xref: /openbmc/linux/kernel/sched/core.c (revision d7a3d85e)
1 /*
2  *  kernel/sched/core.c
3  *
4  *  Kernel scheduler and related syscalls
5  *
6  *  Copyright (C) 1991-2002  Linus Torvalds
7  *
8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
9  *		make semaphores SMP safe
10  *  1998-11-19	Implemented schedule_timeout() and related stuff
11  *		by Andrea Arcangeli
12  *  2002-01-04	New ultra-scalable O(1) scheduler by Ingo Molnar:
13  *		hybrid priority-list and round-robin design with
14  *		an array-switch method of distributing timeslices
15  *		and per-CPU runqueues.  Cleanups and useful suggestions
16  *		by Davide Libenzi, preemptible kernel bits by Robert Love.
17  *  2003-09-03	Interactivity tuning by Con Kolivas.
18  *  2004-04-02	Scheduler domains code by Nick Piggin
19  *  2007-04-15  Work begun on replacing all interactivity tuning with a
20  *              fair scheduling design by Con Kolivas.
21  *  2007-05-05  Load balancing (smp-nice) and other improvements
22  *              by Peter Williams
23  *  2007-05-06  Interactivity improvements to CFS by Mike Galbraith
24  *  2007-07-01  Group scheduling enhancements by Srivatsa Vaddagiri
25  *  2007-11-29  RT balancing improvements by Steven Rostedt, Gregory Haskins,
26  *              Thomas Gleixner, Mike Kravetz
27  */
28 
29 #include <linux/mm.h>
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
77 
78 #include <asm/switch_to.h>
79 #include <asm/tlb.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
84 #endif
85 
86 #include "sched.h"
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
89 
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
92 
93 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
94 {
95 	unsigned long delta;
96 	ktime_t soft, hard, now;
97 
98 	for (;;) {
99 		if (hrtimer_active(period_timer))
100 			break;
101 
102 		now = hrtimer_cb_get_time(period_timer);
103 		hrtimer_forward(period_timer, now, period);
104 
105 		soft = hrtimer_get_softexpires(period_timer);
106 		hard = hrtimer_get_expires(period_timer);
107 		delta = ktime_to_ns(ktime_sub(hard, soft));
108 		__hrtimer_start_range_ns(period_timer, soft, delta,
109 					 HRTIMER_MODE_ABS_PINNED, 0);
110 	}
111 }
112 
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 
118 void update_rq_clock(struct rq *rq)
119 {
120 	s64 delta;
121 
122 	lockdep_assert_held(&rq->lock);
123 
124 	if (rq->clock_skip_update & RQCF_ACT_SKIP)
125 		return;
126 
127 	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
128 	if (delta < 0)
129 		return;
130 	rq->clock += delta;
131 	update_rq_clock_task(rq, delta);
132 }
133 
134 /*
135  * Debugging: various feature bits
136  */
137 
138 #define SCHED_FEAT(name, enabled)	\
139 	(1UL << __SCHED_FEAT_##name) * enabled |
140 
141 const_debug unsigned int sysctl_sched_features =
142 #include "features.h"
143 	0;
144 
145 #undef SCHED_FEAT
146 
147 #ifdef CONFIG_SCHED_DEBUG
148 #define SCHED_FEAT(name, enabled)	\
149 	#name ,
150 
151 static const char * const sched_feat_names[] = {
152 #include "features.h"
153 };
154 
155 #undef SCHED_FEAT
156 
157 static int sched_feat_show(struct seq_file *m, void *v)
158 {
159 	int i;
160 
161 	for (i = 0; i < __SCHED_FEAT_NR; i++) {
162 		if (!(sysctl_sched_features & (1UL << i)))
163 			seq_puts(m, "NO_");
164 		seq_printf(m, "%s ", sched_feat_names[i]);
165 	}
166 	seq_puts(m, "\n");
167 
168 	return 0;
169 }
170 
171 #ifdef HAVE_JUMP_LABEL
172 
173 #define jump_label_key__true  STATIC_KEY_INIT_TRUE
174 #define jump_label_key__false STATIC_KEY_INIT_FALSE
175 
176 #define SCHED_FEAT(name, enabled)	\
177 	jump_label_key__##enabled ,
178 
179 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
180 #include "features.h"
181 };
182 
183 #undef SCHED_FEAT
184 
185 static void sched_feat_disable(int i)
186 {
187 	if (static_key_enabled(&sched_feat_keys[i]))
188 		static_key_slow_dec(&sched_feat_keys[i]);
189 }
190 
191 static void sched_feat_enable(int i)
192 {
193 	if (!static_key_enabled(&sched_feat_keys[i]))
194 		static_key_slow_inc(&sched_feat_keys[i]);
195 }
196 #else
197 static void sched_feat_disable(int i) { };
198 static void sched_feat_enable(int i) { };
199 #endif /* HAVE_JUMP_LABEL */
200 
201 static int sched_feat_set(char *cmp)
202 {
203 	int i;
204 	int neg = 0;
205 
206 	if (strncmp(cmp, "NO_", 3) == 0) {
207 		neg = 1;
208 		cmp += 3;
209 	}
210 
211 	for (i = 0; i < __SCHED_FEAT_NR; i++) {
212 		if (strcmp(cmp, sched_feat_names[i]) == 0) {
213 			if (neg) {
214 				sysctl_sched_features &= ~(1UL << i);
215 				sched_feat_disable(i);
216 			} else {
217 				sysctl_sched_features |= (1UL << i);
218 				sched_feat_enable(i);
219 			}
220 			break;
221 		}
222 	}
223 
224 	return i;
225 }
226 
227 static ssize_t
228 sched_feat_write(struct file *filp, const char __user *ubuf,
229 		size_t cnt, loff_t *ppos)
230 {
231 	char buf[64];
232 	char *cmp;
233 	int i;
234 	struct inode *inode;
235 
236 	if (cnt > 63)
237 		cnt = 63;
238 
239 	if (copy_from_user(&buf, ubuf, cnt))
240 		return -EFAULT;
241 
242 	buf[cnt] = 0;
243 	cmp = strstrip(buf);
244 
245 	/* Ensure the static_key remains in a consistent state */
246 	inode = file_inode(filp);
247 	mutex_lock(&inode->i_mutex);
248 	i = sched_feat_set(cmp);
249 	mutex_unlock(&inode->i_mutex);
250 	if (i == __SCHED_FEAT_NR)
251 		return -EINVAL;
252 
253 	*ppos += cnt;
254 
255 	return cnt;
256 }
257 
258 static int sched_feat_open(struct inode *inode, struct file *filp)
259 {
260 	return single_open(filp, sched_feat_show, NULL);
261 }
262 
263 static const struct file_operations sched_feat_fops = {
264 	.open		= sched_feat_open,
265 	.write		= sched_feat_write,
266 	.read		= seq_read,
267 	.llseek		= seq_lseek,
268 	.release	= single_release,
269 };
270 
271 static __init int sched_init_debug(void)
272 {
273 	debugfs_create_file("sched_features", 0644, NULL, NULL,
274 			&sched_feat_fops);
275 
276 	return 0;
277 }
278 late_initcall(sched_init_debug);
279 #endif /* CONFIG_SCHED_DEBUG */
280 
281 /*
282  * Number of tasks to iterate in a single balance run.
283  * Limited because this is done with IRQs disabled.
284  */
285 const_debug unsigned int sysctl_sched_nr_migrate = 32;
286 
287 /*
288  * period over which we average the RT time consumption, measured
289  * in ms.
290  *
291  * default: 1s
292  */
293 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
294 
295 /*
296  * period over which we measure -rt task cpu usage in us.
297  * default: 1s
298  */
299 unsigned int sysctl_sched_rt_period = 1000000;
300 
301 __read_mostly int scheduler_running;
302 
303 /*
304  * part of the period that we allow rt tasks to run in us.
305  * default: 0.95s
306  */
307 int sysctl_sched_rt_runtime = 950000;
308 
309 /* cpus with isolated domains */
310 cpumask_var_t cpu_isolated_map;
311 
312 /*
313  * this_rq_lock - lock this runqueue and disable interrupts.
314  */
315 static struct rq *this_rq_lock(void)
316 	__acquires(rq->lock)
317 {
318 	struct rq *rq;
319 
320 	local_irq_disable();
321 	rq = this_rq();
322 	raw_spin_lock(&rq->lock);
323 
324 	return rq;
325 }
326 
327 #ifdef CONFIG_SCHED_HRTICK
328 /*
329  * Use HR-timers to deliver accurate preemption points.
330  */
331 
332 static void hrtick_clear(struct rq *rq)
333 {
334 	if (hrtimer_active(&rq->hrtick_timer))
335 		hrtimer_cancel(&rq->hrtick_timer);
336 }
337 
338 /*
339  * High-resolution timer tick.
340  * Runs from hardirq context with interrupts disabled.
341  */
342 static enum hrtimer_restart hrtick(struct hrtimer *timer)
343 {
344 	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
345 
346 	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
347 
348 	raw_spin_lock(&rq->lock);
349 	update_rq_clock(rq);
350 	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
351 	raw_spin_unlock(&rq->lock);
352 
353 	return HRTIMER_NORESTART;
354 }
355 
356 #ifdef CONFIG_SMP
357 
358 static int __hrtick_restart(struct rq *rq)
359 {
360 	struct hrtimer *timer = &rq->hrtick_timer;
361 	ktime_t time = hrtimer_get_softexpires(timer);
362 
363 	return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
364 }
365 
366 /*
367  * called from hardirq (IPI) context
368  */
369 static void __hrtick_start(void *arg)
370 {
371 	struct rq *rq = arg;
372 
373 	raw_spin_lock(&rq->lock);
374 	__hrtick_restart(rq);
375 	rq->hrtick_csd_pending = 0;
376 	raw_spin_unlock(&rq->lock);
377 }
378 
379 /*
380  * Called to set the hrtick timer state.
381  *
382  * called with rq->lock held and irqs disabled
383  */
384 void hrtick_start(struct rq *rq, u64 delay)
385 {
386 	struct hrtimer *timer = &rq->hrtick_timer;
387 	ktime_t time;
388 	s64 delta;
389 
390 	/*
391 	 * Don't schedule slices shorter than 10000ns, that just
392 	 * doesn't make sense and can cause timer DoS.
393 	 */
394 	delta = max_t(s64, delay, 10000LL);
395 	time = ktime_add_ns(timer->base->get_time(), delta);
396 
397 	hrtimer_set_expires(timer, time);
398 
399 	if (rq == this_rq()) {
400 		__hrtick_restart(rq);
401 	} else if (!rq->hrtick_csd_pending) {
402 		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
403 		rq->hrtick_csd_pending = 1;
404 	}
405 }
406 
407 static int
408 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
409 {
410 	int cpu = (int)(long)hcpu;
411 
412 	switch (action) {
413 	case CPU_UP_CANCELED:
414 	case CPU_UP_CANCELED_FROZEN:
415 	case CPU_DOWN_PREPARE:
416 	case CPU_DOWN_PREPARE_FROZEN:
417 	case CPU_DEAD:
418 	case CPU_DEAD_FROZEN:
419 		hrtick_clear(cpu_rq(cpu));
420 		return NOTIFY_OK;
421 	}
422 
423 	return NOTIFY_DONE;
424 }
425 
426 static __init void init_hrtick(void)
427 {
428 	hotcpu_notifier(hotplug_hrtick, 0);
429 }
430 #else
431 /*
432  * Called to set the hrtick timer state.
433  *
434  * called with rq->lock held and irqs disabled
435  */
436 void hrtick_start(struct rq *rq, u64 delay)
437 {
438 	/*
439 	 * Don't schedule slices shorter than 10000ns, that just
440 	 * doesn't make sense. Rely on vruntime for fairness.
441 	 */
442 	delay = max_t(u64, delay, 10000LL);
443 	__hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
444 			HRTIMER_MODE_REL_PINNED, 0);
445 }
446 
447 static inline void init_hrtick(void)
448 {
449 }
450 #endif /* CONFIG_SMP */
451 
452 static void init_rq_hrtick(struct rq *rq)
453 {
454 #ifdef CONFIG_SMP
455 	rq->hrtick_csd_pending = 0;
456 
457 	rq->hrtick_csd.flags = 0;
458 	rq->hrtick_csd.func = __hrtick_start;
459 	rq->hrtick_csd.info = rq;
460 #endif
461 
462 	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
463 	rq->hrtick_timer.function = hrtick;
464 }
465 #else	/* CONFIG_SCHED_HRTICK */
466 static inline void hrtick_clear(struct rq *rq)
467 {
468 }
469 
470 static inline void init_rq_hrtick(struct rq *rq)
471 {
472 }
473 
474 static inline void init_hrtick(void)
475 {
476 }
477 #endif	/* CONFIG_SCHED_HRTICK */
478 
479 /*
480  * cmpxchg based fetch_or, macro so it works for different integer types
481  */
482 #define fetch_or(ptr, val)						\
483 ({	typeof(*(ptr)) __old, __val = *(ptr);				\
484  	for (;;) {							\
485  		__old = cmpxchg((ptr), __val, __val | (val));		\
486  		if (__old == __val)					\
487  			break;						\
488  		__val = __old;						\
489  	}								\
490  	__old;								\
491 })
492 
493 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
494 /*
495  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
496  * this avoids any races wrt polling state changes and thereby avoids
497  * spurious IPIs.
498  */
499 static bool set_nr_and_not_polling(struct task_struct *p)
500 {
501 	struct thread_info *ti = task_thread_info(p);
502 	return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
503 }
504 
505 /*
506  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
507  *
508  * If this returns true, then the idle task promises to call
509  * sched_ttwu_pending() and reschedule soon.
510  */
511 static bool set_nr_if_polling(struct task_struct *p)
512 {
513 	struct thread_info *ti = task_thread_info(p);
514 	typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
515 
516 	for (;;) {
517 		if (!(val & _TIF_POLLING_NRFLAG))
518 			return false;
519 		if (val & _TIF_NEED_RESCHED)
520 			return true;
521 		old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
522 		if (old == val)
523 			break;
524 		val = old;
525 	}
526 	return true;
527 }
528 
529 #else
530 static bool set_nr_and_not_polling(struct task_struct *p)
531 {
532 	set_tsk_need_resched(p);
533 	return true;
534 }
535 
536 #ifdef CONFIG_SMP
537 static bool set_nr_if_polling(struct task_struct *p)
538 {
539 	return false;
540 }
541 #endif
542 #endif
543 
544 /*
545  * resched_curr - mark rq's current task 'to be rescheduled now'.
546  *
547  * On UP this means the setting of the need_resched flag, on SMP it
548  * might also involve a cross-CPU call to trigger the scheduler on
549  * the target CPU.
550  */
551 void resched_curr(struct rq *rq)
552 {
553 	struct task_struct *curr = rq->curr;
554 	int cpu;
555 
556 	lockdep_assert_held(&rq->lock);
557 
558 	if (test_tsk_need_resched(curr))
559 		return;
560 
561 	cpu = cpu_of(rq);
562 
563 	if (cpu == smp_processor_id()) {
564 		set_tsk_need_resched(curr);
565 		set_preempt_need_resched();
566 		return;
567 	}
568 
569 	if (set_nr_and_not_polling(curr))
570 		smp_send_reschedule(cpu);
571 	else
572 		trace_sched_wake_idle_without_ipi(cpu);
573 }
574 
575 void resched_cpu(int cpu)
576 {
577 	struct rq *rq = cpu_rq(cpu);
578 	unsigned long flags;
579 
580 	if (!raw_spin_trylock_irqsave(&rq->lock, flags))
581 		return;
582 	resched_curr(rq);
583 	raw_spin_unlock_irqrestore(&rq->lock, flags);
584 }
585 
586 #ifdef CONFIG_SMP
587 #ifdef CONFIG_NO_HZ_COMMON
588 /*
589  * In the semi idle case, use the nearest busy cpu for migrating timers
590  * from an idle cpu.  This is good for power-savings.
591  *
592  * We don't do similar optimization for completely idle system, as
593  * selecting an idle cpu will add more delays to the timers than intended
594  * (as that cpu's timer base may not be uptodate wrt jiffies etc).
595  */
596 int get_nohz_timer_target(int pinned)
597 {
598 	int cpu = smp_processor_id();
599 	int i;
600 	struct sched_domain *sd;
601 
602 	if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
603 		return cpu;
604 
605 	rcu_read_lock();
606 	for_each_domain(cpu, sd) {
607 		for_each_cpu(i, sched_domain_span(sd)) {
608 			if (!idle_cpu(i)) {
609 				cpu = i;
610 				goto unlock;
611 			}
612 		}
613 	}
614 unlock:
615 	rcu_read_unlock();
616 	return cpu;
617 }
618 /*
619  * When add_timer_on() enqueues a timer into the timer wheel of an
620  * idle CPU then this timer might expire before the next timer event
621  * which is scheduled to wake up that CPU. In case of a completely
622  * idle system the next event might even be infinite time into the
623  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
624  * leaves the inner idle loop so the newly added timer is taken into
625  * account when the CPU goes back to idle and evaluates the timer
626  * wheel for the next timer event.
627  */
628 static void wake_up_idle_cpu(int cpu)
629 {
630 	struct rq *rq = cpu_rq(cpu);
631 
632 	if (cpu == smp_processor_id())
633 		return;
634 
635 	if (set_nr_and_not_polling(rq->idle))
636 		smp_send_reschedule(cpu);
637 	else
638 		trace_sched_wake_idle_without_ipi(cpu);
639 }
640 
641 static bool wake_up_full_nohz_cpu(int cpu)
642 {
643 	/*
644 	 * We just need the target to call irq_exit() and re-evaluate
645 	 * the next tick. The nohz full kick at least implies that.
646 	 * If needed we can still optimize that later with an
647 	 * empty IRQ.
648 	 */
649 	if (tick_nohz_full_cpu(cpu)) {
650 		if (cpu != smp_processor_id() ||
651 		    tick_nohz_tick_stopped())
652 			tick_nohz_full_kick_cpu(cpu);
653 		return true;
654 	}
655 
656 	return false;
657 }
658 
659 void wake_up_nohz_cpu(int cpu)
660 {
661 	if (!wake_up_full_nohz_cpu(cpu))
662 		wake_up_idle_cpu(cpu);
663 }
664 
665 static inline bool got_nohz_idle_kick(void)
666 {
667 	int cpu = smp_processor_id();
668 
669 	if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
670 		return false;
671 
672 	if (idle_cpu(cpu) && !need_resched())
673 		return true;
674 
675 	/*
676 	 * We can't run Idle Load Balance on this CPU for this time so we
677 	 * cancel it and clear NOHZ_BALANCE_KICK
678 	 */
679 	clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
680 	return false;
681 }
682 
683 #else /* CONFIG_NO_HZ_COMMON */
684 
685 static inline bool got_nohz_idle_kick(void)
686 {
687 	return false;
688 }
689 
690 #endif /* CONFIG_NO_HZ_COMMON */
691 
692 #ifdef CONFIG_NO_HZ_FULL
693 bool sched_can_stop_tick(void)
694 {
695 	/*
696 	 * FIFO realtime policy runs the highest priority task. Other runnable
697 	 * tasks are of a lower priority. The scheduler tick does nothing.
698 	 */
699 	if (current->policy == SCHED_FIFO)
700 		return true;
701 
702 	/*
703 	 * Round-robin realtime tasks time slice with other tasks at the same
704 	 * realtime priority. Is this task the only one at this priority?
705 	 */
706 	if (current->policy == SCHED_RR) {
707 		struct sched_rt_entity *rt_se = &current->rt;
708 
709 		return rt_se->run_list.prev == rt_se->run_list.next;
710 	}
711 
712 	/*
713 	 * More than one running task need preemption.
714 	 * nr_running update is assumed to be visible
715 	 * after IPI is sent from wakers.
716 	 */
717 	if (this_rq()->nr_running > 1)
718 		return false;
719 
720 	return true;
721 }
722 #endif /* CONFIG_NO_HZ_FULL */
723 
724 void sched_avg_update(struct rq *rq)
725 {
726 	s64 period = sched_avg_period();
727 
728 	while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
729 		/*
730 		 * Inline assembly required to prevent the compiler
731 		 * optimising this loop into a divmod call.
732 		 * See __iter_div_u64_rem() for another example of this.
733 		 */
734 		asm("" : "+rm" (rq->age_stamp));
735 		rq->age_stamp += period;
736 		rq->rt_avg /= 2;
737 	}
738 }
739 
740 #endif /* CONFIG_SMP */
741 
742 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
743 			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
744 /*
745  * Iterate task_group tree rooted at *from, calling @down when first entering a
746  * node and @up when leaving it for the final time.
747  *
748  * Caller must hold rcu_lock or sufficient equivalent.
749  */
750 int walk_tg_tree_from(struct task_group *from,
751 			     tg_visitor down, tg_visitor up, void *data)
752 {
753 	struct task_group *parent, *child;
754 	int ret;
755 
756 	parent = from;
757 
758 down:
759 	ret = (*down)(parent, data);
760 	if (ret)
761 		goto out;
762 	list_for_each_entry_rcu(child, &parent->children, siblings) {
763 		parent = child;
764 		goto down;
765 
766 up:
767 		continue;
768 	}
769 	ret = (*up)(parent, data);
770 	if (ret || parent == from)
771 		goto out;
772 
773 	child = parent;
774 	parent = parent->parent;
775 	if (parent)
776 		goto up;
777 out:
778 	return ret;
779 }
780 
781 int tg_nop(struct task_group *tg, void *data)
782 {
783 	return 0;
784 }
785 #endif
786 
787 static void set_load_weight(struct task_struct *p)
788 {
789 	int prio = p->static_prio - MAX_RT_PRIO;
790 	struct load_weight *load = &p->se.load;
791 
792 	/*
793 	 * SCHED_IDLE tasks get minimal weight:
794 	 */
795 	if (p->policy == SCHED_IDLE) {
796 		load->weight = scale_load(WEIGHT_IDLEPRIO);
797 		load->inv_weight = WMULT_IDLEPRIO;
798 		return;
799 	}
800 
801 	load->weight = scale_load(prio_to_weight[prio]);
802 	load->inv_weight = prio_to_wmult[prio];
803 }
804 
805 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
806 {
807 	update_rq_clock(rq);
808 	sched_info_queued(rq, p);
809 	p->sched_class->enqueue_task(rq, p, flags);
810 }
811 
812 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
813 {
814 	update_rq_clock(rq);
815 	sched_info_dequeued(rq, p);
816 	p->sched_class->dequeue_task(rq, p, flags);
817 }
818 
819 void activate_task(struct rq *rq, struct task_struct *p, int flags)
820 {
821 	if (task_contributes_to_load(p))
822 		rq->nr_uninterruptible--;
823 
824 	enqueue_task(rq, p, flags);
825 }
826 
827 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
828 {
829 	if (task_contributes_to_load(p))
830 		rq->nr_uninterruptible++;
831 
832 	dequeue_task(rq, p, flags);
833 }
834 
835 static void update_rq_clock_task(struct rq *rq, s64 delta)
836 {
837 /*
838  * In theory, the compile should just see 0 here, and optimize out the call
839  * to sched_rt_avg_update. But I don't trust it...
840  */
841 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
842 	s64 steal = 0, irq_delta = 0;
843 #endif
844 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
845 	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
846 
847 	/*
848 	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
849 	 * this case when a previous update_rq_clock() happened inside a
850 	 * {soft,}irq region.
851 	 *
852 	 * When this happens, we stop ->clock_task and only update the
853 	 * prev_irq_time stamp to account for the part that fit, so that a next
854 	 * update will consume the rest. This ensures ->clock_task is
855 	 * monotonic.
856 	 *
857 	 * It does however cause some slight miss-attribution of {soft,}irq
858 	 * time, a more accurate solution would be to update the irq_time using
859 	 * the current rq->clock timestamp, except that would require using
860 	 * atomic ops.
861 	 */
862 	if (irq_delta > delta)
863 		irq_delta = delta;
864 
865 	rq->prev_irq_time += irq_delta;
866 	delta -= irq_delta;
867 #endif
868 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
869 	if (static_key_false((&paravirt_steal_rq_enabled))) {
870 		steal = paravirt_steal_clock(cpu_of(rq));
871 		steal -= rq->prev_steal_time_rq;
872 
873 		if (unlikely(steal > delta))
874 			steal = delta;
875 
876 		rq->prev_steal_time_rq += steal;
877 		delta -= steal;
878 	}
879 #endif
880 
881 	rq->clock_task += delta;
882 
883 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
884 	if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
885 		sched_rt_avg_update(rq, irq_delta + steal);
886 #endif
887 }
888 
889 void sched_set_stop_task(int cpu, struct task_struct *stop)
890 {
891 	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
892 	struct task_struct *old_stop = cpu_rq(cpu)->stop;
893 
894 	if (stop) {
895 		/*
896 		 * Make it appear like a SCHED_FIFO task, its something
897 		 * userspace knows about and won't get confused about.
898 		 *
899 		 * Also, it will make PI more or less work without too
900 		 * much confusion -- but then, stop work should not
901 		 * rely on PI working anyway.
902 		 */
903 		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
904 
905 		stop->sched_class = &stop_sched_class;
906 	}
907 
908 	cpu_rq(cpu)->stop = stop;
909 
910 	if (old_stop) {
911 		/*
912 		 * Reset it back to a normal scheduling class so that
913 		 * it can die in pieces.
914 		 */
915 		old_stop->sched_class = &rt_sched_class;
916 	}
917 }
918 
919 /*
920  * __normal_prio - return the priority that is based on the static prio
921  */
922 static inline int __normal_prio(struct task_struct *p)
923 {
924 	return p->static_prio;
925 }
926 
927 /*
928  * Calculate the expected normal priority: i.e. priority
929  * without taking RT-inheritance into account. Might be
930  * boosted by interactivity modifiers. Changes upon fork,
931  * setprio syscalls, and whenever the interactivity
932  * estimator recalculates.
933  */
934 static inline int normal_prio(struct task_struct *p)
935 {
936 	int prio;
937 
938 	if (task_has_dl_policy(p))
939 		prio = MAX_DL_PRIO-1;
940 	else if (task_has_rt_policy(p))
941 		prio = MAX_RT_PRIO-1 - p->rt_priority;
942 	else
943 		prio = __normal_prio(p);
944 	return prio;
945 }
946 
947 /*
948  * Calculate the current priority, i.e. the priority
949  * taken into account by the scheduler. This value might
950  * be boosted by RT tasks, or might be boosted by
951  * interactivity modifiers. Will be RT if the task got
952  * RT-boosted. If not then it returns p->normal_prio.
953  */
954 static int effective_prio(struct task_struct *p)
955 {
956 	p->normal_prio = normal_prio(p);
957 	/*
958 	 * If we are RT tasks or we were boosted to RT priority,
959 	 * keep the priority unchanged. Otherwise, update priority
960 	 * to the normal priority:
961 	 */
962 	if (!rt_prio(p->prio))
963 		return p->normal_prio;
964 	return p->prio;
965 }
966 
967 /**
968  * task_curr - is this task currently executing on a CPU?
969  * @p: the task in question.
970  *
971  * Return: 1 if the task is currently executing. 0 otherwise.
972  */
973 inline int task_curr(const struct task_struct *p)
974 {
975 	return cpu_curr(task_cpu(p)) == p;
976 }
977 
978 /*
979  * Can drop rq->lock because from sched_class::switched_from() methods drop it.
980  */
981 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
982 				       const struct sched_class *prev_class,
983 				       int oldprio)
984 {
985 	if (prev_class != p->sched_class) {
986 		if (prev_class->switched_from)
987 			prev_class->switched_from(rq, p);
988 		/* Possble rq->lock 'hole'.  */
989 		p->sched_class->switched_to(rq, p);
990 	} else if (oldprio != p->prio || dl_task(p))
991 		p->sched_class->prio_changed(rq, p, oldprio);
992 }
993 
994 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
995 {
996 	const struct sched_class *class;
997 
998 	if (p->sched_class == rq->curr->sched_class) {
999 		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1000 	} else {
1001 		for_each_class(class) {
1002 			if (class == rq->curr->sched_class)
1003 				break;
1004 			if (class == p->sched_class) {
1005 				resched_curr(rq);
1006 				break;
1007 			}
1008 		}
1009 	}
1010 
1011 	/*
1012 	 * A queue event has occurred, and we're going to schedule.  In
1013 	 * this case, we can save a useless back to back clock update.
1014 	 */
1015 	if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1016 		rq_clock_skip_update(rq, true);
1017 }
1018 
1019 #ifdef CONFIG_SMP
1020 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1021 {
1022 #ifdef CONFIG_SCHED_DEBUG
1023 	/*
1024 	 * We should never call set_task_cpu() on a blocked task,
1025 	 * ttwu() will sort out the placement.
1026 	 */
1027 	WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1028 			!p->on_rq);
1029 
1030 #ifdef CONFIG_LOCKDEP
1031 	/*
1032 	 * The caller should hold either p->pi_lock or rq->lock, when changing
1033 	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1034 	 *
1035 	 * sched_move_task() holds both and thus holding either pins the cgroup,
1036 	 * see task_group().
1037 	 *
1038 	 * Furthermore, all task_rq users should acquire both locks, see
1039 	 * task_rq_lock().
1040 	 */
1041 	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1042 				      lockdep_is_held(&task_rq(p)->lock)));
1043 #endif
1044 #endif
1045 
1046 	trace_sched_migrate_task(p, new_cpu);
1047 
1048 	if (task_cpu(p) != new_cpu) {
1049 		if (p->sched_class->migrate_task_rq)
1050 			p->sched_class->migrate_task_rq(p, new_cpu);
1051 		p->se.nr_migrations++;
1052 		perf_event_task_migrate(p);
1053 	}
1054 
1055 	__set_task_cpu(p, new_cpu);
1056 }
1057 
1058 static void __migrate_swap_task(struct task_struct *p, int cpu)
1059 {
1060 	if (task_on_rq_queued(p)) {
1061 		struct rq *src_rq, *dst_rq;
1062 
1063 		src_rq = task_rq(p);
1064 		dst_rq = cpu_rq(cpu);
1065 
1066 		deactivate_task(src_rq, p, 0);
1067 		set_task_cpu(p, cpu);
1068 		activate_task(dst_rq, p, 0);
1069 		check_preempt_curr(dst_rq, p, 0);
1070 	} else {
1071 		/*
1072 		 * Task isn't running anymore; make it appear like we migrated
1073 		 * it before it went to sleep. This means on wakeup we make the
1074 		 * previous cpu our targer instead of where it really is.
1075 		 */
1076 		p->wake_cpu = cpu;
1077 	}
1078 }
1079 
1080 struct migration_swap_arg {
1081 	struct task_struct *src_task, *dst_task;
1082 	int src_cpu, dst_cpu;
1083 };
1084 
1085 static int migrate_swap_stop(void *data)
1086 {
1087 	struct migration_swap_arg *arg = data;
1088 	struct rq *src_rq, *dst_rq;
1089 	int ret = -EAGAIN;
1090 
1091 	src_rq = cpu_rq(arg->src_cpu);
1092 	dst_rq = cpu_rq(arg->dst_cpu);
1093 
1094 	double_raw_lock(&arg->src_task->pi_lock,
1095 			&arg->dst_task->pi_lock);
1096 	double_rq_lock(src_rq, dst_rq);
1097 	if (task_cpu(arg->dst_task) != arg->dst_cpu)
1098 		goto unlock;
1099 
1100 	if (task_cpu(arg->src_task) != arg->src_cpu)
1101 		goto unlock;
1102 
1103 	if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1104 		goto unlock;
1105 
1106 	if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1107 		goto unlock;
1108 
1109 	__migrate_swap_task(arg->src_task, arg->dst_cpu);
1110 	__migrate_swap_task(arg->dst_task, arg->src_cpu);
1111 
1112 	ret = 0;
1113 
1114 unlock:
1115 	double_rq_unlock(src_rq, dst_rq);
1116 	raw_spin_unlock(&arg->dst_task->pi_lock);
1117 	raw_spin_unlock(&arg->src_task->pi_lock);
1118 
1119 	return ret;
1120 }
1121 
1122 /*
1123  * Cross migrate two tasks
1124  */
1125 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1126 {
1127 	struct migration_swap_arg arg;
1128 	int ret = -EINVAL;
1129 
1130 	arg = (struct migration_swap_arg){
1131 		.src_task = cur,
1132 		.src_cpu = task_cpu(cur),
1133 		.dst_task = p,
1134 		.dst_cpu = task_cpu(p),
1135 	};
1136 
1137 	if (arg.src_cpu == arg.dst_cpu)
1138 		goto out;
1139 
1140 	/*
1141 	 * These three tests are all lockless; this is OK since all of them
1142 	 * will be re-checked with proper locks held further down the line.
1143 	 */
1144 	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1145 		goto out;
1146 
1147 	if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1148 		goto out;
1149 
1150 	if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1151 		goto out;
1152 
1153 	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1154 	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1155 
1156 out:
1157 	return ret;
1158 }
1159 
1160 struct migration_arg {
1161 	struct task_struct *task;
1162 	int dest_cpu;
1163 };
1164 
1165 static int migration_cpu_stop(void *data);
1166 
1167 /*
1168  * wait_task_inactive - wait for a thread to unschedule.
1169  *
1170  * If @match_state is nonzero, it's the @p->state value just checked and
1171  * not expected to change.  If it changes, i.e. @p might have woken up,
1172  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1173  * we return a positive number (its total switch count).  If a second call
1174  * a short while later returns the same number, the caller can be sure that
1175  * @p has remained unscheduled the whole time.
1176  *
1177  * The caller must ensure that the task *will* unschedule sometime soon,
1178  * else this function might spin for a *long* time. This function can't
1179  * be called with interrupts off, or it may introduce deadlock with
1180  * smp_call_function() if an IPI is sent by the same process we are
1181  * waiting to become inactive.
1182  */
1183 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1184 {
1185 	unsigned long flags;
1186 	int running, queued;
1187 	unsigned long ncsw;
1188 	struct rq *rq;
1189 
1190 	for (;;) {
1191 		/*
1192 		 * We do the initial early heuristics without holding
1193 		 * any task-queue locks at all. We'll only try to get
1194 		 * the runqueue lock when things look like they will
1195 		 * work out!
1196 		 */
1197 		rq = task_rq(p);
1198 
1199 		/*
1200 		 * If the task is actively running on another CPU
1201 		 * still, just relax and busy-wait without holding
1202 		 * any locks.
1203 		 *
1204 		 * NOTE! Since we don't hold any locks, it's not
1205 		 * even sure that "rq" stays as the right runqueue!
1206 		 * But we don't care, since "task_running()" will
1207 		 * return false if the runqueue has changed and p
1208 		 * is actually now running somewhere else!
1209 		 */
1210 		while (task_running(rq, p)) {
1211 			if (match_state && unlikely(p->state != match_state))
1212 				return 0;
1213 			cpu_relax();
1214 		}
1215 
1216 		/*
1217 		 * Ok, time to look more closely! We need the rq
1218 		 * lock now, to be *sure*. If we're wrong, we'll
1219 		 * just go back and repeat.
1220 		 */
1221 		rq = task_rq_lock(p, &flags);
1222 		trace_sched_wait_task(p);
1223 		running = task_running(rq, p);
1224 		queued = task_on_rq_queued(p);
1225 		ncsw = 0;
1226 		if (!match_state || p->state == match_state)
1227 			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1228 		task_rq_unlock(rq, p, &flags);
1229 
1230 		/*
1231 		 * If it changed from the expected state, bail out now.
1232 		 */
1233 		if (unlikely(!ncsw))
1234 			break;
1235 
1236 		/*
1237 		 * Was it really running after all now that we
1238 		 * checked with the proper locks actually held?
1239 		 *
1240 		 * Oops. Go back and try again..
1241 		 */
1242 		if (unlikely(running)) {
1243 			cpu_relax();
1244 			continue;
1245 		}
1246 
1247 		/*
1248 		 * It's not enough that it's not actively running,
1249 		 * it must be off the runqueue _entirely_, and not
1250 		 * preempted!
1251 		 *
1252 		 * So if it was still runnable (but just not actively
1253 		 * running right now), it's preempted, and we should
1254 		 * yield - it could be a while.
1255 		 */
1256 		if (unlikely(queued)) {
1257 			ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1258 
1259 			set_current_state(TASK_UNINTERRUPTIBLE);
1260 			schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1261 			continue;
1262 		}
1263 
1264 		/*
1265 		 * Ahh, all good. It wasn't running, and it wasn't
1266 		 * runnable, which means that it will never become
1267 		 * running in the future either. We're all done!
1268 		 */
1269 		break;
1270 	}
1271 
1272 	return ncsw;
1273 }
1274 
1275 /***
1276  * kick_process - kick a running thread to enter/exit the kernel
1277  * @p: the to-be-kicked thread
1278  *
1279  * Cause a process which is running on another CPU to enter
1280  * kernel-mode, without any delay. (to get signals handled.)
1281  *
1282  * NOTE: this function doesn't have to take the runqueue lock,
1283  * because all it wants to ensure is that the remote task enters
1284  * the kernel. If the IPI races and the task has been migrated
1285  * to another CPU then no harm is done and the purpose has been
1286  * achieved as well.
1287  */
1288 void kick_process(struct task_struct *p)
1289 {
1290 	int cpu;
1291 
1292 	preempt_disable();
1293 	cpu = task_cpu(p);
1294 	if ((cpu != smp_processor_id()) && task_curr(p))
1295 		smp_send_reschedule(cpu);
1296 	preempt_enable();
1297 }
1298 EXPORT_SYMBOL_GPL(kick_process);
1299 #endif /* CONFIG_SMP */
1300 
1301 #ifdef CONFIG_SMP
1302 /*
1303  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1304  */
1305 static int select_fallback_rq(int cpu, struct task_struct *p)
1306 {
1307 	int nid = cpu_to_node(cpu);
1308 	const struct cpumask *nodemask = NULL;
1309 	enum { cpuset, possible, fail } state = cpuset;
1310 	int dest_cpu;
1311 
1312 	/*
1313 	 * If the node that the cpu is on has been offlined, cpu_to_node()
1314 	 * will return -1. There is no cpu on the node, and we should
1315 	 * select the cpu on the other node.
1316 	 */
1317 	if (nid != -1) {
1318 		nodemask = cpumask_of_node(nid);
1319 
1320 		/* Look for allowed, online CPU in same node. */
1321 		for_each_cpu(dest_cpu, nodemask) {
1322 			if (!cpu_online(dest_cpu))
1323 				continue;
1324 			if (!cpu_active(dest_cpu))
1325 				continue;
1326 			if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1327 				return dest_cpu;
1328 		}
1329 	}
1330 
1331 	for (;;) {
1332 		/* Any allowed, online CPU? */
1333 		for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1334 			if (!cpu_online(dest_cpu))
1335 				continue;
1336 			if (!cpu_active(dest_cpu))
1337 				continue;
1338 			goto out;
1339 		}
1340 
1341 		switch (state) {
1342 		case cpuset:
1343 			/* No more Mr. Nice Guy. */
1344 			cpuset_cpus_allowed_fallback(p);
1345 			state = possible;
1346 			break;
1347 
1348 		case possible:
1349 			do_set_cpus_allowed(p, cpu_possible_mask);
1350 			state = fail;
1351 			break;
1352 
1353 		case fail:
1354 			BUG();
1355 			break;
1356 		}
1357 	}
1358 
1359 out:
1360 	if (state != cpuset) {
1361 		/*
1362 		 * Don't tell them about moving exiting tasks or
1363 		 * kernel threads (both mm NULL), since they never
1364 		 * leave kernel.
1365 		 */
1366 		if (p->mm && printk_ratelimit()) {
1367 			printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1368 					task_pid_nr(p), p->comm, cpu);
1369 		}
1370 	}
1371 
1372 	return dest_cpu;
1373 }
1374 
1375 /*
1376  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1377  */
1378 static inline
1379 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1380 {
1381 	if (p->nr_cpus_allowed > 1)
1382 		cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1383 
1384 	/*
1385 	 * In order not to call set_task_cpu() on a blocking task we need
1386 	 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1387 	 * cpu.
1388 	 *
1389 	 * Since this is common to all placement strategies, this lives here.
1390 	 *
1391 	 * [ this allows ->select_task() to simply return task_cpu(p) and
1392 	 *   not worry about this generic constraint ]
1393 	 */
1394 	if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1395 		     !cpu_online(cpu)))
1396 		cpu = select_fallback_rq(task_cpu(p), p);
1397 
1398 	return cpu;
1399 }
1400 
1401 static void update_avg(u64 *avg, u64 sample)
1402 {
1403 	s64 diff = sample - *avg;
1404 	*avg += diff >> 3;
1405 }
1406 #endif
1407 
1408 static void
1409 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1410 {
1411 #ifdef CONFIG_SCHEDSTATS
1412 	struct rq *rq = this_rq();
1413 
1414 #ifdef CONFIG_SMP
1415 	int this_cpu = smp_processor_id();
1416 
1417 	if (cpu == this_cpu) {
1418 		schedstat_inc(rq, ttwu_local);
1419 		schedstat_inc(p, se.statistics.nr_wakeups_local);
1420 	} else {
1421 		struct sched_domain *sd;
1422 
1423 		schedstat_inc(p, se.statistics.nr_wakeups_remote);
1424 		rcu_read_lock();
1425 		for_each_domain(this_cpu, sd) {
1426 			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1427 				schedstat_inc(sd, ttwu_wake_remote);
1428 				break;
1429 			}
1430 		}
1431 		rcu_read_unlock();
1432 	}
1433 
1434 	if (wake_flags & WF_MIGRATED)
1435 		schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1436 
1437 #endif /* CONFIG_SMP */
1438 
1439 	schedstat_inc(rq, ttwu_count);
1440 	schedstat_inc(p, se.statistics.nr_wakeups);
1441 
1442 	if (wake_flags & WF_SYNC)
1443 		schedstat_inc(p, se.statistics.nr_wakeups_sync);
1444 
1445 #endif /* CONFIG_SCHEDSTATS */
1446 }
1447 
1448 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1449 {
1450 	activate_task(rq, p, en_flags);
1451 	p->on_rq = TASK_ON_RQ_QUEUED;
1452 
1453 	/* if a worker is waking up, notify workqueue */
1454 	if (p->flags & PF_WQ_WORKER)
1455 		wq_worker_waking_up(p, cpu_of(rq));
1456 }
1457 
1458 /*
1459  * Mark the task runnable and perform wakeup-preemption.
1460  */
1461 static void
1462 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1463 {
1464 	check_preempt_curr(rq, p, wake_flags);
1465 	trace_sched_wakeup(p, true);
1466 
1467 	p->state = TASK_RUNNING;
1468 #ifdef CONFIG_SMP
1469 	if (p->sched_class->task_woken)
1470 		p->sched_class->task_woken(rq, p);
1471 
1472 	if (rq->idle_stamp) {
1473 		u64 delta = rq_clock(rq) - rq->idle_stamp;
1474 		u64 max = 2*rq->max_idle_balance_cost;
1475 
1476 		update_avg(&rq->avg_idle, delta);
1477 
1478 		if (rq->avg_idle > max)
1479 			rq->avg_idle = max;
1480 
1481 		rq->idle_stamp = 0;
1482 	}
1483 #endif
1484 }
1485 
1486 static void
1487 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1488 {
1489 #ifdef CONFIG_SMP
1490 	if (p->sched_contributes_to_load)
1491 		rq->nr_uninterruptible--;
1492 #endif
1493 
1494 	ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1495 	ttwu_do_wakeup(rq, p, wake_flags);
1496 }
1497 
1498 /*
1499  * Called in case the task @p isn't fully descheduled from its runqueue,
1500  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1501  * since all we need to do is flip p->state to TASK_RUNNING, since
1502  * the task is still ->on_rq.
1503  */
1504 static int ttwu_remote(struct task_struct *p, int wake_flags)
1505 {
1506 	struct rq *rq;
1507 	int ret = 0;
1508 
1509 	rq = __task_rq_lock(p);
1510 	if (task_on_rq_queued(p)) {
1511 		/* check_preempt_curr() may use rq clock */
1512 		update_rq_clock(rq);
1513 		ttwu_do_wakeup(rq, p, wake_flags);
1514 		ret = 1;
1515 	}
1516 	__task_rq_unlock(rq);
1517 
1518 	return ret;
1519 }
1520 
1521 #ifdef CONFIG_SMP
1522 void sched_ttwu_pending(void)
1523 {
1524 	struct rq *rq = this_rq();
1525 	struct llist_node *llist = llist_del_all(&rq->wake_list);
1526 	struct task_struct *p;
1527 	unsigned long flags;
1528 
1529 	if (!llist)
1530 		return;
1531 
1532 	raw_spin_lock_irqsave(&rq->lock, flags);
1533 
1534 	while (llist) {
1535 		p = llist_entry(llist, struct task_struct, wake_entry);
1536 		llist = llist_next(llist);
1537 		ttwu_do_activate(rq, p, 0);
1538 	}
1539 
1540 	raw_spin_unlock_irqrestore(&rq->lock, flags);
1541 }
1542 
1543 void scheduler_ipi(void)
1544 {
1545 	/*
1546 	 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1547 	 * TIF_NEED_RESCHED remotely (for the first time) will also send
1548 	 * this IPI.
1549 	 */
1550 	preempt_fold_need_resched();
1551 
1552 	if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1553 		return;
1554 
1555 	/*
1556 	 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1557 	 * traditionally all their work was done from the interrupt return
1558 	 * path. Now that we actually do some work, we need to make sure
1559 	 * we do call them.
1560 	 *
1561 	 * Some archs already do call them, luckily irq_enter/exit nest
1562 	 * properly.
1563 	 *
1564 	 * Arguably we should visit all archs and update all handlers,
1565 	 * however a fair share of IPIs are still resched only so this would
1566 	 * somewhat pessimize the simple resched case.
1567 	 */
1568 	irq_enter();
1569 	sched_ttwu_pending();
1570 
1571 	/*
1572 	 * Check if someone kicked us for doing the nohz idle load balance.
1573 	 */
1574 	if (unlikely(got_nohz_idle_kick())) {
1575 		this_rq()->idle_balance = 1;
1576 		raise_softirq_irqoff(SCHED_SOFTIRQ);
1577 	}
1578 	irq_exit();
1579 }
1580 
1581 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1582 {
1583 	struct rq *rq = cpu_rq(cpu);
1584 
1585 	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1586 		if (!set_nr_if_polling(rq->idle))
1587 			smp_send_reschedule(cpu);
1588 		else
1589 			trace_sched_wake_idle_without_ipi(cpu);
1590 	}
1591 }
1592 
1593 void wake_up_if_idle(int cpu)
1594 {
1595 	struct rq *rq = cpu_rq(cpu);
1596 	unsigned long flags;
1597 
1598 	rcu_read_lock();
1599 
1600 	if (!is_idle_task(rcu_dereference(rq->curr)))
1601 		goto out;
1602 
1603 	if (set_nr_if_polling(rq->idle)) {
1604 		trace_sched_wake_idle_without_ipi(cpu);
1605 	} else {
1606 		raw_spin_lock_irqsave(&rq->lock, flags);
1607 		if (is_idle_task(rq->curr))
1608 			smp_send_reschedule(cpu);
1609 		/* Else cpu is not in idle, do nothing here */
1610 		raw_spin_unlock_irqrestore(&rq->lock, flags);
1611 	}
1612 
1613 out:
1614 	rcu_read_unlock();
1615 }
1616 
1617 bool cpus_share_cache(int this_cpu, int that_cpu)
1618 {
1619 	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1620 }
1621 #endif /* CONFIG_SMP */
1622 
1623 static void ttwu_queue(struct task_struct *p, int cpu)
1624 {
1625 	struct rq *rq = cpu_rq(cpu);
1626 
1627 #if defined(CONFIG_SMP)
1628 	if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1629 		sched_clock_cpu(cpu); /* sync clocks x-cpu */
1630 		ttwu_queue_remote(p, cpu);
1631 		return;
1632 	}
1633 #endif
1634 
1635 	raw_spin_lock(&rq->lock);
1636 	ttwu_do_activate(rq, p, 0);
1637 	raw_spin_unlock(&rq->lock);
1638 }
1639 
1640 /**
1641  * try_to_wake_up - wake up a thread
1642  * @p: the thread to be awakened
1643  * @state: the mask of task states that can be woken
1644  * @wake_flags: wake modifier flags (WF_*)
1645  *
1646  * Put it on the run-queue if it's not already there. The "current"
1647  * thread is always on the run-queue (except when the actual
1648  * re-schedule is in progress), and as such you're allowed to do
1649  * the simpler "current->state = TASK_RUNNING" to mark yourself
1650  * runnable without the overhead of this.
1651  *
1652  * Return: %true if @p was woken up, %false if it was already running.
1653  * or @state didn't match @p's state.
1654  */
1655 static int
1656 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1657 {
1658 	unsigned long flags;
1659 	int cpu, success = 0;
1660 
1661 	/*
1662 	 * If we are going to wake up a thread waiting for CONDITION we
1663 	 * need to ensure that CONDITION=1 done by the caller can not be
1664 	 * reordered with p->state check below. This pairs with mb() in
1665 	 * set_current_state() the waiting thread does.
1666 	 */
1667 	smp_mb__before_spinlock();
1668 	raw_spin_lock_irqsave(&p->pi_lock, flags);
1669 	if (!(p->state & state))
1670 		goto out;
1671 
1672 	success = 1; /* we're going to change ->state */
1673 	cpu = task_cpu(p);
1674 
1675 	if (p->on_rq && ttwu_remote(p, wake_flags))
1676 		goto stat;
1677 
1678 #ifdef CONFIG_SMP
1679 	/*
1680 	 * If the owning (remote) cpu is still in the middle of schedule() with
1681 	 * this task as prev, wait until its done referencing the task.
1682 	 */
1683 	while (p->on_cpu)
1684 		cpu_relax();
1685 	/*
1686 	 * Pairs with the smp_wmb() in finish_lock_switch().
1687 	 */
1688 	smp_rmb();
1689 
1690 	p->sched_contributes_to_load = !!task_contributes_to_load(p);
1691 	p->state = TASK_WAKING;
1692 
1693 	if (p->sched_class->task_waking)
1694 		p->sched_class->task_waking(p);
1695 
1696 	cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1697 	if (task_cpu(p) != cpu) {
1698 		wake_flags |= WF_MIGRATED;
1699 		set_task_cpu(p, cpu);
1700 	}
1701 #endif /* CONFIG_SMP */
1702 
1703 	ttwu_queue(p, cpu);
1704 stat:
1705 	ttwu_stat(p, cpu, wake_flags);
1706 out:
1707 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1708 
1709 	return success;
1710 }
1711 
1712 /**
1713  * try_to_wake_up_local - try to wake up a local task with rq lock held
1714  * @p: the thread to be awakened
1715  *
1716  * Put @p on the run-queue if it's not already there. The caller must
1717  * ensure that this_rq() is locked, @p is bound to this_rq() and not
1718  * the current task.
1719  */
1720 static void try_to_wake_up_local(struct task_struct *p)
1721 {
1722 	struct rq *rq = task_rq(p);
1723 
1724 	if (WARN_ON_ONCE(rq != this_rq()) ||
1725 	    WARN_ON_ONCE(p == current))
1726 		return;
1727 
1728 	lockdep_assert_held(&rq->lock);
1729 
1730 	if (!raw_spin_trylock(&p->pi_lock)) {
1731 		raw_spin_unlock(&rq->lock);
1732 		raw_spin_lock(&p->pi_lock);
1733 		raw_spin_lock(&rq->lock);
1734 	}
1735 
1736 	if (!(p->state & TASK_NORMAL))
1737 		goto out;
1738 
1739 	if (!task_on_rq_queued(p))
1740 		ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1741 
1742 	ttwu_do_wakeup(rq, p, 0);
1743 	ttwu_stat(p, smp_processor_id(), 0);
1744 out:
1745 	raw_spin_unlock(&p->pi_lock);
1746 }
1747 
1748 /**
1749  * wake_up_process - Wake up a specific process
1750  * @p: The process to be woken up.
1751  *
1752  * Attempt to wake up the nominated process and move it to the set of runnable
1753  * processes.
1754  *
1755  * Return: 1 if the process was woken up, 0 if it was already running.
1756  *
1757  * It may be assumed that this function implies a write memory barrier before
1758  * changing the task state if and only if any tasks are woken up.
1759  */
1760 int wake_up_process(struct task_struct *p)
1761 {
1762 	WARN_ON(task_is_stopped_or_traced(p));
1763 	return try_to_wake_up(p, TASK_NORMAL, 0);
1764 }
1765 EXPORT_SYMBOL(wake_up_process);
1766 
1767 int wake_up_state(struct task_struct *p, unsigned int state)
1768 {
1769 	return try_to_wake_up(p, state, 0);
1770 }
1771 
1772 /*
1773  * This function clears the sched_dl_entity static params.
1774  */
1775 void __dl_clear_params(struct task_struct *p)
1776 {
1777 	struct sched_dl_entity *dl_se = &p->dl;
1778 
1779 	dl_se->dl_runtime = 0;
1780 	dl_se->dl_deadline = 0;
1781 	dl_se->dl_period = 0;
1782 	dl_se->flags = 0;
1783 	dl_se->dl_bw = 0;
1784 
1785 	dl_se->dl_throttled = 0;
1786 	dl_se->dl_new = 1;
1787 	dl_se->dl_yielded = 0;
1788 }
1789 
1790 /*
1791  * Perform scheduler related setup for a newly forked process p.
1792  * p is forked by current.
1793  *
1794  * __sched_fork() is basic setup used by init_idle() too:
1795  */
1796 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1797 {
1798 	p->on_rq			= 0;
1799 
1800 	p->se.on_rq			= 0;
1801 	p->se.exec_start		= 0;
1802 	p->se.sum_exec_runtime		= 0;
1803 	p->se.prev_sum_exec_runtime	= 0;
1804 	p->se.nr_migrations		= 0;
1805 	p->se.vruntime			= 0;
1806 #ifdef CONFIG_SMP
1807 	p->se.avg.decay_count		= 0;
1808 #endif
1809 	INIT_LIST_HEAD(&p->se.group_node);
1810 
1811 #ifdef CONFIG_SCHEDSTATS
1812 	memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1813 #endif
1814 
1815 	RB_CLEAR_NODE(&p->dl.rb_node);
1816 	init_dl_task_timer(&p->dl);
1817 	__dl_clear_params(p);
1818 
1819 	INIT_LIST_HEAD(&p->rt.run_list);
1820 
1821 #ifdef CONFIG_PREEMPT_NOTIFIERS
1822 	INIT_HLIST_HEAD(&p->preempt_notifiers);
1823 #endif
1824 
1825 #ifdef CONFIG_NUMA_BALANCING
1826 	if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1827 		p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1828 		p->mm->numa_scan_seq = 0;
1829 	}
1830 
1831 	if (clone_flags & CLONE_VM)
1832 		p->numa_preferred_nid = current->numa_preferred_nid;
1833 	else
1834 		p->numa_preferred_nid = -1;
1835 
1836 	p->node_stamp = 0ULL;
1837 	p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1838 	p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1839 	p->numa_work.next = &p->numa_work;
1840 	p->numa_faults = NULL;
1841 	p->last_task_numa_placement = 0;
1842 	p->last_sum_exec_runtime = 0;
1843 
1844 	p->numa_group = NULL;
1845 #endif /* CONFIG_NUMA_BALANCING */
1846 }
1847 
1848 #ifdef CONFIG_NUMA_BALANCING
1849 #ifdef CONFIG_SCHED_DEBUG
1850 void set_numabalancing_state(bool enabled)
1851 {
1852 	if (enabled)
1853 		sched_feat_set("NUMA");
1854 	else
1855 		sched_feat_set("NO_NUMA");
1856 }
1857 #else
1858 __read_mostly bool numabalancing_enabled;
1859 
1860 void set_numabalancing_state(bool enabled)
1861 {
1862 	numabalancing_enabled = enabled;
1863 }
1864 #endif /* CONFIG_SCHED_DEBUG */
1865 
1866 #ifdef CONFIG_PROC_SYSCTL
1867 int sysctl_numa_balancing(struct ctl_table *table, int write,
1868 			 void __user *buffer, size_t *lenp, loff_t *ppos)
1869 {
1870 	struct ctl_table t;
1871 	int err;
1872 	int state = numabalancing_enabled;
1873 
1874 	if (write && !capable(CAP_SYS_ADMIN))
1875 		return -EPERM;
1876 
1877 	t = *table;
1878 	t.data = &state;
1879 	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1880 	if (err < 0)
1881 		return err;
1882 	if (write)
1883 		set_numabalancing_state(state);
1884 	return err;
1885 }
1886 #endif
1887 #endif
1888 
1889 /*
1890  * fork()/clone()-time setup:
1891  */
1892 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1893 {
1894 	unsigned long flags;
1895 	int cpu = get_cpu();
1896 
1897 	__sched_fork(clone_flags, p);
1898 	/*
1899 	 * We mark the process as running here. This guarantees that
1900 	 * nobody will actually run it, and a signal or other external
1901 	 * event cannot wake it up and insert it on the runqueue either.
1902 	 */
1903 	p->state = TASK_RUNNING;
1904 
1905 	/*
1906 	 * Make sure we do not leak PI boosting priority to the child.
1907 	 */
1908 	p->prio = current->normal_prio;
1909 
1910 	/*
1911 	 * Revert to default priority/policy on fork if requested.
1912 	 */
1913 	if (unlikely(p->sched_reset_on_fork)) {
1914 		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1915 			p->policy = SCHED_NORMAL;
1916 			p->static_prio = NICE_TO_PRIO(0);
1917 			p->rt_priority = 0;
1918 		} else if (PRIO_TO_NICE(p->static_prio) < 0)
1919 			p->static_prio = NICE_TO_PRIO(0);
1920 
1921 		p->prio = p->normal_prio = __normal_prio(p);
1922 		set_load_weight(p);
1923 
1924 		/*
1925 		 * We don't need the reset flag anymore after the fork. It has
1926 		 * fulfilled its duty:
1927 		 */
1928 		p->sched_reset_on_fork = 0;
1929 	}
1930 
1931 	if (dl_prio(p->prio)) {
1932 		put_cpu();
1933 		return -EAGAIN;
1934 	} else if (rt_prio(p->prio)) {
1935 		p->sched_class = &rt_sched_class;
1936 	} else {
1937 		p->sched_class = &fair_sched_class;
1938 	}
1939 
1940 	if (p->sched_class->task_fork)
1941 		p->sched_class->task_fork(p);
1942 
1943 	/*
1944 	 * The child is not yet in the pid-hash so no cgroup attach races,
1945 	 * and the cgroup is pinned to this child due to cgroup_fork()
1946 	 * is ran before sched_fork().
1947 	 *
1948 	 * Silence PROVE_RCU.
1949 	 */
1950 	raw_spin_lock_irqsave(&p->pi_lock, flags);
1951 	set_task_cpu(p, cpu);
1952 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1953 
1954 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1955 	if (likely(sched_info_on()))
1956 		memset(&p->sched_info, 0, sizeof(p->sched_info));
1957 #endif
1958 #if defined(CONFIG_SMP)
1959 	p->on_cpu = 0;
1960 #endif
1961 	init_task_preempt_count(p);
1962 #ifdef CONFIG_SMP
1963 	plist_node_init(&p->pushable_tasks, MAX_PRIO);
1964 	RB_CLEAR_NODE(&p->pushable_dl_tasks);
1965 #endif
1966 
1967 	put_cpu();
1968 	return 0;
1969 }
1970 
1971 unsigned long to_ratio(u64 period, u64 runtime)
1972 {
1973 	if (runtime == RUNTIME_INF)
1974 		return 1ULL << 20;
1975 
1976 	/*
1977 	 * Doing this here saves a lot of checks in all
1978 	 * the calling paths, and returning zero seems
1979 	 * safe for them anyway.
1980 	 */
1981 	if (period == 0)
1982 		return 0;
1983 
1984 	return div64_u64(runtime << 20, period);
1985 }
1986 
1987 #ifdef CONFIG_SMP
1988 inline struct dl_bw *dl_bw_of(int i)
1989 {
1990 	rcu_lockdep_assert(rcu_read_lock_sched_held(),
1991 			   "sched RCU must be held");
1992 	return &cpu_rq(i)->rd->dl_bw;
1993 }
1994 
1995 static inline int dl_bw_cpus(int i)
1996 {
1997 	struct root_domain *rd = cpu_rq(i)->rd;
1998 	int cpus = 0;
1999 
2000 	rcu_lockdep_assert(rcu_read_lock_sched_held(),
2001 			   "sched RCU must be held");
2002 	for_each_cpu_and(i, rd->span, cpu_active_mask)
2003 		cpus++;
2004 
2005 	return cpus;
2006 }
2007 #else
2008 inline struct dl_bw *dl_bw_of(int i)
2009 {
2010 	return &cpu_rq(i)->dl.dl_bw;
2011 }
2012 
2013 static inline int dl_bw_cpus(int i)
2014 {
2015 	return 1;
2016 }
2017 #endif
2018 
2019 /*
2020  * We must be sure that accepting a new task (or allowing changing the
2021  * parameters of an existing one) is consistent with the bandwidth
2022  * constraints. If yes, this function also accordingly updates the currently
2023  * allocated bandwidth to reflect the new situation.
2024  *
2025  * This function is called while holding p's rq->lock.
2026  *
2027  * XXX we should delay bw change until the task's 0-lag point, see
2028  * __setparam_dl().
2029  */
2030 static int dl_overflow(struct task_struct *p, int policy,
2031 		       const struct sched_attr *attr)
2032 {
2033 
2034 	struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2035 	u64 period = attr->sched_period ?: attr->sched_deadline;
2036 	u64 runtime = attr->sched_runtime;
2037 	u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2038 	int cpus, err = -1;
2039 
2040 	if (new_bw == p->dl.dl_bw)
2041 		return 0;
2042 
2043 	/*
2044 	 * Either if a task, enters, leave, or stays -deadline but changes
2045 	 * its parameters, we may need to update accordingly the total
2046 	 * allocated bandwidth of the container.
2047 	 */
2048 	raw_spin_lock(&dl_b->lock);
2049 	cpus = dl_bw_cpus(task_cpu(p));
2050 	if (dl_policy(policy) && !task_has_dl_policy(p) &&
2051 	    !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2052 		__dl_add(dl_b, new_bw);
2053 		err = 0;
2054 	} else if (dl_policy(policy) && task_has_dl_policy(p) &&
2055 		   !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2056 		__dl_clear(dl_b, p->dl.dl_bw);
2057 		__dl_add(dl_b, new_bw);
2058 		err = 0;
2059 	} else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2060 		__dl_clear(dl_b, p->dl.dl_bw);
2061 		err = 0;
2062 	}
2063 	raw_spin_unlock(&dl_b->lock);
2064 
2065 	return err;
2066 }
2067 
2068 extern void init_dl_bw(struct dl_bw *dl_b);
2069 
2070 /*
2071  * wake_up_new_task - wake up a newly created task for the first time.
2072  *
2073  * This function will do some initial scheduler statistics housekeeping
2074  * that must be done for every newly created context, then puts the task
2075  * on the runqueue and wakes it.
2076  */
2077 void wake_up_new_task(struct task_struct *p)
2078 {
2079 	unsigned long flags;
2080 	struct rq *rq;
2081 
2082 	raw_spin_lock_irqsave(&p->pi_lock, flags);
2083 #ifdef CONFIG_SMP
2084 	/*
2085 	 * Fork balancing, do it here and not earlier because:
2086 	 *  - cpus_allowed can change in the fork path
2087 	 *  - any previously selected cpu might disappear through hotplug
2088 	 */
2089 	set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2090 #endif
2091 
2092 	/* Initialize new task's runnable average */
2093 	init_task_runnable_average(p);
2094 	rq = __task_rq_lock(p);
2095 	activate_task(rq, p, 0);
2096 	p->on_rq = TASK_ON_RQ_QUEUED;
2097 	trace_sched_wakeup_new(p, true);
2098 	check_preempt_curr(rq, p, WF_FORK);
2099 #ifdef CONFIG_SMP
2100 	if (p->sched_class->task_woken)
2101 		p->sched_class->task_woken(rq, p);
2102 #endif
2103 	task_rq_unlock(rq, p, &flags);
2104 }
2105 
2106 #ifdef CONFIG_PREEMPT_NOTIFIERS
2107 
2108 /**
2109  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2110  * @notifier: notifier struct to register
2111  */
2112 void preempt_notifier_register(struct preempt_notifier *notifier)
2113 {
2114 	hlist_add_head(&notifier->link, &current->preempt_notifiers);
2115 }
2116 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2117 
2118 /**
2119  * preempt_notifier_unregister - no longer interested in preemption notifications
2120  * @notifier: notifier struct to unregister
2121  *
2122  * This is safe to call from within a preemption notifier.
2123  */
2124 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2125 {
2126 	hlist_del(&notifier->link);
2127 }
2128 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2129 
2130 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2131 {
2132 	struct preempt_notifier *notifier;
2133 
2134 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2135 		notifier->ops->sched_in(notifier, raw_smp_processor_id());
2136 }
2137 
2138 static void
2139 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2140 				 struct task_struct *next)
2141 {
2142 	struct preempt_notifier *notifier;
2143 
2144 	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2145 		notifier->ops->sched_out(notifier, next);
2146 }
2147 
2148 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2149 
2150 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2151 {
2152 }
2153 
2154 static void
2155 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2156 				 struct task_struct *next)
2157 {
2158 }
2159 
2160 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2161 
2162 /**
2163  * prepare_task_switch - prepare to switch tasks
2164  * @rq: the runqueue preparing to switch
2165  * @prev: the current task that is being switched out
2166  * @next: the task we are going to switch to.
2167  *
2168  * This is called with the rq lock held and interrupts off. It must
2169  * be paired with a subsequent finish_task_switch after the context
2170  * switch.
2171  *
2172  * prepare_task_switch sets up locking and calls architecture specific
2173  * hooks.
2174  */
2175 static inline void
2176 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2177 		    struct task_struct *next)
2178 {
2179 	trace_sched_switch(prev, next);
2180 	sched_info_switch(rq, prev, next);
2181 	perf_event_task_sched_out(prev, next);
2182 	fire_sched_out_preempt_notifiers(prev, next);
2183 	prepare_lock_switch(rq, next);
2184 	prepare_arch_switch(next);
2185 }
2186 
2187 /**
2188  * finish_task_switch - clean up after a task-switch
2189  * @prev: the thread we just switched away from.
2190  *
2191  * finish_task_switch must be called after the context switch, paired
2192  * with a prepare_task_switch call before the context switch.
2193  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2194  * and do any other architecture-specific cleanup actions.
2195  *
2196  * Note that we may have delayed dropping an mm in context_switch(). If
2197  * so, we finish that here outside of the runqueue lock. (Doing it
2198  * with the lock held can cause deadlocks; see schedule() for
2199  * details.)
2200  *
2201  * The context switch have flipped the stack from under us and restored the
2202  * local variables which were saved when this task called schedule() in the
2203  * past. prev == current is still correct but we need to recalculate this_rq
2204  * because prev may have moved to another CPU.
2205  */
2206 static struct rq *finish_task_switch(struct task_struct *prev)
2207 	__releases(rq->lock)
2208 {
2209 	struct rq *rq = this_rq();
2210 	struct mm_struct *mm = rq->prev_mm;
2211 	long prev_state;
2212 
2213 	rq->prev_mm = NULL;
2214 
2215 	/*
2216 	 * A task struct has one reference for the use as "current".
2217 	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2218 	 * schedule one last time. The schedule call will never return, and
2219 	 * the scheduled task must drop that reference.
2220 	 * The test for TASK_DEAD must occur while the runqueue locks are
2221 	 * still held, otherwise prev could be scheduled on another cpu, die
2222 	 * there before we look at prev->state, and then the reference would
2223 	 * be dropped twice.
2224 	 *		Manfred Spraul <manfred@colorfullife.com>
2225 	 */
2226 	prev_state = prev->state;
2227 	vtime_task_switch(prev);
2228 	finish_arch_switch(prev);
2229 	perf_event_task_sched_in(prev, current);
2230 	finish_lock_switch(rq, prev);
2231 	finish_arch_post_lock_switch();
2232 
2233 	fire_sched_in_preempt_notifiers(current);
2234 	if (mm)
2235 		mmdrop(mm);
2236 	if (unlikely(prev_state == TASK_DEAD)) {
2237 		if (prev->sched_class->task_dead)
2238 			prev->sched_class->task_dead(prev);
2239 
2240 		/*
2241 		 * Remove function-return probe instances associated with this
2242 		 * task and put them back on the free list.
2243 		 */
2244 		kprobe_flush_task(prev);
2245 		put_task_struct(prev);
2246 	}
2247 
2248 	tick_nohz_task_switch(current);
2249 	return rq;
2250 }
2251 
2252 #ifdef CONFIG_SMP
2253 
2254 /* rq->lock is NOT held, but preemption is disabled */
2255 static inline void post_schedule(struct rq *rq)
2256 {
2257 	if (rq->post_schedule) {
2258 		unsigned long flags;
2259 
2260 		raw_spin_lock_irqsave(&rq->lock, flags);
2261 		if (rq->curr->sched_class->post_schedule)
2262 			rq->curr->sched_class->post_schedule(rq);
2263 		raw_spin_unlock_irqrestore(&rq->lock, flags);
2264 
2265 		rq->post_schedule = 0;
2266 	}
2267 }
2268 
2269 #else
2270 
2271 static inline void post_schedule(struct rq *rq)
2272 {
2273 }
2274 
2275 #endif
2276 
2277 /**
2278  * schedule_tail - first thing a freshly forked thread must call.
2279  * @prev: the thread we just switched away from.
2280  */
2281 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2282 	__releases(rq->lock)
2283 {
2284 	struct rq *rq;
2285 
2286 	/* finish_task_switch() drops rq->lock and enables preemtion */
2287 	preempt_disable();
2288 	rq = finish_task_switch(prev);
2289 	post_schedule(rq);
2290 	preempt_enable();
2291 
2292 	if (current->set_child_tid)
2293 		put_user(task_pid_vnr(current), current->set_child_tid);
2294 }
2295 
2296 /*
2297  * context_switch - switch to the new MM and the new thread's register state.
2298  */
2299 static inline struct rq *
2300 context_switch(struct rq *rq, struct task_struct *prev,
2301 	       struct task_struct *next)
2302 {
2303 	struct mm_struct *mm, *oldmm;
2304 
2305 	prepare_task_switch(rq, prev, next);
2306 
2307 	mm = next->mm;
2308 	oldmm = prev->active_mm;
2309 	/*
2310 	 * For paravirt, this is coupled with an exit in switch_to to
2311 	 * combine the page table reload and the switch backend into
2312 	 * one hypercall.
2313 	 */
2314 	arch_start_context_switch(prev);
2315 
2316 	if (!mm) {
2317 		next->active_mm = oldmm;
2318 		atomic_inc(&oldmm->mm_count);
2319 		enter_lazy_tlb(oldmm, next);
2320 	} else
2321 		switch_mm(oldmm, mm, next);
2322 
2323 	if (!prev->mm) {
2324 		prev->active_mm = NULL;
2325 		rq->prev_mm = oldmm;
2326 	}
2327 	/*
2328 	 * Since the runqueue lock will be released by the next
2329 	 * task (which is an invalid locking op but in the case
2330 	 * of the scheduler it's an obvious special-case), so we
2331 	 * do an early lockdep release here:
2332 	 */
2333 	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2334 
2335 	context_tracking_task_switch(prev, next);
2336 	/* Here we just switch the register state and the stack. */
2337 	switch_to(prev, next, prev);
2338 	barrier();
2339 
2340 	return finish_task_switch(prev);
2341 }
2342 
2343 /*
2344  * nr_running and nr_context_switches:
2345  *
2346  * externally visible scheduler statistics: current number of runnable
2347  * threads, total number of context switches performed since bootup.
2348  */
2349 unsigned long nr_running(void)
2350 {
2351 	unsigned long i, sum = 0;
2352 
2353 	for_each_online_cpu(i)
2354 		sum += cpu_rq(i)->nr_running;
2355 
2356 	return sum;
2357 }
2358 
2359 /*
2360  * Check if only the current task is running on the cpu.
2361  */
2362 bool single_task_running(void)
2363 {
2364 	if (cpu_rq(smp_processor_id())->nr_running == 1)
2365 		return true;
2366 	else
2367 		return false;
2368 }
2369 EXPORT_SYMBOL(single_task_running);
2370 
2371 unsigned long long nr_context_switches(void)
2372 {
2373 	int i;
2374 	unsigned long long sum = 0;
2375 
2376 	for_each_possible_cpu(i)
2377 		sum += cpu_rq(i)->nr_switches;
2378 
2379 	return sum;
2380 }
2381 
2382 unsigned long nr_iowait(void)
2383 {
2384 	unsigned long i, sum = 0;
2385 
2386 	for_each_possible_cpu(i)
2387 		sum += atomic_read(&cpu_rq(i)->nr_iowait);
2388 
2389 	return sum;
2390 }
2391 
2392 unsigned long nr_iowait_cpu(int cpu)
2393 {
2394 	struct rq *this = cpu_rq(cpu);
2395 	return atomic_read(&this->nr_iowait);
2396 }
2397 
2398 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2399 {
2400 	struct rq *this = this_rq();
2401 	*nr_waiters = atomic_read(&this->nr_iowait);
2402 	*load = this->cpu_load[0];
2403 }
2404 
2405 #ifdef CONFIG_SMP
2406 
2407 /*
2408  * sched_exec - execve() is a valuable balancing opportunity, because at
2409  * this point the task has the smallest effective memory and cache footprint.
2410  */
2411 void sched_exec(void)
2412 {
2413 	struct task_struct *p = current;
2414 	unsigned long flags;
2415 	int dest_cpu;
2416 
2417 	raw_spin_lock_irqsave(&p->pi_lock, flags);
2418 	dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2419 	if (dest_cpu == smp_processor_id())
2420 		goto unlock;
2421 
2422 	if (likely(cpu_active(dest_cpu))) {
2423 		struct migration_arg arg = { p, dest_cpu };
2424 
2425 		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2426 		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2427 		return;
2428 	}
2429 unlock:
2430 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2431 }
2432 
2433 #endif
2434 
2435 DEFINE_PER_CPU(struct kernel_stat, kstat);
2436 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2437 
2438 EXPORT_PER_CPU_SYMBOL(kstat);
2439 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2440 
2441 /*
2442  * Return accounted runtime for the task.
2443  * In case the task is currently running, return the runtime plus current's
2444  * pending runtime that have not been accounted yet.
2445  */
2446 unsigned long long task_sched_runtime(struct task_struct *p)
2447 {
2448 	unsigned long flags;
2449 	struct rq *rq;
2450 	u64 ns;
2451 
2452 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2453 	/*
2454 	 * 64-bit doesn't need locks to atomically read a 64bit value.
2455 	 * So we have a optimization chance when the task's delta_exec is 0.
2456 	 * Reading ->on_cpu is racy, but this is ok.
2457 	 *
2458 	 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2459 	 * If we race with it entering cpu, unaccounted time is 0. This is
2460 	 * indistinguishable from the read occurring a few cycles earlier.
2461 	 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2462 	 * been accounted, so we're correct here as well.
2463 	 */
2464 	if (!p->on_cpu || !task_on_rq_queued(p))
2465 		return p->se.sum_exec_runtime;
2466 #endif
2467 
2468 	rq = task_rq_lock(p, &flags);
2469 	/*
2470 	 * Must be ->curr _and_ ->on_rq.  If dequeued, we would
2471 	 * project cycles that may never be accounted to this
2472 	 * thread, breaking clock_gettime().
2473 	 */
2474 	if (task_current(rq, p) && task_on_rq_queued(p)) {
2475 		update_rq_clock(rq);
2476 		p->sched_class->update_curr(rq);
2477 	}
2478 	ns = p->se.sum_exec_runtime;
2479 	task_rq_unlock(rq, p, &flags);
2480 
2481 	return ns;
2482 }
2483 
2484 /*
2485  * This function gets called by the timer code, with HZ frequency.
2486  * We call it with interrupts disabled.
2487  */
2488 void scheduler_tick(void)
2489 {
2490 	int cpu = smp_processor_id();
2491 	struct rq *rq = cpu_rq(cpu);
2492 	struct task_struct *curr = rq->curr;
2493 
2494 	sched_clock_tick();
2495 
2496 	raw_spin_lock(&rq->lock);
2497 	update_rq_clock(rq);
2498 	curr->sched_class->task_tick(rq, curr, 0);
2499 	update_cpu_load_active(rq);
2500 	raw_spin_unlock(&rq->lock);
2501 
2502 	perf_event_task_tick();
2503 
2504 #ifdef CONFIG_SMP
2505 	rq->idle_balance = idle_cpu(cpu);
2506 	trigger_load_balance(rq);
2507 #endif
2508 	rq_last_tick_reset(rq);
2509 }
2510 
2511 #ifdef CONFIG_NO_HZ_FULL
2512 /**
2513  * scheduler_tick_max_deferment
2514  *
2515  * Keep at least one tick per second when a single
2516  * active task is running because the scheduler doesn't
2517  * yet completely support full dynticks environment.
2518  *
2519  * This makes sure that uptime, CFS vruntime, load
2520  * balancing, etc... continue to move forward, even
2521  * with a very low granularity.
2522  *
2523  * Return: Maximum deferment in nanoseconds.
2524  */
2525 u64 scheduler_tick_max_deferment(void)
2526 {
2527 	struct rq *rq = this_rq();
2528 	unsigned long next, now = ACCESS_ONCE(jiffies);
2529 
2530 	next = rq->last_sched_tick + HZ;
2531 
2532 	if (time_before_eq(next, now))
2533 		return 0;
2534 
2535 	return jiffies_to_nsecs(next - now);
2536 }
2537 #endif
2538 
2539 notrace unsigned long get_parent_ip(unsigned long addr)
2540 {
2541 	if (in_lock_functions(addr)) {
2542 		addr = CALLER_ADDR2;
2543 		if (in_lock_functions(addr))
2544 			addr = CALLER_ADDR3;
2545 	}
2546 	return addr;
2547 }
2548 
2549 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2550 				defined(CONFIG_PREEMPT_TRACER))
2551 
2552 void preempt_count_add(int val)
2553 {
2554 #ifdef CONFIG_DEBUG_PREEMPT
2555 	/*
2556 	 * Underflow?
2557 	 */
2558 	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2559 		return;
2560 #endif
2561 	__preempt_count_add(val);
2562 #ifdef CONFIG_DEBUG_PREEMPT
2563 	/*
2564 	 * Spinlock count overflowing soon?
2565 	 */
2566 	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2567 				PREEMPT_MASK - 10);
2568 #endif
2569 	if (preempt_count() == val) {
2570 		unsigned long ip = get_parent_ip(CALLER_ADDR1);
2571 #ifdef CONFIG_DEBUG_PREEMPT
2572 		current->preempt_disable_ip = ip;
2573 #endif
2574 		trace_preempt_off(CALLER_ADDR0, ip);
2575 	}
2576 }
2577 EXPORT_SYMBOL(preempt_count_add);
2578 NOKPROBE_SYMBOL(preempt_count_add);
2579 
2580 void preempt_count_sub(int val)
2581 {
2582 #ifdef CONFIG_DEBUG_PREEMPT
2583 	/*
2584 	 * Underflow?
2585 	 */
2586 	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2587 		return;
2588 	/*
2589 	 * Is the spinlock portion underflowing?
2590 	 */
2591 	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2592 			!(preempt_count() & PREEMPT_MASK)))
2593 		return;
2594 #endif
2595 
2596 	if (preempt_count() == val)
2597 		trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2598 	__preempt_count_sub(val);
2599 }
2600 EXPORT_SYMBOL(preempt_count_sub);
2601 NOKPROBE_SYMBOL(preempt_count_sub);
2602 
2603 #endif
2604 
2605 /*
2606  * Print scheduling while atomic bug:
2607  */
2608 static noinline void __schedule_bug(struct task_struct *prev)
2609 {
2610 	if (oops_in_progress)
2611 		return;
2612 
2613 	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2614 		prev->comm, prev->pid, preempt_count());
2615 
2616 	debug_show_held_locks(prev);
2617 	print_modules();
2618 	if (irqs_disabled())
2619 		print_irqtrace_events(prev);
2620 #ifdef CONFIG_DEBUG_PREEMPT
2621 	if (in_atomic_preempt_off()) {
2622 		pr_err("Preemption disabled at:");
2623 		print_ip_sym(current->preempt_disable_ip);
2624 		pr_cont("\n");
2625 	}
2626 #endif
2627 	dump_stack();
2628 	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2629 }
2630 
2631 /*
2632  * Various schedule()-time debugging checks and statistics:
2633  */
2634 static inline void schedule_debug(struct task_struct *prev)
2635 {
2636 #ifdef CONFIG_SCHED_STACK_END_CHECK
2637 	BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2638 #endif
2639 	/*
2640 	 * Test if we are atomic. Since do_exit() needs to call into
2641 	 * schedule() atomically, we ignore that path. Otherwise whine
2642 	 * if we are scheduling when we should not.
2643 	 */
2644 	if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2645 		__schedule_bug(prev);
2646 	rcu_sleep_check();
2647 
2648 	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2649 
2650 	schedstat_inc(this_rq(), sched_count);
2651 }
2652 
2653 /*
2654  * Pick up the highest-prio task:
2655  */
2656 static inline struct task_struct *
2657 pick_next_task(struct rq *rq, struct task_struct *prev)
2658 {
2659 	const struct sched_class *class = &fair_sched_class;
2660 	struct task_struct *p;
2661 
2662 	/*
2663 	 * Optimization: we know that if all tasks are in
2664 	 * the fair class we can call that function directly:
2665 	 */
2666 	if (likely(prev->sched_class == class &&
2667 		   rq->nr_running == rq->cfs.h_nr_running)) {
2668 		p = fair_sched_class.pick_next_task(rq, prev);
2669 		if (unlikely(p == RETRY_TASK))
2670 			goto again;
2671 
2672 		/* assumes fair_sched_class->next == idle_sched_class */
2673 		if (unlikely(!p))
2674 			p = idle_sched_class.pick_next_task(rq, prev);
2675 
2676 		return p;
2677 	}
2678 
2679 again:
2680 	for_each_class(class) {
2681 		p = class->pick_next_task(rq, prev);
2682 		if (p) {
2683 			if (unlikely(p == RETRY_TASK))
2684 				goto again;
2685 			return p;
2686 		}
2687 	}
2688 
2689 	BUG(); /* the idle class will always have a runnable task */
2690 }
2691 
2692 /*
2693  * __schedule() is the main scheduler function.
2694  *
2695  * The main means of driving the scheduler and thus entering this function are:
2696  *
2697  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2698  *
2699  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2700  *      paths. For example, see arch/x86/entry_64.S.
2701  *
2702  *      To drive preemption between tasks, the scheduler sets the flag in timer
2703  *      interrupt handler scheduler_tick().
2704  *
2705  *   3. Wakeups don't really cause entry into schedule(). They add a
2706  *      task to the run-queue and that's it.
2707  *
2708  *      Now, if the new task added to the run-queue preempts the current
2709  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2710  *      called on the nearest possible occasion:
2711  *
2712  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2713  *
2714  *         - in syscall or exception context, at the next outmost
2715  *           preempt_enable(). (this might be as soon as the wake_up()'s
2716  *           spin_unlock()!)
2717  *
2718  *         - in IRQ context, return from interrupt-handler to
2719  *           preemptible context
2720  *
2721  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2722  *         then at the next:
2723  *
2724  *          - cond_resched() call
2725  *          - explicit schedule() call
2726  *          - return from syscall or exception to user-space
2727  *          - return from interrupt-handler to user-space
2728  *
2729  * WARNING: all callers must re-check need_resched() afterward and reschedule
2730  * accordingly in case an event triggered the need for rescheduling (such as
2731  * an interrupt waking up a task) while preemption was disabled in __schedule().
2732  */
2733 static void __sched __schedule(void)
2734 {
2735 	struct task_struct *prev, *next;
2736 	unsigned long *switch_count;
2737 	struct rq *rq;
2738 	int cpu;
2739 
2740 	preempt_disable();
2741 	cpu = smp_processor_id();
2742 	rq = cpu_rq(cpu);
2743 	rcu_note_context_switch();
2744 	prev = rq->curr;
2745 
2746 	schedule_debug(prev);
2747 
2748 	if (sched_feat(HRTICK))
2749 		hrtick_clear(rq);
2750 
2751 	/*
2752 	 * Make sure that signal_pending_state()->signal_pending() below
2753 	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2754 	 * done by the caller to avoid the race with signal_wake_up().
2755 	 */
2756 	smp_mb__before_spinlock();
2757 	raw_spin_lock_irq(&rq->lock);
2758 
2759 	rq->clock_skip_update <<= 1; /* promote REQ to ACT */
2760 
2761 	switch_count = &prev->nivcsw;
2762 	if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2763 		if (unlikely(signal_pending_state(prev->state, prev))) {
2764 			prev->state = TASK_RUNNING;
2765 		} else {
2766 			deactivate_task(rq, prev, DEQUEUE_SLEEP);
2767 			prev->on_rq = 0;
2768 
2769 			/*
2770 			 * If a worker went to sleep, notify and ask workqueue
2771 			 * whether it wants to wake up a task to maintain
2772 			 * concurrency.
2773 			 */
2774 			if (prev->flags & PF_WQ_WORKER) {
2775 				struct task_struct *to_wakeup;
2776 
2777 				to_wakeup = wq_worker_sleeping(prev, cpu);
2778 				if (to_wakeup)
2779 					try_to_wake_up_local(to_wakeup);
2780 			}
2781 		}
2782 		switch_count = &prev->nvcsw;
2783 	}
2784 
2785 	if (task_on_rq_queued(prev))
2786 		update_rq_clock(rq);
2787 
2788 	next = pick_next_task(rq, prev);
2789 	clear_tsk_need_resched(prev);
2790 	clear_preempt_need_resched();
2791 	rq->clock_skip_update = 0;
2792 
2793 	if (likely(prev != next)) {
2794 		rq->nr_switches++;
2795 		rq->curr = next;
2796 		++*switch_count;
2797 
2798 		rq = context_switch(rq, prev, next); /* unlocks the rq */
2799 		cpu = cpu_of(rq);
2800 	} else
2801 		raw_spin_unlock_irq(&rq->lock);
2802 
2803 	post_schedule(rq);
2804 
2805 	sched_preempt_enable_no_resched();
2806 }
2807 
2808 static inline void sched_submit_work(struct task_struct *tsk)
2809 {
2810 	if (!tsk->state || tsk_is_pi_blocked(tsk))
2811 		return;
2812 	/*
2813 	 * If we are going to sleep and we have plugged IO queued,
2814 	 * make sure to submit it to avoid deadlocks.
2815 	 */
2816 	if (blk_needs_flush_plug(tsk))
2817 		blk_schedule_flush_plug(tsk);
2818 }
2819 
2820 asmlinkage __visible void __sched schedule(void)
2821 {
2822 	struct task_struct *tsk = current;
2823 
2824 	sched_submit_work(tsk);
2825 	do {
2826 		__schedule();
2827 	} while (need_resched());
2828 }
2829 EXPORT_SYMBOL(schedule);
2830 
2831 #ifdef CONFIG_CONTEXT_TRACKING
2832 asmlinkage __visible void __sched schedule_user(void)
2833 {
2834 	/*
2835 	 * If we come here after a random call to set_need_resched(),
2836 	 * or we have been woken up remotely but the IPI has not yet arrived,
2837 	 * we haven't yet exited the RCU idle mode. Do it here manually until
2838 	 * we find a better solution.
2839 	 *
2840 	 * NB: There are buggy callers of this function.  Ideally we
2841 	 * should warn if prev_state != CONTEXT_USER, but that will trigger
2842 	 * too frequently to make sense yet.
2843 	 */
2844 	enum ctx_state prev_state = exception_enter();
2845 	schedule();
2846 	exception_exit(prev_state);
2847 }
2848 #endif
2849 
2850 /**
2851  * schedule_preempt_disabled - called with preemption disabled
2852  *
2853  * Returns with preemption disabled. Note: preempt_count must be 1
2854  */
2855 void __sched schedule_preempt_disabled(void)
2856 {
2857 	sched_preempt_enable_no_resched();
2858 	schedule();
2859 	preempt_disable();
2860 }
2861 
2862 static void __sched notrace preempt_schedule_common(void)
2863 {
2864 	do {
2865 		__preempt_count_add(PREEMPT_ACTIVE);
2866 		__schedule();
2867 		__preempt_count_sub(PREEMPT_ACTIVE);
2868 
2869 		/*
2870 		 * Check again in case we missed a preemption opportunity
2871 		 * between schedule and now.
2872 		 */
2873 		barrier();
2874 	} while (need_resched());
2875 }
2876 
2877 #ifdef CONFIG_PREEMPT
2878 /*
2879  * this is the entry point to schedule() from in-kernel preemption
2880  * off of preempt_enable. Kernel preemptions off return from interrupt
2881  * occur there and call schedule directly.
2882  */
2883 asmlinkage __visible void __sched notrace preempt_schedule(void)
2884 {
2885 	/*
2886 	 * If there is a non-zero preempt_count or interrupts are disabled,
2887 	 * we do not want to preempt the current task. Just return..
2888 	 */
2889 	if (likely(!preemptible()))
2890 		return;
2891 
2892 	preempt_schedule_common();
2893 }
2894 NOKPROBE_SYMBOL(preempt_schedule);
2895 EXPORT_SYMBOL(preempt_schedule);
2896 
2897 #ifdef CONFIG_CONTEXT_TRACKING
2898 /**
2899  * preempt_schedule_context - preempt_schedule called by tracing
2900  *
2901  * The tracing infrastructure uses preempt_enable_notrace to prevent
2902  * recursion and tracing preempt enabling caused by the tracing
2903  * infrastructure itself. But as tracing can happen in areas coming
2904  * from userspace or just about to enter userspace, a preempt enable
2905  * can occur before user_exit() is called. This will cause the scheduler
2906  * to be called when the system is still in usermode.
2907  *
2908  * To prevent this, the preempt_enable_notrace will use this function
2909  * instead of preempt_schedule() to exit user context if needed before
2910  * calling the scheduler.
2911  */
2912 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2913 {
2914 	enum ctx_state prev_ctx;
2915 
2916 	if (likely(!preemptible()))
2917 		return;
2918 
2919 	do {
2920 		__preempt_count_add(PREEMPT_ACTIVE);
2921 		/*
2922 		 * Needs preempt disabled in case user_exit() is traced
2923 		 * and the tracer calls preempt_enable_notrace() causing
2924 		 * an infinite recursion.
2925 		 */
2926 		prev_ctx = exception_enter();
2927 		__schedule();
2928 		exception_exit(prev_ctx);
2929 
2930 		__preempt_count_sub(PREEMPT_ACTIVE);
2931 		barrier();
2932 	} while (need_resched());
2933 }
2934 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2935 #endif /* CONFIG_CONTEXT_TRACKING */
2936 
2937 #endif /* CONFIG_PREEMPT */
2938 
2939 /*
2940  * this is the entry point to schedule() from kernel preemption
2941  * off of irq context.
2942  * Note, that this is called and return with irqs disabled. This will
2943  * protect us against recursive calling from irq.
2944  */
2945 asmlinkage __visible void __sched preempt_schedule_irq(void)
2946 {
2947 	enum ctx_state prev_state;
2948 
2949 	/* Catch callers which need to be fixed */
2950 	BUG_ON(preempt_count() || !irqs_disabled());
2951 
2952 	prev_state = exception_enter();
2953 
2954 	do {
2955 		__preempt_count_add(PREEMPT_ACTIVE);
2956 		local_irq_enable();
2957 		__schedule();
2958 		local_irq_disable();
2959 		__preempt_count_sub(PREEMPT_ACTIVE);
2960 
2961 		/*
2962 		 * Check again in case we missed a preemption opportunity
2963 		 * between schedule and now.
2964 		 */
2965 		barrier();
2966 	} while (need_resched());
2967 
2968 	exception_exit(prev_state);
2969 }
2970 
2971 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2972 			  void *key)
2973 {
2974 	return try_to_wake_up(curr->private, mode, wake_flags);
2975 }
2976 EXPORT_SYMBOL(default_wake_function);
2977 
2978 #ifdef CONFIG_RT_MUTEXES
2979 
2980 /*
2981  * rt_mutex_setprio - set the current priority of a task
2982  * @p: task
2983  * @prio: prio value (kernel-internal form)
2984  *
2985  * This function changes the 'effective' priority of a task. It does
2986  * not touch ->normal_prio like __setscheduler().
2987  *
2988  * Used by the rt_mutex code to implement priority inheritance
2989  * logic. Call site only calls if the priority of the task changed.
2990  */
2991 void rt_mutex_setprio(struct task_struct *p, int prio)
2992 {
2993 	int oldprio, queued, running, enqueue_flag = 0;
2994 	struct rq *rq;
2995 	const struct sched_class *prev_class;
2996 
2997 	BUG_ON(prio > MAX_PRIO);
2998 
2999 	rq = __task_rq_lock(p);
3000 
3001 	/*
3002 	 * Idle task boosting is a nono in general. There is one
3003 	 * exception, when PREEMPT_RT and NOHZ is active:
3004 	 *
3005 	 * The idle task calls get_next_timer_interrupt() and holds
3006 	 * the timer wheel base->lock on the CPU and another CPU wants
3007 	 * to access the timer (probably to cancel it). We can safely
3008 	 * ignore the boosting request, as the idle CPU runs this code
3009 	 * with interrupts disabled and will complete the lock
3010 	 * protected section without being interrupted. So there is no
3011 	 * real need to boost.
3012 	 */
3013 	if (unlikely(p == rq->idle)) {
3014 		WARN_ON(p != rq->curr);
3015 		WARN_ON(p->pi_blocked_on);
3016 		goto out_unlock;
3017 	}
3018 
3019 	trace_sched_pi_setprio(p, prio);
3020 	oldprio = p->prio;
3021 	prev_class = p->sched_class;
3022 	queued = task_on_rq_queued(p);
3023 	running = task_current(rq, p);
3024 	if (queued)
3025 		dequeue_task(rq, p, 0);
3026 	if (running)
3027 		put_prev_task(rq, p);
3028 
3029 	/*
3030 	 * Boosting condition are:
3031 	 * 1. -rt task is running and holds mutex A
3032 	 *      --> -dl task blocks on mutex A
3033 	 *
3034 	 * 2. -dl task is running and holds mutex A
3035 	 *      --> -dl task blocks on mutex A and could preempt the
3036 	 *          running task
3037 	 */
3038 	if (dl_prio(prio)) {
3039 		struct task_struct *pi_task = rt_mutex_get_top_task(p);
3040 		if (!dl_prio(p->normal_prio) ||
3041 		    (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3042 			p->dl.dl_boosted = 1;
3043 			p->dl.dl_throttled = 0;
3044 			enqueue_flag = ENQUEUE_REPLENISH;
3045 		} else
3046 			p->dl.dl_boosted = 0;
3047 		p->sched_class = &dl_sched_class;
3048 	} else if (rt_prio(prio)) {
3049 		if (dl_prio(oldprio))
3050 			p->dl.dl_boosted = 0;
3051 		if (oldprio < prio)
3052 			enqueue_flag = ENQUEUE_HEAD;
3053 		p->sched_class = &rt_sched_class;
3054 	} else {
3055 		if (dl_prio(oldprio))
3056 			p->dl.dl_boosted = 0;
3057 		if (rt_prio(oldprio))
3058 			p->rt.timeout = 0;
3059 		p->sched_class = &fair_sched_class;
3060 	}
3061 
3062 	p->prio = prio;
3063 
3064 	if (running)
3065 		p->sched_class->set_curr_task(rq);
3066 	if (queued)
3067 		enqueue_task(rq, p, enqueue_flag);
3068 
3069 	check_class_changed(rq, p, prev_class, oldprio);
3070 out_unlock:
3071 	__task_rq_unlock(rq);
3072 }
3073 #endif
3074 
3075 void set_user_nice(struct task_struct *p, long nice)
3076 {
3077 	int old_prio, delta, queued;
3078 	unsigned long flags;
3079 	struct rq *rq;
3080 
3081 	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3082 		return;
3083 	/*
3084 	 * We have to be careful, if called from sys_setpriority(),
3085 	 * the task might be in the middle of scheduling on another CPU.
3086 	 */
3087 	rq = task_rq_lock(p, &flags);
3088 	/*
3089 	 * The RT priorities are set via sched_setscheduler(), but we still
3090 	 * allow the 'normal' nice value to be set - but as expected
3091 	 * it wont have any effect on scheduling until the task is
3092 	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3093 	 */
3094 	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3095 		p->static_prio = NICE_TO_PRIO(nice);
3096 		goto out_unlock;
3097 	}
3098 	queued = task_on_rq_queued(p);
3099 	if (queued)
3100 		dequeue_task(rq, p, 0);
3101 
3102 	p->static_prio = NICE_TO_PRIO(nice);
3103 	set_load_weight(p);
3104 	old_prio = p->prio;
3105 	p->prio = effective_prio(p);
3106 	delta = p->prio - old_prio;
3107 
3108 	if (queued) {
3109 		enqueue_task(rq, p, 0);
3110 		/*
3111 		 * If the task increased its priority or is running and
3112 		 * lowered its priority, then reschedule its CPU:
3113 		 */
3114 		if (delta < 0 || (delta > 0 && task_running(rq, p)))
3115 			resched_curr(rq);
3116 	}
3117 out_unlock:
3118 	task_rq_unlock(rq, p, &flags);
3119 }
3120 EXPORT_SYMBOL(set_user_nice);
3121 
3122 /*
3123  * can_nice - check if a task can reduce its nice value
3124  * @p: task
3125  * @nice: nice value
3126  */
3127 int can_nice(const struct task_struct *p, const int nice)
3128 {
3129 	/* convert nice value [19,-20] to rlimit style value [1,40] */
3130 	int nice_rlim = nice_to_rlimit(nice);
3131 
3132 	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3133 		capable(CAP_SYS_NICE));
3134 }
3135 
3136 #ifdef __ARCH_WANT_SYS_NICE
3137 
3138 /*
3139  * sys_nice - change the priority of the current process.
3140  * @increment: priority increment
3141  *
3142  * sys_setpriority is a more generic, but much slower function that
3143  * does similar things.
3144  */
3145 SYSCALL_DEFINE1(nice, int, increment)
3146 {
3147 	long nice, retval;
3148 
3149 	/*
3150 	 * Setpriority might change our priority at the same moment.
3151 	 * We don't have to worry. Conceptually one call occurs first
3152 	 * and we have a single winner.
3153 	 */
3154 	increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3155 	nice = task_nice(current) + increment;
3156 
3157 	nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3158 	if (increment < 0 && !can_nice(current, nice))
3159 		return -EPERM;
3160 
3161 	retval = security_task_setnice(current, nice);
3162 	if (retval)
3163 		return retval;
3164 
3165 	set_user_nice(current, nice);
3166 	return 0;
3167 }
3168 
3169 #endif
3170 
3171 /**
3172  * task_prio - return the priority value of a given task.
3173  * @p: the task in question.
3174  *
3175  * Return: The priority value as seen by users in /proc.
3176  * RT tasks are offset by -200. Normal tasks are centered
3177  * around 0, value goes from -16 to +15.
3178  */
3179 int task_prio(const struct task_struct *p)
3180 {
3181 	return p->prio - MAX_RT_PRIO;
3182 }
3183 
3184 /**
3185  * idle_cpu - is a given cpu idle currently?
3186  * @cpu: the processor in question.
3187  *
3188  * Return: 1 if the CPU is currently idle. 0 otherwise.
3189  */
3190 int idle_cpu(int cpu)
3191 {
3192 	struct rq *rq = cpu_rq(cpu);
3193 
3194 	if (rq->curr != rq->idle)
3195 		return 0;
3196 
3197 	if (rq->nr_running)
3198 		return 0;
3199 
3200 #ifdef CONFIG_SMP
3201 	if (!llist_empty(&rq->wake_list))
3202 		return 0;
3203 #endif
3204 
3205 	return 1;
3206 }
3207 
3208 /**
3209  * idle_task - return the idle task for a given cpu.
3210  * @cpu: the processor in question.
3211  *
3212  * Return: The idle task for the cpu @cpu.
3213  */
3214 struct task_struct *idle_task(int cpu)
3215 {
3216 	return cpu_rq(cpu)->idle;
3217 }
3218 
3219 /**
3220  * find_process_by_pid - find a process with a matching PID value.
3221  * @pid: the pid in question.
3222  *
3223  * The task of @pid, if found. %NULL otherwise.
3224  */
3225 static struct task_struct *find_process_by_pid(pid_t pid)
3226 {
3227 	return pid ? find_task_by_vpid(pid) : current;
3228 }
3229 
3230 /*
3231  * This function initializes the sched_dl_entity of a newly becoming
3232  * SCHED_DEADLINE task.
3233  *
3234  * Only the static values are considered here, the actual runtime and the
3235  * absolute deadline will be properly calculated when the task is enqueued
3236  * for the first time with its new policy.
3237  */
3238 static void
3239 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3240 {
3241 	struct sched_dl_entity *dl_se = &p->dl;
3242 
3243 	dl_se->dl_runtime = attr->sched_runtime;
3244 	dl_se->dl_deadline = attr->sched_deadline;
3245 	dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3246 	dl_se->flags = attr->sched_flags;
3247 	dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3248 
3249 	/*
3250 	 * Changing the parameters of a task is 'tricky' and we're not doing
3251 	 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3252 	 *
3253 	 * What we SHOULD do is delay the bandwidth release until the 0-lag
3254 	 * point. This would include retaining the task_struct until that time
3255 	 * and change dl_overflow() to not immediately decrement the current
3256 	 * amount.
3257 	 *
3258 	 * Instead we retain the current runtime/deadline and let the new
3259 	 * parameters take effect after the current reservation period lapses.
3260 	 * This is safe (albeit pessimistic) because the 0-lag point is always
3261 	 * before the current scheduling deadline.
3262 	 *
3263 	 * We can still have temporary overloads because we do not delay the
3264 	 * change in bandwidth until that time; so admission control is
3265 	 * not on the safe side. It does however guarantee tasks will never
3266 	 * consume more than promised.
3267 	 */
3268 }
3269 
3270 /*
3271  * sched_setparam() passes in -1 for its policy, to let the functions
3272  * it calls know not to change it.
3273  */
3274 #define SETPARAM_POLICY	-1
3275 
3276 static void __setscheduler_params(struct task_struct *p,
3277 		const struct sched_attr *attr)
3278 {
3279 	int policy = attr->sched_policy;
3280 
3281 	if (policy == SETPARAM_POLICY)
3282 		policy = p->policy;
3283 
3284 	p->policy = policy;
3285 
3286 	if (dl_policy(policy))
3287 		__setparam_dl(p, attr);
3288 	else if (fair_policy(policy))
3289 		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3290 
3291 	/*
3292 	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3293 	 * !rt_policy. Always setting this ensures that things like
3294 	 * getparam()/getattr() don't report silly values for !rt tasks.
3295 	 */
3296 	p->rt_priority = attr->sched_priority;
3297 	p->normal_prio = normal_prio(p);
3298 	set_load_weight(p);
3299 }
3300 
3301 /* Actually do priority change: must hold pi & rq lock. */
3302 static void __setscheduler(struct rq *rq, struct task_struct *p,
3303 			   const struct sched_attr *attr, bool keep_boost)
3304 {
3305 	__setscheduler_params(p, attr);
3306 
3307 	/*
3308 	 * Keep a potential priority boosting if called from
3309 	 * sched_setscheduler().
3310 	 */
3311 	if (keep_boost)
3312 		p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3313 	else
3314 		p->prio = normal_prio(p);
3315 
3316 	if (dl_prio(p->prio))
3317 		p->sched_class = &dl_sched_class;
3318 	else if (rt_prio(p->prio))
3319 		p->sched_class = &rt_sched_class;
3320 	else
3321 		p->sched_class = &fair_sched_class;
3322 }
3323 
3324 static void
3325 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3326 {
3327 	struct sched_dl_entity *dl_se = &p->dl;
3328 
3329 	attr->sched_priority = p->rt_priority;
3330 	attr->sched_runtime = dl_se->dl_runtime;
3331 	attr->sched_deadline = dl_se->dl_deadline;
3332 	attr->sched_period = dl_se->dl_period;
3333 	attr->sched_flags = dl_se->flags;
3334 }
3335 
3336 /*
3337  * This function validates the new parameters of a -deadline task.
3338  * We ask for the deadline not being zero, and greater or equal
3339  * than the runtime, as well as the period of being zero or
3340  * greater than deadline. Furthermore, we have to be sure that
3341  * user parameters are above the internal resolution of 1us (we
3342  * check sched_runtime only since it is always the smaller one) and
3343  * below 2^63 ns (we have to check both sched_deadline and
3344  * sched_period, as the latter can be zero).
3345  */
3346 static bool
3347 __checkparam_dl(const struct sched_attr *attr)
3348 {
3349 	/* deadline != 0 */
3350 	if (attr->sched_deadline == 0)
3351 		return false;
3352 
3353 	/*
3354 	 * Since we truncate DL_SCALE bits, make sure we're at least
3355 	 * that big.
3356 	 */
3357 	if (attr->sched_runtime < (1ULL << DL_SCALE))
3358 		return false;
3359 
3360 	/*
3361 	 * Since we use the MSB for wrap-around and sign issues, make
3362 	 * sure it's not set (mind that period can be equal to zero).
3363 	 */
3364 	if (attr->sched_deadline & (1ULL << 63) ||
3365 	    attr->sched_period & (1ULL << 63))
3366 		return false;
3367 
3368 	/* runtime <= deadline <= period (if period != 0) */
3369 	if ((attr->sched_period != 0 &&
3370 	     attr->sched_period < attr->sched_deadline) ||
3371 	    attr->sched_deadline < attr->sched_runtime)
3372 		return false;
3373 
3374 	return true;
3375 }
3376 
3377 /*
3378  * check the target process has a UID that matches the current process's
3379  */
3380 static bool check_same_owner(struct task_struct *p)
3381 {
3382 	const struct cred *cred = current_cred(), *pcred;
3383 	bool match;
3384 
3385 	rcu_read_lock();
3386 	pcred = __task_cred(p);
3387 	match = (uid_eq(cred->euid, pcred->euid) ||
3388 		 uid_eq(cred->euid, pcred->uid));
3389 	rcu_read_unlock();
3390 	return match;
3391 }
3392 
3393 static bool dl_param_changed(struct task_struct *p,
3394 		const struct sched_attr *attr)
3395 {
3396 	struct sched_dl_entity *dl_se = &p->dl;
3397 
3398 	if (dl_se->dl_runtime != attr->sched_runtime ||
3399 		dl_se->dl_deadline != attr->sched_deadline ||
3400 		dl_se->dl_period != attr->sched_period ||
3401 		dl_se->flags != attr->sched_flags)
3402 		return true;
3403 
3404 	return false;
3405 }
3406 
3407 static int __sched_setscheduler(struct task_struct *p,
3408 				const struct sched_attr *attr,
3409 				bool user)
3410 {
3411 	int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3412 		      MAX_RT_PRIO - 1 - attr->sched_priority;
3413 	int retval, oldprio, oldpolicy = -1, queued, running;
3414 	int new_effective_prio, policy = attr->sched_policy;
3415 	unsigned long flags;
3416 	const struct sched_class *prev_class;
3417 	struct rq *rq;
3418 	int reset_on_fork;
3419 
3420 	/* may grab non-irq protected spin_locks */
3421 	BUG_ON(in_interrupt());
3422 recheck:
3423 	/* double check policy once rq lock held */
3424 	if (policy < 0) {
3425 		reset_on_fork = p->sched_reset_on_fork;
3426 		policy = oldpolicy = p->policy;
3427 	} else {
3428 		reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3429 
3430 		if (policy != SCHED_DEADLINE &&
3431 				policy != SCHED_FIFO && policy != SCHED_RR &&
3432 				policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3433 				policy != SCHED_IDLE)
3434 			return -EINVAL;
3435 	}
3436 
3437 	if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3438 		return -EINVAL;
3439 
3440 	/*
3441 	 * Valid priorities for SCHED_FIFO and SCHED_RR are
3442 	 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3443 	 * SCHED_BATCH and SCHED_IDLE is 0.
3444 	 */
3445 	if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3446 	    (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3447 		return -EINVAL;
3448 	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3449 	    (rt_policy(policy) != (attr->sched_priority != 0)))
3450 		return -EINVAL;
3451 
3452 	/*
3453 	 * Allow unprivileged RT tasks to decrease priority:
3454 	 */
3455 	if (user && !capable(CAP_SYS_NICE)) {
3456 		if (fair_policy(policy)) {
3457 			if (attr->sched_nice < task_nice(p) &&
3458 			    !can_nice(p, attr->sched_nice))
3459 				return -EPERM;
3460 		}
3461 
3462 		if (rt_policy(policy)) {
3463 			unsigned long rlim_rtprio =
3464 					task_rlimit(p, RLIMIT_RTPRIO);
3465 
3466 			/* can't set/change the rt policy */
3467 			if (policy != p->policy && !rlim_rtprio)
3468 				return -EPERM;
3469 
3470 			/* can't increase priority */
3471 			if (attr->sched_priority > p->rt_priority &&
3472 			    attr->sched_priority > rlim_rtprio)
3473 				return -EPERM;
3474 		}
3475 
3476 		 /*
3477 		  * Can't set/change SCHED_DEADLINE policy at all for now
3478 		  * (safest behavior); in the future we would like to allow
3479 		  * unprivileged DL tasks to increase their relative deadline
3480 		  * or reduce their runtime (both ways reducing utilization)
3481 		  */
3482 		if (dl_policy(policy))
3483 			return -EPERM;
3484 
3485 		/*
3486 		 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3487 		 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3488 		 */
3489 		if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3490 			if (!can_nice(p, task_nice(p)))
3491 				return -EPERM;
3492 		}
3493 
3494 		/* can't change other user's priorities */
3495 		if (!check_same_owner(p))
3496 			return -EPERM;
3497 
3498 		/* Normal users shall not reset the sched_reset_on_fork flag */
3499 		if (p->sched_reset_on_fork && !reset_on_fork)
3500 			return -EPERM;
3501 	}
3502 
3503 	if (user) {
3504 		retval = security_task_setscheduler(p);
3505 		if (retval)
3506 			return retval;
3507 	}
3508 
3509 	/*
3510 	 * make sure no PI-waiters arrive (or leave) while we are
3511 	 * changing the priority of the task:
3512 	 *
3513 	 * To be able to change p->policy safely, the appropriate
3514 	 * runqueue lock must be held.
3515 	 */
3516 	rq = task_rq_lock(p, &flags);
3517 
3518 	/*
3519 	 * Changing the policy of the stop threads its a very bad idea
3520 	 */
3521 	if (p == rq->stop) {
3522 		task_rq_unlock(rq, p, &flags);
3523 		return -EINVAL;
3524 	}
3525 
3526 	/*
3527 	 * If not changing anything there's no need to proceed further,
3528 	 * but store a possible modification of reset_on_fork.
3529 	 */
3530 	if (unlikely(policy == p->policy)) {
3531 		if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3532 			goto change;
3533 		if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3534 			goto change;
3535 		if (dl_policy(policy) && dl_param_changed(p, attr))
3536 			goto change;
3537 
3538 		p->sched_reset_on_fork = reset_on_fork;
3539 		task_rq_unlock(rq, p, &flags);
3540 		return 0;
3541 	}
3542 change:
3543 
3544 	if (user) {
3545 #ifdef CONFIG_RT_GROUP_SCHED
3546 		/*
3547 		 * Do not allow realtime tasks into groups that have no runtime
3548 		 * assigned.
3549 		 */
3550 		if (rt_bandwidth_enabled() && rt_policy(policy) &&
3551 				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3552 				!task_group_is_autogroup(task_group(p))) {
3553 			task_rq_unlock(rq, p, &flags);
3554 			return -EPERM;
3555 		}
3556 #endif
3557 #ifdef CONFIG_SMP
3558 		if (dl_bandwidth_enabled() && dl_policy(policy)) {
3559 			cpumask_t *span = rq->rd->span;
3560 
3561 			/*
3562 			 * Don't allow tasks with an affinity mask smaller than
3563 			 * the entire root_domain to become SCHED_DEADLINE. We
3564 			 * will also fail if there's no bandwidth available.
3565 			 */
3566 			if (!cpumask_subset(span, &p->cpus_allowed) ||
3567 			    rq->rd->dl_bw.bw == 0) {
3568 				task_rq_unlock(rq, p, &flags);
3569 				return -EPERM;
3570 			}
3571 		}
3572 #endif
3573 	}
3574 
3575 	/* recheck policy now with rq lock held */
3576 	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3577 		policy = oldpolicy = -1;
3578 		task_rq_unlock(rq, p, &flags);
3579 		goto recheck;
3580 	}
3581 
3582 	/*
3583 	 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3584 	 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3585 	 * is available.
3586 	 */
3587 	if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3588 		task_rq_unlock(rq, p, &flags);
3589 		return -EBUSY;
3590 	}
3591 
3592 	p->sched_reset_on_fork = reset_on_fork;
3593 	oldprio = p->prio;
3594 
3595 	/*
3596 	 * Take priority boosted tasks into account. If the new
3597 	 * effective priority is unchanged, we just store the new
3598 	 * normal parameters and do not touch the scheduler class and
3599 	 * the runqueue. This will be done when the task deboost
3600 	 * itself.
3601 	 */
3602 	new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3603 	if (new_effective_prio == oldprio) {
3604 		__setscheduler_params(p, attr);
3605 		task_rq_unlock(rq, p, &flags);
3606 		return 0;
3607 	}
3608 
3609 	queued = task_on_rq_queued(p);
3610 	running = task_current(rq, p);
3611 	if (queued)
3612 		dequeue_task(rq, p, 0);
3613 	if (running)
3614 		put_prev_task(rq, p);
3615 
3616 	prev_class = p->sched_class;
3617 	__setscheduler(rq, p, attr, true);
3618 
3619 	if (running)
3620 		p->sched_class->set_curr_task(rq);
3621 	if (queued) {
3622 		/*
3623 		 * We enqueue to tail when the priority of a task is
3624 		 * increased (user space view).
3625 		 */
3626 		enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3627 	}
3628 
3629 	check_class_changed(rq, p, prev_class, oldprio);
3630 	task_rq_unlock(rq, p, &flags);
3631 
3632 	rt_mutex_adjust_pi(p);
3633 
3634 	return 0;
3635 }
3636 
3637 static int _sched_setscheduler(struct task_struct *p, int policy,
3638 			       const struct sched_param *param, bool check)
3639 {
3640 	struct sched_attr attr = {
3641 		.sched_policy   = policy,
3642 		.sched_priority = param->sched_priority,
3643 		.sched_nice	= PRIO_TO_NICE(p->static_prio),
3644 	};
3645 
3646 	/* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3647 	if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3648 		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3649 		policy &= ~SCHED_RESET_ON_FORK;
3650 		attr.sched_policy = policy;
3651 	}
3652 
3653 	return __sched_setscheduler(p, &attr, check);
3654 }
3655 /**
3656  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3657  * @p: the task in question.
3658  * @policy: new policy.
3659  * @param: structure containing the new RT priority.
3660  *
3661  * Return: 0 on success. An error code otherwise.
3662  *
3663  * NOTE that the task may be already dead.
3664  */
3665 int sched_setscheduler(struct task_struct *p, int policy,
3666 		       const struct sched_param *param)
3667 {
3668 	return _sched_setscheduler(p, policy, param, true);
3669 }
3670 EXPORT_SYMBOL_GPL(sched_setscheduler);
3671 
3672 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3673 {
3674 	return __sched_setscheduler(p, attr, true);
3675 }
3676 EXPORT_SYMBOL_GPL(sched_setattr);
3677 
3678 /**
3679  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3680  * @p: the task in question.
3681  * @policy: new policy.
3682  * @param: structure containing the new RT priority.
3683  *
3684  * Just like sched_setscheduler, only don't bother checking if the
3685  * current context has permission.  For example, this is needed in
3686  * stop_machine(): we create temporary high priority worker threads,
3687  * but our caller might not have that capability.
3688  *
3689  * Return: 0 on success. An error code otherwise.
3690  */
3691 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3692 			       const struct sched_param *param)
3693 {
3694 	return _sched_setscheduler(p, policy, param, false);
3695 }
3696 
3697 static int
3698 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3699 {
3700 	struct sched_param lparam;
3701 	struct task_struct *p;
3702 	int retval;
3703 
3704 	if (!param || pid < 0)
3705 		return -EINVAL;
3706 	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3707 		return -EFAULT;
3708 
3709 	rcu_read_lock();
3710 	retval = -ESRCH;
3711 	p = find_process_by_pid(pid);
3712 	if (p != NULL)
3713 		retval = sched_setscheduler(p, policy, &lparam);
3714 	rcu_read_unlock();
3715 
3716 	return retval;
3717 }
3718 
3719 /*
3720  * Mimics kernel/events/core.c perf_copy_attr().
3721  */
3722 static int sched_copy_attr(struct sched_attr __user *uattr,
3723 			   struct sched_attr *attr)
3724 {
3725 	u32 size;
3726 	int ret;
3727 
3728 	if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3729 		return -EFAULT;
3730 
3731 	/*
3732 	 * zero the full structure, so that a short copy will be nice.
3733 	 */
3734 	memset(attr, 0, sizeof(*attr));
3735 
3736 	ret = get_user(size, &uattr->size);
3737 	if (ret)
3738 		return ret;
3739 
3740 	if (size > PAGE_SIZE)	/* silly large */
3741 		goto err_size;
3742 
3743 	if (!size)		/* abi compat */
3744 		size = SCHED_ATTR_SIZE_VER0;
3745 
3746 	if (size < SCHED_ATTR_SIZE_VER0)
3747 		goto err_size;
3748 
3749 	/*
3750 	 * If we're handed a bigger struct than we know of,
3751 	 * ensure all the unknown bits are 0 - i.e. new
3752 	 * user-space does not rely on any kernel feature
3753 	 * extensions we dont know about yet.
3754 	 */
3755 	if (size > sizeof(*attr)) {
3756 		unsigned char __user *addr;
3757 		unsigned char __user *end;
3758 		unsigned char val;
3759 
3760 		addr = (void __user *)uattr + sizeof(*attr);
3761 		end  = (void __user *)uattr + size;
3762 
3763 		for (; addr < end; addr++) {
3764 			ret = get_user(val, addr);
3765 			if (ret)
3766 				return ret;
3767 			if (val)
3768 				goto err_size;
3769 		}
3770 		size = sizeof(*attr);
3771 	}
3772 
3773 	ret = copy_from_user(attr, uattr, size);
3774 	if (ret)
3775 		return -EFAULT;
3776 
3777 	/*
3778 	 * XXX: do we want to be lenient like existing syscalls; or do we want
3779 	 * to be strict and return an error on out-of-bounds values?
3780 	 */
3781 	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3782 
3783 	return 0;
3784 
3785 err_size:
3786 	put_user(sizeof(*attr), &uattr->size);
3787 	return -E2BIG;
3788 }
3789 
3790 /**
3791  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3792  * @pid: the pid in question.
3793  * @policy: new policy.
3794  * @param: structure containing the new RT priority.
3795  *
3796  * Return: 0 on success. An error code otherwise.
3797  */
3798 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3799 		struct sched_param __user *, param)
3800 {
3801 	/* negative values for policy are not valid */
3802 	if (policy < 0)
3803 		return -EINVAL;
3804 
3805 	return do_sched_setscheduler(pid, policy, param);
3806 }
3807 
3808 /**
3809  * sys_sched_setparam - set/change the RT priority of a thread
3810  * @pid: the pid in question.
3811  * @param: structure containing the new RT priority.
3812  *
3813  * Return: 0 on success. An error code otherwise.
3814  */
3815 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3816 {
3817 	return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3818 }
3819 
3820 /**
3821  * sys_sched_setattr - same as above, but with extended sched_attr
3822  * @pid: the pid in question.
3823  * @uattr: structure containing the extended parameters.
3824  * @flags: for future extension.
3825  */
3826 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3827 			       unsigned int, flags)
3828 {
3829 	struct sched_attr attr;
3830 	struct task_struct *p;
3831 	int retval;
3832 
3833 	if (!uattr || pid < 0 || flags)
3834 		return -EINVAL;
3835 
3836 	retval = sched_copy_attr(uattr, &attr);
3837 	if (retval)
3838 		return retval;
3839 
3840 	if ((int)attr.sched_policy < 0)
3841 		return -EINVAL;
3842 
3843 	rcu_read_lock();
3844 	retval = -ESRCH;
3845 	p = find_process_by_pid(pid);
3846 	if (p != NULL)
3847 		retval = sched_setattr(p, &attr);
3848 	rcu_read_unlock();
3849 
3850 	return retval;
3851 }
3852 
3853 /**
3854  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3855  * @pid: the pid in question.
3856  *
3857  * Return: On success, the policy of the thread. Otherwise, a negative error
3858  * code.
3859  */
3860 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3861 {
3862 	struct task_struct *p;
3863 	int retval;
3864 
3865 	if (pid < 0)
3866 		return -EINVAL;
3867 
3868 	retval = -ESRCH;
3869 	rcu_read_lock();
3870 	p = find_process_by_pid(pid);
3871 	if (p) {
3872 		retval = security_task_getscheduler(p);
3873 		if (!retval)
3874 			retval = p->policy
3875 				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3876 	}
3877 	rcu_read_unlock();
3878 	return retval;
3879 }
3880 
3881 /**
3882  * sys_sched_getparam - get the RT priority of a thread
3883  * @pid: the pid in question.
3884  * @param: structure containing the RT priority.
3885  *
3886  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3887  * code.
3888  */
3889 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3890 {
3891 	struct sched_param lp = { .sched_priority = 0 };
3892 	struct task_struct *p;
3893 	int retval;
3894 
3895 	if (!param || pid < 0)
3896 		return -EINVAL;
3897 
3898 	rcu_read_lock();
3899 	p = find_process_by_pid(pid);
3900 	retval = -ESRCH;
3901 	if (!p)
3902 		goto out_unlock;
3903 
3904 	retval = security_task_getscheduler(p);
3905 	if (retval)
3906 		goto out_unlock;
3907 
3908 	if (task_has_rt_policy(p))
3909 		lp.sched_priority = p->rt_priority;
3910 	rcu_read_unlock();
3911 
3912 	/*
3913 	 * This one might sleep, we cannot do it with a spinlock held ...
3914 	 */
3915 	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3916 
3917 	return retval;
3918 
3919 out_unlock:
3920 	rcu_read_unlock();
3921 	return retval;
3922 }
3923 
3924 static int sched_read_attr(struct sched_attr __user *uattr,
3925 			   struct sched_attr *attr,
3926 			   unsigned int usize)
3927 {
3928 	int ret;
3929 
3930 	if (!access_ok(VERIFY_WRITE, uattr, usize))
3931 		return -EFAULT;
3932 
3933 	/*
3934 	 * If we're handed a smaller struct than we know of,
3935 	 * ensure all the unknown bits are 0 - i.e. old
3936 	 * user-space does not get uncomplete information.
3937 	 */
3938 	if (usize < sizeof(*attr)) {
3939 		unsigned char *addr;
3940 		unsigned char *end;
3941 
3942 		addr = (void *)attr + usize;
3943 		end  = (void *)attr + sizeof(*attr);
3944 
3945 		for (; addr < end; addr++) {
3946 			if (*addr)
3947 				return -EFBIG;
3948 		}
3949 
3950 		attr->size = usize;
3951 	}
3952 
3953 	ret = copy_to_user(uattr, attr, attr->size);
3954 	if (ret)
3955 		return -EFAULT;
3956 
3957 	return 0;
3958 }
3959 
3960 /**
3961  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3962  * @pid: the pid in question.
3963  * @uattr: structure containing the extended parameters.
3964  * @size: sizeof(attr) for fwd/bwd comp.
3965  * @flags: for future extension.
3966  */
3967 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3968 		unsigned int, size, unsigned int, flags)
3969 {
3970 	struct sched_attr attr = {
3971 		.size = sizeof(struct sched_attr),
3972 	};
3973 	struct task_struct *p;
3974 	int retval;
3975 
3976 	if (!uattr || pid < 0 || size > PAGE_SIZE ||
3977 	    size < SCHED_ATTR_SIZE_VER0 || flags)
3978 		return -EINVAL;
3979 
3980 	rcu_read_lock();
3981 	p = find_process_by_pid(pid);
3982 	retval = -ESRCH;
3983 	if (!p)
3984 		goto out_unlock;
3985 
3986 	retval = security_task_getscheduler(p);
3987 	if (retval)
3988 		goto out_unlock;
3989 
3990 	attr.sched_policy = p->policy;
3991 	if (p->sched_reset_on_fork)
3992 		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3993 	if (task_has_dl_policy(p))
3994 		__getparam_dl(p, &attr);
3995 	else if (task_has_rt_policy(p))
3996 		attr.sched_priority = p->rt_priority;
3997 	else
3998 		attr.sched_nice = task_nice(p);
3999 
4000 	rcu_read_unlock();
4001 
4002 	retval = sched_read_attr(uattr, &attr, size);
4003 	return retval;
4004 
4005 out_unlock:
4006 	rcu_read_unlock();
4007 	return retval;
4008 }
4009 
4010 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4011 {
4012 	cpumask_var_t cpus_allowed, new_mask;
4013 	struct task_struct *p;
4014 	int retval;
4015 
4016 	rcu_read_lock();
4017 
4018 	p = find_process_by_pid(pid);
4019 	if (!p) {
4020 		rcu_read_unlock();
4021 		return -ESRCH;
4022 	}
4023 
4024 	/* Prevent p going away */
4025 	get_task_struct(p);
4026 	rcu_read_unlock();
4027 
4028 	if (p->flags & PF_NO_SETAFFINITY) {
4029 		retval = -EINVAL;
4030 		goto out_put_task;
4031 	}
4032 	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4033 		retval = -ENOMEM;
4034 		goto out_put_task;
4035 	}
4036 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4037 		retval = -ENOMEM;
4038 		goto out_free_cpus_allowed;
4039 	}
4040 	retval = -EPERM;
4041 	if (!check_same_owner(p)) {
4042 		rcu_read_lock();
4043 		if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4044 			rcu_read_unlock();
4045 			goto out_free_new_mask;
4046 		}
4047 		rcu_read_unlock();
4048 	}
4049 
4050 	retval = security_task_setscheduler(p);
4051 	if (retval)
4052 		goto out_free_new_mask;
4053 
4054 
4055 	cpuset_cpus_allowed(p, cpus_allowed);
4056 	cpumask_and(new_mask, in_mask, cpus_allowed);
4057 
4058 	/*
4059 	 * Since bandwidth control happens on root_domain basis,
4060 	 * if admission test is enabled, we only admit -deadline
4061 	 * tasks allowed to run on all the CPUs in the task's
4062 	 * root_domain.
4063 	 */
4064 #ifdef CONFIG_SMP
4065 	if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4066 		rcu_read_lock();
4067 		if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4068 			retval = -EBUSY;
4069 			rcu_read_unlock();
4070 			goto out_free_new_mask;
4071 		}
4072 		rcu_read_unlock();
4073 	}
4074 #endif
4075 again:
4076 	retval = set_cpus_allowed_ptr(p, new_mask);
4077 
4078 	if (!retval) {
4079 		cpuset_cpus_allowed(p, cpus_allowed);
4080 		if (!cpumask_subset(new_mask, cpus_allowed)) {
4081 			/*
4082 			 * We must have raced with a concurrent cpuset
4083 			 * update. Just reset the cpus_allowed to the
4084 			 * cpuset's cpus_allowed
4085 			 */
4086 			cpumask_copy(new_mask, cpus_allowed);
4087 			goto again;
4088 		}
4089 	}
4090 out_free_new_mask:
4091 	free_cpumask_var(new_mask);
4092 out_free_cpus_allowed:
4093 	free_cpumask_var(cpus_allowed);
4094 out_put_task:
4095 	put_task_struct(p);
4096 	return retval;
4097 }
4098 
4099 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4100 			     struct cpumask *new_mask)
4101 {
4102 	if (len < cpumask_size())
4103 		cpumask_clear(new_mask);
4104 	else if (len > cpumask_size())
4105 		len = cpumask_size();
4106 
4107 	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4108 }
4109 
4110 /**
4111  * sys_sched_setaffinity - set the cpu affinity of a process
4112  * @pid: pid of the process
4113  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4114  * @user_mask_ptr: user-space pointer to the new cpu mask
4115  *
4116  * Return: 0 on success. An error code otherwise.
4117  */
4118 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4119 		unsigned long __user *, user_mask_ptr)
4120 {
4121 	cpumask_var_t new_mask;
4122 	int retval;
4123 
4124 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4125 		return -ENOMEM;
4126 
4127 	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4128 	if (retval == 0)
4129 		retval = sched_setaffinity(pid, new_mask);
4130 	free_cpumask_var(new_mask);
4131 	return retval;
4132 }
4133 
4134 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4135 {
4136 	struct task_struct *p;
4137 	unsigned long flags;
4138 	int retval;
4139 
4140 	rcu_read_lock();
4141 
4142 	retval = -ESRCH;
4143 	p = find_process_by_pid(pid);
4144 	if (!p)
4145 		goto out_unlock;
4146 
4147 	retval = security_task_getscheduler(p);
4148 	if (retval)
4149 		goto out_unlock;
4150 
4151 	raw_spin_lock_irqsave(&p->pi_lock, flags);
4152 	cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4153 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4154 
4155 out_unlock:
4156 	rcu_read_unlock();
4157 
4158 	return retval;
4159 }
4160 
4161 /**
4162  * sys_sched_getaffinity - get the cpu affinity of a process
4163  * @pid: pid of the process
4164  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4165  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4166  *
4167  * Return: 0 on success. An error code otherwise.
4168  */
4169 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4170 		unsigned long __user *, user_mask_ptr)
4171 {
4172 	int ret;
4173 	cpumask_var_t mask;
4174 
4175 	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4176 		return -EINVAL;
4177 	if (len & (sizeof(unsigned long)-1))
4178 		return -EINVAL;
4179 
4180 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4181 		return -ENOMEM;
4182 
4183 	ret = sched_getaffinity(pid, mask);
4184 	if (ret == 0) {
4185 		size_t retlen = min_t(size_t, len, cpumask_size());
4186 
4187 		if (copy_to_user(user_mask_ptr, mask, retlen))
4188 			ret = -EFAULT;
4189 		else
4190 			ret = retlen;
4191 	}
4192 	free_cpumask_var(mask);
4193 
4194 	return ret;
4195 }
4196 
4197 /**
4198  * sys_sched_yield - yield the current processor to other threads.
4199  *
4200  * This function yields the current CPU to other tasks. If there are no
4201  * other threads running on this CPU then this function will return.
4202  *
4203  * Return: 0.
4204  */
4205 SYSCALL_DEFINE0(sched_yield)
4206 {
4207 	struct rq *rq = this_rq_lock();
4208 
4209 	schedstat_inc(rq, yld_count);
4210 	current->sched_class->yield_task(rq);
4211 
4212 	/*
4213 	 * Since we are going to call schedule() anyway, there's
4214 	 * no need to preempt or enable interrupts:
4215 	 */
4216 	__release(rq->lock);
4217 	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4218 	do_raw_spin_unlock(&rq->lock);
4219 	sched_preempt_enable_no_resched();
4220 
4221 	schedule();
4222 
4223 	return 0;
4224 }
4225 
4226 int __sched _cond_resched(void)
4227 {
4228 	if (should_resched()) {
4229 		preempt_schedule_common();
4230 		return 1;
4231 	}
4232 	return 0;
4233 }
4234 EXPORT_SYMBOL(_cond_resched);
4235 
4236 /*
4237  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4238  * call schedule, and on return reacquire the lock.
4239  *
4240  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4241  * operations here to prevent schedule() from being called twice (once via
4242  * spin_unlock(), once by hand).
4243  */
4244 int __cond_resched_lock(spinlock_t *lock)
4245 {
4246 	int resched = should_resched();
4247 	int ret = 0;
4248 
4249 	lockdep_assert_held(lock);
4250 
4251 	if (spin_needbreak(lock) || resched) {
4252 		spin_unlock(lock);
4253 		if (resched)
4254 			preempt_schedule_common();
4255 		else
4256 			cpu_relax();
4257 		ret = 1;
4258 		spin_lock(lock);
4259 	}
4260 	return ret;
4261 }
4262 EXPORT_SYMBOL(__cond_resched_lock);
4263 
4264 int __sched __cond_resched_softirq(void)
4265 {
4266 	BUG_ON(!in_softirq());
4267 
4268 	if (should_resched()) {
4269 		local_bh_enable();
4270 		preempt_schedule_common();
4271 		local_bh_disable();
4272 		return 1;
4273 	}
4274 	return 0;
4275 }
4276 EXPORT_SYMBOL(__cond_resched_softirq);
4277 
4278 /**
4279  * yield - yield the current processor to other threads.
4280  *
4281  * Do not ever use this function, there's a 99% chance you're doing it wrong.
4282  *
4283  * The scheduler is at all times free to pick the calling task as the most
4284  * eligible task to run, if removing the yield() call from your code breaks
4285  * it, its already broken.
4286  *
4287  * Typical broken usage is:
4288  *
4289  * while (!event)
4290  * 	yield();
4291  *
4292  * where one assumes that yield() will let 'the other' process run that will
4293  * make event true. If the current task is a SCHED_FIFO task that will never
4294  * happen. Never use yield() as a progress guarantee!!
4295  *
4296  * If you want to use yield() to wait for something, use wait_event().
4297  * If you want to use yield() to be 'nice' for others, use cond_resched().
4298  * If you still want to use yield(), do not!
4299  */
4300 void __sched yield(void)
4301 {
4302 	set_current_state(TASK_RUNNING);
4303 	sys_sched_yield();
4304 }
4305 EXPORT_SYMBOL(yield);
4306 
4307 /**
4308  * yield_to - yield the current processor to another thread in
4309  * your thread group, or accelerate that thread toward the
4310  * processor it's on.
4311  * @p: target task
4312  * @preempt: whether task preemption is allowed or not
4313  *
4314  * It's the caller's job to ensure that the target task struct
4315  * can't go away on us before we can do any checks.
4316  *
4317  * Return:
4318  *	true (>0) if we indeed boosted the target task.
4319  *	false (0) if we failed to boost the target.
4320  *	-ESRCH if there's no task to yield to.
4321  */
4322 int __sched yield_to(struct task_struct *p, bool preempt)
4323 {
4324 	struct task_struct *curr = current;
4325 	struct rq *rq, *p_rq;
4326 	unsigned long flags;
4327 	int yielded = 0;
4328 
4329 	local_irq_save(flags);
4330 	rq = this_rq();
4331 
4332 again:
4333 	p_rq = task_rq(p);
4334 	/*
4335 	 * If we're the only runnable task on the rq and target rq also
4336 	 * has only one task, there's absolutely no point in yielding.
4337 	 */
4338 	if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4339 		yielded = -ESRCH;
4340 		goto out_irq;
4341 	}
4342 
4343 	double_rq_lock(rq, p_rq);
4344 	if (task_rq(p) != p_rq) {
4345 		double_rq_unlock(rq, p_rq);
4346 		goto again;
4347 	}
4348 
4349 	if (!curr->sched_class->yield_to_task)
4350 		goto out_unlock;
4351 
4352 	if (curr->sched_class != p->sched_class)
4353 		goto out_unlock;
4354 
4355 	if (task_running(p_rq, p) || p->state)
4356 		goto out_unlock;
4357 
4358 	yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4359 	if (yielded) {
4360 		schedstat_inc(rq, yld_count);
4361 		/*
4362 		 * Make p's CPU reschedule; pick_next_entity takes care of
4363 		 * fairness.
4364 		 */
4365 		if (preempt && rq != p_rq)
4366 			resched_curr(p_rq);
4367 	}
4368 
4369 out_unlock:
4370 	double_rq_unlock(rq, p_rq);
4371 out_irq:
4372 	local_irq_restore(flags);
4373 
4374 	if (yielded > 0)
4375 		schedule();
4376 
4377 	return yielded;
4378 }
4379 EXPORT_SYMBOL_GPL(yield_to);
4380 
4381 /*
4382  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4383  * that process accounting knows that this is a task in IO wait state.
4384  */
4385 long __sched io_schedule_timeout(long timeout)
4386 {
4387 	int old_iowait = current->in_iowait;
4388 	struct rq *rq;
4389 	long ret;
4390 
4391 	current->in_iowait = 1;
4392 	blk_schedule_flush_plug(current);
4393 
4394 	delayacct_blkio_start();
4395 	rq = raw_rq();
4396 	atomic_inc(&rq->nr_iowait);
4397 	ret = schedule_timeout(timeout);
4398 	current->in_iowait = old_iowait;
4399 	atomic_dec(&rq->nr_iowait);
4400 	delayacct_blkio_end();
4401 
4402 	return ret;
4403 }
4404 EXPORT_SYMBOL(io_schedule_timeout);
4405 
4406 /**
4407  * sys_sched_get_priority_max - return maximum RT priority.
4408  * @policy: scheduling class.
4409  *
4410  * Return: On success, this syscall returns the maximum
4411  * rt_priority that can be used by a given scheduling class.
4412  * On failure, a negative error code is returned.
4413  */
4414 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4415 {
4416 	int ret = -EINVAL;
4417 
4418 	switch (policy) {
4419 	case SCHED_FIFO:
4420 	case SCHED_RR:
4421 		ret = MAX_USER_RT_PRIO-1;
4422 		break;
4423 	case SCHED_DEADLINE:
4424 	case SCHED_NORMAL:
4425 	case SCHED_BATCH:
4426 	case SCHED_IDLE:
4427 		ret = 0;
4428 		break;
4429 	}
4430 	return ret;
4431 }
4432 
4433 /**
4434  * sys_sched_get_priority_min - return minimum RT priority.
4435  * @policy: scheduling class.
4436  *
4437  * Return: On success, this syscall returns the minimum
4438  * rt_priority that can be used by a given scheduling class.
4439  * On failure, a negative error code is returned.
4440  */
4441 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4442 {
4443 	int ret = -EINVAL;
4444 
4445 	switch (policy) {
4446 	case SCHED_FIFO:
4447 	case SCHED_RR:
4448 		ret = 1;
4449 		break;
4450 	case SCHED_DEADLINE:
4451 	case SCHED_NORMAL:
4452 	case SCHED_BATCH:
4453 	case SCHED_IDLE:
4454 		ret = 0;
4455 	}
4456 	return ret;
4457 }
4458 
4459 /**
4460  * sys_sched_rr_get_interval - return the default timeslice of a process.
4461  * @pid: pid of the process.
4462  * @interval: userspace pointer to the timeslice value.
4463  *
4464  * this syscall writes the default timeslice value of a given process
4465  * into the user-space timespec buffer. A value of '0' means infinity.
4466  *
4467  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4468  * an error code.
4469  */
4470 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4471 		struct timespec __user *, interval)
4472 {
4473 	struct task_struct *p;
4474 	unsigned int time_slice;
4475 	unsigned long flags;
4476 	struct rq *rq;
4477 	int retval;
4478 	struct timespec t;
4479 
4480 	if (pid < 0)
4481 		return -EINVAL;
4482 
4483 	retval = -ESRCH;
4484 	rcu_read_lock();
4485 	p = find_process_by_pid(pid);
4486 	if (!p)
4487 		goto out_unlock;
4488 
4489 	retval = security_task_getscheduler(p);
4490 	if (retval)
4491 		goto out_unlock;
4492 
4493 	rq = task_rq_lock(p, &flags);
4494 	time_slice = 0;
4495 	if (p->sched_class->get_rr_interval)
4496 		time_slice = p->sched_class->get_rr_interval(rq, p);
4497 	task_rq_unlock(rq, p, &flags);
4498 
4499 	rcu_read_unlock();
4500 	jiffies_to_timespec(time_slice, &t);
4501 	retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4502 	return retval;
4503 
4504 out_unlock:
4505 	rcu_read_unlock();
4506 	return retval;
4507 }
4508 
4509 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4510 
4511 void sched_show_task(struct task_struct *p)
4512 {
4513 	unsigned long free = 0;
4514 	int ppid;
4515 	unsigned long state = p->state;
4516 
4517 	if (state)
4518 		state = __ffs(state) + 1;
4519 	printk(KERN_INFO "%-15.15s %c", p->comm,
4520 		state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4521 #if BITS_PER_LONG == 32
4522 	if (state == TASK_RUNNING)
4523 		printk(KERN_CONT " running  ");
4524 	else
4525 		printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4526 #else
4527 	if (state == TASK_RUNNING)
4528 		printk(KERN_CONT "  running task    ");
4529 	else
4530 		printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4531 #endif
4532 #ifdef CONFIG_DEBUG_STACK_USAGE
4533 	free = stack_not_used(p);
4534 #endif
4535 	ppid = 0;
4536 	rcu_read_lock();
4537 	if (pid_alive(p))
4538 		ppid = task_pid_nr(rcu_dereference(p->real_parent));
4539 	rcu_read_unlock();
4540 	printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4541 		task_pid_nr(p), ppid,
4542 		(unsigned long)task_thread_info(p)->flags);
4543 
4544 	print_worker_info(KERN_INFO, p);
4545 	show_stack(p, NULL);
4546 }
4547 
4548 void show_state_filter(unsigned long state_filter)
4549 {
4550 	struct task_struct *g, *p;
4551 
4552 #if BITS_PER_LONG == 32
4553 	printk(KERN_INFO
4554 		"  task                PC stack   pid father\n");
4555 #else
4556 	printk(KERN_INFO
4557 		"  task                        PC stack   pid father\n");
4558 #endif
4559 	rcu_read_lock();
4560 	for_each_process_thread(g, p) {
4561 		/*
4562 		 * reset the NMI-timeout, listing all files on a slow
4563 		 * console might take a lot of time:
4564 		 */
4565 		touch_nmi_watchdog();
4566 		if (!state_filter || (p->state & state_filter))
4567 			sched_show_task(p);
4568 	}
4569 
4570 	touch_all_softlockup_watchdogs();
4571 
4572 #ifdef CONFIG_SCHED_DEBUG
4573 	sysrq_sched_debug_show();
4574 #endif
4575 	rcu_read_unlock();
4576 	/*
4577 	 * Only show locks if all tasks are dumped:
4578 	 */
4579 	if (!state_filter)
4580 		debug_show_all_locks();
4581 }
4582 
4583 void init_idle_bootup_task(struct task_struct *idle)
4584 {
4585 	idle->sched_class = &idle_sched_class;
4586 }
4587 
4588 /**
4589  * init_idle - set up an idle thread for a given CPU
4590  * @idle: task in question
4591  * @cpu: cpu the idle task belongs to
4592  *
4593  * NOTE: this function does not set the idle thread's NEED_RESCHED
4594  * flag, to make booting more robust.
4595  */
4596 void init_idle(struct task_struct *idle, int cpu)
4597 {
4598 	struct rq *rq = cpu_rq(cpu);
4599 	unsigned long flags;
4600 
4601 	raw_spin_lock_irqsave(&rq->lock, flags);
4602 
4603 	__sched_fork(0, idle);
4604 	idle->state = TASK_RUNNING;
4605 	idle->se.exec_start = sched_clock();
4606 
4607 	do_set_cpus_allowed(idle, cpumask_of(cpu));
4608 	/*
4609 	 * We're having a chicken and egg problem, even though we are
4610 	 * holding rq->lock, the cpu isn't yet set to this cpu so the
4611 	 * lockdep check in task_group() will fail.
4612 	 *
4613 	 * Similar case to sched_fork(). / Alternatively we could
4614 	 * use task_rq_lock() here and obtain the other rq->lock.
4615 	 *
4616 	 * Silence PROVE_RCU
4617 	 */
4618 	rcu_read_lock();
4619 	__set_task_cpu(idle, cpu);
4620 	rcu_read_unlock();
4621 
4622 	rq->curr = rq->idle = idle;
4623 	idle->on_rq = TASK_ON_RQ_QUEUED;
4624 #if defined(CONFIG_SMP)
4625 	idle->on_cpu = 1;
4626 #endif
4627 	raw_spin_unlock_irqrestore(&rq->lock, flags);
4628 
4629 	/* Set the preempt count _outside_ the spinlocks! */
4630 	init_idle_preempt_count(idle, cpu);
4631 
4632 	/*
4633 	 * The idle tasks have their own, simple scheduling class:
4634 	 */
4635 	idle->sched_class = &idle_sched_class;
4636 	ftrace_graph_init_idle_task(idle, cpu);
4637 	vtime_init_idle(idle, cpu);
4638 #if defined(CONFIG_SMP)
4639 	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4640 #endif
4641 }
4642 
4643 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4644 			      const struct cpumask *trial)
4645 {
4646 	int ret = 1, trial_cpus;
4647 	struct dl_bw *cur_dl_b;
4648 	unsigned long flags;
4649 
4650 	if (!cpumask_weight(cur))
4651 		return ret;
4652 
4653 	rcu_read_lock_sched();
4654 	cur_dl_b = dl_bw_of(cpumask_any(cur));
4655 	trial_cpus = cpumask_weight(trial);
4656 
4657 	raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4658 	if (cur_dl_b->bw != -1 &&
4659 	    cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4660 		ret = 0;
4661 	raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4662 	rcu_read_unlock_sched();
4663 
4664 	return ret;
4665 }
4666 
4667 int task_can_attach(struct task_struct *p,
4668 		    const struct cpumask *cs_cpus_allowed)
4669 {
4670 	int ret = 0;
4671 
4672 	/*
4673 	 * Kthreads which disallow setaffinity shouldn't be moved
4674 	 * to a new cpuset; we don't want to change their cpu
4675 	 * affinity and isolating such threads by their set of
4676 	 * allowed nodes is unnecessary.  Thus, cpusets are not
4677 	 * applicable for such threads.  This prevents checking for
4678 	 * success of set_cpus_allowed_ptr() on all attached tasks
4679 	 * before cpus_allowed may be changed.
4680 	 */
4681 	if (p->flags & PF_NO_SETAFFINITY) {
4682 		ret = -EINVAL;
4683 		goto out;
4684 	}
4685 
4686 #ifdef CONFIG_SMP
4687 	if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4688 					      cs_cpus_allowed)) {
4689 		unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4690 							cs_cpus_allowed);
4691 		struct dl_bw *dl_b;
4692 		bool overflow;
4693 		int cpus;
4694 		unsigned long flags;
4695 
4696 		rcu_read_lock_sched();
4697 		dl_b = dl_bw_of(dest_cpu);
4698 		raw_spin_lock_irqsave(&dl_b->lock, flags);
4699 		cpus = dl_bw_cpus(dest_cpu);
4700 		overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4701 		if (overflow)
4702 			ret = -EBUSY;
4703 		else {
4704 			/*
4705 			 * We reserve space for this task in the destination
4706 			 * root_domain, as we can't fail after this point.
4707 			 * We will free resources in the source root_domain
4708 			 * later on (see set_cpus_allowed_dl()).
4709 			 */
4710 			__dl_add(dl_b, p->dl.dl_bw);
4711 		}
4712 		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4713 		rcu_read_unlock_sched();
4714 
4715 	}
4716 #endif
4717 out:
4718 	return ret;
4719 }
4720 
4721 #ifdef CONFIG_SMP
4722 /*
4723  * move_queued_task - move a queued task to new rq.
4724  *
4725  * Returns (locked) new rq. Old rq's lock is released.
4726  */
4727 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4728 {
4729 	struct rq *rq = task_rq(p);
4730 
4731 	lockdep_assert_held(&rq->lock);
4732 
4733 	dequeue_task(rq, p, 0);
4734 	p->on_rq = TASK_ON_RQ_MIGRATING;
4735 	set_task_cpu(p, new_cpu);
4736 	raw_spin_unlock(&rq->lock);
4737 
4738 	rq = cpu_rq(new_cpu);
4739 
4740 	raw_spin_lock(&rq->lock);
4741 	BUG_ON(task_cpu(p) != new_cpu);
4742 	p->on_rq = TASK_ON_RQ_QUEUED;
4743 	enqueue_task(rq, p, 0);
4744 	check_preempt_curr(rq, p, 0);
4745 
4746 	return rq;
4747 }
4748 
4749 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4750 {
4751 	if (p->sched_class->set_cpus_allowed)
4752 		p->sched_class->set_cpus_allowed(p, new_mask);
4753 
4754 	cpumask_copy(&p->cpus_allowed, new_mask);
4755 	p->nr_cpus_allowed = cpumask_weight(new_mask);
4756 }
4757 
4758 /*
4759  * This is how migration works:
4760  *
4761  * 1) we invoke migration_cpu_stop() on the target CPU using
4762  *    stop_one_cpu().
4763  * 2) stopper starts to run (implicitly forcing the migrated thread
4764  *    off the CPU)
4765  * 3) it checks whether the migrated task is still in the wrong runqueue.
4766  * 4) if it's in the wrong runqueue then the migration thread removes
4767  *    it and puts it into the right queue.
4768  * 5) stopper completes and stop_one_cpu() returns and the migration
4769  *    is done.
4770  */
4771 
4772 /*
4773  * Change a given task's CPU affinity. Migrate the thread to a
4774  * proper CPU and schedule it away if the CPU it's executing on
4775  * is removed from the allowed bitmask.
4776  *
4777  * NOTE: the caller must have a valid reference to the task, the
4778  * task must not exit() & deallocate itself prematurely. The
4779  * call is not atomic; no spinlocks may be held.
4780  */
4781 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4782 {
4783 	unsigned long flags;
4784 	struct rq *rq;
4785 	unsigned int dest_cpu;
4786 	int ret = 0;
4787 
4788 	rq = task_rq_lock(p, &flags);
4789 
4790 	if (cpumask_equal(&p->cpus_allowed, new_mask))
4791 		goto out;
4792 
4793 	if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4794 		ret = -EINVAL;
4795 		goto out;
4796 	}
4797 
4798 	do_set_cpus_allowed(p, new_mask);
4799 
4800 	/* Can the task run on the task's current CPU? If so, we're done */
4801 	if (cpumask_test_cpu(task_cpu(p), new_mask))
4802 		goto out;
4803 
4804 	dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4805 	if (task_running(rq, p) || p->state == TASK_WAKING) {
4806 		struct migration_arg arg = { p, dest_cpu };
4807 		/* Need help from migration thread: drop lock and wait. */
4808 		task_rq_unlock(rq, p, &flags);
4809 		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4810 		tlb_migrate_finish(p->mm);
4811 		return 0;
4812 	} else if (task_on_rq_queued(p))
4813 		rq = move_queued_task(p, dest_cpu);
4814 out:
4815 	task_rq_unlock(rq, p, &flags);
4816 
4817 	return ret;
4818 }
4819 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4820 
4821 /*
4822  * Move (not current) task off this cpu, onto dest cpu. We're doing
4823  * this because either it can't run here any more (set_cpus_allowed()
4824  * away from this CPU, or CPU going down), or because we're
4825  * attempting to rebalance this task on exec (sched_exec).
4826  *
4827  * So we race with normal scheduler movements, but that's OK, as long
4828  * as the task is no longer on this CPU.
4829  *
4830  * Returns non-zero if task was successfully migrated.
4831  */
4832 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4833 {
4834 	struct rq *rq;
4835 	int ret = 0;
4836 
4837 	if (unlikely(!cpu_active(dest_cpu)))
4838 		return ret;
4839 
4840 	rq = cpu_rq(src_cpu);
4841 
4842 	raw_spin_lock(&p->pi_lock);
4843 	raw_spin_lock(&rq->lock);
4844 	/* Already moved. */
4845 	if (task_cpu(p) != src_cpu)
4846 		goto done;
4847 
4848 	/* Affinity changed (again). */
4849 	if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4850 		goto fail;
4851 
4852 	/*
4853 	 * If we're not on a rq, the next wake-up will ensure we're
4854 	 * placed properly.
4855 	 */
4856 	if (task_on_rq_queued(p))
4857 		rq = move_queued_task(p, dest_cpu);
4858 done:
4859 	ret = 1;
4860 fail:
4861 	raw_spin_unlock(&rq->lock);
4862 	raw_spin_unlock(&p->pi_lock);
4863 	return ret;
4864 }
4865 
4866 #ifdef CONFIG_NUMA_BALANCING
4867 /* Migrate current task p to target_cpu */
4868 int migrate_task_to(struct task_struct *p, int target_cpu)
4869 {
4870 	struct migration_arg arg = { p, target_cpu };
4871 	int curr_cpu = task_cpu(p);
4872 
4873 	if (curr_cpu == target_cpu)
4874 		return 0;
4875 
4876 	if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4877 		return -EINVAL;
4878 
4879 	/* TODO: This is not properly updating schedstats */
4880 
4881 	trace_sched_move_numa(p, curr_cpu, target_cpu);
4882 	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4883 }
4884 
4885 /*
4886  * Requeue a task on a given node and accurately track the number of NUMA
4887  * tasks on the runqueues
4888  */
4889 void sched_setnuma(struct task_struct *p, int nid)
4890 {
4891 	struct rq *rq;
4892 	unsigned long flags;
4893 	bool queued, running;
4894 
4895 	rq = task_rq_lock(p, &flags);
4896 	queued = task_on_rq_queued(p);
4897 	running = task_current(rq, p);
4898 
4899 	if (queued)
4900 		dequeue_task(rq, p, 0);
4901 	if (running)
4902 		put_prev_task(rq, p);
4903 
4904 	p->numa_preferred_nid = nid;
4905 
4906 	if (running)
4907 		p->sched_class->set_curr_task(rq);
4908 	if (queued)
4909 		enqueue_task(rq, p, 0);
4910 	task_rq_unlock(rq, p, &flags);
4911 }
4912 #endif
4913 
4914 /*
4915  * migration_cpu_stop - this will be executed by a highprio stopper thread
4916  * and performs thread migration by bumping thread off CPU then
4917  * 'pushing' onto another runqueue.
4918  */
4919 static int migration_cpu_stop(void *data)
4920 {
4921 	struct migration_arg *arg = data;
4922 
4923 	/*
4924 	 * The original target cpu might have gone down and we might
4925 	 * be on another cpu but it doesn't matter.
4926 	 */
4927 	local_irq_disable();
4928 	/*
4929 	 * We need to explicitly wake pending tasks before running
4930 	 * __migrate_task() such that we will not miss enforcing cpus_allowed
4931 	 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4932 	 */
4933 	sched_ttwu_pending();
4934 	__migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4935 	local_irq_enable();
4936 	return 0;
4937 }
4938 
4939 #ifdef CONFIG_HOTPLUG_CPU
4940 
4941 /*
4942  * Ensures that the idle task is using init_mm right before its cpu goes
4943  * offline.
4944  */
4945 void idle_task_exit(void)
4946 {
4947 	struct mm_struct *mm = current->active_mm;
4948 
4949 	BUG_ON(cpu_online(smp_processor_id()));
4950 
4951 	if (mm != &init_mm) {
4952 		switch_mm(mm, &init_mm, current);
4953 		finish_arch_post_lock_switch();
4954 	}
4955 	mmdrop(mm);
4956 }
4957 
4958 /*
4959  * Since this CPU is going 'away' for a while, fold any nr_active delta
4960  * we might have. Assumes we're called after migrate_tasks() so that the
4961  * nr_active count is stable.
4962  *
4963  * Also see the comment "Global load-average calculations".
4964  */
4965 static void calc_load_migrate(struct rq *rq)
4966 {
4967 	long delta = calc_load_fold_active(rq);
4968 	if (delta)
4969 		atomic_long_add(delta, &calc_load_tasks);
4970 }
4971 
4972 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4973 {
4974 }
4975 
4976 static const struct sched_class fake_sched_class = {
4977 	.put_prev_task = put_prev_task_fake,
4978 };
4979 
4980 static struct task_struct fake_task = {
4981 	/*
4982 	 * Avoid pull_{rt,dl}_task()
4983 	 */
4984 	.prio = MAX_PRIO + 1,
4985 	.sched_class = &fake_sched_class,
4986 };
4987 
4988 /*
4989  * Migrate all tasks from the rq, sleeping tasks will be migrated by
4990  * try_to_wake_up()->select_task_rq().
4991  *
4992  * Called with rq->lock held even though we'er in stop_machine() and
4993  * there's no concurrency possible, we hold the required locks anyway
4994  * because of lock validation efforts.
4995  */
4996 static void migrate_tasks(unsigned int dead_cpu)
4997 {
4998 	struct rq *rq = cpu_rq(dead_cpu);
4999 	struct task_struct *next, *stop = rq->stop;
5000 	int dest_cpu;
5001 
5002 	/*
5003 	 * Fudge the rq selection such that the below task selection loop
5004 	 * doesn't get stuck on the currently eligible stop task.
5005 	 *
5006 	 * We're currently inside stop_machine() and the rq is either stuck
5007 	 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5008 	 * either way we should never end up calling schedule() until we're
5009 	 * done here.
5010 	 */
5011 	rq->stop = NULL;
5012 
5013 	/*
5014 	 * put_prev_task() and pick_next_task() sched
5015 	 * class method both need to have an up-to-date
5016 	 * value of rq->clock[_task]
5017 	 */
5018 	update_rq_clock(rq);
5019 
5020 	for ( ; ; ) {
5021 		/*
5022 		 * There's this thread running, bail when that's the only
5023 		 * remaining thread.
5024 		 */
5025 		if (rq->nr_running == 1)
5026 			break;
5027 
5028 		next = pick_next_task(rq, &fake_task);
5029 		BUG_ON(!next);
5030 		next->sched_class->put_prev_task(rq, next);
5031 
5032 		/* Find suitable destination for @next, with force if needed. */
5033 		dest_cpu = select_fallback_rq(dead_cpu, next);
5034 		raw_spin_unlock(&rq->lock);
5035 
5036 		__migrate_task(next, dead_cpu, dest_cpu);
5037 
5038 		raw_spin_lock(&rq->lock);
5039 	}
5040 
5041 	rq->stop = stop;
5042 }
5043 
5044 #endif /* CONFIG_HOTPLUG_CPU */
5045 
5046 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5047 
5048 static struct ctl_table sd_ctl_dir[] = {
5049 	{
5050 		.procname	= "sched_domain",
5051 		.mode		= 0555,
5052 	},
5053 	{}
5054 };
5055 
5056 static struct ctl_table sd_ctl_root[] = {
5057 	{
5058 		.procname	= "kernel",
5059 		.mode		= 0555,
5060 		.child		= sd_ctl_dir,
5061 	},
5062 	{}
5063 };
5064 
5065 static struct ctl_table *sd_alloc_ctl_entry(int n)
5066 {
5067 	struct ctl_table *entry =
5068 		kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5069 
5070 	return entry;
5071 }
5072 
5073 static void sd_free_ctl_entry(struct ctl_table **tablep)
5074 {
5075 	struct ctl_table *entry;
5076 
5077 	/*
5078 	 * In the intermediate directories, both the child directory and
5079 	 * procname are dynamically allocated and could fail but the mode
5080 	 * will always be set. In the lowest directory the names are
5081 	 * static strings and all have proc handlers.
5082 	 */
5083 	for (entry = *tablep; entry->mode; entry++) {
5084 		if (entry->child)
5085 			sd_free_ctl_entry(&entry->child);
5086 		if (entry->proc_handler == NULL)
5087 			kfree(entry->procname);
5088 	}
5089 
5090 	kfree(*tablep);
5091 	*tablep = NULL;
5092 }
5093 
5094 static int min_load_idx = 0;
5095 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5096 
5097 static void
5098 set_table_entry(struct ctl_table *entry,
5099 		const char *procname, void *data, int maxlen,
5100 		umode_t mode, proc_handler *proc_handler,
5101 		bool load_idx)
5102 {
5103 	entry->procname = procname;
5104 	entry->data = data;
5105 	entry->maxlen = maxlen;
5106 	entry->mode = mode;
5107 	entry->proc_handler = proc_handler;
5108 
5109 	if (load_idx) {
5110 		entry->extra1 = &min_load_idx;
5111 		entry->extra2 = &max_load_idx;
5112 	}
5113 }
5114 
5115 static struct ctl_table *
5116 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5117 {
5118 	struct ctl_table *table = sd_alloc_ctl_entry(14);
5119 
5120 	if (table == NULL)
5121 		return NULL;
5122 
5123 	set_table_entry(&table[0], "min_interval", &sd->min_interval,
5124 		sizeof(long), 0644, proc_doulongvec_minmax, false);
5125 	set_table_entry(&table[1], "max_interval", &sd->max_interval,
5126 		sizeof(long), 0644, proc_doulongvec_minmax, false);
5127 	set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5128 		sizeof(int), 0644, proc_dointvec_minmax, true);
5129 	set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5130 		sizeof(int), 0644, proc_dointvec_minmax, true);
5131 	set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5132 		sizeof(int), 0644, proc_dointvec_minmax, true);
5133 	set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5134 		sizeof(int), 0644, proc_dointvec_minmax, true);
5135 	set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5136 		sizeof(int), 0644, proc_dointvec_minmax, true);
5137 	set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5138 		sizeof(int), 0644, proc_dointvec_minmax, false);
5139 	set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5140 		sizeof(int), 0644, proc_dointvec_minmax, false);
5141 	set_table_entry(&table[9], "cache_nice_tries",
5142 		&sd->cache_nice_tries,
5143 		sizeof(int), 0644, proc_dointvec_minmax, false);
5144 	set_table_entry(&table[10], "flags", &sd->flags,
5145 		sizeof(int), 0644, proc_dointvec_minmax, false);
5146 	set_table_entry(&table[11], "max_newidle_lb_cost",
5147 		&sd->max_newidle_lb_cost,
5148 		sizeof(long), 0644, proc_doulongvec_minmax, false);
5149 	set_table_entry(&table[12], "name", sd->name,
5150 		CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5151 	/* &table[13] is terminator */
5152 
5153 	return table;
5154 }
5155 
5156 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5157 {
5158 	struct ctl_table *entry, *table;
5159 	struct sched_domain *sd;
5160 	int domain_num = 0, i;
5161 	char buf[32];
5162 
5163 	for_each_domain(cpu, sd)
5164 		domain_num++;
5165 	entry = table = sd_alloc_ctl_entry(domain_num + 1);
5166 	if (table == NULL)
5167 		return NULL;
5168 
5169 	i = 0;
5170 	for_each_domain(cpu, sd) {
5171 		snprintf(buf, 32, "domain%d", i);
5172 		entry->procname = kstrdup(buf, GFP_KERNEL);
5173 		entry->mode = 0555;
5174 		entry->child = sd_alloc_ctl_domain_table(sd);
5175 		entry++;
5176 		i++;
5177 	}
5178 	return table;
5179 }
5180 
5181 static struct ctl_table_header *sd_sysctl_header;
5182 static void register_sched_domain_sysctl(void)
5183 {
5184 	int i, cpu_num = num_possible_cpus();
5185 	struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5186 	char buf[32];
5187 
5188 	WARN_ON(sd_ctl_dir[0].child);
5189 	sd_ctl_dir[0].child = entry;
5190 
5191 	if (entry == NULL)
5192 		return;
5193 
5194 	for_each_possible_cpu(i) {
5195 		snprintf(buf, 32, "cpu%d", i);
5196 		entry->procname = kstrdup(buf, GFP_KERNEL);
5197 		entry->mode = 0555;
5198 		entry->child = sd_alloc_ctl_cpu_table(i);
5199 		entry++;
5200 	}
5201 
5202 	WARN_ON(sd_sysctl_header);
5203 	sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5204 }
5205 
5206 /* may be called multiple times per register */
5207 static void unregister_sched_domain_sysctl(void)
5208 {
5209 	if (sd_sysctl_header)
5210 		unregister_sysctl_table(sd_sysctl_header);
5211 	sd_sysctl_header = NULL;
5212 	if (sd_ctl_dir[0].child)
5213 		sd_free_ctl_entry(&sd_ctl_dir[0].child);
5214 }
5215 #else
5216 static void register_sched_domain_sysctl(void)
5217 {
5218 }
5219 static void unregister_sched_domain_sysctl(void)
5220 {
5221 }
5222 #endif
5223 
5224 static void set_rq_online(struct rq *rq)
5225 {
5226 	if (!rq->online) {
5227 		const struct sched_class *class;
5228 
5229 		cpumask_set_cpu(rq->cpu, rq->rd->online);
5230 		rq->online = 1;
5231 
5232 		for_each_class(class) {
5233 			if (class->rq_online)
5234 				class->rq_online(rq);
5235 		}
5236 	}
5237 }
5238 
5239 static void set_rq_offline(struct rq *rq)
5240 {
5241 	if (rq->online) {
5242 		const struct sched_class *class;
5243 
5244 		for_each_class(class) {
5245 			if (class->rq_offline)
5246 				class->rq_offline(rq);
5247 		}
5248 
5249 		cpumask_clear_cpu(rq->cpu, rq->rd->online);
5250 		rq->online = 0;
5251 	}
5252 }
5253 
5254 /*
5255  * migration_call - callback that gets triggered when a CPU is added.
5256  * Here we can start up the necessary migration thread for the new CPU.
5257  */
5258 static int
5259 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5260 {
5261 	int cpu = (long)hcpu;
5262 	unsigned long flags;
5263 	struct rq *rq = cpu_rq(cpu);
5264 
5265 	switch (action & ~CPU_TASKS_FROZEN) {
5266 
5267 	case CPU_UP_PREPARE:
5268 		rq->calc_load_update = calc_load_update;
5269 		break;
5270 
5271 	case CPU_ONLINE:
5272 		/* Update our root-domain */
5273 		raw_spin_lock_irqsave(&rq->lock, flags);
5274 		if (rq->rd) {
5275 			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5276 
5277 			set_rq_online(rq);
5278 		}
5279 		raw_spin_unlock_irqrestore(&rq->lock, flags);
5280 		break;
5281 
5282 #ifdef CONFIG_HOTPLUG_CPU
5283 	case CPU_DYING:
5284 		sched_ttwu_pending();
5285 		/* Update our root-domain */
5286 		raw_spin_lock_irqsave(&rq->lock, flags);
5287 		if (rq->rd) {
5288 			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5289 			set_rq_offline(rq);
5290 		}
5291 		migrate_tasks(cpu);
5292 		BUG_ON(rq->nr_running != 1); /* the migration thread */
5293 		raw_spin_unlock_irqrestore(&rq->lock, flags);
5294 		break;
5295 
5296 	case CPU_DEAD:
5297 		calc_load_migrate(rq);
5298 		break;
5299 #endif
5300 	}
5301 
5302 	update_max_interval();
5303 
5304 	return NOTIFY_OK;
5305 }
5306 
5307 /*
5308  * Register at high priority so that task migration (migrate_all_tasks)
5309  * happens before everything else.  This has to be lower priority than
5310  * the notifier in the perf_event subsystem, though.
5311  */
5312 static struct notifier_block migration_notifier = {
5313 	.notifier_call = migration_call,
5314 	.priority = CPU_PRI_MIGRATION,
5315 };
5316 
5317 static void __cpuinit set_cpu_rq_start_time(void)
5318 {
5319 	int cpu = smp_processor_id();
5320 	struct rq *rq = cpu_rq(cpu);
5321 	rq->age_stamp = sched_clock_cpu(cpu);
5322 }
5323 
5324 static int sched_cpu_active(struct notifier_block *nfb,
5325 				      unsigned long action, void *hcpu)
5326 {
5327 	switch (action & ~CPU_TASKS_FROZEN) {
5328 	case CPU_STARTING:
5329 		set_cpu_rq_start_time();
5330 		return NOTIFY_OK;
5331 	case CPU_DOWN_FAILED:
5332 		set_cpu_active((long)hcpu, true);
5333 		return NOTIFY_OK;
5334 	default:
5335 		return NOTIFY_DONE;
5336 	}
5337 }
5338 
5339 static int sched_cpu_inactive(struct notifier_block *nfb,
5340 					unsigned long action, void *hcpu)
5341 {
5342 	switch (action & ~CPU_TASKS_FROZEN) {
5343 	case CPU_DOWN_PREPARE:
5344 		set_cpu_active((long)hcpu, false);
5345 		return NOTIFY_OK;
5346 	default:
5347 		return NOTIFY_DONE;
5348 	}
5349 }
5350 
5351 static int __init migration_init(void)
5352 {
5353 	void *cpu = (void *)(long)smp_processor_id();
5354 	int err;
5355 
5356 	/* Initialize migration for the boot CPU */
5357 	err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5358 	BUG_ON(err == NOTIFY_BAD);
5359 	migration_call(&migration_notifier, CPU_ONLINE, cpu);
5360 	register_cpu_notifier(&migration_notifier);
5361 
5362 	/* Register cpu active notifiers */
5363 	cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5364 	cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5365 
5366 	return 0;
5367 }
5368 early_initcall(migration_init);
5369 #endif
5370 
5371 #ifdef CONFIG_SMP
5372 
5373 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5374 
5375 #ifdef CONFIG_SCHED_DEBUG
5376 
5377 static __read_mostly int sched_debug_enabled;
5378 
5379 static int __init sched_debug_setup(char *str)
5380 {
5381 	sched_debug_enabled = 1;
5382 
5383 	return 0;
5384 }
5385 early_param("sched_debug", sched_debug_setup);
5386 
5387 static inline bool sched_debug(void)
5388 {
5389 	return sched_debug_enabled;
5390 }
5391 
5392 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5393 				  struct cpumask *groupmask)
5394 {
5395 	struct sched_group *group = sd->groups;
5396 
5397 	cpumask_clear(groupmask);
5398 
5399 	printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5400 
5401 	if (!(sd->flags & SD_LOAD_BALANCE)) {
5402 		printk("does not load-balance\n");
5403 		if (sd->parent)
5404 			printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5405 					" has parent");
5406 		return -1;
5407 	}
5408 
5409 	printk(KERN_CONT "span %*pbl level %s\n",
5410 	       cpumask_pr_args(sched_domain_span(sd)), sd->name);
5411 
5412 	if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5413 		printk(KERN_ERR "ERROR: domain->span does not contain "
5414 				"CPU%d\n", cpu);
5415 	}
5416 	if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5417 		printk(KERN_ERR "ERROR: domain->groups does not contain"
5418 				" CPU%d\n", cpu);
5419 	}
5420 
5421 	printk(KERN_DEBUG "%*s groups:", level + 1, "");
5422 	do {
5423 		if (!group) {
5424 			printk("\n");
5425 			printk(KERN_ERR "ERROR: group is NULL\n");
5426 			break;
5427 		}
5428 
5429 		if (!cpumask_weight(sched_group_cpus(group))) {
5430 			printk(KERN_CONT "\n");
5431 			printk(KERN_ERR "ERROR: empty group\n");
5432 			break;
5433 		}
5434 
5435 		if (!(sd->flags & SD_OVERLAP) &&
5436 		    cpumask_intersects(groupmask, sched_group_cpus(group))) {
5437 			printk(KERN_CONT "\n");
5438 			printk(KERN_ERR "ERROR: repeated CPUs\n");
5439 			break;
5440 		}
5441 
5442 		cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5443 
5444 		printk(KERN_CONT " %*pbl",
5445 		       cpumask_pr_args(sched_group_cpus(group)));
5446 		if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5447 			printk(KERN_CONT " (cpu_capacity = %d)",
5448 				group->sgc->capacity);
5449 		}
5450 
5451 		group = group->next;
5452 	} while (group != sd->groups);
5453 	printk(KERN_CONT "\n");
5454 
5455 	if (!cpumask_equal(sched_domain_span(sd), groupmask))
5456 		printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5457 
5458 	if (sd->parent &&
5459 	    !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5460 		printk(KERN_ERR "ERROR: parent span is not a superset "
5461 			"of domain->span\n");
5462 	return 0;
5463 }
5464 
5465 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5466 {
5467 	int level = 0;
5468 
5469 	if (!sched_debug_enabled)
5470 		return;
5471 
5472 	if (!sd) {
5473 		printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5474 		return;
5475 	}
5476 
5477 	printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5478 
5479 	for (;;) {
5480 		if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5481 			break;
5482 		level++;
5483 		sd = sd->parent;
5484 		if (!sd)
5485 			break;
5486 	}
5487 }
5488 #else /* !CONFIG_SCHED_DEBUG */
5489 # define sched_domain_debug(sd, cpu) do { } while (0)
5490 static inline bool sched_debug(void)
5491 {
5492 	return false;
5493 }
5494 #endif /* CONFIG_SCHED_DEBUG */
5495 
5496 static int sd_degenerate(struct sched_domain *sd)
5497 {
5498 	if (cpumask_weight(sched_domain_span(sd)) == 1)
5499 		return 1;
5500 
5501 	/* Following flags need at least 2 groups */
5502 	if (sd->flags & (SD_LOAD_BALANCE |
5503 			 SD_BALANCE_NEWIDLE |
5504 			 SD_BALANCE_FORK |
5505 			 SD_BALANCE_EXEC |
5506 			 SD_SHARE_CPUCAPACITY |
5507 			 SD_SHARE_PKG_RESOURCES |
5508 			 SD_SHARE_POWERDOMAIN)) {
5509 		if (sd->groups != sd->groups->next)
5510 			return 0;
5511 	}
5512 
5513 	/* Following flags don't use groups */
5514 	if (sd->flags & (SD_WAKE_AFFINE))
5515 		return 0;
5516 
5517 	return 1;
5518 }
5519 
5520 static int
5521 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5522 {
5523 	unsigned long cflags = sd->flags, pflags = parent->flags;
5524 
5525 	if (sd_degenerate(parent))
5526 		return 1;
5527 
5528 	if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5529 		return 0;
5530 
5531 	/* Flags needing groups don't count if only 1 group in parent */
5532 	if (parent->groups == parent->groups->next) {
5533 		pflags &= ~(SD_LOAD_BALANCE |
5534 				SD_BALANCE_NEWIDLE |
5535 				SD_BALANCE_FORK |
5536 				SD_BALANCE_EXEC |
5537 				SD_SHARE_CPUCAPACITY |
5538 				SD_SHARE_PKG_RESOURCES |
5539 				SD_PREFER_SIBLING |
5540 				SD_SHARE_POWERDOMAIN);
5541 		if (nr_node_ids == 1)
5542 			pflags &= ~SD_SERIALIZE;
5543 	}
5544 	if (~cflags & pflags)
5545 		return 0;
5546 
5547 	return 1;
5548 }
5549 
5550 static void free_rootdomain(struct rcu_head *rcu)
5551 {
5552 	struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5553 
5554 	cpupri_cleanup(&rd->cpupri);
5555 	cpudl_cleanup(&rd->cpudl);
5556 	free_cpumask_var(rd->dlo_mask);
5557 	free_cpumask_var(rd->rto_mask);
5558 	free_cpumask_var(rd->online);
5559 	free_cpumask_var(rd->span);
5560 	kfree(rd);
5561 }
5562 
5563 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5564 {
5565 	struct root_domain *old_rd = NULL;
5566 	unsigned long flags;
5567 
5568 	raw_spin_lock_irqsave(&rq->lock, flags);
5569 
5570 	if (rq->rd) {
5571 		old_rd = rq->rd;
5572 
5573 		if (cpumask_test_cpu(rq->cpu, old_rd->online))
5574 			set_rq_offline(rq);
5575 
5576 		cpumask_clear_cpu(rq->cpu, old_rd->span);
5577 
5578 		/*
5579 		 * If we dont want to free the old_rd yet then
5580 		 * set old_rd to NULL to skip the freeing later
5581 		 * in this function:
5582 		 */
5583 		if (!atomic_dec_and_test(&old_rd->refcount))
5584 			old_rd = NULL;
5585 	}
5586 
5587 	atomic_inc(&rd->refcount);
5588 	rq->rd = rd;
5589 
5590 	cpumask_set_cpu(rq->cpu, rd->span);
5591 	if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5592 		set_rq_online(rq);
5593 
5594 	raw_spin_unlock_irqrestore(&rq->lock, flags);
5595 
5596 	if (old_rd)
5597 		call_rcu_sched(&old_rd->rcu, free_rootdomain);
5598 }
5599 
5600 static int init_rootdomain(struct root_domain *rd)
5601 {
5602 	memset(rd, 0, sizeof(*rd));
5603 
5604 	if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5605 		goto out;
5606 	if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5607 		goto free_span;
5608 	if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5609 		goto free_online;
5610 	if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5611 		goto free_dlo_mask;
5612 
5613 	init_dl_bw(&rd->dl_bw);
5614 	if (cpudl_init(&rd->cpudl) != 0)
5615 		goto free_dlo_mask;
5616 
5617 	if (cpupri_init(&rd->cpupri) != 0)
5618 		goto free_rto_mask;
5619 	return 0;
5620 
5621 free_rto_mask:
5622 	free_cpumask_var(rd->rto_mask);
5623 free_dlo_mask:
5624 	free_cpumask_var(rd->dlo_mask);
5625 free_online:
5626 	free_cpumask_var(rd->online);
5627 free_span:
5628 	free_cpumask_var(rd->span);
5629 out:
5630 	return -ENOMEM;
5631 }
5632 
5633 /*
5634  * By default the system creates a single root-domain with all cpus as
5635  * members (mimicking the global state we have today).
5636  */
5637 struct root_domain def_root_domain;
5638 
5639 static void init_defrootdomain(void)
5640 {
5641 	init_rootdomain(&def_root_domain);
5642 
5643 	atomic_set(&def_root_domain.refcount, 1);
5644 }
5645 
5646 static struct root_domain *alloc_rootdomain(void)
5647 {
5648 	struct root_domain *rd;
5649 
5650 	rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5651 	if (!rd)
5652 		return NULL;
5653 
5654 	if (init_rootdomain(rd) != 0) {
5655 		kfree(rd);
5656 		return NULL;
5657 	}
5658 
5659 	return rd;
5660 }
5661 
5662 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5663 {
5664 	struct sched_group *tmp, *first;
5665 
5666 	if (!sg)
5667 		return;
5668 
5669 	first = sg;
5670 	do {
5671 		tmp = sg->next;
5672 
5673 		if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5674 			kfree(sg->sgc);
5675 
5676 		kfree(sg);
5677 		sg = tmp;
5678 	} while (sg != first);
5679 }
5680 
5681 static void free_sched_domain(struct rcu_head *rcu)
5682 {
5683 	struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5684 
5685 	/*
5686 	 * If its an overlapping domain it has private groups, iterate and
5687 	 * nuke them all.
5688 	 */
5689 	if (sd->flags & SD_OVERLAP) {
5690 		free_sched_groups(sd->groups, 1);
5691 	} else if (atomic_dec_and_test(&sd->groups->ref)) {
5692 		kfree(sd->groups->sgc);
5693 		kfree(sd->groups);
5694 	}
5695 	kfree(sd);
5696 }
5697 
5698 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5699 {
5700 	call_rcu(&sd->rcu, free_sched_domain);
5701 }
5702 
5703 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5704 {
5705 	for (; sd; sd = sd->parent)
5706 		destroy_sched_domain(sd, cpu);
5707 }
5708 
5709 /*
5710  * Keep a special pointer to the highest sched_domain that has
5711  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5712  * allows us to avoid some pointer chasing select_idle_sibling().
5713  *
5714  * Also keep a unique ID per domain (we use the first cpu number in
5715  * the cpumask of the domain), this allows us to quickly tell if
5716  * two cpus are in the same cache domain, see cpus_share_cache().
5717  */
5718 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5719 DEFINE_PER_CPU(int, sd_llc_size);
5720 DEFINE_PER_CPU(int, sd_llc_id);
5721 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5722 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5723 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5724 
5725 static void update_top_cache_domain(int cpu)
5726 {
5727 	struct sched_domain *sd;
5728 	struct sched_domain *busy_sd = NULL;
5729 	int id = cpu;
5730 	int size = 1;
5731 
5732 	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5733 	if (sd) {
5734 		id = cpumask_first(sched_domain_span(sd));
5735 		size = cpumask_weight(sched_domain_span(sd));
5736 		busy_sd = sd->parent; /* sd_busy */
5737 	}
5738 	rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5739 
5740 	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5741 	per_cpu(sd_llc_size, cpu) = size;
5742 	per_cpu(sd_llc_id, cpu) = id;
5743 
5744 	sd = lowest_flag_domain(cpu, SD_NUMA);
5745 	rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5746 
5747 	sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5748 	rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5749 }
5750 
5751 /*
5752  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5753  * hold the hotplug lock.
5754  */
5755 static void
5756 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5757 {
5758 	struct rq *rq = cpu_rq(cpu);
5759 	struct sched_domain *tmp;
5760 
5761 	/* Remove the sched domains which do not contribute to scheduling. */
5762 	for (tmp = sd; tmp; ) {
5763 		struct sched_domain *parent = tmp->parent;
5764 		if (!parent)
5765 			break;
5766 
5767 		if (sd_parent_degenerate(tmp, parent)) {
5768 			tmp->parent = parent->parent;
5769 			if (parent->parent)
5770 				parent->parent->child = tmp;
5771 			/*
5772 			 * Transfer SD_PREFER_SIBLING down in case of a
5773 			 * degenerate parent; the spans match for this
5774 			 * so the property transfers.
5775 			 */
5776 			if (parent->flags & SD_PREFER_SIBLING)
5777 				tmp->flags |= SD_PREFER_SIBLING;
5778 			destroy_sched_domain(parent, cpu);
5779 		} else
5780 			tmp = tmp->parent;
5781 	}
5782 
5783 	if (sd && sd_degenerate(sd)) {
5784 		tmp = sd;
5785 		sd = sd->parent;
5786 		destroy_sched_domain(tmp, cpu);
5787 		if (sd)
5788 			sd->child = NULL;
5789 	}
5790 
5791 	sched_domain_debug(sd, cpu);
5792 
5793 	rq_attach_root(rq, rd);
5794 	tmp = rq->sd;
5795 	rcu_assign_pointer(rq->sd, sd);
5796 	destroy_sched_domains(tmp, cpu);
5797 
5798 	update_top_cache_domain(cpu);
5799 }
5800 
5801 /* Setup the mask of cpus configured for isolated domains */
5802 static int __init isolated_cpu_setup(char *str)
5803 {
5804 	alloc_bootmem_cpumask_var(&cpu_isolated_map);
5805 	cpulist_parse(str, cpu_isolated_map);
5806 	return 1;
5807 }
5808 
5809 __setup("isolcpus=", isolated_cpu_setup);
5810 
5811 struct s_data {
5812 	struct sched_domain ** __percpu sd;
5813 	struct root_domain	*rd;
5814 };
5815 
5816 enum s_alloc {
5817 	sa_rootdomain,
5818 	sa_sd,
5819 	sa_sd_storage,
5820 	sa_none,
5821 };
5822 
5823 /*
5824  * Build an iteration mask that can exclude certain CPUs from the upwards
5825  * domain traversal.
5826  *
5827  * Asymmetric node setups can result in situations where the domain tree is of
5828  * unequal depth, make sure to skip domains that already cover the entire
5829  * range.
5830  *
5831  * In that case build_sched_domains() will have terminated the iteration early
5832  * and our sibling sd spans will be empty. Domains should always include the
5833  * cpu they're built on, so check that.
5834  *
5835  */
5836 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5837 {
5838 	const struct cpumask *span = sched_domain_span(sd);
5839 	struct sd_data *sdd = sd->private;
5840 	struct sched_domain *sibling;
5841 	int i;
5842 
5843 	for_each_cpu(i, span) {
5844 		sibling = *per_cpu_ptr(sdd->sd, i);
5845 		if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5846 			continue;
5847 
5848 		cpumask_set_cpu(i, sched_group_mask(sg));
5849 	}
5850 }
5851 
5852 /*
5853  * Return the canonical balance cpu for this group, this is the first cpu
5854  * of this group that's also in the iteration mask.
5855  */
5856 int group_balance_cpu(struct sched_group *sg)
5857 {
5858 	return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5859 }
5860 
5861 static int
5862 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5863 {
5864 	struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5865 	const struct cpumask *span = sched_domain_span(sd);
5866 	struct cpumask *covered = sched_domains_tmpmask;
5867 	struct sd_data *sdd = sd->private;
5868 	struct sched_domain *sibling;
5869 	int i;
5870 
5871 	cpumask_clear(covered);
5872 
5873 	for_each_cpu(i, span) {
5874 		struct cpumask *sg_span;
5875 
5876 		if (cpumask_test_cpu(i, covered))
5877 			continue;
5878 
5879 		sibling = *per_cpu_ptr(sdd->sd, i);
5880 
5881 		/* See the comment near build_group_mask(). */
5882 		if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5883 			continue;
5884 
5885 		sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5886 				GFP_KERNEL, cpu_to_node(cpu));
5887 
5888 		if (!sg)
5889 			goto fail;
5890 
5891 		sg_span = sched_group_cpus(sg);
5892 		if (sibling->child)
5893 			cpumask_copy(sg_span, sched_domain_span(sibling->child));
5894 		else
5895 			cpumask_set_cpu(i, sg_span);
5896 
5897 		cpumask_or(covered, covered, sg_span);
5898 
5899 		sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5900 		if (atomic_inc_return(&sg->sgc->ref) == 1)
5901 			build_group_mask(sd, sg);
5902 
5903 		/*
5904 		 * Initialize sgc->capacity such that even if we mess up the
5905 		 * domains and no possible iteration will get us here, we won't
5906 		 * die on a /0 trap.
5907 		 */
5908 		sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5909 
5910 		/*
5911 		 * Make sure the first group of this domain contains the
5912 		 * canonical balance cpu. Otherwise the sched_domain iteration
5913 		 * breaks. See update_sg_lb_stats().
5914 		 */
5915 		if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5916 		    group_balance_cpu(sg) == cpu)
5917 			groups = sg;
5918 
5919 		if (!first)
5920 			first = sg;
5921 		if (last)
5922 			last->next = sg;
5923 		last = sg;
5924 		last->next = first;
5925 	}
5926 	sd->groups = groups;
5927 
5928 	return 0;
5929 
5930 fail:
5931 	free_sched_groups(first, 0);
5932 
5933 	return -ENOMEM;
5934 }
5935 
5936 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5937 {
5938 	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5939 	struct sched_domain *child = sd->child;
5940 
5941 	if (child)
5942 		cpu = cpumask_first(sched_domain_span(child));
5943 
5944 	if (sg) {
5945 		*sg = *per_cpu_ptr(sdd->sg, cpu);
5946 		(*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5947 		atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5948 	}
5949 
5950 	return cpu;
5951 }
5952 
5953 /*
5954  * build_sched_groups will build a circular linked list of the groups
5955  * covered by the given span, and will set each group's ->cpumask correctly,
5956  * and ->cpu_capacity to 0.
5957  *
5958  * Assumes the sched_domain tree is fully constructed
5959  */
5960 static int
5961 build_sched_groups(struct sched_domain *sd, int cpu)
5962 {
5963 	struct sched_group *first = NULL, *last = NULL;
5964 	struct sd_data *sdd = sd->private;
5965 	const struct cpumask *span = sched_domain_span(sd);
5966 	struct cpumask *covered;
5967 	int i;
5968 
5969 	get_group(cpu, sdd, &sd->groups);
5970 	atomic_inc(&sd->groups->ref);
5971 
5972 	if (cpu != cpumask_first(span))
5973 		return 0;
5974 
5975 	lockdep_assert_held(&sched_domains_mutex);
5976 	covered = sched_domains_tmpmask;
5977 
5978 	cpumask_clear(covered);
5979 
5980 	for_each_cpu(i, span) {
5981 		struct sched_group *sg;
5982 		int group, j;
5983 
5984 		if (cpumask_test_cpu(i, covered))
5985 			continue;
5986 
5987 		group = get_group(i, sdd, &sg);
5988 		cpumask_setall(sched_group_mask(sg));
5989 
5990 		for_each_cpu(j, span) {
5991 			if (get_group(j, sdd, NULL) != group)
5992 				continue;
5993 
5994 			cpumask_set_cpu(j, covered);
5995 			cpumask_set_cpu(j, sched_group_cpus(sg));
5996 		}
5997 
5998 		if (!first)
5999 			first = sg;
6000 		if (last)
6001 			last->next = sg;
6002 		last = sg;
6003 	}
6004 	last->next = first;
6005 
6006 	return 0;
6007 }
6008 
6009 /*
6010  * Initialize sched groups cpu_capacity.
6011  *
6012  * cpu_capacity indicates the capacity of sched group, which is used while
6013  * distributing the load between different sched groups in a sched domain.
6014  * Typically cpu_capacity for all the groups in a sched domain will be same
6015  * unless there are asymmetries in the topology. If there are asymmetries,
6016  * group having more cpu_capacity will pickup more load compared to the
6017  * group having less cpu_capacity.
6018  */
6019 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6020 {
6021 	struct sched_group *sg = sd->groups;
6022 
6023 	WARN_ON(!sg);
6024 
6025 	do {
6026 		sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6027 		sg = sg->next;
6028 	} while (sg != sd->groups);
6029 
6030 	if (cpu != group_balance_cpu(sg))
6031 		return;
6032 
6033 	update_group_capacity(sd, cpu);
6034 	atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6035 }
6036 
6037 /*
6038  * Initializers for schedule domains
6039  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6040  */
6041 
6042 static int default_relax_domain_level = -1;
6043 int sched_domain_level_max;
6044 
6045 static int __init setup_relax_domain_level(char *str)
6046 {
6047 	if (kstrtoint(str, 0, &default_relax_domain_level))
6048 		pr_warn("Unable to set relax_domain_level\n");
6049 
6050 	return 1;
6051 }
6052 __setup("relax_domain_level=", setup_relax_domain_level);
6053 
6054 static void set_domain_attribute(struct sched_domain *sd,
6055 				 struct sched_domain_attr *attr)
6056 {
6057 	int request;
6058 
6059 	if (!attr || attr->relax_domain_level < 0) {
6060 		if (default_relax_domain_level < 0)
6061 			return;
6062 		else
6063 			request = default_relax_domain_level;
6064 	} else
6065 		request = attr->relax_domain_level;
6066 	if (request < sd->level) {
6067 		/* turn off idle balance on this domain */
6068 		sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6069 	} else {
6070 		/* turn on idle balance on this domain */
6071 		sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6072 	}
6073 }
6074 
6075 static void __sdt_free(const struct cpumask *cpu_map);
6076 static int __sdt_alloc(const struct cpumask *cpu_map);
6077 
6078 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6079 				 const struct cpumask *cpu_map)
6080 {
6081 	switch (what) {
6082 	case sa_rootdomain:
6083 		if (!atomic_read(&d->rd->refcount))
6084 			free_rootdomain(&d->rd->rcu); /* fall through */
6085 	case sa_sd:
6086 		free_percpu(d->sd); /* fall through */
6087 	case sa_sd_storage:
6088 		__sdt_free(cpu_map); /* fall through */
6089 	case sa_none:
6090 		break;
6091 	}
6092 }
6093 
6094 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6095 						   const struct cpumask *cpu_map)
6096 {
6097 	memset(d, 0, sizeof(*d));
6098 
6099 	if (__sdt_alloc(cpu_map))
6100 		return sa_sd_storage;
6101 	d->sd = alloc_percpu(struct sched_domain *);
6102 	if (!d->sd)
6103 		return sa_sd_storage;
6104 	d->rd = alloc_rootdomain();
6105 	if (!d->rd)
6106 		return sa_sd;
6107 	return sa_rootdomain;
6108 }
6109 
6110 /*
6111  * NULL the sd_data elements we've used to build the sched_domain and
6112  * sched_group structure so that the subsequent __free_domain_allocs()
6113  * will not free the data we're using.
6114  */
6115 static void claim_allocations(int cpu, struct sched_domain *sd)
6116 {
6117 	struct sd_data *sdd = sd->private;
6118 
6119 	WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6120 	*per_cpu_ptr(sdd->sd, cpu) = NULL;
6121 
6122 	if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6123 		*per_cpu_ptr(sdd->sg, cpu) = NULL;
6124 
6125 	if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6126 		*per_cpu_ptr(sdd->sgc, cpu) = NULL;
6127 }
6128 
6129 #ifdef CONFIG_NUMA
6130 static int sched_domains_numa_levels;
6131 enum numa_topology_type sched_numa_topology_type;
6132 static int *sched_domains_numa_distance;
6133 int sched_max_numa_distance;
6134 static struct cpumask ***sched_domains_numa_masks;
6135 static int sched_domains_curr_level;
6136 #endif
6137 
6138 /*
6139  * SD_flags allowed in topology descriptions.
6140  *
6141  * SD_SHARE_CPUCAPACITY      - describes SMT topologies
6142  * SD_SHARE_PKG_RESOURCES - describes shared caches
6143  * SD_NUMA                - describes NUMA topologies
6144  * SD_SHARE_POWERDOMAIN   - describes shared power domain
6145  *
6146  * Odd one out:
6147  * SD_ASYM_PACKING        - describes SMT quirks
6148  */
6149 #define TOPOLOGY_SD_FLAGS		\
6150 	(SD_SHARE_CPUCAPACITY |		\
6151 	 SD_SHARE_PKG_RESOURCES |	\
6152 	 SD_NUMA |			\
6153 	 SD_ASYM_PACKING |		\
6154 	 SD_SHARE_POWERDOMAIN)
6155 
6156 static struct sched_domain *
6157 sd_init(struct sched_domain_topology_level *tl, int cpu)
6158 {
6159 	struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6160 	int sd_weight, sd_flags = 0;
6161 
6162 #ifdef CONFIG_NUMA
6163 	/*
6164 	 * Ugly hack to pass state to sd_numa_mask()...
6165 	 */
6166 	sched_domains_curr_level = tl->numa_level;
6167 #endif
6168 
6169 	sd_weight = cpumask_weight(tl->mask(cpu));
6170 
6171 	if (tl->sd_flags)
6172 		sd_flags = (*tl->sd_flags)();
6173 	if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6174 			"wrong sd_flags in topology description\n"))
6175 		sd_flags &= ~TOPOLOGY_SD_FLAGS;
6176 
6177 	*sd = (struct sched_domain){
6178 		.min_interval		= sd_weight,
6179 		.max_interval		= 2*sd_weight,
6180 		.busy_factor		= 32,
6181 		.imbalance_pct		= 125,
6182 
6183 		.cache_nice_tries	= 0,
6184 		.busy_idx		= 0,
6185 		.idle_idx		= 0,
6186 		.newidle_idx		= 0,
6187 		.wake_idx		= 0,
6188 		.forkexec_idx		= 0,
6189 
6190 		.flags			= 1*SD_LOAD_BALANCE
6191 					| 1*SD_BALANCE_NEWIDLE
6192 					| 1*SD_BALANCE_EXEC
6193 					| 1*SD_BALANCE_FORK
6194 					| 0*SD_BALANCE_WAKE
6195 					| 1*SD_WAKE_AFFINE
6196 					| 0*SD_SHARE_CPUCAPACITY
6197 					| 0*SD_SHARE_PKG_RESOURCES
6198 					| 0*SD_SERIALIZE
6199 					| 0*SD_PREFER_SIBLING
6200 					| 0*SD_NUMA
6201 					| sd_flags
6202 					,
6203 
6204 		.last_balance		= jiffies,
6205 		.balance_interval	= sd_weight,
6206 		.smt_gain		= 0,
6207 		.max_newidle_lb_cost	= 0,
6208 		.next_decay_max_lb_cost	= jiffies,
6209 #ifdef CONFIG_SCHED_DEBUG
6210 		.name			= tl->name,
6211 #endif
6212 	};
6213 
6214 	/*
6215 	 * Convert topological properties into behaviour.
6216 	 */
6217 
6218 	if (sd->flags & SD_SHARE_CPUCAPACITY) {
6219 		sd->flags |= SD_PREFER_SIBLING;
6220 		sd->imbalance_pct = 110;
6221 		sd->smt_gain = 1178; /* ~15% */
6222 
6223 	} else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6224 		sd->imbalance_pct = 117;
6225 		sd->cache_nice_tries = 1;
6226 		sd->busy_idx = 2;
6227 
6228 #ifdef CONFIG_NUMA
6229 	} else if (sd->flags & SD_NUMA) {
6230 		sd->cache_nice_tries = 2;
6231 		sd->busy_idx = 3;
6232 		sd->idle_idx = 2;
6233 
6234 		sd->flags |= SD_SERIALIZE;
6235 		if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6236 			sd->flags &= ~(SD_BALANCE_EXEC |
6237 				       SD_BALANCE_FORK |
6238 				       SD_WAKE_AFFINE);
6239 		}
6240 
6241 #endif
6242 	} else {
6243 		sd->flags |= SD_PREFER_SIBLING;
6244 		sd->cache_nice_tries = 1;
6245 		sd->busy_idx = 2;
6246 		sd->idle_idx = 1;
6247 	}
6248 
6249 	sd->private = &tl->data;
6250 
6251 	return sd;
6252 }
6253 
6254 /*
6255  * Topology list, bottom-up.
6256  */
6257 static struct sched_domain_topology_level default_topology[] = {
6258 #ifdef CONFIG_SCHED_SMT
6259 	{ cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6260 #endif
6261 #ifdef CONFIG_SCHED_MC
6262 	{ cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6263 #endif
6264 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
6265 	{ NULL, },
6266 };
6267 
6268 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6269 
6270 #define for_each_sd_topology(tl)			\
6271 	for (tl = sched_domain_topology; tl->mask; tl++)
6272 
6273 void set_sched_topology(struct sched_domain_topology_level *tl)
6274 {
6275 	sched_domain_topology = tl;
6276 }
6277 
6278 #ifdef CONFIG_NUMA
6279 
6280 static const struct cpumask *sd_numa_mask(int cpu)
6281 {
6282 	return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6283 }
6284 
6285 static void sched_numa_warn(const char *str)
6286 {
6287 	static int done = false;
6288 	int i,j;
6289 
6290 	if (done)
6291 		return;
6292 
6293 	done = true;
6294 
6295 	printk(KERN_WARNING "ERROR: %s\n\n", str);
6296 
6297 	for (i = 0; i < nr_node_ids; i++) {
6298 		printk(KERN_WARNING "  ");
6299 		for (j = 0; j < nr_node_ids; j++)
6300 			printk(KERN_CONT "%02d ", node_distance(i,j));
6301 		printk(KERN_CONT "\n");
6302 	}
6303 	printk(KERN_WARNING "\n");
6304 }
6305 
6306 bool find_numa_distance(int distance)
6307 {
6308 	int i;
6309 
6310 	if (distance == node_distance(0, 0))
6311 		return true;
6312 
6313 	for (i = 0; i < sched_domains_numa_levels; i++) {
6314 		if (sched_domains_numa_distance[i] == distance)
6315 			return true;
6316 	}
6317 
6318 	return false;
6319 }
6320 
6321 /*
6322  * A system can have three types of NUMA topology:
6323  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6324  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6325  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6326  *
6327  * The difference between a glueless mesh topology and a backplane
6328  * topology lies in whether communication between not directly
6329  * connected nodes goes through intermediary nodes (where programs
6330  * could run), or through backplane controllers. This affects
6331  * placement of programs.
6332  *
6333  * The type of topology can be discerned with the following tests:
6334  * - If the maximum distance between any nodes is 1 hop, the system
6335  *   is directly connected.
6336  * - If for two nodes A and B, located N > 1 hops away from each other,
6337  *   there is an intermediary node C, which is < N hops away from both
6338  *   nodes A and B, the system is a glueless mesh.
6339  */
6340 static void init_numa_topology_type(void)
6341 {
6342 	int a, b, c, n;
6343 
6344 	n = sched_max_numa_distance;
6345 
6346 	if (n <= 1)
6347 		sched_numa_topology_type = NUMA_DIRECT;
6348 
6349 	for_each_online_node(a) {
6350 		for_each_online_node(b) {
6351 			/* Find two nodes furthest removed from each other. */
6352 			if (node_distance(a, b) < n)
6353 				continue;
6354 
6355 			/* Is there an intermediary node between a and b? */
6356 			for_each_online_node(c) {
6357 				if (node_distance(a, c) < n &&
6358 				    node_distance(b, c) < n) {
6359 					sched_numa_topology_type =
6360 							NUMA_GLUELESS_MESH;
6361 					return;
6362 				}
6363 			}
6364 
6365 			sched_numa_topology_type = NUMA_BACKPLANE;
6366 			return;
6367 		}
6368 	}
6369 }
6370 
6371 static void sched_init_numa(void)
6372 {
6373 	int next_distance, curr_distance = node_distance(0, 0);
6374 	struct sched_domain_topology_level *tl;
6375 	int level = 0;
6376 	int i, j, k;
6377 
6378 	sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6379 	if (!sched_domains_numa_distance)
6380 		return;
6381 
6382 	/*
6383 	 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6384 	 * unique distances in the node_distance() table.
6385 	 *
6386 	 * Assumes node_distance(0,j) includes all distances in
6387 	 * node_distance(i,j) in order to avoid cubic time.
6388 	 */
6389 	next_distance = curr_distance;
6390 	for (i = 0; i < nr_node_ids; i++) {
6391 		for (j = 0; j < nr_node_ids; j++) {
6392 			for (k = 0; k < nr_node_ids; k++) {
6393 				int distance = node_distance(i, k);
6394 
6395 				if (distance > curr_distance &&
6396 				    (distance < next_distance ||
6397 				     next_distance == curr_distance))
6398 					next_distance = distance;
6399 
6400 				/*
6401 				 * While not a strong assumption it would be nice to know
6402 				 * about cases where if node A is connected to B, B is not
6403 				 * equally connected to A.
6404 				 */
6405 				if (sched_debug() && node_distance(k, i) != distance)
6406 					sched_numa_warn("Node-distance not symmetric");
6407 
6408 				if (sched_debug() && i && !find_numa_distance(distance))
6409 					sched_numa_warn("Node-0 not representative");
6410 			}
6411 			if (next_distance != curr_distance) {
6412 				sched_domains_numa_distance[level++] = next_distance;
6413 				sched_domains_numa_levels = level;
6414 				curr_distance = next_distance;
6415 			} else break;
6416 		}
6417 
6418 		/*
6419 		 * In case of sched_debug() we verify the above assumption.
6420 		 */
6421 		if (!sched_debug())
6422 			break;
6423 	}
6424 
6425 	if (!level)
6426 		return;
6427 
6428 	/*
6429 	 * 'level' contains the number of unique distances, excluding the
6430 	 * identity distance node_distance(i,i).
6431 	 *
6432 	 * The sched_domains_numa_distance[] array includes the actual distance
6433 	 * numbers.
6434 	 */
6435 
6436 	/*
6437 	 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6438 	 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6439 	 * the array will contain less then 'level' members. This could be
6440 	 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6441 	 * in other functions.
6442 	 *
6443 	 * We reset it to 'level' at the end of this function.
6444 	 */
6445 	sched_domains_numa_levels = 0;
6446 
6447 	sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6448 	if (!sched_domains_numa_masks)
6449 		return;
6450 
6451 	/*
6452 	 * Now for each level, construct a mask per node which contains all
6453 	 * cpus of nodes that are that many hops away from us.
6454 	 */
6455 	for (i = 0; i < level; i++) {
6456 		sched_domains_numa_masks[i] =
6457 			kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6458 		if (!sched_domains_numa_masks[i])
6459 			return;
6460 
6461 		for (j = 0; j < nr_node_ids; j++) {
6462 			struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6463 			if (!mask)
6464 				return;
6465 
6466 			sched_domains_numa_masks[i][j] = mask;
6467 
6468 			for (k = 0; k < nr_node_ids; k++) {
6469 				if (node_distance(j, k) > sched_domains_numa_distance[i])
6470 					continue;
6471 
6472 				cpumask_or(mask, mask, cpumask_of_node(k));
6473 			}
6474 		}
6475 	}
6476 
6477 	/* Compute default topology size */
6478 	for (i = 0; sched_domain_topology[i].mask; i++);
6479 
6480 	tl = kzalloc((i + level + 1) *
6481 			sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6482 	if (!tl)
6483 		return;
6484 
6485 	/*
6486 	 * Copy the default topology bits..
6487 	 */
6488 	for (i = 0; sched_domain_topology[i].mask; i++)
6489 		tl[i] = sched_domain_topology[i];
6490 
6491 	/*
6492 	 * .. and append 'j' levels of NUMA goodness.
6493 	 */
6494 	for (j = 0; j < level; i++, j++) {
6495 		tl[i] = (struct sched_domain_topology_level){
6496 			.mask = sd_numa_mask,
6497 			.sd_flags = cpu_numa_flags,
6498 			.flags = SDTL_OVERLAP,
6499 			.numa_level = j,
6500 			SD_INIT_NAME(NUMA)
6501 		};
6502 	}
6503 
6504 	sched_domain_topology = tl;
6505 
6506 	sched_domains_numa_levels = level;
6507 	sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6508 
6509 	init_numa_topology_type();
6510 }
6511 
6512 static void sched_domains_numa_masks_set(int cpu)
6513 {
6514 	int i, j;
6515 	int node = cpu_to_node(cpu);
6516 
6517 	for (i = 0; i < sched_domains_numa_levels; i++) {
6518 		for (j = 0; j < nr_node_ids; j++) {
6519 			if (node_distance(j, node) <= sched_domains_numa_distance[i])
6520 				cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6521 		}
6522 	}
6523 }
6524 
6525 static void sched_domains_numa_masks_clear(int cpu)
6526 {
6527 	int i, j;
6528 	for (i = 0; i < sched_domains_numa_levels; i++) {
6529 		for (j = 0; j < nr_node_ids; j++)
6530 			cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6531 	}
6532 }
6533 
6534 /*
6535  * Update sched_domains_numa_masks[level][node] array when new cpus
6536  * are onlined.
6537  */
6538 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6539 					   unsigned long action,
6540 					   void *hcpu)
6541 {
6542 	int cpu = (long)hcpu;
6543 
6544 	switch (action & ~CPU_TASKS_FROZEN) {
6545 	case CPU_ONLINE:
6546 		sched_domains_numa_masks_set(cpu);
6547 		break;
6548 
6549 	case CPU_DEAD:
6550 		sched_domains_numa_masks_clear(cpu);
6551 		break;
6552 
6553 	default:
6554 		return NOTIFY_DONE;
6555 	}
6556 
6557 	return NOTIFY_OK;
6558 }
6559 #else
6560 static inline void sched_init_numa(void)
6561 {
6562 }
6563 
6564 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6565 					   unsigned long action,
6566 					   void *hcpu)
6567 {
6568 	return 0;
6569 }
6570 #endif /* CONFIG_NUMA */
6571 
6572 static int __sdt_alloc(const struct cpumask *cpu_map)
6573 {
6574 	struct sched_domain_topology_level *tl;
6575 	int j;
6576 
6577 	for_each_sd_topology(tl) {
6578 		struct sd_data *sdd = &tl->data;
6579 
6580 		sdd->sd = alloc_percpu(struct sched_domain *);
6581 		if (!sdd->sd)
6582 			return -ENOMEM;
6583 
6584 		sdd->sg = alloc_percpu(struct sched_group *);
6585 		if (!sdd->sg)
6586 			return -ENOMEM;
6587 
6588 		sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6589 		if (!sdd->sgc)
6590 			return -ENOMEM;
6591 
6592 		for_each_cpu(j, cpu_map) {
6593 			struct sched_domain *sd;
6594 			struct sched_group *sg;
6595 			struct sched_group_capacity *sgc;
6596 
6597 		       	sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6598 					GFP_KERNEL, cpu_to_node(j));
6599 			if (!sd)
6600 				return -ENOMEM;
6601 
6602 			*per_cpu_ptr(sdd->sd, j) = sd;
6603 
6604 			sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6605 					GFP_KERNEL, cpu_to_node(j));
6606 			if (!sg)
6607 				return -ENOMEM;
6608 
6609 			sg->next = sg;
6610 
6611 			*per_cpu_ptr(sdd->sg, j) = sg;
6612 
6613 			sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6614 					GFP_KERNEL, cpu_to_node(j));
6615 			if (!sgc)
6616 				return -ENOMEM;
6617 
6618 			*per_cpu_ptr(sdd->sgc, j) = sgc;
6619 		}
6620 	}
6621 
6622 	return 0;
6623 }
6624 
6625 static void __sdt_free(const struct cpumask *cpu_map)
6626 {
6627 	struct sched_domain_topology_level *tl;
6628 	int j;
6629 
6630 	for_each_sd_topology(tl) {
6631 		struct sd_data *sdd = &tl->data;
6632 
6633 		for_each_cpu(j, cpu_map) {
6634 			struct sched_domain *sd;
6635 
6636 			if (sdd->sd) {
6637 				sd = *per_cpu_ptr(sdd->sd, j);
6638 				if (sd && (sd->flags & SD_OVERLAP))
6639 					free_sched_groups(sd->groups, 0);
6640 				kfree(*per_cpu_ptr(sdd->sd, j));
6641 			}
6642 
6643 			if (sdd->sg)
6644 				kfree(*per_cpu_ptr(sdd->sg, j));
6645 			if (sdd->sgc)
6646 				kfree(*per_cpu_ptr(sdd->sgc, j));
6647 		}
6648 		free_percpu(sdd->sd);
6649 		sdd->sd = NULL;
6650 		free_percpu(sdd->sg);
6651 		sdd->sg = NULL;
6652 		free_percpu(sdd->sgc);
6653 		sdd->sgc = NULL;
6654 	}
6655 }
6656 
6657 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6658 		const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6659 		struct sched_domain *child, int cpu)
6660 {
6661 	struct sched_domain *sd = sd_init(tl, cpu);
6662 	if (!sd)
6663 		return child;
6664 
6665 	cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6666 	if (child) {
6667 		sd->level = child->level + 1;
6668 		sched_domain_level_max = max(sched_domain_level_max, sd->level);
6669 		child->parent = sd;
6670 		sd->child = child;
6671 
6672 		if (!cpumask_subset(sched_domain_span(child),
6673 				    sched_domain_span(sd))) {
6674 			pr_err("BUG: arch topology borken\n");
6675 #ifdef CONFIG_SCHED_DEBUG
6676 			pr_err("     the %s domain not a subset of the %s domain\n",
6677 					child->name, sd->name);
6678 #endif
6679 			/* Fixup, ensure @sd has at least @child cpus. */
6680 			cpumask_or(sched_domain_span(sd),
6681 				   sched_domain_span(sd),
6682 				   sched_domain_span(child));
6683 		}
6684 
6685 	}
6686 	set_domain_attribute(sd, attr);
6687 
6688 	return sd;
6689 }
6690 
6691 /*
6692  * Build sched domains for a given set of cpus and attach the sched domains
6693  * to the individual cpus
6694  */
6695 static int build_sched_domains(const struct cpumask *cpu_map,
6696 			       struct sched_domain_attr *attr)
6697 {
6698 	enum s_alloc alloc_state;
6699 	struct sched_domain *sd;
6700 	struct s_data d;
6701 	int i, ret = -ENOMEM;
6702 
6703 	alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6704 	if (alloc_state != sa_rootdomain)
6705 		goto error;
6706 
6707 	/* Set up domains for cpus specified by the cpu_map. */
6708 	for_each_cpu(i, cpu_map) {
6709 		struct sched_domain_topology_level *tl;
6710 
6711 		sd = NULL;
6712 		for_each_sd_topology(tl) {
6713 			sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6714 			if (tl == sched_domain_topology)
6715 				*per_cpu_ptr(d.sd, i) = sd;
6716 			if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6717 				sd->flags |= SD_OVERLAP;
6718 			if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6719 				break;
6720 		}
6721 	}
6722 
6723 	/* Build the groups for the domains */
6724 	for_each_cpu(i, cpu_map) {
6725 		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6726 			sd->span_weight = cpumask_weight(sched_domain_span(sd));
6727 			if (sd->flags & SD_OVERLAP) {
6728 				if (build_overlap_sched_groups(sd, i))
6729 					goto error;
6730 			} else {
6731 				if (build_sched_groups(sd, i))
6732 					goto error;
6733 			}
6734 		}
6735 	}
6736 
6737 	/* Calculate CPU capacity for physical packages and nodes */
6738 	for (i = nr_cpumask_bits-1; i >= 0; i--) {
6739 		if (!cpumask_test_cpu(i, cpu_map))
6740 			continue;
6741 
6742 		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6743 			claim_allocations(i, sd);
6744 			init_sched_groups_capacity(i, sd);
6745 		}
6746 	}
6747 
6748 	/* Attach the domains */
6749 	rcu_read_lock();
6750 	for_each_cpu(i, cpu_map) {
6751 		sd = *per_cpu_ptr(d.sd, i);
6752 		cpu_attach_domain(sd, d.rd, i);
6753 	}
6754 	rcu_read_unlock();
6755 
6756 	ret = 0;
6757 error:
6758 	__free_domain_allocs(&d, alloc_state, cpu_map);
6759 	return ret;
6760 }
6761 
6762 static cpumask_var_t *doms_cur;	/* current sched domains */
6763 static int ndoms_cur;		/* number of sched domains in 'doms_cur' */
6764 static struct sched_domain_attr *dattr_cur;
6765 				/* attribues of custom domains in 'doms_cur' */
6766 
6767 /*
6768  * Special case: If a kmalloc of a doms_cur partition (array of
6769  * cpumask) fails, then fallback to a single sched domain,
6770  * as determined by the single cpumask fallback_doms.
6771  */
6772 static cpumask_var_t fallback_doms;
6773 
6774 /*
6775  * arch_update_cpu_topology lets virtualized architectures update the
6776  * cpu core maps. It is supposed to return 1 if the topology changed
6777  * or 0 if it stayed the same.
6778  */
6779 int __weak arch_update_cpu_topology(void)
6780 {
6781 	return 0;
6782 }
6783 
6784 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6785 {
6786 	int i;
6787 	cpumask_var_t *doms;
6788 
6789 	doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6790 	if (!doms)
6791 		return NULL;
6792 	for (i = 0; i < ndoms; i++) {
6793 		if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6794 			free_sched_domains(doms, i);
6795 			return NULL;
6796 		}
6797 	}
6798 	return doms;
6799 }
6800 
6801 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6802 {
6803 	unsigned int i;
6804 	for (i = 0; i < ndoms; i++)
6805 		free_cpumask_var(doms[i]);
6806 	kfree(doms);
6807 }
6808 
6809 /*
6810  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6811  * For now this just excludes isolated cpus, but could be used to
6812  * exclude other special cases in the future.
6813  */
6814 static int init_sched_domains(const struct cpumask *cpu_map)
6815 {
6816 	int err;
6817 
6818 	arch_update_cpu_topology();
6819 	ndoms_cur = 1;
6820 	doms_cur = alloc_sched_domains(ndoms_cur);
6821 	if (!doms_cur)
6822 		doms_cur = &fallback_doms;
6823 	cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6824 	err = build_sched_domains(doms_cur[0], NULL);
6825 	register_sched_domain_sysctl();
6826 
6827 	return err;
6828 }
6829 
6830 /*
6831  * Detach sched domains from a group of cpus specified in cpu_map
6832  * These cpus will now be attached to the NULL domain
6833  */
6834 static void detach_destroy_domains(const struct cpumask *cpu_map)
6835 {
6836 	int i;
6837 
6838 	rcu_read_lock();
6839 	for_each_cpu(i, cpu_map)
6840 		cpu_attach_domain(NULL, &def_root_domain, i);
6841 	rcu_read_unlock();
6842 }
6843 
6844 /* handle null as "default" */
6845 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6846 			struct sched_domain_attr *new, int idx_new)
6847 {
6848 	struct sched_domain_attr tmp;
6849 
6850 	/* fast path */
6851 	if (!new && !cur)
6852 		return 1;
6853 
6854 	tmp = SD_ATTR_INIT;
6855 	return !memcmp(cur ? (cur + idx_cur) : &tmp,
6856 			new ? (new + idx_new) : &tmp,
6857 			sizeof(struct sched_domain_attr));
6858 }
6859 
6860 /*
6861  * Partition sched domains as specified by the 'ndoms_new'
6862  * cpumasks in the array doms_new[] of cpumasks. This compares
6863  * doms_new[] to the current sched domain partitioning, doms_cur[].
6864  * It destroys each deleted domain and builds each new domain.
6865  *
6866  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6867  * The masks don't intersect (don't overlap.) We should setup one
6868  * sched domain for each mask. CPUs not in any of the cpumasks will
6869  * not be load balanced. If the same cpumask appears both in the
6870  * current 'doms_cur' domains and in the new 'doms_new', we can leave
6871  * it as it is.
6872  *
6873  * The passed in 'doms_new' should be allocated using
6874  * alloc_sched_domains.  This routine takes ownership of it and will
6875  * free_sched_domains it when done with it. If the caller failed the
6876  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6877  * and partition_sched_domains() will fallback to the single partition
6878  * 'fallback_doms', it also forces the domains to be rebuilt.
6879  *
6880  * If doms_new == NULL it will be replaced with cpu_online_mask.
6881  * ndoms_new == 0 is a special case for destroying existing domains,
6882  * and it will not create the default domain.
6883  *
6884  * Call with hotplug lock held
6885  */
6886 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6887 			     struct sched_domain_attr *dattr_new)
6888 {
6889 	int i, j, n;
6890 	int new_topology;
6891 
6892 	mutex_lock(&sched_domains_mutex);
6893 
6894 	/* always unregister in case we don't destroy any domains */
6895 	unregister_sched_domain_sysctl();
6896 
6897 	/* Let architecture update cpu core mappings. */
6898 	new_topology = arch_update_cpu_topology();
6899 
6900 	n = doms_new ? ndoms_new : 0;
6901 
6902 	/* Destroy deleted domains */
6903 	for (i = 0; i < ndoms_cur; i++) {
6904 		for (j = 0; j < n && !new_topology; j++) {
6905 			if (cpumask_equal(doms_cur[i], doms_new[j])
6906 			    && dattrs_equal(dattr_cur, i, dattr_new, j))
6907 				goto match1;
6908 		}
6909 		/* no match - a current sched domain not in new doms_new[] */
6910 		detach_destroy_domains(doms_cur[i]);
6911 match1:
6912 		;
6913 	}
6914 
6915 	n = ndoms_cur;
6916 	if (doms_new == NULL) {
6917 		n = 0;
6918 		doms_new = &fallback_doms;
6919 		cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6920 		WARN_ON_ONCE(dattr_new);
6921 	}
6922 
6923 	/* Build new domains */
6924 	for (i = 0; i < ndoms_new; i++) {
6925 		for (j = 0; j < n && !new_topology; j++) {
6926 			if (cpumask_equal(doms_new[i], doms_cur[j])
6927 			    && dattrs_equal(dattr_new, i, dattr_cur, j))
6928 				goto match2;
6929 		}
6930 		/* no match - add a new doms_new */
6931 		build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6932 match2:
6933 		;
6934 	}
6935 
6936 	/* Remember the new sched domains */
6937 	if (doms_cur != &fallback_doms)
6938 		free_sched_domains(doms_cur, ndoms_cur);
6939 	kfree(dattr_cur);	/* kfree(NULL) is safe */
6940 	doms_cur = doms_new;
6941 	dattr_cur = dattr_new;
6942 	ndoms_cur = ndoms_new;
6943 
6944 	register_sched_domain_sysctl();
6945 
6946 	mutex_unlock(&sched_domains_mutex);
6947 }
6948 
6949 static int num_cpus_frozen;	/* used to mark begin/end of suspend/resume */
6950 
6951 /*
6952  * Update cpusets according to cpu_active mask.  If cpusets are
6953  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6954  * around partition_sched_domains().
6955  *
6956  * If we come here as part of a suspend/resume, don't touch cpusets because we
6957  * want to restore it back to its original state upon resume anyway.
6958  */
6959 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6960 			     void *hcpu)
6961 {
6962 	switch (action) {
6963 	case CPU_ONLINE_FROZEN:
6964 	case CPU_DOWN_FAILED_FROZEN:
6965 
6966 		/*
6967 		 * num_cpus_frozen tracks how many CPUs are involved in suspend
6968 		 * resume sequence. As long as this is not the last online
6969 		 * operation in the resume sequence, just build a single sched
6970 		 * domain, ignoring cpusets.
6971 		 */
6972 		num_cpus_frozen--;
6973 		if (likely(num_cpus_frozen)) {
6974 			partition_sched_domains(1, NULL, NULL);
6975 			break;
6976 		}
6977 
6978 		/*
6979 		 * This is the last CPU online operation. So fall through and
6980 		 * restore the original sched domains by considering the
6981 		 * cpuset configurations.
6982 		 */
6983 
6984 	case CPU_ONLINE:
6985 		cpuset_update_active_cpus(true);
6986 		break;
6987 	default:
6988 		return NOTIFY_DONE;
6989 	}
6990 	return NOTIFY_OK;
6991 }
6992 
6993 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6994 			       void *hcpu)
6995 {
6996 	unsigned long flags;
6997 	long cpu = (long)hcpu;
6998 	struct dl_bw *dl_b;
6999 	bool overflow;
7000 	int cpus;
7001 
7002 	switch (action) {
7003 	case CPU_DOWN_PREPARE:
7004 		rcu_read_lock_sched();
7005 		dl_b = dl_bw_of(cpu);
7006 
7007 		raw_spin_lock_irqsave(&dl_b->lock, flags);
7008 		cpus = dl_bw_cpus(cpu);
7009 		overflow = __dl_overflow(dl_b, cpus, 0, 0);
7010 		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7011 
7012 		rcu_read_unlock_sched();
7013 
7014 		if (overflow)
7015 			return notifier_from_errno(-EBUSY);
7016 		cpuset_update_active_cpus(false);
7017 		break;
7018 	case CPU_DOWN_PREPARE_FROZEN:
7019 		num_cpus_frozen++;
7020 		partition_sched_domains(1, NULL, NULL);
7021 		break;
7022 	default:
7023 		return NOTIFY_DONE;
7024 	}
7025 	return NOTIFY_OK;
7026 }
7027 
7028 void __init sched_init_smp(void)
7029 {
7030 	cpumask_var_t non_isolated_cpus;
7031 
7032 	alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7033 	alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7034 
7035 	sched_init_numa();
7036 
7037 	/*
7038 	 * There's no userspace yet to cause hotplug operations; hence all the
7039 	 * cpu masks are stable and all blatant races in the below code cannot
7040 	 * happen.
7041 	 */
7042 	mutex_lock(&sched_domains_mutex);
7043 	init_sched_domains(cpu_active_mask);
7044 	cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7045 	if (cpumask_empty(non_isolated_cpus))
7046 		cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7047 	mutex_unlock(&sched_domains_mutex);
7048 
7049 	hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7050 	hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7051 	hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7052 
7053 	init_hrtick();
7054 
7055 	/* Move init over to a non-isolated CPU */
7056 	if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7057 		BUG();
7058 	sched_init_granularity();
7059 	free_cpumask_var(non_isolated_cpus);
7060 
7061 	init_sched_rt_class();
7062 	init_sched_dl_class();
7063 }
7064 #else
7065 void __init sched_init_smp(void)
7066 {
7067 	sched_init_granularity();
7068 }
7069 #endif /* CONFIG_SMP */
7070 
7071 const_debug unsigned int sysctl_timer_migration = 1;
7072 
7073 int in_sched_functions(unsigned long addr)
7074 {
7075 	return in_lock_functions(addr) ||
7076 		(addr >= (unsigned long)__sched_text_start
7077 		&& addr < (unsigned long)__sched_text_end);
7078 }
7079 
7080 #ifdef CONFIG_CGROUP_SCHED
7081 /*
7082  * Default task group.
7083  * Every task in system belongs to this group at bootup.
7084  */
7085 struct task_group root_task_group;
7086 LIST_HEAD(task_groups);
7087 #endif
7088 
7089 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7090 
7091 void __init sched_init(void)
7092 {
7093 	int i, j;
7094 	unsigned long alloc_size = 0, ptr;
7095 
7096 #ifdef CONFIG_FAIR_GROUP_SCHED
7097 	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7098 #endif
7099 #ifdef CONFIG_RT_GROUP_SCHED
7100 	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7101 #endif
7102 	if (alloc_size) {
7103 		ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7104 
7105 #ifdef CONFIG_FAIR_GROUP_SCHED
7106 		root_task_group.se = (struct sched_entity **)ptr;
7107 		ptr += nr_cpu_ids * sizeof(void **);
7108 
7109 		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7110 		ptr += nr_cpu_ids * sizeof(void **);
7111 
7112 #endif /* CONFIG_FAIR_GROUP_SCHED */
7113 #ifdef CONFIG_RT_GROUP_SCHED
7114 		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7115 		ptr += nr_cpu_ids * sizeof(void **);
7116 
7117 		root_task_group.rt_rq = (struct rt_rq **)ptr;
7118 		ptr += nr_cpu_ids * sizeof(void **);
7119 
7120 #endif /* CONFIG_RT_GROUP_SCHED */
7121 	}
7122 #ifdef CONFIG_CPUMASK_OFFSTACK
7123 	for_each_possible_cpu(i) {
7124 		per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7125 			cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7126 	}
7127 #endif /* CONFIG_CPUMASK_OFFSTACK */
7128 
7129 	init_rt_bandwidth(&def_rt_bandwidth,
7130 			global_rt_period(), global_rt_runtime());
7131 	init_dl_bandwidth(&def_dl_bandwidth,
7132 			global_rt_period(), global_rt_runtime());
7133 
7134 #ifdef CONFIG_SMP
7135 	init_defrootdomain();
7136 #endif
7137 
7138 #ifdef CONFIG_RT_GROUP_SCHED
7139 	init_rt_bandwidth(&root_task_group.rt_bandwidth,
7140 			global_rt_period(), global_rt_runtime());
7141 #endif /* CONFIG_RT_GROUP_SCHED */
7142 
7143 #ifdef CONFIG_CGROUP_SCHED
7144 	list_add(&root_task_group.list, &task_groups);
7145 	INIT_LIST_HEAD(&root_task_group.children);
7146 	INIT_LIST_HEAD(&root_task_group.siblings);
7147 	autogroup_init(&init_task);
7148 
7149 #endif /* CONFIG_CGROUP_SCHED */
7150 
7151 	for_each_possible_cpu(i) {
7152 		struct rq *rq;
7153 
7154 		rq = cpu_rq(i);
7155 		raw_spin_lock_init(&rq->lock);
7156 		rq->nr_running = 0;
7157 		rq->calc_load_active = 0;
7158 		rq->calc_load_update = jiffies + LOAD_FREQ;
7159 		init_cfs_rq(&rq->cfs);
7160 		init_rt_rq(&rq->rt);
7161 		init_dl_rq(&rq->dl);
7162 #ifdef CONFIG_FAIR_GROUP_SCHED
7163 		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7164 		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7165 		/*
7166 		 * How much cpu bandwidth does root_task_group get?
7167 		 *
7168 		 * In case of task-groups formed thr' the cgroup filesystem, it
7169 		 * gets 100% of the cpu resources in the system. This overall
7170 		 * system cpu resource is divided among the tasks of
7171 		 * root_task_group and its child task-groups in a fair manner,
7172 		 * based on each entity's (task or task-group's) weight
7173 		 * (se->load.weight).
7174 		 *
7175 		 * In other words, if root_task_group has 10 tasks of weight
7176 		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7177 		 * then A0's share of the cpu resource is:
7178 		 *
7179 		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7180 		 *
7181 		 * We achieve this by letting root_task_group's tasks sit
7182 		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7183 		 */
7184 		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7185 		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7186 #endif /* CONFIG_FAIR_GROUP_SCHED */
7187 
7188 		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7189 #ifdef CONFIG_RT_GROUP_SCHED
7190 		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7191 #endif
7192 
7193 		for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7194 			rq->cpu_load[j] = 0;
7195 
7196 		rq->last_load_update_tick = jiffies;
7197 
7198 #ifdef CONFIG_SMP
7199 		rq->sd = NULL;
7200 		rq->rd = NULL;
7201 		rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7202 		rq->post_schedule = 0;
7203 		rq->active_balance = 0;
7204 		rq->next_balance = jiffies;
7205 		rq->push_cpu = 0;
7206 		rq->cpu = i;
7207 		rq->online = 0;
7208 		rq->idle_stamp = 0;
7209 		rq->avg_idle = 2*sysctl_sched_migration_cost;
7210 		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7211 
7212 		INIT_LIST_HEAD(&rq->cfs_tasks);
7213 
7214 		rq_attach_root(rq, &def_root_domain);
7215 #ifdef CONFIG_NO_HZ_COMMON
7216 		rq->nohz_flags = 0;
7217 #endif
7218 #ifdef CONFIG_NO_HZ_FULL
7219 		rq->last_sched_tick = 0;
7220 #endif
7221 #endif
7222 		init_rq_hrtick(rq);
7223 		atomic_set(&rq->nr_iowait, 0);
7224 	}
7225 
7226 	set_load_weight(&init_task);
7227 
7228 #ifdef CONFIG_PREEMPT_NOTIFIERS
7229 	INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7230 #endif
7231 
7232 	/*
7233 	 * The boot idle thread does lazy MMU switching as well:
7234 	 */
7235 	atomic_inc(&init_mm.mm_count);
7236 	enter_lazy_tlb(&init_mm, current);
7237 
7238 	/*
7239 	 * During early bootup we pretend to be a normal task:
7240 	 */
7241 	current->sched_class = &fair_sched_class;
7242 
7243 	/*
7244 	 * Make us the idle thread. Technically, schedule() should not be
7245 	 * called from this thread, however somewhere below it might be,
7246 	 * but because we are the idle thread, we just pick up running again
7247 	 * when this runqueue becomes "idle".
7248 	 */
7249 	init_idle(current, smp_processor_id());
7250 
7251 	calc_load_update = jiffies + LOAD_FREQ;
7252 
7253 #ifdef CONFIG_SMP
7254 	zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7255 	/* May be allocated at isolcpus cmdline parse time */
7256 	if (cpu_isolated_map == NULL)
7257 		zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7258 	idle_thread_set_boot_cpu();
7259 	set_cpu_rq_start_time();
7260 #endif
7261 	init_sched_fair_class();
7262 
7263 	scheduler_running = 1;
7264 }
7265 
7266 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7267 static inline int preempt_count_equals(int preempt_offset)
7268 {
7269 	int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7270 
7271 	return (nested == preempt_offset);
7272 }
7273 
7274 void __might_sleep(const char *file, int line, int preempt_offset)
7275 {
7276 	/*
7277 	 * Blocking primitives will set (and therefore destroy) current->state,
7278 	 * since we will exit with TASK_RUNNING make sure we enter with it,
7279 	 * otherwise we will destroy state.
7280 	 */
7281 	WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7282 			"do not call blocking ops when !TASK_RUNNING; "
7283 			"state=%lx set at [<%p>] %pS\n",
7284 			current->state,
7285 			(void *)current->task_state_change,
7286 			(void *)current->task_state_change);
7287 
7288 	___might_sleep(file, line, preempt_offset);
7289 }
7290 EXPORT_SYMBOL(__might_sleep);
7291 
7292 void ___might_sleep(const char *file, int line, int preempt_offset)
7293 {
7294 	static unsigned long prev_jiffy;	/* ratelimiting */
7295 
7296 	rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7297 	if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7298 	     !is_idle_task(current)) ||
7299 	    system_state != SYSTEM_RUNNING || oops_in_progress)
7300 		return;
7301 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7302 		return;
7303 	prev_jiffy = jiffies;
7304 
7305 	printk(KERN_ERR
7306 		"BUG: sleeping function called from invalid context at %s:%d\n",
7307 			file, line);
7308 	printk(KERN_ERR
7309 		"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7310 			in_atomic(), irqs_disabled(),
7311 			current->pid, current->comm);
7312 
7313 	if (task_stack_end_corrupted(current))
7314 		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7315 
7316 	debug_show_held_locks(current);
7317 	if (irqs_disabled())
7318 		print_irqtrace_events(current);
7319 #ifdef CONFIG_DEBUG_PREEMPT
7320 	if (!preempt_count_equals(preempt_offset)) {
7321 		pr_err("Preemption disabled at:");
7322 		print_ip_sym(current->preempt_disable_ip);
7323 		pr_cont("\n");
7324 	}
7325 #endif
7326 	dump_stack();
7327 }
7328 EXPORT_SYMBOL(___might_sleep);
7329 #endif
7330 
7331 #ifdef CONFIG_MAGIC_SYSRQ
7332 static void normalize_task(struct rq *rq, struct task_struct *p)
7333 {
7334 	const struct sched_class *prev_class = p->sched_class;
7335 	struct sched_attr attr = {
7336 		.sched_policy = SCHED_NORMAL,
7337 	};
7338 	int old_prio = p->prio;
7339 	int queued;
7340 
7341 	queued = task_on_rq_queued(p);
7342 	if (queued)
7343 		dequeue_task(rq, p, 0);
7344 	__setscheduler(rq, p, &attr, false);
7345 	if (queued) {
7346 		enqueue_task(rq, p, 0);
7347 		resched_curr(rq);
7348 	}
7349 
7350 	check_class_changed(rq, p, prev_class, old_prio);
7351 }
7352 
7353 void normalize_rt_tasks(void)
7354 {
7355 	struct task_struct *g, *p;
7356 	unsigned long flags;
7357 	struct rq *rq;
7358 
7359 	read_lock(&tasklist_lock);
7360 	for_each_process_thread(g, p) {
7361 		/*
7362 		 * Only normalize user tasks:
7363 		 */
7364 		if (p->flags & PF_KTHREAD)
7365 			continue;
7366 
7367 		p->se.exec_start		= 0;
7368 #ifdef CONFIG_SCHEDSTATS
7369 		p->se.statistics.wait_start	= 0;
7370 		p->se.statistics.sleep_start	= 0;
7371 		p->se.statistics.block_start	= 0;
7372 #endif
7373 
7374 		if (!dl_task(p) && !rt_task(p)) {
7375 			/*
7376 			 * Renice negative nice level userspace
7377 			 * tasks back to 0:
7378 			 */
7379 			if (task_nice(p) < 0)
7380 				set_user_nice(p, 0);
7381 			continue;
7382 		}
7383 
7384 		rq = task_rq_lock(p, &flags);
7385 		normalize_task(rq, p);
7386 		task_rq_unlock(rq, p, &flags);
7387 	}
7388 	read_unlock(&tasklist_lock);
7389 }
7390 
7391 #endif /* CONFIG_MAGIC_SYSRQ */
7392 
7393 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7394 /*
7395  * These functions are only useful for the IA64 MCA handling, or kdb.
7396  *
7397  * They can only be called when the whole system has been
7398  * stopped - every CPU needs to be quiescent, and no scheduling
7399  * activity can take place. Using them for anything else would
7400  * be a serious bug, and as a result, they aren't even visible
7401  * under any other configuration.
7402  */
7403 
7404 /**
7405  * curr_task - return the current task for a given cpu.
7406  * @cpu: the processor in question.
7407  *
7408  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7409  *
7410  * Return: The current task for @cpu.
7411  */
7412 struct task_struct *curr_task(int cpu)
7413 {
7414 	return cpu_curr(cpu);
7415 }
7416 
7417 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7418 
7419 #ifdef CONFIG_IA64
7420 /**
7421  * set_curr_task - set the current task for a given cpu.
7422  * @cpu: the processor in question.
7423  * @p: the task pointer to set.
7424  *
7425  * Description: This function must only be used when non-maskable interrupts
7426  * are serviced on a separate stack. It allows the architecture to switch the
7427  * notion of the current task on a cpu in a non-blocking manner. This function
7428  * must be called with all CPU's synchronized, and interrupts disabled, the
7429  * and caller must save the original value of the current task (see
7430  * curr_task() above) and restore that value before reenabling interrupts and
7431  * re-starting the system.
7432  *
7433  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7434  */
7435 void set_curr_task(int cpu, struct task_struct *p)
7436 {
7437 	cpu_curr(cpu) = p;
7438 }
7439 
7440 #endif
7441 
7442 #ifdef CONFIG_CGROUP_SCHED
7443 /* task_group_lock serializes the addition/removal of task groups */
7444 static DEFINE_SPINLOCK(task_group_lock);
7445 
7446 static void free_sched_group(struct task_group *tg)
7447 {
7448 	free_fair_sched_group(tg);
7449 	free_rt_sched_group(tg);
7450 	autogroup_free(tg);
7451 	kfree(tg);
7452 }
7453 
7454 /* allocate runqueue etc for a new task group */
7455 struct task_group *sched_create_group(struct task_group *parent)
7456 {
7457 	struct task_group *tg;
7458 
7459 	tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7460 	if (!tg)
7461 		return ERR_PTR(-ENOMEM);
7462 
7463 	if (!alloc_fair_sched_group(tg, parent))
7464 		goto err;
7465 
7466 	if (!alloc_rt_sched_group(tg, parent))
7467 		goto err;
7468 
7469 	return tg;
7470 
7471 err:
7472 	free_sched_group(tg);
7473 	return ERR_PTR(-ENOMEM);
7474 }
7475 
7476 void sched_online_group(struct task_group *tg, struct task_group *parent)
7477 {
7478 	unsigned long flags;
7479 
7480 	spin_lock_irqsave(&task_group_lock, flags);
7481 	list_add_rcu(&tg->list, &task_groups);
7482 
7483 	WARN_ON(!parent); /* root should already exist */
7484 
7485 	tg->parent = parent;
7486 	INIT_LIST_HEAD(&tg->children);
7487 	list_add_rcu(&tg->siblings, &parent->children);
7488 	spin_unlock_irqrestore(&task_group_lock, flags);
7489 }
7490 
7491 /* rcu callback to free various structures associated with a task group */
7492 static void free_sched_group_rcu(struct rcu_head *rhp)
7493 {
7494 	/* now it should be safe to free those cfs_rqs */
7495 	free_sched_group(container_of(rhp, struct task_group, rcu));
7496 }
7497 
7498 /* Destroy runqueue etc associated with a task group */
7499 void sched_destroy_group(struct task_group *tg)
7500 {
7501 	/* wait for possible concurrent references to cfs_rqs complete */
7502 	call_rcu(&tg->rcu, free_sched_group_rcu);
7503 }
7504 
7505 void sched_offline_group(struct task_group *tg)
7506 {
7507 	unsigned long flags;
7508 	int i;
7509 
7510 	/* end participation in shares distribution */
7511 	for_each_possible_cpu(i)
7512 		unregister_fair_sched_group(tg, i);
7513 
7514 	spin_lock_irqsave(&task_group_lock, flags);
7515 	list_del_rcu(&tg->list);
7516 	list_del_rcu(&tg->siblings);
7517 	spin_unlock_irqrestore(&task_group_lock, flags);
7518 }
7519 
7520 /* change task's runqueue when it moves between groups.
7521  *	The caller of this function should have put the task in its new group
7522  *	by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7523  *	reflect its new group.
7524  */
7525 void sched_move_task(struct task_struct *tsk)
7526 {
7527 	struct task_group *tg;
7528 	int queued, running;
7529 	unsigned long flags;
7530 	struct rq *rq;
7531 
7532 	rq = task_rq_lock(tsk, &flags);
7533 
7534 	running = task_current(rq, tsk);
7535 	queued = task_on_rq_queued(tsk);
7536 
7537 	if (queued)
7538 		dequeue_task(rq, tsk, 0);
7539 	if (unlikely(running))
7540 		put_prev_task(rq, tsk);
7541 
7542 	/*
7543 	 * All callers are synchronized by task_rq_lock(); we do not use RCU
7544 	 * which is pointless here. Thus, we pass "true" to task_css_check()
7545 	 * to prevent lockdep warnings.
7546 	 */
7547 	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7548 			  struct task_group, css);
7549 	tg = autogroup_task_group(tsk, tg);
7550 	tsk->sched_task_group = tg;
7551 
7552 #ifdef CONFIG_FAIR_GROUP_SCHED
7553 	if (tsk->sched_class->task_move_group)
7554 		tsk->sched_class->task_move_group(tsk, queued);
7555 	else
7556 #endif
7557 		set_task_rq(tsk, task_cpu(tsk));
7558 
7559 	if (unlikely(running))
7560 		tsk->sched_class->set_curr_task(rq);
7561 	if (queued)
7562 		enqueue_task(rq, tsk, 0);
7563 
7564 	task_rq_unlock(rq, tsk, &flags);
7565 }
7566 #endif /* CONFIG_CGROUP_SCHED */
7567 
7568 #ifdef CONFIG_RT_GROUP_SCHED
7569 /*
7570  * Ensure that the real time constraints are schedulable.
7571  */
7572 static DEFINE_MUTEX(rt_constraints_mutex);
7573 
7574 /* Must be called with tasklist_lock held */
7575 static inline int tg_has_rt_tasks(struct task_group *tg)
7576 {
7577 	struct task_struct *g, *p;
7578 
7579 	/*
7580 	 * Autogroups do not have RT tasks; see autogroup_create().
7581 	 */
7582 	if (task_group_is_autogroup(tg))
7583 		return 0;
7584 
7585 	for_each_process_thread(g, p) {
7586 		if (rt_task(p) && task_group(p) == tg)
7587 			return 1;
7588 	}
7589 
7590 	return 0;
7591 }
7592 
7593 struct rt_schedulable_data {
7594 	struct task_group *tg;
7595 	u64 rt_period;
7596 	u64 rt_runtime;
7597 };
7598 
7599 static int tg_rt_schedulable(struct task_group *tg, void *data)
7600 {
7601 	struct rt_schedulable_data *d = data;
7602 	struct task_group *child;
7603 	unsigned long total, sum = 0;
7604 	u64 period, runtime;
7605 
7606 	period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7607 	runtime = tg->rt_bandwidth.rt_runtime;
7608 
7609 	if (tg == d->tg) {
7610 		period = d->rt_period;
7611 		runtime = d->rt_runtime;
7612 	}
7613 
7614 	/*
7615 	 * Cannot have more runtime than the period.
7616 	 */
7617 	if (runtime > period && runtime != RUNTIME_INF)
7618 		return -EINVAL;
7619 
7620 	/*
7621 	 * Ensure we don't starve existing RT tasks.
7622 	 */
7623 	if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7624 		return -EBUSY;
7625 
7626 	total = to_ratio(period, runtime);
7627 
7628 	/*
7629 	 * Nobody can have more than the global setting allows.
7630 	 */
7631 	if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7632 		return -EINVAL;
7633 
7634 	/*
7635 	 * The sum of our children's runtime should not exceed our own.
7636 	 */
7637 	list_for_each_entry_rcu(child, &tg->children, siblings) {
7638 		period = ktime_to_ns(child->rt_bandwidth.rt_period);
7639 		runtime = child->rt_bandwidth.rt_runtime;
7640 
7641 		if (child == d->tg) {
7642 			period = d->rt_period;
7643 			runtime = d->rt_runtime;
7644 		}
7645 
7646 		sum += to_ratio(period, runtime);
7647 	}
7648 
7649 	if (sum > total)
7650 		return -EINVAL;
7651 
7652 	return 0;
7653 }
7654 
7655 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7656 {
7657 	int ret;
7658 
7659 	struct rt_schedulable_data data = {
7660 		.tg = tg,
7661 		.rt_period = period,
7662 		.rt_runtime = runtime,
7663 	};
7664 
7665 	rcu_read_lock();
7666 	ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7667 	rcu_read_unlock();
7668 
7669 	return ret;
7670 }
7671 
7672 static int tg_set_rt_bandwidth(struct task_group *tg,
7673 		u64 rt_period, u64 rt_runtime)
7674 {
7675 	int i, err = 0;
7676 
7677 	/*
7678 	 * Disallowing the root group RT runtime is BAD, it would disallow the
7679 	 * kernel creating (and or operating) RT threads.
7680 	 */
7681 	if (tg == &root_task_group && rt_runtime == 0)
7682 		return -EINVAL;
7683 
7684 	/* No period doesn't make any sense. */
7685 	if (rt_period == 0)
7686 		return -EINVAL;
7687 
7688 	mutex_lock(&rt_constraints_mutex);
7689 	read_lock(&tasklist_lock);
7690 	err = __rt_schedulable(tg, rt_period, rt_runtime);
7691 	if (err)
7692 		goto unlock;
7693 
7694 	raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7695 	tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7696 	tg->rt_bandwidth.rt_runtime = rt_runtime;
7697 
7698 	for_each_possible_cpu(i) {
7699 		struct rt_rq *rt_rq = tg->rt_rq[i];
7700 
7701 		raw_spin_lock(&rt_rq->rt_runtime_lock);
7702 		rt_rq->rt_runtime = rt_runtime;
7703 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
7704 	}
7705 	raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7706 unlock:
7707 	read_unlock(&tasklist_lock);
7708 	mutex_unlock(&rt_constraints_mutex);
7709 
7710 	return err;
7711 }
7712 
7713 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7714 {
7715 	u64 rt_runtime, rt_period;
7716 
7717 	rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7718 	rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7719 	if (rt_runtime_us < 0)
7720 		rt_runtime = RUNTIME_INF;
7721 
7722 	return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7723 }
7724 
7725 static long sched_group_rt_runtime(struct task_group *tg)
7726 {
7727 	u64 rt_runtime_us;
7728 
7729 	if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7730 		return -1;
7731 
7732 	rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7733 	do_div(rt_runtime_us, NSEC_PER_USEC);
7734 	return rt_runtime_us;
7735 }
7736 
7737 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7738 {
7739 	u64 rt_runtime, rt_period;
7740 
7741 	rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7742 	rt_runtime = tg->rt_bandwidth.rt_runtime;
7743 
7744 	return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7745 }
7746 
7747 static long sched_group_rt_period(struct task_group *tg)
7748 {
7749 	u64 rt_period_us;
7750 
7751 	rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7752 	do_div(rt_period_us, NSEC_PER_USEC);
7753 	return rt_period_us;
7754 }
7755 #endif /* CONFIG_RT_GROUP_SCHED */
7756 
7757 #ifdef CONFIG_RT_GROUP_SCHED
7758 static int sched_rt_global_constraints(void)
7759 {
7760 	int ret = 0;
7761 
7762 	mutex_lock(&rt_constraints_mutex);
7763 	read_lock(&tasklist_lock);
7764 	ret = __rt_schedulable(NULL, 0, 0);
7765 	read_unlock(&tasklist_lock);
7766 	mutex_unlock(&rt_constraints_mutex);
7767 
7768 	return ret;
7769 }
7770 
7771 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7772 {
7773 	/* Don't accept realtime tasks when there is no way for them to run */
7774 	if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7775 		return 0;
7776 
7777 	return 1;
7778 }
7779 
7780 #else /* !CONFIG_RT_GROUP_SCHED */
7781 static int sched_rt_global_constraints(void)
7782 {
7783 	unsigned long flags;
7784 	int i, ret = 0;
7785 
7786 	raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7787 	for_each_possible_cpu(i) {
7788 		struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7789 
7790 		raw_spin_lock(&rt_rq->rt_runtime_lock);
7791 		rt_rq->rt_runtime = global_rt_runtime();
7792 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
7793 	}
7794 	raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7795 
7796 	return ret;
7797 }
7798 #endif /* CONFIG_RT_GROUP_SCHED */
7799 
7800 static int sched_dl_global_validate(void)
7801 {
7802 	u64 runtime = global_rt_runtime();
7803 	u64 period = global_rt_period();
7804 	u64 new_bw = to_ratio(period, runtime);
7805 	struct dl_bw *dl_b;
7806 	int cpu, ret = 0;
7807 	unsigned long flags;
7808 
7809 	/*
7810 	 * Here we want to check the bandwidth not being set to some
7811 	 * value smaller than the currently allocated bandwidth in
7812 	 * any of the root_domains.
7813 	 *
7814 	 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7815 	 * cycling on root_domains... Discussion on different/better
7816 	 * solutions is welcome!
7817 	 */
7818 	for_each_possible_cpu(cpu) {
7819 		rcu_read_lock_sched();
7820 		dl_b = dl_bw_of(cpu);
7821 
7822 		raw_spin_lock_irqsave(&dl_b->lock, flags);
7823 		if (new_bw < dl_b->total_bw)
7824 			ret = -EBUSY;
7825 		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7826 
7827 		rcu_read_unlock_sched();
7828 
7829 		if (ret)
7830 			break;
7831 	}
7832 
7833 	return ret;
7834 }
7835 
7836 static void sched_dl_do_global(void)
7837 {
7838 	u64 new_bw = -1;
7839 	struct dl_bw *dl_b;
7840 	int cpu;
7841 	unsigned long flags;
7842 
7843 	def_dl_bandwidth.dl_period = global_rt_period();
7844 	def_dl_bandwidth.dl_runtime = global_rt_runtime();
7845 
7846 	if (global_rt_runtime() != RUNTIME_INF)
7847 		new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7848 
7849 	/*
7850 	 * FIXME: As above...
7851 	 */
7852 	for_each_possible_cpu(cpu) {
7853 		rcu_read_lock_sched();
7854 		dl_b = dl_bw_of(cpu);
7855 
7856 		raw_spin_lock_irqsave(&dl_b->lock, flags);
7857 		dl_b->bw = new_bw;
7858 		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7859 
7860 		rcu_read_unlock_sched();
7861 	}
7862 }
7863 
7864 static int sched_rt_global_validate(void)
7865 {
7866 	if (sysctl_sched_rt_period <= 0)
7867 		return -EINVAL;
7868 
7869 	if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7870 		(sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7871 		return -EINVAL;
7872 
7873 	return 0;
7874 }
7875 
7876 static void sched_rt_do_global(void)
7877 {
7878 	def_rt_bandwidth.rt_runtime = global_rt_runtime();
7879 	def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7880 }
7881 
7882 int sched_rt_handler(struct ctl_table *table, int write,
7883 		void __user *buffer, size_t *lenp,
7884 		loff_t *ppos)
7885 {
7886 	int old_period, old_runtime;
7887 	static DEFINE_MUTEX(mutex);
7888 	int ret;
7889 
7890 	mutex_lock(&mutex);
7891 	old_period = sysctl_sched_rt_period;
7892 	old_runtime = sysctl_sched_rt_runtime;
7893 
7894 	ret = proc_dointvec(table, write, buffer, lenp, ppos);
7895 
7896 	if (!ret && write) {
7897 		ret = sched_rt_global_validate();
7898 		if (ret)
7899 			goto undo;
7900 
7901 		ret = sched_dl_global_validate();
7902 		if (ret)
7903 			goto undo;
7904 
7905 		ret = sched_rt_global_constraints();
7906 		if (ret)
7907 			goto undo;
7908 
7909 		sched_rt_do_global();
7910 		sched_dl_do_global();
7911 	}
7912 	if (0) {
7913 undo:
7914 		sysctl_sched_rt_period = old_period;
7915 		sysctl_sched_rt_runtime = old_runtime;
7916 	}
7917 	mutex_unlock(&mutex);
7918 
7919 	return ret;
7920 }
7921 
7922 int sched_rr_handler(struct ctl_table *table, int write,
7923 		void __user *buffer, size_t *lenp,
7924 		loff_t *ppos)
7925 {
7926 	int ret;
7927 	static DEFINE_MUTEX(mutex);
7928 
7929 	mutex_lock(&mutex);
7930 	ret = proc_dointvec(table, write, buffer, lenp, ppos);
7931 	/* make sure that internally we keep jiffies */
7932 	/* also, writing zero resets timeslice to default */
7933 	if (!ret && write) {
7934 		sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7935 			RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7936 	}
7937 	mutex_unlock(&mutex);
7938 	return ret;
7939 }
7940 
7941 #ifdef CONFIG_CGROUP_SCHED
7942 
7943 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7944 {
7945 	return css ? container_of(css, struct task_group, css) : NULL;
7946 }
7947 
7948 static struct cgroup_subsys_state *
7949 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7950 {
7951 	struct task_group *parent = css_tg(parent_css);
7952 	struct task_group *tg;
7953 
7954 	if (!parent) {
7955 		/* This is early initialization for the top cgroup */
7956 		return &root_task_group.css;
7957 	}
7958 
7959 	tg = sched_create_group(parent);
7960 	if (IS_ERR(tg))
7961 		return ERR_PTR(-ENOMEM);
7962 
7963 	return &tg->css;
7964 }
7965 
7966 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7967 {
7968 	struct task_group *tg = css_tg(css);
7969 	struct task_group *parent = css_tg(css->parent);
7970 
7971 	if (parent)
7972 		sched_online_group(tg, parent);
7973 	return 0;
7974 }
7975 
7976 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7977 {
7978 	struct task_group *tg = css_tg(css);
7979 
7980 	sched_destroy_group(tg);
7981 }
7982 
7983 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7984 {
7985 	struct task_group *tg = css_tg(css);
7986 
7987 	sched_offline_group(tg);
7988 }
7989 
7990 static void cpu_cgroup_fork(struct task_struct *task)
7991 {
7992 	sched_move_task(task);
7993 }
7994 
7995 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7996 				 struct cgroup_taskset *tset)
7997 {
7998 	struct task_struct *task;
7999 
8000 	cgroup_taskset_for_each(task, tset) {
8001 #ifdef CONFIG_RT_GROUP_SCHED
8002 		if (!sched_rt_can_attach(css_tg(css), task))
8003 			return -EINVAL;
8004 #else
8005 		/* We don't support RT-tasks being in separate groups */
8006 		if (task->sched_class != &fair_sched_class)
8007 			return -EINVAL;
8008 #endif
8009 	}
8010 	return 0;
8011 }
8012 
8013 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8014 			      struct cgroup_taskset *tset)
8015 {
8016 	struct task_struct *task;
8017 
8018 	cgroup_taskset_for_each(task, tset)
8019 		sched_move_task(task);
8020 }
8021 
8022 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8023 			    struct cgroup_subsys_state *old_css,
8024 			    struct task_struct *task)
8025 {
8026 	/*
8027 	 * cgroup_exit() is called in the copy_process() failure path.
8028 	 * Ignore this case since the task hasn't ran yet, this avoids
8029 	 * trying to poke a half freed task state from generic code.
8030 	 */
8031 	if (!(task->flags & PF_EXITING))
8032 		return;
8033 
8034 	sched_move_task(task);
8035 }
8036 
8037 #ifdef CONFIG_FAIR_GROUP_SCHED
8038 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8039 				struct cftype *cftype, u64 shareval)
8040 {
8041 	return sched_group_set_shares(css_tg(css), scale_load(shareval));
8042 }
8043 
8044 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8045 			       struct cftype *cft)
8046 {
8047 	struct task_group *tg = css_tg(css);
8048 
8049 	return (u64) scale_load_down(tg->shares);
8050 }
8051 
8052 #ifdef CONFIG_CFS_BANDWIDTH
8053 static DEFINE_MUTEX(cfs_constraints_mutex);
8054 
8055 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8056 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8057 
8058 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8059 
8060 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8061 {
8062 	int i, ret = 0, runtime_enabled, runtime_was_enabled;
8063 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8064 
8065 	if (tg == &root_task_group)
8066 		return -EINVAL;
8067 
8068 	/*
8069 	 * Ensure we have at some amount of bandwidth every period.  This is
8070 	 * to prevent reaching a state of large arrears when throttled via
8071 	 * entity_tick() resulting in prolonged exit starvation.
8072 	 */
8073 	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8074 		return -EINVAL;
8075 
8076 	/*
8077 	 * Likewise, bound things on the otherside by preventing insane quota
8078 	 * periods.  This also allows us to normalize in computing quota
8079 	 * feasibility.
8080 	 */
8081 	if (period > max_cfs_quota_period)
8082 		return -EINVAL;
8083 
8084 	/*
8085 	 * Prevent race between setting of cfs_rq->runtime_enabled and
8086 	 * unthrottle_offline_cfs_rqs().
8087 	 */
8088 	get_online_cpus();
8089 	mutex_lock(&cfs_constraints_mutex);
8090 	ret = __cfs_schedulable(tg, period, quota);
8091 	if (ret)
8092 		goto out_unlock;
8093 
8094 	runtime_enabled = quota != RUNTIME_INF;
8095 	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8096 	/*
8097 	 * If we need to toggle cfs_bandwidth_used, off->on must occur
8098 	 * before making related changes, and on->off must occur afterwards
8099 	 */
8100 	if (runtime_enabled && !runtime_was_enabled)
8101 		cfs_bandwidth_usage_inc();
8102 	raw_spin_lock_irq(&cfs_b->lock);
8103 	cfs_b->period = ns_to_ktime(period);
8104 	cfs_b->quota = quota;
8105 
8106 	__refill_cfs_bandwidth_runtime(cfs_b);
8107 	/* restart the period timer (if active) to handle new period expiry */
8108 	if (runtime_enabled && cfs_b->timer_active) {
8109 		/* force a reprogram */
8110 		__start_cfs_bandwidth(cfs_b, true);
8111 	}
8112 	raw_spin_unlock_irq(&cfs_b->lock);
8113 
8114 	for_each_online_cpu(i) {
8115 		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8116 		struct rq *rq = cfs_rq->rq;
8117 
8118 		raw_spin_lock_irq(&rq->lock);
8119 		cfs_rq->runtime_enabled = runtime_enabled;
8120 		cfs_rq->runtime_remaining = 0;
8121 
8122 		if (cfs_rq->throttled)
8123 			unthrottle_cfs_rq(cfs_rq);
8124 		raw_spin_unlock_irq(&rq->lock);
8125 	}
8126 	if (runtime_was_enabled && !runtime_enabled)
8127 		cfs_bandwidth_usage_dec();
8128 out_unlock:
8129 	mutex_unlock(&cfs_constraints_mutex);
8130 	put_online_cpus();
8131 
8132 	return ret;
8133 }
8134 
8135 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8136 {
8137 	u64 quota, period;
8138 
8139 	period = ktime_to_ns(tg->cfs_bandwidth.period);
8140 	if (cfs_quota_us < 0)
8141 		quota = RUNTIME_INF;
8142 	else
8143 		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8144 
8145 	return tg_set_cfs_bandwidth(tg, period, quota);
8146 }
8147 
8148 long tg_get_cfs_quota(struct task_group *tg)
8149 {
8150 	u64 quota_us;
8151 
8152 	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8153 		return -1;
8154 
8155 	quota_us = tg->cfs_bandwidth.quota;
8156 	do_div(quota_us, NSEC_PER_USEC);
8157 
8158 	return quota_us;
8159 }
8160 
8161 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8162 {
8163 	u64 quota, period;
8164 
8165 	period = (u64)cfs_period_us * NSEC_PER_USEC;
8166 	quota = tg->cfs_bandwidth.quota;
8167 
8168 	return tg_set_cfs_bandwidth(tg, period, quota);
8169 }
8170 
8171 long tg_get_cfs_period(struct task_group *tg)
8172 {
8173 	u64 cfs_period_us;
8174 
8175 	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8176 	do_div(cfs_period_us, NSEC_PER_USEC);
8177 
8178 	return cfs_period_us;
8179 }
8180 
8181 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8182 				  struct cftype *cft)
8183 {
8184 	return tg_get_cfs_quota(css_tg(css));
8185 }
8186 
8187 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8188 				   struct cftype *cftype, s64 cfs_quota_us)
8189 {
8190 	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8191 }
8192 
8193 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8194 				   struct cftype *cft)
8195 {
8196 	return tg_get_cfs_period(css_tg(css));
8197 }
8198 
8199 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8200 				    struct cftype *cftype, u64 cfs_period_us)
8201 {
8202 	return tg_set_cfs_period(css_tg(css), cfs_period_us);
8203 }
8204 
8205 struct cfs_schedulable_data {
8206 	struct task_group *tg;
8207 	u64 period, quota;
8208 };
8209 
8210 /*
8211  * normalize group quota/period to be quota/max_period
8212  * note: units are usecs
8213  */
8214 static u64 normalize_cfs_quota(struct task_group *tg,
8215 			       struct cfs_schedulable_data *d)
8216 {
8217 	u64 quota, period;
8218 
8219 	if (tg == d->tg) {
8220 		period = d->period;
8221 		quota = d->quota;
8222 	} else {
8223 		period = tg_get_cfs_period(tg);
8224 		quota = tg_get_cfs_quota(tg);
8225 	}
8226 
8227 	/* note: these should typically be equivalent */
8228 	if (quota == RUNTIME_INF || quota == -1)
8229 		return RUNTIME_INF;
8230 
8231 	return to_ratio(period, quota);
8232 }
8233 
8234 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8235 {
8236 	struct cfs_schedulable_data *d = data;
8237 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8238 	s64 quota = 0, parent_quota = -1;
8239 
8240 	if (!tg->parent) {
8241 		quota = RUNTIME_INF;
8242 	} else {
8243 		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8244 
8245 		quota = normalize_cfs_quota(tg, d);
8246 		parent_quota = parent_b->hierarchical_quota;
8247 
8248 		/*
8249 		 * ensure max(child_quota) <= parent_quota, inherit when no
8250 		 * limit is set
8251 		 */
8252 		if (quota == RUNTIME_INF)
8253 			quota = parent_quota;
8254 		else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8255 			return -EINVAL;
8256 	}
8257 	cfs_b->hierarchical_quota = quota;
8258 
8259 	return 0;
8260 }
8261 
8262 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8263 {
8264 	int ret;
8265 	struct cfs_schedulable_data data = {
8266 		.tg = tg,
8267 		.period = period,
8268 		.quota = quota,
8269 	};
8270 
8271 	if (quota != RUNTIME_INF) {
8272 		do_div(data.period, NSEC_PER_USEC);
8273 		do_div(data.quota, NSEC_PER_USEC);
8274 	}
8275 
8276 	rcu_read_lock();
8277 	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8278 	rcu_read_unlock();
8279 
8280 	return ret;
8281 }
8282 
8283 static int cpu_stats_show(struct seq_file *sf, void *v)
8284 {
8285 	struct task_group *tg = css_tg(seq_css(sf));
8286 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8287 
8288 	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8289 	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8290 	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8291 
8292 	return 0;
8293 }
8294 #endif /* CONFIG_CFS_BANDWIDTH */
8295 #endif /* CONFIG_FAIR_GROUP_SCHED */
8296 
8297 #ifdef CONFIG_RT_GROUP_SCHED
8298 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8299 				struct cftype *cft, s64 val)
8300 {
8301 	return sched_group_set_rt_runtime(css_tg(css), val);
8302 }
8303 
8304 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8305 			       struct cftype *cft)
8306 {
8307 	return sched_group_rt_runtime(css_tg(css));
8308 }
8309 
8310 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8311 				    struct cftype *cftype, u64 rt_period_us)
8312 {
8313 	return sched_group_set_rt_period(css_tg(css), rt_period_us);
8314 }
8315 
8316 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8317 				   struct cftype *cft)
8318 {
8319 	return sched_group_rt_period(css_tg(css));
8320 }
8321 #endif /* CONFIG_RT_GROUP_SCHED */
8322 
8323 static struct cftype cpu_files[] = {
8324 #ifdef CONFIG_FAIR_GROUP_SCHED
8325 	{
8326 		.name = "shares",
8327 		.read_u64 = cpu_shares_read_u64,
8328 		.write_u64 = cpu_shares_write_u64,
8329 	},
8330 #endif
8331 #ifdef CONFIG_CFS_BANDWIDTH
8332 	{
8333 		.name = "cfs_quota_us",
8334 		.read_s64 = cpu_cfs_quota_read_s64,
8335 		.write_s64 = cpu_cfs_quota_write_s64,
8336 	},
8337 	{
8338 		.name = "cfs_period_us",
8339 		.read_u64 = cpu_cfs_period_read_u64,
8340 		.write_u64 = cpu_cfs_period_write_u64,
8341 	},
8342 	{
8343 		.name = "stat",
8344 		.seq_show = cpu_stats_show,
8345 	},
8346 #endif
8347 #ifdef CONFIG_RT_GROUP_SCHED
8348 	{
8349 		.name = "rt_runtime_us",
8350 		.read_s64 = cpu_rt_runtime_read,
8351 		.write_s64 = cpu_rt_runtime_write,
8352 	},
8353 	{
8354 		.name = "rt_period_us",
8355 		.read_u64 = cpu_rt_period_read_uint,
8356 		.write_u64 = cpu_rt_period_write_uint,
8357 	},
8358 #endif
8359 	{ }	/* terminate */
8360 };
8361 
8362 struct cgroup_subsys cpu_cgrp_subsys = {
8363 	.css_alloc	= cpu_cgroup_css_alloc,
8364 	.css_free	= cpu_cgroup_css_free,
8365 	.css_online	= cpu_cgroup_css_online,
8366 	.css_offline	= cpu_cgroup_css_offline,
8367 	.fork		= cpu_cgroup_fork,
8368 	.can_attach	= cpu_cgroup_can_attach,
8369 	.attach		= cpu_cgroup_attach,
8370 	.exit		= cpu_cgroup_exit,
8371 	.legacy_cftypes	= cpu_files,
8372 	.early_init	= 1,
8373 };
8374 
8375 #endif	/* CONFIG_CGROUP_SCHED */
8376 
8377 void dump_cpu_task(int cpu)
8378 {
8379 	pr_info("Task dump for CPU %d:\n", cpu);
8380 	sched_show_task(cpu_curr(cpu));
8381 }
8382