core.c revision 52f5684c8e1ec7463192aba8e2916df49807511a
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
93void 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
113DEFINE_MUTEX(sched_domains_mutex);
114DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115
116static void update_rq_clock_task(struct rq *rq, s64 delta);
117
118void update_rq_clock(struct rq *rq)
119{
120	s64 delta;
121
122	if (rq->skip_clock_update > 0)
123		return;
124
125	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126	rq->clock += delta;
127	update_rq_clock_task(rq, delta);
128}
129
130/*
131 * Debugging: various feature bits
132 */
133
134#define SCHED_FEAT(name, enabled)	\
135	(1UL << __SCHED_FEAT_##name) * enabled |
136
137const_debug unsigned int sysctl_sched_features =
138#include "features.h"
139	0;
140
141#undef SCHED_FEAT
142
143#ifdef CONFIG_SCHED_DEBUG
144#define SCHED_FEAT(name, enabled)	\
145	#name ,
146
147static const char * const sched_feat_names[] = {
148#include "features.h"
149};
150
151#undef SCHED_FEAT
152
153static int sched_feat_show(struct seq_file *m, void *v)
154{
155	int i;
156
157	for (i = 0; i < __SCHED_FEAT_NR; i++) {
158		if (!(sysctl_sched_features & (1UL << i)))
159			seq_puts(m, "NO_");
160		seq_printf(m, "%s ", sched_feat_names[i]);
161	}
162	seq_puts(m, "\n");
163
164	return 0;
165}
166
167#ifdef HAVE_JUMP_LABEL
168
169#define jump_label_key__true  STATIC_KEY_INIT_TRUE
170#define jump_label_key__false STATIC_KEY_INIT_FALSE
171
172#define SCHED_FEAT(name, enabled)	\
173	jump_label_key__##enabled ,
174
175struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
176#include "features.h"
177};
178
179#undef SCHED_FEAT
180
181static void sched_feat_disable(int i)
182{
183	if (static_key_enabled(&sched_feat_keys[i]))
184		static_key_slow_dec(&sched_feat_keys[i]);
185}
186
187static void sched_feat_enable(int i)
188{
189	if (!static_key_enabled(&sched_feat_keys[i]))
190		static_key_slow_inc(&sched_feat_keys[i]);
191}
192#else
193static void sched_feat_disable(int i) { };
194static void sched_feat_enable(int i) { };
195#endif /* HAVE_JUMP_LABEL */
196
197static int sched_feat_set(char *cmp)
198{
199	int i;
200	int neg = 0;
201
202	if (strncmp(cmp, "NO_", 3) == 0) {
203		neg = 1;
204		cmp += 3;
205	}
206
207	for (i = 0; i < __SCHED_FEAT_NR; i++) {
208		if (strcmp(cmp, sched_feat_names[i]) == 0) {
209			if (neg) {
210				sysctl_sched_features &= ~(1UL << i);
211				sched_feat_disable(i);
212			} else {
213				sysctl_sched_features |= (1UL << i);
214				sched_feat_enable(i);
215			}
216			break;
217		}
218	}
219
220	return i;
221}
222
223static ssize_t
224sched_feat_write(struct file *filp, const char __user *ubuf,
225		size_t cnt, loff_t *ppos)
226{
227	char buf[64];
228	char *cmp;
229	int i;
230
231	if (cnt > 63)
232		cnt = 63;
233
234	if (copy_from_user(&buf, ubuf, cnt))
235		return -EFAULT;
236
237	buf[cnt] = 0;
238	cmp = strstrip(buf);
239
240	i = sched_feat_set(cmp);
241	if (i == __SCHED_FEAT_NR)
242		return -EINVAL;
243
244	*ppos += cnt;
245
246	return cnt;
247}
248
249static int sched_feat_open(struct inode *inode, struct file *filp)
250{
251	return single_open(filp, sched_feat_show, NULL);
252}
253
254static const struct file_operations sched_feat_fops = {
255	.open		= sched_feat_open,
256	.write		= sched_feat_write,
257	.read		= seq_read,
258	.llseek		= seq_lseek,
259	.release	= single_release,
260};
261
262static __init int sched_init_debug(void)
263{
264	debugfs_create_file("sched_features", 0644, NULL, NULL,
265			&sched_feat_fops);
266
267	return 0;
268}
269late_initcall(sched_init_debug);
270#endif /* CONFIG_SCHED_DEBUG */
271
272/*
273 * Number of tasks to iterate in a single balance run.
274 * Limited because this is done with IRQs disabled.
275 */
276const_debug unsigned int sysctl_sched_nr_migrate = 32;
277
278/*
279 * period over which we average the RT time consumption, measured
280 * in ms.
281 *
282 * default: 1s
283 */
284const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
285
286/*
287 * period over which we measure -rt task cpu usage in us.
288 * default: 1s
289 */
290unsigned int sysctl_sched_rt_period = 1000000;
291
292__read_mostly int scheduler_running;
293
294/*
295 * part of the period that we allow rt tasks to run in us.
296 * default: 0.95s
297 */
298int sysctl_sched_rt_runtime = 950000;
299
300/*
301 * __task_rq_lock - lock the rq @p resides on.
302 */
303static inline struct rq *__task_rq_lock(struct task_struct *p)
304	__acquires(rq->lock)
305{
306	struct rq *rq;
307
308	lockdep_assert_held(&p->pi_lock);
309
310	for (;;) {
311		rq = task_rq(p);
312		raw_spin_lock(&rq->lock);
313		if (likely(rq == task_rq(p)))
314			return rq;
315		raw_spin_unlock(&rq->lock);
316	}
317}
318
319/*
320 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
321 */
322static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
323	__acquires(p->pi_lock)
324	__acquires(rq->lock)
325{
326	struct rq *rq;
327
328	for (;;) {
329		raw_spin_lock_irqsave(&p->pi_lock, *flags);
330		rq = task_rq(p);
331		raw_spin_lock(&rq->lock);
332		if (likely(rq == task_rq(p)))
333			return rq;
334		raw_spin_unlock(&rq->lock);
335		raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
336	}
337}
338
339static void __task_rq_unlock(struct rq *rq)
340	__releases(rq->lock)
341{
342	raw_spin_unlock(&rq->lock);
343}
344
345static inline void
346task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
347	__releases(rq->lock)
348	__releases(p->pi_lock)
349{
350	raw_spin_unlock(&rq->lock);
351	raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
352}
353
354/*
355 * this_rq_lock - lock this runqueue and disable interrupts.
356 */
357static struct rq *this_rq_lock(void)
358	__acquires(rq->lock)
359{
360	struct rq *rq;
361
362	local_irq_disable();
363	rq = this_rq();
364	raw_spin_lock(&rq->lock);
365
366	return rq;
367}
368
369#ifdef CONFIG_SCHED_HRTICK
370/*
371 * Use HR-timers to deliver accurate preemption points.
372 */
373
374static void hrtick_clear(struct rq *rq)
375{
376	if (hrtimer_active(&rq->hrtick_timer))
377		hrtimer_cancel(&rq->hrtick_timer);
378}
379
380/*
381 * High-resolution timer tick.
382 * Runs from hardirq context with interrupts disabled.
383 */
384static enum hrtimer_restart hrtick(struct hrtimer *timer)
385{
386	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
387
388	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
389
390	raw_spin_lock(&rq->lock);
391	update_rq_clock(rq);
392	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
393	raw_spin_unlock(&rq->lock);
394
395	return HRTIMER_NORESTART;
396}
397
398#ifdef CONFIG_SMP
399
400static int __hrtick_restart(struct rq *rq)
401{
402	struct hrtimer *timer = &rq->hrtick_timer;
403	ktime_t time = hrtimer_get_softexpires(timer);
404
405	return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
406}
407
408/*
409 * called from hardirq (IPI) context
410 */
411static void __hrtick_start(void *arg)
412{
413	struct rq *rq = arg;
414
415	raw_spin_lock(&rq->lock);
416	__hrtick_restart(rq);
417	rq->hrtick_csd_pending = 0;
418	raw_spin_unlock(&rq->lock);
419}
420
421/*
422 * Called to set the hrtick timer state.
423 *
424 * called with rq->lock held and irqs disabled
425 */
426void hrtick_start(struct rq *rq, u64 delay)
427{
428	struct hrtimer *timer = &rq->hrtick_timer;
429	ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430
431	hrtimer_set_expires(timer, time);
432
433	if (rq == this_rq()) {
434		__hrtick_restart(rq);
435	} else if (!rq->hrtick_csd_pending) {
436		smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
437		rq->hrtick_csd_pending = 1;
438	}
439}
440
441static int
442hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443{
444	int cpu = (int)(long)hcpu;
445
446	switch (action) {
447	case CPU_UP_CANCELED:
448	case CPU_UP_CANCELED_FROZEN:
449	case CPU_DOWN_PREPARE:
450	case CPU_DOWN_PREPARE_FROZEN:
451	case CPU_DEAD:
452	case CPU_DEAD_FROZEN:
453		hrtick_clear(cpu_rq(cpu));
454		return NOTIFY_OK;
455	}
456
457	return NOTIFY_DONE;
458}
459
460static __init void init_hrtick(void)
461{
462	hotcpu_notifier(hotplug_hrtick, 0);
463}
464#else
465/*
466 * Called to set the hrtick timer state.
467 *
468 * called with rq->lock held and irqs disabled
469 */
470void hrtick_start(struct rq *rq, u64 delay)
471{
472	__hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
473			HRTIMER_MODE_REL_PINNED, 0);
474}
475
476static inline void init_hrtick(void)
477{
478}
479#endif /* CONFIG_SMP */
480
481static void init_rq_hrtick(struct rq *rq)
482{
483#ifdef CONFIG_SMP
484	rq->hrtick_csd_pending = 0;
485
486	rq->hrtick_csd.flags = 0;
487	rq->hrtick_csd.func = __hrtick_start;
488	rq->hrtick_csd.info = rq;
489#endif
490
491	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
492	rq->hrtick_timer.function = hrtick;
493}
494#else	/* CONFIG_SCHED_HRTICK */
495static inline void hrtick_clear(struct rq *rq)
496{
497}
498
499static inline void init_rq_hrtick(struct rq *rq)
500{
501}
502
503static inline void init_hrtick(void)
504{
505}
506#endif	/* CONFIG_SCHED_HRTICK */
507
508/*
509 * resched_task - mark a task 'to be rescheduled now'.
510 *
511 * On UP this means the setting of the need_resched flag, on SMP it
512 * might also involve a cross-CPU call to trigger the scheduler on
513 * the target CPU.
514 */
515void resched_task(struct task_struct *p)
516{
517	int cpu;
518
519	lockdep_assert_held(&task_rq(p)->lock);
520
521	if (test_tsk_need_resched(p))
522		return;
523
524	set_tsk_need_resched(p);
525
526	cpu = task_cpu(p);
527	if (cpu == smp_processor_id()) {
528		set_preempt_need_resched();
529		return;
530	}
531
532	/* NEED_RESCHED must be visible before we test polling */
533	smp_mb();
534	if (!tsk_is_polling(p))
535		smp_send_reschedule(cpu);
536}
537
538void resched_cpu(int cpu)
539{
540	struct rq *rq = cpu_rq(cpu);
541	unsigned long flags;
542
543	if (!raw_spin_trylock_irqsave(&rq->lock, flags))
544		return;
545	resched_task(cpu_curr(cpu));
546	raw_spin_unlock_irqrestore(&rq->lock, flags);
547}
548
549#ifdef CONFIG_SMP
550#ifdef CONFIG_NO_HZ_COMMON
551/*
552 * In the semi idle case, use the nearest busy cpu for migrating timers
553 * from an idle cpu.  This is good for power-savings.
554 *
555 * We don't do similar optimization for completely idle system, as
556 * selecting an idle cpu will add more delays to the timers than intended
557 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
558 */
559int get_nohz_timer_target(int pinned)
560{
561	int cpu = smp_processor_id();
562	int i;
563	struct sched_domain *sd;
564
565	if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
566		return cpu;
567
568	rcu_read_lock();
569	for_each_domain(cpu, sd) {
570		for_each_cpu(i, sched_domain_span(sd)) {
571			if (!idle_cpu(i)) {
572				cpu = i;
573				goto unlock;
574			}
575		}
576	}
577unlock:
578	rcu_read_unlock();
579	return cpu;
580}
581/*
582 * When add_timer_on() enqueues a timer into the timer wheel of an
583 * idle CPU then this timer might expire before the next timer event
584 * which is scheduled to wake up that CPU. In case of a completely
585 * idle system the next event might even be infinite time into the
586 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
587 * leaves the inner idle loop so the newly added timer is taken into
588 * account when the CPU goes back to idle and evaluates the timer
589 * wheel for the next timer event.
590 */
591static void wake_up_idle_cpu(int cpu)
592{
593	struct rq *rq = cpu_rq(cpu);
594
595	if (cpu == smp_processor_id())
596		return;
597
598	/*
599	 * This is safe, as this function is called with the timer
600	 * wheel base lock of (cpu) held. When the CPU is on the way
601	 * to idle and has not yet set rq->curr to idle then it will
602	 * be serialized on the timer wheel base lock and take the new
603	 * timer into account automatically.
604	 */
605	if (rq->curr != rq->idle)
606		return;
607
608	/*
609	 * We can set TIF_RESCHED on the idle task of the other CPU
610	 * lockless. The worst case is that the other CPU runs the
611	 * idle task through an additional NOOP schedule()
612	 */
613	set_tsk_need_resched(rq->idle);
614
615	/* NEED_RESCHED must be visible before we test polling */
616	smp_mb();
617	if (!tsk_is_polling(rq->idle))
618		smp_send_reschedule(cpu);
619}
620
621static bool wake_up_full_nohz_cpu(int cpu)
622{
623	if (tick_nohz_full_cpu(cpu)) {
624		if (cpu != smp_processor_id() ||
625		    tick_nohz_tick_stopped())
626			smp_send_reschedule(cpu);
627		return true;
628	}
629
630	return false;
631}
632
633void wake_up_nohz_cpu(int cpu)
634{
635	if (!wake_up_full_nohz_cpu(cpu))
636		wake_up_idle_cpu(cpu);
637}
638
639static inline bool got_nohz_idle_kick(void)
640{
641	int cpu = smp_processor_id();
642
643	if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
644		return false;
645
646	if (idle_cpu(cpu) && !need_resched())
647		return true;
648
649	/*
650	 * We can't run Idle Load Balance on this CPU for this time so we
651	 * cancel it and clear NOHZ_BALANCE_KICK
652	 */
653	clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
654	return false;
655}
656
657#else /* CONFIG_NO_HZ_COMMON */
658
659static inline bool got_nohz_idle_kick(void)
660{
661	return false;
662}
663
664#endif /* CONFIG_NO_HZ_COMMON */
665
666#ifdef CONFIG_NO_HZ_FULL
667bool sched_can_stop_tick(void)
668{
669       struct rq *rq;
670
671       rq = this_rq();
672
673       /* Make sure rq->nr_running update is visible after the IPI */
674       smp_rmb();
675
676       /* More than one running task need preemption */
677       if (rq->nr_running > 1)
678               return false;
679
680       return true;
681}
682#endif /* CONFIG_NO_HZ_FULL */
683
684void sched_avg_update(struct rq *rq)
685{
686	s64 period = sched_avg_period();
687
688	while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
689		/*
690		 * Inline assembly required to prevent the compiler
691		 * optimising this loop into a divmod call.
692		 * See __iter_div_u64_rem() for another example of this.
693		 */
694		asm("" : "+rm" (rq->age_stamp));
695		rq->age_stamp += period;
696		rq->rt_avg /= 2;
697	}
698}
699
700#endif /* CONFIG_SMP */
701
702#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
704/*
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
707 *
708 * Caller must hold rcu_lock or sufficient equivalent.
709 */
710int walk_tg_tree_from(struct task_group *from,
711			     tg_visitor down, tg_visitor up, void *data)
712{
713	struct task_group *parent, *child;
714	int ret;
715
716	parent = from;
717
718down:
719	ret = (*down)(parent, data);
720	if (ret)
721		goto out;
722	list_for_each_entry_rcu(child, &parent->children, siblings) {
723		parent = child;
724		goto down;
725
726up:
727		continue;
728	}
729	ret = (*up)(parent, data);
730	if (ret || parent == from)
731		goto out;
732
733	child = parent;
734	parent = parent->parent;
735	if (parent)
736		goto up;
737out:
738	return ret;
739}
740
741int tg_nop(struct task_group *tg, void *data)
742{
743	return 0;
744}
745#endif
746
747static void set_load_weight(struct task_struct *p)
748{
749	int prio = p->static_prio - MAX_RT_PRIO;
750	struct load_weight *load = &p->se.load;
751
752	/*
753	 * SCHED_IDLE tasks get minimal weight:
754	 */
755	if (p->policy == SCHED_IDLE) {
756		load->weight = scale_load(WEIGHT_IDLEPRIO);
757		load->inv_weight = WMULT_IDLEPRIO;
758		return;
759	}
760
761	load->weight = scale_load(prio_to_weight[prio]);
762	load->inv_weight = prio_to_wmult[prio];
763}
764
765static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
766{
767	update_rq_clock(rq);
768	sched_info_queued(rq, p);
769	p->sched_class->enqueue_task(rq, p, flags);
770}
771
772static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
773{
774	update_rq_clock(rq);
775	sched_info_dequeued(rq, p);
776	p->sched_class->dequeue_task(rq, p, flags);
777}
778
779void activate_task(struct rq *rq, struct task_struct *p, int flags)
780{
781	if (task_contributes_to_load(p))
782		rq->nr_uninterruptible--;
783
784	enqueue_task(rq, p, flags);
785}
786
787void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
788{
789	if (task_contributes_to_load(p))
790		rq->nr_uninterruptible++;
791
792	dequeue_task(rq, p, flags);
793}
794
795static void update_rq_clock_task(struct rq *rq, s64 delta)
796{
797/*
798 * In theory, the compile should just see 0 here, and optimize out the call
799 * to sched_rt_avg_update. But I don't trust it...
800 */
801#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802	s64 steal = 0, irq_delta = 0;
803#endif
804#ifdef CONFIG_IRQ_TIME_ACCOUNTING
805	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
806
807	/*
808	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
809	 * this case when a previous update_rq_clock() happened inside a
810	 * {soft,}irq region.
811	 *
812	 * When this happens, we stop ->clock_task and only update the
813	 * prev_irq_time stamp to account for the part that fit, so that a next
814	 * update will consume the rest. This ensures ->clock_task is
815	 * monotonic.
816	 *
817	 * It does however cause some slight miss-attribution of {soft,}irq
818	 * time, a more accurate solution would be to update the irq_time using
819	 * the current rq->clock timestamp, except that would require using
820	 * atomic ops.
821	 */
822	if (irq_delta > delta)
823		irq_delta = delta;
824
825	rq->prev_irq_time += irq_delta;
826	delta -= irq_delta;
827#endif
828#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
829	if (static_key_false((&paravirt_steal_rq_enabled))) {
830		steal = paravirt_steal_clock(cpu_of(rq));
831		steal -= rq->prev_steal_time_rq;
832
833		if (unlikely(steal > delta))
834			steal = delta;
835
836		rq->prev_steal_time_rq += steal;
837		delta -= steal;
838	}
839#endif
840
841	rq->clock_task += delta;
842
843#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
844	if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
845		sched_rt_avg_update(rq, irq_delta + steal);
846#endif
847}
848
849void sched_set_stop_task(int cpu, struct task_struct *stop)
850{
851	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
852	struct task_struct *old_stop = cpu_rq(cpu)->stop;
853
854	if (stop) {
855		/*
856		 * Make it appear like a SCHED_FIFO task, its something
857		 * userspace knows about and won't get confused about.
858		 *
859		 * Also, it will make PI more or less work without too
860		 * much confusion -- but then, stop work should not
861		 * rely on PI working anyway.
862		 */
863		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
864
865		stop->sched_class = &stop_sched_class;
866	}
867
868	cpu_rq(cpu)->stop = stop;
869
870	if (old_stop) {
871		/*
872		 * Reset it back to a normal scheduling class so that
873		 * it can die in pieces.
874		 */
875		old_stop->sched_class = &rt_sched_class;
876	}
877}
878
879/*
880 * __normal_prio - return the priority that is based on the static prio
881 */
882static inline int __normal_prio(struct task_struct *p)
883{
884	return p->static_prio;
885}
886
887/*
888 * Calculate the expected normal priority: i.e. priority
889 * without taking RT-inheritance into account. Might be
890 * boosted by interactivity modifiers. Changes upon fork,
891 * setprio syscalls, and whenever the interactivity
892 * estimator recalculates.
893 */
894static inline int normal_prio(struct task_struct *p)
895{
896	int prio;
897
898	if (task_has_dl_policy(p))
899		prio = MAX_DL_PRIO-1;
900	else if (task_has_rt_policy(p))
901		prio = MAX_RT_PRIO-1 - p->rt_priority;
902	else
903		prio = __normal_prio(p);
904	return prio;
905}
906
907/*
908 * Calculate the current priority, i.e. the priority
909 * taken into account by the scheduler. This value might
910 * be boosted by RT tasks, or might be boosted by
911 * interactivity modifiers. Will be RT if the task got
912 * RT-boosted. If not then it returns p->normal_prio.
913 */
914static int effective_prio(struct task_struct *p)
915{
916	p->normal_prio = normal_prio(p);
917	/*
918	 * If we are RT tasks or we were boosted to RT priority,
919	 * keep the priority unchanged. Otherwise, update priority
920	 * to the normal priority:
921	 */
922	if (!rt_prio(p->prio))
923		return p->normal_prio;
924	return p->prio;
925}
926
927/**
928 * task_curr - is this task currently executing on a CPU?
929 * @p: the task in question.
930 *
931 * Return: 1 if the task is currently executing. 0 otherwise.
932 */
933inline int task_curr(const struct task_struct *p)
934{
935	return cpu_curr(task_cpu(p)) == p;
936}
937
938static inline void check_class_changed(struct rq *rq, struct task_struct *p,
939				       const struct sched_class *prev_class,
940				       int oldprio)
941{
942	if (prev_class != p->sched_class) {
943		if (prev_class->switched_from)
944			prev_class->switched_from(rq, p);
945		p->sched_class->switched_to(rq, p);
946	} else if (oldprio != p->prio || dl_task(p))
947		p->sched_class->prio_changed(rq, p, oldprio);
948}
949
950void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
951{
952	const struct sched_class *class;
953
954	if (p->sched_class == rq->curr->sched_class) {
955		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
956	} else {
957		for_each_class(class) {
958			if (class == rq->curr->sched_class)
959				break;
960			if (class == p->sched_class) {
961				resched_task(rq->curr);
962				break;
963			}
964		}
965	}
966
967	/*
968	 * A queue event has occurred, and we're going to schedule.  In
969	 * this case, we can save a useless back to back clock update.
970	 */
971	if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
972		rq->skip_clock_update = 1;
973}
974
975#ifdef CONFIG_SMP
976void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
977{
978#ifdef CONFIG_SCHED_DEBUG
979	/*
980	 * We should never call set_task_cpu() on a blocked task,
981	 * ttwu() will sort out the placement.
982	 */
983	WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
984			!(task_preempt_count(p) & PREEMPT_ACTIVE));
985
986#ifdef CONFIG_LOCKDEP
987	/*
988	 * The caller should hold either p->pi_lock or rq->lock, when changing
989	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
990	 *
991	 * sched_move_task() holds both and thus holding either pins the cgroup,
992	 * see task_group().
993	 *
994	 * Furthermore, all task_rq users should acquire both locks, see
995	 * task_rq_lock().
996	 */
997	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
998				      lockdep_is_held(&task_rq(p)->lock)));
999#endif
1000#endif
1001
1002	trace_sched_migrate_task(p, new_cpu);
1003
1004	if (task_cpu(p) != new_cpu) {
1005		if (p->sched_class->migrate_task_rq)
1006			p->sched_class->migrate_task_rq(p, new_cpu);
1007		p->se.nr_migrations++;
1008		perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1009	}
1010
1011	__set_task_cpu(p, new_cpu);
1012}
1013
1014static void __migrate_swap_task(struct task_struct *p, int cpu)
1015{
1016	if (p->on_rq) {
1017		struct rq *src_rq, *dst_rq;
1018
1019		src_rq = task_rq(p);
1020		dst_rq = cpu_rq(cpu);
1021
1022		deactivate_task(src_rq, p, 0);
1023		set_task_cpu(p, cpu);
1024		activate_task(dst_rq, p, 0);
1025		check_preempt_curr(dst_rq, p, 0);
1026	} else {
1027		/*
1028		 * Task isn't running anymore; make it appear like we migrated
1029		 * it before it went to sleep. This means on wakeup we make the
1030		 * previous cpu our targer instead of where it really is.
1031		 */
1032		p->wake_cpu = cpu;
1033	}
1034}
1035
1036struct migration_swap_arg {
1037	struct task_struct *src_task, *dst_task;
1038	int src_cpu, dst_cpu;
1039};
1040
1041static int migrate_swap_stop(void *data)
1042{
1043	struct migration_swap_arg *arg = data;
1044	struct rq *src_rq, *dst_rq;
1045	int ret = -EAGAIN;
1046
1047	src_rq = cpu_rq(arg->src_cpu);
1048	dst_rq = cpu_rq(arg->dst_cpu);
1049
1050	double_raw_lock(&arg->src_task->pi_lock,
1051			&arg->dst_task->pi_lock);
1052	double_rq_lock(src_rq, dst_rq);
1053	if (task_cpu(arg->dst_task) != arg->dst_cpu)
1054		goto unlock;
1055
1056	if (task_cpu(arg->src_task) != arg->src_cpu)
1057		goto unlock;
1058
1059	if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1060		goto unlock;
1061
1062	if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1063		goto unlock;
1064
1065	__migrate_swap_task(arg->src_task, arg->dst_cpu);
1066	__migrate_swap_task(arg->dst_task, arg->src_cpu);
1067
1068	ret = 0;
1069
1070unlock:
1071	double_rq_unlock(src_rq, dst_rq);
1072	raw_spin_unlock(&arg->dst_task->pi_lock);
1073	raw_spin_unlock(&arg->src_task->pi_lock);
1074
1075	return ret;
1076}
1077
1078/*
1079 * Cross migrate two tasks
1080 */
1081int migrate_swap(struct task_struct *cur, struct task_struct *p)
1082{
1083	struct migration_swap_arg arg;
1084	int ret = -EINVAL;
1085
1086	arg = (struct migration_swap_arg){
1087		.src_task = cur,
1088		.src_cpu = task_cpu(cur),
1089		.dst_task = p,
1090		.dst_cpu = task_cpu(p),
1091	};
1092
1093	if (arg.src_cpu == arg.dst_cpu)
1094		goto out;
1095
1096	/*
1097	 * These three tests are all lockless; this is OK since all of them
1098	 * will be re-checked with proper locks held further down the line.
1099	 */
1100	if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1101		goto out;
1102
1103	if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1104		goto out;
1105
1106	if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1107		goto out;
1108
1109	trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1110	ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1111
1112out:
1113	return ret;
1114}
1115
1116struct migration_arg {
1117	struct task_struct *task;
1118	int dest_cpu;
1119};
1120
1121static int migration_cpu_stop(void *data);
1122
1123/*
1124 * wait_task_inactive - wait for a thread to unschedule.
1125 *
1126 * If @match_state is nonzero, it's the @p->state value just checked and
1127 * not expected to change.  If it changes, i.e. @p might have woken up,
1128 * then return zero.  When we succeed in waiting for @p to be off its CPU,
1129 * we return a positive number (its total switch count).  If a second call
1130 * a short while later returns the same number, the caller can be sure that
1131 * @p has remained unscheduled the whole time.
1132 *
1133 * The caller must ensure that the task *will* unschedule sometime soon,
1134 * else this function might spin for a *long* time. This function can't
1135 * be called with interrupts off, or it may introduce deadlock with
1136 * smp_call_function() if an IPI is sent by the same process we are
1137 * waiting to become inactive.
1138 */
1139unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1140{
1141	unsigned long flags;
1142	int running, on_rq;
1143	unsigned long ncsw;
1144	struct rq *rq;
1145
1146	for (;;) {
1147		/*
1148		 * We do the initial early heuristics without holding
1149		 * any task-queue locks at all. We'll only try to get
1150		 * the runqueue lock when things look like they will
1151		 * work out!
1152		 */
1153		rq = task_rq(p);
1154
1155		/*
1156		 * If the task is actively running on another CPU
1157		 * still, just relax and busy-wait without holding
1158		 * any locks.
1159		 *
1160		 * NOTE! Since we don't hold any locks, it's not
1161		 * even sure that "rq" stays as the right runqueue!
1162		 * But we don't care, since "task_running()" will
1163		 * return false if the runqueue has changed and p
1164		 * is actually now running somewhere else!
1165		 */
1166		while (task_running(rq, p)) {
1167			if (match_state && unlikely(p->state != match_state))
1168				return 0;
1169			cpu_relax();
1170		}
1171
1172		/*
1173		 * Ok, time to look more closely! We need the rq
1174		 * lock now, to be *sure*. If we're wrong, we'll
1175		 * just go back and repeat.
1176		 */
1177		rq = task_rq_lock(p, &flags);
1178		trace_sched_wait_task(p);
1179		running = task_running(rq, p);
1180		on_rq = p->on_rq;
1181		ncsw = 0;
1182		if (!match_state || p->state == match_state)
1183			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1184		task_rq_unlock(rq, p, &flags);
1185
1186		/*
1187		 * If it changed from the expected state, bail out now.
1188		 */
1189		if (unlikely(!ncsw))
1190			break;
1191
1192		/*
1193		 * Was it really running after all now that we
1194		 * checked with the proper locks actually held?
1195		 *
1196		 * Oops. Go back and try again..
1197		 */
1198		if (unlikely(running)) {
1199			cpu_relax();
1200			continue;
1201		}
1202
1203		/*
1204		 * It's not enough that it's not actively running,
1205		 * it must be off the runqueue _entirely_, and not
1206		 * preempted!
1207		 *
1208		 * So if it was still runnable (but just not actively
1209		 * running right now), it's preempted, and we should
1210		 * yield - it could be a while.
1211		 */
1212		if (unlikely(on_rq)) {
1213			ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1214
1215			set_current_state(TASK_UNINTERRUPTIBLE);
1216			schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1217			continue;
1218		}
1219
1220		/*
1221		 * Ahh, all good. It wasn't running, and it wasn't
1222		 * runnable, which means that it will never become
1223		 * running in the future either. We're all done!
1224		 */
1225		break;
1226	}
1227
1228	return ncsw;
1229}
1230
1231/***
1232 * kick_process - kick a running thread to enter/exit the kernel
1233 * @p: the to-be-kicked thread
1234 *
1235 * Cause a process which is running on another CPU to enter
1236 * kernel-mode, without any delay. (to get signals handled.)
1237 *
1238 * NOTE: this function doesn't have to take the runqueue lock,
1239 * because all it wants to ensure is that the remote task enters
1240 * the kernel. If the IPI races and the task has been migrated
1241 * to another CPU then no harm is done and the purpose has been
1242 * achieved as well.
1243 */
1244void kick_process(struct task_struct *p)
1245{
1246	int cpu;
1247
1248	preempt_disable();
1249	cpu = task_cpu(p);
1250	if ((cpu != smp_processor_id()) && task_curr(p))
1251		smp_send_reschedule(cpu);
1252	preempt_enable();
1253}
1254EXPORT_SYMBOL_GPL(kick_process);
1255#endif /* CONFIG_SMP */
1256
1257#ifdef CONFIG_SMP
1258/*
1259 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1260 */
1261static int select_fallback_rq(int cpu, struct task_struct *p)
1262{
1263	int nid = cpu_to_node(cpu);
1264	const struct cpumask *nodemask = NULL;
1265	enum { cpuset, possible, fail } state = cpuset;
1266	int dest_cpu;
1267
1268	/*
1269	 * If the node that the cpu is on has been offlined, cpu_to_node()
1270	 * will return -1. There is no cpu on the node, and we should
1271	 * select the cpu on the other node.
1272	 */
1273	if (nid != -1) {
1274		nodemask = cpumask_of_node(nid);
1275
1276		/* Look for allowed, online CPU in same node. */
1277		for_each_cpu(dest_cpu, nodemask) {
1278			if (!cpu_online(dest_cpu))
1279				continue;
1280			if (!cpu_active(dest_cpu))
1281				continue;
1282			if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1283				return dest_cpu;
1284		}
1285	}
1286
1287	for (;;) {
1288		/* Any allowed, online CPU? */
1289		for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1290			if (!cpu_online(dest_cpu))
1291				continue;
1292			if (!cpu_active(dest_cpu))
1293				continue;
1294			goto out;
1295		}
1296
1297		switch (state) {
1298		case cpuset:
1299			/* No more Mr. Nice Guy. */
1300			cpuset_cpus_allowed_fallback(p);
1301			state = possible;
1302			break;
1303
1304		case possible:
1305			do_set_cpus_allowed(p, cpu_possible_mask);
1306			state = fail;
1307			break;
1308
1309		case fail:
1310			BUG();
1311			break;
1312		}
1313	}
1314
1315out:
1316	if (state != cpuset) {
1317		/*
1318		 * Don't tell them about moving exiting tasks or
1319		 * kernel threads (both mm NULL), since they never
1320		 * leave kernel.
1321		 */
1322		if (p->mm && printk_ratelimit()) {
1323			printk_sched("process %d (%s) no longer affine to cpu%d\n",
1324					task_pid_nr(p), p->comm, cpu);
1325		}
1326	}
1327
1328	return dest_cpu;
1329}
1330
1331/*
1332 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1333 */
1334static inline
1335int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1336{
1337	cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1338
1339	/*
1340	 * In order not to call set_task_cpu() on a blocking task we need
1341	 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1342	 * cpu.
1343	 *
1344	 * Since this is common to all placement strategies, this lives here.
1345	 *
1346	 * [ this allows ->select_task() to simply return task_cpu(p) and
1347	 *   not worry about this generic constraint ]
1348	 */
1349	if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1350		     !cpu_online(cpu)))
1351		cpu = select_fallback_rq(task_cpu(p), p);
1352
1353	return cpu;
1354}
1355
1356static void update_avg(u64 *avg, u64 sample)
1357{
1358	s64 diff = sample - *avg;
1359	*avg += diff >> 3;
1360}
1361#endif
1362
1363static void
1364ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1365{
1366#ifdef CONFIG_SCHEDSTATS
1367	struct rq *rq = this_rq();
1368
1369#ifdef CONFIG_SMP
1370	int this_cpu = smp_processor_id();
1371
1372	if (cpu == this_cpu) {
1373		schedstat_inc(rq, ttwu_local);
1374		schedstat_inc(p, se.statistics.nr_wakeups_local);
1375	} else {
1376		struct sched_domain *sd;
1377
1378		schedstat_inc(p, se.statistics.nr_wakeups_remote);
1379		rcu_read_lock();
1380		for_each_domain(this_cpu, sd) {
1381			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1382				schedstat_inc(sd, ttwu_wake_remote);
1383				break;
1384			}
1385		}
1386		rcu_read_unlock();
1387	}
1388
1389	if (wake_flags & WF_MIGRATED)
1390		schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1391
1392#endif /* CONFIG_SMP */
1393
1394	schedstat_inc(rq, ttwu_count);
1395	schedstat_inc(p, se.statistics.nr_wakeups);
1396
1397	if (wake_flags & WF_SYNC)
1398		schedstat_inc(p, se.statistics.nr_wakeups_sync);
1399
1400#endif /* CONFIG_SCHEDSTATS */
1401}
1402
1403static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1404{
1405	activate_task(rq, p, en_flags);
1406	p->on_rq = 1;
1407
1408	/* if a worker is waking up, notify workqueue */
1409	if (p->flags & PF_WQ_WORKER)
1410		wq_worker_waking_up(p, cpu_of(rq));
1411}
1412
1413/*
1414 * Mark the task runnable and perform wakeup-preemption.
1415 */
1416static void
1417ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1418{
1419	check_preempt_curr(rq, p, wake_flags);
1420	trace_sched_wakeup(p, true);
1421
1422	p->state = TASK_RUNNING;
1423#ifdef CONFIG_SMP
1424	if (p->sched_class->task_woken)
1425		p->sched_class->task_woken(rq, p);
1426
1427	if (rq->idle_stamp) {
1428		u64 delta = rq_clock(rq) - rq->idle_stamp;
1429		u64 max = 2*rq->max_idle_balance_cost;
1430
1431		update_avg(&rq->avg_idle, delta);
1432
1433		if (rq->avg_idle > max)
1434			rq->avg_idle = max;
1435
1436		rq->idle_stamp = 0;
1437	}
1438#endif
1439}
1440
1441static void
1442ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1443{
1444#ifdef CONFIG_SMP
1445	if (p->sched_contributes_to_load)
1446		rq->nr_uninterruptible--;
1447#endif
1448
1449	ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1450	ttwu_do_wakeup(rq, p, wake_flags);
1451}
1452
1453/*
1454 * Called in case the task @p isn't fully descheduled from its runqueue,
1455 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1456 * since all we need to do is flip p->state to TASK_RUNNING, since
1457 * the task is still ->on_rq.
1458 */
1459static int ttwu_remote(struct task_struct *p, int wake_flags)
1460{
1461	struct rq *rq;
1462	int ret = 0;
1463
1464	rq = __task_rq_lock(p);
1465	if (p->on_rq) {
1466		/* check_preempt_curr() may use rq clock */
1467		update_rq_clock(rq);
1468		ttwu_do_wakeup(rq, p, wake_flags);
1469		ret = 1;
1470	}
1471	__task_rq_unlock(rq);
1472
1473	return ret;
1474}
1475
1476#ifdef CONFIG_SMP
1477static void sched_ttwu_pending(void)
1478{
1479	struct rq *rq = this_rq();
1480	struct llist_node *llist = llist_del_all(&rq->wake_list);
1481	struct task_struct *p;
1482
1483	raw_spin_lock(&rq->lock);
1484
1485	while (llist) {
1486		p = llist_entry(llist, struct task_struct, wake_entry);
1487		llist = llist_next(llist);
1488		ttwu_do_activate(rq, p, 0);
1489	}
1490
1491	raw_spin_unlock(&rq->lock);
1492}
1493
1494void scheduler_ipi(void)
1495{
1496	/*
1497	 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1498	 * TIF_NEED_RESCHED remotely (for the first time) will also send
1499	 * this IPI.
1500	 */
1501	preempt_fold_need_resched();
1502
1503	if (llist_empty(&this_rq()->wake_list)
1504			&& !tick_nohz_full_cpu(smp_processor_id())
1505			&& !got_nohz_idle_kick())
1506		return;
1507
1508	/*
1509	 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1510	 * traditionally all their work was done from the interrupt return
1511	 * path. Now that we actually do some work, we need to make sure
1512	 * we do call them.
1513	 *
1514	 * Some archs already do call them, luckily irq_enter/exit nest
1515	 * properly.
1516	 *
1517	 * Arguably we should visit all archs and update all handlers,
1518	 * however a fair share of IPIs are still resched only so this would
1519	 * somewhat pessimize the simple resched case.
1520	 */
1521	irq_enter();
1522	tick_nohz_full_check();
1523	sched_ttwu_pending();
1524
1525	/*
1526	 * Check if someone kicked us for doing the nohz idle load balance.
1527	 */
1528	if (unlikely(got_nohz_idle_kick())) {
1529		this_rq()->idle_balance = 1;
1530		raise_softirq_irqoff(SCHED_SOFTIRQ);
1531	}
1532	irq_exit();
1533}
1534
1535static void ttwu_queue_remote(struct task_struct *p, int cpu)
1536{
1537	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1538		smp_send_reschedule(cpu);
1539}
1540
1541bool cpus_share_cache(int this_cpu, int that_cpu)
1542{
1543	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1544}
1545#endif /* CONFIG_SMP */
1546
1547static void ttwu_queue(struct task_struct *p, int cpu)
1548{
1549	struct rq *rq = cpu_rq(cpu);
1550
1551#if defined(CONFIG_SMP)
1552	if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1553		sched_clock_cpu(cpu); /* sync clocks x-cpu */
1554		ttwu_queue_remote(p, cpu);
1555		return;
1556	}
1557#endif
1558
1559	raw_spin_lock(&rq->lock);
1560	ttwu_do_activate(rq, p, 0);
1561	raw_spin_unlock(&rq->lock);
1562}
1563
1564/**
1565 * try_to_wake_up - wake up a thread
1566 * @p: the thread to be awakened
1567 * @state: the mask of task states that can be woken
1568 * @wake_flags: wake modifier flags (WF_*)
1569 *
1570 * Put it on the run-queue if it's not already there. The "current"
1571 * thread is always on the run-queue (except when the actual
1572 * re-schedule is in progress), and as such you're allowed to do
1573 * the simpler "current->state = TASK_RUNNING" to mark yourself
1574 * runnable without the overhead of this.
1575 *
1576 * Return: %true if @p was woken up, %false if it was already running.
1577 * or @state didn't match @p's state.
1578 */
1579static int
1580try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1581{
1582	unsigned long flags;
1583	int cpu, success = 0;
1584
1585	/*
1586	 * If we are going to wake up a thread waiting for CONDITION we
1587	 * need to ensure that CONDITION=1 done by the caller can not be
1588	 * reordered with p->state check below. This pairs with mb() in
1589	 * set_current_state() the waiting thread does.
1590	 */
1591	smp_mb__before_spinlock();
1592	raw_spin_lock_irqsave(&p->pi_lock, flags);
1593	if (!(p->state & state))
1594		goto out;
1595
1596	success = 1; /* we're going to change ->state */
1597	cpu = task_cpu(p);
1598
1599	if (p->on_rq && ttwu_remote(p, wake_flags))
1600		goto stat;
1601
1602#ifdef CONFIG_SMP
1603	/*
1604	 * If the owning (remote) cpu is still in the middle of schedule() with
1605	 * this task as prev, wait until its done referencing the task.
1606	 */
1607	while (p->on_cpu)
1608		cpu_relax();
1609	/*
1610	 * Pairs with the smp_wmb() in finish_lock_switch().
1611	 */
1612	smp_rmb();
1613
1614	p->sched_contributes_to_load = !!task_contributes_to_load(p);
1615	p->state = TASK_WAKING;
1616
1617	if (p->sched_class->task_waking)
1618		p->sched_class->task_waking(p);
1619
1620	cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1621	if (task_cpu(p) != cpu) {
1622		wake_flags |= WF_MIGRATED;
1623		set_task_cpu(p, cpu);
1624	}
1625#endif /* CONFIG_SMP */
1626
1627	ttwu_queue(p, cpu);
1628stat:
1629	ttwu_stat(p, cpu, wake_flags);
1630out:
1631	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1632
1633	return success;
1634}
1635
1636/**
1637 * try_to_wake_up_local - try to wake up a local task with rq lock held
1638 * @p: the thread to be awakened
1639 *
1640 * Put @p on the run-queue if it's not already there. The caller must
1641 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1642 * the current task.
1643 */
1644static void try_to_wake_up_local(struct task_struct *p)
1645{
1646	struct rq *rq = task_rq(p);
1647
1648	if (WARN_ON_ONCE(rq != this_rq()) ||
1649	    WARN_ON_ONCE(p == current))
1650		return;
1651
1652	lockdep_assert_held(&rq->lock);
1653
1654	if (!raw_spin_trylock(&p->pi_lock)) {
1655		raw_spin_unlock(&rq->lock);
1656		raw_spin_lock(&p->pi_lock);
1657		raw_spin_lock(&rq->lock);
1658	}
1659
1660	if (!(p->state & TASK_NORMAL))
1661		goto out;
1662
1663	if (!p->on_rq)
1664		ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1665
1666	ttwu_do_wakeup(rq, p, 0);
1667	ttwu_stat(p, smp_processor_id(), 0);
1668out:
1669	raw_spin_unlock(&p->pi_lock);
1670}
1671
1672/**
1673 * wake_up_process - Wake up a specific process
1674 * @p: The process to be woken up.
1675 *
1676 * Attempt to wake up the nominated process and move it to the set of runnable
1677 * processes.
1678 *
1679 * Return: 1 if the process was woken up, 0 if it was already running.
1680 *
1681 * It may be assumed that this function implies a write memory barrier before
1682 * changing the task state if and only if any tasks are woken up.
1683 */
1684int wake_up_process(struct task_struct *p)
1685{
1686	WARN_ON(task_is_stopped_or_traced(p));
1687	return try_to_wake_up(p, TASK_NORMAL, 0);
1688}
1689EXPORT_SYMBOL(wake_up_process);
1690
1691int wake_up_state(struct task_struct *p, unsigned int state)
1692{
1693	return try_to_wake_up(p, state, 0);
1694}
1695
1696/*
1697 * Perform scheduler related setup for a newly forked process p.
1698 * p is forked by current.
1699 *
1700 * __sched_fork() is basic setup used by init_idle() too:
1701 */
1702static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1703{
1704	p->on_rq			= 0;
1705
1706	p->se.on_rq			= 0;
1707	p->se.exec_start		= 0;
1708	p->se.sum_exec_runtime		= 0;
1709	p->se.prev_sum_exec_runtime	= 0;
1710	p->se.nr_migrations		= 0;
1711	p->se.vruntime			= 0;
1712	INIT_LIST_HEAD(&p->se.group_node);
1713
1714#ifdef CONFIG_SCHEDSTATS
1715	memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1716#endif
1717
1718	RB_CLEAR_NODE(&p->dl.rb_node);
1719	hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1720	p->dl.dl_runtime = p->dl.runtime = 0;
1721	p->dl.dl_deadline = p->dl.deadline = 0;
1722	p->dl.dl_period = 0;
1723	p->dl.flags = 0;
1724
1725	INIT_LIST_HEAD(&p->rt.run_list);
1726
1727#ifdef CONFIG_PREEMPT_NOTIFIERS
1728	INIT_HLIST_HEAD(&p->preempt_notifiers);
1729#endif
1730
1731#ifdef CONFIG_NUMA_BALANCING
1732	if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1733		p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1734		p->mm->numa_scan_seq = 0;
1735	}
1736
1737	if (clone_flags & CLONE_VM)
1738		p->numa_preferred_nid = current->numa_preferred_nid;
1739	else
1740		p->numa_preferred_nid = -1;
1741
1742	p->node_stamp = 0ULL;
1743	p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1744	p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1745	p->numa_work.next = &p->numa_work;
1746	p->numa_faults_memory = NULL;
1747	p->numa_faults_buffer_memory = NULL;
1748	p->last_task_numa_placement = 0;
1749	p->last_sum_exec_runtime = 0;
1750
1751	INIT_LIST_HEAD(&p->numa_entry);
1752	p->numa_group = NULL;
1753#endif /* CONFIG_NUMA_BALANCING */
1754}
1755
1756#ifdef CONFIG_NUMA_BALANCING
1757#ifdef CONFIG_SCHED_DEBUG
1758void set_numabalancing_state(bool enabled)
1759{
1760	if (enabled)
1761		sched_feat_set("NUMA");
1762	else
1763		sched_feat_set("NO_NUMA");
1764}
1765#else
1766__read_mostly bool numabalancing_enabled;
1767
1768void set_numabalancing_state(bool enabled)
1769{
1770	numabalancing_enabled = enabled;
1771}
1772#endif /* CONFIG_SCHED_DEBUG */
1773
1774#ifdef CONFIG_PROC_SYSCTL
1775int sysctl_numa_balancing(struct ctl_table *table, int write,
1776			 void __user *buffer, size_t *lenp, loff_t *ppos)
1777{
1778	struct ctl_table t;
1779	int err;
1780	int state = numabalancing_enabled;
1781
1782	if (write && !capable(CAP_SYS_ADMIN))
1783		return -EPERM;
1784
1785	t = *table;
1786	t.data = &state;
1787	err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1788	if (err < 0)
1789		return err;
1790	if (write)
1791		set_numabalancing_state(state);
1792	return err;
1793}
1794#endif
1795#endif
1796
1797/*
1798 * fork()/clone()-time setup:
1799 */
1800int sched_fork(unsigned long clone_flags, struct task_struct *p)
1801{
1802	unsigned long flags;
1803	int cpu = get_cpu();
1804
1805	__sched_fork(clone_flags, p);
1806	/*
1807	 * We mark the process as running here. This guarantees that
1808	 * nobody will actually run it, and a signal or other external
1809	 * event cannot wake it up and insert it on the runqueue either.
1810	 */
1811	p->state = TASK_RUNNING;
1812
1813	/*
1814	 * Make sure we do not leak PI boosting priority to the child.
1815	 */
1816	p->prio = current->normal_prio;
1817
1818	/*
1819	 * Revert to default priority/policy on fork if requested.
1820	 */
1821	if (unlikely(p->sched_reset_on_fork)) {
1822		if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1823			p->policy = SCHED_NORMAL;
1824			p->static_prio = NICE_TO_PRIO(0);
1825			p->rt_priority = 0;
1826		} else if (PRIO_TO_NICE(p->static_prio) < 0)
1827			p->static_prio = NICE_TO_PRIO(0);
1828
1829		p->prio = p->normal_prio = __normal_prio(p);
1830		set_load_weight(p);
1831
1832		/*
1833		 * We don't need the reset flag anymore after the fork. It has
1834		 * fulfilled its duty:
1835		 */
1836		p->sched_reset_on_fork = 0;
1837	}
1838
1839	if (dl_prio(p->prio)) {
1840		put_cpu();
1841		return -EAGAIN;
1842	} else if (rt_prio(p->prio)) {
1843		p->sched_class = &rt_sched_class;
1844	} else {
1845		p->sched_class = &fair_sched_class;
1846	}
1847
1848	if (p->sched_class->task_fork)
1849		p->sched_class->task_fork(p);
1850
1851	/*
1852	 * The child is not yet in the pid-hash so no cgroup attach races,
1853	 * and the cgroup is pinned to this child due to cgroup_fork()
1854	 * is ran before sched_fork().
1855	 *
1856	 * Silence PROVE_RCU.
1857	 */
1858	raw_spin_lock_irqsave(&p->pi_lock, flags);
1859	set_task_cpu(p, cpu);
1860	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1861
1862#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1863	if (likely(sched_info_on()))
1864		memset(&p->sched_info, 0, sizeof(p->sched_info));
1865#endif
1866#if defined(CONFIG_SMP)
1867	p->on_cpu = 0;
1868#endif
1869	init_task_preempt_count(p);
1870#ifdef CONFIG_SMP
1871	plist_node_init(&p->pushable_tasks, MAX_PRIO);
1872	RB_CLEAR_NODE(&p->pushable_dl_tasks);
1873#endif
1874
1875	put_cpu();
1876	return 0;
1877}
1878
1879unsigned long to_ratio(u64 period, u64 runtime)
1880{
1881	if (runtime == RUNTIME_INF)
1882		return 1ULL << 20;
1883
1884	/*
1885	 * Doing this here saves a lot of checks in all
1886	 * the calling paths, and returning zero seems
1887	 * safe for them anyway.
1888	 */
1889	if (period == 0)
1890		return 0;
1891
1892	return div64_u64(runtime << 20, period);
1893}
1894
1895#ifdef CONFIG_SMP
1896inline struct dl_bw *dl_bw_of(int i)
1897{
1898	return &cpu_rq(i)->rd->dl_bw;
1899}
1900
1901static inline int dl_bw_cpus(int i)
1902{
1903	struct root_domain *rd = cpu_rq(i)->rd;
1904	int cpus = 0;
1905
1906	for_each_cpu_and(i, rd->span, cpu_active_mask)
1907		cpus++;
1908
1909	return cpus;
1910}
1911#else
1912inline struct dl_bw *dl_bw_of(int i)
1913{
1914	return &cpu_rq(i)->dl.dl_bw;
1915}
1916
1917static inline int dl_bw_cpus(int i)
1918{
1919	return 1;
1920}
1921#endif
1922
1923static inline
1924void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1925{
1926	dl_b->total_bw -= tsk_bw;
1927}
1928
1929static inline
1930void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1931{
1932	dl_b->total_bw += tsk_bw;
1933}
1934
1935static inline
1936bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1937{
1938	return dl_b->bw != -1 &&
1939	       dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1940}
1941
1942/*
1943 * We must be sure that accepting a new task (or allowing changing the
1944 * parameters of an existing one) is consistent with the bandwidth
1945 * constraints. If yes, this function also accordingly updates the currently
1946 * allocated bandwidth to reflect the new situation.
1947 *
1948 * This function is called while holding p's rq->lock.
1949 */
1950static int dl_overflow(struct task_struct *p, int policy,
1951		       const struct sched_attr *attr)
1952{
1953
1954	struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1955	u64 period = attr->sched_period ?: attr->sched_deadline;
1956	u64 runtime = attr->sched_runtime;
1957	u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
1958	int cpus, err = -1;
1959
1960	if (new_bw == p->dl.dl_bw)
1961		return 0;
1962
1963	/*
1964	 * Either if a task, enters, leave, or stays -deadline but changes
1965	 * its parameters, we may need to update accordingly the total
1966	 * allocated bandwidth of the container.
1967	 */
1968	raw_spin_lock(&dl_b->lock);
1969	cpus = dl_bw_cpus(task_cpu(p));
1970	if (dl_policy(policy) && !task_has_dl_policy(p) &&
1971	    !__dl_overflow(dl_b, cpus, 0, new_bw)) {
1972		__dl_add(dl_b, new_bw);
1973		err = 0;
1974	} else if (dl_policy(policy) && task_has_dl_policy(p) &&
1975		   !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
1976		__dl_clear(dl_b, p->dl.dl_bw);
1977		__dl_add(dl_b, new_bw);
1978		err = 0;
1979	} else if (!dl_policy(policy) && task_has_dl_policy(p)) {
1980		__dl_clear(dl_b, p->dl.dl_bw);
1981		err = 0;
1982	}
1983	raw_spin_unlock(&dl_b->lock);
1984
1985	return err;
1986}
1987
1988extern void init_dl_bw(struct dl_bw *dl_b);
1989
1990/*
1991 * wake_up_new_task - wake up a newly created task for the first time.
1992 *
1993 * This function will do some initial scheduler statistics housekeeping
1994 * that must be done for every newly created context, then puts the task
1995 * on the runqueue and wakes it.
1996 */
1997void wake_up_new_task(struct task_struct *p)
1998{
1999	unsigned long flags;
2000	struct rq *rq;
2001
2002	raw_spin_lock_irqsave(&p->pi_lock, flags);
2003#ifdef CONFIG_SMP
2004	/*
2005	 * Fork balancing, do it here and not earlier because:
2006	 *  - cpus_allowed can change in the fork path
2007	 *  - any previously selected cpu might disappear through hotplug
2008	 */
2009	set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2010#endif
2011
2012	/* Initialize new task's runnable average */
2013	init_task_runnable_average(p);
2014	rq = __task_rq_lock(p);
2015	activate_task(rq, p, 0);
2016	p->on_rq = 1;
2017	trace_sched_wakeup_new(p, true);
2018	check_preempt_curr(rq, p, WF_FORK);
2019#ifdef CONFIG_SMP
2020	if (p->sched_class->task_woken)
2021		p->sched_class->task_woken(rq, p);
2022#endif
2023	task_rq_unlock(rq, p, &flags);
2024}
2025
2026#ifdef CONFIG_PREEMPT_NOTIFIERS
2027
2028/**
2029 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2030 * @notifier: notifier struct to register
2031 */
2032void preempt_notifier_register(struct preempt_notifier *notifier)
2033{
2034	hlist_add_head(&notifier->link, &current->preempt_notifiers);
2035}
2036EXPORT_SYMBOL_GPL(preempt_notifier_register);
2037
2038/**
2039 * preempt_notifier_unregister - no longer interested in preemption notifications
2040 * @notifier: notifier struct to unregister
2041 *
2042 * This is safe to call from within a preemption notifier.
2043 */
2044void preempt_notifier_unregister(struct preempt_notifier *notifier)
2045{
2046	hlist_del(&notifier->link);
2047}
2048EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2049
2050static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2051{
2052	struct preempt_notifier *notifier;
2053
2054	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2055		notifier->ops->sched_in(notifier, raw_smp_processor_id());
2056}
2057
2058static void
2059fire_sched_out_preempt_notifiers(struct task_struct *curr,
2060				 struct task_struct *next)
2061{
2062	struct preempt_notifier *notifier;
2063
2064	hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2065		notifier->ops->sched_out(notifier, next);
2066}
2067
2068#else /* !CONFIG_PREEMPT_NOTIFIERS */
2069
2070static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2071{
2072}
2073
2074static void
2075fire_sched_out_preempt_notifiers(struct task_struct *curr,
2076				 struct task_struct *next)
2077{
2078}
2079
2080#endif /* CONFIG_PREEMPT_NOTIFIERS */
2081
2082/**
2083 * prepare_task_switch - prepare to switch tasks
2084 * @rq: the runqueue preparing to switch
2085 * @prev: the current task that is being switched out
2086 * @next: the task we are going to switch to.
2087 *
2088 * This is called with the rq lock held and interrupts off. It must
2089 * be paired with a subsequent finish_task_switch after the context
2090 * switch.
2091 *
2092 * prepare_task_switch sets up locking and calls architecture specific
2093 * hooks.
2094 */
2095static inline void
2096prepare_task_switch(struct rq *rq, struct task_struct *prev,
2097		    struct task_struct *next)
2098{
2099	trace_sched_switch(prev, next);
2100	sched_info_switch(rq, prev, next);
2101	perf_event_task_sched_out(prev, next);
2102	fire_sched_out_preempt_notifiers(prev, next);
2103	prepare_lock_switch(rq, next);
2104	prepare_arch_switch(next);
2105}
2106
2107/**
2108 * finish_task_switch - clean up after a task-switch
2109 * @rq: runqueue associated with task-switch
2110 * @prev: the thread we just switched away from.
2111 *
2112 * finish_task_switch must be called after the context switch, paired
2113 * with a prepare_task_switch call before the context switch.
2114 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2115 * and do any other architecture-specific cleanup actions.
2116 *
2117 * Note that we may have delayed dropping an mm in context_switch(). If
2118 * so, we finish that here outside of the runqueue lock. (Doing it
2119 * with the lock held can cause deadlocks; see schedule() for
2120 * details.)
2121 */
2122static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2123	__releases(rq->lock)
2124{
2125	struct mm_struct *mm = rq->prev_mm;
2126	long prev_state;
2127
2128	rq->prev_mm = NULL;
2129
2130	/*
2131	 * A task struct has one reference for the use as "current".
2132	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2133	 * schedule one last time. The schedule call will never return, and
2134	 * the scheduled task must drop that reference.
2135	 * The test for TASK_DEAD must occur while the runqueue locks are
2136	 * still held, otherwise prev could be scheduled on another cpu, die
2137	 * there before we look at prev->state, and then the reference would
2138	 * be dropped twice.
2139	 *		Manfred Spraul <manfred@colorfullife.com>
2140	 */
2141	prev_state = prev->state;
2142	vtime_task_switch(prev);
2143	finish_arch_switch(prev);
2144	perf_event_task_sched_in(prev, current);
2145	finish_lock_switch(rq, prev);
2146	finish_arch_post_lock_switch();
2147
2148	fire_sched_in_preempt_notifiers(current);
2149	if (mm)
2150		mmdrop(mm);
2151	if (unlikely(prev_state == TASK_DEAD)) {
2152		if (prev->sched_class->task_dead)
2153			prev->sched_class->task_dead(prev);
2154
2155		/*
2156		 * Remove function-return probe instances associated with this
2157		 * task and put them back on the free list.
2158		 */
2159		kprobe_flush_task(prev);
2160		put_task_struct(prev);
2161	}
2162
2163	tick_nohz_task_switch(current);
2164}
2165
2166#ifdef CONFIG_SMP
2167
2168/* rq->lock is NOT held, but preemption is disabled */
2169static inline void post_schedule(struct rq *rq)
2170{
2171	if (rq->post_schedule) {
2172		unsigned long flags;
2173
2174		raw_spin_lock_irqsave(&rq->lock, flags);
2175		if (rq->curr->sched_class->post_schedule)
2176			rq->curr->sched_class->post_schedule(rq);
2177		raw_spin_unlock_irqrestore(&rq->lock, flags);
2178
2179		rq->post_schedule = 0;
2180	}
2181}
2182
2183#else
2184
2185static inline void post_schedule(struct rq *rq)
2186{
2187}
2188
2189#endif
2190
2191/**
2192 * schedule_tail - first thing a freshly forked thread must call.
2193 * @prev: the thread we just switched away from.
2194 */
2195asmlinkage void schedule_tail(struct task_struct *prev)
2196	__releases(rq->lock)
2197{
2198	struct rq *rq = this_rq();
2199
2200	finish_task_switch(rq, prev);
2201
2202	/*
2203	 * FIXME: do we need to worry about rq being invalidated by the
2204	 * task_switch?
2205	 */
2206	post_schedule(rq);
2207
2208#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2209	/* In this case, finish_task_switch does not reenable preemption */
2210	preempt_enable();
2211#endif
2212	if (current->set_child_tid)
2213		put_user(task_pid_vnr(current), current->set_child_tid);
2214}
2215
2216/*
2217 * context_switch - switch to the new MM and the new
2218 * thread's register state.
2219 */
2220static inline void
2221context_switch(struct rq *rq, struct task_struct *prev,
2222	       struct task_struct *next)
2223{
2224	struct mm_struct *mm, *oldmm;
2225
2226	prepare_task_switch(rq, prev, next);
2227
2228	mm = next->mm;
2229	oldmm = prev->active_mm;
2230	/*
2231	 * For paravirt, this is coupled with an exit in switch_to to
2232	 * combine the page table reload and the switch backend into
2233	 * one hypercall.
2234	 */
2235	arch_start_context_switch(prev);
2236
2237	if (!mm) {
2238		next->active_mm = oldmm;
2239		atomic_inc(&oldmm->mm_count);
2240		enter_lazy_tlb(oldmm, next);
2241	} else
2242		switch_mm(oldmm, mm, next);
2243
2244	if (!prev->mm) {
2245		prev->active_mm = NULL;
2246		rq->prev_mm = oldmm;
2247	}
2248	/*
2249	 * Since the runqueue lock will be released by the next
2250	 * task (which is an invalid locking op but in the case
2251	 * of the scheduler it's an obvious special-case), so we
2252	 * do an early lockdep release here:
2253	 */
2254#ifndef __ARCH_WANT_UNLOCKED_CTXSW
2255	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2256#endif
2257
2258	context_tracking_task_switch(prev, next);
2259	/* Here we just switch the register state and the stack. */
2260	switch_to(prev, next, prev);
2261
2262	barrier();
2263	/*
2264	 * this_rq must be evaluated again because prev may have moved
2265	 * CPUs since it called schedule(), thus the 'rq' on its stack
2266	 * frame will be invalid.
2267	 */
2268	finish_task_switch(this_rq(), prev);
2269}
2270
2271/*
2272 * nr_running and nr_context_switches:
2273 *
2274 * externally visible scheduler statistics: current number of runnable
2275 * threads, total number of context switches performed since bootup.
2276 */
2277unsigned long nr_running(void)
2278{
2279	unsigned long i, sum = 0;
2280
2281	for_each_online_cpu(i)
2282		sum += cpu_rq(i)->nr_running;
2283
2284	return sum;
2285}
2286
2287unsigned long long nr_context_switches(void)
2288{
2289	int i;
2290	unsigned long long sum = 0;
2291
2292	for_each_possible_cpu(i)
2293		sum += cpu_rq(i)->nr_switches;
2294
2295	return sum;
2296}
2297
2298unsigned long nr_iowait(void)
2299{
2300	unsigned long i, sum = 0;
2301
2302	for_each_possible_cpu(i)
2303		sum += atomic_read(&cpu_rq(i)->nr_iowait);
2304
2305	return sum;
2306}
2307
2308unsigned long nr_iowait_cpu(int cpu)
2309{
2310	struct rq *this = cpu_rq(cpu);
2311	return atomic_read(&this->nr_iowait);
2312}
2313
2314#ifdef CONFIG_SMP
2315
2316/*
2317 * sched_exec - execve() is a valuable balancing opportunity, because at
2318 * this point the task has the smallest effective memory and cache footprint.
2319 */
2320void sched_exec(void)
2321{
2322	struct task_struct *p = current;
2323	unsigned long flags;
2324	int dest_cpu;
2325
2326	raw_spin_lock_irqsave(&p->pi_lock, flags);
2327	dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2328	if (dest_cpu == smp_processor_id())
2329		goto unlock;
2330
2331	if (likely(cpu_active(dest_cpu))) {
2332		struct migration_arg arg = { p, dest_cpu };
2333
2334		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2335		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2336		return;
2337	}
2338unlock:
2339	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2340}
2341
2342#endif
2343
2344DEFINE_PER_CPU(struct kernel_stat, kstat);
2345DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2346
2347EXPORT_PER_CPU_SYMBOL(kstat);
2348EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2349
2350/*
2351 * Return any ns on the sched_clock that have not yet been accounted in
2352 * @p in case that task is currently running.
2353 *
2354 * Called with task_rq_lock() held on @rq.
2355 */
2356static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2357{
2358	u64 ns = 0;
2359
2360	if (task_current(rq, p)) {
2361		update_rq_clock(rq);
2362		ns = rq_clock_task(rq) - p->se.exec_start;
2363		if ((s64)ns < 0)
2364			ns = 0;
2365	}
2366
2367	return ns;
2368}
2369
2370unsigned long long task_delta_exec(struct task_struct *p)
2371{
2372	unsigned long flags;
2373	struct rq *rq;
2374	u64 ns = 0;
2375
2376	rq = task_rq_lock(p, &flags);
2377	ns = do_task_delta_exec(p, rq);
2378	task_rq_unlock(rq, p, &flags);
2379
2380	return ns;
2381}
2382
2383/*
2384 * Return accounted runtime for the task.
2385 * In case the task is currently running, return the runtime plus current's
2386 * pending runtime that have not been accounted yet.
2387 */
2388unsigned long long task_sched_runtime(struct task_struct *p)
2389{
2390	unsigned long flags;
2391	struct rq *rq;
2392	u64 ns = 0;
2393
2394#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2395	/*
2396	 * 64-bit doesn't need locks to atomically read a 64bit value.
2397	 * So we have a optimization chance when the task's delta_exec is 0.
2398	 * Reading ->on_cpu is racy, but this is ok.
2399	 *
2400	 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2401	 * If we race with it entering cpu, unaccounted time is 0. This is
2402	 * indistinguishable from the read occurring a few cycles earlier.
2403	 */
2404	if (!p->on_cpu)
2405		return p->se.sum_exec_runtime;
2406#endif
2407
2408	rq = task_rq_lock(p, &flags);
2409	ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2410	task_rq_unlock(rq, p, &flags);
2411
2412	return ns;
2413}
2414
2415/*
2416 * This function gets called by the timer code, with HZ frequency.
2417 * We call it with interrupts disabled.
2418 */
2419void scheduler_tick(void)
2420{
2421	int cpu = smp_processor_id();
2422	struct rq *rq = cpu_rq(cpu);
2423	struct task_struct *curr = rq->curr;
2424
2425	sched_clock_tick();
2426
2427	raw_spin_lock(&rq->lock);
2428	update_rq_clock(rq);
2429	curr->sched_class->task_tick(rq, curr, 0);
2430	update_cpu_load_active(rq);
2431	raw_spin_unlock(&rq->lock);
2432
2433	perf_event_task_tick();
2434
2435#ifdef CONFIG_SMP
2436	rq->idle_balance = idle_cpu(cpu);
2437	trigger_load_balance(rq);
2438#endif
2439	rq_last_tick_reset(rq);
2440}
2441
2442#ifdef CONFIG_NO_HZ_FULL
2443/**
2444 * scheduler_tick_max_deferment
2445 *
2446 * Keep at least one tick per second when a single
2447 * active task is running because the scheduler doesn't
2448 * yet completely support full dynticks environment.
2449 *
2450 * This makes sure that uptime, CFS vruntime, load
2451 * balancing, etc... continue to move forward, even
2452 * with a very low granularity.
2453 *
2454 * Return: Maximum deferment in nanoseconds.
2455 */
2456u64 scheduler_tick_max_deferment(void)
2457{
2458	struct rq *rq = this_rq();
2459	unsigned long next, now = ACCESS_ONCE(jiffies);
2460
2461	next = rq->last_sched_tick + HZ;
2462
2463	if (time_before_eq(next, now))
2464		return 0;
2465
2466	return jiffies_to_nsecs(next - now);
2467}
2468#endif
2469
2470notrace unsigned long get_parent_ip(unsigned long addr)
2471{
2472	if (in_lock_functions(addr)) {
2473		addr = CALLER_ADDR2;
2474		if (in_lock_functions(addr))
2475			addr = CALLER_ADDR3;
2476	}
2477	return addr;
2478}
2479
2480#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2481				defined(CONFIG_PREEMPT_TRACER))
2482
2483void __kprobes preempt_count_add(int val)
2484{
2485#ifdef CONFIG_DEBUG_PREEMPT
2486	/*
2487	 * Underflow?
2488	 */
2489	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2490		return;
2491#endif
2492	__preempt_count_add(val);
2493#ifdef CONFIG_DEBUG_PREEMPT
2494	/*
2495	 * Spinlock count overflowing soon?
2496	 */
2497	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2498				PREEMPT_MASK - 10);
2499#endif
2500	if (preempt_count() == val) {
2501		unsigned long ip = get_parent_ip(CALLER_ADDR1);
2502#ifdef CONFIG_DEBUG_PREEMPT
2503		current->preempt_disable_ip = ip;
2504#endif
2505		trace_preempt_off(CALLER_ADDR0, ip);
2506	}
2507}
2508EXPORT_SYMBOL(preempt_count_add);
2509
2510void __kprobes preempt_count_sub(int val)
2511{
2512#ifdef CONFIG_DEBUG_PREEMPT
2513	/*
2514	 * Underflow?
2515	 */
2516	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2517		return;
2518	/*
2519	 * Is the spinlock portion underflowing?
2520	 */
2521	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2522			!(preempt_count() & PREEMPT_MASK)))
2523		return;
2524#endif
2525
2526	if (preempt_count() == val)
2527		trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2528	__preempt_count_sub(val);
2529}
2530EXPORT_SYMBOL(preempt_count_sub);
2531
2532#endif
2533
2534/*
2535 * Print scheduling while atomic bug:
2536 */
2537static noinline void __schedule_bug(struct task_struct *prev)
2538{
2539	if (oops_in_progress)
2540		return;
2541
2542	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2543		prev->comm, prev->pid, preempt_count());
2544
2545	debug_show_held_locks(prev);
2546	print_modules();
2547	if (irqs_disabled())
2548		print_irqtrace_events(prev);
2549#ifdef CONFIG_DEBUG_PREEMPT
2550	if (in_atomic_preempt_off()) {
2551		pr_err("Preemption disabled at:");
2552		print_ip_sym(current->preempt_disable_ip);
2553		pr_cont("\n");
2554	}
2555#endif
2556	dump_stack();
2557	add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2558}
2559
2560/*
2561 * Various schedule()-time debugging checks and statistics:
2562 */
2563static inline void schedule_debug(struct task_struct *prev)
2564{
2565	/*
2566	 * Test if we are atomic. Since do_exit() needs to call into
2567	 * schedule() atomically, we ignore that path. Otherwise whine
2568	 * if we are scheduling when we should not.
2569	 */
2570	if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2571		__schedule_bug(prev);
2572	rcu_sleep_check();
2573
2574	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2575
2576	schedstat_inc(this_rq(), sched_count);
2577}
2578
2579/*
2580 * Pick up the highest-prio task:
2581 */
2582static inline struct task_struct *
2583pick_next_task(struct rq *rq, struct task_struct *prev)
2584{
2585	const struct sched_class *class = &fair_sched_class;
2586	struct task_struct *p;
2587
2588	/*
2589	 * Optimization: we know that if all tasks are in
2590	 * the fair class we can call that function directly:
2591	 */
2592	if (likely(prev->sched_class == class &&
2593		   rq->nr_running == rq->cfs.h_nr_running)) {
2594		p = fair_sched_class.pick_next_task(rq, prev);
2595		if (likely(p && p != RETRY_TASK))
2596			return p;
2597	}
2598
2599again:
2600	for_each_class(class) {
2601		p = class->pick_next_task(rq, prev);
2602		if (p) {
2603			if (unlikely(p == RETRY_TASK))
2604				goto again;
2605			return p;
2606		}
2607	}
2608
2609	BUG(); /* the idle class will always have a runnable task */
2610}
2611
2612/*
2613 * __schedule() is the main scheduler function.
2614 *
2615 * The main means of driving the scheduler and thus entering this function are:
2616 *
2617 *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2618 *
2619 *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2620 *      paths. For example, see arch/x86/entry_64.S.
2621 *
2622 *      To drive preemption between tasks, the scheduler sets the flag in timer
2623 *      interrupt handler scheduler_tick().
2624 *
2625 *   3. Wakeups don't really cause entry into schedule(). They add a
2626 *      task to the run-queue and that's it.
2627 *
2628 *      Now, if the new task added to the run-queue preempts the current
2629 *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2630 *      called on the nearest possible occasion:
2631 *
2632 *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2633 *
2634 *         - in syscall or exception context, at the next outmost
2635 *           preempt_enable(). (this might be as soon as the wake_up()'s
2636 *           spin_unlock()!)
2637 *
2638 *         - in IRQ context, return from interrupt-handler to
2639 *           preemptible context
2640 *
2641 *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2642 *         then at the next:
2643 *
2644 *          - cond_resched() call
2645 *          - explicit schedule() call
2646 *          - return from syscall or exception to user-space
2647 *          - return from interrupt-handler to user-space
2648 */
2649static void __sched __schedule(void)
2650{
2651	struct task_struct *prev, *next;
2652	unsigned long *switch_count;
2653	struct rq *rq;
2654	int cpu;
2655
2656need_resched:
2657	preempt_disable();
2658	cpu = smp_processor_id();
2659	rq = cpu_rq(cpu);
2660	rcu_note_context_switch(cpu);
2661	prev = rq->curr;
2662
2663	schedule_debug(prev);
2664
2665	if (sched_feat(HRTICK))
2666		hrtick_clear(rq);
2667
2668	/*
2669	 * Make sure that signal_pending_state()->signal_pending() below
2670	 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2671	 * done by the caller to avoid the race with signal_wake_up().
2672	 */
2673	smp_mb__before_spinlock();
2674	raw_spin_lock_irq(&rq->lock);
2675
2676	switch_count = &prev->nivcsw;
2677	if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2678		if (unlikely(signal_pending_state(prev->state, prev))) {
2679			prev->state = TASK_RUNNING;
2680		} else {
2681			deactivate_task(rq, prev, DEQUEUE_SLEEP);
2682			prev->on_rq = 0;
2683
2684			/*
2685			 * If a worker went to sleep, notify and ask workqueue
2686			 * whether it wants to wake up a task to maintain
2687			 * concurrency.
2688			 */
2689			if (prev->flags & PF_WQ_WORKER) {
2690				struct task_struct *to_wakeup;
2691
2692				to_wakeup = wq_worker_sleeping(prev, cpu);
2693				if (to_wakeup)
2694					try_to_wake_up_local(to_wakeup);
2695			}
2696		}
2697		switch_count = &prev->nvcsw;
2698	}
2699
2700	if (prev->on_rq || rq->skip_clock_update < 0)
2701		update_rq_clock(rq);
2702
2703	next = pick_next_task(rq, prev);
2704	clear_tsk_need_resched(prev);
2705	clear_preempt_need_resched();
2706	rq->skip_clock_update = 0;
2707
2708	if (likely(prev != next)) {
2709		rq->nr_switches++;
2710		rq->curr = next;
2711		++*switch_count;
2712
2713		context_switch(rq, prev, next); /* unlocks the rq */
2714		/*
2715		 * The context switch have flipped the stack from under us
2716		 * and restored the local variables which were saved when
2717		 * this task called schedule() in the past. prev == current
2718		 * is still correct, but it can be moved to another cpu/rq.
2719		 */
2720		cpu = smp_processor_id();
2721		rq = cpu_rq(cpu);
2722	} else
2723		raw_spin_unlock_irq(&rq->lock);
2724
2725	post_schedule(rq);
2726
2727	sched_preempt_enable_no_resched();
2728	if (need_resched())
2729		goto need_resched;
2730}
2731
2732static inline void sched_submit_work(struct task_struct *tsk)
2733{
2734	if (!tsk->state || tsk_is_pi_blocked(tsk))
2735		return;
2736	/*
2737	 * If we are going to sleep and we have plugged IO queued,
2738	 * make sure to submit it to avoid deadlocks.
2739	 */
2740	if (blk_needs_flush_plug(tsk))
2741		blk_schedule_flush_plug(tsk);
2742}
2743
2744asmlinkage void __sched schedule(void)
2745{
2746	struct task_struct *tsk = current;
2747
2748	sched_submit_work(tsk);
2749	__schedule();
2750}
2751EXPORT_SYMBOL(schedule);
2752
2753#ifdef CONFIG_CONTEXT_TRACKING
2754asmlinkage void __sched schedule_user(void)
2755{
2756	/*
2757	 * If we come here after a random call to set_need_resched(),
2758	 * or we have been woken up remotely but the IPI has not yet arrived,
2759	 * we haven't yet exited the RCU idle mode. Do it here manually until
2760	 * we find a better solution.
2761	 */
2762	user_exit();
2763	schedule();
2764	user_enter();
2765}
2766#endif
2767
2768/**
2769 * schedule_preempt_disabled - called with preemption disabled
2770 *
2771 * Returns with preemption disabled. Note: preempt_count must be 1
2772 */
2773void __sched schedule_preempt_disabled(void)
2774{
2775	sched_preempt_enable_no_resched();
2776	schedule();
2777	preempt_disable();
2778}
2779
2780#ifdef CONFIG_PREEMPT
2781/*
2782 * this is the entry point to schedule() from in-kernel preemption
2783 * off of preempt_enable. Kernel preemptions off return from interrupt
2784 * occur there and call schedule directly.
2785 */
2786asmlinkage void __sched notrace preempt_schedule(void)
2787{
2788	/*
2789	 * If there is a non-zero preempt_count or interrupts are disabled,
2790	 * we do not want to preempt the current task. Just return..
2791	 */
2792	if (likely(!preemptible()))
2793		return;
2794
2795	do {
2796		__preempt_count_add(PREEMPT_ACTIVE);
2797		__schedule();
2798		__preempt_count_sub(PREEMPT_ACTIVE);
2799
2800		/*
2801		 * Check again in case we missed a preemption opportunity
2802		 * between schedule and now.
2803		 */
2804		barrier();
2805	} while (need_resched());
2806}
2807EXPORT_SYMBOL(preempt_schedule);
2808#endif /* CONFIG_PREEMPT */
2809
2810/*
2811 * this is the entry point to schedule() from kernel preemption
2812 * off of irq context.
2813 * Note, that this is called and return with irqs disabled. This will
2814 * protect us against recursive calling from irq.
2815 */
2816asmlinkage void __sched preempt_schedule_irq(void)
2817{
2818	enum ctx_state prev_state;
2819
2820	/* Catch callers which need to be fixed */
2821	BUG_ON(preempt_count() || !irqs_disabled());
2822
2823	prev_state = exception_enter();
2824
2825	do {
2826		__preempt_count_add(PREEMPT_ACTIVE);
2827		local_irq_enable();
2828		__schedule();
2829		local_irq_disable();
2830		__preempt_count_sub(PREEMPT_ACTIVE);
2831
2832		/*
2833		 * Check again in case we missed a preemption opportunity
2834		 * between schedule and now.
2835		 */
2836		barrier();
2837	} while (need_resched());
2838
2839	exception_exit(prev_state);
2840}
2841
2842int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2843			  void *key)
2844{
2845	return try_to_wake_up(curr->private, mode, wake_flags);
2846}
2847EXPORT_SYMBOL(default_wake_function);
2848
2849static long __sched
2850sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2851{
2852	unsigned long flags;
2853	wait_queue_t wait;
2854
2855	init_waitqueue_entry(&wait, current);
2856
2857	__set_current_state(state);
2858
2859	spin_lock_irqsave(&q->lock, flags);
2860	__add_wait_queue(q, &wait);
2861	spin_unlock(&q->lock);
2862	timeout = schedule_timeout(timeout);
2863	spin_lock_irq(&q->lock);
2864	__remove_wait_queue(q, &wait);
2865	spin_unlock_irqrestore(&q->lock, flags);
2866
2867	return timeout;
2868}
2869
2870void __sched interruptible_sleep_on(wait_queue_head_t *q)
2871{
2872	sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2873}
2874EXPORT_SYMBOL(interruptible_sleep_on);
2875
2876long __sched
2877interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2878{
2879	return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2880}
2881EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2882
2883void __sched sleep_on(wait_queue_head_t *q)
2884{
2885	sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2886}
2887EXPORT_SYMBOL(sleep_on);
2888
2889long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2890{
2891	return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2892}
2893EXPORT_SYMBOL(sleep_on_timeout);
2894
2895#ifdef CONFIG_RT_MUTEXES
2896
2897/*
2898 * rt_mutex_setprio - set the current priority of a task
2899 * @p: task
2900 * @prio: prio value (kernel-internal form)
2901 *
2902 * This function changes the 'effective' priority of a task. It does
2903 * not touch ->normal_prio like __setscheduler().
2904 *
2905 * Used by the rt_mutex code to implement priority inheritance
2906 * logic. Call site only calls if the priority of the task changed.
2907 */
2908void rt_mutex_setprio(struct task_struct *p, int prio)
2909{
2910	int oldprio, on_rq, running, enqueue_flag = 0;
2911	struct rq *rq;
2912	const struct sched_class *prev_class;
2913
2914	BUG_ON(prio > MAX_PRIO);
2915
2916	rq = __task_rq_lock(p);
2917
2918	/*
2919	 * Idle task boosting is a nono in general. There is one
2920	 * exception, when PREEMPT_RT and NOHZ is active:
2921	 *
2922	 * The idle task calls get_next_timer_interrupt() and holds
2923	 * the timer wheel base->lock on the CPU and another CPU wants
2924	 * to access the timer (probably to cancel it). We can safely
2925	 * ignore the boosting request, as the idle CPU runs this code
2926	 * with interrupts disabled and will complete the lock
2927	 * protected section without being interrupted. So there is no
2928	 * real need to boost.
2929	 */
2930	if (unlikely(p == rq->idle)) {
2931		WARN_ON(p != rq->curr);
2932		WARN_ON(p->pi_blocked_on);
2933		goto out_unlock;
2934	}
2935
2936	trace_sched_pi_setprio(p, prio);
2937	p->pi_top_task = rt_mutex_get_top_task(p);
2938	oldprio = p->prio;
2939	prev_class = p->sched_class;
2940	on_rq = p->on_rq;
2941	running = task_current(rq, p);
2942	if (on_rq)
2943		dequeue_task(rq, p, 0);
2944	if (running)
2945		p->sched_class->put_prev_task(rq, p);
2946
2947	/*
2948	 * Boosting condition are:
2949	 * 1. -rt task is running and holds mutex A
2950	 *      --> -dl task blocks on mutex A
2951	 *
2952	 * 2. -dl task is running and holds mutex A
2953	 *      --> -dl task blocks on mutex A and could preempt the
2954	 *          running task
2955	 */
2956	if (dl_prio(prio)) {
2957		if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2958			dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2959			p->dl.dl_boosted = 1;
2960			p->dl.dl_throttled = 0;
2961			enqueue_flag = ENQUEUE_REPLENISH;
2962		} else
2963			p->dl.dl_boosted = 0;
2964		p->sched_class = &dl_sched_class;
2965	} else if (rt_prio(prio)) {
2966		if (dl_prio(oldprio))
2967			p->dl.dl_boosted = 0;
2968		if (oldprio < prio)
2969			enqueue_flag = ENQUEUE_HEAD;
2970		p->sched_class = &rt_sched_class;
2971	} else {
2972		if (dl_prio(oldprio))
2973			p->dl.dl_boosted = 0;
2974		p->sched_class = &fair_sched_class;
2975	}
2976
2977	p->prio = prio;
2978
2979	if (running)
2980		p->sched_class->set_curr_task(rq);
2981	if (on_rq)
2982		enqueue_task(rq, p, enqueue_flag);
2983
2984	check_class_changed(rq, p, prev_class, oldprio);
2985out_unlock:
2986	__task_rq_unlock(rq);
2987}
2988#endif
2989
2990void set_user_nice(struct task_struct *p, long nice)
2991{
2992	int old_prio, delta, on_rq;
2993	unsigned long flags;
2994	struct rq *rq;
2995
2996	if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
2997		return;
2998	/*
2999	 * We have to be careful, if called from sys_setpriority(),
3000	 * the task might be in the middle of scheduling on another CPU.
3001	 */
3002	rq = task_rq_lock(p, &flags);
3003	/*
3004	 * The RT priorities are set via sched_setscheduler(), but we still
3005	 * allow the 'normal' nice value to be set - but as expected
3006	 * it wont have any effect on scheduling until the task is
3007	 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3008	 */
3009	if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3010		p->static_prio = NICE_TO_PRIO(nice);
3011		goto out_unlock;
3012	}
3013	on_rq = p->on_rq;
3014	if (on_rq)
3015		dequeue_task(rq, p, 0);
3016
3017	p->static_prio = NICE_TO_PRIO(nice);
3018	set_load_weight(p);
3019	old_prio = p->prio;
3020	p->prio = effective_prio(p);
3021	delta = p->prio - old_prio;
3022
3023	if (on_rq) {
3024		enqueue_task(rq, p, 0);
3025		/*
3026		 * If the task increased its priority or is running and
3027		 * lowered its priority, then reschedule its CPU:
3028		 */
3029		if (delta < 0 || (delta > 0 && task_running(rq, p)))
3030			resched_task(rq->curr);
3031	}
3032out_unlock:
3033	task_rq_unlock(rq, p, &flags);
3034}
3035EXPORT_SYMBOL(set_user_nice);
3036
3037/*
3038 * can_nice - check if a task can reduce its nice value
3039 * @p: task
3040 * @nice: nice value
3041 */
3042int can_nice(const struct task_struct *p, const int nice)
3043{
3044	/* convert nice value [19,-20] to rlimit style value [1,40] */
3045	int nice_rlim = 20 - nice;
3046
3047	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3048		capable(CAP_SYS_NICE));
3049}
3050
3051#ifdef __ARCH_WANT_SYS_NICE
3052
3053/*
3054 * sys_nice - change the priority of the current process.
3055 * @increment: priority increment
3056 *
3057 * sys_setpriority is a more generic, but much slower function that
3058 * does similar things.
3059 */
3060SYSCALL_DEFINE1(nice, int, increment)
3061{
3062	long nice, retval;
3063
3064	/*
3065	 * Setpriority might change our priority at the same moment.
3066	 * We don't have to worry. Conceptually one call occurs first
3067	 * and we have a single winner.
3068	 */
3069	if (increment < -40)
3070		increment = -40;
3071	if (increment > 40)
3072		increment = 40;
3073
3074	nice = task_nice(current) + increment;
3075	if (nice < MIN_NICE)
3076		nice = MIN_NICE;
3077	if (nice > MAX_NICE)
3078		nice = MAX_NICE;
3079
3080	if (increment < 0 && !can_nice(current, nice))
3081		return -EPERM;
3082
3083	retval = security_task_setnice(current, nice);
3084	if (retval)
3085		return retval;
3086
3087	set_user_nice(current, nice);
3088	return 0;
3089}
3090
3091#endif
3092
3093/**
3094 * task_prio - return the priority value of a given task.
3095 * @p: the task in question.
3096 *
3097 * Return: The priority value as seen by users in /proc.
3098 * RT tasks are offset by -200. Normal tasks are centered
3099 * around 0, value goes from -16 to +15.
3100 */
3101int task_prio(const struct task_struct *p)
3102{
3103	return p->prio - MAX_RT_PRIO;
3104}
3105
3106/**
3107 * idle_cpu - is a given cpu idle currently?
3108 * @cpu: the processor in question.
3109 *
3110 * Return: 1 if the CPU is currently idle. 0 otherwise.
3111 */
3112int idle_cpu(int cpu)
3113{
3114	struct rq *rq = cpu_rq(cpu);
3115
3116	if (rq->curr != rq->idle)
3117		return 0;
3118
3119	if (rq->nr_running)
3120		return 0;
3121
3122#ifdef CONFIG_SMP
3123	if (!llist_empty(&rq->wake_list))
3124		return 0;
3125#endif
3126
3127	return 1;
3128}
3129
3130/**
3131 * idle_task - return the idle task for a given cpu.
3132 * @cpu: the processor in question.
3133 *
3134 * Return: The idle task for the cpu @cpu.
3135 */
3136struct task_struct *idle_task(int cpu)
3137{
3138	return cpu_rq(cpu)->idle;
3139}
3140
3141/**
3142 * find_process_by_pid - find a process with a matching PID value.
3143 * @pid: the pid in question.
3144 *
3145 * The task of @pid, if found. %NULL otherwise.
3146 */
3147static struct task_struct *find_process_by_pid(pid_t pid)
3148{
3149	return pid ? find_task_by_vpid(pid) : current;
3150}
3151
3152/*
3153 * This function initializes the sched_dl_entity of a newly becoming
3154 * SCHED_DEADLINE task.
3155 *
3156 * Only the static values are considered here, the actual runtime and the
3157 * absolute deadline will be properly calculated when the task is enqueued
3158 * for the first time with its new policy.
3159 */
3160static void
3161__setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3162{
3163	struct sched_dl_entity *dl_se = &p->dl;
3164
3165	init_dl_task_timer(dl_se);
3166	dl_se->dl_runtime = attr->sched_runtime;
3167	dl_se->dl_deadline = attr->sched_deadline;
3168	dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3169	dl_se->flags = attr->sched_flags;
3170	dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3171	dl_se->dl_throttled = 0;
3172	dl_se->dl_new = 1;
3173}
3174
3175static void __setscheduler_params(struct task_struct *p,
3176		const struct sched_attr *attr)
3177{
3178	int policy = attr->sched_policy;
3179
3180	if (policy == -1) /* setparam */
3181		policy = p->policy;
3182
3183	p->policy = policy;
3184
3185	if (dl_policy(policy))
3186		__setparam_dl(p, attr);
3187	else if (fair_policy(policy))
3188		p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3189
3190	/*
3191	 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3192	 * !rt_policy. Always setting this ensures that things like
3193	 * getparam()/getattr() don't report silly values for !rt tasks.
3194	 */
3195	p->rt_priority = attr->sched_priority;
3196	p->normal_prio = normal_prio(p);
3197	set_load_weight(p);
3198}
3199
3200/* Actually do priority change: must hold pi & rq lock. */
3201static void __setscheduler(struct rq *rq, struct task_struct *p,
3202			   const struct sched_attr *attr)
3203{
3204	__setscheduler_params(p, attr);
3205
3206	/*
3207	 * If we get here, there was no pi waiters boosting the
3208	 * task. It is safe to use the normal prio.
3209	 */
3210	p->prio = normal_prio(p);
3211
3212	if (dl_prio(p->prio))
3213		p->sched_class = &dl_sched_class;
3214	else if (rt_prio(p->prio))
3215		p->sched_class = &rt_sched_class;
3216	else
3217		p->sched_class = &fair_sched_class;
3218}
3219
3220static void
3221__getparam_dl(struct task_struct *p, struct sched_attr *attr)
3222{
3223	struct sched_dl_entity *dl_se = &p->dl;
3224
3225	attr->sched_priority = p->rt_priority;
3226	attr->sched_runtime = dl_se->dl_runtime;
3227	attr->sched_deadline = dl_se->dl_deadline;
3228	attr->sched_period = dl_se->dl_period;
3229	attr->sched_flags = dl_se->flags;
3230}
3231
3232/*
3233 * This function validates the new parameters of a -deadline task.
3234 * We ask for the deadline not being zero, and greater or equal
3235 * than the runtime, as well as the period of being zero or
3236 * greater than deadline. Furthermore, we have to be sure that
3237 * user parameters are above the internal resolution (1us); we
3238 * check sched_runtime only since it is always the smaller one.
3239 */
3240static bool
3241__checkparam_dl(const struct sched_attr *attr)
3242{
3243	return attr && attr->sched_deadline != 0 &&
3244		(attr->sched_period == 0 ||
3245		(s64)(attr->sched_period   - attr->sched_deadline) >= 0) &&
3246		(s64)(attr->sched_deadline - attr->sched_runtime ) >= 0  &&
3247		attr->sched_runtime >= (2 << (DL_SCALE - 1));
3248}
3249
3250/*
3251 * check the target process has a UID that matches the current process's
3252 */
3253static bool check_same_owner(struct task_struct *p)
3254{
3255	const struct cred *cred = current_cred(), *pcred;
3256	bool match;
3257
3258	rcu_read_lock();
3259	pcred = __task_cred(p);
3260	match = (uid_eq(cred->euid, pcred->euid) ||
3261		 uid_eq(cred->euid, pcred->uid));
3262	rcu_read_unlock();
3263	return match;
3264}
3265
3266static int __sched_setscheduler(struct task_struct *p,
3267				const struct sched_attr *attr,
3268				bool user)
3269{
3270	int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3271		      MAX_RT_PRIO - 1 - attr->sched_priority;
3272	int retval, oldprio, oldpolicy = -1, on_rq, running;
3273	int policy = attr->sched_policy;
3274	unsigned long flags;
3275	const struct sched_class *prev_class;
3276	struct rq *rq;
3277	int reset_on_fork;
3278
3279	/* may grab non-irq protected spin_locks */
3280	BUG_ON(in_interrupt());
3281recheck:
3282	/* double check policy once rq lock held */
3283	if (policy < 0) {
3284		reset_on_fork = p->sched_reset_on_fork;
3285		policy = oldpolicy = p->policy;
3286	} else {
3287		reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3288
3289		if (policy != SCHED_DEADLINE &&
3290				policy != SCHED_FIFO && policy != SCHED_RR &&
3291				policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3292				policy != SCHED_IDLE)
3293			return -EINVAL;
3294	}
3295
3296	if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3297		return -EINVAL;
3298
3299	/*
3300	 * Valid priorities for SCHED_FIFO and SCHED_RR are
3301	 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3302	 * SCHED_BATCH and SCHED_IDLE is 0.
3303	 */
3304	if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3305	    (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3306		return -EINVAL;
3307	if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3308	    (rt_policy(policy) != (attr->sched_priority != 0)))
3309		return -EINVAL;
3310
3311	/*
3312	 * Allow unprivileged RT tasks to decrease priority:
3313	 */
3314	if (user && !capable(CAP_SYS_NICE)) {
3315		if (fair_policy(policy)) {
3316			if (attr->sched_nice < task_nice(p) &&
3317			    !can_nice(p, attr->sched_nice))
3318				return -EPERM;
3319		}
3320
3321		if (rt_policy(policy)) {
3322			unsigned long rlim_rtprio =
3323					task_rlimit(p, RLIMIT_RTPRIO);
3324
3325			/* can't set/change the rt policy */
3326			if (policy != p->policy && !rlim_rtprio)
3327				return -EPERM;
3328
3329			/* can't increase priority */
3330			if (attr->sched_priority > p->rt_priority &&
3331			    attr->sched_priority > rlim_rtprio)
3332				return -EPERM;
3333		}
3334
3335		 /*
3336		  * Can't set/change SCHED_DEADLINE policy at all for now
3337		  * (safest behavior); in the future we would like to allow
3338		  * unprivileged DL tasks to increase their relative deadline
3339		  * or reduce their runtime (both ways reducing utilization)
3340		  */
3341		if (dl_policy(policy))
3342			return -EPERM;
3343
3344		/*
3345		 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3346		 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3347		 */
3348		if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3349			if (!can_nice(p, task_nice(p)))
3350				return -EPERM;
3351		}
3352
3353		/* can't change other user's priorities */
3354		if (!check_same_owner(p))
3355			return -EPERM;
3356
3357		/* Normal users shall not reset the sched_reset_on_fork flag */
3358		if (p->sched_reset_on_fork && !reset_on_fork)
3359			return -EPERM;
3360	}
3361
3362	if (user) {
3363		retval = security_task_setscheduler(p);
3364		if (retval)
3365			return retval;
3366	}
3367
3368	/*
3369	 * make sure no PI-waiters arrive (or leave) while we are
3370	 * changing the priority of the task:
3371	 *
3372	 * To be able to change p->policy safely, the appropriate
3373	 * runqueue lock must be held.
3374	 */
3375	rq = task_rq_lock(p, &flags);
3376
3377	/*
3378	 * Changing the policy of the stop threads its a very bad idea
3379	 */
3380	if (p == rq->stop) {
3381		task_rq_unlock(rq, p, &flags);
3382		return -EINVAL;
3383	}
3384
3385	/*
3386	 * If not changing anything there's no need to proceed further,
3387	 * but store a possible modification of reset_on_fork.
3388	 */
3389	if (unlikely(policy == p->policy)) {
3390		if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3391			goto change;
3392		if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3393			goto change;
3394		if (dl_policy(policy))
3395			goto change;
3396
3397		p->sched_reset_on_fork = reset_on_fork;
3398		task_rq_unlock(rq, p, &flags);
3399		return 0;
3400	}
3401change:
3402
3403	if (user) {
3404#ifdef CONFIG_RT_GROUP_SCHED
3405		/*
3406		 * Do not allow realtime tasks into groups that have no runtime
3407		 * assigned.
3408		 */
3409		if (rt_bandwidth_enabled() && rt_policy(policy) &&
3410				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3411				!task_group_is_autogroup(task_group(p))) {
3412			task_rq_unlock(rq, p, &flags);
3413			return -EPERM;
3414		}
3415#endif
3416#ifdef CONFIG_SMP
3417		if (dl_bandwidth_enabled() && dl_policy(policy)) {
3418			cpumask_t *span = rq->rd->span;
3419
3420			/*
3421			 * Don't allow tasks with an affinity mask smaller than
3422			 * the entire root_domain to become SCHED_DEADLINE. We
3423			 * will also fail if there's no bandwidth available.
3424			 */
3425			if (!cpumask_subset(span, &p->cpus_allowed) ||
3426			    rq->rd->dl_bw.bw == 0) {
3427				task_rq_unlock(rq, p, &flags);
3428				return -EPERM;
3429			}
3430		}
3431#endif
3432	}
3433
3434	/* recheck policy now with rq lock held */
3435	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3436		policy = oldpolicy = -1;
3437		task_rq_unlock(rq, p, &flags);
3438		goto recheck;
3439	}
3440
3441	/*
3442	 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3443	 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3444	 * is available.
3445	 */
3446	if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3447		task_rq_unlock(rq, p, &flags);
3448		return -EBUSY;
3449	}
3450
3451	p->sched_reset_on_fork = reset_on_fork;
3452	oldprio = p->prio;
3453
3454	/*
3455	 * Special case for priority boosted tasks.
3456	 *
3457	 * If the new priority is lower or equal (user space view)
3458	 * than the current (boosted) priority, we just store the new
3459	 * normal parameters and do not touch the scheduler class and
3460	 * the runqueue. This will be done when the task deboost
3461	 * itself.
3462	 */
3463	if (rt_mutex_check_prio(p, newprio)) {
3464		__setscheduler_params(p, attr);
3465		task_rq_unlock(rq, p, &flags);
3466		return 0;
3467	}
3468
3469	on_rq = p->on_rq;
3470	running = task_current(rq, p);
3471	if (on_rq)
3472		dequeue_task(rq, p, 0);
3473	if (running)
3474		p->sched_class->put_prev_task(rq, p);
3475
3476	prev_class = p->sched_class;
3477	__setscheduler(rq, p, attr);
3478
3479	if (running)
3480		p->sched_class->set_curr_task(rq);
3481	if (on_rq) {
3482		/*
3483		 * We enqueue to tail when the priority of a task is
3484		 * increased (user space view).
3485		 */
3486		enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3487	}
3488
3489	check_class_changed(rq, p, prev_class, oldprio);
3490	task_rq_unlock(rq, p, &flags);
3491
3492	rt_mutex_adjust_pi(p);
3493
3494	return 0;
3495}
3496
3497static int _sched_setscheduler(struct task_struct *p, int policy,
3498			       const struct sched_param *param, bool check)
3499{
3500	struct sched_attr attr = {
3501		.sched_policy   = policy,
3502		.sched_priority = param->sched_priority,
3503		.sched_nice	= PRIO_TO_NICE(p->static_prio),
3504	};
3505
3506	/*
3507	 * Fixup the legacy SCHED_RESET_ON_FORK hack
3508	 */
3509	if (policy & SCHED_RESET_ON_FORK) {
3510		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3511		policy &= ~SCHED_RESET_ON_FORK;
3512		attr.sched_policy = policy;
3513	}
3514
3515	return __sched_setscheduler(p, &attr, check);
3516}
3517/**
3518 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3519 * @p: the task in question.
3520 * @policy: new policy.
3521 * @param: structure containing the new RT priority.
3522 *
3523 * Return: 0 on success. An error code otherwise.
3524 *
3525 * NOTE that the task may be already dead.
3526 */
3527int sched_setscheduler(struct task_struct *p, int policy,
3528		       const struct sched_param *param)
3529{
3530	return _sched_setscheduler(p, policy, param, true);
3531}
3532EXPORT_SYMBOL_GPL(sched_setscheduler);
3533
3534int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3535{
3536	return __sched_setscheduler(p, attr, true);
3537}
3538EXPORT_SYMBOL_GPL(sched_setattr);
3539
3540/**
3541 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3542 * @p: the task in question.
3543 * @policy: new policy.
3544 * @param: structure containing the new RT priority.
3545 *
3546 * Just like sched_setscheduler, only don't bother checking if the
3547 * current context has permission.  For example, this is needed in
3548 * stop_machine(): we create temporary high priority worker threads,
3549 * but our caller might not have that capability.
3550 *
3551 * Return: 0 on success. An error code otherwise.
3552 */
3553int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3554			       const struct sched_param *param)
3555{
3556	return _sched_setscheduler(p, policy, param, false);
3557}
3558
3559static int
3560do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3561{
3562	struct sched_param lparam;
3563	struct task_struct *p;
3564	int retval;
3565
3566	if (!param || pid < 0)
3567		return -EINVAL;
3568	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3569		return -EFAULT;
3570
3571	rcu_read_lock();
3572	retval = -ESRCH;
3573	p = find_process_by_pid(pid);
3574	if (p != NULL)
3575		retval = sched_setscheduler(p, policy, &lparam);
3576	rcu_read_unlock();
3577
3578	return retval;
3579}
3580
3581/*
3582 * Mimics kernel/events/core.c perf_copy_attr().
3583 */
3584static int sched_copy_attr(struct sched_attr __user *uattr,
3585			   struct sched_attr *attr)
3586{
3587	u32 size;
3588	int ret;
3589
3590	if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3591		return -EFAULT;
3592
3593	/*
3594	 * zero the full structure, so that a short copy will be nice.
3595	 */
3596	memset(attr, 0, sizeof(*attr));
3597
3598	ret = get_user(size, &uattr->size);
3599	if (ret)
3600		return ret;
3601
3602	if (size > PAGE_SIZE)	/* silly large */
3603		goto err_size;
3604
3605	if (!size)		/* abi compat */
3606		size = SCHED_ATTR_SIZE_VER0;
3607
3608	if (size < SCHED_ATTR_SIZE_VER0)
3609		goto err_size;
3610
3611	/*
3612	 * If we're handed a bigger struct than we know of,
3613	 * ensure all the unknown bits are 0 - i.e. new
3614	 * user-space does not rely on any kernel feature
3615	 * extensions we dont know about yet.
3616	 */
3617	if (size > sizeof(*attr)) {
3618		unsigned char __user *addr;
3619		unsigned char __user *end;
3620		unsigned char val;
3621
3622		addr = (void __user *)uattr + sizeof(*attr);
3623		end  = (void __user *)uattr + size;
3624
3625		for (; addr < end; addr++) {
3626			ret = get_user(val, addr);
3627			if (ret)
3628				return ret;
3629			if (val)
3630				goto err_size;
3631		}
3632		size = sizeof(*attr);
3633	}
3634
3635	ret = copy_from_user(attr, uattr, size);
3636	if (ret)
3637		return -EFAULT;
3638
3639	/*
3640	 * XXX: do we want to be lenient like existing syscalls; or do we want
3641	 * to be strict and return an error on out-of-bounds values?
3642	 */
3643	attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3644
3645out:
3646	return ret;
3647
3648err_size:
3649	put_user(sizeof(*attr), &uattr->size);
3650	ret = -E2BIG;
3651	goto out;
3652}
3653
3654/**
3655 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3656 * @pid: the pid in question.
3657 * @policy: new policy.
3658 * @param: structure containing the new RT priority.
3659 *
3660 * Return: 0 on success. An error code otherwise.
3661 */
3662SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3663		struct sched_param __user *, param)
3664{
3665	/* negative values for policy are not valid */
3666	if (policy < 0)
3667		return -EINVAL;
3668
3669	return do_sched_setscheduler(pid, policy, param);
3670}
3671
3672/**
3673 * sys_sched_setparam - set/change the RT priority of a thread
3674 * @pid: the pid in question.
3675 * @param: structure containing the new RT priority.
3676 *
3677 * Return: 0 on success. An error code otherwise.
3678 */
3679SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3680{
3681	return do_sched_setscheduler(pid, -1, param);
3682}
3683
3684/**
3685 * sys_sched_setattr - same as above, but with extended sched_attr
3686 * @pid: the pid in question.
3687 * @uattr: structure containing the extended parameters.
3688 */
3689SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3690			       unsigned int, flags)
3691{
3692	struct sched_attr attr;
3693	struct task_struct *p;
3694	int retval;
3695
3696	if (!uattr || pid < 0 || flags)
3697		return -EINVAL;
3698
3699	if (sched_copy_attr(uattr, &attr))
3700		return -EFAULT;
3701
3702	rcu_read_lock();
3703	retval = -ESRCH;
3704	p = find_process_by_pid(pid);
3705	if (p != NULL)
3706		retval = sched_setattr(p, &attr);
3707	rcu_read_unlock();
3708
3709	return retval;
3710}
3711
3712/**
3713 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3714 * @pid: the pid in question.
3715 *
3716 * Return: On success, the policy of the thread. Otherwise, a negative error
3717 * code.
3718 */
3719SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3720{
3721	struct task_struct *p;
3722	int retval;
3723
3724	if (pid < 0)
3725		return -EINVAL;
3726
3727	retval = -ESRCH;
3728	rcu_read_lock();
3729	p = find_process_by_pid(pid);
3730	if (p) {
3731		retval = security_task_getscheduler(p);
3732		if (!retval)
3733			retval = p->policy
3734				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3735	}
3736	rcu_read_unlock();
3737	return retval;
3738}
3739
3740/**
3741 * sys_sched_getparam - get the RT priority of a thread
3742 * @pid: the pid in question.
3743 * @param: structure containing the RT priority.
3744 *
3745 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3746 * code.
3747 */
3748SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3749{
3750	struct sched_param lp;
3751	struct task_struct *p;
3752	int retval;
3753
3754	if (!param || pid < 0)
3755		return -EINVAL;
3756
3757	rcu_read_lock();
3758	p = find_process_by_pid(pid);
3759	retval = -ESRCH;
3760	if (!p)
3761		goto out_unlock;
3762
3763	retval = security_task_getscheduler(p);
3764	if (retval)
3765		goto out_unlock;
3766
3767	if (task_has_dl_policy(p)) {
3768		retval = -EINVAL;
3769		goto out_unlock;
3770	}
3771	lp.sched_priority = p->rt_priority;
3772	rcu_read_unlock();
3773
3774	/*
3775	 * This one might sleep, we cannot do it with a spinlock held ...
3776	 */
3777	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3778
3779	return retval;
3780
3781out_unlock:
3782	rcu_read_unlock();
3783	return retval;
3784}
3785
3786static int sched_read_attr(struct sched_attr __user *uattr,
3787			   struct sched_attr *attr,
3788			   unsigned int usize)
3789{
3790	int ret;
3791
3792	if (!access_ok(VERIFY_WRITE, uattr, usize))
3793		return -EFAULT;
3794
3795	/*
3796	 * If we're handed a smaller struct than we know of,
3797	 * ensure all the unknown bits are 0 - i.e. old
3798	 * user-space does not get uncomplete information.
3799	 */
3800	if (usize < sizeof(*attr)) {
3801		unsigned char *addr;
3802		unsigned char *end;
3803
3804		addr = (void *)attr + usize;
3805		end  = (void *)attr + sizeof(*attr);
3806
3807		for (; addr < end; addr++) {
3808			if (*addr)
3809				goto err_size;
3810		}
3811
3812		attr->size = usize;
3813	}
3814
3815	ret = copy_to_user(uattr, attr, attr->size);
3816	if (ret)
3817		return -EFAULT;
3818
3819out:
3820	return ret;
3821
3822err_size:
3823	ret = -E2BIG;
3824	goto out;
3825}
3826
3827/**
3828 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3829 * @pid: the pid in question.
3830 * @uattr: structure containing the extended parameters.
3831 * @size: sizeof(attr) for fwd/bwd comp.
3832 */
3833SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3834		unsigned int, size, unsigned int, flags)
3835{
3836	struct sched_attr attr = {
3837		.size = sizeof(struct sched_attr),
3838	};
3839	struct task_struct *p;
3840	int retval;
3841
3842	if (!uattr || pid < 0 || size > PAGE_SIZE ||
3843	    size < SCHED_ATTR_SIZE_VER0 || flags)
3844		return -EINVAL;
3845
3846	rcu_read_lock();
3847	p = find_process_by_pid(pid);
3848	retval = -ESRCH;
3849	if (!p)
3850		goto out_unlock;
3851
3852	retval = security_task_getscheduler(p);
3853	if (retval)
3854		goto out_unlock;
3855
3856	attr.sched_policy = p->policy;
3857	if (p->sched_reset_on_fork)
3858		attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3859	if (task_has_dl_policy(p))
3860		__getparam_dl(p, &attr);
3861	else if (task_has_rt_policy(p))
3862		attr.sched_priority = p->rt_priority;
3863	else
3864		attr.sched_nice = task_nice(p);
3865
3866	rcu_read_unlock();
3867
3868	retval = sched_read_attr(uattr, &attr, size);
3869	return retval;
3870
3871out_unlock:
3872	rcu_read_unlock();
3873	return retval;
3874}
3875
3876long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3877{
3878	cpumask_var_t cpus_allowed, new_mask;
3879	struct task_struct *p;
3880	int retval;
3881
3882	rcu_read_lock();
3883
3884	p = find_process_by_pid(pid);
3885	if (!p) {
3886		rcu_read_unlock();
3887		return -ESRCH;
3888	}
3889
3890	/* Prevent p going away */
3891	get_task_struct(p);
3892	rcu_read_unlock();
3893
3894	if (p->flags & PF_NO_SETAFFINITY) {
3895		retval = -EINVAL;
3896		goto out_put_task;
3897	}
3898	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3899		retval = -ENOMEM;
3900		goto out_put_task;
3901	}
3902	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3903		retval = -ENOMEM;
3904		goto out_free_cpus_allowed;
3905	}
3906	retval = -EPERM;
3907	if (!check_same_owner(p)) {
3908		rcu_read_lock();
3909		if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3910			rcu_read_unlock();
3911			goto out_unlock;
3912		}
3913		rcu_read_unlock();
3914	}
3915
3916	retval = security_task_setscheduler(p);
3917	if (retval)
3918		goto out_unlock;
3919
3920
3921	cpuset_cpus_allowed(p, cpus_allowed);
3922	cpumask_and(new_mask, in_mask, cpus_allowed);
3923
3924	/*
3925	 * Since bandwidth control happens on root_domain basis,
3926	 * if admission test is enabled, we only admit -deadline
3927	 * tasks allowed to run on all the CPUs in the task's
3928	 * root_domain.
3929	 */
3930#ifdef CONFIG_SMP
3931	if (task_has_dl_policy(p)) {
3932		const struct cpumask *span = task_rq(p)->rd->span;
3933
3934		if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3935			retval = -EBUSY;
3936			goto out_unlock;
3937		}
3938	}
3939#endif
3940again:
3941	retval = set_cpus_allowed_ptr(p, new_mask);
3942
3943	if (!retval) {
3944		cpuset_cpus_allowed(p, cpus_allowed);
3945		if (!cpumask_subset(new_mask, cpus_allowed)) {
3946			/*
3947			 * We must have raced with a concurrent cpuset
3948			 * update. Just reset the cpus_allowed to the
3949			 * cpuset's cpus_allowed
3950			 */
3951			cpumask_copy(new_mask, cpus_allowed);
3952			goto again;
3953		}
3954	}
3955out_unlock:
3956	free_cpumask_var(new_mask);
3957out_free_cpus_allowed:
3958	free_cpumask_var(cpus_allowed);
3959out_put_task:
3960	put_task_struct(p);
3961	return retval;
3962}
3963
3964static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3965			     struct cpumask *new_mask)
3966{
3967	if (len < cpumask_size())
3968		cpumask_clear(new_mask);
3969	else if (len > cpumask_size())
3970		len = cpumask_size();
3971
3972	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3973}
3974
3975/**
3976 * sys_sched_setaffinity - set the cpu affinity of a process
3977 * @pid: pid of the process
3978 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3979 * @user_mask_ptr: user-space pointer to the new cpu mask
3980 *
3981 * Return: 0 on success. An error code otherwise.
3982 */
3983SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3984		unsigned long __user *, user_mask_ptr)
3985{
3986	cpumask_var_t new_mask;
3987	int retval;
3988
3989	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3990		return -ENOMEM;
3991
3992	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3993	if (retval == 0)
3994		retval = sched_setaffinity(pid, new_mask);
3995	free_cpumask_var(new_mask);
3996	return retval;
3997}
3998
3999long sched_getaffinity(pid_t pid, struct cpumask *mask)
4000{
4001	struct task_struct *p;
4002	unsigned long flags;
4003	int retval;
4004
4005	rcu_read_lock();
4006
4007	retval = -ESRCH;
4008	p = find_process_by_pid(pid);
4009	if (!p)
4010		goto out_unlock;
4011
4012	retval = security_task_getscheduler(p);
4013	if (retval)
4014		goto out_unlock;
4015
4016	raw_spin_lock_irqsave(&p->pi_lock, flags);
4017	cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4018	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4019
4020out_unlock:
4021	rcu_read_unlock();
4022
4023	return retval;
4024}
4025
4026/**
4027 * sys_sched_getaffinity - get the cpu affinity of a process
4028 * @pid: pid of the process
4029 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4030 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4031 *
4032 * Return: 0 on success. An error code otherwise.
4033 */
4034SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4035		unsigned long __user *, user_mask_ptr)
4036{
4037	int ret;
4038	cpumask_var_t mask;
4039
4040	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4041		return -EINVAL;
4042	if (len & (sizeof(unsigned long)-1))
4043		return -EINVAL;
4044
4045	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4046		return -ENOMEM;
4047
4048	ret = sched_getaffinity(pid, mask);
4049	if (ret == 0) {
4050		size_t retlen = min_t(size_t, len, cpumask_size());
4051
4052		if (copy_to_user(user_mask_ptr, mask, retlen))
4053			ret = -EFAULT;
4054		else
4055			ret = retlen;
4056	}
4057	free_cpumask_var(mask);
4058
4059	return ret;
4060}
4061
4062/**
4063 * sys_sched_yield - yield the current processor to other threads.
4064 *
4065 * This function yields the current CPU to other tasks. If there are no
4066 * other threads running on this CPU then this function will return.
4067 *
4068 * Return: 0.
4069 */
4070SYSCALL_DEFINE0(sched_yield)
4071{
4072	struct rq *rq = this_rq_lock();
4073
4074	schedstat_inc(rq, yld_count);
4075	current->sched_class->yield_task(rq);
4076
4077	/*
4078	 * Since we are going to call schedule() anyway, there's
4079	 * no need to preempt or enable interrupts:
4080	 */
4081	__release(rq->lock);
4082	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4083	do_raw_spin_unlock(&rq->lock);
4084	sched_preempt_enable_no_resched();
4085
4086	schedule();
4087
4088	return 0;
4089}
4090
4091static void __cond_resched(void)
4092{
4093	__preempt_count_add(PREEMPT_ACTIVE);
4094	__schedule();
4095	__preempt_count_sub(PREEMPT_ACTIVE);
4096}
4097
4098int __sched _cond_resched(void)
4099{
4100	if (should_resched()) {
4101		__cond_resched();
4102		return 1;
4103	}
4104	return 0;
4105}
4106EXPORT_SYMBOL(_cond_resched);
4107
4108/*
4109 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4110 * call schedule, and on return reacquire the lock.
4111 *
4112 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4113 * operations here to prevent schedule() from being called twice (once via
4114 * spin_unlock(), once by hand).
4115 */
4116int __cond_resched_lock(spinlock_t *lock)
4117{
4118	int resched = should_resched();
4119	int ret = 0;
4120
4121	lockdep_assert_held(lock);
4122
4123	if (spin_needbreak(lock) || resched) {
4124		spin_unlock(lock);
4125		if (resched)
4126			__cond_resched();
4127		else
4128			cpu_relax();
4129		ret = 1;
4130		spin_lock(lock);
4131	}
4132	return ret;
4133}
4134EXPORT_SYMBOL(__cond_resched_lock);
4135
4136int __sched __cond_resched_softirq(void)
4137{
4138	BUG_ON(!in_softirq());
4139
4140	if (should_resched()) {
4141		local_bh_enable();
4142		__cond_resched();
4143		local_bh_disable();
4144		return 1;
4145	}
4146	return 0;
4147}
4148EXPORT_SYMBOL(__cond_resched_softirq);
4149
4150/**
4151 * yield - yield the current processor to other threads.
4152 *
4153 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4154 *
4155 * The scheduler is at all times free to pick the calling task as the most
4156 * eligible task to run, if removing the yield() call from your code breaks
4157 * it, its already broken.
4158 *
4159 * Typical broken usage is:
4160 *
4161 * while (!event)
4162 * 	yield();
4163 *
4164 * where one assumes that yield() will let 'the other' process run that will
4165 * make event true. If the current task is a SCHED_FIFO task that will never
4166 * happen. Never use yield() as a progress guarantee!!
4167 *
4168 * If you want to use yield() to wait for something, use wait_event().
4169 * If you want to use yield() to be 'nice' for others, use cond_resched().
4170 * If you still want to use yield(), do not!
4171 */
4172void __sched yield(void)
4173{
4174	set_current_state(TASK_RUNNING);
4175	sys_sched_yield();
4176}
4177EXPORT_SYMBOL(yield);
4178
4179/**
4180 * yield_to - yield the current processor to another thread in
4181 * your thread group, or accelerate that thread toward the
4182 * processor it's on.
4183 * @p: target task
4184 * @preempt: whether task preemption is allowed or not
4185 *
4186 * It's the caller's job to ensure that the target task struct
4187 * can't go away on us before we can do any checks.
4188 *
4189 * Return:
4190 *	true (>0) if we indeed boosted the target task.
4191 *	false (0) if we failed to boost the target.
4192 *	-ESRCH if there's no task to yield to.
4193 */
4194bool __sched yield_to(struct task_struct *p, bool preempt)
4195{
4196	struct task_struct *curr = current;
4197	struct rq *rq, *p_rq;
4198	unsigned long flags;
4199	int yielded = 0;
4200
4201	local_irq_save(flags);
4202	rq = this_rq();
4203
4204again:
4205	p_rq = task_rq(p);
4206	/*
4207	 * If we're the only runnable task on the rq and target rq also
4208	 * has only one task, there's absolutely no point in yielding.
4209	 */
4210	if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4211		yielded = -ESRCH;
4212		goto out_irq;
4213	}
4214
4215	double_rq_lock(rq, p_rq);
4216	if (task_rq(p) != p_rq) {
4217		double_rq_unlock(rq, p_rq);
4218		goto again;
4219	}
4220
4221	if (!curr->sched_class->yield_to_task)
4222		goto out_unlock;
4223
4224	if (curr->sched_class != p->sched_class)
4225		goto out_unlock;
4226
4227	if (task_running(p_rq, p) || p->state)
4228		goto out_unlock;
4229
4230	yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4231	if (yielded) {
4232		schedstat_inc(rq, yld_count);
4233		/*
4234		 * Make p's CPU reschedule; pick_next_entity takes care of
4235		 * fairness.
4236		 */
4237		if (preempt && rq != p_rq)
4238			resched_task(p_rq->curr);
4239	}
4240
4241out_unlock:
4242	double_rq_unlock(rq, p_rq);
4243out_irq:
4244	local_irq_restore(flags);
4245
4246	if (yielded > 0)
4247		schedule();
4248
4249	return yielded;
4250}
4251EXPORT_SYMBOL_GPL(yield_to);
4252
4253/*
4254 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4255 * that process accounting knows that this is a task in IO wait state.
4256 */
4257void __sched io_schedule(void)
4258{
4259	struct rq *rq = raw_rq();
4260
4261	delayacct_blkio_start();
4262	atomic_inc(&rq->nr_iowait);
4263	blk_flush_plug(current);
4264	current->in_iowait = 1;
4265	schedule();
4266	current->in_iowait = 0;
4267	atomic_dec(&rq->nr_iowait);
4268	delayacct_blkio_end();
4269}
4270EXPORT_SYMBOL(io_schedule);
4271
4272long __sched io_schedule_timeout(long timeout)
4273{
4274	struct rq *rq = raw_rq();
4275	long ret;
4276
4277	delayacct_blkio_start();
4278	atomic_inc(&rq->nr_iowait);
4279	blk_flush_plug(current);
4280	current->in_iowait = 1;
4281	ret = schedule_timeout(timeout);
4282	current->in_iowait = 0;
4283	atomic_dec(&rq->nr_iowait);
4284	delayacct_blkio_end();
4285	return ret;
4286}
4287
4288/**
4289 * sys_sched_get_priority_max - return maximum RT priority.
4290 * @policy: scheduling class.
4291 *
4292 * Return: On success, this syscall returns the maximum
4293 * rt_priority that can be used by a given scheduling class.
4294 * On failure, a negative error code is returned.
4295 */
4296SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4297{
4298	int ret = -EINVAL;
4299
4300	switch (policy) {
4301	case SCHED_FIFO:
4302	case SCHED_RR:
4303		ret = MAX_USER_RT_PRIO-1;
4304		break;
4305	case SCHED_DEADLINE:
4306	case SCHED_NORMAL:
4307	case SCHED_BATCH:
4308	case SCHED_IDLE:
4309		ret = 0;
4310		break;
4311	}
4312	return ret;
4313}
4314
4315/**
4316 * sys_sched_get_priority_min - return minimum RT priority.
4317 * @policy: scheduling class.
4318 *
4319 * Return: On success, this syscall returns the minimum
4320 * rt_priority that can be used by a given scheduling class.
4321 * On failure, a negative error code is returned.
4322 */
4323SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4324{
4325	int ret = -EINVAL;
4326
4327	switch (policy) {
4328	case SCHED_FIFO:
4329	case SCHED_RR:
4330		ret = 1;
4331		break;
4332	case SCHED_DEADLINE:
4333	case SCHED_NORMAL:
4334	case SCHED_BATCH:
4335	case SCHED_IDLE:
4336		ret = 0;
4337	}
4338	return ret;
4339}
4340
4341/**
4342 * sys_sched_rr_get_interval - return the default timeslice of a process.
4343 * @pid: pid of the process.
4344 * @interval: userspace pointer to the timeslice value.
4345 *
4346 * this syscall writes the default timeslice value of a given process
4347 * into the user-space timespec buffer. A value of '0' means infinity.
4348 *
4349 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4350 * an error code.
4351 */
4352SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4353		struct timespec __user *, interval)
4354{
4355	struct task_struct *p;
4356	unsigned int time_slice;
4357	unsigned long flags;
4358	struct rq *rq;
4359	int retval;
4360	struct timespec t;
4361
4362	if (pid < 0)
4363		return -EINVAL;
4364
4365	retval = -ESRCH;
4366	rcu_read_lock();
4367	p = find_process_by_pid(pid);
4368	if (!p)
4369		goto out_unlock;
4370
4371	retval = security_task_getscheduler(p);
4372	if (retval)
4373		goto out_unlock;
4374
4375	rq = task_rq_lock(p, &flags);
4376	time_slice = 0;
4377	if (p->sched_class->get_rr_interval)
4378		time_slice = p->sched_class->get_rr_interval(rq, p);
4379	task_rq_unlock(rq, p, &flags);
4380
4381	rcu_read_unlock();
4382	jiffies_to_timespec(time_slice, &t);
4383	retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4384	return retval;
4385
4386out_unlock:
4387	rcu_read_unlock();
4388	return retval;
4389}
4390
4391static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4392
4393void sched_show_task(struct task_struct *p)
4394{
4395	unsigned long free = 0;
4396	int ppid;
4397	unsigned state;
4398
4399	state = p->state ? __ffs(p->state) + 1 : 0;
4400	printk(KERN_INFO "%-15.15s %c", p->comm,
4401		state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4402#if BITS_PER_LONG == 32
4403	if (state == TASK_RUNNING)
4404		printk(KERN_CONT " running  ");
4405	else
4406		printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4407#else
4408	if (state == TASK_RUNNING)
4409		printk(KERN_CONT "  running task    ");
4410	else
4411		printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4412#endif
4413#ifdef CONFIG_DEBUG_STACK_USAGE
4414	free = stack_not_used(p);
4415#endif
4416	rcu_read_lock();
4417	ppid = task_pid_nr(rcu_dereference(p->real_parent));
4418	rcu_read_unlock();
4419	printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4420		task_pid_nr(p), ppid,
4421		(unsigned long)task_thread_info(p)->flags);
4422
4423	print_worker_info(KERN_INFO, p);
4424	show_stack(p, NULL);
4425}
4426
4427void show_state_filter(unsigned long state_filter)
4428{
4429	struct task_struct *g, *p;
4430
4431#if BITS_PER_LONG == 32
4432	printk(KERN_INFO
4433		"  task                PC stack   pid father\n");
4434#else
4435	printk(KERN_INFO
4436		"  task                        PC stack   pid father\n");
4437#endif
4438	rcu_read_lock();
4439	do_each_thread(g, p) {
4440		/*
4441		 * reset the NMI-timeout, listing all files on a slow
4442		 * console might take a lot of time:
4443		 */
4444		touch_nmi_watchdog();
4445		if (!state_filter || (p->state & state_filter))
4446			sched_show_task(p);
4447	} while_each_thread(g, p);
4448
4449	touch_all_softlockup_watchdogs();
4450
4451#ifdef CONFIG_SCHED_DEBUG
4452	sysrq_sched_debug_show();
4453#endif
4454	rcu_read_unlock();
4455	/*
4456	 * Only show locks if all tasks are dumped:
4457	 */
4458	if (!state_filter)
4459		debug_show_all_locks();
4460}
4461
4462void init_idle_bootup_task(struct task_struct *idle)
4463{
4464	idle->sched_class = &idle_sched_class;
4465}
4466
4467/**
4468 * init_idle - set up an idle thread for a given CPU
4469 * @idle: task in question
4470 * @cpu: cpu the idle task belongs to
4471 *
4472 * NOTE: this function does not set the idle thread's NEED_RESCHED
4473 * flag, to make booting more robust.
4474 */
4475void init_idle(struct task_struct *idle, int cpu)
4476{
4477	struct rq *rq = cpu_rq(cpu);
4478	unsigned long flags;
4479
4480	raw_spin_lock_irqsave(&rq->lock, flags);
4481
4482	__sched_fork(0, idle);
4483	idle->state = TASK_RUNNING;
4484	idle->se.exec_start = sched_clock();
4485
4486	do_set_cpus_allowed(idle, cpumask_of(cpu));
4487	/*
4488	 * We're having a chicken and egg problem, even though we are
4489	 * holding rq->lock, the cpu isn't yet set to this cpu so the
4490	 * lockdep check in task_group() will fail.
4491	 *
4492	 * Similar case to sched_fork(). / Alternatively we could
4493	 * use task_rq_lock() here and obtain the other rq->lock.
4494	 *
4495	 * Silence PROVE_RCU
4496	 */
4497	rcu_read_lock();
4498	__set_task_cpu(idle, cpu);
4499	rcu_read_unlock();
4500
4501	rq->curr = rq->idle = idle;
4502	idle->on_rq = 1;
4503#if defined(CONFIG_SMP)
4504	idle->on_cpu = 1;
4505#endif
4506	raw_spin_unlock_irqrestore(&rq->lock, flags);
4507
4508	/* Set the preempt count _outside_ the spinlocks! */
4509	init_idle_preempt_count(idle, cpu);
4510
4511	/*
4512	 * The idle tasks have their own, simple scheduling class:
4513	 */
4514	idle->sched_class = &idle_sched_class;
4515	ftrace_graph_init_idle_task(idle, cpu);
4516	vtime_init_idle(idle, cpu);
4517#if defined(CONFIG_SMP)
4518	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4519#endif
4520}
4521
4522#ifdef CONFIG_SMP
4523void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4524{
4525	if (p->sched_class && p->sched_class->set_cpus_allowed)
4526		p->sched_class->set_cpus_allowed(p, new_mask);
4527
4528	cpumask_copy(&p->cpus_allowed, new_mask);
4529	p->nr_cpus_allowed = cpumask_weight(new_mask);
4530}
4531
4532/*
4533 * This is how migration works:
4534 *
4535 * 1) we invoke migration_cpu_stop() on the target CPU using
4536 *    stop_one_cpu().
4537 * 2) stopper starts to run (implicitly forcing the migrated thread
4538 *    off the CPU)
4539 * 3) it checks whether the migrated task is still in the wrong runqueue.
4540 * 4) if it's in the wrong runqueue then the migration thread removes
4541 *    it and puts it into the right queue.
4542 * 5) stopper completes and stop_one_cpu() returns and the migration
4543 *    is done.
4544 */
4545
4546/*
4547 * Change a given task's CPU affinity. Migrate the thread to a
4548 * proper CPU and schedule it away if the CPU it's executing on
4549 * is removed from the allowed bitmask.
4550 *
4551 * NOTE: the caller must have a valid reference to the task, the
4552 * task must not exit() & deallocate itself prematurely. The
4553 * call is not atomic; no spinlocks may be held.
4554 */
4555int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4556{
4557	unsigned long flags;
4558	struct rq *rq;
4559	unsigned int dest_cpu;
4560	int ret = 0;
4561
4562	rq = task_rq_lock(p, &flags);
4563
4564	if (cpumask_equal(&p->cpus_allowed, new_mask))
4565		goto out;
4566
4567	if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4568		ret = -EINVAL;
4569		goto out;
4570	}
4571
4572	do_set_cpus_allowed(p, new_mask);
4573
4574	/* Can the task run on the task's current CPU? If so, we're done */
4575	if (cpumask_test_cpu(task_cpu(p), new_mask))
4576		goto out;
4577
4578	dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4579	if (p->on_rq) {
4580		struct migration_arg arg = { p, dest_cpu };
4581		/* Need help from migration thread: drop lock and wait. */
4582		task_rq_unlock(rq, p, &flags);
4583		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4584		tlb_migrate_finish(p->mm);
4585		return 0;
4586	}
4587out:
4588	task_rq_unlock(rq, p, &flags);
4589
4590	return ret;
4591}
4592EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4593
4594/*
4595 * Move (not current) task off this cpu, onto dest cpu. We're doing
4596 * this because either it can't run here any more (set_cpus_allowed()
4597 * away from this CPU, or CPU going down), or because we're
4598 * attempting to rebalance this task on exec (sched_exec).
4599 *
4600 * So we race with normal scheduler movements, but that's OK, as long
4601 * as the task is no longer on this CPU.
4602 *
4603 * Returns non-zero if task was successfully migrated.
4604 */
4605static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4606{
4607	struct rq *rq_dest, *rq_src;
4608	int ret = 0;
4609
4610	if (unlikely(!cpu_active(dest_cpu)))
4611		return ret;
4612
4613	rq_src = cpu_rq(src_cpu);
4614	rq_dest = cpu_rq(dest_cpu);
4615
4616	raw_spin_lock(&p->pi_lock);
4617	double_rq_lock(rq_src, rq_dest);
4618	/* Already moved. */
4619	if (task_cpu(p) != src_cpu)
4620		goto done;
4621	/* Affinity changed (again). */
4622	if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4623		goto fail;
4624
4625	/*
4626	 * If we're not on a rq, the next wake-up will ensure we're
4627	 * placed properly.
4628	 */
4629	if (p->on_rq) {
4630		dequeue_task(rq_src, p, 0);
4631		set_task_cpu(p, dest_cpu);
4632		enqueue_task(rq_dest, p, 0);
4633		check_preempt_curr(rq_dest, p, 0);
4634	}
4635done:
4636	ret = 1;
4637fail:
4638	double_rq_unlock(rq_src, rq_dest);
4639	raw_spin_unlock(&p->pi_lock);
4640	return ret;
4641}
4642
4643#ifdef CONFIG_NUMA_BALANCING
4644/* Migrate current task p to target_cpu */
4645int migrate_task_to(struct task_struct *p, int target_cpu)
4646{
4647	struct migration_arg arg = { p, target_cpu };
4648	int curr_cpu = task_cpu(p);
4649
4650	if (curr_cpu == target_cpu)
4651		return 0;
4652
4653	if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4654		return -EINVAL;
4655
4656	/* TODO: This is not properly updating schedstats */
4657
4658	trace_sched_move_numa(p, curr_cpu, target_cpu);
4659	return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4660}
4661
4662/*
4663 * Requeue a task on a given node and accurately track the number of NUMA
4664 * tasks on the runqueues
4665 */
4666void sched_setnuma(struct task_struct *p, int nid)
4667{
4668	struct rq *rq;
4669	unsigned long flags;
4670	bool on_rq, running;
4671
4672	rq = task_rq_lock(p, &flags);
4673	on_rq = p->on_rq;
4674	running = task_current(rq, p);
4675
4676	if (on_rq)
4677		dequeue_task(rq, p, 0);
4678	if (running)
4679		p->sched_class->put_prev_task(rq, p);
4680
4681	p->numa_preferred_nid = nid;
4682
4683	if (running)
4684		p->sched_class->set_curr_task(rq);
4685	if (on_rq)
4686		enqueue_task(rq, p, 0);
4687	task_rq_unlock(rq, p, &flags);
4688}
4689#endif
4690
4691/*
4692 * migration_cpu_stop - this will be executed by a highprio stopper thread
4693 * and performs thread migration by bumping thread off CPU then
4694 * 'pushing' onto another runqueue.
4695 */
4696static int migration_cpu_stop(void *data)
4697{
4698	struct migration_arg *arg = data;
4699
4700	/*
4701	 * The original target cpu might have gone down and we might
4702	 * be on another cpu but it doesn't matter.
4703	 */
4704	local_irq_disable();
4705	__migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4706	local_irq_enable();
4707	return 0;
4708}
4709
4710#ifdef CONFIG_HOTPLUG_CPU
4711
4712/*
4713 * Ensures that the idle task is using init_mm right before its cpu goes
4714 * offline.
4715 */
4716void idle_task_exit(void)
4717{
4718	struct mm_struct *mm = current->active_mm;
4719
4720	BUG_ON(cpu_online(smp_processor_id()));
4721
4722	if (mm != &init_mm) {
4723		switch_mm(mm, &init_mm, current);
4724		finish_arch_post_lock_switch();
4725	}
4726	mmdrop(mm);
4727}
4728
4729/*
4730 * Since this CPU is going 'away' for a while, fold any nr_active delta
4731 * we might have. Assumes we're called after migrate_tasks() so that the
4732 * nr_active count is stable.
4733 *
4734 * Also see the comment "Global load-average calculations".
4735 */
4736static void calc_load_migrate(struct rq *rq)
4737{
4738	long delta = calc_load_fold_active(rq);
4739	if (delta)
4740		atomic_long_add(delta, &calc_load_tasks);
4741}
4742
4743static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4744{
4745}
4746
4747static const struct sched_class fake_sched_class = {
4748	.put_prev_task = put_prev_task_fake,
4749};
4750
4751static struct task_struct fake_task = {
4752	/*
4753	 * Avoid pull_{rt,dl}_task()
4754	 */
4755	.prio = MAX_PRIO + 1,
4756	.sched_class = &fake_sched_class,
4757};
4758
4759/*
4760 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4761 * try_to_wake_up()->select_task_rq().
4762 *
4763 * Called with rq->lock held even though we'er in stop_machine() and
4764 * there's no concurrency possible, we hold the required locks anyway
4765 * because of lock validation efforts.
4766 */
4767static void migrate_tasks(unsigned int dead_cpu)
4768{
4769	struct rq *rq = cpu_rq(dead_cpu);
4770	struct task_struct *next, *stop = rq->stop;
4771	int dest_cpu;
4772
4773	/*
4774	 * Fudge the rq selection such that the below task selection loop
4775	 * doesn't get stuck on the currently eligible stop task.
4776	 *
4777	 * We're currently inside stop_machine() and the rq is either stuck
4778	 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4779	 * either way we should never end up calling schedule() until we're
4780	 * done here.
4781	 */
4782	rq->stop = NULL;
4783
4784	/*
4785	 * put_prev_task() and pick_next_task() sched
4786	 * class method both need to have an up-to-date
4787	 * value of rq->clock[_task]
4788	 */
4789	update_rq_clock(rq);
4790
4791	for ( ; ; ) {
4792		/*
4793		 * There's this thread running, bail when that's the only
4794		 * remaining thread.
4795		 */
4796		if (rq->nr_running == 1)
4797			break;
4798
4799		next = pick_next_task(rq, &fake_task);
4800		BUG_ON(!next);
4801		next->sched_class->put_prev_task(rq, next);
4802
4803		/* Find suitable destination for @next, with force if needed. */
4804		dest_cpu = select_fallback_rq(dead_cpu, next);
4805		raw_spin_unlock(&rq->lock);
4806
4807		__migrate_task(next, dead_cpu, dest_cpu);
4808
4809		raw_spin_lock(&rq->lock);
4810	}
4811
4812	rq->stop = stop;
4813}
4814
4815#endif /* CONFIG_HOTPLUG_CPU */
4816
4817#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4818
4819static struct ctl_table sd_ctl_dir[] = {
4820	{
4821		.procname	= "sched_domain",
4822		.mode		= 0555,
4823	},
4824	{}
4825};
4826
4827static struct ctl_table sd_ctl_root[] = {
4828	{
4829		.procname	= "kernel",
4830		.mode		= 0555,
4831		.child		= sd_ctl_dir,
4832	},
4833	{}
4834};
4835
4836static struct ctl_table *sd_alloc_ctl_entry(int n)
4837{
4838	struct ctl_table *entry =
4839		kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4840
4841	return entry;
4842}
4843
4844static void sd_free_ctl_entry(struct ctl_table **tablep)
4845{
4846	struct ctl_table *entry;
4847
4848	/*
4849	 * In the intermediate directories, both the child directory and
4850	 * procname are dynamically allocated and could fail but the mode
4851	 * will always be set. In the lowest directory the names are
4852	 * static strings and all have proc handlers.
4853	 */
4854	for (entry = *tablep; entry->mode; entry++) {
4855		if (entry->child)
4856			sd_free_ctl_entry(&entry->child);
4857		if (entry->proc_handler == NULL)
4858			kfree(entry->procname);
4859	}
4860
4861	kfree(*tablep);
4862	*tablep = NULL;
4863}
4864
4865static int min_load_idx = 0;
4866static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4867
4868static void
4869set_table_entry(struct ctl_table *entry,
4870		const char *procname, void *data, int maxlen,
4871		umode_t mode, proc_handler *proc_handler,
4872		bool load_idx)
4873{
4874	entry->procname = procname;
4875	entry->data = data;
4876	entry->maxlen = maxlen;
4877	entry->mode = mode;
4878	entry->proc_handler = proc_handler;
4879
4880	if (load_idx) {
4881		entry->extra1 = &min_load_idx;
4882		entry->extra2 = &max_load_idx;
4883	}
4884}
4885
4886static struct ctl_table *
4887sd_alloc_ctl_domain_table(struct sched_domain *sd)
4888{
4889	struct ctl_table *table = sd_alloc_ctl_entry(14);
4890
4891	if (table == NULL)
4892		return NULL;
4893
4894	set_table_entry(&table[0], "min_interval", &sd->min_interval,
4895		sizeof(long), 0644, proc_doulongvec_minmax, false);
4896	set_table_entry(&table[1], "max_interval", &sd->max_interval,
4897		sizeof(long), 0644, proc_doulongvec_minmax, false);
4898	set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4899		sizeof(int), 0644, proc_dointvec_minmax, true);
4900	set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4901		sizeof(int), 0644, proc_dointvec_minmax, true);
4902	set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4903		sizeof(int), 0644, proc_dointvec_minmax, true);
4904	set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4905		sizeof(int), 0644, proc_dointvec_minmax, true);
4906	set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4907		sizeof(int), 0644, proc_dointvec_minmax, true);
4908	set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4909		sizeof(int), 0644, proc_dointvec_minmax, false);
4910	set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4911		sizeof(int), 0644, proc_dointvec_minmax, false);
4912	set_table_entry(&table[9], "cache_nice_tries",
4913		&sd->cache_nice_tries,
4914		sizeof(int), 0644, proc_dointvec_minmax, false);
4915	set_table_entry(&table[10], "flags", &sd->flags,
4916		sizeof(int), 0644, proc_dointvec_minmax, false);
4917	set_table_entry(&table[11], "max_newidle_lb_cost",
4918		&sd->max_newidle_lb_cost,
4919		sizeof(long), 0644, proc_doulongvec_minmax, false);
4920	set_table_entry(&table[12], "name", sd->name,
4921		CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4922	/* &table[13] is terminator */
4923
4924	return table;
4925}
4926
4927static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4928{
4929	struct ctl_table *entry, *table;
4930	struct sched_domain *sd;
4931	int domain_num = 0, i;
4932	char buf[32];
4933
4934	for_each_domain(cpu, sd)
4935		domain_num++;
4936	entry = table = sd_alloc_ctl_entry(domain_num + 1);
4937	if (table == NULL)
4938		return NULL;
4939
4940	i = 0;
4941	for_each_domain(cpu, sd) {
4942		snprintf(buf, 32, "domain%d", i);
4943		entry->procname = kstrdup(buf, GFP_KERNEL);
4944		entry->mode = 0555;
4945		entry->child = sd_alloc_ctl_domain_table(sd);
4946		entry++;
4947		i++;
4948	}
4949	return table;
4950}
4951
4952static struct ctl_table_header *sd_sysctl_header;
4953static void register_sched_domain_sysctl(void)
4954{
4955	int i, cpu_num = num_possible_cpus();
4956	struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4957	char buf[32];
4958
4959	WARN_ON(sd_ctl_dir[0].child);
4960	sd_ctl_dir[0].child = entry;
4961
4962	if (entry == NULL)
4963		return;
4964
4965	for_each_possible_cpu(i) {
4966		snprintf(buf, 32, "cpu%d", i);
4967		entry->procname = kstrdup(buf, GFP_KERNEL);
4968		entry->mode = 0555;
4969		entry->child = sd_alloc_ctl_cpu_table(i);
4970		entry++;
4971	}
4972
4973	WARN_ON(sd_sysctl_header);
4974	sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4975}
4976
4977/* may be called multiple times per register */
4978static void unregister_sched_domain_sysctl(void)
4979{
4980	if (sd_sysctl_header)
4981		unregister_sysctl_table(sd_sysctl_header);
4982	sd_sysctl_header = NULL;
4983	if (sd_ctl_dir[0].child)
4984		sd_free_ctl_entry(&sd_ctl_dir[0].child);
4985}
4986#else
4987static void register_sched_domain_sysctl(void)
4988{
4989}
4990static void unregister_sched_domain_sysctl(void)
4991{
4992}
4993#endif
4994
4995static void set_rq_online(struct rq *rq)
4996{
4997	if (!rq->online) {
4998		const struct sched_class *class;
4999
5000		cpumask_set_cpu(rq->cpu, rq->rd->online);
5001		rq->online = 1;
5002
5003		for_each_class(class) {
5004			if (class->rq_online)
5005				class->rq_online(rq);
5006		}
5007	}
5008}
5009
5010static void set_rq_offline(struct rq *rq)
5011{
5012	if (rq->online) {
5013		const struct sched_class *class;
5014
5015		for_each_class(class) {
5016			if (class->rq_offline)
5017				class->rq_offline(rq);
5018		}
5019
5020		cpumask_clear_cpu(rq->cpu, rq->rd->online);
5021		rq->online = 0;
5022	}
5023}
5024
5025/*
5026 * migration_call - callback that gets triggered when a CPU is added.
5027 * Here we can start up the necessary migration thread for the new CPU.
5028 */
5029static int
5030migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5031{
5032	int cpu = (long)hcpu;
5033	unsigned long flags;
5034	struct rq *rq = cpu_rq(cpu);
5035
5036	switch (action & ~CPU_TASKS_FROZEN) {
5037
5038	case CPU_UP_PREPARE:
5039		rq->calc_load_update = calc_load_update;
5040		break;
5041
5042	case CPU_ONLINE:
5043		/* Update our root-domain */
5044		raw_spin_lock_irqsave(&rq->lock, flags);
5045		if (rq->rd) {
5046			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5047
5048			set_rq_online(rq);
5049		}
5050		raw_spin_unlock_irqrestore(&rq->lock, flags);
5051		break;
5052
5053#ifdef CONFIG_HOTPLUG_CPU
5054	case CPU_DYING:
5055		sched_ttwu_pending();
5056		/* Update our root-domain */
5057		raw_spin_lock_irqsave(&rq->lock, flags);
5058		if (rq->rd) {
5059			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5060			set_rq_offline(rq);
5061		}
5062		migrate_tasks(cpu);
5063		BUG_ON(rq->nr_running != 1); /* the migration thread */
5064		raw_spin_unlock_irqrestore(&rq->lock, flags);
5065		break;
5066
5067	case CPU_DEAD:
5068		calc_load_migrate(rq);
5069		break;
5070#endif
5071	}
5072
5073	update_max_interval();
5074
5075	return NOTIFY_OK;
5076}
5077
5078/*
5079 * Register at high priority so that task migration (migrate_all_tasks)
5080 * happens before everything else.  This has to be lower priority than
5081 * the notifier in the perf_event subsystem, though.
5082 */
5083static struct notifier_block migration_notifier = {
5084	.notifier_call = migration_call,
5085	.priority = CPU_PRI_MIGRATION,
5086};
5087
5088static int sched_cpu_active(struct notifier_block *nfb,
5089				      unsigned long action, void *hcpu)
5090{
5091	switch (action & ~CPU_TASKS_FROZEN) {
5092	case CPU_STARTING:
5093	case CPU_DOWN_FAILED:
5094		set_cpu_active((long)hcpu, true);
5095		return NOTIFY_OK;
5096	default:
5097		return NOTIFY_DONE;
5098	}
5099}
5100
5101static int sched_cpu_inactive(struct notifier_block *nfb,
5102					unsigned long action, void *hcpu)
5103{
5104	unsigned long flags;
5105	long cpu = (long)hcpu;
5106
5107	switch (action & ~CPU_TASKS_FROZEN) {
5108	case CPU_DOWN_PREPARE:
5109		set_cpu_active(cpu, false);
5110
5111		/* explicitly allow suspend */
5112		if (!(action & CPU_TASKS_FROZEN)) {
5113			struct dl_bw *dl_b = dl_bw_of(cpu);
5114			bool overflow;
5115			int cpus;
5116
5117			raw_spin_lock_irqsave(&dl_b->lock, flags);
5118			cpus = dl_bw_cpus(cpu);
5119			overflow = __dl_overflow(dl_b, cpus, 0, 0);
5120			raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5121
5122			if (overflow)
5123				return notifier_from_errno(-EBUSY);
5124		}
5125		return NOTIFY_OK;
5126	}
5127
5128	return NOTIFY_DONE;
5129}
5130
5131static int __init migration_init(void)
5132{
5133	void *cpu = (void *)(long)smp_processor_id();
5134	int err;
5135
5136	/* Initialize migration for the boot CPU */
5137	err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5138	BUG_ON(err == NOTIFY_BAD);
5139	migration_call(&migration_notifier, CPU_ONLINE, cpu);
5140	register_cpu_notifier(&migration_notifier);
5141
5142	/* Register cpu active notifiers */
5143	cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5144	cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5145
5146	return 0;
5147}
5148early_initcall(migration_init);
5149#endif
5150
5151#ifdef CONFIG_SMP
5152
5153static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5154
5155#ifdef CONFIG_SCHED_DEBUG
5156
5157static __read_mostly int sched_debug_enabled;
5158
5159static int __init sched_debug_setup(char *str)
5160{
5161	sched_debug_enabled = 1;
5162
5163	return 0;
5164}
5165early_param("sched_debug", sched_debug_setup);
5166
5167static inline bool sched_debug(void)
5168{
5169	return sched_debug_enabled;
5170}
5171
5172static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5173				  struct cpumask *groupmask)
5174{
5175	struct sched_group *group = sd->groups;
5176	char str[256];
5177
5178	cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5179	cpumask_clear(groupmask);
5180
5181	printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5182
5183	if (!(sd->flags & SD_LOAD_BALANCE)) {
5184		printk("does not load-balance\n");
5185		if (sd->parent)
5186			printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5187					" has parent");
5188		return -1;
5189	}
5190
5191	printk(KERN_CONT "span %s level %s\n", str, sd->name);
5192
5193	if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5194		printk(KERN_ERR "ERROR: domain->span does not contain "
5195				"CPU%d\n", cpu);
5196	}
5197	if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5198		printk(KERN_ERR "ERROR: domain->groups does not contain"
5199				" CPU%d\n", cpu);
5200	}
5201
5202	printk(KERN_DEBUG "%*s groups:", level + 1, "");
5203	do {
5204		if (!group) {
5205			printk("\n");
5206			printk(KERN_ERR "ERROR: group is NULL\n");
5207			break;
5208		}
5209
5210		/*
5211		 * Even though we initialize ->power to something semi-sane,
5212		 * we leave power_orig unset. This allows us to detect if
5213		 * domain iteration is still funny without causing /0 traps.
5214		 */
5215		if (!group->sgp->power_orig) {
5216			printk(KERN_CONT "\n");
5217			printk(KERN_ERR "ERROR: domain->cpu_power not "
5218					"set\n");
5219			break;
5220		}
5221
5222		if (!cpumask_weight(sched_group_cpus(group))) {
5223			printk(KERN_CONT "\n");
5224			printk(KERN_ERR "ERROR: empty group\n");
5225			break;
5226		}
5227
5228		if (!(sd->flags & SD_OVERLAP) &&
5229		    cpumask_intersects(groupmask, sched_group_cpus(group))) {
5230			printk(KERN_CONT "\n");
5231			printk(KERN_ERR "ERROR: repeated CPUs\n");
5232			break;
5233		}
5234
5235		cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5236
5237		cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5238
5239		printk(KERN_CONT " %s", str);
5240		if (group->sgp->power != SCHED_POWER_SCALE) {
5241			printk(KERN_CONT " (cpu_power = %d)",
5242				group->sgp->power);
5243		}
5244
5245		group = group->next;
5246	} while (group != sd->groups);
5247	printk(KERN_CONT "\n");
5248
5249	if (!cpumask_equal(sched_domain_span(sd), groupmask))
5250		printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5251
5252	if (sd->parent &&
5253	    !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5254		printk(KERN_ERR "ERROR: parent span is not a superset "
5255			"of domain->span\n");
5256	return 0;
5257}
5258
5259static void sched_domain_debug(struct sched_domain *sd, int cpu)
5260{
5261	int level = 0;
5262
5263	if (!sched_debug_enabled)
5264		return;
5265
5266	if (!sd) {
5267		printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5268		return;
5269	}
5270
5271	printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5272
5273	for (;;) {
5274		if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5275			break;
5276		level++;
5277		sd = sd->parent;
5278		if (!sd)
5279			break;
5280	}
5281}
5282#else /* !CONFIG_SCHED_DEBUG */
5283# define sched_domain_debug(sd, cpu) do { } while (0)
5284static inline bool sched_debug(void)
5285{
5286	return false;
5287}
5288#endif /* CONFIG_SCHED_DEBUG */
5289
5290static int sd_degenerate(struct sched_domain *sd)
5291{
5292	if (cpumask_weight(sched_domain_span(sd)) == 1)
5293		return 1;
5294
5295	/* Following flags need at least 2 groups */
5296	if (sd->flags & (SD_LOAD_BALANCE |
5297			 SD_BALANCE_NEWIDLE |
5298			 SD_BALANCE_FORK |
5299			 SD_BALANCE_EXEC |
5300			 SD_SHARE_CPUPOWER |
5301			 SD_SHARE_PKG_RESOURCES)) {
5302		if (sd->groups != sd->groups->next)
5303			return 0;
5304	}
5305
5306	/* Following flags don't use groups */
5307	if (sd->flags & (SD_WAKE_AFFINE))
5308		return 0;
5309
5310	return 1;
5311}
5312
5313static int
5314sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5315{
5316	unsigned long cflags = sd->flags, pflags = parent->flags;
5317
5318	if (sd_degenerate(parent))
5319		return 1;
5320
5321	if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5322		return 0;
5323
5324	/* Flags needing groups don't count if only 1 group in parent */
5325	if (parent->groups == parent->groups->next) {
5326		pflags &= ~(SD_LOAD_BALANCE |
5327				SD_BALANCE_NEWIDLE |
5328				SD_BALANCE_FORK |
5329				SD_BALANCE_EXEC |
5330				SD_SHARE_CPUPOWER |
5331				SD_SHARE_PKG_RESOURCES |
5332				SD_PREFER_SIBLING);
5333		if (nr_node_ids == 1)
5334			pflags &= ~SD_SERIALIZE;
5335	}
5336	if (~cflags & pflags)
5337		return 0;
5338
5339	return 1;
5340}
5341
5342static void free_rootdomain(struct rcu_head *rcu)
5343{
5344	struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5345
5346	cpupri_cleanup(&rd->cpupri);
5347	cpudl_cleanup(&rd->cpudl);
5348	free_cpumask_var(rd->dlo_mask);
5349	free_cpumask_var(rd->rto_mask);
5350	free_cpumask_var(rd->online);
5351	free_cpumask_var(rd->span);
5352	kfree(rd);
5353}
5354
5355static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5356{
5357	struct root_domain *old_rd = NULL;
5358	unsigned long flags;
5359
5360	raw_spin_lock_irqsave(&rq->lock, flags);
5361
5362	if (rq->rd) {
5363		old_rd = rq->rd;
5364
5365		if (cpumask_test_cpu(rq->cpu, old_rd->online))
5366			set_rq_offline(rq);
5367
5368		cpumask_clear_cpu(rq->cpu, old_rd->span);
5369
5370		/*
5371		 * If we dont want to free the old_rd yet then
5372		 * set old_rd to NULL to skip the freeing later
5373		 * in this function:
5374		 */
5375		if (!atomic_dec_and_test(&old_rd->refcount))
5376			old_rd = NULL;
5377	}
5378
5379	atomic_inc(&rd->refcount);
5380	rq->rd = rd;
5381
5382	cpumask_set_cpu(rq->cpu, rd->span);
5383	if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5384		set_rq_online(rq);
5385
5386	raw_spin_unlock_irqrestore(&rq->lock, flags);
5387
5388	if (old_rd)
5389		call_rcu_sched(&old_rd->rcu, free_rootdomain);
5390}
5391
5392static int init_rootdomain(struct root_domain *rd)
5393{
5394	memset(rd, 0, sizeof(*rd));
5395
5396	if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5397		goto out;
5398	if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5399		goto free_span;
5400	if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5401		goto free_online;
5402	if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5403		goto free_dlo_mask;
5404
5405	init_dl_bw(&rd->dl_bw);
5406	if (cpudl_init(&rd->cpudl) != 0)
5407		goto free_dlo_mask;
5408
5409	if (cpupri_init(&rd->cpupri) != 0)
5410		goto free_rto_mask;
5411	return 0;
5412
5413free_rto_mask:
5414	free_cpumask_var(rd->rto_mask);
5415free_dlo_mask:
5416	free_cpumask_var(rd->dlo_mask);
5417free_online:
5418	free_cpumask_var(rd->online);
5419free_span:
5420	free_cpumask_var(rd->span);
5421out:
5422	return -ENOMEM;
5423}
5424
5425/*
5426 * By default the system creates a single root-domain with all cpus as
5427 * members (mimicking the global state we have today).
5428 */
5429struct root_domain def_root_domain;
5430
5431static void init_defrootdomain(void)
5432{
5433	init_rootdomain(&def_root_domain);
5434
5435	atomic_set(&def_root_domain.refcount, 1);
5436}
5437
5438static struct root_domain *alloc_rootdomain(void)
5439{
5440	struct root_domain *rd;
5441
5442	rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5443	if (!rd)
5444		return NULL;
5445
5446	if (init_rootdomain(rd) != 0) {
5447		kfree(rd);
5448		return NULL;
5449	}
5450
5451	return rd;
5452}
5453
5454static void free_sched_groups(struct sched_group *sg, int free_sgp)
5455{
5456	struct sched_group *tmp, *first;
5457
5458	if (!sg)
5459		return;
5460
5461	first = sg;
5462	do {
5463		tmp = sg->next;
5464
5465		if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5466			kfree(sg->sgp);
5467
5468		kfree(sg);
5469		sg = tmp;
5470	} while (sg != first);
5471}
5472
5473static void free_sched_domain(struct rcu_head *rcu)
5474{
5475	struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5476
5477	/*
5478	 * If its an overlapping domain it has private groups, iterate and
5479	 * nuke them all.
5480	 */
5481	if (sd->flags & SD_OVERLAP) {
5482		free_sched_groups(sd->groups, 1);
5483	} else if (atomic_dec_and_test(&sd->groups->ref)) {
5484		kfree(sd->groups->sgp);
5485		kfree(sd->groups);
5486	}
5487	kfree(sd);
5488}
5489
5490static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5491{
5492	call_rcu(&sd->rcu, free_sched_domain);
5493}
5494
5495static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5496{
5497	for (; sd; sd = sd->parent)
5498		destroy_sched_domain(sd, cpu);
5499}
5500
5501/*
5502 * Keep a special pointer to the highest sched_domain that has
5503 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5504 * allows us to avoid some pointer chasing select_idle_sibling().
5505 *
5506 * Also keep a unique ID per domain (we use the first cpu number in
5507 * the cpumask of the domain), this allows us to quickly tell if
5508 * two cpus are in the same cache domain, see cpus_share_cache().
5509 */
5510DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5511DEFINE_PER_CPU(int, sd_llc_size);
5512DEFINE_PER_CPU(int, sd_llc_id);
5513DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5514DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5515DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5516
5517static void update_top_cache_domain(int cpu)
5518{
5519	struct sched_domain *sd;
5520	struct sched_domain *busy_sd = NULL;
5521	int id = cpu;
5522	int size = 1;
5523
5524	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5525	if (sd) {
5526		id = cpumask_first(sched_domain_span(sd));
5527		size = cpumask_weight(sched_domain_span(sd));
5528		busy_sd = sd->parent; /* sd_busy */
5529	}
5530	rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5531
5532	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5533	per_cpu(sd_llc_size, cpu) = size;
5534	per_cpu(sd_llc_id, cpu) = id;
5535
5536	sd = lowest_flag_domain(cpu, SD_NUMA);
5537	rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5538
5539	sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5540	rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5541}
5542
5543/*
5544 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5545 * hold the hotplug lock.
5546 */
5547static void
5548cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5549{
5550	struct rq *rq = cpu_rq(cpu);
5551	struct sched_domain *tmp;
5552
5553	/* Remove the sched domains which do not contribute to scheduling. */
5554	for (tmp = sd; tmp; ) {
5555		struct sched_domain *parent = tmp->parent;
5556		if (!parent)
5557			break;
5558
5559		if (sd_parent_degenerate(tmp, parent)) {
5560			tmp->parent = parent->parent;
5561			if (parent->parent)
5562				parent->parent->child = tmp;
5563			/*
5564			 * Transfer SD_PREFER_SIBLING down in case of a
5565			 * degenerate parent; the spans match for this
5566			 * so the property transfers.
5567			 */
5568			if (parent->flags & SD_PREFER_SIBLING)
5569				tmp->flags |= SD_PREFER_SIBLING;
5570			destroy_sched_domain(parent, cpu);
5571		} else
5572			tmp = tmp->parent;
5573	}
5574
5575	if (sd && sd_degenerate(sd)) {
5576		tmp = sd;
5577		sd = sd->parent;
5578		destroy_sched_domain(tmp, cpu);
5579		if (sd)
5580			sd->child = NULL;
5581	}
5582
5583	sched_domain_debug(sd, cpu);
5584
5585	rq_attach_root(rq, rd);
5586	tmp = rq->sd;
5587	rcu_assign_pointer(rq->sd, sd);
5588	destroy_sched_domains(tmp, cpu);
5589
5590	update_top_cache_domain(cpu);
5591}
5592
5593/* cpus with isolated domains */
5594static cpumask_var_t cpu_isolated_map;
5595
5596/* Setup the mask of cpus configured for isolated domains */
5597static int __init isolated_cpu_setup(char *str)
5598{
5599	alloc_bootmem_cpumask_var(&cpu_isolated_map);
5600	cpulist_parse(str, cpu_isolated_map);
5601	return 1;
5602}
5603
5604__setup("isolcpus=", isolated_cpu_setup);
5605
5606static const struct cpumask *cpu_cpu_mask(int cpu)
5607{
5608	return cpumask_of_node(cpu_to_node(cpu));
5609}
5610
5611struct sd_data {
5612	struct sched_domain **__percpu sd;
5613	struct sched_group **__percpu sg;
5614	struct sched_group_power **__percpu sgp;
5615};
5616
5617struct s_data {
5618	struct sched_domain ** __percpu sd;
5619	struct root_domain	*rd;
5620};
5621
5622enum s_alloc {
5623	sa_rootdomain,
5624	sa_sd,
5625	sa_sd_storage,
5626	sa_none,
5627};
5628
5629struct sched_domain_topology_level;
5630
5631typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5632typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5633
5634#define SDTL_OVERLAP	0x01
5635
5636struct sched_domain_topology_level {
5637	sched_domain_init_f init;
5638	sched_domain_mask_f mask;
5639	int		    flags;
5640	int		    numa_level;
5641	struct sd_data      data;
5642};
5643
5644/*
5645 * Build an iteration mask that can exclude certain CPUs from the upwards
5646 * domain traversal.
5647 *
5648 * Asymmetric node setups can result in situations where the domain tree is of
5649 * unequal depth, make sure to skip domains that already cover the entire
5650 * range.
5651 *
5652 * In that case build_sched_domains() will have terminated the iteration early
5653 * and our sibling sd spans will be empty. Domains should always include the
5654 * cpu they're built on, so check that.
5655 *
5656 */
5657static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5658{
5659	const struct cpumask *span = sched_domain_span(sd);
5660	struct sd_data *sdd = sd->private;
5661	struct sched_domain *sibling;
5662	int i;
5663
5664	for_each_cpu(i, span) {
5665		sibling = *per_cpu_ptr(sdd->sd, i);
5666		if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5667			continue;
5668
5669		cpumask_set_cpu(i, sched_group_mask(sg));
5670	}
5671}
5672
5673/*
5674 * Return the canonical balance cpu for this group, this is the first cpu
5675 * of this group that's also in the iteration mask.
5676 */
5677int group_balance_cpu(struct sched_group *sg)
5678{
5679	return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5680}
5681
5682static int
5683build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5684{
5685	struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5686	const struct cpumask *span = sched_domain_span(sd);
5687	struct cpumask *covered = sched_domains_tmpmask;
5688	struct sd_data *sdd = sd->private;
5689	struct sched_domain *child;
5690	int i;
5691
5692	cpumask_clear(covered);
5693
5694	for_each_cpu(i, span) {
5695		struct cpumask *sg_span;
5696
5697		if (cpumask_test_cpu(i, covered))
5698			continue;
5699
5700		child = *per_cpu_ptr(sdd->sd, i);
5701
5702		/* See the comment near build_group_mask(). */
5703		if (!cpumask_test_cpu(i, sched_domain_span(child)))
5704			continue;
5705
5706		sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5707				GFP_KERNEL, cpu_to_node(cpu));
5708
5709		if (!sg)
5710			goto fail;
5711
5712		sg_span = sched_group_cpus(sg);
5713		if (child->child) {
5714			child = child->child;
5715			cpumask_copy(sg_span, sched_domain_span(child));
5716		} else
5717			cpumask_set_cpu(i, sg_span);
5718
5719		cpumask_or(covered, covered, sg_span);
5720
5721		sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5722		if (atomic_inc_return(&sg->sgp->ref) == 1)
5723			build_group_mask(sd, sg);
5724
5725		/*
5726		 * Initialize sgp->power such that even if we mess up the
5727		 * domains and no possible iteration will get us here, we won't
5728		 * die on a /0 trap.
5729		 */
5730		sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5731		sg->sgp->power_orig = sg->sgp->power;
5732
5733		/*
5734		 * Make sure the first group of this domain contains the
5735		 * canonical balance cpu. Otherwise the sched_domain iteration
5736		 * breaks. See update_sg_lb_stats().
5737		 */
5738		if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5739		    group_balance_cpu(sg) == cpu)
5740			groups = sg;
5741
5742		if (!first)
5743			first = sg;
5744		if (last)
5745			last->next = sg;
5746		last = sg;
5747		last->next = first;
5748	}
5749	sd->groups = groups;
5750
5751	return 0;
5752
5753fail:
5754	free_sched_groups(first, 0);
5755
5756	return -ENOMEM;
5757}
5758
5759static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5760{
5761	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5762	struct sched_domain *child = sd->child;
5763
5764	if (child)
5765		cpu = cpumask_first(sched_domain_span(child));
5766
5767	if (sg) {
5768		*sg = *per_cpu_ptr(sdd->sg, cpu);
5769		(*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5770		atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5771	}
5772
5773	return cpu;
5774}
5775
5776/*
5777 * build_sched_groups will build a circular linked list of the groups
5778 * covered by the given span, and will set each group's ->cpumask correctly,
5779 * and ->cpu_power to 0.
5780 *
5781 * Assumes the sched_domain tree is fully constructed
5782 */
5783static int
5784build_sched_groups(struct sched_domain *sd, int cpu)
5785{
5786	struct sched_group *first = NULL, *last = NULL;
5787	struct sd_data *sdd = sd->private;
5788	const struct cpumask *span = sched_domain_span(sd);
5789	struct cpumask *covered;
5790	int i;
5791
5792	get_group(cpu, sdd, &sd->groups);
5793	atomic_inc(&sd->groups->ref);
5794
5795	if (cpu != cpumask_first(span))
5796		return 0;
5797
5798	lockdep_assert_held(&sched_domains_mutex);
5799	covered = sched_domains_tmpmask;
5800
5801	cpumask_clear(covered);
5802
5803	for_each_cpu(i, span) {
5804		struct sched_group *sg;
5805		int group, j;
5806
5807		if (cpumask_test_cpu(i, covered))
5808			continue;
5809
5810		group = get_group(i, sdd, &sg);
5811		cpumask_clear(sched_group_cpus(sg));
5812		sg->sgp->power = 0;
5813		cpumask_setall(sched_group_mask(sg));
5814
5815		for_each_cpu(j, span) {
5816			if (get_group(j, sdd, NULL) != group)
5817				continue;
5818
5819			cpumask_set_cpu(j, covered);
5820			cpumask_set_cpu(j, sched_group_cpus(sg));
5821		}
5822
5823		if (!first)
5824			first = sg;
5825		if (last)
5826			last->next = sg;
5827		last = sg;
5828	}
5829	last->next = first;
5830
5831	return 0;
5832}
5833
5834/*
5835 * Initialize sched groups cpu_power.
5836 *
5837 * cpu_power indicates the capacity of sched group, which is used while
5838 * distributing the load between different sched groups in a sched domain.
5839 * Typically cpu_power for all the groups in a sched domain will be same unless
5840 * there are asymmetries in the topology. If there are asymmetries, group
5841 * having more cpu_power will pickup more load compared to the group having
5842 * less cpu_power.
5843 */
5844static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5845{
5846	struct sched_group *sg = sd->groups;
5847
5848	WARN_ON(!sg);
5849
5850	do {
5851		sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5852		sg = sg->next;
5853	} while (sg != sd->groups);
5854
5855	if (cpu != group_balance_cpu(sg))
5856		return;
5857
5858	update_group_power(sd, cpu);
5859	atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5860}
5861
5862int __weak arch_sd_sibling_asym_packing(void)
5863{
5864       return 0*SD_ASYM_PACKING;
5865}
5866
5867/*
5868 * Initializers for schedule domains
5869 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5870 */
5871
5872#ifdef CONFIG_SCHED_DEBUG
5873# define SD_INIT_NAME(sd, type)		sd->name = #type
5874#else
5875# define SD_INIT_NAME(sd, type)		do { } while (0)
5876#endif
5877
5878#define SD_INIT_FUNC(type)						\
5879static noinline struct sched_domain *					\
5880sd_init_##type(struct sched_domain_topology_level *tl, int cpu) 	\
5881{									\
5882	struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);	\
5883	*sd = SD_##type##_INIT;						\
5884	SD_INIT_NAME(sd, type);						\
5885	sd->private = &tl->data;					\
5886	return sd;							\
5887}
5888
5889SD_INIT_FUNC(CPU)
5890#ifdef CONFIG_SCHED_SMT
5891 SD_INIT_FUNC(SIBLING)
5892#endif
5893#ifdef CONFIG_SCHED_MC
5894 SD_INIT_FUNC(MC)
5895#endif
5896#ifdef CONFIG_SCHED_BOOK
5897 SD_INIT_FUNC(BOOK)
5898#endif
5899
5900static int default_relax_domain_level = -1;
5901int sched_domain_level_max;
5902
5903static int __init setup_relax_domain_level(char *str)
5904{
5905	if (kstrtoint(str, 0, &default_relax_domain_level))
5906		pr_warn("Unable to set relax_domain_level\n");
5907
5908	return 1;
5909}
5910__setup("relax_domain_level=", setup_relax_domain_level);
5911
5912static void set_domain_attribute(struct sched_domain *sd,
5913				 struct sched_domain_attr *attr)
5914{
5915	int request;
5916
5917	if (!attr || attr->relax_domain_level < 0) {
5918		if (default_relax_domain_level < 0)
5919			return;
5920		else
5921			request = default_relax_domain_level;
5922	} else
5923		request = attr->relax_domain_level;
5924	if (request < sd->level) {
5925		/* turn off idle balance on this domain */
5926		sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5927	} else {
5928		/* turn on idle balance on this domain */
5929		sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5930	}
5931}
5932
5933static void __sdt_free(const struct cpumask *cpu_map);
5934static int __sdt_alloc(const struct cpumask *cpu_map);
5935
5936static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5937				 const struct cpumask *cpu_map)
5938{
5939	switch (what) {
5940	case sa_rootdomain:
5941		if (!atomic_read(&d->rd->refcount))
5942			free_rootdomain(&d->rd->rcu); /* fall through */
5943	case sa_sd:
5944		free_percpu(d->sd); /* fall through */
5945	case sa_sd_storage:
5946		__sdt_free(cpu_map); /* fall through */
5947	case sa_none:
5948		break;
5949	}
5950}
5951
5952static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5953						   const struct cpumask *cpu_map)
5954{
5955	memset(d, 0, sizeof(*d));
5956
5957	if (__sdt_alloc(cpu_map))
5958		return sa_sd_storage;
5959	d->sd = alloc_percpu(struct sched_domain *);
5960	if (!d->sd)
5961		return sa_sd_storage;
5962	d->rd = alloc_rootdomain();
5963	if (!d->rd)
5964		return sa_sd;
5965	return sa_rootdomain;
5966}
5967
5968/*
5969 * NULL the sd_data elements we've used to build the sched_domain and
5970 * sched_group structure so that the subsequent __free_domain_allocs()
5971 * will not free the data we're using.
5972 */
5973static void claim_allocations(int cpu, struct sched_domain *sd)
5974{
5975	struct sd_data *sdd = sd->private;
5976
5977	WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5978	*per_cpu_ptr(sdd->sd, cpu) = NULL;
5979
5980	if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5981		*per_cpu_ptr(sdd->sg, cpu) = NULL;
5982
5983	if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5984		*per_cpu_ptr(sdd->sgp, cpu) = NULL;
5985}
5986
5987#ifdef CONFIG_SCHED_SMT
5988static const struct cpumask *cpu_smt_mask(int cpu)
5989{
5990	return topology_thread_cpumask(cpu);
5991}
5992#endif
5993
5994/*
5995 * Topology list, bottom-up.
5996 */
5997static struct sched_domain_topology_level default_topology[] = {
5998#ifdef CONFIG_SCHED_SMT
5999	{ sd_init_SIBLING, cpu_smt_mask, },
6000#endif
6001#ifdef CONFIG_SCHED_MC
6002	{ sd_init_MC, cpu_coregroup_mask, },
6003#endif
6004#ifdef CONFIG_SCHED_BOOK
6005	{ sd_init_BOOK, cpu_book_mask, },
6006#endif
6007	{ sd_init_CPU, cpu_cpu_mask, },
6008	{ NULL, },
6009};
6010
6011static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6012
6013#define for_each_sd_topology(tl)			\
6014	for (tl = sched_domain_topology; tl->init; tl++)
6015
6016#ifdef CONFIG_NUMA
6017
6018static int sched_domains_numa_levels;
6019static int *sched_domains_numa_distance;
6020static struct cpumask ***sched_domains_numa_masks;
6021static int sched_domains_curr_level;
6022
6023static inline int sd_local_flags(int level)
6024{
6025	if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6026		return 0;
6027
6028	return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6029}
6030
6031static struct sched_domain *
6032sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6033{
6034	struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6035	int level = tl->numa_level;
6036	int sd_weight = cpumask_weight(
6037			sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6038
6039	*sd = (struct sched_domain){
6040		.min_interval		= sd_weight,
6041		.max_interval		= 2*sd_weight,
6042		.busy_factor		= 32,
6043		.imbalance_pct		= 125,
6044		.cache_nice_tries	= 2,
6045		.busy_idx		= 3,
6046		.idle_idx		= 2,
6047		.newidle_idx		= 0,
6048		.wake_idx		= 0,
6049		.forkexec_idx		= 0,
6050
6051		.flags			= 1*SD_LOAD_BALANCE
6052					| 1*SD_BALANCE_NEWIDLE
6053					| 0*SD_BALANCE_EXEC
6054					| 0*SD_BALANCE_FORK
6055					| 0*SD_BALANCE_WAKE
6056					| 0*SD_WAKE_AFFINE
6057					| 0*SD_SHARE_CPUPOWER
6058					| 0*SD_SHARE_PKG_RESOURCES
6059					| 1*SD_SERIALIZE
6060					| 0*SD_PREFER_SIBLING
6061					| 1*SD_NUMA
6062					| sd_local_flags(level)
6063					,
6064		.last_balance		= jiffies,
6065		.balance_interval	= sd_weight,
6066	};
6067	SD_INIT_NAME(sd, NUMA);
6068	sd->private = &tl->data;
6069
6070	/*
6071	 * Ugly hack to pass state to sd_numa_mask()...
6072	 */
6073	sched_domains_curr_level = tl->numa_level;
6074
6075	return sd;
6076}
6077
6078static const struct cpumask *sd_numa_mask(int cpu)
6079{
6080	return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6081}
6082
6083static void sched_numa_warn(const char *str)
6084{
6085	static int done = false;
6086	int i,j;
6087
6088	if (done)
6089		return;
6090
6091	done = true;
6092
6093	printk(KERN_WARNING "ERROR: %s\n\n", str);
6094
6095	for (i = 0; i < nr_node_ids; i++) {
6096		printk(KERN_WARNING "  ");
6097		for (j = 0; j < nr_node_ids; j++)
6098			printk(KERN_CONT "%02d ", node_distance(i,j));
6099		printk(KERN_CONT "\n");
6100	}
6101	printk(KERN_WARNING "\n");
6102}
6103
6104static bool find_numa_distance(int distance)
6105{
6106	int i;
6107
6108	if (distance == node_distance(0, 0))
6109		return true;
6110
6111	for (i = 0; i < sched_domains_numa_levels; i++) {
6112		if (sched_domains_numa_distance[i] == distance)
6113			return true;
6114	}
6115
6116	return false;
6117}
6118
6119static void sched_init_numa(void)
6120{
6121	int next_distance, curr_distance = node_distance(0, 0);
6122	struct sched_domain_topology_level *tl;
6123	int level = 0;
6124	int i, j, k;
6125
6126	sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6127	if (!sched_domains_numa_distance)
6128		return;
6129
6130	/*
6131	 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6132	 * unique distances in the node_distance() table.
6133	 *
6134	 * Assumes node_distance(0,j) includes all distances in
6135	 * node_distance(i,j) in order to avoid cubic time.
6136	 */
6137	next_distance = curr_distance;
6138	for (i = 0; i < nr_node_ids; i++) {
6139		for (j = 0; j < nr_node_ids; j++) {
6140			for (k = 0; k < nr_node_ids; k++) {
6141				int distance = node_distance(i, k);
6142
6143				if (distance > curr_distance &&
6144				    (distance < next_distance ||
6145				     next_distance == curr_distance))
6146					next_distance = distance;
6147
6148				/*
6149				 * While not a strong assumption it would be nice to know
6150				 * about cases where if node A is connected to B, B is not
6151				 * equally connected to A.
6152				 */
6153				if (sched_debug() && node_distance(k, i) != distance)
6154					sched_numa_warn("Node-distance not symmetric");
6155
6156				if (sched_debug() && i && !find_numa_distance(distance))
6157					sched_numa_warn("Node-0 not representative");
6158			}
6159			if (next_distance != curr_distance) {
6160				sched_domains_numa_distance[level++] = next_distance;
6161				sched_domains_numa_levels = level;
6162				curr_distance = next_distance;
6163			} else break;
6164		}
6165
6166		/*
6167		 * In case of sched_debug() we verify the above assumption.
6168		 */
6169		if (!sched_debug())
6170			break;
6171	}
6172	/*
6173	 * 'level' contains the number of unique distances, excluding the
6174	 * identity distance node_distance(i,i).
6175	 *
6176	 * The sched_domains_numa_distance[] array includes the actual distance
6177	 * numbers.
6178	 */
6179
6180	/*
6181	 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6182	 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6183	 * the array will contain less then 'level' members. This could be
6184	 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6185	 * in other functions.
6186	 *
6187	 * We reset it to 'level' at the end of this function.
6188	 */
6189	sched_domains_numa_levels = 0;
6190
6191	sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6192	if (!sched_domains_numa_masks)
6193		return;
6194
6195	/*
6196	 * Now for each level, construct a mask per node which contains all
6197	 * cpus of nodes that are that many hops away from us.
6198	 */
6199	for (i = 0; i < level; i++) {
6200		sched_domains_numa_masks[i] =
6201			kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6202		if (!sched_domains_numa_masks[i])
6203			return;
6204
6205		for (j = 0; j < nr_node_ids; j++) {
6206			struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6207			if (!mask)
6208				return;
6209
6210			sched_domains_numa_masks[i][j] = mask;
6211
6212			for (k = 0; k < nr_node_ids; k++) {
6213				if (node_distance(j, k) > sched_domains_numa_distance[i])
6214					continue;
6215
6216				cpumask_or(mask, mask, cpumask_of_node(k));
6217			}
6218		}
6219	}
6220
6221	tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6222			sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6223	if (!tl)
6224		return;
6225
6226	/*
6227	 * Copy the default topology bits..
6228	 */
6229	for (i = 0; default_topology[i].init; i++)
6230		tl[i] = default_topology[i];
6231
6232	/*
6233	 * .. and append 'j' levels of NUMA goodness.
6234	 */
6235	for (j = 0; j < level; i++, j++) {
6236		tl[i] = (struct sched_domain_topology_level){
6237			.init = sd_numa_init,
6238			.mask = sd_numa_mask,
6239			.flags = SDTL_OVERLAP,
6240			.numa_level = j,
6241		};
6242	}
6243
6244	sched_domain_topology = tl;
6245
6246	sched_domains_numa_levels = level;
6247}
6248
6249static void sched_domains_numa_masks_set(int cpu)
6250{
6251	int i, j;
6252	int node = cpu_to_node(cpu);
6253
6254	for (i = 0; i < sched_domains_numa_levels; i++) {
6255		for (j = 0; j < nr_node_ids; j++) {
6256			if (node_distance(j, node) <= sched_domains_numa_distance[i])
6257				cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6258		}
6259	}
6260}
6261
6262static void sched_domains_numa_masks_clear(int cpu)
6263{
6264	int i, j;
6265	for (i = 0; i < sched_domains_numa_levels; i++) {
6266		for (j = 0; j < nr_node_ids; j++)
6267			cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6268	}
6269}
6270
6271/*
6272 * Update sched_domains_numa_masks[level][node] array when new cpus
6273 * are onlined.
6274 */
6275static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6276					   unsigned long action,
6277					   void *hcpu)
6278{
6279	int cpu = (long)hcpu;
6280
6281	switch (action & ~CPU_TASKS_FROZEN) {
6282	case CPU_ONLINE:
6283		sched_domains_numa_masks_set(cpu);
6284		break;
6285
6286	case CPU_DEAD:
6287		sched_domains_numa_masks_clear(cpu);
6288		break;
6289
6290	default:
6291		return NOTIFY_DONE;
6292	}
6293
6294	return NOTIFY_OK;
6295}
6296#else
6297static inline void sched_init_numa(void)
6298{
6299}
6300
6301static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6302					   unsigned long action,
6303					   void *hcpu)
6304{
6305	return 0;
6306}
6307#endif /* CONFIG_NUMA */
6308
6309static int __sdt_alloc(const struct cpumask *cpu_map)
6310{
6311	struct sched_domain_topology_level *tl;
6312	int j;
6313
6314	for_each_sd_topology(tl) {
6315		struct sd_data *sdd = &tl->data;
6316
6317		sdd->sd = alloc_percpu(struct sched_domain *);
6318		if (!sdd->sd)
6319			return -ENOMEM;
6320
6321		sdd->sg = alloc_percpu(struct sched_group *);
6322		if (!sdd->sg)
6323			return -ENOMEM;
6324
6325		sdd->sgp = alloc_percpu(struct sched_group_power *);
6326		if (!sdd->sgp)
6327			return -ENOMEM;
6328
6329		for_each_cpu(j, cpu_map) {
6330			struct sched_domain *sd;
6331			struct sched_group *sg;
6332			struct sched_group_power *sgp;
6333
6334		       	sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6335					GFP_KERNEL, cpu_to_node(j));
6336			if (!sd)
6337				return -ENOMEM;
6338
6339			*per_cpu_ptr(sdd->sd, j) = sd;
6340
6341			sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6342					GFP_KERNEL, cpu_to_node(j));
6343			if (!sg)
6344				return -ENOMEM;
6345
6346			sg->next = sg;
6347
6348			*per_cpu_ptr(sdd->sg, j) = sg;
6349
6350			sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6351					GFP_KERNEL, cpu_to_node(j));
6352			if (!sgp)
6353				return -ENOMEM;
6354
6355			*per_cpu_ptr(sdd->sgp, j) = sgp;
6356		}
6357	}
6358
6359	return 0;
6360}
6361
6362static void __sdt_free(const struct cpumask *cpu_map)
6363{
6364	struct sched_domain_topology_level *tl;
6365	int j;
6366
6367	for_each_sd_topology(tl) {
6368		struct sd_data *sdd = &tl->data;
6369
6370		for_each_cpu(j, cpu_map) {
6371			struct sched_domain *sd;
6372
6373			if (sdd->sd) {
6374				sd = *per_cpu_ptr(sdd->sd, j);
6375				if (sd && (sd->flags & SD_OVERLAP))
6376					free_sched_groups(sd->groups, 0);
6377				kfree(*per_cpu_ptr(sdd->sd, j));
6378			}
6379
6380			if (sdd->sg)
6381				kfree(*per_cpu_ptr(sdd->sg, j));
6382			if (sdd->sgp)
6383				kfree(*per_cpu_ptr(sdd->sgp, j));
6384		}
6385		free_percpu(sdd->sd);
6386		sdd->sd = NULL;
6387		free_percpu(sdd->sg);
6388		sdd->sg = NULL;
6389		free_percpu(sdd->sgp);
6390		sdd->sgp = NULL;
6391	}
6392}
6393
6394struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6395		const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6396		struct sched_domain *child, int cpu)
6397{
6398	struct sched_domain *sd = tl->init(tl, cpu);
6399	if (!sd)
6400		return child;
6401
6402	cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6403	if (child) {
6404		sd->level = child->level + 1;
6405		sched_domain_level_max = max(sched_domain_level_max, sd->level);
6406		child->parent = sd;
6407		sd->child = child;
6408	}
6409	set_domain_attribute(sd, attr);
6410
6411	return sd;
6412}
6413
6414/*
6415 * Build sched domains for a given set of cpus and attach the sched domains
6416 * to the individual cpus
6417 */
6418static int build_sched_domains(const struct cpumask *cpu_map,
6419			       struct sched_domain_attr *attr)
6420{
6421	enum s_alloc alloc_state;
6422	struct sched_domain *sd;
6423	struct s_data d;
6424	int i, ret = -ENOMEM;
6425
6426	alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6427	if (alloc_state != sa_rootdomain)
6428		goto error;
6429
6430	/* Set up domains for cpus specified by the cpu_map. */
6431	for_each_cpu(i, cpu_map) {
6432		struct sched_domain_topology_level *tl;
6433
6434		sd = NULL;
6435		for_each_sd_topology(tl) {
6436			sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6437			if (tl == sched_domain_topology)
6438				*per_cpu_ptr(d.sd, i) = sd;
6439			if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6440				sd->flags |= SD_OVERLAP;
6441			if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6442				break;
6443		}
6444	}
6445
6446	/* Build the groups for the domains */
6447	for_each_cpu(i, cpu_map) {
6448		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6449			sd->span_weight = cpumask_weight(sched_domain_span(sd));
6450			if (sd->flags & SD_OVERLAP) {
6451				if (build_overlap_sched_groups(sd, i))
6452					goto error;
6453			} else {
6454				if (build_sched_groups(sd, i))
6455					goto error;
6456			}
6457		}
6458	}
6459
6460	/* Calculate CPU power for physical packages and nodes */
6461	for (i = nr_cpumask_bits-1; i >= 0; i--) {
6462		if (!cpumask_test_cpu(i, cpu_map))
6463			continue;
6464
6465		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6466			claim_allocations(i, sd);
6467			init_sched_groups_power(i, sd);
6468		}
6469	}
6470
6471	/* Attach the domains */
6472	rcu_read_lock();
6473	for_each_cpu(i, cpu_map) {
6474		sd = *per_cpu_ptr(d.sd, i);
6475		cpu_attach_domain(sd, d.rd, i);
6476	}
6477	rcu_read_unlock();
6478
6479	ret = 0;
6480error:
6481	__free_domain_allocs(&d, alloc_state, cpu_map);
6482	return ret;
6483}
6484
6485static cpumask_var_t *doms_cur;	/* current sched domains */
6486static int ndoms_cur;		/* number of sched domains in 'doms_cur' */
6487static struct sched_domain_attr *dattr_cur;
6488				/* attribues of custom domains in 'doms_cur' */
6489
6490/*
6491 * Special case: If a kmalloc of a doms_cur partition (array of
6492 * cpumask) fails, then fallback to a single sched domain,
6493 * as determined by the single cpumask fallback_doms.
6494 */
6495static cpumask_var_t fallback_doms;
6496
6497/*
6498 * arch_update_cpu_topology lets virtualized architectures update the
6499 * cpu core maps. It is supposed to return 1 if the topology changed
6500 * or 0 if it stayed the same.
6501 */
6502int __weak arch_update_cpu_topology(void)
6503{
6504	return 0;
6505}
6506
6507cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6508{
6509	int i;
6510	cpumask_var_t *doms;
6511
6512	doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6513	if (!doms)
6514		return NULL;
6515	for (i = 0; i < ndoms; i++) {
6516		if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6517			free_sched_domains(doms, i);
6518			return NULL;
6519		}
6520	}
6521	return doms;
6522}
6523
6524void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6525{
6526	unsigned int i;
6527	for (i = 0; i < ndoms; i++)
6528		free_cpumask_var(doms[i]);
6529	kfree(doms);
6530}
6531
6532/*
6533 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6534 * For now this just excludes isolated cpus, but could be used to
6535 * exclude other special cases in the future.
6536 */
6537static int init_sched_domains(const struct cpumask *cpu_map)
6538{
6539	int err;
6540
6541	arch_update_cpu_topology();
6542	ndoms_cur = 1;
6543	doms_cur = alloc_sched_domains(ndoms_cur);
6544	if (!doms_cur)
6545		doms_cur = &fallback_doms;
6546	cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6547	err = build_sched_domains(doms_cur[0], NULL);
6548	register_sched_domain_sysctl();
6549
6550	return err;
6551}
6552
6553/*
6554 * Detach sched domains from a group of cpus specified in cpu_map
6555 * These cpus will now be attached to the NULL domain
6556 */
6557static void detach_destroy_domains(const struct cpumask *cpu_map)
6558{
6559	int i;
6560
6561	rcu_read_lock();
6562	for_each_cpu(i, cpu_map)
6563		cpu_attach_domain(NULL, &def_root_domain, i);
6564	rcu_read_unlock();
6565}
6566
6567/* handle null as "default" */
6568static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6569			struct sched_domain_attr *new, int idx_new)
6570{
6571	struct sched_domain_attr tmp;
6572
6573	/* fast path */
6574	if (!new && !cur)
6575		return 1;
6576
6577	tmp = SD_ATTR_INIT;
6578	return !memcmp(cur ? (cur + idx_cur) : &tmp,
6579			new ? (new + idx_new) : &tmp,
6580			sizeof(struct sched_domain_attr));
6581}
6582
6583/*
6584 * Partition sched domains as specified by the 'ndoms_new'
6585 * cpumasks in the array doms_new[] of cpumasks. This compares
6586 * doms_new[] to the current sched domain partitioning, doms_cur[].
6587 * It destroys each deleted domain and builds each new domain.
6588 *
6589 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6590 * The masks don't intersect (don't overlap.) We should setup one
6591 * sched domain for each mask. CPUs not in any of the cpumasks will
6592 * not be load balanced. If the same cpumask appears both in the
6593 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6594 * it as it is.
6595 *
6596 * The passed in 'doms_new' should be allocated using
6597 * alloc_sched_domains.  This routine takes ownership of it and will
6598 * free_sched_domains it when done with it. If the caller failed the
6599 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6600 * and partition_sched_domains() will fallback to the single partition
6601 * 'fallback_doms', it also forces the domains to be rebuilt.
6602 *
6603 * If doms_new == NULL it will be replaced with cpu_online_mask.
6604 * ndoms_new == 0 is a special case for destroying existing domains,
6605 * and it will not create the default domain.
6606 *
6607 * Call with hotplug lock held
6608 */
6609void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6610			     struct sched_domain_attr *dattr_new)
6611{
6612	int i, j, n;
6613	int new_topology;
6614
6615	mutex_lock(&sched_domains_mutex);
6616
6617	/* always unregister in case we don't destroy any domains */
6618	unregister_sched_domain_sysctl();
6619
6620	/* Let architecture update cpu core mappings. */
6621	new_topology = arch_update_cpu_topology();
6622
6623	n = doms_new ? ndoms_new : 0;
6624
6625	/* Destroy deleted domains */
6626	for (i = 0; i < ndoms_cur; i++) {
6627		for (j = 0; j < n && !new_topology; j++) {
6628			if (cpumask_equal(doms_cur[i], doms_new[j])
6629			    && dattrs_equal(dattr_cur, i, dattr_new, j))
6630				goto match1;
6631		}
6632		/* no match - a current sched domain not in new doms_new[] */
6633		detach_destroy_domains(doms_cur[i]);
6634match1:
6635		;
6636	}
6637
6638	n = ndoms_cur;
6639	if (doms_new == NULL) {
6640		n = 0;
6641		doms_new = &fallback_doms;
6642		cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6643		WARN_ON_ONCE(dattr_new);
6644	}
6645
6646	/* Build new domains */
6647	for (i = 0; i < ndoms_new; i++) {
6648		for (j = 0; j < n && !new_topology; j++) {
6649			if (cpumask_equal(doms_new[i], doms_cur[j])
6650			    && dattrs_equal(dattr_new, i, dattr_cur, j))
6651				goto match2;
6652		}
6653		/* no match - add a new doms_new */
6654		build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6655match2:
6656		;
6657	}
6658
6659	/* Remember the new sched domains */
6660	if (doms_cur != &fallback_doms)
6661		free_sched_domains(doms_cur, ndoms_cur);
6662	kfree(dattr_cur);	/* kfree(NULL) is safe */
6663	doms_cur = doms_new;
6664	dattr_cur = dattr_new;
6665	ndoms_cur = ndoms_new;
6666
6667	register_sched_domain_sysctl();
6668
6669	mutex_unlock(&sched_domains_mutex);
6670}
6671
6672static int num_cpus_frozen;	/* used to mark begin/end of suspend/resume */
6673
6674/*
6675 * Update cpusets according to cpu_active mask.  If cpusets are
6676 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6677 * around partition_sched_domains().
6678 *
6679 * If we come here as part of a suspend/resume, don't touch cpusets because we
6680 * want to restore it back to its original state upon resume anyway.
6681 */
6682static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6683			     void *hcpu)
6684{
6685	switch (action) {
6686	case CPU_ONLINE_FROZEN:
6687	case CPU_DOWN_FAILED_FROZEN:
6688
6689		/*
6690		 * num_cpus_frozen tracks how many CPUs are involved in suspend
6691		 * resume sequence. As long as this is not the last online
6692		 * operation in the resume sequence, just build a single sched
6693		 * domain, ignoring cpusets.
6694		 */
6695		num_cpus_frozen--;
6696		if (likely(num_cpus_frozen)) {
6697			partition_sched_domains(1, NULL, NULL);
6698			break;
6699		}
6700
6701		/*
6702		 * This is the last CPU online operation. So fall through and
6703		 * restore the original sched domains by considering the
6704		 * cpuset configurations.
6705		 */
6706
6707	case CPU_ONLINE:
6708	case CPU_DOWN_FAILED:
6709		cpuset_update_active_cpus(true);
6710		break;
6711	default:
6712		return NOTIFY_DONE;
6713	}
6714	return NOTIFY_OK;
6715}
6716
6717static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6718			       void *hcpu)
6719{
6720	switch (action) {
6721	case CPU_DOWN_PREPARE:
6722		cpuset_update_active_cpus(false);
6723		break;
6724	case CPU_DOWN_PREPARE_FROZEN:
6725		num_cpus_frozen++;
6726		partition_sched_domains(1, NULL, NULL);
6727		break;
6728	default:
6729		return NOTIFY_DONE;
6730	}
6731	return NOTIFY_OK;
6732}
6733
6734void __init sched_init_smp(void)
6735{
6736	cpumask_var_t non_isolated_cpus;
6737
6738	alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6739	alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6740
6741	sched_init_numa();
6742
6743	/*
6744	 * There's no userspace yet to cause hotplug operations; hence all the
6745	 * cpu masks are stable and all blatant races in the below code cannot
6746	 * happen.
6747	 */
6748	mutex_lock(&sched_domains_mutex);
6749	init_sched_domains(cpu_active_mask);
6750	cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6751	if (cpumask_empty(non_isolated_cpus))
6752		cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6753	mutex_unlock(&sched_domains_mutex);
6754
6755	hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6756	hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6757	hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6758
6759	init_hrtick();
6760
6761	/* Move init over to a non-isolated CPU */
6762	if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6763		BUG();
6764	sched_init_granularity();
6765	free_cpumask_var(non_isolated_cpus);
6766
6767	init_sched_rt_class();
6768	init_sched_dl_class();
6769}
6770#else
6771void __init sched_init_smp(void)
6772{
6773	sched_init_granularity();
6774}
6775#endif /* CONFIG_SMP */
6776
6777const_debug unsigned int sysctl_timer_migration = 1;
6778
6779int in_sched_functions(unsigned long addr)
6780{
6781	return in_lock_functions(addr) ||
6782		(addr >= (unsigned long)__sched_text_start
6783		&& addr < (unsigned long)__sched_text_end);
6784}
6785
6786#ifdef CONFIG_CGROUP_SCHED
6787/*
6788 * Default task group.
6789 * Every task in system belongs to this group at bootup.
6790 */
6791struct task_group root_task_group;
6792LIST_HEAD(task_groups);
6793#endif
6794
6795DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6796
6797void __init sched_init(void)
6798{
6799	int i, j;
6800	unsigned long alloc_size = 0, ptr;
6801
6802#ifdef CONFIG_FAIR_GROUP_SCHED
6803	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6804#endif
6805#ifdef CONFIG_RT_GROUP_SCHED
6806	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6807#endif
6808#ifdef CONFIG_CPUMASK_OFFSTACK
6809	alloc_size += num_possible_cpus() * cpumask_size();
6810#endif
6811	if (alloc_size) {
6812		ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6813
6814#ifdef CONFIG_FAIR_GROUP_SCHED
6815		root_task_group.se = (struct sched_entity **)ptr;
6816		ptr += nr_cpu_ids * sizeof(void **);
6817
6818		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6819		ptr += nr_cpu_ids * sizeof(void **);
6820
6821#endif /* CONFIG_FAIR_GROUP_SCHED */
6822#ifdef CONFIG_RT_GROUP_SCHED
6823		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6824		ptr += nr_cpu_ids * sizeof(void **);
6825
6826		root_task_group.rt_rq = (struct rt_rq **)ptr;
6827		ptr += nr_cpu_ids * sizeof(void **);
6828
6829#endif /* CONFIG_RT_GROUP_SCHED */
6830#ifdef CONFIG_CPUMASK_OFFSTACK
6831		for_each_possible_cpu(i) {
6832			per_cpu(load_balance_mask, i) = (void *)ptr;
6833			ptr += cpumask_size();
6834		}
6835#endif /* CONFIG_CPUMASK_OFFSTACK */
6836	}
6837
6838	init_rt_bandwidth(&def_rt_bandwidth,
6839			global_rt_period(), global_rt_runtime());
6840	init_dl_bandwidth(&def_dl_bandwidth,
6841			global_rt_period(), global_rt_runtime());
6842
6843#ifdef CONFIG_SMP
6844	init_defrootdomain();
6845#endif
6846
6847#ifdef CONFIG_RT_GROUP_SCHED
6848	init_rt_bandwidth(&root_task_group.rt_bandwidth,
6849			global_rt_period(), global_rt_runtime());
6850#endif /* CONFIG_RT_GROUP_SCHED */
6851
6852#ifdef CONFIG_CGROUP_SCHED
6853	list_add(&root_task_group.list, &task_groups);
6854	INIT_LIST_HEAD(&root_task_group.children);
6855	INIT_LIST_HEAD(&root_task_group.siblings);
6856	autogroup_init(&init_task);
6857
6858#endif /* CONFIG_CGROUP_SCHED */
6859
6860	for_each_possible_cpu(i) {
6861		struct rq *rq;
6862
6863		rq = cpu_rq(i);
6864		raw_spin_lock_init(&rq->lock);
6865		rq->nr_running = 0;
6866		rq->calc_load_active = 0;
6867		rq->calc_load_update = jiffies + LOAD_FREQ;
6868		init_cfs_rq(&rq->cfs);
6869		init_rt_rq(&rq->rt, rq);
6870		init_dl_rq(&rq->dl, rq);
6871#ifdef CONFIG_FAIR_GROUP_SCHED
6872		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6873		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6874		/*
6875		 * How much cpu bandwidth does root_task_group get?
6876		 *
6877		 * In case of task-groups formed thr' the cgroup filesystem, it
6878		 * gets 100% of the cpu resources in the system. This overall
6879		 * system cpu resource is divided among the tasks of
6880		 * root_task_group and its child task-groups in a fair manner,
6881		 * based on each entity's (task or task-group's) weight
6882		 * (se->load.weight).
6883		 *
6884		 * In other words, if root_task_group has 10 tasks of weight
6885		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6886		 * then A0's share of the cpu resource is:
6887		 *
6888		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6889		 *
6890		 * We achieve this by letting root_task_group's tasks sit
6891		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6892		 */
6893		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6894		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6895#endif /* CONFIG_FAIR_GROUP_SCHED */
6896
6897		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6898#ifdef CONFIG_RT_GROUP_SCHED
6899		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6900#endif
6901
6902		for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6903			rq->cpu_load[j] = 0;
6904
6905		rq->last_load_update_tick = jiffies;
6906
6907#ifdef CONFIG_SMP
6908		rq->sd = NULL;
6909		rq->rd = NULL;
6910		rq->cpu_power = SCHED_POWER_SCALE;
6911		rq->post_schedule = 0;
6912		rq->active_balance = 0;
6913		rq->next_balance = jiffies;
6914		rq->push_cpu = 0;
6915		rq->cpu = i;
6916		rq->online = 0;
6917		rq->idle_stamp = 0;
6918		rq->avg_idle = 2*sysctl_sched_migration_cost;
6919		rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6920
6921		INIT_LIST_HEAD(&rq->cfs_tasks);
6922
6923		rq_attach_root(rq, &def_root_domain);
6924#ifdef CONFIG_NO_HZ_COMMON
6925		rq->nohz_flags = 0;
6926#endif
6927#ifdef CONFIG_NO_HZ_FULL
6928		rq->last_sched_tick = 0;
6929#endif
6930#endif
6931		init_rq_hrtick(rq);
6932		atomic_set(&rq->nr_iowait, 0);
6933	}
6934
6935	set_load_weight(&init_task);
6936
6937#ifdef CONFIG_PREEMPT_NOTIFIERS
6938	INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6939#endif
6940
6941	/*
6942	 * The boot idle thread does lazy MMU switching as well:
6943	 */
6944	atomic_inc(&init_mm.mm_count);
6945	enter_lazy_tlb(&init_mm, current);
6946
6947	/*
6948	 * Make us the idle thread. Technically, schedule() should not be
6949	 * called from this thread, however somewhere below it might be,
6950	 * but because we are the idle thread, we just pick up running again
6951	 * when this runqueue becomes "idle".
6952	 */
6953	init_idle(current, smp_processor_id());
6954
6955	calc_load_update = jiffies + LOAD_FREQ;
6956
6957	/*
6958	 * During early bootup we pretend to be a normal task:
6959	 */
6960	current->sched_class = &fair_sched_class;
6961
6962#ifdef CONFIG_SMP
6963	zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6964	/* May be allocated at isolcpus cmdline parse time */
6965	if (cpu_isolated_map == NULL)
6966		zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6967	idle_thread_set_boot_cpu();
6968#endif
6969	init_sched_fair_class();
6970
6971	scheduler_running = 1;
6972}
6973
6974#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6975static inline int preempt_count_equals(int preempt_offset)
6976{
6977	int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6978
6979	return (nested == preempt_offset);
6980}
6981
6982void __might_sleep(const char *file, int line, int preempt_offset)
6983{
6984	static unsigned long prev_jiffy;	/* ratelimiting */
6985
6986	rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6987	if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6988	     !is_idle_task(current)) ||
6989	    system_state != SYSTEM_RUNNING || oops_in_progress)
6990		return;
6991	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6992		return;
6993	prev_jiffy = jiffies;
6994
6995	printk(KERN_ERR
6996		"BUG: sleeping function called from invalid context at %s:%d\n",
6997			file, line);
6998	printk(KERN_ERR
6999		"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7000			in_atomic(), irqs_disabled(),
7001			current->pid, current->comm);
7002
7003	debug_show_held_locks(current);
7004	if (irqs_disabled())
7005		print_irqtrace_events(current);
7006#ifdef CONFIG_DEBUG_PREEMPT
7007	if (!preempt_count_equals(preempt_offset)) {
7008		pr_err("Preemption disabled at:");
7009		print_ip_sym(current->preempt_disable_ip);
7010		pr_cont("\n");
7011	}
7012#endif
7013	dump_stack();
7014}
7015EXPORT_SYMBOL(__might_sleep);
7016#endif
7017
7018#ifdef CONFIG_MAGIC_SYSRQ
7019static void normalize_task(struct rq *rq, struct task_struct *p)
7020{
7021	const struct sched_class *prev_class = p->sched_class;
7022	struct sched_attr attr = {
7023		.sched_policy = SCHED_NORMAL,
7024	};
7025	int old_prio = p->prio;
7026	int on_rq;
7027
7028	on_rq = p->on_rq;
7029	if (on_rq)
7030		dequeue_task(rq, p, 0);
7031	__setscheduler(rq, p, &attr);
7032	if (on_rq) {
7033		enqueue_task(rq, p, 0);
7034		resched_task(rq->curr);
7035	}
7036
7037	check_class_changed(rq, p, prev_class, old_prio);
7038}
7039
7040void normalize_rt_tasks(void)
7041{
7042	struct task_struct *g, *p;
7043	unsigned long flags;
7044	struct rq *rq;
7045
7046	read_lock_irqsave(&tasklist_lock, flags);
7047	do_each_thread(g, p) {
7048		/*
7049		 * Only normalize user tasks:
7050		 */
7051		if (!p->mm)
7052			continue;
7053
7054		p->se.exec_start		= 0;
7055#ifdef CONFIG_SCHEDSTATS
7056		p->se.statistics.wait_start	= 0;
7057		p->se.statistics.sleep_start	= 0;
7058		p->se.statistics.block_start	= 0;
7059#endif
7060
7061		if (!dl_task(p) && !rt_task(p)) {
7062			/*
7063			 * Renice negative nice level userspace
7064			 * tasks back to 0:
7065			 */
7066			if (task_nice(p) < 0 && p->mm)
7067				set_user_nice(p, 0);
7068			continue;
7069		}
7070
7071		raw_spin_lock(&p->pi_lock);
7072		rq = __task_rq_lock(p);
7073
7074		normalize_task(rq, p);
7075
7076		__task_rq_unlock(rq);
7077		raw_spin_unlock(&p->pi_lock);
7078	} while_each_thread(g, p);
7079
7080	read_unlock_irqrestore(&tasklist_lock, flags);
7081}
7082
7083#endif /* CONFIG_MAGIC_SYSRQ */
7084
7085#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7086/*
7087 * These functions are only useful for the IA64 MCA handling, or kdb.
7088 *
7089 * They can only be called when the whole system has been
7090 * stopped - every CPU needs to be quiescent, and no scheduling
7091 * activity can take place. Using them for anything else would
7092 * be a serious bug, and as a result, they aren't even visible
7093 * under any other configuration.
7094 */
7095
7096/**
7097 * curr_task - return the current task for a given cpu.
7098 * @cpu: the processor in question.
7099 *
7100 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7101 *
7102 * Return: The current task for @cpu.
7103 */
7104struct task_struct *curr_task(int cpu)
7105{
7106	return cpu_curr(cpu);
7107}
7108
7109#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7110
7111#ifdef CONFIG_IA64
7112/**
7113 * set_curr_task - set the current task for a given cpu.
7114 * @cpu: the processor in question.
7115 * @p: the task pointer to set.
7116 *
7117 * Description: This function must only be used when non-maskable interrupts
7118 * are serviced on a separate stack. It allows the architecture to switch the
7119 * notion of the current task on a cpu in a non-blocking manner. This function
7120 * must be called with all CPU's synchronized, and interrupts disabled, the
7121 * and caller must save the original value of the current task (see
7122 * curr_task() above) and restore that value before reenabling interrupts and
7123 * re-starting the system.
7124 *
7125 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7126 */
7127void set_curr_task(int cpu, struct task_struct *p)
7128{
7129	cpu_curr(cpu) = p;
7130}
7131
7132#endif
7133
7134#ifdef CONFIG_CGROUP_SCHED
7135/* task_group_lock serializes the addition/removal of task groups */
7136static DEFINE_SPINLOCK(task_group_lock);
7137
7138static void free_sched_group(struct task_group *tg)
7139{
7140	free_fair_sched_group(tg);
7141	free_rt_sched_group(tg);
7142	autogroup_free(tg);
7143	kfree(tg);
7144}
7145
7146/* allocate runqueue etc for a new task group */
7147struct task_group *sched_create_group(struct task_group *parent)
7148{
7149	struct task_group *tg;
7150
7151	tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7152	if (!tg)
7153		return ERR_PTR(-ENOMEM);
7154
7155	if (!alloc_fair_sched_group(tg, parent))
7156		goto err;
7157
7158	if (!alloc_rt_sched_group(tg, parent))
7159		goto err;
7160
7161	return tg;
7162
7163err:
7164	free_sched_group(tg);
7165	return ERR_PTR(-ENOMEM);
7166}
7167
7168void sched_online_group(struct task_group *tg, struct task_group *parent)
7169{
7170	unsigned long flags;
7171
7172	spin_lock_irqsave(&task_group_lock, flags);
7173	list_add_rcu(&tg->list, &task_groups);
7174
7175	WARN_ON(!parent); /* root should already exist */
7176
7177	tg->parent = parent;
7178	INIT_LIST_HEAD(&tg->children);
7179	list_add_rcu(&tg->siblings, &parent->children);
7180	spin_unlock_irqrestore(&task_group_lock, flags);
7181}
7182
7183/* rcu callback to free various structures associated with a task group */
7184static void free_sched_group_rcu(struct rcu_head *rhp)
7185{
7186	/* now it should be safe to free those cfs_rqs */
7187	free_sched_group(container_of(rhp, struct task_group, rcu));
7188}
7189
7190/* Destroy runqueue etc associated with a task group */
7191void sched_destroy_group(struct task_group *tg)
7192{
7193	/* wait for possible concurrent references to cfs_rqs complete */
7194	call_rcu(&tg->rcu, free_sched_group_rcu);
7195}
7196
7197void sched_offline_group(struct task_group *tg)
7198{
7199	unsigned long flags;
7200	int i;
7201
7202	/* end participation in shares distribution */
7203	for_each_possible_cpu(i)
7204		unregister_fair_sched_group(tg, i);
7205
7206	spin_lock_irqsave(&task_group_lock, flags);
7207	list_del_rcu(&tg->list);
7208	list_del_rcu(&tg->siblings);
7209	spin_unlock_irqrestore(&task_group_lock, flags);
7210}
7211
7212/* change task's runqueue when it moves between groups.
7213 *	The caller of this function should have put the task in its new group
7214 *	by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7215 *	reflect its new group.
7216 */
7217void sched_move_task(struct task_struct *tsk)
7218{
7219	struct task_group *tg;
7220	int on_rq, running;
7221	unsigned long flags;
7222	struct rq *rq;
7223
7224	rq = task_rq_lock(tsk, &flags);
7225
7226	running = task_current(rq, tsk);
7227	on_rq = tsk->on_rq;
7228
7229	if (on_rq)
7230		dequeue_task(rq, tsk, 0);
7231	if (unlikely(running))
7232		tsk->sched_class->put_prev_task(rq, tsk);
7233
7234	tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7235				lockdep_is_held(&tsk->sighand->siglock)),
7236			  struct task_group, css);
7237	tg = autogroup_task_group(tsk, tg);
7238	tsk->sched_task_group = tg;
7239
7240#ifdef CONFIG_FAIR_GROUP_SCHED
7241	if (tsk->sched_class->task_move_group)
7242		tsk->sched_class->task_move_group(tsk, on_rq);
7243	else
7244#endif
7245		set_task_rq(tsk, task_cpu(tsk));
7246
7247	if (unlikely(running))
7248		tsk->sched_class->set_curr_task(rq);
7249	if (on_rq)
7250		enqueue_task(rq, tsk, 0);
7251
7252	task_rq_unlock(rq, tsk, &flags);
7253}
7254#endif /* CONFIG_CGROUP_SCHED */
7255
7256#ifdef CONFIG_RT_GROUP_SCHED
7257/*
7258 * Ensure that the real time constraints are schedulable.
7259 */
7260static DEFINE_MUTEX(rt_constraints_mutex);
7261
7262/* Must be called with tasklist_lock held */
7263static inline int tg_has_rt_tasks(struct task_group *tg)
7264{
7265	struct task_struct *g, *p;
7266
7267	do_each_thread(g, p) {
7268		if (rt_task(p) && task_rq(p)->rt.tg == tg)
7269			return 1;
7270	} while_each_thread(g, p);
7271
7272	return 0;
7273}
7274
7275struct rt_schedulable_data {
7276	struct task_group *tg;
7277	u64 rt_period;
7278	u64 rt_runtime;
7279};
7280
7281static int tg_rt_schedulable(struct task_group *tg, void *data)
7282{
7283	struct rt_schedulable_data *d = data;
7284	struct task_group *child;
7285	unsigned long total, sum = 0;
7286	u64 period, runtime;
7287
7288	period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7289	runtime = tg->rt_bandwidth.rt_runtime;
7290
7291	if (tg == d->tg) {
7292		period = d->rt_period;
7293		runtime = d->rt_runtime;
7294	}
7295
7296	/*
7297	 * Cannot have more runtime than the period.
7298	 */
7299	if (runtime > period && runtime != RUNTIME_INF)
7300		return -EINVAL;
7301
7302	/*
7303	 * Ensure we don't starve existing RT tasks.
7304	 */
7305	if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7306		return -EBUSY;
7307
7308	total = to_ratio(period, runtime);
7309
7310	/*
7311	 * Nobody can have more than the global setting allows.
7312	 */
7313	if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7314		return -EINVAL;
7315
7316	/*
7317	 * The sum of our children's runtime should not exceed our own.
7318	 */
7319	list_for_each_entry_rcu(child, &tg->children, siblings) {
7320		period = ktime_to_ns(child->rt_bandwidth.rt_period);
7321		runtime = child->rt_bandwidth.rt_runtime;
7322
7323		if (child == d->tg) {
7324			period = d->rt_period;
7325			runtime = d->rt_runtime;
7326		}
7327
7328		sum += to_ratio(period, runtime);
7329	}
7330
7331	if (sum > total)
7332		return -EINVAL;
7333
7334	return 0;
7335}
7336
7337static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7338{
7339	int ret;
7340
7341	struct rt_schedulable_data data = {
7342		.tg = tg,
7343		.rt_period = period,
7344		.rt_runtime = runtime,
7345	};
7346
7347	rcu_read_lock();
7348	ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7349	rcu_read_unlock();
7350
7351	return ret;
7352}
7353
7354static int tg_set_rt_bandwidth(struct task_group *tg,
7355		u64 rt_period, u64 rt_runtime)
7356{
7357	int i, err = 0;
7358
7359	mutex_lock(&rt_constraints_mutex);
7360	read_lock(&tasklist_lock);
7361	err = __rt_schedulable(tg, rt_period, rt_runtime);
7362	if (err)
7363		goto unlock;
7364
7365	raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7366	tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7367	tg->rt_bandwidth.rt_runtime = rt_runtime;
7368
7369	for_each_possible_cpu(i) {
7370		struct rt_rq *rt_rq = tg->rt_rq[i];
7371
7372		raw_spin_lock(&rt_rq->rt_runtime_lock);
7373		rt_rq->rt_runtime = rt_runtime;
7374		raw_spin_unlock(&rt_rq->rt_runtime_lock);
7375	}
7376	raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7377unlock:
7378	read_unlock(&tasklist_lock);
7379	mutex_unlock(&rt_constraints_mutex);
7380
7381	return err;
7382}
7383
7384static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7385{
7386	u64 rt_runtime, rt_period;
7387
7388	rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7389	rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7390	if (rt_runtime_us < 0)
7391		rt_runtime = RUNTIME_INF;
7392
7393	return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7394}
7395
7396static long sched_group_rt_runtime(struct task_group *tg)
7397{
7398	u64 rt_runtime_us;
7399
7400	if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7401		return -1;
7402
7403	rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7404	do_div(rt_runtime_us, NSEC_PER_USEC);
7405	return rt_runtime_us;
7406}
7407
7408static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7409{
7410	u64 rt_runtime, rt_period;
7411
7412	rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7413	rt_runtime = tg->rt_bandwidth.rt_runtime;
7414
7415	if (rt_period == 0)
7416		return -EINVAL;
7417
7418	return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7419}
7420
7421static long sched_group_rt_period(struct task_group *tg)
7422{
7423	u64 rt_period_us;
7424
7425	rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7426	do_div(rt_period_us, NSEC_PER_USEC);
7427	return rt_period_us;
7428}
7429#endif /* CONFIG_RT_GROUP_SCHED */
7430
7431#ifdef CONFIG_RT_GROUP_SCHED
7432static int sched_rt_global_constraints(void)
7433{
7434	int ret = 0;
7435
7436	mutex_lock(&rt_constraints_mutex);
7437	read_lock(&tasklist_lock);
7438	ret = __rt_schedulable(NULL, 0, 0);
7439	read_unlock(&tasklist_lock);
7440	mutex_unlock(&rt_constraints_mutex);
7441
7442	return ret;
7443}
7444
7445static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7446{
7447	/* Don't accept realtime tasks when there is no way for them to run */
7448	if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7449		return 0;
7450
7451	return 1;
7452}
7453
7454#else /* !CONFIG_RT_GROUP_SCHED */
7455static int sched_rt_global_constraints(void)
7456{
7457	unsigned long flags;
7458	int i, ret = 0;
7459
7460	raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7461	for_each_possible_cpu(i) {
7462		struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7463
7464		raw_spin_lock(&rt_rq->rt_runtime_lock);
7465		rt_rq->rt_runtime = global_rt_runtime();
7466		raw_spin_unlock(&rt_rq->rt_runtime_lock);
7467	}
7468	raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7469
7470	return ret;
7471}
7472#endif /* CONFIG_RT_GROUP_SCHED */
7473
7474static int sched_dl_global_constraints(void)
7475{
7476	u64 runtime = global_rt_runtime();
7477	u64 period = global_rt_period();
7478	u64 new_bw = to_ratio(period, runtime);
7479	int cpu, ret = 0;
7480	unsigned long flags;
7481
7482	/*
7483	 * Here we want to check the bandwidth not being set to some
7484	 * value smaller than the currently allocated bandwidth in
7485	 * any of the root_domains.
7486	 *
7487	 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7488	 * cycling on root_domains... Discussion on different/better
7489	 * solutions is welcome!
7490	 */
7491	for_each_possible_cpu(cpu) {
7492		struct dl_bw *dl_b = dl_bw_of(cpu);
7493
7494		raw_spin_lock_irqsave(&dl_b->lock, flags);
7495		if (new_bw < dl_b->total_bw)
7496			ret = -EBUSY;
7497		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7498
7499		if (ret)
7500			break;
7501	}
7502
7503	return ret;
7504}
7505
7506static void sched_dl_do_global(void)
7507{
7508	u64 new_bw = -1;
7509	int cpu;
7510	unsigned long flags;
7511
7512	def_dl_bandwidth.dl_period = global_rt_period();
7513	def_dl_bandwidth.dl_runtime = global_rt_runtime();
7514
7515	if (global_rt_runtime() != RUNTIME_INF)
7516		new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7517
7518	/*
7519	 * FIXME: As above...
7520	 */
7521	for_each_possible_cpu(cpu) {
7522		struct dl_bw *dl_b = dl_bw_of(cpu);
7523
7524		raw_spin_lock_irqsave(&dl_b->lock, flags);
7525		dl_b->bw = new_bw;
7526		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7527	}
7528}
7529
7530static int sched_rt_global_validate(void)
7531{
7532	if (sysctl_sched_rt_period <= 0)
7533		return -EINVAL;
7534
7535	if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7536		(sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7537		return -EINVAL;
7538
7539	return 0;
7540}
7541
7542static void sched_rt_do_global(void)
7543{
7544	def_rt_bandwidth.rt_runtime = global_rt_runtime();
7545	def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7546}
7547
7548int sched_rt_handler(struct ctl_table *table, int write,
7549		void __user *buffer, size_t *lenp,
7550		loff_t *ppos)
7551{
7552	int old_period, old_runtime;
7553	static DEFINE_MUTEX(mutex);
7554	int ret;
7555
7556	mutex_lock(&mutex);
7557	old_period = sysctl_sched_rt_period;
7558	old_runtime = sysctl_sched_rt_runtime;
7559
7560	ret = proc_dointvec(table, write, buffer, lenp, ppos);
7561
7562	if (!ret && write) {
7563		ret = sched_rt_global_validate();
7564		if (ret)
7565			goto undo;
7566
7567		ret = sched_rt_global_constraints();
7568		if (ret)
7569			goto undo;
7570
7571		ret = sched_dl_global_constraints();
7572		if (ret)
7573			goto undo;
7574
7575		sched_rt_do_global();
7576		sched_dl_do_global();
7577	}
7578	if (0) {
7579undo:
7580		sysctl_sched_rt_period = old_period;
7581		sysctl_sched_rt_runtime = old_runtime;
7582	}
7583	mutex_unlock(&mutex);
7584
7585	return ret;
7586}
7587
7588int sched_rr_handler(struct ctl_table *table, int write,
7589		void __user *buffer, size_t *lenp,
7590		loff_t *ppos)
7591{
7592	int ret;
7593	static DEFINE_MUTEX(mutex);
7594
7595	mutex_lock(&mutex);
7596	ret = proc_dointvec(table, write, buffer, lenp, ppos);
7597	/* make sure that internally we keep jiffies */
7598	/* also, writing zero resets timeslice to default */
7599	if (!ret && write) {
7600		sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7601			RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7602	}
7603	mutex_unlock(&mutex);
7604	return ret;
7605}
7606
7607#ifdef CONFIG_CGROUP_SCHED
7608
7609static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7610{
7611	return css ? container_of(css, struct task_group, css) : NULL;
7612}
7613
7614static struct cgroup_subsys_state *
7615cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7616{
7617	struct task_group *parent = css_tg(parent_css);
7618	struct task_group *tg;
7619
7620	if (!parent) {
7621		/* This is early initialization for the top cgroup */
7622		return &root_task_group.css;
7623	}
7624
7625	tg = sched_create_group(parent);
7626	if (IS_ERR(tg))
7627		return ERR_PTR(-ENOMEM);
7628
7629	return &tg->css;
7630}
7631
7632static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7633{
7634	struct task_group *tg = css_tg(css);
7635	struct task_group *parent = css_tg(css_parent(css));
7636
7637	if (parent)
7638		sched_online_group(tg, parent);
7639	return 0;
7640}
7641
7642static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7643{
7644	struct task_group *tg = css_tg(css);
7645
7646	sched_destroy_group(tg);
7647}
7648
7649static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7650{
7651	struct task_group *tg = css_tg(css);
7652
7653	sched_offline_group(tg);
7654}
7655
7656static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7657				 struct cgroup_taskset *tset)
7658{
7659	struct task_struct *task;
7660
7661	cgroup_taskset_for_each(task, tset) {
7662#ifdef CONFIG_RT_GROUP_SCHED
7663		if (!sched_rt_can_attach(css_tg(css), task))
7664			return -EINVAL;
7665#else
7666		/* We don't support RT-tasks being in separate groups */
7667		if (task->sched_class != &fair_sched_class)
7668			return -EINVAL;
7669#endif
7670	}
7671	return 0;
7672}
7673
7674static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7675			      struct cgroup_taskset *tset)
7676{
7677	struct task_struct *task;
7678
7679	cgroup_taskset_for_each(task, tset)
7680		sched_move_task(task);
7681}
7682
7683static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7684			    struct cgroup_subsys_state *old_css,
7685			    struct task_struct *task)
7686{
7687	/*
7688	 * cgroup_exit() is called in the copy_process() failure path.
7689	 * Ignore this case since the task hasn't ran yet, this avoids
7690	 * trying to poke a half freed task state from generic code.
7691	 */
7692	if (!(task->flags & PF_EXITING))
7693		return;
7694
7695	sched_move_task(task);
7696}
7697
7698#ifdef CONFIG_FAIR_GROUP_SCHED
7699static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7700				struct cftype *cftype, u64 shareval)
7701{
7702	return sched_group_set_shares(css_tg(css), scale_load(shareval));
7703}
7704
7705static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7706			       struct cftype *cft)
7707{
7708	struct task_group *tg = css_tg(css);
7709
7710	return (u64) scale_load_down(tg->shares);
7711}
7712
7713#ifdef CONFIG_CFS_BANDWIDTH
7714static DEFINE_MUTEX(cfs_constraints_mutex);
7715
7716const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7717const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7718
7719static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7720
7721static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7722{
7723	int i, ret = 0, runtime_enabled, runtime_was_enabled;
7724	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7725
7726	if (tg == &root_task_group)
7727		return -EINVAL;
7728
7729	/*
7730	 * Ensure we have at some amount of bandwidth every period.  This is
7731	 * to prevent reaching a state of large arrears when throttled via
7732	 * entity_tick() resulting in prolonged exit starvation.
7733	 */
7734	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7735		return -EINVAL;
7736
7737	/*
7738	 * Likewise, bound things on the otherside by preventing insane quota
7739	 * periods.  This also allows us to normalize in computing quota
7740	 * feasibility.
7741	 */
7742	if (period > max_cfs_quota_period)
7743		return -EINVAL;
7744
7745	mutex_lock(&cfs_constraints_mutex);
7746	ret = __cfs_schedulable(tg, period, quota);
7747	if (ret)
7748		goto out_unlock;
7749
7750	runtime_enabled = quota != RUNTIME_INF;
7751	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7752	/*
7753	 * If we need to toggle cfs_bandwidth_used, off->on must occur
7754	 * before making related changes, and on->off must occur afterwards
7755	 */
7756	if (runtime_enabled && !runtime_was_enabled)
7757		cfs_bandwidth_usage_inc();
7758	raw_spin_lock_irq(&cfs_b->lock);
7759	cfs_b->period = ns_to_ktime(period);
7760	cfs_b->quota = quota;
7761
7762	__refill_cfs_bandwidth_runtime(cfs_b);
7763	/* restart the period timer (if active) to handle new period expiry */
7764	if (runtime_enabled && cfs_b->timer_active) {
7765		/* force a reprogram */
7766		cfs_b->timer_active = 0;
7767		__start_cfs_bandwidth(cfs_b);
7768	}
7769	raw_spin_unlock_irq(&cfs_b->lock);
7770
7771	for_each_possible_cpu(i) {
7772		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7773		struct rq *rq = cfs_rq->rq;
7774
7775		raw_spin_lock_irq(&rq->lock);
7776		cfs_rq->runtime_enabled = runtime_enabled;
7777		cfs_rq->runtime_remaining = 0;
7778
7779		if (cfs_rq->throttled)
7780			unthrottle_cfs_rq(cfs_rq);
7781		raw_spin_unlock_irq(&rq->lock);
7782	}
7783	if (runtime_was_enabled && !runtime_enabled)
7784		cfs_bandwidth_usage_dec();
7785out_unlock:
7786	mutex_unlock(&cfs_constraints_mutex);
7787
7788	return ret;
7789}
7790
7791int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7792{
7793	u64 quota, period;
7794
7795	period = ktime_to_ns(tg->cfs_bandwidth.period);
7796	if (cfs_quota_us < 0)
7797		quota = RUNTIME_INF;
7798	else
7799		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7800
7801	return tg_set_cfs_bandwidth(tg, period, quota);
7802}
7803
7804long tg_get_cfs_quota(struct task_group *tg)
7805{
7806	u64 quota_us;
7807
7808	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7809		return -1;
7810
7811	quota_us = tg->cfs_bandwidth.quota;
7812	do_div(quota_us, NSEC_PER_USEC);
7813
7814	return quota_us;
7815}
7816
7817int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7818{
7819	u64 quota, period;
7820
7821	period = (u64)cfs_period_us * NSEC_PER_USEC;
7822	quota = tg->cfs_bandwidth.quota;
7823
7824	return tg_set_cfs_bandwidth(tg, period, quota);
7825}
7826
7827long tg_get_cfs_period(struct task_group *tg)
7828{
7829	u64 cfs_period_us;
7830
7831	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7832	do_div(cfs_period_us, NSEC_PER_USEC);
7833
7834	return cfs_period_us;
7835}
7836
7837static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7838				  struct cftype *cft)
7839{
7840	return tg_get_cfs_quota(css_tg(css));
7841}
7842
7843static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7844				   struct cftype *cftype, s64 cfs_quota_us)
7845{
7846	return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7847}
7848
7849static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7850				   struct cftype *cft)
7851{
7852	return tg_get_cfs_period(css_tg(css));
7853}
7854
7855static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7856				    struct cftype *cftype, u64 cfs_period_us)
7857{
7858	return tg_set_cfs_period(css_tg(css), cfs_period_us);
7859}
7860
7861struct cfs_schedulable_data {
7862	struct task_group *tg;
7863	u64 period, quota;
7864};
7865
7866/*
7867 * normalize group quota/period to be quota/max_period
7868 * note: units are usecs
7869 */
7870static u64 normalize_cfs_quota(struct task_group *tg,
7871			       struct cfs_schedulable_data *d)
7872{
7873	u64 quota, period;
7874
7875	if (tg == d->tg) {
7876		period = d->period;
7877		quota = d->quota;
7878	} else {
7879		period = tg_get_cfs_period(tg);
7880		quota = tg_get_cfs_quota(tg);
7881	}
7882
7883	/* note: these should typically be equivalent */
7884	if (quota == RUNTIME_INF || quota == -1)
7885		return RUNTIME_INF;
7886
7887	return to_ratio(period, quota);
7888}
7889
7890static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7891{
7892	struct cfs_schedulable_data *d = data;
7893	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7894	s64 quota = 0, parent_quota = -1;
7895
7896	if (!tg->parent) {
7897		quota = RUNTIME_INF;
7898	} else {
7899		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7900
7901		quota = normalize_cfs_quota(tg, d);
7902		parent_quota = parent_b->hierarchal_quota;
7903
7904		/*
7905		 * ensure max(child_quota) <= parent_quota, inherit when no
7906		 * limit is set
7907		 */
7908		if (quota == RUNTIME_INF)
7909			quota = parent_quota;
7910		else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7911			return -EINVAL;
7912	}
7913	cfs_b->hierarchal_quota = quota;
7914
7915	return 0;
7916}
7917
7918static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7919{
7920	int ret;
7921	struct cfs_schedulable_data data = {
7922		.tg = tg,
7923		.period = period,
7924		.quota = quota,
7925	};
7926
7927	if (quota != RUNTIME_INF) {
7928		do_div(data.period, NSEC_PER_USEC);
7929		do_div(data.quota, NSEC_PER_USEC);
7930	}
7931
7932	rcu_read_lock();
7933	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7934	rcu_read_unlock();
7935
7936	return ret;
7937}
7938
7939static int cpu_stats_show(struct seq_file *sf, void *v)
7940{
7941	struct task_group *tg = css_tg(seq_css(sf));
7942	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7943
7944	seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7945	seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7946	seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7947
7948	return 0;
7949}
7950#endif /* CONFIG_CFS_BANDWIDTH */
7951#endif /* CONFIG_FAIR_GROUP_SCHED */
7952
7953#ifdef CONFIG_RT_GROUP_SCHED
7954static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7955				struct cftype *cft, s64 val)
7956{
7957	return sched_group_set_rt_runtime(css_tg(css), val);
7958}
7959
7960static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7961			       struct cftype *cft)
7962{
7963	return sched_group_rt_runtime(css_tg(css));
7964}
7965
7966static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7967				    struct cftype *cftype, u64 rt_period_us)
7968{
7969	return sched_group_set_rt_period(css_tg(css), rt_period_us);
7970}
7971
7972static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7973				   struct cftype *cft)
7974{
7975	return sched_group_rt_period(css_tg(css));
7976}
7977#endif /* CONFIG_RT_GROUP_SCHED */
7978
7979static struct cftype cpu_files[] = {
7980#ifdef CONFIG_FAIR_GROUP_SCHED
7981	{
7982		.name = "shares",
7983		.read_u64 = cpu_shares_read_u64,
7984		.write_u64 = cpu_shares_write_u64,
7985	},
7986#endif
7987#ifdef CONFIG_CFS_BANDWIDTH
7988	{
7989		.name = "cfs_quota_us",
7990		.read_s64 = cpu_cfs_quota_read_s64,
7991		.write_s64 = cpu_cfs_quota_write_s64,
7992	},
7993	{
7994		.name = "cfs_period_us",
7995		.read_u64 = cpu_cfs_period_read_u64,
7996		.write_u64 = cpu_cfs_period_write_u64,
7997	},
7998	{
7999		.name = "stat",
8000		.seq_show = cpu_stats_show,
8001	},
8002#endif
8003#ifdef CONFIG_RT_GROUP_SCHED
8004	{
8005		.name = "rt_runtime_us",
8006		.read_s64 = cpu_rt_runtime_read,
8007		.write_s64 = cpu_rt_runtime_write,
8008	},
8009	{
8010		.name = "rt_period_us",
8011		.read_u64 = cpu_rt_period_read_uint,
8012		.write_u64 = cpu_rt_period_write_uint,
8013	},
8014#endif
8015	{ }	/* terminate */
8016};
8017
8018struct cgroup_subsys cpu_cgrp_subsys = {
8019	.css_alloc	= cpu_cgroup_css_alloc,
8020	.css_free	= cpu_cgroup_css_free,
8021	.css_online	= cpu_cgroup_css_online,
8022	.css_offline	= cpu_cgroup_css_offline,
8023	.can_attach	= cpu_cgroup_can_attach,
8024	.attach		= cpu_cgroup_attach,
8025	.exit		= cpu_cgroup_exit,
8026	.base_cftypes	= cpu_files,
8027	.early_init	= 1,
8028};
8029
8030#endif	/* CONFIG_CGROUP_SCHED */
8031
8032void dump_cpu_task(int cpu)
8033{
8034	pr_info("Task dump for CPU %d:\n", cpu);
8035	sched_show_task(cpu_curr(cpu));
8036}
8037