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 
75 #include <asm/tlb.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
80 #endif
81 
82 #include "sched.h"
83 #include "../workqueue_sched.h"
84 
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/sched.h>
87 
start_bandwidth_timer(struct hrtimer * period_timer,ktime_t period)88 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
89 {
90 	unsigned long delta;
91 	ktime_t soft, hard, now;
92 
93 	for (;;) {
94 		if (hrtimer_active(period_timer))
95 			break;
96 
97 		now = hrtimer_cb_get_time(period_timer);
98 		hrtimer_forward(period_timer, now, period);
99 
100 		soft = hrtimer_get_softexpires(period_timer);
101 		hard = hrtimer_get_expires(period_timer);
102 		delta = ktime_to_ns(ktime_sub(hard, soft));
103 		__hrtimer_start_range_ns(period_timer, soft, delta,
104 					 HRTIMER_MODE_ABS_PINNED, 0);
105 	}
106 }
107 
108 DEFINE_MUTEX(sched_domains_mutex);
109 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
110 
111 static void update_rq_clock_task(struct rq *rq, s64 delta);
112 
update_rq_clock(struct rq * rq)113 void update_rq_clock(struct rq *rq)
114 {
115 	s64 delta;
116 
117 	if (rq->skip_clock_update > 0)
118 		return;
119 
120 	delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
121 	rq->clock += delta;
122 	update_rq_clock_task(rq, delta);
123 }
124 
125 /*
126  * Debugging: various feature bits
127  */
128 
129 #define SCHED_FEAT(name, enabled)	\
130 	(1UL << __SCHED_FEAT_##name) * enabled |
131 
132 const_debug unsigned int sysctl_sched_features =
133 #include "features.h"
134 	0;
135 
136 #undef SCHED_FEAT
137 
138 #ifdef CONFIG_SCHED_DEBUG
139 #define SCHED_FEAT(name, enabled)	\
140 	#name ,
141 
142 static __read_mostly char *sched_feat_names[] = {
143 #include "features.h"
144 	NULL
145 };
146 
147 #undef SCHED_FEAT
148 
sched_feat_show(struct seq_file * m,void * v)149 static int sched_feat_show(struct seq_file *m, void *v)
150 {
151 	int i;
152 
153 	for (i = 0; i < __SCHED_FEAT_NR; i++) {
154 		if (!(sysctl_sched_features & (1UL << i)))
155 			seq_puts(m, "NO_");
156 		seq_printf(m, "%s ", sched_feat_names[i]);
157 	}
158 	seq_puts(m, "\n");
159 
160 	return 0;
161 }
162 
163 #ifdef HAVE_JUMP_LABEL
164 
165 #define jump_label_key__true  jump_label_key_enabled
166 #define jump_label_key__false jump_label_key_disabled
167 
168 #define SCHED_FEAT(name, enabled)	\
169 	jump_label_key__##enabled ,
170 
171 struct jump_label_key sched_feat_keys[__SCHED_FEAT_NR] = {
172 #include "features.h"
173 };
174 
175 #undef SCHED_FEAT
176 
sched_feat_disable(int i)177 static void sched_feat_disable(int i)
178 {
179 	if (jump_label_enabled(&sched_feat_keys[i]))
180 		jump_label_dec(&sched_feat_keys[i]);
181 }
182 
sched_feat_enable(int i)183 static void sched_feat_enable(int i)
184 {
185 	if (!jump_label_enabled(&sched_feat_keys[i]))
186 		jump_label_inc(&sched_feat_keys[i]);
187 }
188 #else
sched_feat_disable(int i)189 static void sched_feat_disable(int i) { };
sched_feat_enable(int i)190 static void sched_feat_enable(int i) { };
191 #endif /* HAVE_JUMP_LABEL */
192 
193 static ssize_t
sched_feat_write(struct file * filp,const char __user * ubuf,size_t cnt,loff_t * ppos)194 sched_feat_write(struct file *filp, const char __user *ubuf,
195 		size_t cnt, loff_t *ppos)
196 {
197 	char buf[64];
198 	char *cmp;
199 	int neg = 0;
200 	int i;
201 
202 	if (cnt > 63)
203 		cnt = 63;
204 
205 	if (copy_from_user(&buf, ubuf, cnt))
206 		return -EFAULT;
207 
208 	buf[cnt] = 0;
209 	cmp = strstrip(buf);
210 
211 	if (strncmp(cmp, "NO_", 3) == 0) {
212 		neg = 1;
213 		cmp += 3;
214 	}
215 
216 	for (i = 0; i < __SCHED_FEAT_NR; i++) {
217 		if (strcmp(cmp, sched_feat_names[i]) == 0) {
218 			if (neg) {
219 				sysctl_sched_features &= ~(1UL << i);
220 				sched_feat_disable(i);
221 			} else {
222 				sysctl_sched_features |= (1UL << i);
223 				sched_feat_enable(i);
224 			}
225 			break;
226 		}
227 	}
228 
229 	if (i == __SCHED_FEAT_NR)
230 		return -EINVAL;
231 
232 	*ppos += cnt;
233 
234 	return cnt;
235 }
236 
sched_feat_open(struct inode * inode,struct file * filp)237 static int sched_feat_open(struct inode *inode, struct file *filp)
238 {
239 	return single_open(filp, sched_feat_show, NULL);
240 }
241 
242 static const struct file_operations sched_feat_fops = {
243 	.open		= sched_feat_open,
244 	.write		= sched_feat_write,
245 	.read		= seq_read,
246 	.llseek		= seq_lseek,
247 	.release	= single_release,
248 };
249 
sched_init_debug(void)250 static __init int sched_init_debug(void)
251 {
252 	debugfs_create_file("sched_features", 0644, NULL, NULL,
253 			&sched_feat_fops);
254 
255 	return 0;
256 }
257 late_initcall(sched_init_debug);
258 #endif /* CONFIG_SCHED_DEBUG */
259 
260 /*
261  * Number of tasks to iterate in a single balance run.
262  * Limited because this is done with IRQs disabled.
263  */
264 const_debug unsigned int sysctl_sched_nr_migrate = 32;
265 
266 /*
267  * period over which we average the RT time consumption, measured
268  * in ms.
269  *
270  * default: 1s
271  */
272 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
273 
274 /*
275  * period over which we measure -rt task cpu usage in us.
276  * default: 1s
277  */
278 unsigned int sysctl_sched_rt_period = 1000000;
279 
280 __read_mostly int scheduler_running;
281 
282 /*
283  * part of the period that we allow rt tasks to run in us.
284  * default: 0.95s
285  */
286 int sysctl_sched_rt_runtime = 950000;
287 
288 
289 
290 /*
291  * __task_rq_lock - lock the rq @p resides on.
292  */
__task_rq_lock(struct task_struct * p)293 static inline struct rq *__task_rq_lock(struct task_struct *p)
294 	__acquires(rq->lock)
295 {
296 	struct rq *rq;
297 
298 	lockdep_assert_held(&p->pi_lock);
299 
300 	for (;;) {
301 		rq = task_rq(p);
302 		raw_spin_lock(&rq->lock);
303 		if (likely(rq == task_rq(p)))
304 			return rq;
305 		raw_spin_unlock(&rq->lock);
306 	}
307 }
308 
309 /*
310  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
311  */
task_rq_lock(struct task_struct * p,unsigned long * flags)312 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
313 	__acquires(p->pi_lock)
314 	__acquires(rq->lock)
315 {
316 	struct rq *rq;
317 
318 	for (;;) {
319 		raw_spin_lock_irqsave(&p->pi_lock, *flags);
320 		rq = task_rq(p);
321 		raw_spin_lock(&rq->lock);
322 		if (likely(rq == task_rq(p)))
323 			return rq;
324 		raw_spin_unlock(&rq->lock);
325 		raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
326 	}
327 }
328 
__task_rq_unlock(struct rq * rq)329 static void __task_rq_unlock(struct rq *rq)
330 	__releases(rq->lock)
331 {
332 	raw_spin_unlock(&rq->lock);
333 }
334 
335 static inline void
task_rq_unlock(struct rq * rq,struct task_struct * p,unsigned long * flags)336 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
337 	__releases(rq->lock)
338 	__releases(p->pi_lock)
339 {
340 	raw_spin_unlock(&rq->lock);
341 	raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
342 }
343 
344 /*
345  * this_rq_lock - lock this runqueue and disable interrupts.
346  */
this_rq_lock(void)347 static struct rq *this_rq_lock(void)
348 	__acquires(rq->lock)
349 {
350 	struct rq *rq;
351 
352 	local_irq_disable();
353 	rq = this_rq();
354 	raw_spin_lock(&rq->lock);
355 
356 	return rq;
357 }
358 
359 #ifdef CONFIG_SCHED_HRTICK
360 /*
361  * Use HR-timers to deliver accurate preemption points.
362  *
363  * Its all a bit involved since we cannot program an hrt while holding the
364  * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
365  * reschedule event.
366  *
367  * When we get rescheduled we reprogram the hrtick_timer outside of the
368  * rq->lock.
369  */
370 
hrtick_clear(struct rq * rq)371 static void hrtick_clear(struct rq *rq)
372 {
373 	if (hrtimer_active(&rq->hrtick_timer))
374 		hrtimer_cancel(&rq->hrtick_timer);
375 }
376 
377 /*
378  * High-resolution timer tick.
379  * Runs from hardirq context with interrupts disabled.
380  */
hrtick(struct hrtimer * timer)381 static enum hrtimer_restart hrtick(struct hrtimer *timer)
382 {
383 	struct rq *rq = container_of(timer, struct rq, hrtick_timer);
384 
385 	WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
386 
387 	raw_spin_lock(&rq->lock);
388 	update_rq_clock(rq);
389 	rq->curr->sched_class->task_tick(rq, rq->curr, 1);
390 	raw_spin_unlock(&rq->lock);
391 
392 	return HRTIMER_NORESTART;
393 }
394 
395 #ifdef CONFIG_SMP
396 /*
397  * called from hardirq (IPI) context
398  */
__hrtick_start(void * arg)399 static void __hrtick_start(void *arg)
400 {
401 	struct rq *rq = arg;
402 
403 	raw_spin_lock(&rq->lock);
404 	hrtimer_restart(&rq->hrtick_timer);
405 	rq->hrtick_csd_pending = 0;
406 	raw_spin_unlock(&rq->lock);
407 }
408 
409 /*
410  * Called to set the hrtick timer state.
411  *
412  * called with rq->lock held and irqs disabled
413  */
hrtick_start(struct rq * rq,u64 delay)414 void hrtick_start(struct rq *rq, u64 delay)
415 {
416 	struct hrtimer *timer = &rq->hrtick_timer;
417 	ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
418 
419 	hrtimer_set_expires(timer, time);
420 
421 	if (rq == this_rq()) {
422 		hrtimer_restart(timer);
423 	} else if (!rq->hrtick_csd_pending) {
424 		__smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
425 		rq->hrtick_csd_pending = 1;
426 	}
427 }
428 
429 static int
hotplug_hrtick(struct notifier_block * nfb,unsigned long action,void * hcpu)430 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
431 {
432 	int cpu = (int)(long)hcpu;
433 
434 	switch (action) {
435 	case CPU_UP_CANCELED:
436 	case CPU_UP_CANCELED_FROZEN:
437 	case CPU_DOWN_PREPARE:
438 	case CPU_DOWN_PREPARE_FROZEN:
439 	case CPU_DEAD:
440 	case CPU_DEAD_FROZEN:
441 		hrtick_clear(cpu_rq(cpu));
442 		return NOTIFY_OK;
443 	}
444 
445 	return NOTIFY_DONE;
446 }
447 
init_hrtick(void)448 static __init void init_hrtick(void)
449 {
450 	hotcpu_notifier(hotplug_hrtick, 0);
451 }
452 #else
453 /*
454  * Called to set the hrtick timer state.
455  *
456  * called with rq->lock held and irqs disabled
457  */
hrtick_start(struct rq * rq,u64 delay)458 void hrtick_start(struct rq *rq, u64 delay)
459 {
460 	__hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
461 			HRTIMER_MODE_REL_PINNED, 0);
462 }
463 
init_hrtick(void)464 static inline void init_hrtick(void)
465 {
466 }
467 #endif /* CONFIG_SMP */
468 
init_rq_hrtick(struct rq * rq)469 static void init_rq_hrtick(struct rq *rq)
470 {
471 #ifdef CONFIG_SMP
472 	rq->hrtick_csd_pending = 0;
473 
474 	rq->hrtick_csd.flags = 0;
475 	rq->hrtick_csd.func = __hrtick_start;
476 	rq->hrtick_csd.info = rq;
477 #endif
478 
479 	hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
480 	rq->hrtick_timer.function = hrtick;
481 }
482 #else	/* CONFIG_SCHED_HRTICK */
hrtick_clear(struct rq * rq)483 static inline void hrtick_clear(struct rq *rq)
484 {
485 }
486 
init_rq_hrtick(struct rq * rq)487 static inline void init_rq_hrtick(struct rq *rq)
488 {
489 }
490 
init_hrtick(void)491 static inline void init_hrtick(void)
492 {
493 }
494 #endif	/* CONFIG_SCHED_HRTICK */
495 
496 /*
497  * resched_task - mark a task 'to be rescheduled now'.
498  *
499  * On UP this means the setting of the need_resched flag, on SMP it
500  * might also involve a cross-CPU call to trigger the scheduler on
501  * the target CPU.
502  */
503 #ifdef CONFIG_SMP
504 
505 #ifndef tsk_is_polling
506 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
507 #endif
508 
resched_task(struct task_struct * p)509 void resched_task(struct task_struct *p)
510 {
511 	int cpu;
512 
513 	assert_raw_spin_locked(&task_rq(p)->lock);
514 
515 	if (test_tsk_need_resched(p))
516 		return;
517 
518 	set_tsk_need_resched(p);
519 
520 	cpu = task_cpu(p);
521 	if (cpu == smp_processor_id())
522 		return;
523 
524 	/* NEED_RESCHED must be visible before we test polling */
525 	smp_mb();
526 	if (!tsk_is_polling(p))
527 		smp_send_reschedule(cpu);
528 }
529 
resched_cpu(int cpu)530 void resched_cpu(int cpu)
531 {
532 	struct rq *rq = cpu_rq(cpu);
533 	unsigned long flags;
534 
535 	if (!raw_spin_trylock_irqsave(&rq->lock, flags))
536 		return;
537 	resched_task(cpu_curr(cpu));
538 	raw_spin_unlock_irqrestore(&rq->lock, flags);
539 }
540 
541 #ifdef CONFIG_NO_HZ
542 /*
543  * In the semi idle case, use the nearest busy cpu for migrating timers
544  * from an idle cpu.  This is good for power-savings.
545  *
546  * We don't do similar optimization for completely idle system, as
547  * selecting an idle cpu will add more delays to the timers than intended
548  * (as that cpu's timer base may not be uptodate wrt jiffies etc).
549  */
get_nohz_timer_target(void)550 int get_nohz_timer_target(void)
551 {
552 	int cpu = smp_processor_id();
553 	int i;
554 	struct sched_domain *sd;
555 
556 	rcu_read_lock();
557 	for_each_domain(cpu, sd) {
558 		for_each_cpu(i, sched_domain_span(sd)) {
559 			if (!idle_cpu(i)) {
560 				cpu = i;
561 				goto unlock;
562 			}
563 		}
564 	}
565 unlock:
566 	rcu_read_unlock();
567 	return cpu;
568 }
569 /*
570  * When add_timer_on() enqueues a timer into the timer wheel of an
571  * idle CPU then this timer might expire before the next timer event
572  * which is scheduled to wake up that CPU. In case of a completely
573  * idle system the next event might even be infinite time into the
574  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
575  * leaves the inner idle loop so the newly added timer is taken into
576  * account when the CPU goes back to idle and evaluates the timer
577  * wheel for the next timer event.
578  */
wake_up_idle_cpu(int cpu)579 void wake_up_idle_cpu(int cpu)
580 {
581 	struct rq *rq = cpu_rq(cpu);
582 
583 	if (cpu == smp_processor_id())
584 		return;
585 
586 	/*
587 	 * This is safe, as this function is called with the timer
588 	 * wheel base lock of (cpu) held. When the CPU is on the way
589 	 * to idle and has not yet set rq->curr to idle then it will
590 	 * be serialized on the timer wheel base lock and take the new
591 	 * timer into account automatically.
592 	 */
593 	if (rq->curr != rq->idle)
594 		return;
595 
596 	/*
597 	 * We can set TIF_RESCHED on the idle task of the other CPU
598 	 * lockless. The worst case is that the other CPU runs the
599 	 * idle task through an additional NOOP schedule()
600 	 */
601 	set_tsk_need_resched(rq->idle);
602 
603 	/* NEED_RESCHED must be visible before we test polling */
604 	smp_mb();
605 	if (!tsk_is_polling(rq->idle))
606 		smp_send_reschedule(cpu);
607 }
608 
got_nohz_idle_kick(void)609 static inline bool got_nohz_idle_kick(void)
610 {
611 	int cpu = smp_processor_id();
612 	return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
613 }
614 
615 #else /* CONFIG_NO_HZ */
616 
got_nohz_idle_kick(void)617 static inline bool got_nohz_idle_kick(void)
618 {
619 	return false;
620 }
621 
622 #endif /* CONFIG_NO_HZ */
623 
sched_avg_update(struct rq * rq)624 void sched_avg_update(struct rq *rq)
625 {
626 	s64 period = sched_avg_period();
627 
628 	while ((s64)(rq->clock - rq->age_stamp) > period) {
629 		/*
630 		 * Inline assembly required to prevent the compiler
631 		 * optimising this loop into a divmod call.
632 		 * See __iter_div_u64_rem() for another example of this.
633 		 */
634 		asm("" : "+rm" (rq->age_stamp));
635 		rq->age_stamp += period;
636 		rq->rt_avg /= 2;
637 	}
638 }
639 
640 #else /* !CONFIG_SMP */
resched_task(struct task_struct * p)641 void resched_task(struct task_struct *p)
642 {
643 	assert_raw_spin_locked(&task_rq(p)->lock);
644 	set_tsk_need_resched(p);
645 }
646 #endif /* CONFIG_SMP */
647 
648 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
649 			(defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
650 /*
651  * Iterate task_group tree rooted at *from, calling @down when first entering a
652  * node and @up when leaving it for the final time.
653  *
654  * Caller must hold rcu_lock or sufficient equivalent.
655  */
walk_tg_tree_from(struct task_group * from,tg_visitor down,tg_visitor up,void * data)656 int walk_tg_tree_from(struct task_group *from,
657 			     tg_visitor down, tg_visitor up, void *data)
658 {
659 	struct task_group *parent, *child;
660 	int ret;
661 
662 	parent = from;
663 
664 down:
665 	ret = (*down)(parent, data);
666 	if (ret)
667 		goto out;
668 	list_for_each_entry_rcu(child, &parent->children, siblings) {
669 		parent = child;
670 		goto down;
671 
672 up:
673 		continue;
674 	}
675 	ret = (*up)(parent, data);
676 	if (ret || parent == from)
677 		goto out;
678 
679 	child = parent;
680 	parent = parent->parent;
681 	if (parent)
682 		goto up;
683 out:
684 	return ret;
685 }
686 
tg_nop(struct task_group * tg,void * data)687 int tg_nop(struct task_group *tg, void *data)
688 {
689 	return 0;
690 }
691 #endif
692 
693 void update_cpu_load(struct rq *this_rq);
694 
set_load_weight(struct task_struct * p)695 static 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 
enqueue_task(struct rq * rq,struct task_struct * p,int flags)713 static 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 
dequeue_task(struct rq * rq,struct task_struct * p,int flags)720 static 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 
activate_task(struct rq * rq,struct task_struct * p,int flags)727 void 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 
deactivate_task(struct rq * rq,struct task_struct * p,int flags)735 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
736 {
737 	if (task_contributes_to_load(p))
738 		rq->nr_uninterruptible++;
739 
740 	dequeue_task(rq, p, flags);
741 }
742 
743 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
744 
745 /*
746  * There are no locks covering percpu hardirq/softirq time.
747  * They are only modified in account_system_vtime, on corresponding CPU
748  * with interrupts disabled. So, writes are safe.
749  * They are read and saved off onto struct rq in update_rq_clock().
750  * This may result in other CPU reading this CPU's irq time and can
751  * race with irq/account_system_vtime on this CPU. We would either get old
752  * or new value with a side effect of accounting a slice of irq time to wrong
753  * task when irq is in progress while we read rq->clock. That is a worthy
754  * compromise in place of having locks on each irq in account_system_time.
755  */
756 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
757 static DEFINE_PER_CPU(u64, cpu_softirq_time);
758 
759 static DEFINE_PER_CPU(u64, irq_start_time);
760 static int sched_clock_irqtime;
761 
enable_sched_clock_irqtime(void)762 void enable_sched_clock_irqtime(void)
763 {
764 	sched_clock_irqtime = 1;
765 }
766 
disable_sched_clock_irqtime(void)767 void disable_sched_clock_irqtime(void)
768 {
769 	sched_clock_irqtime = 0;
770 }
771 
772 #ifndef CONFIG_64BIT
773 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
774 
irq_time_write_begin(void)775 static inline void irq_time_write_begin(void)
776 {
777 	__this_cpu_inc(irq_time_seq.sequence);
778 	smp_wmb();
779 }
780 
irq_time_write_end(void)781 static inline void irq_time_write_end(void)
782 {
783 	smp_wmb();
784 	__this_cpu_inc(irq_time_seq.sequence);
785 }
786 
irq_time_read(int cpu)787 static inline u64 irq_time_read(int cpu)
788 {
789 	u64 irq_time;
790 	unsigned seq;
791 
792 	do {
793 		seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
794 		irq_time = per_cpu(cpu_softirq_time, cpu) +
795 			   per_cpu(cpu_hardirq_time, cpu);
796 	} while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
797 
798 	return irq_time;
799 }
800 #else /* CONFIG_64BIT */
irq_time_write_begin(void)801 static inline void irq_time_write_begin(void)
802 {
803 }
804 
irq_time_write_end(void)805 static inline void irq_time_write_end(void)
806 {
807 }
808 
irq_time_read(int cpu)809 static inline u64 irq_time_read(int cpu)
810 {
811 	return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
812 }
813 #endif /* CONFIG_64BIT */
814 
815 /*
816  * Called before incrementing preempt_count on {soft,}irq_enter
817  * and before decrementing preempt_count on {soft,}irq_exit.
818  */
account_system_vtime(struct task_struct * curr)819 void account_system_vtime(struct task_struct *curr)
820 {
821 	unsigned long flags;
822 	s64 delta;
823 	int cpu;
824 
825 	if (!sched_clock_irqtime)
826 		return;
827 
828 	local_irq_save(flags);
829 
830 	cpu = smp_processor_id();
831 	delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
832 	__this_cpu_add(irq_start_time, delta);
833 
834 	irq_time_write_begin();
835 	/*
836 	 * We do not account for softirq time from ksoftirqd here.
837 	 * We want to continue accounting softirq time to ksoftirqd thread
838 	 * in that case, so as not to confuse scheduler with a special task
839 	 * that do not consume any time, but still wants to run.
840 	 */
841 	if (hardirq_count())
842 		__this_cpu_add(cpu_hardirq_time, delta);
843 	else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
844 		__this_cpu_add(cpu_softirq_time, delta);
845 
846 	irq_time_write_end();
847 	local_irq_restore(flags);
848 }
849 EXPORT_SYMBOL_GPL(account_system_vtime);
850 
851 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
852 
853 #ifdef CONFIG_PARAVIRT
steal_ticks(u64 steal)854 static inline u64 steal_ticks(u64 steal)
855 {
856 	if (unlikely(steal > NSEC_PER_SEC))
857 		return div_u64(steal, TICK_NSEC);
858 
859 	return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
860 }
861 #endif
862 
update_rq_clock_task(struct rq * rq,s64 delta)863 static void update_rq_clock_task(struct rq *rq, s64 delta)
864 {
865 /*
866  * In theory, the compile should just see 0 here, and optimize out the call
867  * to sched_rt_avg_update. But I don't trust it...
868  */
869 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870 	s64 steal = 0, irq_delta = 0;
871 #endif
872 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
873 	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
874 
875 	/*
876 	 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877 	 * this case when a previous update_rq_clock() happened inside a
878 	 * {soft,}irq region.
879 	 *
880 	 * When this happens, we stop ->clock_task and only update the
881 	 * prev_irq_time stamp to account for the part that fit, so that a next
882 	 * update will consume the rest. This ensures ->clock_task is
883 	 * monotonic.
884 	 *
885 	 * It does however cause some slight miss-attribution of {soft,}irq
886 	 * time, a more accurate solution would be to update the irq_time using
887 	 * the current rq->clock timestamp, except that would require using
888 	 * atomic ops.
889 	 */
890 	if (irq_delta > delta)
891 		irq_delta = delta;
892 
893 	rq->prev_irq_time += irq_delta;
894 	delta -= irq_delta;
895 #endif
896 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897 	if (static_branch((&paravirt_steal_rq_enabled))) {
898 		u64 st;
899 
900 		steal = paravirt_steal_clock(cpu_of(rq));
901 		steal -= rq->prev_steal_time_rq;
902 
903 		if (unlikely(steal > delta))
904 			steal = delta;
905 
906 		st = steal_ticks(steal);
907 		steal = st * TICK_NSEC;
908 
909 		rq->prev_steal_time_rq += steal;
910 
911 		delta -= steal;
912 	}
913 #endif
914 
915 	rq->clock_task += delta;
916 
917 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
918 	if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
919 		sched_rt_avg_update(rq, irq_delta + steal);
920 #endif
921 }
922 
923 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
irqtime_account_hi_update(void)924 static int irqtime_account_hi_update(void)
925 {
926 	u64 *cpustat = kcpustat_this_cpu->cpustat;
927 	unsigned long flags;
928 	u64 latest_ns;
929 	int ret = 0;
930 
931 	local_irq_save(flags);
932 	latest_ns = this_cpu_read(cpu_hardirq_time);
933 	if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
934 		ret = 1;
935 	local_irq_restore(flags);
936 	return ret;
937 }
938 
irqtime_account_si_update(void)939 static int irqtime_account_si_update(void)
940 {
941 	u64 *cpustat = kcpustat_this_cpu->cpustat;
942 	unsigned long flags;
943 	u64 latest_ns;
944 	int ret = 0;
945 
946 	local_irq_save(flags);
947 	latest_ns = this_cpu_read(cpu_softirq_time);
948 	if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
949 		ret = 1;
950 	local_irq_restore(flags);
951 	return ret;
952 }
953 
954 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
955 
956 #define sched_clock_irqtime	(0)
957 
958 #endif
959 
sched_set_stop_task(int cpu,struct task_struct * stop)960 void sched_set_stop_task(int cpu, struct task_struct *stop)
961 {
962 	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
963 	struct task_struct *old_stop = cpu_rq(cpu)->stop;
964 
965 	if (stop) {
966 		/*
967 		 * Make it appear like a SCHED_FIFO task, its something
968 		 * userspace knows about and won't get confused about.
969 		 *
970 		 * Also, it will make PI more or less work without too
971 		 * much confusion -- but then, stop work should not
972 		 * rely on PI working anyway.
973 		 */
974 		sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
975 
976 		stop->sched_class = &stop_sched_class;
977 	}
978 
979 	cpu_rq(cpu)->stop = stop;
980 
981 	if (old_stop) {
982 		/*
983 		 * Reset it back to a normal scheduling class so that
984 		 * it can die in pieces.
985 		 */
986 		old_stop->sched_class = &rt_sched_class;
987 	}
988 }
989 
990 /*
991  * __normal_prio - return the priority that is based on the static prio
992  */
__normal_prio(struct task_struct * p)993 static inline int __normal_prio(struct task_struct *p)
994 {
995 	return p->static_prio;
996 }
997 
998 /*
999  * Calculate the expected normal priority: i.e. priority
1000  * without taking RT-inheritance into account. Might be
1001  * boosted by interactivity modifiers. Changes upon fork,
1002  * setprio syscalls, and whenever the interactivity
1003  * estimator recalculates.
1004  */
normal_prio(struct task_struct * p)1005 static inline int normal_prio(struct task_struct *p)
1006 {
1007 	int prio;
1008 
1009 	if (task_has_rt_policy(p))
1010 		prio = MAX_RT_PRIO-1 - p->rt_priority;
1011 	else
1012 		prio = __normal_prio(p);
1013 	return prio;
1014 }
1015 
1016 /*
1017  * Calculate the current priority, i.e. the priority
1018  * taken into account by the scheduler. This value might
1019  * be boosted by RT tasks, or might be boosted by
1020  * interactivity modifiers. Will be RT if the task got
1021  * RT-boosted. If not then it returns p->normal_prio.
1022  */
effective_prio(struct task_struct * p)1023 static int effective_prio(struct task_struct *p)
1024 {
1025 	p->normal_prio = normal_prio(p);
1026 	/*
1027 	 * If we are RT tasks or we were boosted to RT priority,
1028 	 * keep the priority unchanged. Otherwise, update priority
1029 	 * to the normal priority:
1030 	 */
1031 	if (!rt_prio(p->prio))
1032 		return p->normal_prio;
1033 	return p->prio;
1034 }
1035 
1036 /**
1037  * task_curr - is this task currently executing on a CPU?
1038  * @p: the task in question.
1039  */
task_curr(const struct task_struct * p)1040 inline int task_curr(const struct task_struct *p)
1041 {
1042 	return cpu_curr(task_cpu(p)) == p;
1043 }
1044 
check_class_changed(struct rq * rq,struct task_struct * p,const struct sched_class * prev_class,int oldprio)1045 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1046 				       const struct sched_class *prev_class,
1047 				       int oldprio)
1048 {
1049 	if (prev_class != p->sched_class) {
1050 		if (prev_class->switched_from)
1051 			prev_class->switched_from(rq, p);
1052 		p->sched_class->switched_to(rq, p);
1053 	} else if (oldprio != p->prio)
1054 		p->sched_class->prio_changed(rq, p, oldprio);
1055 }
1056 
check_preempt_curr(struct rq * rq,struct task_struct * p,int flags)1057 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1058 {
1059 	const struct sched_class *class;
1060 
1061 	if (p->sched_class == rq->curr->sched_class) {
1062 		rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1063 	} else {
1064 		for_each_class(class) {
1065 			if (class == rq->curr->sched_class)
1066 				break;
1067 			if (class == p->sched_class) {
1068 				resched_task(rq->curr);
1069 				break;
1070 			}
1071 		}
1072 	}
1073 
1074 	/*
1075 	 * A queue event has occurred, and we're going to schedule.  In
1076 	 * this case, we can save a useless back to back clock update.
1077 	 */
1078 	if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1079 		rq->skip_clock_update = 1;
1080 }
1081 
1082 #ifdef CONFIG_SMP
set_task_cpu(struct task_struct * p,unsigned int new_cpu)1083 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1084 {
1085 #ifdef CONFIG_SCHED_DEBUG
1086 	/*
1087 	 * We should never call set_task_cpu() on a blocked task,
1088 	 * ttwu() will sort out the placement.
1089 	 */
1090 	WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1091 			!(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1092 
1093 #ifdef CONFIG_LOCKDEP
1094 	/*
1095 	 * The caller should hold either p->pi_lock or rq->lock, when changing
1096 	 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1097 	 *
1098 	 * sched_move_task() holds both and thus holding either pins the cgroup,
1099 	 * see set_task_rq().
1100 	 *
1101 	 * Furthermore, all task_rq users should acquire both locks, see
1102 	 * task_rq_lock().
1103 	 */
1104 	WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1105 				      lockdep_is_held(&task_rq(p)->lock)));
1106 #endif
1107 #endif
1108 
1109 	trace_sched_migrate_task(p, new_cpu);
1110 
1111 	if (task_cpu(p) != new_cpu) {
1112 		p->se.nr_migrations++;
1113 		perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1114 	}
1115 
1116 	__set_task_cpu(p, new_cpu);
1117 }
1118 
1119 struct migration_arg {
1120 	struct task_struct *task;
1121 	int dest_cpu;
1122 };
1123 
1124 static int migration_cpu_stop(void *data);
1125 
1126 /*
1127  * wait_task_inactive - wait for a thread to unschedule.
1128  *
1129  * If @match_state is nonzero, it's the @p->state value just checked and
1130  * not expected to change.  If it changes, i.e. @p might have woken up,
1131  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1132  * we return a positive number (its total switch count).  If a second call
1133  * a short while later returns the same number, the caller can be sure that
1134  * @p has remained unscheduled the whole time.
1135  *
1136  * The caller must ensure that the task *will* unschedule sometime soon,
1137  * else this function might spin for a *long* time. This function can't
1138  * be called with interrupts off, or it may introduce deadlock with
1139  * smp_call_function() if an IPI is sent by the same process we are
1140  * waiting to become inactive.
1141  */
wait_task_inactive(struct task_struct * p,long match_state)1142 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1143 {
1144 	unsigned long flags;
1145 	int running, on_rq;
1146 	unsigned long ncsw;
1147 	struct rq *rq;
1148 
1149 	for (;;) {
1150 		/*
1151 		 * We do the initial early heuristics without holding
1152 		 * any task-queue locks at all. We'll only try to get
1153 		 * the runqueue lock when things look like they will
1154 		 * work out!
1155 		 */
1156 		rq = task_rq(p);
1157 
1158 		/*
1159 		 * If the task is actively running on another CPU
1160 		 * still, just relax and busy-wait without holding
1161 		 * any locks.
1162 		 *
1163 		 * NOTE! Since we don't hold any locks, it's not
1164 		 * even sure that "rq" stays as the right runqueue!
1165 		 * But we don't care, since "task_running()" will
1166 		 * return false if the runqueue has changed and p
1167 		 * is actually now running somewhere else!
1168 		 */
1169 		while (task_running(rq, p)) {
1170 			if (match_state && unlikely(p->state != match_state))
1171 				return 0;
1172 			cpu_relax();
1173 		}
1174 
1175 		/*
1176 		 * Ok, time to look more closely! We need the rq
1177 		 * lock now, to be *sure*. If we're wrong, we'll
1178 		 * just go back and repeat.
1179 		 */
1180 		rq = task_rq_lock(p, &flags);
1181 		trace_sched_wait_task(p);
1182 		running = task_running(rq, p);
1183 		on_rq = p->on_rq;
1184 		ncsw = 0;
1185 		if (!match_state || p->state == match_state)
1186 			ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1187 		task_rq_unlock(rq, p, &flags);
1188 
1189 		/*
1190 		 * If it changed from the expected state, bail out now.
1191 		 */
1192 		if (unlikely(!ncsw))
1193 			break;
1194 
1195 		/*
1196 		 * Was it really running after all now that we
1197 		 * checked with the proper locks actually held?
1198 		 *
1199 		 * Oops. Go back and try again..
1200 		 */
1201 		if (unlikely(running)) {
1202 			cpu_relax();
1203 			continue;
1204 		}
1205 
1206 		/*
1207 		 * It's not enough that it's not actively running,
1208 		 * it must be off the runqueue _entirely_, and not
1209 		 * preempted!
1210 		 *
1211 		 * So if it was still runnable (but just not actively
1212 		 * running right now), it's preempted, and we should
1213 		 * yield - it could be a while.
1214 		 */
1215 		if (unlikely(on_rq)) {
1216 			ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1217 
1218 			set_current_state(TASK_UNINTERRUPTIBLE);
1219 			schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1220 			continue;
1221 		}
1222 
1223 		/*
1224 		 * Ahh, all good. It wasn't running, and it wasn't
1225 		 * runnable, which means that it will never become
1226 		 * running in the future either. We're all done!
1227 		 */
1228 		break;
1229 	}
1230 
1231 	return ncsw;
1232 }
1233 
1234 /***
1235  * kick_process - kick a running thread to enter/exit the kernel
1236  * @p: the to-be-kicked thread
1237  *
1238  * Cause a process which is running on another CPU to enter
1239  * kernel-mode, without any delay. (to get signals handled.)
1240  *
1241  * NOTE: this function doesn't have to take the runqueue lock,
1242  * because all it wants to ensure is that the remote task enters
1243  * the kernel. If the IPI races and the task has been migrated
1244  * to another CPU then no harm is done and the purpose has been
1245  * achieved as well.
1246  */
kick_process(struct task_struct * p)1247 void kick_process(struct task_struct *p)
1248 {
1249 	int cpu;
1250 
1251 	preempt_disable();
1252 	cpu = task_cpu(p);
1253 	if ((cpu != smp_processor_id()) && task_curr(p))
1254 		smp_send_reschedule(cpu);
1255 	preempt_enable();
1256 }
1257 EXPORT_SYMBOL_GPL(kick_process);
1258 #endif /* CONFIG_SMP */
1259 
1260 #ifdef CONFIG_SMP
1261 /*
1262  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1263  */
select_fallback_rq(int cpu,struct task_struct * p)1264 static int select_fallback_rq(int cpu, struct task_struct *p)
1265 {
1266 	int dest_cpu;
1267 	const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1268 
1269 	/* Look for allowed, online CPU in same node. */
1270 	for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
1271 		if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1272 			return dest_cpu;
1273 
1274 	/* Any allowed, online CPU? */
1275 	dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
1276 	if (dest_cpu < nr_cpu_ids)
1277 		return dest_cpu;
1278 
1279 	/* No more Mr. Nice Guy. */
1280 	dest_cpu = cpuset_cpus_allowed_fallback(p);
1281 	/*
1282 	 * Don't tell them about moving exiting tasks or
1283 	 * kernel threads (both mm NULL), since they never
1284 	 * leave kernel.
1285 	 */
1286 	if (p->mm && printk_ratelimit()) {
1287 		printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
1288 				task_pid_nr(p), p->comm, cpu);
1289 	}
1290 
1291 	return dest_cpu;
1292 }
1293 
1294 /*
1295  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1296  */
1297 static inline
select_task_rq(struct task_struct * p,int sd_flags,int wake_flags)1298 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1299 {
1300 	int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1301 
1302 	/*
1303 	 * In order not to call set_task_cpu() on a blocking task we need
1304 	 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1305 	 * cpu.
1306 	 *
1307 	 * Since this is common to all placement strategies, this lives here.
1308 	 *
1309 	 * [ this allows ->select_task() to simply return task_cpu(p) and
1310 	 *   not worry about this generic constraint ]
1311 	 */
1312 	if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1313 		     !cpu_online(cpu)))
1314 		cpu = select_fallback_rq(task_cpu(p), p);
1315 
1316 	return cpu;
1317 }
1318 
update_avg(u64 * avg,u64 sample)1319 static void update_avg(u64 *avg, u64 sample)
1320 {
1321 	s64 diff = sample - *avg;
1322 	*avg += diff >> 3;
1323 }
1324 #endif
1325 
1326 static void
ttwu_stat(struct task_struct * p,int cpu,int wake_flags)1327 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1328 {
1329 #ifdef CONFIG_SCHEDSTATS
1330 	struct rq *rq = this_rq();
1331 
1332 #ifdef CONFIG_SMP
1333 	int this_cpu = smp_processor_id();
1334 
1335 	if (cpu == this_cpu) {
1336 		schedstat_inc(rq, ttwu_local);
1337 		schedstat_inc(p, se.statistics.nr_wakeups_local);
1338 	} else {
1339 		struct sched_domain *sd;
1340 
1341 		schedstat_inc(p, se.statistics.nr_wakeups_remote);
1342 		rcu_read_lock();
1343 		for_each_domain(this_cpu, sd) {
1344 			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1345 				schedstat_inc(sd, ttwu_wake_remote);
1346 				break;
1347 			}
1348 		}
1349 		rcu_read_unlock();
1350 	}
1351 
1352 	if (wake_flags & WF_MIGRATED)
1353 		schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1354 
1355 #endif /* CONFIG_SMP */
1356 
1357 	schedstat_inc(rq, ttwu_count);
1358 	schedstat_inc(p, se.statistics.nr_wakeups);
1359 
1360 	if (wake_flags & WF_SYNC)
1361 		schedstat_inc(p, se.statistics.nr_wakeups_sync);
1362 
1363 #endif /* CONFIG_SCHEDSTATS */
1364 }
1365 
ttwu_activate(struct rq * rq,struct task_struct * p,int en_flags)1366 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1367 {
1368 	activate_task(rq, p, en_flags);
1369 	p->on_rq = 1;
1370 
1371 	/* if a worker is waking up, notify workqueue */
1372 	if (p->flags & PF_WQ_WORKER)
1373 		wq_worker_waking_up(p, cpu_of(rq));
1374 }
1375 
1376 /*
1377  * Mark the task runnable and perform wakeup-preemption.
1378  */
1379 static void
ttwu_do_wakeup(struct rq * rq,struct task_struct * p,int wake_flags)1380 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1381 {
1382 	trace_sched_wakeup(p, true);
1383 	check_preempt_curr(rq, p, wake_flags);
1384 
1385 	p->state = TASK_RUNNING;
1386 #ifdef CONFIG_SMP
1387 	if (p->sched_class->task_woken)
1388 		p->sched_class->task_woken(rq, p);
1389 
1390 	if (rq->idle_stamp) {
1391 		u64 delta = rq->clock - rq->idle_stamp;
1392 		u64 max = 2*sysctl_sched_migration_cost;
1393 
1394 		if (delta > max)
1395 			rq->avg_idle = max;
1396 		else
1397 			update_avg(&rq->avg_idle, delta);
1398 		rq->idle_stamp = 0;
1399 	}
1400 #endif
1401 }
1402 
1403 static void
ttwu_do_activate(struct rq * rq,struct task_struct * p,int wake_flags)1404 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1405 {
1406 #ifdef CONFIG_SMP
1407 	if (p->sched_contributes_to_load)
1408 		rq->nr_uninterruptible--;
1409 #endif
1410 
1411 	ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1412 	ttwu_do_wakeup(rq, p, wake_flags);
1413 }
1414 
1415 /*
1416  * Called in case the task @p isn't fully descheduled from its runqueue,
1417  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1418  * since all we need to do is flip p->state to TASK_RUNNING, since
1419  * the task is still ->on_rq.
1420  */
ttwu_remote(struct task_struct * p,int wake_flags)1421 static int ttwu_remote(struct task_struct *p, int wake_flags)
1422 {
1423 	struct rq *rq;
1424 	int ret = 0;
1425 
1426 	rq = __task_rq_lock(p);
1427 	if (p->on_rq) {
1428 		ttwu_do_wakeup(rq, p, wake_flags);
1429 		ret = 1;
1430 	}
1431 	__task_rq_unlock(rq);
1432 
1433 	return ret;
1434 }
1435 
1436 #ifdef CONFIG_SMP
sched_ttwu_pending(void)1437 static void sched_ttwu_pending(void)
1438 {
1439 	struct rq *rq = this_rq();
1440 	struct llist_node *llist = llist_del_all(&rq->wake_list);
1441 	struct task_struct *p;
1442 
1443 	raw_spin_lock(&rq->lock);
1444 
1445 	while (llist) {
1446 		p = llist_entry(llist, struct task_struct, wake_entry);
1447 		llist = llist_next(llist);
1448 		ttwu_do_activate(rq, p, 0);
1449 	}
1450 
1451 	raw_spin_unlock(&rq->lock);
1452 }
1453 
scheduler_ipi(void)1454 void scheduler_ipi(void)
1455 {
1456 	if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1457 		return;
1458 
1459 	/*
1460 	 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1461 	 * traditionally all their work was done from the interrupt return
1462 	 * path. Now that we actually do some work, we need to make sure
1463 	 * we do call them.
1464 	 *
1465 	 * Some archs already do call them, luckily irq_enter/exit nest
1466 	 * properly.
1467 	 *
1468 	 * Arguably we should visit all archs and update all handlers,
1469 	 * however a fair share of IPIs are still resched only so this would
1470 	 * somewhat pessimize the simple resched case.
1471 	 */
1472 	irq_enter();
1473 	sched_ttwu_pending();
1474 
1475 	/*
1476 	 * Check if someone kicked us for doing the nohz idle load balance.
1477 	 */
1478 	if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1479 		this_rq()->idle_balance = 1;
1480 		raise_softirq_irqoff(SCHED_SOFTIRQ);
1481 	}
1482 	irq_exit();
1483 }
1484 
ttwu_queue_remote(struct task_struct * p,int cpu)1485 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1486 {
1487 	if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1488 		smp_send_reschedule(cpu);
1489 }
1490 
1491 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
ttwu_activate_remote(struct task_struct * p,int wake_flags)1492 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1493 {
1494 	struct rq *rq;
1495 	int ret = 0;
1496 
1497 	rq = __task_rq_lock(p);
1498 	if (p->on_cpu) {
1499 		ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1500 		ttwu_do_wakeup(rq, p, wake_flags);
1501 		ret = 1;
1502 	}
1503 	__task_rq_unlock(rq);
1504 
1505 	return ret;
1506 
1507 }
1508 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1509 
ttwu_share_cache(int this_cpu,int that_cpu)1510 static inline int ttwu_share_cache(int this_cpu, int that_cpu)
1511 {
1512 	return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1513 }
1514 #endif /* CONFIG_SMP */
1515 
ttwu_queue(struct task_struct * p,int cpu)1516 static void ttwu_queue(struct task_struct *p, int cpu)
1517 {
1518 	struct rq *rq = cpu_rq(cpu);
1519 
1520 #if defined(CONFIG_SMP)
1521 	if (sched_feat(TTWU_QUEUE) && !ttwu_share_cache(smp_processor_id(), cpu)) {
1522 		sched_clock_cpu(cpu); /* sync clocks x-cpu */
1523 		ttwu_queue_remote(p, cpu);
1524 		return;
1525 	}
1526 #endif
1527 
1528 	raw_spin_lock(&rq->lock);
1529 	ttwu_do_activate(rq, p, 0);
1530 	raw_spin_unlock(&rq->lock);
1531 }
1532 
1533 /**
1534  * try_to_wake_up - wake up a thread
1535  * @p: the thread to be awakened
1536  * @state: the mask of task states that can be woken
1537  * @wake_flags: wake modifier flags (WF_*)
1538  *
1539  * Put it on the run-queue if it's not already there. The "current"
1540  * thread is always on the run-queue (except when the actual
1541  * re-schedule is in progress), and as such you're allowed to do
1542  * the simpler "current->state = TASK_RUNNING" to mark yourself
1543  * runnable without the overhead of this.
1544  *
1545  * Returns %true if @p was woken up, %false if it was already running
1546  * or @state didn't match @p's state.
1547  */
1548 static int
try_to_wake_up(struct task_struct * p,unsigned int state,int wake_flags)1549 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1550 {
1551 	unsigned long flags;
1552 	int cpu, success = 0;
1553 
1554 	smp_wmb();
1555 	raw_spin_lock_irqsave(&p->pi_lock, flags);
1556 	if (!(p->state & state))
1557 		goto out;
1558 
1559 	success = 1; /* we're going to change ->state */
1560 	cpu = task_cpu(p);
1561 
1562 	if (p->on_rq && ttwu_remote(p, wake_flags))
1563 		goto stat;
1564 
1565 #ifdef CONFIG_SMP
1566 	/*
1567 	 * If the owning (remote) cpu is still in the middle of schedule() with
1568 	 * this task as prev, wait until its done referencing the task.
1569 	 */
1570 	while (p->on_cpu) {
1571 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1572 		/*
1573 		 * In case the architecture enables interrupts in
1574 		 * context_switch(), we cannot busy wait, since that
1575 		 * would lead to deadlocks when an interrupt hits and
1576 		 * tries to wake up @prev. So bail and do a complete
1577 		 * remote wakeup.
1578 		 */
1579 		if (ttwu_activate_remote(p, wake_flags))
1580 			goto stat;
1581 #else
1582 		cpu_relax();
1583 #endif
1584 	}
1585 	/*
1586 	 * Pairs with the smp_wmb() in finish_lock_switch().
1587 	 */
1588 	smp_rmb();
1589 
1590 	p->sched_contributes_to_load = !!task_contributes_to_load(p);
1591 	p->state = TASK_WAKING;
1592 
1593 	if (p->sched_class->task_waking)
1594 		p->sched_class->task_waking(p);
1595 
1596 	cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1597 	if (task_cpu(p) != cpu) {
1598 		wake_flags |= WF_MIGRATED;
1599 		set_task_cpu(p, cpu);
1600 	}
1601 #endif /* CONFIG_SMP */
1602 
1603 	ttwu_queue(p, cpu);
1604 stat:
1605 	ttwu_stat(p, cpu, wake_flags);
1606 out:
1607 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1608 
1609 	return success;
1610 }
1611 
1612 /**
1613  * try_to_wake_up_local - try to wake up a local task with rq lock held
1614  * @p: the thread to be awakened
1615  *
1616  * Put @p on the run-queue if it's not already there. The caller must
1617  * ensure that this_rq() is locked, @p is bound to this_rq() and not
1618  * the current task.
1619  */
try_to_wake_up_local(struct task_struct * p)1620 static void try_to_wake_up_local(struct task_struct *p)
1621 {
1622 	struct rq *rq = task_rq(p);
1623 
1624 	BUG_ON(rq != this_rq());
1625 	BUG_ON(p == current);
1626 	lockdep_assert_held(&rq->lock);
1627 
1628 	if (!raw_spin_trylock(&p->pi_lock)) {
1629 		raw_spin_unlock(&rq->lock);
1630 		raw_spin_lock(&p->pi_lock);
1631 		raw_spin_lock(&rq->lock);
1632 	}
1633 
1634 	if (!(p->state & TASK_NORMAL))
1635 		goto out;
1636 
1637 	if (!p->on_rq)
1638 		ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1639 
1640 	ttwu_do_wakeup(rq, p, 0);
1641 	ttwu_stat(p, smp_processor_id(), 0);
1642 out:
1643 	raw_spin_unlock(&p->pi_lock);
1644 }
1645 
1646 /**
1647  * wake_up_process - Wake up a specific process
1648  * @p: The process to be woken up.
1649  *
1650  * Attempt to wake up the nominated process and move it to the set of runnable
1651  * processes.  Returns 1 if the process was woken up, 0 if it was already
1652  * running.
1653  *
1654  * It may be assumed that this function implies a write memory barrier before
1655  * changing the task state if and only if any tasks are woken up.
1656  */
wake_up_process(struct task_struct * p)1657 int wake_up_process(struct task_struct *p)
1658 {
1659 	return try_to_wake_up(p, TASK_ALL, 0);
1660 }
1661 EXPORT_SYMBOL(wake_up_process);
1662 
wake_up_state(struct task_struct * p,unsigned int state)1663 int wake_up_state(struct task_struct *p, unsigned int state)
1664 {
1665 	return try_to_wake_up(p, state, 0);
1666 }
1667 
1668 /*
1669  * Perform scheduler related setup for a newly forked process p.
1670  * p is forked by current.
1671  *
1672  * __sched_fork() is basic setup used by init_idle() too:
1673  */
__sched_fork(struct task_struct * p)1674 static void __sched_fork(struct task_struct *p)
1675 {
1676 	p->on_rq			= 0;
1677 
1678 	p->se.on_rq			= 0;
1679 	p->se.exec_start		= 0;
1680 	p->se.sum_exec_runtime		= 0;
1681 	p->se.prev_sum_exec_runtime	= 0;
1682 	p->se.nr_migrations		= 0;
1683 	p->se.vruntime			= 0;
1684 	INIT_LIST_HEAD(&p->se.group_node);
1685 
1686 #ifdef CONFIG_SCHEDSTATS
1687 	memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1688 #endif
1689 
1690 	INIT_LIST_HEAD(&p->rt.run_list);
1691 
1692 #ifdef CONFIG_PREEMPT_NOTIFIERS
1693 	INIT_HLIST_HEAD(&p->preempt_notifiers);
1694 #endif
1695 }
1696 
1697 /*
1698  * fork()/clone()-time setup:
1699  */
sched_fork(struct task_struct * p)1700 void sched_fork(struct task_struct *p)
1701 {
1702 	unsigned long flags;
1703 	int cpu = get_cpu();
1704 
1705 	__sched_fork(p);
1706 	/*
1707 	 * We mark the process as running here. This guarantees that
1708 	 * nobody will actually run it, and a signal or other external
1709 	 * event cannot wake it up and insert it on the runqueue either.
1710 	 */
1711 	p->state = TASK_RUNNING;
1712 
1713 	/*
1714 	 * Make sure we do not leak PI boosting priority to the child.
1715 	 */
1716 	p->prio = current->normal_prio;
1717 
1718 	/*
1719 	 * Revert to default priority/policy on fork if requested.
1720 	 */
1721 	if (unlikely(p->sched_reset_on_fork)) {
1722 		if (task_has_rt_policy(p)) {
1723 			p->policy = SCHED_NORMAL;
1724 			p->static_prio = NICE_TO_PRIO(0);
1725 			p->rt_priority = 0;
1726 		} else if (PRIO_TO_NICE(p->static_prio) < 0)
1727 			p->static_prio = NICE_TO_PRIO(0);
1728 
1729 		p->prio = p->normal_prio = __normal_prio(p);
1730 		set_load_weight(p);
1731 
1732 		/*
1733 		 * We don't need the reset flag anymore after the fork. It has
1734 		 * fulfilled its duty:
1735 		 */
1736 		p->sched_reset_on_fork = 0;
1737 	}
1738 
1739 	if (!rt_prio(p->prio))
1740 		p->sched_class = &fair_sched_class;
1741 
1742 	if (p->sched_class->task_fork)
1743 		p->sched_class->task_fork(p);
1744 
1745 	/*
1746 	 * The child is not yet in the pid-hash so no cgroup attach races,
1747 	 * and the cgroup is pinned to this child due to cgroup_fork()
1748 	 * is ran before sched_fork().
1749 	 *
1750 	 * Silence PROVE_RCU.
1751 	 */
1752 	raw_spin_lock_irqsave(&p->pi_lock, flags);
1753 	set_task_cpu(p, cpu);
1754 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1755 
1756 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1757 	if (likely(sched_info_on()))
1758 		memset(&p->sched_info, 0, sizeof(p->sched_info));
1759 #endif
1760 #if defined(CONFIG_SMP)
1761 	p->on_cpu = 0;
1762 #endif
1763 #ifdef CONFIG_PREEMPT_COUNT
1764 	/* Want to start with kernel preemption disabled. */
1765 	task_thread_info(p)->preempt_count = 1;
1766 #endif
1767 #ifdef CONFIG_SMP
1768 	plist_node_init(&p->pushable_tasks, MAX_PRIO);
1769 #endif
1770 
1771 	put_cpu();
1772 }
1773 
1774 /*
1775  * wake_up_new_task - wake up a newly created task for the first time.
1776  *
1777  * This function will do some initial scheduler statistics housekeeping
1778  * that must be done for every newly created context, then puts the task
1779  * on the runqueue and wakes it.
1780  */
wake_up_new_task(struct task_struct * p)1781 void wake_up_new_task(struct task_struct *p)
1782 {
1783 	unsigned long flags;
1784 	struct rq *rq;
1785 
1786 	raw_spin_lock_irqsave(&p->pi_lock, flags);
1787 #ifdef CONFIG_SMP
1788 	/*
1789 	 * Fork balancing, do it here and not earlier because:
1790 	 *  - cpus_allowed can change in the fork path
1791 	 *  - any previously selected cpu might disappear through hotplug
1792 	 */
1793 	set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1794 #endif
1795 
1796 	rq = __task_rq_lock(p);
1797 	activate_task(rq, p, 0);
1798 	p->on_rq = 1;
1799 	trace_sched_wakeup_new(p, true);
1800 	check_preempt_curr(rq, p, WF_FORK);
1801 #ifdef CONFIG_SMP
1802 	if (p->sched_class->task_woken)
1803 		p->sched_class->task_woken(rq, p);
1804 #endif
1805 	task_rq_unlock(rq, p, &flags);
1806 }
1807 
1808 #ifdef CONFIG_PREEMPT_NOTIFIERS
1809 
1810 /**
1811  * preempt_notifier_register - tell me when current is being preempted & rescheduled
1812  * @notifier: notifier struct to register
1813  */
preempt_notifier_register(struct preempt_notifier * notifier)1814 void preempt_notifier_register(struct preempt_notifier *notifier)
1815 {
1816 	hlist_add_head(&notifier->link, &current->preempt_notifiers);
1817 }
1818 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1819 
1820 /**
1821  * preempt_notifier_unregister - no longer interested in preemption notifications
1822  * @notifier: notifier struct to unregister
1823  *
1824  * This is safe to call from within a preemption notifier.
1825  */
preempt_notifier_unregister(struct preempt_notifier * notifier)1826 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1827 {
1828 	hlist_del(&notifier->link);
1829 }
1830 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1831 
fire_sched_in_preempt_notifiers(struct task_struct * curr)1832 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1833 {
1834 	struct preempt_notifier *notifier;
1835 	struct hlist_node *node;
1836 
1837 	hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1838 		notifier->ops->sched_in(notifier, raw_smp_processor_id());
1839 }
1840 
1841 static void
fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)1842 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1843 				 struct task_struct *next)
1844 {
1845 	struct preempt_notifier *notifier;
1846 	struct hlist_node *node;
1847 
1848 	hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1849 		notifier->ops->sched_out(notifier, next);
1850 }
1851 
1852 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1853 
fire_sched_in_preempt_notifiers(struct task_struct * curr)1854 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1855 {
1856 }
1857 
1858 static void
fire_sched_out_preempt_notifiers(struct task_struct * curr,struct task_struct * next)1859 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1860 				 struct task_struct *next)
1861 {
1862 }
1863 
1864 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1865 
1866 /**
1867  * prepare_task_switch - prepare to switch tasks
1868  * @rq: the runqueue preparing to switch
1869  * @prev: the current task that is being switched out
1870  * @next: the task we are going to switch to.
1871  *
1872  * This is called with the rq lock held and interrupts off. It must
1873  * be paired with a subsequent finish_task_switch after the context
1874  * switch.
1875  *
1876  * prepare_task_switch sets up locking and calls architecture specific
1877  * hooks.
1878  */
1879 static inline void
prepare_task_switch(struct rq * rq,struct task_struct * prev,struct task_struct * next)1880 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1881 		    struct task_struct *next)
1882 {
1883 	sched_info_switch(prev, next);
1884 	perf_event_task_sched_out(prev, next);
1885 	fire_sched_out_preempt_notifiers(prev, next);
1886 	prepare_lock_switch(rq, next);
1887 	prepare_arch_switch(next);
1888 	trace_sched_switch(prev, next);
1889 }
1890 
1891 /**
1892  * finish_task_switch - clean up after a task-switch
1893  * @rq: runqueue associated with task-switch
1894  * @prev: the thread we just switched away from.
1895  *
1896  * finish_task_switch must be called after the context switch, paired
1897  * with a prepare_task_switch call before the context switch.
1898  * finish_task_switch will reconcile locking set up by prepare_task_switch,
1899  * and do any other architecture-specific cleanup actions.
1900  *
1901  * Note that we may have delayed dropping an mm in context_switch(). If
1902  * so, we finish that here outside of the runqueue lock. (Doing it
1903  * with the lock held can cause deadlocks; see schedule() for
1904  * details.)
1905  */
finish_task_switch(struct rq * rq,struct task_struct * prev)1906 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1907 	__releases(rq->lock)
1908 {
1909 	struct mm_struct *mm = rq->prev_mm;
1910 	long prev_state;
1911 
1912 	rq->prev_mm = NULL;
1913 
1914 	/*
1915 	 * A task struct has one reference for the use as "current".
1916 	 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1917 	 * schedule one last time. The schedule call will never return, and
1918 	 * the scheduled task must drop that reference.
1919 	 * The test for TASK_DEAD must occur while the runqueue locks are
1920 	 * still held, otherwise prev could be scheduled on another cpu, die
1921 	 * there before we look at prev->state, and then the reference would
1922 	 * be dropped twice.
1923 	 *		Manfred Spraul <manfred@colorfullife.com>
1924 	 */
1925 	prev_state = prev->state;
1926 	finish_arch_switch(prev);
1927 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1928 	local_irq_disable();
1929 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1930 	perf_event_task_sched_in(prev, current);
1931 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1932 	local_irq_enable();
1933 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1934 	finish_lock_switch(rq, prev);
1935 
1936 	fire_sched_in_preempt_notifiers(current);
1937 	if (mm)
1938 		mmdrop(mm);
1939 	if (unlikely(prev_state == TASK_DEAD)) {
1940 		/*
1941 		 * Remove function-return probe instances associated with this
1942 		 * task and put them back on the free list.
1943 		 */
1944 		kprobe_flush_task(prev);
1945 		put_task_struct(prev);
1946 	}
1947 }
1948 
1949 #ifdef CONFIG_SMP
1950 
1951 /* assumes rq->lock is held */
pre_schedule(struct rq * rq,struct task_struct * prev)1952 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1953 {
1954 	if (prev->sched_class->pre_schedule)
1955 		prev->sched_class->pre_schedule(rq, prev);
1956 }
1957 
1958 /* rq->lock is NOT held, but preemption is disabled */
post_schedule(struct rq * rq)1959 static inline void post_schedule(struct rq *rq)
1960 {
1961 	if (rq->post_schedule) {
1962 		unsigned long flags;
1963 
1964 		raw_spin_lock_irqsave(&rq->lock, flags);
1965 		if (rq->curr->sched_class->post_schedule)
1966 			rq->curr->sched_class->post_schedule(rq);
1967 		raw_spin_unlock_irqrestore(&rq->lock, flags);
1968 
1969 		rq->post_schedule = 0;
1970 	}
1971 }
1972 
1973 #else
1974 
pre_schedule(struct rq * rq,struct task_struct * p)1975 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1976 {
1977 }
1978 
post_schedule(struct rq * rq)1979 static inline void post_schedule(struct rq *rq)
1980 {
1981 }
1982 
1983 #endif
1984 
1985 /**
1986  * schedule_tail - first thing a freshly forked thread must call.
1987  * @prev: the thread we just switched away from.
1988  */
schedule_tail(struct task_struct * prev)1989 asmlinkage void schedule_tail(struct task_struct *prev)
1990 	__releases(rq->lock)
1991 {
1992 	struct rq *rq = this_rq();
1993 
1994 	finish_task_switch(rq, prev);
1995 
1996 	/*
1997 	 * FIXME: do we need to worry about rq being invalidated by the
1998 	 * task_switch?
1999 	 */
2000 	post_schedule(rq);
2001 
2002 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2003 	/* In this case, finish_task_switch does not reenable preemption */
2004 	preempt_enable();
2005 #endif
2006 	if (current->set_child_tid)
2007 		put_user(task_pid_vnr(current), current->set_child_tid);
2008 }
2009 
2010 /*
2011  * context_switch - switch to the new MM and the new
2012  * thread's register state.
2013  */
2014 static inline void
context_switch(struct rq * rq,struct task_struct * prev,struct task_struct * next)2015 context_switch(struct rq *rq, struct task_struct *prev,
2016 	       struct task_struct *next)
2017 {
2018 	struct mm_struct *mm, *oldmm;
2019 
2020 	prepare_task_switch(rq, prev, next);
2021 
2022 	mm = next->mm;
2023 	oldmm = prev->active_mm;
2024 	/*
2025 	 * For paravirt, this is coupled with an exit in switch_to to
2026 	 * combine the page table reload and the switch backend into
2027 	 * one hypercall.
2028 	 */
2029 	arch_start_context_switch(prev);
2030 
2031 	if (!mm) {
2032 		next->active_mm = oldmm;
2033 		atomic_inc(&oldmm->mm_count);
2034 		enter_lazy_tlb(oldmm, next);
2035 	} else
2036 		switch_mm(oldmm, mm, next);
2037 
2038 	if (!prev->mm) {
2039 		prev->active_mm = NULL;
2040 		rq->prev_mm = oldmm;
2041 	}
2042 	/*
2043 	 * Since the runqueue lock will be released by the next
2044 	 * task (which is an invalid locking op but in the case
2045 	 * of the scheduler it's an obvious special-case), so we
2046 	 * do an early lockdep release here:
2047 	 */
2048 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2049 	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2050 #endif
2051 
2052 	/* Here we just switch the register state and the stack. */
2053 	switch_to(prev, next, prev);
2054 
2055 	barrier();
2056 	/*
2057 	 * this_rq must be evaluated again because prev may have moved
2058 	 * CPUs since it called schedule(), thus the 'rq' on its stack
2059 	 * frame will be invalid.
2060 	 */
2061 	finish_task_switch(this_rq(), prev);
2062 }
2063 
2064 /*
2065  * nr_running, nr_uninterruptible and nr_context_switches:
2066  *
2067  * externally visible scheduler statistics: current number of runnable
2068  * threads, current number of uninterruptible-sleeping threads, total
2069  * number of context switches performed since bootup.
2070  */
nr_running(void)2071 unsigned long nr_running(void)
2072 {
2073 	unsigned long i, sum = 0;
2074 
2075 	for_each_online_cpu(i)
2076 		sum += cpu_rq(i)->nr_running;
2077 
2078 	return sum;
2079 }
2080 
nr_uninterruptible(void)2081 unsigned long nr_uninterruptible(void)
2082 {
2083 	unsigned long i, sum = 0;
2084 
2085 	for_each_possible_cpu(i)
2086 		sum += cpu_rq(i)->nr_uninterruptible;
2087 
2088 	/*
2089 	 * Since we read the counters lockless, it might be slightly
2090 	 * inaccurate. Do not allow it to go below zero though:
2091 	 */
2092 	if (unlikely((long)sum < 0))
2093 		sum = 0;
2094 
2095 	return sum;
2096 }
2097 
nr_context_switches(void)2098 unsigned long long nr_context_switches(void)
2099 {
2100 	int i;
2101 	unsigned long long sum = 0;
2102 
2103 	for_each_possible_cpu(i)
2104 		sum += cpu_rq(i)->nr_switches;
2105 
2106 	return sum;
2107 }
2108 
nr_iowait(void)2109 unsigned long nr_iowait(void)
2110 {
2111 	unsigned long i, sum = 0;
2112 
2113 	for_each_possible_cpu(i)
2114 		sum += atomic_read(&cpu_rq(i)->nr_iowait);
2115 
2116 	return sum;
2117 }
2118 
nr_iowait_cpu(int cpu)2119 unsigned long nr_iowait_cpu(int cpu)
2120 {
2121 	struct rq *this = cpu_rq(cpu);
2122 	return atomic_read(&this->nr_iowait);
2123 }
2124 
this_cpu_load(void)2125 unsigned long this_cpu_load(void)
2126 {
2127 	struct rq *this = this_rq();
2128 	return this->cpu_load[0];
2129 }
2130 
2131 
2132 /* Variables and functions for calc_load */
2133 static atomic_long_t calc_load_tasks;
2134 static unsigned long calc_load_update;
2135 unsigned long avenrun[3];
2136 EXPORT_SYMBOL(avenrun);
2137 
calc_load_fold_active(struct rq * this_rq)2138 static long calc_load_fold_active(struct rq *this_rq)
2139 {
2140 	long nr_active, delta = 0;
2141 
2142 	nr_active = this_rq->nr_running;
2143 	nr_active += (long) this_rq->nr_uninterruptible;
2144 
2145 	if (nr_active != this_rq->calc_load_active) {
2146 		delta = nr_active - this_rq->calc_load_active;
2147 		this_rq->calc_load_active = nr_active;
2148 	}
2149 
2150 	return delta;
2151 }
2152 
2153 static unsigned long
calc_load(unsigned long load,unsigned long exp,unsigned long active)2154 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2155 {
2156 	load *= exp;
2157 	load += active * (FIXED_1 - exp);
2158 	load += 1UL << (FSHIFT - 1);
2159 	return load >> FSHIFT;
2160 }
2161 
2162 #ifdef CONFIG_NO_HZ
2163 /*
2164  * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2165  *
2166  * When making the ILB scale, we should try to pull this in as well.
2167  */
2168 static atomic_long_t calc_load_tasks_idle;
2169 
calc_load_account_idle(struct rq * this_rq)2170 void calc_load_account_idle(struct rq *this_rq)
2171 {
2172 	long delta;
2173 
2174 	delta = calc_load_fold_active(this_rq);
2175 	if (delta)
2176 		atomic_long_add(delta, &calc_load_tasks_idle);
2177 }
2178 
calc_load_fold_idle(void)2179 static long calc_load_fold_idle(void)
2180 {
2181 	long delta = 0;
2182 
2183 	/*
2184 	 * Its got a race, we don't care...
2185 	 */
2186 	if (atomic_long_read(&calc_load_tasks_idle))
2187 		delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2188 
2189 	return delta;
2190 }
2191 
2192 /**
2193  * fixed_power_int - compute: x^n, in O(log n) time
2194  *
2195  * @x:         base of the power
2196  * @frac_bits: fractional bits of @x
2197  * @n:         power to raise @x to.
2198  *
2199  * By exploiting the relation between the definition of the natural power
2200  * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2201  * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2202  * (where: n_i \elem {0, 1}, the binary vector representing n),
2203  * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2204  * of course trivially computable in O(log_2 n), the length of our binary
2205  * vector.
2206  */
2207 static unsigned long
fixed_power_int(unsigned long x,unsigned int frac_bits,unsigned int n)2208 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2209 {
2210 	unsigned long result = 1UL << frac_bits;
2211 
2212 	if (n) for (;;) {
2213 		if (n & 1) {
2214 			result *= x;
2215 			result += 1UL << (frac_bits - 1);
2216 			result >>= frac_bits;
2217 		}
2218 		n >>= 1;
2219 		if (!n)
2220 			break;
2221 		x *= x;
2222 		x += 1UL << (frac_bits - 1);
2223 		x >>= frac_bits;
2224 	}
2225 
2226 	return result;
2227 }
2228 
2229 /*
2230  * a1 = a0 * e + a * (1 - e)
2231  *
2232  * a2 = a1 * e + a * (1 - e)
2233  *    = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2234  *    = a0 * e^2 + a * (1 - e) * (1 + e)
2235  *
2236  * a3 = a2 * e + a * (1 - e)
2237  *    = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2238  *    = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2239  *
2240  *  ...
2241  *
2242  * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2243  *    = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2244  *    = a0 * e^n + a * (1 - e^n)
2245  *
2246  * [1] application of the geometric series:
2247  *
2248  *              n         1 - x^(n+1)
2249  *     S_n := \Sum x^i = -------------
2250  *             i=0          1 - x
2251  */
2252 static unsigned long
calc_load_n(unsigned long load,unsigned long exp,unsigned long active,unsigned int n)2253 calc_load_n(unsigned long load, unsigned long exp,
2254 	    unsigned long active, unsigned int n)
2255 {
2256 
2257 	return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2258 }
2259 
2260 /*
2261  * NO_HZ can leave us missing all per-cpu ticks calling
2262  * calc_load_account_active(), but since an idle CPU folds its delta into
2263  * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2264  * in the pending idle delta if our idle period crossed a load cycle boundary.
2265  *
2266  * Once we've updated the global active value, we need to apply the exponential
2267  * weights adjusted to the number of cycles missed.
2268  */
calc_global_nohz(unsigned long ticks)2269 static void calc_global_nohz(unsigned long ticks)
2270 {
2271 	long delta, active, n;
2272 
2273 	if (time_before(jiffies, calc_load_update))
2274 		return;
2275 
2276 	/*
2277 	 * If we crossed a calc_load_update boundary, make sure to fold
2278 	 * any pending idle changes, the respective CPUs might have
2279 	 * missed the tick driven calc_load_account_active() update
2280 	 * due to NO_HZ.
2281 	 */
2282 	delta = calc_load_fold_idle();
2283 	if (delta)
2284 		atomic_long_add(delta, &calc_load_tasks);
2285 
2286 	/*
2287 	 * If we were idle for multiple load cycles, apply them.
2288 	 */
2289 	if (ticks >= LOAD_FREQ) {
2290 		n = ticks / LOAD_FREQ;
2291 
2292 		active = atomic_long_read(&calc_load_tasks);
2293 		active = active > 0 ? active * FIXED_1 : 0;
2294 
2295 		avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2296 		avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2297 		avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2298 
2299 		calc_load_update += n * LOAD_FREQ;
2300 	}
2301 
2302 	/*
2303 	 * Its possible the remainder of the above division also crosses
2304 	 * a LOAD_FREQ period, the regular check in calc_global_load()
2305 	 * which comes after this will take care of that.
2306 	 *
2307 	 * Consider us being 11 ticks before a cycle completion, and us
2308 	 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
2309 	 * age us 4 cycles, and the test in calc_global_load() will
2310 	 * pick up the final one.
2311 	 */
2312 }
2313 #else
calc_load_account_idle(struct rq * this_rq)2314 void calc_load_account_idle(struct rq *this_rq)
2315 {
2316 }
2317 
calc_load_fold_idle(void)2318 static inline long calc_load_fold_idle(void)
2319 {
2320 	return 0;
2321 }
2322 
calc_global_nohz(unsigned long ticks)2323 static void calc_global_nohz(unsigned long ticks)
2324 {
2325 }
2326 #endif
2327 
2328 /**
2329  * get_avenrun - get the load average array
2330  * @loads:	pointer to dest load array
2331  * @offset:	offset to add
2332  * @shift:	shift count to shift the result left
2333  *
2334  * These values are estimates at best, so no need for locking.
2335  */
get_avenrun(unsigned long * loads,unsigned long offset,int shift)2336 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2337 {
2338 	loads[0] = (avenrun[0] + offset) << shift;
2339 	loads[1] = (avenrun[1] + offset) << shift;
2340 	loads[2] = (avenrun[2] + offset) << shift;
2341 }
2342 
2343 /*
2344  * calc_load - update the avenrun load estimates 10 ticks after the
2345  * CPUs have updated calc_load_tasks.
2346  */
calc_global_load(unsigned long ticks)2347 void calc_global_load(unsigned long ticks)
2348 {
2349 	long active;
2350 
2351 	calc_global_nohz(ticks);
2352 
2353 	if (time_before(jiffies, calc_load_update + 10))
2354 		return;
2355 
2356 	active = atomic_long_read(&calc_load_tasks);
2357 	active = active > 0 ? active * FIXED_1 : 0;
2358 
2359 	avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2360 	avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2361 	avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2362 
2363 	calc_load_update += LOAD_FREQ;
2364 }
2365 
2366 /*
2367  * Called from update_cpu_load() to periodically update this CPU's
2368  * active count.
2369  */
calc_load_account_active(struct rq * this_rq)2370 static void calc_load_account_active(struct rq *this_rq)
2371 {
2372 	long delta;
2373 
2374 	if (time_before(jiffies, this_rq->calc_load_update))
2375 		return;
2376 
2377 	delta  = calc_load_fold_active(this_rq);
2378 	delta += calc_load_fold_idle();
2379 	if (delta)
2380 		atomic_long_add(delta, &calc_load_tasks);
2381 
2382 	this_rq->calc_load_update += LOAD_FREQ;
2383 }
2384 
2385 /*
2386  * The exact cpuload at various idx values, calculated at every tick would be
2387  * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2388  *
2389  * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2390  * on nth tick when cpu may be busy, then we have:
2391  * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2392  * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2393  *
2394  * decay_load_missed() below does efficient calculation of
2395  * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2396  * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2397  *
2398  * The calculation is approximated on a 128 point scale.
2399  * degrade_zero_ticks is the number of ticks after which load at any
2400  * particular idx is approximated to be zero.
2401  * degrade_factor is a precomputed table, a row for each load idx.
2402  * Each column corresponds to degradation factor for a power of two ticks,
2403  * based on 128 point scale.
2404  * Example:
2405  * row 2, col 3 (=12) says that the degradation at load idx 2 after
2406  * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2407  *
2408  * With this power of 2 load factors, we can degrade the load n times
2409  * by looking at 1 bits in n and doing as many mult/shift instead of
2410  * n mult/shifts needed by the exact degradation.
2411  */
2412 #define DEGRADE_SHIFT		7
2413 static const unsigned char
2414 		degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2415 static const unsigned char
2416 		degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2417 					{0, 0, 0, 0, 0, 0, 0, 0},
2418 					{64, 32, 8, 0, 0, 0, 0, 0},
2419 					{96, 72, 40, 12, 1, 0, 0},
2420 					{112, 98, 75, 43, 15, 1, 0},
2421 					{120, 112, 98, 76, 45, 16, 2} };
2422 
2423 /*
2424  * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2425  * would be when CPU is idle and so we just decay the old load without
2426  * adding any new load.
2427  */
2428 static unsigned long
decay_load_missed(unsigned long load,unsigned long missed_updates,int idx)2429 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2430 {
2431 	int j = 0;
2432 
2433 	if (!missed_updates)
2434 		return load;
2435 
2436 	if (missed_updates >= degrade_zero_ticks[idx])
2437 		return 0;
2438 
2439 	if (idx == 1)
2440 		return load >> missed_updates;
2441 
2442 	while (missed_updates) {
2443 		if (missed_updates % 2)
2444 			load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2445 
2446 		missed_updates >>= 1;
2447 		j++;
2448 	}
2449 	return load;
2450 }
2451 
2452 /*
2453  * Update rq->cpu_load[] statistics. This function is usually called every
2454  * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2455  * every tick. We fix it up based on jiffies.
2456  */
update_cpu_load(struct rq * this_rq)2457 void update_cpu_load(struct rq *this_rq)
2458 {
2459 	unsigned long this_load = this_rq->load.weight;
2460 	unsigned long curr_jiffies = jiffies;
2461 	unsigned long pending_updates;
2462 	int i, scale;
2463 
2464 	this_rq->nr_load_updates++;
2465 
2466 	/* Avoid repeated calls on same jiffy, when moving in and out of idle */
2467 	if (curr_jiffies == this_rq->last_load_update_tick)
2468 		return;
2469 
2470 	pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2471 	this_rq->last_load_update_tick = curr_jiffies;
2472 
2473 	/* Update our load: */
2474 	this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2475 	for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2476 		unsigned long old_load, new_load;
2477 
2478 		/* scale is effectively 1 << i now, and >> i divides by scale */
2479 
2480 		old_load = this_rq->cpu_load[i];
2481 		old_load = decay_load_missed(old_load, pending_updates - 1, i);
2482 		new_load = this_load;
2483 		/*
2484 		 * Round up the averaging division if load is increasing. This
2485 		 * prevents us from getting stuck on 9 if the load is 10, for
2486 		 * example.
2487 		 */
2488 		if (new_load > old_load)
2489 			new_load += scale - 1;
2490 
2491 		this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2492 	}
2493 
2494 	sched_avg_update(this_rq);
2495 }
2496 
update_cpu_load_active(struct rq * this_rq)2497 static void update_cpu_load_active(struct rq *this_rq)
2498 {
2499 	update_cpu_load(this_rq);
2500 
2501 	calc_load_account_active(this_rq);
2502 }
2503 
2504 #ifdef CONFIG_SMP
2505 
2506 /*
2507  * sched_exec - execve() is a valuable balancing opportunity, because at
2508  * this point the task has the smallest effective memory and cache footprint.
2509  */
sched_exec(void)2510 void sched_exec(void)
2511 {
2512 	struct task_struct *p = current;
2513 	unsigned long flags;
2514 	int dest_cpu;
2515 
2516 	raw_spin_lock_irqsave(&p->pi_lock, flags);
2517 	dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2518 	if (dest_cpu == smp_processor_id())
2519 		goto unlock;
2520 
2521 	if (likely(cpu_active(dest_cpu))) {
2522 		struct migration_arg arg = { p, dest_cpu };
2523 
2524 		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2525 		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2526 		return;
2527 	}
2528 unlock:
2529 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2530 }
2531 
2532 #endif
2533 
2534 DEFINE_PER_CPU(struct kernel_stat, kstat);
2535 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2536 
2537 EXPORT_PER_CPU_SYMBOL(kstat);
2538 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2539 
2540 /*
2541  * Return any ns on the sched_clock that have not yet been accounted in
2542  * @p in case that task is currently running.
2543  *
2544  * Called with task_rq_lock() held on @rq.
2545  */
do_task_delta_exec(struct task_struct * p,struct rq * rq)2546 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2547 {
2548 	u64 ns = 0;
2549 
2550 	if (task_current(rq, p)) {
2551 		update_rq_clock(rq);
2552 		ns = rq->clock_task - p->se.exec_start;
2553 		if ((s64)ns < 0)
2554 			ns = 0;
2555 	}
2556 
2557 	return ns;
2558 }
2559 
task_delta_exec(struct task_struct * p)2560 unsigned long long task_delta_exec(struct task_struct *p)
2561 {
2562 	unsigned long flags;
2563 	struct rq *rq;
2564 	u64 ns = 0;
2565 
2566 	rq = task_rq_lock(p, &flags);
2567 	ns = do_task_delta_exec(p, rq);
2568 	task_rq_unlock(rq, p, &flags);
2569 
2570 	return ns;
2571 }
2572 
2573 /*
2574  * Return accounted runtime for the task.
2575  * In case the task is currently running, return the runtime plus current's
2576  * pending runtime that have not been accounted yet.
2577  */
task_sched_runtime(struct task_struct * p)2578 unsigned long long task_sched_runtime(struct task_struct *p)
2579 {
2580 	unsigned long flags;
2581 	struct rq *rq;
2582 	u64 ns = 0;
2583 
2584 	rq = task_rq_lock(p, &flags);
2585 	ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2586 	task_rq_unlock(rq, p, &flags);
2587 
2588 	return ns;
2589 }
2590 
2591 #ifdef CONFIG_CGROUP_CPUACCT
2592 struct cgroup_subsys cpuacct_subsys;
2593 struct cpuacct root_cpuacct;
2594 #endif
2595 
task_group_account_field(struct task_struct * p,int index,u64 tmp)2596 static inline void task_group_account_field(struct task_struct *p, int index,
2597 					    u64 tmp)
2598 {
2599 #ifdef CONFIG_CGROUP_CPUACCT
2600 	struct kernel_cpustat *kcpustat;
2601 	struct cpuacct *ca;
2602 #endif
2603 	/*
2604 	 * Since all updates are sure to touch the root cgroup, we
2605 	 * get ourselves ahead and touch it first. If the root cgroup
2606 	 * is the only cgroup, then nothing else should be necessary.
2607 	 *
2608 	 */
2609 	__get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2610 
2611 #ifdef CONFIG_CGROUP_CPUACCT
2612 	if (unlikely(!cpuacct_subsys.active))
2613 		return;
2614 
2615 	rcu_read_lock();
2616 	ca = task_ca(p);
2617 	while (ca && (ca != &root_cpuacct)) {
2618 		kcpustat = this_cpu_ptr(ca->cpustat);
2619 		kcpustat->cpustat[index] += tmp;
2620 		ca = parent_ca(ca);
2621 	}
2622 	rcu_read_unlock();
2623 #endif
2624 }
2625 
2626 
2627 /*
2628  * Account user cpu time to a process.
2629  * @p: the process that the cpu time gets accounted to
2630  * @cputime: the cpu time spent in user space since the last update
2631  * @cputime_scaled: cputime scaled by cpu frequency
2632  */
account_user_time(struct task_struct * p,cputime_t cputime,cputime_t cputime_scaled)2633 void account_user_time(struct task_struct *p, cputime_t cputime,
2634 		       cputime_t cputime_scaled)
2635 {
2636 	int index;
2637 
2638 	/* Add user time to process. */
2639 	p->utime += cputime;
2640 	p->utimescaled += cputime_scaled;
2641 	account_group_user_time(p, cputime);
2642 
2643 	index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2644 
2645 	/* Add user time to cpustat. */
2646 	task_group_account_field(p, index, (__force u64) cputime);
2647 
2648 	/* Account for user time used */
2649 	acct_update_integrals(p);
2650 }
2651 
2652 /*
2653  * Account guest cpu time to a process.
2654  * @p: the process that the cpu time gets accounted to
2655  * @cputime: the cpu time spent in virtual machine since the last update
2656  * @cputime_scaled: cputime scaled by cpu frequency
2657  */
account_guest_time(struct task_struct * p,cputime_t cputime,cputime_t cputime_scaled)2658 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2659 			       cputime_t cputime_scaled)
2660 {
2661 	u64 *cpustat = kcpustat_this_cpu->cpustat;
2662 
2663 	/* Add guest time to process. */
2664 	p->utime += cputime;
2665 	p->utimescaled += cputime_scaled;
2666 	account_group_user_time(p, cputime);
2667 	p->gtime += cputime;
2668 
2669 	/* Add guest time to cpustat. */
2670 	if (TASK_NICE(p) > 0) {
2671 		cpustat[CPUTIME_NICE] += (__force u64) cputime;
2672 		cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2673 	} else {
2674 		cpustat[CPUTIME_USER] += (__force u64) cputime;
2675 		cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2676 	}
2677 }
2678 
2679 /*
2680  * Account system cpu time to a process and desired cpustat field
2681  * @p: the process that the cpu time gets accounted to
2682  * @cputime: the cpu time spent in kernel space since the last update
2683  * @cputime_scaled: cputime scaled by cpu frequency
2684  * @target_cputime64: pointer to cpustat field that has to be updated
2685  */
2686 static inline
__account_system_time(struct task_struct * p,cputime_t cputime,cputime_t cputime_scaled,int index)2687 void __account_system_time(struct task_struct *p, cputime_t cputime,
2688 			cputime_t cputime_scaled, int index)
2689 {
2690 	/* Add system time to process. */
2691 	p->stime += cputime;
2692 	p->stimescaled += cputime_scaled;
2693 	account_group_system_time(p, cputime);
2694 
2695 	/* Add system time to cpustat. */
2696 	task_group_account_field(p, index, (__force u64) cputime);
2697 
2698 	/* Account for system time used */
2699 	acct_update_integrals(p);
2700 }
2701 
2702 /*
2703  * Account system cpu time to a process.
2704  * @p: the process that the cpu time gets accounted to
2705  * @hardirq_offset: the offset to subtract from hardirq_count()
2706  * @cputime: the cpu time spent in kernel space since the last update
2707  * @cputime_scaled: cputime scaled by cpu frequency
2708  */
account_system_time(struct task_struct * p,int hardirq_offset,cputime_t cputime,cputime_t cputime_scaled)2709 void account_system_time(struct task_struct *p, int hardirq_offset,
2710 			 cputime_t cputime, cputime_t cputime_scaled)
2711 {
2712 	int index;
2713 
2714 	if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2715 		account_guest_time(p, cputime, cputime_scaled);
2716 		return;
2717 	}
2718 
2719 	if (hardirq_count() - hardirq_offset)
2720 		index = CPUTIME_IRQ;
2721 	else if (in_serving_softirq())
2722 		index = CPUTIME_SOFTIRQ;
2723 	else
2724 		index = CPUTIME_SYSTEM;
2725 
2726 	__account_system_time(p, cputime, cputime_scaled, index);
2727 }
2728 
2729 /*
2730  * Account for involuntary wait time.
2731  * @cputime: the cpu time spent in involuntary wait
2732  */
account_steal_time(cputime_t cputime)2733 void account_steal_time(cputime_t cputime)
2734 {
2735 	u64 *cpustat = kcpustat_this_cpu->cpustat;
2736 
2737 	cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2738 }
2739 
2740 /*
2741  * Account for idle time.
2742  * @cputime: the cpu time spent in idle wait
2743  */
account_idle_time(cputime_t cputime)2744 void account_idle_time(cputime_t cputime)
2745 {
2746 	u64 *cpustat = kcpustat_this_cpu->cpustat;
2747 	struct rq *rq = this_rq();
2748 
2749 	if (atomic_read(&rq->nr_iowait) > 0)
2750 		cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2751 	else
2752 		cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2753 }
2754 
steal_account_process_tick(void)2755 static __always_inline bool steal_account_process_tick(void)
2756 {
2757 #ifdef CONFIG_PARAVIRT
2758 	if (static_branch(&paravirt_steal_enabled)) {
2759 		u64 steal, st = 0;
2760 
2761 		steal = paravirt_steal_clock(smp_processor_id());
2762 		steal -= this_rq()->prev_steal_time;
2763 
2764 		st = steal_ticks(steal);
2765 		this_rq()->prev_steal_time += st * TICK_NSEC;
2766 
2767 		account_steal_time(st);
2768 		return st;
2769 	}
2770 #endif
2771 	return false;
2772 }
2773 
2774 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2775 
2776 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2777 /*
2778  * Account a tick to a process and cpustat
2779  * @p: the process that the cpu time gets accounted to
2780  * @user_tick: is the tick from userspace
2781  * @rq: the pointer to rq
2782  *
2783  * Tick demultiplexing follows the order
2784  * - pending hardirq update
2785  * - pending softirq update
2786  * - user_time
2787  * - idle_time
2788  * - system time
2789  *   - check for guest_time
2790  *   - else account as system_time
2791  *
2792  * Check for hardirq is done both for system and user time as there is
2793  * no timer going off while we are on hardirq and hence we may never get an
2794  * opportunity to update it solely in system time.
2795  * p->stime and friends are only updated on system time and not on irq
2796  * softirq as those do not count in task exec_runtime any more.
2797  */
irqtime_account_process_tick(struct task_struct * p,int user_tick,struct rq * rq)2798 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2799 						struct rq *rq)
2800 {
2801 	cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2802 	u64 *cpustat = kcpustat_this_cpu->cpustat;
2803 
2804 	if (steal_account_process_tick())
2805 		return;
2806 
2807 	if (irqtime_account_hi_update()) {
2808 		cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
2809 	} else if (irqtime_account_si_update()) {
2810 		cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
2811 	} else if (this_cpu_ksoftirqd() == p) {
2812 		/*
2813 		 * ksoftirqd time do not get accounted in cpu_softirq_time.
2814 		 * So, we have to handle it separately here.
2815 		 * Also, p->stime needs to be updated for ksoftirqd.
2816 		 */
2817 		__account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2818 					CPUTIME_SOFTIRQ);
2819 	} else if (user_tick) {
2820 		account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2821 	} else if (p == rq->idle) {
2822 		account_idle_time(cputime_one_jiffy);
2823 	} else if (p->flags & PF_VCPU) { /* System time or guest time */
2824 		account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2825 	} else {
2826 		__account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2827 					CPUTIME_SYSTEM);
2828 	}
2829 }
2830 
irqtime_account_idle_ticks(int ticks)2831 static void irqtime_account_idle_ticks(int ticks)
2832 {
2833 	int i;
2834 	struct rq *rq = this_rq();
2835 
2836 	for (i = 0; i < ticks; i++)
2837 		irqtime_account_process_tick(current, 0, rq);
2838 }
2839 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
irqtime_account_idle_ticks(int ticks)2840 static void irqtime_account_idle_ticks(int ticks) {}
irqtime_account_process_tick(struct task_struct * p,int user_tick,struct rq * rq)2841 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2842 						struct rq *rq) {}
2843 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2844 
2845 /*
2846  * Account a single tick of cpu time.
2847  * @p: the process that the cpu time gets accounted to
2848  * @user_tick: indicates if the tick is a user or a system tick
2849  */
account_process_tick(struct task_struct * p,int user_tick)2850 void account_process_tick(struct task_struct *p, int user_tick)
2851 {
2852 	cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2853 	struct rq *rq = this_rq();
2854 
2855 	if (sched_clock_irqtime) {
2856 		irqtime_account_process_tick(p, user_tick, rq);
2857 		return;
2858 	}
2859 
2860 	if (steal_account_process_tick())
2861 		return;
2862 
2863 	if (user_tick)
2864 		account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2865 	else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2866 		account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2867 				    one_jiffy_scaled);
2868 	else
2869 		account_idle_time(cputime_one_jiffy);
2870 }
2871 
2872 /*
2873  * Account multiple ticks of steal time.
2874  * @p: the process from which the cpu time has been stolen
2875  * @ticks: number of stolen ticks
2876  */
account_steal_ticks(unsigned long ticks)2877 void account_steal_ticks(unsigned long ticks)
2878 {
2879 	account_steal_time(jiffies_to_cputime(ticks));
2880 }
2881 
2882 /*
2883  * Account multiple ticks of idle time.
2884  * @ticks: number of stolen ticks
2885  */
account_idle_ticks(unsigned long ticks)2886 void account_idle_ticks(unsigned long ticks)
2887 {
2888 
2889 	if (sched_clock_irqtime) {
2890 		irqtime_account_idle_ticks(ticks);
2891 		return;
2892 	}
2893 
2894 	account_idle_time(jiffies_to_cputime(ticks));
2895 }
2896 
2897 #endif
2898 
2899 /*
2900  * Use precise platform statistics if available:
2901  */
2902 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
task_times(struct task_struct * p,cputime_t * ut,cputime_t * st)2903 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2904 {
2905 	*ut = p->utime;
2906 	*st = p->stime;
2907 }
2908 
thread_group_times(struct task_struct * p,cputime_t * ut,cputime_t * st)2909 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2910 {
2911 	struct task_cputime cputime;
2912 
2913 	thread_group_cputime(p, &cputime);
2914 
2915 	*ut = cputime.utime;
2916 	*st = cputime.stime;
2917 }
2918 #else
2919 
2920 #ifndef nsecs_to_cputime
2921 # define nsecs_to_cputime(__nsecs)	nsecs_to_jiffies(__nsecs)
2922 #endif
2923 
task_times(struct task_struct * p,cputime_t * ut,cputime_t * st)2924 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2925 {
2926 	cputime_t rtime, utime = p->utime, total = utime + p->stime;
2927 
2928 	/*
2929 	 * Use CFS's precise accounting:
2930 	 */
2931 	rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
2932 
2933 	if (total) {
2934 		u64 temp = (__force u64) rtime;
2935 
2936 		temp *= (__force u64) utime;
2937 		do_div(temp, (__force u32) total);
2938 		utime = (__force cputime_t) temp;
2939 	} else
2940 		utime = rtime;
2941 
2942 	/*
2943 	 * Compare with previous values, to keep monotonicity:
2944 	 */
2945 	p->prev_utime = max(p->prev_utime, utime);
2946 	p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
2947 
2948 	*ut = p->prev_utime;
2949 	*st = p->prev_stime;
2950 }
2951 
2952 /*
2953  * Must be called with siglock held.
2954  */
thread_group_times(struct task_struct * p,cputime_t * ut,cputime_t * st)2955 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2956 {
2957 	struct signal_struct *sig = p->signal;
2958 	struct task_cputime cputime;
2959 	cputime_t rtime, utime, total;
2960 
2961 	thread_group_cputime(p, &cputime);
2962 
2963 	total = cputime.utime + cputime.stime;
2964 	rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
2965 
2966 	if (total) {
2967 		u64 temp = (__force u64) rtime;
2968 
2969 		temp *= (__force u64) cputime.utime;
2970 		do_div(temp, (__force u32) total);
2971 		utime = (__force cputime_t) temp;
2972 	} else
2973 		utime = rtime;
2974 
2975 	sig->prev_utime = max(sig->prev_utime, utime);
2976 	sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
2977 
2978 	*ut = sig->prev_utime;
2979 	*st = sig->prev_stime;
2980 }
2981 #endif
2982 
2983 /*
2984  * This function gets called by the timer code, with HZ frequency.
2985  * We call it with interrupts disabled.
2986  */
scheduler_tick(void)2987 void scheduler_tick(void)
2988 {
2989 	int cpu = smp_processor_id();
2990 	struct rq *rq = cpu_rq(cpu);
2991 	struct task_struct *curr = rq->curr;
2992 
2993 	sched_clock_tick();
2994 
2995 	raw_spin_lock(&rq->lock);
2996 	update_rq_clock(rq);
2997 	update_cpu_load_active(rq);
2998 	curr->sched_class->task_tick(rq, curr, 0);
2999 	raw_spin_unlock(&rq->lock);
3000 
3001 	perf_event_task_tick();
3002 
3003 #ifdef CONFIG_SMP
3004 	rq->idle_balance = idle_cpu(cpu);
3005 	trigger_load_balance(rq, cpu);
3006 #endif
3007 }
3008 
get_parent_ip(unsigned long addr)3009 notrace unsigned long get_parent_ip(unsigned long addr)
3010 {
3011 	if (in_lock_functions(addr)) {
3012 		addr = CALLER_ADDR2;
3013 		if (in_lock_functions(addr))
3014 			addr = CALLER_ADDR3;
3015 	}
3016 	return addr;
3017 }
3018 
3019 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3020 				defined(CONFIG_PREEMPT_TRACER))
3021 
add_preempt_count(int val)3022 void __kprobes add_preempt_count(int val)
3023 {
3024 #ifdef CONFIG_DEBUG_PREEMPT
3025 	/*
3026 	 * Underflow?
3027 	 */
3028 	if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3029 		return;
3030 #endif
3031 	preempt_count() += val;
3032 #ifdef CONFIG_DEBUG_PREEMPT
3033 	/*
3034 	 * Spinlock count overflowing soon?
3035 	 */
3036 	DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3037 				PREEMPT_MASK - 10);
3038 #endif
3039 	if (preempt_count() == val)
3040 		trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3041 }
3042 EXPORT_SYMBOL(add_preempt_count);
3043 
sub_preempt_count(int val)3044 void __kprobes sub_preempt_count(int val)
3045 {
3046 #ifdef CONFIG_DEBUG_PREEMPT
3047 	/*
3048 	 * Underflow?
3049 	 */
3050 	if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3051 		return;
3052 	/*
3053 	 * Is the spinlock portion underflowing?
3054 	 */
3055 	if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3056 			!(preempt_count() & PREEMPT_MASK)))
3057 		return;
3058 #endif
3059 
3060 	if (preempt_count() == val)
3061 		trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3062 	preempt_count() -= val;
3063 }
3064 EXPORT_SYMBOL(sub_preempt_count);
3065 
3066 #endif
3067 
3068 /*
3069  * Print scheduling while atomic bug:
3070  */
__schedule_bug(struct task_struct * prev)3071 static noinline void __schedule_bug(struct task_struct *prev)
3072 {
3073 	struct pt_regs *regs = get_irq_regs();
3074 
3075 	if (oops_in_progress)
3076 		return;
3077 
3078 	printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3079 		prev->comm, prev->pid, preempt_count());
3080 
3081 	debug_show_held_locks(prev);
3082 	print_modules();
3083 	if (irqs_disabled())
3084 		print_irqtrace_events(prev);
3085 
3086 	if (regs)
3087 		show_regs(regs);
3088 	else
3089 		dump_stack();
3090 }
3091 
3092 /*
3093  * Various schedule()-time debugging checks and statistics:
3094  */
schedule_debug(struct task_struct * prev)3095 static inline void schedule_debug(struct task_struct *prev)
3096 {
3097 	/*
3098 	 * Test if we are atomic. Since do_exit() needs to call into
3099 	 * schedule() atomically, we ignore that path for now.
3100 	 * Otherwise, whine if we are scheduling when we should not be.
3101 	 */
3102 	if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3103 		__schedule_bug(prev);
3104 	rcu_sleep_check();
3105 
3106 	profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3107 
3108 	schedstat_inc(this_rq(), sched_count);
3109 }
3110 
put_prev_task(struct rq * rq,struct task_struct * prev)3111 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3112 {
3113 	if (prev->on_rq || rq->skip_clock_update < 0)
3114 		update_rq_clock(rq);
3115 	prev->sched_class->put_prev_task(rq, prev);
3116 }
3117 
3118 /*
3119  * Pick up the highest-prio task:
3120  */
3121 static inline struct task_struct *
pick_next_task(struct rq * rq)3122 pick_next_task(struct rq *rq)
3123 {
3124 	const struct sched_class *class;
3125 	struct task_struct *p;
3126 
3127 	/*
3128 	 * Optimization: we know that if all tasks are in
3129 	 * the fair class we can call that function directly:
3130 	 */
3131 	if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3132 		p = fair_sched_class.pick_next_task(rq);
3133 		if (likely(p))
3134 			return p;
3135 	}
3136 
3137 	for_each_class(class) {
3138 		p = class->pick_next_task(rq);
3139 		if (p)
3140 			return p;
3141 	}
3142 
3143 	BUG(); /* the idle class will always have a runnable task */
3144 }
3145 
3146 /*
3147  * __schedule() is the main scheduler function.
3148  */
__schedule(void)3149 static void __sched __schedule(void)
3150 {
3151 	struct task_struct *prev, *next;
3152 	unsigned long *switch_count;
3153 	struct rq *rq;
3154 	int cpu;
3155 
3156 need_resched:
3157 	preempt_disable();
3158 	cpu = smp_processor_id();
3159 	rq = cpu_rq(cpu);
3160 	rcu_note_context_switch(cpu);
3161 	prev = rq->curr;
3162 
3163 	schedule_debug(prev);
3164 
3165 	if (sched_feat(HRTICK))
3166 		hrtick_clear(rq);
3167 
3168 	raw_spin_lock_irq(&rq->lock);
3169 
3170 	switch_count = &prev->nivcsw;
3171 	if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3172 		if (unlikely(signal_pending_state(prev->state, prev))) {
3173 			prev->state = TASK_RUNNING;
3174 		} else {
3175 			deactivate_task(rq, prev, DEQUEUE_SLEEP);
3176 			prev->on_rq = 0;
3177 
3178 			/*
3179 			 * If a worker went to sleep, notify and ask workqueue
3180 			 * whether it wants to wake up a task to maintain
3181 			 * concurrency.
3182 			 */
3183 			if (prev->flags & PF_WQ_WORKER) {
3184 				struct task_struct *to_wakeup;
3185 
3186 				to_wakeup = wq_worker_sleeping(prev, cpu);
3187 				if (to_wakeup)
3188 					try_to_wake_up_local(to_wakeup);
3189 			}
3190 		}
3191 		switch_count = &prev->nvcsw;
3192 	}
3193 
3194 	pre_schedule(rq, prev);
3195 
3196 	if (unlikely(!rq->nr_running))
3197 		idle_balance(cpu, rq);
3198 
3199 	put_prev_task(rq, prev);
3200 	next = pick_next_task(rq);
3201 	clear_tsk_need_resched(prev);
3202 	rq->skip_clock_update = 0;
3203 
3204 	if (likely(prev != next)) {
3205 		rq->nr_switches++;
3206 		rq->curr = next;
3207 		++*switch_count;
3208 
3209 		context_switch(rq, prev, next); /* unlocks the rq */
3210 		/*
3211 		 * The context switch have flipped the stack from under us
3212 		 * and restored the local variables which were saved when
3213 		 * this task called schedule() in the past. prev == current
3214 		 * is still correct, but it can be moved to another cpu/rq.
3215 		 */
3216 		cpu = smp_processor_id();
3217 		rq = cpu_rq(cpu);
3218 	} else
3219 		raw_spin_unlock_irq(&rq->lock);
3220 
3221 	post_schedule(rq);
3222 
3223 	preempt_enable_no_resched();
3224 	if (need_resched())
3225 		goto need_resched;
3226 }
3227 
sched_submit_work(struct task_struct * tsk)3228 static inline void sched_submit_work(struct task_struct *tsk)
3229 {
3230 	if (!tsk->state)
3231 		return;
3232 	/*
3233 	 * If we are going to sleep and we have plugged IO queued,
3234 	 * make sure to submit it to avoid deadlocks.
3235 	 */
3236 	if (blk_needs_flush_plug(tsk))
3237 		blk_schedule_flush_plug(tsk);
3238 }
3239 
schedule(void)3240 asmlinkage void __sched schedule(void)
3241 {
3242 	struct task_struct *tsk = current;
3243 
3244 	sched_submit_work(tsk);
3245 	__schedule();
3246 }
3247 EXPORT_SYMBOL(schedule);
3248 
3249 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3250 
owner_running(struct mutex * lock,struct task_struct * owner)3251 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3252 {
3253 	if (lock->owner != owner)
3254 		return false;
3255 
3256 	/*
3257 	 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3258 	 * lock->owner still matches owner, if that fails, owner might
3259 	 * point to free()d memory, if it still matches, the rcu_read_lock()
3260 	 * ensures the memory stays valid.
3261 	 */
3262 	barrier();
3263 
3264 	return owner->on_cpu;
3265 }
3266 
3267 /*
3268  * Look out! "owner" is an entirely speculative pointer
3269  * access and not reliable.
3270  */
mutex_spin_on_owner(struct mutex * lock,struct task_struct * owner)3271 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3272 {
3273 	if (!sched_feat(OWNER_SPIN))
3274 		return 0;
3275 
3276 	rcu_read_lock();
3277 	while (owner_running(lock, owner)) {
3278 		if (need_resched())
3279 			break;
3280 
3281 		arch_mutex_cpu_relax();
3282 	}
3283 	rcu_read_unlock();
3284 
3285 	/*
3286 	 * We break out the loop above on need_resched() and when the
3287 	 * owner changed, which is a sign for heavy contention. Return
3288 	 * success only when lock->owner is NULL.
3289 	 */
3290 	return lock->owner == NULL;
3291 }
3292 #endif
3293 
3294 #ifdef CONFIG_PREEMPT
3295 /*
3296  * this is the entry point to schedule() from in-kernel preemption
3297  * off of preempt_enable. Kernel preemptions off return from interrupt
3298  * occur there and call schedule directly.
3299  */
preempt_schedule(void)3300 asmlinkage void __sched notrace preempt_schedule(void)
3301 {
3302 	struct thread_info *ti = current_thread_info();
3303 
3304 	/*
3305 	 * If there is a non-zero preempt_count or interrupts are disabled,
3306 	 * we do not want to preempt the current task. Just return..
3307 	 */
3308 	if (likely(ti->preempt_count || irqs_disabled()))
3309 		return;
3310 
3311 	do {
3312 		add_preempt_count_notrace(PREEMPT_ACTIVE);
3313 		__schedule();
3314 		sub_preempt_count_notrace(PREEMPT_ACTIVE);
3315 
3316 		/*
3317 		 * Check again in case we missed a preemption opportunity
3318 		 * between schedule and now.
3319 		 */
3320 		barrier();
3321 	} while (need_resched());
3322 }
3323 EXPORT_SYMBOL(preempt_schedule);
3324 
3325 /*
3326  * this is the entry point to schedule() from kernel preemption
3327  * off of irq context.
3328  * Note, that this is called and return with irqs disabled. This will
3329  * protect us against recursive calling from irq.
3330  */
preempt_schedule_irq(void)3331 asmlinkage void __sched preempt_schedule_irq(void)
3332 {
3333 	struct thread_info *ti = current_thread_info();
3334 
3335 	/* Catch callers which need to be fixed */
3336 	BUG_ON(ti->preempt_count || !irqs_disabled());
3337 
3338 	do {
3339 		add_preempt_count(PREEMPT_ACTIVE);
3340 		local_irq_enable();
3341 		__schedule();
3342 		local_irq_disable();
3343 		sub_preempt_count(PREEMPT_ACTIVE);
3344 
3345 		/*
3346 		 * Check again in case we missed a preemption opportunity
3347 		 * between schedule and now.
3348 		 */
3349 		barrier();
3350 	} while (need_resched());
3351 }
3352 
3353 #endif /* CONFIG_PREEMPT */
3354 
default_wake_function(wait_queue_t * curr,unsigned mode,int wake_flags,void * key)3355 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3356 			  void *key)
3357 {
3358 	return try_to_wake_up(curr->private, mode, wake_flags);
3359 }
3360 EXPORT_SYMBOL(default_wake_function);
3361 
3362 /*
3363  * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3364  * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3365  * number) then we wake all the non-exclusive tasks and one exclusive task.
3366  *
3367  * There are circumstances in which we can try to wake a task which has already
3368  * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3369  * zero in this (rare) case, and we handle it by continuing to scan the queue.
3370  */
__wake_up_common(wait_queue_head_t * q,unsigned int mode,int nr_exclusive,int wake_flags,void * key)3371 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3372 			int nr_exclusive, int wake_flags, void *key)
3373 {
3374 	wait_queue_t *curr, *next;
3375 
3376 	list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3377 		unsigned flags = curr->flags;
3378 
3379 		if (curr->func(curr, mode, wake_flags, key) &&
3380 				(flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3381 			break;
3382 	}
3383 }
3384 
3385 /**
3386  * __wake_up - wake up threads blocked on a waitqueue.
3387  * @q: the waitqueue
3388  * @mode: which threads
3389  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3390  * @key: is directly passed to the wakeup function
3391  *
3392  * It may be assumed that this function implies a write memory barrier before
3393  * changing the task state if and only if any tasks are woken up.
3394  */
__wake_up(wait_queue_head_t * q,unsigned int mode,int nr_exclusive,void * key)3395 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3396 			int nr_exclusive, void *key)
3397 {
3398 	unsigned long flags;
3399 
3400 	spin_lock_irqsave(&q->lock, flags);
3401 	__wake_up_common(q, mode, nr_exclusive, 0, key);
3402 	spin_unlock_irqrestore(&q->lock, flags);
3403 }
3404 EXPORT_SYMBOL(__wake_up);
3405 
3406 /*
3407  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3408  */
__wake_up_locked(wait_queue_head_t * q,unsigned int mode)3409 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3410 {
3411 	__wake_up_common(q, mode, 1, 0, NULL);
3412 }
3413 EXPORT_SYMBOL_GPL(__wake_up_locked);
3414 
__wake_up_locked_key(wait_queue_head_t * q,unsigned int mode,void * key)3415 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3416 {
3417 	__wake_up_common(q, mode, 1, 0, key);
3418 }
3419 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3420 
3421 /**
3422  * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3423  * @q: the waitqueue
3424  * @mode: which threads
3425  * @nr_exclusive: how many wake-one or wake-many threads to wake up
3426  * @key: opaque value to be passed to wakeup targets
3427  *
3428  * The sync wakeup differs that the waker knows that it will schedule
3429  * away soon, so while the target thread will be woken up, it will not
3430  * be migrated to another CPU - ie. the two threads are 'synchronized'
3431  * with each other. This can prevent needless bouncing between CPUs.
3432  *
3433  * On UP it can prevent extra preemption.
3434  *
3435  * It may be assumed that this function implies a write memory barrier before
3436  * changing the task state if and only if any tasks are woken up.
3437  */
__wake_up_sync_key(wait_queue_head_t * q,unsigned int mode,int nr_exclusive,void * key)3438 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3439 			int nr_exclusive, void *key)
3440 {
3441 	unsigned long flags;
3442 	int wake_flags = WF_SYNC;
3443 
3444 	if (unlikely(!q))
3445 		return;
3446 
3447 	if (unlikely(!nr_exclusive))
3448 		wake_flags = 0;
3449 
3450 	spin_lock_irqsave(&q->lock, flags);
3451 	__wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3452 	spin_unlock_irqrestore(&q->lock, flags);
3453 }
3454 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3455 
3456 /*
3457  * __wake_up_sync - see __wake_up_sync_key()
3458  */
__wake_up_sync(wait_queue_head_t * q,unsigned int mode,int nr_exclusive)3459 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3460 {
3461 	__wake_up_sync_key(q, mode, nr_exclusive, NULL);
3462 }
3463 EXPORT_SYMBOL_GPL(__wake_up_sync);	/* For internal use only */
3464 
3465 /**
3466  * complete: - signals a single thread waiting on this completion
3467  * @x:  holds the state of this particular completion
3468  *
3469  * This will wake up a single thread waiting on this completion. Threads will be
3470  * awakened in the same order in which they were queued.
3471  *
3472  * See also complete_all(), wait_for_completion() and related routines.
3473  *
3474  * It may be assumed that this function implies a write memory barrier before
3475  * changing the task state if and only if any tasks are woken up.
3476  */
complete(struct completion * x)3477 void complete(struct completion *x)
3478 {
3479 	unsigned long flags;
3480 
3481 	spin_lock_irqsave(&x->wait.lock, flags);
3482 	x->done++;
3483 	__wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3484 	spin_unlock_irqrestore(&x->wait.lock, flags);
3485 }
3486 EXPORT_SYMBOL(complete);
3487 
3488 /**
3489  * complete_all: - signals all threads waiting on this completion
3490  * @x:  holds the state of this particular completion
3491  *
3492  * This will wake up all threads waiting on this particular completion event.
3493  *
3494  * It may be assumed that this function implies a write memory barrier before
3495  * changing the task state if and only if any tasks are woken up.
3496  */
complete_all(struct completion * x)3497 void complete_all(struct completion *x)
3498 {
3499 	unsigned long flags;
3500 
3501 	spin_lock_irqsave(&x->wait.lock, flags);
3502 	x->done += UINT_MAX/2;
3503 	__wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3504 	spin_unlock_irqrestore(&x->wait.lock, flags);
3505 }
3506 EXPORT_SYMBOL(complete_all);
3507 
3508 static inline long __sched
do_wait_for_common(struct completion * x,long timeout,int state)3509 do_wait_for_common(struct completion *x, long timeout, int state)
3510 {
3511 	if (!x->done) {
3512 		DECLARE_WAITQUEUE(wait, current);
3513 
3514 		__add_wait_queue_tail_exclusive(&x->wait, &wait);
3515 		do {
3516 			if (signal_pending_state(state, current)) {
3517 				timeout = -ERESTARTSYS;
3518 				break;
3519 			}
3520 			__set_current_state(state);
3521 			spin_unlock_irq(&x->wait.lock);
3522 			timeout = schedule_timeout(timeout);
3523 			spin_lock_irq(&x->wait.lock);
3524 		} while (!x->done && timeout);
3525 		__remove_wait_queue(&x->wait, &wait);
3526 		if (!x->done)
3527 			return timeout;
3528 	}
3529 	x->done--;
3530 	return timeout ?: 1;
3531 }
3532 
3533 static long __sched
wait_for_common(struct completion * x,long timeout,int state)3534 wait_for_common(struct completion *x, long timeout, int state)
3535 {
3536 	might_sleep();
3537 
3538 	spin_lock_irq(&x->wait.lock);
3539 	timeout = do_wait_for_common(x, timeout, state);
3540 	spin_unlock_irq(&x->wait.lock);
3541 	return timeout;
3542 }
3543 
3544 /**
3545  * wait_for_completion: - waits for completion of a task
3546  * @x:  holds the state of this particular completion
3547  *
3548  * This waits to be signaled for completion of a specific task. It is NOT
3549  * interruptible and there is no timeout.
3550  *
3551  * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3552  * and interrupt capability. Also see complete().
3553  */
wait_for_completion(struct completion * x)3554 void __sched wait_for_completion(struct completion *x)
3555 {
3556 	wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3557 }
3558 EXPORT_SYMBOL(wait_for_completion);
3559 
3560 /**
3561  * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3562  * @x:  holds the state of this particular completion
3563  * @timeout:  timeout value in jiffies
3564  *
3565  * This waits for either a completion of a specific task to be signaled or for a
3566  * specified timeout to expire. The timeout is in jiffies. It is not
3567  * interruptible.
3568  *
3569  * The return value is 0 if timed out, and positive (at least 1, or number of
3570  * jiffies left till timeout) if completed.
3571  */
3572 unsigned long __sched
wait_for_completion_timeout(struct completion * x,unsigned long timeout)3573 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3574 {
3575 	return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3576 }
3577 EXPORT_SYMBOL(wait_for_completion_timeout);
3578 
3579 /**
3580  * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3581  * @x:  holds the state of this particular completion
3582  *
3583  * This waits for completion of a specific task to be signaled. It is
3584  * interruptible.
3585  *
3586  * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3587  */
wait_for_completion_interruptible(struct completion * x)3588 int __sched wait_for_completion_interruptible(struct completion *x)
3589 {
3590 	long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3591 	if (t == -ERESTARTSYS)
3592 		return t;
3593 	return 0;
3594 }
3595 EXPORT_SYMBOL(wait_for_completion_interruptible);
3596 
3597 /**
3598  * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3599  * @x:  holds the state of this particular completion
3600  * @timeout:  timeout value in jiffies
3601  *
3602  * This waits for either a completion of a specific task to be signaled or for a
3603  * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3604  *
3605  * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3606  * positive (at least 1, or number of jiffies left till timeout) if completed.
3607  */
3608 long __sched
wait_for_completion_interruptible_timeout(struct completion * x,unsigned long timeout)3609 wait_for_completion_interruptible_timeout(struct completion *x,
3610 					  unsigned long timeout)
3611 {
3612 	return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3613 }
3614 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3615 
3616 /**
3617  * wait_for_completion_killable: - waits for completion of a task (killable)
3618  * @x:  holds the state of this particular completion
3619  *
3620  * This waits to be signaled for completion of a specific task. It can be
3621  * interrupted by a kill signal.
3622  *
3623  * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3624  */
wait_for_completion_killable(struct completion * x)3625 int __sched wait_for_completion_killable(struct completion *x)
3626 {
3627 	long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3628 	if (t == -ERESTARTSYS)
3629 		return t;
3630 	return 0;
3631 }
3632 EXPORT_SYMBOL(wait_for_completion_killable);
3633 
3634 /**
3635  * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3636  * @x:  holds the state of this particular completion
3637  * @timeout:  timeout value in jiffies
3638  *
3639  * This waits for either a completion of a specific task to be
3640  * signaled or for a specified timeout to expire. It can be
3641  * interrupted by a kill signal. The timeout is in jiffies.
3642  *
3643  * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3644  * positive (at least 1, or number of jiffies left till timeout) if completed.
3645  */
3646 long __sched
wait_for_completion_killable_timeout(struct completion * x,unsigned long timeout)3647 wait_for_completion_killable_timeout(struct completion *x,
3648 				     unsigned long timeout)
3649 {
3650 	return wait_for_common(x, timeout, TASK_KILLABLE);
3651 }
3652 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3653 
3654 /**
3655  *	try_wait_for_completion - try to decrement a completion without blocking
3656  *	@x:	completion structure
3657  *
3658  *	Returns: 0 if a decrement cannot be done without blocking
3659  *		 1 if a decrement succeeded.
3660  *
3661  *	If a completion is being used as a counting completion,
3662  *	attempt to decrement the counter without blocking. This
3663  *	enables us to avoid waiting if the resource the completion
3664  *	is protecting is not available.
3665  */
try_wait_for_completion(struct completion * x)3666 bool try_wait_for_completion(struct completion *x)
3667 {
3668 	unsigned long flags;
3669 	int ret = 1;
3670 
3671 	spin_lock_irqsave(&x->wait.lock, flags);
3672 	if (!x->done)
3673 		ret = 0;
3674 	else
3675 		x->done--;
3676 	spin_unlock_irqrestore(&x->wait.lock, flags);
3677 	return ret;
3678 }
3679 EXPORT_SYMBOL(try_wait_for_completion);
3680 
3681 /**
3682  *	completion_done - Test to see if a completion has any waiters
3683  *	@x:	completion structure
3684  *
3685  *	Returns: 0 if there are waiters (wait_for_completion() in progress)
3686  *		 1 if there are no waiters.
3687  *
3688  */
completion_done(struct completion * x)3689 bool completion_done(struct completion *x)
3690 {
3691 	unsigned long flags;
3692 	int ret = 1;
3693 
3694 	spin_lock_irqsave(&x->wait.lock, flags);
3695 	if (!x->done)
3696 		ret = 0;
3697 	spin_unlock_irqrestore(&x->wait.lock, flags);
3698 	return ret;
3699 }
3700 EXPORT_SYMBOL(completion_done);
3701 
3702 static long __sched
sleep_on_common(wait_queue_head_t * q,int state,long timeout)3703 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3704 {
3705 	unsigned long flags;
3706 	wait_queue_t wait;
3707 
3708 	init_waitqueue_entry(&wait, current);
3709 
3710 	__set_current_state(state);
3711 
3712 	spin_lock_irqsave(&q->lock, flags);
3713 	__add_wait_queue(q, &wait);
3714 	spin_unlock(&q->lock);
3715 	timeout = schedule_timeout(timeout);
3716 	spin_lock_irq(&q->lock);
3717 	__remove_wait_queue(q, &wait);
3718 	spin_unlock_irqrestore(&q->lock, flags);
3719 
3720 	return timeout;
3721 }
3722 
interruptible_sleep_on(wait_queue_head_t * q)3723 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3724 {
3725 	sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3726 }
3727 EXPORT_SYMBOL(interruptible_sleep_on);
3728 
3729 long __sched
interruptible_sleep_on_timeout(wait_queue_head_t * q,long timeout)3730 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3731 {
3732 	return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3733 }
3734 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3735 
sleep_on(wait_queue_head_t * q)3736 void __sched sleep_on(wait_queue_head_t *q)
3737 {
3738 	sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3739 }
3740 EXPORT_SYMBOL(sleep_on);
3741 
sleep_on_timeout(wait_queue_head_t * q,long timeout)3742 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3743 {
3744 	return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3745 }
3746 EXPORT_SYMBOL(sleep_on_timeout);
3747 
3748 #ifdef CONFIG_RT_MUTEXES
3749 
3750 /*
3751  * rt_mutex_setprio - set the current priority of a task
3752  * @p: task
3753  * @prio: prio value (kernel-internal form)
3754  *
3755  * This function changes the 'effective' priority of a task. It does
3756  * not touch ->normal_prio like __setscheduler().
3757  *
3758  * Used by the rt_mutex code to implement priority inheritance logic.
3759  */
rt_mutex_setprio(struct task_struct * p,int prio)3760 void rt_mutex_setprio(struct task_struct *p, int prio)
3761 {
3762 	int oldprio, on_rq, running;
3763 	struct rq *rq;
3764 	const struct sched_class *prev_class;
3765 
3766 	BUG_ON(prio < 0 || prio > MAX_PRIO);
3767 
3768 	rq = __task_rq_lock(p);
3769 
3770 	trace_sched_pi_setprio(p, prio);
3771 	oldprio = p->prio;
3772 	prev_class = p->sched_class;
3773 	on_rq = p->on_rq;
3774 	running = task_current(rq, p);
3775 	if (on_rq)
3776 		dequeue_task(rq, p, 0);
3777 	if (running)
3778 		p->sched_class->put_prev_task(rq, p);
3779 
3780 	if (rt_prio(prio))
3781 		p->sched_class = &rt_sched_class;
3782 	else
3783 		p->sched_class = &fair_sched_class;
3784 
3785 	p->prio = prio;
3786 
3787 	if (running)
3788 		p->sched_class->set_curr_task(rq);
3789 	if (on_rq)
3790 		enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3791 
3792 	check_class_changed(rq, p, prev_class, oldprio);
3793 	__task_rq_unlock(rq);
3794 }
3795 
3796 #endif
3797 
set_user_nice(struct task_struct * p,long nice)3798 void set_user_nice(struct task_struct *p, long nice)
3799 {
3800 	int old_prio, delta, on_rq;
3801 	unsigned long flags;
3802 	struct rq *rq;
3803 
3804 	if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3805 		return;
3806 	/*
3807 	 * We have to be careful, if called from sys_setpriority(),
3808 	 * the task might be in the middle of scheduling on another CPU.
3809 	 */
3810 	rq = task_rq_lock(p, &flags);
3811 	/*
3812 	 * The RT priorities are set via sched_setscheduler(), but we still
3813 	 * allow the 'normal' nice value to be set - but as expected
3814 	 * it wont have any effect on scheduling until the task is
3815 	 * SCHED_FIFO/SCHED_RR:
3816 	 */
3817 	if (task_has_rt_policy(p)) {
3818 		p->static_prio = NICE_TO_PRIO(nice);
3819 		goto out_unlock;
3820 	}
3821 	on_rq = p->on_rq;
3822 	if (on_rq)
3823 		dequeue_task(rq, p, 0);
3824 
3825 	p->static_prio = NICE_TO_PRIO(nice);
3826 	set_load_weight(p);
3827 	old_prio = p->prio;
3828 	p->prio = effective_prio(p);
3829 	delta = p->prio - old_prio;
3830 
3831 	if (on_rq) {
3832 		enqueue_task(rq, p, 0);
3833 		/*
3834 		 * If the task increased its priority or is running and
3835 		 * lowered its priority, then reschedule its CPU:
3836 		 */
3837 		if (delta < 0 || (delta > 0 && task_running(rq, p)))
3838 			resched_task(rq->curr);
3839 	}
3840 out_unlock:
3841 	task_rq_unlock(rq, p, &flags);
3842 }
3843 EXPORT_SYMBOL(set_user_nice);
3844 
3845 /*
3846  * can_nice - check if a task can reduce its nice value
3847  * @p: task
3848  * @nice: nice value
3849  */
can_nice(const struct task_struct * p,const int nice)3850 int can_nice(const struct task_struct *p, const int nice)
3851 {
3852 	/* convert nice value [19,-20] to rlimit style value [1,40] */
3853 	int nice_rlim = 20 - nice;
3854 
3855 	return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3856 		capable(CAP_SYS_NICE));
3857 }
3858 
3859 #ifdef __ARCH_WANT_SYS_NICE
3860 
3861 /*
3862  * sys_nice - change the priority of the current process.
3863  * @increment: priority increment
3864  *
3865  * sys_setpriority is a more generic, but much slower function that
3866  * does similar things.
3867  */
SYSCALL_DEFINE1(nice,int,increment)3868 SYSCALL_DEFINE1(nice, int, increment)
3869 {
3870 	long nice, retval;
3871 
3872 	/*
3873 	 * Setpriority might change our priority at the same moment.
3874 	 * We don't have to worry. Conceptually one call occurs first
3875 	 * and we have a single winner.
3876 	 */
3877 	if (increment < -40)
3878 		increment = -40;
3879 	if (increment > 40)
3880 		increment = 40;
3881 
3882 	nice = TASK_NICE(current) + increment;
3883 	if (nice < -20)
3884 		nice = -20;
3885 	if (nice > 19)
3886 		nice = 19;
3887 
3888 	if (increment < 0 && !can_nice(current, nice))
3889 		return -EPERM;
3890 
3891 	retval = security_task_setnice(current, nice);
3892 	if (retval)
3893 		return retval;
3894 
3895 	set_user_nice(current, nice);
3896 	return 0;
3897 }
3898 
3899 #endif
3900 
3901 /**
3902  * task_prio - return the priority value of a given task.
3903  * @p: the task in question.
3904  *
3905  * This is the priority value as seen by users in /proc.
3906  * RT tasks are offset by -200. Normal tasks are centered
3907  * around 0, value goes from -16 to +15.
3908  */
task_prio(const struct task_struct * p)3909 int task_prio(const struct task_struct *p)
3910 {
3911 	return p->prio - MAX_RT_PRIO;
3912 }
3913 
3914 /**
3915  * task_nice - return the nice value of a given task.
3916  * @p: the task in question.
3917  */
task_nice(const struct task_struct * p)3918 int task_nice(const struct task_struct *p)
3919 {
3920 	return TASK_NICE(p);
3921 }
3922 EXPORT_SYMBOL(task_nice);
3923 
3924 /**
3925  * idle_cpu - is a given cpu idle currently?
3926  * @cpu: the processor in question.
3927  */
idle_cpu(int cpu)3928 int idle_cpu(int cpu)
3929 {
3930 	struct rq *rq = cpu_rq(cpu);
3931 
3932 	if (rq->curr != rq->idle)
3933 		return 0;
3934 
3935 	if (rq->nr_running)
3936 		return 0;
3937 
3938 #ifdef CONFIG_SMP
3939 	if (!llist_empty(&rq->wake_list))
3940 		return 0;
3941 #endif
3942 
3943 	return 1;
3944 }
3945 
3946 /**
3947  * idle_task - return the idle task for a given cpu.
3948  * @cpu: the processor in question.
3949  */
idle_task(int cpu)3950 struct task_struct *idle_task(int cpu)
3951 {
3952 	return cpu_rq(cpu)->idle;
3953 }
3954 
3955 /**
3956  * find_process_by_pid - find a process with a matching PID value.
3957  * @pid: the pid in question.
3958  */
find_process_by_pid(pid_t pid)3959 static struct task_struct *find_process_by_pid(pid_t pid)
3960 {
3961 	return pid ? find_task_by_vpid(pid) : current;
3962 }
3963 
3964 /* Actually do priority change: must hold rq lock. */
3965 static void
__setscheduler(struct rq * rq,struct task_struct * p,int policy,int prio)3966 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3967 {
3968 	p->policy = policy;
3969 	p->rt_priority = prio;
3970 	p->normal_prio = normal_prio(p);
3971 	/* we are holding p->pi_lock already */
3972 	p->prio = rt_mutex_getprio(p);
3973 	if (rt_prio(p->prio))
3974 		p->sched_class = &rt_sched_class;
3975 	else
3976 		p->sched_class = &fair_sched_class;
3977 	set_load_weight(p);
3978 }
3979 
3980 /*
3981  * check the target process has a UID that matches the current process's
3982  */
check_same_owner(struct task_struct * p)3983 static bool check_same_owner(struct task_struct *p)
3984 {
3985 	const struct cred *cred = current_cred(), *pcred;
3986 	bool match;
3987 
3988 	rcu_read_lock();
3989 	pcred = __task_cred(p);
3990 	if (cred->user->user_ns == pcred->user->user_ns)
3991 		match = (cred->euid == pcred->euid ||
3992 			 cred->euid == pcred->uid);
3993 	else
3994 		match = false;
3995 	rcu_read_unlock();
3996 	return match;
3997 }
3998 
__sched_setscheduler(struct task_struct * p,int policy,const struct sched_param * param,bool user)3999 static int __sched_setscheduler(struct task_struct *p, int policy,
4000 				const struct sched_param *param, bool user)
4001 {
4002 	int retval, oldprio, oldpolicy = -1, on_rq, running;
4003 	unsigned long flags;
4004 	const struct sched_class *prev_class;
4005 	struct rq *rq;
4006 	int reset_on_fork;
4007 
4008 	/* may grab non-irq protected spin_locks */
4009 	BUG_ON(in_interrupt());
4010 recheck:
4011 	/* double check policy once rq lock held */
4012 	if (policy < 0) {
4013 		reset_on_fork = p->sched_reset_on_fork;
4014 		policy = oldpolicy = p->policy;
4015 	} else {
4016 		reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4017 		policy &= ~SCHED_RESET_ON_FORK;
4018 
4019 		if (policy != SCHED_FIFO && policy != SCHED_RR &&
4020 				policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4021 				policy != SCHED_IDLE)
4022 			return -EINVAL;
4023 	}
4024 
4025 	/*
4026 	 * Valid priorities for SCHED_FIFO and SCHED_RR are
4027 	 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4028 	 * SCHED_BATCH and SCHED_IDLE is 0.
4029 	 */
4030 	if (param->sched_priority < 0 ||
4031 	    (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4032 	    (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4033 		return -EINVAL;
4034 	if (rt_policy(policy) != (param->sched_priority != 0))
4035 		return -EINVAL;
4036 
4037 	/*
4038 	 * Allow unprivileged RT tasks to decrease priority:
4039 	 */
4040 	if (user && !capable(CAP_SYS_NICE)) {
4041 		if (rt_policy(policy)) {
4042 			unsigned long rlim_rtprio =
4043 					task_rlimit(p, RLIMIT_RTPRIO);
4044 
4045 			/* can't set/change the rt policy */
4046 			if (policy != p->policy && !rlim_rtprio)
4047 				return -EPERM;
4048 
4049 			/* can't increase priority */
4050 			if (param->sched_priority > p->rt_priority &&
4051 			    param->sched_priority > rlim_rtprio)
4052 				return -EPERM;
4053 		}
4054 
4055 		/*
4056 		 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4057 		 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4058 		 */
4059 		if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4060 			if (!can_nice(p, TASK_NICE(p)))
4061 				return -EPERM;
4062 		}
4063 
4064 		/* can't change other user's priorities */
4065 		if (!check_same_owner(p))
4066 			return -EPERM;
4067 
4068 		/* Normal users shall not reset the sched_reset_on_fork flag */
4069 		if (p->sched_reset_on_fork && !reset_on_fork)
4070 			return -EPERM;
4071 	}
4072 
4073 	if (user) {
4074 		retval = security_task_setscheduler(p);
4075 		if (retval)
4076 			return retval;
4077 	}
4078 
4079 	/*
4080 	 * make sure no PI-waiters arrive (or leave) while we are
4081 	 * changing the priority of the task:
4082 	 *
4083 	 * To be able to change p->policy safely, the appropriate
4084 	 * runqueue lock must be held.
4085 	 */
4086 	rq = task_rq_lock(p, &flags);
4087 
4088 	/*
4089 	 * Changing the policy of the stop threads its a very bad idea
4090 	 */
4091 	if (p == rq->stop) {
4092 		task_rq_unlock(rq, p, &flags);
4093 		return -EINVAL;
4094 	}
4095 
4096 	/*
4097 	 * If not changing anything there's no need to proceed further:
4098 	 */
4099 	if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4100 			param->sched_priority == p->rt_priority))) {
4101 
4102 		__task_rq_unlock(rq);
4103 		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4104 		return 0;
4105 	}
4106 
4107 #ifdef CONFIG_RT_GROUP_SCHED
4108 	if (user) {
4109 		/*
4110 		 * Do not allow realtime tasks into groups that have no runtime
4111 		 * assigned.
4112 		 */
4113 		if (rt_bandwidth_enabled() && rt_policy(policy) &&
4114 				task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4115 				!task_group_is_autogroup(task_group(p))) {
4116 			task_rq_unlock(rq, p, &flags);
4117 			return -EPERM;
4118 		}
4119 	}
4120 #endif
4121 
4122 	/* recheck policy now with rq lock held */
4123 	if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4124 		policy = oldpolicy = -1;
4125 		task_rq_unlock(rq, p, &flags);
4126 		goto recheck;
4127 	}
4128 	on_rq = p->on_rq;
4129 	running = task_current(rq, p);
4130 	if (on_rq)
4131 		dequeue_task(rq, p, 0);
4132 	if (running)
4133 		p->sched_class->put_prev_task(rq, p);
4134 
4135 	p->sched_reset_on_fork = reset_on_fork;
4136 
4137 	oldprio = p->prio;
4138 	prev_class = p->sched_class;
4139 	__setscheduler(rq, p, policy, param->sched_priority);
4140 
4141 	if (running)
4142 		p->sched_class->set_curr_task(rq);
4143 	if (on_rq)
4144 		enqueue_task(rq, p, 0);
4145 
4146 	check_class_changed(rq, p, prev_class, oldprio);
4147 	task_rq_unlock(rq, p, &flags);
4148 
4149 	rt_mutex_adjust_pi(p);
4150 
4151 	return 0;
4152 }
4153 
4154 /**
4155  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4156  * @p: the task in question.
4157  * @policy: new policy.
4158  * @param: structure containing the new RT priority.
4159  *
4160  * NOTE that the task may be already dead.
4161  */
sched_setscheduler(struct task_struct * p,int policy,const struct sched_param * param)4162 int sched_setscheduler(struct task_struct *p, int policy,
4163 		       const struct sched_param *param)
4164 {
4165 	return __sched_setscheduler(p, policy, param, true);
4166 }
4167 EXPORT_SYMBOL_GPL(sched_setscheduler);
4168 
4169 /**
4170  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4171  * @p: the task in question.
4172  * @policy: new policy.
4173  * @param: structure containing the new RT priority.
4174  *
4175  * Just like sched_setscheduler, only don't bother checking if the
4176  * current context has permission.  For example, this is needed in
4177  * stop_machine(): we create temporary high priority worker threads,
4178  * but our caller might not have that capability.
4179  */
sched_setscheduler_nocheck(struct task_struct * p,int policy,const struct sched_param * param)4180 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4181 			       const struct sched_param *param)
4182 {
4183 	return __sched_setscheduler(p, policy, param, false);
4184 }
4185 
4186 static int
do_sched_setscheduler(pid_t pid,int policy,struct sched_param __user * param)4187 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4188 {
4189 	struct sched_param lparam;
4190 	struct task_struct *p;
4191 	int retval;
4192 
4193 	if (!param || pid < 0)
4194 		return -EINVAL;
4195 	if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4196 		return -EFAULT;
4197 
4198 	rcu_read_lock();
4199 	retval = -ESRCH;
4200 	p = find_process_by_pid(pid);
4201 	if (p != NULL)
4202 		retval = sched_setscheduler(p, policy, &lparam);
4203 	rcu_read_unlock();
4204 
4205 	return retval;
4206 }
4207 
4208 /**
4209  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4210  * @pid: the pid in question.
4211  * @policy: new policy.
4212  * @param: structure containing the new RT priority.
4213  */
SYSCALL_DEFINE3(sched_setscheduler,pid_t,pid,int,policy,struct sched_param __user *,param)4214 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4215 		struct sched_param __user *, param)
4216 {
4217 	/* negative values for policy are not valid */
4218 	if (policy < 0)
4219 		return -EINVAL;
4220 
4221 	return do_sched_setscheduler(pid, policy, param);
4222 }
4223 
4224 /**
4225  * sys_sched_setparam - set/change the RT priority of a thread
4226  * @pid: the pid in question.
4227  * @param: structure containing the new RT priority.
4228  */
SYSCALL_DEFINE2(sched_setparam,pid_t,pid,struct sched_param __user *,param)4229 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4230 {
4231 	return do_sched_setscheduler(pid, -1, param);
4232 }
4233 
4234 /**
4235  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4236  * @pid: the pid in question.
4237  */
SYSCALL_DEFINE1(sched_getscheduler,pid_t,pid)4238 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4239 {
4240 	struct task_struct *p;
4241 	int retval;
4242 
4243 	if (pid < 0)
4244 		return -EINVAL;
4245 
4246 	retval = -ESRCH;
4247 	rcu_read_lock();
4248 	p = find_process_by_pid(pid);
4249 	if (p) {
4250 		retval = security_task_getscheduler(p);
4251 		if (!retval)
4252 			retval = p->policy
4253 				| (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4254 	}
4255 	rcu_read_unlock();
4256 	return retval;
4257 }
4258 
4259 /**
4260  * sys_sched_getparam - get the RT priority of a thread
4261  * @pid: the pid in question.
4262  * @param: structure containing the RT priority.
4263  */
SYSCALL_DEFINE2(sched_getparam,pid_t,pid,struct sched_param __user *,param)4264 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4265 {
4266 	struct sched_param lp;
4267 	struct task_struct *p;
4268 	int retval;
4269 
4270 	if (!param || pid < 0)
4271 		return -EINVAL;
4272 
4273 	rcu_read_lock();
4274 	p = find_process_by_pid(pid);
4275 	retval = -ESRCH;
4276 	if (!p)
4277 		goto out_unlock;
4278 
4279 	retval = security_task_getscheduler(p);
4280 	if (retval)
4281 		goto out_unlock;
4282 
4283 	lp.sched_priority = p->rt_priority;
4284 	rcu_read_unlock();
4285 
4286 	/*
4287 	 * This one might sleep, we cannot do it with a spinlock held ...
4288 	 */
4289 	retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4290 
4291 	return retval;
4292 
4293 out_unlock:
4294 	rcu_read_unlock();
4295 	return retval;
4296 }
4297 
sched_setaffinity(pid_t pid,const struct cpumask * in_mask)4298 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4299 {
4300 	cpumask_var_t cpus_allowed, new_mask;
4301 	struct task_struct *p;
4302 	int retval;
4303 
4304 	get_online_cpus();
4305 	rcu_read_lock();
4306 
4307 	p = find_process_by_pid(pid);
4308 	if (!p) {
4309 		rcu_read_unlock();
4310 		put_online_cpus();
4311 		return -ESRCH;
4312 	}
4313 
4314 	/* Prevent p going away */
4315 	get_task_struct(p);
4316 	rcu_read_unlock();
4317 
4318 	if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4319 		retval = -ENOMEM;
4320 		goto out_put_task;
4321 	}
4322 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4323 		retval = -ENOMEM;
4324 		goto out_free_cpus_allowed;
4325 	}
4326 	retval = -EPERM;
4327 	if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4328 		goto out_unlock;
4329 
4330 	retval = security_task_setscheduler(p);
4331 	if (retval)
4332 		goto out_unlock;
4333 
4334 	cpuset_cpus_allowed(p, cpus_allowed);
4335 	cpumask_and(new_mask, in_mask, cpus_allowed);
4336 again:
4337 	retval = set_cpus_allowed_ptr(p, new_mask);
4338 
4339 	if (!retval) {
4340 		cpuset_cpus_allowed(p, cpus_allowed);
4341 		if (!cpumask_subset(new_mask, cpus_allowed)) {
4342 			/*
4343 			 * We must have raced with a concurrent cpuset
4344 			 * update. Just reset the cpus_allowed to the
4345 			 * cpuset's cpus_allowed
4346 			 */
4347 			cpumask_copy(new_mask, cpus_allowed);
4348 			goto again;
4349 		}
4350 	}
4351 out_unlock:
4352 	free_cpumask_var(new_mask);
4353 out_free_cpus_allowed:
4354 	free_cpumask_var(cpus_allowed);
4355 out_put_task:
4356 	put_task_struct(p);
4357 	put_online_cpus();
4358 	return retval;
4359 }
4360 
get_user_cpu_mask(unsigned long __user * user_mask_ptr,unsigned len,struct cpumask * new_mask)4361 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4362 			     struct cpumask *new_mask)
4363 {
4364 	if (len < cpumask_size())
4365 		cpumask_clear(new_mask);
4366 	else if (len > cpumask_size())
4367 		len = cpumask_size();
4368 
4369 	return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4370 }
4371 
4372 /**
4373  * sys_sched_setaffinity - set the cpu affinity of a process
4374  * @pid: pid of the process
4375  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4376  * @user_mask_ptr: user-space pointer to the new cpu mask
4377  */
SYSCALL_DEFINE3(sched_setaffinity,pid_t,pid,unsigned int,len,unsigned long __user *,user_mask_ptr)4378 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4379 		unsigned long __user *, user_mask_ptr)
4380 {
4381 	cpumask_var_t new_mask;
4382 	int retval;
4383 
4384 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4385 		return -ENOMEM;
4386 
4387 	retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4388 	if (retval == 0)
4389 		retval = sched_setaffinity(pid, new_mask);
4390 	free_cpumask_var(new_mask);
4391 	return retval;
4392 }
4393 
sched_getaffinity(pid_t pid,struct cpumask * mask)4394 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4395 {
4396 	struct task_struct *p;
4397 	unsigned long flags;
4398 	int retval;
4399 
4400 	get_online_cpus();
4401 	rcu_read_lock();
4402 
4403 	retval = -ESRCH;
4404 	p = find_process_by_pid(pid);
4405 	if (!p)
4406 		goto out_unlock;
4407 
4408 	retval = security_task_getscheduler(p);
4409 	if (retval)
4410 		goto out_unlock;
4411 
4412 	raw_spin_lock_irqsave(&p->pi_lock, flags);
4413 	cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4414 	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4415 
4416 out_unlock:
4417 	rcu_read_unlock();
4418 	put_online_cpus();
4419 
4420 	return retval;
4421 }
4422 
4423 /**
4424  * sys_sched_getaffinity - get the cpu affinity of a process
4425  * @pid: pid of the process
4426  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4427  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4428  */
SYSCALL_DEFINE3(sched_getaffinity,pid_t,pid,unsigned int,len,unsigned long __user *,user_mask_ptr)4429 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4430 		unsigned long __user *, user_mask_ptr)
4431 {
4432 	int ret;
4433 	cpumask_var_t mask;
4434 
4435 	if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4436 		return -EINVAL;
4437 	if (len & (sizeof(unsigned long)-1))
4438 		return -EINVAL;
4439 
4440 	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4441 		return -ENOMEM;
4442 
4443 	ret = sched_getaffinity(pid, mask);
4444 	if (ret == 0) {
4445 		size_t retlen = min_t(size_t, len, cpumask_size());
4446 
4447 		if (copy_to_user(user_mask_ptr, mask, retlen))
4448 			ret = -EFAULT;
4449 		else
4450 			ret = retlen;
4451 	}
4452 	free_cpumask_var(mask);
4453 
4454 	return ret;
4455 }
4456 
4457 /**
4458  * sys_sched_yield - yield the current processor to other threads.
4459  *
4460  * This function yields the current CPU to other tasks. If there are no
4461  * other threads running on this CPU then this function will return.
4462  */
SYSCALL_DEFINE0(sched_yield)4463 SYSCALL_DEFINE0(sched_yield)
4464 {
4465 	struct rq *rq = this_rq_lock();
4466 
4467 	schedstat_inc(rq, yld_count);
4468 	current->sched_class->yield_task(rq);
4469 
4470 	/*
4471 	 * Since we are going to call schedule() anyway, there's
4472 	 * no need to preempt or enable interrupts:
4473 	 */
4474 	__release(rq->lock);
4475 	spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4476 	do_raw_spin_unlock(&rq->lock);
4477 	preempt_enable_no_resched();
4478 
4479 	schedule();
4480 
4481 	return 0;
4482 }
4483 
should_resched(void)4484 static inline int should_resched(void)
4485 {
4486 	return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4487 }
4488 
__cond_resched(void)4489 static void __cond_resched(void)
4490 {
4491 	add_preempt_count(PREEMPT_ACTIVE);
4492 	__schedule();
4493 	sub_preempt_count(PREEMPT_ACTIVE);
4494 }
4495 
_cond_resched(void)4496 int __sched _cond_resched(void)
4497 {
4498 	if (should_resched()) {
4499 		__cond_resched();
4500 		return 1;
4501 	}
4502 	return 0;
4503 }
4504 EXPORT_SYMBOL(_cond_resched);
4505 
4506 /*
4507  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4508  * call schedule, and on return reacquire the lock.
4509  *
4510  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4511  * operations here to prevent schedule() from being called twice (once via
4512  * spin_unlock(), once by hand).
4513  */
__cond_resched_lock(spinlock_t * lock)4514 int __cond_resched_lock(spinlock_t *lock)
4515 {
4516 	int resched = should_resched();
4517 	int ret = 0;
4518 
4519 	lockdep_assert_held(lock);
4520 
4521 	if (spin_needbreak(lock) || resched) {
4522 		spin_unlock(lock);
4523 		if (resched)
4524 			__cond_resched();
4525 		else
4526 			cpu_relax();
4527 		ret = 1;
4528 		spin_lock(lock);
4529 	}
4530 	return ret;
4531 }
4532 EXPORT_SYMBOL(__cond_resched_lock);
4533 
__cond_resched_softirq(void)4534 int __sched __cond_resched_softirq(void)
4535 {
4536 	BUG_ON(!in_softirq());
4537 
4538 	if (should_resched()) {
4539 		local_bh_enable();
4540 		__cond_resched();
4541 		local_bh_disable();
4542 		return 1;
4543 	}
4544 	return 0;
4545 }
4546 EXPORT_SYMBOL(__cond_resched_softirq);
4547 
4548 /**
4549  * yield - yield the current processor to other threads.
4550  *
4551  * This is a shortcut for kernel-space yielding - it marks the
4552  * thread runnable and calls sys_sched_yield().
4553  */
yield(void)4554 void __sched yield(void)
4555 {
4556 	set_current_state(TASK_RUNNING);
4557 	sys_sched_yield();
4558 }
4559 EXPORT_SYMBOL(yield);
4560 
4561 /**
4562  * yield_to - yield the current processor to another thread in
4563  * your thread group, or accelerate that thread toward the
4564  * processor it's on.
4565  * @p: target task
4566  * @preempt: whether task preemption is allowed or not
4567  *
4568  * It's the caller's job to ensure that the target task struct
4569  * can't go away on us before we can do any checks.
4570  *
4571  * Returns true if we indeed boosted the target task.
4572  */
yield_to(struct task_struct * p,bool preempt)4573 bool __sched yield_to(struct task_struct *p, bool preempt)
4574 {
4575 	struct task_struct *curr = current;
4576 	struct rq *rq, *p_rq;
4577 	unsigned long flags;
4578 	bool yielded = 0;
4579 
4580 	local_irq_save(flags);
4581 	rq = this_rq();
4582 
4583 again:
4584 	p_rq = task_rq(p);
4585 	double_rq_lock(rq, p_rq);
4586 	while (task_rq(p) != p_rq) {
4587 		double_rq_unlock(rq, p_rq);
4588 		goto again;
4589 	}
4590 
4591 	if (!curr->sched_class->yield_to_task)
4592 		goto out;
4593 
4594 	if (curr->sched_class != p->sched_class)
4595 		goto out;
4596 
4597 	if (task_running(p_rq, p) || p->state)
4598 		goto out;
4599 
4600 	yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4601 	if (yielded) {
4602 		schedstat_inc(rq, yld_count);
4603 		/*
4604 		 * Make p's CPU reschedule; pick_next_entity takes care of
4605 		 * fairness.
4606 		 */
4607 		if (preempt && rq != p_rq)
4608 			resched_task(p_rq->curr);
4609 	} else {
4610 		/*
4611 		 * We might have set it in task_yield_fair(), but are
4612 		 * not going to schedule(), so don't want to skip
4613 		 * the next update.
4614 		 */
4615 		rq->skip_clock_update = 0;
4616 	}
4617 
4618 out:
4619 	double_rq_unlock(rq, p_rq);
4620 	local_irq_restore(flags);
4621 
4622 	if (yielded)
4623 		schedule();
4624 
4625 	return yielded;
4626 }
4627 EXPORT_SYMBOL_GPL(yield_to);
4628 
4629 /*
4630  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4631  * that process accounting knows that this is a task in IO wait state.
4632  */
io_schedule(void)4633 void __sched io_schedule(void)
4634 {
4635 	struct rq *rq = raw_rq();
4636 
4637 	delayacct_blkio_start();
4638 	atomic_inc(&rq->nr_iowait);
4639 	blk_flush_plug(current);
4640 	current->in_iowait = 1;
4641 	schedule();
4642 	current->in_iowait = 0;
4643 	atomic_dec(&rq->nr_iowait);
4644 	delayacct_blkio_end();
4645 }
4646 EXPORT_SYMBOL(io_schedule);
4647 
io_schedule_timeout(long timeout)4648 long __sched io_schedule_timeout(long timeout)
4649 {
4650 	struct rq *rq = raw_rq();
4651 	long ret;
4652 
4653 	delayacct_blkio_start();
4654 	atomic_inc(&rq->nr_iowait);
4655 	blk_flush_plug(current);
4656 	current->in_iowait = 1;
4657 	ret = schedule_timeout(timeout);
4658 	current->in_iowait = 0;
4659 	atomic_dec(&rq->nr_iowait);
4660 	delayacct_blkio_end();
4661 	return ret;
4662 }
4663 
4664 /**
4665  * sys_sched_get_priority_max - return maximum RT priority.
4666  * @policy: scheduling class.
4667  *
4668  * this syscall returns the maximum rt_priority that can be used
4669  * by a given scheduling class.
4670  */
SYSCALL_DEFINE1(sched_get_priority_max,int,policy)4671 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4672 {
4673 	int ret = -EINVAL;
4674 
4675 	switch (policy) {
4676 	case SCHED_FIFO:
4677 	case SCHED_RR:
4678 		ret = MAX_USER_RT_PRIO-1;
4679 		break;
4680 	case SCHED_NORMAL:
4681 	case SCHED_BATCH:
4682 	case SCHED_IDLE:
4683 		ret = 0;
4684 		break;
4685 	}
4686 	return ret;
4687 }
4688 
4689 /**
4690  * sys_sched_get_priority_min - return minimum RT priority.
4691  * @policy: scheduling class.
4692  *
4693  * this syscall returns the minimum rt_priority that can be used
4694  * by a given scheduling class.
4695  */
SYSCALL_DEFINE1(sched_get_priority_min,int,policy)4696 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4697 {
4698 	int ret = -EINVAL;
4699 
4700 	switch (policy) {
4701 	case SCHED_FIFO:
4702 	case SCHED_RR:
4703 		ret = 1;
4704 		break;
4705 	case SCHED_NORMAL:
4706 	case SCHED_BATCH:
4707 	case SCHED_IDLE:
4708 		ret = 0;
4709 	}
4710 	return ret;
4711 }
4712 
4713 /**
4714  * sys_sched_rr_get_interval - return the default timeslice of a process.
4715  * @pid: pid of the process.
4716  * @interval: userspace pointer to the timeslice value.
4717  *
4718  * this syscall writes the default timeslice value of a given process
4719  * into the user-space timespec buffer. A value of '0' means infinity.
4720  */
SYSCALL_DEFINE2(sched_rr_get_interval,pid_t,pid,struct timespec __user *,interval)4721 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4722 		struct timespec __user *, interval)
4723 {
4724 	struct task_struct *p;
4725 	unsigned int time_slice;
4726 	unsigned long flags;
4727 	struct rq *rq;
4728 	int retval;
4729 	struct timespec t;
4730 
4731 	if (pid < 0)
4732 		return -EINVAL;
4733 
4734 	retval = -ESRCH;
4735 	rcu_read_lock();
4736 	p = find_process_by_pid(pid);
4737 	if (!p)
4738 		goto out_unlock;
4739 
4740 	retval = security_task_getscheduler(p);
4741 	if (retval)
4742 		goto out_unlock;
4743 
4744 	rq = task_rq_lock(p, &flags);
4745 	time_slice = p->sched_class->get_rr_interval(rq, p);
4746 	task_rq_unlock(rq, p, &flags);
4747 
4748 	rcu_read_unlock();
4749 	jiffies_to_timespec(time_slice, &t);
4750 	retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4751 	return retval;
4752 
4753 out_unlock:
4754 	rcu_read_unlock();
4755 	return retval;
4756 }
4757 
4758 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4759 
sched_show_task(struct task_struct * p)4760 void sched_show_task(struct task_struct *p)
4761 {
4762 	unsigned long free = 0;
4763 	unsigned state;
4764 
4765 	state = p->state ? __ffs(p->state) + 1 : 0;
4766 	printk(KERN_INFO "%-15.15s %c", p->comm,
4767 		state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4768 #if BITS_PER_LONG == 32
4769 	if (state == TASK_RUNNING)
4770 		printk(KERN_CONT " running  ");
4771 	else
4772 		printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4773 #else
4774 	if (state == TASK_RUNNING)
4775 		printk(KERN_CONT "  running task    ");
4776 	else
4777 		printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4778 #endif
4779 #ifdef CONFIG_DEBUG_STACK_USAGE
4780 	free = stack_not_used(p);
4781 #endif
4782 	printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4783 		task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4784 		(unsigned long)task_thread_info(p)->flags);
4785 
4786 	show_stack(p, NULL);
4787 }
4788 
show_state_filter(unsigned long state_filter)4789 void show_state_filter(unsigned long state_filter)
4790 {
4791 	struct task_struct *g, *p;
4792 
4793 #if BITS_PER_LONG == 32
4794 	printk(KERN_INFO
4795 		"  task                PC stack   pid father\n");
4796 #else
4797 	printk(KERN_INFO
4798 		"  task                        PC stack   pid father\n");
4799 #endif
4800 	rcu_read_lock();
4801 	do_each_thread(g, p) {
4802 		/*
4803 		 * reset the NMI-timeout, listing all files on a slow
4804 		 * console might take a lot of time:
4805 		 */
4806 		touch_nmi_watchdog();
4807 		if (!state_filter || (p->state & state_filter))
4808 			sched_show_task(p);
4809 	} while_each_thread(g, p);
4810 
4811 	touch_all_softlockup_watchdogs();
4812 
4813 #ifdef CONFIG_SCHED_DEBUG
4814 	sysrq_sched_debug_show();
4815 #endif
4816 	rcu_read_unlock();
4817 	/*
4818 	 * Only show locks if all tasks are dumped:
4819 	 */
4820 	if (!state_filter)
4821 		debug_show_all_locks();
4822 }
4823 
init_idle_bootup_task(struct task_struct * idle)4824 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4825 {
4826 	idle->sched_class = &idle_sched_class;
4827 }
4828 
4829 /**
4830  * init_idle - set up an idle thread for a given CPU
4831  * @idle: task in question
4832  * @cpu: cpu the idle task belongs to
4833  *
4834  * NOTE: this function does not set the idle thread's NEED_RESCHED
4835  * flag, to make booting more robust.
4836  */
init_idle(struct task_struct * idle,int cpu)4837 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4838 {
4839 	struct rq *rq = cpu_rq(cpu);
4840 	unsigned long flags;
4841 
4842 	raw_spin_lock_irqsave(&rq->lock, flags);
4843 
4844 	__sched_fork(idle);
4845 	idle->state = TASK_RUNNING;
4846 	idle->se.exec_start = sched_clock();
4847 
4848 	do_set_cpus_allowed(idle, cpumask_of(cpu));
4849 	/*
4850 	 * We're having a chicken and egg problem, even though we are
4851 	 * holding rq->lock, the cpu isn't yet set to this cpu so the
4852 	 * lockdep check in task_group() will fail.
4853 	 *
4854 	 * Similar case to sched_fork(). / Alternatively we could
4855 	 * use task_rq_lock() here and obtain the other rq->lock.
4856 	 *
4857 	 * Silence PROVE_RCU
4858 	 */
4859 	rcu_read_lock();
4860 	__set_task_cpu(idle, cpu);
4861 	rcu_read_unlock();
4862 
4863 	rq->curr = rq->idle = idle;
4864 #if defined(CONFIG_SMP)
4865 	idle->on_cpu = 1;
4866 #endif
4867 	raw_spin_unlock_irqrestore(&rq->lock, flags);
4868 
4869 	/* Set the preempt count _outside_ the spinlocks! */
4870 	task_thread_info(idle)->preempt_count = 0;
4871 
4872 	/*
4873 	 * The idle tasks have their own, simple scheduling class:
4874 	 */
4875 	idle->sched_class = &idle_sched_class;
4876 	ftrace_graph_init_idle_task(idle, cpu);
4877 #if defined(CONFIG_SMP)
4878 	sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4879 #endif
4880 }
4881 
4882 #ifdef CONFIG_SMP
do_set_cpus_allowed(struct task_struct * p,const struct cpumask * new_mask)4883 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4884 {
4885 	if (p->sched_class && p->sched_class->set_cpus_allowed)
4886 		p->sched_class->set_cpus_allowed(p, new_mask);
4887 
4888 	cpumask_copy(&p->cpus_allowed, new_mask);
4889 	p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
4890 }
4891 
4892 /*
4893  * This is how migration works:
4894  *
4895  * 1) we invoke migration_cpu_stop() on the target CPU using
4896  *    stop_one_cpu().
4897  * 2) stopper starts to run (implicitly forcing the migrated thread
4898  *    off the CPU)
4899  * 3) it checks whether the migrated task is still in the wrong runqueue.
4900  * 4) if it's in the wrong runqueue then the migration thread removes
4901  *    it and puts it into the right queue.
4902  * 5) stopper completes and stop_one_cpu() returns and the migration
4903  *    is done.
4904  */
4905 
4906 /*
4907  * Change a given task's CPU affinity. Migrate the thread to a
4908  * proper CPU and schedule it away if the CPU it's executing on
4909  * is removed from the allowed bitmask.
4910  *
4911  * NOTE: the caller must have a valid reference to the task, the
4912  * task must not exit() & deallocate itself prematurely. The
4913  * call is not atomic; no spinlocks may be held.
4914  */
set_cpus_allowed_ptr(struct task_struct * p,const struct cpumask * new_mask)4915 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4916 {
4917 	unsigned long flags;
4918 	struct rq *rq;
4919 	unsigned int dest_cpu;
4920 	int ret = 0;
4921 
4922 	rq = task_rq_lock(p, &flags);
4923 
4924 	if (cpumask_equal(&p->cpus_allowed, new_mask))
4925 		goto out;
4926 
4927 	if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4928 		ret = -EINVAL;
4929 		goto out;
4930 	}
4931 
4932 	if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4933 		ret = -EINVAL;
4934 		goto out;
4935 	}
4936 
4937 	do_set_cpus_allowed(p, new_mask);
4938 
4939 	/* Can the task run on the task's current CPU? If so, we're done */
4940 	if (cpumask_test_cpu(task_cpu(p), new_mask))
4941 		goto out;
4942 
4943 	dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4944 	if (p->on_rq) {
4945 		struct migration_arg arg = { p, dest_cpu };
4946 		/* Need help from migration thread: drop lock and wait. */
4947 		task_rq_unlock(rq, p, &flags);
4948 		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4949 		tlb_migrate_finish(p->mm);
4950 		return 0;
4951 	}
4952 out:
4953 	task_rq_unlock(rq, p, &flags);
4954 
4955 	return ret;
4956 }
4957 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4958 
4959 /*
4960  * Move (not current) task off this cpu, onto dest cpu. We're doing
4961  * this because either it can't run here any more (set_cpus_allowed()
4962  * away from this CPU, or CPU going down), or because we're
4963  * attempting to rebalance this task on exec (sched_exec).
4964  *
4965  * So we race with normal scheduler movements, but that's OK, as long
4966  * as the task is no longer on this CPU.
4967  *
4968  * Returns non-zero if task was successfully migrated.
4969  */
__migrate_task(struct task_struct * p,int src_cpu,int dest_cpu)4970 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4971 {
4972 	struct rq *rq_dest, *rq_src;
4973 	int ret = 0;
4974 
4975 	if (unlikely(!cpu_active(dest_cpu)))
4976 		return ret;
4977 
4978 	rq_src = cpu_rq(src_cpu);
4979 	rq_dest = cpu_rq(dest_cpu);
4980 
4981 	raw_spin_lock(&p->pi_lock);
4982 	double_rq_lock(rq_src, rq_dest);
4983 	/* Already moved. */
4984 	if (task_cpu(p) != src_cpu)
4985 		goto done;
4986 	/* Affinity changed (again). */
4987 	if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4988 		goto fail;
4989 
4990 	/*
4991 	 * If we're not on a rq, the next wake-up will ensure we're
4992 	 * placed properly.
4993 	 */
4994 	if (p->on_rq) {
4995 		dequeue_task(rq_src, p, 0);
4996 		set_task_cpu(p, dest_cpu);
4997 		enqueue_task(rq_dest, p, 0);
4998 		check_preempt_curr(rq_dest, p, 0);
4999 	}
5000 done:
5001 	ret = 1;
5002 fail:
5003 	double_rq_unlock(rq_src, rq_dest);
5004 	raw_spin_unlock(&p->pi_lock);
5005 	return ret;
5006 }
5007 
5008 /*
5009  * migration_cpu_stop - this will be executed by a highprio stopper thread
5010  * and performs thread migration by bumping thread off CPU then
5011  * 'pushing' onto another runqueue.
5012  */
migration_cpu_stop(void * data)5013 static int migration_cpu_stop(void *data)
5014 {
5015 	struct migration_arg *arg = data;
5016 
5017 	/*
5018 	 * The original target cpu might have gone down and we might
5019 	 * be on another cpu but it doesn't matter.
5020 	 */
5021 	local_irq_disable();
5022 	__migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5023 	local_irq_enable();
5024 	return 0;
5025 }
5026 
5027 #ifdef CONFIG_HOTPLUG_CPU
5028 
5029 /*
5030  * Ensures that the idle task is using init_mm right before its cpu goes
5031  * offline.
5032  */
idle_task_exit(void)5033 void idle_task_exit(void)
5034 {
5035 	struct mm_struct *mm = current->active_mm;
5036 
5037 	BUG_ON(cpu_online(smp_processor_id()));
5038 
5039 	if (mm != &init_mm)
5040 		switch_mm(mm, &init_mm, current);
5041 	mmdrop(mm);
5042 }
5043 
5044 /*
5045  * While a dead CPU has no uninterruptible tasks queued at this point,
5046  * it might still have a nonzero ->nr_uninterruptible counter, because
5047  * for performance reasons the counter is not stricly tracking tasks to
5048  * their home CPUs. So we just add the counter to another CPU's counter,
5049  * to keep the global sum constant after CPU-down:
5050  */
migrate_nr_uninterruptible(struct rq * rq_src)5051 static void migrate_nr_uninterruptible(struct rq *rq_src)
5052 {
5053 	struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5054 
5055 	rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5056 	rq_src->nr_uninterruptible = 0;
5057 }
5058 
5059 /*
5060  * remove the tasks which were accounted by rq from calc_load_tasks.
5061  */
calc_global_load_remove(struct rq * rq)5062 static void calc_global_load_remove(struct rq *rq)
5063 {
5064 	atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5065 	rq->calc_load_active = 0;
5066 }
5067 
5068 /*
5069  * Migrate all tasks from the rq, sleeping tasks will be migrated by
5070  * try_to_wake_up()->select_task_rq().
5071  *
5072  * Called with rq->lock held even though we'er in stop_machine() and
5073  * there's no concurrency possible, we hold the required locks anyway
5074  * because of lock validation efforts.
5075  */
migrate_tasks(unsigned int dead_cpu)5076 static void migrate_tasks(unsigned int dead_cpu)
5077 {
5078 	struct rq *rq = cpu_rq(dead_cpu);
5079 	struct task_struct *next, *stop = rq->stop;
5080 	int dest_cpu;
5081 
5082 	/*
5083 	 * Fudge the rq selection such that the below task selection loop
5084 	 * doesn't get stuck on the currently eligible stop task.
5085 	 *
5086 	 * We're currently inside stop_machine() and the rq is either stuck
5087 	 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5088 	 * either way we should never end up calling schedule() until we're
5089 	 * done here.
5090 	 */
5091 	rq->stop = NULL;
5092 
5093 	/* Ensure any throttled groups are reachable by pick_next_task */
5094 	unthrottle_offline_cfs_rqs(rq);
5095 
5096 	for ( ; ; ) {
5097 		/*
5098 		 * There's this thread running, bail when that's the only
5099 		 * remaining thread.
5100 		 */
5101 		if (rq->nr_running == 1)
5102 			break;
5103 
5104 		next = pick_next_task(rq);
5105 		BUG_ON(!next);
5106 		next->sched_class->put_prev_task(rq, next);
5107 
5108 		/* Find suitable destination for @next, with force if needed. */
5109 		dest_cpu = select_fallback_rq(dead_cpu, next);
5110 		raw_spin_unlock(&rq->lock);
5111 
5112 		__migrate_task(next, dead_cpu, dest_cpu);
5113 
5114 		raw_spin_lock(&rq->lock);
5115 	}
5116 
5117 	rq->stop = stop;
5118 }
5119 
5120 #endif /* CONFIG_HOTPLUG_CPU */
5121 
5122 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5123 
5124 static struct ctl_table sd_ctl_dir[] = {
5125 	{
5126 		.procname	= "sched_domain",
5127 		.mode		= 0555,
5128 	},
5129 	{}
5130 };
5131 
5132 static struct ctl_table sd_ctl_root[] = {
5133 	{
5134 		.procname	= "kernel",
5135 		.mode		= 0555,
5136 		.child		= sd_ctl_dir,
5137 	},
5138 	{}
5139 };
5140 
sd_alloc_ctl_entry(int n)5141 static struct ctl_table *sd_alloc_ctl_entry(int n)
5142 {
5143 	struct ctl_table *entry =
5144 		kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5145 
5146 	return entry;
5147 }
5148 
sd_free_ctl_entry(struct ctl_table ** tablep)5149 static void sd_free_ctl_entry(struct ctl_table **tablep)
5150 {
5151 	struct ctl_table *entry;
5152 
5153 	/*
5154 	 * In the intermediate directories, both the child directory and
5155 	 * procname are dynamically allocated and could fail but the mode
5156 	 * will always be set. In the lowest directory the names are
5157 	 * static strings and all have proc handlers.
5158 	 */
5159 	for (entry = *tablep; entry->mode; entry++) {
5160 		if (entry->child)
5161 			sd_free_ctl_entry(&entry->child);
5162 		if (entry->proc_handler == NULL)
5163 			kfree(entry->procname);
5164 	}
5165 
5166 	kfree(*tablep);
5167 	*tablep = NULL;
5168 }
5169 
5170 static void
set_table_entry(struct ctl_table * entry,const char * procname,void * data,int maxlen,umode_t mode,proc_handler * proc_handler)5171 set_table_entry(struct ctl_table *entry,
5172 		const char *procname, void *data, int maxlen,
5173 		umode_t mode, proc_handler *proc_handler)
5174 {
5175 	entry->procname = procname;
5176 	entry->data = data;
5177 	entry->maxlen = maxlen;
5178 	entry->mode = mode;
5179 	entry->proc_handler = proc_handler;
5180 }
5181 
5182 static struct ctl_table *
sd_alloc_ctl_domain_table(struct sched_domain * sd)5183 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5184 {
5185 	struct ctl_table *table = sd_alloc_ctl_entry(13);
5186 
5187 	if (table == NULL)
5188 		return NULL;
5189 
5190 	set_table_entry(&table[0], "min_interval", &sd->min_interval,
5191 		sizeof(long), 0644, proc_doulongvec_minmax);
5192 	set_table_entry(&table[1], "max_interval", &sd->max_interval,
5193 		sizeof(long), 0644, proc_doulongvec_minmax);
5194 	set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5195 		sizeof(int), 0644, proc_dointvec_minmax);
5196 	set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5197 		sizeof(int), 0644, proc_dointvec_minmax);
5198 	set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5199 		sizeof(int), 0644, proc_dointvec_minmax);
5200 	set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5201 		sizeof(int), 0644, proc_dointvec_minmax);
5202 	set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5203 		sizeof(int), 0644, proc_dointvec_minmax);
5204 	set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5205 		sizeof(int), 0644, proc_dointvec_minmax);
5206 	set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5207 		sizeof(int), 0644, proc_dointvec_minmax);
5208 	set_table_entry(&table[9], "cache_nice_tries",
5209 		&sd->cache_nice_tries,
5210 		sizeof(int), 0644, proc_dointvec_minmax);
5211 	set_table_entry(&table[10], "flags", &sd->flags,
5212 		sizeof(int), 0644, proc_dointvec_minmax);
5213 	set_table_entry(&table[11], "name", sd->name,
5214 		CORENAME_MAX_SIZE, 0444, proc_dostring);
5215 	/* &table[12] is terminator */
5216 
5217 	return table;
5218 }
5219 
sd_alloc_ctl_cpu_table(int cpu)5220 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5221 {
5222 	struct ctl_table *entry, *table;
5223 	struct sched_domain *sd;
5224 	int domain_num = 0, i;
5225 	char buf[32];
5226 
5227 	for_each_domain(cpu, sd)
5228 		domain_num++;
5229 	entry = table = sd_alloc_ctl_entry(domain_num + 1);
5230 	if (table == NULL)
5231 		return NULL;
5232 
5233 	i = 0;
5234 	for_each_domain(cpu, sd) {
5235 		snprintf(buf, 32, "domain%d", i);
5236 		entry->procname = kstrdup(buf, GFP_KERNEL);
5237 		entry->mode = 0555;
5238 		entry->child = sd_alloc_ctl_domain_table(sd);
5239 		entry++;
5240 		i++;
5241 	}
5242 	return table;
5243 }
5244 
5245 static struct ctl_table_header *sd_sysctl_header;
register_sched_domain_sysctl(void)5246 static void register_sched_domain_sysctl(void)
5247 {
5248 	int i, cpu_num = num_possible_cpus();
5249 	struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5250 	char buf[32];
5251 
5252 	WARN_ON(sd_ctl_dir[0].child);
5253 	sd_ctl_dir[0].child = entry;
5254 
5255 	if (entry == NULL)
5256 		return;
5257 
5258 	for_each_possible_cpu(i) {
5259 		snprintf(buf, 32, "cpu%d", i);
5260 		entry->procname = kstrdup(buf, GFP_KERNEL);
5261 		entry->mode = 0555;
5262 		entry->child = sd_alloc_ctl_cpu_table(i);
5263 		entry++;
5264 	}
5265 
5266 	WARN_ON(sd_sysctl_header);
5267 	sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5268 }
5269 
5270 /* may be called multiple times per register */
unregister_sched_domain_sysctl(void)5271 static void unregister_sched_domain_sysctl(void)
5272 {
5273 	if (sd_sysctl_header)
5274 		unregister_sysctl_table(sd_sysctl_header);
5275 	sd_sysctl_header = NULL;
5276 	if (sd_ctl_dir[0].child)
5277 		sd_free_ctl_entry(&sd_ctl_dir[0].child);
5278 }
5279 #else
register_sched_domain_sysctl(void)5280 static void register_sched_domain_sysctl(void)
5281 {
5282 }
unregister_sched_domain_sysctl(void)5283 static void unregister_sched_domain_sysctl(void)
5284 {
5285 }
5286 #endif
5287 
set_rq_online(struct rq * rq)5288 static void set_rq_online(struct rq *rq)
5289 {
5290 	if (!rq->online) {
5291 		const struct sched_class *class;
5292 
5293 		cpumask_set_cpu(rq->cpu, rq->rd->online);
5294 		rq->online = 1;
5295 
5296 		for_each_class(class) {
5297 			if (class->rq_online)
5298 				class->rq_online(rq);
5299 		}
5300 	}
5301 }
5302 
set_rq_offline(struct rq * rq)5303 static void set_rq_offline(struct rq *rq)
5304 {
5305 	if (rq->online) {
5306 		const struct sched_class *class;
5307 
5308 		for_each_class(class) {
5309 			if (class->rq_offline)
5310 				class->rq_offline(rq);
5311 		}
5312 
5313 		cpumask_clear_cpu(rq->cpu, rq->rd->online);
5314 		rq->online = 0;
5315 	}
5316 }
5317 
5318 /*
5319  * migration_call - callback that gets triggered when a CPU is added.
5320  * Here we can start up the necessary migration thread for the new CPU.
5321  */
5322 static int __cpuinit
migration_call(struct notifier_block * nfb,unsigned long action,void * hcpu)5323 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5324 {
5325 	int cpu = (long)hcpu;
5326 	unsigned long flags;
5327 	struct rq *rq = cpu_rq(cpu);
5328 
5329 	switch (action & ~CPU_TASKS_FROZEN) {
5330 
5331 	case CPU_UP_PREPARE:
5332 		rq->calc_load_update = calc_load_update;
5333 		break;
5334 
5335 	case CPU_ONLINE:
5336 		/* Update our root-domain */
5337 		raw_spin_lock_irqsave(&rq->lock, flags);
5338 		if (rq->rd) {
5339 			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5340 
5341 			set_rq_online(rq);
5342 		}
5343 		raw_spin_unlock_irqrestore(&rq->lock, flags);
5344 		break;
5345 
5346 #ifdef CONFIG_HOTPLUG_CPU
5347 	case CPU_DYING:
5348 		sched_ttwu_pending();
5349 		/* Update our root-domain */
5350 		raw_spin_lock_irqsave(&rq->lock, flags);
5351 		if (rq->rd) {
5352 			BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5353 			set_rq_offline(rq);
5354 		}
5355 		migrate_tasks(cpu);
5356 		BUG_ON(rq->nr_running != 1); /* the migration thread */
5357 		raw_spin_unlock_irqrestore(&rq->lock, flags);
5358 
5359 		migrate_nr_uninterruptible(rq);
5360 		calc_global_load_remove(rq);
5361 		break;
5362 #endif
5363 	}
5364 
5365 	update_max_interval();
5366 
5367 	return NOTIFY_OK;
5368 }
5369 
5370 /*
5371  * Register at high priority so that task migration (migrate_all_tasks)
5372  * happens before everything else.  This has to be lower priority than
5373  * the notifier in the perf_event subsystem, though.
5374  */
5375 static struct notifier_block __cpuinitdata migration_notifier = {
5376 	.notifier_call = migration_call,
5377 	.priority = CPU_PRI_MIGRATION,
5378 };
5379 
sched_cpu_active(struct notifier_block * nfb,unsigned long action,void * hcpu)5380 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5381 				      unsigned long action, void *hcpu)
5382 {
5383 	switch (action & ~CPU_TASKS_FROZEN) {
5384 	case CPU_ONLINE:
5385 	case CPU_DOWN_FAILED:
5386 		set_cpu_active((long)hcpu, true);
5387 		return NOTIFY_OK;
5388 	default:
5389 		return NOTIFY_DONE;
5390 	}
5391 }
5392 
sched_cpu_inactive(struct notifier_block * nfb,unsigned long action,void * hcpu)5393 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5394 					unsigned long action, void *hcpu)
5395 {
5396 	switch (action & ~CPU_TASKS_FROZEN) {
5397 	case CPU_DOWN_PREPARE:
5398 		set_cpu_active((long)hcpu, false);
5399 		return NOTIFY_OK;
5400 	default:
5401 		return NOTIFY_DONE;
5402 	}
5403 }
5404 
migration_init(void)5405 static int __init migration_init(void)
5406 {
5407 	void *cpu = (void *)(long)smp_processor_id();
5408 	int err;
5409 
5410 	/* Initialize migration for the boot CPU */
5411 	err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5412 	BUG_ON(err == NOTIFY_BAD);
5413 	migration_call(&migration_notifier, CPU_ONLINE, cpu);
5414 	register_cpu_notifier(&migration_notifier);
5415 
5416 	/* Register cpu active notifiers */
5417 	cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5418 	cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5419 
5420 	return 0;
5421 }
5422 early_initcall(migration_init);
5423 #endif
5424 
5425 #ifdef CONFIG_SMP
5426 
5427 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5428 
5429 #ifdef CONFIG_SCHED_DEBUG
5430 
5431 static __read_mostly int sched_domain_debug_enabled;
5432 
sched_domain_debug_setup(char * str)5433 static int __init sched_domain_debug_setup(char *str)
5434 {
5435 	sched_domain_debug_enabled = 1;
5436 
5437 	return 0;
5438 }
5439 early_param("sched_debug", sched_domain_debug_setup);
5440 
sched_domain_debug_one(struct sched_domain * sd,int cpu,int level,struct cpumask * groupmask)5441 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5442 				  struct cpumask *groupmask)
5443 {
5444 	struct sched_group *group = sd->groups;
5445 	char str[256];
5446 
5447 	cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5448 	cpumask_clear(groupmask);
5449 
5450 	printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5451 
5452 	if (!(sd->flags & SD_LOAD_BALANCE)) {
5453 		printk("does not load-balance\n");
5454 		if (sd->parent)
5455 			printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5456 					" has parent");
5457 		return -1;
5458 	}
5459 
5460 	printk(KERN_CONT "span %s level %s\n", str, sd->name);
5461 
5462 	if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5463 		printk(KERN_ERR "ERROR: domain->span does not contain "
5464 				"CPU%d\n", cpu);
5465 	}
5466 	if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5467 		printk(KERN_ERR "ERROR: domain->groups does not contain"
5468 				" CPU%d\n", cpu);
5469 	}
5470 
5471 	printk(KERN_DEBUG "%*s groups:", level + 1, "");
5472 	do {
5473 		if (!group) {
5474 			printk("\n");
5475 			printk(KERN_ERR "ERROR: group is NULL\n");
5476 			break;
5477 		}
5478 
5479 		if (!group->sgp->power) {
5480 			printk(KERN_CONT "\n");
5481 			printk(KERN_ERR "ERROR: domain->cpu_power not "
5482 					"set\n");
5483 			break;
5484 		}
5485 
5486 		if (!cpumask_weight(sched_group_cpus(group))) {
5487 			printk(KERN_CONT "\n");
5488 			printk(KERN_ERR "ERROR: empty group\n");
5489 			break;
5490 		}
5491 
5492 		if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5493 			printk(KERN_CONT "\n");
5494 			printk(KERN_ERR "ERROR: repeated CPUs\n");
5495 			break;
5496 		}
5497 
5498 		cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5499 
5500 		cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5501 
5502 		printk(KERN_CONT " %s", str);
5503 		if (group->sgp->power != SCHED_POWER_SCALE) {
5504 			printk(KERN_CONT " (cpu_power = %d)",
5505 				group->sgp->power);
5506 		}
5507 
5508 		group = group->next;
5509 	} while (group != sd->groups);
5510 	printk(KERN_CONT "\n");
5511 
5512 	if (!cpumask_equal(sched_domain_span(sd), groupmask))
5513 		printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5514 
5515 	if (sd->parent &&
5516 	    !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5517 		printk(KERN_ERR "ERROR: parent span is not a superset "
5518 			"of domain->span\n");
5519 	return 0;
5520 }
5521 
sched_domain_debug(struct sched_domain * sd,int cpu)5522 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5523 {
5524 	int level = 0;
5525 
5526 	if (!sched_domain_debug_enabled)
5527 		return;
5528 
5529 	if (!sd) {
5530 		printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5531 		return;
5532 	}
5533 
5534 	printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5535 
5536 	for (;;) {
5537 		if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5538 			break;
5539 		level++;
5540 		sd = sd->parent;
5541 		if (!sd)
5542 			break;
5543 	}
5544 }
5545 #else /* !CONFIG_SCHED_DEBUG */
5546 # define sched_domain_debug(sd, cpu) do { } while (0)
5547 #endif /* CONFIG_SCHED_DEBUG */
5548 
sd_degenerate(struct sched_domain * sd)5549 static int sd_degenerate(struct sched_domain *sd)
5550 {
5551 	if (cpumask_weight(sched_domain_span(sd)) == 1)
5552 		return 1;
5553 
5554 	/* Following flags need at least 2 groups */
5555 	if (sd->flags & (SD_LOAD_BALANCE |
5556 			 SD_BALANCE_NEWIDLE |
5557 			 SD_BALANCE_FORK |
5558 			 SD_BALANCE_EXEC |
5559 			 SD_SHARE_CPUPOWER |
5560 			 SD_SHARE_PKG_RESOURCES)) {
5561 		if (sd->groups != sd->groups->next)
5562 			return 0;
5563 	}
5564 
5565 	/* Following flags don't use groups */
5566 	if (sd->flags & (SD_WAKE_AFFINE))
5567 		return 0;
5568 
5569 	return 1;
5570 }
5571 
5572 static int
sd_parent_degenerate(struct sched_domain * sd,struct sched_domain * parent)5573 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5574 {
5575 	unsigned long cflags = sd->flags, pflags = parent->flags;
5576 
5577 	if (sd_degenerate(parent))
5578 		return 1;
5579 
5580 	if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5581 		return 0;
5582 
5583 	/* Flags needing groups don't count if only 1 group in parent */
5584 	if (parent->groups == parent->groups->next) {
5585 		pflags &= ~(SD_LOAD_BALANCE |
5586 				SD_BALANCE_NEWIDLE |
5587 				SD_BALANCE_FORK |
5588 				SD_BALANCE_EXEC |
5589 				SD_SHARE_CPUPOWER |
5590 				SD_SHARE_PKG_RESOURCES);
5591 		if (nr_node_ids == 1)
5592 			pflags &= ~SD_SERIALIZE;
5593 	}
5594 	if (~cflags & pflags)
5595 		return 0;
5596 
5597 	return 1;
5598 }
5599 
free_rootdomain(struct rcu_head * rcu)5600 static void free_rootdomain(struct rcu_head *rcu)
5601 {
5602 	struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5603 
5604 	cpupri_cleanup(&rd->cpupri);
5605 	free_cpumask_var(rd->rto_mask);
5606 	free_cpumask_var(rd->online);
5607 	free_cpumask_var(rd->span);
5608 	kfree(rd);
5609 }
5610 
rq_attach_root(struct rq * rq,struct root_domain * rd)5611 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5612 {
5613 	struct root_domain *old_rd = NULL;
5614 	unsigned long flags;
5615 
5616 	raw_spin_lock_irqsave(&rq->lock, flags);
5617 
5618 	if (rq->rd) {
5619 		old_rd = rq->rd;
5620 
5621 		if (cpumask_test_cpu(rq->cpu, old_rd->online))
5622 			set_rq_offline(rq);
5623 
5624 		cpumask_clear_cpu(rq->cpu, old_rd->span);
5625 
5626 		/*
5627 		 * If we dont want to free the old_rt yet then
5628 		 * set old_rd to NULL to skip the freeing later
5629 		 * in this function:
5630 		 */
5631 		if (!atomic_dec_and_test(&old_rd->refcount))
5632 			old_rd = NULL;
5633 	}
5634 
5635 	atomic_inc(&rd->refcount);
5636 	rq->rd = rd;
5637 
5638 	cpumask_set_cpu(rq->cpu, rd->span);
5639 	if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5640 		set_rq_online(rq);
5641 
5642 	raw_spin_unlock_irqrestore(&rq->lock, flags);
5643 
5644 	if (old_rd)
5645 		call_rcu_sched(&old_rd->rcu, free_rootdomain);
5646 }
5647 
init_rootdomain(struct root_domain * rd)5648 static int init_rootdomain(struct root_domain *rd)
5649 {
5650 	memset(rd, 0, sizeof(*rd));
5651 
5652 	if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5653 		goto out;
5654 	if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5655 		goto free_span;
5656 	if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5657 		goto free_online;
5658 
5659 	if (cpupri_init(&rd->cpupri) != 0)
5660 		goto free_rto_mask;
5661 	return 0;
5662 
5663 free_rto_mask:
5664 	free_cpumask_var(rd->rto_mask);
5665 free_online:
5666 	free_cpumask_var(rd->online);
5667 free_span:
5668 	free_cpumask_var(rd->span);
5669 out:
5670 	return -ENOMEM;
5671 }
5672 
5673 /*
5674  * By default the system creates a single root-domain with all cpus as
5675  * members (mimicking the global state we have today).
5676  */
5677 struct root_domain def_root_domain;
5678 
init_defrootdomain(void)5679 static void init_defrootdomain(void)
5680 {
5681 	init_rootdomain(&def_root_domain);
5682 
5683 	atomic_set(&def_root_domain.refcount, 1);
5684 }
5685 
alloc_rootdomain(void)5686 static struct root_domain *alloc_rootdomain(void)
5687 {
5688 	struct root_domain *rd;
5689 
5690 	rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5691 	if (!rd)
5692 		return NULL;
5693 
5694 	if (init_rootdomain(rd) != 0) {
5695 		kfree(rd);
5696 		return NULL;
5697 	}
5698 
5699 	return rd;
5700 }
5701 
free_sched_groups(struct sched_group * sg,int free_sgp)5702 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5703 {
5704 	struct sched_group *tmp, *first;
5705 
5706 	if (!sg)
5707 		return;
5708 
5709 	first = sg;
5710 	do {
5711 		tmp = sg->next;
5712 
5713 		if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5714 			kfree(sg->sgp);
5715 
5716 		kfree(sg);
5717 		sg = tmp;
5718 	} while (sg != first);
5719 }
5720 
free_sched_domain(struct rcu_head * rcu)5721 static void free_sched_domain(struct rcu_head *rcu)
5722 {
5723 	struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5724 
5725 	/*
5726 	 * If its an overlapping domain it has private groups, iterate and
5727 	 * nuke them all.
5728 	 */
5729 	if (sd->flags & SD_OVERLAP) {
5730 		free_sched_groups(sd->groups, 1);
5731 	} else if (atomic_dec_and_test(&sd->groups->ref)) {
5732 		kfree(sd->groups->sgp);
5733 		kfree(sd->groups);
5734 	}
5735 	kfree(sd);
5736 }
5737 
destroy_sched_domain(struct sched_domain * sd,int cpu)5738 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5739 {
5740 	call_rcu(&sd->rcu, free_sched_domain);
5741 }
5742 
destroy_sched_domains(struct sched_domain * sd,int cpu)5743 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5744 {
5745 	for (; sd; sd = sd->parent)
5746 		destroy_sched_domain(sd, cpu);
5747 }
5748 
5749 /*
5750  * Keep a special pointer to the highest sched_domain that has
5751  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5752  * allows us to avoid some pointer chasing select_idle_sibling().
5753  *
5754  * Also keep a unique ID per domain (we use the first cpu number in
5755  * the cpumask of the domain), this allows us to quickly tell if
5756  * two cpus are in the same cache domain, see ttwu_share_cache().
5757  */
5758 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5759 DEFINE_PER_CPU(int, sd_llc_id);
5760 
update_top_cache_domain(int cpu)5761 static void update_top_cache_domain(int cpu)
5762 {
5763 	struct sched_domain *sd;
5764 	int id = cpu;
5765 
5766 	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5767 	if (sd)
5768 		id = cpumask_first(sched_domain_span(sd));
5769 
5770 	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5771 	per_cpu(sd_llc_id, cpu) = id;
5772 }
5773 
5774 /*
5775  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5776  * hold the hotplug lock.
5777  */
5778 static void
cpu_attach_domain(struct sched_domain * sd,struct root_domain * rd,int cpu)5779 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5780 {
5781 	struct rq *rq = cpu_rq(cpu);
5782 	struct sched_domain *tmp;
5783 
5784 	/* Remove the sched domains which do not contribute to scheduling. */
5785 	for (tmp = sd; tmp; ) {
5786 		struct sched_domain *parent = tmp->parent;
5787 		if (!parent)
5788 			break;
5789 
5790 		if (sd_parent_degenerate(tmp, parent)) {
5791 			tmp->parent = parent->parent;
5792 			if (parent->parent)
5793 				parent->parent->child = tmp;
5794 			destroy_sched_domain(parent, cpu);
5795 		} else
5796 			tmp = tmp->parent;
5797 	}
5798 
5799 	if (sd && sd_degenerate(sd)) {
5800 		tmp = sd;
5801 		sd = sd->parent;
5802 		destroy_sched_domain(tmp, cpu);
5803 		if (sd)
5804 			sd->child = NULL;
5805 	}
5806 
5807 	sched_domain_debug(sd, cpu);
5808 
5809 	rq_attach_root(rq, rd);
5810 	tmp = rq->sd;
5811 	rcu_assign_pointer(rq->sd, sd);
5812 	destroy_sched_domains(tmp, cpu);
5813 
5814 	update_top_cache_domain(cpu);
5815 }
5816 
5817 /* cpus with isolated domains */
5818 static cpumask_var_t cpu_isolated_map;
5819 
5820 /* Setup the mask of cpus configured for isolated domains */
isolated_cpu_setup(char * str)5821 static int __init isolated_cpu_setup(char *str)
5822 {
5823 	alloc_bootmem_cpumask_var(&cpu_isolated_map);
5824 	cpulist_parse(str, cpu_isolated_map);
5825 	return 1;
5826 }
5827 
5828 __setup("isolcpus=", isolated_cpu_setup);
5829 
5830 #ifdef CONFIG_NUMA
5831 
5832 /**
5833  * find_next_best_node - find the next node to include in a sched_domain
5834  * @node: node whose sched_domain we're building
5835  * @used_nodes: nodes already in the sched_domain
5836  *
5837  * Find the next node to include in a given scheduling domain. Simply
5838  * finds the closest node not already in the @used_nodes map.
5839  *
5840  * Should use nodemask_t.
5841  */
find_next_best_node(int node,nodemask_t * used_nodes)5842 static int find_next_best_node(int node, nodemask_t *used_nodes)
5843 {
5844 	int i, n, val, min_val, best_node = -1;
5845 
5846 	min_val = INT_MAX;
5847 
5848 	for (i = 0; i < nr_node_ids; i++) {
5849 		/* Start at @node */
5850 		n = (node + i) % nr_node_ids;
5851 
5852 		if (!nr_cpus_node(n))
5853 			continue;
5854 
5855 		/* Skip already used nodes */
5856 		if (node_isset(n, *used_nodes))
5857 			continue;
5858 
5859 		/* Simple min distance search */
5860 		val = node_distance(node, n);
5861 
5862 		if (val < min_val) {
5863 			min_val = val;
5864 			best_node = n;
5865 		}
5866 	}
5867 
5868 	if (best_node != -1)
5869 		node_set(best_node, *used_nodes);
5870 	return best_node;
5871 }
5872 
5873 /**
5874  * sched_domain_node_span - get a cpumask for a node's sched_domain
5875  * @node: node whose cpumask we're constructing
5876  * @span: resulting cpumask
5877  *
5878  * Given a node, construct a good cpumask for its sched_domain to span. It
5879  * should be one that prevents unnecessary balancing, but also spreads tasks
5880  * out optimally.
5881  */
sched_domain_node_span(int node,struct cpumask * span)5882 static void sched_domain_node_span(int node, struct cpumask *span)
5883 {
5884 	nodemask_t used_nodes;
5885 	int i;
5886 
5887 	cpumask_clear(span);
5888 	nodes_clear(used_nodes);
5889 
5890 	cpumask_or(span, span, cpumask_of_node(node));
5891 	node_set(node, used_nodes);
5892 
5893 	for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5894 		int next_node = find_next_best_node(node, &used_nodes);
5895 		if (next_node < 0)
5896 			break;
5897 		cpumask_or(span, span, cpumask_of_node(next_node));
5898 	}
5899 }
5900 
cpu_node_mask(int cpu)5901 static const struct cpumask *cpu_node_mask(int cpu)
5902 {
5903 	lockdep_assert_held(&sched_domains_mutex);
5904 
5905 	sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
5906 
5907 	return sched_domains_tmpmask;
5908 }
5909 
cpu_allnodes_mask(int cpu)5910 static const struct cpumask *cpu_allnodes_mask(int cpu)
5911 {
5912 	return cpu_possible_mask;
5913 }
5914 #endif /* CONFIG_NUMA */
5915 
cpu_cpu_mask(int cpu)5916 static const struct cpumask *cpu_cpu_mask(int cpu)
5917 {
5918 	return cpumask_of_node(cpu_to_node(cpu));
5919 }
5920 
5921 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5922 
5923 struct sd_data {
5924 	struct sched_domain **__percpu sd;
5925 	struct sched_group **__percpu sg;
5926 	struct sched_group_power **__percpu sgp;
5927 };
5928 
5929 struct s_data {
5930 	struct sched_domain ** __percpu sd;
5931 	struct root_domain	*rd;
5932 };
5933 
5934 enum s_alloc {
5935 	sa_rootdomain,
5936 	sa_sd,
5937 	sa_sd_storage,
5938 	sa_none,
5939 };
5940 
5941 struct sched_domain_topology_level;
5942 
5943 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5944 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5945 
5946 #define SDTL_OVERLAP	0x01
5947 
5948 struct sched_domain_topology_level {
5949 	sched_domain_init_f init;
5950 	sched_domain_mask_f mask;
5951 	int		    flags;
5952 	struct sd_data      data;
5953 };
5954 
5955 static int
build_overlap_sched_groups(struct sched_domain * sd,int cpu)5956 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5957 {
5958 	struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5959 	const struct cpumask *span = sched_domain_span(sd);
5960 	struct cpumask *covered = sched_domains_tmpmask;
5961 	struct sd_data *sdd = sd->private;
5962 	struct sched_domain *child;
5963 	int i;
5964 
5965 	cpumask_clear(covered);
5966 
5967 	for_each_cpu(i, span) {
5968 		struct cpumask *sg_span;
5969 
5970 		if (cpumask_test_cpu(i, covered))
5971 			continue;
5972 
5973 		sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5974 				GFP_KERNEL, cpu_to_node(cpu));
5975 
5976 		if (!sg)
5977 			goto fail;
5978 
5979 		sg_span = sched_group_cpus(sg);
5980 
5981 		child = *per_cpu_ptr(sdd->sd, i);
5982 		if (child->child) {
5983 			child = child->child;
5984 			cpumask_copy(sg_span, sched_domain_span(child));
5985 		} else
5986 			cpumask_set_cpu(i, sg_span);
5987 
5988 		cpumask_or(covered, covered, sg_span);
5989 
5990 		sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
5991 		atomic_inc(&sg->sgp->ref);
5992 
5993 		if (cpumask_test_cpu(cpu, sg_span))
5994 			groups = sg;
5995 
5996 		if (!first)
5997 			first = sg;
5998 		if (last)
5999 			last->next = sg;
6000 		last = sg;
6001 		last->next = first;
6002 	}
6003 	sd->groups = groups;
6004 
6005 	return 0;
6006 
6007 fail:
6008 	free_sched_groups(first, 0);
6009 
6010 	return -ENOMEM;
6011 }
6012 
get_group(int cpu,struct sd_data * sdd,struct sched_group ** sg)6013 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6014 {
6015 	struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6016 	struct sched_domain *child = sd->child;
6017 
6018 	if (child)
6019 		cpu = cpumask_first(sched_domain_span(child));
6020 
6021 	if (sg) {
6022 		*sg = *per_cpu_ptr(sdd->sg, cpu);
6023 		(*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6024 		atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6025 	}
6026 
6027 	return cpu;
6028 }
6029 
6030 /*
6031  * build_sched_groups will build a circular linked list of the groups
6032  * covered by the given span, and will set each group's ->cpumask correctly,
6033  * and ->cpu_power to 0.
6034  *
6035  * Assumes the sched_domain tree is fully constructed
6036  */
6037 static int
build_sched_groups(struct sched_domain * sd,int cpu)6038 build_sched_groups(struct sched_domain *sd, int cpu)
6039 {
6040 	struct sched_group *first = NULL, *last = NULL;
6041 	struct sd_data *sdd = sd->private;
6042 	const struct cpumask *span = sched_domain_span(sd);
6043 	struct cpumask *covered;
6044 	int i;
6045 
6046 	get_group(cpu, sdd, &sd->groups);
6047 	atomic_inc(&sd->groups->ref);
6048 
6049 	if (cpu != cpumask_first(sched_domain_span(sd)))
6050 		return 0;
6051 
6052 	lockdep_assert_held(&sched_domains_mutex);
6053 	covered = sched_domains_tmpmask;
6054 
6055 	cpumask_clear(covered);
6056 
6057 	for_each_cpu(i, span) {
6058 		struct sched_group *sg;
6059 		int group = get_group(i, sdd, &sg);
6060 		int j;
6061 
6062 		if (cpumask_test_cpu(i, covered))
6063 			continue;
6064 
6065 		cpumask_clear(sched_group_cpus(sg));
6066 		sg->sgp->power = 0;
6067 
6068 		for_each_cpu(j, span) {
6069 			if (get_group(j, sdd, NULL) != group)
6070 				continue;
6071 
6072 			cpumask_set_cpu(j, covered);
6073 			cpumask_set_cpu(j, sched_group_cpus(sg));
6074 		}
6075 
6076 		if (!first)
6077 			first = sg;
6078 		if (last)
6079 			last->next = sg;
6080 		last = sg;
6081 	}
6082 	last->next = first;
6083 
6084 	return 0;
6085 }
6086 
6087 /*
6088  * Initialize sched groups cpu_power.
6089  *
6090  * cpu_power indicates the capacity of sched group, which is used while
6091  * distributing the load between different sched groups in a sched domain.
6092  * Typically cpu_power for all the groups in a sched domain will be same unless
6093  * there are asymmetries in the topology. If there are asymmetries, group
6094  * having more cpu_power will pickup more load compared to the group having
6095  * less cpu_power.
6096  */
init_sched_groups_power(int cpu,struct sched_domain * sd)6097 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6098 {
6099 	struct sched_group *sg = sd->groups;
6100 
6101 	WARN_ON(!sd || !sg);
6102 
6103 	do {
6104 		sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6105 		sg = sg->next;
6106 	} while (sg != sd->groups);
6107 
6108 	if (cpu != group_first_cpu(sg))
6109 		return;
6110 
6111 	update_group_power(sd, cpu);
6112 	atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6113 }
6114 
arch_sd_sibling_asym_packing(void)6115 int __weak arch_sd_sibling_asym_packing(void)
6116 {
6117        return 0*SD_ASYM_PACKING;
6118 }
6119 
6120 /*
6121  * Initializers for schedule domains
6122  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6123  */
6124 
6125 #ifdef CONFIG_SCHED_DEBUG
6126 # define SD_INIT_NAME(sd, type)		sd->name = #type
6127 #else
6128 # define SD_INIT_NAME(sd, type)		do { } while (0)
6129 #endif
6130 
6131 #define SD_INIT_FUNC(type)						\
6132 static noinline struct sched_domain *					\
6133 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) 	\
6134 {									\
6135 	struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);	\
6136 	*sd = SD_##type##_INIT;						\
6137 	SD_INIT_NAME(sd, type);						\
6138 	sd->private = &tl->data;					\
6139 	return sd;							\
6140 }
6141 
6142 SD_INIT_FUNC(CPU)
6143 #ifdef CONFIG_NUMA
6144  SD_INIT_FUNC(ALLNODES)
6145  SD_INIT_FUNC(NODE)
6146 #endif
6147 #ifdef CONFIG_SCHED_SMT
6148  SD_INIT_FUNC(SIBLING)
6149 #endif
6150 #ifdef CONFIG_SCHED_MC
6151  SD_INIT_FUNC(MC)
6152 #endif
6153 #ifdef CONFIG_SCHED_BOOK
6154  SD_INIT_FUNC(BOOK)
6155 #endif
6156 
6157 static int default_relax_domain_level = -1;
6158 int sched_domain_level_max;
6159 
setup_relax_domain_level(char * str)6160 static int __init setup_relax_domain_level(char *str)
6161 {
6162 	unsigned long val;
6163 
6164 	val = simple_strtoul(str, NULL, 0);
6165 	if (val < sched_domain_level_max)
6166 		default_relax_domain_level = val;
6167 
6168 	return 1;
6169 }
6170 __setup("relax_domain_level=", setup_relax_domain_level);
6171 
set_domain_attribute(struct sched_domain * sd,struct sched_domain_attr * attr)6172 static void set_domain_attribute(struct sched_domain *sd,
6173 				 struct sched_domain_attr *attr)
6174 {
6175 	int request;
6176 
6177 	if (!attr || attr->relax_domain_level < 0) {
6178 		if (default_relax_domain_level < 0)
6179 			return;
6180 		else
6181 			request = default_relax_domain_level;
6182 	} else
6183 		request = attr->relax_domain_level;
6184 	if (request < sd->level) {
6185 		/* turn off idle balance on this domain */
6186 		sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6187 	} else {
6188 		/* turn on idle balance on this domain */
6189 		sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6190 	}
6191 }
6192 
6193 static void __sdt_free(const struct cpumask *cpu_map);
6194 static int __sdt_alloc(const struct cpumask *cpu_map);
6195 
__free_domain_allocs(struct s_data * d,enum s_alloc what,const struct cpumask * cpu_map)6196 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6197 				 const struct cpumask *cpu_map)
6198 {
6199 	switch (what) {
6200 	case sa_rootdomain:
6201 		if (!atomic_read(&d->rd->refcount))
6202 			free_rootdomain(&d->rd->rcu); /* fall through */
6203 	case sa_sd:
6204 		free_percpu(d->sd); /* fall through */
6205 	case sa_sd_storage:
6206 		__sdt_free(cpu_map); /* fall through */
6207 	case sa_none:
6208 		break;
6209 	}
6210 }
6211 
__visit_domain_allocation_hell(struct s_data * d,const struct cpumask * cpu_map)6212 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6213 						   const struct cpumask *cpu_map)
6214 {
6215 	memset(d, 0, sizeof(*d));
6216 
6217 	if (__sdt_alloc(cpu_map))
6218 		return sa_sd_storage;
6219 	d->sd = alloc_percpu(struct sched_domain *);
6220 	if (!d->sd)
6221 		return sa_sd_storage;
6222 	d->rd = alloc_rootdomain();
6223 	if (!d->rd)
6224 		return sa_sd;
6225 	return sa_rootdomain;
6226 }
6227 
6228 /*
6229  * NULL the sd_data elements we've used to build the sched_domain and
6230  * sched_group structure so that the subsequent __free_domain_allocs()
6231  * will not free the data we're using.
6232  */
claim_allocations(int cpu,struct sched_domain * sd)6233 static void claim_allocations(int cpu, struct sched_domain *sd)
6234 {
6235 	struct sd_data *sdd = sd->private;
6236 
6237 	WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6238 	*per_cpu_ptr(sdd->sd, cpu) = NULL;
6239 
6240 	if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6241 		*per_cpu_ptr(sdd->sg, cpu) = NULL;
6242 
6243 	if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6244 		*per_cpu_ptr(sdd->sgp, cpu) = NULL;
6245 }
6246 
6247 #ifdef CONFIG_SCHED_SMT
cpu_smt_mask(int cpu)6248 static const struct cpumask *cpu_smt_mask(int cpu)
6249 {
6250 	return topology_thread_cpumask(cpu);
6251 }
6252 #endif
6253 
6254 /*
6255  * Topology list, bottom-up.
6256  */
6257 static struct sched_domain_topology_level default_topology[] = {
6258 #ifdef CONFIG_SCHED_SMT
6259 	{ sd_init_SIBLING, cpu_smt_mask, },
6260 #endif
6261 #ifdef CONFIG_SCHED_MC
6262 	{ sd_init_MC, cpu_coregroup_mask, },
6263 #endif
6264 #ifdef CONFIG_SCHED_BOOK
6265 	{ sd_init_BOOK, cpu_book_mask, },
6266 #endif
6267 	{ sd_init_CPU, cpu_cpu_mask, },
6268 #ifdef CONFIG_NUMA
6269 	{ sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
6270 	{ sd_init_ALLNODES, cpu_allnodes_mask, },
6271 #endif
6272 	{ NULL, },
6273 };
6274 
6275 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6276 
__sdt_alloc(const struct cpumask * cpu_map)6277 static int __sdt_alloc(const struct cpumask *cpu_map)
6278 {
6279 	struct sched_domain_topology_level *tl;
6280 	int j;
6281 
6282 	for (tl = sched_domain_topology; tl->init; tl++) {
6283 		struct sd_data *sdd = &tl->data;
6284 
6285 		sdd->sd = alloc_percpu(struct sched_domain *);
6286 		if (!sdd->sd)
6287 			return -ENOMEM;
6288 
6289 		sdd->sg = alloc_percpu(struct sched_group *);
6290 		if (!sdd->sg)
6291 			return -ENOMEM;
6292 
6293 		sdd->sgp = alloc_percpu(struct sched_group_power *);
6294 		if (!sdd->sgp)
6295 			return -ENOMEM;
6296 
6297 		for_each_cpu(j, cpu_map) {
6298 			struct sched_domain *sd;
6299 			struct sched_group *sg;
6300 			struct sched_group_power *sgp;
6301 
6302 		       	sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6303 					GFP_KERNEL, cpu_to_node(j));
6304 			if (!sd)
6305 				return -ENOMEM;
6306 
6307 			*per_cpu_ptr(sdd->sd, j) = sd;
6308 
6309 			sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6310 					GFP_KERNEL, cpu_to_node(j));
6311 			if (!sg)
6312 				return -ENOMEM;
6313 
6314 			*per_cpu_ptr(sdd->sg, j) = sg;
6315 
6316 			sgp = kzalloc_node(sizeof(struct sched_group_power),
6317 					GFP_KERNEL, cpu_to_node(j));
6318 			if (!sgp)
6319 				return -ENOMEM;
6320 
6321 			*per_cpu_ptr(sdd->sgp, j) = sgp;
6322 		}
6323 	}
6324 
6325 	return 0;
6326 }
6327 
__sdt_free(const struct cpumask * cpu_map)6328 static void __sdt_free(const struct cpumask *cpu_map)
6329 {
6330 	struct sched_domain_topology_level *tl;
6331 	int j;
6332 
6333 	for (tl = sched_domain_topology; tl->init; tl++) {
6334 		struct sd_data *sdd = &tl->data;
6335 
6336 		for_each_cpu(j, cpu_map) {
6337 			struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
6338 			if (sd && (sd->flags & SD_OVERLAP))
6339 				free_sched_groups(sd->groups, 0);
6340 			kfree(*per_cpu_ptr(sdd->sd, j));
6341 			kfree(*per_cpu_ptr(sdd->sg, j));
6342 			kfree(*per_cpu_ptr(sdd->sgp, j));
6343 		}
6344 		free_percpu(sdd->sd);
6345 		free_percpu(sdd->sg);
6346 		free_percpu(sdd->sgp);
6347 	}
6348 }
6349 
build_sched_domain(struct sched_domain_topology_level * tl,struct s_data * d,const struct cpumask * cpu_map,struct sched_domain_attr * attr,struct sched_domain * child,int cpu)6350 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6351 		struct s_data *d, const struct cpumask *cpu_map,
6352 		struct sched_domain_attr *attr, struct sched_domain *child,
6353 		int cpu)
6354 {
6355 	struct sched_domain *sd = tl->init(tl, cpu);
6356 	if (!sd)
6357 		return child;
6358 
6359 	set_domain_attribute(sd, attr);
6360 	cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6361 	if (child) {
6362 		sd->level = child->level + 1;
6363 		sched_domain_level_max = max(sched_domain_level_max, sd->level);
6364 		child->parent = sd;
6365 	}
6366 	sd->child = child;
6367 
6368 	return sd;
6369 }
6370 
6371 /*
6372  * Build sched domains for a given set of cpus and attach the sched domains
6373  * to the individual cpus
6374  */
build_sched_domains(const struct cpumask * cpu_map,struct sched_domain_attr * attr)6375 static int build_sched_domains(const struct cpumask *cpu_map,
6376 			       struct sched_domain_attr *attr)
6377 {
6378 	enum s_alloc alloc_state = sa_none;
6379 	struct sched_domain *sd;
6380 	struct s_data d;
6381 	int i, ret = -ENOMEM;
6382 
6383 	alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6384 	if (alloc_state != sa_rootdomain)
6385 		goto error;
6386 
6387 	/* Set up domains for cpus specified by the cpu_map. */
6388 	for_each_cpu(i, cpu_map) {
6389 		struct sched_domain_topology_level *tl;
6390 
6391 		sd = NULL;
6392 		for (tl = sched_domain_topology; tl->init; tl++) {
6393 			sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6394 			if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6395 				sd->flags |= SD_OVERLAP;
6396 			if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6397 				break;
6398 		}
6399 
6400 		while (sd->child)
6401 			sd = sd->child;
6402 
6403 		*per_cpu_ptr(d.sd, i) = sd;
6404 	}
6405 
6406 	/* Build the groups for the domains */
6407 	for_each_cpu(i, cpu_map) {
6408 		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6409 			sd->span_weight = cpumask_weight(sched_domain_span(sd));
6410 			if (sd->flags & SD_OVERLAP) {
6411 				if (build_overlap_sched_groups(sd, i))
6412 					goto error;
6413 			} else {
6414 				if (build_sched_groups(sd, i))
6415 					goto error;
6416 			}
6417 		}
6418 	}
6419 
6420 	/* Calculate CPU power for physical packages and nodes */
6421 	for (i = nr_cpumask_bits-1; i >= 0; i--) {
6422 		if (!cpumask_test_cpu(i, cpu_map))
6423 			continue;
6424 
6425 		for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6426 			claim_allocations(i, sd);
6427 			init_sched_groups_power(i, sd);
6428 		}
6429 	}
6430 
6431 	/* Attach the domains */
6432 	rcu_read_lock();
6433 	for_each_cpu(i, cpu_map) {
6434 		sd = *per_cpu_ptr(d.sd, i);
6435 		cpu_attach_domain(sd, d.rd, i);
6436 	}
6437 	rcu_read_unlock();
6438 
6439 	ret = 0;
6440 error:
6441 	__free_domain_allocs(&d, alloc_state, cpu_map);
6442 	return ret;
6443 }
6444 
6445 static cpumask_var_t *doms_cur;	/* current sched domains */
6446 static int ndoms_cur;		/* number of sched domains in 'doms_cur' */
6447 static struct sched_domain_attr *dattr_cur;
6448 				/* attribues of custom domains in 'doms_cur' */
6449 
6450 /*
6451  * Special case: If a kmalloc of a doms_cur partition (array of
6452  * cpumask) fails, then fallback to a single sched domain,
6453  * as determined by the single cpumask fallback_doms.
6454  */
6455 static cpumask_var_t fallback_doms;
6456 
6457 /*
6458  * arch_update_cpu_topology lets virtualized architectures update the
6459  * cpu core maps. It is supposed to return 1 if the topology changed
6460  * or 0 if it stayed the same.
6461  */
arch_update_cpu_topology(void)6462 int __attribute__((weak)) arch_update_cpu_topology(void)
6463 {
6464 	return 0;
6465 }
6466 
alloc_sched_domains(unsigned int ndoms)6467 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6468 {
6469 	int i;
6470 	cpumask_var_t *doms;
6471 
6472 	doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6473 	if (!doms)
6474 		return NULL;
6475 	for (i = 0; i < ndoms; i++) {
6476 		if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6477 			free_sched_domains(doms, i);
6478 			return NULL;
6479 		}
6480 	}
6481 	return doms;
6482 }
6483 
free_sched_domains(cpumask_var_t doms[],unsigned int ndoms)6484 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6485 {
6486 	unsigned int i;
6487 	for (i = 0; i < ndoms; i++)
6488 		free_cpumask_var(doms[i]);
6489 	kfree(doms);
6490 }
6491 
6492 /*
6493  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6494  * For now this just excludes isolated cpus, but could be used to
6495  * exclude other special cases in the future.
6496  */
init_sched_domains(const struct cpumask * cpu_map)6497 static int init_sched_domains(const struct cpumask *cpu_map)
6498 {
6499 	int err;
6500 
6501 	arch_update_cpu_topology();
6502 	ndoms_cur = 1;
6503 	doms_cur = alloc_sched_domains(ndoms_cur);
6504 	if (!doms_cur)
6505 		doms_cur = &fallback_doms;
6506 	cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6507 	dattr_cur = NULL;
6508 	err = build_sched_domains(doms_cur[0], NULL);
6509 	register_sched_domain_sysctl();
6510 
6511 	return err;
6512 }
6513 
6514 /*
6515  * Detach sched domains from a group of cpus specified in cpu_map
6516  * These cpus will now be attached to the NULL domain
6517  */
detach_destroy_domains(const struct cpumask * cpu_map)6518 static void detach_destroy_domains(const struct cpumask *cpu_map)
6519 {
6520 	int i;
6521 
6522 	rcu_read_lock();
6523 	for_each_cpu(i, cpu_map)
6524 		cpu_attach_domain(NULL, &def_root_domain, i);
6525 	rcu_read_unlock();
6526 }
6527 
6528 /* handle null as "default" */
dattrs_equal(struct sched_domain_attr * cur,int idx_cur,struct sched_domain_attr * new,int idx_new)6529 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6530 			struct sched_domain_attr *new, int idx_new)
6531 {
6532 	struct sched_domain_attr tmp;
6533 
6534 	/* fast path */
6535 	if (!new && !cur)
6536 		return 1;
6537 
6538 	tmp = SD_ATTR_INIT;
6539 	return !memcmp(cur ? (cur + idx_cur) : &tmp,
6540 			new ? (new + idx_new) : &tmp,
6541 			sizeof(struct sched_domain_attr));
6542 }
6543 
6544 /*
6545  * Partition sched domains as specified by the 'ndoms_new'
6546  * cpumasks in the array doms_new[] of cpumasks. This compares
6547  * doms_new[] to the current sched domain partitioning, doms_cur[].
6548  * It destroys each deleted domain and builds each new domain.
6549  *
6550  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6551  * The masks don't intersect (don't overlap.) We should setup one
6552  * sched domain for each mask. CPUs not in any of the cpumasks will
6553  * not be load balanced. If the same cpumask appears both in the
6554  * current 'doms_cur' domains and in the new 'doms_new', we can leave
6555  * it as it is.
6556  *
6557  * The passed in 'doms_new' should be allocated using
6558  * alloc_sched_domains.  This routine takes ownership of it and will
6559  * free_sched_domains it when done with it. If the caller failed the
6560  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6561  * and partition_sched_domains() will fallback to the single partition
6562  * 'fallback_doms', it also forces the domains to be rebuilt.
6563  *
6564  * If doms_new == NULL it will be replaced with cpu_online_mask.
6565  * ndoms_new == 0 is a special case for destroying existing domains,
6566  * and it will not create the default domain.
6567  *
6568  * Call with hotplug lock held
6569  */
partition_sched_domains(int ndoms_new,cpumask_var_t doms_new[],struct sched_domain_attr * dattr_new)6570 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6571 			     struct sched_domain_attr *dattr_new)
6572 {
6573 	int i, j, n;
6574 	int new_topology;
6575 
6576 	mutex_lock(&sched_domains_mutex);
6577 
6578 	/* always unregister in case we don't destroy any domains */
6579 	unregister_sched_domain_sysctl();
6580 
6581 	/* Let architecture update cpu core mappings. */
6582 	new_topology = arch_update_cpu_topology();
6583 
6584 	n = doms_new ? ndoms_new : 0;
6585 
6586 	/* Destroy deleted domains */
6587 	for (i = 0; i < ndoms_cur; i++) {
6588 		for (j = 0; j < n && !new_topology; j++) {
6589 			if (cpumask_equal(doms_cur[i], doms_new[j])
6590 			    && dattrs_equal(dattr_cur, i, dattr_new, j))
6591 				goto match1;
6592 		}
6593 		/* no match - a current sched domain not in new doms_new[] */
6594 		detach_destroy_domains(doms_cur[i]);
6595 match1:
6596 		;
6597 	}
6598 
6599 	if (doms_new == NULL) {
6600 		ndoms_cur = 0;
6601 		doms_new = &fallback_doms;
6602 		cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6603 		WARN_ON_ONCE(dattr_new);
6604 	}
6605 
6606 	/* Build new domains */
6607 	for (i = 0; i < ndoms_new; i++) {
6608 		for (j = 0; j < ndoms_cur && !new_topology; j++) {
6609 			if (cpumask_equal(doms_new[i], doms_cur[j])
6610 			    && dattrs_equal(dattr_new, i, dattr_cur, j))
6611 				goto match2;
6612 		}
6613 		/* no match - add a new doms_new */
6614 		build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6615 match2:
6616 		;
6617 	}
6618 
6619 	/* Remember the new sched domains */
6620 	if (doms_cur != &fallback_doms)
6621 		free_sched_domains(doms_cur, ndoms_cur);
6622 	kfree(dattr_cur);	/* kfree(NULL) is safe */
6623 	doms_cur = doms_new;
6624 	dattr_cur = dattr_new;
6625 	ndoms_cur = ndoms_new;
6626 
6627 	register_sched_domain_sysctl();
6628 
6629 	mutex_unlock(&sched_domains_mutex);
6630 }
6631 
6632 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
reinit_sched_domains(void)6633 static void reinit_sched_domains(void)
6634 {
6635 	get_online_cpus();
6636 
6637 	/* Destroy domains first to force the rebuild */
6638 	partition_sched_domains(0, NULL, NULL);
6639 
6640 	rebuild_sched_domains();
6641 	put_online_cpus();
6642 }
6643 
sched_power_savings_store(const char * buf,size_t count,int smt)6644 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6645 {
6646 	unsigned int level = 0;
6647 
6648 	if (sscanf(buf, "%u", &level) != 1)
6649 		return -EINVAL;
6650 
6651 	/*
6652 	 * level is always be positive so don't check for
6653 	 * level < POWERSAVINGS_BALANCE_NONE which is 0
6654 	 * What happens on 0 or 1 byte write,
6655 	 * need to check for count as well?
6656 	 */
6657 
6658 	if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
6659 		return -EINVAL;
6660 
6661 	if (smt)
6662 		sched_smt_power_savings = level;
6663 	else
6664 		sched_mc_power_savings = level;
6665 
6666 	reinit_sched_domains();
6667 
6668 	return count;
6669 }
6670 
6671 #ifdef CONFIG_SCHED_MC
sched_mc_power_savings_show(struct device * dev,struct device_attribute * attr,char * buf)6672 static ssize_t sched_mc_power_savings_show(struct device *dev,
6673 					   struct device_attribute *attr,
6674 					   char *buf)
6675 {
6676 	return sprintf(buf, "%u\n", sched_mc_power_savings);
6677 }
sched_mc_power_savings_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)6678 static ssize_t sched_mc_power_savings_store(struct device *dev,
6679 					    struct device_attribute *attr,
6680 					    const char *buf, size_t count)
6681 {
6682 	return sched_power_savings_store(buf, count, 0);
6683 }
6684 static DEVICE_ATTR(sched_mc_power_savings, 0644,
6685 		   sched_mc_power_savings_show,
6686 		   sched_mc_power_savings_store);
6687 #endif
6688 
6689 #ifdef CONFIG_SCHED_SMT
sched_smt_power_savings_show(struct device * dev,struct device_attribute * attr,char * buf)6690 static ssize_t sched_smt_power_savings_show(struct device *dev,
6691 					    struct device_attribute *attr,
6692 					    char *buf)
6693 {
6694 	return sprintf(buf, "%u\n", sched_smt_power_savings);
6695 }
sched_smt_power_savings_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)6696 static ssize_t sched_smt_power_savings_store(struct device *dev,
6697 					    struct device_attribute *attr,
6698 					     const char *buf, size_t count)
6699 {
6700 	return sched_power_savings_store(buf, count, 1);
6701 }
6702 static DEVICE_ATTR(sched_smt_power_savings, 0644,
6703 		   sched_smt_power_savings_show,
6704 		   sched_smt_power_savings_store);
6705 #endif
6706 
sched_create_sysfs_power_savings_entries(struct device * dev)6707 int __init sched_create_sysfs_power_savings_entries(struct device *dev)
6708 {
6709 	int err = 0;
6710 
6711 #ifdef CONFIG_SCHED_SMT
6712 	if (smt_capable())
6713 		err = device_create_file(dev, &dev_attr_sched_smt_power_savings);
6714 #endif
6715 #ifdef CONFIG_SCHED_MC
6716 	if (!err && mc_capable())
6717 		err = device_create_file(dev, &dev_attr_sched_mc_power_savings);
6718 #endif
6719 	return err;
6720 }
6721 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6722 
6723 /*
6724  * Update cpusets according to cpu_active mask.  If cpusets are
6725  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6726  * around partition_sched_domains().
6727  */
cpuset_cpu_active(struct notifier_block * nfb,unsigned long action,void * hcpu)6728 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6729 			     void *hcpu)
6730 {
6731 	switch (action & ~CPU_TASKS_FROZEN) {
6732 	case CPU_ONLINE:
6733 	case CPU_DOWN_FAILED:
6734 		cpuset_update_active_cpus();
6735 		return NOTIFY_OK;
6736 	default:
6737 		return NOTIFY_DONE;
6738 	}
6739 }
6740 
cpuset_cpu_inactive(struct notifier_block * nfb,unsigned long action,void * hcpu)6741 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6742 			       void *hcpu)
6743 {
6744 	switch (action & ~CPU_TASKS_FROZEN) {
6745 	case CPU_DOWN_PREPARE:
6746 		cpuset_update_active_cpus();
6747 		return NOTIFY_OK;
6748 	default:
6749 		return NOTIFY_DONE;
6750 	}
6751 }
6752 
sched_init_smp(void)6753 void __init sched_init_smp(void)
6754 {
6755 	cpumask_var_t non_isolated_cpus;
6756 
6757 	alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6758 	alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6759 
6760 	get_online_cpus();
6761 	mutex_lock(&sched_domains_mutex);
6762 	init_sched_domains(cpu_active_mask);
6763 	cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6764 	if (cpumask_empty(non_isolated_cpus))
6765 		cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6766 	mutex_unlock(&sched_domains_mutex);
6767 	put_online_cpus();
6768 
6769 	hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6770 	hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6771 
6772 	/* RT runtime code needs to handle some hotplug events */
6773 	hotcpu_notifier(update_runtime, 0);
6774 
6775 	init_hrtick();
6776 
6777 	/* Move init over to a non-isolated CPU */
6778 	if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6779 		BUG();
6780 	sched_init_granularity();
6781 	free_cpumask_var(non_isolated_cpus);
6782 
6783 	init_sched_rt_class();
6784 }
6785 #else
sched_init_smp(void)6786 void __init sched_init_smp(void)
6787 {
6788 	sched_init_granularity();
6789 }
6790 #endif /* CONFIG_SMP */
6791 
6792 const_debug unsigned int sysctl_timer_migration = 1;
6793 
in_sched_functions(unsigned long addr)6794 int in_sched_functions(unsigned long addr)
6795 {
6796 	return in_lock_functions(addr) ||
6797 		(addr >= (unsigned long)__sched_text_start
6798 		&& addr < (unsigned long)__sched_text_end);
6799 }
6800 
6801 #ifdef CONFIG_CGROUP_SCHED
6802 struct task_group root_task_group;
6803 #endif
6804 
6805 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6806 
sched_init(void)6807 void __init sched_init(void)
6808 {
6809 	int i, j;
6810 	unsigned long alloc_size = 0, ptr;
6811 
6812 #ifdef CONFIG_FAIR_GROUP_SCHED
6813 	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6814 #endif
6815 #ifdef CONFIG_RT_GROUP_SCHED
6816 	alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6817 #endif
6818 #ifdef CONFIG_CPUMASK_OFFSTACK
6819 	alloc_size += num_possible_cpus() * cpumask_size();
6820 #endif
6821 	if (alloc_size) {
6822 		ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6823 
6824 #ifdef CONFIG_FAIR_GROUP_SCHED
6825 		root_task_group.se = (struct sched_entity **)ptr;
6826 		ptr += nr_cpu_ids * sizeof(void **);
6827 
6828 		root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6829 		ptr += nr_cpu_ids * sizeof(void **);
6830 
6831 #endif /* CONFIG_FAIR_GROUP_SCHED */
6832 #ifdef CONFIG_RT_GROUP_SCHED
6833 		root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6834 		ptr += nr_cpu_ids * sizeof(void **);
6835 
6836 		root_task_group.rt_rq = (struct rt_rq **)ptr;
6837 		ptr += nr_cpu_ids * sizeof(void **);
6838 
6839 #endif /* CONFIG_RT_GROUP_SCHED */
6840 #ifdef CONFIG_CPUMASK_OFFSTACK
6841 		for_each_possible_cpu(i) {
6842 			per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6843 			ptr += cpumask_size();
6844 		}
6845 #endif /* CONFIG_CPUMASK_OFFSTACK */
6846 	}
6847 
6848 #ifdef CONFIG_SMP
6849 	init_defrootdomain();
6850 #endif
6851 
6852 	init_rt_bandwidth(&def_rt_bandwidth,
6853 			global_rt_period(), global_rt_runtime());
6854 
6855 #ifdef CONFIG_RT_GROUP_SCHED
6856 	init_rt_bandwidth(&root_task_group.rt_bandwidth,
6857 			global_rt_period(), global_rt_runtime());
6858 #endif /* CONFIG_RT_GROUP_SCHED */
6859 
6860 #ifdef CONFIG_CGROUP_SCHED
6861 	list_add(&root_task_group.list, &task_groups);
6862 	INIT_LIST_HEAD(&root_task_group.children);
6863 	INIT_LIST_HEAD(&root_task_group.siblings);
6864 	autogroup_init(&init_task);
6865 
6866 #endif /* CONFIG_CGROUP_SCHED */
6867 
6868 #ifdef CONFIG_CGROUP_CPUACCT
6869 	root_cpuacct.cpustat = &kernel_cpustat;
6870 	root_cpuacct.cpuusage = alloc_percpu(u64);
6871 	/* Too early, not expected to fail */
6872 	BUG_ON(!root_cpuacct.cpuusage);
6873 #endif
6874 	for_each_possible_cpu(i) {
6875 		struct rq *rq;
6876 
6877 		rq = cpu_rq(i);
6878 		raw_spin_lock_init(&rq->lock);
6879 		rq->nr_running = 0;
6880 		rq->calc_load_active = 0;
6881 		rq->calc_load_update = jiffies + LOAD_FREQ;
6882 		init_cfs_rq(&rq->cfs);
6883 		init_rt_rq(&rq->rt, rq);
6884 #ifdef CONFIG_FAIR_GROUP_SCHED
6885 		root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6886 		INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6887 		/*
6888 		 * How much cpu bandwidth does root_task_group get?
6889 		 *
6890 		 * In case of task-groups formed thr' the cgroup filesystem, it
6891 		 * gets 100% of the cpu resources in the system. This overall
6892 		 * system cpu resource is divided among the tasks of
6893 		 * root_task_group and its child task-groups in a fair manner,
6894 		 * based on each entity's (task or task-group's) weight
6895 		 * (se->load.weight).
6896 		 *
6897 		 * In other words, if root_task_group has 10 tasks of weight
6898 		 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6899 		 * then A0's share of the cpu resource is:
6900 		 *
6901 		 *	A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6902 		 *
6903 		 * We achieve this by letting root_task_group's tasks sit
6904 		 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6905 		 */
6906 		init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6907 		init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6908 #endif /* CONFIG_FAIR_GROUP_SCHED */
6909 
6910 		rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6911 #ifdef CONFIG_RT_GROUP_SCHED
6912 		INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6913 		init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6914 #endif
6915 
6916 		for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6917 			rq->cpu_load[j] = 0;
6918 
6919 		rq->last_load_update_tick = jiffies;
6920 
6921 #ifdef CONFIG_SMP
6922 		rq->sd = NULL;
6923 		rq->rd = NULL;
6924 		rq->cpu_power = SCHED_POWER_SCALE;
6925 		rq->post_schedule = 0;
6926 		rq->active_balance = 0;
6927 		rq->next_balance = jiffies;
6928 		rq->push_cpu = 0;
6929 		rq->cpu = i;
6930 		rq->online = 0;
6931 		rq->idle_stamp = 0;
6932 		rq->avg_idle = 2*sysctl_sched_migration_cost;
6933 		rq_attach_root(rq, &def_root_domain);
6934 #ifdef CONFIG_NO_HZ
6935 		rq->nohz_flags = 0;
6936 #endif
6937 #endif
6938 		init_rq_hrtick(rq);
6939 		atomic_set(&rq->nr_iowait, 0);
6940 	}
6941 
6942 	set_load_weight(&init_task);
6943 
6944 #ifdef CONFIG_PREEMPT_NOTIFIERS
6945 	INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6946 #endif
6947 
6948 #ifdef CONFIG_RT_MUTEXES
6949 	plist_head_init(&init_task.pi_waiters);
6950 #endif
6951 
6952 	/*
6953 	 * The boot idle thread does lazy MMU switching as well:
6954 	 */
6955 	atomic_inc(&init_mm.mm_count);
6956 	enter_lazy_tlb(&init_mm, current);
6957 
6958 	/*
6959 	 * Make us the idle thread. Technically, schedule() should not be
6960 	 * called from this thread, however somewhere below it might be,
6961 	 * but because we are the idle thread, we just pick up running again
6962 	 * when this runqueue becomes "idle".
6963 	 */
6964 	init_idle(current, smp_processor_id());
6965 
6966 	calc_load_update = jiffies + LOAD_FREQ;
6967 
6968 	/*
6969 	 * During early bootup we pretend to be a normal task:
6970 	 */
6971 	current->sched_class = &fair_sched_class;
6972 
6973 #ifdef CONFIG_SMP
6974 	zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6975 	/* May be allocated at isolcpus cmdline parse time */
6976 	if (cpu_isolated_map == NULL)
6977 		zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6978 #endif
6979 	init_sched_fair_class();
6980 
6981 	scheduler_running = 1;
6982 }
6983 
6984 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
preempt_count_equals(int preempt_offset)6985 static inline int preempt_count_equals(int preempt_offset)
6986 {
6987 	int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6988 
6989 	return (nested == preempt_offset);
6990 }
6991 
__might_sleep(const char * file,int line,int preempt_offset)6992 void __might_sleep(const char *file, int line, int preempt_offset)
6993 {
6994 	static unsigned long prev_jiffy;	/* ratelimiting */
6995 
6996 	rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6997 	if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6998 	    system_state != SYSTEM_RUNNING || oops_in_progress)
6999 		return;
7000 	if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7001 		return;
7002 	prev_jiffy = jiffies;
7003 
7004 	printk(KERN_ERR
7005 		"BUG: sleeping function called from invalid context at %s:%d\n",
7006 			file, line);
7007 	printk(KERN_ERR
7008 		"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7009 			in_atomic(), irqs_disabled(),
7010 			current->pid, current->comm);
7011 
7012 	debug_show_held_locks(current);
7013 	if (irqs_disabled())
7014 		print_irqtrace_events(current);
7015 	dump_stack();
7016 }
7017 EXPORT_SYMBOL(__might_sleep);
7018 #endif
7019 
7020 #ifdef CONFIG_MAGIC_SYSRQ
normalize_task(struct rq * rq,struct task_struct * p)7021 static void normalize_task(struct rq *rq, struct task_struct *p)
7022 {
7023 	const struct sched_class *prev_class = p->sched_class;
7024 	int old_prio = p->prio;
7025 	int on_rq;
7026 
7027 	on_rq = p->on_rq;
7028 	if (on_rq)
7029 		dequeue_task(rq, p, 0);
7030 	__setscheduler(rq, p, SCHED_NORMAL, 0);
7031 	if (on_rq) {
7032 		enqueue_task(rq, p, 0);
7033 		resched_task(rq->curr);
7034 	}
7035 
7036 	check_class_changed(rq, p, prev_class, old_prio);
7037 }
7038 
normalize_rt_tasks(void)7039 void normalize_rt_tasks(void)
7040 {
7041 	struct task_struct *g, *p;
7042 	unsigned long flags;
7043 	struct rq *rq;
7044 
7045 	read_lock_irqsave(&tasklist_lock, flags);
7046 	do_each_thread(g, p) {
7047 		/*
7048 		 * Only normalize user tasks:
7049 		 */
7050 		if (!p->mm)
7051 			continue;
7052 
7053 		p->se.exec_start		= 0;
7054 #ifdef CONFIG_SCHEDSTATS
7055 		p->se.statistics.wait_start	= 0;
7056 		p->se.statistics.sleep_start	= 0;
7057 		p->se.statistics.block_start	= 0;
7058 #endif
7059 
7060 		if (!rt_task(p)) {
7061 			/*
7062 			 * Renice negative nice level userspace
7063 			 * tasks back to 0:
7064 			 */
7065 			if (TASK_NICE(p) < 0 && p->mm)
7066 				set_user_nice(p, 0);
7067 			continue;
7068 		}
7069 
7070 		raw_spin_lock(&p->pi_lock);
7071 		rq = __task_rq_lock(p);
7072 
7073 		normalize_task(rq, p);
7074 
7075 		__task_rq_unlock(rq);
7076 		raw_spin_unlock(&p->pi_lock);
7077 	} while_each_thread(g, p);
7078 
7079 	read_unlock_irqrestore(&tasklist_lock, flags);
7080 }
7081 
7082 #endif /* CONFIG_MAGIC_SYSRQ */
7083 
7084 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7085 /*
7086  * These functions are only useful for the IA64 MCA handling, or kdb.
7087  *
7088  * They can only be called when the whole system has been
7089  * stopped - every CPU needs to be quiescent, and no scheduling
7090  * activity can take place. Using them for anything else would
7091  * be a serious bug, and as a result, they aren't even visible
7092  * under any other configuration.
7093  */
7094 
7095 /**
7096  * curr_task - return the current task for a given cpu.
7097  * @cpu: the processor in question.
7098  *
7099  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7100  */
curr_task(int cpu)7101 struct task_struct *curr_task(int cpu)
7102 {
7103 	return cpu_curr(cpu);
7104 }
7105 
7106 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7107 
7108 #ifdef CONFIG_IA64
7109 /**
7110  * set_curr_task - set the current task for a given cpu.
7111  * @cpu: the processor in question.
7112  * @p: the task pointer to set.
7113  *
7114  * Description: This function must only be used when non-maskable interrupts
7115  * are serviced on a separate stack. It allows the architecture to switch the
7116  * notion of the current task on a cpu in a non-blocking manner. This function
7117  * must be called with all CPU's synchronized, and interrupts disabled, the
7118  * and caller must save the original value of the current task (see
7119  * curr_task() above) and restore that value before reenabling interrupts and
7120  * re-starting the system.
7121  *
7122  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7123  */
set_curr_task(int cpu,struct task_struct * p)7124 void set_curr_task(int cpu, struct task_struct *p)
7125 {
7126 	cpu_curr(cpu) = p;
7127 }
7128 
7129 #endif
7130 
7131 #ifdef CONFIG_CGROUP_SCHED
7132 /* task_group_lock serializes the addition/removal of task groups */
7133 static DEFINE_SPINLOCK(task_group_lock);
7134 
free_sched_group(struct task_group * tg)7135 static void free_sched_group(struct task_group *tg)
7136 {
7137 	free_fair_sched_group(tg);
7138 	free_rt_sched_group(tg);
7139 	autogroup_free(tg);
7140 	kfree(tg);
7141 }
7142 
7143 /* allocate runqueue etc for a new task group */
sched_create_group(struct task_group * parent)7144 struct task_group *sched_create_group(struct task_group *parent)
7145 {
7146 	struct task_group *tg;
7147 	unsigned long flags;
7148 
7149 	tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7150 	if (!tg)
7151 		return ERR_PTR(-ENOMEM);
7152 
7153 	if (!alloc_fair_sched_group(tg, parent))
7154 		goto err;
7155 
7156 	if (!alloc_rt_sched_group(tg, parent))
7157 		goto err;
7158 
7159 	spin_lock_irqsave(&task_group_lock, flags);
7160 	list_add_rcu(&tg->list, &task_groups);
7161 
7162 	WARN_ON(!parent); /* root should already exist */
7163 
7164 	tg->parent = parent;
7165 	INIT_LIST_HEAD(&tg->children);
7166 	list_add_rcu(&tg->siblings, &parent->children);
7167 	spin_unlock_irqrestore(&task_group_lock, flags);
7168 
7169 	return tg;
7170 
7171 err:
7172 	free_sched_group(tg);
7173 	return ERR_PTR(-ENOMEM);
7174 }
7175 
7176 /* rcu callback to free various structures associated with a task group */
free_sched_group_rcu(struct rcu_head * rhp)7177 static void free_sched_group_rcu(struct rcu_head *rhp)
7178 {
7179 	/* now it should be safe to free those cfs_rqs */
7180 	free_sched_group(container_of(rhp, struct task_group, rcu));
7181 }
7182 
7183 /* Destroy runqueue etc associated with a task group */
sched_destroy_group(struct task_group * tg)7184 void sched_destroy_group(struct task_group *tg)
7185 {
7186 	unsigned long flags;
7187 	int i;
7188 
7189 	/* end participation in shares distribution */
7190 	for_each_possible_cpu(i)
7191 		unregister_fair_sched_group(tg, i);
7192 
7193 	spin_lock_irqsave(&task_group_lock, flags);
7194 	list_del_rcu(&tg->list);
7195 	list_del_rcu(&tg->siblings);
7196 	spin_unlock_irqrestore(&task_group_lock, flags);
7197 
7198 	/* wait for possible concurrent references to cfs_rqs complete */
7199 	call_rcu(&tg->rcu, free_sched_group_rcu);
7200 }
7201 
7202 /* change task's runqueue when it moves between groups.
7203  *	The caller of this function should have put the task in its new group
7204  *	by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7205  *	reflect its new group.
7206  */
sched_move_task(struct task_struct * tsk)7207 void sched_move_task(struct task_struct *tsk)
7208 {
7209 	int on_rq, running;
7210 	unsigned long flags;
7211 	struct rq *rq;
7212 
7213 	rq = task_rq_lock(tsk, &flags);
7214 
7215 	running = task_current(rq, tsk);
7216 	on_rq = tsk->on_rq;
7217 
7218 	if (on_rq)
7219 		dequeue_task(rq, tsk, 0);
7220 	if (unlikely(running))
7221 		tsk->sched_class->put_prev_task(rq, tsk);
7222 
7223 #ifdef CONFIG_FAIR_GROUP_SCHED
7224 	if (tsk->sched_class->task_move_group)
7225 		tsk->sched_class->task_move_group(tsk, on_rq);
7226 	else
7227 #endif
7228 		set_task_rq(tsk, task_cpu(tsk));
7229 
7230 	if (unlikely(running))
7231 		tsk->sched_class->set_curr_task(rq);
7232 	if (on_rq)
7233 		enqueue_task(rq, tsk, 0);
7234 
7235 	task_rq_unlock(rq, tsk, &flags);
7236 }
7237 #endif /* CONFIG_CGROUP_SCHED */
7238 
7239 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
to_ratio(u64 period,u64 runtime)7240 static unsigned long to_ratio(u64 period, u64 runtime)
7241 {
7242 	if (runtime == RUNTIME_INF)
7243 		return 1ULL << 20;
7244 
7245 	return div64_u64(runtime << 20, period);
7246 }
7247 #endif
7248 
7249 #ifdef CONFIG_RT_GROUP_SCHED
7250 /*
7251  * Ensure that the real time constraints are schedulable.
7252  */
7253 static DEFINE_MUTEX(rt_constraints_mutex);
7254 
7255 /* Must be called with tasklist_lock held */
tg_has_rt_tasks(struct task_group * tg)7256 static inline int tg_has_rt_tasks(struct task_group *tg)
7257 {
7258 	struct task_struct *g, *p;
7259 
7260 	do_each_thread(g, p) {
7261 		if (rt_task(p) && task_rq(p)->rt.tg == tg)
7262 			return 1;
7263 	} while_each_thread(g, p);
7264 
7265 	return 0;
7266 }
7267 
7268 struct rt_schedulable_data {
7269 	struct task_group *tg;
7270 	u64 rt_period;
7271 	u64 rt_runtime;
7272 };
7273 
tg_rt_schedulable(struct task_group * tg,void * data)7274 static int tg_rt_schedulable(struct task_group *tg, void *data)
7275 {
7276 	struct rt_schedulable_data *d = data;
7277 	struct task_group *child;
7278 	unsigned long total, sum = 0;
7279 	u64 period, runtime;
7280 
7281 	period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7282 	runtime = tg->rt_bandwidth.rt_runtime;
7283 
7284 	if (tg == d->tg) {
7285 		period = d->rt_period;
7286 		runtime = d->rt_runtime;
7287 	}
7288 
7289 	/*
7290 	 * Cannot have more runtime than the period.
7291 	 */
7292 	if (runtime > period && runtime != RUNTIME_INF)
7293 		return -EINVAL;
7294 
7295 	/*
7296 	 * Ensure we don't starve existing RT tasks.
7297 	 */
7298 	if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7299 		return -EBUSY;
7300 
7301 	total = to_ratio(period, runtime);
7302 
7303 	/*
7304 	 * Nobody can have more than the global setting allows.
7305 	 */
7306 	if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7307 		return -EINVAL;
7308 
7309 	/*
7310 	 * The sum of our children's runtime should not exceed our own.
7311 	 */
7312 	list_for_each_entry_rcu(child, &tg->children, siblings) {
7313 		period = ktime_to_ns(child->rt_bandwidth.rt_period);
7314 		runtime = child->rt_bandwidth.rt_runtime;
7315 
7316 		if (child == d->tg) {
7317 			period = d->rt_period;
7318 			runtime = d->rt_runtime;
7319 		}
7320 
7321 		sum += to_ratio(period, runtime);
7322 	}
7323 
7324 	if (sum > total)
7325 		return -EINVAL;
7326 
7327 	return 0;
7328 }
7329 
__rt_schedulable(struct task_group * tg,u64 period,u64 runtime)7330 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7331 {
7332 	int ret;
7333 
7334 	struct rt_schedulable_data data = {
7335 		.tg = tg,
7336 		.rt_period = period,
7337 		.rt_runtime = runtime,
7338 	};
7339 
7340 	rcu_read_lock();
7341 	ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7342 	rcu_read_unlock();
7343 
7344 	return ret;
7345 }
7346 
tg_set_rt_bandwidth(struct task_group * tg,u64 rt_period,u64 rt_runtime)7347 static int tg_set_rt_bandwidth(struct task_group *tg,
7348 		u64 rt_period, u64 rt_runtime)
7349 {
7350 	int i, err = 0;
7351 
7352 	mutex_lock(&rt_constraints_mutex);
7353 	read_lock(&tasklist_lock);
7354 	err = __rt_schedulable(tg, rt_period, rt_runtime);
7355 	if (err)
7356 		goto unlock;
7357 
7358 	raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7359 	tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7360 	tg->rt_bandwidth.rt_runtime = rt_runtime;
7361 
7362 	for_each_possible_cpu(i) {
7363 		struct rt_rq *rt_rq = tg->rt_rq[i];
7364 
7365 		raw_spin_lock(&rt_rq->rt_runtime_lock);
7366 		rt_rq->rt_runtime = rt_runtime;
7367 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
7368 	}
7369 	raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7370 unlock:
7371 	read_unlock(&tasklist_lock);
7372 	mutex_unlock(&rt_constraints_mutex);
7373 
7374 	return err;
7375 }
7376 
sched_group_set_rt_runtime(struct task_group * tg,long rt_runtime_us)7377 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7378 {
7379 	u64 rt_runtime, rt_period;
7380 
7381 	rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7382 	rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7383 	if (rt_runtime_us < 0)
7384 		rt_runtime = RUNTIME_INF;
7385 
7386 	return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7387 }
7388 
sched_group_rt_runtime(struct task_group * tg)7389 long sched_group_rt_runtime(struct task_group *tg)
7390 {
7391 	u64 rt_runtime_us;
7392 
7393 	if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7394 		return -1;
7395 
7396 	rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7397 	do_div(rt_runtime_us, NSEC_PER_USEC);
7398 	return rt_runtime_us;
7399 }
7400 
sched_group_set_rt_period(struct task_group * tg,long rt_period_us)7401 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7402 {
7403 	u64 rt_runtime, rt_period;
7404 
7405 	rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7406 	rt_runtime = tg->rt_bandwidth.rt_runtime;
7407 
7408 	if (rt_period == 0)
7409 		return -EINVAL;
7410 
7411 	return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7412 }
7413 
sched_group_rt_period(struct task_group * tg)7414 long sched_group_rt_period(struct task_group *tg)
7415 {
7416 	u64 rt_period_us;
7417 
7418 	rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7419 	do_div(rt_period_us, NSEC_PER_USEC);
7420 	return rt_period_us;
7421 }
7422 
sched_rt_global_constraints(void)7423 static int sched_rt_global_constraints(void)
7424 {
7425 	u64 runtime, period;
7426 	int ret = 0;
7427 
7428 	if (sysctl_sched_rt_period <= 0)
7429 		return -EINVAL;
7430 
7431 	runtime = global_rt_runtime();
7432 	period = global_rt_period();
7433 
7434 	/*
7435 	 * Sanity check on the sysctl variables.
7436 	 */
7437 	if (runtime > period && runtime != RUNTIME_INF)
7438 		return -EINVAL;
7439 
7440 	mutex_lock(&rt_constraints_mutex);
7441 	read_lock(&tasklist_lock);
7442 	ret = __rt_schedulable(NULL, 0, 0);
7443 	read_unlock(&tasklist_lock);
7444 	mutex_unlock(&rt_constraints_mutex);
7445 
7446 	return ret;
7447 }
7448 
sched_rt_can_attach(struct task_group * tg,struct task_struct * tsk)7449 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7450 {
7451 	/* Don't accept realtime tasks when there is no way for them to run */
7452 	if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7453 		return 0;
7454 
7455 	return 1;
7456 }
7457 
7458 #else /* !CONFIG_RT_GROUP_SCHED */
sched_rt_global_constraints(void)7459 static int sched_rt_global_constraints(void)
7460 {
7461 	unsigned long flags;
7462 	int i;
7463 
7464 	if (sysctl_sched_rt_period <= 0)
7465 		return -EINVAL;
7466 
7467 	/*
7468 	 * There's always some RT tasks in the root group
7469 	 * -- migration, kstopmachine etc..
7470 	 */
7471 	if (sysctl_sched_rt_runtime == 0)
7472 		return -EBUSY;
7473 
7474 	raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7475 	for_each_possible_cpu(i) {
7476 		struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7477 
7478 		raw_spin_lock(&rt_rq->rt_runtime_lock);
7479 		rt_rq->rt_runtime = global_rt_runtime();
7480 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
7481 	}
7482 	raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7483 
7484 	return 0;
7485 }
7486 #endif /* CONFIG_RT_GROUP_SCHED */
7487 
sched_rt_handler(struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)7488 int sched_rt_handler(struct ctl_table *table, int write,
7489 		void __user *buffer, size_t *lenp,
7490 		loff_t *ppos)
7491 {
7492 	int ret;
7493 	int old_period, old_runtime;
7494 	static DEFINE_MUTEX(mutex);
7495 
7496 	mutex_lock(&mutex);
7497 	old_period = sysctl_sched_rt_period;
7498 	old_runtime = sysctl_sched_rt_runtime;
7499 
7500 	ret = proc_dointvec(table, write, buffer, lenp, ppos);
7501 
7502 	if (!ret && write) {
7503 		ret = sched_rt_global_constraints();
7504 		if (ret) {
7505 			sysctl_sched_rt_period = old_period;
7506 			sysctl_sched_rt_runtime = old_runtime;
7507 		} else {
7508 			def_rt_bandwidth.rt_runtime = global_rt_runtime();
7509 			def_rt_bandwidth.rt_period =
7510 				ns_to_ktime(global_rt_period());
7511 		}
7512 	}
7513 	mutex_unlock(&mutex);
7514 
7515 	return ret;
7516 }
7517 
7518 #ifdef CONFIG_CGROUP_SCHED
7519 
7520 /* return corresponding task_group object of a cgroup */
cgroup_tg(struct cgroup * cgrp)7521 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7522 {
7523 	return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7524 			    struct task_group, css);
7525 }
7526 
7527 static struct cgroup_subsys_state *
cpu_cgroup_create(struct cgroup_subsys * ss,struct cgroup * cgrp)7528 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7529 {
7530 	struct task_group *tg, *parent;
7531 
7532 	if (!cgrp->parent) {
7533 		/* This is early initialization for the top cgroup */
7534 		return &root_task_group.css;
7535 	}
7536 
7537 	parent = cgroup_tg(cgrp->parent);
7538 	tg = sched_create_group(parent);
7539 	if (IS_ERR(tg))
7540 		return ERR_PTR(-ENOMEM);
7541 
7542 	return &tg->css;
7543 }
7544 
7545 static void
cpu_cgroup_destroy(struct cgroup_subsys * ss,struct cgroup * cgrp)7546 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7547 {
7548 	struct task_group *tg = cgroup_tg(cgrp);
7549 
7550 	sched_destroy_group(tg);
7551 }
7552 
cpu_cgroup_can_attach(struct cgroup_subsys * ss,struct cgroup * cgrp,struct cgroup_taskset * tset)7553 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7554 				 struct cgroup_taskset *tset)
7555 {
7556 	struct task_struct *task;
7557 
7558 	cgroup_taskset_for_each(task, cgrp, tset) {
7559 #ifdef CONFIG_RT_GROUP_SCHED
7560 		if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7561 			return -EINVAL;
7562 #else
7563 		/* We don't support RT-tasks being in separate groups */
7564 		if (task->sched_class != &fair_sched_class)
7565 			return -EINVAL;
7566 #endif
7567 	}
7568 	return 0;
7569 }
7570 
cpu_cgroup_attach(struct cgroup_subsys * ss,struct cgroup * cgrp,struct cgroup_taskset * tset)7571 static void cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7572 			      struct cgroup_taskset *tset)
7573 {
7574 	struct task_struct *task;
7575 
7576 	cgroup_taskset_for_each(task, cgrp, tset)
7577 		sched_move_task(task);
7578 }
7579 
7580 static void
cpu_cgroup_exit(struct cgroup_subsys * ss,struct cgroup * cgrp,struct cgroup * old_cgrp,struct task_struct * task)7581 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7582 		struct cgroup *old_cgrp, struct task_struct *task)
7583 {
7584 	/*
7585 	 * cgroup_exit() is called in the copy_process() failure path.
7586 	 * Ignore this case since the task hasn't ran yet, this avoids
7587 	 * trying to poke a half freed task state from generic code.
7588 	 */
7589 	if (!(task->flags & PF_EXITING))
7590 		return;
7591 
7592 	sched_move_task(task);
7593 }
7594 
7595 #ifdef CONFIG_FAIR_GROUP_SCHED
cpu_shares_write_u64(struct cgroup * cgrp,struct cftype * cftype,u64 shareval)7596 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7597 				u64 shareval)
7598 {
7599 	return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7600 }
7601 
cpu_shares_read_u64(struct cgroup * cgrp,struct cftype * cft)7602 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7603 {
7604 	struct task_group *tg = cgroup_tg(cgrp);
7605 
7606 	return (u64) scale_load_down(tg->shares);
7607 }
7608 
7609 #ifdef CONFIG_CFS_BANDWIDTH
7610 static DEFINE_MUTEX(cfs_constraints_mutex);
7611 
7612 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7613 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7614 
7615 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7616 
tg_set_cfs_bandwidth(struct task_group * tg,u64 period,u64 quota)7617 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7618 {
7619 	int i, ret = 0, runtime_enabled, runtime_was_enabled;
7620 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7621 
7622 	if (tg == &root_task_group)
7623 		return -EINVAL;
7624 
7625 	/*
7626 	 * Ensure we have at some amount of bandwidth every period.  This is
7627 	 * to prevent reaching a state of large arrears when throttled via
7628 	 * entity_tick() resulting in prolonged exit starvation.
7629 	 */
7630 	if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7631 		return -EINVAL;
7632 
7633 	/*
7634 	 * Likewise, bound things on the otherside by preventing insane quota
7635 	 * periods.  This also allows us to normalize in computing quota
7636 	 * feasibility.
7637 	 */
7638 	if (period > max_cfs_quota_period)
7639 		return -EINVAL;
7640 
7641 	mutex_lock(&cfs_constraints_mutex);
7642 	ret = __cfs_schedulable(tg, period, quota);
7643 	if (ret)
7644 		goto out_unlock;
7645 
7646 	runtime_enabled = quota != RUNTIME_INF;
7647 	runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7648 	account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7649 	raw_spin_lock_irq(&cfs_b->lock);
7650 	cfs_b->period = ns_to_ktime(period);
7651 	cfs_b->quota = quota;
7652 
7653 	__refill_cfs_bandwidth_runtime(cfs_b);
7654 	/* restart the period timer (if active) to handle new period expiry */
7655 	if (runtime_enabled && cfs_b->timer_active) {
7656 		/* force a reprogram */
7657 		cfs_b->timer_active = 0;
7658 		__start_cfs_bandwidth(cfs_b);
7659 	}
7660 	raw_spin_unlock_irq(&cfs_b->lock);
7661 
7662 	for_each_possible_cpu(i) {
7663 		struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7664 		struct rq *rq = cfs_rq->rq;
7665 
7666 		raw_spin_lock_irq(&rq->lock);
7667 		cfs_rq->runtime_enabled = runtime_enabled;
7668 		cfs_rq->runtime_remaining = 0;
7669 
7670 		if (cfs_rq->throttled)
7671 			unthrottle_cfs_rq(cfs_rq);
7672 		raw_spin_unlock_irq(&rq->lock);
7673 	}
7674 out_unlock:
7675 	mutex_unlock(&cfs_constraints_mutex);
7676 
7677 	return ret;
7678 }
7679 
tg_set_cfs_quota(struct task_group * tg,long cfs_quota_us)7680 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7681 {
7682 	u64 quota, period;
7683 
7684 	period = ktime_to_ns(tg->cfs_bandwidth.period);
7685 	if (cfs_quota_us < 0)
7686 		quota = RUNTIME_INF;
7687 	else
7688 		quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7689 
7690 	return tg_set_cfs_bandwidth(tg, period, quota);
7691 }
7692 
tg_get_cfs_quota(struct task_group * tg)7693 long tg_get_cfs_quota(struct task_group *tg)
7694 {
7695 	u64 quota_us;
7696 
7697 	if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7698 		return -1;
7699 
7700 	quota_us = tg->cfs_bandwidth.quota;
7701 	do_div(quota_us, NSEC_PER_USEC);
7702 
7703 	return quota_us;
7704 }
7705 
tg_set_cfs_period(struct task_group * tg,long cfs_period_us)7706 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7707 {
7708 	u64 quota, period;
7709 
7710 	period = (u64)cfs_period_us * NSEC_PER_USEC;
7711 	quota = tg->cfs_bandwidth.quota;
7712 
7713 	return tg_set_cfs_bandwidth(tg, period, quota);
7714 }
7715 
tg_get_cfs_period(struct task_group * tg)7716 long tg_get_cfs_period(struct task_group *tg)
7717 {
7718 	u64 cfs_period_us;
7719 
7720 	cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7721 	do_div(cfs_period_us, NSEC_PER_USEC);
7722 
7723 	return cfs_period_us;
7724 }
7725 
cpu_cfs_quota_read_s64(struct cgroup * cgrp,struct cftype * cft)7726 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7727 {
7728 	return tg_get_cfs_quota(cgroup_tg(cgrp));
7729 }
7730 
cpu_cfs_quota_write_s64(struct cgroup * cgrp,struct cftype * cftype,s64 cfs_quota_us)7731 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7732 				s64 cfs_quota_us)
7733 {
7734 	return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7735 }
7736 
cpu_cfs_period_read_u64(struct cgroup * cgrp,struct cftype * cft)7737 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7738 {
7739 	return tg_get_cfs_period(cgroup_tg(cgrp));
7740 }
7741 
cpu_cfs_period_write_u64(struct cgroup * cgrp,struct cftype * cftype,u64 cfs_period_us)7742 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7743 				u64 cfs_period_us)
7744 {
7745 	return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7746 }
7747 
7748 struct cfs_schedulable_data {
7749 	struct task_group *tg;
7750 	u64 period, quota;
7751 };
7752 
7753 /*
7754  * normalize group quota/period to be quota/max_period
7755  * note: units are usecs
7756  */
normalize_cfs_quota(struct task_group * tg,struct cfs_schedulable_data * d)7757 static u64 normalize_cfs_quota(struct task_group *tg,
7758 			       struct cfs_schedulable_data *d)
7759 {
7760 	u64 quota, period;
7761 
7762 	if (tg == d->tg) {
7763 		period = d->period;
7764 		quota = d->quota;
7765 	} else {
7766 		period = tg_get_cfs_period(tg);
7767 		quota = tg_get_cfs_quota(tg);
7768 	}
7769 
7770 	/* note: these should typically be equivalent */
7771 	if (quota == RUNTIME_INF || quota == -1)
7772 		return RUNTIME_INF;
7773 
7774 	return to_ratio(period, quota);
7775 }
7776 
tg_cfs_schedulable_down(struct task_group * tg,void * data)7777 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7778 {
7779 	struct cfs_schedulable_data *d = data;
7780 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7781 	s64 quota = 0, parent_quota = -1;
7782 
7783 	if (!tg->parent) {
7784 		quota = RUNTIME_INF;
7785 	} else {
7786 		struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7787 
7788 		quota = normalize_cfs_quota(tg, d);
7789 		parent_quota = parent_b->hierarchal_quota;
7790 
7791 		/*
7792 		 * ensure max(child_quota) <= parent_quota, inherit when no
7793 		 * limit is set
7794 		 */
7795 		if (quota == RUNTIME_INF)
7796 			quota = parent_quota;
7797 		else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7798 			return -EINVAL;
7799 	}
7800 	cfs_b->hierarchal_quota = quota;
7801 
7802 	return 0;
7803 }
7804 
__cfs_schedulable(struct task_group * tg,u64 period,u64 quota)7805 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7806 {
7807 	int ret;
7808 	struct cfs_schedulable_data data = {
7809 		.tg = tg,
7810 		.period = period,
7811 		.quota = quota,
7812 	};
7813 
7814 	if (quota != RUNTIME_INF) {
7815 		do_div(data.period, NSEC_PER_USEC);
7816 		do_div(data.quota, NSEC_PER_USEC);
7817 	}
7818 
7819 	rcu_read_lock();
7820 	ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7821 	rcu_read_unlock();
7822 
7823 	return ret;
7824 }
7825 
cpu_stats_show(struct cgroup * cgrp,struct cftype * cft,struct cgroup_map_cb * cb)7826 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7827 		struct cgroup_map_cb *cb)
7828 {
7829 	struct task_group *tg = cgroup_tg(cgrp);
7830 	struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7831 
7832 	cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7833 	cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7834 	cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7835 
7836 	return 0;
7837 }
7838 #endif /* CONFIG_CFS_BANDWIDTH */
7839 #endif /* CONFIG_FAIR_GROUP_SCHED */
7840 
7841 #ifdef CONFIG_RT_GROUP_SCHED
cpu_rt_runtime_write(struct cgroup * cgrp,struct cftype * cft,s64 val)7842 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7843 				s64 val)
7844 {
7845 	return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7846 }
7847 
cpu_rt_runtime_read(struct cgroup * cgrp,struct cftype * cft)7848 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7849 {
7850 	return sched_group_rt_runtime(cgroup_tg(cgrp));
7851 }
7852 
cpu_rt_period_write_uint(struct cgroup * cgrp,struct cftype * cftype,u64 rt_period_us)7853 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7854 		u64 rt_period_us)
7855 {
7856 	return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7857 }
7858 
cpu_rt_period_read_uint(struct cgroup * cgrp,struct cftype * cft)7859 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7860 {
7861 	return sched_group_rt_period(cgroup_tg(cgrp));
7862 }
7863 #endif /* CONFIG_RT_GROUP_SCHED */
7864 
7865 static struct cftype cpu_files[] = {
7866 #ifdef CONFIG_FAIR_GROUP_SCHED
7867 	{
7868 		.name = "shares",
7869 		.read_u64 = cpu_shares_read_u64,
7870 		.write_u64 = cpu_shares_write_u64,
7871 	},
7872 #endif
7873 #ifdef CONFIG_CFS_BANDWIDTH
7874 	{
7875 		.name = "cfs_quota_us",
7876 		.read_s64 = cpu_cfs_quota_read_s64,
7877 		.write_s64 = cpu_cfs_quota_write_s64,
7878 	},
7879 	{
7880 		.name = "cfs_period_us",
7881 		.read_u64 = cpu_cfs_period_read_u64,
7882 		.write_u64 = cpu_cfs_period_write_u64,
7883 	},
7884 	{
7885 		.name = "stat",
7886 		.read_map = cpu_stats_show,
7887 	},
7888 #endif
7889 #ifdef CONFIG_RT_GROUP_SCHED
7890 	{
7891 		.name = "rt_runtime_us",
7892 		.read_s64 = cpu_rt_runtime_read,
7893 		.write_s64 = cpu_rt_runtime_write,
7894 	},
7895 	{
7896 		.name = "rt_period_us",
7897 		.read_u64 = cpu_rt_period_read_uint,
7898 		.write_u64 = cpu_rt_period_write_uint,
7899 	},
7900 #endif
7901 };
7902 
cpu_cgroup_populate(struct cgroup_subsys * ss,struct cgroup * cont)7903 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7904 {
7905 	return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7906 }
7907 
7908 struct cgroup_subsys cpu_cgroup_subsys = {
7909 	.name		= "cpu",
7910 	.create		= cpu_cgroup_create,
7911 	.destroy	= cpu_cgroup_destroy,
7912 	.can_attach	= cpu_cgroup_can_attach,
7913 	.attach		= cpu_cgroup_attach,
7914 	.exit		= cpu_cgroup_exit,
7915 	.populate	= cpu_cgroup_populate,
7916 	.subsys_id	= cpu_cgroup_subsys_id,
7917 	.early_init	= 1,
7918 };
7919 
7920 #endif	/* CONFIG_CGROUP_SCHED */
7921 
7922 #ifdef CONFIG_CGROUP_CPUACCT
7923 
7924 /*
7925  * CPU accounting code for task groups.
7926  *
7927  * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7928  * (balbir@in.ibm.com).
7929  */
7930 
7931 /* create a new cpu accounting group */
cpuacct_create(struct cgroup_subsys * ss,struct cgroup * cgrp)7932 static struct cgroup_subsys_state *cpuacct_create(
7933 	struct cgroup_subsys *ss, struct cgroup *cgrp)
7934 {
7935 	struct cpuacct *ca;
7936 
7937 	if (!cgrp->parent)
7938 		return &root_cpuacct.css;
7939 
7940 	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7941 	if (!ca)
7942 		goto out;
7943 
7944 	ca->cpuusage = alloc_percpu(u64);
7945 	if (!ca->cpuusage)
7946 		goto out_free_ca;
7947 
7948 	ca->cpustat = alloc_percpu(struct kernel_cpustat);
7949 	if (!ca->cpustat)
7950 		goto out_free_cpuusage;
7951 
7952 	return &ca->css;
7953 
7954 out_free_cpuusage:
7955 	free_percpu(ca->cpuusage);
7956 out_free_ca:
7957 	kfree(ca);
7958 out:
7959 	return ERR_PTR(-ENOMEM);
7960 }
7961 
7962 /* destroy an existing cpu accounting group */
7963 static void
cpuacct_destroy(struct cgroup_subsys * ss,struct cgroup * cgrp)7964 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7965 {
7966 	struct cpuacct *ca = cgroup_ca(cgrp);
7967 
7968 	free_percpu(ca->cpustat);
7969 	free_percpu(ca->cpuusage);
7970 	kfree(ca);
7971 }
7972 
cpuacct_cpuusage_read(struct cpuacct * ca,int cpu)7973 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7974 {
7975 	u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7976 	u64 data;
7977 
7978 #ifndef CONFIG_64BIT
7979 	/*
7980 	 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7981 	 */
7982 	raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7983 	data = *cpuusage;
7984 	raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7985 #else
7986 	data = *cpuusage;
7987 #endif
7988 
7989 	return data;
7990 }
7991 
cpuacct_cpuusage_write(struct cpuacct * ca,int cpu,u64 val)7992 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7993 {
7994 	u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7995 
7996 #ifndef CONFIG_64BIT
7997 	/*
7998 	 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7999 	 */
8000 	raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8001 	*cpuusage = val;
8002 	raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8003 #else
8004 	*cpuusage = val;
8005 #endif
8006 }
8007 
8008 /* return total cpu usage (in nanoseconds) of a group */
cpuusage_read(struct cgroup * cgrp,struct cftype * cft)8009 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8010 {
8011 	struct cpuacct *ca = cgroup_ca(cgrp);
8012 	u64 totalcpuusage = 0;
8013 	int i;
8014 
8015 	for_each_present_cpu(i)
8016 		totalcpuusage += cpuacct_cpuusage_read(ca, i);
8017 
8018 	return totalcpuusage;
8019 }
8020 
cpuusage_write(struct cgroup * cgrp,struct cftype * cftype,u64 reset)8021 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8022 								u64 reset)
8023 {
8024 	struct cpuacct *ca = cgroup_ca(cgrp);
8025 	int err = 0;
8026 	int i;
8027 
8028 	if (reset) {
8029 		err = -EINVAL;
8030 		goto out;
8031 	}
8032 
8033 	for_each_present_cpu(i)
8034 		cpuacct_cpuusage_write(ca, i, 0);
8035 
8036 out:
8037 	return err;
8038 }
8039 
cpuacct_percpu_seq_read(struct cgroup * cgroup,struct cftype * cft,struct seq_file * m)8040 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8041 				   struct seq_file *m)
8042 {
8043 	struct cpuacct *ca = cgroup_ca(cgroup);
8044 	u64 percpu;
8045 	int i;
8046 
8047 	for_each_present_cpu(i) {
8048 		percpu = cpuacct_cpuusage_read(ca, i);
8049 		seq_printf(m, "%llu ", (unsigned long long) percpu);
8050 	}
8051 	seq_printf(m, "\n");
8052 	return 0;
8053 }
8054 
8055 static const char *cpuacct_stat_desc[] = {
8056 	[CPUACCT_STAT_USER] = "user",
8057 	[CPUACCT_STAT_SYSTEM] = "system",
8058 };
8059 
cpuacct_stats_show(struct cgroup * cgrp,struct cftype * cft,struct cgroup_map_cb * cb)8060 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8061 			      struct cgroup_map_cb *cb)
8062 {
8063 	struct cpuacct *ca = cgroup_ca(cgrp);
8064 	int cpu;
8065 	s64 val = 0;
8066 
8067 	for_each_online_cpu(cpu) {
8068 		struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8069 		val += kcpustat->cpustat[CPUTIME_USER];
8070 		val += kcpustat->cpustat[CPUTIME_NICE];
8071 	}
8072 	val = cputime64_to_clock_t(val);
8073 	cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8074 
8075 	val = 0;
8076 	for_each_online_cpu(cpu) {
8077 		struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8078 		val += kcpustat->cpustat[CPUTIME_SYSTEM];
8079 		val += kcpustat->cpustat[CPUTIME_IRQ];
8080 		val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8081 	}
8082 
8083 	val = cputime64_to_clock_t(val);
8084 	cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8085 
8086 	return 0;
8087 }
8088 
8089 static struct cftype files[] = {
8090 	{
8091 		.name = "usage",
8092 		.read_u64 = cpuusage_read,
8093 		.write_u64 = cpuusage_write,
8094 	},
8095 	{
8096 		.name = "usage_percpu",
8097 		.read_seq_string = cpuacct_percpu_seq_read,
8098 	},
8099 	{
8100 		.name = "stat",
8101 		.read_map = cpuacct_stats_show,
8102 	},
8103 };
8104 
cpuacct_populate(struct cgroup_subsys * ss,struct cgroup * cgrp)8105 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8106 {
8107 	return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8108 }
8109 
8110 /*
8111  * charge this task's execution time to its accounting group.
8112  *
8113  * called with rq->lock held.
8114  */
cpuacct_charge(struct task_struct * tsk,u64 cputime)8115 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8116 {
8117 	struct cpuacct *ca;
8118 	int cpu;
8119 
8120 	if (unlikely(!cpuacct_subsys.active))
8121 		return;
8122 
8123 	cpu = task_cpu(tsk);
8124 
8125 	rcu_read_lock();
8126 
8127 	ca = task_ca(tsk);
8128 
8129 	for (; ca; ca = parent_ca(ca)) {
8130 		u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8131 		*cpuusage += cputime;
8132 	}
8133 
8134 	rcu_read_unlock();
8135 }
8136 
8137 struct cgroup_subsys cpuacct_subsys = {
8138 	.name = "cpuacct",
8139 	.create = cpuacct_create,
8140 	.destroy = cpuacct_destroy,
8141 	.populate = cpuacct_populate,
8142 	.subsys_id = cpuacct_subsys_id,
8143 };
8144 #endif	/* CONFIG_CGROUP_CPUACCT */
8145