019 | * This makes sure that uptime, CFS vruntime, load | |
3020 | * balancing, etc... continue to move forward, even | |
3021 | * with a very low granularity. | |
3022 | * | |
3023 | * Return: Maximum deferment in nanoseconds. | |
3024 | */ | |
3025 | u64 scheduler_tick_max_deferment(void) | |
3026 | { | |
3027 | struct rq *rq = this_rq(); | |
3028 | unsigned long next, now = READ_ONCE(jiffies); | |
3029 | ||
3030 | next = rq->last_sched_tick + HZ; | |
3031 | ||
3032 | if (time_before_eq(next, now)) | |
3033 | return 0; | |
3034 | ||
3035 | return jiffies_to_nsecs(next - now); | |
3036 | } | |
3037 | #endif | |
3038 | ||
3039 | #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ | |
3040 | defined(CONFIG_PREEMPT_TRACER)) | |
3041 | /* | |
3042 | * If the value passed in is equal to the current preempt count | |
3043 | * then we just disabled preemption. Start timing the latency. | |
3044 | */ | |
3045 | static inline void preempt_latency_start(int val) | |
3046 | { | |
3047 | if (preempt_count() == val) { | |
3048 | unsigned long ip = get_lock_parent_ip(); | |
3049 | #ifdef CONFIG_DEBUG_PREEMPT | |
3050 | current->preempt_disable_ip = ip; | |
3051 | #endif | |
3052 | trace_preempt_off(CALLER_ADDR0, ip); | |
3053 | } | |
3054 | } | |
3055 | ||
3056 | void preempt_count_add(int val) | |
3057 | { | |
3058 | #ifdef CONFIG_DEBUG_PREEMPT | |
3059 | /* | |
3060 | * Underflow? | |
3061 | */ | |
3062 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) | |
3063 | return; | |
3064 | #endif | |
3065 | __preempt_count_add(val); | |
3066 | #ifdef CONFIG_DEBUG_PREEMPT | |
3067 | /* | |
3068 | * Spinlock count overflowing soon? | |
3069 | */ | |
3070 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= | |
3071 | PREEMPT_MASK - 10); | |
3072 | #endif | |
3073 | preempt_latency_start(val); | |
3074 | } | |
3075 | EXPORT_SYMBOL(preempt_count_add); | |
3076 | NOKPROBE_SYMBOL(preempt_count_add); | |
3077 | ||
3078 | /* | |
3079 | * If the value passed in equals to the current preempt count | |
3080 | * then we just enabled preemption. Stop timing the latency. | |
3081 | */ | |
3082 | static inline void preempt_latency_stop(int val) | |
3083 | { | |
3084 | if (preempt_count() == val) | |
3085 | trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip()); | |
3086 | } | |
3087 | ||
3088 | void preempt_count_sub(int val) | |
3089 | { | |
3090 | #ifdef CONFIG_DEBUG_PREEMPT | |
3091 | /* | |
3092 | * Underflow? | |
3093 | */ | |
3094 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) | |
3095 | return; | |
3096 | /* | |
3097 | * Is the spinlock portion underflowing? | |
3098 | */ | |
3099 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && | |
3100 | !(preempt_count() & PREEMPT_MASK))) | |
3101 | return; | |
3102 | #endif | |
3103 | ||
3104 | preempt_latency_stop(val); | |
3105 | __preempt_count_sub(val); | |
3106 | } | |
3107 | EXPORT_SYMBOL(preempt_count_sub); | |
3108 | NOKPROBE_SYMBOL(preempt_count_sub); | |
3109 | ||
3110 | #else | |
3111 | static inline void preempt_latency_start(int val) { } | |
3112 | static inline void preempt_latency_stop(int val) { } | |
3113 | #endif | |
3114 | ||
3115 | static inline unsigned long get_preempt_disable_ip(struct task_struct *p) | |
3116 | { | |
3117 | #ifdef CONFIG_DEBUG_PREEMPT | |
3118 | return p->preempt_disable_ip; | |
3119 | #else | |
3120 | return 0; | |
3121 | #endif | |
3122 | } | |
3123 | ||
3124 | /* | |
3125 | * Print scheduling while atomic bug: | |
3126 | */ | |
3127 | static noinline void __schedule_bug(struct task_struct *prev) | |
3128 | { | |
3129 | /* Save this before calling printk(), since that will clobber it */ | |
3130 | unsigned long preempt_disable_ip = get_preempt_disable_ip(current); | |
3131 | ||
3132 | if (oops_in_progress) | |
3133 | return; | |
3134 | ||
3135 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", | |
3136 | prev->comm, prev->pid, preempt_count()); | |
3137 | ||
3138 | debug_show_held_locks(prev); | |
3139 | print_modules(); | |
3140 | if (irqs_disabled()) | |
3141 | print_irqtrace_events(prev); | |
3142 | if (IS_ENABLED(CONFIG_DEBUG_PREEMPT) | |
3143 | && in_atomic_preempt_off()) { | |
3144 | pr_err("Preemption disabled at:"); | |
3145 | print_ip_sym(preempt_disable_ip); | |
3146 | pr_cont("\n"); | |
3147 | } | |
3148 | if (panic_on_warn) | |
3149 | panic("scheduling while atomic\n"); | |
3150 | ||
3151 | dump_stack(); | |
3152 | add_taint(TAINT_WARN, LOCKDEP_STILL_OK); | |
3153 | } | |
3154 | ||
3155 | /* | |
3156 | * Various schedule()-time debugging checks and statistics: | |
3157 | */ | |
3158 | static inline void schedule_debug(struct task_struct *prev) | |
3159 | { | |
3160 | #ifdef CONFIG_SCHED_STACK_END_CHECK | |
3161 | if (task_stack_end_corrupted(prev)) | |
3162 | panic("corrupted stack end detected inside scheduler\n"); | |
3163 | #endif | |
3164 | ||
3165 | if (unlikely(in_atomic_preempt_off())) { | |
3166 | __schedule_bug(prev); | |
3167 | preempt_count_set(PREEMPT_DISABLED); | |
3168 | } | |
3169 | rcu_sleep_check(); | |
3170 | ||
3171 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | |
3172 | ||
3173 | schedstat_inc(this_rq()->sched_count); | |
3174 | } | |
3175 | ||
3176 | /* | |
3177 | * Pick up the highest-prio task: | |
3178 | */ | |
3179 | static inline struct task_struct * | |
3180 | pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) | |
3181 | { | |
3182 | const struct sched_class *class; | |
3183 | struct task_struct *p; | |
3184 | ||
3185 | /* | |
3186 | * Optimization: we know that if all tasks are in the fair class we can | |
3187 | * call that function directly, but only if the @prev task wasn't of a | |
3188 | * higher scheduling class, because otherwise those loose the | |
3189 | * opportunity to pull in more work from other CPUs. | |
3190 | */ | |
3191 | if (likely((prev->sched_class == &idle_sched_class || | |
3192 | prev->sched_class == &fair_sched_class) && | |
3193 | rq->nr_running == rq->cfs.h_nr_running)) { | |
3194 | ||
3195 | p = fair_sched_class.pick_next_task(rq, prev, rf); | |
3196 | if (unlikely(p == RETRY_TASK)) | |
3197 | goto again; | |
3198 | ||
3199 | /* Assumes fair_sched_class->next == idle_sched_class */ | |
3200 | if (unlikely(!p)) | |
3201 | p = idle_sched_class.pick_next_task(rq, prev, rf); | |
3202 | ||
3203 | return p; | |
3204 | } | |
3205 | ||
3206 | again: | |
3207 | for_each_class(class) { | |
3208 | p = class->pick_next_task(rq, prev, rf); | |
3209 | if (p) { | |
3210 | if (unlikely(p == RETRY_TASK)) | |
3211 | goto again; | |
3212 | return p; | |
3213 | } | |
3214 | } | |
3215 | ||
3216 | /* The idle class should always have a runnable task: */ | |
3217 | BUG(); | |
3218 | } | |
3219 | ||
3220 | /* | |
3221 | * __schedule() is the main scheduler function. | |
3222 | * | |
3223 | * The main means of driving the scheduler and thus entering this function are: | |
3224 | * | |
3225 | * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. | |
3226 | * | |
3227 | * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return | |
3228 | * paths. For example, see arch/x86/entry_64.S. | |
3229 | * | |
3230 | * To drive preemption between tasks, the scheduler sets the flag in timer | |
3231 | * interrupt handler scheduler_tick(). | |
3232 | * | |
3233 | * 3. Wakeups don't really cause entry into schedule(). They add a | |
3234 | * task to the run-queue and that's it. | |
3235 | * | |
3236 | * Now, if the new task added to the run-queue preempts the current | |
3237 | * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets | |
3238 | * called on the nearest possible occasion: | |
3239 | * | |
3240 | * - If the kernel is preemptible (CONFIG_PREEMPT=y): | |
3241 | * | |
3242 | * - in syscall or exception context, at the next outmost | |
3243 | * preempt_enable(). (this might be as soon as the wake_up()'s | |
3244 | * spin_unlock()!) | |
3245 | * | |
3246 | * - in IRQ context, return from interrupt-handler to | |
3247 | * preemptible context | |
3248 | * | |
3249 | * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) | |
3250 | * then at the next: | |
3251 | * | |
3252 | * - cond_resched() call | |
3253 | * - explicit schedule() call | |
3254 | * - return from syscall or exception to user-space | |
3255 | * - return from interrupt-handler to user-space | |
3256 | * | |
3257 | * WARNING: must be called with preemption disabled! | |
3258 | */ | |
3259 | static void __sched notrace __schedule(bool preempt) | |
3260 | { | |
3261 | struct task_struct *prev, *next; | |
3262 | unsigned long *switch_count; | |
3263 | struct rq_flags rf; | |
3264 | struct rq *rq; | |
3265 | int cpu; | |
3266 | ||
3267 | cpu = smp_processor_id(); | |
3268 | rq = cpu_rq(cpu); | |
3269 | prev = rq->curr; | |
3270 | ||
3271 | schedule_debug(prev); | |
3272 | ||
3273 | if (sched_feat(HRTICK)) | |
3274 | hrtick_clear(rq); | |
3275 | ||
3276 | local_irq_disable(); | |
3277 | rcu_note_context_switch(preempt); | |
3278 | ||
3279 | /* | |
3280 | * Make sure that signal_pending_state()->signal_pending() below | |
3281 | * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) | |
3282 | * done by the caller to avoid the race with signal_wake_up(). | |
3283 | */ | |
3284 | smp_mb__before_spinlock(); | |
3285 | rq_lock(rq, &rf); | |
3286 | ||
3287 | /* Promote REQ to ACT */ | |
3288 | rq->clock_update_flags <<= 1; | |
3289 | update_rq_clock(rq); | |
3290 | ||
3291 | switch_count = &prev->nivcsw; | |
3292 | if (!preempt && prev->state) { | |
3293 | if (unlikely(signal_pending_state(prev->state, prev))) { | |
3294 | prev->state = TASK_RUNNING; | |
3295 | } else { | |
3296 | deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK); | |
3297 | prev->on_rq = 0; | |
3298 | ||
3299 | if (prev->in_iowait) { | |
3300 | atomic_inc(&rq->nr_iowait); | |
3301 | delayacct_blkio_start(); | |
3302 | } | |
3303 | ||
3304 | /* | |
3305 | * If a worker went to sleep, notify and ask workqueue | |
3306 | * whether it wants to wake up a task to maintain | |
3307 | * concurrency. | |
3308 | */ | |
3309 | if (prev->flags & PF_WQ_WORKER) { | |
3310 | struct task_struct *to_wakeup; | |
3311 | ||
3312 | to_wakeup = wq_worker_sleeping(prev); | |
3313 | if (to_wakeup) | |
3314 | try_to_wake_up_local(to_wakeup, &rf); | |
3315 | } | |
3316 | } | |
3317 | switch_count = &prev->nvcsw; | |
3318 | } | |
3319 | ||
3320 | next = pick_next_task(rq, prev, &rf); | |
3321 | clear_tsk_need_resched(prev); | |
3322 | clear_preempt_need_resched(); | |
3323 | ||
3324 | if (likely(prev != next)) { | |
3325 | rq->nr_switches++; | |
3326 | rq->curr = next; | |
3327 | ++*switch_count; | |
3328 | ||
3329 | trace_sched_switch(preempt, prev, next); | |
3330 | ||
3331 | /* Also unlocks the rq: */ | |
3332 | rq = context_switch(rq, prev, next, &rf); | |
3333 | } else { | |
3334 | rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP); | |
3335 | rq_unlock_irq(rq, &rf); | |
3336 | } | |
3337 | ||
3338 | balance_callback(rq); | |
3339 | } | |
3340 | ||
3341 | void __noreturn do_task_dead(void) | |
3342 | { | |
3343 | /* | |
3344 | * The setting of TASK_RUNNING by try_to_wake_up() may be delayed | |
3345 | * when the following two conditions become true. | |
3346 | * - There is race condition of mmap_sem (It is acquired by | |
3347 | * exit_mm()), and | |
3348 | * - SMI occurs before setting TASK_RUNINNG. | |
3349 | * (or hypervisor of virtual machine switches to other guest) | |
3350 | * As a result, we may become TASK_RUNNING after becoming TASK_DEAD | |
3351 | * | |
3352 | * To avoid it, we have to wait for releasing tsk->pi_lock which | |
3353 | * is held by try_to_wake_up() | |
3354 | */ | |
3355 | smp_mb(); | |
3356 | raw_spin_unlock_wait(¤t->pi_lock); | |
3357 | ||
3358 | /* Causes final put_task_struct in finish_task_switch(): */ | |
3359 | __set_current_state(TASK_DEAD); | |
3360 | ||
3361 | /* Tell freezer to ignore us: */ | |
3362 | current->flags |= PF_NOFREEZE; | |
3363 | ||
3364 | __schedule(false); | |
3365 | BUG(); | |
3366 | ||
3367 | /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */ | |
3368 | for (;;) | |
3369 | cpu_relax(); | |
3370 | } | |
3371 | ||
3372 | static inline void sched_submit_work(struct task_struct *tsk) | |
3373 | { | |
3374 | if (!tsk->state || tsk_is_pi_blocked(tsk)) | |
3375 | return; | |
3376 | /* | |
3377 | * If we are going to sleep and we have plugged IO queued, | |
3378 | * make sure to submit it to avoid deadlocks. | |
3379 | */ | |
3380 | if (blk_needs_flush_plug(tsk)) | |
3381 | blk_schedule_flush_plug(tsk); | |
3382 | } | |
3383 | ||
3384 | asmlinkage __visible void __sched schedule(void) | |
3385 | { | |
3386 | struct task_struct *tsk = current; | |
3387 | ||
3388 | sched_submit_work(tsk); | |
3389 | do { | |
3390 | preempt_disable(); | |
3391 | __schedule(false); | |
3392 | sched_preempt_enable_no_resched(); | |
3393 | } while (need_resched()); | |
3394 | } | |
3395 | EXPORT_SYMBOL(schedule); | |
3396 | ||
3397 | /* | |
3398 | * synchronize_rcu_tasks() makes sure that no task is stuck in preempted | |
3399 | * state (have scheduled out non-voluntarily) by making sure that all | |
3400 | * tasks have either left the run queue or have gone into user space. | |
3401 | * As idle tasks do not do either, they must not ever be preempted | |
3402 | * (schedule out non-voluntarily). | |
3403 | * | |
3404 | * schedule_idle() is similar to schedule_preempt_disable() except that it | |
3405 | * never enables preemption because it does not call sched_submit_work(). | |
3406 | */ | |
3407 | void __sched schedule_idle(void) | |
3408 | { | |
3409 | /* | |
3410 | * As this skips calling sched_submit_work(), which the idle task does | |
3411 | * regardless because that function is a nop when the task is in a | |
3412 | * TASK_RUNNING state, make sure this isn't used someplace that the | |
3413 | * current task can be in any other state. Note, idle is always in the | |
3414 | * TASK_RUNNING state. | |
3415 | */ | |
3416 | WARN_ON_ONCE(current->state); | |
3417 | do { | |
3418 | __schedule(false); | |
3419 | } while (need_resched()); | |
3420 | } | |
3421 | ||
3422 | #ifdef CONFIG_CONTEXT_TRACKING | |
3423 | asmlinkage __visible void __sched schedule_user(void) | |
3424 | { | |
3425 | /* | |
3426 | * If we come here after a random call to set_need_resched(), | |
3427 | * or we have been woken up remotely but the IPI has not yet arrived, | |
3428 | * we haven't yet exited the RCU idle mode. Do it here manually until | |
3429 | * we find a better solution. | |
3430 | * | |
3431 | * NB: There are buggy callers of this function. Ideally we | |
3432 | * should warn if prev_state != CONTEXT_USER, but that will trigger | |
3433 | * too frequently to make sense yet. | |
3434 | */ | |
3435 | enum ctx_state prev_state = exception_enter(); | |
3436 | schedule(); | |
3437 | exception_exit(prev_state); | |
3438 | } | |
3439 | #endif | |
3440 | ||
3441 | /** | |
3442 | * schedule_preempt_disabled - called with preemption disabled | |
3443 | * | |
3444 | * Returns with preemption disabled. Note: preempt_count must be 1 | |
3445 | */ | |
3446 | void __sched schedule_preempt_disabled(void) | |
3447 | { | |
3448 | sched_preempt_enable_no_resched(); | |
3449 | schedule(); | |
3450 | preempt_disable(); | |
3451 | } | |
3452 | ||
3453 | static void __sched notrace preempt_schedule_common(void) | |
3454 | { | |
3455 | do { | |
3456 | /* | |
3457 | * Because the function tracer can trace preempt_count_sub() | |
3458 | * and it also uses preempt_enable/disable_notrace(), if | |
3459 | * NEED_RESCHED is set, the preempt_enable_notrace() called | |
3460 | * by the function tracer will call this function again and | |
3461 | * cause infinite recursion. | |
3462 | * | |
3463 | * Preemption must be disabled here before the function | |
3464 | * tracer can trace. Break up preempt_disable() into two | |
3465 | * calls. One to disable preemption without fear of being | |
3466 | * traced. The other to still record the preemption latency, | |
3467 | * which can also be traced by the function tracer. | |
3468 | */ | |
3469 | preempt_disable_notrace(); | |
3470 | preempt_latency_start(1); | |
3471 | __schedule(true); | |
3472 | preempt_latency_stop(1); | |
3473 | preempt_enable_no_resched_notrace(); | |
3474 | ||
3475 | /* | |
3476 | * Check again in case we missed a preemption opportunity | |
3477 | * between schedule and now. | |
3478 | */ | |
3479 | } while (need_resched()); | |
3480 | } | |
3481 | ||
3482 | #ifdef CONFIG_PREEMPT | |
3483 | /* | |
3484 | * this is the entry point to schedule() from in-kernel preemption | |
3485 | * off of preempt_enable. Kernel preemptions off return from interrupt | |
3486 | * occur there and call schedule directly. | |
3487 | */ | |
3488 | asmlinkage __visible void __sched notrace preempt_schedule(void) | |
3489 | { | |
3490 | /* | |
3491 | * If there is a non-zero preempt_count or interrupts are disabled, | |
3492 | * we do not want to preempt the current task. Just return.. | |
3493 | */ | |
3494 | if (likely(!preemptible())) | |
3495 | return; | |
3496 | ||
3497 | preempt_schedule_common(); | |
3498 | } | |
3499 | NOKPROBE_SYMBOL(preempt_schedule); | |
3500 | EXPORT_SYMBOL(preempt_schedule); | |
3501 | ||
3502 | /** | |
3503 | * preempt_schedule_notrace - preempt_schedule called by tracing | |
3504 | * | |
3505 | * The tracing infrastructure uses preempt_enable_notrace to prevent | |
3506 | * recursion and tracing preempt enabling caused by the tracing | |
3507 | * infrastructure itself. But as tracing can happen in areas coming | |
3508 | * from userspace or just about to enter userspace, a preempt enable | |
3509 | * can occur before user_exit() is called. This will cause the scheduler | |
3510 | * to be called when the system is still in usermode. | |
3511 | * | |
3512 | * To prevent this, the preempt_enable_notrace will use this function | |
3513 | * instead of preempt_schedule() to exit user context if needed before | |
3514 | * calling the scheduler. | |
3515 | */ | |
3516 | asmlinkage __visible void __sched notrace preempt_schedule_notrace(void) | |
3517 | { | |
3518 | enum ctx_state prev_ctx; | |
3519 | ||
3520 | if (likely(!preemptible())) | |
3521 | return; | |
3522 | ||
3523 | do { | |
3524 | /* | |
3525 | * Because the function tracer can trace preempt_count_sub() | |
3526 | * and it also uses preempt_enable/disable_notrace(), if | |
3527 | * NEED_RESCHED is set, the preempt_enable_notrace() called | |
3528 | * by the function tracer will call this function again and | |
3529 | * cause infinite recursion. | |
3530 | * | |
3531 | * Preemption must be disabled here before the function | |
3532 | * tracer can trace. Break up preempt_disable() into two | |
3533 | * calls. One to disable preemption without fear of being | |
3534 | * traced. The other to still record the preemption latency, | |
3535 | * which can also be traced by the function tracer. | |
3536 | */ | |
3537 | preempt_disable_notrace(); | |
3538 | preempt_latency_start(1); | |
3539 | /* | |
3540 | * Needs preempt disabled in case user_exit() is traced | |
3541 | * and the tracer calls preempt_enable_notrace() causing | |
3542 | * an infinite recursion. | |
3543 | */ | |
3544 | prev_ctx = exception_enter(); | |
3545 | __schedule(true); | |
3546 | exception_exit(prev_ctx); | |
3547 | ||
3548 | preempt_latency_stop(1); | |
3549 | preempt_enable_no_resched_notrace(); | |
3550 | } while (need_resched()); | |
3551 | } | |
3552 | EXPORT_SYMBOL_GPL(preempt_schedule_notrace); | |
3553 | ||
3554 | #endif /* CONFIG_PREEMPT */ | |
3555 | ||
3556 | /* | |
3557 | * this is the entry point to schedule() from kernel preemption | |
3558 | * off of irq context. | |
3559 | * Note, that this is called and return with irqs disabled. This will | |
3560 | * protect us against recursive calling from irq. | |
3561 | */ | |
3562 | asmlinkage __visible void __sched preempt_schedule_irq(void) | |
3563 | { | |
3564 | enum ctx_state prev_state; | |
3565 | ||
3566 | /* Catch callers which need to be fixed */ | |
3567 | BUG_ON(preempt_count() || !irqs_disabled()); | |
3568 | ||
3569 | prev_state = exception_enter(); | |
3570 | ||
3571 | do { | |
3572 | preempt_disable(); | |
3573 | local_irq_enable(); | |
3574 | __schedule(true); | |
3575 | local_irq_disable(); | |
3576 | sched_preempt_enable_no_resched(); | |
3577 | } while (need_resched()); | |
3578 | ||
3579 | exception_exit(prev_state); | |
3580 | } | |
3581 | ||
3582 | int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags, | |
3583 | void *key) | |
3584 | { | |
3585 | return try_to_wake_up(curr->private, mode, wake_flags); | |
3586 | } | |
3587 | EXPORT_SYMBOL(default_wake_function); | |
3588 | ||
3589 | #ifdef CONFIG_RT_MUTEXES | |
3590 | ||
3591 | static inline int __rt_effective_prio(struct task_struct *pi_task, int prio) | |
3592 | { | |
3593 | if (pi_task) | |
3594 | prio = min(prio, pi_task->prio); | |
3595 | ||
3596 | return prio; | |
3597 | } | |
3598 | ||
3599 | static inline int rt_effective_prio(struct task_struct *p, int prio) | |
3600 | { | |
3601 | struct task_struct *pi_task = rt_mutex_get_top_task(p); | |
3602 | ||
3603 | return __rt_effective_prio(pi_task, prio); | |
3604 | } | |
3605 | ||
3606 | /* | |
3607 | * rt_mutex_setprio - set the current priority of a task | |
3608 | * @p: task to boost | |
3609 | * @pi_task: donor task | |
3610 | * | |
3611 | * This function changes the 'effective' priority of a task. It does | |
3612 | * not touch ->normal_prio like __setscheduler(). | |
3613 | * | |
3614 | * Used by the rt_mutex code to implement priority inheritance | |
3615 | * logic. Call site only calls if the priority of the task changed. | |
3616 | */ | |
3617 | void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task) | |
3618 | { | |
3619 | int prio, oldprio, queued, running, queue_flag = | |
3620 | DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK; | |
3621 | const struct sched_class *prev_class; | |
3622 | struct rq_flags rf; | |
3623 | struct rq *rq; | |
3624 | ||
3625 | /* XXX used to be waiter->prio, not waiter->task->prio */ | |
3626 | prio = __rt_effective_prio(pi_task, p->normal_prio); | |
3627 | ||
3628 | /* | |
3629 | * If nothing changed; bail early. | |
3630 | */ | |
3631 | if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio)) | |
3632 | return; | |
3633 | ||
3634 | rq = __task_rq_lock(p, &rf); | |
3635 | update_rq_clock(rq); | |
3636 | /* | |
3637 | * Set under pi_lock && rq->lock, such that the value can be used under | |
3638 | * either lock. | |
3639 | * | |
3640 | * Note that there is loads of tricky to make this pointer cache work | |
3641 | * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to | |
3642 | * ensure a task is de-boosted (pi_task is set to NULL) before the | |
3643 | * task is allowed to run again (and can exit). This ensures the pointer | |
3644 | * points to a blocked task -- which guaratees the task is present. | |
3645 | */ | |
3646 | p->pi_top_task = pi_task; | |
3647 | ||
3648 | /* | |
3649 | * For FIFO/RR we only need to set prio, if that matches we're done. | |
3650 | */ | |
3651 | if (prio == p->prio && !dl_prio(prio)) | |
3652 | goto out_unlock; | |
3653 | ||
3654 | /* | |
3655 | * Idle task boosting is a nono in general. There is one | |
3656 | * exception, when PREEMPT_RT and NOHZ is active: | |
3657 | * | |
3658 | * The idle task calls get_next_timer_interrupt() and holds | |
3659 | * the timer wheel base->lock on the CPU and another CPU wants | |
3660 | * to access the timer (probably to cancel it). We can safely | |
3661 | * ignore the boosting request, as the idle CPU runs this code | |
3662 | * with interrupts disabled and will complete the lock | |
3663 | * protected section without being interrupted. So there is no | |
3664 | * real need to boost. | |
3665 | */ | |
3666 | if (unlikely(p == rq->idle)) { | |
3667 | WARN_ON(p != rq->curr); | |
3668 | WARN_ON(p->pi_blocked_on); | |
3669 | goto out_unlock; | |
3670 | } | |
3671 | ||
3672 | trace_sched_pi_setprio(p, pi_task); | |
3673 | oldprio = p->prio; | |
3674 | ||
3675 | if (oldprio == prio) | |
3676 | queue_flag &= ~DEQUEUE_MOVE; | |
3677 | ||
3678 | prev_class = p->sched_class; | |
3679 | queued = task_on_rq_queued(p); | |
3680 | running = task_current(rq, p); | |
3681 | if (queued) | |
3682 | dequeue_task(rq, p, queue_flag); | |
3683 | if (running) | |
3684 | put_prev_task(rq, p); | |
3685 | ||
3686 | /* | |
3687 | * Boosting condition are: | |
3688 | * 1. -rt task is running and holds mutex A | |
3689 | * --> -dl task blocks on mutex A | |
3690 | * | |
3691 | * 2. -dl task is running and holds mutex A | |
3692 | * --> -dl task blocks on mutex A and could preempt the | |
3693 | * running task | |
3694 | */ | |
3695 | if (dl_prio(prio)) { | |
3696 | if (!dl_prio(p->normal_prio) || | |
3697 | (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) { | |
3698 | p->dl.dl_boosted = 1; | |
3699 | queue_flag |= ENQUEUE_REPLENISH; | |
3700 | } else | |
3701 | p->dl.dl_boosted = 0; | |
3702 | p->sched_class = &dl_sched_class; | |
3703 | } else if (rt_prio(prio)) { | |
3704 | if (dl_prio(oldprio)) | |
3705 | p->dl.dl_boosted = 0; | |
3706 | if (oldprio < prio) | |
3707 | queue_flag |= ENQUEUE_HEAD; | |
3708 | p->sched_class = &rt_sched_class; | |
3709 | } else { | |
3710 | if (dl_prio(oldprio)) | |
3711 | p->dl.dl_boosted = 0; | |
3712 | if (rt_prio(oldprio)) | |
3713 | p->rt.timeout = 0; | |
3714 | p->sched_class = &fair_sched_class; | |
3715 | } | |
3716 | ||
3717 | p->prio = prio; | |
3718 | ||
3719 | if (queued) | |
3720 | enqueue_task(rq, p, queue_flag); | |
3721 | if (running) | |
3722 | set_curr_task(rq, p); | |
3723 | ||
3724 | check_class_changed(rq, p, prev_class, oldprio); | |
3725 | out_unlock: | |
3726 | /* Avoid rq from going away on us: */ | |
3727 | preempt_disable(); | |
3728 | __task_rq_unlock(rq, &rf); | |
3729 | ||
3730 | balance_callback(rq); | |
3731 | preempt_enable(); | |
3732 | } | |
3733 | #else | |
3734 | static inline int rt_effective_prio(struct task_struct *p, int prio) | |
3735 | { | |
3736 | return prio; | |
3737 | } | |
3738 | #endif | |
3739 | ||
3740 | void set_user_nice(struct task_struct *p, long nice) | |
3741 | { | |
3742 | bool queued, running; | |
3743 | int old_prio, delta; | |
3744 | struct rq_flags rf; | |
3745 | struct rq *rq; | |
3746 | ||
3747 | if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) | |
3748 | return; | |
3749 | /* | |
3750 | * We have to be careful, if called from sys_setpriority(), | |
3751 | * the task might be in the middle of scheduling on another CPU. | |
3752 | */ | |
3753 | rq = task_rq_lock(p, &rf); | |
3754 | update_rq_clock(rq); | |
3755 | ||
3756 | /* | |
3757 | * The RT priorities are set via sched_setscheduler(), but we still | |
3758 | * allow the 'normal' nice value to be set - but as expected | |
3759 | * it wont have any effect on scheduling until the task is | |
3760 | * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: | |
3761 | */ | |
3762 | if (task_has_dl_policy(p) || task_has_rt_policy(p)) { | |
3763 | p->static_prio = NICE_TO_PRIO(nice); | |
3764 | goto out_unlock; | |
3765 | } | |
3766 | queued = task_on_rq_queued(p); | |
3767 | running = task_current(rq, p); | |
3768 | if (queued) | |
3769 | dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK); | |
3770 | if (running) | |
3771 | put_prev_task(rq, p); | |
3772 | ||
3773 | p->static_prio = NICE_TO_PRIO(nice); | |
3774 | set_load_weight(p); | |
3775 | old_prio = p->prio; | |
3776 | p->prio = effective_prio(p); | |
3777 | delta = p->prio - old_prio; | |
3778 | ||
3779 | if (queued) { | |
3780 | enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK); | |
3781 | /* | |
3782 | * If the task increased its priority or is running and | |
3783 | * lowered its priority, then reschedule its CPU: | |
3784 | */ | |
3785 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | |
3786 | resched_curr(rq); | |
3787 | } | |
3788 | if (running) | |
3789 | set_curr_task(rq, p); | |
3790 | out_unlock: | |
3791 | task_rq_unlock(rq, p, &rf); | |
3792 | } | |
3793 | EXPORT_SYMBOL(set_user_nice); | |
3794 | ||
3795 | /* | |
3796 | * can_nice - check if a task can reduce its nice value | |
3797 | * @p: task | |
3798 | * @nice: nice value | |
3799 | */ | |
3800 | int can_nice(const struct task_struct *p, const int nice) | |
3801 | { | |
3802 | /* Convert nice value [19,-20] to rlimit style value [1,40]: */ | |
3803 | int nice_rlim = nice_to_rlimit(nice); | |
3804 | ||
3805 | return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || | |
3806 | capable(CAP_SYS_NICE)); | |
3807 | } | |
3808 | ||
3809 | #ifdef __ARCH_WANT_SYS_NICE | |
3810 | ||
3811 | /* | |
3812 | * sys_nice - change the priority of the current process. | |
3813 | * @increment: priority increment | |
3814 | * | |
3815 | * sys_setpriority is a more generic, but much slower function that | |
3816 | * does similar things. | |
3817 | */ | |
3818 | SYSCALL_DEFINE1(nice, int, increment) | |
3819 | { | |
3820 | long nice, retval; | |
3821 | ||
3822 | /* | |
3823 | * Setpriority might change our priority at the same moment. | |
3824 | * We don't have to worry. Conceptually one call occurs first | |
3825 | * and we have a single winner. | |
3826 | */ | |
3827 | increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); | |
3828 | nice = task_nice(current) + increment; | |
3829 | ||
3830 | nice = clamp_val(nice, MIN_NICE, MAX_NICE); | |
3831 | if (increment < 0 && !can_nice(current, nice)) | |
3832 | return -EPERM; | |
3833 | ||
3834 | retval = security_task_setnice(current, nice); | |
3835 | if (retval) | |
3836 | return retval; | |
3837 | ||
3838 | set_user_nice(current, nice); | |
3839 | return 0; | |
3840 | } | |
3841 | ||
3842 | #endif | |
3843 | ||
3844 | /** | |
3845 | * task_prio - return the priority value of a given task. | |
3846 | * @p: the task in question. | |
3847 | * | |
3848 | * Return: The priority value as seen by users in /proc. | |
3849 | * RT tasks are offset by -200. Normal tasks are centered | |
3850 | * around 0, value goes from -16 to +15. | |
3851 | */ | |
3852 | int task_prio(const struct task_struct *p) | |
3853 | { | |
3854 | return p->prio - MAX_RT_PRIO; | |
3855 | } | |
3856 | ||
3857 | /** | |
3858 | * idle_cpu - is a given CPU idle currently? | |
3859 | * @cpu: the processor in question. | |
3860 | * | |
3861 | * Return: 1 if the CPU is currently idle. 0 otherwise. | |
3862 | */ | |
3863 | int idle_cpu(int cpu) | |
3864 | { | |
3865 | struct rq *rq = cpu_rq(cpu); | |
3866 | ||
3867 | if (rq->curr != rq->idle) | |
3868 | return 0; | |
3869 | ||
3870 | if (rq->nr_running) | |
3871 | return 0; | |
3872 | ||
3873 | #ifdef CONFIG_SMP | |
3874 | if (!llist_empty(&rq->wake_list)) | |
3875 | return 0; | |
3876 | #endif | |
3877 | ||
3878 | return 1; | |
3879 | } | |
3880 | ||
3881 | /** | |
3882 | * idle_task - return the idle task for a given CPU. | |
3883 | * @cpu: the processor in question. | |
3884 | * | |
3885 | * Return: The idle task for the CPU @cpu. | |
3886 | */ | |
3887 | struct task_struct *idle_task(int cpu) | |
3888 | { | |
3889 | return cpu_rq(cpu)->idle; | |
3890 | } | |
3891 | ||
3892 | /** | |
3893 | * find_process_by_pid - find a process with a matching PID value. | |
3894 | * @pid: the pid in question. | |
3895 | * | |
3896 | * The task of @pid, if found. %NULL otherwise. | |
3897 | */ | |
3898 | static struct task_struct *find_process_by_pid(pid_t pid) | |
3899 | { | |
3900 | return pid ? find_task_by_vpid(pid) : current; | |
3901 | } | |
3902 | ||
3903 | /* | |
3904 | * sched_setparam() passes in -1 for its policy, to let the functions | |
3905 | * it calls know not to change it. | |
3906 | */ | |
3907 | #define SETPARAM_POLICY -1 | |
3908 | ||
3909 | static void __setscheduler_params(struct task_struct *p, | |
3910 | const struct sched_attr *attr) | |
3911 | { | |
3912 | int policy = attr->sched_policy; | |
3913 | ||
3914 | if (policy == SETPARAM_POLICY) | |
3915 | policy = p->policy; | |
3916 | ||
3917 | p->policy = policy; | |
3918 | ||
3919 | if (dl_policy(policy)) | |
3920 | __setparam_dl(p, attr); | |
3921 | else if (fair_policy(policy)) | |
3922 | p->static_prio = NICE_TO_PRIO(attr->sched_nice); | |
3923 | ||
3924 | /* | |
3925 | * __sched_setscheduler() ensures attr->sched_priority == 0 when | |
3926 | * !rt_policy. Always setting this ensures that things like | |
3927 | * getparam()/getattr() don't report silly values for !rt tasks. | |
3928 | */ | |
3929 | p->rt_priority = attr->sched_priority; | |
3930 | p->normal_prio = normal_prio(p); | |
3931 | set_load_weight(p); | |
3932 | } | |
3933 | ||
3934 | /* Actually do priority change: must hold pi & rq lock. */ | |
3935 | static void __setscheduler(struct rq *rq, struct task_struct *p, | |
3936 | const struct sched_attr *attr, bool keep_boost) | |
3937 | { | |
3938 | __setscheduler_params(p, attr); | |
3939 | ||
3940 | /* | |
3941 | * Keep a potential priority boosting if called from | |
3942 | * sched_setscheduler(). | |
3943 | */ | |
3944 | p->prio = normal_prio(p); | |
3945 | if (keep_boost) | |
3946 | p->prio = rt_effective_prio(p, p->prio); | |
3947 | ||
3948 | if (dl_prio(p->prio)) | |
3949 | p->sched_class = &dl_sched_class; | |
3950 | else if (rt_prio(p->prio)) | |
3951 | p->sched_class = &rt_sched_class; | |
3952 | else | |
3953 | p->sched_class = &fair_sched_class; | |
3954 | } | |
3955 | ||
3956 | /* | |
3957 | * Check the target process has a UID that matches the current process's: | |
3958 | */ | |
3959 | static bool check_same_owner(struct task_struct *p) | |
3960 | { | |
3961 | const struct cred *cred = current_cred(), *pcred; | |
3962 | bool match; | |
3963 | ||
3964 | rcu_read_lock(); | |
3965 | pcred = __task_cred(p); | |
3966 | match = (uid_eq(cred->euid, pcred->euid) || | |
3967 | uid_eq(cred->euid, pcred->uid)); | |
3968 | rcu_read_unlock(); | |
3969 | return match; | |
3970 | } | |
3971 | ||
3972 | static int __sched_setscheduler(struct task_struct *p, | |
3973 | const struct sched_attr *attr, | |
3974 | bool user, bool pi) | |
3975 | { | |
3976 | int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : | |
3977 | MAX_RT_PRIO - 1 - attr->sched_priority; | |
3978 | int retval, oldprio, oldpolicy = -1, queued, running; | |
3979 | int new_effective_prio, policy = attr->sched_policy; | |
3980 | const struct sched_class *prev_class; | |
3981 | struct rq_flags rf; | |
3982 | int reset_on_fork; | |
3983 | int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK; | |
3984 | struct rq *rq; | |
3985 | ||
3986 | /* The pi code expects interrupts enabled */ | |
3987 | BUG_ON(pi && in_interrupt()); | |
3988 | recheck: | |
3989 | /* Double check policy once rq lock held: */ | |
3990 | if (policy < 0) { | |
3991 | reset_on_fork = p->sched_reset_on_fork; | |
3992 | policy = oldpolicy = p->policy; | |
3993 | } else { | |
3994 | reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); | |
3995 | ||
3996 | if (!valid_policy(policy)) | |
3997 | return -EINVAL; | |
3998 | } | |
3999 | ||
4000 | if (attr->sched_flags & | |
4001 | ~(SCHED_FLAG_RESET_ON_FORK | SCHED_FLAG_RECLAIM)) | |
4002 | return -EINVAL; | |
4003 | ||
4004 | /* | |
4005 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
4006 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, | |
4007 | * SCHED_BATCH and SCHED_IDLE is 0. | |
4008 | */ | |
4009 | if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || | |
4010 | (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) | |
4011 | return -EINVAL; | |
4012 | if ((dl_policy(policy) && !__checkparam_dl(attr)) || | |
4013 | (rt_policy(policy) != (attr->sched_priority != 0))) | |
4014 | return -EINVAL; | |
4015 | ||
4016 | /* | |
4017 | * Allow unprivileged RT tasks to decrease priority: | |
4018 | */ | |
4019 | if (user && !capable(CAP_SYS_NICE)) { | |
4020 | if (fair_policy(policy)) { | |
4021 | if (attr->sched_nice < task_nice(p) && | |
4022 | !can_nice(p, attr->sched_nice)) | |
4023 | return -EPERM; | |
4024 | } | |
4025 | ||
4026 | if (rt_policy(policy)) { | |
4027 | unsigned long rlim_rtprio = | |
4028 | task_rlimit(p, RLIMIT_RTPRIO); | |
4029 | ||
4030 | /* Can't set/change the rt policy: */ | |
4031 | if (policy != p->policy && !rlim_rtprio) | |
4032 | return -EPERM; | |
4033 | ||
4034 | /* Can't increase priority: */ | |
4035 | if (attr->sched_priority > p->rt_priority && | |
4036 | attr->sched_priority > rlim_rtprio) | |
4037 | return -EPERM; | |
4038 | } | |
4039 | ||
4040 | /* | |
4041 | * Can't set/change SCHED_DEADLINE policy at all for now | |
4042 | * (safest behavior); in the future we would like to allow | |
4043 | * unprivileged DL tasks to increase their relative deadline | |
4044 | * or reduce their runtime (both ways reducing utilization) | |
4045 | */ | |
4046 | if (dl_policy(policy)) | |
4047 | return -EPERM; | |
4048 | ||
4049 | /* | |
4050 | * Treat SCHED_IDLE as nice 20. Only allow a switch to | |
4051 | * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. | |
4052 | */ | |
4053 | if (idle_policy(p->policy) && !idle_policy(policy)) { | |
4054 | if (!can_nice(p, task_nice(p))) | |
4055 | return -EPERM; | |
4056 | } | |
4057 | ||
4058 | /* Can't change other user's priorities: */ | |
4059 | if (!check_same_owner(p)) | |
4060 | return -EPERM; | |
4061 | ||
4062 | /* Normal users shall not reset the sched_reset_on_fork flag: */ | |
4063 | if (p->sched_reset_on_fork && !reset_on_fork) | |
4064 | return -EPERM; | |
4065 | } | |
4066 | ||
4067 | if (user) { | |
4068 | retval = security_task_setscheduler(p); | |
4069 | if (retval) | |
4070 | return retval; | |
4071 | } | |
4072 | ||
4073 | /* | |
4074 | * Make sure no PI-waiters arrive (or leave) while we are | |
4075 | * changing the priority of the task: | |
4076 | * | |
4077 | * To be able to change p->policy safely, the appropriate | |
4078 | * runqueue lock must be held. | |
4079 | */ | |
4080 | rq = task_rq_lock(p, &rf); | |
4081 | update_rq_clock(rq); | |
4082 | ||
4083 | /* | |
4084 | * Changing the policy of the stop threads its a very bad idea: | |
4085 | */ | |
4086 | if (p == rq->stop) { | |
4087 | task_rq_unlock(rq, p, &rf); | |
4088 | return -EINVAL; | |
4089 | } | |
4090 | ||
4091 | /* | |
4092 | * If not changing anything there's no need to proceed further, | |
4093 | * but store a possible modification of reset_on_fork. | |
4094 | */ | |
4095 | if (unlikely(policy == p->policy)) { | |
4096 | if (fair_policy(policy) && attr->sched_nice != task_nice(p)) | |
4097 | goto change; | |
4098 | if (rt_policy(policy) && attr->sched_priority != p->rt_priority) | |
4099 | goto change; | |
4100 | if (dl_policy(policy) && dl_param_changed(p, attr)) | |
4101 | goto change; | |
4102 | ||
4103 | p->sched_reset_on_fork = reset_on_fork; | |
4104 | task_rq_unlock(rq, p, &rf); | |
4105 | return 0; | |
4106 | } | |
4107 | change: | |
4108 | ||
4109 | if (user) { | |
4110 | #ifdef CONFIG_RT_GROUP_SCHED | |
4111 | /* | |
4112 | * Do not allow realtime tasks into groups that have no runtime | |
4113 | * assigned. | |
4114 | */ | |
4115 | if (rt_bandwidth_enabled() && rt_policy(policy) && | |
4116 | task_group(p)->rt_bandwidth.rt_runtime == 0 && | |
4117 | !task_group_is_autogroup(task_group(p))) { | |
4118 | task_rq_unlock(rq, p, &rf); | |
4119 | return -EPERM; | |
4120 | } | |
4121 | #endif | |
4122 | #ifdef CONFIG_SMP | |
4123 | if (dl_bandwidth_enabled() && dl_policy(policy)) { | |
4124 | cpumask_t *span = rq->rd->span; | |
4125 | ||
4126 | /* | |
4127 | * Don't allow tasks with an affinity mask smaller than | |
4128 | * the entire root_domain to become SCHED_DEADLINE. We | |
4129 | * will also fail if there's no bandwidth available. | |
4130 | */ | |
4131 | if (!cpumask_subset(span, &p->cpus_allowed) || | |
4132 | rq->rd->dl_bw.bw == 0) { | |
4133 | task_rq_unlock(rq, p, &rf); | |
4134 | return -EPERM; | |
4135 | } | |
4136 | } | |
4137 | #endif | |
4138 | } | |
4139 | ||
4140 | /* Re-check policy now with rq lock held: */ | |
4141 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
4142 | policy = oldpolicy = -1; | |
4143 | task_rq_unlock(rq, p, &rf); | |
4144 | goto recheck; | |
4145 | } | |
4146 | ||
4147 | /* | |
4148 | * If setscheduling to SCHED_DEADLINE (or changing the parameters | |
4149 | * of a SCHED_DEADLINE task) we need to check if enough bandwidth | |
4150 | * is available. | |
4151 | */ | |
4152 | if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) { | |
4153 | task_rq_unlock(rq, p, &rf); | |
4154 | return -EBUSY; | |
4155 | } | |
4156 | ||
4157 | p->sched_reset_on_fork = reset_on_fork; | |
4158 | oldprio = p->prio; | |
4159 | ||
4160 | if (pi) { | |
4161 | /* | |
4162 | * Take priority boosted tasks into account. If the new | |
4163 | * effective priority is unchanged, we just store the new | |
4164 | * normal parameters and do not touch the scheduler class and | |
4165 | * the runqueue. This will be done when the task deboost | |
4166 | * itself. | |
4167 | */ | |
4168 | new_effective_prio = rt_effective_prio(p, newprio); | |
4169 | if (new_effective_prio == oldprio) | |
4170 | queue_flags &= ~DEQUEUE_MOVE; | |
4171 | } | |
4172 | ||
4173 | queued = task_on_rq_queued(p); | |
4174 | running = task_current(rq, p); | |
4175 | if (queued) | |
4176 | dequeue_task(rq, p, queue_flags); | |
4177 | if (running) | |
4178 | put_prev_task(rq, p); | |
4179 | ||
4180 | prev_class = p->sched_class; | |
4181 | __setscheduler(rq, p, attr, pi); | |
4182 | ||
4183 | if (queued) { | |
4184 | /* | |
4185 | * We enqueue to tail when the priority of a task is | |
4186 | * increased (user space view). | |
4187 | */ | |
4188 | if (oldprio < p->prio) | |
4189 | queue_flags |= ENQUEUE_HEAD; | |
4190 | ||
4191 | enqueue_task(rq, p, queue_flags); | |
4192 | } | |
4193 | if (running) | |
4194 | set_curr_task(rq, p); | |
4195 | ||
4196 | check_class_changed(rq, p, prev_class, oldprio); | |
4197 | ||
4198 | /* Avoid rq from going away on us: */ | |
4199 | preempt_disable(); | |
4200 | task_rq_unlock(rq, p, &rf); | |
4201 | ||
4202 | if (pi) | |
4203 | rt_mutex_adjust_pi(p); | |
4204 | ||
4205 | /* Run balance callbacks after we've adjusted the PI chain: */ | |
4206 | balance_callback(rq); | |
4207 | preempt_enable(); | |
4208 | ||
4209 | return 0; | |
4210 | } | |
4211 | ||
4212 | static int _sched_setscheduler(struct task_struct *p, int policy, | |
4213 | const struct sched_param *param, bool check) | |
4214 | { | |
4215 | struct sched_attr attr = { | |
4216 | .sched_policy = policy, | |
4217 | .sched_priority = param->sched_priority, | |
4218 | .sched_nice = PRIO_TO_NICE(p->static_prio), | |
4219 | }; | |
4220 | ||
4221 | /* Fixup the legacy SCHED_RESET_ON_FORK hack. */ | |
4222 | if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) { | |
4223 | attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; | |
4224 | policy &= ~SCHED_RESET_ON_FORK; | |
4225 | attr.sched_policy = policy; | |
4226 | } | |
4227 | ||
4228 | return __sched_setscheduler(p, &attr, check, true); | |
4229 | } | |
4230 | /** | |
4231 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. | |
4232 | * @p: the task in question. | |
4233 | * @policy: new policy. | |
4234 | * @param: structure containing the new RT priority. | |
4235 | * | |
4236 | * Return: 0 on success. An error code otherwise. | |
4237 | * | |
4238 | * NOTE that the task may be already dead. | |
4239 | */ | |
4240 | int sched_setscheduler(struct task_struct *p, int policy, | |
4241 | const struct sched_param *param) | |
4242 | { | |
4243 | return _sched_setscheduler(p, policy, param, true); | |
4244 | } | |
4245 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
4246 | ||
4247 | int sched_setattr(struct task_struct *p, const struct sched_attr *attr) | |
4248 | { | |
4249 | return __sched_setscheduler(p, attr, true, true); | |
4250 | } | |
4251 | EXPORT_SYMBOL_GPL(sched_setattr); | |
4252 | ||
4253 | /** | |
4254 | * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. | |
4255 | * @p: the task in question. | |
4256 | * @policy: new policy. | |
4257 | * @param: structure containing the new RT priority. | |
4258 | * | |
4259 | * Just like sched_setscheduler, only don't bother checking if the | |
4260 | * current context has permission. For example, this is needed in | |
4261 | * stop_machine(): we create temporary high priority worker threads, | |
4262 | * but our caller might not have that capability. | |
4263 | * | |
4264 | * Return: 0 on success. An error code otherwise. | |
4265 | */ | |
4266 | int sched_setscheduler_nocheck(struct task_struct *p, int policy, | |
4267 | const struct sched_param *param) | |
4268 | { | |
4269 | return _sched_setscheduler(p, policy, param, false); | |
4270 | } | |
4271 | EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck); | |
4272 | ||
4273 | static int | |
4274 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
4275 | { | |
4276 | struct sched_param lparam; | |
4277 | struct task_struct *p; | |
4278 | int retval; | |
4279 | ||
4280 | if (!param || pid < 0) | |
4281 | return -EINVAL; | |
4282 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
4283 | return -EFAULT; | |
4284 | ||
4285 | rcu_read_lock(); | |
4286 | retval = -ESRCH; | |
4287 | p = find_process_by_pid(pid); | |
4288 | if (p != NULL) | |
4289 | retval = sched_setscheduler(p, policy, &lparam); | |
4290 | rcu_read_unlock(); | |
4291 | ||
4292 | return retval; | |
4293 | } | |
4294 | ||
4295 | /* | |
4296 | * Mimics kernel/events/core.c perf_copy_attr(). | |
4297 | */ | |
4298 | static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr) | |
4299 | { | |
4300 | u32 size; | |
4301 | int ret; | |
4302 | ||
4303 | if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) | |
4304 | return -EFAULT; | |
4305 | ||
4306 | /* Zero the full structure, so that a short copy will be nice: */ | |
4307 | memset(attr, 0, sizeof(*attr)); | |
4308 | ||
4309 | ret = get_user(size, &uattr->size); | |
4310 | if (ret) | |
4311 | return ret; | |
4312 | ||
4313 | /* Bail out on silly large: */ | |
4314 | if (size > PAGE_SIZE) | |
4315 | goto err_size; | |
4316 | ||
4317 | /* ABI compatibility quirk: */ | |
4318 | if (!size) | |
4319 | size = SCHED_ATTR_SIZE_VER0; | |
4320 | ||
4321 | if (size < SCHED_ATTR_SIZE_VER0) | |
4322 | goto err_size; | |
4323 | ||
4324 | /* | |
4325 | * If we're handed a bigger struct than we know of, | |
4326 | * ensure all the unknown bits are 0 - i.e. new | |
4327 | * user-space does not rely on any kernel feature | |
4328 | * extensions we dont know about yet. | |
4329 | */ | |
4330 | if (size > sizeof(*attr)) { | |
4331 | unsigned char __user *addr; | |
4332 | unsigned char __user *end; | |
4333 | unsigned char val; | |
4334 | ||
4335 | addr = (void __user *)uattr + sizeof(*attr); | |
4336 | end = (void __user *)uattr + size; | |
4337 | ||
4338 | for (; addr < end; addr++) { | |
4339 | ret = get_user(val, addr); | |
4340 | if (ret) | |
4341 | return ret; | |
4342 | if (val) | |
4343 | goto err_size; | |
4344 | } | |
4345 | size = sizeof(*attr); | |
4346 | } | |
4347 | ||
4348 | ret = copy_from_user(attr, uattr, size); | |
4349 | if (ret) | |
4350 | return -EFAULT; | |
4351 | ||
4352 | /* | |
4353 | * XXX: Do we want to be lenient like existing syscalls; or do we want | |
4354 | * to be strict and return an error on out-of-bounds values? | |
4355 | */ | |
4356 | attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); | |
4357 | ||
4358 | return 0; | |
4359 | ||
4360 | err_size: | |
4361 | put_user(sizeof(*attr), &uattr->size); | |
4362 | return -E2BIG; | |
4363 | } | |
4364 | ||
4365 | /** | |
4366 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
4367 | * @pid: the pid in question. | |
4368 | * @policy: new policy. | |
4369 | * @param: structure containing the new RT priority. | |
4370 | * | |
4371 | * Return: 0 on success. An error code otherwise. | |
4372 | */ | |
4373 | SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param) | |
4374 | { | |
4375 | if (policy < 0) | |
4376 | return -EINVAL; | |
4377 | ||
4378 | return do_sched_setscheduler(pid, policy, param); | |
4379 | } | |
4380 | ||
4381 | /** | |
4382 | * sys_sched_setparam - set/change the RT priority of a thread | |
4383 | * @pid: the pid in question. | |
4384 | * @param: structure containing the new RT priority. | |
4385 | * | |
4386 | * Return: 0 on success. An error code otherwise. | |
4387 | */ | |
4388 | SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) | |
4389 | { | |
4390 | return do_sched_setscheduler(pid, SETPARAM_POLICY, param); | |
4391 | } | |
4392 | ||
4393 | /** | |
4394 | * sys_sched_setattr - same as above, but with extended sched_attr | |
4395 | * @pid: the pid in question. | |
4396 | * @uattr: structure containing the extended parameters. | |
4397 | * @flags: for future extension. | |
4398 | */ | |
4399 | SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, | |
4400 | unsigned int, flags) | |
4401 | { | |
4402 | struct sched_attr attr; | |
4403 | struct task_struct *p; | |
4404 | int retval; | |
4405 | ||
4406 | if (!uattr || pid < 0 || flags) | |
4407 | return -EINVAL; | |
4408 | ||
4409 | retval = sched_copy_attr(uattr, &attr); | |
4410 | if (retval) | |
4411 | return retval; | |
4412 | ||
4413 | if ((int)attr.sched_policy < 0) | |
4414 | return -EINVAL; | |
4415 | ||
4416 | rcu_read_lock(); | |
4417 | retval = -ESRCH; | |
4418 | p = find_process_by_pid(pid); | |
4419 | if (p != NULL) | |
4420 | retval = sched_setattr(p, &attr); | |
4421 | rcu_read_unlock(); | |
4422 | ||
4423 | return retval; | |
4424 | } | |
4425 | ||
4426 | /** | |
4427 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
4428 | * @pid: the pid in question. | |
4429 | * | |
4430 | * Return: On success, the policy of the thread. Otherwise, a negative error | |
4431 | * code. | |
4432 | */ | |
4433 | SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) | |
4434 | { | |
4435 | struct task_struct *p; | |
4436 | int retval; | |
4437 | ||
4438 | if (pid < 0) | |
4439 | return -EINVAL; | |
4440 | ||
4441 | retval = -ESRCH; | |
4442 | rcu_read_lock(); | |
4443 | p = find_process_by_pid(pid); | |
4444 | if (p) { | |
4445 | retval = security_task_getscheduler(p); | |
4446 | if (!retval) | |
4447 | retval = p->policy | |
4448 | | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); | |
4449 | } | |
4450 | rcu_read_unlock(); | |
4451 | return retval; | |
4452 | } | |
4453 | ||
4454 | /** | |
4455 | * sys_sched_getparam - get the RT priority of a thread | |
4456 | * @pid: the pid in question. | |
4457 | * @param: structure containing the RT priority. | |
4458 | * | |
4459 | * Return: On success, 0 and the RT priority is in @param. Otherwise, an error | |
4460 | * code. | |
4461 | */ | |
4462 | SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) | |
4463 | { | |
4464 | struct sched_param lp = { .sched_priority = 0 }; | |
4465 | struct task_struct *p; | |
4466 | int retval; | |
4467 | ||
4468 | if (!param || pid < 0) | |
4469 | return -EINVAL; | |
4470 | ||
4471 | rcu_read_lock(); | |
4472 | p = find_process_by_pid(pid); | |
4473 | retval = -ESRCH; | |
4474 | if (!p) | |
4475 | goto out_unlock; | |
4476 | ||
4477 | retval = security_task_getscheduler(p); | |
4478 | if (retval) | |
4479 | goto out_unlock; | |
4480 | ||
4481 | if (task_has_rt_policy(p)) | |
4482 | lp.sched_priority = p->rt_priority; | |
4483 | rcu_read_unlock(); | |
4484 | ||
4485 | /* | |
4486 | * This one might sleep, we cannot do it with a spinlock held ... | |
4487 | */ | |
4488 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
4489 | ||
4490 | return retval; | |
4491 | ||
4492 | out_unlock: | |
4493 | rcu_read_unlock(); | |
4494 | return retval; | |
4495 | } | |
4496 | ||
4497 | static int sched_read_attr(struct sched_attr __user *uattr, | |
4498 | struct sched_attr *attr, | |
4499 | unsigned int usize) | |
4500 | { | |
4501 | int ret; | |
4502 | ||
4503 | if (!access_ok(VERIFY_WRITE, uattr, usize)) | |
4504 | return -EFAULT; | |
4505 | ||
4506 | /* | |
4507 | * If we're handed a smaller struct than we know of, | |
4508 | * ensure all the unknown bits are 0 - i.e. old | |
4509 | * user-space does not get uncomplete information. | |
4510 | */ | |
4511 | if (usize < sizeof(*attr)) { | |
4512 | unsigned char *addr; | |
4513 | unsigned char *end; | |
4514 | ||
4515 | addr = (void *)attr + usize; | |
4516 | end = (void *)attr + sizeof(*attr); | |
4517 | ||
4518 | for (; addr < end; addr++) { | |
4519 | if (*addr) | |
4520 | return -EFBIG; | |
4521 | } | |
4522 | ||
4523 | attr->size = usize; | |
4524 | } | |
4525 | ||
4526 | ret = copy_to_user(uattr, attr, attr->size); | |
4527 | if (ret) | |
4528 | return -EFAULT; | |
4529 | ||
4530 | return 0; | |
4531 | } | |
4532 | ||
4533 | /** | |
4534 | * sys_sched_getattr - similar to sched_getparam, but with sched_attr | |
4535 | * @pid: the pid in question. | |
4536 | * @uattr: structure containing the extended parameters. | |
4537 | * @size: sizeof(attr) for fwd/bwd comp. | |
4538 | * @flags: for future extension. | |
4539 | */ | |
4540 | SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, | |
4541 | unsigned int, size, unsigned int, flags) | |
4542 | { | |
4543 | struct sched_attr attr = { | |
4544 | .size = sizeof(struct sched_attr), | |
4545 | }; | |
4546 | struct task_struct *p; | |
4547 | int retval; | |
4548 | ||
4549 | if (!uattr || pid < 0 || size > PAGE_SIZE || | |
4550 | size < SCHED_ATTR_SIZE_VER0 || flags) | |
4551 | return -EINVAL; | |
4552 | ||
4553 | rcu_read_lock(); | |
4554 | p = find_process_by_pid(pid); | |
4555 | retval = -ESRCH; | |
4556 | if (!p) | |
4557 | goto out_unlock; | |
4558 | ||
4559 | retval = security_task_getscheduler(p); | |
4560 | if (retval) | |
4561 | goto out_unlock; | |
4562 | ||
4563 | attr.sched_policy = p->policy; | |
4564 | if (p->sched_reset_on_fork) | |
4565 | attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; | |
4566 | if (task_has_dl_policy(p)) | |
4567 | __getparam_dl(p, &attr); | |
4568 | else if (task_has_rt_policy(p)) | |
4569 | attr.sched_priority = p->rt_priority; | |
4570 | else | |
4571 | attr.sched_nice = task_nice(p); | |
4572 | ||
4573 | rcu_read_unlock(); | |
4574 | ||
4575 | retval = sched_read_attr(uattr, &attr, size); | |
4576 | return retval; | |
4577 | ||
4578 | out_unlock: | |
4579 | rcu_read_unlock(); | |
4580 | return retval; | |
4581 | } | |
4582 | ||
4583 | long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) | |
4584 | { | |
4585 | cpumask_var_t cpus_allowed, new_mask; | |
4586 | struct task_struct *p; | |
4587 | int retval; | |
4588 | ||
4589 | rcu_read_lock(); | |
4590 | ||
4591 | p = find_process_by_pid(pid); | |
4592 | if (!p) { | |
4593 | rcu_read_unlock(); | |
4594 | return -ESRCH; | |
4595 | } | |
4596 | ||
4597 | /* Prevent p going away */ | |
4598 | get_task_struct(p); | |
4599 | rcu_read_unlock(); | |
4600 | ||
4601 | if (p->flags & PF_NO_SETAFFINITY) { | |
4602 | retval = -EINVAL; | |
4603 | goto out_put_task; | |
4604 | } | |
4605 | if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { | |
4606 | retval = -ENOMEM; | |
4607 | goto out_put_task; | |
4608 | } | |
4609 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { | |
4610 | retval = -ENOMEM; | |
4611 | goto out_free_cpus_allowed; | |
4612 | } | |
4613 | retval = -EPERM; | |
4614 | if (!check_same_owner(p)) { | |
4615 | rcu_read_lock(); | |
4616 | if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { | |
4617 | rcu_read_unlock(); | |
4618 | goto out_free_new_mask; | |
4619 | } | |
4620 | rcu_read_unlock(); | |
4621 | } | |
4622 | ||
4623 | retval = security_task_setscheduler(p); | |
4624 | if (retval) | |
4625 | goto out_free_new_mask; | |
4626 | ||
4627 | ||
4628 | cpuset_cpus_allowed(p, cpus_allowed); | |
4629 | cpumask_and(new_mask, in_mask, cpus_allowed); | |
4630 | ||
4631 | /* | |
4632 | * Since bandwidth control happens on root_domain basis, | |
4633 | * if admission test is enabled, we only admit -deadline | |
4634 | * tasks allowed to run on all the CPUs in the task's | |
4635 | * root_domain. | |
4636 | */ | |
4637 | #ifdef CONFIG_SMP | |
4638 | if (task_has_dl_policy(p) && dl_bandwidth_enabled()) { | |
4639 | rcu_read_lock(); | |
4640 | if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) { | |
4641 | retval = -EBUSY; | |
4642 | rcu_read_unlock(); | |
4643 | goto out_free_new_mask; | |
4644 | } | |
4645 | rcu_read_unlock(); | |
4646 | } | |
4647 | #endif | |
4648 | again: | |
4649 | retval = __set_cpus_allowed_ptr(p, new_mask, true); | |
4650 | ||
4651 | if (!retval) { | |
4652 | cpuset_cpus_allowed(p, cpus_allowed); | |
4653 | if (!cpumask_subset(new_mask, cpus_allowed)) { | |
4654 | /* | |
4655 | * We must have raced with a concurrent cpuset | |
4656 | * update. Just reset the cpus_allowed to the | |
4657 | * cpuset's cpus_allowed | |
4658 | */ | |
4659 | cpumask_copy(new_mask, cpus_allowed); | |
4660 | goto again; | |
4661 | } | |
4662 | } | |
4663 | out_free_new_mask: | |
4664 | free_cpumask_var(new_mask); | |
4665 | out_free_cpus_allowed: | |
4666 | free_cpumask_var(cpus_allowed); | |
4667 | out_put_task: | |
4668 | put_task_struct(p); | |
4669 | return retval; | |
4670 | } | |
4671 | ||
4672 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
4673 | struct cpumask *new_mask) | |
4674 | { | |
4675 | if (len < cpumask_size()) | |
4676 | cpumask_clear(new_mask); | |
4677 | else if (len > cpumask_size()) | |
4678 | len = cpumask_size(); | |
4679 | ||
4680 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
4681 | } | |
4682 | ||
4683 | /** | |
4684 | * sys_sched_setaffinity - set the CPU affinity of a process | |
4685 | * @pid: pid of the process | |
4686 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4687 | * @user_mask_ptr: user-space pointer to the new CPU mask | |
4688 | * | |
4689 | * Return: 0 on success. An error code otherwise. | |
4690 | */ | |
4691 | SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, | |
4692 | unsigned long __user *, user_mask_ptr) | |
4693 | { | |
4694 | cpumask_var_t new_mask; | |
4695 | int retval; | |
4696 | ||
4697 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
4698 | return -ENOMEM; | |
4699 | ||
4700 | retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); | |
4701 | if (retval == 0) | |
4702 | retval = sched_setaffinity(pid, new_mask); | |
4703 | free_cpumask_var(new_mask); | |
4704 | return retval; | |
4705 | } | |
4706 | ||
4707 | long sched_getaffinity(pid_t pid, struct cpumask *mask) | |
4708 | { | |
4709 | struct task_struct *p; | |
4710 | unsigned long flags; | |
4711 | int retval; | |
4712 | ||
4713 | rcu_read_lock(); | |
4714 | ||
4715 | retval = -ESRCH; | |
4716 | p = find_process_by_pid(pid); | |
4717 | if (!p) | |
4718 | goto out_unlock; | |
4719 | ||
4720 | retval = security_task_getscheduler(p); | |
4721 | if (retval) | |
4722 | goto out_unlock; | |
4723 | ||
4724 | raw_spin_lock_irqsave(&p->pi_lock, flags); | |
4725 | cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); | |
4726 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | |
4727 | ||
4728 | out_unlock: | |
4729 | rcu_read_unlock(); | |
4730 | ||
4731 | return retval; | |
4732 | } | |
4733 | ||
4734 | /** | |
4735 | * sys_sched_getaffinity - get the CPU affinity of a process | |
4736 | * @pid: pid of the process | |
4737 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
4738 | * @user_mask_ptr: user-space pointer to hold the current CPU mask | |
4739 | * | |
4740 | * Return: size of CPU mask copied to user_mask_ptr on success. An | |
4741 | * error code otherwise. | |
4742 | */ | |
4743 | SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, | |
4744 | unsigned long __user *, user_mask_ptr) | |
4745 | { | |
4746 | int ret; | |
4747 | cpumask_var_t mask; | |
4748 | ||
4749 | if ((len * BITS_PER_BYTE) < nr_cpu_ids) | |
4750 | return -EINVAL; | |
4751 | if (len & (sizeof(unsigned long)-1)) | |
4752 | return -EINVAL; | |
4753 | ||
4754 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) | |
4755 | return -ENOMEM; | |
4756 | ||
4757 | ret = sched_getaffinity(pid, mask); | |
4758 | if (ret == 0) { | |
4759 | size_t retlen = min_t(size_t, len, cpumask_size()); | |
4760 | ||
4761 | if (copy_to_user(user_mask_ptr, mask, retlen)) | |
4762 | ret = -EFAULT; | |
4763 | else | |
4764 | ret = retlen; | |
4765 | } | |
4766 | free_cpumask_var(mask); | |
4767 | ||
4768 | return ret; | |
4769 | } | |
4770 | ||
4771 | /** | |
4772 | * sys_sched_yield - yield the current processor to other threads. | |
4773 | * | |
4774 | * This function yields the current CPU to other tasks. If there are no | |
4775 | * other threads running on this CPU then this function will return. | |
4776 | * | |
4777 | * Return: 0. | |
4778 | */ | |
4779 | SYSCALL_DEFINE0(sched_yield) | |
4780 | { | |
4781 | struct rq_flags rf; | |
4782 | struct rq *rq; | |
4783 | ||
4784 | local_irq_disable(); | |
4785 | rq = this_rq(); | |
4786 | rq_lock(rq, &rf); | |
4787 | ||
4788 | schedstat_inc(rq->yld_count); | |
4789 | current->sched_class->yield_task(rq); | |
4790 | ||
4791 | /* | |
4792 | * Since we are going to call schedule() anyway, there's | |
4793 | * no need to preempt or enable interrupts: | |
4794 | */ | |
4795 | preempt_disable(); | |
4796 | rq_unlock(rq, &rf); | |
4797 | sched_preempt_enable_no_resched(); | |
4798 | ||
4799 | schedule(); | |
4800 | ||
4801 | return 0; | |
4802 | } | |
4803 | ||
4804 | #ifndef CONFIG_PREEMPT | |
4805 | int __sched _cond_resched(void) | |
4806 | { | |
4807 | if (should_resched(0)) { | |
4808 | preempt_schedule_common(); | |
4809 | return 1; | |
4810 | } | |
4811 | return 0; | |
4812 | } | |
4813 | EXPORT_SYMBOL(_cond_resched); | |
4814 | #endif | |
4815 | ||
4816 | /* | |
4817 | * __cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
4818 | * call schedule, and on return reacquire the lock. | |
4819 | * | |
4820 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | |
4821 | * operations here to prevent schedule() from being called twice (once via | |
4822 | * spin_unlock(), once by hand). | |
4823 | */ | |
4824 | int __cond_resched_lock(spinlock_t *lock) | |
4825 | { | |
4826 | int resched = should_resched(PREEMPT_LOCK_OFFSET); | |
4827 | int ret = 0; | |
4828 | ||
4829 | lockdep_assert_held(lock); | |
4830 | ||
4831 | if (spin_needbreak(lock) || resched) { | |
4832 | spin_unlock(lock); | |
4833 | if (resched) | |
4834 | preempt_schedule_common(); | |
4835 | else | |
4836 | cpu_relax(); | |
4837 | ret = 1; | |
4838 | spin_lock(lock); | |
4839 | } | |
4840 | return ret; | |
4841 | } | |
4842 | EXPORT_SYMBOL(__cond_resched_lock); | |
4843 | ||
4844 | int __sched __cond_resched_softirq(void) | |
4845 | { | |
4846 | BUG_ON(!in_softirq()); | |
4847 | ||
4848 | if (should_resched(SOFTIRQ_DISABLE_OFFSET)) { | |
4849 | local_bh_enable(); | |
4850 | preempt_schedule_common(); | |
4851 | local_bh_disable(); | |
4852 | return 1; | |
4853 | } | |
4854 | return 0; | |
4855 | } | |
4856 | EXPORT_SYMBOL(__cond_resched_softirq); | |
4857 | ||
4858 | /** | |
4859 | * yield - yield the current processor to other threads. | |
4860 | * | |
4861 | * Do not ever use this function, there's a 99% chance you're doing it wrong. | |
4862 | * | |
4863 | * The scheduler is at all times free to pick the calling task as the most | |
4864 | * eligible task to run, if removing the yield() call from your code breaks | |
4865 | * it, its already broken. | |
4866 | * | |
4867 | * Typical broken usage is: | |
4868 | * | |
4869 | * while (!event) | |
4870 | * yield(); | |
4871 | * | |
4872 | * where one assumes that yield() will let 'the other' process run that will | |
4873 | * make event true. If the current task is a SCHED_FIFO task that will never | |
4874 | * happen. Never use yield() as a progress guarantee!! | |
4875 | * | |
4876 | * If you want to use yield() to wait for something, use wait_event(). | |
4877 | * If you want to use yield() to be 'nice' for others, use cond_resched(). | |
4878 | * If you still want to use yield(), do not! | |
4879 | */ | |
4880 | void __sched yield(void) | |
4881 | { | |
4882 | set_current_state(TASK_RUNNING); | |
4883 | sys_sched_yield(); | |
4884 | } | |
4885 | EXPORT_SYMBOL(yield); | |
4886 | ||
4887 | /** | |
4888 | * yield_to - yield the current processor to another thread in | |
4889 | * your thread group, or accelerate that thread toward the | |
4890 | * processor it's on. | |
4891 | * @p: target task | |
4892 | * @preempt: whether task preemption is allowed or not | |
4893 | * | |
4894 | * It's the caller's job to ensure that the target task struct | |
4895 | * can't go away on us before we can do any checks. | |
4896 | * | |
4897 | * Return: | |
4898 | * true (>0) if we indeed boosted the target task. | |
4899 | * false (0) if we failed to boost the target. | |
4900 | * -ESRCH if there's no task to yield to. | |
4901 | */ | |
4902 | int __sched yield_to(struct task_struct *p, bool preempt) | |
4903 | { | |
4904 | struct task_struct *curr = current; | |
4905 | struct rq *rq, *p_rq; | |
4906 | unsigned long flags; | |
4907 | int yielded = 0; | |
4908 | ||
4909 | local_irq_save(flags); | |
4910 | rq = this_rq(); | |
4911 | ||
4912 | again: | |
4913 | p_rq = task_rq(p); | |
4914 | /* | |
4915 | * If we're the only runnable task on the rq and target rq also | |
4916 | * has only one task, there's absolutely no point in yielding. | |
4917 | */ | |
4918 | if (rq->nr_running == 1 && p_rq->nr_running == 1) { | |
4919 | yielded = -ESRCH; | |
4920 | goto out_irq; | |
4921 | } | |
4922 | ||
4923 | double_rq_lock(rq, p_rq); | |
4924 | if (task_rq(p) != p_rq) { | |
4925 | double_rq_unlock(rq, p_rq); | |
4926 | goto again; | |
4927 | } | |
4928 | ||
4929 | if (!curr->sched_class->yield_to_task) | |
4930 | goto out_unlock; | |
4931 | ||
4932 | if (curr->sched_class != p->sched_class) | |
4933 | goto out_unlock; | |
4934 | ||
4935 | if (task_running(p_rq, p) || p->state) | |
4936 | goto out_unlock; | |
4937 | ||
4938 | yielded = curr->sched_class->yield_to_task(rq, p, preempt); | |
4939 | if (yielded) { | |
4940 | schedstat_inc(rq->yld_count); | |
4941 | /* | |
4942 | * Make p's CPU reschedule; pick_next_entity takes care of | |
4943 | * fairness. | |
4944 | */ | |
4945 | if (preempt && rq != p_rq) | |
4946 | resched_curr(p_rq); | |
4947 | } | |
4948 | ||
4949 | out_unlock: | |
4950 | double_rq_unlock(rq, p_rq); | |
4951 | out_irq: | |
4952 | local_irq_restore(flags); | |
4953 | ||
4954 | if (yielded > 0) | |
4955 | schedule(); | |
4956 | ||
4957 | return yielded; | |
4958 | } | |
4959 | EXPORT_SYMBOL_GPL(yield_to); | |
4960 | ||
4961 | int io_schedule_prepare(void) | |
4962 | { | |
4963 | int old_iowait = current->in_iowait; | |
4964 | ||
4965 | current->in_iowait = 1; | |
4966 | blk_schedule_flush_plug(current); | |
4967 | ||
4968 | return old_iowait; | |
4969 | } | |
4970 | ||
4971 | void io_schedule_finish(int token) | |
4972 | { | |
4973 | current->in_iowait = token; | |
4974 | } | |
4975 | ||
4976 | /* | |
4977 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | |
4978 | * that process accounting knows that this is a task in IO wait state. | |
4979 | */ | |
4980 | long __sched io_schedule_timeout(long timeout) | |
4981 | { | |
4982 | int token; | |
4983 | long ret; | |
4984 | ||
4985 | token = io_schedule_prepare(); | |
4986 | ret = schedule_timeout(timeout); | |
4987 | io_schedule_finish(token); | |
4988 | ||
4989 | return ret; | |
4990 | } | |
4991 | EXPORT_SYMBOL(io_schedule_timeout); | |
4992 | ||
4993 | void io_schedule(void) | |
4994 | { | |
4995 | int token; | |
4996 | ||
4997 | token = io_schedule_prepare(); | |
4998 | schedule(); | |
4999 | io_schedule_finish(token); | |
5000 | } | |
5001 | EXPORT_SYMBOL(io_schedule); | |
5002 | ||
5003 | /** | |
5004 | * sys_sched_get_priority_max - return maximum RT priority. | |
5005 | * @policy: scheduling class. | |
5006 | * | |
5007 | * Return: On success, this syscall returns the maximum | |
5008 | * rt_priority that can be used by a given scheduling class. | |
5009 | * On failure, a negative error code is returned. | |
5010 | */ | |
5011 | SYSCALL_DEFINE1(sched_get_priority_max, int, policy) | |
5012 | { | |
5013 | int ret = -EINVAL; | |
5014 | ||
5015 | switch (policy) { | |
5016 | case SCHED_FIFO: | |
5017 | case SCHED_RR: | |
5018 | ret = MAX_USER_RT_PRIO-1; | |
5019 | break; | |
5020 | case SCHED_DEADLINE: | |
5021 | case SCHED_NORMAL: | |
5022 | case SCHED_BATCH: | |
5023 | case SCHED_IDLE: | |
5024 | ret = 0; | |
5025 | break; | |
5026 | } | |
5027 | return ret; | |
5028 | } | |
5029 | ||
5030 | /** | |
5031 | * sys_sched_get_priority_min - return minimum RT priority. | |
5032 | * @policy: scheduling class. | |
5033 | * | |
5034 | * Return: On success, this syscall returns the minimum | |
5035 | * rt_priority that can be used by a given scheduling class. | |
5036 | * On failure, a negative error code is returned. | |
5037 | */ | |
5038 | SYSCALL_DEFINE1(sched_get_priority_min, int, policy) | |
5039 | { | |
5040 | int ret = -EINVAL; | |
5041 | ||
5042 | switch (policy) { | |
5043 | case SCHED_FIFO: | |
5044 | case SCHED_RR: | |
5045 | ret = 1; | |
5046 | break; | |
5047 | case SCHED_DEADLINE: | |
5048 | case SCHED_NORMAL: | |
5049 | case SCHED_BATCH: | |
5050 | case SCHED_IDLE: | |
5051 | ret = 0; | |
5052 | } | |
5053 | return ret; | |
5054 | } | |
5055 | ||
5056 | /** | |
5057 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
5058 | * @pid: pid of the process. | |
5059 | * @interval: userspace pointer to the timeslice value. | |
5060 | * | |
5061 | * this syscall writes the default timeslice value of a given process | |
5062 | * into the user-space timespec buffer. A value of '0' means infinity. | |
5063 | * | |
5064 | * Return: On success, 0 and the timeslice is in @interval. Otherwise, | |
5065 | * an error code. | |
5066 | */ | |
5067 | SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, | |
5068 | struct timespec __user *, interval) | |
5069 | { | |
5070 | struct task_struct *p; | |
5071 | unsigned int time_slice; | |
5072 | struct rq_flags rf; | |
5073 | struct timespec t; | |
5074 | struct rq *rq; | |
5075 | int retval; | |
5076 | ||
5077 | if (pid < 0) | |
5078 | return -EINVAL; | |
5079 | ||
5080 | retval = -ESRCH; | |
5081 | rcu_read_lock(); | |
5082 | p = find_process_by_pid(pid); | |
5083 | if (!p) | |
5084 | goto out_unlock; | |
5085 | ||
5086 | retval = security_task_getscheduler(p); | |
5087 | if (retval) | |
5088 | goto out_unlock; | |
5089 | ||
5090 | rq = task_rq_lock(p, &rf); | |
5091 | time_slice = 0; | |
5092 | if (p->sched_class->get_rr_interval) | |
5093 | time_slice = p->sched_class->get_rr_interval(rq, p); | |
5094 | task_rq_unlock(rq, p, &rf); | |
5095 | ||
5096 | rcu_read_unlock(); | |
5097 | jiffies_to_timespec(time_slice, &t); | |
5098 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | |
5099 | return retval; | |
5100 | ||
5101 | out_unlock: | |
5102 | rcu_read_unlock(); | |
5103 | return retval; | |
5104 | } | |
5105 | ||
5106 | static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; | |
5107 | ||
5108 | void sched_show_task(struct task_struct *p) | |
5109 | { | |
5110 | unsigned long free = 0; | |
5111 | int ppid; | |
5112 | unsigned long state = p->state; | |
5113 | ||
5114 | /* Make sure the string lines up properly with the number of task states: */ | |
5115 | BUILD_BUG_ON(sizeof(TASK_STATE_TO_CHAR_STR)-1 != ilog2(TASK_STATE_MAX)+1); | |
5116 | ||
5117 | if (!try_get_task_stack(p)) | |
5118 | return; | |
5119 | if (state) | |
5120 | state = __ffs(state) + 1; | |
5121 | printk(KERN_INFO "%-15.15s %c", p->comm, | |
5122 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); | |
5123 | if (state == TASK_RUNNING) | |
5124 | printk(KERN_CONT " running task "); | |
5125 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
5126 | free = stack_not_used(p); | |
5127 | #endif | |
5128 | ppid = 0; | |
5129 | rcu_read_lock(); | |
5130 | if (pid_alive(p)) | |
5131 | ppid = task_pid_nr(rcu_dereference(p->real_parent)); | |
5132 | rcu_read_unlock(); | |
5133 | printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, | |
5134 | task_pid_nr(p), ppid, | |
5135 | (unsigned long)task_thread_info(p)->flags); | |
5136 | ||
5137 | print_worker_info(KERN_INFO, p); | |
5138 | show_stack(p, NULL); | |
5139 | put_task_stack(p); | |
5140 | } | |
5141 | ||
5142 | void show_state_filter(unsigned long state_filter) | |
5143 | { | |
5144 | struct task_struct *g, *p; | |
5145 | ||
5146 | #if BITS_PER_LONG == 32 | |
5147 | printk(KERN_INFO | |
5148 | " task PC stack pid father\n"); | |
5149 | #else | |
5150 | printk(KERN_INFO | |
5151 | " task PC stack pid father\n"); | |
5152 | #endif | |
5153 | rcu_read_lock(); | |
5154 | for_each_process_thread(g, p) { | |
5155 | /* | |
5156 | * reset the NMI-timeout, listing all files on a slow | |
5157 | * console might take a lot of time: | |
5158 | * Also, reset softlockup watchdogs on all CPUs, because | |
5159 | * another CPU might be blocked waiting for us to process | |
5160 | * an IPI. | |
5161 | */ | |
5162 | touch_nmi_watchdog(); | |
5163 | touch_all_softlockup_watchdogs(); | |
5164 | if (!state_filter || (p->state & state_filter)) | |
5165 | sched_show_task(p); | |
5166 | } | |
5167 | ||
5168 | #ifdef CONFIG_SCHED_DEBUG | |
5169 | if (!state_filter) | |
5170 | sysrq_sched_debug_show(); | |
5171 | #endif | |
5172 | rcu_read_unlock(); | |
5173 | /* | |
5174 | * Only show locks if all tasks are dumped: | |
5175 | */ | |
5176 | if (!state_filter) | |
5177 | debug_show_all_locks(); | |
5178 | } | |
5179 | ||
5180 | void init_idle_bootup_task(struct task_struct *idle) | |
5181 | { | |
5182 | idle->sched_class = &idle_sched_class; | |
5183 | } | |
5184 | ||
5185 | /** | |
5186 | * init_idle - set up an idle thread for a given CPU | |
5187 | * @idle: task in question | |
5188 | * @cpu: CPU the idle task belongs to | |
5189 | * | |
5190 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
5191 | * flag, to make booting more robust. | |
5192 | */ | |
5193 | void init_idle(struct task_struct *idle, int cpu) | |
5194 | { | |
5195 | struct rq *rq = cpu_rq(cpu); | |
5196 | unsigned long flags; | |
5197 | ||
5198 | raw_spin_lock_irqsave(&idle->pi_lock, flags); | |
5199 | raw_spin_lock(&rq->lock); | |
5200 | ||
5201 | __sched_fork(0, idle); | |
5202 | idle->state = TASK_RUNNING; | |
5203 | idle->se.exec_start = sched_clock(); | |
5204 | idle->flags |= PF_IDLE; | |
5205 | ||
5206 | kasan_unpoison_task_stack(idle); | |
5207 | ||
5208 | #ifdef CONFIG_SMP | |
5209 | /* | |
5210 | * Its possible that init_idle() gets called multiple times on a task, | |
5211 | * in that case do_set_cpus_allowed() will not do the right thing. | |
5212 | * | |
5213 | * And since this is boot we can forgo the serialization. | |
5214 | */ | |
5215 | set_cpus_allowed_common(idle, cpumask_of(cpu)); | |
5216 | #endif | |
5217 | /* | |
5218 | * We're having a chicken and egg problem, even though we are | |
5219 | * holding rq->lock, the CPU isn't yet set to this CPU so the | |
5220 | * lockdep check in task_group() will fail. | |
5221 | * | |
5222 | * Similar case to sched_fork(). / Alternatively we could | |
5223 | * use task_rq_lock() here and obtain the other rq->lock. | |
5224 | * | |
5225 | * Silence PROVE_RCU | |
5226 | */ | |
5227 | rcu_read_lock(); | |
5228 | __set_task_cpu(idle, cpu); | |
5229 | rcu_read_unlock(); | |
5230 | ||
5231 | rq->curr = rq->idle = idle; | |
5232 | idle->on_rq = TASK_ON_RQ_QUEUED; | |
5233 | #ifdef CONFIG_SMP | |
5234 | idle->on_cpu = 1; | |
5235 | #endif | |
5236 | raw_spin_unlock(&rq->lock); | |
5237 | raw_spin_unlock_irqrestore(&idle->pi_lock, flags); | |
5238 | ||
5239 | /* Set the preempt count _outside_ the spinlocks! */ | |
5240 | init_idle_preempt_count(idle, cpu); | |
5241 | ||
5242 | /* | |
5243 | * The idle tasks have their own, simple scheduling class: | |
5244 | */ | |
5245 | idle->sched_class = &idle_sched_class; | |
5246 | ftrace_graph_init_idle_task(idle, cpu); | |
5247 | vtime_init_idle(idle, cpu); | |
5248 | #ifdef CONFIG_SMP | |
5249 | sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); | |
5250 | #endif | |
5251 | } | |
5252 | ||
5253 | #ifdef CONFIG_SMP | |
5254 | ||
5255 | int cpuset_cpumask_can_shrink(const struct cpumask *cur, | |
5256 | const struct cpumask *trial) | |
5257 | { | |
5258 | int ret = 1; | |
5259 | ||
5260 | if (!cpumask_weight(cur)) | |
5261 | return ret; | |
5262 | ||
5263 | ret = dl_cpuset_cpumask_can_shrink(cur, trial); | |
5264 | ||
5265 | return ret; | |
5266 | } | |
5267 | ||
5268 | int task_can_attach(struct task_struct *p, | |
5269 | const struct cpumask *cs_cpus_allowed) | |
5270 | { | |
5271 | int ret = 0; | |
5272 | ||
5273 | /* | |
5274 | * Kthreads which disallow setaffinity shouldn't be moved | |
5275 | * to a new cpuset; we don't want to change their CPU | |
5276 | * affinity and isolating such threads by their set of | |
5277 | * allowed nodes is unnecessary. Thus, cpusets are not | |
5278 | * applicable for such threads. This prevents checking for | |
5279 | * success of set_cpus_allowed_ptr() on all attached tasks | |
5280 | * before cpus_allowed may be changed. | |
5281 | */ | |
5282 | if (p->flags & PF_NO_SETAFFINITY) { | |
5283 | ret = -EINVAL; | |
5284 | goto out; | |
5285 | } | |
5286 | ||
5287 | if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span, | |
5288 | cs_cpus_allowed)) | |
5289 | ret = dl_task_can_attach(p, cs_cpus_allowed); | |
5290 | ||
5291 | out: | |
5292 | return ret; | |
5293 | } | |
5294 | ||
5295 | bool sched_smp_initialized __read_mostly; | |
5296 | ||
5297 | #ifdef CONFIG_NUMA_BALANCING | |
5298 | /* Migrate current task p to target_cpu */ | |
5299 | int migrate_task_to(struct task_struct *p, int target_cpu) | |
5300 | { | |
5301 | struct migration_arg arg = { p, target_cpu }; | |
5302 | int curr_cpu = task_cpu(p); | |
5303 | ||
5304 | if (curr_cpu == target_cpu) | |
5305 | return 0; | |
5306 | ||
5307 | if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed)) | |
5308 | return -EINVAL; | |
5309 | ||
5310 | /* TODO: This is not properly updating schedstats */ | |
5311 | ||
5312 | trace_sched_move_numa(p, curr_cpu, target_cpu); | |
5313 | return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); | |
5314 | } | |
5315 | ||
5316 | /* | |
5317 | * Requeue a task on a given node and accurately track the number of NUMA | |
5318 | * tasks on the runqueues | |
5319 | */ | |
5320 | void sched_setnuma(struct task_struct *p, int nid) | |
5321 | { | |
5322 | bool queued, running; | |
5323 | struct rq_flags rf; | |
5324 | struct rq *rq; | |
5325 | ||
5326 | rq = task_rq_lock(p, &rf); | |
5327 | queued = task_on_rq_queued(p); | |
5328 | running = task_current(rq, p); | |
5329 | ||
5330 | if (queued) | |
5331 | dequeue_task(rq, p, DEQUEUE_SAVE); | |
5332 | if (running) | |
5333 | put_prev_task(rq, p); | |
5334 | ||
5335 | p->numa_preferred_nid = nid; | |
5336 | ||
5337 | if (queued) | |
5338 | enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK); | |
5339 | if (running) | |
5340 | set_curr_task(rq, p); | |
5341 | task_rq_unlock(rq, p, &rf); | |
5342 | } | |
5343 | #endif /* CONFIG_NUMA_BALANCING */ | |
5344 | ||
5345 | #ifdef CONFIG_HOTPLUG_CPU | |
5346 | /* | |
5347 | * Ensure that the idle task is using init_mm right before its CPU goes | |
5348 | * offline. | |
5349 | */ | |
5350 | void idle_task_exit(void) | |
5351 | { | |
5352 | struct mm_struct *mm = current->active_mm; | |
5353 | ||
5354 | BUG_ON(cpu_online(smp_processor_id())); | |
5355 | ||
5356 | if (mm != &init_mm) { | |
5357 | switch_mm(mm, &init_mm, current); | |
5358 | finish_arch_post_lock_switch(); | |
5359 | } | |
5360 | mmdrop(mm); | |
5361 | } | |
5362 | ||
5363 | /* | |
5364 | * Since this CPU is going 'away' for a while, fold any nr_active delta | |
5365 | * we might have. Assumes we're called after migrate_tasks() so that the | |
5366 | * nr_active count is stable. We need to take the teardown thread which | |
5367 | * is calling this into account, so we hand in adjust = 1 to the load | |
5368 | * calculation. | |
5369 | * | |
5370 | * Also see the comment "Global load-average calculations". | |
5371 | */ | |
5372 | static void calc_load_migrate(struct rq *rq) | |
5373 | { | |
5374 | long delta = calc_load_fold_active(rq, 1); | |
5375 | if (delta) | |
5376 | atomic_long_add(delta, &calc_load_tasks); | |
5377 | } | |
5378 | ||
5379 | static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) | |
5380 | { | |
5381 | } | |
5382 | ||
5383 | static const struct sched_class fake_sched_class = { | |
5384 | .put_prev_task = put_prev_task_fake, | |
5385 | }; | |
5386 | ||
5387 | static struct task_struct fake_task = { | |
5388 | /* | |
5389 | * Avoid pull_{rt,dl}_task() | |
5390 | */ | |
5391 | .prio = MAX_PRIO + 1, | |
5392 | .sched_class = &fake_sched_class, | |
5393 | }; | |
5394 | ||
5395 | /* | |
5396 | * Migrate all tasks from the rq, sleeping tasks will be migrated by | |
5397 | * try_to_wake_up()->select_task_rq(). | |
5398 | * | |
5399 | * Called with rq->lock held even though we'er in stop_machine() and | |
5400 | * there's no concurrency possible, we hold the required locks anyway | |
5401 | * because of lock validation efforts. | |
5402 | */ | |
5403 | static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf) | |
5404 | { | |
5405 | struct rq *rq = dead_rq; | |
5406 | struct task_struct *next, *stop = rq->stop; | |
5407 | struct rq_flags orf = *rf; | |
5408 | int dest_cpu; | |
5409 | ||
5410 | /* | |
5411 | * Fudge the rq selection such that the below task selection loop | |
5412 | * doesn't get stuck on the currently eligible stop task. | |
5413 | * | |
5414 | * We're currently inside stop_machine() and the rq is either stuck | |
5415 | * in the stop_machine_cpu_stop() loop, or we're executing this code, | |
5416 | * either way we should never end up calling schedule() until we're | |
5417 | * done here. | |
5418 | */ | |
5419 | rq->stop = NULL; | |
5420 | ||
5421 | /* | |
5422 | * put_prev_task() and pick_next_task() sched | |
5423 | * class method both need to have an up-to-date | |
5424 | * value of rq->clock[_task] | |
5425 | */ | |
5426 | update_rq_clock(rq); | |
5427 | ||
5428 | for (;;) { | |
5429 | /* | |
5430 | * There's this thread running, bail when that's the only | |
5431 | * remaining thread: | |
5432 | */ | |
5433 | if (rq->nr_running == 1) | |
5434 | break; | |
5435 | ||
5436 | /* | |
5437 | * pick_next_task() assumes pinned rq->lock: | |
5438 | */ | |
5439 | next = pick_next_task(rq, &fake_task, rf); | |
5440 | BUG_ON(!next); | |
5441 | next->sched_class->put_prev_task(rq, next); | |
5442 | ||
5443 | /* | |
5444 | * Rules for changing task_struct::cpus_allowed are holding | |
5445 | * both pi_lock and rq->lock, such that holding either | |
5446 | * stabilizes the mask. | |
5447 | * | |
5448 | * Drop rq->lock is not quite as disastrous as it usually is | |
5449 | * because !cpu_active at this point, which means load-balance | |
5450 | * will not interfere. Also, stop-machine. | |
5451 | */ | |
5452 | rq_unlock(rq, rf); | |
5453 | raw_spin_lock(&next->pi_lock); | |
5454 | rq_relock(rq, rf); | |
5455 | ||
5456 | /* | |
5457 | * Since we're inside stop-machine, _nothing_ should have | |
5458 | * changed the task, WARN if weird stuff happened, because in | |
5459 | * that case the above rq->lock drop is a fail too. | |
5460 | */ | |
5461 | if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) { | |
5462 | raw_spin_unlock(&next->pi_lock); | |
5463 | continue; | |
5464 | } | |
5465 | ||
5466 | /* Find suitable destination for @next, with force if needed. */ | |
5467 | dest_cpu = select_fallback_rq(dead_rq->cpu, next); | |
5468 | rq = __migrate_task(rq, rf, next, dest_cpu); | |
5469 | if (rq != dead_rq) { | |
5470 | rq_unlock(rq, rf); | |
5471 | rq = dead_rq; | |
5472 | *rf = orf; | |
5473 | rq_relock(rq, rf); | |
5474 | } | |
5475 | raw_spin_unlock(&next->pi_lock); | |
5476 | } | |
5477 | ||
5478 | rq->stop = stop; | |
5479 | } | |
5480 | #endif /* CONFIG_HOTPLUG_CPU */ | |
5481 | ||
5482 | void set_rq_online(struct rq *rq) | |
5483 | { | |
5484 | if (!rq->online) { | |
5485 | const struct sched_class *class; | |
5486 | ||
5487 | cpumask_set_cpu(rq->cpu, rq->rd->online); | |
5488 | rq->online = 1; | |
5489 | ||
5490 | for_each_class(class) { | |
5491 | if (class->rq_online) | |
5492 | class->rq_online(rq); | |
5493 | } | |
5494 | } | |
5495 | } | |
5496 | ||
5497 | void set_rq_offline(struct rq *rq) | |
5498 | { | |
5499 | if (rq->online) { | |
5500 | const struct sched_class *class; | |
5501 | ||
5502 | for_each_class(class) { | |
5503 | if (class->rq_offline) | |
5504 | class->rq_offline(rq); | |
5505 | } | |
5506 | ||
5507 | cpumask_clear_cpu(rq->cpu, rq->rd->online); | |
5508 | rq->online = 0; | |
5509 | } | |
5510 | } | |
5511 | ||
5512 | static void set_cpu_rq_start_time(unsigned int cpu) | |
5513 | { | |
5514 | struct rq *rq = cpu_rq(cpu); | |
5515 | ||
5516 | rq->age_stamp = sched_clock_cpu(cpu); | |
5517 | } | |
5518 | ||
5519 | /* | |
5520 | * used to mark begin/end of suspend/resume: | |
5521 | */ | |
5522 | static int num_cpus_frozen; | |
5523 | ||
5524 | /* | |
5525 | * Update cpusets according to cpu_active mask. If cpusets are | |
5526 | * disabled, cpuset_update_active_cpus() becomes a simple wrapper | |
5527 | * around partition_sched_domains(). | |
5528 | * | |
5529 | * If we come here as part of a suspend/resume, don't touch cpusets because we | |
5530 | * want to restore it back to its original state upon resume anyway. | |
5531 | */ | |
5532 | static void cpuset_cpu_active(void) | |
5533 | { | |
5534 | if (cpuhp_tasks_frozen) { | |
5535 | /* | |
5536 | * num_cpus_frozen tracks how many CPUs are involved in suspend | |
5537 | * resume sequence. As long as this is not the last online | |
5538 | * operation in the resume sequence, just build a single sched | |
5539 | * domain, ignoring cpusets. | |
5540 | */ | |
5541 | num_cpus_frozen--; | |
5542 | if (likely(num_cpus_frozen)) { | |
5543 | partition_sched_domains(1, NULL, NULL); | |
5544 | return; | |
5545 | } | |
5546 | /* | |
5547 | * This is the last CPU online operation. So fall through and | |
5548 | * restore the original sched domains by considering the | |
5549 | * cpuset configurations. | |
5550 | */ | |
5551 | } | |
5552 | cpuset_update_active_cpus(); | |
5553 | } | |
5554 | ||
5555 | static int cpuset_cpu_inactive(unsigned int cpu) | |
5556 | { | |
5557 | if (!cpuhp_tasks_frozen) { | |
5558 | if (dl_cpu_busy(cpu)) | |
5559 | return -EBUSY; | |
5560 | cpuset_update_active_cpus(); | |
5561 | } else { | |
5562 | num_cpus_frozen++; | |
5563 | partition_sched_domains(1, NULL, NULL); | |
5564 | } | |
5565 | return 0; | |
5566 | } | |
5567 | ||
5568 | int sched_cpu_activate(unsigned int cpu) | |
5569 | { | |
5570 | struct rq *rq = cpu_rq(cpu); | |
5571 | struct rq_flags rf; | |
5572 | ||
5573 | set_cpu_active(cpu, true); | |
5574 | ||
5575 | if (sched_smp_initialized) { | |
5576 | sched_domains_numa_masks_set(cpu); | |
5577 | cpuset_cpu_active(); | |
5578 | } | |
5579 | ||
5580 | /* | |
5581 | * Put the rq online, if not already. This happens: | |
5582 | * | |
5583 | * 1) In the early boot process, because we build the real domains | |
5584 | * after all CPUs have been brought up. | |
5585 | * | |
5586 | * 2) At runtime, if cpuset_cpu_active() fails to rebuild the | |
5587 | * domains. | |
5588 | */ | |
5589 | rq_lock_irqsave(rq, &rf); | |
5590 | if (rq->rd) { | |
5591 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); | |
5592 | set_rq_online(rq); | |
5593 | } | |
5594 | rq_unlock_irqrestore(rq, &rf); | |
5595 | ||
5596 | update_max_interval(); | |
5597 | ||
5598 | return 0; | |
5599 | } | |
5600 | ||
5601 | int sched_cpu_deactivate(unsigned int cpu) | |
5602 | { | |
5603 | int ret; | |
5604 | ||
5605 | set_cpu_active(cpu, false); | |
5606 | /* | |
5607 | * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU | |
5608 | * users of this state to go away such that all new such users will | |
5609 | * observe it. | |
5610 | * | |
5611 | * Do sync before park smpboot threads to take care the rcu boost case. | |
5612 | */ | |
5613 | synchronize_rcu_mult(call_rcu, call_rcu_sched); | |
5614 | ||
5615 | if (!sched_smp_initialized) | |
5616 | return 0; | |
5617 | ||
5618 | ret = cpuset_cpu_inactive(cpu); | |
5619 | if (ret) { | |
5620 | set_cpu_active(cpu, true); | |
5621 | return ret; | |
5622 | } | |
5623 | sched_domains_numa_masks_clear(cpu); | |
5624 | return 0; | |
5625 | } | |
5626 | ||
5627 | static void sched_rq_cpu_starting(unsigned int cpu) | |
5628 | { | |
5629 | struct rq *rq = cpu_rq(cpu); | |
5630 | ||
5631 | rq->calc_load_update = calc_load_update; | |
5632 | update_max_interval(); | |
5633 | } | |
5634 | ||
5635 | int sched_cpu_starting(unsigned int cpu) | |
5636 | { | |
5637 | set_cpu_rq_start_time(cpu); | |
5638 | sched_rq_cpu_starting(cpu); | |
5639 | return 0; | |
5640 | } | |
5641 | ||
5642 | #ifdef CONFIG_HOTPLUG_CPU | |
5643 | int sched_cpu_dying(unsigned int cpu) | |
5644 | { | |
5645 | struct rq *rq = cpu_rq(cpu); | |
5646 | struct rq_flags rf; | |
5647 | ||
5648 | /* Handle pending wakeups and then migrate everything off */ | |
5649 | sched_ttwu_pending(); | |
5650 | ||
5651 | rq_lock_irqsave(rq, &rf); | |
5652 | if (rq->rd) { | |
5653 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); | |
5654 | set_rq_offline(rq); | |
5655 | } | |
5656 | migrate_tasks(rq, &rf); | |
5657 | BUG_ON(rq->nr_running != 1); | |
5658 | rq_unlock_irqrestore(rq, &rf); | |
5659 | ||
5660 | calc_load_migrate(rq); | |
5661 | update_max_interval(); | |
5662 | nohz_balance_exit_idle(cpu); | |
5663 | hrtick_clear(rq); | |
5664 | return 0; | |
5665 | } | |
5666 | #endif | |
5667 | ||
5668 | #ifdef CONFIG_SCHED_SMT | |
5669 | DEFINE_STATIC_KEY_FALSE(sched_smt_present); | |
5670 | ||
5671 | static void sched_init_smt(void) | |
5672 | { | |
5673 | /* | |
5674 | * We've enumerated all CPUs and will assume that if any CPU | |
5675 | * has SMT siblings, CPU0 will too. | |
5676 | */ | |
5677 | if (cpumask_weight(cpu_smt_mask(0)) > 1) | |
5678 | static_branch_enable(&sched_smt_present); | |
5679 | } | |
5680 | #else | |
5681 | static inline void sched_init_smt(void) { } | |
5682 | #endif | |
5683 | ||
5684 | void __init sched_init_smp(void) | |
5685 | { | |
5686 | cpumask_var_t non_isolated_cpus; | |
5687 | ||
5688 | alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); | |
5689 | ||
5690 | sched_init_numa(); | |
5691 | ||
5692 | /* | |
5693 | * There's no userspace yet to cause hotplug operations; hence all the | |
5694 | * CPU masks are stable and all blatant races in the below code cannot | |
5695 | * happen. | |
5696 | */ | |
5697 | mutex_lock(&sched_domains_mutex); | |
5698 | sched_init_domains(cpu_active_mask); | |
5699 | cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); | |
5700 | if (cpumask_empty(non_isolated_cpus)) | |
5701 | cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); | |
5702 | mutex_unlock(&sched_domains_mutex); | |
5703 | ||
5704 | /* Move init over to a non-isolated CPU */ | |
5705 | if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) | |
5706 | BUG(); | |
5707 | sched_init_granularity(); | |
5708 | free_cpumask_var(non_isolated_cpus); | |
5709 | ||
5710 | init_sched_rt_class(); | |
5711 | init_sched_dl_class(); | |
5712 | ||
5713 | sched_init_smt(); | |
5714 | ||
5715 | sched_smp_initialized = true; | |
5716 | } | |
5717 | ||
5718 | static int __init migration_init(void) | |
5719 | { | |
5720 | sched_rq_cpu_starting(smp_processor_id()); | |
5721 | return 0; | |
5722 | } | |
5723 | early_initcall(migration_init); | |
5724 | ||
5725 | #else | |
5726 | void __init sched_init_smp(void) | |
5727 | { | |
5728 | sched_init_granularity(); | |
5729 | } | |
5730 | #endif /* CONFIG_SMP */ | |
5731 | ||
5732 | int in_sched_functions(unsigned long addr) | |
5733 | { | |
5734 | return in_lock_functions(addr) || | |
5735 | (addr >= (unsigned long)__sched_text_start | |
5736 | && addr < (unsigned long)__sched_text_end); | |
5737 | } | |
5738 | ||
5739 | #ifdef CONFIG_CGROUP_SCHED | |
5740 | /* | |
5741 | * Default task group. | |
5742 | * Every task in system belongs to this group at bootup. | |
5743 | */ | |
5744 | struct task_group root_task_group; | |
5745 | LIST_HEAD(task_groups); | |
5746 | ||
5747 | /* Cacheline aligned slab cache for task_group */ | |
5748 | static struct kmem_cache *task_group_cache __read_mostly; | |
5749 | #endif | |
5750 | ||
5751 | DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); | |
5752 | DECLARE_PER_CPU(cpumask_var_t, select_idle_mask); | |
5753 | ||
5754 | void __init sched_init(void) | |
5755 | { | |
5756 | int i, j; | |
5757 | unsigned long alloc_size = 0, ptr; | |
5758 | ||
5759 | sched_clock_init(); | |
5760 | wait_bit_init(); | |
5761 | ||
5762 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5763 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
5764 | #endif | |
5765 | #ifdef CONFIG_RT_GROUP_SCHED | |
5766 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
5767 | #endif | |
5768 | if (alloc_size) { | |
5769 | ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); | |
5770 | ||
5771 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5772 | root_task_group.se = (struct sched_entity **)ptr; | |
5773 | ptr += nr_cpu_ids * sizeof(void **); | |
5774 | ||
5775 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
5776 | ptr += nr_cpu_ids * sizeof(void **); | |
5777 | ||
5778 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
5779 | #ifdef CONFIG_RT_GROUP_SCHED | |
5780 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
5781 | ptr += nr_cpu_ids * sizeof(void **); | |
5782 | ||
5783 | root_task_group.rt_rq = (struct rt_rq **)ptr; | |
5784 | ptr += nr_cpu_ids * sizeof(void **); | |
5785 | ||
5786 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
5787 | } | |
5788 | #ifdef CONFIG_CPUMASK_OFFSTACK | |
5789 | for_each_possible_cpu(i) { | |
5790 | per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node( | |
5791 | cpumask_size(), GFP_KERNEL, cpu_to_node(i)); | |
5792 | per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node( | |
5793 | cpumask_size(), GFP_KERNEL, cpu_to_node(i)); | |
5794 | } | |
5795 | #endif /* CONFIG_CPUMASK_OFFSTACK */ | |
5796 | ||
5797 | init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime()); | |
5798 | init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime()); | |
5799 | ||
5800 | #ifdef CONFIG_SMP | |
5801 | init_defrootdomain(); | |
5802 | #endif | |
5803 | ||
5804 | #ifdef CONFIG_RT_GROUP_SCHED | |
5805 | init_rt_bandwidth(&root_task_group.rt_bandwidth, | |
5806 | global_rt_period(), global_rt_runtime()); | |
5807 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
5808 | ||
5809 | #ifdef CONFIG_CGROUP_SCHED | |
5810 | task_group_cache = KMEM_CACHE(task_group, 0); | |
5811 | ||
5812 | list_add(&root_task_group.list, &task_groups); | |
5813 | INIT_LIST_HEAD(&root_task_group.children); | |
5814 | INIT_LIST_HEAD(&root_task_group.siblings); | |
5815 | autogroup_init(&init_task); | |
5816 | #endif /* CONFIG_CGROUP_SCHED */ | |
5817 | ||
5818 | for_each_possible_cpu(i) { | |
5819 | struct rq *rq; | |
5820 | ||
5821 | rq = cpu_rq(i); | |
5822 | raw_spin_lock_init(&rq->lock); | |
5823 | rq->nr_running = 0; | |
5824 | rq->calc_load_active = 0; | |
5825 | rq->calc_load_update = jiffies + LOAD_FREQ; | |
5826 | init_cfs_rq(&rq->cfs); | |
5827 | init_rt_rq(&rq->rt); | |
5828 | init_dl_rq(&rq->dl); | |
5829 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5830 | root_task_group.shares = ROOT_TASK_GROUP_LOAD; | |
5831 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); | |
5832 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
5833 | /* | |
5834 | * How much CPU bandwidth does root_task_group get? | |
5835 | * | |
5836 | * In case of task-groups formed thr' the cgroup filesystem, it | |
5837 | * gets 100% of the CPU resources in the system. This overall | |
5838 | * system CPU resource is divided among the tasks of | |
5839 | * root_task_group and its child task-groups in a fair manner, | |
5840 | * based on each entity's (task or task-group's) weight | |
5841 | * (se->load.weight). | |
5842 | * | |
5843 | * In other words, if root_task_group has 10 tasks of weight | |
5844 | * 1024) and two child groups A0 and A1 (of weight 1024 each), | |
5845 | * then A0's share of the CPU resource is: | |
5846 | * | |
5847 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% | |
5848 | * | |
5849 | * We achieve this by letting root_task_group's tasks sit | |
5850 | * directly in rq->cfs (i.e root_task_group->se[] = NULL). | |
5851 | */ | |
5852 | init_cfs_bandwidth(&root_task_group.cfs_bandwidth); | |
5853 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); | |
5854 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
5855 | ||
5856 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; | |
5857 | #ifdef CONFIG_RT_GROUP_SCHED | |
5858 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); | |
5859 | #endif | |
5860 | ||
5861 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) | |
5862 | rq->cpu_load[j] = 0; | |
5863 | ||
5864 | #ifdef CONFIG_SMP | |
5865 | rq->sd = NULL; | |
5866 | rq->rd = NULL; | |
5867 | rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE; | |
5868 | rq->balance_callback = NULL; | |
5869 | rq->active_balance = 0; | |
5870 | rq->next_balance = jiffies; | |
5871 | rq->push_cpu = 0; | |
5872 | rq->cpu = i; | |
5873 | rq->online = 0; | |
5874 | rq->idle_stamp = 0; | |
5875 | rq->avg_idle = 2*sysctl_sched_migration_cost; | |
5876 | rq->max_idle_balance_cost = sysctl_sched_migration_cost; | |
5877 | ||
5878 | INIT_LIST_HEAD(&rq->cfs_tasks); | |
5879 | ||
5880 | rq_attach_root(rq, &def_root_domain); | |
5881 | #ifdef CONFIG_NO_HZ_COMMON | |
5882 | rq->last_load_update_tick = jiffies; | |
5883 | rq->nohz_flags = 0; | |
5884 | #endif | |
5885 | #ifdef CONFIG_NO_HZ_FULL | |
5886 | rq->last_sched_tick = 0; | |
5887 | #endif | |
5888 | #endif /* CONFIG_SMP */ | |
5889 | init_rq_hrtick(rq); | |
5890 | atomic_set(&rq->nr_iowait, 0); | |
5891 | } | |
5892 | ||
5893 | set_load_weight(&init_task); | |
5894 | ||
5895 | /* | |
5896 | * The boot idle thread does lazy MMU switching as well: | |
5897 | */ | |
5898 | mmgrab(&init_mm); | |
5899 | enter_lazy_tlb(&init_mm, current); | |
5900 | ||
5901 | /* | |
5902 | * Make us the idle thread. Technically, schedule() should not be | |
5903 | * called from this thread, however somewhere below it might be, | |
5904 | * but because we are the idle thread, we just pick up running again | |
5905 | * when this runqueue becomes "idle". | |
5906 | */ | |
5907 | init_idle(current, smp_processor_id()); | |
5908 | ||
5909 | calc_load_update = jiffies + LOAD_FREQ; | |
5910 | ||
5911 | #ifdef CONFIG_SMP | |
5912 | /* May be allocated at isolcpus cmdline parse time */ | |
5913 | if (cpu_isolated_map == NULL) | |
5914 | zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); | |
5915 | idle_thread_set_boot_cpu(); | |
5916 | set_cpu_rq_start_time(smp_processor_id()); | |
5917 | #endif | |
5918 | init_sched_fair_class(); | |
5919 | ||
5920 | init_schedstats(); | |
5921 | ||
5922 | scheduler_running = 1; | |
5923 | } | |
5924 | ||
5925 | #ifdef CONFIG_DEBUG_ATOMIC_SLEEP | |
5926 | static inline int preempt_count_equals(int preempt_offset) | |
5927 | { | |
5928 | int nested = preempt_count() + rcu_preempt_depth(); | |
5929 | ||
5930 | return (nested == preempt_offset); | |
5931 | } | |
5932 | ||
5933 | void __might_sleep(const char *file, int line, int preempt_offset) | |
5934 | { | |
5935 | /* | |
5936 | * Blocking primitives will set (and therefore destroy) current->state, | |
5937 | * since we will exit with TASK_RUNNING make sure we enter with it, | |
5938 | * otherwise we will destroy state. | |
5939 | */ | |
5940 | WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change, | |
5941 | "do not call blocking ops when !TASK_RUNNING; " | |
5942 | "state=%lx set at [<%p>] %pS\n", | |
5943 | current->state, | |
5944 | (void *)current->task_state_change, | |
5945 | (void *)current->task_state_change); | |
5946 | ||
5947 | ___might_sleep(file, line, preempt_offset); | |
5948 | } | |
5949 | EXPORT_SYMBOL(__might_sleep); | |
5950 | ||
5951 | void ___might_sleep(const char *file, int line, int preempt_offset) | |
5952 | { | |
5953 | /* Ratelimiting timestamp: */ | |
5954 | static unsigned long prev_jiffy; | |
5955 | ||
5956 | unsigned long preempt_disable_ip; | |
5957 | ||
5958 | /* WARN_ON_ONCE() by default, no rate limit required: */ | |
5959 | rcu_sleep_check(); | |
5960 | ||
5961 | if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && | |
5962 | !is_idle_task(current)) || | |
5963 | system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING || | |
5964 | oops_in_progress) | |
5965 | return; | |
5966 | ||
5967 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
5968 | return; | |
5969 | prev_jiffy = jiffies; | |
5970 | ||
5971 | /* Save this before calling printk(), since that will clobber it: */ | |
5972 | preempt_disable_ip = get_preempt_disable_ip(current); | |
5973 | ||
5974 | printk(KERN_ERR | |
5975 | "BUG: sleeping function called from invalid context at %s:%d\n", | |
5976 | file, line); | |
5977 | printk(KERN_ERR | |
5978 | "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", | |
5979 | in_atomic(), irqs_disabled(), | |
5980 | current->pid, current->comm); | |
5981 | ||
5982 | if (task_stack_end_corrupted(current)) | |
5983 | printk(KERN_EMERG "Thread overran stack, or stack corrupted\n"); | |
5984 | ||
5985 | debug_show_held_locks(current); | |
5986 | if (irqs_disabled()) | |
5987 | print_irqtrace_events(current); | |
5988 | if (IS_ENABLED(CONFIG_DEBUG_PREEMPT) | |
5989 | && !preempt_count_equals(preempt_offset)) { | |
5990 | pr_err("Preemption disabled at:"); | |
5991 | print_ip_sym(preempt_disable_ip); | |
5992 | pr_cont("\n"); | |
5993 | } | |
5994 | dump_stack(); | |
5995 | add_taint(TAINT_WARN, LOCKDEP_STILL_OK); | |
5996 | } | |
5997 | EXPORT_SYMBOL(___might_sleep); | |
5998 | #endif | |
5999 | ||
6000 | #ifdef CONFIG_MAGIC_SYSRQ | |
6001 | void normalize_rt_tasks(void) | |
6002 | { | |
6003 | struct task_struct *g, *p; | |
6004 | struct sched_attr attr = { | |
6005 | .sched_policy = SCHED_NORMAL, | |
6006 | }; | |
6007 | ||
6008 | read_lock(&tasklist_lock); | |
6009 | for_each_process_thread(g, p) { | |
6010 | /* | |
6011 | * Only normalize user tasks: | |
6012 | */ | |
6013 | if (p->flags & PF_KTHREAD) | |
6014 | continue; | |
6015 | ||
6016 | p->se.exec_start = 0; | |
6017 | schedstat_set(p->se.statistics.wait_start, 0); | |
6018 | schedstat_set(p->se.statistics.sleep_start, 0); | |
6019 | schedstat_set(p->se.statistics.block_start, 0); | |
6020 | ||
6021 | if (!dl_task(p) && !rt_task(p)) { | |
6022 | /* | |
6023 | * Renice negative nice level userspace | |
6024 | * tasks back to 0: | |
6025 | */ | |
6026 | if (task_nice(p) < 0) | |
6027 | set_user_nice(p, 0); | |
6028 | continue; | |
6029 | } | |
6030 | ||
6031 | __sched_setscheduler(p, &attr, false, false); | |
6032 | } | |
6033 | read_unlock(&tasklist_lock); | |
6034 | } | |
6035 | ||
6036 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
6037 | ||
6038 | #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) | |
6039 | /* | |
6040 | * These functions are only useful for the IA64 MCA handling, or kdb. | |
6041 | * | |
6042 | * They can only be called when the whole system has been | |
6043 | * stopped - every CPU needs to be quiescent, and no scheduling | |
6044 | * activity can take place. Using them for anything else would | |
6045 | * be a serious bug, and as a result, they aren't even visible | |
6046 | * under any other configuration. | |
6047 | */ | |
6048 | ||
6049 | /** | |
6050 | * curr_task - return the current task for a given CPU. | |
6051 | * @cpu: the processor in question. | |
6052 | * | |
6053 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
6054 | * | |
6055 | * Return: The current task for @cpu. | |
6056 | */ | |
6057 | struct task_struct *curr_task(int cpu) | |
6058 | { | |
6059 | return cpu_curr(cpu); | |
6060 | } | |
6061 | ||
6062 | #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ | |
6063 | ||
6064 | #ifdef CONFIG_IA64 | |
6065 | /** | |
6066 | * set_curr_task - set the current task for a given CPU. | |
6067 | * @cpu: the processor in question. | |
6068 | * @p: the task pointer to set. | |
6069 | * | |
6070 | * Description: This function must only be used when non-maskable interrupts | |
6071 | * are serviced on a separate stack. It allows the architecture to switch the | |
6072 | * notion of the current task on a CPU in a non-blocking manner. This function | |
6073 | * must be called with all CPU's synchronized, and interrupts disabled, the | |
6074 | * and caller must save the original value of the current task (see | |
6075 | * curr_task() above) and restore that value before reenabling interrupts and | |
6076 | * re-starting the system. | |
6077 | * | |
6078 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
6079 | */ | |
6080 | void ia64_set_curr_task(int cpu, struct task_struct *p) | |
6081 | { | |
6082 | cpu_curr(cpu) = p; | |
6083 | } | |
6084 | ||
6085 | #endif | |
6086 | ||
6087 | #ifdef CONFIG_CGROUP_SCHED | |
6088 | /* task_group_lock serializes the addition/removal of task groups */ | |
6089 | static DEFINE_SPINLOCK(task_group_lock); | |
6090 | ||
6091 | static void sched_free_group(struct task_group *tg) | |
6092 | { | |
6093 | free_fair_sched_group(tg); | |
6094 | free_rt_sched_group(tg); | |
6095 | autogroup_free(tg); | |
6096 | kmem_cache_free(task_group_cache, tg); | |
6097 | } | |
6098 | ||
6099 | /* allocate runqueue etc for a new task group */ | |
6100 | struct task_group *sched_create_group(struct task_group *parent) | |
6101 | { | |
6102 | struct task_group *tg; | |
6103 | ||
6104 | tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO); | |
6105 | if (!tg) | |
6106 | return ERR_PTR(-ENOMEM); | |
6107 | ||
6108 | if (!alloc_fair_sched_group(tg, parent)) | |
6109 | goto err; | |
6110 | ||
6111 | if (!alloc_rt_sched_group(tg, parent)) | |
6112 | goto err; | |
6113 | ||
6114 | return tg; | |
6115 | ||
6116 | err: | |
6117 | sched_free_group(tg); | |
6118 | return ERR_PTR(-ENOMEM); | |
6119 | } | |
6120 | ||
6121 | void sched_online_group(struct task_group *tg, struct task_group *parent) | |
6122 | { | |
6123 | unsigned long flags; | |
6124 | ||
6125 | spin_lock_irqsave(&task_group_lock, flags); | |
6126 | list_add_rcu(&tg->list, &task_groups); | |
6127 | ||
6128 | /* Root should already exist: */ | |
6129 | WARN_ON(!parent); | |
6130 | ||
6131 | tg->parent = parent; | |
6132 | INIT_LIST_HEAD(&tg->children); | |
6133 | list_add_rcu(&tg->siblings, &parent->children); | |
6134 | spin_unlock_irqrestore(&task_group_lock, flags); | |
6135 | ||
6136 | online_fair_sched_group(tg); | |
6137 | } | |
6138 | ||
6139 | /* rcu callback to free various structures associated with a task group */ | |
6140 | static void sched_free_group_rcu(struct rcu_head *rhp) | |
6141 | { | |
6142 | /* Now it should be safe to free those cfs_rqs: */ | |
6143 | sched_free_group(container_of(rhp, struct task_group, rcu)); | |
6144 | } | |
6145 | ||
6146 | void sched_destroy_group(struct task_group *tg) | |
6147 | { | |
6148 | /* Wait for possible concurrent references to cfs_rqs complete: */ | |
6149 | call_rcu(&tg->rcu, sched_free_group_rcu); | |
6150 | } | |
6151 | ||
6152 | void sched_offline_group(struct task_group *tg) | |
6153 | { | |
6154 | unsigned long flags; | |
6155 | ||
6156 | /* End participation in shares distribution: */ | |
6157 | unregister_fair_sched_group(tg); | |
6158 | ||
6159 | spin_lock_irqsave(&task_group_lock, flags); | |
6160 | list_del_rcu(&tg->list); | |
6161 | list_del_rcu(&tg->siblings); | |
6162 | spin_unlock_irqrestore(&task_group_lock, flags); | |
6163 | } | |
6164 | ||
6165 | static void sched_change_group(struct task_struct *tsk, int type) | |
6166 | { | |
6167 | struct task_group *tg; | |
6168 | ||
6169 | /* | |
6170 | * All callers are synchronized by task_rq_lock(); we do not use RCU | |
6171 | * which is pointless here. Thus, we pass "true" to task_css_check() | |
6172 | * to prevent lockdep warnings. | |
6173 | */ | |
6174 | tg = container_of(task_css_check(tsk, cpu_cgrp_id, true), | |
6175 | struct task_group, css); | |
6176 | tg = autogroup_task_group(tsk, tg); | |
6177 | tsk->sched_task_group = tg; | |
6178 | ||
6179 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
6180 | if (tsk->sched_class->task_change_group) | |
6181 | tsk->sched_class->task_change_group(tsk, type); | |
6182 | else | |
6183 | #endif | |
6184 | set_task_rq(tsk, task_cpu(tsk)); | |
6185 | } | |
6186 | ||
6187 | /* | |
6188 | * Change task's runqueue when it moves between groups. | |
6189 | * | |
6190 | * The caller of this function should have put the task in its new group by | |
6191 | * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect | |
6192 | * its new group. | |
6193 | */ | |
6194 | void sched_move_task(struct task_struct *tsk) | |
6195 | { | |
6196 | int queued, running, queue_flags = | |
6197 | DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK; | |
6198 | struct rq_flags rf; | |
6199 | struct rq *rq; | |
6200 | ||
6201 | rq = task_rq_lock(tsk, &rf); | |
6202 | update_rq_clock(rq); | |
6203 | ||
6204 | running = task_current(rq, tsk); | |
6205 | queued = task_on_rq_queued(tsk); | |
6206 | ||
6207 | if (queued) | |
6208 | dequeue_task(rq, tsk, queue_flags); | |
6209 | if (running) | |
6210 | put_prev_task(rq, tsk); | |
6211 | ||
6212 | sched_change_group(tsk, TASK_MOVE_GROUP); | |
6213 | ||
6214 | if (queued) | |
6215 | enqueue_task(rq, tsk, queue_flags); | |
6216 | if (running) | |
6217 | set_curr_task(rq, tsk); | |
6218 | ||
6219 | task_rq_unlock(rq, tsk, &rf); | |
6220 | } | |
6221 | ||
6222 | static inline struct task_group *css_tg(struct cgroup_subsys_state *css) | |
6223 | { | |
6224 | return css ? container_of(css, struct task_group, css) : NULL; | |
6225 | } | |
6226 | ||
6227 | static struct cgroup_subsys_state * | |
6228 | cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
6229 | { | |
6230 | struct task_group *parent = css_tg(parent_css); | |
6231 | struct task_group *tg; | |
6232 | ||
6233 | if (!parent) { | |
6234 | /* This is early initialization for the top cgroup */ | |
6235 | return &root_task_group.css; | |
6236 | } | |
6237 | ||
6238 | tg = sched_create_group(parent); | |
6239 | if (IS_ERR(tg)) | |
6240 | return ERR_PTR(-ENOMEM); | |
6241 | ||
6242 | return &tg->css; | |
6243 | } | |
6244 | ||
6245 | /* Expose task group only after completing cgroup initialization */ | |
6246 | static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) | |
6247 | { | |
6248 | struct task_group *tg = css_tg(css); | |
6249 | struct task_group *parent = css_tg(css->parent); | |
6250 | ||
6251 | if (parent) | |
6252 | sched_online_group(tg, parent); | |
6253 | return 0; | |
6254 | } | |
6255 | ||
6256 | static void cpu_cgroup_css_released(struct cgroup_subsys_state *css) | |
6257 | { | |
6258 | struct task_group *tg = css_tg(css); | |
6259 | ||
6260 | sched_offline_group(tg); | |
6261 | } | |
6262 | ||
6263 | static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) | |
6264 | { | |
6265 | struct task_group *tg = css_tg(css); | |
6266 | ||
6267 | /* | |
6268 | * Relies on the RCU grace period between css_released() and this. | |
6269 | */ | |
6270 | sched_free_group(tg); | |
6271 | } | |
6272 | ||
6273 | /* | |
6274 | * This is called before wake_up_new_task(), therefore we really only | |
6275 | * have to set its group bits, all the other stuff does not apply. | |
6276 | */ | |
6277 | static void cpu_cgroup_fork(struct task_struct *task) | |
6278 | { | |
6279 | struct rq_flags rf; | |
6280 | struct rq *rq; | |
6281 | ||
6282 | rq = task_rq_lock(task, &rf); | |
6283 | ||
6284 | update_rq_clock(rq); | |
6285 | sched_change_group(task, TASK_SET_GROUP); | |
6286 | ||
6287 | task_rq_unlock(rq, task, &rf); | |
6288 | } | |
6289 | ||
6290 | static int cpu_cgroup_can_attach(struct cgroup_taskset *tset) | |
6291 | { | |
6292 | struct task_struct *task; | |
6293 | struct cgroup_subsys_state *css; | |
6294 | int ret = 0; | |
6295 | ||
6296 | cgroup_taskset_for_each(task, css, tset) { | |
6297 | #ifdef CONFIG_RT_GROUP_SCHED | |
6298 | if (!sched_rt_can_attach(css_tg(css), task)) | |
6299 | return -EINVAL; | |
6300 | #else | |
6301 | /* We don't support RT-tasks being in separate groups */ | |
6302 | if (task->sched_class != &fair_sched_class) | |
6303 | return -EINVAL; | |
6304 | #endif | |
6305 | /* | |
6306 | * Serialize against wake_up_new_task() such that if its | |
6307 | * running, we're sure to observe its full state. | |
6308 | */ | |
6309 | raw_spin_lock_irq(&task->pi_lock); | |
6310 | /* | |
6311 | * Avoid calling sched_move_task() before wake_up_new_task() | |
6312 | * has happened. This would lead to problems with PELT, due to | |
6313 | * move wanting to detach+attach while we're not attached yet. | |
6314 | */ | |
6315 | if (task->state == TASK_NEW) | |
6316 | ret = -EINVAL; | |
6317 | raw_spin_unlock_irq(&task->pi_lock); | |
6318 | ||
6319 | if (ret) | |
6320 | break; | |
6321 | } | |
6322 | return ret; | |
6323 | } | |
6324 | ||
6325 | static void cpu_cgroup_attach(struct cgroup_taskset *tset) | |
6326 | { | |
6327 | struct task_struct *task; | |
6328 | struct cgroup_subsys_state *css; | |
6329 | ||
6330 | cgroup_taskset_for_each(task, css, tset) | |
6331 | sched_move_task(task); | |
6332 | } | |
6333 | ||
6334 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
6335 | static int cpu_shares_write_u64(struct cgroup_subsys_state *css, | |
6336 | struct cftype *cftype, u64 shareval) | |
6337 | { | |
6338 | return sched_group_set_shares(css_tg(css), scale_load(shareval)); | |
6339 | } | |
6340 | ||
6341 | static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, | |
6342 | struct cftype *cft) | |
6343 | { | |
6344 | struct task_group *tg = css_tg(css); | |
6345 | ||
6346 | return (u64) scale_load_down(tg->shares); | |
6347 | } | |
6348 | ||
6349 | #ifdef CONFIG_CFS_BANDWIDTH | |
6350 | static DEFINE_MUTEX(cfs_constraints_mutex); | |
6351 | ||
6352 | const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ | |
6353 | const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ | |
6354 | ||
6355 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); | |
6356 | ||
6357 | static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) | |
6358 | { | |
6359 | int i, ret = 0, runtime_enabled, runtime_was_enabled; | |
6360 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
6361 | ||
6362 | if (tg == &root_task_group) | |
6363 | return -EINVAL; | |
6364 | ||
6365 | /* | |
6366 | * Ensure we have at some amount of bandwidth every period. This is | |
6367 | * to prevent reaching a state of large arrears when throttled via | |
6368 | * entity_tick() resulting in prolonged exit starvation. | |
6369 | */ | |
6370 | if (quota < min_cfs_quota_period || period < min_cfs_quota_period) | |
6371 | return -EINVAL; | |
6372 | ||
6373 | /* | |
6374 | * Likewise, bound things on the otherside by preventing insane quota | |
6375 | * periods. This also allows us to normalize in computing quota | |
6376 | * feasibility. | |
6377 | */ | |
6378 | if (period > max_cfs_quota_period) | |
6379 | return -EINVAL; | |
6380 | ||
6381 | /* | |
6382 | * Prevent race between setting of cfs_rq->runtime_enabled and | |
6383 | * unthrottle_offline_cfs_rqs(). | |
6384 | */ | |
6385 | get_online_cpus(); | |
6386 | mutex_lock(&cfs_constraints_mutex); | |
6387 | ret = __cfs_schedulable(tg, period, quota); | |
6388 | if (ret) | |
6389 | goto out_unlock; | |
6390 | ||
6391 | runtime_enabled = quota != RUNTIME_INF; | |
6392 | runtime_was_enabled = cfs_b->quota != RUNTIME_INF; | |
6393 | /* | |
6394 | * If we need to toggle cfs_bandwidth_used, off->on must occur | |
6395 | * before making related changes, and on->off must occur afterwards | |
6396 | */ | |
6397 | if (runtime_enabled && !runtime_was_enabled) | |
6398 | cfs_bandwidth_usage_inc(); | |
6399 | raw_spin_lock_irq(&cfs_b->lock); | |
6400 | cfs_b->period = ns_to_ktime(period); | |
6401 | cfs_b->quota = quota; | |
6402 | ||
6403 | __refill_cfs_bandwidth_runtime(cfs_b); | |
6404 | ||
6405 | /* Restart the period timer (if active) to handle new period expiry: */ | |
6406 | if (runtime_enabled) | |
6407 | start_cfs_bandwidth(cfs_b); | |
6408 | ||
6409 | raw_spin_unlock_irq(&cfs_b->lock); | |
6410 | ||
6411 | for_each_online_cpu(i) { | |
6412 | struct cfs_rq *cfs_rq = tg->cfs_rq[i]; | |
6413 | struct rq *rq = cfs_rq->rq; | |
6414 | struct rq_flags rf; | |
6415 | ||
6416 | rq_lock_irq(rq, &rf); | |
6417 | cfs_rq->runtime_enabled = runtime_enabled; | |
6418 | cfs_rq->runtime_remaining = 0; | |
6419 | ||
6420 | if (cfs_rq->throttled) | |
6421 | unthrottle_cfs_rq(cfs_rq); | |
6422 | rq_unlock_irq(rq, &rf); | |
6423 | } | |
6424 | if (runtime_was_enabled && !runtime_enabled) | |
6425 | cfs_bandwidth_usage_dec(); | |
6426 | out_unlock: | |
6427 | mutex_unlock(&cfs_constraints_mutex); | |
6428 | put_online_cpus(); | |
6429 | ||
6430 | return ret; | |
6431 | } | |
6432 | ||
6433 | int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) | |
6434 | { | |
6435 | u64 quota, period; | |
6436 | ||
6437 | period = ktime_to_ns(tg->cfs_bandwidth.period); | |
6438 | if (cfs_quota_us < 0) | |
6439 | quota = RUNTIME_INF; | |
6440 | else | |
6441 | quota = (u64)cfs_quota_us * NSEC_PER_USEC; | |
6442 | ||
6443 | return tg_set_cfs_bandwidth(tg, period, quota); | |
6444 | } | |
6445 | ||
6446 | long tg_get_cfs_quota(struct task_group *tg) | |
6447 | { | |
6448 | u64 quota_us; | |
6449 | ||
6450 | if (tg->cfs_bandwidth.quota == RUNTIME_INF) | |
6451 | return -1; | |
6452 | ||
6453 | quota_us = tg->cfs_bandwidth.quota; | |
6454 | do_div(quota_us, NSEC_PER_USEC); | |
6455 | ||
6456 | return quota_us; | |
6457 | } | |
6458 | ||
6459 | int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) | |
6460 | { | |
6461 | u64 quota, period; | |
6462 | ||
6463 | period = (u64)cfs_period_us * NSEC_PER_USEC; | |
6464 | quota = tg->cfs_bandwidth.quota; | |
6465 | ||
6466 | return tg_set_cfs_bandwidth(tg, period, quota); | |
6467 | } | |
6468 | ||
6469 | long tg_get_cfs_period(struct task_group *tg) | |
6470 | { | |
6471 | u64 cfs_period_us; | |
6472 | ||
6473 | cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); | |
6474 | do_div(cfs_period_us, NSEC_PER_USEC); | |
6475 | ||
6476 | return cfs_period_us; | |
6477 | } | |
6478 | ||
6479 | static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, | |
6480 | struct cftype *cft) | |
6481 | { | |
6482 | return tg_get_cfs_quota(css_tg(css)); | |
6483 | } | |
6484 | ||
6485 | static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, | |
6486 | struct cftype *cftype, s64 cfs_quota_us) | |
6487 | { | |
6488 | return tg_set_cfs_quota(css_tg(css), cfs_quota_us); | |
6489 | } | |
6490 | ||
6491 | static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, | |
6492 | struct cftype *cft) | |
6493 | { | |
6494 | return tg_get_cfs_period(css_tg(css)); | |
6495 | } | |
6496 | ||
6497 | static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, | |
6498 | struct cftype *cftype, u64 cfs_period_us) | |
6499 | { | |
6500 | return tg_set_cfs_period(css_tg(css), cfs_period_us); | |
6501 | } | |
6502 | ||
6503 | struct cfs_schedulable_data { | |
6504 | struct task_group *tg; | |
6505 | u64 period, quota; | |
6506 | }; | |
6507 | ||
6508 | /* | |
6509 | * normalize group quota/period to be quota/max_period | |
6510 | * note: units are usecs | |
6511 | */ | |
6512 | static u64 normalize_cfs_quota(struct task_group *tg, | |
6513 | struct cfs_schedulable_data *d) | |
6514 | { | |
6515 | u64 quota, period; | |
6516 | ||
6517 | if (tg == d->tg) { | |
6518 | period = d->period; | |
6519 | quota = d->quota; | |
6520 | } else { | |
6521 | period = tg_get_cfs_period(tg); | |
6522 | quota = tg_get_cfs_quota(tg); | |
6523 | } | |
6524 | ||
6525 | /* note: these should typically be equivalent */ | |
6526 | if (quota == RUNTIME_INF || quota == -1) | |
6527 | return RUNTIME_INF; | |
6528 | ||
6529 | return to_ratio(period, quota); | |
6530 | } | |
6531 | ||
6532 | static int tg_cfs_schedulable_down(struct task_group *tg, void *data) | |
6533 | { | |
6534 | struct cfs_schedulable_data *d = data; | |
6535 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
6536 | s64 quota = 0, parent_quota = -1; | |
6537 | ||
6538 | if (!tg->parent) { | |
6539 | quota = RUNTIME_INF; | |
6540 | } else { | |
6541 | struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; | |
6542 | ||
6543 | quota = normalize_cfs_quota(tg, d); | |
6544 | parent_quota = parent_b->hierarchical_quota; | |
6545 | ||
6546 | /* | |
6547 | * Ensure max(child_quota) <= parent_quota, inherit when no | |
6548 | * limit is set: | |
6549 | */ | |
6550 | if (quota == RUNTIME_INF) | |
6551 | quota = parent_quota; | |
6552 | else if (parent_quota != RUNTIME_INF && quota > parent_quota) | |
6553 | return -EINVAL; | |
6554 | } | |
6555 | cfs_b->hierarchical_quota = quota; | |
6556 | ||
6557 | return 0; | |
6558 | } | |
6559 | ||
6560 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) | |
6561 | { | |
6562 | int ret; | |
6563 | struct cfs_schedulable_data data = { | |
6564 | .tg = tg, | |
6565 | .period = period, | |
6566 | .quota = quota, | |
6567 | }; | |
6568 | ||
6569 | if (quota != RUNTIME_INF) { | |
6570 | do_div(data.period, NSEC_PER_USEC); | |
6571 | do_div(data.quota, NSEC_PER_USEC); | |
6572 | } | |
6573 | ||
6574 | rcu_read_lock(); | |
6575 | ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); | |
6576 | rcu_read_unlock(); | |
6577 | ||
6578 | return ret; | |
6579 | } | |
6580 | ||
6581 | static int cpu_stats_show(struct seq_file *sf, void *v) | |
6582 | { | |
6583 | struct task_group *tg = css_tg(seq_css(sf)); | |
6584 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
6585 | ||
6586 | seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); | |
6587 | seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); | |
6588 | seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); | |
6589 | ||
6590 | return 0; | |
6591 | } | |
6592 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
6593 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
6594 | ||
6595 | #ifdef CONFIG_RT_GROUP_SCHED | |
6596 | static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, | |
6597 | struct cftype *cft, s64 val) | |
6598 | { | |
6599 | return sched_group_set_rt_runtime(css_tg(css), val); | |
6600 | } | |
6601 | ||
6602 | static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, | |
6603 | struct cftype *cft) | |
6604 | { | |
6605 | return sched_group_rt_runtime(css_tg(css)); | |
6606 | } | |
6607 | ||
6608 | static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, | |
6609 | struct cftype *cftype, u64 rt_period_us) | |
6610 | { | |
6611 | return sched_group_set_rt_period(css_tg(css), rt_period_us); | |
6612 | } | |
6613 | ||
6614 | static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, | |
6615 | struct cftype *cft) | |
6616 | { | |
6617 | return sched_group_rt_period(css_tg(css)); | |
6618 | } | |
6619 | #endif /* CONFIG_RT_GROUP_SCHED */ | |
6620 | ||
6621 | static struct cftype cpu_files[] = { | |
6622 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
6623 | { | |
6624 | .name = "shares", | |
6625 | .read_u64 = cpu_shares_read_u64, | |
6626 | .write_u64 = cpu_shares_write_u64, | |
6627 | }, | |
6628 | #endif | |
6629 | #ifdef CONFIG_CFS_BANDWIDTH | |
6630 | { | |
6631 | .name = "cfs_quota_us", | |
6632 | .read_s64 = cpu_cfs_quota_read_s64, | |
6633 | .write_s64 = cpu_cfs_quota_write_s64, | |
6634 | }, | |
6635 | { | |
6636 | .name = "cfs_period_us", | |
6637 | .read_u64 = cpu_cfs_period_read_u64, | |
6638 | .write_u64 = cpu_cfs_period_write_u64, | |
6639 | }, | |
6640 | { | |
6641 | .name = "stat", | |
6642 | .seq_show = cpu_stats_show, | |
6643 | }, | |
6644 | #endif | |
6645 | #ifdef CONFIG_RT_GROUP_SCHED | |
6646 | { | |
6647 | .name = "rt_runtime_us", | |
6648 | .read_s64 = cpu_rt_runtime_read, | |
6649 | .write_s64 = cpu_rt_runtime_write, | |
6650 | }, | |
6651 | { | |
6652 | .name = "rt_period_us", | |
6653 | .read_u64 = cpu_rt_period_read_uint, | |
6654 | .write_u64 = cpu_rt_period_write_uint, | |
6655 | }, | |
6656 | #endif | |
6657 | { } /* Terminate */ | |
6658 | }; | |
6659 | ||
6660 | struct cgroup_subsys cpu_cgrp_subsys = { | |
6661 | .css_alloc = cpu_cgroup_css_alloc, | |
6662 | .css_online = cpu_cgroup_css_online, | |
6663 | .css_released = cpu_cgroup_css_released, | |
6664 | .css_free = cpu_cgroup_css_free, | |
6665 | .fork = cpu_cgroup_fork, | |
6666 | .can_attach = cpu_cgroup_can_attach, | |
6667 | .attach = cpu_cgroup_attach, | |
6668 | .legacy_cftypes = cpu_files, | |
6669 | .early_init = true, | |
6670 | }; | |
6671 | ||
6672 | #endif /* CONFIG_CGROUP_SCHED */ | |
6673 | ||
6674 | void dump_cpu_task(int cpu) | |
6675 | { | |
6676 | pr_info("Task dump for CPU %d:\n", cpu); | |
6677 | sched_show_task(cpu_curr(cpu)); | |
6678 | } | |
6679 | ||
6680 | /* | |
6681 | * Nice levels are multiplicative, with a gentle 10% change for every | |
6682 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to | |
6683 | * nice 1, it will get ~10% less CPU time than another CPU-bound task | |
6684 | * that remained on nice 0. | |
6685 | * | |
6686 | * The "10% effect" is relative and cumulative: from _any_ nice level, | |
6687 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level | |
6688 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. | |
6689 | * If a task goes up by ~10% and another task goes down by ~10% then | |
6690 | * the relative distance between them is ~25%.) | |
6691 | */ | |
6692 | const int sched_prio_to_weight[40] = { | |
6693 | /* -20 */ 88761, 71755, 56483, 46273, 36291, | |
6694 | /* -15 */ 29154, 23254, 18705, 14949, 11916, | |
6695 | /* -10 */ 9548, 7620, 6100, 4904, 3906, | |
6696 | /* -5 */ 3121, 2501, 1991, 1586, 1277, | |
6697 | /* 0 */ 1024, 820, 655, 526, 423, | |
6698 | /* 5 */ 335, 272, 215, 172, 137, | |
6699 | /* 10 */ 110, 87, 70, 56, 45, | |
6700 | /* 15 */ 36, 29, 23, 18, 15, | |
6701 | }; | |
6702 | ||
6703 | /* | |
6704 | * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated. | |
6705 | * | |
6706 | * In cases where the weight does not change often, we can use the | |
6707 | * precalculated inverse to speed up arithmetics by turning divisions | |
6708 | * into multiplications: | |
6709 | */ | |
6710 | const u32 sched_prio_to_wmult[40] = { | |
6711 | /* -20 */ 48388, 59856, 76040, 92818, 118348, | |
6712 | /* -15 */ 147320, 184698, 229616, 287308, 360437, | |
6713 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, | |
6714 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, | |
6715 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, | |
6716 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, | |
6717 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, | |
6718 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, | |
6719 | }; |