core.c revision d35be8bab9b0ce44bed4b9453f86ebf64062721e
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
76#include <asm/switch_to.h>
77#include <asm/tlb.h>
78#include <asm/irq_regs.h>
79#include <asm/mutex.h>
80#ifdef CONFIG_PARAVIRT
81#include <asm/paravirt.h>
82#endif
83
84#include "sched.h"
85#include "../workqueue_sched.h"
86#include "../smpboot.h"
87
88#define CREATE_TRACE_POINTS
89#include <trace/events/sched.h>
90
91void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
92{
93	unsigned long delta;
94	ktime_t soft, hard, now;
95
96	for (;;) {
97		if (hrtimer_active(period_timer))
98			break;
99
100		now = hrtimer_cb_get_time(period_timer);
101		hrtimer_forward(period_timer, now, period);
102
103		soft = hrtimer_get_softexpires(period_timer);
104		hard = hrtimer_get_expires(period_timer);
105		delta = ktime_to_ns(ktime_sub(hard, soft));
106		__hrtimer_start_range_ns(period_timer, soft, delta,
107					 HRTIMER_MODE_ABS_PINNED, 0);
108	}
109}
110
111DEFINE_MUTEX(sched_domains_mutex);
112DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
113
114static void update_rq_clock_task(struct rq *rq, s64 delta);
115
116void update_rq_clock(struct rq *rq)
117{
118	s64 delta;
119
120	if (rq->skip_clock_update > 0)
121		return;
122
123	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
124	rq->clock += delta;
125	update_rq_clock_task(rq, delta);
126}
127
128/*
129 * Debugging: various feature bits
130 */
131
132#define SCHED_FEAT(name, enabled)	\
133	(1UL << __SCHED_FEAT_##name) * enabled |
134
135const_debug unsigned int sysctl_sched_features =
136#include "features.h"
137	0;
138
139#undef SCHED_FEAT
140
141#ifdef CONFIG_SCHED_DEBUG
142#define SCHED_FEAT(name, enabled)	\
143	#name ,
144
145static const char * const sched_feat_names[] = {
146#include "features.h"
147};
148
149#undef SCHED_FEAT
150
151static int sched_feat_show(struct seq_file *m, void *v)
152{
153	int i;
154
155	for (i = 0; i < __SCHED_FEAT_NR; i++) {
156		if (!(sysctl_sched_features & (1UL << i)))
157			seq_puts(m, "NO_");
158		seq_printf(m, "%s ", sched_feat_names[i]);
159	}
160	seq_puts(m, "\n");
161
162	return 0;
163}
164
165#ifdef HAVE_JUMP_LABEL
166
167#define jump_label_key__true  STATIC_KEY_INIT_TRUE
168#define jump_label_key__false STATIC_KEY_INIT_FALSE
169
170#define SCHED_FEAT(name, enabled)	\
171	jump_label_key__##enabled ,
172
173struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
174#include "features.h"
175};
176
177#undef SCHED_FEAT
178
179static void sched_feat_disable(int i)
180{
181	if (static_key_enabled(&sched_feat_keys[i]))
182		static_key_slow_dec(&sched_feat_keys[i]);
183}
184
185static void sched_feat_enable(int i)
186{
187	if (!static_key_enabled(&sched_feat_keys[i]))
188		static_key_slow_inc(&sched_feat_keys[i]);
189}
190#else
191static void sched_feat_disable(int i) { };
192static void sched_feat_enable(int i) { };
193#endif /* HAVE_JUMP_LABEL */
194
195static ssize_t
196sched_feat_write(struct file *filp, const char __user *ubuf,
197		size_t cnt, loff_t *ppos)
198{
199	char buf[64];
200	char *cmp;
201	int neg = 0;
202	int i;
203
204	if (cnt > 63)
205		cnt = 63;
206
207	if (copy_from_user(&buf, ubuf, cnt))
208		return -EFAULT;
209
210	buf[cnt] = 0;
211	cmp = strstrip(buf);
212
213	if (strncmp(cmp, "NO_", 3) == 0) {
214		neg = 1;
215		cmp += 3;
216	}
217
218	for (i = 0; i < __SCHED_FEAT_NR; i++) {
219		if (strcmp(cmp, sched_feat_names[i]) == 0) {
220			if (neg) {
221				sysctl_sched_features &= ~(1UL << i);
222				sched_feat_disable(i);
223			} else {
224				sysctl_sched_features |= (1UL << i);
225				sched_feat_enable(i);
226			}
227			break;
228		}
229	}
230
231	if (i == __SCHED_FEAT_NR)
232		return -EINVAL;
233
234	*ppos += cnt;
235
236	return cnt;
237}
238
239static int sched_feat_open(struct inode *inode, struct file *filp)
240{
241	return single_open(filp, sched_feat_show, NULL);
242}
243
244static const struct file_operations sched_feat_fops = {
245	.open		= sched_feat_open,
246	.write		= sched_feat_write,
247	.read		= seq_read,
248	.llseek		= seq_lseek,
249	.release	= single_release,
250};
251
252static __init int sched_init_debug(void)
253{
254	debugfs_create_file("sched_features", 0644, NULL, NULL,
255			&sched_feat_fops);
256
257	return 0;
258}
259late_initcall(sched_init_debug);
260#endif /* CONFIG_SCHED_DEBUG */
261
262/*
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
265 */
266const_debug unsigned int sysctl_sched_nr_migrate = 32;
267
268/*
269 * period over which we average the RT time consumption, measured
270 * in ms.
271 *
272 * default: 1s
273 */
274const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
275
276/*
277 * period over which we measure -rt task cpu usage in us.
278 * default: 1s
279 */
280unsigned int sysctl_sched_rt_period = 1000000;
281
282__read_mostly int scheduler_running;
283
284/*
285 * part of the period that we allow rt tasks to run in us.
286 * default: 0.95s
287 */
288int sysctl_sched_rt_runtime = 950000;
289
290
291
292/*
293 * __task_rq_lock - lock the rq @p resides on.
294 */
295static inline struct rq *__task_rq_lock(struct task_struct *p)
296	__acquires(rq->lock)
297{
298	struct rq *rq;
299
300	lockdep_assert_held(&p->pi_lock);
301
302	for (;;) {
303		rq = task_rq(p);
304		raw_spin_lock(&rq->lock);
305		if (likely(rq == task_rq(p)))
306			return rq;
307		raw_spin_unlock(&rq->lock);
308	}
309}
310
311/*
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
313 */
314static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
315	__acquires(p->pi_lock)
316	__acquires(rq->lock)
317{
318	struct rq *rq;
319
320	for (;;) {
321		raw_spin_lock_irqsave(&p->pi_lock, *flags);
322		rq = task_rq(p);
323		raw_spin_lock(&rq->lock);
324		if (likely(rq == task_rq(p)))
325			return rq;
326		raw_spin_unlock(&rq->lock);
327		raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
328	}
329}
330
331static void __task_rq_unlock(struct rq *rq)
332	__releases(rq->lock)
333{
334	raw_spin_unlock(&rq->lock);
335}
336
337static inline void
338task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
339	__releases(rq->lock)
340	__releases(p->pi_lock)
341{
342	raw_spin_unlock(&rq->lock);
343	raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
344}
345
346/*
347 * this_rq_lock - lock this runqueue and disable interrupts.
348 */
349static struct rq *this_rq_lock(void)
350	__acquires(rq->lock)
351{
352	struct rq *rq;
353
354	local_irq_disable();
355	rq = this_rq();
356	raw_spin_lock(&rq->lock);
357
358	return rq;
359}
360
361#ifdef CONFIG_SCHED_HRTICK
362/*
363 * Use HR-timers to deliver accurate preemption points.
364 *
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
367 * reschedule event.
368 *
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
370 * rq->lock.
371 */
372
373static void hrtick_clear(struct rq *rq)
374{
375	if (hrtimer_active(&rq->hrtick_timer))
376		hrtimer_cancel(&rq->hrtick_timer);
377}
378
379/*
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
382 */
383static enum hrtimer_restart hrtick(struct hrtimer *timer)
384{
385	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
386
387	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
388
389	raw_spin_lock(&rq->lock);
390	update_rq_clock(rq);
391	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392	raw_spin_unlock(&rq->lock);
393
394	return HRTIMER_NORESTART;
395}
396
397#ifdef CONFIG_SMP
398/*
399 * called from hardirq (IPI) context
400 */
401static void __hrtick_start(void *arg)
402{
403	struct rq *rq = arg;
404
405	raw_spin_lock(&rq->lock);
406	hrtimer_restart(&rq->hrtick_timer);
407	rq->hrtick_csd_pending = 0;
408	raw_spin_unlock(&rq->lock);
409}
410
411/*
412 * Called to set the hrtick timer state.
413 *
414 * called with rq->lock held and irqs disabled
415 */
416void hrtick_start(struct rq *rq, u64 delay)
417{
418	struct hrtimer *timer = &rq->hrtick_timer;
419	ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
420
421	hrtimer_set_expires(timer, time);
422
423	if (rq == this_rq()) {
424		hrtimer_restart(timer);
425	} else if (!rq->hrtick_csd_pending) {
426		__smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
427		rq->hrtick_csd_pending = 1;
428	}
429}
430
431static int
432hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
433{
434	int cpu = (int)(long)hcpu;
435
436	switch (action) {
437	case CPU_UP_CANCELED:
438	case CPU_UP_CANCELED_FROZEN:
439	case CPU_DOWN_PREPARE:
440	case CPU_DOWN_PREPARE_FROZEN:
441	case CPU_DEAD:
442	case CPU_DEAD_FROZEN:
443		hrtick_clear(cpu_rq(cpu));
444		return NOTIFY_OK;
445	}
446
447	return NOTIFY_DONE;
448}
449
450static __init void init_hrtick(void)
451{
452	hotcpu_notifier(hotplug_hrtick, 0);
453}
454#else
455/*
456 * Called to set the hrtick timer state.
457 *
458 * called with rq->lock held and irqs disabled
459 */
460void hrtick_start(struct rq *rq, u64 delay)
461{
462	__hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
463			HRTIMER_MODE_REL_PINNED, 0);
464}
465
466static inline void init_hrtick(void)
467{
468}
469#endif /* CONFIG_SMP */
470
471static void init_rq_hrtick(struct rq *rq)
472{
473#ifdef CONFIG_SMP
474	rq->hrtick_csd_pending = 0;
475
476	rq->hrtick_csd.flags = 0;
477	rq->hrtick_csd.func = __hrtick_start;
478	rq->hrtick_csd.info = rq;
479#endif
480
481	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
482	rq->hrtick_timer.function = hrtick;
483}
484#else	/* CONFIG_SCHED_HRTICK */
485static inline void hrtick_clear(struct rq *rq)
486{
487}
488
489static inline void init_rq_hrtick(struct rq *rq)
490{
491}
492
493static inline void init_hrtick(void)
494{
495}
496#endif	/* CONFIG_SCHED_HRTICK */
497
498/*
499 * resched_task - mark a task 'to be rescheduled now'.
500 *
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
503 * the target CPU.
504 */
505#ifdef CONFIG_SMP
506
507#ifndef tsk_is_polling
508#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
509#endif
510
511void resched_task(struct task_struct *p)
512{
513	int cpu;
514
515	assert_raw_spin_locked(&task_rq(p)->lock);
516
517	if (test_tsk_need_resched(p))
518		return;
519
520	set_tsk_need_resched(p);
521
522	cpu = task_cpu(p);
523	if (cpu == smp_processor_id())
524		return;
525
526	/* NEED_RESCHED must be visible before we test polling */
527	smp_mb();
528	if (!tsk_is_polling(p))
529		smp_send_reschedule(cpu);
530}
531
532void resched_cpu(int cpu)
533{
534	struct rq *rq = cpu_rq(cpu);
535	unsigned long flags;
536
537	if (!raw_spin_trylock_irqsave(&rq->lock, flags))
538		return;
539	resched_task(cpu_curr(cpu));
540	raw_spin_unlock_irqrestore(&rq->lock, flags);
541}
542
543#ifdef CONFIG_NO_HZ
544/*
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu.  This is good for power-savings.
547 *
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
551 */
552int get_nohz_timer_target(void)
553{
554	int cpu = smp_processor_id();
555	int i;
556	struct sched_domain *sd;
557
558	rcu_read_lock();
559	for_each_domain(cpu, sd) {
560		for_each_cpu(i, sched_domain_span(sd)) {
561			if (!idle_cpu(i)) {
562				cpu = i;
563				goto unlock;
564			}
565		}
566	}
567unlock:
568	rcu_read_unlock();
569	return cpu;
570}
571/*
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
580 */
581void wake_up_idle_cpu(int cpu)
582{
583	struct rq *rq = cpu_rq(cpu);
584
585	if (cpu == smp_processor_id())
586		return;
587
588	/*
589	 * This is safe, as this function is called with the timer
590	 * wheel base lock of (cpu) held. When the CPU is on the way
591	 * to idle and has not yet set rq->curr to idle then it will
592	 * be serialized on the timer wheel base lock and take the new
593	 * timer into account automatically.
594	 */
595	if (rq->curr != rq->idle)
596		return;
597
598	/*
599	 * We can set TIF_RESCHED on the idle task of the other CPU
600	 * lockless. The worst case is that the other CPU runs the
601	 * idle task through an additional NOOP schedule()
602	 */
603	set_tsk_need_resched(rq->idle);
604
605	/* NEED_RESCHED must be visible before we test polling */
606	smp_mb();
607	if (!tsk_is_polling(rq->idle))
608		smp_send_reschedule(cpu);
609}
610
611static inline bool got_nohz_idle_kick(void)
612{
613	int cpu = smp_processor_id();
614	return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
615}
616
617#else /* CONFIG_NO_HZ */
618
619static inline bool got_nohz_idle_kick(void)
620{
621	return false;
622}
623
624#endif /* CONFIG_NO_HZ */
625
626void sched_avg_update(struct rq *rq)
627{
628	s64 period = sched_avg_period();
629
630	while ((s64)(rq->clock - rq->age_stamp) > period) {
631		/*
632		 * Inline assembly required to prevent the compiler
633		 * optimising this loop into a divmod call.
634		 * See __iter_div_u64_rem() for another example of this.
635		 */
636		asm("" : "+rm" (rq->age_stamp));
637		rq->age_stamp += period;
638		rq->rt_avg /= 2;
639	}
640}
641
642#else /* !CONFIG_SMP */
643void resched_task(struct task_struct *p)
644{
645	assert_raw_spin_locked(&task_rq(p)->lock);
646	set_tsk_need_resched(p);
647}
648#endif /* CONFIG_SMP */
649
650#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
652/*
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
655 *
656 * Caller must hold rcu_lock or sufficient equivalent.
657 */
658int walk_tg_tree_from(struct task_group *from,
659			     tg_visitor down, tg_visitor up, void *data)
660{
661	struct task_group *parent, *child;
662	int ret;
663
664	parent = from;
665
666down:
667	ret = (*down)(parent, data);
668	if (ret)
669		goto out;
670	list_for_each_entry_rcu(child, &parent->children, siblings) {
671		parent = child;
672		goto down;
673
674up:
675		continue;
676	}
677	ret = (*up)(parent, data);
678	if (ret || parent == from)
679		goto out;
680
681	child = parent;
682	parent = parent->parent;
683	if (parent)
684		goto up;
685out:
686	return ret;
687}
688
689int tg_nop(struct task_group *tg, void *data)
690{
691	return 0;
692}
693#endif
694
695static void set_load_weight(struct task_struct *p)
696{
697	int prio = p->static_prio - MAX_RT_PRIO;
698	struct load_weight *load = &p->se.load;
699
700	/*
701	 * SCHED_IDLE tasks get minimal weight:
702	 */
703	if (p->policy == SCHED_IDLE) {
704		load->weight = scale_load(WEIGHT_IDLEPRIO);
705		load->inv_weight = WMULT_IDLEPRIO;
706		return;
707	}
708
709	load->weight = scale_load(prio_to_weight[prio]);
710	load->inv_weight = prio_to_wmult[prio];
711}
712
713static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
714{
715	update_rq_clock(rq);
716	sched_info_queued(p);
717	p->sched_class->enqueue_task(rq, p, flags);
718}
719
720static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
721{
722	update_rq_clock(rq);
723	sched_info_dequeued(p);
724	p->sched_class->dequeue_task(rq, p, flags);
725}
726
727void activate_task(struct rq *rq, struct task_struct *p, int flags)
728{
729	if (task_contributes_to_load(p))
730		rq->nr_uninterruptible--;
731
732	enqueue_task(rq, p, flags);
733}
734
735void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
736{
737	if (task_contributes_to_load(p))
738		rq->nr_uninterruptible++;
739
740	dequeue_task(rq, p, flags);
741}
742
743#ifdef CONFIG_IRQ_TIME_ACCOUNTING
744
745/*
746 * There are no locks covering percpu hardirq/softirq time.
747 * They are only modified in account_system_vtime, on corresponding CPU
748 * with interrupts disabled. So, writes are safe.
749 * They are read and saved off onto struct rq in update_rq_clock().
750 * This may result in other CPU reading this CPU's irq time and can
751 * race with irq/account_system_vtime on this CPU. We would either get old
752 * or new value with a side effect of accounting a slice of irq time to wrong
753 * task when irq is in progress while we read rq->clock. That is a worthy
754 * compromise in place of having locks on each irq in account_system_time.
755 */
756static DEFINE_PER_CPU(u64, cpu_hardirq_time);
757static DEFINE_PER_CPU(u64, cpu_softirq_time);
758
759static DEFINE_PER_CPU(u64, irq_start_time);
760static int sched_clock_irqtime;
761
762void enable_sched_clock_irqtime(void)
763{
764	sched_clock_irqtime = 1;
765}
766
767void disable_sched_clock_irqtime(void)
768{
769	sched_clock_irqtime = 0;
770}
771
772#ifndef CONFIG_64BIT
773static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
774
775static inline void irq_time_write_begin(void)
776{
777	__this_cpu_inc(irq_time_seq.sequence);
778	smp_wmb();
779}
780
781static inline void irq_time_write_end(void)
782{
783	smp_wmb();
784	__this_cpu_inc(irq_time_seq.sequence);
785}
786
787static inline u64 irq_time_read(int cpu)
788{
789	u64 irq_time;
790	unsigned seq;
791
792	do {
793		seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
794		irq_time = per_cpu(cpu_softirq_time, cpu) +
795			   per_cpu(cpu_hardirq_time, cpu);
796	} while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
797
798	return irq_time;
799}
800#else /* CONFIG_64BIT */
801static inline void irq_time_write_begin(void)
802{
803}
804
805static inline void irq_time_write_end(void)
806{
807}
808
809static inline u64 irq_time_read(int cpu)
810{
811	return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
812}
813#endif /* CONFIG_64BIT */
814
815/*
816 * Called before incrementing preempt_count on {soft,}irq_enter
817 * and before decrementing preempt_count on {soft,}irq_exit.
818 */
819void account_system_vtime(struct task_struct *curr)
820{
821	unsigned long flags;
822	s64 delta;
823	int cpu;
824
825	if (!sched_clock_irqtime)
826		return;
827
828	local_irq_save(flags);
829
830	cpu = smp_processor_id();
831	delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
832	__this_cpu_add(irq_start_time, delta);
833
834	irq_time_write_begin();
835	/*
836	 * We do not account for softirq time from ksoftirqd here.
837	 * We want to continue accounting softirq time to ksoftirqd thread
838	 * in that case, so as not to confuse scheduler with a special task
839	 * that do not consume any time, but still wants to run.
840	 */
841	if (hardirq_count())
842		__this_cpu_add(cpu_hardirq_time, delta);
843	else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
844		__this_cpu_add(cpu_softirq_time, delta);
845
846	irq_time_write_end();
847	local_irq_restore(flags);
848}
849EXPORT_SYMBOL_GPL(account_system_vtime);
850
851#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
852
853#ifdef CONFIG_PARAVIRT
854static inline u64 steal_ticks(u64 steal)
855{
856	if (unlikely(steal > NSEC_PER_SEC))
857		return div_u64(steal, TICK_NSEC);
858
859	return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
860}
861#endif
862
863static void update_rq_clock_task(struct rq *rq, s64 delta)
864{
865/*
866 * In theory, the compile should just see 0 here, and optimize out the call
867 * to sched_rt_avg_update. But I don't trust it...
868 */
869#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870	s64 steal = 0, irq_delta = 0;
871#endif
872#ifdef CONFIG_IRQ_TIME_ACCOUNTING
873	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
874
875	/*
876	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877	 * this case when a previous update_rq_clock() happened inside a
878	 * {soft,}irq region.
879	 *
880	 * When this happens, we stop ->clock_task and only update the
881	 * prev_irq_time stamp to account for the part that fit, so that a next
882	 * update will consume the rest. This ensures ->clock_task is
883	 * monotonic.
884	 *
885	 * It does however cause some slight miss-attribution of {soft,}irq
886	 * time, a more accurate solution would be to update the irq_time using
887	 * the current rq->clock timestamp, except that would require using
888	 * atomic ops.
889	 */
890	if (irq_delta > delta)
891		irq_delta = delta;
892
893	rq->prev_irq_time += irq_delta;
894	delta -= irq_delta;
895#endif
896#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897	if (static_key_false((&paravirt_steal_rq_enabled))) {
898		u64 st;
899
900		steal = paravirt_steal_clock(cpu_of(rq));
901		steal -= rq->prev_steal_time_rq;
902
903		if (unlikely(steal > delta))
904			steal = delta;
905
906		st = steal_ticks(steal);
907		steal = st * TICK_NSEC;
908
909		rq->prev_steal_time_rq += steal;
910
911		delta -= steal;
912	}
913#endif
914
915	rq->clock_task += delta;
916
917#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
918	if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
919		sched_rt_avg_update(rq, irq_delta + steal);
920#endif
921}
922
923#ifdef CONFIG_IRQ_TIME_ACCOUNTING
924static int irqtime_account_hi_update(void)
925{
926	u64 *cpustat = kcpustat_this_cpu->cpustat;
927	unsigned long flags;
928	u64 latest_ns;
929	int ret = 0;
930
931	local_irq_save(flags);
932	latest_ns = this_cpu_read(cpu_hardirq_time);
933	if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
934		ret = 1;
935	local_irq_restore(flags);
936	return ret;
937}
938
939static int irqtime_account_si_update(void)
940{
941	u64 *cpustat = kcpustat_this_cpu->cpustat;
942	unsigned long flags;
943	u64 latest_ns;
944	int ret = 0;
945
946	local_irq_save(flags);
947	latest_ns = this_cpu_read(cpu_softirq_time);
948	if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
949		ret = 1;
950	local_irq_restore(flags);
951	return ret;
952}
953
954#else /* CONFIG_IRQ_TIME_ACCOUNTING */
955
956#define sched_clock_irqtime	(0)
957
958#endif
959
960void sched_set_stop_task(int cpu, struct task_struct *stop)
961{
962	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
963	struct task_struct *old_stop = cpu_rq(cpu)->stop;
964
965	if (stop) {
966		/*
967		 * Make it appear like a SCHED_FIFO task, its something
968		 * userspace knows about and won't get confused about.
969		 *
970		 * Also, it will make PI more or less work without too
971		 * much confusion -- but then, stop work should not
972		 * rely on PI working anyway.
973		 */
974		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
975
976		stop->sched_class = &stop_sched_class;
977	}
978
979	cpu_rq(cpu)->stop = stop;
980
981	if (old_stop) {
982		/*
983		 * Reset it back to a normal scheduling class so that
984		 * it can die in pieces.
985		 */
986		old_stop->sched_class = &rt_sched_class;
987	}
988}
989
990/*
991 * __normal_prio - return the priority that is based on the static prio
992 */
993static inline int __normal_prio(struct task_struct *p)
994{
995	return p->static_prio;
996}
997
998/*
999 * Calculate the expected normal priority: i.e. priority
1000 * without taking RT-inheritance into account. Might be
1001 * boosted by interactivity modifiers. Changes upon fork,
1002 * setprio syscalls, and whenever the interactivity
1003 * estimator recalculates.
1004 */
1005static inline int normal_prio(struct task_struct *p)
1006{
1007	int prio;
1008
1009	if (task_has_rt_policy(p))
1010		prio = MAX_RT_PRIO-1 - p->rt_priority;
1011	else
1012		prio = __normal_prio(p);
1013	return prio;
1014}
1015
1016/*
1017 * Calculate the current priority, i.e. the priority
1018 * taken into account by the scheduler. This value might
1019 * be boosted by RT tasks, or might be boosted by
1020 * interactivity modifiers. Will be RT if the task got
1021 * RT-boosted. If not then it returns p->normal_prio.
1022 */
1023static int effective_prio(struct task_struct *p)
1024{
1025	p->normal_prio = normal_prio(p);
1026	/*
1027	 * If we are RT tasks or we were boosted to RT priority,
1028	 * keep the priority unchanged. Otherwise, update priority
1029	 * to the normal priority:
1030	 */
1031	if (!rt_prio(p->prio))
1032		return p->normal_prio;
1033	return p->prio;
1034}
1035
1036/**
1037 * task_curr - is this task currently executing on a CPU?
1038 * @p: the task in question.
1039 */
1040inline int task_curr(const struct task_struct *p)
1041{
1042	return cpu_curr(task_cpu(p)) == p;
1043}
1044
1045static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1046				       const struct sched_class *prev_class,
1047				       int oldprio)
1048{
1049	if (prev_class != p->sched_class) {
1050		if (prev_class->switched_from)
1051			prev_class->switched_from(rq, p);
1052		p->sched_class->switched_to(rq, p);
1053	} else if (oldprio != p->prio)
1054		p->sched_class->prio_changed(rq, p, oldprio);
1055}
1056
1057void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1058{
1059	const struct sched_class *class;
1060
1061	if (p->sched_class == rq->curr->sched_class) {
1062		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1063	} else {
1064		for_each_class(class) {
1065			if (class == rq->curr->sched_class)
1066				break;
1067			if (class == p->sched_class) {
1068				resched_task(rq->curr);
1069				break;
1070			}
1071		}
1072	}
1073
1074	/*
1075	 * A queue event has occurred, and we're going to schedule.  In
1076	 * this case, we can save a useless back to back clock update.
1077	 */
1078	if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1079		rq->skip_clock_update = 1;
1080}
1081
1082#ifdef CONFIG_SMP
1083void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1084{
1085#ifdef CONFIG_SCHED_DEBUG
1086	/*
1087	 * We should never call set_task_cpu() on a blocked task,
1088	 * ttwu() will sort out the placement.
1089	 */
1090	WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1091			!(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1092
1093#ifdef CONFIG_LOCKDEP
1094	/*
1095	 * The caller should hold either p->pi_lock or rq->lock, when changing
1096	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1097	 *
1098	 * sched_move_task() holds both and thus holding either pins the cgroup,
1099	 * see set_task_rq().
1100	 *
1101	 * Furthermore, all task_rq users should acquire both locks, see
1102	 * task_rq_lock().
1103	 */
1104	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1105				      lockdep_is_held(&task_rq(p)->lock)));
1106#endif
1107#endif
1108
1109	trace_sched_migrate_task(p, new_cpu);
1110
1111	if (task_cpu(p) != new_cpu) {
1112		p->se.nr_migrations++;
1113		perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1114	}
1115
1116	__set_task_cpu(p, new_cpu);
1117}
1118
1119struct migration_arg {
1120	struct task_struct *task;
1121	int dest_cpu;
1122};
1123
1124static int migration_cpu_stop(void *data);
1125
1126/*
1127 * wait_task_inactive - wait for a thread to unschedule.
1128 *
1129 * If @match_state is nonzero, it's the @p->state value just checked and
1130 * not expected to change.  If it changes, i.e. @p might have woken up,
1131 * then return zero.  When we succeed in waiting for @p to be off its CPU,
1132 * we return a positive number (its total switch count).  If a second call
1133 * a short while later returns the same number, the caller can be sure that
1134 * @p has remained unscheduled the whole time.
1135 *
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1141 */
1142unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1143{
1144	unsigned long flags;
1145	int running, on_rq;
1146	unsigned long ncsw;
1147	struct rq *rq;
1148
1149	for (;;) {
1150		/*
1151		 * We do the initial early heuristics without holding
1152		 * any task-queue locks at all. We'll only try to get
1153		 * the runqueue lock when things look like they will
1154		 * work out!
1155		 */
1156		rq = task_rq(p);
1157
1158		/*
1159		 * If the task is actively running on another CPU
1160		 * still, just relax and busy-wait without holding
1161		 * any locks.
1162		 *
1163		 * NOTE! Since we don't hold any locks, it's not
1164		 * even sure that "rq" stays as the right runqueue!
1165		 * But we don't care, since "task_running()" will
1166		 * return false if the runqueue has changed and p
1167		 * is actually now running somewhere else!
1168		 */
1169		while (task_running(rq, p)) {
1170			if (match_state && unlikely(p->state != match_state))
1171				return 0;
1172			cpu_relax();
1173		}
1174
1175		/*
1176		 * Ok, time to look more closely! We need the rq
1177		 * lock now, to be *sure*. If we're wrong, we'll
1178		 * just go back and repeat.
1179		 */
1180		rq = task_rq_lock(p, &flags);
1181		trace_sched_wait_task(p);
1182		running = task_running(rq, p);
1183		on_rq = p->on_rq;
1184		ncsw = 0;
1185		if (!match_state || p->state == match_state)
1186			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1187		task_rq_unlock(rq, p, &flags);
1188
1189		/*
1190		 * If it changed from the expected state, bail out now.
1191		 */
1192		if (unlikely(!ncsw))
1193			break;
1194
1195		/*
1196		 * Was it really running after all now that we
1197		 * checked with the proper locks actually held?
1198		 *
1199		 * Oops. Go back and try again..
1200		 */
1201		if (unlikely(running)) {
1202			cpu_relax();
1203			continue;
1204		}
1205
1206		/*
1207		 * It's not enough that it's not actively running,
1208		 * it must be off the runqueue _entirely_, and not
1209		 * preempted!
1210		 *
1211		 * So if it was still runnable (but just not actively
1212		 * running right now), it's preempted, and we should
1213		 * yield - it could be a while.
1214		 */
1215		if (unlikely(on_rq)) {
1216			ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1217
1218			set_current_state(TASK_UNINTERRUPTIBLE);
1219			schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1220			continue;
1221		}
1222
1223		/*
1224		 * Ahh, all good. It wasn't running, and it wasn't
1225		 * runnable, which means that it will never become
1226		 * running in the future either. We're all done!
1227		 */
1228		break;
1229	}
1230
1231	return ncsw;
1232}
1233
1234/***
1235 * kick_process - kick a running thread to enter/exit the kernel
1236 * @p: the to-be-kicked thread
1237 *
1238 * Cause a process which is running on another CPU to enter
1239 * kernel-mode, without any delay. (to get signals handled.)
1240 *
1241 * NOTE: this function doesn't have to take the runqueue lock,
1242 * because all it wants to ensure is that the remote task enters
1243 * the kernel. If the IPI races and the task has been migrated
1244 * to another CPU then no harm is done and the purpose has been
1245 * achieved as well.
1246 */
1247void kick_process(struct task_struct *p)
1248{
1249	int cpu;
1250
1251	preempt_disable();
1252	cpu = task_cpu(p);
1253	if ((cpu != smp_processor_id()) && task_curr(p))
1254		smp_send_reschedule(cpu);
1255	preempt_enable();
1256}
1257EXPORT_SYMBOL_GPL(kick_process);
1258#endif /* CONFIG_SMP */
1259
1260#ifdef CONFIG_SMP
1261/*
1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1263 */
1264static int select_fallback_rq(int cpu, struct task_struct *p)
1265{
1266	const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1267	enum { cpuset, possible, fail } state = cpuset;
1268	int dest_cpu;
1269
1270	/* Look for allowed, online CPU in same node. */
1271	for_each_cpu(dest_cpu, nodemask) {
1272		if (!cpu_online(dest_cpu))
1273			continue;
1274		if (!cpu_active(dest_cpu))
1275			continue;
1276		if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1277			return dest_cpu;
1278	}
1279
1280	for (;;) {
1281		/* Any allowed, online CPU? */
1282		for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1283			if (!cpu_online(dest_cpu))
1284				continue;
1285			if (!cpu_active(dest_cpu))
1286				continue;
1287			goto out;
1288		}
1289
1290		switch (state) {
1291		case cpuset:
1292			/* No more Mr. Nice Guy. */
1293			cpuset_cpus_allowed_fallback(p);
1294			state = possible;
1295			break;
1296
1297		case possible:
1298			do_set_cpus_allowed(p, cpu_possible_mask);
1299			state = fail;
1300			break;
1301
1302		case fail:
1303			BUG();
1304			break;
1305		}
1306	}
1307
1308out:
1309	if (state != cpuset) {
1310		/*
1311		 * Don't tell them about moving exiting tasks or
1312		 * kernel threads (both mm NULL), since they never
1313		 * leave kernel.
1314		 */
1315		if (p->mm && printk_ratelimit()) {
1316			printk_sched("process %d (%s) no longer affine to cpu%d\n",
1317					task_pid_nr(p), p->comm, cpu);
1318		}
1319	}
1320
1321	return dest_cpu;
1322}
1323
1324/*
1325 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1326 */
1327static inline
1328int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1329{
1330	int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1331
1332	/*
1333	 * In order not to call set_task_cpu() on a blocking task we need
1334	 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1335	 * cpu.
1336	 *
1337	 * Since this is common to all placement strategies, this lives here.
1338	 *
1339	 * [ this allows ->select_task() to simply return task_cpu(p) and
1340	 *   not worry about this generic constraint ]
1341	 */
1342	if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1343		     !cpu_online(cpu)))
1344		cpu = select_fallback_rq(task_cpu(p), p);
1345
1346	return cpu;
1347}
1348
1349static void update_avg(u64 *avg, u64 sample)
1350{
1351	s64 diff = sample - *avg;
1352	*avg += diff >> 3;
1353}
1354#endif
1355
1356static void
1357ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1358{
1359#ifdef CONFIG_SCHEDSTATS
1360	struct rq *rq = this_rq();
1361
1362#ifdef CONFIG_SMP
1363	int this_cpu = smp_processor_id();
1364
1365	if (cpu == this_cpu) {
1366		schedstat_inc(rq, ttwu_local);
1367		schedstat_inc(p, se.statistics.nr_wakeups_local);
1368	} else {
1369		struct sched_domain *sd;
1370
1371		schedstat_inc(p, se.statistics.nr_wakeups_remote);
1372		rcu_read_lock();
1373		for_each_domain(this_cpu, sd) {
1374			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1375				schedstat_inc(sd, ttwu_wake_remote);
1376				break;
1377			}
1378		}
1379		rcu_read_unlock();
1380	}
1381
1382	if (wake_flags & WF_MIGRATED)
1383		schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1384
1385#endif /* CONFIG_SMP */
1386
1387	schedstat_inc(rq, ttwu_count);
1388	schedstat_inc(p, se.statistics.nr_wakeups);
1389
1390	if (wake_flags & WF_SYNC)
1391		schedstat_inc(p, se.statistics.nr_wakeups_sync);
1392
1393#endif /* CONFIG_SCHEDSTATS */
1394}
1395
1396static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1397{
1398	activate_task(rq, p, en_flags);
1399	p->on_rq = 1;
1400
1401	/* if a worker is waking up, notify workqueue */
1402	if (p->flags & PF_WQ_WORKER)
1403		wq_worker_waking_up(p, cpu_of(rq));
1404}
1405
1406/*
1407 * Mark the task runnable and perform wakeup-preemption.
1408 */
1409static void
1410ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1411{
1412	trace_sched_wakeup(p, true);
1413	check_preempt_curr(rq, p, wake_flags);
1414
1415	p->state = TASK_RUNNING;
1416#ifdef CONFIG_SMP
1417	if (p->sched_class->task_woken)
1418		p->sched_class->task_woken(rq, p);
1419
1420	if (rq->idle_stamp) {
1421		u64 delta = rq->clock - rq->idle_stamp;
1422		u64 max = 2*sysctl_sched_migration_cost;
1423
1424		if (delta > max)
1425			rq->avg_idle = max;
1426		else
1427			update_avg(&rq->avg_idle, delta);
1428		rq->idle_stamp = 0;
1429	}
1430#endif
1431}
1432
1433static void
1434ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1435{
1436#ifdef CONFIG_SMP
1437	if (p->sched_contributes_to_load)
1438		rq->nr_uninterruptible--;
1439#endif
1440
1441	ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1442	ttwu_do_wakeup(rq, p, wake_flags);
1443}
1444
1445/*
1446 * Called in case the task @p isn't fully descheduled from its runqueue,
1447 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1448 * since all we need to do is flip p->state to TASK_RUNNING, since
1449 * the task is still ->on_rq.
1450 */
1451static int ttwu_remote(struct task_struct *p, int wake_flags)
1452{
1453	struct rq *rq;
1454	int ret = 0;
1455
1456	rq = __task_rq_lock(p);
1457	if (p->on_rq) {
1458		ttwu_do_wakeup(rq, p, wake_flags);
1459		ret = 1;
1460	}
1461	__task_rq_unlock(rq);
1462
1463	return ret;
1464}
1465
1466#ifdef CONFIG_SMP
1467static void sched_ttwu_pending(void)
1468{
1469	struct rq *rq = this_rq();
1470	struct llist_node *llist = llist_del_all(&rq->wake_list);
1471	struct task_struct *p;
1472
1473	raw_spin_lock(&rq->lock);
1474
1475	while (llist) {
1476		p = llist_entry(llist, struct task_struct, wake_entry);
1477		llist = llist_next(llist);
1478		ttwu_do_activate(rq, p, 0);
1479	}
1480
1481	raw_spin_unlock(&rq->lock);
1482}
1483
1484void scheduler_ipi(void)
1485{
1486	if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1487		return;
1488
1489	/*
1490	 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1491	 * traditionally all their work was done from the interrupt return
1492	 * path. Now that we actually do some work, we need to make sure
1493	 * we do call them.
1494	 *
1495	 * Some archs already do call them, luckily irq_enter/exit nest
1496	 * properly.
1497	 *
1498	 * Arguably we should visit all archs and update all handlers,
1499	 * however a fair share of IPIs are still resched only so this would
1500	 * somewhat pessimize the simple resched case.
1501	 */
1502	irq_enter();
1503	sched_ttwu_pending();
1504
1505	/*
1506	 * Check if someone kicked us for doing the nohz idle load balance.
1507	 */
1508	if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1509		this_rq()->idle_balance = 1;
1510		raise_softirq_irqoff(SCHED_SOFTIRQ);
1511	}
1512	irq_exit();
1513}
1514
1515static void ttwu_queue_remote(struct task_struct *p, int cpu)
1516{
1517	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1518		smp_send_reschedule(cpu);
1519}
1520
1521#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1522static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1523{
1524	struct rq *rq;
1525	int ret = 0;
1526
1527	rq = __task_rq_lock(p);
1528	if (p->on_cpu) {
1529		ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1530		ttwu_do_wakeup(rq, p, wake_flags);
1531		ret = 1;
1532	}
1533	__task_rq_unlock(rq);
1534
1535	return ret;
1536
1537}
1538#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1539
1540bool cpus_share_cache(int this_cpu, int that_cpu)
1541{
1542	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1543}
1544#endif /* CONFIG_SMP */
1545
1546static void ttwu_queue(struct task_struct *p, int cpu)
1547{
1548	struct rq *rq = cpu_rq(cpu);
1549
1550#if defined(CONFIG_SMP)
1551	if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1552		sched_clock_cpu(cpu); /* sync clocks x-cpu */
1553		ttwu_queue_remote(p, cpu);
1554		return;
1555	}
1556#endif
1557
1558	raw_spin_lock(&rq->lock);
1559	ttwu_do_activate(rq, p, 0);
1560	raw_spin_unlock(&rq->lock);
1561}
1562
1563/**
1564 * try_to_wake_up - wake up a thread
1565 * @p: the thread to be awakened
1566 * @state: the mask of task states that can be woken
1567 * @wake_flags: wake modifier flags (WF_*)
1568 *
1569 * Put it on the run-queue if it's not already there. The "current"
1570 * thread is always on the run-queue (except when the actual
1571 * re-schedule is in progress), and as such you're allowed to do
1572 * the simpler "current->state = TASK_RUNNING" to mark yourself
1573 * runnable without the overhead of this.
1574 *
1575 * Returns %true if @p was woken up, %false if it was already running
1576 * or @state didn't match @p's state.
1577 */
1578static int
1579try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1580{
1581	unsigned long flags;
1582	int cpu, success = 0;
1583
1584	smp_wmb();
1585	raw_spin_lock_irqsave(&p->pi_lock, flags);
1586	if (!(p->state & state))
1587		goto out;
1588
1589	success = 1; /* we're going to change ->state */
1590	cpu = task_cpu(p);
1591
1592	if (p->on_rq && ttwu_remote(p, wake_flags))
1593		goto stat;
1594
1595#ifdef CONFIG_SMP
1596	/*
1597	 * If the owning (remote) cpu is still in the middle of schedule() with
1598	 * this task as prev, wait until its done referencing the task.
1599	 */
1600	while (p->on_cpu) {
1601#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1602		/*
1603		 * In case the architecture enables interrupts in
1604		 * context_switch(), we cannot busy wait, since that
1605		 * would lead to deadlocks when an interrupt hits and
1606		 * tries to wake up @prev. So bail and do a complete
1607		 * remote wakeup.
1608		 */
1609		if (ttwu_activate_remote(p, wake_flags))
1610			goto stat;
1611#else
1612		cpu_relax();
1613#endif
1614	}
1615	/*
1616	 * Pairs with the smp_wmb() in finish_lock_switch().
1617	 */
1618	smp_rmb();
1619
1620	p->sched_contributes_to_load = !!task_contributes_to_load(p);
1621	p->state = TASK_WAKING;
1622
1623	if (p->sched_class->task_waking)
1624		p->sched_class->task_waking(p);
1625
1626	cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1627	if (task_cpu(p) != cpu) {
1628		wake_flags |= WF_MIGRATED;
1629		set_task_cpu(p, cpu);
1630	}
1631#endif /* CONFIG_SMP */
1632
1633	ttwu_queue(p, cpu);
1634stat:
1635	ttwu_stat(p, cpu, wake_flags);
1636out:
1637	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1638
1639	return success;
1640}
1641
1642/**
1643 * try_to_wake_up_local - try to wake up a local task with rq lock held
1644 * @p: the thread to be awakened
1645 *
1646 * Put @p on the run-queue if it's not already there. The caller must
1647 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1648 * the current task.
1649 */
1650static void try_to_wake_up_local(struct task_struct *p)
1651{
1652	struct rq *rq = task_rq(p);
1653
1654	BUG_ON(rq != this_rq());
1655	BUG_ON(p == current);
1656	lockdep_assert_held(&rq->lock);
1657
1658	if (!raw_spin_trylock(&p->pi_lock)) {
1659		raw_spin_unlock(&rq->lock);
1660		raw_spin_lock(&p->pi_lock);
1661		raw_spin_lock(&rq->lock);
1662	}
1663
1664	if (!(p->state & TASK_NORMAL))
1665		goto out;
1666
1667	if (!p->on_rq)
1668		ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1669
1670	ttwu_do_wakeup(rq, p, 0);
1671	ttwu_stat(p, smp_processor_id(), 0);
1672out:
1673	raw_spin_unlock(&p->pi_lock);
1674}
1675
1676/**
1677 * wake_up_process - Wake up a specific process
1678 * @p: The process to be woken up.
1679 *
1680 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * processes.  Returns 1 if the process was woken up, 0 if it was already
1682 * running.
1683 *
1684 * It may be assumed that this function implies a write memory barrier before
1685 * changing the task state if and only if any tasks are woken up.
1686 */
1687int wake_up_process(struct task_struct *p)
1688{
1689	return try_to_wake_up(p, TASK_ALL, 0);
1690}
1691EXPORT_SYMBOL(wake_up_process);
1692
1693int wake_up_state(struct task_struct *p, unsigned int state)
1694{
1695	return try_to_wake_up(p, state, 0);
1696}
1697
1698/*
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1701 *
1702 * __sched_fork() is basic setup used by init_idle() too:
1703 */
1704static void __sched_fork(struct task_struct *p)
1705{
1706	p->on_rq			= 0;
1707
1708	p->se.on_rq			= 0;
1709	p->se.exec_start		= 0;
1710	p->se.sum_exec_runtime		= 0;
1711	p->se.prev_sum_exec_runtime	= 0;
1712	p->se.nr_migrations		= 0;
1713	p->se.vruntime			= 0;
1714	INIT_LIST_HEAD(&p->se.group_node);
1715
1716#ifdef CONFIG_SCHEDSTATS
1717	memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1718#endif
1719
1720	INIT_LIST_HEAD(&p->rt.run_list);
1721
1722#ifdef CONFIG_PREEMPT_NOTIFIERS
1723	INIT_HLIST_HEAD(&p->preempt_notifiers);
1724#endif
1725}
1726
1727/*
1728 * fork()/clone()-time setup:
1729 */
1730void sched_fork(struct task_struct *p)
1731{
1732	unsigned long flags;
1733	int cpu = get_cpu();
1734
1735	__sched_fork(p);
1736	/*
1737	 * We mark the process as running here. This guarantees that
1738	 * nobody will actually run it, and a signal or other external
1739	 * event cannot wake it up and insert it on the runqueue either.
1740	 */
1741	p->state = TASK_RUNNING;
1742
1743	/*
1744	 * Make sure we do not leak PI boosting priority to the child.
1745	 */
1746	p->prio = current->normal_prio;
1747
1748	/*
1749	 * Revert to default priority/policy on fork if requested.
1750	 */
1751	if (unlikely(p->sched_reset_on_fork)) {
1752		if (task_has_rt_policy(p)) {
1753			p->policy = SCHED_NORMAL;
1754			p->static_prio = NICE_TO_PRIO(0);
1755			p->rt_priority = 0;
1756		} else if (PRIO_TO_NICE(p->static_prio) < 0)
1757			p->static_prio = NICE_TO_PRIO(0);
1758
1759		p->prio = p->normal_prio = __normal_prio(p);
1760		set_load_weight(p);
1761
1762		/*
1763		 * We don't need the reset flag anymore after the fork. It has
1764		 * fulfilled its duty:
1765		 */
1766		p->sched_reset_on_fork = 0;
1767	}
1768
1769	if (!rt_prio(p->prio))
1770		p->sched_class = &fair_sched_class;
1771
1772	if (p->sched_class->task_fork)
1773		p->sched_class->task_fork(p);
1774
1775	/*
1776	 * The child is not yet in the pid-hash so no cgroup attach races,
1777	 * and the cgroup is pinned to this child due to cgroup_fork()
1778	 * is ran before sched_fork().
1779	 *
1780	 * Silence PROVE_RCU.
1781	 */
1782	raw_spin_lock_irqsave(&p->pi_lock, flags);
1783	set_task_cpu(p, cpu);
1784	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1785
1786#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1787	if (likely(sched_info_on()))
1788		memset(&p->sched_info, 0, sizeof(p->sched_info));
1789#endif
1790#if defined(CONFIG_SMP)
1791	p->on_cpu = 0;
1792#endif
1793#ifdef CONFIG_PREEMPT_COUNT
1794	/* Want to start with kernel preemption disabled. */
1795	task_thread_info(p)->preempt_count = 1;
1796#endif
1797#ifdef CONFIG_SMP
1798	plist_node_init(&p->pushable_tasks, MAX_PRIO);
1799#endif
1800
1801	put_cpu();
1802}
1803
1804/*
1805 * wake_up_new_task - wake up a newly created task for the first time.
1806 *
1807 * This function will do some initial scheduler statistics housekeeping
1808 * that must be done for every newly created context, then puts the task
1809 * on the runqueue and wakes it.
1810 */
1811void wake_up_new_task(struct task_struct *p)
1812{
1813	unsigned long flags;
1814	struct rq *rq;
1815
1816	raw_spin_lock_irqsave(&p->pi_lock, flags);
1817#ifdef CONFIG_SMP
1818	/*
1819	 * Fork balancing, do it here and not earlier because:
1820	 *  - cpus_allowed can change in the fork path
1821	 *  - any previously selected cpu might disappear through hotplug
1822	 */
1823	set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1824#endif
1825
1826	rq = __task_rq_lock(p);
1827	activate_task(rq, p, 0);
1828	p->on_rq = 1;
1829	trace_sched_wakeup_new(p, true);
1830	check_preempt_curr(rq, p, WF_FORK);
1831#ifdef CONFIG_SMP
1832	if (p->sched_class->task_woken)
1833		p->sched_class->task_woken(rq, p);
1834#endif
1835	task_rq_unlock(rq, p, &flags);
1836}
1837
1838#ifdef CONFIG_PREEMPT_NOTIFIERS
1839
1840/**
1841 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1842 * @notifier: notifier struct to register
1843 */
1844void preempt_notifier_register(struct preempt_notifier *notifier)
1845{
1846	hlist_add_head(&notifier->link, &current->preempt_notifiers);
1847}
1848EXPORT_SYMBOL_GPL(preempt_notifier_register);
1849
1850/**
1851 * preempt_notifier_unregister - no longer interested in preemption notifications
1852 * @notifier: notifier struct to unregister
1853 *
1854 * This is safe to call from within a preemption notifier.
1855 */
1856void preempt_notifier_unregister(struct preempt_notifier *notifier)
1857{
1858	hlist_del(&notifier->link);
1859}
1860EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1861
1862static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1863{
1864	struct preempt_notifier *notifier;
1865	struct hlist_node *node;
1866
1867	hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1868		notifier->ops->sched_in(notifier, raw_smp_processor_id());
1869}
1870
1871static void
1872fire_sched_out_preempt_notifiers(struct task_struct *curr,
1873				 struct task_struct *next)
1874{
1875	struct preempt_notifier *notifier;
1876	struct hlist_node *node;
1877
1878	hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1879		notifier->ops->sched_out(notifier, next);
1880}
1881
1882#else /* !CONFIG_PREEMPT_NOTIFIERS */
1883
1884static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1885{
1886}
1887
1888static void
1889fire_sched_out_preempt_notifiers(struct task_struct *curr,
1890				 struct task_struct *next)
1891{
1892}
1893
1894#endif /* CONFIG_PREEMPT_NOTIFIERS */
1895
1896/**
1897 * prepare_task_switch - prepare to switch tasks
1898 * @rq: the runqueue preparing to switch
1899 * @prev: the current task that is being switched out
1900 * @next: the task we are going to switch to.
1901 *
1902 * This is called with the rq lock held and interrupts off. It must
1903 * be paired with a subsequent finish_task_switch after the context
1904 * switch.
1905 *
1906 * prepare_task_switch sets up locking and calls architecture specific
1907 * hooks.
1908 */
1909static inline void
1910prepare_task_switch(struct rq *rq, struct task_struct *prev,
1911		    struct task_struct *next)
1912{
1913	sched_info_switch(prev, next);
1914	perf_event_task_sched_out(prev, next);
1915	fire_sched_out_preempt_notifiers(prev, next);
1916	prepare_lock_switch(rq, next);
1917	prepare_arch_switch(next);
1918	trace_sched_switch(prev, next);
1919}
1920
1921/**
1922 * finish_task_switch - clean up after a task-switch
1923 * @rq: runqueue associated with task-switch
1924 * @prev: the thread we just switched away from.
1925 *
1926 * finish_task_switch must be called after the context switch, paired
1927 * with a prepare_task_switch call before the context switch.
1928 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1929 * and do any other architecture-specific cleanup actions.
1930 *
1931 * Note that we may have delayed dropping an mm in context_switch(). If
1932 * so, we finish that here outside of the runqueue lock. (Doing it
1933 * with the lock held can cause deadlocks; see schedule() for
1934 * details.)
1935 */
1936static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1937	__releases(rq->lock)
1938{
1939	struct mm_struct *mm = rq->prev_mm;
1940	long prev_state;
1941
1942	rq->prev_mm = NULL;
1943
1944	/*
1945	 * A task struct has one reference for the use as "current".
1946	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1947	 * schedule one last time. The schedule call will never return, and
1948	 * the scheduled task must drop that reference.
1949	 * The test for TASK_DEAD must occur while the runqueue locks are
1950	 * still held, otherwise prev could be scheduled on another cpu, die
1951	 * there before we look at prev->state, and then the reference would
1952	 * be dropped twice.
1953	 *		Manfred Spraul <manfred@colorfullife.com>
1954	 */
1955	prev_state = prev->state;
1956	finish_arch_switch(prev);
1957#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1958	local_irq_disable();
1959#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1960	perf_event_task_sched_in(prev, current);
1961#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1962	local_irq_enable();
1963#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1964	finish_lock_switch(rq, prev);
1965	finish_arch_post_lock_switch();
1966
1967	fire_sched_in_preempt_notifiers(current);
1968	if (mm)
1969		mmdrop(mm);
1970	if (unlikely(prev_state == TASK_DEAD)) {
1971		/*
1972		 * Remove function-return probe instances associated with this
1973		 * task and put them back on the free list.
1974		 */
1975		kprobe_flush_task(prev);
1976		put_task_struct(prev);
1977	}
1978}
1979
1980#ifdef CONFIG_SMP
1981
1982/* assumes rq->lock is held */
1983static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1984{
1985	if (prev->sched_class->pre_schedule)
1986		prev->sched_class->pre_schedule(rq, prev);
1987}
1988
1989/* rq->lock is NOT held, but preemption is disabled */
1990static inline void post_schedule(struct rq *rq)
1991{
1992	if (rq->post_schedule) {
1993		unsigned long flags;
1994
1995		raw_spin_lock_irqsave(&rq->lock, flags);
1996		if (rq->curr->sched_class->post_schedule)
1997			rq->curr->sched_class->post_schedule(rq);
1998		raw_spin_unlock_irqrestore(&rq->lock, flags);
1999
2000		rq->post_schedule = 0;
2001	}
2002}
2003
2004#else
2005
2006static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2007{
2008}
2009
2010static inline void post_schedule(struct rq *rq)
2011{
2012}
2013
2014#endif
2015
2016/**
2017 * schedule_tail - first thing a freshly forked thread must call.
2018 * @prev: the thread we just switched away from.
2019 */
2020asmlinkage void schedule_tail(struct task_struct *prev)
2021	__releases(rq->lock)
2022{
2023	struct rq *rq = this_rq();
2024
2025	finish_task_switch(rq, prev);
2026
2027	/*
2028	 * FIXME: do we need to worry about rq being invalidated by the
2029	 * task_switch?
2030	 */
2031	post_schedule(rq);
2032
2033#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2034	/* In this case, finish_task_switch does not reenable preemption */
2035	preempt_enable();
2036#endif
2037	if (current->set_child_tid)
2038		put_user(task_pid_vnr(current), current->set_child_tid);
2039}
2040
2041/*
2042 * context_switch - switch to the new MM and the new
2043 * thread's register state.
2044 */
2045static inline void
2046context_switch(struct rq *rq, struct task_struct *prev,
2047	       struct task_struct *next)
2048{
2049	struct mm_struct *mm, *oldmm;
2050
2051	prepare_task_switch(rq, prev, next);
2052
2053	mm = next->mm;
2054	oldmm = prev->active_mm;
2055	/*
2056	 * For paravirt, this is coupled with an exit in switch_to to
2057	 * combine the page table reload and the switch backend into
2058	 * one hypercall.
2059	 */
2060	arch_start_context_switch(prev);
2061
2062	if (!mm) {
2063		next->active_mm = oldmm;
2064		atomic_inc(&oldmm->mm_count);
2065		enter_lazy_tlb(oldmm, next);
2066	} else
2067		switch_mm(oldmm, mm, next);
2068
2069	if (!prev->mm) {
2070		prev->active_mm = NULL;
2071		rq->prev_mm = oldmm;
2072	}
2073	/*
2074	 * Since the runqueue lock will be released by the next
2075	 * task (which is an invalid locking op but in the case
2076	 * of the scheduler it's an obvious special-case), so we
2077	 * do an early lockdep release here:
2078	 */
2079#ifndef __ARCH_WANT_UNLOCKED_CTXSW
2080	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2081#endif
2082
2083	/* Here we just switch the register state and the stack. */
2084	switch_to(prev, next, prev);
2085
2086	barrier();
2087	/*
2088	 * this_rq must be evaluated again because prev may have moved
2089	 * CPUs since it called schedule(), thus the 'rq' on its stack
2090	 * frame will be invalid.
2091	 */
2092	finish_task_switch(this_rq(), prev);
2093}
2094
2095/*
2096 * nr_running, nr_uninterruptible and nr_context_switches:
2097 *
2098 * externally visible scheduler statistics: current number of runnable
2099 * threads, current number of uninterruptible-sleeping threads, total
2100 * number of context switches performed since bootup.
2101 */
2102unsigned long nr_running(void)
2103{
2104	unsigned long i, sum = 0;
2105
2106	for_each_online_cpu(i)
2107		sum += cpu_rq(i)->nr_running;
2108
2109	return sum;
2110}
2111
2112unsigned long nr_uninterruptible(void)
2113{
2114	unsigned long i, sum = 0;
2115
2116	for_each_possible_cpu(i)
2117		sum += cpu_rq(i)->nr_uninterruptible;
2118
2119	/*
2120	 * Since we read the counters lockless, it might be slightly
2121	 * inaccurate. Do not allow it to go below zero though:
2122	 */
2123	if (unlikely((long)sum < 0))
2124		sum = 0;
2125
2126	return sum;
2127}
2128
2129unsigned long long nr_context_switches(void)
2130{
2131	int i;
2132	unsigned long long sum = 0;
2133
2134	for_each_possible_cpu(i)
2135		sum += cpu_rq(i)->nr_switches;
2136
2137	return sum;
2138}
2139
2140unsigned long nr_iowait(void)
2141{
2142	unsigned long i, sum = 0;
2143
2144	for_each_possible_cpu(i)
2145		sum += atomic_read(&cpu_rq(i)->nr_iowait);
2146
2147	return sum;
2148}
2149
2150unsigned long nr_iowait_cpu(int cpu)
2151{
2152	struct rq *this = cpu_rq(cpu);
2153	return atomic_read(&this->nr_iowait);
2154}
2155
2156unsigned long this_cpu_load(void)
2157{
2158	struct rq *this = this_rq();
2159	return this->cpu_load[0];
2160}
2161
2162
2163/*
2164 * Global load-average calculations
2165 *
2166 * We take a distributed and async approach to calculating the global load-avg
2167 * in order to minimize overhead.
2168 *
2169 * The global load average is an exponentially decaying average of nr_running +
2170 * nr_uninterruptible.
2171 *
2172 * Once every LOAD_FREQ:
2173 *
2174 *   nr_active = 0;
2175 *   for_each_possible_cpu(cpu)
2176 *   	nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2177 *
2178 *   avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2179 *
2180 * Due to a number of reasons the above turns in the mess below:
2181 *
2182 *  - for_each_possible_cpu() is prohibitively expensive on machines with
2183 *    serious number of cpus, therefore we need to take a distributed approach
2184 *    to calculating nr_active.
2185 *
2186 *        \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2187 *                      = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2188 *
2189 *    So assuming nr_active := 0 when we start out -- true per definition, we
2190 *    can simply take per-cpu deltas and fold those into a global accumulate
2191 *    to obtain the same result. See calc_load_fold_active().
2192 *
2193 *    Furthermore, in order to avoid synchronizing all per-cpu delta folding
2194 *    across the machine, we assume 10 ticks is sufficient time for every
2195 *    cpu to have completed this task.
2196 *
2197 *    This places an upper-bound on the IRQ-off latency of the machine. Then
2198 *    again, being late doesn't loose the delta, just wrecks the sample.
2199 *
2200 *  - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2201 *    this would add another cross-cpu cacheline miss and atomic operation
2202 *    to the wakeup path. Instead we increment on whatever cpu the task ran
2203 *    when it went into uninterruptible state and decrement on whatever cpu
2204 *    did the wakeup. This means that only the sum of nr_uninterruptible over
2205 *    all cpus yields the correct result.
2206 *
2207 *  This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2208 */
2209
2210/* Variables and functions for calc_load */
2211static atomic_long_t calc_load_tasks;
2212static unsigned long calc_load_update;
2213unsigned long avenrun[3];
2214EXPORT_SYMBOL(avenrun); /* should be removed */
2215
2216/**
2217 * get_avenrun - get the load average array
2218 * @loads:	pointer to dest load array
2219 * @offset:	offset to add
2220 * @shift:	shift count to shift the result left
2221 *
2222 * These values are estimates at best, so no need for locking.
2223 */
2224void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2225{
2226	loads[0] = (avenrun[0] + offset) << shift;
2227	loads[1] = (avenrun[1] + offset) << shift;
2228	loads[2] = (avenrun[2] + offset) << shift;
2229}
2230
2231static long calc_load_fold_active(struct rq *this_rq)
2232{
2233	long nr_active, delta = 0;
2234
2235	nr_active = this_rq->nr_running;
2236	nr_active += (long) this_rq->nr_uninterruptible;
2237
2238	if (nr_active != this_rq->calc_load_active) {
2239		delta = nr_active - this_rq->calc_load_active;
2240		this_rq->calc_load_active = nr_active;
2241	}
2242
2243	return delta;
2244}
2245
2246/*
2247 * a1 = a0 * e + a * (1 - e)
2248 */
2249static unsigned long
2250calc_load(unsigned long load, unsigned long exp, unsigned long active)
2251{
2252	load *= exp;
2253	load += active * (FIXED_1 - exp);
2254	load += 1UL << (FSHIFT - 1);
2255	return load >> FSHIFT;
2256}
2257
2258#ifdef CONFIG_NO_HZ
2259/*
2260 * Handle NO_HZ for the global load-average.
2261 *
2262 * Since the above described distributed algorithm to compute the global
2263 * load-average relies on per-cpu sampling from the tick, it is affected by
2264 * NO_HZ.
2265 *
2266 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2267 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2268 * when we read the global state.
2269 *
2270 * Obviously reality has to ruin such a delightfully simple scheme:
2271 *
2272 *  - When we go NO_HZ idle during the window, we can negate our sample
2273 *    contribution, causing under-accounting.
2274 *
2275 *    We avoid this by keeping two idle-delta counters and flipping them
2276 *    when the window starts, thus separating old and new NO_HZ load.
2277 *
2278 *    The only trick is the slight shift in index flip for read vs write.
2279 *
2280 *        0s            5s            10s           15s
2281 *          +10           +10           +10           +10
2282 *        |-|-----------|-|-----------|-|-----------|-|
2283 *    r:0 0 1           1 0           0 1           1 0
2284 *    w:0 1 1           0 0           1 1           0 0
2285 *
2286 *    This ensures we'll fold the old idle contribution in this window while
2287 *    accumlating the new one.
2288 *
2289 *  - When we wake up from NO_HZ idle during the window, we push up our
2290 *    contribution, since we effectively move our sample point to a known
2291 *    busy state.
2292 *
2293 *    This is solved by pushing the window forward, and thus skipping the
2294 *    sample, for this cpu (effectively using the idle-delta for this cpu which
2295 *    was in effect at the time the window opened). This also solves the issue
2296 *    of having to deal with a cpu having been in NOHZ idle for multiple
2297 *    LOAD_FREQ intervals.
2298 *
2299 * When making the ILB scale, we should try to pull this in as well.
2300 */
2301static atomic_long_t calc_load_idle[2];
2302static int calc_load_idx;
2303
2304static inline int calc_load_write_idx(void)
2305{
2306	int idx = calc_load_idx;
2307
2308	/*
2309	 * See calc_global_nohz(), if we observe the new index, we also
2310	 * need to observe the new update time.
2311	 */
2312	smp_rmb();
2313
2314	/*
2315	 * If the folding window started, make sure we start writing in the
2316	 * next idle-delta.
2317	 */
2318	if (!time_before(jiffies, calc_load_update))
2319		idx++;
2320
2321	return idx & 1;
2322}
2323
2324static inline int calc_load_read_idx(void)
2325{
2326	return calc_load_idx & 1;
2327}
2328
2329void calc_load_enter_idle(void)
2330{
2331	struct rq *this_rq = this_rq();
2332	long delta;
2333
2334	/*
2335	 * We're going into NOHZ mode, if there's any pending delta, fold it
2336	 * into the pending idle delta.
2337	 */
2338	delta = calc_load_fold_active(this_rq);
2339	if (delta) {
2340		int idx = calc_load_write_idx();
2341		atomic_long_add(delta, &calc_load_idle[idx]);
2342	}
2343}
2344
2345void calc_load_exit_idle(void)
2346{
2347	struct rq *this_rq = this_rq();
2348
2349	/*
2350	 * If we're still before the sample window, we're done.
2351	 */
2352	if (time_before(jiffies, this_rq->calc_load_update))
2353		return;
2354
2355	/*
2356	 * We woke inside or after the sample window, this means we're already
2357	 * accounted through the nohz accounting, so skip the entire deal and
2358	 * sync up for the next window.
2359	 */
2360	this_rq->calc_load_update = calc_load_update;
2361	if (time_before(jiffies, this_rq->calc_load_update + 10))
2362		this_rq->calc_load_update += LOAD_FREQ;
2363}
2364
2365static long calc_load_fold_idle(void)
2366{
2367	int idx = calc_load_read_idx();
2368	long delta = 0;
2369
2370	if (atomic_long_read(&calc_load_idle[idx]))
2371		delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2372
2373	return delta;
2374}
2375
2376/**
2377 * fixed_power_int - compute: x^n, in O(log n) time
2378 *
2379 * @x:         base of the power
2380 * @frac_bits: fractional bits of @x
2381 * @n:         power to raise @x to.
2382 *
2383 * By exploiting the relation between the definition of the natural power
2384 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2385 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2386 * (where: n_i \elem {0, 1}, the binary vector representing n),
2387 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2388 * of course trivially computable in O(log_2 n), the length of our binary
2389 * vector.
2390 */
2391static unsigned long
2392fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2393{
2394	unsigned long result = 1UL << frac_bits;
2395
2396	if (n) for (;;) {
2397		if (n & 1) {
2398			result *= x;
2399			result += 1UL << (frac_bits - 1);
2400			result >>= frac_bits;
2401		}
2402		n >>= 1;
2403		if (!n)
2404			break;
2405		x *= x;
2406		x += 1UL << (frac_bits - 1);
2407		x >>= frac_bits;
2408	}
2409
2410	return result;
2411}
2412
2413/*
2414 * a1 = a0 * e + a * (1 - e)
2415 *
2416 * a2 = a1 * e + a * (1 - e)
2417 *    = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2418 *    = a0 * e^2 + a * (1 - e) * (1 + e)
2419 *
2420 * a3 = a2 * e + a * (1 - e)
2421 *    = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2422 *    = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2423 *
2424 *  ...
2425 *
2426 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2427 *    = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2428 *    = a0 * e^n + a * (1 - e^n)
2429 *
2430 * [1] application of the geometric series:
2431 *
2432 *              n         1 - x^(n+1)
2433 *     S_n := \Sum x^i = -------------
2434 *             i=0          1 - x
2435 */
2436static unsigned long
2437calc_load_n(unsigned long load, unsigned long exp,
2438	    unsigned long active, unsigned int n)
2439{
2440
2441	return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2442}
2443
2444/*
2445 * NO_HZ can leave us missing all per-cpu ticks calling
2446 * calc_load_account_active(), but since an idle CPU folds its delta into
2447 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2448 * in the pending idle delta if our idle period crossed a load cycle boundary.
2449 *
2450 * Once we've updated the global active value, we need to apply the exponential
2451 * weights adjusted to the number of cycles missed.
2452 */
2453static void calc_global_nohz(void)
2454{
2455	long delta, active, n;
2456
2457	if (!time_before(jiffies, calc_load_update + 10)) {
2458		/*
2459		 * Catch-up, fold however many we are behind still
2460		 */
2461		delta = jiffies - calc_load_update - 10;
2462		n = 1 + (delta / LOAD_FREQ);
2463
2464		active = atomic_long_read(&calc_load_tasks);
2465		active = active > 0 ? active * FIXED_1 : 0;
2466
2467		avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2468		avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2469		avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2470
2471		calc_load_update += n * LOAD_FREQ;
2472	}
2473
2474	/*
2475	 * Flip the idle index...
2476	 *
2477	 * Make sure we first write the new time then flip the index, so that
2478	 * calc_load_write_idx() will see the new time when it reads the new
2479	 * index, this avoids a double flip messing things up.
2480	 */
2481	smp_wmb();
2482	calc_load_idx++;
2483}
2484#else /* !CONFIG_NO_HZ */
2485
2486static inline long calc_load_fold_idle(void) { return 0; }
2487static inline void calc_global_nohz(void) { }
2488
2489#endif /* CONFIG_NO_HZ */
2490
2491/*
2492 * calc_load - update the avenrun load estimates 10 ticks after the
2493 * CPUs have updated calc_load_tasks.
2494 */
2495void calc_global_load(unsigned long ticks)
2496{
2497	long active, delta;
2498
2499	if (time_before(jiffies, calc_load_update + 10))
2500		return;
2501
2502	/*
2503	 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2504	 */
2505	delta = calc_load_fold_idle();
2506	if (delta)
2507		atomic_long_add(delta, &calc_load_tasks);
2508
2509	active = atomic_long_read(&calc_load_tasks);
2510	active = active > 0 ? active * FIXED_1 : 0;
2511
2512	avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2513	avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2514	avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2515
2516	calc_load_update += LOAD_FREQ;
2517
2518	/*
2519	 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2520	 */
2521	calc_global_nohz();
2522}
2523
2524/*
2525 * Called from update_cpu_load() to periodically update this CPU's
2526 * active count.
2527 */
2528static void calc_load_account_active(struct rq *this_rq)
2529{
2530	long delta;
2531
2532	if (time_before(jiffies, this_rq->calc_load_update))
2533		return;
2534
2535	delta  = calc_load_fold_active(this_rq);
2536	if (delta)
2537		atomic_long_add(delta, &calc_load_tasks);
2538
2539	this_rq->calc_load_update += LOAD_FREQ;
2540}
2541
2542/*
2543 * End of global load-average stuff
2544 */
2545
2546/*
2547 * The exact cpuload at various idx values, calculated at every tick would be
2548 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2549 *
2550 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2551 * on nth tick when cpu may be busy, then we have:
2552 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2553 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2554 *
2555 * decay_load_missed() below does efficient calculation of
2556 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2557 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2558 *
2559 * The calculation is approximated on a 128 point scale.
2560 * degrade_zero_ticks is the number of ticks after which load at any
2561 * particular idx is approximated to be zero.
2562 * degrade_factor is a precomputed table, a row for each load idx.
2563 * Each column corresponds to degradation factor for a power of two ticks,
2564 * based on 128 point scale.
2565 * Example:
2566 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2567 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2568 *
2569 * With this power of 2 load factors, we can degrade the load n times
2570 * by looking at 1 bits in n and doing as many mult/shift instead of
2571 * n mult/shifts needed by the exact degradation.
2572 */
2573#define DEGRADE_SHIFT		7
2574static const unsigned char
2575		degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2576static const unsigned char
2577		degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2578					{0, 0, 0, 0, 0, 0, 0, 0},
2579					{64, 32, 8, 0, 0, 0, 0, 0},
2580					{96, 72, 40, 12, 1, 0, 0},
2581					{112, 98, 75, 43, 15, 1, 0},
2582					{120, 112, 98, 76, 45, 16, 2} };
2583
2584/*
2585 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2586 * would be when CPU is idle and so we just decay the old load without
2587 * adding any new load.
2588 */
2589static unsigned long
2590decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2591{
2592	int j = 0;
2593
2594	if (!missed_updates)
2595		return load;
2596
2597	if (missed_updates >= degrade_zero_ticks[idx])
2598		return 0;
2599
2600	if (idx == 1)
2601		return load >> missed_updates;
2602
2603	while (missed_updates) {
2604		if (missed_updates % 2)
2605			load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2606
2607		missed_updates >>= 1;
2608		j++;
2609	}
2610	return load;
2611}
2612
2613/*
2614 * Update rq->cpu_load[] statistics. This function is usually called every
2615 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2616 * every tick. We fix it up based on jiffies.
2617 */
2618static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2619			      unsigned long pending_updates)
2620{
2621	int i, scale;
2622
2623	this_rq->nr_load_updates++;
2624
2625	/* Update our load: */
2626	this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2627	for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2628		unsigned long old_load, new_load;
2629
2630		/* scale is effectively 1 << i now, and >> i divides by scale */
2631
2632		old_load = this_rq->cpu_load[i];
2633		old_load = decay_load_missed(old_load, pending_updates - 1, i);
2634		new_load = this_load;
2635		/*
2636		 * Round up the averaging division if load is increasing. This
2637		 * prevents us from getting stuck on 9 if the load is 10, for
2638		 * example.
2639		 */
2640		if (new_load > old_load)
2641			new_load += scale - 1;
2642
2643		this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2644	}
2645
2646	sched_avg_update(this_rq);
2647}
2648
2649#ifdef CONFIG_NO_HZ
2650/*
2651 * There is no sane way to deal with nohz on smp when using jiffies because the
2652 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2653 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2654 *
2655 * Therefore we cannot use the delta approach from the regular tick since that
2656 * would seriously skew the load calculation. However we'll make do for those
2657 * updates happening while idle (nohz_idle_balance) or coming out of idle
2658 * (tick_nohz_idle_exit).
2659 *
2660 * This means we might still be one tick off for nohz periods.
2661 */
2662
2663/*
2664 * Called from nohz_idle_balance() to update the load ratings before doing the
2665 * idle balance.
2666 */
2667void update_idle_cpu_load(struct rq *this_rq)
2668{
2669	unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2670	unsigned long load = this_rq->load.weight;
2671	unsigned long pending_updates;
2672
2673	/*
2674	 * bail if there's load or we're actually up-to-date.
2675	 */
2676	if (load || curr_jiffies == this_rq->last_load_update_tick)
2677		return;
2678
2679	pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2680	this_rq->last_load_update_tick = curr_jiffies;
2681
2682	__update_cpu_load(this_rq, load, pending_updates);
2683}
2684
2685/*
2686 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2687 */
2688void update_cpu_load_nohz(void)
2689{
2690	struct rq *this_rq = this_rq();
2691	unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2692	unsigned long pending_updates;
2693
2694	if (curr_jiffies == this_rq->last_load_update_tick)
2695		return;
2696
2697	raw_spin_lock(&this_rq->lock);
2698	pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2699	if (pending_updates) {
2700		this_rq->last_load_update_tick = curr_jiffies;
2701		/*
2702		 * We were idle, this means load 0, the current load might be
2703		 * !0 due to remote wakeups and the sort.
2704		 */
2705		__update_cpu_load(this_rq, 0, pending_updates);
2706	}
2707	raw_spin_unlock(&this_rq->lock);
2708}
2709#endif /* CONFIG_NO_HZ */
2710
2711/*
2712 * Called from scheduler_tick()
2713 */
2714static void update_cpu_load_active(struct rq *this_rq)
2715{
2716	/*
2717	 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2718	 */
2719	this_rq->last_load_update_tick = jiffies;
2720	__update_cpu_load(this_rq, this_rq->load.weight, 1);
2721
2722	calc_load_account_active(this_rq);
2723}
2724
2725#ifdef CONFIG_SMP
2726
2727/*
2728 * sched_exec - execve() is a valuable balancing opportunity, because at
2729 * this point the task has the smallest effective memory and cache footprint.
2730 */
2731void sched_exec(void)
2732{
2733	struct task_struct *p = current;
2734	unsigned long flags;
2735	int dest_cpu;
2736
2737	raw_spin_lock_irqsave(&p->pi_lock, flags);
2738	dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2739	if (dest_cpu == smp_processor_id())
2740		goto unlock;
2741
2742	if (likely(cpu_active(dest_cpu))) {
2743		struct migration_arg arg = { p, dest_cpu };
2744
2745		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2746		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2747		return;
2748	}
2749unlock:
2750	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2751}
2752
2753#endif
2754
2755DEFINE_PER_CPU(struct kernel_stat, kstat);
2756DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2757
2758EXPORT_PER_CPU_SYMBOL(kstat);
2759EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2760
2761/*
2762 * Return any ns on the sched_clock that have not yet been accounted in
2763 * @p in case that task is currently running.
2764 *
2765 * Called with task_rq_lock() held on @rq.
2766 */
2767static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2768{
2769	u64 ns = 0;
2770
2771	if (task_current(rq, p)) {
2772		update_rq_clock(rq);
2773		ns = rq->clock_task - p->se.exec_start;
2774		if ((s64)ns < 0)
2775			ns = 0;
2776	}
2777
2778	return ns;
2779}
2780
2781unsigned long long task_delta_exec(struct task_struct *p)
2782{
2783	unsigned long flags;
2784	struct rq *rq;
2785	u64 ns = 0;
2786
2787	rq = task_rq_lock(p, &flags);
2788	ns = do_task_delta_exec(p, rq);
2789	task_rq_unlock(rq, p, &flags);
2790
2791	return ns;
2792}
2793
2794/*
2795 * Return accounted runtime for the task.
2796 * In case the task is currently running, return the runtime plus current's
2797 * pending runtime that have not been accounted yet.
2798 */
2799unsigned long long task_sched_runtime(struct task_struct *p)
2800{
2801	unsigned long flags;
2802	struct rq *rq;
2803	u64 ns = 0;
2804
2805	rq = task_rq_lock(p, &flags);
2806	ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2807	task_rq_unlock(rq, p, &flags);
2808
2809	return ns;
2810}
2811
2812#ifdef CONFIG_CGROUP_CPUACCT
2813struct cgroup_subsys cpuacct_subsys;
2814struct cpuacct root_cpuacct;
2815#endif
2816
2817static inline void task_group_account_field(struct task_struct *p, int index,
2818					    u64 tmp)
2819{
2820#ifdef CONFIG_CGROUP_CPUACCT
2821	struct kernel_cpustat *kcpustat;
2822	struct cpuacct *ca;
2823#endif
2824	/*
2825	 * Since all updates are sure to touch the root cgroup, we
2826	 * get ourselves ahead and touch it first. If the root cgroup
2827	 * is the only cgroup, then nothing else should be necessary.
2828	 *
2829	 */
2830	__get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2831
2832#ifdef CONFIG_CGROUP_CPUACCT
2833	if (unlikely(!cpuacct_subsys.active))
2834		return;
2835
2836	rcu_read_lock();
2837	ca = task_ca(p);
2838	while (ca && (ca != &root_cpuacct)) {
2839		kcpustat = this_cpu_ptr(ca->cpustat);
2840		kcpustat->cpustat[index] += tmp;
2841		ca = parent_ca(ca);
2842	}
2843	rcu_read_unlock();
2844#endif
2845}
2846
2847
2848/*
2849 * Account user cpu time to a process.
2850 * @p: the process that the cpu time gets accounted to
2851 * @cputime: the cpu time spent in user space since the last update
2852 * @cputime_scaled: cputime scaled by cpu frequency
2853 */
2854void account_user_time(struct task_struct *p, cputime_t cputime,
2855		       cputime_t cputime_scaled)
2856{
2857	int index;
2858
2859	/* Add user time to process. */
2860	p->utime += cputime;
2861	p->utimescaled += cputime_scaled;
2862	account_group_user_time(p, cputime);
2863
2864	index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2865
2866	/* Add user time to cpustat. */
2867	task_group_account_field(p, index, (__force u64) cputime);
2868
2869	/* Account for user time used */
2870	acct_update_integrals(p);
2871}
2872
2873/*
2874 * Account guest cpu time to a process.
2875 * @p: the process that the cpu time gets accounted to
2876 * @cputime: the cpu time spent in virtual machine since the last update
2877 * @cputime_scaled: cputime scaled by cpu frequency
2878 */
2879static void account_guest_time(struct task_struct *p, cputime_t cputime,
2880			       cputime_t cputime_scaled)
2881{
2882	u64 *cpustat = kcpustat_this_cpu->cpustat;
2883
2884	/* Add guest time to process. */
2885	p->utime += cputime;
2886	p->utimescaled += cputime_scaled;
2887	account_group_user_time(p, cputime);
2888	p->gtime += cputime;
2889
2890	/* Add guest time to cpustat. */
2891	if (TASK_NICE(p) > 0) {
2892		cpustat[CPUTIME_NICE] += (__force u64) cputime;
2893		cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2894	} else {
2895		cpustat[CPUTIME_USER] += (__force u64) cputime;
2896		cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2897	}
2898}
2899
2900/*
2901 * Account system cpu time to a process and desired cpustat field
2902 * @p: the process that the cpu time gets accounted to
2903 * @cputime: the cpu time spent in kernel space since the last update
2904 * @cputime_scaled: cputime scaled by cpu frequency
2905 * @target_cputime64: pointer to cpustat field that has to be updated
2906 */
2907static inline
2908void __account_system_time(struct task_struct *p, cputime_t cputime,
2909			cputime_t cputime_scaled, int index)
2910{
2911	/* Add system time to process. */
2912	p->stime += cputime;
2913	p->stimescaled += cputime_scaled;
2914	account_group_system_time(p, cputime);
2915
2916	/* Add system time to cpustat. */
2917	task_group_account_field(p, index, (__force u64) cputime);
2918
2919	/* Account for system time used */
2920	acct_update_integrals(p);
2921}
2922
2923/*
2924 * Account system cpu time to a process.
2925 * @p: the process that the cpu time gets accounted to
2926 * @hardirq_offset: the offset to subtract from hardirq_count()
2927 * @cputime: the cpu time spent in kernel space since the last update
2928 * @cputime_scaled: cputime scaled by cpu frequency
2929 */
2930void account_system_time(struct task_struct *p, int hardirq_offset,
2931			 cputime_t cputime, cputime_t cputime_scaled)
2932{
2933	int index;
2934
2935	if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2936		account_guest_time(p, cputime, cputime_scaled);
2937		return;
2938	}
2939
2940	if (hardirq_count() - hardirq_offset)
2941		index = CPUTIME_IRQ;
2942	else if (in_serving_softirq())
2943		index = CPUTIME_SOFTIRQ;
2944	else
2945		index = CPUTIME_SYSTEM;
2946
2947	__account_system_time(p, cputime, cputime_scaled, index);
2948}
2949
2950/*
2951 * Account for involuntary wait time.
2952 * @cputime: the cpu time spent in involuntary wait
2953 */
2954void account_steal_time(cputime_t cputime)
2955{
2956	u64 *cpustat = kcpustat_this_cpu->cpustat;
2957
2958	cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2959}
2960
2961/*
2962 * Account for idle time.
2963 * @cputime: the cpu time spent in idle wait
2964 */
2965void account_idle_time(cputime_t cputime)
2966{
2967	u64 *cpustat = kcpustat_this_cpu->cpustat;
2968	struct rq *rq = this_rq();
2969
2970	if (atomic_read(&rq->nr_iowait) > 0)
2971		cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2972	else
2973		cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2974}
2975
2976static __always_inline bool steal_account_process_tick(void)
2977{
2978#ifdef CONFIG_PARAVIRT
2979	if (static_key_false(&paravirt_steal_enabled)) {
2980		u64 steal, st = 0;
2981
2982		steal = paravirt_steal_clock(smp_processor_id());
2983		steal -= this_rq()->prev_steal_time;
2984
2985		st = steal_ticks(steal);
2986		this_rq()->prev_steal_time += st * TICK_NSEC;
2987
2988		account_steal_time(st);
2989		return st;
2990	}
2991#endif
2992	return false;
2993}
2994
2995#ifndef CONFIG_VIRT_CPU_ACCOUNTING
2996
2997#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2998/*
2999 * Account a tick to a process and cpustat
3000 * @p: the process that the cpu time gets accounted to
3001 * @user_tick: is the tick from userspace
3002 * @rq: the pointer to rq
3003 *
3004 * Tick demultiplexing follows the order
3005 * - pending hardirq update
3006 * - pending softirq update
3007 * - user_time
3008 * - idle_time
3009 * - system time
3010 *   - check for guest_time
3011 *   - else account as system_time
3012 *
3013 * Check for hardirq is done both for system and user time as there is
3014 * no timer going off while we are on hardirq and hence we may never get an
3015 * opportunity to update it solely in system time.
3016 * p->stime and friends are only updated on system time and not on irq
3017 * softirq as those do not count in task exec_runtime any more.
3018 */
3019static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3020						struct rq *rq)
3021{
3022	cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3023	u64 *cpustat = kcpustat_this_cpu->cpustat;
3024
3025	if (steal_account_process_tick())
3026		return;
3027
3028	if (irqtime_account_hi_update()) {
3029		cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
3030	} else if (irqtime_account_si_update()) {
3031		cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
3032	} else if (this_cpu_ksoftirqd() == p) {
3033		/*
3034		 * ksoftirqd time do not get accounted in cpu_softirq_time.
3035		 * So, we have to handle it separately here.
3036		 * Also, p->stime needs to be updated for ksoftirqd.
3037		 */
3038		__account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3039					CPUTIME_SOFTIRQ);
3040	} else if (user_tick) {
3041		account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3042	} else if (p == rq->idle) {
3043		account_idle_time(cputime_one_jiffy);
3044	} else if (p->flags & PF_VCPU) { /* System time or guest time */
3045		account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3046	} else {
3047		__account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3048					CPUTIME_SYSTEM);
3049	}
3050}
3051
3052static void irqtime_account_idle_ticks(int ticks)
3053{
3054	int i;
3055	struct rq *rq = this_rq();
3056
3057	for (i = 0; i < ticks; i++)
3058		irqtime_account_process_tick(current, 0, rq);
3059}
3060#else /* CONFIG_IRQ_TIME_ACCOUNTING */
3061static void irqtime_account_idle_ticks(int ticks) {}
3062static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3063						struct rq *rq) {}
3064#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3065
3066/*
3067 * Account a single tick of cpu time.
3068 * @p: the process that the cpu time gets accounted to
3069 * @user_tick: indicates if the tick is a user or a system tick
3070 */
3071void account_process_tick(struct task_struct *p, int user_tick)
3072{
3073	cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3074	struct rq *rq = this_rq();
3075
3076	if (sched_clock_irqtime) {
3077		irqtime_account_process_tick(p, user_tick, rq);
3078		return;
3079	}
3080
3081	if (steal_account_process_tick())
3082		return;
3083
3084	if (user_tick)
3085		account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3086	else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3087		account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3088				    one_jiffy_scaled);
3089	else
3090		account_idle_time(cputime_one_jiffy);
3091}
3092
3093/*
3094 * Account multiple ticks of steal time.
3095 * @p: the process from which the cpu time has been stolen
3096 * @ticks: number of stolen ticks
3097 */
3098void account_steal_ticks(unsigned long ticks)
3099{
3100	account_steal_time(jiffies_to_cputime(ticks));
3101}
3102
3103/*
3104 * Account multiple ticks of idle time.
3105 * @ticks: number of stolen ticks
3106 */
3107void account_idle_ticks(unsigned long ticks)
3108{
3109
3110	if (sched_clock_irqtime) {
3111		irqtime_account_idle_ticks(ticks);
3112		return;
3113	}
3114
3115	account_idle_time(jiffies_to_cputime(ticks));
3116}
3117
3118#endif
3119
3120/*
3121 * Use precise platform statistics if available:
3122 */
3123#ifdef CONFIG_VIRT_CPU_ACCOUNTING
3124void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3125{
3126	*ut = p->utime;
3127	*st = p->stime;
3128}
3129
3130void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3131{
3132	struct task_cputime cputime;
3133
3134	thread_group_cputime(p, &cputime);
3135
3136	*ut = cputime.utime;
3137	*st = cputime.stime;
3138}
3139#else
3140
3141#ifndef nsecs_to_cputime
3142# define nsecs_to_cputime(__nsecs)	nsecs_to_jiffies(__nsecs)
3143#endif
3144
3145void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3146{
3147	cputime_t rtime, utime = p->utime, total = utime + p->stime;
3148
3149	/*
3150	 * Use CFS's precise accounting:
3151	 */
3152	rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3153
3154	if (total) {
3155		u64 temp = (__force u64) rtime;
3156
3157		temp *= (__force u64) utime;
3158		do_div(temp, (__force u32) total);
3159		utime = (__force cputime_t) temp;
3160	} else
3161		utime = rtime;
3162
3163	/*
3164	 * Compare with previous values, to keep monotonicity:
3165	 */
3166	p->prev_utime = max(p->prev_utime, utime);
3167	p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
3168
3169	*ut = p->prev_utime;
3170	*st = p->prev_stime;
3171}
3172
3173/*
3174 * Must be called with siglock held.
3175 */
3176void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3177{
3178	struct signal_struct *sig = p->signal;
3179	struct task_cputime cputime;
3180	cputime_t rtime, utime, total;
3181
3182	thread_group_cputime(p, &cputime);
3183
3184	total = cputime.utime + cputime.stime;
3185	rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3186
3187	if (total) {
3188		u64 temp = (__force u64) rtime;
3189
3190		temp *= (__force u64) cputime.utime;
3191		do_div(temp, (__force u32) total);
3192		utime = (__force cputime_t) temp;
3193	} else
3194		utime = rtime;
3195
3196	sig->prev_utime = max(sig->prev_utime, utime);
3197	sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3198
3199	*ut = sig->prev_utime;
3200	*st = sig->prev_stime;
3201}
3202#endif
3203
3204/*
3205 * This function gets called by the timer code, with HZ frequency.
3206 * We call it with interrupts disabled.
3207 */
3208void scheduler_tick(void)
3209{
3210	int cpu = smp_processor_id();
3211	struct rq *rq = cpu_rq(cpu);
3212	struct task_struct *curr = rq->curr;
3213
3214	sched_clock_tick();
3215
3216	raw_spin_lock(&rq->lock);
3217	update_rq_clock(rq);
3218	update_cpu_load_active(rq);
3219	curr->sched_class->task_tick(rq, curr, 0);
3220	raw_spin_unlock(&rq->lock);
3221
3222	perf_event_task_tick();
3223
3224#ifdef CONFIG_SMP
3225	rq->idle_balance = idle_cpu(cpu);
3226	trigger_load_balance(rq, cpu);
3227#endif
3228}
3229
3230notrace unsigned long get_parent_ip(unsigned long addr)
3231{
3232	if (in_lock_functions(addr)) {
3233		addr = CALLER_ADDR2;
3234		if (in_lock_functions(addr))
3235			addr = CALLER_ADDR3;
3236	}
3237	return addr;
3238}
3239
3240#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3241				defined(CONFIG_PREEMPT_TRACER))
3242
3243void __kprobes add_preempt_count(int val)
3244{
3245#ifdef CONFIG_DEBUG_PREEMPT
3246	/*
3247	 * Underflow?
3248	 */
3249	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3250		return;
3251#endif
3252	preempt_count() += val;
3253#ifdef CONFIG_DEBUG_PREEMPT
3254	/*
3255	 * Spinlock count overflowing soon?
3256	 */
3257	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3258				PREEMPT_MASK - 10);
3259#endif
3260	if (preempt_count() == val)
3261		trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3262}
3263EXPORT_SYMBOL(add_preempt_count);
3264
3265void __kprobes sub_preempt_count(int val)
3266{
3267#ifdef CONFIG_DEBUG_PREEMPT
3268	/*
3269	 * Underflow?
3270	 */
3271	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3272		return;
3273	/*
3274	 * Is the spinlock portion underflowing?
3275	 */
3276	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3277			!(preempt_count() & PREEMPT_MASK)))
3278		return;
3279#endif
3280
3281	if (preempt_count() == val)
3282		trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3283	preempt_count() -= val;
3284}
3285EXPORT_SYMBOL(sub_preempt_count);
3286
3287#endif
3288
3289/*
3290 * Print scheduling while atomic bug:
3291 */
3292static noinline void __schedule_bug(struct task_struct *prev)
3293{
3294	if (oops_in_progress)
3295		return;
3296
3297	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3298		prev->comm, prev->pid, preempt_count());
3299
3300	debug_show_held_locks(prev);
3301	print_modules();
3302	if (irqs_disabled())
3303		print_irqtrace_events(prev);
3304	dump_stack();
3305	add_taint(TAINT_WARN);
3306}
3307
3308/*
3309 * Various schedule()-time debugging checks and statistics:
3310 */
3311static inline void schedule_debug(struct task_struct *prev)
3312{
3313	/*
3314	 * Test if we are atomic. Since do_exit() needs to call into
3315	 * schedule() atomically, we ignore that path for now.
3316	 * Otherwise, whine if we are scheduling when we should not be.
3317	 */
3318	if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3319		__schedule_bug(prev);
3320	rcu_sleep_check();
3321
3322	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3323
3324	schedstat_inc(this_rq(), sched_count);
3325}
3326
3327static void put_prev_task(struct rq *rq, struct task_struct *prev)
3328{
3329	if (prev->on_rq || rq->skip_clock_update < 0)
3330		update_rq_clock(rq);
3331	prev->sched_class->put_prev_task(rq, prev);
3332}
3333
3334/*
3335 * Pick up the highest-prio task:
3336 */
3337static inline struct task_struct *
3338pick_next_task(struct rq *rq)
3339{
3340	const struct sched_class *class;
3341	struct task_struct *p;
3342
3343	/*
3344	 * Optimization: we know that if all tasks are in
3345	 * the fair class we can call that function directly:
3346	 */
3347	if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3348		p = fair_sched_class.pick_next_task(rq);
3349		if (likely(p))
3350			return p;
3351	}
3352
3353	for_each_class(class) {
3354		p = class->pick_next_task(rq);
3355		if (p)
3356			return p;
3357	}
3358
3359	BUG(); /* the idle class will always have a runnable task */
3360}
3361
3362/*
3363 * __schedule() is the main scheduler function.
3364 */
3365static void __sched __schedule(void)
3366{
3367	struct task_struct *prev, *next;
3368	unsigned long *switch_count;
3369	struct rq *rq;
3370	int cpu;
3371
3372need_resched:
3373	preempt_disable();
3374	cpu = smp_processor_id();
3375	rq = cpu_rq(cpu);
3376	rcu_note_context_switch(cpu);
3377	prev = rq->curr;
3378
3379	schedule_debug(prev);
3380
3381	if (sched_feat(HRTICK))
3382		hrtick_clear(rq);
3383
3384	raw_spin_lock_irq(&rq->lock);
3385
3386	switch_count = &prev->nivcsw;
3387	if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3388		if (unlikely(signal_pending_state(prev->state, prev))) {
3389			prev->state = TASK_RUNNING;
3390		} else {
3391			deactivate_task(rq, prev, DEQUEUE_SLEEP);
3392			prev->on_rq = 0;
3393
3394			/*
3395			 * If a worker went to sleep, notify and ask workqueue
3396			 * whether it wants to wake up a task to maintain
3397			 * concurrency.
3398			 */
3399			if (prev->flags & PF_WQ_WORKER) {
3400				struct task_struct *to_wakeup;
3401
3402				to_wakeup = wq_worker_sleeping(prev, cpu);
3403				if (to_wakeup)
3404					try_to_wake_up_local(to_wakeup);
3405			}
3406		}
3407		switch_count = &prev->nvcsw;
3408	}
3409
3410	pre_schedule(rq, prev);
3411
3412	if (unlikely(!rq->nr_running))
3413		idle_balance(cpu, rq);
3414
3415	put_prev_task(rq, prev);
3416	next = pick_next_task(rq);
3417	clear_tsk_need_resched(prev);
3418	rq->skip_clock_update = 0;
3419
3420	if (likely(prev != next)) {
3421		rq->nr_switches++;
3422		rq->curr = next;
3423		++*switch_count;
3424
3425		context_switch(rq, prev, next); /* unlocks the rq */
3426		/*
3427		 * The context switch have flipped the stack from under us
3428		 * and restored the local variables which were saved when
3429		 * this task called schedule() in the past. prev == current
3430		 * is still correct, but it can be moved to another cpu/rq.
3431		 */
3432		cpu = smp_processor_id();
3433		rq = cpu_rq(cpu);
3434	} else
3435		raw_spin_unlock_irq(&rq->lock);
3436
3437	post_schedule(rq);
3438
3439	sched_preempt_enable_no_resched();
3440	if (need_resched())
3441		goto need_resched;
3442}
3443
3444static inline void sched_submit_work(struct task_struct *tsk)
3445{
3446	if (!tsk->state || tsk_is_pi_blocked(tsk))
3447		return;
3448	/*
3449	 * If we are going to sleep and we have plugged IO queued,
3450	 * make sure to submit it to avoid deadlocks.
3451	 */
3452	if (blk_needs_flush_plug(tsk))
3453		blk_schedule_flush_plug(tsk);
3454}
3455
3456asmlinkage void __sched schedule(void)
3457{
3458	struct task_struct *tsk = current;
3459
3460	sched_submit_work(tsk);
3461	__schedule();
3462}
3463EXPORT_SYMBOL(schedule);
3464
3465/**
3466 * schedule_preempt_disabled - called with preemption disabled
3467 *
3468 * Returns with preemption disabled. Note: preempt_count must be 1
3469 */
3470void __sched schedule_preempt_disabled(void)
3471{
3472	sched_preempt_enable_no_resched();
3473	schedule();
3474	preempt_disable();
3475}
3476
3477#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3478
3479static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3480{
3481	if (lock->owner != owner)
3482		return false;
3483
3484	/*
3485	 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3486	 * lock->owner still matches owner, if that fails, owner might
3487	 * point to free()d memory, if it still matches, the rcu_read_lock()
3488	 * ensures the memory stays valid.
3489	 */
3490	barrier();
3491
3492	return owner->on_cpu;
3493}
3494
3495/*
3496 * Look out! "owner" is an entirely speculative pointer
3497 * access and not reliable.
3498 */
3499int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3500{
3501	if (!sched_feat(OWNER_SPIN))
3502		return 0;
3503
3504	rcu_read_lock();
3505	while (owner_running(lock, owner)) {
3506		if (need_resched())
3507			break;
3508
3509		arch_mutex_cpu_relax();
3510	}
3511	rcu_read_unlock();
3512
3513	/*
3514	 * We break out the loop above on need_resched() and when the
3515	 * owner changed, which is a sign for heavy contention. Return
3516	 * success only when lock->owner is NULL.
3517	 */
3518	return lock->owner == NULL;
3519}
3520#endif
3521
3522#ifdef CONFIG_PREEMPT
3523/*
3524 * this is the entry point to schedule() from in-kernel preemption
3525 * off of preempt_enable. Kernel preemptions off return from interrupt
3526 * occur there and call schedule directly.
3527 */
3528asmlinkage void __sched notrace preempt_schedule(void)
3529{
3530	struct thread_info *ti = current_thread_info();
3531
3532	/*
3533	 * If there is a non-zero preempt_count or interrupts are disabled,
3534	 * we do not want to preempt the current task. Just return..
3535	 */
3536	if (likely(ti->preempt_count || irqs_disabled()))
3537		return;
3538
3539	do {
3540		add_preempt_count_notrace(PREEMPT_ACTIVE);
3541		__schedule();
3542		sub_preempt_count_notrace(PREEMPT_ACTIVE);
3543
3544		/*
3545		 * Check again in case we missed a preemption opportunity
3546		 * between schedule and now.
3547		 */
3548		barrier();
3549	} while (need_resched());
3550}
3551EXPORT_SYMBOL(preempt_schedule);
3552
3553/*
3554 * this is the entry point to schedule() from kernel preemption
3555 * off of irq context.
3556 * Note, that this is called and return with irqs disabled. This will
3557 * protect us against recursive calling from irq.
3558 */
3559asmlinkage void __sched preempt_schedule_irq(void)
3560{
3561	struct thread_info *ti = current_thread_info();
3562
3563	/* Catch callers which need to be fixed */
3564	BUG_ON(ti->preempt_count || !irqs_disabled());
3565
3566	do {
3567		add_preempt_count(PREEMPT_ACTIVE);
3568		local_irq_enable();
3569		__schedule();
3570		local_irq_disable();
3571		sub_preempt_count(PREEMPT_ACTIVE);
3572
3573		/*
3574		 * Check again in case we missed a preemption opportunity
3575		 * between schedule and now.
3576		 */
3577		barrier();
3578	} while (need_resched());
3579}
3580
3581#endif /* CONFIG_PREEMPT */
3582
3583int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3584			  void *key)
3585{
3586	return try_to_wake_up(curr->private, mode, wake_flags);
3587}
3588EXPORT_SYMBOL(default_wake_function);
3589
3590/*
3591 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3592 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3593 * number) then we wake all the non-exclusive tasks and one exclusive task.
3594 *
3595 * There are circumstances in which we can try to wake a task which has already
3596 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3597 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3598 */
3599static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3600			int nr_exclusive, int wake_flags, void *key)
3601{
3602	wait_queue_t *curr, *next;
3603
3604	list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3605		unsigned flags = curr->flags;
3606
3607		if (curr->func(curr, mode, wake_flags, key) &&
3608				(flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3609			break;
3610	}
3611}
3612
3613/**
3614 * __wake_up - wake up threads blocked on a waitqueue.
3615 * @q: the waitqueue
3616 * @mode: which threads
3617 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3618 * @key: is directly passed to the wakeup function
3619 *
3620 * It may be assumed that this function implies a write memory barrier before
3621 * changing the task state if and only if any tasks are woken up.
3622 */
3623void __wake_up(wait_queue_head_t *q, unsigned int mode,
3624			int nr_exclusive, void *key)
3625{
3626	unsigned long flags;
3627
3628	spin_lock_irqsave(&q->lock, flags);
3629	__wake_up_common(q, mode, nr_exclusive, 0, key);
3630	spin_unlock_irqrestore(&q->lock, flags);
3631}
3632EXPORT_SYMBOL(__wake_up);
3633
3634/*
3635 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3636 */
3637void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3638{
3639	__wake_up_common(q, mode, nr, 0, NULL);
3640}
3641EXPORT_SYMBOL_GPL(__wake_up_locked);
3642
3643void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3644{
3645	__wake_up_common(q, mode, 1, 0, key);
3646}
3647EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3648
3649/**
3650 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3651 * @q: the waitqueue
3652 * @mode: which threads
3653 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3654 * @key: opaque value to be passed to wakeup targets
3655 *
3656 * The sync wakeup differs that the waker knows that it will schedule
3657 * away soon, so while the target thread will be woken up, it will not
3658 * be migrated to another CPU - ie. the two threads are 'synchronized'
3659 * with each other. This can prevent needless bouncing between CPUs.
3660 *
3661 * On UP it can prevent extra preemption.
3662 *
3663 * It may be assumed that this function implies a write memory barrier before
3664 * changing the task state if and only if any tasks are woken up.
3665 */
3666void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3667			int nr_exclusive, void *key)
3668{
3669	unsigned long flags;
3670	int wake_flags = WF_SYNC;
3671
3672	if (unlikely(!q))
3673		return;
3674
3675	if (unlikely(!nr_exclusive))
3676		wake_flags = 0;
3677
3678	spin_lock_irqsave(&q->lock, flags);
3679	__wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3680	spin_unlock_irqrestore(&q->lock, flags);
3681}
3682EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3683
3684/*
3685 * __wake_up_sync - see __wake_up_sync_key()
3686 */
3687void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3688{
3689	__wake_up_sync_key(q, mode, nr_exclusive, NULL);
3690}
3691EXPORT_SYMBOL_GPL(__wake_up_sync);	/* For internal use only */
3692
3693/**
3694 * complete: - signals a single thread waiting on this completion
3695 * @x:  holds the state of this particular completion
3696 *
3697 * This will wake up a single thread waiting on this completion. Threads will be
3698 * awakened in the same order in which they were queued.
3699 *
3700 * See also complete_all(), wait_for_completion() and related routines.
3701 *
3702 * It may be assumed that this function implies a write memory barrier before
3703 * changing the task state if and only if any tasks are woken up.
3704 */
3705void complete(struct completion *x)
3706{
3707	unsigned long flags;
3708
3709	spin_lock_irqsave(&x->wait.lock, flags);
3710	x->done++;
3711	__wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3712	spin_unlock_irqrestore(&x->wait.lock, flags);
3713}
3714EXPORT_SYMBOL(complete);
3715
3716/**
3717 * complete_all: - signals all threads waiting on this completion
3718 * @x:  holds the state of this particular completion
3719 *
3720 * This will wake up all threads waiting on this particular completion event.
3721 *
3722 * It may be assumed that this function implies a write memory barrier before
3723 * changing the task state if and only if any tasks are woken up.
3724 */
3725void complete_all(struct completion *x)
3726{
3727	unsigned long flags;
3728
3729	spin_lock_irqsave(&x->wait.lock, flags);
3730	x->done += UINT_MAX/2;
3731	__wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3732	spin_unlock_irqrestore(&x->wait.lock, flags);
3733}
3734EXPORT_SYMBOL(complete_all);
3735
3736static inline long __sched
3737do_wait_for_common(struct completion *x, long timeout, int state)
3738{
3739	if (!x->done) {
3740		DECLARE_WAITQUEUE(wait, current);
3741
3742		__add_wait_queue_tail_exclusive(&x->wait, &wait);
3743		do {
3744			if (signal_pending_state(state, current)) {
3745				timeout = -ERESTARTSYS;
3746				break;
3747			}
3748			__set_current_state(state);
3749			spin_unlock_irq(&x->wait.lock);
3750			timeout = schedule_timeout(timeout);
3751			spin_lock_irq(&x->wait.lock);
3752		} while (!x->done && timeout);
3753		__remove_wait_queue(&x->wait, &wait);
3754		if (!x->done)
3755			return timeout;
3756	}
3757	x->done--;
3758	return timeout ?: 1;
3759}
3760
3761static long __sched
3762wait_for_common(struct completion *x, long timeout, int state)
3763{
3764	might_sleep();
3765
3766	spin_lock_irq(&x->wait.lock);
3767	timeout = do_wait_for_common(x, timeout, state);
3768	spin_unlock_irq(&x->wait.lock);
3769	return timeout;
3770}
3771
3772/**
3773 * wait_for_completion: - waits for completion of a task
3774 * @x:  holds the state of this particular completion
3775 *
3776 * This waits to be signaled for completion of a specific task. It is NOT
3777 * interruptible and there is no timeout.
3778 *
3779 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3780 * and interrupt capability. Also see complete().
3781 */
3782void __sched wait_for_completion(struct completion *x)
3783{
3784	wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3785}
3786EXPORT_SYMBOL(wait_for_completion);
3787
3788/**
3789 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3790 * @x:  holds the state of this particular completion
3791 * @timeout:  timeout value in jiffies
3792 *
3793 * This waits for either a completion of a specific task to be signaled or for a
3794 * specified timeout to expire. The timeout is in jiffies. It is not
3795 * interruptible.
3796 *
3797 * The return value is 0 if timed out, and positive (at least 1, or number of
3798 * jiffies left till timeout) if completed.
3799 */
3800unsigned long __sched
3801wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3802{
3803	return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3804}
3805EXPORT_SYMBOL(wait_for_completion_timeout);
3806
3807/**
3808 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3809 * @x:  holds the state of this particular completion
3810 *
3811 * This waits for completion of a specific task to be signaled. It is
3812 * interruptible.
3813 *
3814 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3815 */
3816int __sched wait_for_completion_interruptible(struct completion *x)
3817{
3818	long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3819	if (t == -ERESTARTSYS)
3820		return t;
3821	return 0;
3822}
3823EXPORT_SYMBOL(wait_for_completion_interruptible);
3824
3825/**
3826 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3827 * @x:  holds the state of this particular completion
3828 * @timeout:  timeout value in jiffies
3829 *
3830 * This waits for either a completion of a specific task to be signaled or for a
3831 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3832 *
3833 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3834 * positive (at least 1, or number of jiffies left till timeout) if completed.
3835 */
3836long __sched
3837wait_for_completion_interruptible_timeout(struct completion *x,
3838					  unsigned long timeout)
3839{
3840	return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3841}
3842EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3843
3844/**
3845 * wait_for_completion_killable: - waits for completion of a task (killable)
3846 * @x:  holds the state of this particular completion
3847 *
3848 * This waits to be signaled for completion of a specific task. It can be
3849 * interrupted by a kill signal.
3850 *
3851 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3852 */
3853int __sched wait_for_completion_killable(struct completion *x)
3854{
3855	long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3856	if (t == -ERESTARTSYS)
3857		return t;
3858	return 0;
3859}
3860EXPORT_SYMBOL(wait_for_completion_killable);
3861
3862/**
3863 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3864 * @x:  holds the state of this particular completion
3865 * @timeout:  timeout value in jiffies
3866 *
3867 * This waits for either a completion of a specific task to be
3868 * signaled or for a specified timeout to expire. It can be
3869 * interrupted by a kill signal. The timeout is in jiffies.
3870 *
3871 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3872 * positive (at least 1, or number of jiffies left till timeout) if completed.
3873 */
3874long __sched
3875wait_for_completion_killable_timeout(struct completion *x,
3876				     unsigned long timeout)
3877{
3878	return wait_for_common(x, timeout, TASK_KILLABLE);
3879}
3880EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3881
3882/**
3883 *	try_wait_for_completion - try to decrement a completion without blocking
3884 *	@x:	completion structure
3885 *
3886 *	Returns: 0 if a decrement cannot be done without blocking
3887 *		 1 if a decrement succeeded.
3888 *
3889 *	If a completion is being used as a counting completion,
3890 *	attempt to decrement the counter without blocking. This
3891 *	enables us to avoid waiting if the resource the completion
3892 *	is protecting is not available.
3893 */
3894bool try_wait_for_completion(struct completion *x)
3895{
3896	unsigned long flags;
3897	int ret = 1;
3898
3899	spin_lock_irqsave(&x->wait.lock, flags);
3900	if (!x->done)
3901		ret = 0;
3902	else
3903		x->done--;
3904	spin_unlock_irqrestore(&x->wait.lock, flags);
3905	return ret;
3906}
3907EXPORT_SYMBOL(try_wait_for_completion);
3908
3909/**
3910 *	completion_done - Test to see if a completion has any waiters
3911 *	@x:	completion structure
3912 *
3913 *	Returns: 0 if there are waiters (wait_for_completion() in progress)
3914 *		 1 if there are no waiters.
3915 *
3916 */
3917bool completion_done(struct completion *x)
3918{
3919	unsigned long flags;
3920	int ret = 1;
3921
3922	spin_lock_irqsave(&x->wait.lock, flags);
3923	if (!x->done)
3924		ret = 0;
3925	spin_unlock_irqrestore(&x->wait.lock, flags);
3926	return ret;
3927}
3928EXPORT_SYMBOL(completion_done);
3929
3930static long __sched
3931sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3932{
3933	unsigned long flags;
3934	wait_queue_t wait;
3935
3936	init_waitqueue_entry(&wait, current);
3937
3938	__set_current_state(state);
3939
3940	spin_lock_irqsave(&q->lock, flags);
3941	__add_wait_queue(q, &wait);
3942	spin_unlock(&q->lock);
3943	timeout = schedule_timeout(timeout);
3944	spin_lock_irq(&q->lock);
3945	__remove_wait_queue(q, &wait);
3946	spin_unlock_irqrestore(&q->lock, flags);
3947
3948	return timeout;
3949}
3950
3951void __sched interruptible_sleep_on(wait_queue_head_t *q)
3952{
3953	sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3954}
3955EXPORT_SYMBOL(interruptible_sleep_on);
3956
3957long __sched
3958interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3959{
3960	return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3961}
3962EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3963
3964void __sched sleep_on(wait_queue_head_t *q)
3965{
3966	sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3967}
3968EXPORT_SYMBOL(sleep_on);
3969
3970long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3971{
3972	return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3973}
3974EXPORT_SYMBOL(sleep_on_timeout);
3975
3976#ifdef CONFIG_RT_MUTEXES
3977
3978/*
3979 * rt_mutex_setprio - set the current priority of a task
3980 * @p: task
3981 * @prio: prio value (kernel-internal form)
3982 *
3983 * This function changes the 'effective' priority of a task. It does
3984 * not touch ->normal_prio like __setscheduler().
3985 *
3986 * Used by the rt_mutex code to implement priority inheritance logic.
3987 */
3988void rt_mutex_setprio(struct task_struct *p, int prio)
3989{
3990	int oldprio, on_rq, running;
3991	struct rq *rq;
3992	const struct sched_class *prev_class;
3993
3994	BUG_ON(prio < 0 || prio > MAX_PRIO);
3995
3996	rq = __task_rq_lock(p);
3997
3998	/*
3999	 * Idle task boosting is a nono in general. There is one
4000	 * exception, when PREEMPT_RT and NOHZ is active:
4001	 *
4002	 * The idle task calls get_next_timer_interrupt() and holds
4003	 * the timer wheel base->lock on the CPU and another CPU wants
4004	 * to access the timer (probably to cancel it). We can safely
4005	 * ignore the boosting request, as the idle CPU runs this code
4006	 * with interrupts disabled and will complete the lock
4007	 * protected section without being interrupted. So there is no
4008	 * real need to boost.
4009	 */
4010	if (unlikely(p == rq->idle)) {
4011		WARN_ON(p != rq->curr);
4012		WARN_ON(p->pi_blocked_on);
4013		goto out_unlock;
4014	}
4015
4016	trace_sched_pi_setprio(p, prio);
4017	oldprio = p->prio;
4018	prev_class = p->sched_class;
4019	on_rq = p->on_rq;
4020	running = task_current(rq, p);
4021	if (on_rq)
4022		dequeue_task(rq, p, 0);
4023	if (running)
4024		p->sched_class->put_prev_task(rq, p);
4025
4026	if (rt_prio(prio))
4027		p->sched_class = &rt_sched_class;
4028	else
4029		p->sched_class = &fair_sched_class;
4030
4031	p->prio = prio;
4032
4033	if (running)
4034		p->sched_class->set_curr_task(rq);
4035	if (on_rq)
4036		enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4037
4038	check_class_changed(rq, p, prev_class, oldprio);
4039out_unlock:
4040	__task_rq_unlock(rq);
4041}
4042#endif
4043void set_user_nice(struct task_struct *p, long nice)
4044{
4045	int old_prio, delta, on_rq;
4046	unsigned long flags;
4047	struct rq *rq;
4048
4049	if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4050		return;
4051	/*
4052	 * We have to be careful, if called from sys_setpriority(),
4053	 * the task might be in the middle of scheduling on another CPU.
4054	 */
4055	rq = task_rq_lock(p, &flags);
4056	/*
4057	 * The RT priorities are set via sched_setscheduler(), but we still
4058	 * allow the 'normal' nice value to be set - but as expected
4059	 * it wont have any effect on scheduling until the task is
4060	 * SCHED_FIFO/SCHED_RR:
4061	 */
4062	if (task_has_rt_policy(p)) {
4063		p->static_prio = NICE_TO_PRIO(nice);
4064		goto out_unlock;
4065	}
4066	on_rq = p->on_rq;
4067	if (on_rq)
4068		dequeue_task(rq, p, 0);
4069
4070	p->static_prio = NICE_TO_PRIO(nice);
4071	set_load_weight(p);
4072	old_prio = p->prio;
4073	p->prio = effective_prio(p);
4074	delta = p->prio - old_prio;
4075
4076	if (on_rq) {
4077		enqueue_task(rq, p, 0);
4078		/*
4079		 * If the task increased its priority or is running and
4080		 * lowered its priority, then reschedule its CPU:
4081		 */
4082		if (delta < 0 || (delta > 0 && task_running(rq, p)))
4083			resched_task(rq->curr);
4084	}
4085out_unlock:
4086	task_rq_unlock(rq, p, &flags);
4087}
4088EXPORT_SYMBOL(set_user_nice);
4089
4090/*
4091 * can_nice - check if a task can reduce its nice value
4092 * @p: task
4093 * @nice: nice value
4094 */
4095int can_nice(const struct task_struct *p, const int nice)
4096{
4097	/* convert nice value [19,-20] to rlimit style value [1,40] */
4098	int nice_rlim = 20 - nice;
4099
4100	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4101		capable(CAP_SYS_NICE));
4102}
4103
4104#ifdef __ARCH_WANT_SYS_NICE
4105
4106/*
4107 * sys_nice - change the priority of the current process.
4108 * @increment: priority increment
4109 *
4110 * sys_setpriority is a more generic, but much slower function that
4111 * does similar things.
4112 */
4113SYSCALL_DEFINE1(nice, int, increment)
4114{
4115	long nice, retval;
4116
4117	/*
4118	 * Setpriority might change our priority at the same moment.
4119	 * We don't have to worry. Conceptually one call occurs first
4120	 * and we have a single winner.
4121	 */
4122	if (increment < -40)
4123		increment = -40;
4124	if (increment > 40)
4125		increment = 40;
4126
4127	nice = TASK_NICE(current) + increment;
4128	if (nice < -20)
4129		nice = -20;
4130	if (nice > 19)
4131		nice = 19;
4132
4133	if (increment < 0 && !can_nice(current, nice))
4134		return -EPERM;
4135
4136	retval = security_task_setnice(current, nice);
4137	if (retval)
4138		return retval;
4139
4140	set_user_nice(current, nice);
4141	return 0;
4142}
4143
4144#endif
4145
4146/**
4147 * task_prio - return the priority value of a given task.
4148 * @p: the task in question.
4149 *
4150 * This is the priority value as seen by users in /proc.
4151 * RT tasks are offset by -200. Normal tasks are centered
4152 * around 0, value goes from -16 to +15.
4153 */
4154int task_prio(const struct task_struct *p)
4155{
4156	return p->prio - MAX_RT_PRIO;
4157}
4158
4159/**
4160 * task_nice - return the nice value of a given task.
4161 * @p: the task in question.
4162 */
4163int task_nice(const struct task_struct *p)
4164{
4165	return TASK_NICE(p);
4166}
4167EXPORT_SYMBOL(task_nice);
4168
4169/**
4170 * idle_cpu - is a given cpu idle currently?
4171 * @cpu: the processor in question.
4172 */
4173int idle_cpu(int cpu)
4174{
4175	struct rq *rq = cpu_rq(cpu);
4176
4177	if (rq->curr != rq->idle)
4178		return 0;
4179
4180	if (rq->nr_running)
4181		return 0;
4182
4183#ifdef CONFIG_SMP
4184	if (!llist_empty(&rq->wake_list))
4185		return 0;
4186#endif
4187
4188	return 1;
4189}
4190
4191/**
4192 * idle_task - return the idle task for a given cpu.
4193 * @cpu: the processor in question.
4194 */
4195struct task_struct *idle_task(int cpu)
4196{
4197	return cpu_rq(cpu)->idle;
4198}
4199
4200/**
4201 * find_process_by_pid - find a process with a matching PID value.
4202 * @pid: the pid in question.
4203 */
4204static struct task_struct *find_process_by_pid(pid_t pid)
4205{
4206	return pid ? find_task_by_vpid(pid) : current;
4207}
4208
4209/* Actually do priority change: must hold rq lock. */
4210static void
4211__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4212{
4213	p->policy = policy;
4214	p->rt_priority = prio;
4215	p->normal_prio = normal_prio(p);
4216	/* we are holding p->pi_lock already */
4217	p->prio = rt_mutex_getprio(p);
4218	if (rt_prio(p->prio))
4219		p->sched_class = &rt_sched_class;
4220	else
4221		p->sched_class = &fair_sched_class;
4222	set_load_weight(p);
4223}
4224
4225/*
4226 * check the target process has a UID that matches the current process's
4227 */
4228static bool check_same_owner(struct task_struct *p)
4229{
4230	const struct cred *cred = current_cred(), *pcred;
4231	bool match;
4232
4233	rcu_read_lock();
4234	pcred = __task_cred(p);
4235	match = (uid_eq(cred->euid, pcred->euid) ||
4236		 uid_eq(cred->euid, pcred->uid));
4237	rcu_read_unlock();
4238	return match;
4239}
4240
4241static int __sched_setscheduler(struct task_struct *p, int policy,
4242				const struct sched_param *param, bool user)
4243{
4244	int retval, oldprio, oldpolicy = -1, on_rq, running;
4245	unsigned long flags;
4246	const struct sched_class *prev_class;
4247	struct rq *rq;
4248	int reset_on_fork;
4249
4250	/* may grab non-irq protected spin_locks */
4251	BUG_ON(in_interrupt());
4252recheck:
4253	/* double check policy once rq lock held */
4254	if (policy < 0) {
4255		reset_on_fork = p->sched_reset_on_fork;
4256		policy = oldpolicy = p->policy;
4257	} else {
4258		reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4259		policy &= ~SCHED_RESET_ON_FORK;
4260
4261		if (policy != SCHED_FIFO && policy != SCHED_RR &&
4262				policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4263				policy != SCHED_IDLE)
4264			return -EINVAL;
4265	}
4266
4267	/*
4268	 * Valid priorities for SCHED_FIFO and SCHED_RR are
4269	 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4270	 * SCHED_BATCH and SCHED_IDLE is 0.
4271	 */
4272	if (param->sched_priority < 0 ||
4273	    (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4274	    (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4275		return -EINVAL;
4276	if (rt_policy(policy) != (param->sched_priority != 0))
4277		return -EINVAL;
4278
4279	/*
4280	 * Allow unprivileged RT tasks to decrease priority:
4281	 */
4282	if (user && !capable(CAP_SYS_NICE)) {
4283		if (rt_policy(policy)) {
4284			unsigned long rlim_rtprio =
4285					task_rlimit(p, RLIMIT_RTPRIO);
4286
4287			/* can't set/change the rt policy */
4288			if (policy != p->policy && !rlim_rtprio)
4289				return -EPERM;
4290
4291			/* can't increase priority */
4292			if (param->sched_priority > p->rt_priority &&
4293			    param->sched_priority > rlim_rtprio)
4294				return -EPERM;
4295		}
4296
4297		/*
4298		 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4299		 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4300		 */
4301		if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4302			if (!can_nice(p, TASK_NICE(p)))
4303				return -EPERM;
4304		}
4305
4306		/* can't change other user's priorities */
4307		if (!check_same_owner(p))
4308			return -EPERM;
4309
4310		/* Normal users shall not reset the sched_reset_on_fork flag */
4311		if (p->sched_reset_on_fork && !reset_on_fork)
4312			return -EPERM;
4313	}
4314
4315	if (user) {
4316		retval = security_task_setscheduler(p);
4317		if (retval)
4318			return retval;
4319	}
4320
4321	/*
4322	 * make sure no PI-waiters arrive (or leave) while we are
4323	 * changing the priority of the task:
4324	 *
4325	 * To be able to change p->policy safely, the appropriate
4326	 * runqueue lock must be held.
4327	 */
4328	rq = task_rq_lock(p, &flags);
4329
4330	/*
4331	 * Changing the policy of the stop threads its a very bad idea
4332	 */
4333	if (p == rq->stop) {
4334		task_rq_unlock(rq, p, &flags);
4335		return -EINVAL;
4336	}
4337
4338	/*
4339	 * If not changing anything there's no need to proceed further:
4340	 */
4341	if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4342			param->sched_priority == p->rt_priority))) {
4343
4344		__task_rq_unlock(rq);
4345		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4346		return 0;
4347	}
4348
4349#ifdef CONFIG_RT_GROUP_SCHED
4350	if (user) {
4351		/*
4352		 * Do not allow realtime tasks into groups that have no runtime
4353		 * assigned.
4354		 */
4355		if (rt_bandwidth_enabled() && rt_policy(policy) &&
4356				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4357				!task_group_is_autogroup(task_group(p))) {
4358			task_rq_unlock(rq, p, &flags);
4359			return -EPERM;
4360		}
4361	}
4362#endif
4363
4364	/* recheck policy now with rq lock held */
4365	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4366		policy = oldpolicy = -1;
4367		task_rq_unlock(rq, p, &flags);
4368		goto recheck;
4369	}
4370	on_rq = p->on_rq;
4371	running = task_current(rq, p);
4372	if (on_rq)
4373		dequeue_task(rq, p, 0);
4374	if (running)
4375		p->sched_class->put_prev_task(rq, p);
4376
4377	p->sched_reset_on_fork = reset_on_fork;
4378
4379	oldprio = p->prio;
4380	prev_class = p->sched_class;
4381	__setscheduler(rq, p, policy, param->sched_priority);
4382
4383	if (running)
4384		p->sched_class->set_curr_task(rq);
4385	if (on_rq)
4386		enqueue_task(rq, p, 0);
4387
4388	check_class_changed(rq, p, prev_class, oldprio);
4389	task_rq_unlock(rq, p, &flags);
4390
4391	rt_mutex_adjust_pi(p);
4392
4393	return 0;
4394}
4395
4396/**
4397 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4398 * @p: the task in question.
4399 * @policy: new policy.
4400 * @param: structure containing the new RT priority.
4401 *
4402 * NOTE that the task may be already dead.
4403 */
4404int sched_setscheduler(struct task_struct *p, int policy,
4405		       const struct sched_param *param)
4406{
4407	return __sched_setscheduler(p, policy, param, true);
4408}
4409EXPORT_SYMBOL_GPL(sched_setscheduler);
4410
4411/**
4412 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4413 * @p: the task in question.
4414 * @policy: new policy.
4415 * @param: structure containing the new RT priority.
4416 *
4417 * Just like sched_setscheduler, only don't bother checking if the
4418 * current context has permission.  For example, this is needed in
4419 * stop_machine(): we create temporary high priority worker threads,
4420 * but our caller might not have that capability.
4421 */
4422int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4423			       const struct sched_param *param)
4424{
4425	return __sched_setscheduler(p, policy, param, false);
4426}
4427
4428static int
4429do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4430{
4431	struct sched_param lparam;
4432	struct task_struct *p;
4433	int retval;
4434
4435	if (!param || pid < 0)
4436		return -EINVAL;
4437	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4438		return -EFAULT;
4439
4440	rcu_read_lock();
4441	retval = -ESRCH;
4442	p = find_process_by_pid(pid);
4443	if (p != NULL)
4444		retval = sched_setscheduler(p, policy, &lparam);
4445	rcu_read_unlock();
4446
4447	return retval;
4448}
4449
4450/**
4451 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4452 * @pid: the pid in question.
4453 * @policy: new policy.
4454 * @param: structure containing the new RT priority.
4455 */
4456SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4457		struct sched_param __user *, param)
4458{
4459	/* negative values for policy are not valid */
4460	if (policy < 0)
4461		return -EINVAL;
4462
4463	return do_sched_setscheduler(pid, policy, param);
4464}
4465
4466/**
4467 * sys_sched_setparam - set/change the RT priority of a thread
4468 * @pid: the pid in question.
4469 * @param: structure containing the new RT priority.
4470 */
4471SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4472{
4473	return do_sched_setscheduler(pid, -1, param);
4474}
4475
4476/**
4477 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4478 * @pid: the pid in question.
4479 */
4480SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4481{
4482	struct task_struct *p;
4483	int retval;
4484
4485	if (pid < 0)
4486		return -EINVAL;
4487
4488	retval = -ESRCH;
4489	rcu_read_lock();
4490	p = find_process_by_pid(pid);
4491	if (p) {
4492		retval = security_task_getscheduler(p);
4493		if (!retval)
4494			retval = p->policy
4495				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4496	}
4497	rcu_read_unlock();
4498	return retval;
4499}
4500
4501/**
4502 * sys_sched_getparam - get the RT priority of a thread
4503 * @pid: the pid in question.
4504 * @param: structure containing the RT priority.
4505 */
4506SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4507{
4508	struct sched_param lp;
4509	struct task_struct *p;
4510	int retval;
4511
4512	if (!param || pid < 0)
4513		return -EINVAL;
4514
4515	rcu_read_lock();
4516	p = find_process_by_pid(pid);
4517	retval = -ESRCH;
4518	if (!p)
4519		goto out_unlock;
4520
4521	retval = security_task_getscheduler(p);
4522	if (retval)
4523		goto out_unlock;
4524
4525	lp.sched_priority = p->rt_priority;
4526	rcu_read_unlock();
4527
4528	/*
4529	 * This one might sleep, we cannot do it with a spinlock held ...
4530	 */
4531	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4532
4533	return retval;
4534
4535out_unlock:
4536	rcu_read_unlock();
4537	return retval;
4538}
4539
4540long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4541{
4542	cpumask_var_t cpus_allowed, new_mask;
4543	struct task_struct *p;
4544	int retval;
4545
4546	get_online_cpus();
4547	rcu_read_lock();
4548
4549	p = find_process_by_pid(pid);
4550	if (!p) {
4551		rcu_read_unlock();
4552		put_online_cpus();
4553		return -ESRCH;
4554	}
4555
4556	/* Prevent p going away */
4557	get_task_struct(p);
4558	rcu_read_unlock();
4559
4560	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4561		retval = -ENOMEM;
4562		goto out_put_task;
4563	}
4564	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4565		retval = -ENOMEM;
4566		goto out_free_cpus_allowed;
4567	}
4568	retval = -EPERM;
4569	if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4570		goto out_unlock;
4571
4572	retval = security_task_setscheduler(p);
4573	if (retval)
4574		goto out_unlock;
4575
4576	cpuset_cpus_allowed(p, cpus_allowed);
4577	cpumask_and(new_mask, in_mask, cpus_allowed);
4578again:
4579	retval = set_cpus_allowed_ptr(p, new_mask);
4580
4581	if (!retval) {
4582		cpuset_cpus_allowed(p, cpus_allowed);
4583		if (!cpumask_subset(new_mask, cpus_allowed)) {
4584			/*
4585			 * We must have raced with a concurrent cpuset
4586			 * update. Just reset the cpus_allowed to the
4587			 * cpuset's cpus_allowed
4588			 */
4589			cpumask_copy(new_mask, cpus_allowed);
4590			goto again;
4591		}
4592	}
4593out_unlock:
4594	free_cpumask_var(new_mask);
4595out_free_cpus_allowed:
4596	free_cpumask_var(cpus_allowed);
4597out_put_task:
4598	put_task_struct(p);
4599	put_online_cpus();
4600	return retval;
4601}
4602
4603static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4604			     struct cpumask *new_mask)
4605{
4606	if (len < cpumask_size())
4607		cpumask_clear(new_mask);
4608	else if (len > cpumask_size())
4609		len = cpumask_size();
4610
4611	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4612}
4613
4614/**
4615 * sys_sched_setaffinity - set the cpu affinity of a process
4616 * @pid: pid of the process
4617 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4618 * @user_mask_ptr: user-space pointer to the new cpu mask
4619 */
4620SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4621		unsigned long __user *, user_mask_ptr)
4622{
4623	cpumask_var_t new_mask;
4624	int retval;
4625
4626	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4627		return -ENOMEM;
4628
4629	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4630	if (retval == 0)
4631		retval = sched_setaffinity(pid, new_mask);
4632	free_cpumask_var(new_mask);
4633	return retval;
4634}
4635
4636long sched_getaffinity(pid_t pid, struct cpumask *mask)
4637{
4638	struct task_struct *p;
4639	unsigned long flags;
4640	int retval;
4641
4642	get_online_cpus();
4643	rcu_read_lock();
4644
4645	retval = -ESRCH;
4646	p = find_process_by_pid(pid);
4647	if (!p)
4648		goto out_unlock;
4649
4650	retval = security_task_getscheduler(p);
4651	if (retval)
4652		goto out_unlock;
4653
4654	raw_spin_lock_irqsave(&p->pi_lock, flags);
4655	cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4656	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4657
4658out_unlock:
4659	rcu_read_unlock();
4660	put_online_cpus();
4661
4662	return retval;
4663}
4664
4665/**
4666 * sys_sched_getaffinity - get the cpu affinity of a process
4667 * @pid: pid of the process
4668 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4669 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4670 */
4671SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4672		unsigned long __user *, user_mask_ptr)
4673{
4674	int ret;
4675	cpumask_var_t mask;
4676
4677	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4678		return -EINVAL;
4679	if (len & (sizeof(unsigned long)-1))
4680		return -EINVAL;
4681
4682	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4683		return -ENOMEM;
4684
4685	ret = sched_getaffinity(pid, mask);
4686	if (ret == 0) {
4687		size_t retlen = min_t(size_t, len, cpumask_size());
4688
4689		if (copy_to_user(user_mask_ptr, mask, retlen))
4690			ret = -EFAULT;
4691		else
4692			ret = retlen;
4693	}
4694	free_cpumask_var(mask);
4695
4696	return ret;
4697}
4698
4699/**
4700 * sys_sched_yield - yield the current processor to other threads.
4701 *
4702 * This function yields the current CPU to other tasks. If there are no
4703 * other threads running on this CPU then this function will return.
4704 */
4705SYSCALL_DEFINE0(sched_yield)
4706{
4707	struct rq *rq = this_rq_lock();
4708
4709	schedstat_inc(rq, yld_count);
4710	current->sched_class->yield_task(rq);
4711
4712	/*
4713	 * Since we are going to call schedule() anyway, there's
4714	 * no need to preempt or enable interrupts:
4715	 */
4716	__release(rq->lock);
4717	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4718	do_raw_spin_unlock(&rq->lock);
4719	sched_preempt_enable_no_resched();
4720
4721	schedule();
4722
4723	return 0;
4724}
4725
4726static inline int should_resched(void)
4727{
4728	return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4729}
4730
4731static void __cond_resched(void)
4732{
4733	add_preempt_count(PREEMPT_ACTIVE);
4734	__schedule();
4735	sub_preempt_count(PREEMPT_ACTIVE);
4736}
4737
4738int __sched _cond_resched(void)
4739{
4740	if (should_resched()) {
4741		__cond_resched();
4742		return 1;
4743	}
4744	return 0;
4745}
4746EXPORT_SYMBOL(_cond_resched);
4747
4748/*
4749 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4750 * call schedule, and on return reacquire the lock.
4751 *
4752 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4753 * operations here to prevent schedule() from being called twice (once via
4754 * spin_unlock(), once by hand).
4755 */
4756int __cond_resched_lock(spinlock_t *lock)
4757{
4758	int resched = should_resched();
4759	int ret = 0;
4760
4761	lockdep_assert_held(lock);
4762
4763	if (spin_needbreak(lock) || resched) {
4764		spin_unlock(lock);
4765		if (resched)
4766			__cond_resched();
4767		else
4768			cpu_relax();
4769		ret = 1;
4770		spin_lock(lock);
4771	}
4772	return ret;
4773}
4774EXPORT_SYMBOL(__cond_resched_lock);
4775
4776int __sched __cond_resched_softirq(void)
4777{
4778	BUG_ON(!in_softirq());
4779
4780	if (should_resched()) {
4781		local_bh_enable();
4782		__cond_resched();
4783		local_bh_disable();
4784		return 1;
4785	}
4786	return 0;
4787}
4788EXPORT_SYMBOL(__cond_resched_softirq);
4789
4790/**
4791 * yield - yield the current processor to other threads.
4792 *
4793 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4794 *
4795 * The scheduler is at all times free to pick the calling task as the most
4796 * eligible task to run, if removing the yield() call from your code breaks
4797 * it, its already broken.
4798 *
4799 * Typical broken usage is:
4800 *
4801 * while (!event)
4802 * 	yield();
4803 *
4804 * where one assumes that yield() will let 'the other' process run that will
4805 * make event true. If the current task is a SCHED_FIFO task that will never
4806 * happen. Never use yield() as a progress guarantee!!
4807 *
4808 * If you want to use yield() to wait for something, use wait_event().
4809 * If you want to use yield() to be 'nice' for others, use cond_resched().
4810 * If you still want to use yield(), do not!
4811 */
4812void __sched yield(void)
4813{
4814	set_current_state(TASK_RUNNING);
4815	sys_sched_yield();
4816}
4817EXPORT_SYMBOL(yield);
4818
4819/**
4820 * yield_to - yield the current processor to another thread in
4821 * your thread group, or accelerate that thread toward the
4822 * processor it's on.
4823 * @p: target task
4824 * @preempt: whether task preemption is allowed or not
4825 *
4826 * It's the caller's job to ensure that the target task struct
4827 * can't go away on us before we can do any checks.
4828 *
4829 * Returns true if we indeed boosted the target task.
4830 */
4831bool __sched yield_to(struct task_struct *p, bool preempt)
4832{
4833	struct task_struct *curr = current;
4834	struct rq *rq, *p_rq;
4835	unsigned long flags;
4836	bool yielded = 0;
4837
4838	local_irq_save(flags);
4839	rq = this_rq();
4840
4841again:
4842	p_rq = task_rq(p);
4843	double_rq_lock(rq, p_rq);
4844	while (task_rq(p) != p_rq) {
4845		double_rq_unlock(rq, p_rq);
4846		goto again;
4847	}
4848
4849	if (!curr->sched_class->yield_to_task)
4850		goto out;
4851
4852	if (curr->sched_class != p->sched_class)
4853		goto out;
4854
4855	if (task_running(p_rq, p) || p->state)
4856		goto out;
4857
4858	yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4859	if (yielded) {
4860		schedstat_inc(rq, yld_count);
4861		/*
4862		 * Make p's CPU reschedule; pick_next_entity takes care of
4863		 * fairness.
4864		 */
4865		if (preempt && rq != p_rq)
4866			resched_task(p_rq->curr);
4867	} else {
4868		/*
4869		 * We might have set it in task_yield_fair(), but are
4870		 * not going to schedule(), so don't want to skip
4871		 * the next update.
4872		 */
4873		rq->skip_clock_update = 0;
4874	}
4875
4876out:
4877	double_rq_unlock(rq, p_rq);
4878	local_irq_restore(flags);
4879
4880	if (yielded)
4881		schedule();
4882
4883	return yielded;
4884}
4885EXPORT_SYMBOL_GPL(yield_to);
4886
4887/*
4888 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4889 * that process accounting knows that this is a task in IO wait state.
4890 */
4891void __sched io_schedule(void)
4892{
4893	struct rq *rq = raw_rq();
4894
4895	delayacct_blkio_start();
4896	atomic_inc(&rq->nr_iowait);
4897	blk_flush_plug(current);
4898	current->in_iowait = 1;
4899	schedule();
4900	current->in_iowait = 0;
4901	atomic_dec(&rq->nr_iowait);
4902	delayacct_blkio_end();
4903}
4904EXPORT_SYMBOL(io_schedule);
4905
4906long __sched io_schedule_timeout(long timeout)
4907{
4908	struct rq *rq = raw_rq();
4909	long ret;
4910
4911	delayacct_blkio_start();
4912	atomic_inc(&rq->nr_iowait);
4913	blk_flush_plug(current);
4914	current->in_iowait = 1;
4915	ret = schedule_timeout(timeout);
4916	current->in_iowait = 0;
4917	atomic_dec(&rq->nr_iowait);
4918	delayacct_blkio_end();
4919	return ret;
4920}
4921
4922/**
4923 * sys_sched_get_priority_max - return maximum RT priority.
4924 * @policy: scheduling class.
4925 *
4926 * this syscall returns the maximum rt_priority that can be used
4927 * by a given scheduling class.
4928 */
4929SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4930{
4931	int ret = -EINVAL;
4932
4933	switch (policy) {
4934	case SCHED_FIFO:
4935	case SCHED_RR:
4936		ret = MAX_USER_RT_PRIO-1;
4937		break;
4938	case SCHED_NORMAL:
4939	case SCHED_BATCH:
4940	case SCHED_IDLE:
4941		ret = 0;
4942		break;
4943	}
4944	return ret;
4945}
4946
4947/**
4948 * sys_sched_get_priority_min - return minimum RT priority.
4949 * @policy: scheduling class.
4950 *
4951 * this syscall returns the minimum rt_priority that can be used
4952 * by a given scheduling class.
4953 */
4954SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4955{
4956	int ret = -EINVAL;
4957
4958	switch (policy) {
4959	case SCHED_FIFO:
4960	case SCHED_RR:
4961		ret = 1;
4962		break;
4963	case SCHED_NORMAL:
4964	case SCHED_BATCH:
4965	case SCHED_IDLE:
4966		ret = 0;
4967	}
4968	return ret;
4969}
4970
4971/**
4972 * sys_sched_rr_get_interval - return the default timeslice of a process.
4973 * @pid: pid of the process.
4974 * @interval: userspace pointer to the timeslice value.
4975 *
4976 * this syscall writes the default timeslice value of a given process
4977 * into the user-space timespec buffer. A value of '0' means infinity.
4978 */
4979SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4980		struct timespec __user *, interval)
4981{
4982	struct task_struct *p;
4983	unsigned int time_slice;
4984	unsigned long flags;
4985	struct rq *rq;
4986	int retval;
4987	struct timespec t;
4988
4989	if (pid < 0)
4990		return -EINVAL;
4991
4992	retval = -ESRCH;
4993	rcu_read_lock();
4994	p = find_process_by_pid(pid);
4995	if (!p)
4996		goto out_unlock;
4997
4998	retval = security_task_getscheduler(p);
4999	if (retval)
5000		goto out_unlock;
5001
5002	rq = task_rq_lock(p, &flags);
5003	time_slice = p->sched_class->get_rr_interval(rq, p);
5004	task_rq_unlock(rq, p, &flags);
5005
5006	rcu_read_unlock();
5007	jiffies_to_timespec(time_slice, &t);
5008	retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5009	return retval;
5010
5011out_unlock:
5012	rcu_read_unlock();
5013	return retval;
5014}
5015
5016static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5017
5018void sched_show_task(struct task_struct *p)
5019{
5020	unsigned long free = 0;
5021	unsigned state;
5022
5023	state = p->state ? __ffs(p->state) + 1 : 0;
5024	printk(KERN_INFO "%-15.15s %c", p->comm,
5025		state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5026#if BITS_PER_LONG == 32
5027	if (state == TASK_RUNNING)
5028		printk(KERN_CONT " running  ");
5029	else
5030		printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5031#else
5032	if (state == TASK_RUNNING)
5033		printk(KERN_CONT "  running task    ");
5034	else
5035		printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5036#endif
5037#ifdef CONFIG_DEBUG_STACK_USAGE
5038	free = stack_not_used(p);
5039#endif
5040	printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5041		task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
5042		(unsigned long)task_thread_info(p)->flags);
5043
5044	show_stack(p, NULL);
5045}
5046
5047void show_state_filter(unsigned long state_filter)
5048{
5049	struct task_struct *g, *p;
5050
5051#if BITS_PER_LONG == 32
5052	printk(KERN_INFO
5053		"  task                PC stack   pid father\n");
5054#else
5055	printk(KERN_INFO
5056		"  task                        PC stack   pid father\n");
5057#endif
5058	rcu_read_lock();
5059	do_each_thread(g, p) {
5060		/*
5061		 * reset the NMI-timeout, listing all files on a slow
5062		 * console might take a lot of time:
5063		 */
5064		touch_nmi_watchdog();
5065		if (!state_filter || (p->state & state_filter))
5066			sched_show_task(p);
5067	} while_each_thread(g, p);
5068
5069	touch_all_softlockup_watchdogs();
5070
5071#ifdef CONFIG_SCHED_DEBUG
5072	sysrq_sched_debug_show();
5073#endif
5074	rcu_read_unlock();
5075	/*
5076	 * Only show locks if all tasks are dumped:
5077	 */
5078	if (!state_filter)
5079		debug_show_all_locks();
5080}
5081
5082void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5083{
5084	idle->sched_class = &idle_sched_class;
5085}
5086
5087/**
5088 * init_idle - set up an idle thread for a given CPU
5089 * @idle: task in question
5090 * @cpu: cpu the idle task belongs to
5091 *
5092 * NOTE: this function does not set the idle thread's NEED_RESCHED
5093 * flag, to make booting more robust.
5094 */
5095void __cpuinit init_idle(struct task_struct *idle, int cpu)
5096{
5097	struct rq *rq = cpu_rq(cpu);
5098	unsigned long flags;
5099
5100	raw_spin_lock_irqsave(&rq->lock, flags);
5101
5102	__sched_fork(idle);
5103	idle->state = TASK_RUNNING;
5104	idle->se.exec_start = sched_clock();
5105
5106	do_set_cpus_allowed(idle, cpumask_of(cpu));
5107	/*
5108	 * We're having a chicken and egg problem, even though we are
5109	 * holding rq->lock, the cpu isn't yet set to this cpu so the
5110	 * lockdep check in task_group() will fail.
5111	 *
5112	 * Similar case to sched_fork(). / Alternatively we could
5113	 * use task_rq_lock() here and obtain the other rq->lock.
5114	 *
5115	 * Silence PROVE_RCU
5116	 */
5117	rcu_read_lock();
5118	__set_task_cpu(idle, cpu);
5119	rcu_read_unlock();
5120
5121	rq->curr = rq->idle = idle;
5122#if defined(CONFIG_SMP)
5123	idle->on_cpu = 1;
5124#endif
5125	raw_spin_unlock_irqrestore(&rq->lock, flags);
5126
5127	/* Set the preempt count _outside_ the spinlocks! */
5128	task_thread_info(idle)->preempt_count = 0;
5129
5130	/*
5131	 * The idle tasks have their own, simple scheduling class:
5132	 */
5133	idle->sched_class = &idle_sched_class;
5134	ftrace_graph_init_idle_task(idle, cpu);
5135#if defined(CONFIG_SMP)
5136	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5137#endif
5138}
5139
5140#ifdef CONFIG_SMP
5141void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5142{
5143	if (p->sched_class && p->sched_class->set_cpus_allowed)
5144		p->sched_class->set_cpus_allowed(p, new_mask);
5145
5146	cpumask_copy(&p->cpus_allowed, new_mask);
5147	p->nr_cpus_allowed = cpumask_weight(new_mask);
5148}
5149
5150/*
5151 * This is how migration works:
5152 *
5153 * 1) we invoke migration_cpu_stop() on the target CPU using
5154 *    stop_one_cpu().
5155 * 2) stopper starts to run (implicitly forcing the migrated thread
5156 *    off the CPU)
5157 * 3) it checks whether the migrated task is still in the wrong runqueue.
5158 * 4) if it's in the wrong runqueue then the migration thread removes
5159 *    it and puts it into the right queue.
5160 * 5) stopper completes and stop_one_cpu() returns and the migration
5161 *    is done.
5162 */
5163
5164/*
5165 * Change a given task's CPU affinity. Migrate the thread to a
5166 * proper CPU and schedule it away if the CPU it's executing on
5167 * is removed from the allowed bitmask.
5168 *
5169 * NOTE: the caller must have a valid reference to the task, the
5170 * task must not exit() & deallocate itself prematurely. The
5171 * call is not atomic; no spinlocks may be held.
5172 */
5173int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5174{
5175	unsigned long flags;
5176	struct rq *rq;
5177	unsigned int dest_cpu;
5178	int ret = 0;
5179
5180	rq = task_rq_lock(p, &flags);
5181
5182	if (cpumask_equal(&p->cpus_allowed, new_mask))
5183		goto out;
5184
5185	if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5186		ret = -EINVAL;
5187		goto out;
5188	}
5189
5190	if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5191		ret = -EINVAL;
5192		goto out;
5193	}
5194
5195	do_set_cpus_allowed(p, new_mask);
5196
5197	/* Can the task run on the task's current CPU? If so, we're done */
5198	if (cpumask_test_cpu(task_cpu(p), new_mask))
5199		goto out;
5200
5201	dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5202	if (p->on_rq) {
5203		struct migration_arg arg = { p, dest_cpu };
5204		/* Need help from migration thread: drop lock and wait. */
5205		task_rq_unlock(rq, p, &flags);
5206		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5207		tlb_migrate_finish(p->mm);
5208		return 0;
5209	}
5210out:
5211	task_rq_unlock(rq, p, &flags);
5212
5213	return ret;
5214}
5215EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5216
5217/*
5218 * Move (not current) task off this cpu, onto dest cpu. We're doing
5219 * this because either it can't run here any more (set_cpus_allowed()
5220 * away from this CPU, or CPU going down), or because we're
5221 * attempting to rebalance this task on exec (sched_exec).
5222 *
5223 * So we race with normal scheduler movements, but that's OK, as long
5224 * as the task is no longer on this CPU.
5225 *
5226 * Returns non-zero if task was successfully migrated.
5227 */
5228static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5229{
5230	struct rq *rq_dest, *rq_src;
5231	int ret = 0;
5232
5233	if (unlikely(!cpu_active(dest_cpu)))
5234		return ret;
5235
5236	rq_src = cpu_rq(src_cpu);
5237	rq_dest = cpu_rq(dest_cpu);
5238
5239	raw_spin_lock(&p->pi_lock);
5240	double_rq_lock(rq_src, rq_dest);
5241	/* Already moved. */
5242	if (task_cpu(p) != src_cpu)
5243		goto done;
5244	/* Affinity changed (again). */
5245	if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5246		goto fail;
5247
5248	/*
5249	 * If we're not on a rq, the next wake-up will ensure we're
5250	 * placed properly.
5251	 */
5252	if (p->on_rq) {
5253		dequeue_task(rq_src, p, 0);
5254		set_task_cpu(p, dest_cpu);
5255		enqueue_task(rq_dest, p, 0);
5256		check_preempt_curr(rq_dest, p, 0);
5257	}
5258done:
5259	ret = 1;
5260fail:
5261	double_rq_unlock(rq_src, rq_dest);
5262	raw_spin_unlock(&p->pi_lock);
5263	return ret;
5264}
5265
5266/*
5267 * migration_cpu_stop - this will be executed by a highprio stopper thread
5268 * and performs thread migration by bumping thread off CPU then
5269 * 'pushing' onto another runqueue.
5270 */
5271static int migration_cpu_stop(void *data)
5272{
5273	struct migration_arg *arg = data;
5274
5275	/*
5276	 * The original target cpu might have gone down and we might
5277	 * be on another cpu but it doesn't matter.
5278	 */
5279	local_irq_disable();
5280	__migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5281	local_irq_enable();
5282	return 0;
5283}
5284
5285#ifdef CONFIG_HOTPLUG_CPU
5286
5287/*
5288 * Ensures that the idle task is using init_mm right before its cpu goes
5289 * offline.
5290 */
5291void idle_task_exit(void)
5292{
5293	struct mm_struct *mm = current->active_mm;
5294
5295	BUG_ON(cpu_online(smp_processor_id()));
5296
5297	if (mm != &init_mm)
5298		switch_mm(mm, &init_mm, current);
5299	mmdrop(mm);
5300}
5301
5302/*
5303 * While a dead CPU has no uninterruptible tasks queued at this point,
5304 * it might still have a nonzero ->nr_uninterruptible counter, because
5305 * for performance reasons the counter is not stricly tracking tasks to
5306 * their home CPUs. So we just add the counter to another CPU's counter,
5307 * to keep the global sum constant after CPU-down:
5308 */
5309static void migrate_nr_uninterruptible(struct rq *rq_src)
5310{
5311	struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5312
5313	rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5314	rq_src->nr_uninterruptible = 0;
5315}
5316
5317/*
5318 * remove the tasks which were accounted by rq from calc_load_tasks.
5319 */
5320static void calc_global_load_remove(struct rq *rq)
5321{
5322	atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5323	rq->calc_load_active = 0;
5324}
5325
5326/*
5327 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5328 * try_to_wake_up()->select_task_rq().
5329 *
5330 * Called with rq->lock held even though we'er in stop_machine() and
5331 * there's no concurrency possible, we hold the required locks anyway
5332 * because of lock validation efforts.
5333 */
5334static void migrate_tasks(unsigned int dead_cpu)
5335{
5336	struct rq *rq = cpu_rq(dead_cpu);
5337	struct task_struct *next, *stop = rq->stop;
5338	int dest_cpu;
5339
5340	/*
5341	 * Fudge the rq selection such that the below task selection loop
5342	 * doesn't get stuck on the currently eligible stop task.
5343	 *
5344	 * We're currently inside stop_machine() and the rq is either stuck
5345	 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5346	 * either way we should never end up calling schedule() until we're
5347	 * done here.
5348	 */
5349	rq->stop = NULL;
5350
5351	/* Ensure any throttled groups are reachable by pick_next_task */
5352	unthrottle_offline_cfs_rqs(rq);
5353
5354	for ( ; ; ) {
5355		/*
5356		 * There's this thread running, bail when that's the only
5357		 * remaining thread.
5358		 */
5359		if (rq->nr_running == 1)
5360			break;
5361
5362		next = pick_next_task(rq);
5363		BUG_ON(!next);
5364		next->sched_class->put_prev_task(rq, next);
5365
5366		/* Find suitable destination for @next, with force if needed. */
5367		dest_cpu = select_fallback_rq(dead_cpu, next);
5368		raw_spin_unlock(&rq->lock);
5369
5370		__migrate_task(next, dead_cpu, dest_cpu);
5371
5372		raw_spin_lock(&rq->lock);
5373	}
5374
5375	rq->stop = stop;
5376}
5377
5378#endif /* CONFIG_HOTPLUG_CPU */
5379
5380#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5381
5382static struct ctl_table sd_ctl_dir[] = {
5383	{
5384		.procname	= "sched_domain",
5385		.mode		= 0555,
5386	},
5387	{}
5388};
5389
5390static struct ctl_table sd_ctl_root[] = {
5391	{
5392		.procname	= "kernel",
5393		.mode		= 0555,
5394		.child		= sd_ctl_dir,
5395	},
5396	{}
5397};
5398
5399static struct ctl_table *sd_alloc_ctl_entry(int n)
5400{
5401	struct ctl_table *entry =
5402		kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5403
5404	return entry;
5405}
5406
5407static void sd_free_ctl_entry(struct ctl_table **tablep)
5408{
5409	struct ctl_table *entry;
5410
5411	/*
5412	 * In the intermediate directories, both the child directory and
5413	 * procname are dynamically allocated and could fail but the mode
5414	 * will always be set. In the lowest directory the names are
5415	 * static strings and all have proc handlers.
5416	 */
5417	for (entry = *tablep; entry->mode; entry++) {
5418		if (entry->child)
5419			sd_free_ctl_entry(&entry->child);
5420		if (entry->proc_handler == NULL)
5421			kfree(entry->procname);
5422	}
5423
5424	kfree(*tablep);
5425	*tablep = NULL;
5426}
5427
5428static void
5429set_table_entry(struct ctl_table *entry,
5430		const char *procname, void *data, int maxlen,
5431		umode_t mode, proc_handler *proc_handler)
5432{
5433	entry->procname = procname;
5434	entry->data = data;
5435	entry->maxlen = maxlen;
5436	entry->mode = mode;
5437	entry->proc_handler = proc_handler;
5438}
5439
5440static struct ctl_table *
5441sd_alloc_ctl_domain_table(struct sched_domain *sd)
5442{
5443	struct ctl_table *table = sd_alloc_ctl_entry(13);
5444
5445	if (table == NULL)
5446		return NULL;
5447
5448	set_table_entry(&table[0], "min_interval", &sd->min_interval,
5449		sizeof(long), 0644, proc_doulongvec_minmax);
5450	set_table_entry(&table[1], "max_interval", &sd->max_interval,
5451		sizeof(long), 0644, proc_doulongvec_minmax);
5452	set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5453		sizeof(int), 0644, proc_dointvec_minmax);
5454	set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5455		sizeof(int), 0644, proc_dointvec_minmax);
5456	set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5457		sizeof(int), 0644, proc_dointvec_minmax);
5458	set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5459		sizeof(int), 0644, proc_dointvec_minmax);
5460	set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5461		sizeof(int), 0644, proc_dointvec_minmax);
5462	set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5463		sizeof(int), 0644, proc_dointvec_minmax);
5464	set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5465		sizeof(int), 0644, proc_dointvec_minmax);
5466	set_table_entry(&table[9], "cache_nice_tries",
5467		&sd->cache_nice_tries,
5468		sizeof(int), 0644, proc_dointvec_minmax);
5469	set_table_entry(&table[10], "flags", &sd->flags,
5470		sizeof(int), 0644, proc_dointvec_minmax);
5471	set_table_entry(&table[11], "name", sd->name,
5472		CORENAME_MAX_SIZE, 0444, proc_dostring);
5473	/* &table[12] is terminator */
5474
5475	return table;
5476}
5477
5478static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5479{
5480	struct ctl_table *entry, *table;
5481	struct sched_domain *sd;
5482	int domain_num = 0, i;
5483	char buf[32];
5484
5485	for_each_domain(cpu, sd)
5486		domain_num++;
5487	entry = table = sd_alloc_ctl_entry(domain_num + 1);
5488	if (table == NULL)
5489		return NULL;
5490
5491	i = 0;
5492	for_each_domain(cpu, sd) {
5493		snprintf(buf, 32, "domain%d", i);
5494		entry->procname = kstrdup(buf, GFP_KERNEL);
5495		entry->mode = 0555;
5496		entry->child = sd_alloc_ctl_domain_table(sd);
5497		entry++;
5498		i++;
5499	}
5500	return table;
5501}
5502
5503static struct ctl_table_header *sd_sysctl_header;
5504static void register_sched_domain_sysctl(void)
5505{
5506	int i, cpu_num = num_possible_cpus();
5507	struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5508	char buf[32];
5509
5510	WARN_ON(sd_ctl_dir[0].child);
5511	sd_ctl_dir[0].child = entry;
5512
5513	if (entry == NULL)
5514		return;
5515
5516	for_each_possible_cpu(i) {
5517		snprintf(buf, 32, "cpu%d", i);
5518		entry->procname = kstrdup(buf, GFP_KERNEL);
5519		entry->mode = 0555;
5520		entry->child = sd_alloc_ctl_cpu_table(i);
5521		entry++;
5522	}
5523
5524	WARN_ON(sd_sysctl_header);
5525	sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5526}
5527
5528/* may be called multiple times per register */
5529static void unregister_sched_domain_sysctl(void)
5530{
5531	if (sd_sysctl_header)
5532		unregister_sysctl_table(sd_sysctl_header);
5533	sd_sysctl_header = NULL;
5534	if (sd_ctl_dir[0].child)
5535		sd_free_ctl_entry(&sd_ctl_dir[0].child);
5536}
5537#else
5538static void register_sched_domain_sysctl(void)
5539{
5540}
5541static void unregister_sched_domain_sysctl(void)
5542{
5543}
5544#endif
5545
5546static void set_rq_online(struct rq *rq)
5547{
5548	if (!rq->online) {
5549		const struct sched_class *class;
5550
5551		cpumask_set_cpu(rq->cpu, rq->rd->online);
5552		rq->online = 1;
5553
5554		for_each_class(class) {
5555			if (class->rq_online)
5556				class->rq_online(rq);
5557		}
5558	}
5559}
5560
5561static void set_rq_offline(struct rq *rq)
5562{
5563	if (rq->online) {
5564		const struct sched_class *class;
5565
5566		for_each_class(class) {
5567			if (class->rq_offline)
5568				class->rq_offline(rq);
5569		}
5570
5571		cpumask_clear_cpu(rq->cpu, rq->rd->online);
5572		rq->online = 0;
5573	}
5574}
5575
5576/*
5577 * migration_call - callback that gets triggered when a CPU is added.
5578 * Here we can start up the necessary migration thread for the new CPU.
5579 */
5580static int __cpuinit
5581migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5582{
5583	int cpu = (long)hcpu;
5584	unsigned long flags;
5585	struct rq *rq = cpu_rq(cpu);
5586
5587	switch (action & ~CPU_TASKS_FROZEN) {
5588
5589	case CPU_UP_PREPARE:
5590		rq->calc_load_update = calc_load_update;
5591		break;
5592
5593	case CPU_ONLINE:
5594		/* Update our root-domain */
5595		raw_spin_lock_irqsave(&rq->lock, flags);
5596		if (rq->rd) {
5597			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5598
5599			set_rq_online(rq);
5600		}
5601		raw_spin_unlock_irqrestore(&rq->lock, flags);
5602		break;
5603
5604#ifdef CONFIG_HOTPLUG_CPU
5605	case CPU_DYING:
5606		sched_ttwu_pending();
5607		/* Update our root-domain */
5608		raw_spin_lock_irqsave(&rq->lock, flags);
5609		if (rq->rd) {
5610			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5611			set_rq_offline(rq);
5612		}
5613		migrate_tasks(cpu);
5614		BUG_ON(rq->nr_running != 1); /* the migration thread */
5615		raw_spin_unlock_irqrestore(&rq->lock, flags);
5616
5617		migrate_nr_uninterruptible(rq);
5618		calc_global_load_remove(rq);
5619		break;
5620#endif
5621	}
5622
5623	update_max_interval();
5624
5625	return NOTIFY_OK;
5626}
5627
5628/*
5629 * Register at high priority so that task migration (migrate_all_tasks)
5630 * happens before everything else.  This has to be lower priority than
5631 * the notifier in the perf_event subsystem, though.
5632 */
5633static struct notifier_block __cpuinitdata migration_notifier = {
5634	.notifier_call = migration_call,
5635	.priority = CPU_PRI_MIGRATION,
5636};
5637
5638static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5639				      unsigned long action, void *hcpu)
5640{
5641	switch (action & ~CPU_TASKS_FROZEN) {
5642	case CPU_STARTING:
5643	case CPU_DOWN_FAILED:
5644		set_cpu_active((long)hcpu, true);
5645		return NOTIFY_OK;
5646	default:
5647		return NOTIFY_DONE;
5648	}
5649}
5650
5651static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5652					unsigned long action, void *hcpu)
5653{
5654	switch (action & ~CPU_TASKS_FROZEN) {
5655	case CPU_DOWN_PREPARE:
5656		set_cpu_active((long)hcpu, false);
5657		return NOTIFY_OK;
5658	default:
5659		return NOTIFY_DONE;
5660	}
5661}
5662
5663static int __init migration_init(void)
5664{
5665	void *cpu = (void *)(long)smp_processor_id();
5666	int err;
5667
5668	/* Initialize migration for the boot CPU */
5669	err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5670	BUG_ON(err == NOTIFY_BAD);
5671	migration_call(&migration_notifier, CPU_ONLINE, cpu);
5672	register_cpu_notifier(&migration_notifier);
5673
5674	/* Register cpu active notifiers */
5675	cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5676	cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5677
5678	return 0;
5679}
5680early_initcall(migration_init);
5681#endif
5682
5683#ifdef CONFIG_SMP
5684
5685static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5686
5687#ifdef CONFIG_SCHED_DEBUG
5688
5689static __read_mostly int sched_debug_enabled;
5690
5691static int __init sched_debug_setup(char *str)
5692{
5693	sched_debug_enabled = 1;
5694
5695	return 0;
5696}
5697early_param("sched_debug", sched_debug_setup);
5698
5699static inline bool sched_debug(void)
5700{
5701	return sched_debug_enabled;
5702}
5703
5704static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5705				  struct cpumask *groupmask)
5706{
5707	struct sched_group *group = sd->groups;
5708	char str[256];
5709
5710	cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5711	cpumask_clear(groupmask);
5712
5713	printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5714
5715	if (!(sd->flags & SD_LOAD_BALANCE)) {
5716		printk("does not load-balance\n");
5717		if (sd->parent)
5718			printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5719					" has parent");
5720		return -1;
5721	}
5722
5723	printk(KERN_CONT "span %s level %s\n", str, sd->name);
5724
5725	if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5726		printk(KERN_ERR "ERROR: domain->span does not contain "
5727				"CPU%d\n", cpu);
5728	}
5729	if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5730		printk(KERN_ERR "ERROR: domain->groups does not contain"
5731				" CPU%d\n", cpu);
5732	}
5733
5734	printk(KERN_DEBUG "%*s groups:", level + 1, "");
5735	do {
5736		if (!group) {
5737			printk("\n");
5738			printk(KERN_ERR "ERROR: group is NULL\n");
5739			break;
5740		}
5741
5742		/*
5743		 * Even though we initialize ->power to something semi-sane,
5744		 * we leave power_orig unset. This allows us to detect if
5745		 * domain iteration is still funny without causing /0 traps.
5746		 */
5747		if (!group->sgp->power_orig) {
5748			printk(KERN_CONT "\n");
5749			printk(KERN_ERR "ERROR: domain->cpu_power not "
5750					"set\n");
5751			break;
5752		}
5753
5754		if (!cpumask_weight(sched_group_cpus(group))) {
5755			printk(KERN_CONT "\n");
5756			printk(KERN_ERR "ERROR: empty group\n");
5757			break;
5758		}
5759
5760		if (!(sd->flags & SD_OVERLAP) &&
5761		    cpumask_intersects(groupmask, sched_group_cpus(group))) {
5762			printk(KERN_CONT "\n");
5763			printk(KERN_ERR "ERROR: repeated CPUs\n");
5764			break;
5765		}
5766
5767		cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5768
5769		cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5770
5771		printk(KERN_CONT " %s", str);
5772		if (group->sgp->power != SCHED_POWER_SCALE) {
5773			printk(KERN_CONT " (cpu_power = %d)",
5774				group->sgp->power);
5775		}
5776
5777		group = group->next;
5778	} while (group != sd->groups);
5779	printk(KERN_CONT "\n");
5780
5781	if (!cpumask_equal(sched_domain_span(sd), groupmask))
5782		printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5783
5784	if (sd->parent &&
5785	    !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5786		printk(KERN_ERR "ERROR: parent span is not a superset "
5787			"of domain->span\n");
5788	return 0;
5789}
5790
5791static void sched_domain_debug(struct sched_domain *sd, int cpu)
5792{
5793	int level = 0;
5794
5795	if (!sched_debug_enabled)
5796		return;
5797
5798	if (!sd) {
5799		printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5800		return;
5801	}
5802
5803	printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5804
5805	for (;;) {
5806		if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5807			break;
5808		level++;
5809		sd = sd->parent;
5810		if (!sd)
5811			break;
5812	}
5813}
5814#else /* !CONFIG_SCHED_DEBUG */
5815# define sched_domain_debug(sd, cpu) do { } while (0)
5816static inline bool sched_debug(void)
5817{
5818	return false;
5819}
5820#endif /* CONFIG_SCHED_DEBUG */
5821
5822static int sd_degenerate(struct sched_domain *sd)
5823{
5824	if (cpumask_weight(sched_domain_span(sd)) == 1)
5825		return 1;
5826
5827	/* Following flags need at least 2 groups */
5828	if (sd->flags & (SD_LOAD_BALANCE |
5829			 SD_BALANCE_NEWIDLE |
5830			 SD_BALANCE_FORK |
5831			 SD_BALANCE_EXEC |
5832			 SD_SHARE_CPUPOWER |
5833			 SD_SHARE_PKG_RESOURCES)) {
5834		if (sd->groups != sd->groups->next)
5835			return 0;
5836	}
5837
5838	/* Following flags don't use groups */
5839	if (sd->flags & (SD_WAKE_AFFINE))
5840		return 0;
5841
5842	return 1;
5843}
5844
5845static int
5846sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5847{
5848	unsigned long cflags = sd->flags, pflags = parent->flags;
5849
5850	if (sd_degenerate(parent))
5851		return 1;
5852
5853	if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5854		return 0;
5855
5856	/* Flags needing groups don't count if only 1 group in parent */
5857	if (parent->groups == parent->groups->next) {
5858		pflags &= ~(SD_LOAD_BALANCE |
5859				SD_BALANCE_NEWIDLE |
5860				SD_BALANCE_FORK |
5861				SD_BALANCE_EXEC |
5862				SD_SHARE_CPUPOWER |
5863				SD_SHARE_PKG_RESOURCES);
5864		if (nr_node_ids == 1)
5865			pflags &= ~SD_SERIALIZE;
5866	}
5867	if (~cflags & pflags)
5868		return 0;
5869
5870	return 1;
5871}
5872
5873static void free_rootdomain(struct rcu_head *rcu)
5874{
5875	struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5876
5877	cpupri_cleanup(&rd->cpupri);
5878	free_cpumask_var(rd->rto_mask);
5879	free_cpumask_var(rd->online);
5880	free_cpumask_var(rd->span);
5881	kfree(rd);
5882}
5883
5884static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5885{
5886	struct root_domain *old_rd = NULL;
5887	unsigned long flags;
5888
5889	raw_spin_lock_irqsave(&rq->lock, flags);
5890
5891	if (rq->rd) {
5892		old_rd = rq->rd;
5893
5894		if (cpumask_test_cpu(rq->cpu, old_rd->online))
5895			set_rq_offline(rq);
5896
5897		cpumask_clear_cpu(rq->cpu, old_rd->span);
5898
5899		/*
5900		 * If we dont want to free the old_rt yet then
5901		 * set old_rd to NULL to skip the freeing later
5902		 * in this function:
5903		 */
5904		if (!atomic_dec_and_test(&old_rd->refcount))
5905			old_rd = NULL;
5906	}
5907
5908	atomic_inc(&rd->refcount);
5909	rq->rd = rd;
5910
5911	cpumask_set_cpu(rq->cpu, rd->span);
5912	if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5913		set_rq_online(rq);
5914
5915	raw_spin_unlock_irqrestore(&rq->lock, flags);
5916
5917	if (old_rd)
5918		call_rcu_sched(&old_rd->rcu, free_rootdomain);
5919}
5920
5921static int init_rootdomain(struct root_domain *rd)
5922{
5923	memset(rd, 0, sizeof(*rd));
5924
5925	if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5926		goto out;
5927	if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5928		goto free_span;
5929	if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5930		goto free_online;
5931
5932	if (cpupri_init(&rd->cpupri) != 0)
5933		goto free_rto_mask;
5934	return 0;
5935
5936free_rto_mask:
5937	free_cpumask_var(rd->rto_mask);
5938free_online:
5939	free_cpumask_var(rd->online);
5940free_span:
5941	free_cpumask_var(rd->span);
5942out:
5943	return -ENOMEM;
5944}
5945
5946/*
5947 * By default the system creates a single root-domain with all cpus as
5948 * members (mimicking the global state we have today).
5949 */
5950struct root_domain def_root_domain;
5951
5952static void init_defrootdomain(void)
5953{
5954	init_rootdomain(&def_root_domain);
5955
5956	atomic_set(&def_root_domain.refcount, 1);
5957}
5958
5959static struct root_domain *alloc_rootdomain(void)
5960{
5961	struct root_domain *rd;
5962
5963	rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5964	if (!rd)
5965		return NULL;
5966
5967	if (init_rootdomain(rd) != 0) {
5968		kfree(rd);
5969		return NULL;
5970	}
5971
5972	return rd;
5973}
5974
5975static void free_sched_groups(struct sched_group *sg, int free_sgp)
5976{
5977	struct sched_group *tmp, *first;
5978
5979	if (!sg)
5980		return;
5981
5982	first = sg;
5983	do {
5984		tmp = sg->next;
5985
5986		if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5987			kfree(sg->sgp);
5988
5989		kfree(sg);
5990		sg = tmp;
5991	} while (sg != first);
5992}
5993
5994static void free_sched_domain(struct rcu_head *rcu)
5995{
5996	struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5997
5998	/*
5999	 * If its an overlapping domain it has private groups, iterate and
6000	 * nuke them all.
6001	 */
6002	if (sd->flags & SD_OVERLAP) {
6003		free_sched_groups(sd->groups, 1);
6004	} else if (atomic_dec_and_test(&sd->groups->ref)) {
6005		kfree(sd->groups->sgp);
6006		kfree(sd->groups);
6007	}
6008	kfree(sd);
6009}
6010
6011static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6012{
6013	call_rcu(&sd->rcu, free_sched_domain);
6014}
6015
6016static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6017{
6018	for (; sd; sd = sd->parent)
6019		destroy_sched_domain(sd, cpu);
6020}
6021
6022/*
6023 * Keep a special pointer to the highest sched_domain that has
6024 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6025 * allows us to avoid some pointer chasing select_idle_sibling().
6026 *
6027 * Also keep a unique ID per domain (we use the first cpu number in
6028 * the cpumask of the domain), this allows us to quickly tell if
6029 * two cpus are in the same cache domain, see cpus_share_cache().
6030 */
6031DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6032DEFINE_PER_CPU(int, sd_llc_id);
6033
6034static void update_top_cache_domain(int cpu)
6035{
6036	struct sched_domain *sd;
6037	int id = cpu;
6038
6039	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6040	if (sd)
6041		id = cpumask_first(sched_domain_span(sd));
6042
6043	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6044	per_cpu(sd_llc_id, cpu) = id;
6045}
6046
6047/*
6048 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6049 * hold the hotplug lock.
6050 */
6051static void
6052cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6053{
6054	struct rq *rq = cpu_rq(cpu);
6055	struct sched_domain *tmp;
6056
6057	/* Remove the sched domains which do not contribute to scheduling. */
6058	for (tmp = sd; tmp; ) {
6059		struct sched_domain *parent = tmp->parent;
6060		if (!parent)
6061			break;
6062
6063		if (sd_parent_degenerate(tmp, parent)) {
6064			tmp->parent = parent->parent;
6065			if (parent->parent)
6066				parent->parent->child = tmp;
6067			destroy_sched_domain(parent, cpu);
6068		} else
6069			tmp = tmp->parent;
6070	}
6071
6072	if (sd && sd_degenerate(sd)) {
6073		tmp = sd;
6074		sd = sd->parent;
6075		destroy_sched_domain(tmp, cpu);
6076		if (sd)
6077			sd->child = NULL;
6078	}
6079
6080	sched_domain_debug(sd, cpu);
6081
6082	rq_attach_root(rq, rd);
6083	tmp = rq->sd;
6084	rcu_assign_pointer(rq->sd, sd);
6085	destroy_sched_domains(tmp, cpu);
6086
6087	update_top_cache_domain(cpu);
6088}
6089
6090/* cpus with isolated domains */
6091static cpumask_var_t cpu_isolated_map;
6092
6093/* Setup the mask of cpus configured for isolated domains */
6094static int __init isolated_cpu_setup(char *str)
6095{
6096	alloc_bootmem_cpumask_var(&cpu_isolated_map);
6097	cpulist_parse(str, cpu_isolated_map);
6098	return 1;
6099}
6100
6101__setup("isolcpus=", isolated_cpu_setup);
6102
6103static const struct cpumask *cpu_cpu_mask(int cpu)
6104{
6105	return cpumask_of_node(cpu_to_node(cpu));
6106}
6107
6108struct sd_data {
6109	struct sched_domain **__percpu sd;
6110	struct sched_group **__percpu sg;
6111	struct sched_group_power **__percpu sgp;
6112};
6113
6114struct s_data {
6115	struct sched_domain ** __percpu sd;
6116	struct root_domain	*rd;
6117};
6118
6119enum s_alloc {
6120	sa_rootdomain,
6121	sa_sd,
6122	sa_sd_storage,
6123	sa_none,
6124};
6125
6126struct sched_domain_topology_level;
6127
6128typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6129typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6130
6131#define SDTL_OVERLAP	0x01
6132
6133struct sched_domain_topology_level {
6134	sched_domain_init_f init;
6135	sched_domain_mask_f mask;
6136	int		    flags;
6137	int		    numa_level;
6138	struct sd_data      data;
6139};
6140
6141/*
6142 * Build an iteration mask that can exclude certain CPUs from the upwards
6143 * domain traversal.
6144 *
6145 * Asymmetric node setups can result in situations where the domain tree is of
6146 * unequal depth, make sure to skip domains that already cover the entire
6147 * range.
6148 *
6149 * In that case build_sched_domains() will have terminated the iteration early
6150 * and our sibling sd spans will be empty. Domains should always include the
6151 * cpu they're built on, so check that.
6152 *
6153 */
6154static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6155{
6156	const struct cpumask *span = sched_domain_span(sd);
6157	struct sd_data *sdd = sd->private;
6158	struct sched_domain *sibling;
6159	int i;
6160
6161	for_each_cpu(i, span) {
6162		sibling = *per_cpu_ptr(sdd->sd, i);
6163		if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6164			continue;
6165
6166		cpumask_set_cpu(i, sched_group_mask(sg));
6167	}
6168}
6169
6170/*
6171 * Return the canonical balance cpu for this group, this is the first cpu
6172 * of this group that's also in the iteration mask.
6173 */
6174int group_balance_cpu(struct sched_group *sg)
6175{
6176	return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6177}
6178
6179static int
6180build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6181{
6182	struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6183	const struct cpumask *span = sched_domain_span(sd);
6184	struct cpumask *covered = sched_domains_tmpmask;
6185	struct sd_data *sdd = sd->private;
6186	struct sched_domain *child;
6187	int i;
6188
6189	cpumask_clear(covered);
6190
6191	for_each_cpu(i, span) {
6192		struct cpumask *sg_span;
6193
6194		if (cpumask_test_cpu(i, covered))
6195			continue;
6196
6197		child = *per_cpu_ptr(sdd->sd, i);
6198
6199		/* See the comment near build_group_mask(). */
6200		if (!cpumask_test_cpu(i, sched_domain_span(child)))
6201			continue;
6202
6203		sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6204				GFP_KERNEL, cpu_to_node(cpu));
6205
6206		if (!sg)
6207			goto fail;
6208
6209		sg_span = sched_group_cpus(sg);
6210		if (child->child) {
6211			child = child->child;
6212			cpumask_copy(sg_span, sched_domain_span(child));
6213		} else
6214			cpumask_set_cpu(i, sg_span);
6215
6216		cpumask_or(covered, covered, sg_span);
6217
6218		sg->sgp = *per_cpu_ptr(sdd->sgp, i);
6219		if (atomic_inc_return(&sg->sgp->ref) == 1)
6220			build_group_mask(sd, sg);
6221
6222		/*
6223		 * Initialize sgp->power such that even if we mess up the
6224		 * domains and no possible iteration will get us here, we won't
6225		 * die on a /0 trap.
6226		 */
6227		sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
6228
6229		/*
6230		 * Make sure the first group of this domain contains the
6231		 * canonical balance cpu. Otherwise the sched_domain iteration
6232		 * breaks. See update_sg_lb_stats().
6233		 */
6234		if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6235		    group_balance_cpu(sg) == cpu)
6236			groups = sg;
6237
6238		if (!first)
6239			first = sg;
6240		if (last)
6241			last->next = sg;
6242		last = sg;
6243		last->next = first;
6244	}
6245	sd->groups = groups;
6246
6247	return 0;
6248
6249fail:
6250	free_sched_groups(first, 0);
6251
6252	return -ENOMEM;
6253}
6254
6255static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6256{
6257	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6258	struct sched_domain *child = sd->child;
6259
6260	if (child)
6261		cpu = cpumask_first(sched_domain_span(child));
6262
6263	if (sg) {
6264		*sg = *per_cpu_ptr(sdd->sg, cpu);
6265		(*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6266		atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6267	}
6268
6269	return cpu;
6270}
6271
6272/*
6273 * build_sched_groups will build a circular linked list of the groups
6274 * covered by the given span, and will set each group's ->cpumask correctly,
6275 * and ->cpu_power to 0.
6276 *
6277 * Assumes the sched_domain tree is fully constructed
6278 */
6279static int
6280build_sched_groups(struct sched_domain *sd, int cpu)
6281{
6282	struct sched_group *first = NULL, *last = NULL;
6283	struct sd_data *sdd = sd->private;
6284	const struct cpumask *span = sched_domain_span(sd);
6285	struct cpumask *covered;
6286	int i;
6287
6288	get_group(cpu, sdd, &sd->groups);
6289	atomic_inc(&sd->groups->ref);
6290
6291	if (cpu != cpumask_first(sched_domain_span(sd)))
6292		return 0;
6293
6294	lockdep_assert_held(&sched_domains_mutex);
6295	covered = sched_domains_tmpmask;
6296
6297	cpumask_clear(covered);
6298
6299	for_each_cpu(i, span) {
6300		struct sched_group *sg;
6301		int group = get_group(i, sdd, &sg);
6302		int j;
6303
6304		if (cpumask_test_cpu(i, covered))
6305			continue;
6306
6307		cpumask_clear(sched_group_cpus(sg));
6308		sg->sgp->power = 0;
6309		cpumask_setall(sched_group_mask(sg));
6310
6311		for_each_cpu(j, span) {
6312			if (get_group(j, sdd, NULL) != group)
6313				continue;
6314
6315			cpumask_set_cpu(j, covered);
6316			cpumask_set_cpu(j, sched_group_cpus(sg));
6317		}
6318
6319		if (!first)
6320			first = sg;
6321		if (last)
6322			last->next = sg;
6323		last = sg;
6324	}
6325	last->next = first;
6326
6327	return 0;
6328}
6329
6330/*
6331 * Initialize sched groups cpu_power.
6332 *
6333 * cpu_power indicates the capacity of sched group, which is used while
6334 * distributing the load between different sched groups in a sched domain.
6335 * Typically cpu_power for all the groups in a sched domain will be same unless
6336 * there are asymmetries in the topology. If there are asymmetries, group
6337 * having more cpu_power will pickup more load compared to the group having
6338 * less cpu_power.
6339 */
6340static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6341{
6342	struct sched_group *sg = sd->groups;
6343
6344	WARN_ON(!sd || !sg);
6345
6346	do {
6347		sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6348		sg = sg->next;
6349	} while (sg != sd->groups);
6350
6351	if (cpu != group_balance_cpu(sg))
6352		return;
6353
6354	update_group_power(sd, cpu);
6355	atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6356}
6357
6358int __weak arch_sd_sibling_asym_packing(void)
6359{
6360       return 0*SD_ASYM_PACKING;
6361}
6362
6363/*
6364 * Initializers for schedule domains
6365 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6366 */
6367
6368#ifdef CONFIG_SCHED_DEBUG
6369# define SD_INIT_NAME(sd, type)		sd->name = #type
6370#else
6371# define SD_INIT_NAME(sd, type)		do { } while (0)
6372#endif
6373
6374#define SD_INIT_FUNC(type)						\
6375static noinline struct sched_domain *					\
6376sd_init_##type(struct sched_domain_topology_level *tl, int cpu) 	\
6377{									\
6378	struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);	\
6379	*sd = SD_##type##_INIT;						\
6380	SD_INIT_NAME(sd, type);						\
6381	sd->private = &tl->data;					\
6382	return sd;							\
6383}
6384
6385SD_INIT_FUNC(CPU)
6386#ifdef CONFIG_SCHED_SMT
6387 SD_INIT_FUNC(SIBLING)
6388#endif
6389#ifdef CONFIG_SCHED_MC
6390 SD_INIT_FUNC(MC)
6391#endif
6392#ifdef CONFIG_SCHED_BOOK
6393 SD_INIT_FUNC(BOOK)
6394#endif
6395
6396static int default_relax_domain_level = -1;
6397int sched_domain_level_max;
6398
6399static int __init setup_relax_domain_level(char *str)
6400{
6401	if (kstrtoint(str, 0, &default_relax_domain_level))
6402		pr_warn("Unable to set relax_domain_level\n");
6403
6404	return 1;
6405}
6406__setup("relax_domain_level=", setup_relax_domain_level);
6407
6408static void set_domain_attribute(struct sched_domain *sd,
6409				 struct sched_domain_attr *attr)
6410{
6411	int request;
6412
6413	if (!attr || attr->relax_domain_level < 0) {
6414		if (default_relax_domain_level < 0)
6415			return;
6416		else
6417			request = default_relax_domain_level;
6418	} else
6419		request = attr->relax_domain_level;
6420	if (request < sd->level) {
6421		/* turn off idle balance on this domain */
6422		sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6423	} else {
6424		/* turn on idle balance on this domain */
6425		sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6426	}
6427}
6428
6429static void __sdt_free(const struct cpumask *cpu_map);
6430static int __sdt_alloc(const struct cpumask *cpu_map);
6431
6432static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6433				 const struct cpumask *cpu_map)
6434{
6435	switch (what) {
6436	case sa_rootdomain:
6437		if (!atomic_read(&d->rd->refcount))
6438			free_rootdomain(&d->rd->rcu); /* fall through */
6439	case sa_sd:
6440		free_percpu(d->sd); /* fall through */
6441	case sa_sd_storage:
6442		__sdt_free(cpu_map); /* fall through */
6443	case sa_none:
6444		break;
6445	}
6446}
6447
6448static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6449						   const struct cpumask *cpu_map)
6450{
6451	memset(d, 0, sizeof(*d));
6452
6453	if (__sdt_alloc(cpu_map))
6454		return sa_sd_storage;
6455	d->sd = alloc_percpu(struct sched_domain *);
6456	if (!d->sd)
6457		return sa_sd_storage;
6458	d->rd = alloc_rootdomain();
6459	if (!d->rd)
6460		return sa_sd;
6461	return sa_rootdomain;
6462}
6463
6464/*
6465 * NULL the sd_data elements we've used to build the sched_domain and
6466 * sched_group structure so that the subsequent __free_domain_allocs()
6467 * will not free the data we're using.
6468 */
6469static void claim_allocations(int cpu, struct sched_domain *sd)
6470{
6471	struct sd_data *sdd = sd->private;
6472
6473	WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6474	*per_cpu_ptr(sdd->sd, cpu) = NULL;
6475
6476	if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6477		*per_cpu_ptr(sdd->sg, cpu) = NULL;
6478
6479	if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6480		*per_cpu_ptr(sdd->sgp, cpu) = NULL;
6481}
6482
6483#ifdef CONFIG_SCHED_SMT
6484static const struct cpumask *cpu_smt_mask(int cpu)
6485{
6486	return topology_thread_cpumask(cpu);
6487}
6488#endif
6489
6490/*
6491 * Topology list, bottom-up.
6492 */
6493static struct sched_domain_topology_level default_topology[] = {
6494#ifdef CONFIG_SCHED_SMT
6495	{ sd_init_SIBLING, cpu_smt_mask, },
6496#endif
6497#ifdef CONFIG_SCHED_MC
6498	{ sd_init_MC, cpu_coregroup_mask, },
6499#endif
6500#ifdef CONFIG_SCHED_BOOK
6501	{ sd_init_BOOK, cpu_book_mask, },
6502#endif
6503	{ sd_init_CPU, cpu_cpu_mask, },
6504	{ NULL, },
6505};
6506
6507static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6508
6509#ifdef CONFIG_NUMA
6510
6511static int sched_domains_numa_levels;
6512static int *sched_domains_numa_distance;
6513static struct cpumask ***sched_domains_numa_masks;
6514static int sched_domains_curr_level;
6515
6516static inline int sd_local_flags(int level)
6517{
6518	if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6519		return 0;
6520
6521	return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6522}
6523
6524static struct sched_domain *
6525sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6526{
6527	struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6528	int level = tl->numa_level;
6529	int sd_weight = cpumask_weight(
6530			sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6531
6532	*sd = (struct sched_domain){
6533		.min_interval		= sd_weight,
6534		.max_interval		= 2*sd_weight,
6535		.busy_factor		= 32,
6536		.imbalance_pct		= 125,
6537		.cache_nice_tries	= 2,
6538		.busy_idx		= 3,
6539		.idle_idx		= 2,
6540		.newidle_idx		= 0,
6541		.wake_idx		= 0,
6542		.forkexec_idx		= 0,
6543
6544		.flags			= 1*SD_LOAD_BALANCE
6545					| 1*SD_BALANCE_NEWIDLE
6546					| 0*SD_BALANCE_EXEC
6547					| 0*SD_BALANCE_FORK
6548					| 0*SD_BALANCE_WAKE
6549					| 0*SD_WAKE_AFFINE
6550					| 0*SD_PREFER_LOCAL
6551					| 0*SD_SHARE_CPUPOWER
6552					| 0*SD_SHARE_PKG_RESOURCES
6553					| 1*SD_SERIALIZE
6554					| 0*SD_PREFER_SIBLING
6555					| sd_local_flags(level)
6556					,
6557		.last_balance		= jiffies,
6558		.balance_interval	= sd_weight,
6559	};
6560	SD_INIT_NAME(sd, NUMA);
6561	sd->private = &tl->data;
6562
6563	/*
6564	 * Ugly hack to pass state to sd_numa_mask()...
6565	 */
6566	sched_domains_curr_level = tl->numa_level;
6567
6568	return sd;
6569}
6570
6571static const struct cpumask *sd_numa_mask(int cpu)
6572{
6573	return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6574}
6575
6576static void sched_numa_warn(const char *str)
6577{
6578	static int done = false;
6579	int i,j;
6580
6581	if (done)
6582		return;
6583
6584	done = true;
6585
6586	printk(KERN_WARNING "ERROR: %s\n\n", str);
6587
6588	for (i = 0; i < nr_node_ids; i++) {
6589		printk(KERN_WARNING "  ");
6590		for (j = 0; j < nr_node_ids; j++)
6591			printk(KERN_CONT "%02d ", node_distance(i,j));
6592		printk(KERN_CONT "\n");
6593	}
6594	printk(KERN_WARNING "\n");
6595}
6596
6597static bool find_numa_distance(int distance)
6598{
6599	int i;
6600
6601	if (distance == node_distance(0, 0))
6602		return true;
6603
6604	for (i = 0; i < sched_domains_numa_levels; i++) {
6605		if (sched_domains_numa_distance[i] == distance)
6606			return true;
6607	}
6608
6609	return false;
6610}
6611
6612static void sched_init_numa(void)
6613{
6614	int next_distance, curr_distance = node_distance(0, 0);
6615	struct sched_domain_topology_level *tl;
6616	int level = 0;
6617	int i, j, k;
6618
6619	sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6620	if (!sched_domains_numa_distance)
6621		return;
6622
6623	/*
6624	 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6625	 * unique distances in the node_distance() table.
6626	 *
6627	 * Assumes node_distance(0,j) includes all distances in
6628	 * node_distance(i,j) in order to avoid cubic time.
6629	 */
6630	next_distance = curr_distance;
6631	for (i = 0; i < nr_node_ids; i++) {
6632		for (j = 0; j < nr_node_ids; j++) {
6633			for (k = 0; k < nr_node_ids; k++) {
6634				int distance = node_distance(i, k);
6635
6636				if (distance > curr_distance &&
6637				    (distance < next_distance ||
6638				     next_distance == curr_distance))
6639					next_distance = distance;
6640
6641				/*
6642				 * While not a strong assumption it would be nice to know
6643				 * about cases where if node A is connected to B, B is not
6644				 * equally connected to A.
6645				 */
6646				if (sched_debug() && node_distance(k, i) != distance)
6647					sched_numa_warn("Node-distance not symmetric");
6648
6649				if (sched_debug() && i && !find_numa_distance(distance))
6650					sched_numa_warn("Node-0 not representative");
6651			}
6652			if (next_distance != curr_distance) {
6653				sched_domains_numa_distance[level++] = next_distance;
6654				sched_domains_numa_levels = level;
6655				curr_distance = next_distance;
6656			} else break;
6657		}
6658
6659		/*
6660		 * In case of sched_debug() we verify the above assumption.
6661		 */
6662		if (!sched_debug())
6663			break;
6664	}
6665	/*
6666	 * 'level' contains the number of unique distances, excluding the
6667	 * identity distance node_distance(i,i).
6668	 *
6669	 * The sched_domains_nume_distance[] array includes the actual distance
6670	 * numbers.
6671	 */
6672
6673	sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6674	if (!sched_domains_numa_masks)
6675		return;
6676
6677	/*
6678	 * Now for each level, construct a mask per node which contains all
6679	 * cpus of nodes that are that many hops away from us.
6680	 */
6681	for (i = 0; i < level; i++) {
6682		sched_domains_numa_masks[i] =
6683			kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6684		if (!sched_domains_numa_masks[i])
6685			return;
6686
6687		for (j = 0; j < nr_node_ids; j++) {
6688			struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6689			if (!mask)
6690				return;
6691
6692			sched_domains_numa_masks[i][j] = mask;
6693
6694			for (k = 0; k < nr_node_ids; k++) {
6695				if (node_distance(j, k) > sched_domains_numa_distance[i])
6696					continue;
6697
6698				cpumask_or(mask, mask, cpumask_of_node(k));
6699			}
6700		}
6701	}
6702
6703	tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6704			sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6705	if (!tl)
6706		return;
6707
6708	/*
6709	 * Copy the default topology bits..
6710	 */
6711	for (i = 0; default_topology[i].init; i++)
6712		tl[i] = default_topology[i];
6713
6714	/*
6715	 * .. and append 'j' levels of NUMA goodness.
6716	 */
6717	for (j = 0; j < level; i++, j++) {
6718		tl[i] = (struct sched_domain_topology_level){
6719			.init = sd_numa_init,
6720			.mask = sd_numa_mask,
6721			.flags = SDTL_OVERLAP,
6722			.numa_level = j,
6723		};
6724	}
6725
6726	sched_domain_topology = tl;
6727}
6728#else
6729static inline void sched_init_numa(void)
6730{
6731}
6732#endif /* CONFIG_NUMA */
6733
6734static int __sdt_alloc(const struct cpumask *cpu_map)
6735{
6736	struct sched_domain_topology_level *tl;
6737	int j;
6738
6739	for (tl = sched_domain_topology; tl->init; tl++) {
6740		struct sd_data *sdd = &tl->data;
6741
6742		sdd->sd = alloc_percpu(struct sched_domain *);
6743		if (!sdd->sd)
6744			return -ENOMEM;
6745
6746		sdd->sg = alloc_percpu(struct sched_group *);
6747		if (!sdd->sg)
6748			return -ENOMEM;
6749
6750		sdd->sgp = alloc_percpu(struct sched_group_power *);
6751		if (!sdd->sgp)
6752			return -ENOMEM;
6753
6754		for_each_cpu(j, cpu_map) {
6755			struct sched_domain *sd;
6756			struct sched_group *sg;
6757			struct sched_group_power *sgp;
6758
6759		       	sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6760					GFP_KERNEL, cpu_to_node(j));
6761			if (!sd)
6762				return -ENOMEM;
6763
6764			*per_cpu_ptr(sdd->sd, j) = sd;
6765
6766			sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6767					GFP_KERNEL, cpu_to_node(j));
6768			if (!sg)
6769				return -ENOMEM;
6770
6771			sg->next = sg;
6772
6773			*per_cpu_ptr(sdd->sg, j) = sg;
6774
6775			sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6776					GFP_KERNEL, cpu_to_node(j));
6777			if (!sgp)
6778				return -ENOMEM;
6779
6780			*per_cpu_ptr(sdd->sgp, j) = sgp;
6781		}
6782	}
6783
6784	return 0;
6785}
6786
6787static void __sdt_free(const struct cpumask *cpu_map)
6788{
6789	struct sched_domain_topology_level *tl;
6790	int j;
6791
6792	for (tl = sched_domain_topology; tl->init; tl++) {
6793		struct sd_data *sdd = &tl->data;
6794
6795		for_each_cpu(j, cpu_map) {
6796			struct sched_domain *sd;
6797
6798			if (sdd->sd) {
6799				sd = *per_cpu_ptr(sdd->sd, j);
6800				if (sd && (sd->flags & SD_OVERLAP))
6801					free_sched_groups(sd->groups, 0);
6802				kfree(*per_cpu_ptr(sdd->sd, j));
6803			}
6804
6805			if (sdd->sg)
6806				kfree(*per_cpu_ptr(sdd->sg, j));
6807			if (sdd->sgp)
6808				kfree(*per_cpu_ptr(sdd->sgp, j));
6809		}
6810		free_percpu(sdd->sd);
6811		sdd->sd = NULL;
6812		free_percpu(sdd->sg);
6813		sdd->sg = NULL;
6814		free_percpu(sdd->sgp);
6815		sdd->sgp = NULL;
6816	}
6817}
6818
6819struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6820		struct s_data *d, const struct cpumask *cpu_map,
6821		struct sched_domain_attr *attr, struct sched_domain *child,
6822		int cpu)
6823{
6824	struct sched_domain *sd = tl->init(tl, cpu);
6825	if (!sd)
6826		return child;
6827
6828	cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6829	if (child) {
6830		sd->level = child->level + 1;
6831		sched_domain_level_max = max(sched_domain_level_max, sd->level);
6832		child->parent = sd;
6833	}
6834	sd->child = child;
6835	set_domain_attribute(sd, attr);
6836
6837	return sd;
6838}
6839
6840/*
6841 * Build sched domains for a given set of cpus and attach the sched domains
6842 * to the individual cpus
6843 */
6844static int build_sched_domains(const struct cpumask *cpu_map,
6845			       struct sched_domain_attr *attr)
6846{
6847	enum s_alloc alloc_state = sa_none;
6848	struct sched_domain *sd;
6849	struct s_data d;
6850	int i, ret = -ENOMEM;
6851
6852	alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6853	if (alloc_state != sa_rootdomain)
6854		goto error;
6855
6856	/* Set up domains for cpus specified by the cpu_map. */
6857	for_each_cpu(i, cpu_map) {
6858		struct sched_domain_topology_level *tl;
6859
6860		sd = NULL;
6861		for (tl = sched_domain_topology; tl->init; tl++) {
6862			sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6863			if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6864				sd->flags |= SD_OVERLAP;
6865			if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6866				break;
6867		}
6868
6869		while (sd->child)
6870			sd = sd->child;
6871
6872		*per_cpu_ptr(d.sd, i) = sd;
6873	}
6874
6875	/* Build the groups for the domains */
6876	for_each_cpu(i, cpu_map) {
6877		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6878			sd->span_weight = cpumask_weight(sched_domain_span(sd));
6879			if (sd->flags & SD_OVERLAP) {
6880				if (build_overlap_sched_groups(sd, i))
6881					goto error;
6882			} else {
6883				if (build_sched_groups(sd, i))
6884					goto error;
6885			}
6886		}
6887	}
6888
6889	/* Calculate CPU power for physical packages and nodes */
6890	for (i = nr_cpumask_bits-1; i >= 0; i--) {
6891		if (!cpumask_test_cpu(i, cpu_map))
6892			continue;
6893
6894		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6895			claim_allocations(i, sd);
6896			init_sched_groups_power(i, sd);
6897		}
6898	}
6899
6900	/* Attach the domains */
6901	rcu_read_lock();
6902	for_each_cpu(i, cpu_map) {
6903		sd = *per_cpu_ptr(d.sd, i);
6904		cpu_attach_domain(sd, d.rd, i);
6905	}
6906	rcu_read_unlock();
6907
6908	ret = 0;
6909error:
6910	__free_domain_allocs(&d, alloc_state, cpu_map);
6911	return ret;
6912}
6913
6914static cpumask_var_t *doms_cur;	/* current sched domains */
6915static int ndoms_cur;		/* number of sched domains in 'doms_cur' */
6916static struct sched_domain_attr *dattr_cur;
6917				/* attribues of custom domains in 'doms_cur' */
6918
6919/*
6920 * Special case: If a kmalloc of a doms_cur partition (array of
6921 * cpumask) fails, then fallback to a single sched domain,
6922 * as determined by the single cpumask fallback_doms.
6923 */
6924static cpumask_var_t fallback_doms;
6925
6926/*
6927 * arch_update_cpu_topology lets virtualized architectures update the
6928 * cpu core maps. It is supposed to return 1 if the topology changed
6929 * or 0 if it stayed the same.
6930 */
6931int __attribute__((weak)) arch_update_cpu_topology(void)
6932{
6933	return 0;
6934}
6935
6936cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6937{
6938	int i;
6939	cpumask_var_t *doms;
6940
6941	doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6942	if (!doms)
6943		return NULL;
6944	for (i = 0; i < ndoms; i++) {
6945		if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6946			free_sched_domains(doms, i);
6947			return NULL;
6948		}
6949	}
6950	return doms;
6951}
6952
6953void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6954{
6955	unsigned int i;
6956	for (i = 0; i < ndoms; i++)
6957		free_cpumask_var(doms[i]);
6958	kfree(doms);
6959}
6960
6961/*
6962 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6963 * For now this just excludes isolated cpus, but could be used to
6964 * exclude other special cases in the future.
6965 */
6966static int init_sched_domains(const struct cpumask *cpu_map)
6967{
6968	int err;
6969
6970	arch_update_cpu_topology();
6971	ndoms_cur = 1;
6972	doms_cur = alloc_sched_domains(ndoms_cur);
6973	if (!doms_cur)
6974		doms_cur = &fallback_doms;
6975	cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6976	err = build_sched_domains(doms_cur[0], NULL);
6977	register_sched_domain_sysctl();
6978
6979	return err;
6980}
6981
6982/*
6983 * Detach sched domains from a group of cpus specified in cpu_map
6984 * These cpus will now be attached to the NULL domain
6985 */
6986static void detach_destroy_domains(const struct cpumask *cpu_map)
6987{
6988	int i;
6989
6990	rcu_read_lock();
6991	for_each_cpu(i, cpu_map)
6992		cpu_attach_domain(NULL, &def_root_domain, i);
6993	rcu_read_unlock();
6994}
6995
6996/* handle null as "default" */
6997static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6998			struct sched_domain_attr *new, int idx_new)
6999{
7000	struct sched_domain_attr tmp;
7001
7002	/* fast path */
7003	if (!new && !cur)
7004		return 1;
7005
7006	tmp = SD_ATTR_INIT;
7007	return !memcmp(cur ? (cur + idx_cur) : &tmp,
7008			new ? (new + idx_new) : &tmp,
7009			sizeof(struct sched_domain_attr));
7010}
7011
7012/*
7013 * Partition sched domains as specified by the 'ndoms_new'
7014 * cpumasks in the array doms_new[] of cpumasks. This compares
7015 * doms_new[] to the current sched domain partitioning, doms_cur[].
7016 * It destroys each deleted domain and builds each new domain.
7017 *
7018 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7019 * The masks don't intersect (don't overlap.) We should setup one
7020 * sched domain for each mask. CPUs not in any of the cpumasks will
7021 * not be load balanced. If the same cpumask appears both in the
7022 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7023 * it as it is.
7024 *
7025 * The passed in 'doms_new' should be allocated using
7026 * alloc_sched_domains.  This routine takes ownership of it and will
7027 * free_sched_domains it when done with it. If the caller failed the
7028 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7029 * and partition_sched_domains() will fallback to the single partition
7030 * 'fallback_doms', it also forces the domains to be rebuilt.
7031 *
7032 * If doms_new == NULL it will be replaced with cpu_online_mask.
7033 * ndoms_new == 0 is a special case for destroying existing domains,
7034 * and it will not create the default domain.
7035 *
7036 * Call with hotplug lock held
7037 */
7038void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7039			     struct sched_domain_attr *dattr_new)
7040{
7041	int i, j, n;
7042	int new_topology;
7043
7044	mutex_lock(&sched_domains_mutex);
7045
7046	/* always unregister in case we don't destroy any domains */
7047	unregister_sched_domain_sysctl();
7048
7049	/* Let architecture update cpu core mappings. */
7050	new_topology = arch_update_cpu_topology();
7051
7052	n = doms_new ? ndoms_new : 0;
7053
7054	/* Destroy deleted domains */
7055	for (i = 0; i < ndoms_cur; i++) {
7056		for (j = 0; j < n && !new_topology; j++) {
7057			if (cpumask_equal(doms_cur[i], doms_new[j])
7058			    && dattrs_equal(dattr_cur, i, dattr_new, j))
7059				goto match1;
7060		}
7061		/* no match - a current sched domain not in new doms_new[] */
7062		detach_destroy_domains(doms_cur[i]);
7063match1:
7064		;
7065	}
7066
7067	if (doms_new == NULL) {
7068		ndoms_cur = 0;
7069		doms_new = &fallback_doms;
7070		cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7071		WARN_ON_ONCE(dattr_new);
7072	}
7073
7074	/* Build new domains */
7075	for (i = 0; i < ndoms_new; i++) {
7076		for (j = 0; j < ndoms_cur && !new_topology; j++) {
7077			if (cpumask_equal(doms_new[i], doms_cur[j])
7078			    && dattrs_equal(dattr_new, i, dattr_cur, j))
7079				goto match2;
7080		}
7081		/* no match - add a new doms_new */
7082		build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7083match2:
7084		;
7085	}
7086
7087	/* Remember the new sched domains */
7088	if (doms_cur != &fallback_doms)
7089		free_sched_domains(doms_cur, ndoms_cur);
7090	kfree(dattr_cur);	/* kfree(NULL) is safe */
7091	doms_cur = doms_new;
7092	dattr_cur = dattr_new;
7093	ndoms_cur = ndoms_new;
7094
7095	register_sched_domain_sysctl();
7096
7097	mutex_unlock(&sched_domains_mutex);
7098}
7099
7100static int num_cpus_frozen;	/* used to mark begin/end of suspend/resume */
7101
7102/*
7103 * Update cpusets according to cpu_active mask.  If cpusets are
7104 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7105 * around partition_sched_domains().
7106 *
7107 * If we come here as part of a suspend/resume, don't touch cpusets because we
7108 * want to restore it back to its original state upon resume anyway.
7109 */
7110static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7111			     void *hcpu)
7112{
7113	switch (action) {
7114	case CPU_ONLINE_FROZEN:
7115	case CPU_DOWN_FAILED_FROZEN:
7116
7117		/*
7118		 * num_cpus_frozen tracks how many CPUs are involved in suspend
7119		 * resume sequence. As long as this is not the last online
7120		 * operation in the resume sequence, just build a single sched
7121		 * domain, ignoring cpusets.
7122		 */
7123		num_cpus_frozen--;
7124		if (likely(num_cpus_frozen)) {
7125			partition_sched_domains(1, NULL, NULL);
7126			break;
7127		}
7128
7129		/*
7130		 * This is the last CPU online operation. So fall through and
7131		 * restore the original sched domains by considering the
7132		 * cpuset configurations.
7133		 */
7134
7135	case CPU_ONLINE:
7136	case CPU_DOWN_FAILED:
7137		cpuset_update_active_cpus();
7138		break;
7139	default:
7140		return NOTIFY_DONE;
7141	}
7142	return NOTIFY_OK;
7143}
7144
7145static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7146			       void *hcpu)
7147{
7148	switch (action) {
7149	case CPU_DOWN_PREPARE:
7150		cpuset_update_active_cpus();
7151		break;
7152	case CPU_DOWN_PREPARE_FROZEN:
7153		num_cpus_frozen++;
7154		partition_sched_domains(1, NULL, NULL);
7155		break;
7156	default:
7157		return NOTIFY_DONE;
7158	}
7159	return NOTIFY_OK;
7160}
7161
7162void __init sched_init_smp(void)
7163{
7164	cpumask_var_t non_isolated_cpus;
7165
7166	alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7167	alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7168
7169	sched_init_numa();
7170
7171	get_online_cpus();
7172	mutex_lock(&sched_domains_mutex);
7173	init_sched_domains(cpu_active_mask);
7174	cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7175	if (cpumask_empty(non_isolated_cpus))
7176		cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7177	mutex_unlock(&sched_domains_mutex);
7178	put_online_cpus();
7179
7180	hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7181	hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7182
7183	/* RT runtime code needs to handle some hotplug events */
7184	hotcpu_notifier(update_runtime, 0);
7185
7186	init_hrtick();
7187
7188	/* Move init over to a non-isolated CPU */
7189	if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7190		BUG();
7191	sched_init_granularity();
7192	free_cpumask_var(non_isolated_cpus);
7193
7194	init_sched_rt_class();
7195}
7196#else
7197void __init sched_init_smp(void)
7198{
7199	sched_init_granularity();
7200}
7201#endif /* CONFIG_SMP */
7202
7203const_debug unsigned int sysctl_timer_migration = 1;
7204
7205int in_sched_functions(unsigned long addr)
7206{
7207	return in_lock_functions(addr) ||
7208		(addr >= (unsigned long)__sched_text_start
7209		&& addr < (unsigned long)__sched_text_end);
7210}
7211
7212#ifdef CONFIG_CGROUP_SCHED
7213struct task_group root_task_group;
7214#endif
7215
7216DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
7217
7218void __init sched_init(void)
7219{
7220	int i, j;
7221	unsigned long alloc_size = 0, ptr;
7222
7223#ifdef CONFIG_FAIR_GROUP_SCHED
7224	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7225#endif
7226#ifdef CONFIG_RT_GROUP_SCHED
7227	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7228#endif
7229#ifdef CONFIG_CPUMASK_OFFSTACK
7230	alloc_size += num_possible_cpus() * cpumask_size();
7231#endif
7232	if (alloc_size) {
7233		ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7234
7235#ifdef CONFIG_FAIR_GROUP_SCHED
7236		root_task_group.se = (struct sched_entity **)ptr;
7237		ptr += nr_cpu_ids * sizeof(void **);
7238
7239		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7240		ptr += nr_cpu_ids * sizeof(void **);
7241
7242#endif /* CONFIG_FAIR_GROUP_SCHED */
7243#ifdef CONFIG_RT_GROUP_SCHED
7244		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7245		ptr += nr_cpu_ids * sizeof(void **);
7246
7247		root_task_group.rt_rq = (struct rt_rq **)ptr;
7248		ptr += nr_cpu_ids * sizeof(void **);
7249
7250#endif /* CONFIG_RT_GROUP_SCHED */
7251#ifdef CONFIG_CPUMASK_OFFSTACK
7252		for_each_possible_cpu(i) {
7253			per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7254			ptr += cpumask_size();
7255		}
7256#endif /* CONFIG_CPUMASK_OFFSTACK */
7257	}
7258
7259#ifdef CONFIG_SMP
7260	init_defrootdomain();
7261#endif
7262
7263	init_rt_bandwidth(&def_rt_bandwidth,
7264			global_rt_period(), global_rt_runtime());
7265
7266#ifdef CONFIG_RT_GROUP_SCHED
7267	init_rt_bandwidth(&root_task_group.rt_bandwidth,
7268			global_rt_period(), global_rt_runtime());
7269#endif /* CONFIG_RT_GROUP_SCHED */
7270
7271#ifdef CONFIG_CGROUP_SCHED
7272	list_add(&root_task_group.list, &task_groups);
7273	INIT_LIST_HEAD(&root_task_group.children);
7274	INIT_LIST_HEAD(&root_task_group.siblings);
7275	autogroup_init(&init_task);
7276
7277#endif /* CONFIG_CGROUP_SCHED */
7278
7279#ifdef CONFIG_CGROUP_CPUACCT
7280	root_cpuacct.cpustat = &kernel_cpustat;
7281	root_cpuacct.cpuusage = alloc_percpu(u64);
7282	/* Too early, not expected to fail */
7283	BUG_ON(!root_cpuacct.cpuusage);
7284#endif
7285	for_each_possible_cpu(i) {
7286		struct rq *rq;
7287
7288		rq = cpu_rq(i);
7289		raw_spin_lock_init(&rq->lock);
7290		rq->nr_running = 0;
7291		rq->calc_load_active = 0;
7292		rq->calc_load_update = jiffies + LOAD_FREQ;
7293		init_cfs_rq(&rq->cfs);
7294		init_rt_rq(&rq->rt, rq);
7295#ifdef CONFIG_FAIR_GROUP_SCHED
7296		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7297		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7298		/*
7299		 * How much cpu bandwidth does root_task_group get?
7300		 *
7301		 * In case of task-groups formed thr' the cgroup filesystem, it
7302		 * gets 100% of the cpu resources in the system. This overall
7303		 * system cpu resource is divided among the tasks of
7304		 * root_task_group and its child task-groups in a fair manner,
7305		 * based on each entity's (task or task-group's) weight
7306		 * (se->load.weight).
7307		 *
7308		 * In other words, if root_task_group has 10 tasks of weight
7309		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7310		 * then A0's share of the cpu resource is:
7311		 *
7312		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7313		 *
7314		 * We achieve this by letting root_task_group's tasks sit
7315		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7316		 */
7317		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7318		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7319#endif /* CONFIG_FAIR_GROUP_SCHED */
7320
7321		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7322#ifdef CONFIG_RT_GROUP_SCHED
7323		INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7324		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7325#endif
7326
7327		for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7328			rq->cpu_load[j] = 0;
7329
7330		rq->last_load_update_tick = jiffies;
7331
7332#ifdef CONFIG_SMP
7333		rq->sd = NULL;
7334		rq->rd = NULL;
7335		rq->cpu_power = SCHED_POWER_SCALE;
7336		rq->post_schedule = 0;
7337		rq->active_balance = 0;
7338		rq->next_balance = jiffies;
7339		rq->push_cpu = 0;
7340		rq->cpu = i;
7341		rq->online = 0;
7342		rq->idle_stamp = 0;
7343		rq->avg_idle = 2*sysctl_sched_migration_cost;
7344
7345		INIT_LIST_HEAD(&rq->cfs_tasks);
7346
7347		rq_attach_root(rq, &def_root_domain);
7348#ifdef CONFIG_NO_HZ
7349		rq->nohz_flags = 0;
7350#endif
7351#endif
7352		init_rq_hrtick(rq);
7353		atomic_set(&rq->nr_iowait, 0);
7354	}
7355
7356	set_load_weight(&init_task);
7357
7358#ifdef CONFIG_PREEMPT_NOTIFIERS
7359	INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7360#endif
7361
7362#ifdef CONFIG_RT_MUTEXES
7363	plist_head_init(&init_task.pi_waiters);
7364#endif
7365
7366	/*
7367	 * The boot idle thread does lazy MMU switching as well:
7368	 */
7369	atomic_inc(&init_mm.mm_count);
7370	enter_lazy_tlb(&init_mm, current);
7371
7372	/*
7373	 * Make us the idle thread. Technically, schedule() should not be
7374	 * called from this thread, however somewhere below it might be,
7375	 * but because we are the idle thread, we just pick up running again
7376	 * when this runqueue becomes "idle".
7377	 */
7378	init_idle(current, smp_processor_id());
7379
7380	calc_load_update = jiffies + LOAD_FREQ;
7381
7382	/*
7383	 * During early bootup we pretend to be a normal task:
7384	 */
7385	current->sched_class = &fair_sched_class;
7386
7387#ifdef CONFIG_SMP
7388	zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7389	/* May be allocated at isolcpus cmdline parse time */
7390	if (cpu_isolated_map == NULL)
7391		zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7392	idle_thread_set_boot_cpu();
7393#endif
7394	init_sched_fair_class();
7395
7396	scheduler_running = 1;
7397}
7398
7399#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7400static inline int preempt_count_equals(int preempt_offset)
7401{
7402	int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7403
7404	return (nested == preempt_offset);
7405}
7406
7407void __might_sleep(const char *file, int line, int preempt_offset)
7408{
7409	static unsigned long prev_jiffy;	/* ratelimiting */
7410
7411	rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7412	if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7413	    system_state != SYSTEM_RUNNING || oops_in_progress)
7414		return;
7415	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7416		return;
7417	prev_jiffy = jiffies;
7418
7419	printk(KERN_ERR
7420		"BUG: sleeping function called from invalid context at %s:%d\n",
7421			file, line);
7422	printk(KERN_ERR
7423		"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7424			in_atomic(), irqs_disabled(),
7425			current->pid, current->comm);
7426
7427	debug_show_held_locks(current);
7428	if (irqs_disabled())
7429		print_irqtrace_events(current);
7430	dump_stack();
7431}
7432EXPORT_SYMBOL(__might_sleep);
7433#endif
7434
7435#ifdef CONFIG_MAGIC_SYSRQ
7436static void normalize_task(struct rq *rq, struct task_struct *p)
7437{
7438	const struct sched_class *prev_class = p->sched_class;
7439	int old_prio = p->prio;
7440	int on_rq;
7441
7442	on_rq = p->on_rq;
7443	if (on_rq)
7444		dequeue_task(rq, p, 0);
7445	__setscheduler(rq, p, SCHED_NORMAL, 0);
7446	if (on_rq) {
7447		enqueue_task(rq, p, 0);
7448		resched_task(rq->curr);
7449	}
7450
7451	check_class_changed(rq, p, prev_class, old_prio);
7452}
7453
7454void normalize_rt_tasks(void)
7455{
7456	struct task_struct *g, *p;
7457	unsigned long flags;
7458	struct rq *rq;
7459
7460	read_lock_irqsave(&tasklist_lock, flags);
7461	do_each_thread(g, p) {
7462		/*
7463		 * Only normalize user tasks:
7464		 */
7465		if (!p->mm)
7466			continue;
7467
7468		p->se.exec_start		= 0;
7469#ifdef CONFIG_SCHEDSTATS
7470		p->se.statistics.wait_start	= 0;
7471		p->se.statistics.sleep_start	= 0;
7472		p->se.statistics.block_start	= 0;
7473#endif
7474
7475		if (!rt_task(p)) {
7476			/*
7477			 * Renice negative nice level userspace
7478			 * tasks back to 0:
7479			 */
7480			if (TASK_NICE(p) < 0 && p->mm)
7481				set_user_nice(p, 0);
7482			continue;
7483		}
7484
7485		raw_spin_lock(&p->pi_lock);
7486		rq = __task_rq_lock(p);
7487
7488		normalize_task(rq, p);
7489
7490		__task_rq_unlock(rq);
7491		raw_spin_unlock(&p->pi_lock);
7492	} while_each_thread(g, p);
7493
7494	read_unlock_irqrestore(&tasklist_lock, flags);
7495}
7496
7497#endif /* CONFIG_MAGIC_SYSRQ */
7498
7499#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7500/*
7501 * These functions are only useful for the IA64 MCA handling, or kdb.
7502 *
7503 * They can only be called when the whole system has been
7504 * stopped - every CPU needs to be quiescent, and no scheduling
7505 * activity can take place. Using them for anything else would
7506 * be a serious bug, and as a result, they aren't even visible
7507 * under any other configuration.
7508 */
7509
7510/**
7511 * curr_task - return the current task for a given cpu.
7512 * @cpu: the processor in question.
7513 *
7514 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7515 */
7516struct task_struct *curr_task(int cpu)
7517{
7518	return cpu_curr(cpu);
7519}
7520
7521#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7522
7523#ifdef CONFIG_IA64
7524/**
7525 * set_curr_task - set the current task for a given cpu.
7526 * @cpu: the processor in question.
7527 * @p: the task pointer to set.
7528 *
7529 * Description: This function must only be used when non-maskable interrupts
7530 * are serviced on a separate stack. It allows the architecture to switch the
7531 * notion of the current task on a cpu in a non-blocking manner. This function
7532 * must be called with all CPU's synchronized, and interrupts disabled, the
7533 * and caller must save the original value of the current task (see
7534 * curr_task() above) and restore that value before reenabling interrupts and
7535 * re-starting the system.
7536 *
7537 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7538 */
7539void set_curr_task(int cpu, struct task_struct *p)
7540{
7541	cpu_curr(cpu) = p;
7542}
7543
7544#endif
7545
7546#ifdef CONFIG_CGROUP_SCHED
7547/* task_group_lock serializes the addition/removal of task groups */
7548static DEFINE_SPINLOCK(task_group_lock);
7549
7550static void free_sched_group(struct task_group *tg)
7551{
7552	free_fair_sched_group(tg);
7553	free_rt_sched_group(tg);
7554	autogroup_free(tg);
7555	kfree(tg);
7556}
7557
7558/* allocate runqueue etc for a new task group */
7559struct task_group *sched_create_group(struct task_group *parent)
7560{
7561	struct task_group *tg;
7562	unsigned long flags;
7563
7564	tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7565	if (!tg)
7566		return ERR_PTR(-ENOMEM);
7567
7568	if (!alloc_fair_sched_group(tg, parent))
7569		goto err;
7570
7571	if (!alloc_rt_sched_group(tg, parent))
7572		goto err;
7573
7574	spin_lock_irqsave(&task_group_lock, flags);
7575	list_add_rcu(&tg->list, &task_groups);
7576
7577	WARN_ON(!parent); /* root should already exist */
7578
7579	tg->parent = parent;
7580	INIT_LIST_HEAD(&tg->children);
7581	list_add_rcu(&tg->siblings, &parent->children);
7582	spin_unlock_irqrestore(&task_group_lock, flags);
7583
7584	return tg;
7585
7586err:
7587	free_sched_group(tg);
7588	return ERR_PTR(-ENOMEM);
7589}
7590
7591/* rcu callback to free various structures associated with a task group */
7592static void free_sched_group_rcu(struct rcu_head *rhp)
7593{
7594	/* now it should be safe to free those cfs_rqs */
7595	free_sched_group(container_of(rhp, struct task_group, rcu));
7596}
7597
7598/* Destroy runqueue etc associated with a task group */
7599void sched_destroy_group(struct task_group *tg)
7600{
7601	unsigned long flags;
7602	int i;
7603
7604	/* end participation in shares distribution */
7605	for_each_possible_cpu(i)
7606		unregister_fair_sched_group(tg, i);
7607
7608	spin_lock_irqsave(&task_group_lock, flags);
7609	list_del_rcu(&tg->list);
7610	list_del_rcu(&tg->siblings);
7611	spin_unlock_irqrestore(&task_group_lock, flags);
7612
7613	/* wait for possible concurrent references to cfs_rqs complete */
7614	call_rcu(&tg->rcu, free_sched_group_rcu);
7615}
7616
7617/* change task's runqueue when it moves between groups.
7618 *	The caller of this function should have put the task in its new group
7619 *	by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7620 *	reflect its new group.
7621 */
7622void sched_move_task(struct task_struct *tsk)
7623{
7624	int on_rq, running;
7625	unsigned long flags;
7626	struct rq *rq;
7627
7628	rq = task_rq_lock(tsk, &flags);
7629
7630	running = task_current(rq, tsk);
7631	on_rq = tsk->on_rq;
7632
7633	if (on_rq)
7634		dequeue_task(rq, tsk, 0);
7635	if (unlikely(running))
7636		tsk->sched_class->put_prev_task(rq, tsk);
7637
7638#ifdef CONFIG_FAIR_GROUP_SCHED
7639	if (tsk->sched_class->task_move_group)
7640		tsk->sched_class->task_move_group(tsk, on_rq);
7641	else
7642#endif
7643		set_task_rq(tsk, task_cpu(tsk));
7644
7645	if (unlikely(running))
7646		tsk->sched_class->set_curr_task(rq);
7647	if (on_rq)
7648		enqueue_task(rq, tsk, 0);
7649
7650	task_rq_unlock(rq, tsk, &flags);
7651}
7652#endif /* CONFIG_CGROUP_SCHED */
7653
7654#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7655static unsigned long to_ratio(u64 period, u64 runtime)
7656{
7657	if (runtime == RUNTIME_INF)
7658		return 1ULL << 20;
7659
7660	return div64_u64(runtime << 20, period);
7661}
7662#endif
7663
7664#ifdef CONFIG_RT_GROUP_SCHED
7665/*
7666 * Ensure that the real time constraints are schedulable.
7667 */
7668static DEFINE_MUTEX(rt_constraints_mutex);
7669
7670/* Must be called with tasklist_lock held */
7671static inline int tg_has_rt_tasks(struct task_group *tg)
7672{
7673	struct task_struct *g, *p;
7674
7675	do_each_thread(g, p) {
7676		if (rt_task(p) && task_rq(p)->rt.tg == tg)
7677			return 1;
7678	} while_each_thread(g, p);
7679
7680	return 0;
7681}
7682
7683struct rt_schedulable_data {
7684	struct task_group *tg;
7685	u64 rt_period;
7686	u64 rt_runtime;
7687};
7688
7689static int tg_rt_schedulable(struct task_group *tg, void *data)
7690{
7691	struct rt_schedulable_data *d = data;
7692	struct task_group *child;
7693	unsigned long total, sum = 0;
7694	u64 period, runtime;
7695
7696	period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7697	runtime = tg->rt_bandwidth.rt_runtime;
7698
7699	if (tg == d->tg) {
7700		period = d->rt_period;
7701		runtime = d->rt_runtime;
7702	}
7703
7704	/*
7705	 * Cannot have more runtime than the period.
7706	 */
7707	if (runtime > period && runtime != RUNTIME_INF)
7708		return -EINVAL;
7709
7710	/*
7711	 * Ensure we don't starve existing RT tasks.
7712	 */
7713	if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7714		return -EBUSY;
7715
7716	total = to_ratio(period, runtime);
7717
7718	/*
7719	 * Nobody can have more than the global setting allows.
7720	 */
7721	if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7722		return -EINVAL;
7723
7724	/*
7725	 * The sum of our children's runtime should not exceed our own.
7726	 */
7727	list_for_each_entry_rcu(child, &tg->children, siblings) {
7728		period = ktime_to_ns(child->rt_bandwidth.rt_period);
7729		runtime = child->rt_bandwidth.rt_runtime;
7730
7731		if (child == d->tg) {
7732			period = d->rt_period;
7733			runtime = d->rt_runtime;
7734		}
7735
7736		sum += to_ratio(period, runtime);
7737	}
7738
7739	if (sum > total)
7740		return -EINVAL;
7741
7742	return 0;
7743}
7744
7745static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7746{
7747	int ret;
7748
7749	struct rt_schedulable_data data = {
7750		.tg = tg,
7751		.rt_period = period,
7752		.rt_runtime = runtime,
7753	};
7754
7755	rcu_read_lock();
7756	ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7757	rcu_read_unlock();
7758
7759	return ret;
7760}
7761
7762static int tg_set_rt_bandwidth(struct task_group *tg,
7763		u64 rt_period, u64 rt_runtime)
7764{
7765	int i, err = 0;
7766
7767	mutex_lock(&rt_constraints_mutex);
7768	read_lock(&tasklist_lock);
7769	err = __rt_schedulable(tg, rt_period, rt_runtime);
7770	if (err)
7771		goto unlock;
7772
7773	raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7774	tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7775	tg->rt_bandwidth.rt_runtime = rt_runtime;
7776
7777	for_each_possible_cpu(i) {
7778		struct rt_rq *rt_rq = tg->rt_rq[i];
7779
7780		raw_spin_lock(&rt_rq->rt_runtime_lock);
7781		rt_rq->rt_runtime = rt_runtime;
7782		raw_spin_unlock(&rt_rq->rt_runtime_lock);
7783	}
7784	raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7785unlock:
7786	read_unlock(&tasklist_lock);
7787	mutex_unlock(&rt_constraints_mutex);
7788
7789	return err;
7790}
7791
7792int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7793{
7794	u64 rt_runtime, rt_period;
7795
7796	rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7797	rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7798	if (rt_runtime_us < 0)
7799		rt_runtime = RUNTIME_INF;
7800
7801	return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7802}
7803
7804long sched_group_rt_runtime(struct task_group *tg)
7805{
7806	u64 rt_runtime_us;
7807
7808	if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7809		return -1;
7810
7811	rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7812	do_div(rt_runtime_us, NSEC_PER_USEC);
7813	return rt_runtime_us;
7814}
7815
7816int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7817{
7818	u64 rt_runtime, rt_period;
7819
7820	rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7821	rt_runtime = tg->rt_bandwidth.rt_runtime;
7822
7823	if (rt_period == 0)
7824		return -EINVAL;
7825
7826	return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7827}
7828
7829long sched_group_rt_period(struct task_group *tg)
7830{
7831	u64 rt_period_us;
7832
7833	rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7834	do_div(rt_period_us, NSEC_PER_USEC);
7835	return rt_period_us;
7836}
7837
7838static int sched_rt_global_constraints(void)
7839{
7840	u64 runtime, period;
7841	int ret = 0;
7842
7843	if (sysctl_sched_rt_period <= 0)
7844		return -EINVAL;
7845
7846	runtime = global_rt_runtime();
7847	period = global_rt_period();
7848
7849	/*
7850	 * Sanity check on the sysctl variables.
7851	 */
7852	if (runtime > period && runtime != RUNTIME_INF)
7853		return -EINVAL;
7854
7855	mutex_lock(&rt_constraints_mutex);
7856	read_lock(&tasklist_lock);
7857	ret = __rt_schedulable(NULL, 0, 0);
7858	read_unlock(&tasklist_lock);
7859	mutex_unlock(&rt_constraints_mutex);
7860
7861	return ret;
7862}
7863
7864int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7865{
7866	/* Don't accept realtime tasks when there is no way for them to run */
7867	if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7868		return 0;
7869
7870	return 1;
7871}
7872
7873#else /* !CONFIG_RT_GROUP_SCHED */
7874static int sched_rt_global_constraints(void)
7875{
7876	unsigned long flags;
7877	int i;
7878
7879	if (sysctl_sched_rt_period <= 0)
7880		return -EINVAL;
7881
7882	/*
7883	 * There's always some RT tasks in the root group
7884	 * -- migration, kstopmachine etc..
7885	 */
7886	if (sysctl_sched_rt_runtime == 0)
7887		return -EBUSY;
7888
7889	raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7890	for_each_possible_cpu(i) {
7891		struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7892
7893		raw_spin_lock(&rt_rq->rt_runtime_lock);
7894		rt_rq->rt_runtime = global_rt_runtime();
7895		raw_spin_unlock(&rt_rq->rt_runtime_lock);
7896	}
7897	raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7898
7899	return 0;
7900}
7901#endif /* CONFIG_RT_GROUP_SCHED */
7902
7903int sched_rt_handler(struct ctl_table *table, int write,
7904		void __user *buffer, size_t *lenp,
7905		loff_t *ppos)
7906{
7907	int ret;
7908	int old_period, old_runtime;
7909	static DEFINE_MUTEX(mutex);
7910
7911	mutex_lock(&mutex);
7912	old_period = sysctl_sched_rt_period;
7913	old_runtime = sysctl_sched_rt_runtime;
7914
7915	ret = proc_dointvec(table, write, buffer, lenp, ppos);
7916
7917	if (!ret && write) {
7918		ret = sched_rt_global_constraints();
7919		if (ret) {
7920			sysctl_sched_rt_period = old_period;
7921			sysctl_sched_rt_runtime = old_runtime;
7922		} else {
7923			def_rt_bandwidth.rt_runtime = global_rt_runtime();
7924			def_rt_bandwidth.rt_period =
7925				ns_to_ktime(global_rt_period());
7926		}
7927	}
7928	mutex_unlock(&mutex);
7929
7930	return ret;
7931}
7932
7933#ifdef CONFIG_CGROUP_SCHED
7934
7935/* return corresponding task_group object of a cgroup */
7936static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7937{
7938	return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7939			    struct task_group, css);
7940}
7941
7942static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7943{
7944	struct task_group *tg, *parent;
7945
7946	if (!cgrp->parent) {
7947		/* This is early initialization for the top cgroup */
7948		return &root_task_group.css;
7949	}
7950
7951	parent = cgroup_tg(cgrp->parent);
7952	tg = sched_create_group(parent);
7953	if (IS_ERR(tg))
7954		return ERR_PTR(-ENOMEM);
7955
7956	return &tg->css;
7957}
7958
7959static void cpu_cgroup_destroy(struct cgroup *cgrp)
7960{
7961	struct task_group *tg = cgroup_tg(cgrp);
7962
7963	sched_destroy_group(tg);
7964}
7965
7966static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7967				 struct cgroup_taskset *tset)
7968{
7969	struct task_struct *task;
7970
7971	cgroup_taskset_for_each(task, cgrp, tset) {
7972#ifdef CONFIG_RT_GROUP_SCHED
7973		if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7974			return -EINVAL;
7975#else
7976		/* We don't support RT-tasks being in separate groups */
7977		if (task->sched_class != &fair_sched_class)
7978			return -EINVAL;
7979#endif
7980	}
7981	return 0;
7982}
7983
7984static void cpu_cgroup_attach(struct cgroup *cgrp,
7985			      struct cgroup_taskset *tset)
7986{
7987	struct task_struct *task;
7988
7989	cgroup_taskset_for_each(task, cgrp, tset)
7990		sched_move_task(task);
7991}
7992
7993static void
7994cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7995		struct task_struct *task)
7996{
7997	/*
7998	 * cgroup_exit() is called in the copy_process() failure path.
7999	 * Ignore this case since the task hasn't ran yet, this avoids
8000	 * trying to poke a half freed task state from generic code.
8001	 */
8002	if (!(task->flags & PF_EXITING))
8003		return;
8004
8005	sched_move_task(task);
8006}
8007
8008#ifdef CONFIG_FAIR_GROUP_SCHED
8009static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8010				u64 shareval)
8011{
8012	return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
8013}
8014
8015static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8016{
8017	struct task_group *tg = cgroup_tg(cgrp);
8018
8019	return (u64) scale_load_down(tg->shares);
8020}
8021
8022#ifdef CONFIG_CFS_BANDWIDTH
8023static DEFINE_MUTEX(cfs_constraints_mutex);
8024
8025const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8026const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8027
8028static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8029
8030static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8031{
8032	int i, ret = 0, runtime_enabled, runtime_was_enabled;
8033	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8034
8035	if (tg == &root_task_group)
8036		return -EINVAL;
8037
8038	/*
8039	 * Ensure we have at some amount of bandwidth every period.  This is
8040	 * to prevent reaching a state of large arrears when throttled via
8041	 * entity_tick() resulting in prolonged exit starvation.
8042	 */
8043	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8044		return -EINVAL;
8045
8046	/*
8047	 * Likewise, bound things on the otherside by preventing insane quota
8048	 * periods.  This also allows us to normalize in computing quota
8049	 * feasibility.
8050	 */
8051	if (period > max_cfs_quota_period)
8052		return -EINVAL;
8053
8054	mutex_lock(&cfs_constraints_mutex);
8055	ret = __cfs_schedulable(tg, period, quota);
8056	if (ret)
8057		goto out_unlock;
8058
8059	runtime_enabled = quota != RUNTIME_INF;
8060	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8061	account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
8062	raw_spin_lock_irq(&cfs_b->lock);
8063	cfs_b->period = ns_to_ktime(period);
8064	cfs_b->quota = quota;
8065
8066	__refill_cfs_bandwidth_runtime(cfs_b);
8067	/* restart the period timer (if active) to handle new period expiry */
8068	if (runtime_enabled && cfs_b->timer_active) {
8069		/* force a reprogram */
8070		cfs_b->timer_active = 0;
8071		__start_cfs_bandwidth(cfs_b);
8072	}
8073	raw_spin_unlock_irq(&cfs_b->lock);
8074
8075	for_each_possible_cpu(i) {
8076		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8077		struct rq *rq = cfs_rq->rq;
8078
8079		raw_spin_lock_irq(&rq->lock);
8080		cfs_rq->runtime_enabled = runtime_enabled;
8081		cfs_rq->runtime_remaining = 0;
8082
8083		if (cfs_rq->throttled)
8084			unthrottle_cfs_rq(cfs_rq);
8085		raw_spin_unlock_irq(&rq->lock);
8086	}
8087out_unlock:
8088	mutex_unlock(&cfs_constraints_mutex);
8089
8090	return ret;
8091}
8092
8093int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8094{
8095	u64 quota, period;
8096
8097	period = ktime_to_ns(tg->cfs_bandwidth.period);
8098	if (cfs_quota_us < 0)
8099		quota = RUNTIME_INF;
8100	else
8101		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8102
8103	return tg_set_cfs_bandwidth(tg, period, quota);
8104}
8105
8106long tg_get_cfs_quota(struct task_group *tg)
8107{
8108	u64 quota_us;
8109
8110	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8111		return -1;
8112
8113	quota_us = tg->cfs_bandwidth.quota;
8114	do_div(quota_us, NSEC_PER_USEC);
8115
8116	return quota_us;
8117}
8118
8119int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8120{
8121	u64 quota, period;
8122
8123	period = (u64)cfs_period_us * NSEC_PER_USEC;
8124	quota = tg->cfs_bandwidth.quota;
8125
8126	return tg_set_cfs_bandwidth(tg, period, quota);
8127}
8128
8129long tg_get_cfs_period(struct task_group *tg)
8130{
8131	u64 cfs_period_us;
8132
8133	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8134	do_div(cfs_period_us, NSEC_PER_USEC);
8135
8136	return cfs_period_us;
8137}
8138
8139static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
8140{
8141	return tg_get_cfs_quota(cgroup_tg(cgrp));
8142}
8143
8144static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
8145				s64 cfs_quota_us)
8146{
8147	return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
8148}
8149
8150static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
8151{
8152	return tg_get_cfs_period(cgroup_tg(cgrp));
8153}
8154
8155static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8156				u64 cfs_period_us)
8157{
8158	return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
8159}
8160
8161struct cfs_schedulable_data {
8162	struct task_group *tg;
8163	u64 period, quota;
8164};
8165
8166/*
8167 * normalize group quota/period to be quota/max_period
8168 * note: units are usecs
8169 */
8170static u64 normalize_cfs_quota(struct task_group *tg,
8171			       struct cfs_schedulable_data *d)
8172{
8173	u64 quota, period;
8174
8175	if (tg == d->tg) {
8176		period = d->period;
8177		quota = d->quota;
8178	} else {
8179		period = tg_get_cfs_period(tg);
8180		quota = tg_get_cfs_quota(tg);
8181	}
8182
8183	/* note: these should typically be equivalent */
8184	if (quota == RUNTIME_INF || quota == -1)
8185		return RUNTIME_INF;
8186
8187	return to_ratio(period, quota);
8188}
8189
8190static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8191{
8192	struct cfs_schedulable_data *d = data;
8193	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8194	s64 quota = 0, parent_quota = -1;
8195
8196	if (!tg->parent) {
8197		quota = RUNTIME_INF;
8198	} else {
8199		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8200
8201		quota = normalize_cfs_quota(tg, d);
8202		parent_quota = parent_b->hierarchal_quota;
8203
8204		/*
8205		 * ensure max(child_quota) <= parent_quota, inherit when no
8206		 * limit is set
8207		 */
8208		if (quota == RUNTIME_INF)
8209			quota = parent_quota;
8210		else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8211			return -EINVAL;
8212	}
8213	cfs_b->hierarchal_quota = quota;
8214
8215	return 0;
8216}
8217
8218static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8219{
8220	int ret;
8221	struct cfs_schedulable_data data = {
8222		.tg = tg,
8223		.period = period,
8224		.quota = quota,
8225	};
8226
8227	if (quota != RUNTIME_INF) {
8228		do_div(data.period, NSEC_PER_USEC);
8229		do_div(data.quota, NSEC_PER_USEC);
8230	}
8231
8232	rcu_read_lock();
8233	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8234	rcu_read_unlock();
8235
8236	return ret;
8237}
8238
8239static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
8240		struct cgroup_map_cb *cb)
8241{
8242	struct task_group *tg = cgroup_tg(cgrp);
8243	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8244
8245	cb->fill(cb, "nr_periods", cfs_b->nr_periods);
8246	cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
8247	cb->fill(cb, "throttled_time", cfs_b->throttled_time);
8248
8249	return 0;
8250}
8251#endif /* CONFIG_CFS_BANDWIDTH */
8252#endif /* CONFIG_FAIR_GROUP_SCHED */
8253
8254#ifdef CONFIG_RT_GROUP_SCHED
8255static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8256				s64 val)
8257{
8258	return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8259}
8260
8261static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8262{
8263	return sched_group_rt_runtime(cgroup_tg(cgrp));
8264}
8265
8266static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8267		u64 rt_period_us)
8268{
8269	return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8270}
8271
8272static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8273{
8274	return sched_group_rt_period(cgroup_tg(cgrp));
8275}
8276#endif /* CONFIG_RT_GROUP_SCHED */
8277
8278static struct cftype cpu_files[] = {
8279#ifdef CONFIG_FAIR_GROUP_SCHED
8280	{
8281		.name = "shares",
8282		.read_u64 = cpu_shares_read_u64,
8283		.write_u64 = cpu_shares_write_u64,
8284	},
8285#endif
8286#ifdef CONFIG_CFS_BANDWIDTH
8287	{
8288		.name = "cfs_quota_us",
8289		.read_s64 = cpu_cfs_quota_read_s64,
8290		.write_s64 = cpu_cfs_quota_write_s64,
8291	},
8292	{
8293		.name = "cfs_period_us",
8294		.read_u64 = cpu_cfs_period_read_u64,
8295		.write_u64 = cpu_cfs_period_write_u64,
8296	},
8297	{
8298		.name = "stat",
8299		.read_map = cpu_stats_show,
8300	},
8301#endif
8302#ifdef CONFIG_RT_GROUP_SCHED
8303	{
8304		.name = "rt_runtime_us",
8305		.read_s64 = cpu_rt_runtime_read,
8306		.write_s64 = cpu_rt_runtime_write,
8307	},
8308	{
8309		.name = "rt_period_us",
8310		.read_u64 = cpu_rt_period_read_uint,
8311		.write_u64 = cpu_rt_period_write_uint,
8312	},
8313#endif
8314	{ }	/* terminate */
8315};
8316
8317struct cgroup_subsys cpu_cgroup_subsys = {
8318	.name		= "cpu",
8319	.create		= cpu_cgroup_create,
8320	.destroy	= cpu_cgroup_destroy,
8321	.can_attach	= cpu_cgroup_can_attach,
8322	.attach		= cpu_cgroup_attach,
8323	.exit		= cpu_cgroup_exit,
8324	.subsys_id	= cpu_cgroup_subsys_id,
8325	.base_cftypes	= cpu_files,
8326	.early_init	= 1,
8327};
8328
8329#endif	/* CONFIG_CGROUP_SCHED */
8330
8331#ifdef CONFIG_CGROUP_CPUACCT
8332
8333/*
8334 * CPU accounting code for task groups.
8335 *
8336 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8337 * (balbir@in.ibm.com).
8338 */
8339
8340/* create a new cpu accounting group */
8341static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
8342{
8343	struct cpuacct *ca;
8344
8345	if (!cgrp->parent)
8346		return &root_cpuacct.css;
8347
8348	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8349	if (!ca)
8350		goto out;
8351
8352	ca->cpuusage = alloc_percpu(u64);
8353	if (!ca->cpuusage)
8354		goto out_free_ca;
8355
8356	ca->cpustat = alloc_percpu(struct kernel_cpustat);
8357	if (!ca->cpustat)
8358		goto out_free_cpuusage;
8359
8360	return &ca->css;
8361
8362out_free_cpuusage:
8363	free_percpu(ca->cpuusage);
8364out_free_ca:
8365	kfree(ca);
8366out:
8367	return ERR_PTR(-ENOMEM);
8368}
8369
8370/* destroy an existing cpu accounting group */
8371static void cpuacct_destroy(struct cgroup *cgrp)
8372{
8373	struct cpuacct *ca = cgroup_ca(cgrp);
8374
8375	free_percpu(ca->cpustat);
8376	free_percpu(ca->cpuusage);
8377	kfree(ca);
8378}
8379
8380static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8381{
8382	u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8383	u64 data;
8384
8385#ifndef CONFIG_64BIT
8386	/*
8387	 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8388	 */
8389	raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8390	data = *cpuusage;
8391	raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8392#else
8393	data = *cpuusage;
8394#endif
8395
8396	return data;
8397}
8398
8399static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8400{
8401	u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8402
8403#ifndef CONFIG_64BIT
8404	/*
8405	 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8406	 */
8407	raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8408	*cpuusage = val;
8409	raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8410#else
8411	*cpuusage = val;
8412#endif
8413}
8414
8415/* return total cpu usage (in nanoseconds) of a group */
8416static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8417{
8418	struct cpuacct *ca = cgroup_ca(cgrp);
8419	u64 totalcpuusage = 0;
8420	int i;
8421
8422	for_each_present_cpu(i)
8423		totalcpuusage += cpuacct_cpuusage_read(ca, i);
8424
8425	return totalcpuusage;
8426}
8427
8428static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8429								u64 reset)
8430{
8431	struct cpuacct *ca = cgroup_ca(cgrp);
8432	int err = 0;
8433	int i;
8434
8435	if (reset) {
8436		err = -EINVAL;
8437		goto out;
8438	}
8439
8440	for_each_present_cpu(i)
8441		cpuacct_cpuusage_write(ca, i, 0);
8442
8443out:
8444	return err;
8445}
8446
8447static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8448				   struct seq_file *m)
8449{
8450	struct cpuacct *ca = cgroup_ca(cgroup);
8451	u64 percpu;
8452	int i;
8453
8454	for_each_present_cpu(i) {
8455		percpu = cpuacct_cpuusage_read(ca, i);
8456		seq_printf(m, "%llu ", (unsigned long long) percpu);
8457	}
8458	seq_printf(m, "\n");
8459	return 0;
8460}
8461
8462static const char *cpuacct_stat_desc[] = {
8463	[CPUACCT_STAT_USER] = "user",
8464	[CPUACCT_STAT_SYSTEM] = "system",
8465};
8466
8467static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8468			      struct cgroup_map_cb *cb)
8469{
8470	struct cpuacct *ca = cgroup_ca(cgrp);
8471	int cpu;
8472	s64 val = 0;
8473
8474	for_each_online_cpu(cpu) {
8475		struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8476		val += kcpustat->cpustat[CPUTIME_USER];
8477		val += kcpustat->cpustat[CPUTIME_NICE];
8478	}
8479	val = cputime64_to_clock_t(val);
8480	cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8481
8482	val = 0;
8483	for_each_online_cpu(cpu) {
8484		struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8485		val += kcpustat->cpustat[CPUTIME_SYSTEM];
8486		val += kcpustat->cpustat[CPUTIME_IRQ];
8487		val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8488	}
8489
8490	val = cputime64_to_clock_t(val);
8491	cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8492
8493	return 0;
8494}
8495
8496static struct cftype files[] = {
8497	{
8498		.name = "usage",
8499		.read_u64 = cpuusage_read,
8500		.write_u64 = cpuusage_write,
8501	},
8502	{
8503		.name = "usage_percpu",
8504		.read_seq_string = cpuacct_percpu_seq_read,
8505	},
8506	{
8507		.name = "stat",
8508		.read_map = cpuacct_stats_show,
8509	},
8510	{ }	/* terminate */
8511};
8512
8513/*
8514 * charge this task's execution time to its accounting group.
8515 *
8516 * called with rq->lock held.
8517 */
8518void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8519{
8520	struct cpuacct *ca;
8521	int cpu;
8522
8523	if (unlikely(!cpuacct_subsys.active))
8524		return;
8525
8526	cpu = task_cpu(tsk);
8527
8528	rcu_read_lock();
8529
8530	ca = task_ca(tsk);
8531
8532	for (; ca; ca = parent_ca(ca)) {
8533		u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8534		*cpuusage += cputime;
8535	}
8536
8537	rcu_read_unlock();
8538}
8539
8540struct cgroup_subsys cpuacct_subsys = {
8541	.name = "cpuacct",
8542	.create = cpuacct_create,
8543	.destroy = cpuacct_destroy,
8544	.subsys_id = cpuacct_subsys_id,
8545	.base_cftypes = files,
8546};
8547#endif	/* CONFIG_CGROUP_CPUACCT */
8548