core.c revision b8593bfda1652755136333cdd362de125b283a9c
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) 0
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
743static void update_rq_clock_task(struct rq *rq, s64 delta)
744{
745/*
746 * In theory, the compile should just see 0 here, and optimize out the call
747 * to sched_rt_avg_update. But I don't trust it...
748 */
749#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
750	s64 steal = 0, irq_delta = 0;
751#endif
752#ifdef CONFIG_IRQ_TIME_ACCOUNTING
753	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
754
755	/*
756	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
757	 * this case when a previous update_rq_clock() happened inside a
758	 * {soft,}irq region.
759	 *
760	 * When this happens, we stop ->clock_task and only update the
761	 * prev_irq_time stamp to account for the part that fit, so that a next
762	 * update will consume the rest. This ensures ->clock_task is
763	 * monotonic.
764	 *
765	 * It does however cause some slight miss-attribution of {soft,}irq
766	 * time, a more accurate solution would be to update the irq_time using
767	 * the current rq->clock timestamp, except that would require using
768	 * atomic ops.
769	 */
770	if (irq_delta > delta)
771		irq_delta = delta;
772
773	rq->prev_irq_time += irq_delta;
774	delta -= irq_delta;
775#endif
776#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
777	if (static_key_false((&paravirt_steal_rq_enabled))) {
778		u64 st;
779
780		steal = paravirt_steal_clock(cpu_of(rq));
781		steal -= rq->prev_steal_time_rq;
782
783		if (unlikely(steal > delta))
784			steal = delta;
785
786		st = steal_ticks(steal);
787		steal = st * TICK_NSEC;
788
789		rq->prev_steal_time_rq += steal;
790
791		delta -= steal;
792	}
793#endif
794
795	rq->clock_task += delta;
796
797#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
798	if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
799		sched_rt_avg_update(rq, irq_delta + steal);
800#endif
801}
802
803void sched_set_stop_task(int cpu, struct task_struct *stop)
804{
805	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
806	struct task_struct *old_stop = cpu_rq(cpu)->stop;
807
808	if (stop) {
809		/*
810		 * Make it appear like a SCHED_FIFO task, its something
811		 * userspace knows about and won't get confused about.
812		 *
813		 * Also, it will make PI more or less work without too
814		 * much confusion -- but then, stop work should not
815		 * rely on PI working anyway.
816		 */
817		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
818
819		stop->sched_class = &stop_sched_class;
820	}
821
822	cpu_rq(cpu)->stop = stop;
823
824	if (old_stop) {
825		/*
826		 * Reset it back to a normal scheduling class so that
827		 * it can die in pieces.
828		 */
829		old_stop->sched_class = &rt_sched_class;
830	}
831}
832
833/*
834 * __normal_prio - return the priority that is based on the static prio
835 */
836static inline int __normal_prio(struct task_struct *p)
837{
838	return p->static_prio;
839}
840
841/*
842 * Calculate the expected normal priority: i.e. priority
843 * without taking RT-inheritance into account. Might be
844 * boosted by interactivity modifiers. Changes upon fork,
845 * setprio syscalls, and whenever the interactivity
846 * estimator recalculates.
847 */
848static inline int normal_prio(struct task_struct *p)
849{
850	int prio;
851
852	if (task_has_rt_policy(p))
853		prio = MAX_RT_PRIO-1 - p->rt_priority;
854	else
855		prio = __normal_prio(p);
856	return prio;
857}
858
859/*
860 * Calculate the current priority, i.e. the priority
861 * taken into account by the scheduler. This value might
862 * be boosted by RT tasks, or might be boosted by
863 * interactivity modifiers. Will be RT if the task got
864 * RT-boosted. If not then it returns p->normal_prio.
865 */
866static int effective_prio(struct task_struct *p)
867{
868	p->normal_prio = normal_prio(p);
869	/*
870	 * If we are RT tasks or we were boosted to RT priority,
871	 * keep the priority unchanged. Otherwise, update priority
872	 * to the normal priority:
873	 */
874	if (!rt_prio(p->prio))
875		return p->normal_prio;
876	return p->prio;
877}
878
879/**
880 * task_curr - is this task currently executing on a CPU?
881 * @p: the task in question.
882 */
883inline int task_curr(const struct task_struct *p)
884{
885	return cpu_curr(task_cpu(p)) == p;
886}
887
888static inline void check_class_changed(struct rq *rq, struct task_struct *p,
889				       const struct sched_class *prev_class,
890				       int oldprio)
891{
892	if (prev_class != p->sched_class) {
893		if (prev_class->switched_from)
894			prev_class->switched_from(rq, p);
895		p->sched_class->switched_to(rq, p);
896	} else if (oldprio != p->prio)
897		p->sched_class->prio_changed(rq, p, oldprio);
898}
899
900void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
901{
902	const struct sched_class *class;
903
904	if (p->sched_class == rq->curr->sched_class) {
905		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
906	} else {
907		for_each_class(class) {
908			if (class == rq->curr->sched_class)
909				break;
910			if (class == p->sched_class) {
911				resched_task(rq->curr);
912				break;
913			}
914		}
915	}
916
917	/*
918	 * A queue event has occurred, and we're going to schedule.  In
919	 * this case, we can save a useless back to back clock update.
920	 */
921	if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
922		rq->skip_clock_update = 1;
923}
924
925#ifdef CONFIG_SMP
926void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
927{
928#ifdef CONFIG_SCHED_DEBUG
929	/*
930	 * We should never call set_task_cpu() on a blocked task,
931	 * ttwu() will sort out the placement.
932	 */
933	WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
934			!(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
935
936#ifdef CONFIG_LOCKDEP
937	/*
938	 * The caller should hold either p->pi_lock or rq->lock, when changing
939	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
940	 *
941	 * sched_move_task() holds both and thus holding either pins the cgroup,
942	 * see task_group().
943	 *
944	 * Furthermore, all task_rq users should acquire both locks, see
945	 * task_rq_lock().
946	 */
947	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
948				      lockdep_is_held(&task_rq(p)->lock)));
949#endif
950#endif
951
952	trace_sched_migrate_task(p, new_cpu);
953
954	if (task_cpu(p) != new_cpu) {
955		p->se.nr_migrations++;
956		perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
957	}
958
959	__set_task_cpu(p, new_cpu);
960}
961
962struct migration_arg {
963	struct task_struct *task;
964	int dest_cpu;
965};
966
967static int migration_cpu_stop(void *data);
968
969/*
970 * wait_task_inactive - wait for a thread to unschedule.
971 *
972 * If @match_state is nonzero, it's the @p->state value just checked and
973 * not expected to change.  If it changes, i.e. @p might have woken up,
974 * then return zero.  When we succeed in waiting for @p to be off its CPU,
975 * we return a positive number (its total switch count).  If a second call
976 * a short while later returns the same number, the caller can be sure that
977 * @p has remained unscheduled the whole time.
978 *
979 * The caller must ensure that the task *will* unschedule sometime soon,
980 * else this function might spin for a *long* time. This function can't
981 * be called with interrupts off, or it may introduce deadlock with
982 * smp_call_function() if an IPI is sent by the same process we are
983 * waiting to become inactive.
984 */
985unsigned long wait_task_inactive(struct task_struct *p, long match_state)
986{
987	unsigned long flags;
988	int running, on_rq;
989	unsigned long ncsw;
990	struct rq *rq;
991
992	for (;;) {
993		/*
994		 * We do the initial early heuristics without holding
995		 * any task-queue locks at all. We'll only try to get
996		 * the runqueue lock when things look like they will
997		 * work out!
998		 */
999		rq = task_rq(p);
1000
1001		/*
1002		 * If the task is actively running on another CPU
1003		 * still, just relax and busy-wait without holding
1004		 * any locks.
1005		 *
1006		 * NOTE! Since we don't hold any locks, it's not
1007		 * even sure that "rq" stays as the right runqueue!
1008		 * But we don't care, since "task_running()" will
1009		 * return false if the runqueue has changed and p
1010		 * is actually now running somewhere else!
1011		 */
1012		while (task_running(rq, p)) {
1013			if (match_state && unlikely(p->state != match_state))
1014				return 0;
1015			cpu_relax();
1016		}
1017
1018		/*
1019		 * Ok, time to look more closely! We need the rq
1020		 * lock now, to be *sure*. If we're wrong, we'll
1021		 * just go back and repeat.
1022		 */
1023		rq = task_rq_lock(p, &flags);
1024		trace_sched_wait_task(p);
1025		running = task_running(rq, p);
1026		on_rq = p->on_rq;
1027		ncsw = 0;
1028		if (!match_state || p->state == match_state)
1029			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1030		task_rq_unlock(rq, p, &flags);
1031
1032		/*
1033		 * If it changed from the expected state, bail out now.
1034		 */
1035		if (unlikely(!ncsw))
1036			break;
1037
1038		/*
1039		 * Was it really running after all now that we
1040		 * checked with the proper locks actually held?
1041		 *
1042		 * Oops. Go back and try again..
1043		 */
1044		if (unlikely(running)) {
1045			cpu_relax();
1046			continue;
1047		}
1048
1049		/*
1050		 * It's not enough that it's not actively running,
1051		 * it must be off the runqueue _entirely_, and not
1052		 * preempted!
1053		 *
1054		 * So if it was still runnable (but just not actively
1055		 * running right now), it's preempted, and we should
1056		 * yield - it could be a while.
1057		 */
1058		if (unlikely(on_rq)) {
1059			ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1060
1061			set_current_state(TASK_UNINTERRUPTIBLE);
1062			schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1063			continue;
1064		}
1065
1066		/*
1067		 * Ahh, all good. It wasn't running, and it wasn't
1068		 * runnable, which means that it will never become
1069		 * running in the future either. We're all done!
1070		 */
1071		break;
1072	}
1073
1074	return ncsw;
1075}
1076
1077/***
1078 * kick_process - kick a running thread to enter/exit the kernel
1079 * @p: the to-be-kicked thread
1080 *
1081 * Cause a process which is running on another CPU to enter
1082 * kernel-mode, without any delay. (to get signals handled.)
1083 *
1084 * NOTE: this function doesn't have to take the runqueue lock,
1085 * because all it wants to ensure is that the remote task enters
1086 * the kernel. If the IPI races and the task has been migrated
1087 * to another CPU then no harm is done and the purpose has been
1088 * achieved as well.
1089 */
1090void kick_process(struct task_struct *p)
1091{
1092	int cpu;
1093
1094	preempt_disable();
1095	cpu = task_cpu(p);
1096	if ((cpu != smp_processor_id()) && task_curr(p))
1097		smp_send_reschedule(cpu);
1098	preempt_enable();
1099}
1100EXPORT_SYMBOL_GPL(kick_process);
1101#endif /* CONFIG_SMP */
1102
1103#ifdef CONFIG_SMP
1104/*
1105 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1106 */
1107static int select_fallback_rq(int cpu, struct task_struct *p)
1108{
1109	const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1110	enum { cpuset, possible, fail } state = cpuset;
1111	int dest_cpu;
1112
1113	/* Look for allowed, online CPU in same node. */
1114	for_each_cpu(dest_cpu, nodemask) {
1115		if (!cpu_online(dest_cpu))
1116			continue;
1117		if (!cpu_active(dest_cpu))
1118			continue;
1119		if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1120			return dest_cpu;
1121	}
1122
1123	for (;;) {
1124		/* Any allowed, online CPU? */
1125		for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1126			if (!cpu_online(dest_cpu))
1127				continue;
1128			if (!cpu_active(dest_cpu))
1129				continue;
1130			goto out;
1131		}
1132
1133		switch (state) {
1134		case cpuset:
1135			/* No more Mr. Nice Guy. */
1136			cpuset_cpus_allowed_fallback(p);
1137			state = possible;
1138			break;
1139
1140		case possible:
1141			do_set_cpus_allowed(p, cpu_possible_mask);
1142			state = fail;
1143			break;
1144
1145		case fail:
1146			BUG();
1147			break;
1148		}
1149	}
1150
1151out:
1152	if (state != cpuset) {
1153		/*
1154		 * Don't tell them about moving exiting tasks or
1155		 * kernel threads (both mm NULL), since they never
1156		 * leave kernel.
1157		 */
1158		if (p->mm && printk_ratelimit()) {
1159			printk_sched("process %d (%s) no longer affine to cpu%d\n",
1160					task_pid_nr(p), p->comm, cpu);
1161		}
1162	}
1163
1164	return dest_cpu;
1165}
1166
1167/*
1168 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1169 */
1170static inline
1171int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1172{
1173	int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1174
1175	/*
1176	 * In order not to call set_task_cpu() on a blocking task we need
1177	 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1178	 * cpu.
1179	 *
1180	 * Since this is common to all placement strategies, this lives here.
1181	 *
1182	 * [ this allows ->select_task() to simply return task_cpu(p) and
1183	 *   not worry about this generic constraint ]
1184	 */
1185	if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1186		     !cpu_online(cpu)))
1187		cpu = select_fallback_rq(task_cpu(p), p);
1188
1189	return cpu;
1190}
1191
1192static void update_avg(u64 *avg, u64 sample)
1193{
1194	s64 diff = sample - *avg;
1195	*avg += diff >> 3;
1196}
1197#endif
1198
1199static void
1200ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1201{
1202#ifdef CONFIG_SCHEDSTATS
1203	struct rq *rq = this_rq();
1204
1205#ifdef CONFIG_SMP
1206	int this_cpu = smp_processor_id();
1207
1208	if (cpu == this_cpu) {
1209		schedstat_inc(rq, ttwu_local);
1210		schedstat_inc(p, se.statistics.nr_wakeups_local);
1211	} else {
1212		struct sched_domain *sd;
1213
1214		schedstat_inc(p, se.statistics.nr_wakeups_remote);
1215		rcu_read_lock();
1216		for_each_domain(this_cpu, sd) {
1217			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1218				schedstat_inc(sd, ttwu_wake_remote);
1219				break;
1220			}
1221		}
1222		rcu_read_unlock();
1223	}
1224
1225	if (wake_flags & WF_MIGRATED)
1226		schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1227
1228#endif /* CONFIG_SMP */
1229
1230	schedstat_inc(rq, ttwu_count);
1231	schedstat_inc(p, se.statistics.nr_wakeups);
1232
1233	if (wake_flags & WF_SYNC)
1234		schedstat_inc(p, se.statistics.nr_wakeups_sync);
1235
1236#endif /* CONFIG_SCHEDSTATS */
1237}
1238
1239static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1240{
1241	activate_task(rq, p, en_flags);
1242	p->on_rq = 1;
1243
1244	/* if a worker is waking up, notify workqueue */
1245	if (p->flags & PF_WQ_WORKER)
1246		wq_worker_waking_up(p, cpu_of(rq));
1247}
1248
1249/*
1250 * Mark the task runnable and perform wakeup-preemption.
1251 */
1252static void
1253ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1254{
1255	trace_sched_wakeup(p, true);
1256	check_preempt_curr(rq, p, wake_flags);
1257
1258	p->state = TASK_RUNNING;
1259#ifdef CONFIG_SMP
1260	if (p->sched_class->task_woken)
1261		p->sched_class->task_woken(rq, p);
1262
1263	if (rq->idle_stamp) {
1264		u64 delta = rq->clock - rq->idle_stamp;
1265		u64 max = 2*sysctl_sched_migration_cost;
1266
1267		if (delta > max)
1268			rq->avg_idle = max;
1269		else
1270			update_avg(&rq->avg_idle, delta);
1271		rq->idle_stamp = 0;
1272	}
1273#endif
1274}
1275
1276static void
1277ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1278{
1279#ifdef CONFIG_SMP
1280	if (p->sched_contributes_to_load)
1281		rq->nr_uninterruptible--;
1282#endif
1283
1284	ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1285	ttwu_do_wakeup(rq, p, wake_flags);
1286}
1287
1288/*
1289 * Called in case the task @p isn't fully descheduled from its runqueue,
1290 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1291 * since all we need to do is flip p->state to TASK_RUNNING, since
1292 * the task is still ->on_rq.
1293 */
1294static int ttwu_remote(struct task_struct *p, int wake_flags)
1295{
1296	struct rq *rq;
1297	int ret = 0;
1298
1299	rq = __task_rq_lock(p);
1300	if (p->on_rq) {
1301		ttwu_do_wakeup(rq, p, wake_flags);
1302		ret = 1;
1303	}
1304	__task_rq_unlock(rq);
1305
1306	return ret;
1307}
1308
1309#ifdef CONFIG_SMP
1310static void sched_ttwu_pending(void)
1311{
1312	struct rq *rq = this_rq();
1313	struct llist_node *llist = llist_del_all(&rq->wake_list);
1314	struct task_struct *p;
1315
1316	raw_spin_lock(&rq->lock);
1317
1318	while (llist) {
1319		p = llist_entry(llist, struct task_struct, wake_entry);
1320		llist = llist_next(llist);
1321		ttwu_do_activate(rq, p, 0);
1322	}
1323
1324	raw_spin_unlock(&rq->lock);
1325}
1326
1327void scheduler_ipi(void)
1328{
1329	if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1330		return;
1331
1332	/*
1333	 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1334	 * traditionally all their work was done from the interrupt return
1335	 * path. Now that we actually do some work, we need to make sure
1336	 * we do call them.
1337	 *
1338	 * Some archs already do call them, luckily irq_enter/exit nest
1339	 * properly.
1340	 *
1341	 * Arguably we should visit all archs and update all handlers,
1342	 * however a fair share of IPIs are still resched only so this would
1343	 * somewhat pessimize the simple resched case.
1344	 */
1345	irq_enter();
1346	sched_ttwu_pending();
1347
1348	/*
1349	 * Check if someone kicked us for doing the nohz idle load balance.
1350	 */
1351	if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1352		this_rq()->idle_balance = 1;
1353		raise_softirq_irqoff(SCHED_SOFTIRQ);
1354	}
1355	irq_exit();
1356}
1357
1358static void ttwu_queue_remote(struct task_struct *p, int cpu)
1359{
1360	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1361		smp_send_reschedule(cpu);
1362}
1363
1364bool cpus_share_cache(int this_cpu, int that_cpu)
1365{
1366	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1367}
1368#endif /* CONFIG_SMP */
1369
1370static void ttwu_queue(struct task_struct *p, int cpu)
1371{
1372	struct rq *rq = cpu_rq(cpu);
1373
1374#if defined(CONFIG_SMP)
1375	if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1376		sched_clock_cpu(cpu); /* sync clocks x-cpu */
1377		ttwu_queue_remote(p, cpu);
1378		return;
1379	}
1380#endif
1381
1382	raw_spin_lock(&rq->lock);
1383	ttwu_do_activate(rq, p, 0);
1384	raw_spin_unlock(&rq->lock);
1385}
1386
1387/**
1388 * try_to_wake_up - wake up a thread
1389 * @p: the thread to be awakened
1390 * @state: the mask of task states that can be woken
1391 * @wake_flags: wake modifier flags (WF_*)
1392 *
1393 * Put it on the run-queue if it's not already there. The "current"
1394 * thread is always on the run-queue (except when the actual
1395 * re-schedule is in progress), and as such you're allowed to do
1396 * the simpler "current->state = TASK_RUNNING" to mark yourself
1397 * runnable without the overhead of this.
1398 *
1399 * Returns %true if @p was woken up, %false if it was already running
1400 * or @state didn't match @p's state.
1401 */
1402static int
1403try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1404{
1405	unsigned long flags;
1406	int cpu, success = 0;
1407
1408	smp_wmb();
1409	raw_spin_lock_irqsave(&p->pi_lock, flags);
1410	if (!(p->state & state))
1411		goto out;
1412
1413	success = 1; /* we're going to change ->state */
1414	cpu = task_cpu(p);
1415
1416	if (p->on_rq && ttwu_remote(p, wake_flags))
1417		goto stat;
1418
1419#ifdef CONFIG_SMP
1420	/*
1421	 * If the owning (remote) cpu is still in the middle of schedule() with
1422	 * this task as prev, wait until its done referencing the task.
1423	 */
1424	while (p->on_cpu)
1425		cpu_relax();
1426	/*
1427	 * Pairs with the smp_wmb() in finish_lock_switch().
1428	 */
1429	smp_rmb();
1430
1431	p->sched_contributes_to_load = !!task_contributes_to_load(p);
1432	p->state = TASK_WAKING;
1433
1434	if (p->sched_class->task_waking)
1435		p->sched_class->task_waking(p);
1436
1437	cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1438	if (task_cpu(p) != cpu) {
1439		wake_flags |= WF_MIGRATED;
1440		set_task_cpu(p, cpu);
1441	}
1442#endif /* CONFIG_SMP */
1443
1444	ttwu_queue(p, cpu);
1445stat:
1446	ttwu_stat(p, cpu, wake_flags);
1447out:
1448	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1449
1450	return success;
1451}
1452
1453/**
1454 * try_to_wake_up_local - try to wake up a local task with rq lock held
1455 * @p: the thread to be awakened
1456 *
1457 * Put @p on the run-queue if it's not already there. The caller must
1458 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1459 * the current task.
1460 */
1461static void try_to_wake_up_local(struct task_struct *p)
1462{
1463	struct rq *rq = task_rq(p);
1464
1465	BUG_ON(rq != this_rq());
1466	BUG_ON(p == current);
1467	lockdep_assert_held(&rq->lock);
1468
1469	if (!raw_spin_trylock(&p->pi_lock)) {
1470		raw_spin_unlock(&rq->lock);
1471		raw_spin_lock(&p->pi_lock);
1472		raw_spin_lock(&rq->lock);
1473	}
1474
1475	if (!(p->state & TASK_NORMAL))
1476		goto out;
1477
1478	if (!p->on_rq)
1479		ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1480
1481	ttwu_do_wakeup(rq, p, 0);
1482	ttwu_stat(p, smp_processor_id(), 0);
1483out:
1484	raw_spin_unlock(&p->pi_lock);
1485}
1486
1487/**
1488 * wake_up_process - Wake up a specific process
1489 * @p: The process to be woken up.
1490 *
1491 * Attempt to wake up the nominated process and move it to the set of runnable
1492 * processes.  Returns 1 if the process was woken up, 0 if it was already
1493 * running.
1494 *
1495 * It may be assumed that this function implies a write memory barrier before
1496 * changing the task state if and only if any tasks are woken up.
1497 */
1498int wake_up_process(struct task_struct *p)
1499{
1500	return try_to_wake_up(p, TASK_ALL, 0);
1501}
1502EXPORT_SYMBOL(wake_up_process);
1503
1504int wake_up_state(struct task_struct *p, unsigned int state)
1505{
1506	return try_to_wake_up(p, state, 0);
1507}
1508
1509/*
1510 * Perform scheduler related setup for a newly forked process p.
1511 * p is forked by current.
1512 *
1513 * __sched_fork() is basic setup used by init_idle() too:
1514 */
1515static void __sched_fork(struct task_struct *p)
1516{
1517	p->on_rq			= 0;
1518
1519	p->se.on_rq			= 0;
1520	p->se.exec_start		= 0;
1521	p->se.sum_exec_runtime		= 0;
1522	p->se.prev_sum_exec_runtime	= 0;
1523	p->se.nr_migrations		= 0;
1524	p->se.vruntime			= 0;
1525	INIT_LIST_HEAD(&p->se.group_node);
1526
1527#ifdef CONFIG_SCHEDSTATS
1528	memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1529#endif
1530
1531	INIT_LIST_HEAD(&p->rt.run_list);
1532
1533#ifdef CONFIG_PREEMPT_NOTIFIERS
1534	INIT_HLIST_HEAD(&p->preempt_notifiers);
1535#endif
1536
1537#ifdef CONFIG_NUMA_BALANCING
1538	if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1539		p->mm->numa_next_scan = jiffies;
1540		p->mm->numa_next_reset = jiffies;
1541		p->mm->numa_scan_seq = 0;
1542	}
1543
1544	p->node_stamp = 0ULL;
1545	p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1546	p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1547	p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1548	p->numa_work.next = &p->numa_work;
1549#endif /* CONFIG_NUMA_BALANCING */
1550}
1551
1552/*
1553 * fork()/clone()-time setup:
1554 */
1555void sched_fork(struct task_struct *p)
1556{
1557	unsigned long flags;
1558	int cpu = get_cpu();
1559
1560	__sched_fork(p);
1561	/*
1562	 * We mark the process as running here. This guarantees that
1563	 * nobody will actually run it, and a signal or other external
1564	 * event cannot wake it up and insert it on the runqueue either.
1565	 */
1566	p->state = TASK_RUNNING;
1567
1568	/*
1569	 * Make sure we do not leak PI boosting priority to the child.
1570	 */
1571	p->prio = current->normal_prio;
1572
1573	/*
1574	 * Revert to default priority/policy on fork if requested.
1575	 */
1576	if (unlikely(p->sched_reset_on_fork)) {
1577		if (task_has_rt_policy(p)) {
1578			p->policy = SCHED_NORMAL;
1579			p->static_prio = NICE_TO_PRIO(0);
1580			p->rt_priority = 0;
1581		} else if (PRIO_TO_NICE(p->static_prio) < 0)
1582			p->static_prio = NICE_TO_PRIO(0);
1583
1584		p->prio = p->normal_prio = __normal_prio(p);
1585		set_load_weight(p);
1586
1587		/*
1588		 * We don't need the reset flag anymore after the fork. It has
1589		 * fulfilled its duty:
1590		 */
1591		p->sched_reset_on_fork = 0;
1592	}
1593
1594	if (!rt_prio(p->prio))
1595		p->sched_class = &fair_sched_class;
1596
1597	if (p->sched_class->task_fork)
1598		p->sched_class->task_fork(p);
1599
1600	/*
1601	 * The child is not yet in the pid-hash so no cgroup attach races,
1602	 * and the cgroup is pinned to this child due to cgroup_fork()
1603	 * is ran before sched_fork().
1604	 *
1605	 * Silence PROVE_RCU.
1606	 */
1607	raw_spin_lock_irqsave(&p->pi_lock, flags);
1608	set_task_cpu(p, cpu);
1609	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1610
1611#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1612	if (likely(sched_info_on()))
1613		memset(&p->sched_info, 0, sizeof(p->sched_info));
1614#endif
1615#if defined(CONFIG_SMP)
1616	p->on_cpu = 0;
1617#endif
1618#ifdef CONFIG_PREEMPT_COUNT
1619	/* Want to start with kernel preemption disabled. */
1620	task_thread_info(p)->preempt_count = 1;
1621#endif
1622#ifdef CONFIG_SMP
1623	plist_node_init(&p->pushable_tasks, MAX_PRIO);
1624#endif
1625
1626	put_cpu();
1627}
1628
1629/*
1630 * wake_up_new_task - wake up a newly created task for the first time.
1631 *
1632 * This function will do some initial scheduler statistics housekeeping
1633 * that must be done for every newly created context, then puts the task
1634 * on the runqueue and wakes it.
1635 */
1636void wake_up_new_task(struct task_struct *p)
1637{
1638	unsigned long flags;
1639	struct rq *rq;
1640
1641	raw_spin_lock_irqsave(&p->pi_lock, flags);
1642#ifdef CONFIG_SMP
1643	/*
1644	 * Fork balancing, do it here and not earlier because:
1645	 *  - cpus_allowed can change in the fork path
1646	 *  - any previously selected cpu might disappear through hotplug
1647	 */
1648	set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1649#endif
1650
1651	rq = __task_rq_lock(p);
1652	activate_task(rq, p, 0);
1653	p->on_rq = 1;
1654	trace_sched_wakeup_new(p, true);
1655	check_preempt_curr(rq, p, WF_FORK);
1656#ifdef CONFIG_SMP
1657	if (p->sched_class->task_woken)
1658		p->sched_class->task_woken(rq, p);
1659#endif
1660	task_rq_unlock(rq, p, &flags);
1661}
1662
1663#ifdef CONFIG_PREEMPT_NOTIFIERS
1664
1665/**
1666 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1667 * @notifier: notifier struct to register
1668 */
1669void preempt_notifier_register(struct preempt_notifier *notifier)
1670{
1671	hlist_add_head(&notifier->link, &current->preempt_notifiers);
1672}
1673EXPORT_SYMBOL_GPL(preempt_notifier_register);
1674
1675/**
1676 * preempt_notifier_unregister - no longer interested in preemption notifications
1677 * @notifier: notifier struct to unregister
1678 *
1679 * This is safe to call from within a preemption notifier.
1680 */
1681void preempt_notifier_unregister(struct preempt_notifier *notifier)
1682{
1683	hlist_del(&notifier->link);
1684}
1685EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1686
1687static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1688{
1689	struct preempt_notifier *notifier;
1690	struct hlist_node *node;
1691
1692	hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1693		notifier->ops->sched_in(notifier, raw_smp_processor_id());
1694}
1695
1696static void
1697fire_sched_out_preempt_notifiers(struct task_struct *curr,
1698				 struct task_struct *next)
1699{
1700	struct preempt_notifier *notifier;
1701	struct hlist_node *node;
1702
1703	hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1704		notifier->ops->sched_out(notifier, next);
1705}
1706
1707#else /* !CONFIG_PREEMPT_NOTIFIERS */
1708
1709static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1710{
1711}
1712
1713static void
1714fire_sched_out_preempt_notifiers(struct task_struct *curr,
1715				 struct task_struct *next)
1716{
1717}
1718
1719#endif /* CONFIG_PREEMPT_NOTIFIERS */
1720
1721/**
1722 * prepare_task_switch - prepare to switch tasks
1723 * @rq: the runqueue preparing to switch
1724 * @prev: the current task that is being switched out
1725 * @next: the task we are going to switch to.
1726 *
1727 * This is called with the rq lock held and interrupts off. It must
1728 * be paired with a subsequent finish_task_switch after the context
1729 * switch.
1730 *
1731 * prepare_task_switch sets up locking and calls architecture specific
1732 * hooks.
1733 */
1734static inline void
1735prepare_task_switch(struct rq *rq, struct task_struct *prev,
1736		    struct task_struct *next)
1737{
1738	trace_sched_switch(prev, next);
1739	sched_info_switch(prev, next);
1740	perf_event_task_sched_out(prev, next);
1741	fire_sched_out_preempt_notifiers(prev, next);
1742	prepare_lock_switch(rq, next);
1743	prepare_arch_switch(next);
1744}
1745
1746/**
1747 * finish_task_switch - clean up after a task-switch
1748 * @rq: runqueue associated with task-switch
1749 * @prev: the thread we just switched away from.
1750 *
1751 * finish_task_switch must be called after the context switch, paired
1752 * with a prepare_task_switch call before the context switch.
1753 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1754 * and do any other architecture-specific cleanup actions.
1755 *
1756 * Note that we may have delayed dropping an mm in context_switch(). If
1757 * so, we finish that here outside of the runqueue lock. (Doing it
1758 * with the lock held can cause deadlocks; see schedule() for
1759 * details.)
1760 */
1761static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1762	__releases(rq->lock)
1763{
1764	struct mm_struct *mm = rq->prev_mm;
1765	long prev_state;
1766
1767	rq->prev_mm = NULL;
1768
1769	/*
1770	 * A task struct has one reference for the use as "current".
1771	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1772	 * schedule one last time. The schedule call will never return, and
1773	 * the scheduled task must drop that reference.
1774	 * The test for TASK_DEAD must occur while the runqueue locks are
1775	 * still held, otherwise prev could be scheduled on another cpu, die
1776	 * there before we look at prev->state, and then the reference would
1777	 * be dropped twice.
1778	 *		Manfred Spraul <manfred@colorfullife.com>
1779	 */
1780	prev_state = prev->state;
1781	vtime_task_switch(prev);
1782	finish_arch_switch(prev);
1783	perf_event_task_sched_in(prev, current);
1784	finish_lock_switch(rq, prev);
1785	finish_arch_post_lock_switch();
1786
1787	fire_sched_in_preempt_notifiers(current);
1788	if (mm)
1789		mmdrop(mm);
1790	if (unlikely(prev_state == TASK_DEAD)) {
1791		/*
1792		 * Remove function-return probe instances associated with this
1793		 * task and put them back on the free list.
1794		 */
1795		kprobe_flush_task(prev);
1796		put_task_struct(prev);
1797	}
1798}
1799
1800#ifdef CONFIG_SMP
1801
1802/* assumes rq->lock is held */
1803static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1804{
1805	if (prev->sched_class->pre_schedule)
1806		prev->sched_class->pre_schedule(rq, prev);
1807}
1808
1809/* rq->lock is NOT held, but preemption is disabled */
1810static inline void post_schedule(struct rq *rq)
1811{
1812	if (rq->post_schedule) {
1813		unsigned long flags;
1814
1815		raw_spin_lock_irqsave(&rq->lock, flags);
1816		if (rq->curr->sched_class->post_schedule)
1817			rq->curr->sched_class->post_schedule(rq);
1818		raw_spin_unlock_irqrestore(&rq->lock, flags);
1819
1820		rq->post_schedule = 0;
1821	}
1822}
1823
1824#else
1825
1826static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1827{
1828}
1829
1830static inline void post_schedule(struct rq *rq)
1831{
1832}
1833
1834#endif
1835
1836/**
1837 * schedule_tail - first thing a freshly forked thread must call.
1838 * @prev: the thread we just switched away from.
1839 */
1840asmlinkage void schedule_tail(struct task_struct *prev)
1841	__releases(rq->lock)
1842{
1843	struct rq *rq = this_rq();
1844
1845	finish_task_switch(rq, prev);
1846
1847	/*
1848	 * FIXME: do we need to worry about rq being invalidated by the
1849	 * task_switch?
1850	 */
1851	post_schedule(rq);
1852
1853#ifdef __ARCH_WANT_UNLOCKED_CTXSW
1854	/* In this case, finish_task_switch does not reenable preemption */
1855	preempt_enable();
1856#endif
1857	if (current->set_child_tid)
1858		put_user(task_pid_vnr(current), current->set_child_tid);
1859}
1860
1861/*
1862 * context_switch - switch to the new MM and the new
1863 * thread's register state.
1864 */
1865static inline void
1866context_switch(struct rq *rq, struct task_struct *prev,
1867	       struct task_struct *next)
1868{
1869	struct mm_struct *mm, *oldmm;
1870
1871	prepare_task_switch(rq, prev, next);
1872
1873	mm = next->mm;
1874	oldmm = prev->active_mm;
1875	/*
1876	 * For paravirt, this is coupled with an exit in switch_to to
1877	 * combine the page table reload and the switch backend into
1878	 * one hypercall.
1879	 */
1880	arch_start_context_switch(prev);
1881
1882	if (!mm) {
1883		next->active_mm = oldmm;
1884		atomic_inc(&oldmm->mm_count);
1885		enter_lazy_tlb(oldmm, next);
1886	} else
1887		switch_mm(oldmm, mm, next);
1888
1889	if (!prev->mm) {
1890		prev->active_mm = NULL;
1891		rq->prev_mm = oldmm;
1892	}
1893	/*
1894	 * Since the runqueue lock will be released by the next
1895	 * task (which is an invalid locking op but in the case
1896	 * of the scheduler it's an obvious special-case), so we
1897	 * do an early lockdep release here:
1898	 */
1899#ifndef __ARCH_WANT_UNLOCKED_CTXSW
1900	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1901#endif
1902
1903	/* Here we just switch the register state and the stack. */
1904	rcu_switch(prev, next);
1905	switch_to(prev, next, prev);
1906
1907	barrier();
1908	/*
1909	 * this_rq must be evaluated again because prev may have moved
1910	 * CPUs since it called schedule(), thus the 'rq' on its stack
1911	 * frame will be invalid.
1912	 */
1913	finish_task_switch(this_rq(), prev);
1914}
1915
1916/*
1917 * nr_running, nr_uninterruptible and nr_context_switches:
1918 *
1919 * externally visible scheduler statistics: current number of runnable
1920 * threads, current number of uninterruptible-sleeping threads, total
1921 * number of context switches performed since bootup.
1922 */
1923unsigned long nr_running(void)
1924{
1925	unsigned long i, sum = 0;
1926
1927	for_each_online_cpu(i)
1928		sum += cpu_rq(i)->nr_running;
1929
1930	return sum;
1931}
1932
1933unsigned long nr_uninterruptible(void)
1934{
1935	unsigned long i, sum = 0;
1936
1937	for_each_possible_cpu(i)
1938		sum += cpu_rq(i)->nr_uninterruptible;
1939
1940	/*
1941	 * Since we read the counters lockless, it might be slightly
1942	 * inaccurate. Do not allow it to go below zero though:
1943	 */
1944	if (unlikely((long)sum < 0))
1945		sum = 0;
1946
1947	return sum;
1948}
1949
1950unsigned long long nr_context_switches(void)
1951{
1952	int i;
1953	unsigned long long sum = 0;
1954
1955	for_each_possible_cpu(i)
1956		sum += cpu_rq(i)->nr_switches;
1957
1958	return sum;
1959}
1960
1961unsigned long nr_iowait(void)
1962{
1963	unsigned long i, sum = 0;
1964
1965	for_each_possible_cpu(i)
1966		sum += atomic_read(&cpu_rq(i)->nr_iowait);
1967
1968	return sum;
1969}
1970
1971unsigned long nr_iowait_cpu(int cpu)
1972{
1973	struct rq *this = cpu_rq(cpu);
1974	return atomic_read(&this->nr_iowait);
1975}
1976
1977unsigned long this_cpu_load(void)
1978{
1979	struct rq *this = this_rq();
1980	return this->cpu_load[0];
1981}
1982
1983
1984/*
1985 * Global load-average calculations
1986 *
1987 * We take a distributed and async approach to calculating the global load-avg
1988 * in order to minimize overhead.
1989 *
1990 * The global load average is an exponentially decaying average of nr_running +
1991 * nr_uninterruptible.
1992 *
1993 * Once every LOAD_FREQ:
1994 *
1995 *   nr_active = 0;
1996 *   for_each_possible_cpu(cpu)
1997 *   	nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
1998 *
1999 *   avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2000 *
2001 * Due to a number of reasons the above turns in the mess below:
2002 *
2003 *  - for_each_possible_cpu() is prohibitively expensive on machines with
2004 *    serious number of cpus, therefore we need to take a distributed approach
2005 *    to calculating nr_active.
2006 *
2007 *        \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2008 *                      = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2009 *
2010 *    So assuming nr_active := 0 when we start out -- true per definition, we
2011 *    can simply take per-cpu deltas and fold those into a global accumulate
2012 *    to obtain the same result. See calc_load_fold_active().
2013 *
2014 *    Furthermore, in order to avoid synchronizing all per-cpu delta folding
2015 *    across the machine, we assume 10 ticks is sufficient time for every
2016 *    cpu to have completed this task.
2017 *
2018 *    This places an upper-bound on the IRQ-off latency of the machine. Then
2019 *    again, being late doesn't loose the delta, just wrecks the sample.
2020 *
2021 *  - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2022 *    this would add another cross-cpu cacheline miss and atomic operation
2023 *    to the wakeup path. Instead we increment on whatever cpu the task ran
2024 *    when it went into uninterruptible state and decrement on whatever cpu
2025 *    did the wakeup. This means that only the sum of nr_uninterruptible over
2026 *    all cpus yields the correct result.
2027 *
2028 *  This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2029 */
2030
2031/* Variables and functions for calc_load */
2032static atomic_long_t calc_load_tasks;
2033static unsigned long calc_load_update;
2034unsigned long avenrun[3];
2035EXPORT_SYMBOL(avenrun); /* should be removed */
2036
2037/**
2038 * get_avenrun - get the load average array
2039 * @loads:	pointer to dest load array
2040 * @offset:	offset to add
2041 * @shift:	shift count to shift the result left
2042 *
2043 * These values are estimates at best, so no need for locking.
2044 */
2045void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2046{
2047	loads[0] = (avenrun[0] + offset) << shift;
2048	loads[1] = (avenrun[1] + offset) << shift;
2049	loads[2] = (avenrun[2] + offset) << shift;
2050}
2051
2052static long calc_load_fold_active(struct rq *this_rq)
2053{
2054	long nr_active, delta = 0;
2055
2056	nr_active = this_rq->nr_running;
2057	nr_active += (long) this_rq->nr_uninterruptible;
2058
2059	if (nr_active != this_rq->calc_load_active) {
2060		delta = nr_active - this_rq->calc_load_active;
2061		this_rq->calc_load_active = nr_active;
2062	}
2063
2064	return delta;
2065}
2066
2067/*
2068 * a1 = a0 * e + a * (1 - e)
2069 */
2070static unsigned long
2071calc_load(unsigned long load, unsigned long exp, unsigned long active)
2072{
2073	load *= exp;
2074	load += active * (FIXED_1 - exp);
2075	load += 1UL << (FSHIFT - 1);
2076	return load >> FSHIFT;
2077}
2078
2079#ifdef CONFIG_NO_HZ
2080/*
2081 * Handle NO_HZ for the global load-average.
2082 *
2083 * Since the above described distributed algorithm to compute the global
2084 * load-average relies on per-cpu sampling from the tick, it is affected by
2085 * NO_HZ.
2086 *
2087 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2088 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2089 * when we read the global state.
2090 *
2091 * Obviously reality has to ruin such a delightfully simple scheme:
2092 *
2093 *  - When we go NO_HZ idle during the window, we can negate our sample
2094 *    contribution, causing under-accounting.
2095 *
2096 *    We avoid this by keeping two idle-delta counters and flipping them
2097 *    when the window starts, thus separating old and new NO_HZ load.
2098 *
2099 *    The only trick is the slight shift in index flip for read vs write.
2100 *
2101 *        0s            5s            10s           15s
2102 *          +10           +10           +10           +10
2103 *        |-|-----------|-|-----------|-|-----------|-|
2104 *    r:0 0 1           1 0           0 1           1 0
2105 *    w:0 1 1           0 0           1 1           0 0
2106 *
2107 *    This ensures we'll fold the old idle contribution in this window while
2108 *    accumlating the new one.
2109 *
2110 *  - When we wake up from NO_HZ idle during the window, we push up our
2111 *    contribution, since we effectively move our sample point to a known
2112 *    busy state.
2113 *
2114 *    This is solved by pushing the window forward, and thus skipping the
2115 *    sample, for this cpu (effectively using the idle-delta for this cpu which
2116 *    was in effect at the time the window opened). This also solves the issue
2117 *    of having to deal with a cpu having been in NOHZ idle for multiple
2118 *    LOAD_FREQ intervals.
2119 *
2120 * When making the ILB scale, we should try to pull this in as well.
2121 */
2122static atomic_long_t calc_load_idle[2];
2123static int calc_load_idx;
2124
2125static inline int calc_load_write_idx(void)
2126{
2127	int idx = calc_load_idx;
2128
2129	/*
2130	 * See calc_global_nohz(), if we observe the new index, we also
2131	 * need to observe the new update time.
2132	 */
2133	smp_rmb();
2134
2135	/*
2136	 * If the folding window started, make sure we start writing in the
2137	 * next idle-delta.
2138	 */
2139	if (!time_before(jiffies, calc_load_update))
2140		idx++;
2141
2142	return idx & 1;
2143}
2144
2145static inline int calc_load_read_idx(void)
2146{
2147	return calc_load_idx & 1;
2148}
2149
2150void calc_load_enter_idle(void)
2151{
2152	struct rq *this_rq = this_rq();
2153	long delta;
2154
2155	/*
2156	 * We're going into NOHZ mode, if there's any pending delta, fold it
2157	 * into the pending idle delta.
2158	 */
2159	delta = calc_load_fold_active(this_rq);
2160	if (delta) {
2161		int idx = calc_load_write_idx();
2162		atomic_long_add(delta, &calc_load_idle[idx]);
2163	}
2164}
2165
2166void calc_load_exit_idle(void)
2167{
2168	struct rq *this_rq = this_rq();
2169
2170	/*
2171	 * If we're still before the sample window, we're done.
2172	 */
2173	if (time_before(jiffies, this_rq->calc_load_update))
2174		return;
2175
2176	/*
2177	 * We woke inside or after the sample window, this means we're already
2178	 * accounted through the nohz accounting, so skip the entire deal and
2179	 * sync up for the next window.
2180	 */
2181	this_rq->calc_load_update = calc_load_update;
2182	if (time_before(jiffies, this_rq->calc_load_update + 10))
2183		this_rq->calc_load_update += LOAD_FREQ;
2184}
2185
2186static long calc_load_fold_idle(void)
2187{
2188	int idx = calc_load_read_idx();
2189	long delta = 0;
2190
2191	if (atomic_long_read(&calc_load_idle[idx]))
2192		delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2193
2194	return delta;
2195}
2196
2197/**
2198 * fixed_power_int - compute: x^n, in O(log n) time
2199 *
2200 * @x:         base of the power
2201 * @frac_bits: fractional bits of @x
2202 * @n:         power to raise @x to.
2203 *
2204 * By exploiting the relation between the definition of the natural power
2205 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2206 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2207 * (where: n_i \elem {0, 1}, the binary vector representing n),
2208 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2209 * of course trivially computable in O(log_2 n), the length of our binary
2210 * vector.
2211 */
2212static unsigned long
2213fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2214{
2215	unsigned long result = 1UL << frac_bits;
2216
2217	if (n) for (;;) {
2218		if (n & 1) {
2219			result *= x;
2220			result += 1UL << (frac_bits - 1);
2221			result >>= frac_bits;
2222		}
2223		n >>= 1;
2224		if (!n)
2225			break;
2226		x *= x;
2227		x += 1UL << (frac_bits - 1);
2228		x >>= frac_bits;
2229	}
2230
2231	return result;
2232}
2233
2234/*
2235 * a1 = a0 * e + a * (1 - e)
2236 *
2237 * a2 = a1 * e + a * (1 - e)
2238 *    = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2239 *    = a0 * e^2 + a * (1 - e) * (1 + e)
2240 *
2241 * a3 = a2 * e + a * (1 - e)
2242 *    = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2243 *    = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2244 *
2245 *  ...
2246 *
2247 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2248 *    = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2249 *    = a0 * e^n + a * (1 - e^n)
2250 *
2251 * [1] application of the geometric series:
2252 *
2253 *              n         1 - x^(n+1)
2254 *     S_n := \Sum x^i = -------------
2255 *             i=0          1 - x
2256 */
2257static unsigned long
2258calc_load_n(unsigned long load, unsigned long exp,
2259	    unsigned long active, unsigned int n)
2260{
2261
2262	return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2263}
2264
2265/*
2266 * NO_HZ can leave us missing all per-cpu ticks calling
2267 * calc_load_account_active(), but since an idle CPU folds its delta into
2268 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2269 * in the pending idle delta if our idle period crossed a load cycle boundary.
2270 *
2271 * Once we've updated the global active value, we need to apply the exponential
2272 * weights adjusted to the number of cycles missed.
2273 */
2274static void calc_global_nohz(void)
2275{
2276	long delta, active, n;
2277
2278	if (!time_before(jiffies, calc_load_update + 10)) {
2279		/*
2280		 * Catch-up, fold however many we are behind still
2281		 */
2282		delta = jiffies - calc_load_update - 10;
2283		n = 1 + (delta / LOAD_FREQ);
2284
2285		active = atomic_long_read(&calc_load_tasks);
2286		active = active > 0 ? active * FIXED_1 : 0;
2287
2288		avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2289		avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2290		avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2291
2292		calc_load_update += n * LOAD_FREQ;
2293	}
2294
2295	/*
2296	 * Flip the idle index...
2297	 *
2298	 * Make sure we first write the new time then flip the index, so that
2299	 * calc_load_write_idx() will see the new time when it reads the new
2300	 * index, this avoids a double flip messing things up.
2301	 */
2302	smp_wmb();
2303	calc_load_idx++;
2304}
2305#else /* !CONFIG_NO_HZ */
2306
2307static inline long calc_load_fold_idle(void) { return 0; }
2308static inline void calc_global_nohz(void) { }
2309
2310#endif /* CONFIG_NO_HZ */
2311
2312/*
2313 * calc_load - update the avenrun load estimates 10 ticks after the
2314 * CPUs have updated calc_load_tasks.
2315 */
2316void calc_global_load(unsigned long ticks)
2317{
2318	long active, delta;
2319
2320	if (time_before(jiffies, calc_load_update + 10))
2321		return;
2322
2323	/*
2324	 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2325	 */
2326	delta = calc_load_fold_idle();
2327	if (delta)
2328		atomic_long_add(delta, &calc_load_tasks);
2329
2330	active = atomic_long_read(&calc_load_tasks);
2331	active = active > 0 ? active * FIXED_1 : 0;
2332
2333	avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2334	avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2335	avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2336
2337	calc_load_update += LOAD_FREQ;
2338
2339	/*
2340	 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2341	 */
2342	calc_global_nohz();
2343}
2344
2345/*
2346 * Called from update_cpu_load() to periodically update this CPU's
2347 * active count.
2348 */
2349static void calc_load_account_active(struct rq *this_rq)
2350{
2351	long delta;
2352
2353	if (time_before(jiffies, this_rq->calc_load_update))
2354		return;
2355
2356	delta  = calc_load_fold_active(this_rq);
2357	if (delta)
2358		atomic_long_add(delta, &calc_load_tasks);
2359
2360	this_rq->calc_load_update += LOAD_FREQ;
2361}
2362
2363/*
2364 * End of global load-average stuff
2365 */
2366
2367/*
2368 * The exact cpuload at various idx values, calculated at every tick would be
2369 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2370 *
2371 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2372 * on nth tick when cpu may be busy, then we have:
2373 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2374 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2375 *
2376 * decay_load_missed() below does efficient calculation of
2377 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2378 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2379 *
2380 * The calculation is approximated on a 128 point scale.
2381 * degrade_zero_ticks is the number of ticks after which load at any
2382 * particular idx is approximated to be zero.
2383 * degrade_factor is a precomputed table, a row for each load idx.
2384 * Each column corresponds to degradation factor for a power of two ticks,
2385 * based on 128 point scale.
2386 * Example:
2387 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2388 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2389 *
2390 * With this power of 2 load factors, we can degrade the load n times
2391 * by looking at 1 bits in n and doing as many mult/shift instead of
2392 * n mult/shifts needed by the exact degradation.
2393 */
2394#define DEGRADE_SHIFT		7
2395static const unsigned char
2396		degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2397static const unsigned char
2398		degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2399					{0, 0, 0, 0, 0, 0, 0, 0},
2400					{64, 32, 8, 0, 0, 0, 0, 0},
2401					{96, 72, 40, 12, 1, 0, 0},
2402					{112, 98, 75, 43, 15, 1, 0},
2403					{120, 112, 98, 76, 45, 16, 2} };
2404
2405/*
2406 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2407 * would be when CPU is idle and so we just decay the old load without
2408 * adding any new load.
2409 */
2410static unsigned long
2411decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2412{
2413	int j = 0;
2414
2415	if (!missed_updates)
2416		return load;
2417
2418	if (missed_updates >= degrade_zero_ticks[idx])
2419		return 0;
2420
2421	if (idx == 1)
2422		return load >> missed_updates;
2423
2424	while (missed_updates) {
2425		if (missed_updates % 2)
2426			load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2427
2428		missed_updates >>= 1;
2429		j++;
2430	}
2431	return load;
2432}
2433
2434/*
2435 * Update rq->cpu_load[] statistics. This function is usually called every
2436 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2437 * every tick. We fix it up based on jiffies.
2438 */
2439static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2440			      unsigned long pending_updates)
2441{
2442	int i, scale;
2443
2444	this_rq->nr_load_updates++;
2445
2446	/* Update our load: */
2447	this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2448	for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2449		unsigned long old_load, new_load;
2450
2451		/* scale is effectively 1 << i now, and >> i divides by scale */
2452
2453		old_load = this_rq->cpu_load[i];
2454		old_load = decay_load_missed(old_load, pending_updates - 1, i);
2455		new_load = this_load;
2456		/*
2457		 * Round up the averaging division if load is increasing. This
2458		 * prevents us from getting stuck on 9 if the load is 10, for
2459		 * example.
2460		 */
2461		if (new_load > old_load)
2462			new_load += scale - 1;
2463
2464		this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2465	}
2466
2467	sched_avg_update(this_rq);
2468}
2469
2470#ifdef CONFIG_NO_HZ
2471/*
2472 * There is no sane way to deal with nohz on smp when using jiffies because the
2473 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2474 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2475 *
2476 * Therefore we cannot use the delta approach from the regular tick since that
2477 * would seriously skew the load calculation. However we'll make do for those
2478 * updates happening while idle (nohz_idle_balance) or coming out of idle
2479 * (tick_nohz_idle_exit).
2480 *
2481 * This means we might still be one tick off for nohz periods.
2482 */
2483
2484/*
2485 * Called from nohz_idle_balance() to update the load ratings before doing the
2486 * idle balance.
2487 */
2488void update_idle_cpu_load(struct rq *this_rq)
2489{
2490	unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2491	unsigned long load = this_rq->load.weight;
2492	unsigned long pending_updates;
2493
2494	/*
2495	 * bail if there's load or we're actually up-to-date.
2496	 */
2497	if (load || curr_jiffies == this_rq->last_load_update_tick)
2498		return;
2499
2500	pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2501	this_rq->last_load_update_tick = curr_jiffies;
2502
2503	__update_cpu_load(this_rq, load, pending_updates);
2504}
2505
2506/*
2507 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2508 */
2509void update_cpu_load_nohz(void)
2510{
2511	struct rq *this_rq = this_rq();
2512	unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2513	unsigned long pending_updates;
2514
2515	if (curr_jiffies == this_rq->last_load_update_tick)
2516		return;
2517
2518	raw_spin_lock(&this_rq->lock);
2519	pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2520	if (pending_updates) {
2521		this_rq->last_load_update_tick = curr_jiffies;
2522		/*
2523		 * We were idle, this means load 0, the current load might be
2524		 * !0 due to remote wakeups and the sort.
2525		 */
2526		__update_cpu_load(this_rq, 0, pending_updates);
2527	}
2528	raw_spin_unlock(&this_rq->lock);
2529}
2530#endif /* CONFIG_NO_HZ */
2531
2532/*
2533 * Called from scheduler_tick()
2534 */
2535static void update_cpu_load_active(struct rq *this_rq)
2536{
2537	/*
2538	 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2539	 */
2540	this_rq->last_load_update_tick = jiffies;
2541	__update_cpu_load(this_rq, this_rq->load.weight, 1);
2542
2543	calc_load_account_active(this_rq);
2544}
2545
2546#ifdef CONFIG_SMP
2547
2548/*
2549 * sched_exec - execve() is a valuable balancing opportunity, because at
2550 * this point the task has the smallest effective memory and cache footprint.
2551 */
2552void sched_exec(void)
2553{
2554	struct task_struct *p = current;
2555	unsigned long flags;
2556	int dest_cpu;
2557
2558	raw_spin_lock_irqsave(&p->pi_lock, flags);
2559	dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2560	if (dest_cpu == smp_processor_id())
2561		goto unlock;
2562
2563	if (likely(cpu_active(dest_cpu))) {
2564		struct migration_arg arg = { p, dest_cpu };
2565
2566		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2567		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2568		return;
2569	}
2570unlock:
2571	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2572}
2573
2574#endif
2575
2576DEFINE_PER_CPU(struct kernel_stat, kstat);
2577DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2578
2579EXPORT_PER_CPU_SYMBOL(kstat);
2580EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2581
2582/*
2583 * Return any ns on the sched_clock that have not yet been accounted in
2584 * @p in case that task is currently running.
2585 *
2586 * Called with task_rq_lock() held on @rq.
2587 */
2588static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2589{
2590	u64 ns = 0;
2591
2592	if (task_current(rq, p)) {
2593		update_rq_clock(rq);
2594		ns = rq->clock_task - p->se.exec_start;
2595		if ((s64)ns < 0)
2596			ns = 0;
2597	}
2598
2599	return ns;
2600}
2601
2602unsigned long long task_delta_exec(struct task_struct *p)
2603{
2604	unsigned long flags;
2605	struct rq *rq;
2606	u64 ns = 0;
2607
2608	rq = task_rq_lock(p, &flags);
2609	ns = do_task_delta_exec(p, rq);
2610	task_rq_unlock(rq, p, &flags);
2611
2612	return ns;
2613}
2614
2615/*
2616 * Return accounted runtime for the task.
2617 * In case the task is currently running, return the runtime plus current's
2618 * pending runtime that have not been accounted yet.
2619 */
2620unsigned long long task_sched_runtime(struct task_struct *p)
2621{
2622	unsigned long flags;
2623	struct rq *rq;
2624	u64 ns = 0;
2625
2626	rq = task_rq_lock(p, &flags);
2627	ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2628	task_rq_unlock(rq, p, &flags);
2629
2630	return ns;
2631}
2632
2633/*
2634 * This function gets called by the timer code, with HZ frequency.
2635 * We call it with interrupts disabled.
2636 */
2637void scheduler_tick(void)
2638{
2639	int cpu = smp_processor_id();
2640	struct rq *rq = cpu_rq(cpu);
2641	struct task_struct *curr = rq->curr;
2642
2643	sched_clock_tick();
2644
2645	raw_spin_lock(&rq->lock);
2646	update_rq_clock(rq);
2647	update_cpu_load_active(rq);
2648	curr->sched_class->task_tick(rq, curr, 0);
2649	raw_spin_unlock(&rq->lock);
2650
2651	perf_event_task_tick();
2652
2653#ifdef CONFIG_SMP
2654	rq->idle_balance = idle_cpu(cpu);
2655	trigger_load_balance(rq, cpu);
2656#endif
2657}
2658
2659notrace unsigned long get_parent_ip(unsigned long addr)
2660{
2661	if (in_lock_functions(addr)) {
2662		addr = CALLER_ADDR2;
2663		if (in_lock_functions(addr))
2664			addr = CALLER_ADDR3;
2665	}
2666	return addr;
2667}
2668
2669#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2670				defined(CONFIG_PREEMPT_TRACER))
2671
2672void __kprobes add_preempt_count(int val)
2673{
2674#ifdef CONFIG_DEBUG_PREEMPT
2675	/*
2676	 * Underflow?
2677	 */
2678	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2679		return;
2680#endif
2681	preempt_count() += val;
2682#ifdef CONFIG_DEBUG_PREEMPT
2683	/*
2684	 * Spinlock count overflowing soon?
2685	 */
2686	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2687				PREEMPT_MASK - 10);
2688#endif
2689	if (preempt_count() == val)
2690		trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2691}
2692EXPORT_SYMBOL(add_preempt_count);
2693
2694void __kprobes sub_preempt_count(int val)
2695{
2696#ifdef CONFIG_DEBUG_PREEMPT
2697	/*
2698	 * Underflow?
2699	 */
2700	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2701		return;
2702	/*
2703	 * Is the spinlock portion underflowing?
2704	 */
2705	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2706			!(preempt_count() & PREEMPT_MASK)))
2707		return;
2708#endif
2709
2710	if (preempt_count() == val)
2711		trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2712	preempt_count() -= val;
2713}
2714EXPORT_SYMBOL(sub_preempt_count);
2715
2716#endif
2717
2718/*
2719 * Print scheduling while atomic bug:
2720 */
2721static noinline void __schedule_bug(struct task_struct *prev)
2722{
2723	if (oops_in_progress)
2724		return;
2725
2726	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2727		prev->comm, prev->pid, preempt_count());
2728
2729	debug_show_held_locks(prev);
2730	print_modules();
2731	if (irqs_disabled())
2732		print_irqtrace_events(prev);
2733	dump_stack();
2734	add_taint(TAINT_WARN);
2735}
2736
2737/*
2738 * Various schedule()-time debugging checks and statistics:
2739 */
2740static inline void schedule_debug(struct task_struct *prev)
2741{
2742	/*
2743	 * Test if we are atomic. Since do_exit() needs to call into
2744	 * schedule() atomically, we ignore that path for now.
2745	 * Otherwise, whine if we are scheduling when we should not be.
2746	 */
2747	if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2748		__schedule_bug(prev);
2749	rcu_sleep_check();
2750
2751	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2752
2753	schedstat_inc(this_rq(), sched_count);
2754}
2755
2756static void put_prev_task(struct rq *rq, struct task_struct *prev)
2757{
2758	if (prev->on_rq || rq->skip_clock_update < 0)
2759		update_rq_clock(rq);
2760	prev->sched_class->put_prev_task(rq, prev);
2761}
2762
2763/*
2764 * Pick up the highest-prio task:
2765 */
2766static inline struct task_struct *
2767pick_next_task(struct rq *rq)
2768{
2769	const struct sched_class *class;
2770	struct task_struct *p;
2771
2772	/*
2773	 * Optimization: we know that if all tasks are in
2774	 * the fair class we can call that function directly:
2775	 */
2776	if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2777		p = fair_sched_class.pick_next_task(rq);
2778		if (likely(p))
2779			return p;
2780	}
2781
2782	for_each_class(class) {
2783		p = class->pick_next_task(rq);
2784		if (p)
2785			return p;
2786	}
2787
2788	BUG(); /* the idle class will always have a runnable task */
2789}
2790
2791/*
2792 * __schedule() is the main scheduler function.
2793 *
2794 * The main means of driving the scheduler and thus entering this function are:
2795 *
2796 *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2797 *
2798 *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2799 *      paths. For example, see arch/x86/entry_64.S.
2800 *
2801 *      To drive preemption between tasks, the scheduler sets the flag in timer
2802 *      interrupt handler scheduler_tick().
2803 *
2804 *   3. Wakeups don't really cause entry into schedule(). They add a
2805 *      task to the run-queue and that's it.
2806 *
2807 *      Now, if the new task added to the run-queue preempts the current
2808 *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2809 *      called on the nearest possible occasion:
2810 *
2811 *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2812 *
2813 *         - in syscall or exception context, at the next outmost
2814 *           preempt_enable(). (this might be as soon as the wake_up()'s
2815 *           spin_unlock()!)
2816 *
2817 *         - in IRQ context, return from interrupt-handler to
2818 *           preemptible context
2819 *
2820 *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2821 *         then at the next:
2822 *
2823 *          - cond_resched() call
2824 *          - explicit schedule() call
2825 *          - return from syscall or exception to user-space
2826 *          - return from interrupt-handler to user-space
2827 */
2828static void __sched __schedule(void)
2829{
2830	struct task_struct *prev, *next;
2831	unsigned long *switch_count;
2832	struct rq *rq;
2833	int cpu;
2834
2835need_resched:
2836	preempt_disable();
2837	cpu = smp_processor_id();
2838	rq = cpu_rq(cpu);
2839	rcu_note_context_switch(cpu);
2840	prev = rq->curr;
2841
2842	schedule_debug(prev);
2843
2844	if (sched_feat(HRTICK))
2845		hrtick_clear(rq);
2846
2847	raw_spin_lock_irq(&rq->lock);
2848
2849	switch_count = &prev->nivcsw;
2850	if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2851		if (unlikely(signal_pending_state(prev->state, prev))) {
2852			prev->state = TASK_RUNNING;
2853		} else {
2854			deactivate_task(rq, prev, DEQUEUE_SLEEP);
2855			prev->on_rq = 0;
2856
2857			/*
2858			 * If a worker went to sleep, notify and ask workqueue
2859			 * whether it wants to wake up a task to maintain
2860			 * concurrency.
2861			 */
2862			if (prev->flags & PF_WQ_WORKER) {
2863				struct task_struct *to_wakeup;
2864
2865				to_wakeup = wq_worker_sleeping(prev, cpu);
2866				if (to_wakeup)
2867					try_to_wake_up_local(to_wakeup);
2868			}
2869		}
2870		switch_count = &prev->nvcsw;
2871	}
2872
2873	pre_schedule(rq, prev);
2874
2875	if (unlikely(!rq->nr_running))
2876		idle_balance(cpu, rq);
2877
2878	put_prev_task(rq, prev);
2879	next = pick_next_task(rq);
2880	clear_tsk_need_resched(prev);
2881	rq->skip_clock_update = 0;
2882
2883	if (likely(prev != next)) {
2884		rq->nr_switches++;
2885		rq->curr = next;
2886		++*switch_count;
2887
2888		context_switch(rq, prev, next); /* unlocks the rq */
2889		/*
2890		 * The context switch have flipped the stack from under us
2891		 * and restored the local variables which were saved when
2892		 * this task called schedule() in the past. prev == current
2893		 * is still correct, but it can be moved to another cpu/rq.
2894		 */
2895		cpu = smp_processor_id();
2896		rq = cpu_rq(cpu);
2897	} else
2898		raw_spin_unlock_irq(&rq->lock);
2899
2900	post_schedule(rq);
2901
2902	sched_preempt_enable_no_resched();
2903	if (need_resched())
2904		goto need_resched;
2905}
2906
2907static inline void sched_submit_work(struct task_struct *tsk)
2908{
2909	if (!tsk->state || tsk_is_pi_blocked(tsk))
2910		return;
2911	/*
2912	 * If we are going to sleep and we have plugged IO queued,
2913	 * make sure to submit it to avoid deadlocks.
2914	 */
2915	if (blk_needs_flush_plug(tsk))
2916		blk_schedule_flush_plug(tsk);
2917}
2918
2919asmlinkage void __sched schedule(void)
2920{
2921	struct task_struct *tsk = current;
2922
2923	sched_submit_work(tsk);
2924	__schedule();
2925}
2926EXPORT_SYMBOL(schedule);
2927
2928#ifdef CONFIG_RCU_USER_QS
2929asmlinkage void __sched schedule_user(void)
2930{
2931	/*
2932	 * If we come here after a random call to set_need_resched(),
2933	 * or we have been woken up remotely but the IPI has not yet arrived,
2934	 * we haven't yet exited the RCU idle mode. Do it here manually until
2935	 * we find a better solution.
2936	 */
2937	rcu_user_exit();
2938	schedule();
2939	rcu_user_enter();
2940}
2941#endif
2942
2943/**
2944 * schedule_preempt_disabled - called with preemption disabled
2945 *
2946 * Returns with preemption disabled. Note: preempt_count must be 1
2947 */
2948void __sched schedule_preempt_disabled(void)
2949{
2950	sched_preempt_enable_no_resched();
2951	schedule();
2952	preempt_disable();
2953}
2954
2955#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2956
2957static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2958{
2959	if (lock->owner != owner)
2960		return false;
2961
2962	/*
2963	 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2964	 * lock->owner still matches owner, if that fails, owner might
2965	 * point to free()d memory, if it still matches, the rcu_read_lock()
2966	 * ensures the memory stays valid.
2967	 */
2968	barrier();
2969
2970	return owner->on_cpu;
2971}
2972
2973/*
2974 * Look out! "owner" is an entirely speculative pointer
2975 * access and not reliable.
2976 */
2977int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
2978{
2979	if (!sched_feat(OWNER_SPIN))
2980		return 0;
2981
2982	rcu_read_lock();
2983	while (owner_running(lock, owner)) {
2984		if (need_resched())
2985			break;
2986
2987		arch_mutex_cpu_relax();
2988	}
2989	rcu_read_unlock();
2990
2991	/*
2992	 * We break out the loop above on need_resched() and when the
2993	 * owner changed, which is a sign for heavy contention. Return
2994	 * success only when lock->owner is NULL.
2995	 */
2996	return lock->owner == NULL;
2997}
2998#endif
2999
3000#ifdef CONFIG_PREEMPT
3001/*
3002 * this is the entry point to schedule() from in-kernel preemption
3003 * off of preempt_enable. Kernel preemptions off return from interrupt
3004 * occur there and call schedule directly.
3005 */
3006asmlinkage void __sched notrace preempt_schedule(void)
3007{
3008	struct thread_info *ti = current_thread_info();
3009
3010	/*
3011	 * If there is a non-zero preempt_count or interrupts are disabled,
3012	 * we do not want to preempt the current task. Just return..
3013	 */
3014	if (likely(ti->preempt_count || irqs_disabled()))
3015		return;
3016
3017	do {
3018		add_preempt_count_notrace(PREEMPT_ACTIVE);
3019		__schedule();
3020		sub_preempt_count_notrace(PREEMPT_ACTIVE);
3021
3022		/*
3023		 * Check again in case we missed a preemption opportunity
3024		 * between schedule and now.
3025		 */
3026		barrier();
3027	} while (need_resched());
3028}
3029EXPORT_SYMBOL(preempt_schedule);
3030
3031/*
3032 * this is the entry point to schedule() from kernel preemption
3033 * off of irq context.
3034 * Note, that this is called and return with irqs disabled. This will
3035 * protect us against recursive calling from irq.
3036 */
3037asmlinkage void __sched preempt_schedule_irq(void)
3038{
3039	struct thread_info *ti = current_thread_info();
3040
3041	/* Catch callers which need to be fixed */
3042	BUG_ON(ti->preempt_count || !irqs_disabled());
3043
3044	rcu_user_exit();
3045	do {
3046		add_preempt_count(PREEMPT_ACTIVE);
3047		local_irq_enable();
3048		__schedule();
3049		local_irq_disable();
3050		sub_preempt_count(PREEMPT_ACTIVE);
3051
3052		/*
3053		 * Check again in case we missed a preemption opportunity
3054		 * between schedule and now.
3055		 */
3056		barrier();
3057	} while (need_resched());
3058}
3059
3060#endif /* CONFIG_PREEMPT */
3061
3062int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3063			  void *key)
3064{
3065	return try_to_wake_up(curr->private, mode, wake_flags);
3066}
3067EXPORT_SYMBOL(default_wake_function);
3068
3069/*
3070 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3071 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3072 * number) then we wake all the non-exclusive tasks and one exclusive task.
3073 *
3074 * There are circumstances in which we can try to wake a task which has already
3075 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3076 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3077 */
3078static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3079			int nr_exclusive, int wake_flags, void *key)
3080{
3081	wait_queue_t *curr, *next;
3082
3083	list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3084		unsigned flags = curr->flags;
3085
3086		if (curr->func(curr, mode, wake_flags, key) &&
3087				(flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3088			break;
3089	}
3090}
3091
3092/**
3093 * __wake_up - wake up threads blocked on a waitqueue.
3094 * @q: the waitqueue
3095 * @mode: which threads
3096 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3097 * @key: is directly passed to the wakeup function
3098 *
3099 * It may be assumed that this function implies a write memory barrier before
3100 * changing the task state if and only if any tasks are woken up.
3101 */
3102void __wake_up(wait_queue_head_t *q, unsigned int mode,
3103			int nr_exclusive, void *key)
3104{
3105	unsigned long flags;
3106
3107	spin_lock_irqsave(&q->lock, flags);
3108	__wake_up_common(q, mode, nr_exclusive, 0, key);
3109	spin_unlock_irqrestore(&q->lock, flags);
3110}
3111EXPORT_SYMBOL(__wake_up);
3112
3113/*
3114 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3115 */
3116void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3117{
3118	__wake_up_common(q, mode, nr, 0, NULL);
3119}
3120EXPORT_SYMBOL_GPL(__wake_up_locked);
3121
3122void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3123{
3124	__wake_up_common(q, mode, 1, 0, key);
3125}
3126EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3127
3128/**
3129 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3130 * @q: the waitqueue
3131 * @mode: which threads
3132 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3133 * @key: opaque value to be passed to wakeup targets
3134 *
3135 * The sync wakeup differs that the waker knows that it will schedule
3136 * away soon, so while the target thread will be woken up, it will not
3137 * be migrated to another CPU - ie. the two threads are 'synchronized'
3138 * with each other. This can prevent needless bouncing between CPUs.
3139 *
3140 * On UP it can prevent extra preemption.
3141 *
3142 * It may be assumed that this function implies a write memory barrier before
3143 * changing the task state if and only if any tasks are woken up.
3144 */
3145void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3146			int nr_exclusive, void *key)
3147{
3148	unsigned long flags;
3149	int wake_flags = WF_SYNC;
3150
3151	if (unlikely(!q))
3152		return;
3153
3154	if (unlikely(!nr_exclusive))
3155		wake_flags = 0;
3156
3157	spin_lock_irqsave(&q->lock, flags);
3158	__wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3159	spin_unlock_irqrestore(&q->lock, flags);
3160}
3161EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3162
3163/*
3164 * __wake_up_sync - see __wake_up_sync_key()
3165 */
3166void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3167{
3168	__wake_up_sync_key(q, mode, nr_exclusive, NULL);
3169}
3170EXPORT_SYMBOL_GPL(__wake_up_sync);	/* For internal use only */
3171
3172/**
3173 * complete: - signals a single thread waiting on this completion
3174 * @x:  holds the state of this particular completion
3175 *
3176 * This will wake up a single thread waiting on this completion. Threads will be
3177 * awakened in the same order in which they were queued.
3178 *
3179 * See also complete_all(), wait_for_completion() and related routines.
3180 *
3181 * It may be assumed that this function implies a write memory barrier before
3182 * changing the task state if and only if any tasks are woken up.
3183 */
3184void complete(struct completion *x)
3185{
3186	unsigned long flags;
3187
3188	spin_lock_irqsave(&x->wait.lock, flags);
3189	x->done++;
3190	__wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3191	spin_unlock_irqrestore(&x->wait.lock, flags);
3192}
3193EXPORT_SYMBOL(complete);
3194
3195/**
3196 * complete_all: - signals all threads waiting on this completion
3197 * @x:  holds the state of this particular completion
3198 *
3199 * This will wake up all threads waiting on this particular completion event.
3200 *
3201 * It may be assumed that this function implies a write memory barrier before
3202 * changing the task state if and only if any tasks are woken up.
3203 */
3204void complete_all(struct completion *x)
3205{
3206	unsigned long flags;
3207
3208	spin_lock_irqsave(&x->wait.lock, flags);
3209	x->done += UINT_MAX/2;
3210	__wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3211	spin_unlock_irqrestore(&x->wait.lock, flags);
3212}
3213EXPORT_SYMBOL(complete_all);
3214
3215static inline long __sched
3216do_wait_for_common(struct completion *x, long timeout, int state)
3217{
3218	if (!x->done) {
3219		DECLARE_WAITQUEUE(wait, current);
3220
3221		__add_wait_queue_tail_exclusive(&x->wait, &wait);
3222		do {
3223			if (signal_pending_state(state, current)) {
3224				timeout = -ERESTARTSYS;
3225				break;
3226			}
3227			__set_current_state(state);
3228			spin_unlock_irq(&x->wait.lock);
3229			timeout = schedule_timeout(timeout);
3230			spin_lock_irq(&x->wait.lock);
3231		} while (!x->done && timeout);
3232		__remove_wait_queue(&x->wait, &wait);
3233		if (!x->done)
3234			return timeout;
3235	}
3236	x->done--;
3237	return timeout ?: 1;
3238}
3239
3240static long __sched
3241wait_for_common(struct completion *x, long timeout, int state)
3242{
3243	might_sleep();
3244
3245	spin_lock_irq(&x->wait.lock);
3246	timeout = do_wait_for_common(x, timeout, state);
3247	spin_unlock_irq(&x->wait.lock);
3248	return timeout;
3249}
3250
3251/**
3252 * wait_for_completion: - waits for completion of a task
3253 * @x:  holds the state of this particular completion
3254 *
3255 * This waits to be signaled for completion of a specific task. It is NOT
3256 * interruptible and there is no timeout.
3257 *
3258 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3259 * and interrupt capability. Also see complete().
3260 */
3261void __sched wait_for_completion(struct completion *x)
3262{
3263	wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3264}
3265EXPORT_SYMBOL(wait_for_completion);
3266
3267/**
3268 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3269 * @x:  holds the state of this particular completion
3270 * @timeout:  timeout value in jiffies
3271 *
3272 * This waits for either a completion of a specific task to be signaled or for a
3273 * specified timeout to expire. The timeout is in jiffies. It is not
3274 * interruptible.
3275 *
3276 * The return value is 0 if timed out, and positive (at least 1, or number of
3277 * jiffies left till timeout) if completed.
3278 */
3279unsigned long __sched
3280wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3281{
3282	return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3283}
3284EXPORT_SYMBOL(wait_for_completion_timeout);
3285
3286/**
3287 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3288 * @x:  holds the state of this particular completion
3289 *
3290 * This waits for completion of a specific task to be signaled. It is
3291 * interruptible.
3292 *
3293 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3294 */
3295int __sched wait_for_completion_interruptible(struct completion *x)
3296{
3297	long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3298	if (t == -ERESTARTSYS)
3299		return t;
3300	return 0;
3301}
3302EXPORT_SYMBOL(wait_for_completion_interruptible);
3303
3304/**
3305 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3306 * @x:  holds the state of this particular completion
3307 * @timeout:  timeout value in jiffies
3308 *
3309 * This waits for either a completion of a specific task to be signaled or for a
3310 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3311 *
3312 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3313 * positive (at least 1, or number of jiffies left till timeout) if completed.
3314 */
3315long __sched
3316wait_for_completion_interruptible_timeout(struct completion *x,
3317					  unsigned long timeout)
3318{
3319	return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3320}
3321EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3322
3323/**
3324 * wait_for_completion_killable: - waits for completion of a task (killable)
3325 * @x:  holds the state of this particular completion
3326 *
3327 * This waits to be signaled for completion of a specific task. It can be
3328 * interrupted by a kill signal.
3329 *
3330 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3331 */
3332int __sched wait_for_completion_killable(struct completion *x)
3333{
3334	long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3335	if (t == -ERESTARTSYS)
3336		return t;
3337	return 0;
3338}
3339EXPORT_SYMBOL(wait_for_completion_killable);
3340
3341/**
3342 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3343 * @x:  holds the state of this particular completion
3344 * @timeout:  timeout value in jiffies
3345 *
3346 * This waits for either a completion of a specific task to be
3347 * signaled or for a specified timeout to expire. It can be
3348 * interrupted by a kill signal. The timeout is in jiffies.
3349 *
3350 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3351 * positive (at least 1, or number of jiffies left till timeout) if completed.
3352 */
3353long __sched
3354wait_for_completion_killable_timeout(struct completion *x,
3355				     unsigned long timeout)
3356{
3357	return wait_for_common(x, timeout, TASK_KILLABLE);
3358}
3359EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3360
3361/**
3362 *	try_wait_for_completion - try to decrement a completion without blocking
3363 *	@x:	completion structure
3364 *
3365 *	Returns: 0 if a decrement cannot be done without blocking
3366 *		 1 if a decrement succeeded.
3367 *
3368 *	If a completion is being used as a counting completion,
3369 *	attempt to decrement the counter without blocking. This
3370 *	enables us to avoid waiting if the resource the completion
3371 *	is protecting is not available.
3372 */
3373bool try_wait_for_completion(struct completion *x)
3374{
3375	unsigned long flags;
3376	int ret = 1;
3377
3378	spin_lock_irqsave(&x->wait.lock, flags);
3379	if (!x->done)
3380		ret = 0;
3381	else
3382		x->done--;
3383	spin_unlock_irqrestore(&x->wait.lock, flags);
3384	return ret;
3385}
3386EXPORT_SYMBOL(try_wait_for_completion);
3387
3388/**
3389 *	completion_done - Test to see if a completion has any waiters
3390 *	@x:	completion structure
3391 *
3392 *	Returns: 0 if there are waiters (wait_for_completion() in progress)
3393 *		 1 if there are no waiters.
3394 *
3395 */
3396bool completion_done(struct completion *x)
3397{
3398	unsigned long flags;
3399	int ret = 1;
3400
3401	spin_lock_irqsave(&x->wait.lock, flags);
3402	if (!x->done)
3403		ret = 0;
3404	spin_unlock_irqrestore(&x->wait.lock, flags);
3405	return ret;
3406}
3407EXPORT_SYMBOL(completion_done);
3408
3409static long __sched
3410sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3411{
3412	unsigned long flags;
3413	wait_queue_t wait;
3414
3415	init_waitqueue_entry(&wait, current);
3416
3417	__set_current_state(state);
3418
3419	spin_lock_irqsave(&q->lock, flags);
3420	__add_wait_queue(q, &wait);
3421	spin_unlock(&q->lock);
3422	timeout = schedule_timeout(timeout);
3423	spin_lock_irq(&q->lock);
3424	__remove_wait_queue(q, &wait);
3425	spin_unlock_irqrestore(&q->lock, flags);
3426
3427	return timeout;
3428}
3429
3430void __sched interruptible_sleep_on(wait_queue_head_t *q)
3431{
3432	sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3433}
3434EXPORT_SYMBOL(interruptible_sleep_on);
3435
3436long __sched
3437interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3438{
3439	return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3440}
3441EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3442
3443void __sched sleep_on(wait_queue_head_t *q)
3444{
3445	sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3446}
3447EXPORT_SYMBOL(sleep_on);
3448
3449long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3450{
3451	return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3452}
3453EXPORT_SYMBOL(sleep_on_timeout);
3454
3455#ifdef CONFIG_RT_MUTEXES
3456
3457/*
3458 * rt_mutex_setprio - set the current priority of a task
3459 * @p: task
3460 * @prio: prio value (kernel-internal form)
3461 *
3462 * This function changes the 'effective' priority of a task. It does
3463 * not touch ->normal_prio like __setscheduler().
3464 *
3465 * Used by the rt_mutex code to implement priority inheritance logic.
3466 */
3467void rt_mutex_setprio(struct task_struct *p, int prio)
3468{
3469	int oldprio, on_rq, running;
3470	struct rq *rq;
3471	const struct sched_class *prev_class;
3472
3473	BUG_ON(prio < 0 || prio > MAX_PRIO);
3474
3475	rq = __task_rq_lock(p);
3476
3477	/*
3478	 * Idle task boosting is a nono in general. There is one
3479	 * exception, when PREEMPT_RT and NOHZ is active:
3480	 *
3481	 * The idle task calls get_next_timer_interrupt() and holds
3482	 * the timer wheel base->lock on the CPU and another CPU wants
3483	 * to access the timer (probably to cancel it). We can safely
3484	 * ignore the boosting request, as the idle CPU runs this code
3485	 * with interrupts disabled and will complete the lock
3486	 * protected section without being interrupted. So there is no
3487	 * real need to boost.
3488	 */
3489	if (unlikely(p == rq->idle)) {
3490		WARN_ON(p != rq->curr);
3491		WARN_ON(p->pi_blocked_on);
3492		goto out_unlock;
3493	}
3494
3495	trace_sched_pi_setprio(p, prio);
3496	oldprio = p->prio;
3497	prev_class = p->sched_class;
3498	on_rq = p->on_rq;
3499	running = task_current(rq, p);
3500	if (on_rq)
3501		dequeue_task(rq, p, 0);
3502	if (running)
3503		p->sched_class->put_prev_task(rq, p);
3504
3505	if (rt_prio(prio))
3506		p->sched_class = &rt_sched_class;
3507	else
3508		p->sched_class = &fair_sched_class;
3509
3510	p->prio = prio;
3511
3512	if (running)
3513		p->sched_class->set_curr_task(rq);
3514	if (on_rq)
3515		enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3516
3517	check_class_changed(rq, p, prev_class, oldprio);
3518out_unlock:
3519	__task_rq_unlock(rq);
3520}
3521#endif
3522void set_user_nice(struct task_struct *p, long nice)
3523{
3524	int old_prio, delta, on_rq;
3525	unsigned long flags;
3526	struct rq *rq;
3527
3528	if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3529		return;
3530	/*
3531	 * We have to be careful, if called from sys_setpriority(),
3532	 * the task might be in the middle of scheduling on another CPU.
3533	 */
3534	rq = task_rq_lock(p, &flags);
3535	/*
3536	 * The RT priorities are set via sched_setscheduler(), but we still
3537	 * allow the 'normal' nice value to be set - but as expected
3538	 * it wont have any effect on scheduling until the task is
3539	 * SCHED_FIFO/SCHED_RR:
3540	 */
3541	if (task_has_rt_policy(p)) {
3542		p->static_prio = NICE_TO_PRIO(nice);
3543		goto out_unlock;
3544	}
3545	on_rq = p->on_rq;
3546	if (on_rq)
3547		dequeue_task(rq, p, 0);
3548
3549	p->static_prio = NICE_TO_PRIO(nice);
3550	set_load_weight(p);
3551	old_prio = p->prio;
3552	p->prio = effective_prio(p);
3553	delta = p->prio - old_prio;
3554
3555	if (on_rq) {
3556		enqueue_task(rq, p, 0);
3557		/*
3558		 * If the task increased its priority or is running and
3559		 * lowered its priority, then reschedule its CPU:
3560		 */
3561		if (delta < 0 || (delta > 0 && task_running(rq, p)))
3562			resched_task(rq->curr);
3563	}
3564out_unlock:
3565	task_rq_unlock(rq, p, &flags);
3566}
3567EXPORT_SYMBOL(set_user_nice);
3568
3569/*
3570 * can_nice - check if a task can reduce its nice value
3571 * @p: task
3572 * @nice: nice value
3573 */
3574int can_nice(const struct task_struct *p, const int nice)
3575{
3576	/* convert nice value [19,-20] to rlimit style value [1,40] */
3577	int nice_rlim = 20 - nice;
3578
3579	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3580		capable(CAP_SYS_NICE));
3581}
3582
3583#ifdef __ARCH_WANT_SYS_NICE
3584
3585/*
3586 * sys_nice - change the priority of the current process.
3587 * @increment: priority increment
3588 *
3589 * sys_setpriority is a more generic, but much slower function that
3590 * does similar things.
3591 */
3592SYSCALL_DEFINE1(nice, int, increment)
3593{
3594	long nice, retval;
3595
3596	/*
3597	 * Setpriority might change our priority at the same moment.
3598	 * We don't have to worry. Conceptually one call occurs first
3599	 * and we have a single winner.
3600	 */
3601	if (increment < -40)
3602		increment = -40;
3603	if (increment > 40)
3604		increment = 40;
3605
3606	nice = TASK_NICE(current) + increment;
3607	if (nice < -20)
3608		nice = -20;
3609	if (nice > 19)
3610		nice = 19;
3611
3612	if (increment < 0 && !can_nice(current, nice))
3613		return -EPERM;
3614
3615	retval = security_task_setnice(current, nice);
3616	if (retval)
3617		return retval;
3618
3619	set_user_nice(current, nice);
3620	return 0;
3621}
3622
3623#endif
3624
3625/**
3626 * task_prio - return the priority value of a given task.
3627 * @p: the task in question.
3628 *
3629 * This is the priority value as seen by users in /proc.
3630 * RT tasks are offset by -200. Normal tasks are centered
3631 * around 0, value goes from -16 to +15.
3632 */
3633int task_prio(const struct task_struct *p)
3634{
3635	return p->prio - MAX_RT_PRIO;
3636}
3637
3638/**
3639 * task_nice - return the nice value of a given task.
3640 * @p: the task in question.
3641 */
3642int task_nice(const struct task_struct *p)
3643{
3644	return TASK_NICE(p);
3645}
3646EXPORT_SYMBOL(task_nice);
3647
3648/**
3649 * idle_cpu - is a given cpu idle currently?
3650 * @cpu: the processor in question.
3651 */
3652int idle_cpu(int cpu)
3653{
3654	struct rq *rq = cpu_rq(cpu);
3655
3656	if (rq->curr != rq->idle)
3657		return 0;
3658
3659	if (rq->nr_running)
3660		return 0;
3661
3662#ifdef CONFIG_SMP
3663	if (!llist_empty(&rq->wake_list))
3664		return 0;
3665#endif
3666
3667	return 1;
3668}
3669
3670/**
3671 * idle_task - return the idle task for a given cpu.
3672 * @cpu: the processor in question.
3673 */
3674struct task_struct *idle_task(int cpu)
3675{
3676	return cpu_rq(cpu)->idle;
3677}
3678
3679/**
3680 * find_process_by_pid - find a process with a matching PID value.
3681 * @pid: the pid in question.
3682 */
3683static struct task_struct *find_process_by_pid(pid_t pid)
3684{
3685	return pid ? find_task_by_vpid(pid) : current;
3686}
3687
3688/* Actually do priority change: must hold rq lock. */
3689static void
3690__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3691{
3692	p->policy = policy;
3693	p->rt_priority = prio;
3694	p->normal_prio = normal_prio(p);
3695	/* we are holding p->pi_lock already */
3696	p->prio = rt_mutex_getprio(p);
3697	if (rt_prio(p->prio))
3698		p->sched_class = &rt_sched_class;
3699	else
3700		p->sched_class = &fair_sched_class;
3701	set_load_weight(p);
3702}
3703
3704/*
3705 * check the target process has a UID that matches the current process's
3706 */
3707static bool check_same_owner(struct task_struct *p)
3708{
3709	const struct cred *cred = current_cred(), *pcred;
3710	bool match;
3711
3712	rcu_read_lock();
3713	pcred = __task_cred(p);
3714	match = (uid_eq(cred->euid, pcred->euid) ||
3715		 uid_eq(cred->euid, pcred->uid));
3716	rcu_read_unlock();
3717	return match;
3718}
3719
3720static int __sched_setscheduler(struct task_struct *p, int policy,
3721				const struct sched_param *param, bool user)
3722{
3723	int retval, oldprio, oldpolicy = -1, on_rq, running;
3724	unsigned long flags;
3725	const struct sched_class *prev_class;
3726	struct rq *rq;
3727	int reset_on_fork;
3728
3729	/* may grab non-irq protected spin_locks */
3730	BUG_ON(in_interrupt());
3731recheck:
3732	/* double check policy once rq lock held */
3733	if (policy < 0) {
3734		reset_on_fork = p->sched_reset_on_fork;
3735		policy = oldpolicy = p->policy;
3736	} else {
3737		reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3738		policy &= ~SCHED_RESET_ON_FORK;
3739
3740		if (policy != SCHED_FIFO && policy != SCHED_RR &&
3741				policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3742				policy != SCHED_IDLE)
3743			return -EINVAL;
3744	}
3745
3746	/*
3747	 * Valid priorities for SCHED_FIFO and SCHED_RR are
3748	 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3749	 * SCHED_BATCH and SCHED_IDLE is 0.
3750	 */
3751	if (param->sched_priority < 0 ||
3752	    (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3753	    (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3754		return -EINVAL;
3755	if (rt_policy(policy) != (param->sched_priority != 0))
3756		return -EINVAL;
3757
3758	/*
3759	 * Allow unprivileged RT tasks to decrease priority:
3760	 */
3761	if (user && !capable(CAP_SYS_NICE)) {
3762		if (rt_policy(policy)) {
3763			unsigned long rlim_rtprio =
3764					task_rlimit(p, RLIMIT_RTPRIO);
3765
3766			/* can't set/change the rt policy */
3767			if (policy != p->policy && !rlim_rtprio)
3768				return -EPERM;
3769
3770			/* can't increase priority */
3771			if (param->sched_priority > p->rt_priority &&
3772			    param->sched_priority > rlim_rtprio)
3773				return -EPERM;
3774		}
3775
3776		/*
3777		 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3778		 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3779		 */
3780		if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3781			if (!can_nice(p, TASK_NICE(p)))
3782				return -EPERM;
3783		}
3784
3785		/* can't change other user's priorities */
3786		if (!check_same_owner(p))
3787			return -EPERM;
3788
3789		/* Normal users shall not reset the sched_reset_on_fork flag */
3790		if (p->sched_reset_on_fork && !reset_on_fork)
3791			return -EPERM;
3792	}
3793
3794	if (user) {
3795		retval = security_task_setscheduler(p);
3796		if (retval)
3797			return retval;
3798	}
3799
3800	/*
3801	 * make sure no PI-waiters arrive (or leave) while we are
3802	 * changing the priority of the task:
3803	 *
3804	 * To be able to change p->policy safely, the appropriate
3805	 * runqueue lock must be held.
3806	 */
3807	rq = task_rq_lock(p, &flags);
3808
3809	/*
3810	 * Changing the policy of the stop threads its a very bad idea
3811	 */
3812	if (p == rq->stop) {
3813		task_rq_unlock(rq, p, &flags);
3814		return -EINVAL;
3815	}
3816
3817	/*
3818	 * If not changing anything there's no need to proceed further:
3819	 */
3820	if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3821			param->sched_priority == p->rt_priority))) {
3822		task_rq_unlock(rq, p, &flags);
3823		return 0;
3824	}
3825
3826#ifdef CONFIG_RT_GROUP_SCHED
3827	if (user) {
3828		/*
3829		 * Do not allow realtime tasks into groups that have no runtime
3830		 * assigned.
3831		 */
3832		if (rt_bandwidth_enabled() && rt_policy(policy) &&
3833				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3834				!task_group_is_autogroup(task_group(p))) {
3835			task_rq_unlock(rq, p, &flags);
3836			return -EPERM;
3837		}
3838	}
3839#endif
3840
3841	/* recheck policy now with rq lock held */
3842	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3843		policy = oldpolicy = -1;
3844		task_rq_unlock(rq, p, &flags);
3845		goto recheck;
3846	}
3847	on_rq = p->on_rq;
3848	running = task_current(rq, p);
3849	if (on_rq)
3850		dequeue_task(rq, p, 0);
3851	if (running)
3852		p->sched_class->put_prev_task(rq, p);
3853
3854	p->sched_reset_on_fork = reset_on_fork;
3855
3856	oldprio = p->prio;
3857	prev_class = p->sched_class;
3858	__setscheduler(rq, p, policy, param->sched_priority);
3859
3860	if (running)
3861		p->sched_class->set_curr_task(rq);
3862	if (on_rq)
3863		enqueue_task(rq, p, 0);
3864
3865	check_class_changed(rq, p, prev_class, oldprio);
3866	task_rq_unlock(rq, p, &flags);
3867
3868	rt_mutex_adjust_pi(p);
3869
3870	return 0;
3871}
3872
3873/**
3874 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3875 * @p: the task in question.
3876 * @policy: new policy.
3877 * @param: structure containing the new RT priority.
3878 *
3879 * NOTE that the task may be already dead.
3880 */
3881int sched_setscheduler(struct task_struct *p, int policy,
3882		       const struct sched_param *param)
3883{
3884	return __sched_setscheduler(p, policy, param, true);
3885}
3886EXPORT_SYMBOL_GPL(sched_setscheduler);
3887
3888/**
3889 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3890 * @p: the task in question.
3891 * @policy: new policy.
3892 * @param: structure containing the new RT priority.
3893 *
3894 * Just like sched_setscheduler, only don't bother checking if the
3895 * current context has permission.  For example, this is needed in
3896 * stop_machine(): we create temporary high priority worker threads,
3897 * but our caller might not have that capability.
3898 */
3899int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3900			       const struct sched_param *param)
3901{
3902	return __sched_setscheduler(p, policy, param, false);
3903}
3904
3905static int
3906do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3907{
3908	struct sched_param lparam;
3909	struct task_struct *p;
3910	int retval;
3911
3912	if (!param || pid < 0)
3913		return -EINVAL;
3914	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3915		return -EFAULT;
3916
3917	rcu_read_lock();
3918	retval = -ESRCH;
3919	p = find_process_by_pid(pid);
3920	if (p != NULL)
3921		retval = sched_setscheduler(p, policy, &lparam);
3922	rcu_read_unlock();
3923
3924	return retval;
3925}
3926
3927/**
3928 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3929 * @pid: the pid in question.
3930 * @policy: new policy.
3931 * @param: structure containing the new RT priority.
3932 */
3933SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3934		struct sched_param __user *, param)
3935{
3936	/* negative values for policy are not valid */
3937	if (policy < 0)
3938		return -EINVAL;
3939
3940	return do_sched_setscheduler(pid, policy, param);
3941}
3942
3943/**
3944 * sys_sched_setparam - set/change the RT priority of a thread
3945 * @pid: the pid in question.
3946 * @param: structure containing the new RT priority.
3947 */
3948SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3949{
3950	return do_sched_setscheduler(pid, -1, param);
3951}
3952
3953/**
3954 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3955 * @pid: the pid in question.
3956 */
3957SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3958{
3959	struct task_struct *p;
3960	int retval;
3961
3962	if (pid < 0)
3963		return -EINVAL;
3964
3965	retval = -ESRCH;
3966	rcu_read_lock();
3967	p = find_process_by_pid(pid);
3968	if (p) {
3969		retval = security_task_getscheduler(p);
3970		if (!retval)
3971			retval = p->policy
3972				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3973	}
3974	rcu_read_unlock();
3975	return retval;
3976}
3977
3978/**
3979 * sys_sched_getparam - get the RT priority of a thread
3980 * @pid: the pid in question.
3981 * @param: structure containing the RT priority.
3982 */
3983SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3984{
3985	struct sched_param lp;
3986	struct task_struct *p;
3987	int retval;
3988
3989	if (!param || pid < 0)
3990		return -EINVAL;
3991
3992	rcu_read_lock();
3993	p = find_process_by_pid(pid);
3994	retval = -ESRCH;
3995	if (!p)
3996		goto out_unlock;
3997
3998	retval = security_task_getscheduler(p);
3999	if (retval)
4000		goto out_unlock;
4001
4002	lp.sched_priority = p->rt_priority;
4003	rcu_read_unlock();
4004
4005	/*
4006	 * This one might sleep, we cannot do it with a spinlock held ...
4007	 */
4008	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4009
4010	return retval;
4011
4012out_unlock:
4013	rcu_read_unlock();
4014	return retval;
4015}
4016
4017long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4018{
4019	cpumask_var_t cpus_allowed, new_mask;
4020	struct task_struct *p;
4021	int retval;
4022
4023	get_online_cpus();
4024	rcu_read_lock();
4025
4026	p = find_process_by_pid(pid);
4027	if (!p) {
4028		rcu_read_unlock();
4029		put_online_cpus();
4030		return -ESRCH;
4031	}
4032
4033	/* Prevent p going away */
4034	get_task_struct(p);
4035	rcu_read_unlock();
4036
4037	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4038		retval = -ENOMEM;
4039		goto out_put_task;
4040	}
4041	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4042		retval = -ENOMEM;
4043		goto out_free_cpus_allowed;
4044	}
4045	retval = -EPERM;
4046	if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4047		goto out_unlock;
4048
4049	retval = security_task_setscheduler(p);
4050	if (retval)
4051		goto out_unlock;
4052
4053	cpuset_cpus_allowed(p, cpus_allowed);
4054	cpumask_and(new_mask, in_mask, cpus_allowed);
4055again:
4056	retval = set_cpus_allowed_ptr(p, new_mask);
4057
4058	if (!retval) {
4059		cpuset_cpus_allowed(p, cpus_allowed);
4060		if (!cpumask_subset(new_mask, cpus_allowed)) {
4061			/*
4062			 * We must have raced with a concurrent cpuset
4063			 * update. Just reset the cpus_allowed to the
4064			 * cpuset's cpus_allowed
4065			 */
4066			cpumask_copy(new_mask, cpus_allowed);
4067			goto again;
4068		}
4069	}
4070out_unlock:
4071	free_cpumask_var(new_mask);
4072out_free_cpus_allowed:
4073	free_cpumask_var(cpus_allowed);
4074out_put_task:
4075	put_task_struct(p);
4076	put_online_cpus();
4077	return retval;
4078}
4079
4080static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4081			     struct cpumask *new_mask)
4082{
4083	if (len < cpumask_size())
4084		cpumask_clear(new_mask);
4085	else if (len > cpumask_size())
4086		len = cpumask_size();
4087
4088	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4089}
4090
4091/**
4092 * sys_sched_setaffinity - set the cpu affinity of a process
4093 * @pid: pid of the process
4094 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4095 * @user_mask_ptr: user-space pointer to the new cpu mask
4096 */
4097SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4098		unsigned long __user *, user_mask_ptr)
4099{
4100	cpumask_var_t new_mask;
4101	int retval;
4102
4103	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4104		return -ENOMEM;
4105
4106	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4107	if (retval == 0)
4108		retval = sched_setaffinity(pid, new_mask);
4109	free_cpumask_var(new_mask);
4110	return retval;
4111}
4112
4113long sched_getaffinity(pid_t pid, struct cpumask *mask)
4114{
4115	struct task_struct *p;
4116	unsigned long flags;
4117	int retval;
4118
4119	get_online_cpus();
4120	rcu_read_lock();
4121
4122	retval = -ESRCH;
4123	p = find_process_by_pid(pid);
4124	if (!p)
4125		goto out_unlock;
4126
4127	retval = security_task_getscheduler(p);
4128	if (retval)
4129		goto out_unlock;
4130
4131	raw_spin_lock_irqsave(&p->pi_lock, flags);
4132	cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4133	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4134
4135out_unlock:
4136	rcu_read_unlock();
4137	put_online_cpus();
4138
4139	return retval;
4140}
4141
4142/**
4143 * sys_sched_getaffinity - get the cpu affinity of a process
4144 * @pid: pid of the process
4145 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4146 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4147 */
4148SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4149		unsigned long __user *, user_mask_ptr)
4150{
4151	int ret;
4152	cpumask_var_t mask;
4153
4154	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4155		return -EINVAL;
4156	if (len & (sizeof(unsigned long)-1))
4157		return -EINVAL;
4158
4159	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4160		return -ENOMEM;
4161
4162	ret = sched_getaffinity(pid, mask);
4163	if (ret == 0) {
4164		size_t retlen = min_t(size_t, len, cpumask_size());
4165
4166		if (copy_to_user(user_mask_ptr, mask, retlen))
4167			ret = -EFAULT;
4168		else
4169			ret = retlen;
4170	}
4171	free_cpumask_var(mask);
4172
4173	return ret;
4174}
4175
4176/**
4177 * sys_sched_yield - yield the current processor to other threads.
4178 *
4179 * This function yields the current CPU to other tasks. If there are no
4180 * other threads running on this CPU then this function will return.
4181 */
4182SYSCALL_DEFINE0(sched_yield)
4183{
4184	struct rq *rq = this_rq_lock();
4185
4186	schedstat_inc(rq, yld_count);
4187	current->sched_class->yield_task(rq);
4188
4189	/*
4190	 * Since we are going to call schedule() anyway, there's
4191	 * no need to preempt or enable interrupts:
4192	 */
4193	__release(rq->lock);
4194	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4195	do_raw_spin_unlock(&rq->lock);
4196	sched_preempt_enable_no_resched();
4197
4198	schedule();
4199
4200	return 0;
4201}
4202
4203static inline int should_resched(void)
4204{
4205	return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4206}
4207
4208static void __cond_resched(void)
4209{
4210	add_preempt_count(PREEMPT_ACTIVE);
4211	__schedule();
4212	sub_preempt_count(PREEMPT_ACTIVE);
4213}
4214
4215int __sched _cond_resched(void)
4216{
4217	if (should_resched()) {
4218		__cond_resched();
4219		return 1;
4220	}
4221	return 0;
4222}
4223EXPORT_SYMBOL(_cond_resched);
4224
4225/*
4226 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4227 * call schedule, and on return reacquire the lock.
4228 *
4229 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4230 * operations here to prevent schedule() from being called twice (once via
4231 * spin_unlock(), once by hand).
4232 */
4233int __cond_resched_lock(spinlock_t *lock)
4234{
4235	int resched = should_resched();
4236	int ret = 0;
4237
4238	lockdep_assert_held(lock);
4239
4240	if (spin_needbreak(lock) || resched) {
4241		spin_unlock(lock);
4242		if (resched)
4243			__cond_resched();
4244		else
4245			cpu_relax();
4246		ret = 1;
4247		spin_lock(lock);
4248	}
4249	return ret;
4250}
4251EXPORT_SYMBOL(__cond_resched_lock);
4252
4253int __sched __cond_resched_softirq(void)
4254{
4255	BUG_ON(!in_softirq());
4256
4257	if (should_resched()) {
4258		local_bh_enable();
4259		__cond_resched();
4260		local_bh_disable();
4261		return 1;
4262	}
4263	return 0;
4264}
4265EXPORT_SYMBOL(__cond_resched_softirq);
4266
4267/**
4268 * yield - yield the current processor to other threads.
4269 *
4270 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4271 *
4272 * The scheduler is at all times free to pick the calling task as the most
4273 * eligible task to run, if removing the yield() call from your code breaks
4274 * it, its already broken.
4275 *
4276 * Typical broken usage is:
4277 *
4278 * while (!event)
4279 * 	yield();
4280 *
4281 * where one assumes that yield() will let 'the other' process run that will
4282 * make event true. If the current task is a SCHED_FIFO task that will never
4283 * happen. Never use yield() as a progress guarantee!!
4284 *
4285 * If you want to use yield() to wait for something, use wait_event().
4286 * If you want to use yield() to be 'nice' for others, use cond_resched().
4287 * If you still want to use yield(), do not!
4288 */
4289void __sched yield(void)
4290{
4291	set_current_state(TASK_RUNNING);
4292	sys_sched_yield();
4293}
4294EXPORT_SYMBOL(yield);
4295
4296/**
4297 * yield_to - yield the current processor to another thread in
4298 * your thread group, or accelerate that thread toward the
4299 * processor it's on.
4300 * @p: target task
4301 * @preempt: whether task preemption is allowed or not
4302 *
4303 * It's the caller's job to ensure that the target task struct
4304 * can't go away on us before we can do any checks.
4305 *
4306 * Returns true if we indeed boosted the target task.
4307 */
4308bool __sched yield_to(struct task_struct *p, bool preempt)
4309{
4310	struct task_struct *curr = current;
4311	struct rq *rq, *p_rq;
4312	unsigned long flags;
4313	bool yielded = 0;
4314
4315	local_irq_save(flags);
4316	rq = this_rq();
4317
4318again:
4319	p_rq = task_rq(p);
4320	double_rq_lock(rq, p_rq);
4321	while (task_rq(p) != p_rq) {
4322		double_rq_unlock(rq, p_rq);
4323		goto again;
4324	}
4325
4326	if (!curr->sched_class->yield_to_task)
4327		goto out;
4328
4329	if (curr->sched_class != p->sched_class)
4330		goto out;
4331
4332	if (task_running(p_rq, p) || p->state)
4333		goto out;
4334
4335	yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4336	if (yielded) {
4337		schedstat_inc(rq, yld_count);
4338		/*
4339		 * Make p's CPU reschedule; pick_next_entity takes care of
4340		 * fairness.
4341		 */
4342		if (preempt && rq != p_rq)
4343			resched_task(p_rq->curr);
4344	}
4345
4346out:
4347	double_rq_unlock(rq, p_rq);
4348	local_irq_restore(flags);
4349
4350	if (yielded)
4351		schedule();
4352
4353	return yielded;
4354}
4355EXPORT_SYMBOL_GPL(yield_to);
4356
4357/*
4358 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4359 * that process accounting knows that this is a task in IO wait state.
4360 */
4361void __sched io_schedule(void)
4362{
4363	struct rq *rq = raw_rq();
4364
4365	delayacct_blkio_start();
4366	atomic_inc(&rq->nr_iowait);
4367	blk_flush_plug(current);
4368	current->in_iowait = 1;
4369	schedule();
4370	current->in_iowait = 0;
4371	atomic_dec(&rq->nr_iowait);
4372	delayacct_blkio_end();
4373}
4374EXPORT_SYMBOL(io_schedule);
4375
4376long __sched io_schedule_timeout(long timeout)
4377{
4378	struct rq *rq = raw_rq();
4379	long ret;
4380
4381	delayacct_blkio_start();
4382	atomic_inc(&rq->nr_iowait);
4383	blk_flush_plug(current);
4384	current->in_iowait = 1;
4385	ret = schedule_timeout(timeout);
4386	current->in_iowait = 0;
4387	atomic_dec(&rq->nr_iowait);
4388	delayacct_blkio_end();
4389	return ret;
4390}
4391
4392/**
4393 * sys_sched_get_priority_max - return maximum RT priority.
4394 * @policy: scheduling class.
4395 *
4396 * this syscall returns the maximum rt_priority that can be used
4397 * by a given scheduling class.
4398 */
4399SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4400{
4401	int ret = -EINVAL;
4402
4403	switch (policy) {
4404	case SCHED_FIFO:
4405	case SCHED_RR:
4406		ret = MAX_USER_RT_PRIO-1;
4407		break;
4408	case SCHED_NORMAL:
4409	case SCHED_BATCH:
4410	case SCHED_IDLE:
4411		ret = 0;
4412		break;
4413	}
4414	return ret;
4415}
4416
4417/**
4418 * sys_sched_get_priority_min - return minimum RT priority.
4419 * @policy: scheduling class.
4420 *
4421 * this syscall returns the minimum rt_priority that can be used
4422 * by a given scheduling class.
4423 */
4424SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4425{
4426	int ret = -EINVAL;
4427
4428	switch (policy) {
4429	case SCHED_FIFO:
4430	case SCHED_RR:
4431		ret = 1;
4432		break;
4433	case SCHED_NORMAL:
4434	case SCHED_BATCH:
4435	case SCHED_IDLE:
4436		ret = 0;
4437	}
4438	return ret;
4439}
4440
4441/**
4442 * sys_sched_rr_get_interval - return the default timeslice of a process.
4443 * @pid: pid of the process.
4444 * @interval: userspace pointer to the timeslice value.
4445 *
4446 * this syscall writes the default timeslice value of a given process
4447 * into the user-space timespec buffer. A value of '0' means infinity.
4448 */
4449SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4450		struct timespec __user *, interval)
4451{
4452	struct task_struct *p;
4453	unsigned int time_slice;
4454	unsigned long flags;
4455	struct rq *rq;
4456	int retval;
4457	struct timespec t;
4458
4459	if (pid < 0)
4460		return -EINVAL;
4461
4462	retval = -ESRCH;
4463	rcu_read_lock();
4464	p = find_process_by_pid(pid);
4465	if (!p)
4466		goto out_unlock;
4467
4468	retval = security_task_getscheduler(p);
4469	if (retval)
4470		goto out_unlock;
4471
4472	rq = task_rq_lock(p, &flags);
4473	time_slice = p->sched_class->get_rr_interval(rq, p);
4474	task_rq_unlock(rq, p, &flags);
4475
4476	rcu_read_unlock();
4477	jiffies_to_timespec(time_slice, &t);
4478	retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4479	return retval;
4480
4481out_unlock:
4482	rcu_read_unlock();
4483	return retval;
4484}
4485
4486static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4487
4488void sched_show_task(struct task_struct *p)
4489{
4490	unsigned long free = 0;
4491	unsigned state;
4492
4493	state = p->state ? __ffs(p->state) + 1 : 0;
4494	printk(KERN_INFO "%-15.15s %c", p->comm,
4495		state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4496#if BITS_PER_LONG == 32
4497	if (state == TASK_RUNNING)
4498		printk(KERN_CONT " running  ");
4499	else
4500		printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4501#else
4502	if (state == TASK_RUNNING)
4503		printk(KERN_CONT "  running task    ");
4504	else
4505		printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4506#endif
4507#ifdef CONFIG_DEBUG_STACK_USAGE
4508	free = stack_not_used(p);
4509#endif
4510	printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4511		task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4512		(unsigned long)task_thread_info(p)->flags);
4513
4514	show_stack(p, NULL);
4515}
4516
4517void show_state_filter(unsigned long state_filter)
4518{
4519	struct task_struct *g, *p;
4520
4521#if BITS_PER_LONG == 32
4522	printk(KERN_INFO
4523		"  task                PC stack   pid father\n");
4524#else
4525	printk(KERN_INFO
4526		"  task                        PC stack   pid father\n");
4527#endif
4528	rcu_read_lock();
4529	do_each_thread(g, p) {
4530		/*
4531		 * reset the NMI-timeout, listing all files on a slow
4532		 * console might take a lot of time:
4533		 */
4534		touch_nmi_watchdog();
4535		if (!state_filter || (p->state & state_filter))
4536			sched_show_task(p);
4537	} while_each_thread(g, p);
4538
4539	touch_all_softlockup_watchdogs();
4540
4541#ifdef CONFIG_SCHED_DEBUG
4542	sysrq_sched_debug_show();
4543#endif
4544	rcu_read_unlock();
4545	/*
4546	 * Only show locks if all tasks are dumped:
4547	 */
4548	if (!state_filter)
4549		debug_show_all_locks();
4550}
4551
4552void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4553{
4554	idle->sched_class = &idle_sched_class;
4555}
4556
4557/**
4558 * init_idle - set up an idle thread for a given CPU
4559 * @idle: task in question
4560 * @cpu: cpu the idle task belongs to
4561 *
4562 * NOTE: this function does not set the idle thread's NEED_RESCHED
4563 * flag, to make booting more robust.
4564 */
4565void __cpuinit init_idle(struct task_struct *idle, int cpu)
4566{
4567	struct rq *rq = cpu_rq(cpu);
4568	unsigned long flags;
4569
4570	raw_spin_lock_irqsave(&rq->lock, flags);
4571
4572	__sched_fork(idle);
4573	idle->state = TASK_RUNNING;
4574	idle->se.exec_start = sched_clock();
4575
4576	do_set_cpus_allowed(idle, cpumask_of(cpu));
4577	/*
4578	 * We're having a chicken and egg problem, even though we are
4579	 * holding rq->lock, the cpu isn't yet set to this cpu so the
4580	 * lockdep check in task_group() will fail.
4581	 *
4582	 * Similar case to sched_fork(). / Alternatively we could
4583	 * use task_rq_lock() here and obtain the other rq->lock.
4584	 *
4585	 * Silence PROVE_RCU
4586	 */
4587	rcu_read_lock();
4588	__set_task_cpu(idle, cpu);
4589	rcu_read_unlock();
4590
4591	rq->curr = rq->idle = idle;
4592#if defined(CONFIG_SMP)
4593	idle->on_cpu = 1;
4594#endif
4595	raw_spin_unlock_irqrestore(&rq->lock, flags);
4596
4597	/* Set the preempt count _outside_ the spinlocks! */
4598	task_thread_info(idle)->preempt_count = 0;
4599
4600	/*
4601	 * The idle tasks have their own, simple scheduling class:
4602	 */
4603	idle->sched_class = &idle_sched_class;
4604	ftrace_graph_init_idle_task(idle, cpu);
4605#if defined(CONFIG_SMP)
4606	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4607#endif
4608}
4609
4610#ifdef CONFIG_SMP
4611void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4612{
4613	if (p->sched_class && p->sched_class->set_cpus_allowed)
4614		p->sched_class->set_cpus_allowed(p, new_mask);
4615
4616	cpumask_copy(&p->cpus_allowed, new_mask);
4617	p->nr_cpus_allowed = cpumask_weight(new_mask);
4618}
4619
4620/*
4621 * This is how migration works:
4622 *
4623 * 1) we invoke migration_cpu_stop() on the target CPU using
4624 *    stop_one_cpu().
4625 * 2) stopper starts to run (implicitly forcing the migrated thread
4626 *    off the CPU)
4627 * 3) it checks whether the migrated task is still in the wrong runqueue.
4628 * 4) if it's in the wrong runqueue then the migration thread removes
4629 *    it and puts it into the right queue.
4630 * 5) stopper completes and stop_one_cpu() returns and the migration
4631 *    is done.
4632 */
4633
4634/*
4635 * Change a given task's CPU affinity. Migrate the thread to a
4636 * proper CPU and schedule it away if the CPU it's executing on
4637 * is removed from the allowed bitmask.
4638 *
4639 * NOTE: the caller must have a valid reference to the task, the
4640 * task must not exit() & deallocate itself prematurely. The
4641 * call is not atomic; no spinlocks may be held.
4642 */
4643int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4644{
4645	unsigned long flags;
4646	struct rq *rq;
4647	unsigned int dest_cpu;
4648	int ret = 0;
4649
4650	rq = task_rq_lock(p, &flags);
4651
4652	if (cpumask_equal(&p->cpus_allowed, new_mask))
4653		goto out;
4654
4655	if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4656		ret = -EINVAL;
4657		goto out;
4658	}
4659
4660	if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4661		ret = -EINVAL;
4662		goto out;
4663	}
4664
4665	do_set_cpus_allowed(p, new_mask);
4666
4667	/* Can the task run on the task's current CPU? If so, we're done */
4668	if (cpumask_test_cpu(task_cpu(p), new_mask))
4669		goto out;
4670
4671	dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4672	if (p->on_rq) {
4673		struct migration_arg arg = { p, dest_cpu };
4674		/* Need help from migration thread: drop lock and wait. */
4675		task_rq_unlock(rq, p, &flags);
4676		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4677		tlb_migrate_finish(p->mm);
4678		return 0;
4679	}
4680out:
4681	task_rq_unlock(rq, p, &flags);
4682
4683	return ret;
4684}
4685EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4686
4687/*
4688 * Move (not current) task off this cpu, onto dest cpu. We're doing
4689 * this because either it can't run here any more (set_cpus_allowed()
4690 * away from this CPU, or CPU going down), or because we're
4691 * attempting to rebalance this task on exec (sched_exec).
4692 *
4693 * So we race with normal scheduler movements, but that's OK, as long
4694 * as the task is no longer on this CPU.
4695 *
4696 * Returns non-zero if task was successfully migrated.
4697 */
4698static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4699{
4700	struct rq *rq_dest, *rq_src;
4701	int ret = 0;
4702
4703	if (unlikely(!cpu_active(dest_cpu)))
4704		return ret;
4705
4706	rq_src = cpu_rq(src_cpu);
4707	rq_dest = cpu_rq(dest_cpu);
4708
4709	raw_spin_lock(&p->pi_lock);
4710	double_rq_lock(rq_src, rq_dest);
4711	/* Already moved. */
4712	if (task_cpu(p) != src_cpu)
4713		goto done;
4714	/* Affinity changed (again). */
4715	if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4716		goto fail;
4717
4718	/*
4719	 * If we're not on a rq, the next wake-up will ensure we're
4720	 * placed properly.
4721	 */
4722	if (p->on_rq) {
4723		dequeue_task(rq_src, p, 0);
4724		set_task_cpu(p, dest_cpu);
4725		enqueue_task(rq_dest, p, 0);
4726		check_preempt_curr(rq_dest, p, 0);
4727	}
4728done:
4729	ret = 1;
4730fail:
4731	double_rq_unlock(rq_src, rq_dest);
4732	raw_spin_unlock(&p->pi_lock);
4733	return ret;
4734}
4735
4736/*
4737 * migration_cpu_stop - this will be executed by a highprio stopper thread
4738 * and performs thread migration by bumping thread off CPU then
4739 * 'pushing' onto another runqueue.
4740 */
4741static int migration_cpu_stop(void *data)
4742{
4743	struct migration_arg *arg = data;
4744
4745	/*
4746	 * The original target cpu might have gone down and we might
4747	 * be on another cpu but it doesn't matter.
4748	 */
4749	local_irq_disable();
4750	__migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4751	local_irq_enable();
4752	return 0;
4753}
4754
4755#ifdef CONFIG_HOTPLUG_CPU
4756
4757/*
4758 * Ensures that the idle task is using init_mm right before its cpu goes
4759 * offline.
4760 */
4761void idle_task_exit(void)
4762{
4763	struct mm_struct *mm = current->active_mm;
4764
4765	BUG_ON(cpu_online(smp_processor_id()));
4766
4767	if (mm != &init_mm)
4768		switch_mm(mm, &init_mm, current);
4769	mmdrop(mm);
4770}
4771
4772/*
4773 * Since this CPU is going 'away' for a while, fold any nr_active delta
4774 * we might have. Assumes we're called after migrate_tasks() so that the
4775 * nr_active count is stable.
4776 *
4777 * Also see the comment "Global load-average calculations".
4778 */
4779static void calc_load_migrate(struct rq *rq)
4780{
4781	long delta = calc_load_fold_active(rq);
4782	if (delta)
4783		atomic_long_add(delta, &calc_load_tasks);
4784}
4785
4786/*
4787 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4788 * try_to_wake_up()->select_task_rq().
4789 *
4790 * Called with rq->lock held even though we'er in stop_machine() and
4791 * there's no concurrency possible, we hold the required locks anyway
4792 * because of lock validation efforts.
4793 */
4794static void migrate_tasks(unsigned int dead_cpu)
4795{
4796	struct rq *rq = cpu_rq(dead_cpu);
4797	struct task_struct *next, *stop = rq->stop;
4798	int dest_cpu;
4799
4800	/*
4801	 * Fudge the rq selection such that the below task selection loop
4802	 * doesn't get stuck on the currently eligible stop task.
4803	 *
4804	 * We're currently inside stop_machine() and the rq is either stuck
4805	 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4806	 * either way we should never end up calling schedule() until we're
4807	 * done here.
4808	 */
4809	rq->stop = NULL;
4810
4811	for ( ; ; ) {
4812		/*
4813		 * There's this thread running, bail when that's the only
4814		 * remaining thread.
4815		 */
4816		if (rq->nr_running == 1)
4817			break;
4818
4819		next = pick_next_task(rq);
4820		BUG_ON(!next);
4821		next->sched_class->put_prev_task(rq, next);
4822
4823		/* Find suitable destination for @next, with force if needed. */
4824		dest_cpu = select_fallback_rq(dead_cpu, next);
4825		raw_spin_unlock(&rq->lock);
4826
4827		__migrate_task(next, dead_cpu, dest_cpu);
4828
4829		raw_spin_lock(&rq->lock);
4830	}
4831
4832	rq->stop = stop;
4833}
4834
4835#endif /* CONFIG_HOTPLUG_CPU */
4836
4837#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4838
4839static struct ctl_table sd_ctl_dir[] = {
4840	{
4841		.procname	= "sched_domain",
4842		.mode		= 0555,
4843	},
4844	{}
4845};
4846
4847static struct ctl_table sd_ctl_root[] = {
4848	{
4849		.procname	= "kernel",
4850		.mode		= 0555,
4851		.child		= sd_ctl_dir,
4852	},
4853	{}
4854};
4855
4856static struct ctl_table *sd_alloc_ctl_entry(int n)
4857{
4858	struct ctl_table *entry =
4859		kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4860
4861	return entry;
4862}
4863
4864static void sd_free_ctl_entry(struct ctl_table **tablep)
4865{
4866	struct ctl_table *entry;
4867
4868	/*
4869	 * In the intermediate directories, both the child directory and
4870	 * procname are dynamically allocated and could fail but the mode
4871	 * will always be set. In the lowest directory the names are
4872	 * static strings and all have proc handlers.
4873	 */
4874	for (entry = *tablep; entry->mode; entry++) {
4875		if (entry->child)
4876			sd_free_ctl_entry(&entry->child);
4877		if (entry->proc_handler == NULL)
4878			kfree(entry->procname);
4879	}
4880
4881	kfree(*tablep);
4882	*tablep = NULL;
4883}
4884
4885static int min_load_idx = 0;
4886static int max_load_idx = CPU_LOAD_IDX_MAX;
4887
4888static void
4889set_table_entry(struct ctl_table *entry,
4890		const char *procname, void *data, int maxlen,
4891		umode_t mode, proc_handler *proc_handler,
4892		bool load_idx)
4893{
4894	entry->procname = procname;
4895	entry->data = data;
4896	entry->maxlen = maxlen;
4897	entry->mode = mode;
4898	entry->proc_handler = proc_handler;
4899
4900	if (load_idx) {
4901		entry->extra1 = &min_load_idx;
4902		entry->extra2 = &max_load_idx;
4903	}
4904}
4905
4906static struct ctl_table *
4907sd_alloc_ctl_domain_table(struct sched_domain *sd)
4908{
4909	struct ctl_table *table = sd_alloc_ctl_entry(13);
4910
4911	if (table == NULL)
4912		return NULL;
4913
4914	set_table_entry(&table[0], "min_interval", &sd->min_interval,
4915		sizeof(long), 0644, proc_doulongvec_minmax, false);
4916	set_table_entry(&table[1], "max_interval", &sd->max_interval,
4917		sizeof(long), 0644, proc_doulongvec_minmax, false);
4918	set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4919		sizeof(int), 0644, proc_dointvec_minmax, true);
4920	set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4921		sizeof(int), 0644, proc_dointvec_minmax, true);
4922	set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4923		sizeof(int), 0644, proc_dointvec_minmax, true);
4924	set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4925		sizeof(int), 0644, proc_dointvec_minmax, true);
4926	set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4927		sizeof(int), 0644, proc_dointvec_minmax, true);
4928	set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4929		sizeof(int), 0644, proc_dointvec_minmax, false);
4930	set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4931		sizeof(int), 0644, proc_dointvec_minmax, false);
4932	set_table_entry(&table[9], "cache_nice_tries",
4933		&sd->cache_nice_tries,
4934		sizeof(int), 0644, proc_dointvec_minmax, false);
4935	set_table_entry(&table[10], "flags", &sd->flags,
4936		sizeof(int), 0644, proc_dointvec_minmax, false);
4937	set_table_entry(&table[11], "name", sd->name,
4938		CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4939	/* &table[12] is terminator */
4940
4941	return table;
4942}
4943
4944static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4945{
4946	struct ctl_table *entry, *table;
4947	struct sched_domain *sd;
4948	int domain_num = 0, i;
4949	char buf[32];
4950
4951	for_each_domain(cpu, sd)
4952		domain_num++;
4953	entry = table = sd_alloc_ctl_entry(domain_num + 1);
4954	if (table == NULL)
4955		return NULL;
4956
4957	i = 0;
4958	for_each_domain(cpu, sd) {
4959		snprintf(buf, 32, "domain%d", i);
4960		entry->procname = kstrdup(buf, GFP_KERNEL);
4961		entry->mode = 0555;
4962		entry->child = sd_alloc_ctl_domain_table(sd);
4963		entry++;
4964		i++;
4965	}
4966	return table;
4967}
4968
4969static struct ctl_table_header *sd_sysctl_header;
4970static void register_sched_domain_sysctl(void)
4971{
4972	int i, cpu_num = num_possible_cpus();
4973	struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4974	char buf[32];
4975
4976	WARN_ON(sd_ctl_dir[0].child);
4977	sd_ctl_dir[0].child = entry;
4978
4979	if (entry == NULL)
4980		return;
4981
4982	for_each_possible_cpu(i) {
4983		snprintf(buf, 32, "cpu%d", i);
4984		entry->procname = kstrdup(buf, GFP_KERNEL);
4985		entry->mode = 0555;
4986		entry->child = sd_alloc_ctl_cpu_table(i);
4987		entry++;
4988	}
4989
4990	WARN_ON(sd_sysctl_header);
4991	sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4992}
4993
4994/* may be called multiple times per register */
4995static void unregister_sched_domain_sysctl(void)
4996{
4997	if (sd_sysctl_header)
4998		unregister_sysctl_table(sd_sysctl_header);
4999	sd_sysctl_header = NULL;
5000	if (sd_ctl_dir[0].child)
5001		sd_free_ctl_entry(&sd_ctl_dir[0].child);
5002}
5003#else
5004static void register_sched_domain_sysctl(void)
5005{
5006}
5007static void unregister_sched_domain_sysctl(void)
5008{
5009}
5010#endif
5011
5012static void set_rq_online(struct rq *rq)
5013{
5014	if (!rq->online) {
5015		const struct sched_class *class;
5016
5017		cpumask_set_cpu(rq->cpu, rq->rd->online);
5018		rq->online = 1;
5019
5020		for_each_class(class) {
5021			if (class->rq_online)
5022				class->rq_online(rq);
5023		}
5024	}
5025}
5026
5027static void set_rq_offline(struct rq *rq)
5028{
5029	if (rq->online) {
5030		const struct sched_class *class;
5031
5032		for_each_class(class) {
5033			if (class->rq_offline)
5034				class->rq_offline(rq);
5035		}
5036
5037		cpumask_clear_cpu(rq->cpu, rq->rd->online);
5038		rq->online = 0;
5039	}
5040}
5041
5042/*
5043 * migration_call - callback that gets triggered when a CPU is added.
5044 * Here we can start up the necessary migration thread for the new CPU.
5045 */
5046static int __cpuinit
5047migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5048{
5049	int cpu = (long)hcpu;
5050	unsigned long flags;
5051	struct rq *rq = cpu_rq(cpu);
5052
5053	switch (action & ~CPU_TASKS_FROZEN) {
5054
5055	case CPU_UP_PREPARE:
5056		rq->calc_load_update = calc_load_update;
5057		break;
5058
5059	case CPU_ONLINE:
5060		/* Update our root-domain */
5061		raw_spin_lock_irqsave(&rq->lock, flags);
5062		if (rq->rd) {
5063			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5064
5065			set_rq_online(rq);
5066		}
5067		raw_spin_unlock_irqrestore(&rq->lock, flags);
5068		break;
5069
5070#ifdef CONFIG_HOTPLUG_CPU
5071	case CPU_DYING:
5072		sched_ttwu_pending();
5073		/* Update our root-domain */
5074		raw_spin_lock_irqsave(&rq->lock, flags);
5075		if (rq->rd) {
5076			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5077			set_rq_offline(rq);
5078		}
5079		migrate_tasks(cpu);
5080		BUG_ON(rq->nr_running != 1); /* the migration thread */
5081		raw_spin_unlock_irqrestore(&rq->lock, flags);
5082		break;
5083
5084	case CPU_DEAD:
5085		calc_load_migrate(rq);
5086		break;
5087#endif
5088	}
5089
5090	update_max_interval();
5091
5092	return NOTIFY_OK;
5093}
5094
5095/*
5096 * Register at high priority so that task migration (migrate_all_tasks)
5097 * happens before everything else.  This has to be lower priority than
5098 * the notifier in the perf_event subsystem, though.
5099 */
5100static struct notifier_block __cpuinitdata migration_notifier = {
5101	.notifier_call = migration_call,
5102	.priority = CPU_PRI_MIGRATION,
5103};
5104
5105static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5106				      unsigned long action, void *hcpu)
5107{
5108	switch (action & ~CPU_TASKS_FROZEN) {
5109	case CPU_STARTING:
5110	case CPU_DOWN_FAILED:
5111		set_cpu_active((long)hcpu, true);
5112		return NOTIFY_OK;
5113	default:
5114		return NOTIFY_DONE;
5115	}
5116}
5117
5118static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5119					unsigned long action, void *hcpu)
5120{
5121	switch (action & ~CPU_TASKS_FROZEN) {
5122	case CPU_DOWN_PREPARE:
5123		set_cpu_active((long)hcpu, false);
5124		return NOTIFY_OK;
5125	default:
5126		return NOTIFY_DONE;
5127	}
5128}
5129
5130static int __init migration_init(void)
5131{
5132	void *cpu = (void *)(long)smp_processor_id();
5133	int err;
5134
5135	/* Initialize migration for the boot CPU */
5136	err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5137	BUG_ON(err == NOTIFY_BAD);
5138	migration_call(&migration_notifier, CPU_ONLINE, cpu);
5139	register_cpu_notifier(&migration_notifier);
5140
5141	/* Register cpu active notifiers */
5142	cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5143	cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5144
5145	return 0;
5146}
5147early_initcall(migration_init);
5148#endif
5149
5150#ifdef CONFIG_SMP
5151
5152static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5153
5154#ifdef CONFIG_SCHED_DEBUG
5155
5156static __read_mostly int sched_debug_enabled;
5157
5158static int __init sched_debug_setup(char *str)
5159{
5160	sched_debug_enabled = 1;
5161
5162	return 0;
5163}
5164early_param("sched_debug", sched_debug_setup);
5165
5166static inline bool sched_debug(void)
5167{
5168	return sched_debug_enabled;
5169}
5170
5171static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5172				  struct cpumask *groupmask)
5173{
5174	struct sched_group *group = sd->groups;
5175	char str[256];
5176
5177	cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5178	cpumask_clear(groupmask);
5179
5180	printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5181
5182	if (!(sd->flags & SD_LOAD_BALANCE)) {
5183		printk("does not load-balance\n");
5184		if (sd->parent)
5185			printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5186					" has parent");
5187		return -1;
5188	}
5189
5190	printk(KERN_CONT "span %s level %s\n", str, sd->name);
5191
5192	if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5193		printk(KERN_ERR "ERROR: domain->span does not contain "
5194				"CPU%d\n", cpu);
5195	}
5196	if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5197		printk(KERN_ERR "ERROR: domain->groups does not contain"
5198				" CPU%d\n", cpu);
5199	}
5200
5201	printk(KERN_DEBUG "%*s groups:", level + 1, "");
5202	do {
5203		if (!group) {
5204			printk("\n");
5205			printk(KERN_ERR "ERROR: group is NULL\n");
5206			break;
5207		}
5208
5209		/*
5210		 * Even though we initialize ->power to something semi-sane,
5211		 * we leave power_orig unset. This allows us to detect if
5212		 * domain iteration is still funny without causing /0 traps.
5213		 */
5214		if (!group->sgp->power_orig) {
5215			printk(KERN_CONT "\n");
5216			printk(KERN_ERR "ERROR: domain->cpu_power not "
5217					"set\n");
5218			break;
5219		}
5220
5221		if (!cpumask_weight(sched_group_cpus(group))) {
5222			printk(KERN_CONT "\n");
5223			printk(KERN_ERR "ERROR: empty group\n");
5224			break;
5225		}
5226
5227		if (!(sd->flags & SD_OVERLAP) &&
5228		    cpumask_intersects(groupmask, sched_group_cpus(group))) {
5229			printk(KERN_CONT "\n");
5230			printk(KERN_ERR "ERROR: repeated CPUs\n");
5231			break;
5232		}
5233
5234		cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5235
5236		cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5237
5238		printk(KERN_CONT " %s", str);
5239		if (group->sgp->power != SCHED_POWER_SCALE) {
5240			printk(KERN_CONT " (cpu_power = %d)",
5241				group->sgp->power);
5242		}
5243
5244		group = group->next;
5245	} while (group != sd->groups);
5246	printk(KERN_CONT "\n");
5247
5248	if (!cpumask_equal(sched_domain_span(sd), groupmask))
5249		printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5250
5251	if (sd->parent &&
5252	    !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5253		printk(KERN_ERR "ERROR: parent span is not a superset "
5254			"of domain->span\n");
5255	return 0;
5256}
5257
5258static void sched_domain_debug(struct sched_domain *sd, int cpu)
5259{
5260	int level = 0;
5261
5262	if (!sched_debug_enabled)
5263		return;
5264
5265	if (!sd) {
5266		printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5267		return;
5268	}
5269
5270	printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5271
5272	for (;;) {
5273		if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5274			break;
5275		level++;
5276		sd = sd->parent;
5277		if (!sd)
5278			break;
5279	}
5280}
5281#else /* !CONFIG_SCHED_DEBUG */
5282# define sched_domain_debug(sd, cpu) do { } while (0)
5283static inline bool sched_debug(void)
5284{
5285	return false;
5286}
5287#endif /* CONFIG_SCHED_DEBUG */
5288
5289static int sd_degenerate(struct sched_domain *sd)
5290{
5291	if (cpumask_weight(sched_domain_span(sd)) == 1)
5292		return 1;
5293
5294	/* Following flags need at least 2 groups */
5295	if (sd->flags & (SD_LOAD_BALANCE |
5296			 SD_BALANCE_NEWIDLE |
5297			 SD_BALANCE_FORK |
5298			 SD_BALANCE_EXEC |
5299			 SD_SHARE_CPUPOWER |
5300			 SD_SHARE_PKG_RESOURCES)) {
5301		if (sd->groups != sd->groups->next)
5302			return 0;
5303	}
5304
5305	/* Following flags don't use groups */
5306	if (sd->flags & (SD_WAKE_AFFINE))
5307		return 0;
5308
5309	return 1;
5310}
5311
5312static int
5313sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5314{
5315	unsigned long cflags = sd->flags, pflags = parent->flags;
5316
5317	if (sd_degenerate(parent))
5318		return 1;
5319
5320	if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5321		return 0;
5322
5323	/* Flags needing groups don't count if only 1 group in parent */
5324	if (parent->groups == parent->groups->next) {
5325		pflags &= ~(SD_LOAD_BALANCE |
5326				SD_BALANCE_NEWIDLE |
5327				SD_BALANCE_FORK |
5328				SD_BALANCE_EXEC |
5329				SD_SHARE_CPUPOWER |
5330				SD_SHARE_PKG_RESOURCES);
5331		if (nr_node_ids == 1)
5332			pflags &= ~SD_SERIALIZE;
5333	}
5334	if (~cflags & pflags)
5335		return 0;
5336
5337	return 1;
5338}
5339
5340static void free_rootdomain(struct rcu_head *rcu)
5341{
5342	struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5343
5344	cpupri_cleanup(&rd->cpupri);
5345	free_cpumask_var(rd->rto_mask);
5346	free_cpumask_var(rd->online);
5347	free_cpumask_var(rd->span);
5348	kfree(rd);
5349}
5350
5351static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5352{
5353	struct root_domain *old_rd = NULL;
5354	unsigned long flags;
5355
5356	raw_spin_lock_irqsave(&rq->lock, flags);
5357
5358	if (rq->rd) {
5359		old_rd = rq->rd;
5360
5361		if (cpumask_test_cpu(rq->cpu, old_rd->online))
5362			set_rq_offline(rq);
5363
5364		cpumask_clear_cpu(rq->cpu, old_rd->span);
5365
5366		/*
5367		 * If we dont want to free the old_rt yet then
5368		 * set old_rd to NULL to skip the freeing later
5369		 * in this function:
5370		 */
5371		if (!atomic_dec_and_test(&old_rd->refcount))
5372			old_rd = NULL;
5373	}
5374
5375	atomic_inc(&rd->refcount);
5376	rq->rd = rd;
5377
5378	cpumask_set_cpu(rq->cpu, rd->span);
5379	if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5380		set_rq_online(rq);
5381
5382	raw_spin_unlock_irqrestore(&rq->lock, flags);
5383
5384	if (old_rd)
5385		call_rcu_sched(&old_rd->rcu, free_rootdomain);
5386}
5387
5388static int init_rootdomain(struct root_domain *rd)
5389{
5390	memset(rd, 0, sizeof(*rd));
5391
5392	if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5393		goto out;
5394	if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5395		goto free_span;
5396	if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5397		goto free_online;
5398
5399	if (cpupri_init(&rd->cpupri) != 0)
5400		goto free_rto_mask;
5401	return 0;
5402
5403free_rto_mask:
5404	free_cpumask_var(rd->rto_mask);
5405free_online:
5406	free_cpumask_var(rd->online);
5407free_span:
5408	free_cpumask_var(rd->span);
5409out:
5410	return -ENOMEM;
5411}
5412
5413/*
5414 * By default the system creates a single root-domain with all cpus as
5415 * members (mimicking the global state we have today).
5416 */
5417struct root_domain def_root_domain;
5418
5419static void init_defrootdomain(void)
5420{
5421	init_rootdomain(&def_root_domain);
5422
5423	atomic_set(&def_root_domain.refcount, 1);
5424}
5425
5426static struct root_domain *alloc_rootdomain(void)
5427{
5428	struct root_domain *rd;
5429
5430	rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5431	if (!rd)
5432		return NULL;
5433
5434	if (init_rootdomain(rd) != 0) {
5435		kfree(rd);
5436		return NULL;
5437	}
5438
5439	return rd;
5440}
5441
5442static void free_sched_groups(struct sched_group *sg, int free_sgp)
5443{
5444	struct sched_group *tmp, *first;
5445
5446	if (!sg)
5447		return;
5448
5449	first = sg;
5450	do {
5451		tmp = sg->next;
5452
5453		if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5454			kfree(sg->sgp);
5455
5456		kfree(sg);
5457		sg = tmp;
5458	} while (sg != first);
5459}
5460
5461static void free_sched_domain(struct rcu_head *rcu)
5462{
5463	struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5464
5465	/*
5466	 * If its an overlapping domain it has private groups, iterate and
5467	 * nuke them all.
5468	 */
5469	if (sd->flags & SD_OVERLAP) {
5470		free_sched_groups(sd->groups, 1);
5471	} else if (atomic_dec_and_test(&sd->groups->ref)) {
5472		kfree(sd->groups->sgp);
5473		kfree(sd->groups);
5474	}
5475	kfree(sd);
5476}
5477
5478static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5479{
5480	call_rcu(&sd->rcu, free_sched_domain);
5481}
5482
5483static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5484{
5485	for (; sd; sd = sd->parent)
5486		destroy_sched_domain(sd, cpu);
5487}
5488
5489/*
5490 * Keep a special pointer to the highest sched_domain that has
5491 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5492 * allows us to avoid some pointer chasing select_idle_sibling().
5493 *
5494 * Also keep a unique ID per domain (we use the first cpu number in
5495 * the cpumask of the domain), this allows us to quickly tell if
5496 * two cpus are in the same cache domain, see cpus_share_cache().
5497 */
5498DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5499DEFINE_PER_CPU(int, sd_llc_id);
5500
5501static void update_top_cache_domain(int cpu)
5502{
5503	struct sched_domain *sd;
5504	int id = cpu;
5505
5506	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5507	if (sd)
5508		id = cpumask_first(sched_domain_span(sd));
5509
5510	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5511	per_cpu(sd_llc_id, cpu) = id;
5512}
5513
5514/*
5515 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5516 * hold the hotplug lock.
5517 */
5518static void
5519cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5520{
5521	struct rq *rq = cpu_rq(cpu);
5522	struct sched_domain *tmp;
5523
5524	/* Remove the sched domains which do not contribute to scheduling. */
5525	for (tmp = sd; tmp; ) {
5526		struct sched_domain *parent = tmp->parent;
5527		if (!parent)
5528			break;
5529
5530		if (sd_parent_degenerate(tmp, parent)) {
5531			tmp->parent = parent->parent;
5532			if (parent->parent)
5533				parent->parent->child = tmp;
5534			destroy_sched_domain(parent, cpu);
5535		} else
5536			tmp = tmp->parent;
5537	}
5538
5539	if (sd && sd_degenerate(sd)) {
5540		tmp = sd;
5541		sd = sd->parent;
5542		destroy_sched_domain(tmp, cpu);
5543		if (sd)
5544			sd->child = NULL;
5545	}
5546
5547	sched_domain_debug(sd, cpu);
5548
5549	rq_attach_root(rq, rd);
5550	tmp = rq->sd;
5551	rcu_assign_pointer(rq->sd, sd);
5552	destroy_sched_domains(tmp, cpu);
5553
5554	update_top_cache_domain(cpu);
5555}
5556
5557/* cpus with isolated domains */
5558static cpumask_var_t cpu_isolated_map;
5559
5560/* Setup the mask of cpus configured for isolated domains */
5561static int __init isolated_cpu_setup(char *str)
5562{
5563	alloc_bootmem_cpumask_var(&cpu_isolated_map);
5564	cpulist_parse(str, cpu_isolated_map);
5565	return 1;
5566}
5567
5568__setup("isolcpus=", isolated_cpu_setup);
5569
5570static const struct cpumask *cpu_cpu_mask(int cpu)
5571{
5572	return cpumask_of_node(cpu_to_node(cpu));
5573}
5574
5575struct sd_data {
5576	struct sched_domain **__percpu sd;
5577	struct sched_group **__percpu sg;
5578	struct sched_group_power **__percpu sgp;
5579};
5580
5581struct s_data {
5582	struct sched_domain ** __percpu sd;
5583	struct root_domain	*rd;
5584};
5585
5586enum s_alloc {
5587	sa_rootdomain,
5588	sa_sd,
5589	sa_sd_storage,
5590	sa_none,
5591};
5592
5593struct sched_domain_topology_level;
5594
5595typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5596typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5597
5598#define SDTL_OVERLAP	0x01
5599
5600struct sched_domain_topology_level {
5601	sched_domain_init_f init;
5602	sched_domain_mask_f mask;
5603	int		    flags;
5604	int		    numa_level;
5605	struct sd_data      data;
5606};
5607
5608/*
5609 * Build an iteration mask that can exclude certain CPUs from the upwards
5610 * domain traversal.
5611 *
5612 * Asymmetric node setups can result in situations where the domain tree is of
5613 * unequal depth, make sure to skip domains that already cover the entire
5614 * range.
5615 *
5616 * In that case build_sched_domains() will have terminated the iteration early
5617 * and our sibling sd spans will be empty. Domains should always include the
5618 * cpu they're built on, so check that.
5619 *
5620 */
5621static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5622{
5623	const struct cpumask *span = sched_domain_span(sd);
5624	struct sd_data *sdd = sd->private;
5625	struct sched_domain *sibling;
5626	int i;
5627
5628	for_each_cpu(i, span) {
5629		sibling = *per_cpu_ptr(sdd->sd, i);
5630		if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5631			continue;
5632
5633		cpumask_set_cpu(i, sched_group_mask(sg));
5634	}
5635}
5636
5637/*
5638 * Return the canonical balance cpu for this group, this is the first cpu
5639 * of this group that's also in the iteration mask.
5640 */
5641int group_balance_cpu(struct sched_group *sg)
5642{
5643	return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5644}
5645
5646static int
5647build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5648{
5649	struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5650	const struct cpumask *span = sched_domain_span(sd);
5651	struct cpumask *covered = sched_domains_tmpmask;
5652	struct sd_data *sdd = sd->private;
5653	struct sched_domain *child;
5654	int i;
5655
5656	cpumask_clear(covered);
5657
5658	for_each_cpu(i, span) {
5659		struct cpumask *sg_span;
5660
5661		if (cpumask_test_cpu(i, covered))
5662			continue;
5663
5664		child = *per_cpu_ptr(sdd->sd, i);
5665
5666		/* See the comment near build_group_mask(). */
5667		if (!cpumask_test_cpu(i, sched_domain_span(child)))
5668			continue;
5669
5670		sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5671				GFP_KERNEL, cpu_to_node(cpu));
5672
5673		if (!sg)
5674			goto fail;
5675
5676		sg_span = sched_group_cpus(sg);
5677		if (child->child) {
5678			child = child->child;
5679			cpumask_copy(sg_span, sched_domain_span(child));
5680		} else
5681			cpumask_set_cpu(i, sg_span);
5682
5683		cpumask_or(covered, covered, sg_span);
5684
5685		sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5686		if (atomic_inc_return(&sg->sgp->ref) == 1)
5687			build_group_mask(sd, sg);
5688
5689		/*
5690		 * Initialize sgp->power such that even if we mess up the
5691		 * domains and no possible iteration will get us here, we won't
5692		 * die on a /0 trap.
5693		 */
5694		sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5695
5696		/*
5697		 * Make sure the first group of this domain contains the
5698		 * canonical balance cpu. Otherwise the sched_domain iteration
5699		 * breaks. See update_sg_lb_stats().
5700		 */
5701		if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5702		    group_balance_cpu(sg) == cpu)
5703			groups = sg;
5704
5705		if (!first)
5706			first = sg;
5707		if (last)
5708			last->next = sg;
5709		last = sg;
5710		last->next = first;
5711	}
5712	sd->groups = groups;
5713
5714	return 0;
5715
5716fail:
5717	free_sched_groups(first, 0);
5718
5719	return -ENOMEM;
5720}
5721
5722static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5723{
5724	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5725	struct sched_domain *child = sd->child;
5726
5727	if (child)
5728		cpu = cpumask_first(sched_domain_span(child));
5729
5730	if (sg) {
5731		*sg = *per_cpu_ptr(sdd->sg, cpu);
5732		(*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5733		atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5734	}
5735
5736	return cpu;
5737}
5738
5739/*
5740 * build_sched_groups will build a circular linked list of the groups
5741 * covered by the given span, and will set each group's ->cpumask correctly,
5742 * and ->cpu_power to 0.
5743 *
5744 * Assumes the sched_domain tree is fully constructed
5745 */
5746static int
5747build_sched_groups(struct sched_domain *sd, int cpu)
5748{
5749	struct sched_group *first = NULL, *last = NULL;
5750	struct sd_data *sdd = sd->private;
5751	const struct cpumask *span = sched_domain_span(sd);
5752	struct cpumask *covered;
5753	int i;
5754
5755	get_group(cpu, sdd, &sd->groups);
5756	atomic_inc(&sd->groups->ref);
5757
5758	if (cpu != cpumask_first(sched_domain_span(sd)))
5759		return 0;
5760
5761	lockdep_assert_held(&sched_domains_mutex);
5762	covered = sched_domains_tmpmask;
5763
5764	cpumask_clear(covered);
5765
5766	for_each_cpu(i, span) {
5767		struct sched_group *sg;
5768		int group = get_group(i, sdd, &sg);
5769		int j;
5770
5771		if (cpumask_test_cpu(i, covered))
5772			continue;
5773
5774		cpumask_clear(sched_group_cpus(sg));
5775		sg->sgp->power = 0;
5776		cpumask_setall(sched_group_mask(sg));
5777
5778		for_each_cpu(j, span) {
5779			if (get_group(j, sdd, NULL) != group)
5780				continue;
5781
5782			cpumask_set_cpu(j, covered);
5783			cpumask_set_cpu(j, sched_group_cpus(sg));
5784		}
5785
5786		if (!first)
5787			first = sg;
5788		if (last)
5789			last->next = sg;
5790		last = sg;
5791	}
5792	last->next = first;
5793
5794	return 0;
5795}
5796
5797/*
5798 * Initialize sched groups cpu_power.
5799 *
5800 * cpu_power indicates the capacity of sched group, which is used while
5801 * distributing the load between different sched groups in a sched domain.
5802 * Typically cpu_power for all the groups in a sched domain will be same unless
5803 * there are asymmetries in the topology. If there are asymmetries, group
5804 * having more cpu_power will pickup more load compared to the group having
5805 * less cpu_power.
5806 */
5807static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5808{
5809	struct sched_group *sg = sd->groups;
5810
5811	WARN_ON(!sd || !sg);
5812
5813	do {
5814		sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5815		sg = sg->next;
5816	} while (sg != sd->groups);
5817
5818	if (cpu != group_balance_cpu(sg))
5819		return;
5820
5821	update_group_power(sd, cpu);
5822	atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5823}
5824
5825int __weak arch_sd_sibling_asym_packing(void)
5826{
5827       return 0*SD_ASYM_PACKING;
5828}
5829
5830/*
5831 * Initializers for schedule domains
5832 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5833 */
5834
5835#ifdef CONFIG_SCHED_DEBUG
5836# define SD_INIT_NAME(sd, type)		sd->name = #type
5837#else
5838# define SD_INIT_NAME(sd, type)		do { } while (0)
5839#endif
5840
5841#define SD_INIT_FUNC(type)						\
5842static noinline struct sched_domain *					\
5843sd_init_##type(struct sched_domain_topology_level *tl, int cpu) 	\
5844{									\
5845	struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);	\
5846	*sd = SD_##type##_INIT;						\
5847	SD_INIT_NAME(sd, type);						\
5848	sd->private = &tl->data;					\
5849	return sd;							\
5850}
5851
5852SD_INIT_FUNC(CPU)
5853#ifdef CONFIG_SCHED_SMT
5854 SD_INIT_FUNC(SIBLING)
5855#endif
5856#ifdef CONFIG_SCHED_MC
5857 SD_INIT_FUNC(MC)
5858#endif
5859#ifdef CONFIG_SCHED_BOOK
5860 SD_INIT_FUNC(BOOK)
5861#endif
5862
5863static int default_relax_domain_level = -1;
5864int sched_domain_level_max;
5865
5866static int __init setup_relax_domain_level(char *str)
5867{
5868	if (kstrtoint(str, 0, &default_relax_domain_level))
5869		pr_warn("Unable to set relax_domain_level\n");
5870
5871	return 1;
5872}
5873__setup("relax_domain_level=", setup_relax_domain_level);
5874
5875static void set_domain_attribute(struct sched_domain *sd,
5876				 struct sched_domain_attr *attr)
5877{
5878	int request;
5879
5880	if (!attr || attr->relax_domain_level < 0) {
5881		if (default_relax_domain_level < 0)
5882			return;
5883		else
5884			request = default_relax_domain_level;
5885	} else
5886		request = attr->relax_domain_level;
5887	if (request < sd->level) {
5888		/* turn off idle balance on this domain */
5889		sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5890	} else {
5891		/* turn on idle balance on this domain */
5892		sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5893	}
5894}
5895
5896static void __sdt_free(const struct cpumask *cpu_map);
5897static int __sdt_alloc(const struct cpumask *cpu_map);
5898
5899static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5900				 const struct cpumask *cpu_map)
5901{
5902	switch (what) {
5903	case sa_rootdomain:
5904		if (!atomic_read(&d->rd->refcount))
5905			free_rootdomain(&d->rd->rcu); /* fall through */
5906	case sa_sd:
5907		free_percpu(d->sd); /* fall through */
5908	case sa_sd_storage:
5909		__sdt_free(cpu_map); /* fall through */
5910	case sa_none:
5911		break;
5912	}
5913}
5914
5915static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5916						   const struct cpumask *cpu_map)
5917{
5918	memset(d, 0, sizeof(*d));
5919
5920	if (__sdt_alloc(cpu_map))
5921		return sa_sd_storage;
5922	d->sd = alloc_percpu(struct sched_domain *);
5923	if (!d->sd)
5924		return sa_sd_storage;
5925	d->rd = alloc_rootdomain();
5926	if (!d->rd)
5927		return sa_sd;
5928	return sa_rootdomain;
5929}
5930
5931/*
5932 * NULL the sd_data elements we've used to build the sched_domain and
5933 * sched_group structure so that the subsequent __free_domain_allocs()
5934 * will not free the data we're using.
5935 */
5936static void claim_allocations(int cpu, struct sched_domain *sd)
5937{
5938	struct sd_data *sdd = sd->private;
5939
5940	WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5941	*per_cpu_ptr(sdd->sd, cpu) = NULL;
5942
5943	if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5944		*per_cpu_ptr(sdd->sg, cpu) = NULL;
5945
5946	if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5947		*per_cpu_ptr(sdd->sgp, cpu) = NULL;
5948}
5949
5950#ifdef CONFIG_SCHED_SMT
5951static const struct cpumask *cpu_smt_mask(int cpu)
5952{
5953	return topology_thread_cpumask(cpu);
5954}
5955#endif
5956
5957/*
5958 * Topology list, bottom-up.
5959 */
5960static struct sched_domain_topology_level default_topology[] = {
5961#ifdef CONFIG_SCHED_SMT
5962	{ sd_init_SIBLING, cpu_smt_mask, },
5963#endif
5964#ifdef CONFIG_SCHED_MC
5965	{ sd_init_MC, cpu_coregroup_mask, },
5966#endif
5967#ifdef CONFIG_SCHED_BOOK
5968	{ sd_init_BOOK, cpu_book_mask, },
5969#endif
5970	{ sd_init_CPU, cpu_cpu_mask, },
5971	{ NULL, },
5972};
5973
5974static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5975
5976#ifdef CONFIG_NUMA
5977
5978static int sched_domains_numa_levels;
5979static int *sched_domains_numa_distance;
5980static struct cpumask ***sched_domains_numa_masks;
5981static int sched_domains_curr_level;
5982
5983static inline int sd_local_flags(int level)
5984{
5985	if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5986		return 0;
5987
5988	return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5989}
5990
5991static struct sched_domain *
5992sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5993{
5994	struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5995	int level = tl->numa_level;
5996	int sd_weight = cpumask_weight(
5997			sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5998
5999	*sd = (struct sched_domain){
6000		.min_interval		= sd_weight,
6001		.max_interval		= 2*sd_weight,
6002		.busy_factor		= 32,
6003		.imbalance_pct		= 125,
6004		.cache_nice_tries	= 2,
6005		.busy_idx		= 3,
6006		.idle_idx		= 2,
6007		.newidle_idx		= 0,
6008		.wake_idx		= 0,
6009		.forkexec_idx		= 0,
6010
6011		.flags			= 1*SD_LOAD_BALANCE
6012					| 1*SD_BALANCE_NEWIDLE
6013					| 0*SD_BALANCE_EXEC
6014					| 0*SD_BALANCE_FORK
6015					| 0*SD_BALANCE_WAKE
6016					| 0*SD_WAKE_AFFINE
6017					| 0*SD_SHARE_CPUPOWER
6018					| 0*SD_SHARE_PKG_RESOURCES
6019					| 1*SD_SERIALIZE
6020					| 0*SD_PREFER_SIBLING
6021					| sd_local_flags(level)
6022					,
6023		.last_balance		= jiffies,
6024		.balance_interval	= sd_weight,
6025	};
6026	SD_INIT_NAME(sd, NUMA);
6027	sd->private = &tl->data;
6028
6029	/*
6030	 * Ugly hack to pass state to sd_numa_mask()...
6031	 */
6032	sched_domains_curr_level = tl->numa_level;
6033
6034	return sd;
6035}
6036
6037static const struct cpumask *sd_numa_mask(int cpu)
6038{
6039	return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6040}
6041
6042static void sched_numa_warn(const char *str)
6043{
6044	static int done = false;
6045	int i,j;
6046
6047	if (done)
6048		return;
6049
6050	done = true;
6051
6052	printk(KERN_WARNING "ERROR: %s\n\n", str);
6053
6054	for (i = 0; i < nr_node_ids; i++) {
6055		printk(KERN_WARNING "  ");
6056		for (j = 0; j < nr_node_ids; j++)
6057			printk(KERN_CONT "%02d ", node_distance(i,j));
6058		printk(KERN_CONT "\n");
6059	}
6060	printk(KERN_WARNING "\n");
6061}
6062
6063static bool find_numa_distance(int distance)
6064{
6065	int i;
6066
6067	if (distance == node_distance(0, 0))
6068		return true;
6069
6070	for (i = 0; i < sched_domains_numa_levels; i++) {
6071		if (sched_domains_numa_distance[i] == distance)
6072			return true;
6073	}
6074
6075	return false;
6076}
6077
6078static void sched_init_numa(void)
6079{
6080	int next_distance, curr_distance = node_distance(0, 0);
6081	struct sched_domain_topology_level *tl;
6082	int level = 0;
6083	int i, j, k;
6084
6085	sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6086	if (!sched_domains_numa_distance)
6087		return;
6088
6089	/*
6090	 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6091	 * unique distances in the node_distance() table.
6092	 *
6093	 * Assumes node_distance(0,j) includes all distances in
6094	 * node_distance(i,j) in order to avoid cubic time.
6095	 */
6096	next_distance = curr_distance;
6097	for (i = 0; i < nr_node_ids; i++) {
6098		for (j = 0; j < nr_node_ids; j++) {
6099			for (k = 0; k < nr_node_ids; k++) {
6100				int distance = node_distance(i, k);
6101
6102				if (distance > curr_distance &&
6103				    (distance < next_distance ||
6104				     next_distance == curr_distance))
6105					next_distance = distance;
6106
6107				/*
6108				 * While not a strong assumption it would be nice to know
6109				 * about cases where if node A is connected to B, B is not
6110				 * equally connected to A.
6111				 */
6112				if (sched_debug() && node_distance(k, i) != distance)
6113					sched_numa_warn("Node-distance not symmetric");
6114
6115				if (sched_debug() && i && !find_numa_distance(distance))
6116					sched_numa_warn("Node-0 not representative");
6117			}
6118			if (next_distance != curr_distance) {
6119				sched_domains_numa_distance[level++] = next_distance;
6120				sched_domains_numa_levels = level;
6121				curr_distance = next_distance;
6122			} else break;
6123		}
6124
6125		/*
6126		 * In case of sched_debug() we verify the above assumption.
6127		 */
6128		if (!sched_debug())
6129			break;
6130	}
6131	/*
6132	 * 'level' contains the number of unique distances, excluding the
6133	 * identity distance node_distance(i,i).
6134	 *
6135	 * The sched_domains_nume_distance[] array includes the actual distance
6136	 * numbers.
6137	 */
6138
6139	/*
6140	 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6141	 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6142	 * the array will contain less then 'level' members. This could be
6143	 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6144	 * in other functions.
6145	 *
6146	 * We reset it to 'level' at the end of this function.
6147	 */
6148	sched_domains_numa_levels = 0;
6149
6150	sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6151	if (!sched_domains_numa_masks)
6152		return;
6153
6154	/*
6155	 * Now for each level, construct a mask per node which contains all
6156	 * cpus of nodes that are that many hops away from us.
6157	 */
6158	for (i = 0; i < level; i++) {
6159		sched_domains_numa_masks[i] =
6160			kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6161		if (!sched_domains_numa_masks[i])
6162			return;
6163
6164		for (j = 0; j < nr_node_ids; j++) {
6165			struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6166			if (!mask)
6167				return;
6168
6169			sched_domains_numa_masks[i][j] = mask;
6170
6171			for (k = 0; k < nr_node_ids; k++) {
6172				if (node_distance(j, k) > sched_domains_numa_distance[i])
6173					continue;
6174
6175				cpumask_or(mask, mask, cpumask_of_node(k));
6176			}
6177		}
6178	}
6179
6180	tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6181			sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6182	if (!tl)
6183		return;
6184
6185	/*
6186	 * Copy the default topology bits..
6187	 */
6188	for (i = 0; default_topology[i].init; i++)
6189		tl[i] = default_topology[i];
6190
6191	/*
6192	 * .. and append 'j' levels of NUMA goodness.
6193	 */
6194	for (j = 0; j < level; i++, j++) {
6195		tl[i] = (struct sched_domain_topology_level){
6196			.init = sd_numa_init,
6197			.mask = sd_numa_mask,
6198			.flags = SDTL_OVERLAP,
6199			.numa_level = j,
6200		};
6201	}
6202
6203	sched_domain_topology = tl;
6204
6205	sched_domains_numa_levels = level;
6206}
6207
6208static void sched_domains_numa_masks_set(int cpu)
6209{
6210	int i, j;
6211	int node = cpu_to_node(cpu);
6212
6213	for (i = 0; i < sched_domains_numa_levels; i++) {
6214		for (j = 0; j < nr_node_ids; j++) {
6215			if (node_distance(j, node) <= sched_domains_numa_distance[i])
6216				cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6217		}
6218	}
6219}
6220
6221static void sched_domains_numa_masks_clear(int cpu)
6222{
6223	int i, j;
6224	for (i = 0; i < sched_domains_numa_levels; i++) {
6225		for (j = 0; j < nr_node_ids; j++)
6226			cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6227	}
6228}
6229
6230/*
6231 * Update sched_domains_numa_masks[level][node] array when new cpus
6232 * are onlined.
6233 */
6234static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6235					   unsigned long action,
6236					   void *hcpu)
6237{
6238	int cpu = (long)hcpu;
6239
6240	switch (action & ~CPU_TASKS_FROZEN) {
6241	case CPU_ONLINE:
6242		sched_domains_numa_masks_set(cpu);
6243		break;
6244
6245	case CPU_DEAD:
6246		sched_domains_numa_masks_clear(cpu);
6247		break;
6248
6249	default:
6250		return NOTIFY_DONE;
6251	}
6252
6253	return NOTIFY_OK;
6254}
6255#else
6256static inline void sched_init_numa(void)
6257{
6258}
6259
6260static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6261					   unsigned long action,
6262					   void *hcpu)
6263{
6264	return 0;
6265}
6266#endif /* CONFIG_NUMA */
6267
6268static int __sdt_alloc(const struct cpumask *cpu_map)
6269{
6270	struct sched_domain_topology_level *tl;
6271	int j;
6272
6273	for (tl = sched_domain_topology; tl->init; tl++) {
6274		struct sd_data *sdd = &tl->data;
6275
6276		sdd->sd = alloc_percpu(struct sched_domain *);
6277		if (!sdd->sd)
6278			return -ENOMEM;
6279
6280		sdd->sg = alloc_percpu(struct sched_group *);
6281		if (!sdd->sg)
6282			return -ENOMEM;
6283
6284		sdd->sgp = alloc_percpu(struct sched_group_power *);
6285		if (!sdd->sgp)
6286			return -ENOMEM;
6287
6288		for_each_cpu(j, cpu_map) {
6289			struct sched_domain *sd;
6290			struct sched_group *sg;
6291			struct sched_group_power *sgp;
6292
6293		       	sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6294					GFP_KERNEL, cpu_to_node(j));
6295			if (!sd)
6296				return -ENOMEM;
6297
6298			*per_cpu_ptr(sdd->sd, j) = sd;
6299
6300			sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6301					GFP_KERNEL, cpu_to_node(j));
6302			if (!sg)
6303				return -ENOMEM;
6304
6305			sg->next = sg;
6306
6307			*per_cpu_ptr(sdd->sg, j) = sg;
6308
6309			sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6310					GFP_KERNEL, cpu_to_node(j));
6311			if (!sgp)
6312				return -ENOMEM;
6313
6314			*per_cpu_ptr(sdd->sgp, j) = sgp;
6315		}
6316	}
6317
6318	return 0;
6319}
6320
6321static void __sdt_free(const struct cpumask *cpu_map)
6322{
6323	struct sched_domain_topology_level *tl;
6324	int j;
6325
6326	for (tl = sched_domain_topology; tl->init; tl++) {
6327		struct sd_data *sdd = &tl->data;
6328
6329		for_each_cpu(j, cpu_map) {
6330			struct sched_domain *sd;
6331
6332			if (sdd->sd) {
6333				sd = *per_cpu_ptr(sdd->sd, j);
6334				if (sd && (sd->flags & SD_OVERLAP))
6335					free_sched_groups(sd->groups, 0);
6336				kfree(*per_cpu_ptr(sdd->sd, j));
6337			}
6338
6339			if (sdd->sg)
6340				kfree(*per_cpu_ptr(sdd->sg, j));
6341			if (sdd->sgp)
6342				kfree(*per_cpu_ptr(sdd->sgp, j));
6343		}
6344		free_percpu(sdd->sd);
6345		sdd->sd = NULL;
6346		free_percpu(sdd->sg);
6347		sdd->sg = NULL;
6348		free_percpu(sdd->sgp);
6349		sdd->sgp = NULL;
6350	}
6351}
6352
6353struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6354		struct s_data *d, const struct cpumask *cpu_map,
6355		struct sched_domain_attr *attr, struct sched_domain *child,
6356		int cpu)
6357{
6358	struct sched_domain *sd = tl->init(tl, cpu);
6359	if (!sd)
6360		return child;
6361
6362	cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6363	if (child) {
6364		sd->level = child->level + 1;
6365		sched_domain_level_max = max(sched_domain_level_max, sd->level);
6366		child->parent = sd;
6367	}
6368	sd->child = child;
6369	set_domain_attribute(sd, attr);
6370
6371	return sd;
6372}
6373
6374/*
6375 * Build sched domains for a given set of cpus and attach the sched domains
6376 * to the individual cpus
6377 */
6378static int build_sched_domains(const struct cpumask *cpu_map,
6379			       struct sched_domain_attr *attr)
6380{
6381	enum s_alloc alloc_state = sa_none;
6382	struct sched_domain *sd;
6383	struct s_data d;
6384	int i, ret = -ENOMEM;
6385
6386	alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6387	if (alloc_state != sa_rootdomain)
6388		goto error;
6389
6390	/* Set up domains for cpus specified by the cpu_map. */
6391	for_each_cpu(i, cpu_map) {
6392		struct sched_domain_topology_level *tl;
6393
6394		sd = NULL;
6395		for (tl = sched_domain_topology; tl->init; tl++) {
6396			sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6397			if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6398				sd->flags |= SD_OVERLAP;
6399			if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6400				break;
6401		}
6402
6403		while (sd->child)
6404			sd = sd->child;
6405
6406		*per_cpu_ptr(d.sd, i) = sd;
6407	}
6408
6409	/* Build the groups for the domains */
6410	for_each_cpu(i, cpu_map) {
6411		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6412			sd->span_weight = cpumask_weight(sched_domain_span(sd));
6413			if (sd->flags & SD_OVERLAP) {
6414				if (build_overlap_sched_groups(sd, i))
6415					goto error;
6416			} else {
6417				if (build_sched_groups(sd, i))
6418					goto error;
6419			}
6420		}
6421	}
6422
6423	/* Calculate CPU power for physical packages and nodes */
6424	for (i = nr_cpumask_bits-1; i >= 0; i--) {
6425		if (!cpumask_test_cpu(i, cpu_map))
6426			continue;
6427
6428		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6429			claim_allocations(i, sd);
6430			init_sched_groups_power(i, sd);
6431		}
6432	}
6433
6434	/* Attach the domains */
6435	rcu_read_lock();
6436	for_each_cpu(i, cpu_map) {
6437		sd = *per_cpu_ptr(d.sd, i);
6438		cpu_attach_domain(sd, d.rd, i);
6439	}
6440	rcu_read_unlock();
6441
6442	ret = 0;
6443error:
6444	__free_domain_allocs(&d, alloc_state, cpu_map);
6445	return ret;
6446}
6447
6448static cpumask_var_t *doms_cur;	/* current sched domains */
6449static int ndoms_cur;		/* number of sched domains in 'doms_cur' */
6450static struct sched_domain_attr *dattr_cur;
6451				/* attribues of custom domains in 'doms_cur' */
6452
6453/*
6454 * Special case: If a kmalloc of a doms_cur partition (array of
6455 * cpumask) fails, then fallback to a single sched domain,
6456 * as determined by the single cpumask fallback_doms.
6457 */
6458static cpumask_var_t fallback_doms;
6459
6460/*
6461 * arch_update_cpu_topology lets virtualized architectures update the
6462 * cpu core maps. It is supposed to return 1 if the topology changed
6463 * or 0 if it stayed the same.
6464 */
6465int __attribute__((weak)) arch_update_cpu_topology(void)
6466{
6467	return 0;
6468}
6469
6470cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6471{
6472	int i;
6473	cpumask_var_t *doms;
6474
6475	doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6476	if (!doms)
6477		return NULL;
6478	for (i = 0; i < ndoms; i++) {
6479		if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6480			free_sched_domains(doms, i);
6481			return NULL;
6482		}
6483	}
6484	return doms;
6485}
6486
6487void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6488{
6489	unsigned int i;
6490	for (i = 0; i < ndoms; i++)
6491		free_cpumask_var(doms[i]);
6492	kfree(doms);
6493}
6494
6495/*
6496 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6497 * For now this just excludes isolated cpus, but could be used to
6498 * exclude other special cases in the future.
6499 */
6500static int init_sched_domains(const struct cpumask *cpu_map)
6501{
6502	int err;
6503
6504	arch_update_cpu_topology();
6505	ndoms_cur = 1;
6506	doms_cur = alloc_sched_domains(ndoms_cur);
6507	if (!doms_cur)
6508		doms_cur = &fallback_doms;
6509	cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6510	err = build_sched_domains(doms_cur[0], NULL);
6511	register_sched_domain_sysctl();
6512
6513	return err;
6514}
6515
6516/*
6517 * Detach sched domains from a group of cpus specified in cpu_map
6518 * These cpus will now be attached to the NULL domain
6519 */
6520static void detach_destroy_domains(const struct cpumask *cpu_map)
6521{
6522	int i;
6523
6524	rcu_read_lock();
6525	for_each_cpu(i, cpu_map)
6526		cpu_attach_domain(NULL, &def_root_domain, i);
6527	rcu_read_unlock();
6528}
6529
6530/* handle null as "default" */
6531static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6532			struct sched_domain_attr *new, int idx_new)
6533{
6534	struct sched_domain_attr tmp;
6535
6536	/* fast path */
6537	if (!new && !cur)
6538		return 1;
6539
6540	tmp = SD_ATTR_INIT;
6541	return !memcmp(cur ? (cur + idx_cur) : &tmp,
6542			new ? (new + idx_new) : &tmp,
6543			sizeof(struct sched_domain_attr));
6544}
6545
6546/*
6547 * Partition sched domains as specified by the 'ndoms_new'
6548 * cpumasks in the array doms_new[] of cpumasks. This compares
6549 * doms_new[] to the current sched domain partitioning, doms_cur[].
6550 * It destroys each deleted domain and builds each new domain.
6551 *
6552 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6553 * The masks don't intersect (don't overlap.) We should setup one
6554 * sched domain for each mask. CPUs not in any of the cpumasks will
6555 * not be load balanced. If the same cpumask appears both in the
6556 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6557 * it as it is.
6558 *
6559 * The passed in 'doms_new' should be allocated using
6560 * alloc_sched_domains.  This routine takes ownership of it and will
6561 * free_sched_domains it when done with it. If the caller failed the
6562 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6563 * and partition_sched_domains() will fallback to the single partition
6564 * 'fallback_doms', it also forces the domains to be rebuilt.
6565 *
6566 * If doms_new == NULL it will be replaced with cpu_online_mask.
6567 * ndoms_new == 0 is a special case for destroying existing domains,
6568 * and it will not create the default domain.
6569 *
6570 * Call with hotplug lock held
6571 */
6572void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6573			     struct sched_domain_attr *dattr_new)
6574{
6575	int i, j, n;
6576	int new_topology;
6577
6578	mutex_lock(&sched_domains_mutex);
6579
6580	/* always unregister in case we don't destroy any domains */
6581	unregister_sched_domain_sysctl();
6582
6583	/* Let architecture update cpu core mappings. */
6584	new_topology = arch_update_cpu_topology();
6585
6586	n = doms_new ? ndoms_new : 0;
6587
6588	/* Destroy deleted domains */
6589	for (i = 0; i < ndoms_cur; i++) {
6590		for (j = 0; j < n && !new_topology; j++) {
6591			if (cpumask_equal(doms_cur[i], doms_new[j])
6592			    && dattrs_equal(dattr_cur, i, dattr_new, j))
6593				goto match1;
6594		}
6595		/* no match - a current sched domain not in new doms_new[] */
6596		detach_destroy_domains(doms_cur[i]);
6597match1:
6598		;
6599	}
6600
6601	if (doms_new == NULL) {
6602		ndoms_cur = 0;
6603		doms_new = &fallback_doms;
6604		cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6605		WARN_ON_ONCE(dattr_new);
6606	}
6607
6608	/* Build new domains */
6609	for (i = 0; i < ndoms_new; i++) {
6610		for (j = 0; j < ndoms_cur && !new_topology; j++) {
6611			if (cpumask_equal(doms_new[i], doms_cur[j])
6612			    && dattrs_equal(dattr_new, i, dattr_cur, j))
6613				goto match2;
6614		}
6615		/* no match - add a new doms_new */
6616		build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6617match2:
6618		;
6619	}
6620
6621	/* Remember the new sched domains */
6622	if (doms_cur != &fallback_doms)
6623		free_sched_domains(doms_cur, ndoms_cur);
6624	kfree(dattr_cur);	/* kfree(NULL) is safe */
6625	doms_cur = doms_new;
6626	dattr_cur = dattr_new;
6627	ndoms_cur = ndoms_new;
6628
6629	register_sched_domain_sysctl();
6630
6631	mutex_unlock(&sched_domains_mutex);
6632}
6633
6634static int num_cpus_frozen;	/* used to mark begin/end of suspend/resume */
6635
6636/*
6637 * Update cpusets according to cpu_active mask.  If cpusets are
6638 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6639 * around partition_sched_domains().
6640 *
6641 * If we come here as part of a suspend/resume, don't touch cpusets because we
6642 * want to restore it back to its original state upon resume anyway.
6643 */
6644static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6645			     void *hcpu)
6646{
6647	switch (action) {
6648	case CPU_ONLINE_FROZEN:
6649	case CPU_DOWN_FAILED_FROZEN:
6650
6651		/*
6652		 * num_cpus_frozen tracks how many CPUs are involved in suspend
6653		 * resume sequence. As long as this is not the last online
6654		 * operation in the resume sequence, just build a single sched
6655		 * domain, ignoring cpusets.
6656		 */
6657		num_cpus_frozen--;
6658		if (likely(num_cpus_frozen)) {
6659			partition_sched_domains(1, NULL, NULL);
6660			break;
6661		}
6662
6663		/*
6664		 * This is the last CPU online operation. So fall through and
6665		 * restore the original sched domains by considering the
6666		 * cpuset configurations.
6667		 */
6668
6669	case CPU_ONLINE:
6670	case CPU_DOWN_FAILED:
6671		cpuset_update_active_cpus(true);
6672		break;
6673	default:
6674		return NOTIFY_DONE;
6675	}
6676	return NOTIFY_OK;
6677}
6678
6679static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6680			       void *hcpu)
6681{
6682	switch (action) {
6683	case CPU_DOWN_PREPARE:
6684		cpuset_update_active_cpus(false);
6685		break;
6686	case CPU_DOWN_PREPARE_FROZEN:
6687		num_cpus_frozen++;
6688		partition_sched_domains(1, NULL, NULL);
6689		break;
6690	default:
6691		return NOTIFY_DONE;
6692	}
6693	return NOTIFY_OK;
6694}
6695
6696void __init sched_init_smp(void)
6697{
6698	cpumask_var_t non_isolated_cpus;
6699
6700	alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6701	alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6702
6703	sched_init_numa();
6704
6705	get_online_cpus();
6706	mutex_lock(&sched_domains_mutex);
6707	init_sched_domains(cpu_active_mask);
6708	cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6709	if (cpumask_empty(non_isolated_cpus))
6710		cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6711	mutex_unlock(&sched_domains_mutex);
6712	put_online_cpus();
6713
6714	hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6715	hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6716	hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6717
6718	/* RT runtime code needs to handle some hotplug events */
6719	hotcpu_notifier(update_runtime, 0);
6720
6721	init_hrtick();
6722
6723	/* Move init over to a non-isolated CPU */
6724	if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6725		BUG();
6726	sched_init_granularity();
6727	free_cpumask_var(non_isolated_cpus);
6728
6729	init_sched_rt_class();
6730}
6731#else
6732void __init sched_init_smp(void)
6733{
6734	sched_init_granularity();
6735}
6736#endif /* CONFIG_SMP */
6737
6738const_debug unsigned int sysctl_timer_migration = 1;
6739
6740int in_sched_functions(unsigned long addr)
6741{
6742	return in_lock_functions(addr) ||
6743		(addr >= (unsigned long)__sched_text_start
6744		&& addr < (unsigned long)__sched_text_end);
6745}
6746
6747#ifdef CONFIG_CGROUP_SCHED
6748struct task_group root_task_group;
6749LIST_HEAD(task_groups);
6750#endif
6751
6752DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6753
6754void __init sched_init(void)
6755{
6756	int i, j;
6757	unsigned long alloc_size = 0, ptr;
6758
6759#ifdef CONFIG_FAIR_GROUP_SCHED
6760	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6761#endif
6762#ifdef CONFIG_RT_GROUP_SCHED
6763	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6764#endif
6765#ifdef CONFIG_CPUMASK_OFFSTACK
6766	alloc_size += num_possible_cpus() * cpumask_size();
6767#endif
6768	if (alloc_size) {
6769		ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6770
6771#ifdef CONFIG_FAIR_GROUP_SCHED
6772		root_task_group.se = (struct sched_entity **)ptr;
6773		ptr += nr_cpu_ids * sizeof(void **);
6774
6775		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6776		ptr += nr_cpu_ids * sizeof(void **);
6777
6778#endif /* CONFIG_FAIR_GROUP_SCHED */
6779#ifdef CONFIG_RT_GROUP_SCHED
6780		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6781		ptr += nr_cpu_ids * sizeof(void **);
6782
6783		root_task_group.rt_rq = (struct rt_rq **)ptr;
6784		ptr += nr_cpu_ids * sizeof(void **);
6785
6786#endif /* CONFIG_RT_GROUP_SCHED */
6787#ifdef CONFIG_CPUMASK_OFFSTACK
6788		for_each_possible_cpu(i) {
6789			per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6790			ptr += cpumask_size();
6791		}
6792#endif /* CONFIG_CPUMASK_OFFSTACK */
6793	}
6794
6795#ifdef CONFIG_SMP
6796	init_defrootdomain();
6797#endif
6798
6799	init_rt_bandwidth(&def_rt_bandwidth,
6800			global_rt_period(), global_rt_runtime());
6801
6802#ifdef CONFIG_RT_GROUP_SCHED
6803	init_rt_bandwidth(&root_task_group.rt_bandwidth,
6804			global_rt_period(), global_rt_runtime());
6805#endif /* CONFIG_RT_GROUP_SCHED */
6806
6807#ifdef CONFIG_CGROUP_SCHED
6808	list_add(&root_task_group.list, &task_groups);
6809	INIT_LIST_HEAD(&root_task_group.children);
6810	INIT_LIST_HEAD(&root_task_group.siblings);
6811	autogroup_init(&init_task);
6812
6813#endif /* CONFIG_CGROUP_SCHED */
6814
6815#ifdef CONFIG_CGROUP_CPUACCT
6816	root_cpuacct.cpustat = &kernel_cpustat;
6817	root_cpuacct.cpuusage = alloc_percpu(u64);
6818	/* Too early, not expected to fail */
6819	BUG_ON(!root_cpuacct.cpuusage);
6820#endif
6821	for_each_possible_cpu(i) {
6822		struct rq *rq;
6823
6824		rq = cpu_rq(i);
6825		raw_spin_lock_init(&rq->lock);
6826		rq->nr_running = 0;
6827		rq->calc_load_active = 0;
6828		rq->calc_load_update = jiffies + LOAD_FREQ;
6829		init_cfs_rq(&rq->cfs);
6830		init_rt_rq(&rq->rt, rq);
6831#ifdef CONFIG_FAIR_GROUP_SCHED
6832		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6833		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6834		/*
6835		 * How much cpu bandwidth does root_task_group get?
6836		 *
6837		 * In case of task-groups formed thr' the cgroup filesystem, it
6838		 * gets 100% of the cpu resources in the system. This overall
6839		 * system cpu resource is divided among the tasks of
6840		 * root_task_group and its child task-groups in a fair manner,
6841		 * based on each entity's (task or task-group's) weight
6842		 * (se->load.weight).
6843		 *
6844		 * In other words, if root_task_group has 10 tasks of weight
6845		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6846		 * then A0's share of the cpu resource is:
6847		 *
6848		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6849		 *
6850		 * We achieve this by letting root_task_group's tasks sit
6851		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6852		 */
6853		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6854		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6855#endif /* CONFIG_FAIR_GROUP_SCHED */
6856
6857		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6858#ifdef CONFIG_RT_GROUP_SCHED
6859		INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6860		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6861#endif
6862
6863		for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6864			rq->cpu_load[j] = 0;
6865
6866		rq->last_load_update_tick = jiffies;
6867
6868#ifdef CONFIG_SMP
6869		rq->sd = NULL;
6870		rq->rd = NULL;
6871		rq->cpu_power = SCHED_POWER_SCALE;
6872		rq->post_schedule = 0;
6873		rq->active_balance = 0;
6874		rq->next_balance = jiffies;
6875		rq->push_cpu = 0;
6876		rq->cpu = i;
6877		rq->online = 0;
6878		rq->idle_stamp = 0;
6879		rq->avg_idle = 2*sysctl_sched_migration_cost;
6880
6881		INIT_LIST_HEAD(&rq->cfs_tasks);
6882
6883		rq_attach_root(rq, &def_root_domain);
6884#ifdef CONFIG_NO_HZ
6885		rq->nohz_flags = 0;
6886#endif
6887#endif
6888		init_rq_hrtick(rq);
6889		atomic_set(&rq->nr_iowait, 0);
6890	}
6891
6892	set_load_weight(&init_task);
6893
6894#ifdef CONFIG_PREEMPT_NOTIFIERS
6895	INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6896#endif
6897
6898#ifdef CONFIG_RT_MUTEXES
6899	plist_head_init(&init_task.pi_waiters);
6900#endif
6901
6902	/*
6903	 * The boot idle thread does lazy MMU switching as well:
6904	 */
6905	atomic_inc(&init_mm.mm_count);
6906	enter_lazy_tlb(&init_mm, current);
6907
6908	/*
6909	 * Make us the idle thread. Technically, schedule() should not be
6910	 * called from this thread, however somewhere below it might be,
6911	 * but because we are the idle thread, we just pick up running again
6912	 * when this runqueue becomes "idle".
6913	 */
6914	init_idle(current, smp_processor_id());
6915
6916	calc_load_update = jiffies + LOAD_FREQ;
6917
6918	/*
6919	 * During early bootup we pretend to be a normal task:
6920	 */
6921	current->sched_class = &fair_sched_class;
6922
6923#ifdef CONFIG_SMP
6924	zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6925	/* May be allocated at isolcpus cmdline parse time */
6926	if (cpu_isolated_map == NULL)
6927		zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6928	idle_thread_set_boot_cpu();
6929#endif
6930	init_sched_fair_class();
6931
6932	scheduler_running = 1;
6933}
6934
6935#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6936static inline int preempt_count_equals(int preempt_offset)
6937{
6938	int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6939
6940	return (nested == preempt_offset);
6941}
6942
6943void __might_sleep(const char *file, int line, int preempt_offset)
6944{
6945	static unsigned long prev_jiffy;	/* ratelimiting */
6946
6947	rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6948	if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6949	    system_state != SYSTEM_RUNNING || oops_in_progress)
6950		return;
6951	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6952		return;
6953	prev_jiffy = jiffies;
6954
6955	printk(KERN_ERR
6956		"BUG: sleeping function called from invalid context at %s:%d\n",
6957			file, line);
6958	printk(KERN_ERR
6959		"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6960			in_atomic(), irqs_disabled(),
6961			current->pid, current->comm);
6962
6963	debug_show_held_locks(current);
6964	if (irqs_disabled())
6965		print_irqtrace_events(current);
6966	dump_stack();
6967}
6968EXPORT_SYMBOL(__might_sleep);
6969#endif
6970
6971#ifdef CONFIG_MAGIC_SYSRQ
6972static void normalize_task(struct rq *rq, struct task_struct *p)
6973{
6974	const struct sched_class *prev_class = p->sched_class;
6975	int old_prio = p->prio;
6976	int on_rq;
6977
6978	on_rq = p->on_rq;
6979	if (on_rq)
6980		dequeue_task(rq, p, 0);
6981	__setscheduler(rq, p, SCHED_NORMAL, 0);
6982	if (on_rq) {
6983		enqueue_task(rq, p, 0);
6984		resched_task(rq->curr);
6985	}
6986
6987	check_class_changed(rq, p, prev_class, old_prio);
6988}
6989
6990void normalize_rt_tasks(void)
6991{
6992	struct task_struct *g, *p;
6993	unsigned long flags;
6994	struct rq *rq;
6995
6996	read_lock_irqsave(&tasklist_lock, flags);
6997	do_each_thread(g, p) {
6998		/*
6999		 * Only normalize user tasks:
7000		 */
7001		if (!p->mm)
7002			continue;
7003
7004		p->se.exec_start		= 0;
7005#ifdef CONFIG_SCHEDSTATS
7006		p->se.statistics.wait_start	= 0;
7007		p->se.statistics.sleep_start	= 0;
7008		p->se.statistics.block_start	= 0;
7009#endif
7010
7011		if (!rt_task(p)) {
7012			/*
7013			 * Renice negative nice level userspace
7014			 * tasks back to 0:
7015			 */
7016			if (TASK_NICE(p) < 0 && p->mm)
7017				set_user_nice(p, 0);
7018			continue;
7019		}
7020
7021		raw_spin_lock(&p->pi_lock);
7022		rq = __task_rq_lock(p);
7023
7024		normalize_task(rq, p);
7025
7026		__task_rq_unlock(rq);
7027		raw_spin_unlock(&p->pi_lock);
7028	} while_each_thread(g, p);
7029
7030	read_unlock_irqrestore(&tasklist_lock, flags);
7031}
7032
7033#endif /* CONFIG_MAGIC_SYSRQ */
7034
7035#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7036/*
7037 * These functions are only useful for the IA64 MCA handling, or kdb.
7038 *
7039 * They can only be called when the whole system has been
7040 * stopped - every CPU needs to be quiescent, and no scheduling
7041 * activity can take place. Using them for anything else would
7042 * be a serious bug, and as a result, they aren't even visible
7043 * under any other configuration.
7044 */
7045
7046/**
7047 * curr_task - return the current task for a given cpu.
7048 * @cpu: the processor in question.
7049 *
7050 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7051 */
7052struct task_struct *curr_task(int cpu)
7053{
7054	return cpu_curr(cpu);
7055}
7056
7057#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7058
7059#ifdef CONFIG_IA64
7060/**
7061 * set_curr_task - set the current task for a given cpu.
7062 * @cpu: the processor in question.
7063 * @p: the task pointer to set.
7064 *
7065 * Description: This function must only be used when non-maskable interrupts
7066 * are serviced on a separate stack. It allows the architecture to switch the
7067 * notion of the current task on a cpu in a non-blocking manner. This function
7068 * must be called with all CPU's synchronized, and interrupts disabled, the
7069 * and caller must save the original value of the current task (see
7070 * curr_task() above) and restore that value before reenabling interrupts and
7071 * re-starting the system.
7072 *
7073 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7074 */
7075void set_curr_task(int cpu, struct task_struct *p)
7076{
7077	cpu_curr(cpu) = p;
7078}
7079
7080#endif
7081
7082#ifdef CONFIG_CGROUP_SCHED
7083/* task_group_lock serializes the addition/removal of task groups */
7084static DEFINE_SPINLOCK(task_group_lock);
7085
7086static void free_sched_group(struct task_group *tg)
7087{
7088	free_fair_sched_group(tg);
7089	free_rt_sched_group(tg);
7090	autogroup_free(tg);
7091	kfree(tg);
7092}
7093
7094/* allocate runqueue etc for a new task group */
7095struct task_group *sched_create_group(struct task_group *parent)
7096{
7097	struct task_group *tg;
7098	unsigned long flags;
7099
7100	tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7101	if (!tg)
7102		return ERR_PTR(-ENOMEM);
7103
7104	if (!alloc_fair_sched_group(tg, parent))
7105		goto err;
7106
7107	if (!alloc_rt_sched_group(tg, parent))
7108		goto err;
7109
7110	spin_lock_irqsave(&task_group_lock, flags);
7111	list_add_rcu(&tg->list, &task_groups);
7112
7113	WARN_ON(!parent); /* root should already exist */
7114
7115	tg->parent = parent;
7116	INIT_LIST_HEAD(&tg->children);
7117	list_add_rcu(&tg->siblings, &parent->children);
7118	spin_unlock_irqrestore(&task_group_lock, flags);
7119
7120	return tg;
7121
7122err:
7123	free_sched_group(tg);
7124	return ERR_PTR(-ENOMEM);
7125}
7126
7127/* rcu callback to free various structures associated with a task group */
7128static void free_sched_group_rcu(struct rcu_head *rhp)
7129{
7130	/* now it should be safe to free those cfs_rqs */
7131	free_sched_group(container_of(rhp, struct task_group, rcu));
7132}
7133
7134/* Destroy runqueue etc associated with a task group */
7135void sched_destroy_group(struct task_group *tg)
7136{
7137	unsigned long flags;
7138	int i;
7139
7140	/* end participation in shares distribution */
7141	for_each_possible_cpu(i)
7142		unregister_fair_sched_group(tg, i);
7143
7144	spin_lock_irqsave(&task_group_lock, flags);
7145	list_del_rcu(&tg->list);
7146	list_del_rcu(&tg->siblings);
7147	spin_unlock_irqrestore(&task_group_lock, flags);
7148
7149	/* wait for possible concurrent references to cfs_rqs complete */
7150	call_rcu(&tg->rcu, free_sched_group_rcu);
7151}
7152
7153/* change task's runqueue when it moves between groups.
7154 *	The caller of this function should have put the task in its new group
7155 *	by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7156 *	reflect its new group.
7157 */
7158void sched_move_task(struct task_struct *tsk)
7159{
7160	struct task_group *tg;
7161	int on_rq, running;
7162	unsigned long flags;
7163	struct rq *rq;
7164
7165	rq = task_rq_lock(tsk, &flags);
7166
7167	running = task_current(rq, tsk);
7168	on_rq = tsk->on_rq;
7169
7170	if (on_rq)
7171		dequeue_task(rq, tsk, 0);
7172	if (unlikely(running))
7173		tsk->sched_class->put_prev_task(rq, tsk);
7174
7175	tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7176				lockdep_is_held(&tsk->sighand->siglock)),
7177			  struct task_group, css);
7178	tg = autogroup_task_group(tsk, tg);
7179	tsk->sched_task_group = tg;
7180
7181#ifdef CONFIG_FAIR_GROUP_SCHED
7182	if (tsk->sched_class->task_move_group)
7183		tsk->sched_class->task_move_group(tsk, on_rq);
7184	else
7185#endif
7186		set_task_rq(tsk, task_cpu(tsk));
7187
7188	if (unlikely(running))
7189		tsk->sched_class->set_curr_task(rq);
7190	if (on_rq)
7191		enqueue_task(rq, tsk, 0);
7192
7193	task_rq_unlock(rq, tsk, &flags);
7194}
7195#endif /* CONFIG_CGROUP_SCHED */
7196
7197#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7198static unsigned long to_ratio(u64 period, u64 runtime)
7199{
7200	if (runtime == RUNTIME_INF)
7201		return 1ULL << 20;
7202
7203	return div64_u64(runtime << 20, period);
7204}
7205#endif
7206
7207#ifdef CONFIG_RT_GROUP_SCHED
7208/*
7209 * Ensure that the real time constraints are schedulable.
7210 */
7211static DEFINE_MUTEX(rt_constraints_mutex);
7212
7213/* Must be called with tasklist_lock held */
7214static inline int tg_has_rt_tasks(struct task_group *tg)
7215{
7216	struct task_struct *g, *p;
7217
7218	do_each_thread(g, p) {
7219		if (rt_task(p) && task_rq(p)->rt.tg == tg)
7220			return 1;
7221	} while_each_thread(g, p);
7222
7223	return 0;
7224}
7225
7226struct rt_schedulable_data {
7227	struct task_group *tg;
7228	u64 rt_period;
7229	u64 rt_runtime;
7230};
7231
7232static int tg_rt_schedulable(struct task_group *tg, void *data)
7233{
7234	struct rt_schedulable_data *d = data;
7235	struct task_group *child;
7236	unsigned long total, sum = 0;
7237	u64 period, runtime;
7238
7239	period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7240	runtime = tg->rt_bandwidth.rt_runtime;
7241
7242	if (tg == d->tg) {
7243		period = d->rt_period;
7244		runtime = d->rt_runtime;
7245	}
7246
7247	/*
7248	 * Cannot have more runtime than the period.
7249	 */
7250	if (runtime > period && runtime != RUNTIME_INF)
7251		return -EINVAL;
7252
7253	/*
7254	 * Ensure we don't starve existing RT tasks.
7255	 */
7256	if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7257		return -EBUSY;
7258
7259	total = to_ratio(period, runtime);
7260
7261	/*
7262	 * Nobody can have more than the global setting allows.
7263	 */
7264	if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7265		return -EINVAL;
7266
7267	/*
7268	 * The sum of our children's runtime should not exceed our own.
7269	 */
7270	list_for_each_entry_rcu(child, &tg->children, siblings) {
7271		period = ktime_to_ns(child->rt_bandwidth.rt_period);
7272		runtime = child->rt_bandwidth.rt_runtime;
7273
7274		if (child == d->tg) {
7275			period = d->rt_period;
7276			runtime = d->rt_runtime;
7277		}
7278
7279		sum += to_ratio(period, runtime);
7280	}
7281
7282	if (sum > total)
7283		return -EINVAL;
7284
7285	return 0;
7286}
7287
7288static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7289{
7290	int ret;
7291
7292	struct rt_schedulable_data data = {
7293		.tg = tg,
7294		.rt_period = period,
7295		.rt_runtime = runtime,
7296	};
7297
7298	rcu_read_lock();
7299	ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7300	rcu_read_unlock();
7301
7302	return ret;
7303}
7304
7305static int tg_set_rt_bandwidth(struct task_group *tg,
7306		u64 rt_period, u64 rt_runtime)
7307{
7308	int i, err = 0;
7309
7310	mutex_lock(&rt_constraints_mutex);
7311	read_lock(&tasklist_lock);
7312	err = __rt_schedulable(tg, rt_period, rt_runtime);
7313	if (err)
7314		goto unlock;
7315
7316	raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7317	tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7318	tg->rt_bandwidth.rt_runtime = rt_runtime;
7319
7320	for_each_possible_cpu(i) {
7321		struct rt_rq *rt_rq = tg->rt_rq[i];
7322
7323		raw_spin_lock(&rt_rq->rt_runtime_lock);
7324		rt_rq->rt_runtime = rt_runtime;
7325		raw_spin_unlock(&rt_rq->rt_runtime_lock);
7326	}
7327	raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7328unlock:
7329	read_unlock(&tasklist_lock);
7330	mutex_unlock(&rt_constraints_mutex);
7331
7332	return err;
7333}
7334
7335int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7336{
7337	u64 rt_runtime, rt_period;
7338
7339	rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7340	rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7341	if (rt_runtime_us < 0)
7342		rt_runtime = RUNTIME_INF;
7343
7344	return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7345}
7346
7347long sched_group_rt_runtime(struct task_group *tg)
7348{
7349	u64 rt_runtime_us;
7350
7351	if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7352		return -1;
7353
7354	rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7355	do_div(rt_runtime_us, NSEC_PER_USEC);
7356	return rt_runtime_us;
7357}
7358
7359int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7360{
7361	u64 rt_runtime, rt_period;
7362
7363	rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7364	rt_runtime = tg->rt_bandwidth.rt_runtime;
7365
7366	if (rt_period == 0)
7367		return -EINVAL;
7368
7369	return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7370}
7371
7372long sched_group_rt_period(struct task_group *tg)
7373{
7374	u64 rt_period_us;
7375
7376	rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7377	do_div(rt_period_us, NSEC_PER_USEC);
7378	return rt_period_us;
7379}
7380
7381static int sched_rt_global_constraints(void)
7382{
7383	u64 runtime, period;
7384	int ret = 0;
7385
7386	if (sysctl_sched_rt_period <= 0)
7387		return -EINVAL;
7388
7389	runtime = global_rt_runtime();
7390	period = global_rt_period();
7391
7392	/*
7393	 * Sanity check on the sysctl variables.
7394	 */
7395	if (runtime > period && runtime != RUNTIME_INF)
7396		return -EINVAL;
7397
7398	mutex_lock(&rt_constraints_mutex);
7399	read_lock(&tasklist_lock);
7400	ret = __rt_schedulable(NULL, 0, 0);
7401	read_unlock(&tasklist_lock);
7402	mutex_unlock(&rt_constraints_mutex);
7403
7404	return ret;
7405}
7406
7407int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7408{
7409	/* Don't accept realtime tasks when there is no way for them to run */
7410	if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7411		return 0;
7412
7413	return 1;
7414}
7415
7416#else /* !CONFIG_RT_GROUP_SCHED */
7417static int sched_rt_global_constraints(void)
7418{
7419	unsigned long flags;
7420	int i;
7421
7422	if (sysctl_sched_rt_period <= 0)
7423		return -EINVAL;
7424
7425	/*
7426	 * There's always some RT tasks in the root group
7427	 * -- migration, kstopmachine etc..
7428	 */
7429	if (sysctl_sched_rt_runtime == 0)
7430		return -EBUSY;
7431
7432	raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7433	for_each_possible_cpu(i) {
7434		struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7435
7436		raw_spin_lock(&rt_rq->rt_runtime_lock);
7437		rt_rq->rt_runtime = global_rt_runtime();
7438		raw_spin_unlock(&rt_rq->rt_runtime_lock);
7439	}
7440	raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7441
7442	return 0;
7443}
7444#endif /* CONFIG_RT_GROUP_SCHED */
7445
7446int sched_rt_handler(struct ctl_table *table, int write,
7447		void __user *buffer, size_t *lenp,
7448		loff_t *ppos)
7449{
7450	int ret;
7451	int old_period, old_runtime;
7452	static DEFINE_MUTEX(mutex);
7453
7454	mutex_lock(&mutex);
7455	old_period = sysctl_sched_rt_period;
7456	old_runtime = sysctl_sched_rt_runtime;
7457
7458	ret = proc_dointvec(table, write, buffer, lenp, ppos);
7459
7460	if (!ret && write) {
7461		ret = sched_rt_global_constraints();
7462		if (ret) {
7463			sysctl_sched_rt_period = old_period;
7464			sysctl_sched_rt_runtime = old_runtime;
7465		} else {
7466			def_rt_bandwidth.rt_runtime = global_rt_runtime();
7467			def_rt_bandwidth.rt_period =
7468				ns_to_ktime(global_rt_period());
7469		}
7470	}
7471	mutex_unlock(&mutex);
7472
7473	return ret;
7474}
7475
7476#ifdef CONFIG_CGROUP_SCHED
7477
7478/* return corresponding task_group object of a cgroup */
7479static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7480{
7481	return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7482			    struct task_group, css);
7483}
7484
7485static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7486{
7487	struct task_group *tg, *parent;
7488
7489	if (!cgrp->parent) {
7490		/* This is early initialization for the top cgroup */
7491		return &root_task_group.css;
7492	}
7493
7494	parent = cgroup_tg(cgrp->parent);
7495	tg = sched_create_group(parent);
7496	if (IS_ERR(tg))
7497		return ERR_PTR(-ENOMEM);
7498
7499	return &tg->css;
7500}
7501
7502static void cpu_cgroup_destroy(struct cgroup *cgrp)
7503{
7504	struct task_group *tg = cgroup_tg(cgrp);
7505
7506	sched_destroy_group(tg);
7507}
7508
7509static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7510				 struct cgroup_taskset *tset)
7511{
7512	struct task_struct *task;
7513
7514	cgroup_taskset_for_each(task, cgrp, tset) {
7515#ifdef CONFIG_RT_GROUP_SCHED
7516		if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7517			return -EINVAL;
7518#else
7519		/* We don't support RT-tasks being in separate groups */
7520		if (task->sched_class != &fair_sched_class)
7521			return -EINVAL;
7522#endif
7523	}
7524	return 0;
7525}
7526
7527static void cpu_cgroup_attach(struct cgroup *cgrp,
7528			      struct cgroup_taskset *tset)
7529{
7530	struct task_struct *task;
7531
7532	cgroup_taskset_for_each(task, cgrp, tset)
7533		sched_move_task(task);
7534}
7535
7536static void
7537cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7538		struct task_struct *task)
7539{
7540	/*
7541	 * cgroup_exit() is called in the copy_process() failure path.
7542	 * Ignore this case since the task hasn't ran yet, this avoids
7543	 * trying to poke a half freed task state from generic code.
7544	 */
7545	if (!(task->flags & PF_EXITING))
7546		return;
7547
7548	sched_move_task(task);
7549}
7550
7551#ifdef CONFIG_FAIR_GROUP_SCHED
7552static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7553				u64 shareval)
7554{
7555	return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7556}
7557
7558static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7559{
7560	struct task_group *tg = cgroup_tg(cgrp);
7561
7562	return (u64) scale_load_down(tg->shares);
7563}
7564
7565#ifdef CONFIG_CFS_BANDWIDTH
7566static DEFINE_MUTEX(cfs_constraints_mutex);
7567
7568const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7569const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7570
7571static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7572
7573static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7574{
7575	int i, ret = 0, runtime_enabled, runtime_was_enabled;
7576	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7577
7578	if (tg == &root_task_group)
7579		return -EINVAL;
7580
7581	/*
7582	 * Ensure we have at some amount of bandwidth every period.  This is
7583	 * to prevent reaching a state of large arrears when throttled via
7584	 * entity_tick() resulting in prolonged exit starvation.
7585	 */
7586	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7587		return -EINVAL;
7588
7589	/*
7590	 * Likewise, bound things on the otherside by preventing insane quota
7591	 * periods.  This also allows us to normalize in computing quota
7592	 * feasibility.
7593	 */
7594	if (period > max_cfs_quota_period)
7595		return -EINVAL;
7596
7597	mutex_lock(&cfs_constraints_mutex);
7598	ret = __cfs_schedulable(tg, period, quota);
7599	if (ret)
7600		goto out_unlock;
7601
7602	runtime_enabled = quota != RUNTIME_INF;
7603	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7604	account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7605	raw_spin_lock_irq(&cfs_b->lock);
7606	cfs_b->period = ns_to_ktime(period);
7607	cfs_b->quota = quota;
7608
7609	__refill_cfs_bandwidth_runtime(cfs_b);
7610	/* restart the period timer (if active) to handle new period expiry */
7611	if (runtime_enabled && cfs_b->timer_active) {
7612		/* force a reprogram */
7613		cfs_b->timer_active = 0;
7614		__start_cfs_bandwidth(cfs_b);
7615	}
7616	raw_spin_unlock_irq(&cfs_b->lock);
7617
7618	for_each_possible_cpu(i) {
7619		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7620		struct rq *rq = cfs_rq->rq;
7621
7622		raw_spin_lock_irq(&rq->lock);
7623		cfs_rq->runtime_enabled = runtime_enabled;
7624		cfs_rq->runtime_remaining = 0;
7625
7626		if (cfs_rq->throttled)
7627			unthrottle_cfs_rq(cfs_rq);
7628		raw_spin_unlock_irq(&rq->lock);
7629	}
7630out_unlock:
7631	mutex_unlock(&cfs_constraints_mutex);
7632
7633	return ret;
7634}
7635
7636int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7637{
7638	u64 quota, period;
7639
7640	period = ktime_to_ns(tg->cfs_bandwidth.period);
7641	if (cfs_quota_us < 0)
7642		quota = RUNTIME_INF;
7643	else
7644		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7645
7646	return tg_set_cfs_bandwidth(tg, period, quota);
7647}
7648
7649long tg_get_cfs_quota(struct task_group *tg)
7650{
7651	u64 quota_us;
7652
7653	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7654		return -1;
7655
7656	quota_us = tg->cfs_bandwidth.quota;
7657	do_div(quota_us, NSEC_PER_USEC);
7658
7659	return quota_us;
7660}
7661
7662int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7663{
7664	u64 quota, period;
7665
7666	period = (u64)cfs_period_us * NSEC_PER_USEC;
7667	quota = tg->cfs_bandwidth.quota;
7668
7669	return tg_set_cfs_bandwidth(tg, period, quota);
7670}
7671
7672long tg_get_cfs_period(struct task_group *tg)
7673{
7674	u64 cfs_period_us;
7675
7676	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7677	do_div(cfs_period_us, NSEC_PER_USEC);
7678
7679	return cfs_period_us;
7680}
7681
7682static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7683{
7684	return tg_get_cfs_quota(cgroup_tg(cgrp));
7685}
7686
7687static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7688				s64 cfs_quota_us)
7689{
7690	return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7691}
7692
7693static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7694{
7695	return tg_get_cfs_period(cgroup_tg(cgrp));
7696}
7697
7698static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7699				u64 cfs_period_us)
7700{
7701	return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7702}
7703
7704struct cfs_schedulable_data {
7705	struct task_group *tg;
7706	u64 period, quota;
7707};
7708
7709/*
7710 * normalize group quota/period to be quota/max_period
7711 * note: units are usecs
7712 */
7713static u64 normalize_cfs_quota(struct task_group *tg,
7714			       struct cfs_schedulable_data *d)
7715{
7716	u64 quota, period;
7717
7718	if (tg == d->tg) {
7719		period = d->period;
7720		quota = d->quota;
7721	} else {
7722		period = tg_get_cfs_period(tg);
7723		quota = tg_get_cfs_quota(tg);
7724	}
7725
7726	/* note: these should typically be equivalent */
7727	if (quota == RUNTIME_INF || quota == -1)
7728		return RUNTIME_INF;
7729
7730	return to_ratio(period, quota);
7731}
7732
7733static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7734{
7735	struct cfs_schedulable_data *d = data;
7736	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7737	s64 quota = 0, parent_quota = -1;
7738
7739	if (!tg->parent) {
7740		quota = RUNTIME_INF;
7741	} else {
7742		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7743
7744		quota = normalize_cfs_quota(tg, d);
7745		parent_quota = parent_b->hierarchal_quota;
7746
7747		/*
7748		 * ensure max(child_quota) <= parent_quota, inherit when no
7749		 * limit is set
7750		 */
7751		if (quota == RUNTIME_INF)
7752			quota = parent_quota;
7753		else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7754			return -EINVAL;
7755	}
7756	cfs_b->hierarchal_quota = quota;
7757
7758	return 0;
7759}
7760
7761static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7762{
7763	int ret;
7764	struct cfs_schedulable_data data = {
7765		.tg = tg,
7766		.period = period,
7767		.quota = quota,
7768	};
7769
7770	if (quota != RUNTIME_INF) {
7771		do_div(data.period, NSEC_PER_USEC);
7772		do_div(data.quota, NSEC_PER_USEC);
7773	}
7774
7775	rcu_read_lock();
7776	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7777	rcu_read_unlock();
7778
7779	return ret;
7780}
7781
7782static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7783		struct cgroup_map_cb *cb)
7784{
7785	struct task_group *tg = cgroup_tg(cgrp);
7786	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7787
7788	cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7789	cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7790	cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7791
7792	return 0;
7793}
7794#endif /* CONFIG_CFS_BANDWIDTH */
7795#endif /* CONFIG_FAIR_GROUP_SCHED */
7796
7797#ifdef CONFIG_RT_GROUP_SCHED
7798static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7799				s64 val)
7800{
7801	return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7802}
7803
7804static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7805{
7806	return sched_group_rt_runtime(cgroup_tg(cgrp));
7807}
7808
7809static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7810		u64 rt_period_us)
7811{
7812	return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7813}
7814
7815static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7816{
7817	return sched_group_rt_period(cgroup_tg(cgrp));
7818}
7819#endif /* CONFIG_RT_GROUP_SCHED */
7820
7821static struct cftype cpu_files[] = {
7822#ifdef CONFIG_FAIR_GROUP_SCHED
7823	{
7824		.name = "shares",
7825		.read_u64 = cpu_shares_read_u64,
7826		.write_u64 = cpu_shares_write_u64,
7827	},
7828#endif
7829#ifdef CONFIG_CFS_BANDWIDTH
7830	{
7831		.name = "cfs_quota_us",
7832		.read_s64 = cpu_cfs_quota_read_s64,
7833		.write_s64 = cpu_cfs_quota_write_s64,
7834	},
7835	{
7836		.name = "cfs_period_us",
7837		.read_u64 = cpu_cfs_period_read_u64,
7838		.write_u64 = cpu_cfs_period_write_u64,
7839	},
7840	{
7841		.name = "stat",
7842		.read_map = cpu_stats_show,
7843	},
7844#endif
7845#ifdef CONFIG_RT_GROUP_SCHED
7846	{
7847		.name = "rt_runtime_us",
7848		.read_s64 = cpu_rt_runtime_read,
7849		.write_s64 = cpu_rt_runtime_write,
7850	},
7851	{
7852		.name = "rt_period_us",
7853		.read_u64 = cpu_rt_period_read_uint,
7854		.write_u64 = cpu_rt_period_write_uint,
7855	},
7856#endif
7857	{ }	/* terminate */
7858};
7859
7860struct cgroup_subsys cpu_cgroup_subsys = {
7861	.name		= "cpu",
7862	.create		= cpu_cgroup_create,
7863	.destroy	= cpu_cgroup_destroy,
7864	.can_attach	= cpu_cgroup_can_attach,
7865	.attach		= cpu_cgroup_attach,
7866	.exit		= cpu_cgroup_exit,
7867	.subsys_id	= cpu_cgroup_subsys_id,
7868	.base_cftypes	= cpu_files,
7869	.early_init	= 1,
7870};
7871
7872#endif	/* CONFIG_CGROUP_SCHED */
7873
7874#ifdef CONFIG_CGROUP_CPUACCT
7875
7876/*
7877 * CPU accounting code for task groups.
7878 *
7879 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7880 * (balbir@in.ibm.com).
7881 */
7882
7883struct cpuacct root_cpuacct;
7884
7885/* create a new cpu accounting group */
7886static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
7887{
7888	struct cpuacct *ca;
7889
7890	if (!cgrp->parent)
7891		return &root_cpuacct.css;
7892
7893	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7894	if (!ca)
7895		goto out;
7896
7897	ca->cpuusage = alloc_percpu(u64);
7898	if (!ca->cpuusage)
7899		goto out_free_ca;
7900
7901	ca->cpustat = alloc_percpu(struct kernel_cpustat);
7902	if (!ca->cpustat)
7903		goto out_free_cpuusage;
7904
7905	return &ca->css;
7906
7907out_free_cpuusage:
7908	free_percpu(ca->cpuusage);
7909out_free_ca:
7910	kfree(ca);
7911out:
7912	return ERR_PTR(-ENOMEM);
7913}
7914
7915/* destroy an existing cpu accounting group */
7916static void cpuacct_destroy(struct cgroup *cgrp)
7917{
7918	struct cpuacct *ca = cgroup_ca(cgrp);
7919
7920	free_percpu(ca->cpustat);
7921	free_percpu(ca->cpuusage);
7922	kfree(ca);
7923}
7924
7925static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7926{
7927	u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7928	u64 data;
7929
7930#ifndef CONFIG_64BIT
7931	/*
7932	 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7933	 */
7934	raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7935	data = *cpuusage;
7936	raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7937#else
7938	data = *cpuusage;
7939#endif
7940
7941	return data;
7942}
7943
7944static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7945{
7946	u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7947
7948#ifndef CONFIG_64BIT
7949	/*
7950	 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7951	 */
7952	raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7953	*cpuusage = val;
7954	raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7955#else
7956	*cpuusage = val;
7957#endif
7958}
7959
7960/* return total cpu usage (in nanoseconds) of a group */
7961static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
7962{
7963	struct cpuacct *ca = cgroup_ca(cgrp);
7964	u64 totalcpuusage = 0;
7965	int i;
7966
7967	for_each_present_cpu(i)
7968		totalcpuusage += cpuacct_cpuusage_read(ca, i);
7969
7970	return totalcpuusage;
7971}
7972
7973static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
7974								u64 reset)
7975{
7976	struct cpuacct *ca = cgroup_ca(cgrp);
7977	int err = 0;
7978	int i;
7979
7980	if (reset) {
7981		err = -EINVAL;
7982		goto out;
7983	}
7984
7985	for_each_present_cpu(i)
7986		cpuacct_cpuusage_write(ca, i, 0);
7987
7988out:
7989	return err;
7990}
7991
7992static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
7993				   struct seq_file *m)
7994{
7995	struct cpuacct *ca = cgroup_ca(cgroup);
7996	u64 percpu;
7997	int i;
7998
7999	for_each_present_cpu(i) {
8000		percpu = cpuacct_cpuusage_read(ca, i);
8001		seq_printf(m, "%llu ", (unsigned long long) percpu);
8002	}
8003	seq_printf(m, "\n");
8004	return 0;
8005}
8006
8007static const char *cpuacct_stat_desc[] = {
8008	[CPUACCT_STAT_USER] = "user",
8009	[CPUACCT_STAT_SYSTEM] = "system",
8010};
8011
8012static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8013			      struct cgroup_map_cb *cb)
8014{
8015	struct cpuacct *ca = cgroup_ca(cgrp);
8016	int cpu;
8017	s64 val = 0;
8018
8019	for_each_online_cpu(cpu) {
8020		struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8021		val += kcpustat->cpustat[CPUTIME_USER];
8022		val += kcpustat->cpustat[CPUTIME_NICE];
8023	}
8024	val = cputime64_to_clock_t(val);
8025	cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8026
8027	val = 0;
8028	for_each_online_cpu(cpu) {
8029		struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8030		val += kcpustat->cpustat[CPUTIME_SYSTEM];
8031		val += kcpustat->cpustat[CPUTIME_IRQ];
8032		val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8033	}
8034
8035	val = cputime64_to_clock_t(val);
8036	cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8037
8038	return 0;
8039}
8040
8041static struct cftype files[] = {
8042	{
8043		.name = "usage",
8044		.read_u64 = cpuusage_read,
8045		.write_u64 = cpuusage_write,
8046	},
8047	{
8048		.name = "usage_percpu",
8049		.read_seq_string = cpuacct_percpu_seq_read,
8050	},
8051	{
8052		.name = "stat",
8053		.read_map = cpuacct_stats_show,
8054	},
8055	{ }	/* terminate */
8056};
8057
8058/*
8059 * charge this task's execution time to its accounting group.
8060 *
8061 * called with rq->lock held.
8062 */
8063void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8064{
8065	struct cpuacct *ca;
8066	int cpu;
8067
8068	if (unlikely(!cpuacct_subsys.active))
8069		return;
8070
8071	cpu = task_cpu(tsk);
8072
8073	rcu_read_lock();
8074
8075	ca = task_ca(tsk);
8076
8077	for (; ca; ca = parent_ca(ca)) {
8078		u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8079		*cpuusage += cputime;
8080	}
8081
8082	rcu_read_unlock();
8083}
8084
8085struct cgroup_subsys cpuacct_subsys = {
8086	.name = "cpuacct",
8087	.create = cpuacct_create,
8088	.destroy = cpuacct_destroy,
8089	.subsys_id = cpuacct_subsys_id,
8090	.base_cftypes = files,
8091};
8092#endif	/* CONFIG_CGROUP_CPUACCT */
8093