Loading...
1/*
2 * kernel/sched/core.c
3 *
4 * Kernel scheduler and related syscalls
5 *
6 * Copyright (C) 1991-2002 Linus Torvalds
7 *
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
27 */
28
29#include <linux/mm.h>
30#include <linux/module.h>
31#include <linux/nmi.h>
32#include <linux/init.h>
33#include <linux/uaccess.h>
34#include <linux/highmem.h>
35#include <asm/mmu_context.h>
36#include <linux/interrupt.h>
37#include <linux/capability.h>
38#include <linux/completion.h>
39#include <linux/kernel_stat.h>
40#include <linux/debug_locks.h>
41#include <linux/perf_event.h>
42#include <linux/security.h>
43#include <linux/notifier.h>
44#include <linux/profile.h>
45#include <linux/freezer.h>
46#include <linux/vmalloc.h>
47#include <linux/blkdev.h>
48#include <linux/delay.h>
49#include <linux/pid_namespace.h>
50#include <linux/smp.h>
51#include <linux/threads.h>
52#include <linux/timer.h>
53#include <linux/rcupdate.h>
54#include <linux/cpu.h>
55#include <linux/cpuset.h>
56#include <linux/percpu.h>
57#include <linux/proc_fs.h>
58#include <linux/seq_file.h>
59#include <linux/sysctl.h>
60#include <linux/syscalls.h>
61#include <linux/times.h>
62#include <linux/tsacct_kern.h>
63#include <linux/kprobes.h>
64#include <linux/delayacct.h>
65#include <linux/unistd.h>
66#include <linux/pagemap.h>
67#include <linux/hrtimer.h>
68#include <linux/tick.h>
69#include <linux/debugfs.h>
70#include <linux/ctype.h>
71#include <linux/ftrace.h>
72#include <linux/slab.h>
73#include <linux/init_task.h>
74#include <linux/binfmts.h>
75
76#include <asm/switch_to.h>
77#include <asm/tlb.h>
78#include <asm/irq_regs.h>
79#include <asm/mutex.h>
80#ifdef CONFIG_PARAVIRT
81#include <asm/paravirt.h>
82#endif
83
84#include "sched.h"
85#include "../workqueue_sched.h"
86#include "../smpboot.h"
87
88#define CREATE_TRACE_POINTS
89#include <trace/events/sched.h>
90
91void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
92{
93 unsigned long delta;
94 ktime_t soft, hard, now;
95
96 for (;;) {
97 if (hrtimer_active(period_timer))
98 break;
99
100 now = hrtimer_cb_get_time(period_timer);
101 hrtimer_forward(period_timer, now, period);
102
103 soft = hrtimer_get_softexpires(period_timer);
104 hard = hrtimer_get_expires(period_timer);
105 delta = ktime_to_ns(ktime_sub(hard, soft));
106 __hrtimer_start_range_ns(period_timer, soft, delta,
107 HRTIMER_MODE_ABS_PINNED, 0);
108 }
109}
110
111DEFINE_MUTEX(sched_domains_mutex);
112DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
113
114static void update_rq_clock_task(struct rq *rq, s64 delta);
115
116void update_rq_clock(struct rq *rq)
117{
118 s64 delta;
119
120 if (rq->skip_clock_update > 0)
121 return;
122
123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
124 rq->clock += delta;
125 update_rq_clock_task(rq, delta);
126}
127
128/*
129 * Debugging: various feature bits
130 */
131
132#define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
134
135const_debug unsigned int sysctl_sched_features =
136#include "features.h"
137 0;
138
139#undef SCHED_FEAT
140
141#ifdef CONFIG_SCHED_DEBUG
142#define SCHED_FEAT(name, enabled) \
143 #name ,
144
145static const char * const sched_feat_names[] = {
146#include "features.h"
147};
148
149#undef SCHED_FEAT
150
151static int sched_feat_show(struct seq_file *m, void *v)
152{
153 int i;
154
155 for (i = 0; i < __SCHED_FEAT_NR; i++) {
156 if (!(sysctl_sched_features & (1UL << i)))
157 seq_puts(m, "NO_");
158 seq_printf(m, "%s ", sched_feat_names[i]);
159 }
160 seq_puts(m, "\n");
161
162 return 0;
163}
164
165#ifdef HAVE_JUMP_LABEL
166
167#define jump_label_key__true STATIC_KEY_INIT_TRUE
168#define jump_label_key__false STATIC_KEY_INIT_FALSE
169
170#define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
172
173struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
174#include "features.h"
175};
176
177#undef SCHED_FEAT
178
179static void sched_feat_disable(int i)
180{
181 if (static_key_enabled(&sched_feat_keys[i]))
182 static_key_slow_dec(&sched_feat_keys[i]);
183}
184
185static void sched_feat_enable(int i)
186{
187 if (!static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_inc(&sched_feat_keys[i]);
189}
190#else
191static void sched_feat_disable(int i) { };
192static void sched_feat_enable(int i) { };
193#endif /* HAVE_JUMP_LABEL */
194
195static ssize_t
196sched_feat_write(struct file *filp, const char __user *ubuf,
197 size_t cnt, loff_t *ppos)
198{
199 char buf[64];
200 char *cmp;
201 int neg = 0;
202 int i;
203
204 if (cnt > 63)
205 cnt = 63;
206
207 if (copy_from_user(&buf, ubuf, cnt))
208 return -EFAULT;
209
210 buf[cnt] = 0;
211 cmp = strstrip(buf);
212
213 if (strncmp(cmp, "NO_", 3) == 0) {
214 neg = 1;
215 cmp += 3;
216 }
217
218 for (i = 0; i < __SCHED_FEAT_NR; i++) {
219 if (strcmp(cmp, sched_feat_names[i]) == 0) {
220 if (neg) {
221 sysctl_sched_features &= ~(1UL << i);
222 sched_feat_disable(i);
223 } else {
224 sysctl_sched_features |= (1UL << i);
225 sched_feat_enable(i);
226 }
227 break;
228 }
229 }
230
231 if (i == __SCHED_FEAT_NR)
232 return -EINVAL;
233
234 *ppos += cnt;
235
236 return cnt;
237}
238
239static int sched_feat_open(struct inode *inode, struct file *filp)
240{
241 return single_open(filp, sched_feat_show, NULL);
242}
243
244static const struct file_operations sched_feat_fops = {
245 .open = sched_feat_open,
246 .write = sched_feat_write,
247 .read = seq_read,
248 .llseek = seq_lseek,
249 .release = single_release,
250};
251
252static __init int sched_init_debug(void)
253{
254 debugfs_create_file("sched_features", 0644, NULL, NULL,
255 &sched_feat_fops);
256
257 return 0;
258}
259late_initcall(sched_init_debug);
260#endif /* CONFIG_SCHED_DEBUG */
261
262/*
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
265 */
266const_debug unsigned int sysctl_sched_nr_migrate = 32;
267
268/*
269 * period over which we average the RT time consumption, measured
270 * in ms.
271 *
272 * default: 1s
273 */
274const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
275
276/*
277 * period over which we measure -rt task cpu usage in us.
278 * default: 1s
279 */
280unsigned int sysctl_sched_rt_period = 1000000;
281
282__read_mostly int scheduler_running;
283
284/*
285 * part of the period that we allow rt tasks to run in us.
286 * default: 0.95s
287 */
288int sysctl_sched_rt_runtime = 950000;
289
290
291
292/*
293 * __task_rq_lock - lock the rq @p resides on.
294 */
295static inline struct rq *__task_rq_lock(struct task_struct *p)
296 __acquires(rq->lock)
297{
298 struct rq *rq;
299
300 lockdep_assert_held(&p->pi_lock);
301
302 for (;;) {
303 rq = task_rq(p);
304 raw_spin_lock(&rq->lock);
305 if (likely(rq == task_rq(p)))
306 return rq;
307 raw_spin_unlock(&rq->lock);
308 }
309}
310
311/*
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
313 */
314static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
315 __acquires(p->pi_lock)
316 __acquires(rq->lock)
317{
318 struct rq *rq;
319
320 for (;;) {
321 raw_spin_lock_irqsave(&p->pi_lock, *flags);
322 rq = task_rq(p);
323 raw_spin_lock(&rq->lock);
324 if (likely(rq == task_rq(p)))
325 return rq;
326 raw_spin_unlock(&rq->lock);
327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
328 }
329}
330
331static void __task_rq_unlock(struct rq *rq)
332 __releases(rq->lock)
333{
334 raw_spin_unlock(&rq->lock);
335}
336
337static inline void
338task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
339 __releases(rq->lock)
340 __releases(p->pi_lock)
341{
342 raw_spin_unlock(&rq->lock);
343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
344}
345
346/*
347 * this_rq_lock - lock this runqueue and disable interrupts.
348 */
349static struct rq *this_rq_lock(void)
350 __acquires(rq->lock)
351{
352 struct rq *rq;
353
354 local_irq_disable();
355 rq = this_rq();
356 raw_spin_lock(&rq->lock);
357
358 return rq;
359}
360
361#ifdef CONFIG_SCHED_HRTICK
362/*
363 * Use HR-timers to deliver accurate preemption points.
364 *
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
367 * reschedule event.
368 *
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
370 * rq->lock.
371 */
372
373static void hrtick_clear(struct rq *rq)
374{
375 if (hrtimer_active(&rq->hrtick_timer))
376 hrtimer_cancel(&rq->hrtick_timer);
377}
378
379/*
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
382 */
383static enum hrtimer_restart hrtick(struct hrtimer *timer)
384{
385 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
386
387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
388
389 raw_spin_lock(&rq->lock);
390 update_rq_clock(rq);
391 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392 raw_spin_unlock(&rq->lock);
393
394 return HRTIMER_NORESTART;
395}
396
397#ifdef CONFIG_SMP
398/*
399 * called from hardirq (IPI) context
400 */
401static void __hrtick_start(void *arg)
402{
403 struct rq *rq = arg;
404
405 raw_spin_lock(&rq->lock);
406 hrtimer_restart(&rq->hrtick_timer);
407 rq->hrtick_csd_pending = 0;
408 raw_spin_unlock(&rq->lock);
409}
410
411/*
412 * Called to set the hrtick timer state.
413 *
414 * called with rq->lock held and irqs disabled
415 */
416void hrtick_start(struct rq *rq, u64 delay)
417{
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
420
421 hrtimer_set_expires(timer, time);
422
423 if (rq == this_rq()) {
424 hrtimer_restart(timer);
425 } else if (!rq->hrtick_csd_pending) {
426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
427 rq->hrtick_csd_pending = 1;
428 }
429}
430
431static int
432hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
433{
434 int cpu = (int)(long)hcpu;
435
436 switch (action) {
437 case CPU_UP_CANCELED:
438 case CPU_UP_CANCELED_FROZEN:
439 case CPU_DOWN_PREPARE:
440 case CPU_DOWN_PREPARE_FROZEN:
441 case CPU_DEAD:
442 case CPU_DEAD_FROZEN:
443 hrtick_clear(cpu_rq(cpu));
444 return NOTIFY_OK;
445 }
446
447 return NOTIFY_DONE;
448}
449
450static __init void init_hrtick(void)
451{
452 hotcpu_notifier(hotplug_hrtick, 0);
453}
454#else
455/*
456 * Called to set the hrtick timer state.
457 *
458 * called with rq->lock held and irqs disabled
459 */
460void hrtick_start(struct rq *rq, u64 delay)
461{
462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
463 HRTIMER_MODE_REL_PINNED, 0);
464}
465
466static inline void init_hrtick(void)
467{
468}
469#endif /* CONFIG_SMP */
470
471static void init_rq_hrtick(struct rq *rq)
472{
473#ifdef CONFIG_SMP
474 rq->hrtick_csd_pending = 0;
475
476 rq->hrtick_csd.flags = 0;
477 rq->hrtick_csd.func = __hrtick_start;
478 rq->hrtick_csd.info = rq;
479#endif
480
481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
482 rq->hrtick_timer.function = hrtick;
483}
484#else /* CONFIG_SCHED_HRTICK */
485static inline void hrtick_clear(struct rq *rq)
486{
487}
488
489static inline void init_rq_hrtick(struct rq *rq)
490{
491}
492
493static inline void init_hrtick(void)
494{
495}
496#endif /* CONFIG_SCHED_HRTICK */
497
498/*
499 * resched_task - mark a task 'to be rescheduled now'.
500 *
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
503 * the target CPU.
504 */
505#ifdef CONFIG_SMP
506
507#ifndef tsk_is_polling
508#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
509#endif
510
511void resched_task(struct task_struct *p)
512{
513 int cpu;
514
515 assert_raw_spin_locked(&task_rq(p)->lock);
516
517 if (test_tsk_need_resched(p))
518 return;
519
520 set_tsk_need_resched(p);
521
522 cpu = task_cpu(p);
523 if (cpu == smp_processor_id())
524 return;
525
526 /* NEED_RESCHED must be visible before we test polling */
527 smp_mb();
528 if (!tsk_is_polling(p))
529 smp_send_reschedule(cpu);
530}
531
532void resched_cpu(int cpu)
533{
534 struct rq *rq = cpu_rq(cpu);
535 unsigned long flags;
536
537 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
538 return;
539 resched_task(cpu_curr(cpu));
540 raw_spin_unlock_irqrestore(&rq->lock, flags);
541}
542
543#ifdef CONFIG_NO_HZ
544/*
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
547 *
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
551 */
552int get_nohz_timer_target(void)
553{
554 int cpu = smp_processor_id();
555 int i;
556 struct sched_domain *sd;
557
558 rcu_read_lock();
559 for_each_domain(cpu, sd) {
560 for_each_cpu(i, sched_domain_span(sd)) {
561 if (!idle_cpu(i)) {
562 cpu = i;
563 goto unlock;
564 }
565 }
566 }
567unlock:
568 rcu_read_unlock();
569 return cpu;
570}
571/*
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
580 */
581void wake_up_idle_cpu(int cpu)
582{
583 struct rq *rq = cpu_rq(cpu);
584
585 if (cpu == smp_processor_id())
586 return;
587
588 /*
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
594 */
595 if (rq->curr != rq->idle)
596 return;
597
598 /*
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
602 */
603 set_tsk_need_resched(rq->idle);
604
605 /* NEED_RESCHED must be visible before we test polling */
606 smp_mb();
607 if (!tsk_is_polling(rq->idle))
608 smp_send_reschedule(cpu);
609}
610
611static inline bool got_nohz_idle_kick(void)
612{
613 int cpu = smp_processor_id();
614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
615}
616
617#else /* CONFIG_NO_HZ */
618
619static inline bool got_nohz_idle_kick(void)
620{
621 return false;
622}
623
624#endif /* CONFIG_NO_HZ */
625
626void sched_avg_update(struct rq *rq)
627{
628 s64 period = sched_avg_period();
629
630 while ((s64)(rq->clock - rq->age_stamp) > period) {
631 /*
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
635 */
636 asm("" : "+rm" (rq->age_stamp));
637 rq->age_stamp += period;
638 rq->rt_avg /= 2;
639 }
640}
641
642#else /* !CONFIG_SMP */
643void resched_task(struct task_struct *p)
644{
645 assert_raw_spin_locked(&task_rq(p)->lock);
646 set_tsk_need_resched(p);
647}
648#endif /* CONFIG_SMP */
649
650#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
652/*
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
655 *
656 * Caller must hold rcu_lock or sufficient equivalent.
657 */
658int walk_tg_tree_from(struct task_group *from,
659 tg_visitor down, tg_visitor up, void *data)
660{
661 struct task_group *parent, *child;
662 int ret;
663
664 parent = from;
665
666down:
667 ret = (*down)(parent, data);
668 if (ret)
669 goto out;
670 list_for_each_entry_rcu(child, &parent->children, siblings) {
671 parent = child;
672 goto down;
673
674up:
675 continue;
676 }
677 ret = (*up)(parent, data);
678 if (ret || parent == from)
679 goto out;
680
681 child = parent;
682 parent = parent->parent;
683 if (parent)
684 goto up;
685out:
686 return ret;
687}
688
689int tg_nop(struct task_group *tg, void *data)
690{
691 return 0;
692}
693#endif
694
695static void set_load_weight(struct task_struct *p)
696{
697 int prio = p->static_prio - MAX_RT_PRIO;
698 struct load_weight *load = &p->se.load;
699
700 /*
701 * SCHED_IDLE tasks get minimal weight:
702 */
703 if (p->policy == SCHED_IDLE) {
704 load->weight = scale_load(WEIGHT_IDLEPRIO);
705 load->inv_weight = WMULT_IDLEPRIO;
706 return;
707 }
708
709 load->weight = scale_load(prio_to_weight[prio]);
710 load->inv_weight = prio_to_wmult[prio];
711}
712
713static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
714{
715 update_rq_clock(rq);
716 sched_info_queued(p);
717 p->sched_class->enqueue_task(rq, p, flags);
718}
719
720static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
721{
722 update_rq_clock(rq);
723 sched_info_dequeued(p);
724 p->sched_class->dequeue_task(rq, p, flags);
725}
726
727void activate_task(struct rq *rq, struct task_struct *p, int flags)
728{
729 if (task_contributes_to_load(p))
730 rq->nr_uninterruptible--;
731
732 enqueue_task(rq, p, flags);
733}
734
735void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
736{
737 if (task_contributes_to_load(p))
738 rq->nr_uninterruptible++;
739
740 dequeue_task(rq, p, flags);
741}
742
743#ifdef CONFIG_IRQ_TIME_ACCOUNTING
744
745/*
746 * There are no locks covering percpu hardirq/softirq time.
747 * They are only modified in account_system_vtime, on corresponding CPU
748 * with interrupts disabled. So, writes are safe.
749 * They are read and saved off onto struct rq in update_rq_clock().
750 * This may result in other CPU reading this CPU's irq time and can
751 * race with irq/account_system_vtime on this CPU. We would either get old
752 * or new value with a side effect of accounting a slice of irq time to wrong
753 * task when irq is in progress while we read rq->clock. That is a worthy
754 * compromise in place of having locks on each irq in account_system_time.
755 */
756static DEFINE_PER_CPU(u64, cpu_hardirq_time);
757static DEFINE_PER_CPU(u64, cpu_softirq_time);
758
759static DEFINE_PER_CPU(u64, irq_start_time);
760static int sched_clock_irqtime;
761
762void enable_sched_clock_irqtime(void)
763{
764 sched_clock_irqtime = 1;
765}
766
767void disable_sched_clock_irqtime(void)
768{
769 sched_clock_irqtime = 0;
770}
771
772#ifndef CONFIG_64BIT
773static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
774
775static inline void irq_time_write_begin(void)
776{
777 __this_cpu_inc(irq_time_seq.sequence);
778 smp_wmb();
779}
780
781static inline void irq_time_write_end(void)
782{
783 smp_wmb();
784 __this_cpu_inc(irq_time_seq.sequence);
785}
786
787static inline u64 irq_time_read(int cpu)
788{
789 u64 irq_time;
790 unsigned seq;
791
792 do {
793 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
794 irq_time = per_cpu(cpu_softirq_time, cpu) +
795 per_cpu(cpu_hardirq_time, cpu);
796 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
797
798 return irq_time;
799}
800#else /* CONFIG_64BIT */
801static inline void irq_time_write_begin(void)
802{
803}
804
805static inline void irq_time_write_end(void)
806{
807}
808
809static inline u64 irq_time_read(int cpu)
810{
811 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
812}
813#endif /* CONFIG_64BIT */
814
815/*
816 * Called before incrementing preempt_count on {soft,}irq_enter
817 * and before decrementing preempt_count on {soft,}irq_exit.
818 */
819void account_system_vtime(struct task_struct *curr)
820{
821 unsigned long flags;
822 s64 delta;
823 int cpu;
824
825 if (!sched_clock_irqtime)
826 return;
827
828 local_irq_save(flags);
829
830 cpu = smp_processor_id();
831 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
832 __this_cpu_add(irq_start_time, delta);
833
834 irq_time_write_begin();
835 /*
836 * We do not account for softirq time from ksoftirqd here.
837 * We want to continue accounting softirq time to ksoftirqd thread
838 * in that case, so as not to confuse scheduler with a special task
839 * that do not consume any time, but still wants to run.
840 */
841 if (hardirq_count())
842 __this_cpu_add(cpu_hardirq_time, delta);
843 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
844 __this_cpu_add(cpu_softirq_time, delta);
845
846 irq_time_write_end();
847 local_irq_restore(flags);
848}
849EXPORT_SYMBOL_GPL(account_system_vtime);
850
851#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
852
853#ifdef CONFIG_PARAVIRT
854static inline u64 steal_ticks(u64 steal)
855{
856 if (unlikely(steal > NSEC_PER_SEC))
857 return div_u64(steal, TICK_NSEC);
858
859 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
860}
861#endif
862
863static void update_rq_clock_task(struct rq *rq, s64 delta)
864{
865/*
866 * In theory, the compile should just see 0 here, and optimize out the call
867 * to sched_rt_avg_update. But I don't trust it...
868 */
869#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870 s64 steal = 0, irq_delta = 0;
871#endif
872#ifdef CONFIG_IRQ_TIME_ACCOUNTING
873 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
874
875 /*
876 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877 * this case when a previous update_rq_clock() happened inside a
878 * {soft,}irq region.
879 *
880 * When this happens, we stop ->clock_task and only update the
881 * prev_irq_time stamp to account for the part that fit, so that a next
882 * update will consume the rest. This ensures ->clock_task is
883 * monotonic.
884 *
885 * It does however cause some slight miss-attribution of {soft,}irq
886 * time, a more accurate solution would be to update the irq_time using
887 * the current rq->clock timestamp, except that would require using
888 * atomic ops.
889 */
890 if (irq_delta > delta)
891 irq_delta = delta;
892
893 rq->prev_irq_time += irq_delta;
894 delta -= irq_delta;
895#endif
896#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897 if (static_key_false((¶virt_steal_rq_enabled))) {
898 u64 st;
899
900 steal = paravirt_steal_clock(cpu_of(rq));
901 steal -= rq->prev_steal_time_rq;
902
903 if (unlikely(steal > delta))
904 steal = delta;
905
906 st = steal_ticks(steal);
907 steal = st * TICK_NSEC;
908
909 rq->prev_steal_time_rq += steal;
910
911 delta -= steal;
912 }
913#endif
914
915 rq->clock_task += delta;
916
917#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
918 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
919 sched_rt_avg_update(rq, irq_delta + steal);
920#endif
921}
922
923#ifdef CONFIG_IRQ_TIME_ACCOUNTING
924static int irqtime_account_hi_update(void)
925{
926 u64 *cpustat = kcpustat_this_cpu->cpustat;
927 unsigned long flags;
928 u64 latest_ns;
929 int ret = 0;
930
931 local_irq_save(flags);
932 latest_ns = this_cpu_read(cpu_hardirq_time);
933 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
934 ret = 1;
935 local_irq_restore(flags);
936 return ret;
937}
938
939static int irqtime_account_si_update(void)
940{
941 u64 *cpustat = kcpustat_this_cpu->cpustat;
942 unsigned long flags;
943 u64 latest_ns;
944 int ret = 0;
945
946 local_irq_save(flags);
947 latest_ns = this_cpu_read(cpu_softirq_time);
948 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
949 ret = 1;
950 local_irq_restore(flags);
951 return ret;
952}
953
954#else /* CONFIG_IRQ_TIME_ACCOUNTING */
955
956#define sched_clock_irqtime (0)
957
958#endif
959
960void sched_set_stop_task(int cpu, struct task_struct *stop)
961{
962 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
963 struct task_struct *old_stop = cpu_rq(cpu)->stop;
964
965 if (stop) {
966 /*
967 * Make it appear like a SCHED_FIFO task, its something
968 * userspace knows about and won't get confused about.
969 *
970 * Also, it will make PI more or less work without too
971 * much confusion -- but then, stop work should not
972 * rely on PI working anyway.
973 */
974 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
975
976 stop->sched_class = &stop_sched_class;
977 }
978
979 cpu_rq(cpu)->stop = stop;
980
981 if (old_stop) {
982 /*
983 * Reset it back to a normal scheduling class so that
984 * it can die in pieces.
985 */
986 old_stop->sched_class = &rt_sched_class;
987 }
988}
989
990/*
991 * __normal_prio - return the priority that is based on the static prio
992 */
993static inline int __normal_prio(struct task_struct *p)
994{
995 return p->static_prio;
996}
997
998/*
999 * Calculate the expected normal priority: i.e. priority
1000 * without taking RT-inheritance into account. Might be
1001 * boosted by interactivity modifiers. Changes upon fork,
1002 * setprio syscalls, and whenever the interactivity
1003 * estimator recalculates.
1004 */
1005static inline int normal_prio(struct task_struct *p)
1006{
1007 int prio;
1008
1009 if (task_has_rt_policy(p))
1010 prio = MAX_RT_PRIO-1 - p->rt_priority;
1011 else
1012 prio = __normal_prio(p);
1013 return prio;
1014}
1015
1016/*
1017 * Calculate the current priority, i.e. the priority
1018 * taken into account by the scheduler. This value might
1019 * be boosted by RT tasks, or might be boosted by
1020 * interactivity modifiers. Will be RT if the task got
1021 * RT-boosted. If not then it returns p->normal_prio.
1022 */
1023static int effective_prio(struct task_struct *p)
1024{
1025 p->normal_prio = normal_prio(p);
1026 /*
1027 * If we are RT tasks or we were boosted to RT priority,
1028 * keep the priority unchanged. Otherwise, update priority
1029 * to the normal priority:
1030 */
1031 if (!rt_prio(p->prio))
1032 return p->normal_prio;
1033 return p->prio;
1034}
1035
1036/**
1037 * task_curr - is this task currently executing on a CPU?
1038 * @p: the task in question.
1039 */
1040inline int task_curr(const struct task_struct *p)
1041{
1042 return cpu_curr(task_cpu(p)) == p;
1043}
1044
1045static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1046 const struct sched_class *prev_class,
1047 int oldprio)
1048{
1049 if (prev_class != p->sched_class) {
1050 if (prev_class->switched_from)
1051 prev_class->switched_from(rq, p);
1052 p->sched_class->switched_to(rq, p);
1053 } else if (oldprio != p->prio)
1054 p->sched_class->prio_changed(rq, p, oldprio);
1055}
1056
1057void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1058{
1059 const struct sched_class *class;
1060
1061 if (p->sched_class == rq->curr->sched_class) {
1062 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1063 } else {
1064 for_each_class(class) {
1065 if (class == rq->curr->sched_class)
1066 break;
1067 if (class == p->sched_class) {
1068 resched_task(rq->curr);
1069 break;
1070 }
1071 }
1072 }
1073
1074 /*
1075 * A queue event has occurred, and we're going to schedule. In
1076 * this case, we can save a useless back to back clock update.
1077 */
1078 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1079 rq->skip_clock_update = 1;
1080}
1081
1082#ifdef CONFIG_SMP
1083void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1084{
1085#ifdef CONFIG_SCHED_DEBUG
1086 /*
1087 * We should never call set_task_cpu() on a blocked task,
1088 * ttwu() will sort out the placement.
1089 */
1090 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1091 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1092
1093#ifdef CONFIG_LOCKDEP
1094 /*
1095 * The caller should hold either p->pi_lock or rq->lock, when changing
1096 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1097 *
1098 * sched_move_task() holds both and thus holding either pins the cgroup,
1099 * see task_group().
1100 *
1101 * Furthermore, all task_rq users should acquire both locks, see
1102 * task_rq_lock().
1103 */
1104 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1105 lockdep_is_held(&task_rq(p)->lock)));
1106#endif
1107#endif
1108
1109 trace_sched_migrate_task(p, new_cpu);
1110
1111 if (task_cpu(p) != new_cpu) {
1112 p->se.nr_migrations++;
1113 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1114 }
1115
1116 __set_task_cpu(p, new_cpu);
1117}
1118
1119struct migration_arg {
1120 struct task_struct *task;
1121 int dest_cpu;
1122};
1123
1124static int migration_cpu_stop(void *data);
1125
1126/*
1127 * wait_task_inactive - wait for a thread to unschedule.
1128 *
1129 * If @match_state is nonzero, it's the @p->state value just checked and
1130 * not expected to change. If it changes, i.e. @p might have woken up,
1131 * then return zero. When we succeed in waiting for @p to be off its CPU,
1132 * we return a positive number (its total switch count). If a second call
1133 * a short while later returns the same number, the caller can be sure that
1134 * @p has remained unscheduled the whole time.
1135 *
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1141 */
1142unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1143{
1144 unsigned long flags;
1145 int running, on_rq;
1146 unsigned long ncsw;
1147 struct rq *rq;
1148
1149 for (;;) {
1150 /*
1151 * We do the initial early heuristics without holding
1152 * any task-queue locks at all. We'll only try to get
1153 * the runqueue lock when things look like they will
1154 * work out!
1155 */
1156 rq = task_rq(p);
1157
1158 /*
1159 * If the task is actively running on another CPU
1160 * still, just relax and busy-wait without holding
1161 * any locks.
1162 *
1163 * NOTE! Since we don't hold any locks, it's not
1164 * even sure that "rq" stays as the right runqueue!
1165 * But we don't care, since "task_running()" will
1166 * return false if the runqueue has changed and p
1167 * is actually now running somewhere else!
1168 */
1169 while (task_running(rq, p)) {
1170 if (match_state && unlikely(p->state != match_state))
1171 return 0;
1172 cpu_relax();
1173 }
1174
1175 /*
1176 * Ok, time to look more closely! We need the rq
1177 * lock now, to be *sure*. If we're wrong, we'll
1178 * just go back and repeat.
1179 */
1180 rq = task_rq_lock(p, &flags);
1181 trace_sched_wait_task(p);
1182 running = task_running(rq, p);
1183 on_rq = p->on_rq;
1184 ncsw = 0;
1185 if (!match_state || p->state == match_state)
1186 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1187 task_rq_unlock(rq, p, &flags);
1188
1189 /*
1190 * If it changed from the expected state, bail out now.
1191 */
1192 if (unlikely(!ncsw))
1193 break;
1194
1195 /*
1196 * Was it really running after all now that we
1197 * checked with the proper locks actually held?
1198 *
1199 * Oops. Go back and try again..
1200 */
1201 if (unlikely(running)) {
1202 cpu_relax();
1203 continue;
1204 }
1205
1206 /*
1207 * It's not enough that it's not actively running,
1208 * it must be off the runqueue _entirely_, and not
1209 * preempted!
1210 *
1211 * So if it was still runnable (but just not actively
1212 * running right now), it's preempted, and we should
1213 * yield - it could be a while.
1214 */
1215 if (unlikely(on_rq)) {
1216 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1217
1218 set_current_state(TASK_UNINTERRUPTIBLE);
1219 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1220 continue;
1221 }
1222
1223 /*
1224 * Ahh, all good. It wasn't running, and it wasn't
1225 * runnable, which means that it will never become
1226 * running in the future either. We're all done!
1227 */
1228 break;
1229 }
1230
1231 return ncsw;
1232}
1233
1234/***
1235 * kick_process - kick a running thread to enter/exit the kernel
1236 * @p: the to-be-kicked thread
1237 *
1238 * Cause a process which is running on another CPU to enter
1239 * kernel-mode, without any delay. (to get signals handled.)
1240 *
1241 * NOTE: this function doesn't have to take the runqueue lock,
1242 * because all it wants to ensure is that the remote task enters
1243 * the kernel. If the IPI races and the task has been migrated
1244 * to another CPU then no harm is done and the purpose has been
1245 * achieved as well.
1246 */
1247void kick_process(struct task_struct *p)
1248{
1249 int cpu;
1250
1251 preempt_disable();
1252 cpu = task_cpu(p);
1253 if ((cpu != smp_processor_id()) && task_curr(p))
1254 smp_send_reschedule(cpu);
1255 preempt_enable();
1256}
1257EXPORT_SYMBOL_GPL(kick_process);
1258#endif /* CONFIG_SMP */
1259
1260#ifdef CONFIG_SMP
1261/*
1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1263 */
1264static int select_fallback_rq(int cpu, struct task_struct *p)
1265{
1266 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1267 enum { cpuset, possible, fail } state = cpuset;
1268 int dest_cpu;
1269
1270 /* Look for allowed, online CPU in same node. */
1271 for_each_cpu(dest_cpu, nodemask) {
1272 if (!cpu_online(dest_cpu))
1273 continue;
1274 if (!cpu_active(dest_cpu))
1275 continue;
1276 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1277 return dest_cpu;
1278 }
1279
1280 for (;;) {
1281 /* Any allowed, online CPU? */
1282 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1283 if (!cpu_online(dest_cpu))
1284 continue;
1285 if (!cpu_active(dest_cpu))
1286 continue;
1287 goto out;
1288 }
1289
1290 switch (state) {
1291 case cpuset:
1292 /* No more Mr. Nice Guy. */
1293 cpuset_cpus_allowed_fallback(p);
1294 state = possible;
1295 break;
1296
1297 case possible:
1298 do_set_cpus_allowed(p, cpu_possible_mask);
1299 state = fail;
1300 break;
1301
1302 case fail:
1303 BUG();
1304 break;
1305 }
1306 }
1307
1308out:
1309 if (state != cpuset) {
1310 /*
1311 * Don't tell them about moving exiting tasks or
1312 * kernel threads (both mm NULL), since they never
1313 * leave kernel.
1314 */
1315 if (p->mm && printk_ratelimit()) {
1316 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1317 task_pid_nr(p), p->comm, cpu);
1318 }
1319 }
1320
1321 return dest_cpu;
1322}
1323
1324/*
1325 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1326 */
1327static inline
1328int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1329{
1330 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1331
1332 /*
1333 * In order not to call set_task_cpu() on a blocking task we need
1334 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1335 * cpu.
1336 *
1337 * Since this is common to all placement strategies, this lives here.
1338 *
1339 * [ this allows ->select_task() to simply return task_cpu(p) and
1340 * not worry about this generic constraint ]
1341 */
1342 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1343 !cpu_online(cpu)))
1344 cpu = select_fallback_rq(task_cpu(p), p);
1345
1346 return cpu;
1347}
1348
1349static void update_avg(u64 *avg, u64 sample)
1350{
1351 s64 diff = sample - *avg;
1352 *avg += diff >> 3;
1353}
1354#endif
1355
1356static void
1357ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1358{
1359#ifdef CONFIG_SCHEDSTATS
1360 struct rq *rq = this_rq();
1361
1362#ifdef CONFIG_SMP
1363 int this_cpu = smp_processor_id();
1364
1365 if (cpu == this_cpu) {
1366 schedstat_inc(rq, ttwu_local);
1367 schedstat_inc(p, se.statistics.nr_wakeups_local);
1368 } else {
1369 struct sched_domain *sd;
1370
1371 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1372 rcu_read_lock();
1373 for_each_domain(this_cpu, sd) {
1374 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1375 schedstat_inc(sd, ttwu_wake_remote);
1376 break;
1377 }
1378 }
1379 rcu_read_unlock();
1380 }
1381
1382 if (wake_flags & WF_MIGRATED)
1383 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1384
1385#endif /* CONFIG_SMP */
1386
1387 schedstat_inc(rq, ttwu_count);
1388 schedstat_inc(p, se.statistics.nr_wakeups);
1389
1390 if (wake_flags & WF_SYNC)
1391 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1392
1393#endif /* CONFIG_SCHEDSTATS */
1394}
1395
1396static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1397{
1398 activate_task(rq, p, en_flags);
1399 p->on_rq = 1;
1400
1401 /* if a worker is waking up, notify workqueue */
1402 if (p->flags & PF_WQ_WORKER)
1403 wq_worker_waking_up(p, cpu_of(rq));
1404}
1405
1406/*
1407 * Mark the task runnable and perform wakeup-preemption.
1408 */
1409static void
1410ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1411{
1412 trace_sched_wakeup(p, true);
1413 check_preempt_curr(rq, p, wake_flags);
1414
1415 p->state = TASK_RUNNING;
1416#ifdef CONFIG_SMP
1417 if (p->sched_class->task_woken)
1418 p->sched_class->task_woken(rq, p);
1419
1420 if (rq->idle_stamp) {
1421 u64 delta = rq->clock - rq->idle_stamp;
1422 u64 max = 2*sysctl_sched_migration_cost;
1423
1424 if (delta > max)
1425 rq->avg_idle = max;
1426 else
1427 update_avg(&rq->avg_idle, delta);
1428 rq->idle_stamp = 0;
1429 }
1430#endif
1431}
1432
1433static void
1434ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1435{
1436#ifdef CONFIG_SMP
1437 if (p->sched_contributes_to_load)
1438 rq->nr_uninterruptible--;
1439#endif
1440
1441 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1442 ttwu_do_wakeup(rq, p, wake_flags);
1443}
1444
1445/*
1446 * Called in case the task @p isn't fully descheduled from its runqueue,
1447 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1448 * since all we need to do is flip p->state to TASK_RUNNING, since
1449 * the task is still ->on_rq.
1450 */
1451static int ttwu_remote(struct task_struct *p, int wake_flags)
1452{
1453 struct rq *rq;
1454 int ret = 0;
1455
1456 rq = __task_rq_lock(p);
1457 if (p->on_rq) {
1458 ttwu_do_wakeup(rq, p, wake_flags);
1459 ret = 1;
1460 }
1461 __task_rq_unlock(rq);
1462
1463 return ret;
1464}
1465
1466#ifdef CONFIG_SMP
1467static void sched_ttwu_pending(void)
1468{
1469 struct rq *rq = this_rq();
1470 struct llist_node *llist = llist_del_all(&rq->wake_list);
1471 struct task_struct *p;
1472
1473 raw_spin_lock(&rq->lock);
1474
1475 while (llist) {
1476 p = llist_entry(llist, struct task_struct, wake_entry);
1477 llist = llist_next(llist);
1478 ttwu_do_activate(rq, p, 0);
1479 }
1480
1481 raw_spin_unlock(&rq->lock);
1482}
1483
1484void scheduler_ipi(void)
1485{
1486 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1487 return;
1488
1489 /*
1490 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1491 * traditionally all their work was done from the interrupt return
1492 * path. Now that we actually do some work, we need to make sure
1493 * we do call them.
1494 *
1495 * Some archs already do call them, luckily irq_enter/exit nest
1496 * properly.
1497 *
1498 * Arguably we should visit all archs and update all handlers,
1499 * however a fair share of IPIs are still resched only so this would
1500 * somewhat pessimize the simple resched case.
1501 */
1502 irq_enter();
1503 sched_ttwu_pending();
1504
1505 /*
1506 * Check if someone kicked us for doing the nohz idle load balance.
1507 */
1508 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1509 this_rq()->idle_balance = 1;
1510 raise_softirq_irqoff(SCHED_SOFTIRQ);
1511 }
1512 irq_exit();
1513}
1514
1515static void ttwu_queue_remote(struct task_struct *p, int cpu)
1516{
1517 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1518 smp_send_reschedule(cpu);
1519}
1520
1521#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1522static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1523{
1524 struct rq *rq;
1525 int ret = 0;
1526
1527 rq = __task_rq_lock(p);
1528 if (p->on_cpu) {
1529 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1530 ttwu_do_wakeup(rq, p, wake_flags);
1531 ret = 1;
1532 }
1533 __task_rq_unlock(rq);
1534
1535 return ret;
1536
1537}
1538#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1539
1540bool cpus_share_cache(int this_cpu, int that_cpu)
1541{
1542 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1543}
1544#endif /* CONFIG_SMP */
1545
1546static void ttwu_queue(struct task_struct *p, int cpu)
1547{
1548 struct rq *rq = cpu_rq(cpu);
1549
1550#if defined(CONFIG_SMP)
1551 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1552 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1553 ttwu_queue_remote(p, cpu);
1554 return;
1555 }
1556#endif
1557
1558 raw_spin_lock(&rq->lock);
1559 ttwu_do_activate(rq, p, 0);
1560 raw_spin_unlock(&rq->lock);
1561}
1562
1563/**
1564 * try_to_wake_up - wake up a thread
1565 * @p: the thread to be awakened
1566 * @state: the mask of task states that can be woken
1567 * @wake_flags: wake modifier flags (WF_*)
1568 *
1569 * Put it on the run-queue if it's not already there. The "current"
1570 * thread is always on the run-queue (except when the actual
1571 * re-schedule is in progress), and as such you're allowed to do
1572 * the simpler "current->state = TASK_RUNNING" to mark yourself
1573 * runnable without the overhead of this.
1574 *
1575 * Returns %true if @p was woken up, %false if it was already running
1576 * or @state didn't match @p's state.
1577 */
1578static int
1579try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1580{
1581 unsigned long flags;
1582 int cpu, success = 0;
1583
1584 smp_wmb();
1585 raw_spin_lock_irqsave(&p->pi_lock, flags);
1586 if (!(p->state & state))
1587 goto out;
1588
1589 success = 1; /* we're going to change ->state */
1590 cpu = task_cpu(p);
1591
1592 if (p->on_rq && ttwu_remote(p, wake_flags))
1593 goto stat;
1594
1595#ifdef CONFIG_SMP
1596 /*
1597 * If the owning (remote) cpu is still in the middle of schedule() with
1598 * this task as prev, wait until its done referencing the task.
1599 */
1600 while (p->on_cpu) {
1601#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1602 /*
1603 * In case the architecture enables interrupts in
1604 * context_switch(), we cannot busy wait, since that
1605 * would lead to deadlocks when an interrupt hits and
1606 * tries to wake up @prev. So bail and do a complete
1607 * remote wakeup.
1608 */
1609 if (ttwu_activate_remote(p, wake_flags))
1610 goto stat;
1611#else
1612 cpu_relax();
1613#endif
1614 }
1615 /*
1616 * Pairs with the smp_wmb() in finish_lock_switch().
1617 */
1618 smp_rmb();
1619
1620 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1621 p->state = TASK_WAKING;
1622
1623 if (p->sched_class->task_waking)
1624 p->sched_class->task_waking(p);
1625
1626 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1627 if (task_cpu(p) != cpu) {
1628 wake_flags |= WF_MIGRATED;
1629 set_task_cpu(p, cpu);
1630 }
1631#endif /* CONFIG_SMP */
1632
1633 ttwu_queue(p, cpu);
1634stat:
1635 ttwu_stat(p, cpu, wake_flags);
1636out:
1637 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1638
1639 return success;
1640}
1641
1642/**
1643 * try_to_wake_up_local - try to wake up a local task with rq lock held
1644 * @p: the thread to be awakened
1645 *
1646 * Put @p on the run-queue if it's not already there. The caller must
1647 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1648 * the current task.
1649 */
1650static void try_to_wake_up_local(struct task_struct *p)
1651{
1652 struct rq *rq = task_rq(p);
1653
1654 BUG_ON(rq != this_rq());
1655 BUG_ON(p == current);
1656 lockdep_assert_held(&rq->lock);
1657
1658 if (!raw_spin_trylock(&p->pi_lock)) {
1659 raw_spin_unlock(&rq->lock);
1660 raw_spin_lock(&p->pi_lock);
1661 raw_spin_lock(&rq->lock);
1662 }
1663
1664 if (!(p->state & TASK_NORMAL))
1665 goto out;
1666
1667 if (!p->on_rq)
1668 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1669
1670 ttwu_do_wakeup(rq, p, 0);
1671 ttwu_stat(p, smp_processor_id(), 0);
1672out:
1673 raw_spin_unlock(&p->pi_lock);
1674}
1675
1676/**
1677 * wake_up_process - Wake up a specific process
1678 * @p: The process to be woken up.
1679 *
1680 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * processes. Returns 1 if the process was woken up, 0 if it was already
1682 * running.
1683 *
1684 * It may be assumed that this function implies a write memory barrier before
1685 * changing the task state if and only if any tasks are woken up.
1686 */
1687int wake_up_process(struct task_struct *p)
1688{
1689 return try_to_wake_up(p, TASK_ALL, 0);
1690}
1691EXPORT_SYMBOL(wake_up_process);
1692
1693int wake_up_state(struct task_struct *p, unsigned int state)
1694{
1695 return try_to_wake_up(p, state, 0);
1696}
1697
1698/*
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1701 *
1702 * __sched_fork() is basic setup used by init_idle() too:
1703 */
1704static void __sched_fork(struct task_struct *p)
1705{
1706 p->on_rq = 0;
1707
1708 p->se.on_rq = 0;
1709 p->se.exec_start = 0;
1710 p->se.sum_exec_runtime = 0;
1711 p->se.prev_sum_exec_runtime = 0;
1712 p->se.nr_migrations = 0;
1713 p->se.vruntime = 0;
1714 INIT_LIST_HEAD(&p->se.group_node);
1715
1716#ifdef CONFIG_SCHEDSTATS
1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1718#endif
1719
1720 INIT_LIST_HEAD(&p->rt.run_list);
1721
1722#ifdef CONFIG_PREEMPT_NOTIFIERS
1723 INIT_HLIST_HEAD(&p->preempt_notifiers);
1724#endif
1725}
1726
1727/*
1728 * fork()/clone()-time setup:
1729 */
1730void sched_fork(struct task_struct *p)
1731{
1732 unsigned long flags;
1733 int cpu = get_cpu();
1734
1735 __sched_fork(p);
1736 /*
1737 * We mark the process as running here. This guarantees that
1738 * nobody will actually run it, and a signal or other external
1739 * event cannot wake it up and insert it on the runqueue either.
1740 */
1741 p->state = TASK_RUNNING;
1742
1743 /*
1744 * Make sure we do not leak PI boosting priority to the child.
1745 */
1746 p->prio = current->normal_prio;
1747
1748 /*
1749 * Revert to default priority/policy on fork if requested.
1750 */
1751 if (unlikely(p->sched_reset_on_fork)) {
1752 if (task_has_rt_policy(p)) {
1753 p->policy = SCHED_NORMAL;
1754 p->static_prio = NICE_TO_PRIO(0);
1755 p->rt_priority = 0;
1756 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1757 p->static_prio = NICE_TO_PRIO(0);
1758
1759 p->prio = p->normal_prio = __normal_prio(p);
1760 set_load_weight(p);
1761
1762 /*
1763 * We don't need the reset flag anymore after the fork. It has
1764 * fulfilled its duty:
1765 */
1766 p->sched_reset_on_fork = 0;
1767 }
1768
1769 if (!rt_prio(p->prio))
1770 p->sched_class = &fair_sched_class;
1771
1772 if (p->sched_class->task_fork)
1773 p->sched_class->task_fork(p);
1774
1775 /*
1776 * The child is not yet in the pid-hash so no cgroup attach races,
1777 * and the cgroup is pinned to this child due to cgroup_fork()
1778 * is ran before sched_fork().
1779 *
1780 * Silence PROVE_RCU.
1781 */
1782 raw_spin_lock_irqsave(&p->pi_lock, flags);
1783 set_task_cpu(p, cpu);
1784 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1785
1786#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1787 if (likely(sched_info_on()))
1788 memset(&p->sched_info, 0, sizeof(p->sched_info));
1789#endif
1790#if defined(CONFIG_SMP)
1791 p->on_cpu = 0;
1792#endif
1793#ifdef CONFIG_PREEMPT_COUNT
1794 /* Want to start with kernel preemption disabled. */
1795 task_thread_info(p)->preempt_count = 1;
1796#endif
1797#ifdef CONFIG_SMP
1798 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1799#endif
1800
1801 put_cpu();
1802}
1803
1804/*
1805 * wake_up_new_task - wake up a newly created task for the first time.
1806 *
1807 * This function will do some initial scheduler statistics housekeeping
1808 * that must be done for every newly created context, then puts the task
1809 * on the runqueue and wakes it.
1810 */
1811void wake_up_new_task(struct task_struct *p)
1812{
1813 unsigned long flags;
1814 struct rq *rq;
1815
1816 raw_spin_lock_irqsave(&p->pi_lock, flags);
1817#ifdef CONFIG_SMP
1818 /*
1819 * Fork balancing, do it here and not earlier because:
1820 * - cpus_allowed can change in the fork path
1821 * - any previously selected cpu might disappear through hotplug
1822 */
1823 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1824#endif
1825
1826 rq = __task_rq_lock(p);
1827 activate_task(rq, p, 0);
1828 p->on_rq = 1;
1829 trace_sched_wakeup_new(p, true);
1830 check_preempt_curr(rq, p, WF_FORK);
1831#ifdef CONFIG_SMP
1832 if (p->sched_class->task_woken)
1833 p->sched_class->task_woken(rq, p);
1834#endif
1835 task_rq_unlock(rq, p, &flags);
1836}
1837
1838#ifdef CONFIG_PREEMPT_NOTIFIERS
1839
1840/**
1841 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1842 * @notifier: notifier struct to register
1843 */
1844void preempt_notifier_register(struct preempt_notifier *notifier)
1845{
1846 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1847}
1848EXPORT_SYMBOL_GPL(preempt_notifier_register);
1849
1850/**
1851 * preempt_notifier_unregister - no longer interested in preemption notifications
1852 * @notifier: notifier struct to unregister
1853 *
1854 * This is safe to call from within a preemption notifier.
1855 */
1856void preempt_notifier_unregister(struct preempt_notifier *notifier)
1857{
1858 hlist_del(¬ifier->link);
1859}
1860EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1861
1862static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1863{
1864 struct preempt_notifier *notifier;
1865 struct hlist_node *node;
1866
1867 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1868 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1869}
1870
1871static void
1872fire_sched_out_preempt_notifiers(struct task_struct *curr,
1873 struct task_struct *next)
1874{
1875 struct preempt_notifier *notifier;
1876 struct hlist_node *node;
1877
1878 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1879 notifier->ops->sched_out(notifier, next);
1880}
1881
1882#else /* !CONFIG_PREEMPT_NOTIFIERS */
1883
1884static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1885{
1886}
1887
1888static void
1889fire_sched_out_preempt_notifiers(struct task_struct *curr,
1890 struct task_struct *next)
1891{
1892}
1893
1894#endif /* CONFIG_PREEMPT_NOTIFIERS */
1895
1896/**
1897 * prepare_task_switch - prepare to switch tasks
1898 * @rq: the runqueue preparing to switch
1899 * @prev: the current task that is being switched out
1900 * @next: the task we are going to switch to.
1901 *
1902 * This is called with the rq lock held and interrupts off. It must
1903 * be paired with a subsequent finish_task_switch after the context
1904 * switch.
1905 *
1906 * prepare_task_switch sets up locking and calls architecture specific
1907 * hooks.
1908 */
1909static inline void
1910prepare_task_switch(struct rq *rq, struct task_struct *prev,
1911 struct task_struct *next)
1912{
1913 sched_info_switch(prev, next);
1914 perf_event_task_sched_out(prev, next);
1915 fire_sched_out_preempt_notifiers(prev, next);
1916 prepare_lock_switch(rq, next);
1917 prepare_arch_switch(next);
1918 trace_sched_switch(prev, next);
1919}
1920
1921/**
1922 * finish_task_switch - clean up after a task-switch
1923 * @rq: runqueue associated with task-switch
1924 * @prev: the thread we just switched away from.
1925 *
1926 * finish_task_switch must be called after the context switch, paired
1927 * with a prepare_task_switch call before the context switch.
1928 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1929 * and do any other architecture-specific cleanup actions.
1930 *
1931 * Note that we may have delayed dropping an mm in context_switch(). If
1932 * so, we finish that here outside of the runqueue lock. (Doing it
1933 * with the lock held can cause deadlocks; see schedule() for
1934 * details.)
1935 */
1936static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1937 __releases(rq->lock)
1938{
1939 struct mm_struct *mm = rq->prev_mm;
1940 long prev_state;
1941
1942 rq->prev_mm = NULL;
1943
1944 /*
1945 * A task struct has one reference for the use as "current".
1946 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1947 * schedule one last time. The schedule call will never return, and
1948 * the scheduled task must drop that reference.
1949 * The test for TASK_DEAD must occur while the runqueue locks are
1950 * still held, otherwise prev could be scheduled on another cpu, die
1951 * there before we look at prev->state, and then the reference would
1952 * be dropped twice.
1953 * Manfred Spraul <manfred@colorfullife.com>
1954 */
1955 prev_state = prev->state;
1956 finish_arch_switch(prev);
1957#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1958 local_irq_disable();
1959#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1960 perf_event_task_sched_in(prev, current);
1961#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1962 local_irq_enable();
1963#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1964 finish_lock_switch(rq, prev);
1965 finish_arch_post_lock_switch();
1966
1967 fire_sched_in_preempt_notifiers(current);
1968 if (mm)
1969 mmdrop(mm);
1970 if (unlikely(prev_state == TASK_DEAD)) {
1971 /*
1972 * Remove function-return probe instances associated with this
1973 * task and put them back on the free list.
1974 */
1975 kprobe_flush_task(prev);
1976 put_task_struct(prev);
1977 }
1978}
1979
1980#ifdef CONFIG_SMP
1981
1982/* assumes rq->lock is held */
1983static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1984{
1985 if (prev->sched_class->pre_schedule)
1986 prev->sched_class->pre_schedule(rq, prev);
1987}
1988
1989/* rq->lock is NOT held, but preemption is disabled */
1990static inline void post_schedule(struct rq *rq)
1991{
1992 if (rq->post_schedule) {
1993 unsigned long flags;
1994
1995 raw_spin_lock_irqsave(&rq->lock, flags);
1996 if (rq->curr->sched_class->post_schedule)
1997 rq->curr->sched_class->post_schedule(rq);
1998 raw_spin_unlock_irqrestore(&rq->lock, flags);
1999
2000 rq->post_schedule = 0;
2001 }
2002}
2003
2004#else
2005
2006static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2007{
2008}
2009
2010static inline void post_schedule(struct rq *rq)
2011{
2012}
2013
2014#endif
2015
2016/**
2017 * schedule_tail - first thing a freshly forked thread must call.
2018 * @prev: the thread we just switched away from.
2019 */
2020asmlinkage void schedule_tail(struct task_struct *prev)
2021 __releases(rq->lock)
2022{
2023 struct rq *rq = this_rq();
2024
2025 finish_task_switch(rq, prev);
2026
2027 /*
2028 * FIXME: do we need to worry about rq being invalidated by the
2029 * task_switch?
2030 */
2031 post_schedule(rq);
2032
2033#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2034 /* In this case, finish_task_switch does not reenable preemption */
2035 preempt_enable();
2036#endif
2037 if (current->set_child_tid)
2038 put_user(task_pid_vnr(current), current->set_child_tid);
2039}
2040
2041/*
2042 * context_switch - switch to the new MM and the new
2043 * thread's register state.
2044 */
2045static inline void
2046context_switch(struct rq *rq, struct task_struct *prev,
2047 struct task_struct *next)
2048{
2049 struct mm_struct *mm, *oldmm;
2050
2051 prepare_task_switch(rq, prev, next);
2052
2053 mm = next->mm;
2054 oldmm = prev->active_mm;
2055 /*
2056 * For paravirt, this is coupled with an exit in switch_to to
2057 * combine the page table reload and the switch backend into
2058 * one hypercall.
2059 */
2060 arch_start_context_switch(prev);
2061
2062 if (!mm) {
2063 next->active_mm = oldmm;
2064 atomic_inc(&oldmm->mm_count);
2065 enter_lazy_tlb(oldmm, next);
2066 } else
2067 switch_mm(oldmm, mm, next);
2068
2069 if (!prev->mm) {
2070 prev->active_mm = NULL;
2071 rq->prev_mm = oldmm;
2072 }
2073 /*
2074 * Since the runqueue lock will be released by the next
2075 * task (which is an invalid locking op but in the case
2076 * of the scheduler it's an obvious special-case), so we
2077 * do an early lockdep release here:
2078 */
2079#ifndef __ARCH_WANT_UNLOCKED_CTXSW
2080 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2081#endif
2082
2083 /* Here we just switch the register state and the stack. */
2084 switch_to(prev, next, prev);
2085
2086 barrier();
2087 /*
2088 * this_rq must be evaluated again because prev may have moved
2089 * CPUs since it called schedule(), thus the 'rq' on its stack
2090 * frame will be invalid.
2091 */
2092 finish_task_switch(this_rq(), prev);
2093}
2094
2095/*
2096 * nr_running, nr_uninterruptible and nr_context_switches:
2097 *
2098 * externally visible scheduler statistics: current number of runnable
2099 * threads, current number of uninterruptible-sleeping threads, total
2100 * number of context switches performed since bootup.
2101 */
2102unsigned long nr_running(void)
2103{
2104 unsigned long i, sum = 0;
2105
2106 for_each_online_cpu(i)
2107 sum += cpu_rq(i)->nr_running;
2108
2109 return sum;
2110}
2111
2112unsigned long nr_uninterruptible(void)
2113{
2114 unsigned long i, sum = 0;
2115
2116 for_each_possible_cpu(i)
2117 sum += cpu_rq(i)->nr_uninterruptible;
2118
2119 /*
2120 * Since we read the counters lockless, it might be slightly
2121 * inaccurate. Do not allow it to go below zero though:
2122 */
2123 if (unlikely((long)sum < 0))
2124 sum = 0;
2125
2126 return sum;
2127}
2128
2129unsigned long long nr_context_switches(void)
2130{
2131 int i;
2132 unsigned long long sum = 0;
2133
2134 for_each_possible_cpu(i)
2135 sum += cpu_rq(i)->nr_switches;
2136
2137 return sum;
2138}
2139
2140unsigned long nr_iowait(void)
2141{
2142 unsigned long i, sum = 0;
2143
2144 for_each_possible_cpu(i)
2145 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2146
2147 return sum;
2148}
2149
2150unsigned long nr_iowait_cpu(int cpu)
2151{
2152 struct rq *this = cpu_rq(cpu);
2153 return atomic_read(&this->nr_iowait);
2154}
2155
2156unsigned long this_cpu_load(void)
2157{
2158 struct rq *this = this_rq();
2159 return this->cpu_load[0];
2160}
2161
2162
2163/*
2164 * Global load-average calculations
2165 *
2166 * We take a distributed and async approach to calculating the global load-avg
2167 * in order to minimize overhead.
2168 *
2169 * The global load average is an exponentially decaying average of nr_running +
2170 * nr_uninterruptible.
2171 *
2172 * Once every LOAD_FREQ:
2173 *
2174 * nr_active = 0;
2175 * for_each_possible_cpu(cpu)
2176 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2177 *
2178 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2179 *
2180 * Due to a number of reasons the above turns in the mess below:
2181 *
2182 * - for_each_possible_cpu() is prohibitively expensive on machines with
2183 * serious number of cpus, therefore we need to take a distributed approach
2184 * to calculating nr_active.
2185 *
2186 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2187 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2188 *
2189 * So assuming nr_active := 0 when we start out -- true per definition, we
2190 * can simply take per-cpu deltas and fold those into a global accumulate
2191 * to obtain the same result. See calc_load_fold_active().
2192 *
2193 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2194 * across the machine, we assume 10 ticks is sufficient time for every
2195 * cpu to have completed this task.
2196 *
2197 * This places an upper-bound on the IRQ-off latency of the machine. Then
2198 * again, being late doesn't loose the delta, just wrecks the sample.
2199 *
2200 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2201 * this would add another cross-cpu cacheline miss and atomic operation
2202 * to the wakeup path. Instead we increment on whatever cpu the task ran
2203 * when it went into uninterruptible state and decrement on whatever cpu
2204 * did the wakeup. This means that only the sum of nr_uninterruptible over
2205 * all cpus yields the correct result.
2206 *
2207 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2208 */
2209
2210/* Variables and functions for calc_load */
2211static atomic_long_t calc_load_tasks;
2212static unsigned long calc_load_update;
2213unsigned long avenrun[3];
2214EXPORT_SYMBOL(avenrun); /* should be removed */
2215
2216/**
2217 * get_avenrun - get the load average array
2218 * @loads: pointer to dest load array
2219 * @offset: offset to add
2220 * @shift: shift count to shift the result left
2221 *
2222 * These values are estimates at best, so no need for locking.
2223 */
2224void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2225{
2226 loads[0] = (avenrun[0] + offset) << shift;
2227 loads[1] = (avenrun[1] + offset) << shift;
2228 loads[2] = (avenrun[2] + offset) << shift;
2229}
2230
2231static long calc_load_fold_active(struct rq *this_rq)
2232{
2233 long nr_active, delta = 0;
2234
2235 nr_active = this_rq->nr_running;
2236 nr_active += (long) this_rq->nr_uninterruptible;
2237
2238 if (nr_active != this_rq->calc_load_active) {
2239 delta = nr_active - this_rq->calc_load_active;
2240 this_rq->calc_load_active = nr_active;
2241 }
2242
2243 return delta;
2244}
2245
2246/*
2247 * a1 = a0 * e + a * (1 - e)
2248 */
2249static unsigned long
2250calc_load(unsigned long load, unsigned long exp, unsigned long active)
2251{
2252 load *= exp;
2253 load += active * (FIXED_1 - exp);
2254 load += 1UL << (FSHIFT - 1);
2255 return load >> FSHIFT;
2256}
2257
2258#ifdef CONFIG_NO_HZ
2259/*
2260 * Handle NO_HZ for the global load-average.
2261 *
2262 * Since the above described distributed algorithm to compute the global
2263 * load-average relies on per-cpu sampling from the tick, it is affected by
2264 * NO_HZ.
2265 *
2266 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2267 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2268 * when we read the global state.
2269 *
2270 * Obviously reality has to ruin such a delightfully simple scheme:
2271 *
2272 * - When we go NO_HZ idle during the window, we can negate our sample
2273 * contribution, causing under-accounting.
2274 *
2275 * We avoid this by keeping two idle-delta counters and flipping them
2276 * when the window starts, thus separating old and new NO_HZ load.
2277 *
2278 * The only trick is the slight shift in index flip for read vs write.
2279 *
2280 * 0s 5s 10s 15s
2281 * +10 +10 +10 +10
2282 * |-|-----------|-|-----------|-|-----------|-|
2283 * r:0 0 1 1 0 0 1 1 0
2284 * w:0 1 1 0 0 1 1 0 0
2285 *
2286 * This ensures we'll fold the old idle contribution in this window while
2287 * accumlating the new one.
2288 *
2289 * - When we wake up from NO_HZ idle during the window, we push up our
2290 * contribution, since we effectively move our sample point to a known
2291 * busy state.
2292 *
2293 * This is solved by pushing the window forward, and thus skipping the
2294 * sample, for this cpu (effectively using the idle-delta for this cpu which
2295 * was in effect at the time the window opened). This also solves the issue
2296 * of having to deal with a cpu having been in NOHZ idle for multiple
2297 * LOAD_FREQ intervals.
2298 *
2299 * When making the ILB scale, we should try to pull this in as well.
2300 */
2301static atomic_long_t calc_load_idle[2];
2302static int calc_load_idx;
2303
2304static inline int calc_load_write_idx(void)
2305{
2306 int idx = calc_load_idx;
2307
2308 /*
2309 * See calc_global_nohz(), if we observe the new index, we also
2310 * need to observe the new update time.
2311 */
2312 smp_rmb();
2313
2314 /*
2315 * If the folding window started, make sure we start writing in the
2316 * next idle-delta.
2317 */
2318 if (!time_before(jiffies, calc_load_update))
2319 idx++;
2320
2321 return idx & 1;
2322}
2323
2324static inline int calc_load_read_idx(void)
2325{
2326 return calc_load_idx & 1;
2327}
2328
2329void calc_load_enter_idle(void)
2330{
2331 struct rq *this_rq = this_rq();
2332 long delta;
2333
2334 /*
2335 * We're going into NOHZ mode, if there's any pending delta, fold it
2336 * into the pending idle delta.
2337 */
2338 delta = calc_load_fold_active(this_rq);
2339 if (delta) {
2340 int idx = calc_load_write_idx();
2341 atomic_long_add(delta, &calc_load_idle[idx]);
2342 }
2343}
2344
2345void calc_load_exit_idle(void)
2346{
2347 struct rq *this_rq = this_rq();
2348
2349 /*
2350 * If we're still before the sample window, we're done.
2351 */
2352 if (time_before(jiffies, this_rq->calc_load_update))
2353 return;
2354
2355 /*
2356 * We woke inside or after the sample window, this means we're already
2357 * accounted through the nohz accounting, so skip the entire deal and
2358 * sync up for the next window.
2359 */
2360 this_rq->calc_load_update = calc_load_update;
2361 if (time_before(jiffies, this_rq->calc_load_update + 10))
2362 this_rq->calc_load_update += LOAD_FREQ;
2363}
2364
2365static long calc_load_fold_idle(void)
2366{
2367 int idx = calc_load_read_idx();
2368 long delta = 0;
2369
2370 if (atomic_long_read(&calc_load_idle[idx]))
2371 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2372
2373 return delta;
2374}
2375
2376/**
2377 * fixed_power_int - compute: x^n, in O(log n) time
2378 *
2379 * @x: base of the power
2380 * @frac_bits: fractional bits of @x
2381 * @n: power to raise @x to.
2382 *
2383 * By exploiting the relation between the definition of the natural power
2384 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2385 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2386 * (where: n_i \elem {0, 1}, the binary vector representing n),
2387 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2388 * of course trivially computable in O(log_2 n), the length of our binary
2389 * vector.
2390 */
2391static unsigned long
2392fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2393{
2394 unsigned long result = 1UL << frac_bits;
2395
2396 if (n) for (;;) {
2397 if (n & 1) {
2398 result *= x;
2399 result += 1UL << (frac_bits - 1);
2400 result >>= frac_bits;
2401 }
2402 n >>= 1;
2403 if (!n)
2404 break;
2405 x *= x;
2406 x += 1UL << (frac_bits - 1);
2407 x >>= frac_bits;
2408 }
2409
2410 return result;
2411}
2412
2413/*
2414 * a1 = a0 * e + a * (1 - e)
2415 *
2416 * a2 = a1 * e + a * (1 - e)
2417 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2418 * = a0 * e^2 + a * (1 - e) * (1 + e)
2419 *
2420 * a3 = a2 * e + a * (1 - e)
2421 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2422 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2423 *
2424 * ...
2425 *
2426 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2427 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2428 * = a0 * e^n + a * (1 - e^n)
2429 *
2430 * [1] application of the geometric series:
2431 *
2432 * n 1 - x^(n+1)
2433 * S_n := \Sum x^i = -------------
2434 * i=0 1 - x
2435 */
2436static unsigned long
2437calc_load_n(unsigned long load, unsigned long exp,
2438 unsigned long active, unsigned int n)
2439{
2440
2441 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2442}
2443
2444/*
2445 * NO_HZ can leave us missing all per-cpu ticks calling
2446 * calc_load_account_active(), but since an idle CPU folds its delta into
2447 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2448 * in the pending idle delta if our idle period crossed a load cycle boundary.
2449 *
2450 * Once we've updated the global active value, we need to apply the exponential
2451 * weights adjusted to the number of cycles missed.
2452 */
2453static void calc_global_nohz(void)
2454{
2455 long delta, active, n;
2456
2457 if (!time_before(jiffies, calc_load_update + 10)) {
2458 /*
2459 * Catch-up, fold however many we are behind still
2460 */
2461 delta = jiffies - calc_load_update - 10;
2462 n = 1 + (delta / LOAD_FREQ);
2463
2464 active = atomic_long_read(&calc_load_tasks);
2465 active = active > 0 ? active * FIXED_1 : 0;
2466
2467 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2468 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2469 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2470
2471 calc_load_update += n * LOAD_FREQ;
2472 }
2473
2474 /*
2475 * Flip the idle index...
2476 *
2477 * Make sure we first write the new time then flip the index, so that
2478 * calc_load_write_idx() will see the new time when it reads the new
2479 * index, this avoids a double flip messing things up.
2480 */
2481 smp_wmb();
2482 calc_load_idx++;
2483}
2484#else /* !CONFIG_NO_HZ */
2485
2486static inline long calc_load_fold_idle(void) { return 0; }
2487static inline void calc_global_nohz(void) { }
2488
2489#endif /* CONFIG_NO_HZ */
2490
2491/*
2492 * calc_load - update the avenrun load estimates 10 ticks after the
2493 * CPUs have updated calc_load_tasks.
2494 */
2495void calc_global_load(unsigned long ticks)
2496{
2497 long active, delta;
2498
2499 if (time_before(jiffies, calc_load_update + 10))
2500 return;
2501
2502 /*
2503 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2504 */
2505 delta = calc_load_fold_idle();
2506 if (delta)
2507 atomic_long_add(delta, &calc_load_tasks);
2508
2509 active = atomic_long_read(&calc_load_tasks);
2510 active = active > 0 ? active * FIXED_1 : 0;
2511
2512 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2513 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2514 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2515
2516 calc_load_update += LOAD_FREQ;
2517
2518 /*
2519 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2520 */
2521 calc_global_nohz();
2522}
2523
2524/*
2525 * Called from update_cpu_load() to periodically update this CPU's
2526 * active count.
2527 */
2528static void calc_load_account_active(struct rq *this_rq)
2529{
2530 long delta;
2531
2532 if (time_before(jiffies, this_rq->calc_load_update))
2533 return;
2534
2535 delta = calc_load_fold_active(this_rq);
2536 if (delta)
2537 atomic_long_add(delta, &calc_load_tasks);
2538
2539 this_rq->calc_load_update += LOAD_FREQ;
2540}
2541
2542/*
2543 * End of global load-average stuff
2544 */
2545
2546/*
2547 * The exact cpuload at various idx values, calculated at every tick would be
2548 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2549 *
2550 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2551 * on nth tick when cpu may be busy, then we have:
2552 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2553 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2554 *
2555 * decay_load_missed() below does efficient calculation of
2556 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2557 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2558 *
2559 * The calculation is approximated on a 128 point scale.
2560 * degrade_zero_ticks is the number of ticks after which load at any
2561 * particular idx is approximated to be zero.
2562 * degrade_factor is a precomputed table, a row for each load idx.
2563 * Each column corresponds to degradation factor for a power of two ticks,
2564 * based on 128 point scale.
2565 * Example:
2566 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2567 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2568 *
2569 * With this power of 2 load factors, we can degrade the load n times
2570 * by looking at 1 bits in n and doing as many mult/shift instead of
2571 * n mult/shifts needed by the exact degradation.
2572 */
2573#define DEGRADE_SHIFT 7
2574static const unsigned char
2575 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2576static const unsigned char
2577 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2578 {0, 0, 0, 0, 0, 0, 0, 0},
2579 {64, 32, 8, 0, 0, 0, 0, 0},
2580 {96, 72, 40, 12, 1, 0, 0},
2581 {112, 98, 75, 43, 15, 1, 0},
2582 {120, 112, 98, 76, 45, 16, 2} };
2583
2584/*
2585 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2586 * would be when CPU is idle and so we just decay the old load without
2587 * adding any new load.
2588 */
2589static unsigned long
2590decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2591{
2592 int j = 0;
2593
2594 if (!missed_updates)
2595 return load;
2596
2597 if (missed_updates >= degrade_zero_ticks[idx])
2598 return 0;
2599
2600 if (idx == 1)
2601 return load >> missed_updates;
2602
2603 while (missed_updates) {
2604 if (missed_updates % 2)
2605 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2606
2607 missed_updates >>= 1;
2608 j++;
2609 }
2610 return load;
2611}
2612
2613/*
2614 * Update rq->cpu_load[] statistics. This function is usually called every
2615 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2616 * every tick. We fix it up based on jiffies.
2617 */
2618static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2619 unsigned long pending_updates)
2620{
2621 int i, scale;
2622
2623 this_rq->nr_load_updates++;
2624
2625 /* Update our load: */
2626 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2627 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2628 unsigned long old_load, new_load;
2629
2630 /* scale is effectively 1 << i now, and >> i divides by scale */
2631
2632 old_load = this_rq->cpu_load[i];
2633 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2634 new_load = this_load;
2635 /*
2636 * Round up the averaging division if load is increasing. This
2637 * prevents us from getting stuck on 9 if the load is 10, for
2638 * example.
2639 */
2640 if (new_load > old_load)
2641 new_load += scale - 1;
2642
2643 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2644 }
2645
2646 sched_avg_update(this_rq);
2647}
2648
2649#ifdef CONFIG_NO_HZ
2650/*
2651 * There is no sane way to deal with nohz on smp when using jiffies because the
2652 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2653 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2654 *
2655 * Therefore we cannot use the delta approach from the regular tick since that
2656 * would seriously skew the load calculation. However we'll make do for those
2657 * updates happening while idle (nohz_idle_balance) or coming out of idle
2658 * (tick_nohz_idle_exit).
2659 *
2660 * This means we might still be one tick off for nohz periods.
2661 */
2662
2663/*
2664 * Called from nohz_idle_balance() to update the load ratings before doing the
2665 * idle balance.
2666 */
2667void update_idle_cpu_load(struct rq *this_rq)
2668{
2669 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2670 unsigned long load = this_rq->load.weight;
2671 unsigned long pending_updates;
2672
2673 /*
2674 * bail if there's load or we're actually up-to-date.
2675 */
2676 if (load || curr_jiffies == this_rq->last_load_update_tick)
2677 return;
2678
2679 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2680 this_rq->last_load_update_tick = curr_jiffies;
2681
2682 __update_cpu_load(this_rq, load, pending_updates);
2683}
2684
2685/*
2686 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2687 */
2688void update_cpu_load_nohz(void)
2689{
2690 struct rq *this_rq = this_rq();
2691 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2692 unsigned long pending_updates;
2693
2694 if (curr_jiffies == this_rq->last_load_update_tick)
2695 return;
2696
2697 raw_spin_lock(&this_rq->lock);
2698 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2699 if (pending_updates) {
2700 this_rq->last_load_update_tick = curr_jiffies;
2701 /*
2702 * We were idle, this means load 0, the current load might be
2703 * !0 due to remote wakeups and the sort.
2704 */
2705 __update_cpu_load(this_rq, 0, pending_updates);
2706 }
2707 raw_spin_unlock(&this_rq->lock);
2708}
2709#endif /* CONFIG_NO_HZ */
2710
2711/*
2712 * Called from scheduler_tick()
2713 */
2714static void update_cpu_load_active(struct rq *this_rq)
2715{
2716 /*
2717 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2718 */
2719 this_rq->last_load_update_tick = jiffies;
2720 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2721
2722 calc_load_account_active(this_rq);
2723}
2724
2725#ifdef CONFIG_SMP
2726
2727/*
2728 * sched_exec - execve() is a valuable balancing opportunity, because at
2729 * this point the task has the smallest effective memory and cache footprint.
2730 */
2731void sched_exec(void)
2732{
2733 struct task_struct *p = current;
2734 unsigned long flags;
2735 int dest_cpu;
2736
2737 raw_spin_lock_irqsave(&p->pi_lock, flags);
2738 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2739 if (dest_cpu == smp_processor_id())
2740 goto unlock;
2741
2742 if (likely(cpu_active(dest_cpu))) {
2743 struct migration_arg arg = { p, dest_cpu };
2744
2745 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2746 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2747 return;
2748 }
2749unlock:
2750 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2751}
2752
2753#endif
2754
2755DEFINE_PER_CPU(struct kernel_stat, kstat);
2756DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2757
2758EXPORT_PER_CPU_SYMBOL(kstat);
2759EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2760
2761/*
2762 * Return any ns on the sched_clock that have not yet been accounted in
2763 * @p in case that task is currently running.
2764 *
2765 * Called with task_rq_lock() held on @rq.
2766 */
2767static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2768{
2769 u64 ns = 0;
2770
2771 if (task_current(rq, p)) {
2772 update_rq_clock(rq);
2773 ns = rq->clock_task - p->se.exec_start;
2774 if ((s64)ns < 0)
2775 ns = 0;
2776 }
2777
2778 return ns;
2779}
2780
2781unsigned long long task_delta_exec(struct task_struct *p)
2782{
2783 unsigned long flags;
2784 struct rq *rq;
2785 u64 ns = 0;
2786
2787 rq = task_rq_lock(p, &flags);
2788 ns = do_task_delta_exec(p, rq);
2789 task_rq_unlock(rq, p, &flags);
2790
2791 return ns;
2792}
2793
2794/*
2795 * Return accounted runtime for the task.
2796 * In case the task is currently running, return the runtime plus current's
2797 * pending runtime that have not been accounted yet.
2798 */
2799unsigned long long task_sched_runtime(struct task_struct *p)
2800{
2801 unsigned long flags;
2802 struct rq *rq;
2803 u64 ns = 0;
2804
2805 rq = task_rq_lock(p, &flags);
2806 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2807 task_rq_unlock(rq, p, &flags);
2808
2809 return ns;
2810}
2811
2812#ifdef CONFIG_CGROUP_CPUACCT
2813struct cgroup_subsys cpuacct_subsys;
2814struct cpuacct root_cpuacct;
2815#endif
2816
2817static inline void task_group_account_field(struct task_struct *p, int index,
2818 u64 tmp)
2819{
2820#ifdef CONFIG_CGROUP_CPUACCT
2821 struct kernel_cpustat *kcpustat;
2822 struct cpuacct *ca;
2823#endif
2824 /*
2825 * Since all updates are sure to touch the root cgroup, we
2826 * get ourselves ahead and touch it first. If the root cgroup
2827 * is the only cgroup, then nothing else should be necessary.
2828 *
2829 */
2830 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2831
2832#ifdef CONFIG_CGROUP_CPUACCT
2833 if (unlikely(!cpuacct_subsys.active))
2834 return;
2835
2836 rcu_read_lock();
2837 ca = task_ca(p);
2838 while (ca && (ca != &root_cpuacct)) {
2839 kcpustat = this_cpu_ptr(ca->cpustat);
2840 kcpustat->cpustat[index] += tmp;
2841 ca = parent_ca(ca);
2842 }
2843 rcu_read_unlock();
2844#endif
2845}
2846
2847
2848/*
2849 * Account user cpu time to a process.
2850 * @p: the process that the cpu time gets accounted to
2851 * @cputime: the cpu time spent in user space since the last update
2852 * @cputime_scaled: cputime scaled by cpu frequency
2853 */
2854void account_user_time(struct task_struct *p, cputime_t cputime,
2855 cputime_t cputime_scaled)
2856{
2857 int index;
2858
2859 /* Add user time to process. */
2860 p->utime += cputime;
2861 p->utimescaled += cputime_scaled;
2862 account_group_user_time(p, cputime);
2863
2864 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2865
2866 /* Add user time to cpustat. */
2867 task_group_account_field(p, index, (__force u64) cputime);
2868
2869 /* Account for user time used */
2870 acct_update_integrals(p);
2871}
2872
2873/*
2874 * Account guest cpu time to a process.
2875 * @p: the process that the cpu time gets accounted to
2876 * @cputime: the cpu time spent in virtual machine since the last update
2877 * @cputime_scaled: cputime scaled by cpu frequency
2878 */
2879static void account_guest_time(struct task_struct *p, cputime_t cputime,
2880 cputime_t cputime_scaled)
2881{
2882 u64 *cpustat = kcpustat_this_cpu->cpustat;
2883
2884 /* Add guest time to process. */
2885 p->utime += cputime;
2886 p->utimescaled += cputime_scaled;
2887 account_group_user_time(p, cputime);
2888 p->gtime += cputime;
2889
2890 /* Add guest time to cpustat. */
2891 if (TASK_NICE(p) > 0) {
2892 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2893 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2894 } else {
2895 cpustat[CPUTIME_USER] += (__force u64) cputime;
2896 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2897 }
2898}
2899
2900/*
2901 * Account system cpu time to a process and desired cpustat field
2902 * @p: the process that the cpu time gets accounted to
2903 * @cputime: the cpu time spent in kernel space since the last update
2904 * @cputime_scaled: cputime scaled by cpu frequency
2905 * @target_cputime64: pointer to cpustat field that has to be updated
2906 */
2907static inline
2908void __account_system_time(struct task_struct *p, cputime_t cputime,
2909 cputime_t cputime_scaled, int index)
2910{
2911 /* Add system time to process. */
2912 p->stime += cputime;
2913 p->stimescaled += cputime_scaled;
2914 account_group_system_time(p, cputime);
2915
2916 /* Add system time to cpustat. */
2917 task_group_account_field(p, index, (__force u64) cputime);
2918
2919 /* Account for system time used */
2920 acct_update_integrals(p);
2921}
2922
2923/*
2924 * Account system cpu time to a process.
2925 * @p: the process that the cpu time gets accounted to
2926 * @hardirq_offset: the offset to subtract from hardirq_count()
2927 * @cputime: the cpu time spent in kernel space since the last update
2928 * @cputime_scaled: cputime scaled by cpu frequency
2929 */
2930void account_system_time(struct task_struct *p, int hardirq_offset,
2931 cputime_t cputime, cputime_t cputime_scaled)
2932{
2933 int index;
2934
2935 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2936 account_guest_time(p, cputime, cputime_scaled);
2937 return;
2938 }
2939
2940 if (hardirq_count() - hardirq_offset)
2941 index = CPUTIME_IRQ;
2942 else if (in_serving_softirq())
2943 index = CPUTIME_SOFTIRQ;
2944 else
2945 index = CPUTIME_SYSTEM;
2946
2947 __account_system_time(p, cputime, cputime_scaled, index);
2948}
2949
2950/*
2951 * Account for involuntary wait time.
2952 * @cputime: the cpu time spent in involuntary wait
2953 */
2954void account_steal_time(cputime_t cputime)
2955{
2956 u64 *cpustat = kcpustat_this_cpu->cpustat;
2957
2958 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2959}
2960
2961/*
2962 * Account for idle time.
2963 * @cputime: the cpu time spent in idle wait
2964 */
2965void account_idle_time(cputime_t cputime)
2966{
2967 u64 *cpustat = kcpustat_this_cpu->cpustat;
2968 struct rq *rq = this_rq();
2969
2970 if (atomic_read(&rq->nr_iowait) > 0)
2971 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2972 else
2973 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2974}
2975
2976static __always_inline bool steal_account_process_tick(void)
2977{
2978#ifdef CONFIG_PARAVIRT
2979 if (static_key_false(¶virt_steal_enabled)) {
2980 u64 steal, st = 0;
2981
2982 steal = paravirt_steal_clock(smp_processor_id());
2983 steal -= this_rq()->prev_steal_time;
2984
2985 st = steal_ticks(steal);
2986 this_rq()->prev_steal_time += st * TICK_NSEC;
2987
2988 account_steal_time(st);
2989 return st;
2990 }
2991#endif
2992 return false;
2993}
2994
2995#ifndef CONFIG_VIRT_CPU_ACCOUNTING
2996
2997#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2998/*
2999 * Account a tick to a process and cpustat
3000 * @p: the process that the cpu time gets accounted to
3001 * @user_tick: is the tick from userspace
3002 * @rq: the pointer to rq
3003 *
3004 * Tick demultiplexing follows the order
3005 * - pending hardirq update
3006 * - pending softirq update
3007 * - user_time
3008 * - idle_time
3009 * - system time
3010 * - check for guest_time
3011 * - else account as system_time
3012 *
3013 * Check for hardirq is done both for system and user time as there is
3014 * no timer going off while we are on hardirq and hence we may never get an
3015 * opportunity to update it solely in system time.
3016 * p->stime and friends are only updated on system time and not on irq
3017 * softirq as those do not count in task exec_runtime any more.
3018 */
3019static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3020 struct rq *rq)
3021{
3022 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3023 u64 *cpustat = kcpustat_this_cpu->cpustat;
3024
3025 if (steal_account_process_tick())
3026 return;
3027
3028 if (irqtime_account_hi_update()) {
3029 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
3030 } else if (irqtime_account_si_update()) {
3031 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
3032 } else if (this_cpu_ksoftirqd() == p) {
3033 /*
3034 * ksoftirqd time do not get accounted in cpu_softirq_time.
3035 * So, we have to handle it separately here.
3036 * Also, p->stime needs to be updated for ksoftirqd.
3037 */
3038 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3039 CPUTIME_SOFTIRQ);
3040 } else if (user_tick) {
3041 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3042 } else if (p == rq->idle) {
3043 account_idle_time(cputime_one_jiffy);
3044 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3045 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3046 } else {
3047 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3048 CPUTIME_SYSTEM);
3049 }
3050}
3051
3052static void irqtime_account_idle_ticks(int ticks)
3053{
3054 int i;
3055 struct rq *rq = this_rq();
3056
3057 for (i = 0; i < ticks; i++)
3058 irqtime_account_process_tick(current, 0, rq);
3059}
3060#else /* CONFIG_IRQ_TIME_ACCOUNTING */
3061static void irqtime_account_idle_ticks(int ticks) {}
3062static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3063 struct rq *rq) {}
3064#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3065
3066/*
3067 * Account a single tick of cpu time.
3068 * @p: the process that the cpu time gets accounted to
3069 * @user_tick: indicates if the tick is a user or a system tick
3070 */
3071void account_process_tick(struct task_struct *p, int user_tick)
3072{
3073 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3074 struct rq *rq = this_rq();
3075
3076 if (sched_clock_irqtime) {
3077 irqtime_account_process_tick(p, user_tick, rq);
3078 return;
3079 }
3080
3081 if (steal_account_process_tick())
3082 return;
3083
3084 if (user_tick)
3085 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3086 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3087 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3088 one_jiffy_scaled);
3089 else
3090 account_idle_time(cputime_one_jiffy);
3091}
3092
3093/*
3094 * Account multiple ticks of steal time.
3095 * @p: the process from which the cpu time has been stolen
3096 * @ticks: number of stolen ticks
3097 */
3098void account_steal_ticks(unsigned long ticks)
3099{
3100 account_steal_time(jiffies_to_cputime(ticks));
3101}
3102
3103/*
3104 * Account multiple ticks of idle time.
3105 * @ticks: number of stolen ticks
3106 */
3107void account_idle_ticks(unsigned long ticks)
3108{
3109
3110 if (sched_clock_irqtime) {
3111 irqtime_account_idle_ticks(ticks);
3112 return;
3113 }
3114
3115 account_idle_time(jiffies_to_cputime(ticks));
3116}
3117
3118#endif
3119
3120/*
3121 * Use precise platform statistics if available:
3122 */
3123#ifdef CONFIG_VIRT_CPU_ACCOUNTING
3124void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3125{
3126 *ut = p->utime;
3127 *st = p->stime;
3128}
3129
3130void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3131{
3132 struct task_cputime cputime;
3133
3134 thread_group_cputime(p, &cputime);
3135
3136 *ut = cputime.utime;
3137 *st = cputime.stime;
3138}
3139#else
3140
3141#ifndef nsecs_to_cputime
3142# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3143#endif
3144
3145static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
3146{
3147 u64 temp = (__force u64) rtime;
3148
3149 temp *= (__force u64) utime;
3150
3151 if (sizeof(cputime_t) == 4)
3152 temp = div_u64(temp, (__force u32) total);
3153 else
3154 temp = div64_u64(temp, (__force u64) total);
3155
3156 return (__force cputime_t) temp;
3157}
3158
3159void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3160{
3161 cputime_t rtime, utime = p->utime, total = utime + p->stime;
3162
3163 /*
3164 * Use CFS's precise accounting:
3165 */
3166 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3167
3168 if (total)
3169 utime = scale_utime(utime, rtime, total);
3170 else
3171 utime = rtime;
3172
3173 /*
3174 * Compare with previous values, to keep monotonicity:
3175 */
3176 p->prev_utime = max(p->prev_utime, utime);
3177 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
3178
3179 *ut = p->prev_utime;
3180 *st = p->prev_stime;
3181}
3182
3183/*
3184 * Must be called with siglock held.
3185 */
3186void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3187{
3188 struct signal_struct *sig = p->signal;
3189 struct task_cputime cputime;
3190 cputime_t rtime, utime, total;
3191
3192 thread_group_cputime(p, &cputime);
3193
3194 total = cputime.utime + cputime.stime;
3195 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3196
3197 if (total)
3198 utime = scale_utime(cputime.utime, rtime, total);
3199 else
3200 utime = rtime;
3201
3202 sig->prev_utime = max(sig->prev_utime, utime);
3203 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3204
3205 *ut = sig->prev_utime;
3206 *st = sig->prev_stime;
3207}
3208#endif
3209
3210/*
3211 * This function gets called by the timer code, with HZ frequency.
3212 * We call it with interrupts disabled.
3213 */
3214void scheduler_tick(void)
3215{
3216 int cpu = smp_processor_id();
3217 struct rq *rq = cpu_rq(cpu);
3218 struct task_struct *curr = rq->curr;
3219
3220 sched_clock_tick();
3221
3222 raw_spin_lock(&rq->lock);
3223 update_rq_clock(rq);
3224 update_cpu_load_active(rq);
3225 curr->sched_class->task_tick(rq, curr, 0);
3226 raw_spin_unlock(&rq->lock);
3227
3228 perf_event_task_tick();
3229
3230#ifdef CONFIG_SMP
3231 rq->idle_balance = idle_cpu(cpu);
3232 trigger_load_balance(rq, cpu);
3233#endif
3234}
3235
3236notrace unsigned long get_parent_ip(unsigned long addr)
3237{
3238 if (in_lock_functions(addr)) {
3239 addr = CALLER_ADDR2;
3240 if (in_lock_functions(addr))
3241 addr = CALLER_ADDR3;
3242 }
3243 return addr;
3244}
3245
3246#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3247 defined(CONFIG_PREEMPT_TRACER))
3248
3249void __kprobes add_preempt_count(int val)
3250{
3251#ifdef CONFIG_DEBUG_PREEMPT
3252 /*
3253 * Underflow?
3254 */
3255 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3256 return;
3257#endif
3258 preempt_count() += val;
3259#ifdef CONFIG_DEBUG_PREEMPT
3260 /*
3261 * Spinlock count overflowing soon?
3262 */
3263 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3264 PREEMPT_MASK - 10);
3265#endif
3266 if (preempt_count() == val)
3267 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3268}
3269EXPORT_SYMBOL(add_preempt_count);
3270
3271void __kprobes sub_preempt_count(int val)
3272{
3273#ifdef CONFIG_DEBUG_PREEMPT
3274 /*
3275 * Underflow?
3276 */
3277 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3278 return;
3279 /*
3280 * Is the spinlock portion underflowing?
3281 */
3282 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3283 !(preempt_count() & PREEMPT_MASK)))
3284 return;
3285#endif
3286
3287 if (preempt_count() == val)
3288 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3289 preempt_count() -= val;
3290}
3291EXPORT_SYMBOL(sub_preempt_count);
3292
3293#endif
3294
3295/*
3296 * Print scheduling while atomic bug:
3297 */
3298static noinline void __schedule_bug(struct task_struct *prev)
3299{
3300 if (oops_in_progress)
3301 return;
3302
3303 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3304 prev->comm, prev->pid, preempt_count());
3305
3306 debug_show_held_locks(prev);
3307 print_modules();
3308 if (irqs_disabled())
3309 print_irqtrace_events(prev);
3310 dump_stack();
3311 add_taint(TAINT_WARN);
3312}
3313
3314/*
3315 * Various schedule()-time debugging checks and statistics:
3316 */
3317static inline void schedule_debug(struct task_struct *prev)
3318{
3319 /*
3320 * Test if we are atomic. Since do_exit() needs to call into
3321 * schedule() atomically, we ignore that path for now.
3322 * Otherwise, whine if we are scheduling when we should not be.
3323 */
3324 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3325 __schedule_bug(prev);
3326 rcu_sleep_check();
3327
3328 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3329
3330 schedstat_inc(this_rq(), sched_count);
3331}
3332
3333static void put_prev_task(struct rq *rq, struct task_struct *prev)
3334{
3335 if (prev->on_rq || rq->skip_clock_update < 0)
3336 update_rq_clock(rq);
3337 prev->sched_class->put_prev_task(rq, prev);
3338}
3339
3340/*
3341 * Pick up the highest-prio task:
3342 */
3343static inline struct task_struct *
3344pick_next_task(struct rq *rq)
3345{
3346 const struct sched_class *class;
3347 struct task_struct *p;
3348
3349 /*
3350 * Optimization: we know that if all tasks are in
3351 * the fair class we can call that function directly:
3352 */
3353 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3354 p = fair_sched_class.pick_next_task(rq);
3355 if (likely(p))
3356 return p;
3357 }
3358
3359 for_each_class(class) {
3360 p = class->pick_next_task(rq);
3361 if (p)
3362 return p;
3363 }
3364
3365 BUG(); /* the idle class will always have a runnable task */
3366}
3367
3368/*
3369 * __schedule() is the main scheduler function.
3370 */
3371static void __sched __schedule(void)
3372{
3373 struct task_struct *prev, *next;
3374 unsigned long *switch_count;
3375 struct rq *rq;
3376 int cpu;
3377
3378need_resched:
3379 preempt_disable();
3380 cpu = smp_processor_id();
3381 rq = cpu_rq(cpu);
3382 rcu_note_context_switch(cpu);
3383 prev = rq->curr;
3384
3385 schedule_debug(prev);
3386
3387 if (sched_feat(HRTICK))
3388 hrtick_clear(rq);
3389
3390 raw_spin_lock_irq(&rq->lock);
3391
3392 switch_count = &prev->nivcsw;
3393 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3394 if (unlikely(signal_pending_state(prev->state, prev))) {
3395 prev->state = TASK_RUNNING;
3396 } else {
3397 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3398 prev->on_rq = 0;
3399
3400 /*
3401 * If a worker went to sleep, notify and ask workqueue
3402 * whether it wants to wake up a task to maintain
3403 * concurrency.
3404 */
3405 if (prev->flags & PF_WQ_WORKER) {
3406 struct task_struct *to_wakeup;
3407
3408 to_wakeup = wq_worker_sleeping(prev, cpu);
3409 if (to_wakeup)
3410 try_to_wake_up_local(to_wakeup);
3411 }
3412 }
3413 switch_count = &prev->nvcsw;
3414 }
3415
3416 pre_schedule(rq, prev);
3417
3418 if (unlikely(!rq->nr_running))
3419 idle_balance(cpu, rq);
3420
3421 put_prev_task(rq, prev);
3422 next = pick_next_task(rq);
3423 clear_tsk_need_resched(prev);
3424 rq->skip_clock_update = 0;
3425
3426 if (likely(prev != next)) {
3427 rq->nr_switches++;
3428 rq->curr = next;
3429 ++*switch_count;
3430
3431 context_switch(rq, prev, next); /* unlocks the rq */
3432 /*
3433 * The context switch have flipped the stack from under us
3434 * and restored the local variables which were saved when
3435 * this task called schedule() in the past. prev == current
3436 * is still correct, but it can be moved to another cpu/rq.
3437 */
3438 cpu = smp_processor_id();
3439 rq = cpu_rq(cpu);
3440 } else
3441 raw_spin_unlock_irq(&rq->lock);
3442
3443 post_schedule(rq);
3444
3445 sched_preempt_enable_no_resched();
3446 if (need_resched())
3447 goto need_resched;
3448}
3449
3450static inline void sched_submit_work(struct task_struct *tsk)
3451{
3452 if (!tsk->state || tsk_is_pi_blocked(tsk))
3453 return;
3454 /*
3455 * If we are going to sleep and we have plugged IO queued,
3456 * make sure to submit it to avoid deadlocks.
3457 */
3458 if (blk_needs_flush_plug(tsk))
3459 blk_schedule_flush_plug(tsk);
3460}
3461
3462asmlinkage void __sched schedule(void)
3463{
3464 struct task_struct *tsk = current;
3465
3466 sched_submit_work(tsk);
3467 __schedule();
3468}
3469EXPORT_SYMBOL(schedule);
3470
3471/**
3472 * schedule_preempt_disabled - called with preemption disabled
3473 *
3474 * Returns with preemption disabled. Note: preempt_count must be 1
3475 */
3476void __sched schedule_preempt_disabled(void)
3477{
3478 sched_preempt_enable_no_resched();
3479 schedule();
3480 preempt_disable();
3481}
3482
3483#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3484
3485static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3486{
3487 if (lock->owner != owner)
3488 return false;
3489
3490 /*
3491 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3492 * lock->owner still matches owner, if that fails, owner might
3493 * point to free()d memory, if it still matches, the rcu_read_lock()
3494 * ensures the memory stays valid.
3495 */
3496 barrier();
3497
3498 return owner->on_cpu;
3499}
3500
3501/*
3502 * Look out! "owner" is an entirely speculative pointer
3503 * access and not reliable.
3504 */
3505int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3506{
3507 if (!sched_feat(OWNER_SPIN))
3508 return 0;
3509
3510 rcu_read_lock();
3511 while (owner_running(lock, owner)) {
3512 if (need_resched())
3513 break;
3514
3515 arch_mutex_cpu_relax();
3516 }
3517 rcu_read_unlock();
3518
3519 /*
3520 * We break out the loop above on need_resched() and when the
3521 * owner changed, which is a sign for heavy contention. Return
3522 * success only when lock->owner is NULL.
3523 */
3524 return lock->owner == NULL;
3525}
3526#endif
3527
3528#ifdef CONFIG_PREEMPT
3529/*
3530 * this is the entry point to schedule() from in-kernel preemption
3531 * off of preempt_enable. Kernel preemptions off return from interrupt
3532 * occur there and call schedule directly.
3533 */
3534asmlinkage void __sched notrace preempt_schedule(void)
3535{
3536 struct thread_info *ti = current_thread_info();
3537
3538 /*
3539 * If there is a non-zero preempt_count or interrupts are disabled,
3540 * we do not want to preempt the current task. Just return..
3541 */
3542 if (likely(ti->preempt_count || irqs_disabled()))
3543 return;
3544
3545 do {
3546 add_preempt_count_notrace(PREEMPT_ACTIVE);
3547 __schedule();
3548 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3549
3550 /*
3551 * Check again in case we missed a preemption opportunity
3552 * between schedule and now.
3553 */
3554 barrier();
3555 } while (need_resched());
3556}
3557EXPORT_SYMBOL(preempt_schedule);
3558
3559/*
3560 * this is the entry point to schedule() from kernel preemption
3561 * off of irq context.
3562 * Note, that this is called and return with irqs disabled. This will
3563 * protect us against recursive calling from irq.
3564 */
3565asmlinkage void __sched preempt_schedule_irq(void)
3566{
3567 struct thread_info *ti = current_thread_info();
3568
3569 /* Catch callers which need to be fixed */
3570 BUG_ON(ti->preempt_count || !irqs_disabled());
3571
3572 do {
3573 add_preempt_count(PREEMPT_ACTIVE);
3574 local_irq_enable();
3575 __schedule();
3576 local_irq_disable();
3577 sub_preempt_count(PREEMPT_ACTIVE);
3578
3579 /*
3580 * Check again in case we missed a preemption opportunity
3581 * between schedule and now.
3582 */
3583 barrier();
3584 } while (need_resched());
3585}
3586
3587#endif /* CONFIG_PREEMPT */
3588
3589int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3590 void *key)
3591{
3592 return try_to_wake_up(curr->private, mode, wake_flags);
3593}
3594EXPORT_SYMBOL(default_wake_function);
3595
3596/*
3597 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3598 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3599 * number) then we wake all the non-exclusive tasks and one exclusive task.
3600 *
3601 * There are circumstances in which we can try to wake a task which has already
3602 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3603 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3604 */
3605static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3606 int nr_exclusive, int wake_flags, void *key)
3607{
3608 wait_queue_t *curr, *next;
3609
3610 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3611 unsigned flags = curr->flags;
3612
3613 if (curr->func(curr, mode, wake_flags, key) &&
3614 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3615 break;
3616 }
3617}
3618
3619/**
3620 * __wake_up - wake up threads blocked on a waitqueue.
3621 * @q: the waitqueue
3622 * @mode: which threads
3623 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3624 * @key: is directly passed to the wakeup function
3625 *
3626 * It may be assumed that this function implies a write memory barrier before
3627 * changing the task state if and only if any tasks are woken up.
3628 */
3629void __wake_up(wait_queue_head_t *q, unsigned int mode,
3630 int nr_exclusive, void *key)
3631{
3632 unsigned long flags;
3633
3634 spin_lock_irqsave(&q->lock, flags);
3635 __wake_up_common(q, mode, nr_exclusive, 0, key);
3636 spin_unlock_irqrestore(&q->lock, flags);
3637}
3638EXPORT_SYMBOL(__wake_up);
3639
3640/*
3641 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3642 */
3643void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3644{
3645 __wake_up_common(q, mode, nr, 0, NULL);
3646}
3647EXPORT_SYMBOL_GPL(__wake_up_locked);
3648
3649void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3650{
3651 __wake_up_common(q, mode, 1, 0, key);
3652}
3653EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3654
3655/**
3656 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3657 * @q: the waitqueue
3658 * @mode: which threads
3659 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3660 * @key: opaque value to be passed to wakeup targets
3661 *
3662 * The sync wakeup differs that the waker knows that it will schedule
3663 * away soon, so while the target thread will be woken up, it will not
3664 * be migrated to another CPU - ie. the two threads are 'synchronized'
3665 * with each other. This can prevent needless bouncing between CPUs.
3666 *
3667 * On UP it can prevent extra preemption.
3668 *
3669 * It may be assumed that this function implies a write memory barrier before
3670 * changing the task state if and only if any tasks are woken up.
3671 */
3672void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3673 int nr_exclusive, void *key)
3674{
3675 unsigned long flags;
3676 int wake_flags = WF_SYNC;
3677
3678 if (unlikely(!q))
3679 return;
3680
3681 if (unlikely(!nr_exclusive))
3682 wake_flags = 0;
3683
3684 spin_lock_irqsave(&q->lock, flags);
3685 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3686 spin_unlock_irqrestore(&q->lock, flags);
3687}
3688EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3689
3690/*
3691 * __wake_up_sync - see __wake_up_sync_key()
3692 */
3693void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3694{
3695 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3696}
3697EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3698
3699/**
3700 * complete: - signals a single thread waiting on this completion
3701 * @x: holds the state of this particular completion
3702 *
3703 * This will wake up a single thread waiting on this completion. Threads will be
3704 * awakened in the same order in which they were queued.
3705 *
3706 * See also complete_all(), wait_for_completion() and related routines.
3707 *
3708 * It may be assumed that this function implies a write memory barrier before
3709 * changing the task state if and only if any tasks are woken up.
3710 */
3711void complete(struct completion *x)
3712{
3713 unsigned long flags;
3714
3715 spin_lock_irqsave(&x->wait.lock, flags);
3716 x->done++;
3717 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3718 spin_unlock_irqrestore(&x->wait.lock, flags);
3719}
3720EXPORT_SYMBOL(complete);
3721
3722/**
3723 * complete_all: - signals all threads waiting on this completion
3724 * @x: holds the state of this particular completion
3725 *
3726 * This will wake up all threads waiting on this particular completion event.
3727 *
3728 * It may be assumed that this function implies a write memory barrier before
3729 * changing the task state if and only if any tasks are woken up.
3730 */
3731void complete_all(struct completion *x)
3732{
3733 unsigned long flags;
3734
3735 spin_lock_irqsave(&x->wait.lock, flags);
3736 x->done += UINT_MAX/2;
3737 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3738 spin_unlock_irqrestore(&x->wait.lock, flags);
3739}
3740EXPORT_SYMBOL(complete_all);
3741
3742static inline long __sched
3743do_wait_for_common(struct completion *x, long timeout, int state)
3744{
3745 if (!x->done) {
3746 DECLARE_WAITQUEUE(wait, current);
3747
3748 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3749 do {
3750 if (signal_pending_state(state, current)) {
3751 timeout = -ERESTARTSYS;
3752 break;
3753 }
3754 __set_current_state(state);
3755 spin_unlock_irq(&x->wait.lock);
3756 timeout = schedule_timeout(timeout);
3757 spin_lock_irq(&x->wait.lock);
3758 } while (!x->done && timeout);
3759 __remove_wait_queue(&x->wait, &wait);
3760 if (!x->done)
3761 return timeout;
3762 }
3763 x->done--;
3764 return timeout ?: 1;
3765}
3766
3767static long __sched
3768wait_for_common(struct completion *x, long timeout, int state)
3769{
3770 might_sleep();
3771
3772 spin_lock_irq(&x->wait.lock);
3773 timeout = do_wait_for_common(x, timeout, state);
3774 spin_unlock_irq(&x->wait.lock);
3775 return timeout;
3776}
3777
3778/**
3779 * wait_for_completion: - waits for completion of a task
3780 * @x: holds the state of this particular completion
3781 *
3782 * This waits to be signaled for completion of a specific task. It is NOT
3783 * interruptible and there is no timeout.
3784 *
3785 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3786 * and interrupt capability. Also see complete().
3787 */
3788void __sched wait_for_completion(struct completion *x)
3789{
3790 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3791}
3792EXPORT_SYMBOL(wait_for_completion);
3793
3794/**
3795 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3796 * @x: holds the state of this particular completion
3797 * @timeout: timeout value in jiffies
3798 *
3799 * This waits for either a completion of a specific task to be signaled or for a
3800 * specified timeout to expire. The timeout is in jiffies. It is not
3801 * interruptible.
3802 *
3803 * The return value is 0 if timed out, and positive (at least 1, or number of
3804 * jiffies left till timeout) if completed.
3805 */
3806unsigned long __sched
3807wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3808{
3809 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3810}
3811EXPORT_SYMBOL(wait_for_completion_timeout);
3812
3813/**
3814 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3815 * @x: holds the state of this particular completion
3816 *
3817 * This waits for completion of a specific task to be signaled. It is
3818 * interruptible.
3819 *
3820 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3821 */
3822int __sched wait_for_completion_interruptible(struct completion *x)
3823{
3824 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3825 if (t == -ERESTARTSYS)
3826 return t;
3827 return 0;
3828}
3829EXPORT_SYMBOL(wait_for_completion_interruptible);
3830
3831/**
3832 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3833 * @x: holds the state of this particular completion
3834 * @timeout: timeout value in jiffies
3835 *
3836 * This waits for either a completion of a specific task to be signaled or for a
3837 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3838 *
3839 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3840 * positive (at least 1, or number of jiffies left till timeout) if completed.
3841 */
3842long __sched
3843wait_for_completion_interruptible_timeout(struct completion *x,
3844 unsigned long timeout)
3845{
3846 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3847}
3848EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3849
3850/**
3851 * wait_for_completion_killable: - waits for completion of a task (killable)
3852 * @x: holds the state of this particular completion
3853 *
3854 * This waits to be signaled for completion of a specific task. It can be
3855 * interrupted by a kill signal.
3856 *
3857 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3858 */
3859int __sched wait_for_completion_killable(struct completion *x)
3860{
3861 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3862 if (t == -ERESTARTSYS)
3863 return t;
3864 return 0;
3865}
3866EXPORT_SYMBOL(wait_for_completion_killable);
3867
3868/**
3869 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3870 * @x: holds the state of this particular completion
3871 * @timeout: timeout value in jiffies
3872 *
3873 * This waits for either a completion of a specific task to be
3874 * signaled or for a specified timeout to expire. It can be
3875 * interrupted by a kill signal. The timeout is in jiffies.
3876 *
3877 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3878 * positive (at least 1, or number of jiffies left till timeout) if completed.
3879 */
3880long __sched
3881wait_for_completion_killable_timeout(struct completion *x,
3882 unsigned long timeout)
3883{
3884 return wait_for_common(x, timeout, TASK_KILLABLE);
3885}
3886EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3887
3888/**
3889 * try_wait_for_completion - try to decrement a completion without blocking
3890 * @x: completion structure
3891 *
3892 * Returns: 0 if a decrement cannot be done without blocking
3893 * 1 if a decrement succeeded.
3894 *
3895 * If a completion is being used as a counting completion,
3896 * attempt to decrement the counter without blocking. This
3897 * enables us to avoid waiting if the resource the completion
3898 * is protecting is not available.
3899 */
3900bool try_wait_for_completion(struct completion *x)
3901{
3902 unsigned long flags;
3903 int ret = 1;
3904
3905 spin_lock_irqsave(&x->wait.lock, flags);
3906 if (!x->done)
3907 ret = 0;
3908 else
3909 x->done--;
3910 spin_unlock_irqrestore(&x->wait.lock, flags);
3911 return ret;
3912}
3913EXPORT_SYMBOL(try_wait_for_completion);
3914
3915/**
3916 * completion_done - Test to see if a completion has any waiters
3917 * @x: completion structure
3918 *
3919 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3920 * 1 if there are no waiters.
3921 *
3922 */
3923bool completion_done(struct completion *x)
3924{
3925 unsigned long flags;
3926 int ret = 1;
3927
3928 spin_lock_irqsave(&x->wait.lock, flags);
3929 if (!x->done)
3930 ret = 0;
3931 spin_unlock_irqrestore(&x->wait.lock, flags);
3932 return ret;
3933}
3934EXPORT_SYMBOL(completion_done);
3935
3936static long __sched
3937sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3938{
3939 unsigned long flags;
3940 wait_queue_t wait;
3941
3942 init_waitqueue_entry(&wait, current);
3943
3944 __set_current_state(state);
3945
3946 spin_lock_irqsave(&q->lock, flags);
3947 __add_wait_queue(q, &wait);
3948 spin_unlock(&q->lock);
3949 timeout = schedule_timeout(timeout);
3950 spin_lock_irq(&q->lock);
3951 __remove_wait_queue(q, &wait);
3952 spin_unlock_irqrestore(&q->lock, flags);
3953
3954 return timeout;
3955}
3956
3957void __sched interruptible_sleep_on(wait_queue_head_t *q)
3958{
3959 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3960}
3961EXPORT_SYMBOL(interruptible_sleep_on);
3962
3963long __sched
3964interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3965{
3966 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3967}
3968EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3969
3970void __sched sleep_on(wait_queue_head_t *q)
3971{
3972 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3973}
3974EXPORT_SYMBOL(sleep_on);
3975
3976long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3977{
3978 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3979}
3980EXPORT_SYMBOL(sleep_on_timeout);
3981
3982#ifdef CONFIG_RT_MUTEXES
3983
3984/*
3985 * rt_mutex_setprio - set the current priority of a task
3986 * @p: task
3987 * @prio: prio value (kernel-internal form)
3988 *
3989 * This function changes the 'effective' priority of a task. It does
3990 * not touch ->normal_prio like __setscheduler().
3991 *
3992 * Used by the rt_mutex code to implement priority inheritance logic.
3993 */
3994void rt_mutex_setprio(struct task_struct *p, int prio)
3995{
3996 int oldprio, on_rq, running;
3997 struct rq *rq;
3998 const struct sched_class *prev_class;
3999
4000 BUG_ON(prio < 0 || prio > MAX_PRIO);
4001
4002 rq = __task_rq_lock(p);
4003
4004 /*
4005 * Idle task boosting is a nono in general. There is one
4006 * exception, when PREEMPT_RT and NOHZ is active:
4007 *
4008 * The idle task calls get_next_timer_interrupt() and holds
4009 * the timer wheel base->lock on the CPU and another CPU wants
4010 * to access the timer (probably to cancel it). We can safely
4011 * ignore the boosting request, as the idle CPU runs this code
4012 * with interrupts disabled and will complete the lock
4013 * protected section without being interrupted. So there is no
4014 * real need to boost.
4015 */
4016 if (unlikely(p == rq->idle)) {
4017 WARN_ON(p != rq->curr);
4018 WARN_ON(p->pi_blocked_on);
4019 goto out_unlock;
4020 }
4021
4022 trace_sched_pi_setprio(p, prio);
4023 oldprio = p->prio;
4024 prev_class = p->sched_class;
4025 on_rq = p->on_rq;
4026 running = task_current(rq, p);
4027 if (on_rq)
4028 dequeue_task(rq, p, 0);
4029 if (running)
4030 p->sched_class->put_prev_task(rq, p);
4031
4032 if (rt_prio(prio))
4033 p->sched_class = &rt_sched_class;
4034 else
4035 p->sched_class = &fair_sched_class;
4036
4037 p->prio = prio;
4038
4039 if (running)
4040 p->sched_class->set_curr_task(rq);
4041 if (on_rq)
4042 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4043
4044 check_class_changed(rq, p, prev_class, oldprio);
4045out_unlock:
4046 __task_rq_unlock(rq);
4047}
4048#endif
4049void set_user_nice(struct task_struct *p, long nice)
4050{
4051 int old_prio, delta, on_rq;
4052 unsigned long flags;
4053 struct rq *rq;
4054
4055 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4056 return;
4057 /*
4058 * We have to be careful, if called from sys_setpriority(),
4059 * the task might be in the middle of scheduling on another CPU.
4060 */
4061 rq = task_rq_lock(p, &flags);
4062 /*
4063 * The RT priorities are set via sched_setscheduler(), but we still
4064 * allow the 'normal' nice value to be set - but as expected
4065 * it wont have any effect on scheduling until the task is
4066 * SCHED_FIFO/SCHED_RR:
4067 */
4068 if (task_has_rt_policy(p)) {
4069 p->static_prio = NICE_TO_PRIO(nice);
4070 goto out_unlock;
4071 }
4072 on_rq = p->on_rq;
4073 if (on_rq)
4074 dequeue_task(rq, p, 0);
4075
4076 p->static_prio = NICE_TO_PRIO(nice);
4077 set_load_weight(p);
4078 old_prio = p->prio;
4079 p->prio = effective_prio(p);
4080 delta = p->prio - old_prio;
4081
4082 if (on_rq) {
4083 enqueue_task(rq, p, 0);
4084 /*
4085 * If the task increased its priority or is running and
4086 * lowered its priority, then reschedule its CPU:
4087 */
4088 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4089 resched_task(rq->curr);
4090 }
4091out_unlock:
4092 task_rq_unlock(rq, p, &flags);
4093}
4094EXPORT_SYMBOL(set_user_nice);
4095
4096/*
4097 * can_nice - check if a task can reduce its nice value
4098 * @p: task
4099 * @nice: nice value
4100 */
4101int can_nice(const struct task_struct *p, const int nice)
4102{
4103 /* convert nice value [19,-20] to rlimit style value [1,40] */
4104 int nice_rlim = 20 - nice;
4105
4106 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4107 capable(CAP_SYS_NICE));
4108}
4109
4110#ifdef __ARCH_WANT_SYS_NICE
4111
4112/*
4113 * sys_nice - change the priority of the current process.
4114 * @increment: priority increment
4115 *
4116 * sys_setpriority is a more generic, but much slower function that
4117 * does similar things.
4118 */
4119SYSCALL_DEFINE1(nice, int, increment)
4120{
4121 long nice, retval;
4122
4123 /*
4124 * Setpriority might change our priority at the same moment.
4125 * We don't have to worry. Conceptually one call occurs first
4126 * and we have a single winner.
4127 */
4128 if (increment < -40)
4129 increment = -40;
4130 if (increment > 40)
4131 increment = 40;
4132
4133 nice = TASK_NICE(current) + increment;
4134 if (nice < -20)
4135 nice = -20;
4136 if (nice > 19)
4137 nice = 19;
4138
4139 if (increment < 0 && !can_nice(current, nice))
4140 return -EPERM;
4141
4142 retval = security_task_setnice(current, nice);
4143 if (retval)
4144 return retval;
4145
4146 set_user_nice(current, nice);
4147 return 0;
4148}
4149
4150#endif
4151
4152/**
4153 * task_prio - return the priority value of a given task.
4154 * @p: the task in question.
4155 *
4156 * This is the priority value as seen by users in /proc.
4157 * RT tasks are offset by -200. Normal tasks are centered
4158 * around 0, value goes from -16 to +15.
4159 */
4160int task_prio(const struct task_struct *p)
4161{
4162 return p->prio - MAX_RT_PRIO;
4163}
4164
4165/**
4166 * task_nice - return the nice value of a given task.
4167 * @p: the task in question.
4168 */
4169int task_nice(const struct task_struct *p)
4170{
4171 return TASK_NICE(p);
4172}
4173EXPORT_SYMBOL(task_nice);
4174
4175/**
4176 * idle_cpu - is a given cpu idle currently?
4177 * @cpu: the processor in question.
4178 */
4179int idle_cpu(int cpu)
4180{
4181 struct rq *rq = cpu_rq(cpu);
4182
4183 if (rq->curr != rq->idle)
4184 return 0;
4185
4186 if (rq->nr_running)
4187 return 0;
4188
4189#ifdef CONFIG_SMP
4190 if (!llist_empty(&rq->wake_list))
4191 return 0;
4192#endif
4193
4194 return 1;
4195}
4196
4197/**
4198 * idle_task - return the idle task for a given cpu.
4199 * @cpu: the processor in question.
4200 */
4201struct task_struct *idle_task(int cpu)
4202{
4203 return cpu_rq(cpu)->idle;
4204}
4205
4206/**
4207 * find_process_by_pid - find a process with a matching PID value.
4208 * @pid: the pid in question.
4209 */
4210static struct task_struct *find_process_by_pid(pid_t pid)
4211{
4212 return pid ? find_task_by_vpid(pid) : current;
4213}
4214
4215/* Actually do priority change: must hold rq lock. */
4216static void
4217__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4218{
4219 p->policy = policy;
4220 p->rt_priority = prio;
4221 p->normal_prio = normal_prio(p);
4222 /* we are holding p->pi_lock already */
4223 p->prio = rt_mutex_getprio(p);
4224 if (rt_prio(p->prio))
4225 p->sched_class = &rt_sched_class;
4226 else
4227 p->sched_class = &fair_sched_class;
4228 set_load_weight(p);
4229}
4230
4231/*
4232 * check the target process has a UID that matches the current process's
4233 */
4234static bool check_same_owner(struct task_struct *p)
4235{
4236 const struct cred *cred = current_cred(), *pcred;
4237 bool match;
4238
4239 rcu_read_lock();
4240 pcred = __task_cred(p);
4241 match = (uid_eq(cred->euid, pcred->euid) ||
4242 uid_eq(cred->euid, pcred->uid));
4243 rcu_read_unlock();
4244 return match;
4245}
4246
4247static int __sched_setscheduler(struct task_struct *p, int policy,
4248 const struct sched_param *param, bool user)
4249{
4250 int retval, oldprio, oldpolicy = -1, on_rq, running;
4251 unsigned long flags;
4252 const struct sched_class *prev_class;
4253 struct rq *rq;
4254 int reset_on_fork;
4255
4256 /* may grab non-irq protected spin_locks */
4257 BUG_ON(in_interrupt());
4258recheck:
4259 /* double check policy once rq lock held */
4260 if (policy < 0) {
4261 reset_on_fork = p->sched_reset_on_fork;
4262 policy = oldpolicy = p->policy;
4263 } else {
4264 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4265 policy &= ~SCHED_RESET_ON_FORK;
4266
4267 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4268 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4269 policy != SCHED_IDLE)
4270 return -EINVAL;
4271 }
4272
4273 /*
4274 * Valid priorities for SCHED_FIFO and SCHED_RR are
4275 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4276 * SCHED_BATCH and SCHED_IDLE is 0.
4277 */
4278 if (param->sched_priority < 0 ||
4279 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4280 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4281 return -EINVAL;
4282 if (rt_policy(policy) != (param->sched_priority != 0))
4283 return -EINVAL;
4284
4285 /*
4286 * Allow unprivileged RT tasks to decrease priority:
4287 */
4288 if (user && !capable(CAP_SYS_NICE)) {
4289 if (rt_policy(policy)) {
4290 unsigned long rlim_rtprio =
4291 task_rlimit(p, RLIMIT_RTPRIO);
4292
4293 /* can't set/change the rt policy */
4294 if (policy != p->policy && !rlim_rtprio)
4295 return -EPERM;
4296
4297 /* can't increase priority */
4298 if (param->sched_priority > p->rt_priority &&
4299 param->sched_priority > rlim_rtprio)
4300 return -EPERM;
4301 }
4302
4303 /*
4304 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4305 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4306 */
4307 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4308 if (!can_nice(p, TASK_NICE(p)))
4309 return -EPERM;
4310 }
4311
4312 /* can't change other user's priorities */
4313 if (!check_same_owner(p))
4314 return -EPERM;
4315
4316 /* Normal users shall not reset the sched_reset_on_fork flag */
4317 if (p->sched_reset_on_fork && !reset_on_fork)
4318 return -EPERM;
4319 }
4320
4321 if (user) {
4322 retval = security_task_setscheduler(p);
4323 if (retval)
4324 return retval;
4325 }
4326
4327 /*
4328 * make sure no PI-waiters arrive (or leave) while we are
4329 * changing the priority of the task:
4330 *
4331 * To be able to change p->policy safely, the appropriate
4332 * runqueue lock must be held.
4333 */
4334 rq = task_rq_lock(p, &flags);
4335
4336 /*
4337 * Changing the policy of the stop threads its a very bad idea
4338 */
4339 if (p == rq->stop) {
4340 task_rq_unlock(rq, p, &flags);
4341 return -EINVAL;
4342 }
4343
4344 /*
4345 * If not changing anything there's no need to proceed further:
4346 */
4347 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4348 param->sched_priority == p->rt_priority))) {
4349
4350 __task_rq_unlock(rq);
4351 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4352 return 0;
4353 }
4354
4355#ifdef CONFIG_RT_GROUP_SCHED
4356 if (user) {
4357 /*
4358 * Do not allow realtime tasks into groups that have no runtime
4359 * assigned.
4360 */
4361 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4362 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4363 !task_group_is_autogroup(task_group(p))) {
4364 task_rq_unlock(rq, p, &flags);
4365 return -EPERM;
4366 }
4367 }
4368#endif
4369
4370 /* recheck policy now with rq lock held */
4371 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4372 policy = oldpolicy = -1;
4373 task_rq_unlock(rq, p, &flags);
4374 goto recheck;
4375 }
4376 on_rq = p->on_rq;
4377 running = task_current(rq, p);
4378 if (on_rq)
4379 dequeue_task(rq, p, 0);
4380 if (running)
4381 p->sched_class->put_prev_task(rq, p);
4382
4383 p->sched_reset_on_fork = reset_on_fork;
4384
4385 oldprio = p->prio;
4386 prev_class = p->sched_class;
4387 __setscheduler(rq, p, policy, param->sched_priority);
4388
4389 if (running)
4390 p->sched_class->set_curr_task(rq);
4391 if (on_rq)
4392 enqueue_task(rq, p, 0);
4393
4394 check_class_changed(rq, p, prev_class, oldprio);
4395 task_rq_unlock(rq, p, &flags);
4396
4397 rt_mutex_adjust_pi(p);
4398
4399 return 0;
4400}
4401
4402/**
4403 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4404 * @p: the task in question.
4405 * @policy: new policy.
4406 * @param: structure containing the new RT priority.
4407 *
4408 * NOTE that the task may be already dead.
4409 */
4410int sched_setscheduler(struct task_struct *p, int policy,
4411 const struct sched_param *param)
4412{
4413 return __sched_setscheduler(p, policy, param, true);
4414}
4415EXPORT_SYMBOL_GPL(sched_setscheduler);
4416
4417/**
4418 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4419 * @p: the task in question.
4420 * @policy: new policy.
4421 * @param: structure containing the new RT priority.
4422 *
4423 * Just like sched_setscheduler, only don't bother checking if the
4424 * current context has permission. For example, this is needed in
4425 * stop_machine(): we create temporary high priority worker threads,
4426 * but our caller might not have that capability.
4427 */
4428int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4429 const struct sched_param *param)
4430{
4431 return __sched_setscheduler(p, policy, param, false);
4432}
4433
4434static int
4435do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4436{
4437 struct sched_param lparam;
4438 struct task_struct *p;
4439 int retval;
4440
4441 if (!param || pid < 0)
4442 return -EINVAL;
4443 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4444 return -EFAULT;
4445
4446 rcu_read_lock();
4447 retval = -ESRCH;
4448 p = find_process_by_pid(pid);
4449 if (p != NULL)
4450 retval = sched_setscheduler(p, policy, &lparam);
4451 rcu_read_unlock();
4452
4453 return retval;
4454}
4455
4456/**
4457 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4458 * @pid: the pid in question.
4459 * @policy: new policy.
4460 * @param: structure containing the new RT priority.
4461 */
4462SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4463 struct sched_param __user *, param)
4464{
4465 /* negative values for policy are not valid */
4466 if (policy < 0)
4467 return -EINVAL;
4468
4469 return do_sched_setscheduler(pid, policy, param);
4470}
4471
4472/**
4473 * sys_sched_setparam - set/change the RT priority of a thread
4474 * @pid: the pid in question.
4475 * @param: structure containing the new RT priority.
4476 */
4477SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4478{
4479 return do_sched_setscheduler(pid, -1, param);
4480}
4481
4482/**
4483 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4484 * @pid: the pid in question.
4485 */
4486SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4487{
4488 struct task_struct *p;
4489 int retval;
4490
4491 if (pid < 0)
4492 return -EINVAL;
4493
4494 retval = -ESRCH;
4495 rcu_read_lock();
4496 p = find_process_by_pid(pid);
4497 if (p) {
4498 retval = security_task_getscheduler(p);
4499 if (!retval)
4500 retval = p->policy
4501 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4502 }
4503 rcu_read_unlock();
4504 return retval;
4505}
4506
4507/**
4508 * sys_sched_getparam - get the RT priority of a thread
4509 * @pid: the pid in question.
4510 * @param: structure containing the RT priority.
4511 */
4512SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4513{
4514 struct sched_param lp;
4515 struct task_struct *p;
4516 int retval;
4517
4518 if (!param || pid < 0)
4519 return -EINVAL;
4520
4521 rcu_read_lock();
4522 p = find_process_by_pid(pid);
4523 retval = -ESRCH;
4524 if (!p)
4525 goto out_unlock;
4526
4527 retval = security_task_getscheduler(p);
4528 if (retval)
4529 goto out_unlock;
4530
4531 lp.sched_priority = p->rt_priority;
4532 rcu_read_unlock();
4533
4534 /*
4535 * This one might sleep, we cannot do it with a spinlock held ...
4536 */
4537 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4538
4539 return retval;
4540
4541out_unlock:
4542 rcu_read_unlock();
4543 return retval;
4544}
4545
4546long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4547{
4548 cpumask_var_t cpus_allowed, new_mask;
4549 struct task_struct *p;
4550 int retval;
4551
4552 get_online_cpus();
4553 rcu_read_lock();
4554
4555 p = find_process_by_pid(pid);
4556 if (!p) {
4557 rcu_read_unlock();
4558 put_online_cpus();
4559 return -ESRCH;
4560 }
4561
4562 /* Prevent p going away */
4563 get_task_struct(p);
4564 rcu_read_unlock();
4565
4566 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4567 retval = -ENOMEM;
4568 goto out_put_task;
4569 }
4570 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4571 retval = -ENOMEM;
4572 goto out_free_cpus_allowed;
4573 }
4574 retval = -EPERM;
4575 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4576 goto out_unlock;
4577
4578 retval = security_task_setscheduler(p);
4579 if (retval)
4580 goto out_unlock;
4581
4582 cpuset_cpus_allowed(p, cpus_allowed);
4583 cpumask_and(new_mask, in_mask, cpus_allowed);
4584again:
4585 retval = set_cpus_allowed_ptr(p, new_mask);
4586
4587 if (!retval) {
4588 cpuset_cpus_allowed(p, cpus_allowed);
4589 if (!cpumask_subset(new_mask, cpus_allowed)) {
4590 /*
4591 * We must have raced with a concurrent cpuset
4592 * update. Just reset the cpus_allowed to the
4593 * cpuset's cpus_allowed
4594 */
4595 cpumask_copy(new_mask, cpus_allowed);
4596 goto again;
4597 }
4598 }
4599out_unlock:
4600 free_cpumask_var(new_mask);
4601out_free_cpus_allowed:
4602 free_cpumask_var(cpus_allowed);
4603out_put_task:
4604 put_task_struct(p);
4605 put_online_cpus();
4606 return retval;
4607}
4608
4609static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4610 struct cpumask *new_mask)
4611{
4612 if (len < cpumask_size())
4613 cpumask_clear(new_mask);
4614 else if (len > cpumask_size())
4615 len = cpumask_size();
4616
4617 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4618}
4619
4620/**
4621 * sys_sched_setaffinity - set the cpu affinity of a process
4622 * @pid: pid of the process
4623 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4624 * @user_mask_ptr: user-space pointer to the new cpu mask
4625 */
4626SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4627 unsigned long __user *, user_mask_ptr)
4628{
4629 cpumask_var_t new_mask;
4630 int retval;
4631
4632 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4633 return -ENOMEM;
4634
4635 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4636 if (retval == 0)
4637 retval = sched_setaffinity(pid, new_mask);
4638 free_cpumask_var(new_mask);
4639 return retval;
4640}
4641
4642long sched_getaffinity(pid_t pid, struct cpumask *mask)
4643{
4644 struct task_struct *p;
4645 unsigned long flags;
4646 int retval;
4647
4648 get_online_cpus();
4649 rcu_read_lock();
4650
4651 retval = -ESRCH;
4652 p = find_process_by_pid(pid);
4653 if (!p)
4654 goto out_unlock;
4655
4656 retval = security_task_getscheduler(p);
4657 if (retval)
4658 goto out_unlock;
4659
4660 raw_spin_lock_irqsave(&p->pi_lock, flags);
4661 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4662 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4663
4664out_unlock:
4665 rcu_read_unlock();
4666 put_online_cpus();
4667
4668 return retval;
4669}
4670
4671/**
4672 * sys_sched_getaffinity - get the cpu affinity of a process
4673 * @pid: pid of the process
4674 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4675 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4676 */
4677SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4678 unsigned long __user *, user_mask_ptr)
4679{
4680 int ret;
4681 cpumask_var_t mask;
4682
4683 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4684 return -EINVAL;
4685 if (len & (sizeof(unsigned long)-1))
4686 return -EINVAL;
4687
4688 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4689 return -ENOMEM;
4690
4691 ret = sched_getaffinity(pid, mask);
4692 if (ret == 0) {
4693 size_t retlen = min_t(size_t, len, cpumask_size());
4694
4695 if (copy_to_user(user_mask_ptr, mask, retlen))
4696 ret = -EFAULT;
4697 else
4698 ret = retlen;
4699 }
4700 free_cpumask_var(mask);
4701
4702 return ret;
4703}
4704
4705/**
4706 * sys_sched_yield - yield the current processor to other threads.
4707 *
4708 * This function yields the current CPU to other tasks. If there are no
4709 * other threads running on this CPU then this function will return.
4710 */
4711SYSCALL_DEFINE0(sched_yield)
4712{
4713 struct rq *rq = this_rq_lock();
4714
4715 schedstat_inc(rq, yld_count);
4716 current->sched_class->yield_task(rq);
4717
4718 /*
4719 * Since we are going to call schedule() anyway, there's
4720 * no need to preempt or enable interrupts:
4721 */
4722 __release(rq->lock);
4723 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4724 do_raw_spin_unlock(&rq->lock);
4725 sched_preempt_enable_no_resched();
4726
4727 schedule();
4728
4729 return 0;
4730}
4731
4732static inline int should_resched(void)
4733{
4734 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4735}
4736
4737static void __cond_resched(void)
4738{
4739 add_preempt_count(PREEMPT_ACTIVE);
4740 __schedule();
4741 sub_preempt_count(PREEMPT_ACTIVE);
4742}
4743
4744int __sched _cond_resched(void)
4745{
4746 if (should_resched()) {
4747 __cond_resched();
4748 return 1;
4749 }
4750 return 0;
4751}
4752EXPORT_SYMBOL(_cond_resched);
4753
4754/*
4755 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4756 * call schedule, and on return reacquire the lock.
4757 *
4758 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4759 * operations here to prevent schedule() from being called twice (once via
4760 * spin_unlock(), once by hand).
4761 */
4762int __cond_resched_lock(spinlock_t *lock)
4763{
4764 int resched = should_resched();
4765 int ret = 0;
4766
4767 lockdep_assert_held(lock);
4768
4769 if (spin_needbreak(lock) || resched) {
4770 spin_unlock(lock);
4771 if (resched)
4772 __cond_resched();
4773 else
4774 cpu_relax();
4775 ret = 1;
4776 spin_lock(lock);
4777 }
4778 return ret;
4779}
4780EXPORT_SYMBOL(__cond_resched_lock);
4781
4782int __sched __cond_resched_softirq(void)
4783{
4784 BUG_ON(!in_softirq());
4785
4786 if (should_resched()) {
4787 local_bh_enable();
4788 __cond_resched();
4789 local_bh_disable();
4790 return 1;
4791 }
4792 return 0;
4793}
4794EXPORT_SYMBOL(__cond_resched_softirq);
4795
4796/**
4797 * yield - yield the current processor to other threads.
4798 *
4799 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4800 *
4801 * The scheduler is at all times free to pick the calling task as the most
4802 * eligible task to run, if removing the yield() call from your code breaks
4803 * it, its already broken.
4804 *
4805 * Typical broken usage is:
4806 *
4807 * while (!event)
4808 * yield();
4809 *
4810 * where one assumes that yield() will let 'the other' process run that will
4811 * make event true. If the current task is a SCHED_FIFO task that will never
4812 * happen. Never use yield() as a progress guarantee!!
4813 *
4814 * If you want to use yield() to wait for something, use wait_event().
4815 * If you want to use yield() to be 'nice' for others, use cond_resched().
4816 * If you still want to use yield(), do not!
4817 */
4818void __sched yield(void)
4819{
4820 set_current_state(TASK_RUNNING);
4821 sys_sched_yield();
4822}
4823EXPORT_SYMBOL(yield);
4824
4825/**
4826 * yield_to - yield the current processor to another thread in
4827 * your thread group, or accelerate that thread toward the
4828 * processor it's on.
4829 * @p: target task
4830 * @preempt: whether task preemption is allowed or not
4831 *
4832 * It's the caller's job to ensure that the target task struct
4833 * can't go away on us before we can do any checks.
4834 *
4835 * Returns true if we indeed boosted the target task.
4836 */
4837bool __sched yield_to(struct task_struct *p, bool preempt)
4838{
4839 struct task_struct *curr = current;
4840 struct rq *rq, *p_rq;
4841 unsigned long flags;
4842 bool yielded = 0;
4843
4844 local_irq_save(flags);
4845 rq = this_rq();
4846
4847again:
4848 p_rq = task_rq(p);
4849 double_rq_lock(rq, p_rq);
4850 while (task_rq(p) != p_rq) {
4851 double_rq_unlock(rq, p_rq);
4852 goto again;
4853 }
4854
4855 if (!curr->sched_class->yield_to_task)
4856 goto out;
4857
4858 if (curr->sched_class != p->sched_class)
4859 goto out;
4860
4861 if (task_running(p_rq, p) || p->state)
4862 goto out;
4863
4864 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4865 if (yielded) {
4866 schedstat_inc(rq, yld_count);
4867 /*
4868 * Make p's CPU reschedule; pick_next_entity takes care of
4869 * fairness.
4870 */
4871 if (preempt && rq != p_rq)
4872 resched_task(p_rq->curr);
4873 } else {
4874 /*
4875 * We might have set it in task_yield_fair(), but are
4876 * not going to schedule(), so don't want to skip
4877 * the next update.
4878 */
4879 rq->skip_clock_update = 0;
4880 }
4881
4882out:
4883 double_rq_unlock(rq, p_rq);
4884 local_irq_restore(flags);
4885
4886 if (yielded)
4887 schedule();
4888
4889 return yielded;
4890}
4891EXPORT_SYMBOL_GPL(yield_to);
4892
4893/*
4894 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4895 * that process accounting knows that this is a task in IO wait state.
4896 */
4897void __sched io_schedule(void)
4898{
4899 struct rq *rq = raw_rq();
4900
4901 delayacct_blkio_start();
4902 atomic_inc(&rq->nr_iowait);
4903 blk_flush_plug(current);
4904 current->in_iowait = 1;
4905 schedule();
4906 current->in_iowait = 0;
4907 atomic_dec(&rq->nr_iowait);
4908 delayacct_blkio_end();
4909}
4910EXPORT_SYMBOL(io_schedule);
4911
4912long __sched io_schedule_timeout(long timeout)
4913{
4914 struct rq *rq = raw_rq();
4915 long ret;
4916
4917 delayacct_blkio_start();
4918 atomic_inc(&rq->nr_iowait);
4919 blk_flush_plug(current);
4920 current->in_iowait = 1;
4921 ret = schedule_timeout(timeout);
4922 current->in_iowait = 0;
4923 atomic_dec(&rq->nr_iowait);
4924 delayacct_blkio_end();
4925 return ret;
4926}
4927
4928/**
4929 * sys_sched_get_priority_max - return maximum RT priority.
4930 * @policy: scheduling class.
4931 *
4932 * this syscall returns the maximum rt_priority that can be used
4933 * by a given scheduling class.
4934 */
4935SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4936{
4937 int ret = -EINVAL;
4938
4939 switch (policy) {
4940 case SCHED_FIFO:
4941 case SCHED_RR:
4942 ret = MAX_USER_RT_PRIO-1;
4943 break;
4944 case SCHED_NORMAL:
4945 case SCHED_BATCH:
4946 case SCHED_IDLE:
4947 ret = 0;
4948 break;
4949 }
4950 return ret;
4951}
4952
4953/**
4954 * sys_sched_get_priority_min - return minimum RT priority.
4955 * @policy: scheduling class.
4956 *
4957 * this syscall returns the minimum rt_priority that can be used
4958 * by a given scheduling class.
4959 */
4960SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4961{
4962 int ret = -EINVAL;
4963
4964 switch (policy) {
4965 case SCHED_FIFO:
4966 case SCHED_RR:
4967 ret = 1;
4968 break;
4969 case SCHED_NORMAL:
4970 case SCHED_BATCH:
4971 case SCHED_IDLE:
4972 ret = 0;
4973 }
4974 return ret;
4975}
4976
4977/**
4978 * sys_sched_rr_get_interval - return the default timeslice of a process.
4979 * @pid: pid of the process.
4980 * @interval: userspace pointer to the timeslice value.
4981 *
4982 * this syscall writes the default timeslice value of a given process
4983 * into the user-space timespec buffer. A value of '0' means infinity.
4984 */
4985SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4986 struct timespec __user *, interval)
4987{
4988 struct task_struct *p;
4989 unsigned int time_slice;
4990 unsigned long flags;
4991 struct rq *rq;
4992 int retval;
4993 struct timespec t;
4994
4995 if (pid < 0)
4996 return -EINVAL;
4997
4998 retval = -ESRCH;
4999 rcu_read_lock();
5000 p = find_process_by_pid(pid);
5001 if (!p)
5002 goto out_unlock;
5003
5004 retval = security_task_getscheduler(p);
5005 if (retval)
5006 goto out_unlock;
5007
5008 rq = task_rq_lock(p, &flags);
5009 time_slice = p->sched_class->get_rr_interval(rq, p);
5010 task_rq_unlock(rq, p, &flags);
5011
5012 rcu_read_unlock();
5013 jiffies_to_timespec(time_slice, &t);
5014 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5015 return retval;
5016
5017out_unlock:
5018 rcu_read_unlock();
5019 return retval;
5020}
5021
5022static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5023
5024void sched_show_task(struct task_struct *p)
5025{
5026 unsigned long free = 0;
5027 unsigned state;
5028
5029 state = p->state ? __ffs(p->state) + 1 : 0;
5030 printk(KERN_INFO "%-15.15s %c", p->comm,
5031 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5032#if BITS_PER_LONG == 32
5033 if (state == TASK_RUNNING)
5034 printk(KERN_CONT " running ");
5035 else
5036 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5037#else
5038 if (state == TASK_RUNNING)
5039 printk(KERN_CONT " running task ");
5040 else
5041 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5042#endif
5043#ifdef CONFIG_DEBUG_STACK_USAGE
5044 free = stack_not_used(p);
5045#endif
5046 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5047 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
5048 (unsigned long)task_thread_info(p)->flags);
5049
5050 show_stack(p, NULL);
5051}
5052
5053void show_state_filter(unsigned long state_filter)
5054{
5055 struct task_struct *g, *p;
5056
5057#if BITS_PER_LONG == 32
5058 printk(KERN_INFO
5059 " task PC stack pid father\n");
5060#else
5061 printk(KERN_INFO
5062 " task PC stack pid father\n");
5063#endif
5064 rcu_read_lock();
5065 do_each_thread(g, p) {
5066 /*
5067 * reset the NMI-timeout, listing all files on a slow
5068 * console might take a lot of time:
5069 */
5070 touch_nmi_watchdog();
5071 if (!state_filter || (p->state & state_filter))
5072 sched_show_task(p);
5073 } while_each_thread(g, p);
5074
5075 touch_all_softlockup_watchdogs();
5076
5077#ifdef CONFIG_SCHED_DEBUG
5078 sysrq_sched_debug_show();
5079#endif
5080 rcu_read_unlock();
5081 /*
5082 * Only show locks if all tasks are dumped:
5083 */
5084 if (!state_filter)
5085 debug_show_all_locks();
5086}
5087
5088void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5089{
5090 idle->sched_class = &idle_sched_class;
5091}
5092
5093/**
5094 * init_idle - set up an idle thread for a given CPU
5095 * @idle: task in question
5096 * @cpu: cpu the idle task belongs to
5097 *
5098 * NOTE: this function does not set the idle thread's NEED_RESCHED
5099 * flag, to make booting more robust.
5100 */
5101void __cpuinit init_idle(struct task_struct *idle, int cpu)
5102{
5103 struct rq *rq = cpu_rq(cpu);
5104 unsigned long flags;
5105
5106 raw_spin_lock_irqsave(&rq->lock, flags);
5107
5108 __sched_fork(idle);
5109 idle->state = TASK_RUNNING;
5110 idle->se.exec_start = sched_clock();
5111
5112 do_set_cpus_allowed(idle, cpumask_of(cpu));
5113 /*
5114 * We're having a chicken and egg problem, even though we are
5115 * holding rq->lock, the cpu isn't yet set to this cpu so the
5116 * lockdep check in task_group() will fail.
5117 *
5118 * Similar case to sched_fork(). / Alternatively we could
5119 * use task_rq_lock() here and obtain the other rq->lock.
5120 *
5121 * Silence PROVE_RCU
5122 */
5123 rcu_read_lock();
5124 __set_task_cpu(idle, cpu);
5125 rcu_read_unlock();
5126
5127 rq->curr = rq->idle = idle;
5128#if defined(CONFIG_SMP)
5129 idle->on_cpu = 1;
5130#endif
5131 raw_spin_unlock_irqrestore(&rq->lock, flags);
5132
5133 /* Set the preempt count _outside_ the spinlocks! */
5134 task_thread_info(idle)->preempt_count = 0;
5135
5136 /*
5137 * The idle tasks have their own, simple scheduling class:
5138 */
5139 idle->sched_class = &idle_sched_class;
5140 ftrace_graph_init_idle_task(idle, cpu);
5141#if defined(CONFIG_SMP)
5142 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5143#endif
5144}
5145
5146#ifdef CONFIG_SMP
5147void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5148{
5149 if (p->sched_class && p->sched_class->set_cpus_allowed)
5150 p->sched_class->set_cpus_allowed(p, new_mask);
5151
5152 cpumask_copy(&p->cpus_allowed, new_mask);
5153 p->nr_cpus_allowed = cpumask_weight(new_mask);
5154}
5155
5156/*
5157 * This is how migration works:
5158 *
5159 * 1) we invoke migration_cpu_stop() on the target CPU using
5160 * stop_one_cpu().
5161 * 2) stopper starts to run (implicitly forcing the migrated thread
5162 * off the CPU)
5163 * 3) it checks whether the migrated task is still in the wrong runqueue.
5164 * 4) if it's in the wrong runqueue then the migration thread removes
5165 * it and puts it into the right queue.
5166 * 5) stopper completes and stop_one_cpu() returns and the migration
5167 * is done.
5168 */
5169
5170/*
5171 * Change a given task's CPU affinity. Migrate the thread to a
5172 * proper CPU and schedule it away if the CPU it's executing on
5173 * is removed from the allowed bitmask.
5174 *
5175 * NOTE: the caller must have a valid reference to the task, the
5176 * task must not exit() & deallocate itself prematurely. The
5177 * call is not atomic; no spinlocks may be held.
5178 */
5179int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5180{
5181 unsigned long flags;
5182 struct rq *rq;
5183 unsigned int dest_cpu;
5184 int ret = 0;
5185
5186 rq = task_rq_lock(p, &flags);
5187
5188 if (cpumask_equal(&p->cpus_allowed, new_mask))
5189 goto out;
5190
5191 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5192 ret = -EINVAL;
5193 goto out;
5194 }
5195
5196 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5197 ret = -EINVAL;
5198 goto out;
5199 }
5200
5201 do_set_cpus_allowed(p, new_mask);
5202
5203 /* Can the task run on the task's current CPU? If so, we're done */
5204 if (cpumask_test_cpu(task_cpu(p), new_mask))
5205 goto out;
5206
5207 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5208 if (p->on_rq) {
5209 struct migration_arg arg = { p, dest_cpu };
5210 /* Need help from migration thread: drop lock and wait. */
5211 task_rq_unlock(rq, p, &flags);
5212 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5213 tlb_migrate_finish(p->mm);
5214 return 0;
5215 }
5216out:
5217 task_rq_unlock(rq, p, &flags);
5218
5219 return ret;
5220}
5221EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5222
5223/*
5224 * Move (not current) task off this cpu, onto dest cpu. We're doing
5225 * this because either it can't run here any more (set_cpus_allowed()
5226 * away from this CPU, or CPU going down), or because we're
5227 * attempting to rebalance this task on exec (sched_exec).
5228 *
5229 * So we race with normal scheduler movements, but that's OK, as long
5230 * as the task is no longer on this CPU.
5231 *
5232 * Returns non-zero if task was successfully migrated.
5233 */
5234static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5235{
5236 struct rq *rq_dest, *rq_src;
5237 int ret = 0;
5238
5239 if (unlikely(!cpu_active(dest_cpu)))
5240 return ret;
5241
5242 rq_src = cpu_rq(src_cpu);
5243 rq_dest = cpu_rq(dest_cpu);
5244
5245 raw_spin_lock(&p->pi_lock);
5246 double_rq_lock(rq_src, rq_dest);
5247 /* Already moved. */
5248 if (task_cpu(p) != src_cpu)
5249 goto done;
5250 /* Affinity changed (again). */
5251 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5252 goto fail;
5253
5254 /*
5255 * If we're not on a rq, the next wake-up will ensure we're
5256 * placed properly.
5257 */
5258 if (p->on_rq) {
5259 dequeue_task(rq_src, p, 0);
5260 set_task_cpu(p, dest_cpu);
5261 enqueue_task(rq_dest, p, 0);
5262 check_preempt_curr(rq_dest, p, 0);
5263 }
5264done:
5265 ret = 1;
5266fail:
5267 double_rq_unlock(rq_src, rq_dest);
5268 raw_spin_unlock(&p->pi_lock);
5269 return ret;
5270}
5271
5272/*
5273 * migration_cpu_stop - this will be executed by a highprio stopper thread
5274 * and performs thread migration by bumping thread off CPU then
5275 * 'pushing' onto another runqueue.
5276 */
5277static int migration_cpu_stop(void *data)
5278{
5279 struct migration_arg *arg = data;
5280
5281 /*
5282 * The original target cpu might have gone down and we might
5283 * be on another cpu but it doesn't matter.
5284 */
5285 local_irq_disable();
5286 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5287 local_irq_enable();
5288 return 0;
5289}
5290
5291#ifdef CONFIG_HOTPLUG_CPU
5292
5293/*
5294 * Ensures that the idle task is using init_mm right before its cpu goes
5295 * offline.
5296 */
5297void idle_task_exit(void)
5298{
5299 struct mm_struct *mm = current->active_mm;
5300
5301 BUG_ON(cpu_online(smp_processor_id()));
5302
5303 if (mm != &init_mm)
5304 switch_mm(mm, &init_mm, current);
5305 mmdrop(mm);
5306}
5307
5308/*
5309 * While a dead CPU has no uninterruptible tasks queued at this point,
5310 * it might still have a nonzero ->nr_uninterruptible counter, because
5311 * for performance reasons the counter is not stricly tracking tasks to
5312 * their home CPUs. So we just add the counter to another CPU's counter,
5313 * to keep the global sum constant after CPU-down:
5314 */
5315static void migrate_nr_uninterruptible(struct rq *rq_src)
5316{
5317 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5318
5319 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5320 rq_src->nr_uninterruptible = 0;
5321}
5322
5323/*
5324 * remove the tasks which were accounted by rq from calc_load_tasks.
5325 */
5326static void calc_global_load_remove(struct rq *rq)
5327{
5328 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5329 rq->calc_load_active = 0;
5330}
5331
5332/*
5333 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5334 * try_to_wake_up()->select_task_rq().
5335 *
5336 * Called with rq->lock held even though we'er in stop_machine() and
5337 * there's no concurrency possible, we hold the required locks anyway
5338 * because of lock validation efforts.
5339 */
5340static void migrate_tasks(unsigned int dead_cpu)
5341{
5342 struct rq *rq = cpu_rq(dead_cpu);
5343 struct task_struct *next, *stop = rq->stop;
5344 int dest_cpu;
5345
5346 /*
5347 * Fudge the rq selection such that the below task selection loop
5348 * doesn't get stuck on the currently eligible stop task.
5349 *
5350 * We're currently inside stop_machine() and the rq is either stuck
5351 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5352 * either way we should never end up calling schedule() until we're
5353 * done here.
5354 */
5355 rq->stop = NULL;
5356
5357 /* Ensure any throttled groups are reachable by pick_next_task */
5358 unthrottle_offline_cfs_rqs(rq);
5359
5360 for ( ; ; ) {
5361 /*
5362 * There's this thread running, bail when that's the only
5363 * remaining thread.
5364 */
5365 if (rq->nr_running == 1)
5366 break;
5367
5368 next = pick_next_task(rq);
5369 BUG_ON(!next);
5370 next->sched_class->put_prev_task(rq, next);
5371
5372 /* Find suitable destination for @next, with force if needed. */
5373 dest_cpu = select_fallback_rq(dead_cpu, next);
5374 raw_spin_unlock(&rq->lock);
5375
5376 __migrate_task(next, dead_cpu, dest_cpu);
5377
5378 raw_spin_lock(&rq->lock);
5379 }
5380
5381 rq->stop = stop;
5382}
5383
5384#endif /* CONFIG_HOTPLUG_CPU */
5385
5386#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5387
5388static struct ctl_table sd_ctl_dir[] = {
5389 {
5390 .procname = "sched_domain",
5391 .mode = 0555,
5392 },
5393 {}
5394};
5395
5396static struct ctl_table sd_ctl_root[] = {
5397 {
5398 .procname = "kernel",
5399 .mode = 0555,
5400 .child = sd_ctl_dir,
5401 },
5402 {}
5403};
5404
5405static struct ctl_table *sd_alloc_ctl_entry(int n)
5406{
5407 struct ctl_table *entry =
5408 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5409
5410 return entry;
5411}
5412
5413static void sd_free_ctl_entry(struct ctl_table **tablep)
5414{
5415 struct ctl_table *entry;
5416
5417 /*
5418 * In the intermediate directories, both the child directory and
5419 * procname are dynamically allocated and could fail but the mode
5420 * will always be set. In the lowest directory the names are
5421 * static strings and all have proc handlers.
5422 */
5423 for (entry = *tablep; entry->mode; entry++) {
5424 if (entry->child)
5425 sd_free_ctl_entry(&entry->child);
5426 if (entry->proc_handler == NULL)
5427 kfree(entry->procname);
5428 }
5429
5430 kfree(*tablep);
5431 *tablep = NULL;
5432}
5433
5434static void
5435set_table_entry(struct ctl_table *entry,
5436 const char *procname, void *data, int maxlen,
5437 umode_t mode, proc_handler *proc_handler)
5438{
5439 entry->procname = procname;
5440 entry->data = data;
5441 entry->maxlen = maxlen;
5442 entry->mode = mode;
5443 entry->proc_handler = proc_handler;
5444}
5445
5446static struct ctl_table *
5447sd_alloc_ctl_domain_table(struct sched_domain *sd)
5448{
5449 struct ctl_table *table = sd_alloc_ctl_entry(13);
5450
5451 if (table == NULL)
5452 return NULL;
5453
5454 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5455 sizeof(long), 0644, proc_doulongvec_minmax);
5456 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5457 sizeof(long), 0644, proc_doulongvec_minmax);
5458 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5459 sizeof(int), 0644, proc_dointvec_minmax);
5460 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5461 sizeof(int), 0644, proc_dointvec_minmax);
5462 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5463 sizeof(int), 0644, proc_dointvec_minmax);
5464 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5465 sizeof(int), 0644, proc_dointvec_minmax);
5466 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5467 sizeof(int), 0644, proc_dointvec_minmax);
5468 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 set_table_entry(&table[9], "cache_nice_tries",
5473 &sd->cache_nice_tries,
5474 sizeof(int), 0644, proc_dointvec_minmax);
5475 set_table_entry(&table[10], "flags", &sd->flags,
5476 sizeof(int), 0644, proc_dointvec_minmax);
5477 set_table_entry(&table[11], "name", sd->name,
5478 CORENAME_MAX_SIZE, 0444, proc_dostring);
5479 /* &table[12] is terminator */
5480
5481 return table;
5482}
5483
5484static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5485{
5486 struct ctl_table *entry, *table;
5487 struct sched_domain *sd;
5488 int domain_num = 0, i;
5489 char buf[32];
5490
5491 for_each_domain(cpu, sd)
5492 domain_num++;
5493 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5494 if (table == NULL)
5495 return NULL;
5496
5497 i = 0;
5498 for_each_domain(cpu, sd) {
5499 snprintf(buf, 32, "domain%d", i);
5500 entry->procname = kstrdup(buf, GFP_KERNEL);
5501 entry->mode = 0555;
5502 entry->child = sd_alloc_ctl_domain_table(sd);
5503 entry++;
5504 i++;
5505 }
5506 return table;
5507}
5508
5509static struct ctl_table_header *sd_sysctl_header;
5510static void register_sched_domain_sysctl(void)
5511{
5512 int i, cpu_num = num_possible_cpus();
5513 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5514 char buf[32];
5515
5516 WARN_ON(sd_ctl_dir[0].child);
5517 sd_ctl_dir[0].child = entry;
5518
5519 if (entry == NULL)
5520 return;
5521
5522 for_each_possible_cpu(i) {
5523 snprintf(buf, 32, "cpu%d", i);
5524 entry->procname = kstrdup(buf, GFP_KERNEL);
5525 entry->mode = 0555;
5526 entry->child = sd_alloc_ctl_cpu_table(i);
5527 entry++;
5528 }
5529
5530 WARN_ON(sd_sysctl_header);
5531 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5532}
5533
5534/* may be called multiple times per register */
5535static void unregister_sched_domain_sysctl(void)
5536{
5537 if (sd_sysctl_header)
5538 unregister_sysctl_table(sd_sysctl_header);
5539 sd_sysctl_header = NULL;
5540 if (sd_ctl_dir[0].child)
5541 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5542}
5543#else
5544static void register_sched_domain_sysctl(void)
5545{
5546}
5547static void unregister_sched_domain_sysctl(void)
5548{
5549}
5550#endif
5551
5552static void set_rq_online(struct rq *rq)
5553{
5554 if (!rq->online) {
5555 const struct sched_class *class;
5556
5557 cpumask_set_cpu(rq->cpu, rq->rd->online);
5558 rq->online = 1;
5559
5560 for_each_class(class) {
5561 if (class->rq_online)
5562 class->rq_online(rq);
5563 }
5564 }
5565}
5566
5567static void set_rq_offline(struct rq *rq)
5568{
5569 if (rq->online) {
5570 const struct sched_class *class;
5571
5572 for_each_class(class) {
5573 if (class->rq_offline)
5574 class->rq_offline(rq);
5575 }
5576
5577 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5578 rq->online = 0;
5579 }
5580}
5581
5582/*
5583 * migration_call - callback that gets triggered when a CPU is added.
5584 * Here we can start up the necessary migration thread for the new CPU.
5585 */
5586static int __cpuinit
5587migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5588{
5589 int cpu = (long)hcpu;
5590 unsigned long flags;
5591 struct rq *rq = cpu_rq(cpu);
5592
5593 switch (action & ~CPU_TASKS_FROZEN) {
5594
5595 case CPU_UP_PREPARE:
5596 rq->calc_load_update = calc_load_update;
5597 break;
5598
5599 case CPU_ONLINE:
5600 /* Update our root-domain */
5601 raw_spin_lock_irqsave(&rq->lock, flags);
5602 if (rq->rd) {
5603 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5604
5605 set_rq_online(rq);
5606 }
5607 raw_spin_unlock_irqrestore(&rq->lock, flags);
5608 break;
5609
5610#ifdef CONFIG_HOTPLUG_CPU
5611 case CPU_DYING:
5612 sched_ttwu_pending();
5613 /* Update our root-domain */
5614 raw_spin_lock_irqsave(&rq->lock, flags);
5615 if (rq->rd) {
5616 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5617 set_rq_offline(rq);
5618 }
5619 migrate_tasks(cpu);
5620 BUG_ON(rq->nr_running != 1); /* the migration thread */
5621 raw_spin_unlock_irqrestore(&rq->lock, flags);
5622
5623 migrate_nr_uninterruptible(rq);
5624 calc_global_load_remove(rq);
5625 break;
5626#endif
5627 }
5628
5629 update_max_interval();
5630
5631 return NOTIFY_OK;
5632}
5633
5634/*
5635 * Register at high priority so that task migration (migrate_all_tasks)
5636 * happens before everything else. This has to be lower priority than
5637 * the notifier in the perf_event subsystem, though.
5638 */
5639static struct notifier_block __cpuinitdata migration_notifier = {
5640 .notifier_call = migration_call,
5641 .priority = CPU_PRI_MIGRATION,
5642};
5643
5644static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5645 unsigned long action, void *hcpu)
5646{
5647 switch (action & ~CPU_TASKS_FROZEN) {
5648 case CPU_STARTING:
5649 case CPU_DOWN_FAILED:
5650 set_cpu_active((long)hcpu, true);
5651 return NOTIFY_OK;
5652 default:
5653 return NOTIFY_DONE;
5654 }
5655}
5656
5657static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5658 unsigned long action, void *hcpu)
5659{
5660 switch (action & ~CPU_TASKS_FROZEN) {
5661 case CPU_DOWN_PREPARE:
5662 set_cpu_active((long)hcpu, false);
5663 return NOTIFY_OK;
5664 default:
5665 return NOTIFY_DONE;
5666 }
5667}
5668
5669static int __init migration_init(void)
5670{
5671 void *cpu = (void *)(long)smp_processor_id();
5672 int err;
5673
5674 /* Initialize migration for the boot CPU */
5675 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5676 BUG_ON(err == NOTIFY_BAD);
5677 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5678 register_cpu_notifier(&migration_notifier);
5679
5680 /* Register cpu active notifiers */
5681 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5682 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5683
5684 return 0;
5685}
5686early_initcall(migration_init);
5687#endif
5688
5689#ifdef CONFIG_SMP
5690
5691static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5692
5693#ifdef CONFIG_SCHED_DEBUG
5694
5695static __read_mostly int sched_debug_enabled;
5696
5697static int __init sched_debug_setup(char *str)
5698{
5699 sched_debug_enabled = 1;
5700
5701 return 0;
5702}
5703early_param("sched_debug", sched_debug_setup);
5704
5705static inline bool sched_debug(void)
5706{
5707 return sched_debug_enabled;
5708}
5709
5710static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5711 struct cpumask *groupmask)
5712{
5713 struct sched_group *group = sd->groups;
5714 char str[256];
5715
5716 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5717 cpumask_clear(groupmask);
5718
5719 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5720
5721 if (!(sd->flags & SD_LOAD_BALANCE)) {
5722 printk("does not load-balance\n");
5723 if (sd->parent)
5724 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5725 " has parent");
5726 return -1;
5727 }
5728
5729 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5730
5731 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5732 printk(KERN_ERR "ERROR: domain->span does not contain "
5733 "CPU%d\n", cpu);
5734 }
5735 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5736 printk(KERN_ERR "ERROR: domain->groups does not contain"
5737 " CPU%d\n", cpu);
5738 }
5739
5740 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5741 do {
5742 if (!group) {
5743 printk("\n");
5744 printk(KERN_ERR "ERROR: group is NULL\n");
5745 break;
5746 }
5747
5748 /*
5749 * Even though we initialize ->power to something semi-sane,
5750 * we leave power_orig unset. This allows us to detect if
5751 * domain iteration is still funny without causing /0 traps.
5752 */
5753 if (!group->sgp->power_orig) {
5754 printk(KERN_CONT "\n");
5755 printk(KERN_ERR "ERROR: domain->cpu_power not "
5756 "set\n");
5757 break;
5758 }
5759
5760 if (!cpumask_weight(sched_group_cpus(group))) {
5761 printk(KERN_CONT "\n");
5762 printk(KERN_ERR "ERROR: empty group\n");
5763 break;
5764 }
5765
5766 if (!(sd->flags & SD_OVERLAP) &&
5767 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5768 printk(KERN_CONT "\n");
5769 printk(KERN_ERR "ERROR: repeated CPUs\n");
5770 break;
5771 }
5772
5773 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5774
5775 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5776
5777 printk(KERN_CONT " %s", str);
5778 if (group->sgp->power != SCHED_POWER_SCALE) {
5779 printk(KERN_CONT " (cpu_power = %d)",
5780 group->sgp->power);
5781 }
5782
5783 group = group->next;
5784 } while (group != sd->groups);
5785 printk(KERN_CONT "\n");
5786
5787 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5788 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5789
5790 if (sd->parent &&
5791 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5792 printk(KERN_ERR "ERROR: parent span is not a superset "
5793 "of domain->span\n");
5794 return 0;
5795}
5796
5797static void sched_domain_debug(struct sched_domain *sd, int cpu)
5798{
5799 int level = 0;
5800
5801 if (!sched_debug_enabled)
5802 return;
5803
5804 if (!sd) {
5805 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5806 return;
5807 }
5808
5809 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5810
5811 for (;;) {
5812 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5813 break;
5814 level++;
5815 sd = sd->parent;
5816 if (!sd)
5817 break;
5818 }
5819}
5820#else /* !CONFIG_SCHED_DEBUG */
5821# define sched_domain_debug(sd, cpu) do { } while (0)
5822static inline bool sched_debug(void)
5823{
5824 return false;
5825}
5826#endif /* CONFIG_SCHED_DEBUG */
5827
5828static int sd_degenerate(struct sched_domain *sd)
5829{
5830 if (cpumask_weight(sched_domain_span(sd)) == 1)
5831 return 1;
5832
5833 /* Following flags need at least 2 groups */
5834 if (sd->flags & (SD_LOAD_BALANCE |
5835 SD_BALANCE_NEWIDLE |
5836 SD_BALANCE_FORK |
5837 SD_BALANCE_EXEC |
5838 SD_SHARE_CPUPOWER |
5839 SD_SHARE_PKG_RESOURCES)) {
5840 if (sd->groups != sd->groups->next)
5841 return 0;
5842 }
5843
5844 /* Following flags don't use groups */
5845 if (sd->flags & (SD_WAKE_AFFINE))
5846 return 0;
5847
5848 return 1;
5849}
5850
5851static int
5852sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5853{
5854 unsigned long cflags = sd->flags, pflags = parent->flags;
5855
5856 if (sd_degenerate(parent))
5857 return 1;
5858
5859 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5860 return 0;
5861
5862 /* Flags needing groups don't count if only 1 group in parent */
5863 if (parent->groups == parent->groups->next) {
5864 pflags &= ~(SD_LOAD_BALANCE |
5865 SD_BALANCE_NEWIDLE |
5866 SD_BALANCE_FORK |
5867 SD_BALANCE_EXEC |
5868 SD_SHARE_CPUPOWER |
5869 SD_SHARE_PKG_RESOURCES);
5870 if (nr_node_ids == 1)
5871 pflags &= ~SD_SERIALIZE;
5872 }
5873 if (~cflags & pflags)
5874 return 0;
5875
5876 return 1;
5877}
5878
5879static void free_rootdomain(struct rcu_head *rcu)
5880{
5881 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5882
5883 cpupri_cleanup(&rd->cpupri);
5884 free_cpumask_var(rd->rto_mask);
5885 free_cpumask_var(rd->online);
5886 free_cpumask_var(rd->span);
5887 kfree(rd);
5888}
5889
5890static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5891{
5892 struct root_domain *old_rd = NULL;
5893 unsigned long flags;
5894
5895 raw_spin_lock_irqsave(&rq->lock, flags);
5896
5897 if (rq->rd) {
5898 old_rd = rq->rd;
5899
5900 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5901 set_rq_offline(rq);
5902
5903 cpumask_clear_cpu(rq->cpu, old_rd->span);
5904
5905 /*
5906 * If we dont want to free the old_rt yet then
5907 * set old_rd to NULL to skip the freeing later
5908 * in this function:
5909 */
5910 if (!atomic_dec_and_test(&old_rd->refcount))
5911 old_rd = NULL;
5912 }
5913
5914 atomic_inc(&rd->refcount);
5915 rq->rd = rd;
5916
5917 cpumask_set_cpu(rq->cpu, rd->span);
5918 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5919 set_rq_online(rq);
5920
5921 raw_spin_unlock_irqrestore(&rq->lock, flags);
5922
5923 if (old_rd)
5924 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5925}
5926
5927static int init_rootdomain(struct root_domain *rd)
5928{
5929 memset(rd, 0, sizeof(*rd));
5930
5931 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5932 goto out;
5933 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5934 goto free_span;
5935 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5936 goto free_online;
5937
5938 if (cpupri_init(&rd->cpupri) != 0)
5939 goto free_rto_mask;
5940 return 0;
5941
5942free_rto_mask:
5943 free_cpumask_var(rd->rto_mask);
5944free_online:
5945 free_cpumask_var(rd->online);
5946free_span:
5947 free_cpumask_var(rd->span);
5948out:
5949 return -ENOMEM;
5950}
5951
5952/*
5953 * By default the system creates a single root-domain with all cpus as
5954 * members (mimicking the global state we have today).
5955 */
5956struct root_domain def_root_domain;
5957
5958static void init_defrootdomain(void)
5959{
5960 init_rootdomain(&def_root_domain);
5961
5962 atomic_set(&def_root_domain.refcount, 1);
5963}
5964
5965static struct root_domain *alloc_rootdomain(void)
5966{
5967 struct root_domain *rd;
5968
5969 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5970 if (!rd)
5971 return NULL;
5972
5973 if (init_rootdomain(rd) != 0) {
5974 kfree(rd);
5975 return NULL;
5976 }
5977
5978 return rd;
5979}
5980
5981static void free_sched_groups(struct sched_group *sg, int free_sgp)
5982{
5983 struct sched_group *tmp, *first;
5984
5985 if (!sg)
5986 return;
5987
5988 first = sg;
5989 do {
5990 tmp = sg->next;
5991
5992 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5993 kfree(sg->sgp);
5994
5995 kfree(sg);
5996 sg = tmp;
5997 } while (sg != first);
5998}
5999
6000static void free_sched_domain(struct rcu_head *rcu)
6001{
6002 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6003
6004 /*
6005 * If its an overlapping domain it has private groups, iterate and
6006 * nuke them all.
6007 */
6008 if (sd->flags & SD_OVERLAP) {
6009 free_sched_groups(sd->groups, 1);
6010 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6011 kfree(sd->groups->sgp);
6012 kfree(sd->groups);
6013 }
6014 kfree(sd);
6015}
6016
6017static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6018{
6019 call_rcu(&sd->rcu, free_sched_domain);
6020}
6021
6022static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6023{
6024 for (; sd; sd = sd->parent)
6025 destroy_sched_domain(sd, cpu);
6026}
6027
6028/*
6029 * Keep a special pointer to the highest sched_domain that has
6030 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6031 * allows us to avoid some pointer chasing select_idle_sibling().
6032 *
6033 * Also keep a unique ID per domain (we use the first cpu number in
6034 * the cpumask of the domain), this allows us to quickly tell if
6035 * two cpus are in the same cache domain, see cpus_share_cache().
6036 */
6037DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6038DEFINE_PER_CPU(int, sd_llc_id);
6039
6040static void update_top_cache_domain(int cpu)
6041{
6042 struct sched_domain *sd;
6043 int id = cpu;
6044
6045 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6046 if (sd)
6047 id = cpumask_first(sched_domain_span(sd));
6048
6049 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6050 per_cpu(sd_llc_id, cpu) = id;
6051}
6052
6053/*
6054 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6055 * hold the hotplug lock.
6056 */
6057static void
6058cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6059{
6060 struct rq *rq = cpu_rq(cpu);
6061 struct sched_domain *tmp;
6062
6063 /* Remove the sched domains which do not contribute to scheduling. */
6064 for (tmp = sd; tmp; ) {
6065 struct sched_domain *parent = tmp->parent;
6066 if (!parent)
6067 break;
6068
6069 if (sd_parent_degenerate(tmp, parent)) {
6070 tmp->parent = parent->parent;
6071 if (parent->parent)
6072 parent->parent->child = tmp;
6073 destroy_sched_domain(parent, cpu);
6074 } else
6075 tmp = tmp->parent;
6076 }
6077
6078 if (sd && sd_degenerate(sd)) {
6079 tmp = sd;
6080 sd = sd->parent;
6081 destroy_sched_domain(tmp, cpu);
6082 if (sd)
6083 sd->child = NULL;
6084 }
6085
6086 sched_domain_debug(sd, cpu);
6087
6088 rq_attach_root(rq, rd);
6089 tmp = rq->sd;
6090 rcu_assign_pointer(rq->sd, sd);
6091 destroy_sched_domains(tmp, cpu);
6092
6093 update_top_cache_domain(cpu);
6094}
6095
6096/* cpus with isolated domains */
6097static cpumask_var_t cpu_isolated_map;
6098
6099/* Setup the mask of cpus configured for isolated domains */
6100static int __init isolated_cpu_setup(char *str)
6101{
6102 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6103 cpulist_parse(str, cpu_isolated_map);
6104 return 1;
6105}
6106
6107__setup("isolcpus=", isolated_cpu_setup);
6108
6109static const struct cpumask *cpu_cpu_mask(int cpu)
6110{
6111 return cpumask_of_node(cpu_to_node(cpu));
6112}
6113
6114struct sd_data {
6115 struct sched_domain **__percpu sd;
6116 struct sched_group **__percpu sg;
6117 struct sched_group_power **__percpu sgp;
6118};
6119
6120struct s_data {
6121 struct sched_domain ** __percpu sd;
6122 struct root_domain *rd;
6123};
6124
6125enum s_alloc {
6126 sa_rootdomain,
6127 sa_sd,
6128 sa_sd_storage,
6129 sa_none,
6130};
6131
6132struct sched_domain_topology_level;
6133
6134typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6135typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6136
6137#define SDTL_OVERLAP 0x01
6138
6139struct sched_domain_topology_level {
6140 sched_domain_init_f init;
6141 sched_domain_mask_f mask;
6142 int flags;
6143 int numa_level;
6144 struct sd_data data;
6145};
6146
6147/*
6148 * Build an iteration mask that can exclude certain CPUs from the upwards
6149 * domain traversal.
6150 *
6151 * Asymmetric node setups can result in situations where the domain tree is of
6152 * unequal depth, make sure to skip domains that already cover the entire
6153 * range.
6154 *
6155 * In that case build_sched_domains() will have terminated the iteration early
6156 * and our sibling sd spans will be empty. Domains should always include the
6157 * cpu they're built on, so check that.
6158 *
6159 */
6160static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6161{
6162 const struct cpumask *span = sched_domain_span(sd);
6163 struct sd_data *sdd = sd->private;
6164 struct sched_domain *sibling;
6165 int i;
6166
6167 for_each_cpu(i, span) {
6168 sibling = *per_cpu_ptr(sdd->sd, i);
6169 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6170 continue;
6171
6172 cpumask_set_cpu(i, sched_group_mask(sg));
6173 }
6174}
6175
6176/*
6177 * Return the canonical balance cpu for this group, this is the first cpu
6178 * of this group that's also in the iteration mask.
6179 */
6180int group_balance_cpu(struct sched_group *sg)
6181{
6182 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6183}
6184
6185static int
6186build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6187{
6188 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6189 const struct cpumask *span = sched_domain_span(sd);
6190 struct cpumask *covered = sched_domains_tmpmask;
6191 struct sd_data *sdd = sd->private;
6192 struct sched_domain *child;
6193 int i;
6194
6195 cpumask_clear(covered);
6196
6197 for_each_cpu(i, span) {
6198 struct cpumask *sg_span;
6199
6200 if (cpumask_test_cpu(i, covered))
6201 continue;
6202
6203 child = *per_cpu_ptr(sdd->sd, i);
6204
6205 /* See the comment near build_group_mask(). */
6206 if (!cpumask_test_cpu(i, sched_domain_span(child)))
6207 continue;
6208
6209 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6210 GFP_KERNEL, cpu_to_node(cpu));
6211
6212 if (!sg)
6213 goto fail;
6214
6215 sg_span = sched_group_cpus(sg);
6216 if (child->child) {
6217 child = child->child;
6218 cpumask_copy(sg_span, sched_domain_span(child));
6219 } else
6220 cpumask_set_cpu(i, sg_span);
6221
6222 cpumask_or(covered, covered, sg_span);
6223
6224 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
6225 if (atomic_inc_return(&sg->sgp->ref) == 1)
6226 build_group_mask(sd, sg);
6227
6228 /*
6229 * Initialize sgp->power such that even if we mess up the
6230 * domains and no possible iteration will get us here, we won't
6231 * die on a /0 trap.
6232 */
6233 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
6234
6235 /*
6236 * Make sure the first group of this domain contains the
6237 * canonical balance cpu. Otherwise the sched_domain iteration
6238 * breaks. See update_sg_lb_stats().
6239 */
6240 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6241 group_balance_cpu(sg) == cpu)
6242 groups = sg;
6243
6244 if (!first)
6245 first = sg;
6246 if (last)
6247 last->next = sg;
6248 last = sg;
6249 last->next = first;
6250 }
6251 sd->groups = groups;
6252
6253 return 0;
6254
6255fail:
6256 free_sched_groups(first, 0);
6257
6258 return -ENOMEM;
6259}
6260
6261static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6262{
6263 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6264 struct sched_domain *child = sd->child;
6265
6266 if (child)
6267 cpu = cpumask_first(sched_domain_span(child));
6268
6269 if (sg) {
6270 *sg = *per_cpu_ptr(sdd->sg, cpu);
6271 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6272 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6273 }
6274
6275 return cpu;
6276}
6277
6278/*
6279 * build_sched_groups will build a circular linked list of the groups
6280 * covered by the given span, and will set each group's ->cpumask correctly,
6281 * and ->cpu_power to 0.
6282 *
6283 * Assumes the sched_domain tree is fully constructed
6284 */
6285static int
6286build_sched_groups(struct sched_domain *sd, int cpu)
6287{
6288 struct sched_group *first = NULL, *last = NULL;
6289 struct sd_data *sdd = sd->private;
6290 const struct cpumask *span = sched_domain_span(sd);
6291 struct cpumask *covered;
6292 int i;
6293
6294 get_group(cpu, sdd, &sd->groups);
6295 atomic_inc(&sd->groups->ref);
6296
6297 if (cpu != cpumask_first(sched_domain_span(sd)))
6298 return 0;
6299
6300 lockdep_assert_held(&sched_domains_mutex);
6301 covered = sched_domains_tmpmask;
6302
6303 cpumask_clear(covered);
6304
6305 for_each_cpu(i, span) {
6306 struct sched_group *sg;
6307 int group = get_group(i, sdd, &sg);
6308 int j;
6309
6310 if (cpumask_test_cpu(i, covered))
6311 continue;
6312
6313 cpumask_clear(sched_group_cpus(sg));
6314 sg->sgp->power = 0;
6315 cpumask_setall(sched_group_mask(sg));
6316
6317 for_each_cpu(j, span) {
6318 if (get_group(j, sdd, NULL) != group)
6319 continue;
6320
6321 cpumask_set_cpu(j, covered);
6322 cpumask_set_cpu(j, sched_group_cpus(sg));
6323 }
6324
6325 if (!first)
6326 first = sg;
6327 if (last)
6328 last->next = sg;
6329 last = sg;
6330 }
6331 last->next = first;
6332
6333 return 0;
6334}
6335
6336/*
6337 * Initialize sched groups cpu_power.
6338 *
6339 * cpu_power indicates the capacity of sched group, which is used while
6340 * distributing the load between different sched groups in a sched domain.
6341 * Typically cpu_power for all the groups in a sched domain will be same unless
6342 * there are asymmetries in the topology. If there are asymmetries, group
6343 * having more cpu_power will pickup more load compared to the group having
6344 * less cpu_power.
6345 */
6346static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6347{
6348 struct sched_group *sg = sd->groups;
6349
6350 WARN_ON(!sd || !sg);
6351
6352 do {
6353 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6354 sg = sg->next;
6355 } while (sg != sd->groups);
6356
6357 if (cpu != group_balance_cpu(sg))
6358 return;
6359
6360 update_group_power(sd, cpu);
6361 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6362}
6363
6364int __weak arch_sd_sibling_asym_packing(void)
6365{
6366 return 0*SD_ASYM_PACKING;
6367}
6368
6369/*
6370 * Initializers for schedule domains
6371 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6372 */
6373
6374#ifdef CONFIG_SCHED_DEBUG
6375# define SD_INIT_NAME(sd, type) sd->name = #type
6376#else
6377# define SD_INIT_NAME(sd, type) do { } while (0)
6378#endif
6379
6380#define SD_INIT_FUNC(type) \
6381static noinline struct sched_domain * \
6382sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6383{ \
6384 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6385 *sd = SD_##type##_INIT; \
6386 SD_INIT_NAME(sd, type); \
6387 sd->private = &tl->data; \
6388 return sd; \
6389}
6390
6391SD_INIT_FUNC(CPU)
6392#ifdef CONFIG_SCHED_SMT
6393 SD_INIT_FUNC(SIBLING)
6394#endif
6395#ifdef CONFIG_SCHED_MC
6396 SD_INIT_FUNC(MC)
6397#endif
6398#ifdef CONFIG_SCHED_BOOK
6399 SD_INIT_FUNC(BOOK)
6400#endif
6401
6402static int default_relax_domain_level = -1;
6403int sched_domain_level_max;
6404
6405static int __init setup_relax_domain_level(char *str)
6406{
6407 if (kstrtoint(str, 0, &default_relax_domain_level))
6408 pr_warn("Unable to set relax_domain_level\n");
6409
6410 return 1;
6411}
6412__setup("relax_domain_level=", setup_relax_domain_level);
6413
6414static void set_domain_attribute(struct sched_domain *sd,
6415 struct sched_domain_attr *attr)
6416{
6417 int request;
6418
6419 if (!attr || attr->relax_domain_level < 0) {
6420 if (default_relax_domain_level < 0)
6421 return;
6422 else
6423 request = default_relax_domain_level;
6424 } else
6425 request = attr->relax_domain_level;
6426 if (request < sd->level) {
6427 /* turn off idle balance on this domain */
6428 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6429 } else {
6430 /* turn on idle balance on this domain */
6431 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6432 }
6433}
6434
6435static void __sdt_free(const struct cpumask *cpu_map);
6436static int __sdt_alloc(const struct cpumask *cpu_map);
6437
6438static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6439 const struct cpumask *cpu_map)
6440{
6441 switch (what) {
6442 case sa_rootdomain:
6443 if (!atomic_read(&d->rd->refcount))
6444 free_rootdomain(&d->rd->rcu); /* fall through */
6445 case sa_sd:
6446 free_percpu(d->sd); /* fall through */
6447 case sa_sd_storage:
6448 __sdt_free(cpu_map); /* fall through */
6449 case sa_none:
6450 break;
6451 }
6452}
6453
6454static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6455 const struct cpumask *cpu_map)
6456{
6457 memset(d, 0, sizeof(*d));
6458
6459 if (__sdt_alloc(cpu_map))
6460 return sa_sd_storage;
6461 d->sd = alloc_percpu(struct sched_domain *);
6462 if (!d->sd)
6463 return sa_sd_storage;
6464 d->rd = alloc_rootdomain();
6465 if (!d->rd)
6466 return sa_sd;
6467 return sa_rootdomain;
6468}
6469
6470/*
6471 * NULL the sd_data elements we've used to build the sched_domain and
6472 * sched_group structure so that the subsequent __free_domain_allocs()
6473 * will not free the data we're using.
6474 */
6475static void claim_allocations(int cpu, struct sched_domain *sd)
6476{
6477 struct sd_data *sdd = sd->private;
6478
6479 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6480 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6481
6482 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6483 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6484
6485 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6486 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6487}
6488
6489#ifdef CONFIG_SCHED_SMT
6490static const struct cpumask *cpu_smt_mask(int cpu)
6491{
6492 return topology_thread_cpumask(cpu);
6493}
6494#endif
6495
6496/*
6497 * Topology list, bottom-up.
6498 */
6499static struct sched_domain_topology_level default_topology[] = {
6500#ifdef CONFIG_SCHED_SMT
6501 { sd_init_SIBLING, cpu_smt_mask, },
6502#endif
6503#ifdef CONFIG_SCHED_MC
6504 { sd_init_MC, cpu_coregroup_mask, },
6505#endif
6506#ifdef CONFIG_SCHED_BOOK
6507 { sd_init_BOOK, cpu_book_mask, },
6508#endif
6509 { sd_init_CPU, cpu_cpu_mask, },
6510 { NULL, },
6511};
6512
6513static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6514
6515#ifdef CONFIG_NUMA
6516
6517static int sched_domains_numa_levels;
6518static int *sched_domains_numa_distance;
6519static struct cpumask ***sched_domains_numa_masks;
6520static int sched_domains_curr_level;
6521
6522static inline int sd_local_flags(int level)
6523{
6524 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6525 return 0;
6526
6527 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6528}
6529
6530static struct sched_domain *
6531sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6532{
6533 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6534 int level = tl->numa_level;
6535 int sd_weight = cpumask_weight(
6536 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6537
6538 *sd = (struct sched_domain){
6539 .min_interval = sd_weight,
6540 .max_interval = 2*sd_weight,
6541 .busy_factor = 32,
6542 .imbalance_pct = 125,
6543 .cache_nice_tries = 2,
6544 .busy_idx = 3,
6545 .idle_idx = 2,
6546 .newidle_idx = 0,
6547 .wake_idx = 0,
6548 .forkexec_idx = 0,
6549
6550 .flags = 1*SD_LOAD_BALANCE
6551 | 1*SD_BALANCE_NEWIDLE
6552 | 0*SD_BALANCE_EXEC
6553 | 0*SD_BALANCE_FORK
6554 | 0*SD_BALANCE_WAKE
6555 | 0*SD_WAKE_AFFINE
6556 | 0*SD_PREFER_LOCAL
6557 | 0*SD_SHARE_CPUPOWER
6558 | 0*SD_SHARE_PKG_RESOURCES
6559 | 1*SD_SERIALIZE
6560 | 0*SD_PREFER_SIBLING
6561 | sd_local_flags(level)
6562 ,
6563 .last_balance = jiffies,
6564 .balance_interval = sd_weight,
6565 };
6566 SD_INIT_NAME(sd, NUMA);
6567 sd->private = &tl->data;
6568
6569 /*
6570 * Ugly hack to pass state to sd_numa_mask()...
6571 */
6572 sched_domains_curr_level = tl->numa_level;
6573
6574 return sd;
6575}
6576
6577static const struct cpumask *sd_numa_mask(int cpu)
6578{
6579 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6580}
6581
6582static void sched_numa_warn(const char *str)
6583{
6584 static int done = false;
6585 int i,j;
6586
6587 if (done)
6588 return;
6589
6590 done = true;
6591
6592 printk(KERN_WARNING "ERROR: %s\n\n", str);
6593
6594 for (i = 0; i < nr_node_ids; i++) {
6595 printk(KERN_WARNING " ");
6596 for (j = 0; j < nr_node_ids; j++)
6597 printk(KERN_CONT "%02d ", node_distance(i,j));
6598 printk(KERN_CONT "\n");
6599 }
6600 printk(KERN_WARNING "\n");
6601}
6602
6603static bool find_numa_distance(int distance)
6604{
6605 int i;
6606
6607 if (distance == node_distance(0, 0))
6608 return true;
6609
6610 for (i = 0; i < sched_domains_numa_levels; i++) {
6611 if (sched_domains_numa_distance[i] == distance)
6612 return true;
6613 }
6614
6615 return false;
6616}
6617
6618static void sched_init_numa(void)
6619{
6620 int next_distance, curr_distance = node_distance(0, 0);
6621 struct sched_domain_topology_level *tl;
6622 int level = 0;
6623 int i, j, k;
6624
6625 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6626 if (!sched_domains_numa_distance)
6627 return;
6628
6629 /*
6630 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6631 * unique distances in the node_distance() table.
6632 *
6633 * Assumes node_distance(0,j) includes all distances in
6634 * node_distance(i,j) in order to avoid cubic time.
6635 */
6636 next_distance = curr_distance;
6637 for (i = 0; i < nr_node_ids; i++) {
6638 for (j = 0; j < nr_node_ids; j++) {
6639 for (k = 0; k < nr_node_ids; k++) {
6640 int distance = node_distance(i, k);
6641
6642 if (distance > curr_distance &&
6643 (distance < next_distance ||
6644 next_distance == curr_distance))
6645 next_distance = distance;
6646
6647 /*
6648 * While not a strong assumption it would be nice to know
6649 * about cases where if node A is connected to B, B is not
6650 * equally connected to A.
6651 */
6652 if (sched_debug() && node_distance(k, i) != distance)
6653 sched_numa_warn("Node-distance not symmetric");
6654
6655 if (sched_debug() && i && !find_numa_distance(distance))
6656 sched_numa_warn("Node-0 not representative");
6657 }
6658 if (next_distance != curr_distance) {
6659 sched_domains_numa_distance[level++] = next_distance;
6660 sched_domains_numa_levels = level;
6661 curr_distance = next_distance;
6662 } else break;
6663 }
6664
6665 /*
6666 * In case of sched_debug() we verify the above assumption.
6667 */
6668 if (!sched_debug())
6669 break;
6670 }
6671 /*
6672 * 'level' contains the number of unique distances, excluding the
6673 * identity distance node_distance(i,i).
6674 *
6675 * The sched_domains_nume_distance[] array includes the actual distance
6676 * numbers.
6677 */
6678
6679 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6680 if (!sched_domains_numa_masks)
6681 return;
6682
6683 /*
6684 * Now for each level, construct a mask per node which contains all
6685 * cpus of nodes that are that many hops away from us.
6686 */
6687 for (i = 0; i < level; i++) {
6688 sched_domains_numa_masks[i] =
6689 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6690 if (!sched_domains_numa_masks[i])
6691 return;
6692
6693 for (j = 0; j < nr_node_ids; j++) {
6694 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6695 if (!mask)
6696 return;
6697
6698 sched_domains_numa_masks[i][j] = mask;
6699
6700 for (k = 0; k < nr_node_ids; k++) {
6701 if (node_distance(j, k) > sched_domains_numa_distance[i])
6702 continue;
6703
6704 cpumask_or(mask, mask, cpumask_of_node(k));
6705 }
6706 }
6707 }
6708
6709 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6710 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6711 if (!tl)
6712 return;
6713
6714 /*
6715 * Copy the default topology bits..
6716 */
6717 for (i = 0; default_topology[i].init; i++)
6718 tl[i] = default_topology[i];
6719
6720 /*
6721 * .. and append 'j' levels of NUMA goodness.
6722 */
6723 for (j = 0; j < level; i++, j++) {
6724 tl[i] = (struct sched_domain_topology_level){
6725 .init = sd_numa_init,
6726 .mask = sd_numa_mask,
6727 .flags = SDTL_OVERLAP,
6728 .numa_level = j,
6729 };
6730 }
6731
6732 sched_domain_topology = tl;
6733}
6734#else
6735static inline void sched_init_numa(void)
6736{
6737}
6738#endif /* CONFIG_NUMA */
6739
6740static int __sdt_alloc(const struct cpumask *cpu_map)
6741{
6742 struct sched_domain_topology_level *tl;
6743 int j;
6744
6745 for (tl = sched_domain_topology; tl->init; tl++) {
6746 struct sd_data *sdd = &tl->data;
6747
6748 sdd->sd = alloc_percpu(struct sched_domain *);
6749 if (!sdd->sd)
6750 return -ENOMEM;
6751
6752 sdd->sg = alloc_percpu(struct sched_group *);
6753 if (!sdd->sg)
6754 return -ENOMEM;
6755
6756 sdd->sgp = alloc_percpu(struct sched_group_power *);
6757 if (!sdd->sgp)
6758 return -ENOMEM;
6759
6760 for_each_cpu(j, cpu_map) {
6761 struct sched_domain *sd;
6762 struct sched_group *sg;
6763 struct sched_group_power *sgp;
6764
6765 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6766 GFP_KERNEL, cpu_to_node(j));
6767 if (!sd)
6768 return -ENOMEM;
6769
6770 *per_cpu_ptr(sdd->sd, j) = sd;
6771
6772 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6773 GFP_KERNEL, cpu_to_node(j));
6774 if (!sg)
6775 return -ENOMEM;
6776
6777 sg->next = sg;
6778
6779 *per_cpu_ptr(sdd->sg, j) = sg;
6780
6781 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6782 GFP_KERNEL, cpu_to_node(j));
6783 if (!sgp)
6784 return -ENOMEM;
6785
6786 *per_cpu_ptr(sdd->sgp, j) = sgp;
6787 }
6788 }
6789
6790 return 0;
6791}
6792
6793static void __sdt_free(const struct cpumask *cpu_map)
6794{
6795 struct sched_domain_topology_level *tl;
6796 int j;
6797
6798 for (tl = sched_domain_topology; tl->init; tl++) {
6799 struct sd_data *sdd = &tl->data;
6800
6801 for_each_cpu(j, cpu_map) {
6802 struct sched_domain *sd;
6803
6804 if (sdd->sd) {
6805 sd = *per_cpu_ptr(sdd->sd, j);
6806 if (sd && (sd->flags & SD_OVERLAP))
6807 free_sched_groups(sd->groups, 0);
6808 kfree(*per_cpu_ptr(sdd->sd, j));
6809 }
6810
6811 if (sdd->sg)
6812 kfree(*per_cpu_ptr(sdd->sg, j));
6813 if (sdd->sgp)
6814 kfree(*per_cpu_ptr(sdd->sgp, j));
6815 }
6816 free_percpu(sdd->sd);
6817 sdd->sd = NULL;
6818 free_percpu(sdd->sg);
6819 sdd->sg = NULL;
6820 free_percpu(sdd->sgp);
6821 sdd->sgp = NULL;
6822 }
6823}
6824
6825struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6826 struct s_data *d, const struct cpumask *cpu_map,
6827 struct sched_domain_attr *attr, struct sched_domain *child,
6828 int cpu)
6829{
6830 struct sched_domain *sd = tl->init(tl, cpu);
6831 if (!sd)
6832 return child;
6833
6834 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6835 if (child) {
6836 sd->level = child->level + 1;
6837 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6838 child->parent = sd;
6839 }
6840 sd->child = child;
6841 set_domain_attribute(sd, attr);
6842
6843 return sd;
6844}
6845
6846/*
6847 * Build sched domains for a given set of cpus and attach the sched domains
6848 * to the individual cpus
6849 */
6850static int build_sched_domains(const struct cpumask *cpu_map,
6851 struct sched_domain_attr *attr)
6852{
6853 enum s_alloc alloc_state = sa_none;
6854 struct sched_domain *sd;
6855 struct s_data d;
6856 int i, ret = -ENOMEM;
6857
6858 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6859 if (alloc_state != sa_rootdomain)
6860 goto error;
6861
6862 /* Set up domains for cpus specified by the cpu_map. */
6863 for_each_cpu(i, cpu_map) {
6864 struct sched_domain_topology_level *tl;
6865
6866 sd = NULL;
6867 for (tl = sched_domain_topology; tl->init; tl++) {
6868 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6869 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6870 sd->flags |= SD_OVERLAP;
6871 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6872 break;
6873 }
6874
6875 while (sd->child)
6876 sd = sd->child;
6877
6878 *per_cpu_ptr(d.sd, i) = sd;
6879 }
6880
6881 /* Build the groups for the domains */
6882 for_each_cpu(i, cpu_map) {
6883 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6884 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6885 if (sd->flags & SD_OVERLAP) {
6886 if (build_overlap_sched_groups(sd, i))
6887 goto error;
6888 } else {
6889 if (build_sched_groups(sd, i))
6890 goto error;
6891 }
6892 }
6893 }
6894
6895 /* Calculate CPU power for physical packages and nodes */
6896 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6897 if (!cpumask_test_cpu(i, cpu_map))
6898 continue;
6899
6900 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6901 claim_allocations(i, sd);
6902 init_sched_groups_power(i, sd);
6903 }
6904 }
6905
6906 /* Attach the domains */
6907 rcu_read_lock();
6908 for_each_cpu(i, cpu_map) {
6909 sd = *per_cpu_ptr(d.sd, i);
6910 cpu_attach_domain(sd, d.rd, i);
6911 }
6912 rcu_read_unlock();
6913
6914 ret = 0;
6915error:
6916 __free_domain_allocs(&d, alloc_state, cpu_map);
6917 return ret;
6918}
6919
6920static cpumask_var_t *doms_cur; /* current sched domains */
6921static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6922static struct sched_domain_attr *dattr_cur;
6923 /* attribues of custom domains in 'doms_cur' */
6924
6925/*
6926 * Special case: If a kmalloc of a doms_cur partition (array of
6927 * cpumask) fails, then fallback to a single sched domain,
6928 * as determined by the single cpumask fallback_doms.
6929 */
6930static cpumask_var_t fallback_doms;
6931
6932/*
6933 * arch_update_cpu_topology lets virtualized architectures update the
6934 * cpu core maps. It is supposed to return 1 if the topology changed
6935 * or 0 if it stayed the same.
6936 */
6937int __attribute__((weak)) arch_update_cpu_topology(void)
6938{
6939 return 0;
6940}
6941
6942cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6943{
6944 int i;
6945 cpumask_var_t *doms;
6946
6947 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6948 if (!doms)
6949 return NULL;
6950 for (i = 0; i < ndoms; i++) {
6951 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6952 free_sched_domains(doms, i);
6953 return NULL;
6954 }
6955 }
6956 return doms;
6957}
6958
6959void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6960{
6961 unsigned int i;
6962 for (i = 0; i < ndoms; i++)
6963 free_cpumask_var(doms[i]);
6964 kfree(doms);
6965}
6966
6967/*
6968 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6969 * For now this just excludes isolated cpus, but could be used to
6970 * exclude other special cases in the future.
6971 */
6972static int init_sched_domains(const struct cpumask *cpu_map)
6973{
6974 int err;
6975
6976 arch_update_cpu_topology();
6977 ndoms_cur = 1;
6978 doms_cur = alloc_sched_domains(ndoms_cur);
6979 if (!doms_cur)
6980 doms_cur = &fallback_doms;
6981 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6982 err = build_sched_domains(doms_cur[0], NULL);
6983 register_sched_domain_sysctl();
6984
6985 return err;
6986}
6987
6988/*
6989 * Detach sched domains from a group of cpus specified in cpu_map
6990 * These cpus will now be attached to the NULL domain
6991 */
6992static void detach_destroy_domains(const struct cpumask *cpu_map)
6993{
6994 int i;
6995
6996 rcu_read_lock();
6997 for_each_cpu(i, cpu_map)
6998 cpu_attach_domain(NULL, &def_root_domain, i);
6999 rcu_read_unlock();
7000}
7001
7002/* handle null as "default" */
7003static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7004 struct sched_domain_attr *new, int idx_new)
7005{
7006 struct sched_domain_attr tmp;
7007
7008 /* fast path */
7009 if (!new && !cur)
7010 return 1;
7011
7012 tmp = SD_ATTR_INIT;
7013 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7014 new ? (new + idx_new) : &tmp,
7015 sizeof(struct sched_domain_attr));
7016}
7017
7018/*
7019 * Partition sched domains as specified by the 'ndoms_new'
7020 * cpumasks in the array doms_new[] of cpumasks. This compares
7021 * doms_new[] to the current sched domain partitioning, doms_cur[].
7022 * It destroys each deleted domain and builds each new domain.
7023 *
7024 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7025 * The masks don't intersect (don't overlap.) We should setup one
7026 * sched domain for each mask. CPUs not in any of the cpumasks will
7027 * not be load balanced. If the same cpumask appears both in the
7028 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7029 * it as it is.
7030 *
7031 * The passed in 'doms_new' should be allocated using
7032 * alloc_sched_domains. This routine takes ownership of it and will
7033 * free_sched_domains it when done with it. If the caller failed the
7034 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7035 * and partition_sched_domains() will fallback to the single partition
7036 * 'fallback_doms', it also forces the domains to be rebuilt.
7037 *
7038 * If doms_new == NULL it will be replaced with cpu_online_mask.
7039 * ndoms_new == 0 is a special case for destroying existing domains,
7040 * and it will not create the default domain.
7041 *
7042 * Call with hotplug lock held
7043 */
7044void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7045 struct sched_domain_attr *dattr_new)
7046{
7047 int i, j, n;
7048 int new_topology;
7049
7050 mutex_lock(&sched_domains_mutex);
7051
7052 /* always unregister in case we don't destroy any domains */
7053 unregister_sched_domain_sysctl();
7054
7055 /* Let architecture update cpu core mappings. */
7056 new_topology = arch_update_cpu_topology();
7057
7058 n = doms_new ? ndoms_new : 0;
7059
7060 /* Destroy deleted domains */
7061 for (i = 0; i < ndoms_cur; i++) {
7062 for (j = 0; j < n && !new_topology; j++) {
7063 if (cpumask_equal(doms_cur[i], doms_new[j])
7064 && dattrs_equal(dattr_cur, i, dattr_new, j))
7065 goto match1;
7066 }
7067 /* no match - a current sched domain not in new doms_new[] */
7068 detach_destroy_domains(doms_cur[i]);
7069match1:
7070 ;
7071 }
7072
7073 if (doms_new == NULL) {
7074 ndoms_cur = 0;
7075 doms_new = &fallback_doms;
7076 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7077 WARN_ON_ONCE(dattr_new);
7078 }
7079
7080 /* Build new domains */
7081 for (i = 0; i < ndoms_new; i++) {
7082 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7083 if (cpumask_equal(doms_new[i], doms_cur[j])
7084 && dattrs_equal(dattr_new, i, dattr_cur, j))
7085 goto match2;
7086 }
7087 /* no match - add a new doms_new */
7088 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7089match2:
7090 ;
7091 }
7092
7093 /* Remember the new sched domains */
7094 if (doms_cur != &fallback_doms)
7095 free_sched_domains(doms_cur, ndoms_cur);
7096 kfree(dattr_cur); /* kfree(NULL) is safe */
7097 doms_cur = doms_new;
7098 dattr_cur = dattr_new;
7099 ndoms_cur = ndoms_new;
7100
7101 register_sched_domain_sysctl();
7102
7103 mutex_unlock(&sched_domains_mutex);
7104}
7105
7106/*
7107 * Update cpusets according to cpu_active mask. If cpusets are
7108 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7109 * around partition_sched_domains().
7110 */
7111static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7112 void *hcpu)
7113{
7114 switch (action & ~CPU_TASKS_FROZEN) {
7115 case CPU_ONLINE:
7116 case CPU_DOWN_FAILED:
7117 cpuset_update_active_cpus();
7118 return NOTIFY_OK;
7119 default:
7120 return NOTIFY_DONE;
7121 }
7122}
7123
7124static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7125 void *hcpu)
7126{
7127 switch (action & ~CPU_TASKS_FROZEN) {
7128 case CPU_DOWN_PREPARE:
7129 cpuset_update_active_cpus();
7130 return NOTIFY_OK;
7131 default:
7132 return NOTIFY_DONE;
7133 }
7134}
7135
7136void __init sched_init_smp(void)
7137{
7138 cpumask_var_t non_isolated_cpus;
7139
7140 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7141 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7142
7143 sched_init_numa();
7144
7145 get_online_cpus();
7146 mutex_lock(&sched_domains_mutex);
7147 init_sched_domains(cpu_active_mask);
7148 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7149 if (cpumask_empty(non_isolated_cpus))
7150 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7151 mutex_unlock(&sched_domains_mutex);
7152 put_online_cpus();
7153
7154 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7155 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7156
7157 /* RT runtime code needs to handle some hotplug events */
7158 hotcpu_notifier(update_runtime, 0);
7159
7160 init_hrtick();
7161
7162 /* Move init over to a non-isolated CPU */
7163 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7164 BUG();
7165 sched_init_granularity();
7166 free_cpumask_var(non_isolated_cpus);
7167
7168 init_sched_rt_class();
7169}
7170#else
7171void __init sched_init_smp(void)
7172{
7173 sched_init_granularity();
7174}
7175#endif /* CONFIG_SMP */
7176
7177const_debug unsigned int sysctl_timer_migration = 1;
7178
7179int in_sched_functions(unsigned long addr)
7180{
7181 return in_lock_functions(addr) ||
7182 (addr >= (unsigned long)__sched_text_start
7183 && addr < (unsigned long)__sched_text_end);
7184}
7185
7186#ifdef CONFIG_CGROUP_SCHED
7187struct task_group root_task_group;
7188LIST_HEAD(task_groups);
7189#endif
7190
7191DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
7192
7193void __init sched_init(void)
7194{
7195 int i, j;
7196 unsigned long alloc_size = 0, ptr;
7197
7198#ifdef CONFIG_FAIR_GROUP_SCHED
7199 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7200#endif
7201#ifdef CONFIG_RT_GROUP_SCHED
7202 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7203#endif
7204#ifdef CONFIG_CPUMASK_OFFSTACK
7205 alloc_size += num_possible_cpus() * cpumask_size();
7206#endif
7207 if (alloc_size) {
7208 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7209
7210#ifdef CONFIG_FAIR_GROUP_SCHED
7211 root_task_group.se = (struct sched_entity **)ptr;
7212 ptr += nr_cpu_ids * sizeof(void **);
7213
7214 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7215 ptr += nr_cpu_ids * sizeof(void **);
7216
7217#endif /* CONFIG_FAIR_GROUP_SCHED */
7218#ifdef CONFIG_RT_GROUP_SCHED
7219 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7220 ptr += nr_cpu_ids * sizeof(void **);
7221
7222 root_task_group.rt_rq = (struct rt_rq **)ptr;
7223 ptr += nr_cpu_ids * sizeof(void **);
7224
7225#endif /* CONFIG_RT_GROUP_SCHED */
7226#ifdef CONFIG_CPUMASK_OFFSTACK
7227 for_each_possible_cpu(i) {
7228 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7229 ptr += cpumask_size();
7230 }
7231#endif /* CONFIG_CPUMASK_OFFSTACK */
7232 }
7233
7234#ifdef CONFIG_SMP
7235 init_defrootdomain();
7236#endif
7237
7238 init_rt_bandwidth(&def_rt_bandwidth,
7239 global_rt_period(), global_rt_runtime());
7240
7241#ifdef CONFIG_RT_GROUP_SCHED
7242 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7243 global_rt_period(), global_rt_runtime());
7244#endif /* CONFIG_RT_GROUP_SCHED */
7245
7246#ifdef CONFIG_CGROUP_SCHED
7247 list_add(&root_task_group.list, &task_groups);
7248 INIT_LIST_HEAD(&root_task_group.children);
7249 INIT_LIST_HEAD(&root_task_group.siblings);
7250 autogroup_init(&init_task);
7251
7252#endif /* CONFIG_CGROUP_SCHED */
7253
7254#ifdef CONFIG_CGROUP_CPUACCT
7255 root_cpuacct.cpustat = &kernel_cpustat;
7256 root_cpuacct.cpuusage = alloc_percpu(u64);
7257 /* Too early, not expected to fail */
7258 BUG_ON(!root_cpuacct.cpuusage);
7259#endif
7260 for_each_possible_cpu(i) {
7261 struct rq *rq;
7262
7263 rq = cpu_rq(i);
7264 raw_spin_lock_init(&rq->lock);
7265 rq->nr_running = 0;
7266 rq->calc_load_active = 0;
7267 rq->calc_load_update = jiffies + LOAD_FREQ;
7268 init_cfs_rq(&rq->cfs);
7269 init_rt_rq(&rq->rt, rq);
7270#ifdef CONFIG_FAIR_GROUP_SCHED
7271 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7272 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7273 /*
7274 * How much cpu bandwidth does root_task_group get?
7275 *
7276 * In case of task-groups formed thr' the cgroup filesystem, it
7277 * gets 100% of the cpu resources in the system. This overall
7278 * system cpu resource is divided among the tasks of
7279 * root_task_group and its child task-groups in a fair manner,
7280 * based on each entity's (task or task-group's) weight
7281 * (se->load.weight).
7282 *
7283 * In other words, if root_task_group has 10 tasks of weight
7284 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7285 * then A0's share of the cpu resource is:
7286 *
7287 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7288 *
7289 * We achieve this by letting root_task_group's tasks sit
7290 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7291 */
7292 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7293 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7294#endif /* CONFIG_FAIR_GROUP_SCHED */
7295
7296 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7297#ifdef CONFIG_RT_GROUP_SCHED
7298 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7299 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7300#endif
7301
7302 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7303 rq->cpu_load[j] = 0;
7304
7305 rq->last_load_update_tick = jiffies;
7306
7307#ifdef CONFIG_SMP
7308 rq->sd = NULL;
7309 rq->rd = NULL;
7310 rq->cpu_power = SCHED_POWER_SCALE;
7311 rq->post_schedule = 0;
7312 rq->active_balance = 0;
7313 rq->next_balance = jiffies;
7314 rq->push_cpu = 0;
7315 rq->cpu = i;
7316 rq->online = 0;
7317 rq->idle_stamp = 0;
7318 rq->avg_idle = 2*sysctl_sched_migration_cost;
7319
7320 INIT_LIST_HEAD(&rq->cfs_tasks);
7321
7322 rq_attach_root(rq, &def_root_domain);
7323#ifdef CONFIG_NO_HZ
7324 rq->nohz_flags = 0;
7325#endif
7326#endif
7327 init_rq_hrtick(rq);
7328 atomic_set(&rq->nr_iowait, 0);
7329 }
7330
7331 set_load_weight(&init_task);
7332
7333#ifdef CONFIG_PREEMPT_NOTIFIERS
7334 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7335#endif
7336
7337#ifdef CONFIG_RT_MUTEXES
7338 plist_head_init(&init_task.pi_waiters);
7339#endif
7340
7341 /*
7342 * The boot idle thread does lazy MMU switching as well:
7343 */
7344 atomic_inc(&init_mm.mm_count);
7345 enter_lazy_tlb(&init_mm, current);
7346
7347 /*
7348 * Make us the idle thread. Technically, schedule() should not be
7349 * called from this thread, however somewhere below it might be,
7350 * but because we are the idle thread, we just pick up running again
7351 * when this runqueue becomes "idle".
7352 */
7353 init_idle(current, smp_processor_id());
7354
7355 calc_load_update = jiffies + LOAD_FREQ;
7356
7357 /*
7358 * During early bootup we pretend to be a normal task:
7359 */
7360 current->sched_class = &fair_sched_class;
7361
7362#ifdef CONFIG_SMP
7363 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7364 /* May be allocated at isolcpus cmdline parse time */
7365 if (cpu_isolated_map == NULL)
7366 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7367 idle_thread_set_boot_cpu();
7368#endif
7369 init_sched_fair_class();
7370
7371 scheduler_running = 1;
7372}
7373
7374#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7375static inline int preempt_count_equals(int preempt_offset)
7376{
7377 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7378
7379 return (nested == preempt_offset);
7380}
7381
7382void __might_sleep(const char *file, int line, int preempt_offset)
7383{
7384 static unsigned long prev_jiffy; /* ratelimiting */
7385
7386 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7387 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7388 system_state != SYSTEM_RUNNING || oops_in_progress)
7389 return;
7390 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7391 return;
7392 prev_jiffy = jiffies;
7393
7394 printk(KERN_ERR
7395 "BUG: sleeping function called from invalid context at %s:%d\n",
7396 file, line);
7397 printk(KERN_ERR
7398 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7399 in_atomic(), irqs_disabled(),
7400 current->pid, current->comm);
7401
7402 debug_show_held_locks(current);
7403 if (irqs_disabled())
7404 print_irqtrace_events(current);
7405 dump_stack();
7406}
7407EXPORT_SYMBOL(__might_sleep);
7408#endif
7409
7410#ifdef CONFIG_MAGIC_SYSRQ
7411static void normalize_task(struct rq *rq, struct task_struct *p)
7412{
7413 const struct sched_class *prev_class = p->sched_class;
7414 int old_prio = p->prio;
7415 int on_rq;
7416
7417 on_rq = p->on_rq;
7418 if (on_rq)
7419 dequeue_task(rq, p, 0);
7420 __setscheduler(rq, p, SCHED_NORMAL, 0);
7421 if (on_rq) {
7422 enqueue_task(rq, p, 0);
7423 resched_task(rq->curr);
7424 }
7425
7426 check_class_changed(rq, p, prev_class, old_prio);
7427}
7428
7429void normalize_rt_tasks(void)
7430{
7431 struct task_struct *g, *p;
7432 unsigned long flags;
7433 struct rq *rq;
7434
7435 read_lock_irqsave(&tasklist_lock, flags);
7436 do_each_thread(g, p) {
7437 /*
7438 * Only normalize user tasks:
7439 */
7440 if (!p->mm)
7441 continue;
7442
7443 p->se.exec_start = 0;
7444#ifdef CONFIG_SCHEDSTATS
7445 p->se.statistics.wait_start = 0;
7446 p->se.statistics.sleep_start = 0;
7447 p->se.statistics.block_start = 0;
7448#endif
7449
7450 if (!rt_task(p)) {
7451 /*
7452 * Renice negative nice level userspace
7453 * tasks back to 0:
7454 */
7455 if (TASK_NICE(p) < 0 && p->mm)
7456 set_user_nice(p, 0);
7457 continue;
7458 }
7459
7460 raw_spin_lock(&p->pi_lock);
7461 rq = __task_rq_lock(p);
7462
7463 normalize_task(rq, p);
7464
7465 __task_rq_unlock(rq);
7466 raw_spin_unlock(&p->pi_lock);
7467 } while_each_thread(g, p);
7468
7469 read_unlock_irqrestore(&tasklist_lock, flags);
7470}
7471
7472#endif /* CONFIG_MAGIC_SYSRQ */
7473
7474#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7475/*
7476 * These functions are only useful for the IA64 MCA handling, or kdb.
7477 *
7478 * They can only be called when the whole system has been
7479 * stopped - every CPU needs to be quiescent, and no scheduling
7480 * activity can take place. Using them for anything else would
7481 * be a serious bug, and as a result, they aren't even visible
7482 * under any other configuration.
7483 */
7484
7485/**
7486 * curr_task - return the current task for a given cpu.
7487 * @cpu: the processor in question.
7488 *
7489 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7490 */
7491struct task_struct *curr_task(int cpu)
7492{
7493 return cpu_curr(cpu);
7494}
7495
7496#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7497
7498#ifdef CONFIG_IA64
7499/**
7500 * set_curr_task - set the current task for a given cpu.
7501 * @cpu: the processor in question.
7502 * @p: the task pointer to set.
7503 *
7504 * Description: This function must only be used when non-maskable interrupts
7505 * are serviced on a separate stack. It allows the architecture to switch the
7506 * notion of the current task on a cpu in a non-blocking manner. This function
7507 * must be called with all CPU's synchronized, and interrupts disabled, the
7508 * and caller must save the original value of the current task (see
7509 * curr_task() above) and restore that value before reenabling interrupts and
7510 * re-starting the system.
7511 *
7512 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7513 */
7514void set_curr_task(int cpu, struct task_struct *p)
7515{
7516 cpu_curr(cpu) = p;
7517}
7518
7519#endif
7520
7521#ifdef CONFIG_CGROUP_SCHED
7522/* task_group_lock serializes the addition/removal of task groups */
7523static DEFINE_SPINLOCK(task_group_lock);
7524
7525static void free_sched_group(struct task_group *tg)
7526{
7527 free_fair_sched_group(tg);
7528 free_rt_sched_group(tg);
7529 autogroup_free(tg);
7530 kfree(tg);
7531}
7532
7533/* allocate runqueue etc for a new task group */
7534struct task_group *sched_create_group(struct task_group *parent)
7535{
7536 struct task_group *tg;
7537 unsigned long flags;
7538
7539 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7540 if (!tg)
7541 return ERR_PTR(-ENOMEM);
7542
7543 if (!alloc_fair_sched_group(tg, parent))
7544 goto err;
7545
7546 if (!alloc_rt_sched_group(tg, parent))
7547 goto err;
7548
7549 spin_lock_irqsave(&task_group_lock, flags);
7550 list_add_rcu(&tg->list, &task_groups);
7551
7552 WARN_ON(!parent); /* root should already exist */
7553
7554 tg->parent = parent;
7555 INIT_LIST_HEAD(&tg->children);
7556 list_add_rcu(&tg->siblings, &parent->children);
7557 spin_unlock_irqrestore(&task_group_lock, flags);
7558
7559 return tg;
7560
7561err:
7562 free_sched_group(tg);
7563 return ERR_PTR(-ENOMEM);
7564}
7565
7566/* rcu callback to free various structures associated with a task group */
7567static void free_sched_group_rcu(struct rcu_head *rhp)
7568{
7569 /* now it should be safe to free those cfs_rqs */
7570 free_sched_group(container_of(rhp, struct task_group, rcu));
7571}
7572
7573/* Destroy runqueue etc associated with a task group */
7574void sched_destroy_group(struct task_group *tg)
7575{
7576 unsigned long flags;
7577 int i;
7578
7579 /* end participation in shares distribution */
7580 for_each_possible_cpu(i)
7581 unregister_fair_sched_group(tg, i);
7582
7583 spin_lock_irqsave(&task_group_lock, flags);
7584 list_del_rcu(&tg->list);
7585 list_del_rcu(&tg->siblings);
7586 spin_unlock_irqrestore(&task_group_lock, flags);
7587
7588 /* wait for possible concurrent references to cfs_rqs complete */
7589 call_rcu(&tg->rcu, free_sched_group_rcu);
7590}
7591
7592/* change task's runqueue when it moves between groups.
7593 * The caller of this function should have put the task in its new group
7594 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7595 * reflect its new group.
7596 */
7597void sched_move_task(struct task_struct *tsk)
7598{
7599 struct task_group *tg;
7600 int on_rq, running;
7601 unsigned long flags;
7602 struct rq *rq;
7603
7604 rq = task_rq_lock(tsk, &flags);
7605
7606 running = task_current(rq, tsk);
7607 on_rq = tsk->on_rq;
7608
7609 if (on_rq)
7610 dequeue_task(rq, tsk, 0);
7611 if (unlikely(running))
7612 tsk->sched_class->put_prev_task(rq, tsk);
7613
7614 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7615 lockdep_is_held(&tsk->sighand->siglock)),
7616 struct task_group, css);
7617 tg = autogroup_task_group(tsk, tg);
7618 tsk->sched_task_group = tg;
7619
7620#ifdef CONFIG_FAIR_GROUP_SCHED
7621 if (tsk->sched_class->task_move_group)
7622 tsk->sched_class->task_move_group(tsk, on_rq);
7623 else
7624#endif
7625 set_task_rq(tsk, task_cpu(tsk));
7626
7627 if (unlikely(running))
7628 tsk->sched_class->set_curr_task(rq);
7629 if (on_rq)
7630 enqueue_task(rq, tsk, 0);
7631
7632 task_rq_unlock(rq, tsk, &flags);
7633}
7634#endif /* CONFIG_CGROUP_SCHED */
7635
7636#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7637static unsigned long to_ratio(u64 period, u64 runtime)
7638{
7639 if (runtime == RUNTIME_INF)
7640 return 1ULL << 20;
7641
7642 return div64_u64(runtime << 20, period);
7643}
7644#endif
7645
7646#ifdef CONFIG_RT_GROUP_SCHED
7647/*
7648 * Ensure that the real time constraints are schedulable.
7649 */
7650static DEFINE_MUTEX(rt_constraints_mutex);
7651
7652/* Must be called with tasklist_lock held */
7653static inline int tg_has_rt_tasks(struct task_group *tg)
7654{
7655 struct task_struct *g, *p;
7656
7657 do_each_thread(g, p) {
7658 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7659 return 1;
7660 } while_each_thread(g, p);
7661
7662 return 0;
7663}
7664
7665struct rt_schedulable_data {
7666 struct task_group *tg;
7667 u64 rt_period;
7668 u64 rt_runtime;
7669};
7670
7671static int tg_rt_schedulable(struct task_group *tg, void *data)
7672{
7673 struct rt_schedulable_data *d = data;
7674 struct task_group *child;
7675 unsigned long total, sum = 0;
7676 u64 period, runtime;
7677
7678 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7679 runtime = tg->rt_bandwidth.rt_runtime;
7680
7681 if (tg == d->tg) {
7682 period = d->rt_period;
7683 runtime = d->rt_runtime;
7684 }
7685
7686 /*
7687 * Cannot have more runtime than the period.
7688 */
7689 if (runtime > period && runtime != RUNTIME_INF)
7690 return -EINVAL;
7691
7692 /*
7693 * Ensure we don't starve existing RT tasks.
7694 */
7695 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7696 return -EBUSY;
7697
7698 total = to_ratio(period, runtime);
7699
7700 /*
7701 * Nobody can have more than the global setting allows.
7702 */
7703 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7704 return -EINVAL;
7705
7706 /*
7707 * The sum of our children's runtime should not exceed our own.
7708 */
7709 list_for_each_entry_rcu(child, &tg->children, siblings) {
7710 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7711 runtime = child->rt_bandwidth.rt_runtime;
7712
7713 if (child == d->tg) {
7714 period = d->rt_period;
7715 runtime = d->rt_runtime;
7716 }
7717
7718 sum += to_ratio(period, runtime);
7719 }
7720
7721 if (sum > total)
7722 return -EINVAL;
7723
7724 return 0;
7725}
7726
7727static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7728{
7729 int ret;
7730
7731 struct rt_schedulable_data data = {
7732 .tg = tg,
7733 .rt_period = period,
7734 .rt_runtime = runtime,
7735 };
7736
7737 rcu_read_lock();
7738 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7739 rcu_read_unlock();
7740
7741 return ret;
7742}
7743
7744static int tg_set_rt_bandwidth(struct task_group *tg,
7745 u64 rt_period, u64 rt_runtime)
7746{
7747 int i, err = 0;
7748
7749 mutex_lock(&rt_constraints_mutex);
7750 read_lock(&tasklist_lock);
7751 err = __rt_schedulable(tg, rt_period, rt_runtime);
7752 if (err)
7753 goto unlock;
7754
7755 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7756 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7757 tg->rt_bandwidth.rt_runtime = rt_runtime;
7758
7759 for_each_possible_cpu(i) {
7760 struct rt_rq *rt_rq = tg->rt_rq[i];
7761
7762 raw_spin_lock(&rt_rq->rt_runtime_lock);
7763 rt_rq->rt_runtime = rt_runtime;
7764 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7765 }
7766 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7767unlock:
7768 read_unlock(&tasklist_lock);
7769 mutex_unlock(&rt_constraints_mutex);
7770
7771 return err;
7772}
7773
7774int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7775{
7776 u64 rt_runtime, rt_period;
7777
7778 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7779 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7780 if (rt_runtime_us < 0)
7781 rt_runtime = RUNTIME_INF;
7782
7783 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7784}
7785
7786long sched_group_rt_runtime(struct task_group *tg)
7787{
7788 u64 rt_runtime_us;
7789
7790 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7791 return -1;
7792
7793 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7794 do_div(rt_runtime_us, NSEC_PER_USEC);
7795 return rt_runtime_us;
7796}
7797
7798int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7799{
7800 u64 rt_runtime, rt_period;
7801
7802 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7803 rt_runtime = tg->rt_bandwidth.rt_runtime;
7804
7805 if (rt_period == 0)
7806 return -EINVAL;
7807
7808 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7809}
7810
7811long sched_group_rt_period(struct task_group *tg)
7812{
7813 u64 rt_period_us;
7814
7815 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7816 do_div(rt_period_us, NSEC_PER_USEC);
7817 return rt_period_us;
7818}
7819
7820static int sched_rt_global_constraints(void)
7821{
7822 u64 runtime, period;
7823 int ret = 0;
7824
7825 if (sysctl_sched_rt_period <= 0)
7826 return -EINVAL;
7827
7828 runtime = global_rt_runtime();
7829 period = global_rt_period();
7830
7831 /*
7832 * Sanity check on the sysctl variables.
7833 */
7834 if (runtime > period && runtime != RUNTIME_INF)
7835 return -EINVAL;
7836
7837 mutex_lock(&rt_constraints_mutex);
7838 read_lock(&tasklist_lock);
7839 ret = __rt_schedulable(NULL, 0, 0);
7840 read_unlock(&tasklist_lock);
7841 mutex_unlock(&rt_constraints_mutex);
7842
7843 return ret;
7844}
7845
7846int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7847{
7848 /* Don't accept realtime tasks when there is no way for them to run */
7849 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7850 return 0;
7851
7852 return 1;
7853}
7854
7855#else /* !CONFIG_RT_GROUP_SCHED */
7856static int sched_rt_global_constraints(void)
7857{
7858 unsigned long flags;
7859 int i;
7860
7861 if (sysctl_sched_rt_period <= 0)
7862 return -EINVAL;
7863
7864 /*
7865 * There's always some RT tasks in the root group
7866 * -- migration, kstopmachine etc..
7867 */
7868 if (sysctl_sched_rt_runtime == 0)
7869 return -EBUSY;
7870
7871 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7872 for_each_possible_cpu(i) {
7873 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7874
7875 raw_spin_lock(&rt_rq->rt_runtime_lock);
7876 rt_rq->rt_runtime = global_rt_runtime();
7877 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7878 }
7879 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7880
7881 return 0;
7882}
7883#endif /* CONFIG_RT_GROUP_SCHED */
7884
7885int sched_rt_handler(struct ctl_table *table, int write,
7886 void __user *buffer, size_t *lenp,
7887 loff_t *ppos)
7888{
7889 int ret;
7890 int old_period, old_runtime;
7891 static DEFINE_MUTEX(mutex);
7892
7893 mutex_lock(&mutex);
7894 old_period = sysctl_sched_rt_period;
7895 old_runtime = sysctl_sched_rt_runtime;
7896
7897 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7898
7899 if (!ret && write) {
7900 ret = sched_rt_global_constraints();
7901 if (ret) {
7902 sysctl_sched_rt_period = old_period;
7903 sysctl_sched_rt_runtime = old_runtime;
7904 } else {
7905 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7906 def_rt_bandwidth.rt_period =
7907 ns_to_ktime(global_rt_period());
7908 }
7909 }
7910 mutex_unlock(&mutex);
7911
7912 return ret;
7913}
7914
7915#ifdef CONFIG_CGROUP_SCHED
7916
7917/* return corresponding task_group object of a cgroup */
7918static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7919{
7920 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7921 struct task_group, css);
7922}
7923
7924static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7925{
7926 struct task_group *tg, *parent;
7927
7928 if (!cgrp->parent) {
7929 /* This is early initialization for the top cgroup */
7930 return &root_task_group.css;
7931 }
7932
7933 parent = cgroup_tg(cgrp->parent);
7934 tg = sched_create_group(parent);
7935 if (IS_ERR(tg))
7936 return ERR_PTR(-ENOMEM);
7937
7938 return &tg->css;
7939}
7940
7941static void cpu_cgroup_destroy(struct cgroup *cgrp)
7942{
7943 struct task_group *tg = cgroup_tg(cgrp);
7944
7945 sched_destroy_group(tg);
7946}
7947
7948static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7949 struct cgroup_taskset *tset)
7950{
7951 struct task_struct *task;
7952
7953 cgroup_taskset_for_each(task, cgrp, tset) {
7954#ifdef CONFIG_RT_GROUP_SCHED
7955 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7956 return -EINVAL;
7957#else
7958 /* We don't support RT-tasks being in separate groups */
7959 if (task->sched_class != &fair_sched_class)
7960 return -EINVAL;
7961#endif
7962 }
7963 return 0;
7964}
7965
7966static void cpu_cgroup_attach(struct cgroup *cgrp,
7967 struct cgroup_taskset *tset)
7968{
7969 struct task_struct *task;
7970
7971 cgroup_taskset_for_each(task, cgrp, tset)
7972 sched_move_task(task);
7973}
7974
7975static void
7976cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7977 struct task_struct *task)
7978{
7979 /*
7980 * cgroup_exit() is called in the copy_process() failure path.
7981 * Ignore this case since the task hasn't ran yet, this avoids
7982 * trying to poke a half freed task state from generic code.
7983 */
7984 if (!(task->flags & PF_EXITING))
7985 return;
7986
7987 sched_move_task(task);
7988}
7989
7990#ifdef CONFIG_FAIR_GROUP_SCHED
7991static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7992 u64 shareval)
7993{
7994 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7995}
7996
7997static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7998{
7999 struct task_group *tg = cgroup_tg(cgrp);
8000
8001 return (u64) scale_load_down(tg->shares);
8002}
8003
8004#ifdef CONFIG_CFS_BANDWIDTH
8005static DEFINE_MUTEX(cfs_constraints_mutex);
8006
8007const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8008const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8009
8010static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8011
8012static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8013{
8014 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8015 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8016
8017 if (tg == &root_task_group)
8018 return -EINVAL;
8019
8020 /*
8021 * Ensure we have at some amount of bandwidth every period. This is
8022 * to prevent reaching a state of large arrears when throttled via
8023 * entity_tick() resulting in prolonged exit starvation.
8024 */
8025 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8026 return -EINVAL;
8027
8028 /*
8029 * Likewise, bound things on the otherside by preventing insane quota
8030 * periods. This also allows us to normalize in computing quota
8031 * feasibility.
8032 */
8033 if (period > max_cfs_quota_period)
8034 return -EINVAL;
8035
8036 mutex_lock(&cfs_constraints_mutex);
8037 ret = __cfs_schedulable(tg, period, quota);
8038 if (ret)
8039 goto out_unlock;
8040
8041 runtime_enabled = quota != RUNTIME_INF;
8042 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8043 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
8044 raw_spin_lock_irq(&cfs_b->lock);
8045 cfs_b->period = ns_to_ktime(period);
8046 cfs_b->quota = quota;
8047
8048 __refill_cfs_bandwidth_runtime(cfs_b);
8049 /* restart the period timer (if active) to handle new period expiry */
8050 if (runtime_enabled && cfs_b->timer_active) {
8051 /* force a reprogram */
8052 cfs_b->timer_active = 0;
8053 __start_cfs_bandwidth(cfs_b);
8054 }
8055 raw_spin_unlock_irq(&cfs_b->lock);
8056
8057 for_each_possible_cpu(i) {
8058 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8059 struct rq *rq = cfs_rq->rq;
8060
8061 raw_spin_lock_irq(&rq->lock);
8062 cfs_rq->runtime_enabled = runtime_enabled;
8063 cfs_rq->runtime_remaining = 0;
8064
8065 if (cfs_rq->throttled)
8066 unthrottle_cfs_rq(cfs_rq);
8067 raw_spin_unlock_irq(&rq->lock);
8068 }
8069out_unlock:
8070 mutex_unlock(&cfs_constraints_mutex);
8071
8072 return ret;
8073}
8074
8075int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8076{
8077 u64 quota, period;
8078
8079 period = ktime_to_ns(tg->cfs_bandwidth.period);
8080 if (cfs_quota_us < 0)
8081 quota = RUNTIME_INF;
8082 else
8083 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8084
8085 return tg_set_cfs_bandwidth(tg, period, quota);
8086}
8087
8088long tg_get_cfs_quota(struct task_group *tg)
8089{
8090 u64 quota_us;
8091
8092 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8093 return -1;
8094
8095 quota_us = tg->cfs_bandwidth.quota;
8096 do_div(quota_us, NSEC_PER_USEC);
8097
8098 return quota_us;
8099}
8100
8101int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8102{
8103 u64 quota, period;
8104
8105 period = (u64)cfs_period_us * NSEC_PER_USEC;
8106 quota = tg->cfs_bandwidth.quota;
8107
8108 return tg_set_cfs_bandwidth(tg, period, quota);
8109}
8110
8111long tg_get_cfs_period(struct task_group *tg)
8112{
8113 u64 cfs_period_us;
8114
8115 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8116 do_div(cfs_period_us, NSEC_PER_USEC);
8117
8118 return cfs_period_us;
8119}
8120
8121static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
8122{
8123 return tg_get_cfs_quota(cgroup_tg(cgrp));
8124}
8125
8126static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
8127 s64 cfs_quota_us)
8128{
8129 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
8130}
8131
8132static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
8133{
8134 return tg_get_cfs_period(cgroup_tg(cgrp));
8135}
8136
8137static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8138 u64 cfs_period_us)
8139{
8140 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
8141}
8142
8143struct cfs_schedulable_data {
8144 struct task_group *tg;
8145 u64 period, quota;
8146};
8147
8148/*
8149 * normalize group quota/period to be quota/max_period
8150 * note: units are usecs
8151 */
8152static u64 normalize_cfs_quota(struct task_group *tg,
8153 struct cfs_schedulable_data *d)
8154{
8155 u64 quota, period;
8156
8157 if (tg == d->tg) {
8158 period = d->period;
8159 quota = d->quota;
8160 } else {
8161 period = tg_get_cfs_period(tg);
8162 quota = tg_get_cfs_quota(tg);
8163 }
8164
8165 /* note: these should typically be equivalent */
8166 if (quota == RUNTIME_INF || quota == -1)
8167 return RUNTIME_INF;
8168
8169 return to_ratio(period, quota);
8170}
8171
8172static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8173{
8174 struct cfs_schedulable_data *d = data;
8175 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8176 s64 quota = 0, parent_quota = -1;
8177
8178 if (!tg->parent) {
8179 quota = RUNTIME_INF;
8180 } else {
8181 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8182
8183 quota = normalize_cfs_quota(tg, d);
8184 parent_quota = parent_b->hierarchal_quota;
8185
8186 /*
8187 * ensure max(child_quota) <= parent_quota, inherit when no
8188 * limit is set
8189 */
8190 if (quota == RUNTIME_INF)
8191 quota = parent_quota;
8192 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8193 return -EINVAL;
8194 }
8195 cfs_b->hierarchal_quota = quota;
8196
8197 return 0;
8198}
8199
8200static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8201{
8202 int ret;
8203 struct cfs_schedulable_data data = {
8204 .tg = tg,
8205 .period = period,
8206 .quota = quota,
8207 };
8208
8209 if (quota != RUNTIME_INF) {
8210 do_div(data.period, NSEC_PER_USEC);
8211 do_div(data.quota, NSEC_PER_USEC);
8212 }
8213
8214 rcu_read_lock();
8215 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8216 rcu_read_unlock();
8217
8218 return ret;
8219}
8220
8221static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
8222 struct cgroup_map_cb *cb)
8223{
8224 struct task_group *tg = cgroup_tg(cgrp);
8225 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8226
8227 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
8228 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
8229 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
8230
8231 return 0;
8232}
8233#endif /* CONFIG_CFS_BANDWIDTH */
8234#endif /* CONFIG_FAIR_GROUP_SCHED */
8235
8236#ifdef CONFIG_RT_GROUP_SCHED
8237static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8238 s64 val)
8239{
8240 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8241}
8242
8243static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8244{
8245 return sched_group_rt_runtime(cgroup_tg(cgrp));
8246}
8247
8248static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8249 u64 rt_period_us)
8250{
8251 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8252}
8253
8254static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8255{
8256 return sched_group_rt_period(cgroup_tg(cgrp));
8257}
8258#endif /* CONFIG_RT_GROUP_SCHED */
8259
8260static struct cftype cpu_files[] = {
8261#ifdef CONFIG_FAIR_GROUP_SCHED
8262 {
8263 .name = "shares",
8264 .read_u64 = cpu_shares_read_u64,
8265 .write_u64 = cpu_shares_write_u64,
8266 },
8267#endif
8268#ifdef CONFIG_CFS_BANDWIDTH
8269 {
8270 .name = "cfs_quota_us",
8271 .read_s64 = cpu_cfs_quota_read_s64,
8272 .write_s64 = cpu_cfs_quota_write_s64,
8273 },
8274 {
8275 .name = "cfs_period_us",
8276 .read_u64 = cpu_cfs_period_read_u64,
8277 .write_u64 = cpu_cfs_period_write_u64,
8278 },
8279 {
8280 .name = "stat",
8281 .read_map = cpu_stats_show,
8282 },
8283#endif
8284#ifdef CONFIG_RT_GROUP_SCHED
8285 {
8286 .name = "rt_runtime_us",
8287 .read_s64 = cpu_rt_runtime_read,
8288 .write_s64 = cpu_rt_runtime_write,
8289 },
8290 {
8291 .name = "rt_period_us",
8292 .read_u64 = cpu_rt_period_read_uint,
8293 .write_u64 = cpu_rt_period_write_uint,
8294 },
8295#endif
8296 { } /* terminate */
8297};
8298
8299struct cgroup_subsys cpu_cgroup_subsys = {
8300 .name = "cpu",
8301 .create = cpu_cgroup_create,
8302 .destroy = cpu_cgroup_destroy,
8303 .can_attach = cpu_cgroup_can_attach,
8304 .attach = cpu_cgroup_attach,
8305 .exit = cpu_cgroup_exit,
8306 .subsys_id = cpu_cgroup_subsys_id,
8307 .base_cftypes = cpu_files,
8308 .early_init = 1,
8309};
8310
8311#endif /* CONFIG_CGROUP_SCHED */
8312
8313#ifdef CONFIG_CGROUP_CPUACCT
8314
8315/*
8316 * CPU accounting code for task groups.
8317 *
8318 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8319 * (balbir@in.ibm.com).
8320 */
8321
8322/* create a new cpu accounting group */
8323static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
8324{
8325 struct cpuacct *ca;
8326
8327 if (!cgrp->parent)
8328 return &root_cpuacct.css;
8329
8330 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8331 if (!ca)
8332 goto out;
8333
8334 ca->cpuusage = alloc_percpu(u64);
8335 if (!ca->cpuusage)
8336 goto out_free_ca;
8337
8338 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8339 if (!ca->cpustat)
8340 goto out_free_cpuusage;
8341
8342 return &ca->css;
8343
8344out_free_cpuusage:
8345 free_percpu(ca->cpuusage);
8346out_free_ca:
8347 kfree(ca);
8348out:
8349 return ERR_PTR(-ENOMEM);
8350}
8351
8352/* destroy an existing cpu accounting group */
8353static void cpuacct_destroy(struct cgroup *cgrp)
8354{
8355 struct cpuacct *ca = cgroup_ca(cgrp);
8356
8357 free_percpu(ca->cpustat);
8358 free_percpu(ca->cpuusage);
8359 kfree(ca);
8360}
8361
8362static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8363{
8364 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8365 u64 data;
8366
8367#ifndef CONFIG_64BIT
8368 /*
8369 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8370 */
8371 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8372 data = *cpuusage;
8373 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8374#else
8375 data = *cpuusage;
8376#endif
8377
8378 return data;
8379}
8380
8381static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8382{
8383 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8384
8385#ifndef CONFIG_64BIT
8386 /*
8387 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8388 */
8389 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8390 *cpuusage = val;
8391 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8392#else
8393 *cpuusage = val;
8394#endif
8395}
8396
8397/* return total cpu usage (in nanoseconds) of a group */
8398static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8399{
8400 struct cpuacct *ca = cgroup_ca(cgrp);
8401 u64 totalcpuusage = 0;
8402 int i;
8403
8404 for_each_present_cpu(i)
8405 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8406
8407 return totalcpuusage;
8408}
8409
8410static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8411 u64 reset)
8412{
8413 struct cpuacct *ca = cgroup_ca(cgrp);
8414 int err = 0;
8415 int i;
8416
8417 if (reset) {
8418 err = -EINVAL;
8419 goto out;
8420 }
8421
8422 for_each_present_cpu(i)
8423 cpuacct_cpuusage_write(ca, i, 0);
8424
8425out:
8426 return err;
8427}
8428
8429static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8430 struct seq_file *m)
8431{
8432 struct cpuacct *ca = cgroup_ca(cgroup);
8433 u64 percpu;
8434 int i;
8435
8436 for_each_present_cpu(i) {
8437 percpu = cpuacct_cpuusage_read(ca, i);
8438 seq_printf(m, "%llu ", (unsigned long long) percpu);
8439 }
8440 seq_printf(m, "\n");
8441 return 0;
8442}
8443
8444static const char *cpuacct_stat_desc[] = {
8445 [CPUACCT_STAT_USER] = "user",
8446 [CPUACCT_STAT_SYSTEM] = "system",
8447};
8448
8449static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8450 struct cgroup_map_cb *cb)
8451{
8452 struct cpuacct *ca = cgroup_ca(cgrp);
8453 int cpu;
8454 s64 val = 0;
8455
8456 for_each_online_cpu(cpu) {
8457 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8458 val += kcpustat->cpustat[CPUTIME_USER];
8459 val += kcpustat->cpustat[CPUTIME_NICE];
8460 }
8461 val = cputime64_to_clock_t(val);
8462 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8463
8464 val = 0;
8465 for_each_online_cpu(cpu) {
8466 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8467 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8468 val += kcpustat->cpustat[CPUTIME_IRQ];
8469 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8470 }
8471
8472 val = cputime64_to_clock_t(val);
8473 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8474
8475 return 0;
8476}
8477
8478static struct cftype files[] = {
8479 {
8480 .name = "usage",
8481 .read_u64 = cpuusage_read,
8482 .write_u64 = cpuusage_write,
8483 },
8484 {
8485 .name = "usage_percpu",
8486 .read_seq_string = cpuacct_percpu_seq_read,
8487 },
8488 {
8489 .name = "stat",
8490 .read_map = cpuacct_stats_show,
8491 },
8492 { } /* terminate */
8493};
8494
8495/*
8496 * charge this task's execution time to its accounting group.
8497 *
8498 * called with rq->lock held.
8499 */
8500void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8501{
8502 struct cpuacct *ca;
8503 int cpu;
8504
8505 if (unlikely(!cpuacct_subsys.active))
8506 return;
8507
8508 cpu = task_cpu(tsk);
8509
8510 rcu_read_lock();
8511
8512 ca = task_ca(tsk);
8513
8514 for (; ca; ca = parent_ca(ca)) {
8515 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8516 *cpuusage += cputime;
8517 }
8518
8519 rcu_read_unlock();
8520}
8521
8522struct cgroup_subsys cpuacct_subsys = {
8523 .name = "cpuacct",
8524 .create = cpuacct_create,
8525 .destroy = cpuacct_destroy,
8526 .subsys_id = cpuacct_subsys_id,
8527 .base_cftypes = files,
8528};
8529#endif /* CONFIG_CGROUP_CPUACCT */
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/kasan.h>
30#include <linux/mm.h>
31#include <linux/module.h>
32#include <linux/nmi.h>
33#include <linux/init.h>
34#include <linux/uaccess.h>
35#include <linux/highmem.h>
36#include <asm/mmu_context.h>
37#include <linux/interrupt.h>
38#include <linux/capability.h>
39#include <linux/completion.h>
40#include <linux/kernel_stat.h>
41#include <linux/debug_locks.h>
42#include <linux/perf_event.h>
43#include <linux/security.h>
44#include <linux/notifier.h>
45#include <linux/profile.h>
46#include <linux/freezer.h>
47#include <linux/vmalloc.h>
48#include <linux/blkdev.h>
49#include <linux/delay.h>
50#include <linux/pid_namespace.h>
51#include <linux/smp.h>
52#include <linux/threads.h>
53#include <linux/timer.h>
54#include <linux/rcupdate.h>
55#include <linux/cpu.h>
56#include <linux/cpuset.h>
57#include <linux/percpu.h>
58#include <linux/proc_fs.h>
59#include <linux/seq_file.h>
60#include <linux/sysctl.h>
61#include <linux/syscalls.h>
62#include <linux/times.h>
63#include <linux/tsacct_kern.h>
64#include <linux/kprobes.h>
65#include <linux/delayacct.h>
66#include <linux/unistd.h>
67#include <linux/pagemap.h>
68#include <linux/hrtimer.h>
69#include <linux/tick.h>
70#include <linux/ctype.h>
71#include <linux/ftrace.h>
72#include <linux/slab.h>
73#include <linux/init_task.h>
74#include <linux/context_tracking.h>
75#include <linux/compiler.h>
76#include <linux/frame.h>
77
78#include <asm/switch_to.h>
79#include <asm/tlb.h>
80#include <asm/irq_regs.h>
81#include <asm/mutex.h>
82#ifdef CONFIG_PARAVIRT
83#include <asm/paravirt.h>
84#endif
85
86#include "sched.h"
87#include "../workqueue_internal.h"
88#include "../smpboot.h"
89
90#define CREATE_TRACE_POINTS
91#include <trace/events/sched.h>
92
93DEFINE_MUTEX(sched_domains_mutex);
94DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
95
96static void update_rq_clock_task(struct rq *rq, s64 delta);
97
98void update_rq_clock(struct rq *rq)
99{
100 s64 delta;
101
102 lockdep_assert_held(&rq->lock);
103
104 if (rq->clock_skip_update & RQCF_ACT_SKIP)
105 return;
106
107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
108 if (delta < 0)
109 return;
110 rq->clock += delta;
111 update_rq_clock_task(rq, delta);
112}
113
114/*
115 * Debugging: various feature bits
116 */
117
118#define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
120
121const_debug unsigned int sysctl_sched_features =
122#include "features.h"
123 0;
124
125#undef SCHED_FEAT
126
127/*
128 * Number of tasks to iterate in a single balance run.
129 * Limited because this is done with IRQs disabled.
130 */
131const_debug unsigned int sysctl_sched_nr_migrate = 32;
132
133/*
134 * period over which we average the RT time consumption, measured
135 * in ms.
136 *
137 * default: 1s
138 */
139const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
140
141/*
142 * period over which we measure -rt task cpu usage in us.
143 * default: 1s
144 */
145unsigned int sysctl_sched_rt_period = 1000000;
146
147__read_mostly int scheduler_running;
148
149/*
150 * part of the period that we allow rt tasks to run in us.
151 * default: 0.95s
152 */
153int sysctl_sched_rt_runtime = 950000;
154
155/* cpus with isolated domains */
156cpumask_var_t cpu_isolated_map;
157
158/*
159 * this_rq_lock - lock this runqueue and disable interrupts.
160 */
161static struct rq *this_rq_lock(void)
162 __acquires(rq->lock)
163{
164 struct rq *rq;
165
166 local_irq_disable();
167 rq = this_rq();
168 raw_spin_lock(&rq->lock);
169
170 return rq;
171}
172
173#ifdef CONFIG_SCHED_HRTICK
174/*
175 * Use HR-timers to deliver accurate preemption points.
176 */
177
178static void hrtick_clear(struct rq *rq)
179{
180 if (hrtimer_active(&rq->hrtick_timer))
181 hrtimer_cancel(&rq->hrtick_timer);
182}
183
184/*
185 * High-resolution timer tick.
186 * Runs from hardirq context with interrupts disabled.
187 */
188static enum hrtimer_restart hrtick(struct hrtimer *timer)
189{
190 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
191
192 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
193
194 raw_spin_lock(&rq->lock);
195 update_rq_clock(rq);
196 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
197 raw_spin_unlock(&rq->lock);
198
199 return HRTIMER_NORESTART;
200}
201
202#ifdef CONFIG_SMP
203
204static void __hrtick_restart(struct rq *rq)
205{
206 struct hrtimer *timer = &rq->hrtick_timer;
207
208 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
209}
210
211/*
212 * called from hardirq (IPI) context
213 */
214static void __hrtick_start(void *arg)
215{
216 struct rq *rq = arg;
217
218 raw_spin_lock(&rq->lock);
219 __hrtick_restart(rq);
220 rq->hrtick_csd_pending = 0;
221 raw_spin_unlock(&rq->lock);
222}
223
224/*
225 * Called to set the hrtick timer state.
226 *
227 * called with rq->lock held and irqs disabled
228 */
229void hrtick_start(struct rq *rq, u64 delay)
230{
231 struct hrtimer *timer = &rq->hrtick_timer;
232 ktime_t time;
233 s64 delta;
234
235 /*
236 * Don't schedule slices shorter than 10000ns, that just
237 * doesn't make sense and can cause timer DoS.
238 */
239 delta = max_t(s64, delay, 10000LL);
240 time = ktime_add_ns(timer->base->get_time(), delta);
241
242 hrtimer_set_expires(timer, time);
243
244 if (rq == this_rq()) {
245 __hrtick_restart(rq);
246 } else if (!rq->hrtick_csd_pending) {
247 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
248 rq->hrtick_csd_pending = 1;
249 }
250}
251
252static int
253hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
254{
255 int cpu = (int)(long)hcpu;
256
257 switch (action) {
258 case CPU_UP_CANCELED:
259 case CPU_UP_CANCELED_FROZEN:
260 case CPU_DOWN_PREPARE:
261 case CPU_DOWN_PREPARE_FROZEN:
262 case CPU_DEAD:
263 case CPU_DEAD_FROZEN:
264 hrtick_clear(cpu_rq(cpu));
265 return NOTIFY_OK;
266 }
267
268 return NOTIFY_DONE;
269}
270
271static __init void init_hrtick(void)
272{
273 hotcpu_notifier(hotplug_hrtick, 0);
274}
275#else
276/*
277 * Called to set the hrtick timer state.
278 *
279 * called with rq->lock held and irqs disabled
280 */
281void hrtick_start(struct rq *rq, u64 delay)
282{
283 /*
284 * Don't schedule slices shorter than 10000ns, that just
285 * doesn't make sense. Rely on vruntime for fairness.
286 */
287 delay = max_t(u64, delay, 10000LL);
288 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
289 HRTIMER_MODE_REL_PINNED);
290}
291
292static inline void init_hrtick(void)
293{
294}
295#endif /* CONFIG_SMP */
296
297static void init_rq_hrtick(struct rq *rq)
298{
299#ifdef CONFIG_SMP
300 rq->hrtick_csd_pending = 0;
301
302 rq->hrtick_csd.flags = 0;
303 rq->hrtick_csd.func = __hrtick_start;
304 rq->hrtick_csd.info = rq;
305#endif
306
307 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
308 rq->hrtick_timer.function = hrtick;
309}
310#else /* CONFIG_SCHED_HRTICK */
311static inline void hrtick_clear(struct rq *rq)
312{
313}
314
315static inline void init_rq_hrtick(struct rq *rq)
316{
317}
318
319static inline void init_hrtick(void)
320{
321}
322#endif /* CONFIG_SCHED_HRTICK */
323
324/*
325 * cmpxchg based fetch_or, macro so it works for different integer types
326 */
327#define fetch_or(ptr, mask) \
328 ({ \
329 typeof(ptr) _ptr = (ptr); \
330 typeof(mask) _mask = (mask); \
331 typeof(*_ptr) _old, _val = *_ptr; \
332 \
333 for (;;) { \
334 _old = cmpxchg(_ptr, _val, _val | _mask); \
335 if (_old == _val) \
336 break; \
337 _val = _old; \
338 } \
339 _old; \
340})
341
342#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
343/*
344 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
345 * this avoids any races wrt polling state changes and thereby avoids
346 * spurious IPIs.
347 */
348static bool set_nr_and_not_polling(struct task_struct *p)
349{
350 struct thread_info *ti = task_thread_info(p);
351 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
352}
353
354/*
355 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
356 *
357 * If this returns true, then the idle task promises to call
358 * sched_ttwu_pending() and reschedule soon.
359 */
360static bool set_nr_if_polling(struct task_struct *p)
361{
362 struct thread_info *ti = task_thread_info(p);
363 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
364
365 for (;;) {
366 if (!(val & _TIF_POLLING_NRFLAG))
367 return false;
368 if (val & _TIF_NEED_RESCHED)
369 return true;
370 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
371 if (old == val)
372 break;
373 val = old;
374 }
375 return true;
376}
377
378#else
379static bool set_nr_and_not_polling(struct task_struct *p)
380{
381 set_tsk_need_resched(p);
382 return true;
383}
384
385#ifdef CONFIG_SMP
386static bool set_nr_if_polling(struct task_struct *p)
387{
388 return false;
389}
390#endif
391#endif
392
393void wake_q_add(struct wake_q_head *head, struct task_struct *task)
394{
395 struct wake_q_node *node = &task->wake_q;
396
397 /*
398 * Atomically grab the task, if ->wake_q is !nil already it means
399 * its already queued (either by us or someone else) and will get the
400 * wakeup due to that.
401 *
402 * This cmpxchg() implies a full barrier, which pairs with the write
403 * barrier implied by the wakeup in wake_up_list().
404 */
405 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
406 return;
407
408 get_task_struct(task);
409
410 /*
411 * The head is context local, there can be no concurrency.
412 */
413 *head->lastp = node;
414 head->lastp = &node->next;
415}
416
417void wake_up_q(struct wake_q_head *head)
418{
419 struct wake_q_node *node = head->first;
420
421 while (node != WAKE_Q_TAIL) {
422 struct task_struct *task;
423
424 task = container_of(node, struct task_struct, wake_q);
425 BUG_ON(!task);
426 /* task can safely be re-inserted now */
427 node = node->next;
428 task->wake_q.next = NULL;
429
430 /*
431 * wake_up_process() implies a wmb() to pair with the queueing
432 * in wake_q_add() so as not to miss wakeups.
433 */
434 wake_up_process(task);
435 put_task_struct(task);
436 }
437}
438
439/*
440 * resched_curr - mark rq's current task 'to be rescheduled now'.
441 *
442 * On UP this means the setting of the need_resched flag, on SMP it
443 * might also involve a cross-CPU call to trigger the scheduler on
444 * the target CPU.
445 */
446void resched_curr(struct rq *rq)
447{
448 struct task_struct *curr = rq->curr;
449 int cpu;
450
451 lockdep_assert_held(&rq->lock);
452
453 if (test_tsk_need_resched(curr))
454 return;
455
456 cpu = cpu_of(rq);
457
458 if (cpu == smp_processor_id()) {
459 set_tsk_need_resched(curr);
460 set_preempt_need_resched();
461 return;
462 }
463
464 if (set_nr_and_not_polling(curr))
465 smp_send_reschedule(cpu);
466 else
467 trace_sched_wake_idle_without_ipi(cpu);
468}
469
470void resched_cpu(int cpu)
471{
472 struct rq *rq = cpu_rq(cpu);
473 unsigned long flags;
474
475 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
476 return;
477 resched_curr(rq);
478 raw_spin_unlock_irqrestore(&rq->lock, flags);
479}
480
481#ifdef CONFIG_SMP
482#ifdef CONFIG_NO_HZ_COMMON
483/*
484 * In the semi idle case, use the nearest busy cpu for migrating timers
485 * from an idle cpu. This is good for power-savings.
486 *
487 * We don't do similar optimization for completely idle system, as
488 * selecting an idle cpu will add more delays to the timers than intended
489 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
490 */
491int get_nohz_timer_target(void)
492{
493 int i, cpu = smp_processor_id();
494 struct sched_domain *sd;
495
496 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
497 return cpu;
498
499 rcu_read_lock();
500 for_each_domain(cpu, sd) {
501 for_each_cpu(i, sched_domain_span(sd)) {
502 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
503 cpu = i;
504 goto unlock;
505 }
506 }
507 }
508
509 if (!is_housekeeping_cpu(cpu))
510 cpu = housekeeping_any_cpu();
511unlock:
512 rcu_read_unlock();
513 return cpu;
514}
515/*
516 * When add_timer_on() enqueues a timer into the timer wheel of an
517 * idle CPU then this timer might expire before the next timer event
518 * which is scheduled to wake up that CPU. In case of a completely
519 * idle system the next event might even be infinite time into the
520 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
521 * leaves the inner idle loop so the newly added timer is taken into
522 * account when the CPU goes back to idle and evaluates the timer
523 * wheel for the next timer event.
524 */
525static void wake_up_idle_cpu(int cpu)
526{
527 struct rq *rq = cpu_rq(cpu);
528
529 if (cpu == smp_processor_id())
530 return;
531
532 if (set_nr_and_not_polling(rq->idle))
533 smp_send_reschedule(cpu);
534 else
535 trace_sched_wake_idle_without_ipi(cpu);
536}
537
538static bool wake_up_full_nohz_cpu(int cpu)
539{
540 /*
541 * We just need the target to call irq_exit() and re-evaluate
542 * the next tick. The nohz full kick at least implies that.
543 * If needed we can still optimize that later with an
544 * empty IRQ.
545 */
546 if (tick_nohz_full_cpu(cpu)) {
547 if (cpu != smp_processor_id() ||
548 tick_nohz_tick_stopped())
549 tick_nohz_full_kick_cpu(cpu);
550 return true;
551 }
552
553 return false;
554}
555
556void wake_up_nohz_cpu(int cpu)
557{
558 if (!wake_up_full_nohz_cpu(cpu))
559 wake_up_idle_cpu(cpu);
560}
561
562static inline bool got_nohz_idle_kick(void)
563{
564 int cpu = smp_processor_id();
565
566 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
567 return false;
568
569 if (idle_cpu(cpu) && !need_resched())
570 return true;
571
572 /*
573 * We can't run Idle Load Balance on this CPU for this time so we
574 * cancel it and clear NOHZ_BALANCE_KICK
575 */
576 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
577 return false;
578}
579
580#else /* CONFIG_NO_HZ_COMMON */
581
582static inline bool got_nohz_idle_kick(void)
583{
584 return false;
585}
586
587#endif /* CONFIG_NO_HZ_COMMON */
588
589#ifdef CONFIG_NO_HZ_FULL
590bool sched_can_stop_tick(struct rq *rq)
591{
592 int fifo_nr_running;
593
594 /* Deadline tasks, even if single, need the tick */
595 if (rq->dl.dl_nr_running)
596 return false;
597
598 /*
599 * If there are more than one RR tasks, we need the tick to effect the
600 * actual RR behaviour.
601 */
602 if (rq->rt.rr_nr_running) {
603 if (rq->rt.rr_nr_running == 1)
604 return true;
605 else
606 return false;
607 }
608
609 /*
610 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
611 * forced preemption between FIFO tasks.
612 */
613 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
614 if (fifo_nr_running)
615 return true;
616
617 /*
618 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
619 * if there's more than one we need the tick for involuntary
620 * preemption.
621 */
622 if (rq->nr_running > 1)
623 return false;
624
625 return true;
626}
627#endif /* CONFIG_NO_HZ_FULL */
628
629void sched_avg_update(struct rq *rq)
630{
631 s64 period = sched_avg_period();
632
633 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
634 /*
635 * Inline assembly required to prevent the compiler
636 * optimising this loop into a divmod call.
637 * See __iter_div_u64_rem() for another example of this.
638 */
639 asm("" : "+rm" (rq->age_stamp));
640 rq->age_stamp += period;
641 rq->rt_avg /= 2;
642 }
643}
644
645#endif /* CONFIG_SMP */
646
647#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
648 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
649/*
650 * Iterate task_group tree rooted at *from, calling @down when first entering a
651 * node and @up when leaving it for the final time.
652 *
653 * Caller must hold rcu_lock or sufficient equivalent.
654 */
655int walk_tg_tree_from(struct task_group *from,
656 tg_visitor down, tg_visitor up, void *data)
657{
658 struct task_group *parent, *child;
659 int ret;
660
661 parent = from;
662
663down:
664 ret = (*down)(parent, data);
665 if (ret)
666 goto out;
667 list_for_each_entry_rcu(child, &parent->children, siblings) {
668 parent = child;
669 goto down;
670
671up:
672 continue;
673 }
674 ret = (*up)(parent, data);
675 if (ret || parent == from)
676 goto out;
677
678 child = parent;
679 parent = parent->parent;
680 if (parent)
681 goto up;
682out:
683 return ret;
684}
685
686int tg_nop(struct task_group *tg, void *data)
687{
688 return 0;
689}
690#endif
691
692static void set_load_weight(struct task_struct *p)
693{
694 int prio = p->static_prio - MAX_RT_PRIO;
695 struct load_weight *load = &p->se.load;
696
697 /*
698 * SCHED_IDLE tasks get minimal weight:
699 */
700 if (idle_policy(p->policy)) {
701 load->weight = scale_load(WEIGHT_IDLEPRIO);
702 load->inv_weight = WMULT_IDLEPRIO;
703 return;
704 }
705
706 load->weight = scale_load(sched_prio_to_weight[prio]);
707 load->inv_weight = sched_prio_to_wmult[prio];
708}
709
710static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
711{
712 update_rq_clock(rq);
713 if (!(flags & ENQUEUE_RESTORE))
714 sched_info_queued(rq, p);
715 p->sched_class->enqueue_task(rq, p, flags);
716}
717
718static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
719{
720 update_rq_clock(rq);
721 if (!(flags & DEQUEUE_SAVE))
722 sched_info_dequeued(rq, p);
723 p->sched_class->dequeue_task(rq, p, flags);
724}
725
726void activate_task(struct rq *rq, struct task_struct *p, int flags)
727{
728 if (task_contributes_to_load(p))
729 rq->nr_uninterruptible--;
730
731 enqueue_task(rq, p, flags);
732}
733
734void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
735{
736 if (task_contributes_to_load(p))
737 rq->nr_uninterruptible++;
738
739 dequeue_task(rq, p, flags);
740}
741
742static void update_rq_clock_task(struct rq *rq, s64 delta)
743{
744/*
745 * In theory, the compile should just see 0 here, and optimize out the call
746 * to sched_rt_avg_update. But I don't trust it...
747 */
748#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
749 s64 steal = 0, irq_delta = 0;
750#endif
751#ifdef CONFIG_IRQ_TIME_ACCOUNTING
752 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
753
754 /*
755 * Since irq_time is only updated on {soft,}irq_exit, we might run into
756 * this case when a previous update_rq_clock() happened inside a
757 * {soft,}irq region.
758 *
759 * When this happens, we stop ->clock_task and only update the
760 * prev_irq_time stamp to account for the part that fit, so that a next
761 * update will consume the rest. This ensures ->clock_task is
762 * monotonic.
763 *
764 * It does however cause some slight miss-attribution of {soft,}irq
765 * time, a more accurate solution would be to update the irq_time using
766 * the current rq->clock timestamp, except that would require using
767 * atomic ops.
768 */
769 if (irq_delta > delta)
770 irq_delta = delta;
771
772 rq->prev_irq_time += irq_delta;
773 delta -= irq_delta;
774#endif
775#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
776 if (static_key_false((¶virt_steal_rq_enabled))) {
777 steal = paravirt_steal_clock(cpu_of(rq));
778 steal -= rq->prev_steal_time_rq;
779
780 if (unlikely(steal > delta))
781 steal = delta;
782
783 rq->prev_steal_time_rq += steal;
784 delta -= steal;
785 }
786#endif
787
788 rq->clock_task += delta;
789
790#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
791 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
792 sched_rt_avg_update(rq, irq_delta + steal);
793#endif
794}
795
796void sched_set_stop_task(int cpu, struct task_struct *stop)
797{
798 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
799 struct task_struct *old_stop = cpu_rq(cpu)->stop;
800
801 if (stop) {
802 /*
803 * Make it appear like a SCHED_FIFO task, its something
804 * userspace knows about and won't get confused about.
805 *
806 * Also, it will make PI more or less work without too
807 * much confusion -- but then, stop work should not
808 * rely on PI working anyway.
809 */
810 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
811
812 stop->sched_class = &stop_sched_class;
813 }
814
815 cpu_rq(cpu)->stop = stop;
816
817 if (old_stop) {
818 /*
819 * Reset it back to a normal scheduling class so that
820 * it can die in pieces.
821 */
822 old_stop->sched_class = &rt_sched_class;
823 }
824}
825
826/*
827 * __normal_prio - return the priority that is based on the static prio
828 */
829static inline int __normal_prio(struct task_struct *p)
830{
831 return p->static_prio;
832}
833
834/*
835 * Calculate the expected normal priority: i.e. priority
836 * without taking RT-inheritance into account. Might be
837 * boosted by interactivity modifiers. Changes upon fork,
838 * setprio syscalls, and whenever the interactivity
839 * estimator recalculates.
840 */
841static inline int normal_prio(struct task_struct *p)
842{
843 int prio;
844
845 if (task_has_dl_policy(p))
846 prio = MAX_DL_PRIO-1;
847 else if (task_has_rt_policy(p))
848 prio = MAX_RT_PRIO-1 - p->rt_priority;
849 else
850 prio = __normal_prio(p);
851 return prio;
852}
853
854/*
855 * Calculate the current priority, i.e. the priority
856 * taken into account by the scheduler. This value might
857 * be boosted by RT tasks, or might be boosted by
858 * interactivity modifiers. Will be RT if the task got
859 * RT-boosted. If not then it returns p->normal_prio.
860 */
861static int effective_prio(struct task_struct *p)
862{
863 p->normal_prio = normal_prio(p);
864 /*
865 * If we are RT tasks or we were boosted to RT priority,
866 * keep the priority unchanged. Otherwise, update priority
867 * to the normal priority:
868 */
869 if (!rt_prio(p->prio))
870 return p->normal_prio;
871 return p->prio;
872}
873
874/**
875 * task_curr - is this task currently executing on a CPU?
876 * @p: the task in question.
877 *
878 * Return: 1 if the task is currently executing. 0 otherwise.
879 */
880inline int task_curr(const struct task_struct *p)
881{
882 return cpu_curr(task_cpu(p)) == p;
883}
884
885/*
886 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
887 * use the balance_callback list if you want balancing.
888 *
889 * this means any call to check_class_changed() must be followed by a call to
890 * balance_callback().
891 */
892static inline void check_class_changed(struct rq *rq, struct task_struct *p,
893 const struct sched_class *prev_class,
894 int oldprio)
895{
896 if (prev_class != p->sched_class) {
897 if (prev_class->switched_from)
898 prev_class->switched_from(rq, p);
899
900 p->sched_class->switched_to(rq, p);
901 } else if (oldprio != p->prio || dl_task(p))
902 p->sched_class->prio_changed(rq, p, oldprio);
903}
904
905void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
906{
907 const struct sched_class *class;
908
909 if (p->sched_class == rq->curr->sched_class) {
910 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
911 } else {
912 for_each_class(class) {
913 if (class == rq->curr->sched_class)
914 break;
915 if (class == p->sched_class) {
916 resched_curr(rq);
917 break;
918 }
919 }
920 }
921
922 /*
923 * A queue event has occurred, and we're going to schedule. In
924 * this case, we can save a useless back to back clock update.
925 */
926 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
927 rq_clock_skip_update(rq, true);
928}
929
930#ifdef CONFIG_SMP
931/*
932 * This is how migration works:
933 *
934 * 1) we invoke migration_cpu_stop() on the target CPU using
935 * stop_one_cpu().
936 * 2) stopper starts to run (implicitly forcing the migrated thread
937 * off the CPU)
938 * 3) it checks whether the migrated task is still in the wrong runqueue.
939 * 4) if it's in the wrong runqueue then the migration thread removes
940 * it and puts it into the right queue.
941 * 5) stopper completes and stop_one_cpu() returns and the migration
942 * is done.
943 */
944
945/*
946 * move_queued_task - move a queued task to new rq.
947 *
948 * Returns (locked) new rq. Old rq's lock is released.
949 */
950static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
951{
952 lockdep_assert_held(&rq->lock);
953
954 p->on_rq = TASK_ON_RQ_MIGRATING;
955 dequeue_task(rq, p, 0);
956 set_task_cpu(p, new_cpu);
957 raw_spin_unlock(&rq->lock);
958
959 rq = cpu_rq(new_cpu);
960
961 raw_spin_lock(&rq->lock);
962 BUG_ON(task_cpu(p) != new_cpu);
963 enqueue_task(rq, p, 0);
964 p->on_rq = TASK_ON_RQ_QUEUED;
965 check_preempt_curr(rq, p, 0);
966
967 return rq;
968}
969
970struct migration_arg {
971 struct task_struct *task;
972 int dest_cpu;
973};
974
975/*
976 * Move (not current) task off this cpu, onto dest cpu. We're doing
977 * this because either it can't run here any more (set_cpus_allowed()
978 * away from this CPU, or CPU going down), or because we're
979 * attempting to rebalance this task on exec (sched_exec).
980 *
981 * So we race with normal scheduler movements, but that's OK, as long
982 * as the task is no longer on this CPU.
983 */
984static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
985{
986 if (unlikely(!cpu_active(dest_cpu)))
987 return rq;
988
989 /* Affinity changed (again). */
990 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
991 return rq;
992
993 rq = move_queued_task(rq, p, dest_cpu);
994
995 return rq;
996}
997
998/*
999 * migration_cpu_stop - this will be executed by a highprio stopper thread
1000 * and performs thread migration by bumping thread off CPU then
1001 * 'pushing' onto another runqueue.
1002 */
1003static int migration_cpu_stop(void *data)
1004{
1005 struct migration_arg *arg = data;
1006 struct task_struct *p = arg->task;
1007 struct rq *rq = this_rq();
1008
1009 /*
1010 * The original target cpu might have gone down and we might
1011 * be on another cpu but it doesn't matter.
1012 */
1013 local_irq_disable();
1014 /*
1015 * We need to explicitly wake pending tasks before running
1016 * __migrate_task() such that we will not miss enforcing cpus_allowed
1017 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1018 */
1019 sched_ttwu_pending();
1020
1021 raw_spin_lock(&p->pi_lock);
1022 raw_spin_lock(&rq->lock);
1023 /*
1024 * If task_rq(p) != rq, it cannot be migrated here, because we're
1025 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1026 * we're holding p->pi_lock.
1027 */
1028 if (task_rq(p) == rq && task_on_rq_queued(p))
1029 rq = __migrate_task(rq, p, arg->dest_cpu);
1030 raw_spin_unlock(&rq->lock);
1031 raw_spin_unlock(&p->pi_lock);
1032
1033 local_irq_enable();
1034 return 0;
1035}
1036
1037/*
1038 * sched_class::set_cpus_allowed must do the below, but is not required to
1039 * actually call this function.
1040 */
1041void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1042{
1043 cpumask_copy(&p->cpus_allowed, new_mask);
1044 p->nr_cpus_allowed = cpumask_weight(new_mask);
1045}
1046
1047void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1048{
1049 struct rq *rq = task_rq(p);
1050 bool queued, running;
1051
1052 lockdep_assert_held(&p->pi_lock);
1053
1054 queued = task_on_rq_queued(p);
1055 running = task_current(rq, p);
1056
1057 if (queued) {
1058 /*
1059 * Because __kthread_bind() calls this on blocked tasks without
1060 * holding rq->lock.
1061 */
1062 lockdep_assert_held(&rq->lock);
1063 dequeue_task(rq, p, DEQUEUE_SAVE);
1064 }
1065 if (running)
1066 put_prev_task(rq, p);
1067
1068 p->sched_class->set_cpus_allowed(p, new_mask);
1069
1070 if (running)
1071 p->sched_class->set_curr_task(rq);
1072 if (queued)
1073 enqueue_task(rq, p, ENQUEUE_RESTORE);
1074}
1075
1076/*
1077 * Change a given task's CPU affinity. Migrate the thread to a
1078 * proper CPU and schedule it away if the CPU it's executing on
1079 * is removed from the allowed bitmask.
1080 *
1081 * NOTE: the caller must have a valid reference to the task, the
1082 * task must not exit() & deallocate itself prematurely. The
1083 * call is not atomic; no spinlocks may be held.
1084 */
1085static int __set_cpus_allowed_ptr(struct task_struct *p,
1086 const struct cpumask *new_mask, bool check)
1087{
1088 unsigned long flags;
1089 struct rq *rq;
1090 unsigned int dest_cpu;
1091 int ret = 0;
1092
1093 rq = task_rq_lock(p, &flags);
1094
1095 /*
1096 * Must re-check here, to close a race against __kthread_bind(),
1097 * sched_setaffinity() is not guaranteed to observe the flag.
1098 */
1099 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1100 ret = -EINVAL;
1101 goto out;
1102 }
1103
1104 if (cpumask_equal(&p->cpus_allowed, new_mask))
1105 goto out;
1106
1107 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1108 ret = -EINVAL;
1109 goto out;
1110 }
1111
1112 do_set_cpus_allowed(p, new_mask);
1113
1114 /* Can the task run on the task's current CPU? If so, we're done */
1115 if (cpumask_test_cpu(task_cpu(p), new_mask))
1116 goto out;
1117
1118 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1119 if (task_running(rq, p) || p->state == TASK_WAKING) {
1120 struct migration_arg arg = { p, dest_cpu };
1121 /* Need help from migration thread: drop lock and wait. */
1122 task_rq_unlock(rq, p, &flags);
1123 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1124 tlb_migrate_finish(p->mm);
1125 return 0;
1126 } else if (task_on_rq_queued(p)) {
1127 /*
1128 * OK, since we're going to drop the lock immediately
1129 * afterwards anyway.
1130 */
1131 lockdep_unpin_lock(&rq->lock);
1132 rq = move_queued_task(rq, p, dest_cpu);
1133 lockdep_pin_lock(&rq->lock);
1134 }
1135out:
1136 task_rq_unlock(rq, p, &flags);
1137
1138 return ret;
1139}
1140
1141int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1142{
1143 return __set_cpus_allowed_ptr(p, new_mask, false);
1144}
1145EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1146
1147void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1148{
1149#ifdef CONFIG_SCHED_DEBUG
1150 /*
1151 * We should never call set_task_cpu() on a blocked task,
1152 * ttwu() will sort out the placement.
1153 */
1154 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1155 !p->on_rq);
1156
1157 /*
1158 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1159 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1160 * time relying on p->on_rq.
1161 */
1162 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1163 p->sched_class == &fair_sched_class &&
1164 (p->on_rq && !task_on_rq_migrating(p)));
1165
1166#ifdef CONFIG_LOCKDEP
1167 /*
1168 * The caller should hold either p->pi_lock or rq->lock, when changing
1169 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1170 *
1171 * sched_move_task() holds both and thus holding either pins the cgroup,
1172 * see task_group().
1173 *
1174 * Furthermore, all task_rq users should acquire both locks, see
1175 * task_rq_lock().
1176 */
1177 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1178 lockdep_is_held(&task_rq(p)->lock)));
1179#endif
1180#endif
1181
1182 trace_sched_migrate_task(p, new_cpu);
1183
1184 if (task_cpu(p) != new_cpu) {
1185 if (p->sched_class->migrate_task_rq)
1186 p->sched_class->migrate_task_rq(p);
1187 p->se.nr_migrations++;
1188 perf_event_task_migrate(p);
1189 }
1190
1191 __set_task_cpu(p, new_cpu);
1192}
1193
1194static void __migrate_swap_task(struct task_struct *p, int cpu)
1195{
1196 if (task_on_rq_queued(p)) {
1197 struct rq *src_rq, *dst_rq;
1198
1199 src_rq = task_rq(p);
1200 dst_rq = cpu_rq(cpu);
1201
1202 p->on_rq = TASK_ON_RQ_MIGRATING;
1203 deactivate_task(src_rq, p, 0);
1204 set_task_cpu(p, cpu);
1205 activate_task(dst_rq, p, 0);
1206 p->on_rq = TASK_ON_RQ_QUEUED;
1207 check_preempt_curr(dst_rq, p, 0);
1208 } else {
1209 /*
1210 * Task isn't running anymore; make it appear like we migrated
1211 * it before it went to sleep. This means on wakeup we make the
1212 * previous cpu our targer instead of where it really is.
1213 */
1214 p->wake_cpu = cpu;
1215 }
1216}
1217
1218struct migration_swap_arg {
1219 struct task_struct *src_task, *dst_task;
1220 int src_cpu, dst_cpu;
1221};
1222
1223static int migrate_swap_stop(void *data)
1224{
1225 struct migration_swap_arg *arg = data;
1226 struct rq *src_rq, *dst_rq;
1227 int ret = -EAGAIN;
1228
1229 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1230 return -EAGAIN;
1231
1232 src_rq = cpu_rq(arg->src_cpu);
1233 dst_rq = cpu_rq(arg->dst_cpu);
1234
1235 double_raw_lock(&arg->src_task->pi_lock,
1236 &arg->dst_task->pi_lock);
1237 double_rq_lock(src_rq, dst_rq);
1238
1239 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1240 goto unlock;
1241
1242 if (task_cpu(arg->src_task) != arg->src_cpu)
1243 goto unlock;
1244
1245 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1246 goto unlock;
1247
1248 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1249 goto unlock;
1250
1251 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1252 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1253
1254 ret = 0;
1255
1256unlock:
1257 double_rq_unlock(src_rq, dst_rq);
1258 raw_spin_unlock(&arg->dst_task->pi_lock);
1259 raw_spin_unlock(&arg->src_task->pi_lock);
1260
1261 return ret;
1262}
1263
1264/*
1265 * Cross migrate two tasks
1266 */
1267int migrate_swap(struct task_struct *cur, struct task_struct *p)
1268{
1269 struct migration_swap_arg arg;
1270 int ret = -EINVAL;
1271
1272 arg = (struct migration_swap_arg){
1273 .src_task = cur,
1274 .src_cpu = task_cpu(cur),
1275 .dst_task = p,
1276 .dst_cpu = task_cpu(p),
1277 };
1278
1279 if (arg.src_cpu == arg.dst_cpu)
1280 goto out;
1281
1282 /*
1283 * These three tests are all lockless; this is OK since all of them
1284 * will be re-checked with proper locks held further down the line.
1285 */
1286 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1287 goto out;
1288
1289 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1290 goto out;
1291
1292 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1293 goto out;
1294
1295 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1296 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1297
1298out:
1299 return ret;
1300}
1301
1302/*
1303 * wait_task_inactive - wait for a thread to unschedule.
1304 *
1305 * If @match_state is nonzero, it's the @p->state value just checked and
1306 * not expected to change. If it changes, i.e. @p might have woken up,
1307 * then return zero. When we succeed in waiting for @p to be off its CPU,
1308 * we return a positive number (its total switch count). If a second call
1309 * a short while later returns the same number, the caller can be sure that
1310 * @p has remained unscheduled the whole time.
1311 *
1312 * The caller must ensure that the task *will* unschedule sometime soon,
1313 * else this function might spin for a *long* time. This function can't
1314 * be called with interrupts off, or it may introduce deadlock with
1315 * smp_call_function() if an IPI is sent by the same process we are
1316 * waiting to become inactive.
1317 */
1318unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1319{
1320 unsigned long flags;
1321 int running, queued;
1322 unsigned long ncsw;
1323 struct rq *rq;
1324
1325 for (;;) {
1326 /*
1327 * We do the initial early heuristics without holding
1328 * any task-queue locks at all. We'll only try to get
1329 * the runqueue lock when things look like they will
1330 * work out!
1331 */
1332 rq = task_rq(p);
1333
1334 /*
1335 * If the task is actively running on another CPU
1336 * still, just relax and busy-wait without holding
1337 * any locks.
1338 *
1339 * NOTE! Since we don't hold any locks, it's not
1340 * even sure that "rq" stays as the right runqueue!
1341 * But we don't care, since "task_running()" will
1342 * return false if the runqueue has changed and p
1343 * is actually now running somewhere else!
1344 */
1345 while (task_running(rq, p)) {
1346 if (match_state && unlikely(p->state != match_state))
1347 return 0;
1348 cpu_relax();
1349 }
1350
1351 /*
1352 * Ok, time to look more closely! We need the rq
1353 * lock now, to be *sure*. If we're wrong, we'll
1354 * just go back and repeat.
1355 */
1356 rq = task_rq_lock(p, &flags);
1357 trace_sched_wait_task(p);
1358 running = task_running(rq, p);
1359 queued = task_on_rq_queued(p);
1360 ncsw = 0;
1361 if (!match_state || p->state == match_state)
1362 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1363 task_rq_unlock(rq, p, &flags);
1364
1365 /*
1366 * If it changed from the expected state, bail out now.
1367 */
1368 if (unlikely(!ncsw))
1369 break;
1370
1371 /*
1372 * Was it really running after all now that we
1373 * checked with the proper locks actually held?
1374 *
1375 * Oops. Go back and try again..
1376 */
1377 if (unlikely(running)) {
1378 cpu_relax();
1379 continue;
1380 }
1381
1382 /*
1383 * It's not enough that it's not actively running,
1384 * it must be off the runqueue _entirely_, and not
1385 * preempted!
1386 *
1387 * So if it was still runnable (but just not actively
1388 * running right now), it's preempted, and we should
1389 * yield - it could be a while.
1390 */
1391 if (unlikely(queued)) {
1392 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1393
1394 set_current_state(TASK_UNINTERRUPTIBLE);
1395 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1396 continue;
1397 }
1398
1399 /*
1400 * Ahh, all good. It wasn't running, and it wasn't
1401 * runnable, which means that it will never become
1402 * running in the future either. We're all done!
1403 */
1404 break;
1405 }
1406
1407 return ncsw;
1408}
1409
1410/***
1411 * kick_process - kick a running thread to enter/exit the kernel
1412 * @p: the to-be-kicked thread
1413 *
1414 * Cause a process which is running on another CPU to enter
1415 * kernel-mode, without any delay. (to get signals handled.)
1416 *
1417 * NOTE: this function doesn't have to take the runqueue lock,
1418 * because all it wants to ensure is that the remote task enters
1419 * the kernel. If the IPI races and the task has been migrated
1420 * to another CPU then no harm is done and the purpose has been
1421 * achieved as well.
1422 */
1423void kick_process(struct task_struct *p)
1424{
1425 int cpu;
1426
1427 preempt_disable();
1428 cpu = task_cpu(p);
1429 if ((cpu != smp_processor_id()) && task_curr(p))
1430 smp_send_reschedule(cpu);
1431 preempt_enable();
1432}
1433EXPORT_SYMBOL_GPL(kick_process);
1434
1435/*
1436 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1437 */
1438static int select_fallback_rq(int cpu, struct task_struct *p)
1439{
1440 int nid = cpu_to_node(cpu);
1441 const struct cpumask *nodemask = NULL;
1442 enum { cpuset, possible, fail } state = cpuset;
1443 int dest_cpu;
1444
1445 /*
1446 * If the node that the cpu is on has been offlined, cpu_to_node()
1447 * will return -1. There is no cpu on the node, and we should
1448 * select the cpu on the other node.
1449 */
1450 if (nid != -1) {
1451 nodemask = cpumask_of_node(nid);
1452
1453 /* Look for allowed, online CPU in same node. */
1454 for_each_cpu(dest_cpu, nodemask) {
1455 if (!cpu_online(dest_cpu))
1456 continue;
1457 if (!cpu_active(dest_cpu))
1458 continue;
1459 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1460 return dest_cpu;
1461 }
1462 }
1463
1464 for (;;) {
1465 /* Any allowed, online CPU? */
1466 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1467 if (!cpu_online(dest_cpu))
1468 continue;
1469 if (!cpu_active(dest_cpu))
1470 continue;
1471 goto out;
1472 }
1473
1474 /* No more Mr. Nice Guy. */
1475 switch (state) {
1476 case cpuset:
1477 if (IS_ENABLED(CONFIG_CPUSETS)) {
1478 cpuset_cpus_allowed_fallback(p);
1479 state = possible;
1480 break;
1481 }
1482 /* fall-through */
1483 case possible:
1484 do_set_cpus_allowed(p, cpu_possible_mask);
1485 state = fail;
1486 break;
1487
1488 case fail:
1489 BUG();
1490 break;
1491 }
1492 }
1493
1494out:
1495 if (state != cpuset) {
1496 /*
1497 * Don't tell them about moving exiting tasks or
1498 * kernel threads (both mm NULL), since they never
1499 * leave kernel.
1500 */
1501 if (p->mm && printk_ratelimit()) {
1502 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1503 task_pid_nr(p), p->comm, cpu);
1504 }
1505 }
1506
1507 return dest_cpu;
1508}
1509
1510/*
1511 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1512 */
1513static inline
1514int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1515{
1516 lockdep_assert_held(&p->pi_lock);
1517
1518 if (p->nr_cpus_allowed > 1)
1519 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1520
1521 /*
1522 * In order not to call set_task_cpu() on a blocking task we need
1523 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1524 * cpu.
1525 *
1526 * Since this is common to all placement strategies, this lives here.
1527 *
1528 * [ this allows ->select_task() to simply return task_cpu(p) and
1529 * not worry about this generic constraint ]
1530 */
1531 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1532 !cpu_online(cpu)))
1533 cpu = select_fallback_rq(task_cpu(p), p);
1534
1535 return cpu;
1536}
1537
1538static void update_avg(u64 *avg, u64 sample)
1539{
1540 s64 diff = sample - *avg;
1541 *avg += diff >> 3;
1542}
1543
1544#else
1545
1546static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1547 const struct cpumask *new_mask, bool check)
1548{
1549 return set_cpus_allowed_ptr(p, new_mask);
1550}
1551
1552#endif /* CONFIG_SMP */
1553
1554static void
1555ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1556{
1557#ifdef CONFIG_SCHEDSTATS
1558 struct rq *rq = this_rq();
1559
1560#ifdef CONFIG_SMP
1561 int this_cpu = smp_processor_id();
1562
1563 if (cpu == this_cpu) {
1564 schedstat_inc(rq, ttwu_local);
1565 schedstat_inc(p, se.statistics.nr_wakeups_local);
1566 } else {
1567 struct sched_domain *sd;
1568
1569 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1570 rcu_read_lock();
1571 for_each_domain(this_cpu, sd) {
1572 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1573 schedstat_inc(sd, ttwu_wake_remote);
1574 break;
1575 }
1576 }
1577 rcu_read_unlock();
1578 }
1579
1580 if (wake_flags & WF_MIGRATED)
1581 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1582
1583#endif /* CONFIG_SMP */
1584
1585 schedstat_inc(rq, ttwu_count);
1586 schedstat_inc(p, se.statistics.nr_wakeups);
1587
1588 if (wake_flags & WF_SYNC)
1589 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1590
1591#endif /* CONFIG_SCHEDSTATS */
1592}
1593
1594static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1595{
1596 activate_task(rq, p, en_flags);
1597 p->on_rq = TASK_ON_RQ_QUEUED;
1598
1599 /* if a worker is waking up, notify workqueue */
1600 if (p->flags & PF_WQ_WORKER)
1601 wq_worker_waking_up(p, cpu_of(rq));
1602}
1603
1604/*
1605 * Mark the task runnable and perform wakeup-preemption.
1606 */
1607static void
1608ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1609{
1610 check_preempt_curr(rq, p, wake_flags);
1611 p->state = TASK_RUNNING;
1612 trace_sched_wakeup(p);
1613
1614#ifdef CONFIG_SMP
1615 if (p->sched_class->task_woken) {
1616 /*
1617 * Our task @p is fully woken up and running; so its safe to
1618 * drop the rq->lock, hereafter rq is only used for statistics.
1619 */
1620 lockdep_unpin_lock(&rq->lock);
1621 p->sched_class->task_woken(rq, p);
1622 lockdep_pin_lock(&rq->lock);
1623 }
1624
1625 if (rq->idle_stamp) {
1626 u64 delta = rq_clock(rq) - rq->idle_stamp;
1627 u64 max = 2*rq->max_idle_balance_cost;
1628
1629 update_avg(&rq->avg_idle, delta);
1630
1631 if (rq->avg_idle > max)
1632 rq->avg_idle = max;
1633
1634 rq->idle_stamp = 0;
1635 }
1636#endif
1637}
1638
1639static void
1640ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1641{
1642 lockdep_assert_held(&rq->lock);
1643
1644#ifdef CONFIG_SMP
1645 if (p->sched_contributes_to_load)
1646 rq->nr_uninterruptible--;
1647#endif
1648
1649 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1650 ttwu_do_wakeup(rq, p, wake_flags);
1651}
1652
1653/*
1654 * Called in case the task @p isn't fully descheduled from its runqueue,
1655 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1656 * since all we need to do is flip p->state to TASK_RUNNING, since
1657 * the task is still ->on_rq.
1658 */
1659static int ttwu_remote(struct task_struct *p, int wake_flags)
1660{
1661 struct rq *rq;
1662 int ret = 0;
1663
1664 rq = __task_rq_lock(p);
1665 if (task_on_rq_queued(p)) {
1666 /* check_preempt_curr() may use rq clock */
1667 update_rq_clock(rq);
1668 ttwu_do_wakeup(rq, p, wake_flags);
1669 ret = 1;
1670 }
1671 __task_rq_unlock(rq);
1672
1673 return ret;
1674}
1675
1676#ifdef CONFIG_SMP
1677void sched_ttwu_pending(void)
1678{
1679 struct rq *rq = this_rq();
1680 struct llist_node *llist = llist_del_all(&rq->wake_list);
1681 struct task_struct *p;
1682 unsigned long flags;
1683
1684 if (!llist)
1685 return;
1686
1687 raw_spin_lock_irqsave(&rq->lock, flags);
1688 lockdep_pin_lock(&rq->lock);
1689
1690 while (llist) {
1691 p = llist_entry(llist, struct task_struct, wake_entry);
1692 llist = llist_next(llist);
1693 ttwu_do_activate(rq, p, 0);
1694 }
1695
1696 lockdep_unpin_lock(&rq->lock);
1697 raw_spin_unlock_irqrestore(&rq->lock, flags);
1698}
1699
1700void scheduler_ipi(void)
1701{
1702 /*
1703 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1704 * TIF_NEED_RESCHED remotely (for the first time) will also send
1705 * this IPI.
1706 */
1707 preempt_fold_need_resched();
1708
1709 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1710 return;
1711
1712 /*
1713 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1714 * traditionally all their work was done from the interrupt return
1715 * path. Now that we actually do some work, we need to make sure
1716 * we do call them.
1717 *
1718 * Some archs already do call them, luckily irq_enter/exit nest
1719 * properly.
1720 *
1721 * Arguably we should visit all archs and update all handlers,
1722 * however a fair share of IPIs are still resched only so this would
1723 * somewhat pessimize the simple resched case.
1724 */
1725 irq_enter();
1726 sched_ttwu_pending();
1727
1728 /*
1729 * Check if someone kicked us for doing the nohz idle load balance.
1730 */
1731 if (unlikely(got_nohz_idle_kick())) {
1732 this_rq()->idle_balance = 1;
1733 raise_softirq_irqoff(SCHED_SOFTIRQ);
1734 }
1735 irq_exit();
1736}
1737
1738static void ttwu_queue_remote(struct task_struct *p, int cpu)
1739{
1740 struct rq *rq = cpu_rq(cpu);
1741
1742 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1743 if (!set_nr_if_polling(rq->idle))
1744 smp_send_reschedule(cpu);
1745 else
1746 trace_sched_wake_idle_without_ipi(cpu);
1747 }
1748}
1749
1750void wake_up_if_idle(int cpu)
1751{
1752 struct rq *rq = cpu_rq(cpu);
1753 unsigned long flags;
1754
1755 rcu_read_lock();
1756
1757 if (!is_idle_task(rcu_dereference(rq->curr)))
1758 goto out;
1759
1760 if (set_nr_if_polling(rq->idle)) {
1761 trace_sched_wake_idle_without_ipi(cpu);
1762 } else {
1763 raw_spin_lock_irqsave(&rq->lock, flags);
1764 if (is_idle_task(rq->curr))
1765 smp_send_reschedule(cpu);
1766 /* Else cpu is not in idle, do nothing here */
1767 raw_spin_unlock_irqrestore(&rq->lock, flags);
1768 }
1769
1770out:
1771 rcu_read_unlock();
1772}
1773
1774bool cpus_share_cache(int this_cpu, int that_cpu)
1775{
1776 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1777}
1778#endif /* CONFIG_SMP */
1779
1780static void ttwu_queue(struct task_struct *p, int cpu)
1781{
1782 struct rq *rq = cpu_rq(cpu);
1783
1784#if defined(CONFIG_SMP)
1785 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1786 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1787 ttwu_queue_remote(p, cpu);
1788 return;
1789 }
1790#endif
1791
1792 raw_spin_lock(&rq->lock);
1793 lockdep_pin_lock(&rq->lock);
1794 ttwu_do_activate(rq, p, 0);
1795 lockdep_unpin_lock(&rq->lock);
1796 raw_spin_unlock(&rq->lock);
1797}
1798
1799/*
1800 * Notes on Program-Order guarantees on SMP systems.
1801 *
1802 * MIGRATION
1803 *
1804 * The basic program-order guarantee on SMP systems is that when a task [t]
1805 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1806 * execution on its new cpu [c1].
1807 *
1808 * For migration (of runnable tasks) this is provided by the following means:
1809 *
1810 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1811 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1812 * rq(c1)->lock (if not at the same time, then in that order).
1813 * C) LOCK of the rq(c1)->lock scheduling in task
1814 *
1815 * Transitivity guarantees that B happens after A and C after B.
1816 * Note: we only require RCpc transitivity.
1817 * Note: the cpu doing B need not be c0 or c1
1818 *
1819 * Example:
1820 *
1821 * CPU0 CPU1 CPU2
1822 *
1823 * LOCK rq(0)->lock
1824 * sched-out X
1825 * sched-in Y
1826 * UNLOCK rq(0)->lock
1827 *
1828 * LOCK rq(0)->lock // orders against CPU0
1829 * dequeue X
1830 * UNLOCK rq(0)->lock
1831 *
1832 * LOCK rq(1)->lock
1833 * enqueue X
1834 * UNLOCK rq(1)->lock
1835 *
1836 * LOCK rq(1)->lock // orders against CPU2
1837 * sched-out Z
1838 * sched-in X
1839 * UNLOCK rq(1)->lock
1840 *
1841 *
1842 * BLOCKING -- aka. SLEEP + WAKEUP
1843 *
1844 * For blocking we (obviously) need to provide the same guarantee as for
1845 * migration. However the means are completely different as there is no lock
1846 * chain to provide order. Instead we do:
1847 *
1848 * 1) smp_store_release(X->on_cpu, 0)
1849 * 2) smp_cond_acquire(!X->on_cpu)
1850 *
1851 * Example:
1852 *
1853 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1854 *
1855 * LOCK rq(0)->lock LOCK X->pi_lock
1856 * dequeue X
1857 * sched-out X
1858 * smp_store_release(X->on_cpu, 0);
1859 *
1860 * smp_cond_acquire(!X->on_cpu);
1861 * X->state = WAKING
1862 * set_task_cpu(X,2)
1863 *
1864 * LOCK rq(2)->lock
1865 * enqueue X
1866 * X->state = RUNNING
1867 * UNLOCK rq(2)->lock
1868 *
1869 * LOCK rq(2)->lock // orders against CPU1
1870 * sched-out Z
1871 * sched-in X
1872 * UNLOCK rq(2)->lock
1873 *
1874 * UNLOCK X->pi_lock
1875 * UNLOCK rq(0)->lock
1876 *
1877 *
1878 * However; for wakeups there is a second guarantee we must provide, namely we
1879 * must observe the state that lead to our wakeup. That is, not only must our
1880 * task observe its own prior state, it must also observe the stores prior to
1881 * its wakeup.
1882 *
1883 * This means that any means of doing remote wakeups must order the CPU doing
1884 * the wakeup against the CPU the task is going to end up running on. This,
1885 * however, is already required for the regular Program-Order guarantee above,
1886 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1887 *
1888 */
1889
1890/**
1891 * try_to_wake_up - wake up a thread
1892 * @p: the thread to be awakened
1893 * @state: the mask of task states that can be woken
1894 * @wake_flags: wake modifier flags (WF_*)
1895 *
1896 * Put it on the run-queue if it's not already there. The "current"
1897 * thread is always on the run-queue (except when the actual
1898 * re-schedule is in progress), and as such you're allowed to do
1899 * the simpler "current->state = TASK_RUNNING" to mark yourself
1900 * runnable without the overhead of this.
1901 *
1902 * Return: %true if @p was woken up, %false if it was already running.
1903 * or @state didn't match @p's state.
1904 */
1905static int
1906try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1907{
1908 unsigned long flags;
1909 int cpu, success = 0;
1910
1911 /*
1912 * If we are going to wake up a thread waiting for CONDITION we
1913 * need to ensure that CONDITION=1 done by the caller can not be
1914 * reordered with p->state check below. This pairs with mb() in
1915 * set_current_state() the waiting thread does.
1916 */
1917 smp_mb__before_spinlock();
1918 raw_spin_lock_irqsave(&p->pi_lock, flags);
1919 if (!(p->state & state))
1920 goto out;
1921
1922 trace_sched_waking(p);
1923
1924 success = 1; /* we're going to change ->state */
1925 cpu = task_cpu(p);
1926
1927 if (p->on_rq && ttwu_remote(p, wake_flags))
1928 goto stat;
1929
1930#ifdef CONFIG_SMP
1931 /*
1932 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1933 * possible to, falsely, observe p->on_cpu == 0.
1934 *
1935 * One must be running (->on_cpu == 1) in order to remove oneself
1936 * from the runqueue.
1937 *
1938 * [S] ->on_cpu = 1; [L] ->on_rq
1939 * UNLOCK rq->lock
1940 * RMB
1941 * LOCK rq->lock
1942 * [S] ->on_rq = 0; [L] ->on_cpu
1943 *
1944 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1945 * from the consecutive calls to schedule(); the first switching to our
1946 * task, the second putting it to sleep.
1947 */
1948 smp_rmb();
1949
1950 /*
1951 * If the owning (remote) cpu is still in the middle of schedule() with
1952 * this task as prev, wait until its done referencing the task.
1953 *
1954 * Pairs with the smp_store_release() in finish_lock_switch().
1955 *
1956 * This ensures that tasks getting woken will be fully ordered against
1957 * their previous state and preserve Program Order.
1958 */
1959 smp_cond_acquire(!p->on_cpu);
1960
1961 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1962 p->state = TASK_WAKING;
1963
1964 if (p->sched_class->task_waking)
1965 p->sched_class->task_waking(p);
1966
1967 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1968 if (task_cpu(p) != cpu) {
1969 wake_flags |= WF_MIGRATED;
1970 set_task_cpu(p, cpu);
1971 }
1972#endif /* CONFIG_SMP */
1973
1974 ttwu_queue(p, cpu);
1975stat:
1976 if (schedstat_enabled())
1977 ttwu_stat(p, cpu, wake_flags);
1978out:
1979 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1980
1981 return success;
1982}
1983
1984/**
1985 * try_to_wake_up_local - try to wake up a local task with rq lock held
1986 * @p: the thread to be awakened
1987 *
1988 * Put @p on the run-queue if it's not already there. The caller must
1989 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1990 * the current task.
1991 */
1992static void try_to_wake_up_local(struct task_struct *p)
1993{
1994 struct rq *rq = task_rq(p);
1995
1996 if (WARN_ON_ONCE(rq != this_rq()) ||
1997 WARN_ON_ONCE(p == current))
1998 return;
1999
2000 lockdep_assert_held(&rq->lock);
2001
2002 if (!raw_spin_trylock(&p->pi_lock)) {
2003 /*
2004 * This is OK, because current is on_cpu, which avoids it being
2005 * picked for load-balance and preemption/IRQs are still
2006 * disabled avoiding further scheduler activity on it and we've
2007 * not yet picked a replacement task.
2008 */
2009 lockdep_unpin_lock(&rq->lock);
2010 raw_spin_unlock(&rq->lock);
2011 raw_spin_lock(&p->pi_lock);
2012 raw_spin_lock(&rq->lock);
2013 lockdep_pin_lock(&rq->lock);
2014 }
2015
2016 if (!(p->state & TASK_NORMAL))
2017 goto out;
2018
2019 trace_sched_waking(p);
2020
2021 if (!task_on_rq_queued(p))
2022 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2023
2024 ttwu_do_wakeup(rq, p, 0);
2025 if (schedstat_enabled())
2026 ttwu_stat(p, smp_processor_id(), 0);
2027out:
2028 raw_spin_unlock(&p->pi_lock);
2029}
2030
2031/**
2032 * wake_up_process - Wake up a specific process
2033 * @p: The process to be woken up.
2034 *
2035 * Attempt to wake up the nominated process and move it to the set of runnable
2036 * processes.
2037 *
2038 * Return: 1 if the process was woken up, 0 if it was already running.
2039 *
2040 * It may be assumed that this function implies a write memory barrier before
2041 * changing the task state if and only if any tasks are woken up.
2042 */
2043int wake_up_process(struct task_struct *p)
2044{
2045 return try_to_wake_up(p, TASK_NORMAL, 0);
2046}
2047EXPORT_SYMBOL(wake_up_process);
2048
2049int wake_up_state(struct task_struct *p, unsigned int state)
2050{
2051 return try_to_wake_up(p, state, 0);
2052}
2053
2054/*
2055 * This function clears the sched_dl_entity static params.
2056 */
2057void __dl_clear_params(struct task_struct *p)
2058{
2059 struct sched_dl_entity *dl_se = &p->dl;
2060
2061 dl_se->dl_runtime = 0;
2062 dl_se->dl_deadline = 0;
2063 dl_se->dl_period = 0;
2064 dl_se->flags = 0;
2065 dl_se->dl_bw = 0;
2066
2067 dl_se->dl_throttled = 0;
2068 dl_se->dl_yielded = 0;
2069}
2070
2071/*
2072 * Perform scheduler related setup for a newly forked process p.
2073 * p is forked by current.
2074 *
2075 * __sched_fork() is basic setup used by init_idle() too:
2076 */
2077static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2078{
2079 p->on_rq = 0;
2080
2081 p->se.on_rq = 0;
2082 p->se.exec_start = 0;
2083 p->se.sum_exec_runtime = 0;
2084 p->se.prev_sum_exec_runtime = 0;
2085 p->se.nr_migrations = 0;
2086 p->se.vruntime = 0;
2087 INIT_LIST_HEAD(&p->se.group_node);
2088
2089#ifdef CONFIG_FAIR_GROUP_SCHED
2090 p->se.cfs_rq = NULL;
2091#endif
2092
2093#ifdef CONFIG_SCHEDSTATS
2094 /* Even if schedstat is disabled, there should not be garbage */
2095 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2096#endif
2097
2098 RB_CLEAR_NODE(&p->dl.rb_node);
2099 init_dl_task_timer(&p->dl);
2100 __dl_clear_params(p);
2101
2102 INIT_LIST_HEAD(&p->rt.run_list);
2103 p->rt.timeout = 0;
2104 p->rt.time_slice = sched_rr_timeslice;
2105 p->rt.on_rq = 0;
2106 p->rt.on_list = 0;
2107
2108#ifdef CONFIG_PREEMPT_NOTIFIERS
2109 INIT_HLIST_HEAD(&p->preempt_notifiers);
2110#endif
2111
2112#ifdef CONFIG_NUMA_BALANCING
2113 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2114 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2115 p->mm->numa_scan_seq = 0;
2116 }
2117
2118 if (clone_flags & CLONE_VM)
2119 p->numa_preferred_nid = current->numa_preferred_nid;
2120 else
2121 p->numa_preferred_nid = -1;
2122
2123 p->node_stamp = 0ULL;
2124 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2125 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2126 p->numa_work.next = &p->numa_work;
2127 p->numa_faults = NULL;
2128 p->last_task_numa_placement = 0;
2129 p->last_sum_exec_runtime = 0;
2130
2131 p->numa_group = NULL;
2132#endif /* CONFIG_NUMA_BALANCING */
2133}
2134
2135DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2136
2137#ifdef CONFIG_NUMA_BALANCING
2138
2139void set_numabalancing_state(bool enabled)
2140{
2141 if (enabled)
2142 static_branch_enable(&sched_numa_balancing);
2143 else
2144 static_branch_disable(&sched_numa_balancing);
2145}
2146
2147#ifdef CONFIG_PROC_SYSCTL
2148int sysctl_numa_balancing(struct ctl_table *table, int write,
2149 void __user *buffer, size_t *lenp, loff_t *ppos)
2150{
2151 struct ctl_table t;
2152 int err;
2153 int state = static_branch_likely(&sched_numa_balancing);
2154
2155 if (write && !capable(CAP_SYS_ADMIN))
2156 return -EPERM;
2157
2158 t = *table;
2159 t.data = &state;
2160 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2161 if (err < 0)
2162 return err;
2163 if (write)
2164 set_numabalancing_state(state);
2165 return err;
2166}
2167#endif
2168#endif
2169
2170DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2171
2172#ifdef CONFIG_SCHEDSTATS
2173static void set_schedstats(bool enabled)
2174{
2175 if (enabled)
2176 static_branch_enable(&sched_schedstats);
2177 else
2178 static_branch_disable(&sched_schedstats);
2179}
2180
2181void force_schedstat_enabled(void)
2182{
2183 if (!schedstat_enabled()) {
2184 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2185 static_branch_enable(&sched_schedstats);
2186 }
2187}
2188
2189static int __init setup_schedstats(char *str)
2190{
2191 int ret = 0;
2192 if (!str)
2193 goto out;
2194
2195 if (!strcmp(str, "enable")) {
2196 set_schedstats(true);
2197 ret = 1;
2198 } else if (!strcmp(str, "disable")) {
2199 set_schedstats(false);
2200 ret = 1;
2201 }
2202out:
2203 if (!ret)
2204 pr_warn("Unable to parse schedstats=\n");
2205
2206 return ret;
2207}
2208__setup("schedstats=", setup_schedstats);
2209
2210#ifdef CONFIG_PROC_SYSCTL
2211int sysctl_schedstats(struct ctl_table *table, int write,
2212 void __user *buffer, size_t *lenp, loff_t *ppos)
2213{
2214 struct ctl_table t;
2215 int err;
2216 int state = static_branch_likely(&sched_schedstats);
2217
2218 if (write && !capable(CAP_SYS_ADMIN))
2219 return -EPERM;
2220
2221 t = *table;
2222 t.data = &state;
2223 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2224 if (err < 0)
2225 return err;
2226 if (write)
2227 set_schedstats(state);
2228 return err;
2229}
2230#endif
2231#endif
2232
2233/*
2234 * fork()/clone()-time setup:
2235 */
2236int sched_fork(unsigned long clone_flags, struct task_struct *p)
2237{
2238 unsigned long flags;
2239 int cpu = get_cpu();
2240
2241 __sched_fork(clone_flags, p);
2242 /*
2243 * We mark the process as running here. This guarantees that
2244 * nobody will actually run it, and a signal or other external
2245 * event cannot wake it up and insert it on the runqueue either.
2246 */
2247 p->state = TASK_RUNNING;
2248
2249 /*
2250 * Make sure we do not leak PI boosting priority to the child.
2251 */
2252 p->prio = current->normal_prio;
2253
2254 /*
2255 * Revert to default priority/policy on fork if requested.
2256 */
2257 if (unlikely(p->sched_reset_on_fork)) {
2258 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2259 p->policy = SCHED_NORMAL;
2260 p->static_prio = NICE_TO_PRIO(0);
2261 p->rt_priority = 0;
2262 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2263 p->static_prio = NICE_TO_PRIO(0);
2264
2265 p->prio = p->normal_prio = __normal_prio(p);
2266 set_load_weight(p);
2267
2268 /*
2269 * We don't need the reset flag anymore after the fork. It has
2270 * fulfilled its duty:
2271 */
2272 p->sched_reset_on_fork = 0;
2273 }
2274
2275 if (dl_prio(p->prio)) {
2276 put_cpu();
2277 return -EAGAIN;
2278 } else if (rt_prio(p->prio)) {
2279 p->sched_class = &rt_sched_class;
2280 } else {
2281 p->sched_class = &fair_sched_class;
2282 }
2283
2284 if (p->sched_class->task_fork)
2285 p->sched_class->task_fork(p);
2286
2287 /*
2288 * The child is not yet in the pid-hash so no cgroup attach races,
2289 * and the cgroup is pinned to this child due to cgroup_fork()
2290 * is ran before sched_fork().
2291 *
2292 * Silence PROVE_RCU.
2293 */
2294 raw_spin_lock_irqsave(&p->pi_lock, flags);
2295 set_task_cpu(p, cpu);
2296 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2297
2298#ifdef CONFIG_SCHED_INFO
2299 if (likely(sched_info_on()))
2300 memset(&p->sched_info, 0, sizeof(p->sched_info));
2301#endif
2302#if defined(CONFIG_SMP)
2303 p->on_cpu = 0;
2304#endif
2305 init_task_preempt_count(p);
2306#ifdef CONFIG_SMP
2307 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2308 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2309#endif
2310
2311 put_cpu();
2312 return 0;
2313}
2314
2315unsigned long to_ratio(u64 period, u64 runtime)
2316{
2317 if (runtime == RUNTIME_INF)
2318 return 1ULL << 20;
2319
2320 /*
2321 * Doing this here saves a lot of checks in all
2322 * the calling paths, and returning zero seems
2323 * safe for them anyway.
2324 */
2325 if (period == 0)
2326 return 0;
2327
2328 return div64_u64(runtime << 20, period);
2329}
2330
2331#ifdef CONFIG_SMP
2332inline struct dl_bw *dl_bw_of(int i)
2333{
2334 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2335 "sched RCU must be held");
2336 return &cpu_rq(i)->rd->dl_bw;
2337}
2338
2339static inline int dl_bw_cpus(int i)
2340{
2341 struct root_domain *rd = cpu_rq(i)->rd;
2342 int cpus = 0;
2343
2344 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2345 "sched RCU must be held");
2346 for_each_cpu_and(i, rd->span, cpu_active_mask)
2347 cpus++;
2348
2349 return cpus;
2350}
2351#else
2352inline struct dl_bw *dl_bw_of(int i)
2353{
2354 return &cpu_rq(i)->dl.dl_bw;
2355}
2356
2357static inline int dl_bw_cpus(int i)
2358{
2359 return 1;
2360}
2361#endif
2362
2363/*
2364 * We must be sure that accepting a new task (or allowing changing the
2365 * parameters of an existing one) is consistent with the bandwidth
2366 * constraints. If yes, this function also accordingly updates the currently
2367 * allocated bandwidth to reflect the new situation.
2368 *
2369 * This function is called while holding p's rq->lock.
2370 *
2371 * XXX we should delay bw change until the task's 0-lag point, see
2372 * __setparam_dl().
2373 */
2374static int dl_overflow(struct task_struct *p, int policy,
2375 const struct sched_attr *attr)
2376{
2377
2378 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2379 u64 period = attr->sched_period ?: attr->sched_deadline;
2380 u64 runtime = attr->sched_runtime;
2381 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2382 int cpus, err = -1;
2383
2384 if (new_bw == p->dl.dl_bw)
2385 return 0;
2386
2387 /*
2388 * Either if a task, enters, leave, or stays -deadline but changes
2389 * its parameters, we may need to update accordingly the total
2390 * allocated bandwidth of the container.
2391 */
2392 raw_spin_lock(&dl_b->lock);
2393 cpus = dl_bw_cpus(task_cpu(p));
2394 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2395 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2396 __dl_add(dl_b, new_bw);
2397 err = 0;
2398 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2399 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2400 __dl_clear(dl_b, p->dl.dl_bw);
2401 __dl_add(dl_b, new_bw);
2402 err = 0;
2403 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2404 __dl_clear(dl_b, p->dl.dl_bw);
2405 err = 0;
2406 }
2407 raw_spin_unlock(&dl_b->lock);
2408
2409 return err;
2410}
2411
2412extern void init_dl_bw(struct dl_bw *dl_b);
2413
2414/*
2415 * wake_up_new_task - wake up a newly created task for the first time.
2416 *
2417 * This function will do some initial scheduler statistics housekeeping
2418 * that must be done for every newly created context, then puts the task
2419 * on the runqueue and wakes it.
2420 */
2421void wake_up_new_task(struct task_struct *p)
2422{
2423 unsigned long flags;
2424 struct rq *rq;
2425
2426 raw_spin_lock_irqsave(&p->pi_lock, flags);
2427 /* Initialize new task's runnable average */
2428 init_entity_runnable_average(&p->se);
2429#ifdef CONFIG_SMP
2430 /*
2431 * Fork balancing, do it here and not earlier because:
2432 * - cpus_allowed can change in the fork path
2433 * - any previously selected cpu might disappear through hotplug
2434 */
2435 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2436#endif
2437
2438 rq = __task_rq_lock(p);
2439 activate_task(rq, p, 0);
2440 p->on_rq = TASK_ON_RQ_QUEUED;
2441 trace_sched_wakeup_new(p);
2442 check_preempt_curr(rq, p, WF_FORK);
2443#ifdef CONFIG_SMP
2444 if (p->sched_class->task_woken) {
2445 /*
2446 * Nothing relies on rq->lock after this, so its fine to
2447 * drop it.
2448 */
2449 lockdep_unpin_lock(&rq->lock);
2450 p->sched_class->task_woken(rq, p);
2451 lockdep_pin_lock(&rq->lock);
2452 }
2453#endif
2454 task_rq_unlock(rq, p, &flags);
2455}
2456
2457#ifdef CONFIG_PREEMPT_NOTIFIERS
2458
2459static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2460
2461void preempt_notifier_inc(void)
2462{
2463 static_key_slow_inc(&preempt_notifier_key);
2464}
2465EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2466
2467void preempt_notifier_dec(void)
2468{
2469 static_key_slow_dec(&preempt_notifier_key);
2470}
2471EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2472
2473/**
2474 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2475 * @notifier: notifier struct to register
2476 */
2477void preempt_notifier_register(struct preempt_notifier *notifier)
2478{
2479 if (!static_key_false(&preempt_notifier_key))
2480 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2481
2482 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2483}
2484EXPORT_SYMBOL_GPL(preempt_notifier_register);
2485
2486/**
2487 * preempt_notifier_unregister - no longer interested in preemption notifications
2488 * @notifier: notifier struct to unregister
2489 *
2490 * This is *not* safe to call from within a preemption notifier.
2491 */
2492void preempt_notifier_unregister(struct preempt_notifier *notifier)
2493{
2494 hlist_del(¬ifier->link);
2495}
2496EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2497
2498static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2499{
2500 struct preempt_notifier *notifier;
2501
2502 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2503 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2504}
2505
2506static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2507{
2508 if (static_key_false(&preempt_notifier_key))
2509 __fire_sched_in_preempt_notifiers(curr);
2510}
2511
2512static void
2513__fire_sched_out_preempt_notifiers(struct task_struct *curr,
2514 struct task_struct *next)
2515{
2516 struct preempt_notifier *notifier;
2517
2518 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2519 notifier->ops->sched_out(notifier, next);
2520}
2521
2522static __always_inline void
2523fire_sched_out_preempt_notifiers(struct task_struct *curr,
2524 struct task_struct *next)
2525{
2526 if (static_key_false(&preempt_notifier_key))
2527 __fire_sched_out_preempt_notifiers(curr, next);
2528}
2529
2530#else /* !CONFIG_PREEMPT_NOTIFIERS */
2531
2532static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2533{
2534}
2535
2536static inline void
2537fire_sched_out_preempt_notifiers(struct task_struct *curr,
2538 struct task_struct *next)
2539{
2540}
2541
2542#endif /* CONFIG_PREEMPT_NOTIFIERS */
2543
2544/**
2545 * prepare_task_switch - prepare to switch tasks
2546 * @rq: the runqueue preparing to switch
2547 * @prev: the current task that is being switched out
2548 * @next: the task we are going to switch to.
2549 *
2550 * This is called with the rq lock held and interrupts off. It must
2551 * be paired with a subsequent finish_task_switch after the context
2552 * switch.
2553 *
2554 * prepare_task_switch sets up locking and calls architecture specific
2555 * hooks.
2556 */
2557static inline void
2558prepare_task_switch(struct rq *rq, struct task_struct *prev,
2559 struct task_struct *next)
2560{
2561 sched_info_switch(rq, prev, next);
2562 perf_event_task_sched_out(prev, next);
2563 fire_sched_out_preempt_notifiers(prev, next);
2564 prepare_lock_switch(rq, next);
2565 prepare_arch_switch(next);
2566}
2567
2568/**
2569 * finish_task_switch - clean up after a task-switch
2570 * @prev: the thread we just switched away from.
2571 *
2572 * finish_task_switch must be called after the context switch, paired
2573 * with a prepare_task_switch call before the context switch.
2574 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2575 * and do any other architecture-specific cleanup actions.
2576 *
2577 * Note that we may have delayed dropping an mm in context_switch(). If
2578 * so, we finish that here outside of the runqueue lock. (Doing it
2579 * with the lock held can cause deadlocks; see schedule() for
2580 * details.)
2581 *
2582 * The context switch have flipped the stack from under us and restored the
2583 * local variables which were saved when this task called schedule() in the
2584 * past. prev == current is still correct but we need to recalculate this_rq
2585 * because prev may have moved to another CPU.
2586 */
2587static struct rq *finish_task_switch(struct task_struct *prev)
2588 __releases(rq->lock)
2589{
2590 struct rq *rq = this_rq();
2591 struct mm_struct *mm = rq->prev_mm;
2592 long prev_state;
2593
2594 /*
2595 * The previous task will have left us with a preempt_count of 2
2596 * because it left us after:
2597 *
2598 * schedule()
2599 * preempt_disable(); // 1
2600 * __schedule()
2601 * raw_spin_lock_irq(&rq->lock) // 2
2602 *
2603 * Also, see FORK_PREEMPT_COUNT.
2604 */
2605 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2606 "corrupted preempt_count: %s/%d/0x%x\n",
2607 current->comm, current->pid, preempt_count()))
2608 preempt_count_set(FORK_PREEMPT_COUNT);
2609
2610 rq->prev_mm = NULL;
2611
2612 /*
2613 * A task struct has one reference for the use as "current".
2614 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2615 * schedule one last time. The schedule call will never return, and
2616 * the scheduled task must drop that reference.
2617 *
2618 * We must observe prev->state before clearing prev->on_cpu (in
2619 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2620 * running on another CPU and we could rave with its RUNNING -> DEAD
2621 * transition, resulting in a double drop.
2622 */
2623 prev_state = prev->state;
2624 vtime_task_switch(prev);
2625 perf_event_task_sched_in(prev, current);
2626 finish_lock_switch(rq, prev);
2627 finish_arch_post_lock_switch();
2628
2629 fire_sched_in_preempt_notifiers(current);
2630 if (mm)
2631 mmdrop(mm);
2632 if (unlikely(prev_state == TASK_DEAD)) {
2633 if (prev->sched_class->task_dead)
2634 prev->sched_class->task_dead(prev);
2635
2636 /*
2637 * Remove function-return probe instances associated with this
2638 * task and put them back on the free list.
2639 */
2640 kprobe_flush_task(prev);
2641 put_task_struct(prev);
2642 }
2643
2644 tick_nohz_task_switch();
2645 return rq;
2646}
2647
2648#ifdef CONFIG_SMP
2649
2650/* rq->lock is NOT held, but preemption is disabled */
2651static void __balance_callback(struct rq *rq)
2652{
2653 struct callback_head *head, *next;
2654 void (*func)(struct rq *rq);
2655 unsigned long flags;
2656
2657 raw_spin_lock_irqsave(&rq->lock, flags);
2658 head = rq->balance_callback;
2659 rq->balance_callback = NULL;
2660 while (head) {
2661 func = (void (*)(struct rq *))head->func;
2662 next = head->next;
2663 head->next = NULL;
2664 head = next;
2665
2666 func(rq);
2667 }
2668 raw_spin_unlock_irqrestore(&rq->lock, flags);
2669}
2670
2671static inline void balance_callback(struct rq *rq)
2672{
2673 if (unlikely(rq->balance_callback))
2674 __balance_callback(rq);
2675}
2676
2677#else
2678
2679static inline void balance_callback(struct rq *rq)
2680{
2681}
2682
2683#endif
2684
2685/**
2686 * schedule_tail - first thing a freshly forked thread must call.
2687 * @prev: the thread we just switched away from.
2688 */
2689asmlinkage __visible void schedule_tail(struct task_struct *prev)
2690 __releases(rq->lock)
2691{
2692 struct rq *rq;
2693
2694 /*
2695 * New tasks start with FORK_PREEMPT_COUNT, see there and
2696 * finish_task_switch() for details.
2697 *
2698 * finish_task_switch() will drop rq->lock() and lower preempt_count
2699 * and the preempt_enable() will end up enabling preemption (on
2700 * PREEMPT_COUNT kernels).
2701 */
2702
2703 rq = finish_task_switch(prev);
2704 balance_callback(rq);
2705 preempt_enable();
2706
2707 if (current->set_child_tid)
2708 put_user(task_pid_vnr(current), current->set_child_tid);
2709}
2710
2711/*
2712 * context_switch - switch to the new MM and the new thread's register state.
2713 */
2714static __always_inline struct rq *
2715context_switch(struct rq *rq, struct task_struct *prev,
2716 struct task_struct *next)
2717{
2718 struct mm_struct *mm, *oldmm;
2719
2720 prepare_task_switch(rq, prev, next);
2721
2722 mm = next->mm;
2723 oldmm = prev->active_mm;
2724 /*
2725 * For paravirt, this is coupled with an exit in switch_to to
2726 * combine the page table reload and the switch backend into
2727 * one hypercall.
2728 */
2729 arch_start_context_switch(prev);
2730
2731 if (!mm) {
2732 next->active_mm = oldmm;
2733 atomic_inc(&oldmm->mm_count);
2734 enter_lazy_tlb(oldmm, next);
2735 } else
2736 switch_mm(oldmm, mm, next);
2737
2738 if (!prev->mm) {
2739 prev->active_mm = NULL;
2740 rq->prev_mm = oldmm;
2741 }
2742 /*
2743 * Since the runqueue lock will be released by the next
2744 * task (which is an invalid locking op but in the case
2745 * of the scheduler it's an obvious special-case), so we
2746 * do an early lockdep release here:
2747 */
2748 lockdep_unpin_lock(&rq->lock);
2749 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2750
2751 /* Here we just switch the register state and the stack. */
2752 switch_to(prev, next, prev);
2753 barrier();
2754
2755 return finish_task_switch(prev);
2756}
2757
2758/*
2759 * nr_running and nr_context_switches:
2760 *
2761 * externally visible scheduler statistics: current number of runnable
2762 * threads, total number of context switches performed since bootup.
2763 */
2764unsigned long nr_running(void)
2765{
2766 unsigned long i, sum = 0;
2767
2768 for_each_online_cpu(i)
2769 sum += cpu_rq(i)->nr_running;
2770
2771 return sum;
2772}
2773
2774/*
2775 * Check if only the current task is running on the cpu.
2776 *
2777 * Caution: this function does not check that the caller has disabled
2778 * preemption, thus the result might have a time-of-check-to-time-of-use
2779 * race. The caller is responsible to use it correctly, for example:
2780 *
2781 * - from a non-preemptable section (of course)
2782 *
2783 * - from a thread that is bound to a single CPU
2784 *
2785 * - in a loop with very short iterations (e.g. a polling loop)
2786 */
2787bool single_task_running(void)
2788{
2789 return raw_rq()->nr_running == 1;
2790}
2791EXPORT_SYMBOL(single_task_running);
2792
2793unsigned long long nr_context_switches(void)
2794{
2795 int i;
2796 unsigned long long sum = 0;
2797
2798 for_each_possible_cpu(i)
2799 sum += cpu_rq(i)->nr_switches;
2800
2801 return sum;
2802}
2803
2804unsigned long nr_iowait(void)
2805{
2806 unsigned long i, sum = 0;
2807
2808 for_each_possible_cpu(i)
2809 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2810
2811 return sum;
2812}
2813
2814unsigned long nr_iowait_cpu(int cpu)
2815{
2816 struct rq *this = cpu_rq(cpu);
2817 return atomic_read(&this->nr_iowait);
2818}
2819
2820void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2821{
2822 struct rq *rq = this_rq();
2823 *nr_waiters = atomic_read(&rq->nr_iowait);
2824 *load = rq->load.weight;
2825}
2826
2827#ifdef CONFIG_SMP
2828
2829/*
2830 * sched_exec - execve() is a valuable balancing opportunity, because at
2831 * this point the task has the smallest effective memory and cache footprint.
2832 */
2833void sched_exec(void)
2834{
2835 struct task_struct *p = current;
2836 unsigned long flags;
2837 int dest_cpu;
2838
2839 raw_spin_lock_irqsave(&p->pi_lock, flags);
2840 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2841 if (dest_cpu == smp_processor_id())
2842 goto unlock;
2843
2844 if (likely(cpu_active(dest_cpu))) {
2845 struct migration_arg arg = { p, dest_cpu };
2846
2847 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2848 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2849 return;
2850 }
2851unlock:
2852 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2853}
2854
2855#endif
2856
2857DEFINE_PER_CPU(struct kernel_stat, kstat);
2858DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2859
2860EXPORT_PER_CPU_SYMBOL(kstat);
2861EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2862
2863/*
2864 * Return accounted runtime for the task.
2865 * In case the task is currently running, return the runtime plus current's
2866 * pending runtime that have not been accounted yet.
2867 */
2868unsigned long long task_sched_runtime(struct task_struct *p)
2869{
2870 unsigned long flags;
2871 struct rq *rq;
2872 u64 ns;
2873
2874#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2875 /*
2876 * 64-bit doesn't need locks to atomically read a 64bit value.
2877 * So we have a optimization chance when the task's delta_exec is 0.
2878 * Reading ->on_cpu is racy, but this is ok.
2879 *
2880 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2881 * If we race with it entering cpu, unaccounted time is 0. This is
2882 * indistinguishable from the read occurring a few cycles earlier.
2883 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2884 * been accounted, so we're correct here as well.
2885 */
2886 if (!p->on_cpu || !task_on_rq_queued(p))
2887 return p->se.sum_exec_runtime;
2888#endif
2889
2890 rq = task_rq_lock(p, &flags);
2891 /*
2892 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2893 * project cycles that may never be accounted to this
2894 * thread, breaking clock_gettime().
2895 */
2896 if (task_current(rq, p) && task_on_rq_queued(p)) {
2897 update_rq_clock(rq);
2898 p->sched_class->update_curr(rq);
2899 }
2900 ns = p->se.sum_exec_runtime;
2901 task_rq_unlock(rq, p, &flags);
2902
2903 return ns;
2904}
2905
2906/*
2907 * This function gets called by the timer code, with HZ frequency.
2908 * We call it with interrupts disabled.
2909 */
2910void scheduler_tick(void)
2911{
2912 int cpu = smp_processor_id();
2913 struct rq *rq = cpu_rq(cpu);
2914 struct task_struct *curr = rq->curr;
2915
2916 sched_clock_tick();
2917
2918 raw_spin_lock(&rq->lock);
2919 update_rq_clock(rq);
2920 curr->sched_class->task_tick(rq, curr, 0);
2921 update_cpu_load_active(rq);
2922 calc_global_load_tick(rq);
2923 raw_spin_unlock(&rq->lock);
2924
2925 perf_event_task_tick();
2926
2927#ifdef CONFIG_SMP
2928 rq->idle_balance = idle_cpu(cpu);
2929 trigger_load_balance(rq);
2930#endif
2931 rq_last_tick_reset(rq);
2932}
2933
2934#ifdef CONFIG_NO_HZ_FULL
2935/**
2936 * scheduler_tick_max_deferment
2937 *
2938 * Keep at least one tick per second when a single
2939 * active task is running because the scheduler doesn't
2940 * yet completely support full dynticks environment.
2941 *
2942 * This makes sure that uptime, CFS vruntime, load
2943 * balancing, etc... continue to move forward, even
2944 * with a very low granularity.
2945 *
2946 * Return: Maximum deferment in nanoseconds.
2947 */
2948u64 scheduler_tick_max_deferment(void)
2949{
2950 struct rq *rq = this_rq();
2951 unsigned long next, now = READ_ONCE(jiffies);
2952
2953 next = rq->last_sched_tick + HZ;
2954
2955 if (time_before_eq(next, now))
2956 return 0;
2957
2958 return jiffies_to_nsecs(next - now);
2959}
2960#endif
2961
2962#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2963 defined(CONFIG_PREEMPT_TRACER))
2964
2965void preempt_count_add(int val)
2966{
2967#ifdef CONFIG_DEBUG_PREEMPT
2968 /*
2969 * Underflow?
2970 */
2971 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2972 return;
2973#endif
2974 __preempt_count_add(val);
2975#ifdef CONFIG_DEBUG_PREEMPT
2976 /*
2977 * Spinlock count overflowing soon?
2978 */
2979 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2980 PREEMPT_MASK - 10);
2981#endif
2982 if (preempt_count() == val) {
2983 unsigned long ip = get_lock_parent_ip();
2984#ifdef CONFIG_DEBUG_PREEMPT
2985 current->preempt_disable_ip = ip;
2986#endif
2987 trace_preempt_off(CALLER_ADDR0, ip);
2988 }
2989}
2990EXPORT_SYMBOL(preempt_count_add);
2991NOKPROBE_SYMBOL(preempt_count_add);
2992
2993void preempt_count_sub(int val)
2994{
2995#ifdef CONFIG_DEBUG_PREEMPT
2996 /*
2997 * Underflow?
2998 */
2999 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3000 return;
3001 /*
3002 * Is the spinlock portion underflowing?
3003 */
3004 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3005 !(preempt_count() & PREEMPT_MASK)))
3006 return;
3007#endif
3008
3009 if (preempt_count() == val)
3010 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3011 __preempt_count_sub(val);
3012}
3013EXPORT_SYMBOL(preempt_count_sub);
3014NOKPROBE_SYMBOL(preempt_count_sub);
3015
3016#endif
3017
3018/*
3019 * Print scheduling while atomic bug:
3020 */
3021static noinline void __schedule_bug(struct task_struct *prev)
3022{
3023 if (oops_in_progress)
3024 return;
3025
3026 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3027 prev->comm, prev->pid, preempt_count());
3028
3029 debug_show_held_locks(prev);
3030 print_modules();
3031 if (irqs_disabled())
3032 print_irqtrace_events(prev);
3033#ifdef CONFIG_DEBUG_PREEMPT
3034 if (in_atomic_preempt_off()) {
3035 pr_err("Preemption disabled at:");
3036 print_ip_sym(current->preempt_disable_ip);
3037 pr_cont("\n");
3038 }
3039#endif
3040 dump_stack();
3041 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3042}
3043
3044/*
3045 * Various schedule()-time debugging checks and statistics:
3046 */
3047static inline void schedule_debug(struct task_struct *prev)
3048{
3049#ifdef CONFIG_SCHED_STACK_END_CHECK
3050 BUG_ON(task_stack_end_corrupted(prev));
3051#endif
3052
3053 if (unlikely(in_atomic_preempt_off())) {
3054 __schedule_bug(prev);
3055 preempt_count_set(PREEMPT_DISABLED);
3056 }
3057 rcu_sleep_check();
3058
3059 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3060
3061 schedstat_inc(this_rq(), sched_count);
3062}
3063
3064/*
3065 * Pick up the highest-prio task:
3066 */
3067static inline struct task_struct *
3068pick_next_task(struct rq *rq, struct task_struct *prev)
3069{
3070 const struct sched_class *class = &fair_sched_class;
3071 struct task_struct *p;
3072
3073 /*
3074 * Optimization: we know that if all tasks are in
3075 * the fair class we can call that function directly:
3076 */
3077 if (likely(prev->sched_class == class &&
3078 rq->nr_running == rq->cfs.h_nr_running)) {
3079 p = fair_sched_class.pick_next_task(rq, prev);
3080 if (unlikely(p == RETRY_TASK))
3081 goto again;
3082
3083 /* assumes fair_sched_class->next == idle_sched_class */
3084 if (unlikely(!p))
3085 p = idle_sched_class.pick_next_task(rq, prev);
3086
3087 return p;
3088 }
3089
3090again:
3091 for_each_class(class) {
3092 p = class->pick_next_task(rq, prev);
3093 if (p) {
3094 if (unlikely(p == RETRY_TASK))
3095 goto again;
3096 return p;
3097 }
3098 }
3099
3100 BUG(); /* the idle class will always have a runnable task */
3101}
3102
3103/*
3104 * __schedule() is the main scheduler function.
3105 *
3106 * The main means of driving the scheduler and thus entering this function are:
3107 *
3108 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3109 *
3110 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3111 * paths. For example, see arch/x86/entry_64.S.
3112 *
3113 * To drive preemption between tasks, the scheduler sets the flag in timer
3114 * interrupt handler scheduler_tick().
3115 *
3116 * 3. Wakeups don't really cause entry into schedule(). They add a
3117 * task to the run-queue and that's it.
3118 *
3119 * Now, if the new task added to the run-queue preempts the current
3120 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3121 * called on the nearest possible occasion:
3122 *
3123 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3124 *
3125 * - in syscall or exception context, at the next outmost
3126 * preempt_enable(). (this might be as soon as the wake_up()'s
3127 * spin_unlock()!)
3128 *
3129 * - in IRQ context, return from interrupt-handler to
3130 * preemptible context
3131 *
3132 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3133 * then at the next:
3134 *
3135 * - cond_resched() call
3136 * - explicit schedule() call
3137 * - return from syscall or exception to user-space
3138 * - return from interrupt-handler to user-space
3139 *
3140 * WARNING: must be called with preemption disabled!
3141 */
3142static void __sched notrace __schedule(bool preempt)
3143{
3144 struct task_struct *prev, *next;
3145 unsigned long *switch_count;
3146 struct rq *rq;
3147 int cpu;
3148
3149 cpu = smp_processor_id();
3150 rq = cpu_rq(cpu);
3151 prev = rq->curr;
3152
3153 /*
3154 * do_exit() calls schedule() with preemption disabled as an exception;
3155 * however we must fix that up, otherwise the next task will see an
3156 * inconsistent (higher) preempt count.
3157 *
3158 * It also avoids the below schedule_debug() test from complaining
3159 * about this.
3160 */
3161 if (unlikely(prev->state == TASK_DEAD))
3162 preempt_enable_no_resched_notrace();
3163
3164 schedule_debug(prev);
3165
3166 if (sched_feat(HRTICK))
3167 hrtick_clear(rq);
3168
3169 local_irq_disable();
3170 rcu_note_context_switch();
3171
3172 /*
3173 * Make sure that signal_pending_state()->signal_pending() below
3174 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3175 * done by the caller to avoid the race with signal_wake_up().
3176 */
3177 smp_mb__before_spinlock();
3178 raw_spin_lock(&rq->lock);
3179 lockdep_pin_lock(&rq->lock);
3180
3181 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3182
3183 switch_count = &prev->nivcsw;
3184 if (!preempt && prev->state) {
3185 if (unlikely(signal_pending_state(prev->state, prev))) {
3186 prev->state = TASK_RUNNING;
3187 } else {
3188 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3189 prev->on_rq = 0;
3190
3191 /*
3192 * If a worker went to sleep, notify and ask workqueue
3193 * whether it wants to wake up a task to maintain
3194 * concurrency.
3195 */
3196 if (prev->flags & PF_WQ_WORKER) {
3197 struct task_struct *to_wakeup;
3198
3199 to_wakeup = wq_worker_sleeping(prev);
3200 if (to_wakeup)
3201 try_to_wake_up_local(to_wakeup);
3202 }
3203 }
3204 switch_count = &prev->nvcsw;
3205 }
3206
3207 if (task_on_rq_queued(prev))
3208 update_rq_clock(rq);
3209
3210 next = pick_next_task(rq, prev);
3211 clear_tsk_need_resched(prev);
3212 clear_preempt_need_resched();
3213 rq->clock_skip_update = 0;
3214
3215 if (likely(prev != next)) {
3216 rq->nr_switches++;
3217 rq->curr = next;
3218 ++*switch_count;
3219
3220 trace_sched_switch(preempt, prev, next);
3221 rq = context_switch(rq, prev, next); /* unlocks the rq */
3222 } else {
3223 lockdep_unpin_lock(&rq->lock);
3224 raw_spin_unlock_irq(&rq->lock);
3225 }
3226
3227 balance_callback(rq);
3228}
3229STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3230
3231static inline void sched_submit_work(struct task_struct *tsk)
3232{
3233 if (!tsk->state || tsk_is_pi_blocked(tsk))
3234 return;
3235 /*
3236 * If we are going to sleep and we have plugged IO queued,
3237 * make sure to submit it to avoid deadlocks.
3238 */
3239 if (blk_needs_flush_plug(tsk))
3240 blk_schedule_flush_plug(tsk);
3241}
3242
3243asmlinkage __visible void __sched schedule(void)
3244{
3245 struct task_struct *tsk = current;
3246
3247 sched_submit_work(tsk);
3248 do {
3249 preempt_disable();
3250 __schedule(false);
3251 sched_preempt_enable_no_resched();
3252 } while (need_resched());
3253}
3254EXPORT_SYMBOL(schedule);
3255
3256#ifdef CONFIG_CONTEXT_TRACKING
3257asmlinkage __visible void __sched schedule_user(void)
3258{
3259 /*
3260 * If we come here after a random call to set_need_resched(),
3261 * or we have been woken up remotely but the IPI has not yet arrived,
3262 * we haven't yet exited the RCU idle mode. Do it here manually until
3263 * we find a better solution.
3264 *
3265 * NB: There are buggy callers of this function. Ideally we
3266 * should warn if prev_state != CONTEXT_USER, but that will trigger
3267 * too frequently to make sense yet.
3268 */
3269 enum ctx_state prev_state = exception_enter();
3270 schedule();
3271 exception_exit(prev_state);
3272}
3273#endif
3274
3275/**
3276 * schedule_preempt_disabled - called with preemption disabled
3277 *
3278 * Returns with preemption disabled. Note: preempt_count must be 1
3279 */
3280void __sched schedule_preempt_disabled(void)
3281{
3282 sched_preempt_enable_no_resched();
3283 schedule();
3284 preempt_disable();
3285}
3286
3287static void __sched notrace preempt_schedule_common(void)
3288{
3289 do {
3290 preempt_disable_notrace();
3291 __schedule(true);
3292 preempt_enable_no_resched_notrace();
3293
3294 /*
3295 * Check again in case we missed a preemption opportunity
3296 * between schedule and now.
3297 */
3298 } while (need_resched());
3299}
3300
3301#ifdef CONFIG_PREEMPT
3302/*
3303 * this is the entry point to schedule() from in-kernel preemption
3304 * off of preempt_enable. Kernel preemptions off return from interrupt
3305 * occur there and call schedule directly.
3306 */
3307asmlinkage __visible void __sched notrace preempt_schedule(void)
3308{
3309 /*
3310 * If there is a non-zero preempt_count or interrupts are disabled,
3311 * we do not want to preempt the current task. Just return..
3312 */
3313 if (likely(!preemptible()))
3314 return;
3315
3316 preempt_schedule_common();
3317}
3318NOKPROBE_SYMBOL(preempt_schedule);
3319EXPORT_SYMBOL(preempt_schedule);
3320
3321/**
3322 * preempt_schedule_notrace - preempt_schedule called by tracing
3323 *
3324 * The tracing infrastructure uses preempt_enable_notrace to prevent
3325 * recursion and tracing preempt enabling caused by the tracing
3326 * infrastructure itself. But as tracing can happen in areas coming
3327 * from userspace or just about to enter userspace, a preempt enable
3328 * can occur before user_exit() is called. This will cause the scheduler
3329 * to be called when the system is still in usermode.
3330 *
3331 * To prevent this, the preempt_enable_notrace will use this function
3332 * instead of preempt_schedule() to exit user context if needed before
3333 * calling the scheduler.
3334 */
3335asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3336{
3337 enum ctx_state prev_ctx;
3338
3339 if (likely(!preemptible()))
3340 return;
3341
3342 do {
3343 preempt_disable_notrace();
3344 /*
3345 * Needs preempt disabled in case user_exit() is traced
3346 * and the tracer calls preempt_enable_notrace() causing
3347 * an infinite recursion.
3348 */
3349 prev_ctx = exception_enter();
3350 __schedule(true);
3351 exception_exit(prev_ctx);
3352
3353 preempt_enable_no_resched_notrace();
3354 } while (need_resched());
3355}
3356EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3357
3358#endif /* CONFIG_PREEMPT */
3359
3360/*
3361 * this is the entry point to schedule() from kernel preemption
3362 * off of irq context.
3363 * Note, that this is called and return with irqs disabled. This will
3364 * protect us against recursive calling from irq.
3365 */
3366asmlinkage __visible void __sched preempt_schedule_irq(void)
3367{
3368 enum ctx_state prev_state;
3369
3370 /* Catch callers which need to be fixed */
3371 BUG_ON(preempt_count() || !irqs_disabled());
3372
3373 prev_state = exception_enter();
3374
3375 do {
3376 preempt_disable();
3377 local_irq_enable();
3378 __schedule(true);
3379 local_irq_disable();
3380 sched_preempt_enable_no_resched();
3381 } while (need_resched());
3382
3383 exception_exit(prev_state);
3384}
3385
3386int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3387 void *key)
3388{
3389 return try_to_wake_up(curr->private, mode, wake_flags);
3390}
3391EXPORT_SYMBOL(default_wake_function);
3392
3393#ifdef CONFIG_RT_MUTEXES
3394
3395/*
3396 * rt_mutex_setprio - set the current priority of a task
3397 * @p: task
3398 * @prio: prio value (kernel-internal form)
3399 *
3400 * This function changes the 'effective' priority of a task. It does
3401 * not touch ->normal_prio like __setscheduler().
3402 *
3403 * Used by the rt_mutex code to implement priority inheritance
3404 * logic. Call site only calls if the priority of the task changed.
3405 */
3406void rt_mutex_setprio(struct task_struct *p, int prio)
3407{
3408 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3409 struct rq *rq;
3410 const struct sched_class *prev_class;
3411
3412 BUG_ON(prio > MAX_PRIO);
3413
3414 rq = __task_rq_lock(p);
3415
3416 /*
3417 * Idle task boosting is a nono in general. There is one
3418 * exception, when PREEMPT_RT and NOHZ is active:
3419 *
3420 * The idle task calls get_next_timer_interrupt() and holds
3421 * the timer wheel base->lock on the CPU and another CPU wants
3422 * to access the timer (probably to cancel it). We can safely
3423 * ignore the boosting request, as the idle CPU runs this code
3424 * with interrupts disabled and will complete the lock
3425 * protected section without being interrupted. So there is no
3426 * real need to boost.
3427 */
3428 if (unlikely(p == rq->idle)) {
3429 WARN_ON(p != rq->curr);
3430 WARN_ON(p->pi_blocked_on);
3431 goto out_unlock;
3432 }
3433
3434 trace_sched_pi_setprio(p, prio);
3435 oldprio = p->prio;
3436
3437 if (oldprio == prio)
3438 queue_flag &= ~DEQUEUE_MOVE;
3439
3440 prev_class = p->sched_class;
3441 queued = task_on_rq_queued(p);
3442 running = task_current(rq, p);
3443 if (queued)
3444 dequeue_task(rq, p, queue_flag);
3445 if (running)
3446 put_prev_task(rq, p);
3447
3448 /*
3449 * Boosting condition are:
3450 * 1. -rt task is running and holds mutex A
3451 * --> -dl task blocks on mutex A
3452 *
3453 * 2. -dl task is running and holds mutex A
3454 * --> -dl task blocks on mutex A and could preempt the
3455 * running task
3456 */
3457 if (dl_prio(prio)) {
3458 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3459 if (!dl_prio(p->normal_prio) ||
3460 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3461 p->dl.dl_boosted = 1;
3462 queue_flag |= ENQUEUE_REPLENISH;
3463 } else
3464 p->dl.dl_boosted = 0;
3465 p->sched_class = &dl_sched_class;
3466 } else if (rt_prio(prio)) {
3467 if (dl_prio(oldprio))
3468 p->dl.dl_boosted = 0;
3469 if (oldprio < prio)
3470 queue_flag |= ENQUEUE_HEAD;
3471 p->sched_class = &rt_sched_class;
3472 } else {
3473 if (dl_prio(oldprio))
3474 p->dl.dl_boosted = 0;
3475 if (rt_prio(oldprio))
3476 p->rt.timeout = 0;
3477 p->sched_class = &fair_sched_class;
3478 }
3479
3480 p->prio = prio;
3481
3482 if (running)
3483 p->sched_class->set_curr_task(rq);
3484 if (queued)
3485 enqueue_task(rq, p, queue_flag);
3486
3487 check_class_changed(rq, p, prev_class, oldprio);
3488out_unlock:
3489 preempt_disable(); /* avoid rq from going away on us */
3490 __task_rq_unlock(rq);
3491
3492 balance_callback(rq);
3493 preempt_enable();
3494}
3495#endif
3496
3497void set_user_nice(struct task_struct *p, long nice)
3498{
3499 int old_prio, delta, queued;
3500 unsigned long flags;
3501 struct rq *rq;
3502
3503 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3504 return;
3505 /*
3506 * We have to be careful, if called from sys_setpriority(),
3507 * the task might be in the middle of scheduling on another CPU.
3508 */
3509 rq = task_rq_lock(p, &flags);
3510 /*
3511 * The RT priorities are set via sched_setscheduler(), but we still
3512 * allow the 'normal' nice value to be set - but as expected
3513 * it wont have any effect on scheduling until the task is
3514 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3515 */
3516 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3517 p->static_prio = NICE_TO_PRIO(nice);
3518 goto out_unlock;
3519 }
3520 queued = task_on_rq_queued(p);
3521 if (queued)
3522 dequeue_task(rq, p, DEQUEUE_SAVE);
3523
3524 p->static_prio = NICE_TO_PRIO(nice);
3525 set_load_weight(p);
3526 old_prio = p->prio;
3527 p->prio = effective_prio(p);
3528 delta = p->prio - old_prio;
3529
3530 if (queued) {
3531 enqueue_task(rq, p, ENQUEUE_RESTORE);
3532 /*
3533 * If the task increased its priority or is running and
3534 * lowered its priority, then reschedule its CPU:
3535 */
3536 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3537 resched_curr(rq);
3538 }
3539out_unlock:
3540 task_rq_unlock(rq, p, &flags);
3541}
3542EXPORT_SYMBOL(set_user_nice);
3543
3544/*
3545 * can_nice - check if a task can reduce its nice value
3546 * @p: task
3547 * @nice: nice value
3548 */
3549int can_nice(const struct task_struct *p, const int nice)
3550{
3551 /* convert nice value [19,-20] to rlimit style value [1,40] */
3552 int nice_rlim = nice_to_rlimit(nice);
3553
3554 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3555 capable(CAP_SYS_NICE));
3556}
3557
3558#ifdef __ARCH_WANT_SYS_NICE
3559
3560/*
3561 * sys_nice - change the priority of the current process.
3562 * @increment: priority increment
3563 *
3564 * sys_setpriority is a more generic, but much slower function that
3565 * does similar things.
3566 */
3567SYSCALL_DEFINE1(nice, int, increment)
3568{
3569 long nice, retval;
3570
3571 /*
3572 * Setpriority might change our priority at the same moment.
3573 * We don't have to worry. Conceptually one call occurs first
3574 * and we have a single winner.
3575 */
3576 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3577 nice = task_nice(current) + increment;
3578
3579 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3580 if (increment < 0 && !can_nice(current, nice))
3581 return -EPERM;
3582
3583 retval = security_task_setnice(current, nice);
3584 if (retval)
3585 return retval;
3586
3587 set_user_nice(current, nice);
3588 return 0;
3589}
3590
3591#endif
3592
3593/**
3594 * task_prio - return the priority value of a given task.
3595 * @p: the task in question.
3596 *
3597 * Return: The priority value as seen by users in /proc.
3598 * RT tasks are offset by -200. Normal tasks are centered
3599 * around 0, value goes from -16 to +15.
3600 */
3601int task_prio(const struct task_struct *p)
3602{
3603 return p->prio - MAX_RT_PRIO;
3604}
3605
3606/**
3607 * idle_cpu - is a given cpu idle currently?
3608 * @cpu: the processor in question.
3609 *
3610 * Return: 1 if the CPU is currently idle. 0 otherwise.
3611 */
3612int idle_cpu(int cpu)
3613{
3614 struct rq *rq = cpu_rq(cpu);
3615
3616 if (rq->curr != rq->idle)
3617 return 0;
3618
3619 if (rq->nr_running)
3620 return 0;
3621
3622#ifdef CONFIG_SMP
3623 if (!llist_empty(&rq->wake_list))
3624 return 0;
3625#endif
3626
3627 return 1;
3628}
3629
3630/**
3631 * idle_task - return the idle task for a given cpu.
3632 * @cpu: the processor in question.
3633 *
3634 * Return: The idle task for the cpu @cpu.
3635 */
3636struct task_struct *idle_task(int cpu)
3637{
3638 return cpu_rq(cpu)->idle;
3639}
3640
3641/**
3642 * find_process_by_pid - find a process with a matching PID value.
3643 * @pid: the pid in question.
3644 *
3645 * The task of @pid, if found. %NULL otherwise.
3646 */
3647static struct task_struct *find_process_by_pid(pid_t pid)
3648{
3649 return pid ? find_task_by_vpid(pid) : current;
3650}
3651
3652/*
3653 * This function initializes the sched_dl_entity of a newly becoming
3654 * SCHED_DEADLINE task.
3655 *
3656 * Only the static values are considered here, the actual runtime and the
3657 * absolute deadline will be properly calculated when the task is enqueued
3658 * for the first time with its new policy.
3659 */
3660static void
3661__setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3662{
3663 struct sched_dl_entity *dl_se = &p->dl;
3664
3665 dl_se->dl_runtime = attr->sched_runtime;
3666 dl_se->dl_deadline = attr->sched_deadline;
3667 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3668 dl_se->flags = attr->sched_flags;
3669 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3670
3671 /*
3672 * Changing the parameters of a task is 'tricky' and we're not doing
3673 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3674 *
3675 * What we SHOULD do is delay the bandwidth release until the 0-lag
3676 * point. This would include retaining the task_struct until that time
3677 * and change dl_overflow() to not immediately decrement the current
3678 * amount.
3679 *
3680 * Instead we retain the current runtime/deadline and let the new
3681 * parameters take effect after the current reservation period lapses.
3682 * This is safe (albeit pessimistic) because the 0-lag point is always
3683 * before the current scheduling deadline.
3684 *
3685 * We can still have temporary overloads because we do not delay the
3686 * change in bandwidth until that time; so admission control is
3687 * not on the safe side. It does however guarantee tasks will never
3688 * consume more than promised.
3689 */
3690}
3691
3692/*
3693 * sched_setparam() passes in -1 for its policy, to let the functions
3694 * it calls know not to change it.
3695 */
3696#define SETPARAM_POLICY -1
3697
3698static void __setscheduler_params(struct task_struct *p,
3699 const struct sched_attr *attr)
3700{
3701 int policy = attr->sched_policy;
3702
3703 if (policy == SETPARAM_POLICY)
3704 policy = p->policy;
3705
3706 p->policy = policy;
3707
3708 if (dl_policy(policy))
3709 __setparam_dl(p, attr);
3710 else if (fair_policy(policy))
3711 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3712
3713 /*
3714 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3715 * !rt_policy. Always setting this ensures that things like
3716 * getparam()/getattr() don't report silly values for !rt tasks.
3717 */
3718 p->rt_priority = attr->sched_priority;
3719 p->normal_prio = normal_prio(p);
3720 set_load_weight(p);
3721}
3722
3723/* Actually do priority change: must hold pi & rq lock. */
3724static void __setscheduler(struct rq *rq, struct task_struct *p,
3725 const struct sched_attr *attr, bool keep_boost)
3726{
3727 __setscheduler_params(p, attr);
3728
3729 /*
3730 * Keep a potential priority boosting if called from
3731 * sched_setscheduler().
3732 */
3733 if (keep_boost)
3734 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3735 else
3736 p->prio = normal_prio(p);
3737
3738 if (dl_prio(p->prio))
3739 p->sched_class = &dl_sched_class;
3740 else if (rt_prio(p->prio))
3741 p->sched_class = &rt_sched_class;
3742 else
3743 p->sched_class = &fair_sched_class;
3744}
3745
3746static void
3747__getparam_dl(struct task_struct *p, struct sched_attr *attr)
3748{
3749 struct sched_dl_entity *dl_se = &p->dl;
3750
3751 attr->sched_priority = p->rt_priority;
3752 attr->sched_runtime = dl_se->dl_runtime;
3753 attr->sched_deadline = dl_se->dl_deadline;
3754 attr->sched_period = dl_se->dl_period;
3755 attr->sched_flags = dl_se->flags;
3756}
3757
3758/*
3759 * This function validates the new parameters of a -deadline task.
3760 * We ask for the deadline not being zero, and greater or equal
3761 * than the runtime, as well as the period of being zero or
3762 * greater than deadline. Furthermore, we have to be sure that
3763 * user parameters are above the internal resolution of 1us (we
3764 * check sched_runtime only since it is always the smaller one) and
3765 * below 2^63 ns (we have to check both sched_deadline and
3766 * sched_period, as the latter can be zero).
3767 */
3768static bool
3769__checkparam_dl(const struct sched_attr *attr)
3770{
3771 /* deadline != 0 */
3772 if (attr->sched_deadline == 0)
3773 return false;
3774
3775 /*
3776 * Since we truncate DL_SCALE bits, make sure we're at least
3777 * that big.
3778 */
3779 if (attr->sched_runtime < (1ULL << DL_SCALE))
3780 return false;
3781
3782 /*
3783 * Since we use the MSB for wrap-around and sign issues, make
3784 * sure it's not set (mind that period can be equal to zero).
3785 */
3786 if (attr->sched_deadline & (1ULL << 63) ||
3787 attr->sched_period & (1ULL << 63))
3788 return false;
3789
3790 /* runtime <= deadline <= period (if period != 0) */
3791 if ((attr->sched_period != 0 &&
3792 attr->sched_period < attr->sched_deadline) ||
3793 attr->sched_deadline < attr->sched_runtime)
3794 return false;
3795
3796 return true;
3797}
3798
3799/*
3800 * check the target process has a UID that matches the current process's
3801 */
3802static bool check_same_owner(struct task_struct *p)
3803{
3804 const struct cred *cred = current_cred(), *pcred;
3805 bool match;
3806
3807 rcu_read_lock();
3808 pcred = __task_cred(p);
3809 match = (uid_eq(cred->euid, pcred->euid) ||
3810 uid_eq(cred->euid, pcred->uid));
3811 rcu_read_unlock();
3812 return match;
3813}
3814
3815static bool dl_param_changed(struct task_struct *p,
3816 const struct sched_attr *attr)
3817{
3818 struct sched_dl_entity *dl_se = &p->dl;
3819
3820 if (dl_se->dl_runtime != attr->sched_runtime ||
3821 dl_se->dl_deadline != attr->sched_deadline ||
3822 dl_se->dl_period != attr->sched_period ||
3823 dl_se->flags != attr->sched_flags)
3824 return true;
3825
3826 return false;
3827}
3828
3829static int __sched_setscheduler(struct task_struct *p,
3830 const struct sched_attr *attr,
3831 bool user, bool pi)
3832{
3833 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3834 MAX_RT_PRIO - 1 - attr->sched_priority;
3835 int retval, oldprio, oldpolicy = -1, queued, running;
3836 int new_effective_prio, policy = attr->sched_policy;
3837 unsigned long flags;
3838 const struct sched_class *prev_class;
3839 struct rq *rq;
3840 int reset_on_fork;
3841 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
3842
3843 /* may grab non-irq protected spin_locks */
3844 BUG_ON(in_interrupt());
3845recheck:
3846 /* double check policy once rq lock held */
3847 if (policy < 0) {
3848 reset_on_fork = p->sched_reset_on_fork;
3849 policy = oldpolicy = p->policy;
3850 } else {
3851 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3852
3853 if (!valid_policy(policy))
3854 return -EINVAL;
3855 }
3856
3857 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3858 return -EINVAL;
3859
3860 /*
3861 * Valid priorities for SCHED_FIFO and SCHED_RR are
3862 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3863 * SCHED_BATCH and SCHED_IDLE is 0.
3864 */
3865 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3866 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3867 return -EINVAL;
3868 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3869 (rt_policy(policy) != (attr->sched_priority != 0)))
3870 return -EINVAL;
3871
3872 /*
3873 * Allow unprivileged RT tasks to decrease priority:
3874 */
3875 if (user && !capable(CAP_SYS_NICE)) {
3876 if (fair_policy(policy)) {
3877 if (attr->sched_nice < task_nice(p) &&
3878 !can_nice(p, attr->sched_nice))
3879 return -EPERM;
3880 }
3881
3882 if (rt_policy(policy)) {
3883 unsigned long rlim_rtprio =
3884 task_rlimit(p, RLIMIT_RTPRIO);
3885
3886 /* can't set/change the rt policy */
3887 if (policy != p->policy && !rlim_rtprio)
3888 return -EPERM;
3889
3890 /* can't increase priority */
3891 if (attr->sched_priority > p->rt_priority &&
3892 attr->sched_priority > rlim_rtprio)
3893 return -EPERM;
3894 }
3895
3896 /*
3897 * Can't set/change SCHED_DEADLINE policy at all for now
3898 * (safest behavior); in the future we would like to allow
3899 * unprivileged DL tasks to increase their relative deadline
3900 * or reduce their runtime (both ways reducing utilization)
3901 */
3902 if (dl_policy(policy))
3903 return -EPERM;
3904
3905 /*
3906 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3907 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3908 */
3909 if (idle_policy(p->policy) && !idle_policy(policy)) {
3910 if (!can_nice(p, task_nice(p)))
3911 return -EPERM;
3912 }
3913
3914 /* can't change other user's priorities */
3915 if (!check_same_owner(p))
3916 return -EPERM;
3917
3918 /* Normal users shall not reset the sched_reset_on_fork flag */
3919 if (p->sched_reset_on_fork && !reset_on_fork)
3920 return -EPERM;
3921 }
3922
3923 if (user) {
3924 retval = security_task_setscheduler(p);
3925 if (retval)
3926 return retval;
3927 }
3928
3929 /*
3930 * make sure no PI-waiters arrive (or leave) while we are
3931 * changing the priority of the task:
3932 *
3933 * To be able to change p->policy safely, the appropriate
3934 * runqueue lock must be held.
3935 */
3936 rq = task_rq_lock(p, &flags);
3937
3938 /*
3939 * Changing the policy of the stop threads its a very bad idea
3940 */
3941 if (p == rq->stop) {
3942 task_rq_unlock(rq, p, &flags);
3943 return -EINVAL;
3944 }
3945
3946 /*
3947 * If not changing anything there's no need to proceed further,
3948 * but store a possible modification of reset_on_fork.
3949 */
3950 if (unlikely(policy == p->policy)) {
3951 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3952 goto change;
3953 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3954 goto change;
3955 if (dl_policy(policy) && dl_param_changed(p, attr))
3956 goto change;
3957
3958 p->sched_reset_on_fork = reset_on_fork;
3959 task_rq_unlock(rq, p, &flags);
3960 return 0;
3961 }
3962change:
3963
3964 if (user) {
3965#ifdef CONFIG_RT_GROUP_SCHED
3966 /*
3967 * Do not allow realtime tasks into groups that have no runtime
3968 * assigned.
3969 */
3970 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3971 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3972 !task_group_is_autogroup(task_group(p))) {
3973 task_rq_unlock(rq, p, &flags);
3974 return -EPERM;
3975 }
3976#endif
3977#ifdef CONFIG_SMP
3978 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3979 cpumask_t *span = rq->rd->span;
3980
3981 /*
3982 * Don't allow tasks with an affinity mask smaller than
3983 * the entire root_domain to become SCHED_DEADLINE. We
3984 * will also fail if there's no bandwidth available.
3985 */
3986 if (!cpumask_subset(span, &p->cpus_allowed) ||
3987 rq->rd->dl_bw.bw == 0) {
3988 task_rq_unlock(rq, p, &flags);
3989 return -EPERM;
3990 }
3991 }
3992#endif
3993 }
3994
3995 /* recheck policy now with rq lock held */
3996 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3997 policy = oldpolicy = -1;
3998 task_rq_unlock(rq, p, &flags);
3999 goto recheck;
4000 }
4001
4002 /*
4003 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4004 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4005 * is available.
4006 */
4007 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4008 task_rq_unlock(rq, p, &flags);
4009 return -EBUSY;
4010 }
4011
4012 p->sched_reset_on_fork = reset_on_fork;
4013 oldprio = p->prio;
4014
4015 if (pi) {
4016 /*
4017 * Take priority boosted tasks into account. If the new
4018 * effective priority is unchanged, we just store the new
4019 * normal parameters and do not touch the scheduler class and
4020 * the runqueue. This will be done when the task deboost
4021 * itself.
4022 */
4023 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4024 if (new_effective_prio == oldprio)
4025 queue_flags &= ~DEQUEUE_MOVE;
4026 }
4027
4028 queued = task_on_rq_queued(p);
4029 running = task_current(rq, p);
4030 if (queued)
4031 dequeue_task(rq, p, queue_flags);
4032 if (running)
4033 put_prev_task(rq, p);
4034
4035 prev_class = p->sched_class;
4036 __setscheduler(rq, p, attr, pi);
4037
4038 if (running)
4039 p->sched_class->set_curr_task(rq);
4040 if (queued) {
4041 /*
4042 * We enqueue to tail when the priority of a task is
4043 * increased (user space view).
4044 */
4045 if (oldprio < p->prio)
4046 queue_flags |= ENQUEUE_HEAD;
4047
4048 enqueue_task(rq, p, queue_flags);
4049 }
4050
4051 check_class_changed(rq, p, prev_class, oldprio);
4052 preempt_disable(); /* avoid rq from going away on us */
4053 task_rq_unlock(rq, p, &flags);
4054
4055 if (pi)
4056 rt_mutex_adjust_pi(p);
4057
4058 /*
4059 * Run balance callbacks after we've adjusted the PI chain.
4060 */
4061 balance_callback(rq);
4062 preempt_enable();
4063
4064 return 0;
4065}
4066
4067static int _sched_setscheduler(struct task_struct *p, int policy,
4068 const struct sched_param *param, bool check)
4069{
4070 struct sched_attr attr = {
4071 .sched_policy = policy,
4072 .sched_priority = param->sched_priority,
4073 .sched_nice = PRIO_TO_NICE(p->static_prio),
4074 };
4075
4076 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4077 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4078 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4079 policy &= ~SCHED_RESET_ON_FORK;
4080 attr.sched_policy = policy;
4081 }
4082
4083 return __sched_setscheduler(p, &attr, check, true);
4084}
4085/**
4086 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4087 * @p: the task in question.
4088 * @policy: new policy.
4089 * @param: structure containing the new RT priority.
4090 *
4091 * Return: 0 on success. An error code otherwise.
4092 *
4093 * NOTE that the task may be already dead.
4094 */
4095int sched_setscheduler(struct task_struct *p, int policy,
4096 const struct sched_param *param)
4097{
4098 return _sched_setscheduler(p, policy, param, true);
4099}
4100EXPORT_SYMBOL_GPL(sched_setscheduler);
4101
4102int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4103{
4104 return __sched_setscheduler(p, attr, true, true);
4105}
4106EXPORT_SYMBOL_GPL(sched_setattr);
4107
4108/**
4109 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4110 * @p: the task in question.
4111 * @policy: new policy.
4112 * @param: structure containing the new RT priority.
4113 *
4114 * Just like sched_setscheduler, only don't bother checking if the
4115 * current context has permission. For example, this is needed in
4116 * stop_machine(): we create temporary high priority worker threads,
4117 * but our caller might not have that capability.
4118 *
4119 * Return: 0 on success. An error code otherwise.
4120 */
4121int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4122 const struct sched_param *param)
4123{
4124 return _sched_setscheduler(p, policy, param, false);
4125}
4126EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4127
4128static int
4129do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4130{
4131 struct sched_param lparam;
4132 struct task_struct *p;
4133 int retval;
4134
4135 if (!param || pid < 0)
4136 return -EINVAL;
4137 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4138 return -EFAULT;
4139
4140 rcu_read_lock();
4141 retval = -ESRCH;
4142 p = find_process_by_pid(pid);
4143 if (p != NULL)
4144 retval = sched_setscheduler(p, policy, &lparam);
4145 rcu_read_unlock();
4146
4147 return retval;
4148}
4149
4150/*
4151 * Mimics kernel/events/core.c perf_copy_attr().
4152 */
4153static int sched_copy_attr(struct sched_attr __user *uattr,
4154 struct sched_attr *attr)
4155{
4156 u32 size;
4157 int ret;
4158
4159 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4160 return -EFAULT;
4161
4162 /*
4163 * zero the full structure, so that a short copy will be nice.
4164 */
4165 memset(attr, 0, sizeof(*attr));
4166
4167 ret = get_user(size, &uattr->size);
4168 if (ret)
4169 return ret;
4170
4171 if (size > PAGE_SIZE) /* silly large */
4172 goto err_size;
4173
4174 if (!size) /* abi compat */
4175 size = SCHED_ATTR_SIZE_VER0;
4176
4177 if (size < SCHED_ATTR_SIZE_VER0)
4178 goto err_size;
4179
4180 /*
4181 * If we're handed a bigger struct than we know of,
4182 * ensure all the unknown bits are 0 - i.e. new
4183 * user-space does not rely on any kernel feature
4184 * extensions we dont know about yet.
4185 */
4186 if (size > sizeof(*attr)) {
4187 unsigned char __user *addr;
4188 unsigned char __user *end;
4189 unsigned char val;
4190
4191 addr = (void __user *)uattr + sizeof(*attr);
4192 end = (void __user *)uattr + size;
4193
4194 for (; addr < end; addr++) {
4195 ret = get_user(val, addr);
4196 if (ret)
4197 return ret;
4198 if (val)
4199 goto err_size;
4200 }
4201 size = sizeof(*attr);
4202 }
4203
4204 ret = copy_from_user(attr, uattr, size);
4205 if (ret)
4206 return -EFAULT;
4207
4208 /*
4209 * XXX: do we want to be lenient like existing syscalls; or do we want
4210 * to be strict and return an error on out-of-bounds values?
4211 */
4212 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4213
4214 return 0;
4215
4216err_size:
4217 put_user(sizeof(*attr), &uattr->size);
4218 return -E2BIG;
4219}
4220
4221/**
4222 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4223 * @pid: the pid in question.
4224 * @policy: new policy.
4225 * @param: structure containing the new RT priority.
4226 *
4227 * Return: 0 on success. An error code otherwise.
4228 */
4229SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4230 struct sched_param __user *, param)
4231{
4232 /* negative values for policy are not valid */
4233 if (policy < 0)
4234 return -EINVAL;
4235
4236 return do_sched_setscheduler(pid, policy, param);
4237}
4238
4239/**
4240 * sys_sched_setparam - set/change the RT priority of a thread
4241 * @pid: the pid in question.
4242 * @param: structure containing the new RT priority.
4243 *
4244 * Return: 0 on success. An error code otherwise.
4245 */
4246SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4247{
4248 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4249}
4250
4251/**
4252 * sys_sched_setattr - same as above, but with extended sched_attr
4253 * @pid: the pid in question.
4254 * @uattr: structure containing the extended parameters.
4255 * @flags: for future extension.
4256 */
4257SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4258 unsigned int, flags)
4259{
4260 struct sched_attr attr;
4261 struct task_struct *p;
4262 int retval;
4263
4264 if (!uattr || pid < 0 || flags)
4265 return -EINVAL;
4266
4267 retval = sched_copy_attr(uattr, &attr);
4268 if (retval)
4269 return retval;
4270
4271 if ((int)attr.sched_policy < 0)
4272 return -EINVAL;
4273
4274 rcu_read_lock();
4275 retval = -ESRCH;
4276 p = find_process_by_pid(pid);
4277 if (p != NULL)
4278 retval = sched_setattr(p, &attr);
4279 rcu_read_unlock();
4280
4281 return retval;
4282}
4283
4284/**
4285 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4286 * @pid: the pid in question.
4287 *
4288 * Return: On success, the policy of the thread. Otherwise, a negative error
4289 * code.
4290 */
4291SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4292{
4293 struct task_struct *p;
4294 int retval;
4295
4296 if (pid < 0)
4297 return -EINVAL;
4298
4299 retval = -ESRCH;
4300 rcu_read_lock();
4301 p = find_process_by_pid(pid);
4302 if (p) {
4303 retval = security_task_getscheduler(p);
4304 if (!retval)
4305 retval = p->policy
4306 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4307 }
4308 rcu_read_unlock();
4309 return retval;
4310}
4311
4312/**
4313 * sys_sched_getparam - get the RT priority of a thread
4314 * @pid: the pid in question.
4315 * @param: structure containing the RT priority.
4316 *
4317 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4318 * code.
4319 */
4320SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4321{
4322 struct sched_param lp = { .sched_priority = 0 };
4323 struct task_struct *p;
4324 int retval;
4325
4326 if (!param || pid < 0)
4327 return -EINVAL;
4328
4329 rcu_read_lock();
4330 p = find_process_by_pid(pid);
4331 retval = -ESRCH;
4332 if (!p)
4333 goto out_unlock;
4334
4335 retval = security_task_getscheduler(p);
4336 if (retval)
4337 goto out_unlock;
4338
4339 if (task_has_rt_policy(p))
4340 lp.sched_priority = p->rt_priority;
4341 rcu_read_unlock();
4342
4343 /*
4344 * This one might sleep, we cannot do it with a spinlock held ...
4345 */
4346 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4347
4348 return retval;
4349
4350out_unlock:
4351 rcu_read_unlock();
4352 return retval;
4353}
4354
4355static int sched_read_attr(struct sched_attr __user *uattr,
4356 struct sched_attr *attr,
4357 unsigned int usize)
4358{
4359 int ret;
4360
4361 if (!access_ok(VERIFY_WRITE, uattr, usize))
4362 return -EFAULT;
4363
4364 /*
4365 * If we're handed a smaller struct than we know of,
4366 * ensure all the unknown bits are 0 - i.e. old
4367 * user-space does not get uncomplete information.
4368 */
4369 if (usize < sizeof(*attr)) {
4370 unsigned char *addr;
4371 unsigned char *end;
4372
4373 addr = (void *)attr + usize;
4374 end = (void *)attr + sizeof(*attr);
4375
4376 for (; addr < end; addr++) {
4377 if (*addr)
4378 return -EFBIG;
4379 }
4380
4381 attr->size = usize;
4382 }
4383
4384 ret = copy_to_user(uattr, attr, attr->size);
4385 if (ret)
4386 return -EFAULT;
4387
4388 return 0;
4389}
4390
4391/**
4392 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4393 * @pid: the pid in question.
4394 * @uattr: structure containing the extended parameters.
4395 * @size: sizeof(attr) for fwd/bwd comp.
4396 * @flags: for future extension.
4397 */
4398SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4399 unsigned int, size, unsigned int, flags)
4400{
4401 struct sched_attr attr = {
4402 .size = sizeof(struct sched_attr),
4403 };
4404 struct task_struct *p;
4405 int retval;
4406
4407 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4408 size < SCHED_ATTR_SIZE_VER0 || flags)
4409 return -EINVAL;
4410
4411 rcu_read_lock();
4412 p = find_process_by_pid(pid);
4413 retval = -ESRCH;
4414 if (!p)
4415 goto out_unlock;
4416
4417 retval = security_task_getscheduler(p);
4418 if (retval)
4419 goto out_unlock;
4420
4421 attr.sched_policy = p->policy;
4422 if (p->sched_reset_on_fork)
4423 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4424 if (task_has_dl_policy(p))
4425 __getparam_dl(p, &attr);
4426 else if (task_has_rt_policy(p))
4427 attr.sched_priority = p->rt_priority;
4428 else
4429 attr.sched_nice = task_nice(p);
4430
4431 rcu_read_unlock();
4432
4433 retval = sched_read_attr(uattr, &attr, size);
4434 return retval;
4435
4436out_unlock:
4437 rcu_read_unlock();
4438 return retval;
4439}
4440
4441long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4442{
4443 cpumask_var_t cpus_allowed, new_mask;
4444 struct task_struct *p;
4445 int retval;
4446
4447 rcu_read_lock();
4448
4449 p = find_process_by_pid(pid);
4450 if (!p) {
4451 rcu_read_unlock();
4452 return -ESRCH;
4453 }
4454
4455 /* Prevent p going away */
4456 get_task_struct(p);
4457 rcu_read_unlock();
4458
4459 if (p->flags & PF_NO_SETAFFINITY) {
4460 retval = -EINVAL;
4461 goto out_put_task;
4462 }
4463 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4464 retval = -ENOMEM;
4465 goto out_put_task;
4466 }
4467 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4468 retval = -ENOMEM;
4469 goto out_free_cpus_allowed;
4470 }
4471 retval = -EPERM;
4472 if (!check_same_owner(p)) {
4473 rcu_read_lock();
4474 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4475 rcu_read_unlock();
4476 goto out_free_new_mask;
4477 }
4478 rcu_read_unlock();
4479 }
4480
4481 retval = security_task_setscheduler(p);
4482 if (retval)
4483 goto out_free_new_mask;
4484
4485
4486 cpuset_cpus_allowed(p, cpus_allowed);
4487 cpumask_and(new_mask, in_mask, cpus_allowed);
4488
4489 /*
4490 * Since bandwidth control happens on root_domain basis,
4491 * if admission test is enabled, we only admit -deadline
4492 * tasks allowed to run on all the CPUs in the task's
4493 * root_domain.
4494 */
4495#ifdef CONFIG_SMP
4496 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4497 rcu_read_lock();
4498 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4499 retval = -EBUSY;
4500 rcu_read_unlock();
4501 goto out_free_new_mask;
4502 }
4503 rcu_read_unlock();
4504 }
4505#endif
4506again:
4507 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4508
4509 if (!retval) {
4510 cpuset_cpus_allowed(p, cpus_allowed);
4511 if (!cpumask_subset(new_mask, cpus_allowed)) {
4512 /*
4513 * We must have raced with a concurrent cpuset
4514 * update. Just reset the cpus_allowed to the
4515 * cpuset's cpus_allowed
4516 */
4517 cpumask_copy(new_mask, cpus_allowed);
4518 goto again;
4519 }
4520 }
4521out_free_new_mask:
4522 free_cpumask_var(new_mask);
4523out_free_cpus_allowed:
4524 free_cpumask_var(cpus_allowed);
4525out_put_task:
4526 put_task_struct(p);
4527 return retval;
4528}
4529
4530static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4531 struct cpumask *new_mask)
4532{
4533 if (len < cpumask_size())
4534 cpumask_clear(new_mask);
4535 else if (len > cpumask_size())
4536 len = cpumask_size();
4537
4538 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4539}
4540
4541/**
4542 * sys_sched_setaffinity - set the cpu affinity of a process
4543 * @pid: pid of the process
4544 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4545 * @user_mask_ptr: user-space pointer to the new cpu mask
4546 *
4547 * Return: 0 on success. An error code otherwise.
4548 */
4549SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4550 unsigned long __user *, user_mask_ptr)
4551{
4552 cpumask_var_t new_mask;
4553 int retval;
4554
4555 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4556 return -ENOMEM;
4557
4558 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4559 if (retval == 0)
4560 retval = sched_setaffinity(pid, new_mask);
4561 free_cpumask_var(new_mask);
4562 return retval;
4563}
4564
4565long sched_getaffinity(pid_t pid, struct cpumask *mask)
4566{
4567 struct task_struct *p;
4568 unsigned long flags;
4569 int retval;
4570
4571 rcu_read_lock();
4572
4573 retval = -ESRCH;
4574 p = find_process_by_pid(pid);
4575 if (!p)
4576 goto out_unlock;
4577
4578 retval = security_task_getscheduler(p);
4579 if (retval)
4580 goto out_unlock;
4581
4582 raw_spin_lock_irqsave(&p->pi_lock, flags);
4583 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4584 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4585
4586out_unlock:
4587 rcu_read_unlock();
4588
4589 return retval;
4590}
4591
4592/**
4593 * sys_sched_getaffinity - get the cpu affinity of a process
4594 * @pid: pid of the process
4595 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4596 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4597 *
4598 * Return: 0 on success. An error code otherwise.
4599 */
4600SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4601 unsigned long __user *, user_mask_ptr)
4602{
4603 int ret;
4604 cpumask_var_t mask;
4605
4606 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4607 return -EINVAL;
4608 if (len & (sizeof(unsigned long)-1))
4609 return -EINVAL;
4610
4611 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4612 return -ENOMEM;
4613
4614 ret = sched_getaffinity(pid, mask);
4615 if (ret == 0) {
4616 size_t retlen = min_t(size_t, len, cpumask_size());
4617
4618 if (copy_to_user(user_mask_ptr, mask, retlen))
4619 ret = -EFAULT;
4620 else
4621 ret = retlen;
4622 }
4623 free_cpumask_var(mask);
4624
4625 return ret;
4626}
4627
4628/**
4629 * sys_sched_yield - yield the current processor to other threads.
4630 *
4631 * This function yields the current CPU to other tasks. If there are no
4632 * other threads running on this CPU then this function will return.
4633 *
4634 * Return: 0.
4635 */
4636SYSCALL_DEFINE0(sched_yield)
4637{
4638 struct rq *rq = this_rq_lock();
4639
4640 schedstat_inc(rq, yld_count);
4641 current->sched_class->yield_task(rq);
4642
4643 /*
4644 * Since we are going to call schedule() anyway, there's
4645 * no need to preempt or enable interrupts:
4646 */
4647 __release(rq->lock);
4648 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4649 do_raw_spin_unlock(&rq->lock);
4650 sched_preempt_enable_no_resched();
4651
4652 schedule();
4653
4654 return 0;
4655}
4656
4657int __sched _cond_resched(void)
4658{
4659 if (should_resched(0)) {
4660 preempt_schedule_common();
4661 return 1;
4662 }
4663 return 0;
4664}
4665EXPORT_SYMBOL(_cond_resched);
4666
4667/*
4668 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4669 * call schedule, and on return reacquire the lock.
4670 *
4671 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4672 * operations here to prevent schedule() from being called twice (once via
4673 * spin_unlock(), once by hand).
4674 */
4675int __cond_resched_lock(spinlock_t *lock)
4676{
4677 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4678 int ret = 0;
4679
4680 lockdep_assert_held(lock);
4681
4682 if (spin_needbreak(lock) || resched) {
4683 spin_unlock(lock);
4684 if (resched)
4685 preempt_schedule_common();
4686 else
4687 cpu_relax();
4688 ret = 1;
4689 spin_lock(lock);
4690 }
4691 return ret;
4692}
4693EXPORT_SYMBOL(__cond_resched_lock);
4694
4695int __sched __cond_resched_softirq(void)
4696{
4697 BUG_ON(!in_softirq());
4698
4699 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4700 local_bh_enable();
4701 preempt_schedule_common();
4702 local_bh_disable();
4703 return 1;
4704 }
4705 return 0;
4706}
4707EXPORT_SYMBOL(__cond_resched_softirq);
4708
4709/**
4710 * yield - yield the current processor to other threads.
4711 *
4712 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4713 *
4714 * The scheduler is at all times free to pick the calling task as the most
4715 * eligible task to run, if removing the yield() call from your code breaks
4716 * it, its already broken.
4717 *
4718 * Typical broken usage is:
4719 *
4720 * while (!event)
4721 * yield();
4722 *
4723 * where one assumes that yield() will let 'the other' process run that will
4724 * make event true. If the current task is a SCHED_FIFO task that will never
4725 * happen. Never use yield() as a progress guarantee!!
4726 *
4727 * If you want to use yield() to wait for something, use wait_event().
4728 * If you want to use yield() to be 'nice' for others, use cond_resched().
4729 * If you still want to use yield(), do not!
4730 */
4731void __sched yield(void)
4732{
4733 set_current_state(TASK_RUNNING);
4734 sys_sched_yield();
4735}
4736EXPORT_SYMBOL(yield);
4737
4738/**
4739 * yield_to - yield the current processor to another thread in
4740 * your thread group, or accelerate that thread toward the
4741 * processor it's on.
4742 * @p: target task
4743 * @preempt: whether task preemption is allowed or not
4744 *
4745 * It's the caller's job to ensure that the target task struct
4746 * can't go away on us before we can do any checks.
4747 *
4748 * Return:
4749 * true (>0) if we indeed boosted the target task.
4750 * false (0) if we failed to boost the target.
4751 * -ESRCH if there's no task to yield to.
4752 */
4753int __sched yield_to(struct task_struct *p, bool preempt)
4754{
4755 struct task_struct *curr = current;
4756 struct rq *rq, *p_rq;
4757 unsigned long flags;
4758 int yielded = 0;
4759
4760 local_irq_save(flags);
4761 rq = this_rq();
4762
4763again:
4764 p_rq = task_rq(p);
4765 /*
4766 * If we're the only runnable task on the rq and target rq also
4767 * has only one task, there's absolutely no point in yielding.
4768 */
4769 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4770 yielded = -ESRCH;
4771 goto out_irq;
4772 }
4773
4774 double_rq_lock(rq, p_rq);
4775 if (task_rq(p) != p_rq) {
4776 double_rq_unlock(rq, p_rq);
4777 goto again;
4778 }
4779
4780 if (!curr->sched_class->yield_to_task)
4781 goto out_unlock;
4782
4783 if (curr->sched_class != p->sched_class)
4784 goto out_unlock;
4785
4786 if (task_running(p_rq, p) || p->state)
4787 goto out_unlock;
4788
4789 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4790 if (yielded) {
4791 schedstat_inc(rq, yld_count);
4792 /*
4793 * Make p's CPU reschedule; pick_next_entity takes care of
4794 * fairness.
4795 */
4796 if (preempt && rq != p_rq)
4797 resched_curr(p_rq);
4798 }
4799
4800out_unlock:
4801 double_rq_unlock(rq, p_rq);
4802out_irq:
4803 local_irq_restore(flags);
4804
4805 if (yielded > 0)
4806 schedule();
4807
4808 return yielded;
4809}
4810EXPORT_SYMBOL_GPL(yield_to);
4811
4812/*
4813 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4814 * that process accounting knows that this is a task in IO wait state.
4815 */
4816long __sched io_schedule_timeout(long timeout)
4817{
4818 int old_iowait = current->in_iowait;
4819 struct rq *rq;
4820 long ret;
4821
4822 current->in_iowait = 1;
4823 blk_schedule_flush_plug(current);
4824
4825 delayacct_blkio_start();
4826 rq = raw_rq();
4827 atomic_inc(&rq->nr_iowait);
4828 ret = schedule_timeout(timeout);
4829 current->in_iowait = old_iowait;
4830 atomic_dec(&rq->nr_iowait);
4831 delayacct_blkio_end();
4832
4833 return ret;
4834}
4835EXPORT_SYMBOL(io_schedule_timeout);
4836
4837/**
4838 * sys_sched_get_priority_max - return maximum RT priority.
4839 * @policy: scheduling class.
4840 *
4841 * Return: On success, this syscall returns the maximum
4842 * rt_priority that can be used by a given scheduling class.
4843 * On failure, a negative error code is returned.
4844 */
4845SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4846{
4847 int ret = -EINVAL;
4848
4849 switch (policy) {
4850 case SCHED_FIFO:
4851 case SCHED_RR:
4852 ret = MAX_USER_RT_PRIO-1;
4853 break;
4854 case SCHED_DEADLINE:
4855 case SCHED_NORMAL:
4856 case SCHED_BATCH:
4857 case SCHED_IDLE:
4858 ret = 0;
4859 break;
4860 }
4861 return ret;
4862}
4863
4864/**
4865 * sys_sched_get_priority_min - return minimum RT priority.
4866 * @policy: scheduling class.
4867 *
4868 * Return: On success, this syscall returns the minimum
4869 * rt_priority that can be used by a given scheduling class.
4870 * On failure, a negative error code is returned.
4871 */
4872SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4873{
4874 int ret = -EINVAL;
4875
4876 switch (policy) {
4877 case SCHED_FIFO:
4878 case SCHED_RR:
4879 ret = 1;
4880 break;
4881 case SCHED_DEADLINE:
4882 case SCHED_NORMAL:
4883 case SCHED_BATCH:
4884 case SCHED_IDLE:
4885 ret = 0;
4886 }
4887 return ret;
4888}
4889
4890/**
4891 * sys_sched_rr_get_interval - return the default timeslice of a process.
4892 * @pid: pid of the process.
4893 * @interval: userspace pointer to the timeslice value.
4894 *
4895 * this syscall writes the default timeslice value of a given process
4896 * into the user-space timespec buffer. A value of '0' means infinity.
4897 *
4898 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4899 * an error code.
4900 */
4901SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4902 struct timespec __user *, interval)
4903{
4904 struct task_struct *p;
4905 unsigned int time_slice;
4906 unsigned long flags;
4907 struct rq *rq;
4908 int retval;
4909 struct timespec t;
4910
4911 if (pid < 0)
4912 return -EINVAL;
4913
4914 retval = -ESRCH;
4915 rcu_read_lock();
4916 p = find_process_by_pid(pid);
4917 if (!p)
4918 goto out_unlock;
4919
4920 retval = security_task_getscheduler(p);
4921 if (retval)
4922 goto out_unlock;
4923
4924 rq = task_rq_lock(p, &flags);
4925 time_slice = 0;
4926 if (p->sched_class->get_rr_interval)
4927 time_slice = p->sched_class->get_rr_interval(rq, p);
4928 task_rq_unlock(rq, p, &flags);
4929
4930 rcu_read_unlock();
4931 jiffies_to_timespec(time_slice, &t);
4932 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4933 return retval;
4934
4935out_unlock:
4936 rcu_read_unlock();
4937 return retval;
4938}
4939
4940static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4941
4942void sched_show_task(struct task_struct *p)
4943{
4944 unsigned long free = 0;
4945 int ppid;
4946 unsigned long state = p->state;
4947
4948 if (state)
4949 state = __ffs(state) + 1;
4950 printk(KERN_INFO "%-15.15s %c", p->comm,
4951 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4952#if BITS_PER_LONG == 32
4953 if (state == TASK_RUNNING)
4954 printk(KERN_CONT " running ");
4955 else
4956 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4957#else
4958 if (state == TASK_RUNNING)
4959 printk(KERN_CONT " running task ");
4960 else
4961 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4962#endif
4963#ifdef CONFIG_DEBUG_STACK_USAGE
4964 free = stack_not_used(p);
4965#endif
4966 ppid = 0;
4967 rcu_read_lock();
4968 if (pid_alive(p))
4969 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4970 rcu_read_unlock();
4971 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4972 task_pid_nr(p), ppid,
4973 (unsigned long)task_thread_info(p)->flags);
4974
4975 print_worker_info(KERN_INFO, p);
4976 show_stack(p, NULL);
4977}
4978
4979void show_state_filter(unsigned long state_filter)
4980{
4981 struct task_struct *g, *p;
4982
4983#if BITS_PER_LONG == 32
4984 printk(KERN_INFO
4985 " task PC stack pid father\n");
4986#else
4987 printk(KERN_INFO
4988 " task PC stack pid father\n");
4989#endif
4990 rcu_read_lock();
4991 for_each_process_thread(g, p) {
4992 /*
4993 * reset the NMI-timeout, listing all files on a slow
4994 * console might take a lot of time:
4995 */
4996 touch_nmi_watchdog();
4997 if (!state_filter || (p->state & state_filter))
4998 sched_show_task(p);
4999 }
5000
5001 touch_all_softlockup_watchdogs();
5002
5003#ifdef CONFIG_SCHED_DEBUG
5004 sysrq_sched_debug_show();
5005#endif
5006 rcu_read_unlock();
5007 /*
5008 * Only show locks if all tasks are dumped:
5009 */
5010 if (!state_filter)
5011 debug_show_all_locks();
5012}
5013
5014void init_idle_bootup_task(struct task_struct *idle)
5015{
5016 idle->sched_class = &idle_sched_class;
5017}
5018
5019/**
5020 * init_idle - set up an idle thread for a given CPU
5021 * @idle: task in question
5022 * @cpu: cpu the idle task belongs to
5023 *
5024 * NOTE: this function does not set the idle thread's NEED_RESCHED
5025 * flag, to make booting more robust.
5026 */
5027void init_idle(struct task_struct *idle, int cpu)
5028{
5029 struct rq *rq = cpu_rq(cpu);
5030 unsigned long flags;
5031
5032 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5033 raw_spin_lock(&rq->lock);
5034
5035 __sched_fork(0, idle);
5036 idle->state = TASK_RUNNING;
5037 idle->se.exec_start = sched_clock();
5038
5039 kasan_unpoison_task_stack(idle);
5040
5041#ifdef CONFIG_SMP
5042 /*
5043 * Its possible that init_idle() gets called multiple times on a task,
5044 * in that case do_set_cpus_allowed() will not do the right thing.
5045 *
5046 * And since this is boot we can forgo the serialization.
5047 */
5048 set_cpus_allowed_common(idle, cpumask_of(cpu));
5049#endif
5050 /*
5051 * We're having a chicken and egg problem, even though we are
5052 * holding rq->lock, the cpu isn't yet set to this cpu so the
5053 * lockdep check in task_group() will fail.
5054 *
5055 * Similar case to sched_fork(). / Alternatively we could
5056 * use task_rq_lock() here and obtain the other rq->lock.
5057 *
5058 * Silence PROVE_RCU
5059 */
5060 rcu_read_lock();
5061 __set_task_cpu(idle, cpu);
5062 rcu_read_unlock();
5063
5064 rq->curr = rq->idle = idle;
5065 idle->on_rq = TASK_ON_RQ_QUEUED;
5066#ifdef CONFIG_SMP
5067 idle->on_cpu = 1;
5068#endif
5069 raw_spin_unlock(&rq->lock);
5070 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5071
5072 /* Set the preempt count _outside_ the spinlocks! */
5073 init_idle_preempt_count(idle, cpu);
5074
5075 /*
5076 * The idle tasks have their own, simple scheduling class:
5077 */
5078 idle->sched_class = &idle_sched_class;
5079 ftrace_graph_init_idle_task(idle, cpu);
5080 vtime_init_idle(idle, cpu);
5081#ifdef CONFIG_SMP
5082 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5083#endif
5084}
5085
5086int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5087 const struct cpumask *trial)
5088{
5089 int ret = 1, trial_cpus;
5090 struct dl_bw *cur_dl_b;
5091 unsigned long flags;
5092
5093 if (!cpumask_weight(cur))
5094 return ret;
5095
5096 rcu_read_lock_sched();
5097 cur_dl_b = dl_bw_of(cpumask_any(cur));
5098 trial_cpus = cpumask_weight(trial);
5099
5100 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5101 if (cur_dl_b->bw != -1 &&
5102 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5103 ret = 0;
5104 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5105 rcu_read_unlock_sched();
5106
5107 return ret;
5108}
5109
5110int task_can_attach(struct task_struct *p,
5111 const struct cpumask *cs_cpus_allowed)
5112{
5113 int ret = 0;
5114
5115 /*
5116 * Kthreads which disallow setaffinity shouldn't be moved
5117 * to a new cpuset; we don't want to change their cpu
5118 * affinity and isolating such threads by their set of
5119 * allowed nodes is unnecessary. Thus, cpusets are not
5120 * applicable for such threads. This prevents checking for
5121 * success of set_cpus_allowed_ptr() on all attached tasks
5122 * before cpus_allowed may be changed.
5123 */
5124 if (p->flags & PF_NO_SETAFFINITY) {
5125 ret = -EINVAL;
5126 goto out;
5127 }
5128
5129#ifdef CONFIG_SMP
5130 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5131 cs_cpus_allowed)) {
5132 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5133 cs_cpus_allowed);
5134 struct dl_bw *dl_b;
5135 bool overflow;
5136 int cpus;
5137 unsigned long flags;
5138
5139 rcu_read_lock_sched();
5140 dl_b = dl_bw_of(dest_cpu);
5141 raw_spin_lock_irqsave(&dl_b->lock, flags);
5142 cpus = dl_bw_cpus(dest_cpu);
5143 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5144 if (overflow)
5145 ret = -EBUSY;
5146 else {
5147 /*
5148 * We reserve space for this task in the destination
5149 * root_domain, as we can't fail after this point.
5150 * We will free resources in the source root_domain
5151 * later on (see set_cpus_allowed_dl()).
5152 */
5153 __dl_add(dl_b, p->dl.dl_bw);
5154 }
5155 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5156 rcu_read_unlock_sched();
5157
5158 }
5159#endif
5160out:
5161 return ret;
5162}
5163
5164#ifdef CONFIG_SMP
5165
5166#ifdef CONFIG_NUMA_BALANCING
5167/* Migrate current task p to target_cpu */
5168int migrate_task_to(struct task_struct *p, int target_cpu)
5169{
5170 struct migration_arg arg = { p, target_cpu };
5171 int curr_cpu = task_cpu(p);
5172
5173 if (curr_cpu == target_cpu)
5174 return 0;
5175
5176 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5177 return -EINVAL;
5178
5179 /* TODO: This is not properly updating schedstats */
5180
5181 trace_sched_move_numa(p, curr_cpu, target_cpu);
5182 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5183}
5184
5185/*
5186 * Requeue a task on a given node and accurately track the number of NUMA
5187 * tasks on the runqueues
5188 */
5189void sched_setnuma(struct task_struct *p, int nid)
5190{
5191 struct rq *rq;
5192 unsigned long flags;
5193 bool queued, running;
5194
5195 rq = task_rq_lock(p, &flags);
5196 queued = task_on_rq_queued(p);
5197 running = task_current(rq, p);
5198
5199 if (queued)
5200 dequeue_task(rq, p, DEQUEUE_SAVE);
5201 if (running)
5202 put_prev_task(rq, p);
5203
5204 p->numa_preferred_nid = nid;
5205
5206 if (running)
5207 p->sched_class->set_curr_task(rq);
5208 if (queued)
5209 enqueue_task(rq, p, ENQUEUE_RESTORE);
5210 task_rq_unlock(rq, p, &flags);
5211}
5212#endif /* CONFIG_NUMA_BALANCING */
5213
5214#ifdef CONFIG_HOTPLUG_CPU
5215/*
5216 * Ensures that the idle task is using init_mm right before its cpu goes
5217 * offline.
5218 */
5219void idle_task_exit(void)
5220{
5221 struct mm_struct *mm = current->active_mm;
5222
5223 BUG_ON(cpu_online(smp_processor_id()));
5224
5225 if (mm != &init_mm) {
5226 switch_mm(mm, &init_mm, current);
5227 finish_arch_post_lock_switch();
5228 }
5229 mmdrop(mm);
5230}
5231
5232/*
5233 * Since this CPU is going 'away' for a while, fold any nr_active delta
5234 * we might have. Assumes we're called after migrate_tasks() so that the
5235 * nr_active count is stable.
5236 *
5237 * Also see the comment "Global load-average calculations".
5238 */
5239static void calc_load_migrate(struct rq *rq)
5240{
5241 long delta = calc_load_fold_active(rq);
5242 if (delta)
5243 atomic_long_add(delta, &calc_load_tasks);
5244}
5245
5246static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5247{
5248}
5249
5250static const struct sched_class fake_sched_class = {
5251 .put_prev_task = put_prev_task_fake,
5252};
5253
5254static struct task_struct fake_task = {
5255 /*
5256 * Avoid pull_{rt,dl}_task()
5257 */
5258 .prio = MAX_PRIO + 1,
5259 .sched_class = &fake_sched_class,
5260};
5261
5262/*
5263 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5264 * try_to_wake_up()->select_task_rq().
5265 *
5266 * Called with rq->lock held even though we'er in stop_machine() and
5267 * there's no concurrency possible, we hold the required locks anyway
5268 * because of lock validation efforts.
5269 */
5270static void migrate_tasks(struct rq *dead_rq)
5271{
5272 struct rq *rq = dead_rq;
5273 struct task_struct *next, *stop = rq->stop;
5274 int dest_cpu;
5275
5276 /*
5277 * Fudge the rq selection such that the below task selection loop
5278 * doesn't get stuck on the currently eligible stop task.
5279 *
5280 * We're currently inside stop_machine() and the rq is either stuck
5281 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5282 * either way we should never end up calling schedule() until we're
5283 * done here.
5284 */
5285 rq->stop = NULL;
5286
5287 /*
5288 * put_prev_task() and pick_next_task() sched
5289 * class method both need to have an up-to-date
5290 * value of rq->clock[_task]
5291 */
5292 update_rq_clock(rq);
5293
5294 for (;;) {
5295 /*
5296 * There's this thread running, bail when that's the only
5297 * remaining thread.
5298 */
5299 if (rq->nr_running == 1)
5300 break;
5301
5302 /*
5303 * pick_next_task assumes pinned rq->lock.
5304 */
5305 lockdep_pin_lock(&rq->lock);
5306 next = pick_next_task(rq, &fake_task);
5307 BUG_ON(!next);
5308 next->sched_class->put_prev_task(rq, next);
5309
5310 /*
5311 * Rules for changing task_struct::cpus_allowed are holding
5312 * both pi_lock and rq->lock, such that holding either
5313 * stabilizes the mask.
5314 *
5315 * Drop rq->lock is not quite as disastrous as it usually is
5316 * because !cpu_active at this point, which means load-balance
5317 * will not interfere. Also, stop-machine.
5318 */
5319 lockdep_unpin_lock(&rq->lock);
5320 raw_spin_unlock(&rq->lock);
5321 raw_spin_lock(&next->pi_lock);
5322 raw_spin_lock(&rq->lock);
5323
5324 /*
5325 * Since we're inside stop-machine, _nothing_ should have
5326 * changed the task, WARN if weird stuff happened, because in
5327 * that case the above rq->lock drop is a fail too.
5328 */
5329 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5330 raw_spin_unlock(&next->pi_lock);
5331 continue;
5332 }
5333
5334 /* Find suitable destination for @next, with force if needed. */
5335 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5336
5337 rq = __migrate_task(rq, next, dest_cpu);
5338 if (rq != dead_rq) {
5339 raw_spin_unlock(&rq->lock);
5340 rq = dead_rq;
5341 raw_spin_lock(&rq->lock);
5342 }
5343 raw_spin_unlock(&next->pi_lock);
5344 }
5345
5346 rq->stop = stop;
5347}
5348#endif /* CONFIG_HOTPLUG_CPU */
5349
5350static void set_rq_online(struct rq *rq)
5351{
5352 if (!rq->online) {
5353 const struct sched_class *class;
5354
5355 cpumask_set_cpu(rq->cpu, rq->rd->online);
5356 rq->online = 1;
5357
5358 for_each_class(class) {
5359 if (class->rq_online)
5360 class->rq_online(rq);
5361 }
5362 }
5363}
5364
5365static void set_rq_offline(struct rq *rq)
5366{
5367 if (rq->online) {
5368 const struct sched_class *class;
5369
5370 for_each_class(class) {
5371 if (class->rq_offline)
5372 class->rq_offline(rq);
5373 }
5374
5375 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5376 rq->online = 0;
5377 }
5378}
5379
5380/*
5381 * migration_call - callback that gets triggered when a CPU is added.
5382 * Here we can start up the necessary migration thread for the new CPU.
5383 */
5384static int
5385migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5386{
5387 int cpu = (long)hcpu;
5388 unsigned long flags;
5389 struct rq *rq = cpu_rq(cpu);
5390
5391 switch (action & ~CPU_TASKS_FROZEN) {
5392
5393 case CPU_UP_PREPARE:
5394 rq->calc_load_update = calc_load_update;
5395 account_reset_rq(rq);
5396 break;
5397
5398 case CPU_ONLINE:
5399 /* Update our root-domain */
5400 raw_spin_lock_irqsave(&rq->lock, flags);
5401 if (rq->rd) {
5402 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5403
5404 set_rq_online(rq);
5405 }
5406 raw_spin_unlock_irqrestore(&rq->lock, flags);
5407 break;
5408
5409#ifdef CONFIG_HOTPLUG_CPU
5410 case CPU_DYING:
5411 sched_ttwu_pending();
5412 /* Update our root-domain */
5413 raw_spin_lock_irqsave(&rq->lock, flags);
5414 if (rq->rd) {
5415 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5416 set_rq_offline(rq);
5417 }
5418 migrate_tasks(rq);
5419 BUG_ON(rq->nr_running != 1); /* the migration thread */
5420 raw_spin_unlock_irqrestore(&rq->lock, flags);
5421 break;
5422
5423 case CPU_DEAD:
5424 calc_load_migrate(rq);
5425 break;
5426#endif
5427 }
5428
5429 update_max_interval();
5430
5431 return NOTIFY_OK;
5432}
5433
5434/*
5435 * Register at high priority so that task migration (migrate_all_tasks)
5436 * happens before everything else. This has to be lower priority than
5437 * the notifier in the perf_event subsystem, though.
5438 */
5439static struct notifier_block migration_notifier = {
5440 .notifier_call = migration_call,
5441 .priority = CPU_PRI_MIGRATION,
5442};
5443
5444static void set_cpu_rq_start_time(void)
5445{
5446 int cpu = smp_processor_id();
5447 struct rq *rq = cpu_rq(cpu);
5448 rq->age_stamp = sched_clock_cpu(cpu);
5449}
5450
5451static int sched_cpu_active(struct notifier_block *nfb,
5452 unsigned long action, void *hcpu)
5453{
5454 int cpu = (long)hcpu;
5455
5456 switch (action & ~CPU_TASKS_FROZEN) {
5457 case CPU_STARTING:
5458 set_cpu_rq_start_time();
5459 return NOTIFY_OK;
5460
5461 case CPU_DOWN_FAILED:
5462 set_cpu_active(cpu, true);
5463 return NOTIFY_OK;
5464
5465 default:
5466 return NOTIFY_DONE;
5467 }
5468}
5469
5470static int sched_cpu_inactive(struct notifier_block *nfb,
5471 unsigned long action, void *hcpu)
5472{
5473 switch (action & ~CPU_TASKS_FROZEN) {
5474 case CPU_DOWN_PREPARE:
5475 set_cpu_active((long)hcpu, false);
5476 return NOTIFY_OK;
5477 default:
5478 return NOTIFY_DONE;
5479 }
5480}
5481
5482static int __init migration_init(void)
5483{
5484 void *cpu = (void *)(long)smp_processor_id();
5485 int err;
5486
5487 /* Initialize migration for the boot CPU */
5488 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5489 BUG_ON(err == NOTIFY_BAD);
5490 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5491 register_cpu_notifier(&migration_notifier);
5492
5493 /* Register cpu active notifiers */
5494 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5495 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5496
5497 return 0;
5498}
5499early_initcall(migration_init);
5500
5501static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5502
5503#ifdef CONFIG_SCHED_DEBUG
5504
5505static __read_mostly int sched_debug_enabled;
5506
5507static int __init sched_debug_setup(char *str)
5508{
5509 sched_debug_enabled = 1;
5510
5511 return 0;
5512}
5513early_param("sched_debug", sched_debug_setup);
5514
5515static inline bool sched_debug(void)
5516{
5517 return sched_debug_enabled;
5518}
5519
5520static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5521 struct cpumask *groupmask)
5522{
5523 struct sched_group *group = sd->groups;
5524
5525 cpumask_clear(groupmask);
5526
5527 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5528
5529 if (!(sd->flags & SD_LOAD_BALANCE)) {
5530 printk("does not load-balance\n");
5531 if (sd->parent)
5532 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5533 " has parent");
5534 return -1;
5535 }
5536
5537 printk(KERN_CONT "span %*pbl level %s\n",
5538 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5539
5540 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5541 printk(KERN_ERR "ERROR: domain->span does not contain "
5542 "CPU%d\n", cpu);
5543 }
5544 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5545 printk(KERN_ERR "ERROR: domain->groups does not contain"
5546 " CPU%d\n", cpu);
5547 }
5548
5549 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5550 do {
5551 if (!group) {
5552 printk("\n");
5553 printk(KERN_ERR "ERROR: group is NULL\n");
5554 break;
5555 }
5556
5557 if (!cpumask_weight(sched_group_cpus(group))) {
5558 printk(KERN_CONT "\n");
5559 printk(KERN_ERR "ERROR: empty group\n");
5560 break;
5561 }
5562
5563 if (!(sd->flags & SD_OVERLAP) &&
5564 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5565 printk(KERN_CONT "\n");
5566 printk(KERN_ERR "ERROR: repeated CPUs\n");
5567 break;
5568 }
5569
5570 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5571
5572 printk(KERN_CONT " %*pbl",
5573 cpumask_pr_args(sched_group_cpus(group)));
5574 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5575 printk(KERN_CONT " (cpu_capacity = %d)",
5576 group->sgc->capacity);
5577 }
5578
5579 group = group->next;
5580 } while (group != sd->groups);
5581 printk(KERN_CONT "\n");
5582
5583 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5584 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5585
5586 if (sd->parent &&
5587 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5588 printk(KERN_ERR "ERROR: parent span is not a superset "
5589 "of domain->span\n");
5590 return 0;
5591}
5592
5593static void sched_domain_debug(struct sched_domain *sd, int cpu)
5594{
5595 int level = 0;
5596
5597 if (!sched_debug_enabled)
5598 return;
5599
5600 if (!sd) {
5601 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5602 return;
5603 }
5604
5605 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5606
5607 for (;;) {
5608 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5609 break;
5610 level++;
5611 sd = sd->parent;
5612 if (!sd)
5613 break;
5614 }
5615}
5616#else /* !CONFIG_SCHED_DEBUG */
5617# define sched_domain_debug(sd, cpu) do { } while (0)
5618static inline bool sched_debug(void)
5619{
5620 return false;
5621}
5622#endif /* CONFIG_SCHED_DEBUG */
5623
5624static int sd_degenerate(struct sched_domain *sd)
5625{
5626 if (cpumask_weight(sched_domain_span(sd)) == 1)
5627 return 1;
5628
5629 /* Following flags need at least 2 groups */
5630 if (sd->flags & (SD_LOAD_BALANCE |
5631 SD_BALANCE_NEWIDLE |
5632 SD_BALANCE_FORK |
5633 SD_BALANCE_EXEC |
5634 SD_SHARE_CPUCAPACITY |
5635 SD_SHARE_PKG_RESOURCES |
5636 SD_SHARE_POWERDOMAIN)) {
5637 if (sd->groups != sd->groups->next)
5638 return 0;
5639 }
5640
5641 /* Following flags don't use groups */
5642 if (sd->flags & (SD_WAKE_AFFINE))
5643 return 0;
5644
5645 return 1;
5646}
5647
5648static int
5649sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5650{
5651 unsigned long cflags = sd->flags, pflags = parent->flags;
5652
5653 if (sd_degenerate(parent))
5654 return 1;
5655
5656 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5657 return 0;
5658
5659 /* Flags needing groups don't count if only 1 group in parent */
5660 if (parent->groups == parent->groups->next) {
5661 pflags &= ~(SD_LOAD_BALANCE |
5662 SD_BALANCE_NEWIDLE |
5663 SD_BALANCE_FORK |
5664 SD_BALANCE_EXEC |
5665 SD_SHARE_CPUCAPACITY |
5666 SD_SHARE_PKG_RESOURCES |
5667 SD_PREFER_SIBLING |
5668 SD_SHARE_POWERDOMAIN);
5669 if (nr_node_ids == 1)
5670 pflags &= ~SD_SERIALIZE;
5671 }
5672 if (~cflags & pflags)
5673 return 0;
5674
5675 return 1;
5676}
5677
5678static void free_rootdomain(struct rcu_head *rcu)
5679{
5680 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5681
5682 cpupri_cleanup(&rd->cpupri);
5683 cpudl_cleanup(&rd->cpudl);
5684 free_cpumask_var(rd->dlo_mask);
5685 free_cpumask_var(rd->rto_mask);
5686 free_cpumask_var(rd->online);
5687 free_cpumask_var(rd->span);
5688 kfree(rd);
5689}
5690
5691static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5692{
5693 struct root_domain *old_rd = NULL;
5694 unsigned long flags;
5695
5696 raw_spin_lock_irqsave(&rq->lock, flags);
5697
5698 if (rq->rd) {
5699 old_rd = rq->rd;
5700
5701 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5702 set_rq_offline(rq);
5703
5704 cpumask_clear_cpu(rq->cpu, old_rd->span);
5705
5706 /*
5707 * If we dont want to free the old_rd yet then
5708 * set old_rd to NULL to skip the freeing later
5709 * in this function:
5710 */
5711 if (!atomic_dec_and_test(&old_rd->refcount))
5712 old_rd = NULL;
5713 }
5714
5715 atomic_inc(&rd->refcount);
5716 rq->rd = rd;
5717
5718 cpumask_set_cpu(rq->cpu, rd->span);
5719 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5720 set_rq_online(rq);
5721
5722 raw_spin_unlock_irqrestore(&rq->lock, flags);
5723
5724 if (old_rd)
5725 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5726}
5727
5728static int init_rootdomain(struct root_domain *rd)
5729{
5730 memset(rd, 0, sizeof(*rd));
5731
5732 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5733 goto out;
5734 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5735 goto free_span;
5736 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5737 goto free_online;
5738 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5739 goto free_dlo_mask;
5740
5741 init_dl_bw(&rd->dl_bw);
5742 if (cpudl_init(&rd->cpudl) != 0)
5743 goto free_dlo_mask;
5744
5745 if (cpupri_init(&rd->cpupri) != 0)
5746 goto free_rto_mask;
5747 return 0;
5748
5749free_rto_mask:
5750 free_cpumask_var(rd->rto_mask);
5751free_dlo_mask:
5752 free_cpumask_var(rd->dlo_mask);
5753free_online:
5754 free_cpumask_var(rd->online);
5755free_span:
5756 free_cpumask_var(rd->span);
5757out:
5758 return -ENOMEM;
5759}
5760
5761/*
5762 * By default the system creates a single root-domain with all cpus as
5763 * members (mimicking the global state we have today).
5764 */
5765struct root_domain def_root_domain;
5766
5767static void init_defrootdomain(void)
5768{
5769 init_rootdomain(&def_root_domain);
5770
5771 atomic_set(&def_root_domain.refcount, 1);
5772}
5773
5774static struct root_domain *alloc_rootdomain(void)
5775{
5776 struct root_domain *rd;
5777
5778 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5779 if (!rd)
5780 return NULL;
5781
5782 if (init_rootdomain(rd) != 0) {
5783 kfree(rd);
5784 return NULL;
5785 }
5786
5787 return rd;
5788}
5789
5790static void free_sched_groups(struct sched_group *sg, int free_sgc)
5791{
5792 struct sched_group *tmp, *first;
5793
5794 if (!sg)
5795 return;
5796
5797 first = sg;
5798 do {
5799 tmp = sg->next;
5800
5801 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5802 kfree(sg->sgc);
5803
5804 kfree(sg);
5805 sg = tmp;
5806 } while (sg != first);
5807}
5808
5809static void free_sched_domain(struct rcu_head *rcu)
5810{
5811 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5812
5813 /*
5814 * If its an overlapping domain it has private groups, iterate and
5815 * nuke them all.
5816 */
5817 if (sd->flags & SD_OVERLAP) {
5818 free_sched_groups(sd->groups, 1);
5819 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5820 kfree(sd->groups->sgc);
5821 kfree(sd->groups);
5822 }
5823 kfree(sd);
5824}
5825
5826static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5827{
5828 call_rcu(&sd->rcu, free_sched_domain);
5829}
5830
5831static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5832{
5833 for (; sd; sd = sd->parent)
5834 destroy_sched_domain(sd, cpu);
5835}
5836
5837/*
5838 * Keep a special pointer to the highest sched_domain that has
5839 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5840 * allows us to avoid some pointer chasing select_idle_sibling().
5841 *
5842 * Also keep a unique ID per domain (we use the first cpu number in
5843 * the cpumask of the domain), this allows us to quickly tell if
5844 * two cpus are in the same cache domain, see cpus_share_cache().
5845 */
5846DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5847DEFINE_PER_CPU(int, sd_llc_size);
5848DEFINE_PER_CPU(int, sd_llc_id);
5849DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5850DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5851DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5852
5853static void update_top_cache_domain(int cpu)
5854{
5855 struct sched_domain *sd;
5856 struct sched_domain *busy_sd = NULL;
5857 int id = cpu;
5858 int size = 1;
5859
5860 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5861 if (sd) {
5862 id = cpumask_first(sched_domain_span(sd));
5863 size = cpumask_weight(sched_domain_span(sd));
5864 busy_sd = sd->parent; /* sd_busy */
5865 }
5866 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5867
5868 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5869 per_cpu(sd_llc_size, cpu) = size;
5870 per_cpu(sd_llc_id, cpu) = id;
5871
5872 sd = lowest_flag_domain(cpu, SD_NUMA);
5873 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5874
5875 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5876 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5877}
5878
5879/*
5880 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5881 * hold the hotplug lock.
5882 */
5883static void
5884cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5885{
5886 struct rq *rq = cpu_rq(cpu);
5887 struct sched_domain *tmp;
5888
5889 /* Remove the sched domains which do not contribute to scheduling. */
5890 for (tmp = sd; tmp; ) {
5891 struct sched_domain *parent = tmp->parent;
5892 if (!parent)
5893 break;
5894
5895 if (sd_parent_degenerate(tmp, parent)) {
5896 tmp->parent = parent->parent;
5897 if (parent->parent)
5898 parent->parent->child = tmp;
5899 /*
5900 * Transfer SD_PREFER_SIBLING down in case of a
5901 * degenerate parent; the spans match for this
5902 * so the property transfers.
5903 */
5904 if (parent->flags & SD_PREFER_SIBLING)
5905 tmp->flags |= SD_PREFER_SIBLING;
5906 destroy_sched_domain(parent, cpu);
5907 } else
5908 tmp = tmp->parent;
5909 }
5910
5911 if (sd && sd_degenerate(sd)) {
5912 tmp = sd;
5913 sd = sd->parent;
5914 destroy_sched_domain(tmp, cpu);
5915 if (sd)
5916 sd->child = NULL;
5917 }
5918
5919 sched_domain_debug(sd, cpu);
5920
5921 rq_attach_root(rq, rd);
5922 tmp = rq->sd;
5923 rcu_assign_pointer(rq->sd, sd);
5924 destroy_sched_domains(tmp, cpu);
5925
5926 update_top_cache_domain(cpu);
5927}
5928
5929/* Setup the mask of cpus configured for isolated domains */
5930static int __init isolated_cpu_setup(char *str)
5931{
5932 int ret;
5933
5934 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5935 ret = cpulist_parse(str, cpu_isolated_map);
5936 if (ret) {
5937 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
5938 return 0;
5939 }
5940 return 1;
5941}
5942__setup("isolcpus=", isolated_cpu_setup);
5943
5944struct s_data {
5945 struct sched_domain ** __percpu sd;
5946 struct root_domain *rd;
5947};
5948
5949enum s_alloc {
5950 sa_rootdomain,
5951 sa_sd,
5952 sa_sd_storage,
5953 sa_none,
5954};
5955
5956/*
5957 * Build an iteration mask that can exclude certain CPUs from the upwards
5958 * domain traversal.
5959 *
5960 * Asymmetric node setups can result in situations where the domain tree is of
5961 * unequal depth, make sure to skip domains that already cover the entire
5962 * range.
5963 *
5964 * In that case build_sched_domains() will have terminated the iteration early
5965 * and our sibling sd spans will be empty. Domains should always include the
5966 * cpu they're built on, so check that.
5967 *
5968 */
5969static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5970{
5971 const struct cpumask *span = sched_domain_span(sd);
5972 struct sd_data *sdd = sd->private;
5973 struct sched_domain *sibling;
5974 int i;
5975
5976 for_each_cpu(i, span) {
5977 sibling = *per_cpu_ptr(sdd->sd, i);
5978 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5979 continue;
5980
5981 cpumask_set_cpu(i, sched_group_mask(sg));
5982 }
5983}
5984
5985/*
5986 * Return the canonical balance cpu for this group, this is the first cpu
5987 * of this group that's also in the iteration mask.
5988 */
5989int group_balance_cpu(struct sched_group *sg)
5990{
5991 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5992}
5993
5994static int
5995build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5996{
5997 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5998 const struct cpumask *span = sched_domain_span(sd);
5999 struct cpumask *covered = sched_domains_tmpmask;
6000 struct sd_data *sdd = sd->private;
6001 struct sched_domain *sibling;
6002 int i;
6003
6004 cpumask_clear(covered);
6005
6006 for_each_cpu(i, span) {
6007 struct cpumask *sg_span;
6008
6009 if (cpumask_test_cpu(i, covered))
6010 continue;
6011
6012 sibling = *per_cpu_ptr(sdd->sd, i);
6013
6014 /* See the comment near build_group_mask(). */
6015 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6016 continue;
6017
6018 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6019 GFP_KERNEL, cpu_to_node(cpu));
6020
6021 if (!sg)
6022 goto fail;
6023
6024 sg_span = sched_group_cpus(sg);
6025 if (sibling->child)
6026 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6027 else
6028 cpumask_set_cpu(i, sg_span);
6029
6030 cpumask_or(covered, covered, sg_span);
6031
6032 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6033 if (atomic_inc_return(&sg->sgc->ref) == 1)
6034 build_group_mask(sd, sg);
6035
6036 /*
6037 * Initialize sgc->capacity such that even if we mess up the
6038 * domains and no possible iteration will get us here, we won't
6039 * die on a /0 trap.
6040 */
6041 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6042
6043 /*
6044 * Make sure the first group of this domain contains the
6045 * canonical balance cpu. Otherwise the sched_domain iteration
6046 * breaks. See update_sg_lb_stats().
6047 */
6048 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6049 group_balance_cpu(sg) == cpu)
6050 groups = sg;
6051
6052 if (!first)
6053 first = sg;
6054 if (last)
6055 last->next = sg;
6056 last = sg;
6057 last->next = first;
6058 }
6059 sd->groups = groups;
6060
6061 return 0;
6062
6063fail:
6064 free_sched_groups(first, 0);
6065
6066 return -ENOMEM;
6067}
6068
6069static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6070{
6071 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6072 struct sched_domain *child = sd->child;
6073
6074 if (child)
6075 cpu = cpumask_first(sched_domain_span(child));
6076
6077 if (sg) {
6078 *sg = *per_cpu_ptr(sdd->sg, cpu);
6079 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6080 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6081 }
6082
6083 return cpu;
6084}
6085
6086/*
6087 * build_sched_groups will build a circular linked list of the groups
6088 * covered by the given span, and will set each group's ->cpumask correctly,
6089 * and ->cpu_capacity to 0.
6090 *
6091 * Assumes the sched_domain tree is fully constructed
6092 */
6093static int
6094build_sched_groups(struct sched_domain *sd, int cpu)
6095{
6096 struct sched_group *first = NULL, *last = NULL;
6097 struct sd_data *sdd = sd->private;
6098 const struct cpumask *span = sched_domain_span(sd);
6099 struct cpumask *covered;
6100 int i;
6101
6102 get_group(cpu, sdd, &sd->groups);
6103 atomic_inc(&sd->groups->ref);
6104
6105 if (cpu != cpumask_first(span))
6106 return 0;
6107
6108 lockdep_assert_held(&sched_domains_mutex);
6109 covered = sched_domains_tmpmask;
6110
6111 cpumask_clear(covered);
6112
6113 for_each_cpu(i, span) {
6114 struct sched_group *sg;
6115 int group, j;
6116
6117 if (cpumask_test_cpu(i, covered))
6118 continue;
6119
6120 group = get_group(i, sdd, &sg);
6121 cpumask_setall(sched_group_mask(sg));
6122
6123 for_each_cpu(j, span) {
6124 if (get_group(j, sdd, NULL) != group)
6125 continue;
6126
6127 cpumask_set_cpu(j, covered);
6128 cpumask_set_cpu(j, sched_group_cpus(sg));
6129 }
6130
6131 if (!first)
6132 first = sg;
6133 if (last)
6134 last->next = sg;
6135 last = sg;
6136 }
6137 last->next = first;
6138
6139 return 0;
6140}
6141
6142/*
6143 * Initialize sched groups cpu_capacity.
6144 *
6145 * cpu_capacity indicates the capacity of sched group, which is used while
6146 * distributing the load between different sched groups in a sched domain.
6147 * Typically cpu_capacity for all the groups in a sched domain will be same
6148 * unless there are asymmetries in the topology. If there are asymmetries,
6149 * group having more cpu_capacity will pickup more load compared to the
6150 * group having less cpu_capacity.
6151 */
6152static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6153{
6154 struct sched_group *sg = sd->groups;
6155
6156 WARN_ON(!sg);
6157
6158 do {
6159 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6160 sg = sg->next;
6161 } while (sg != sd->groups);
6162
6163 if (cpu != group_balance_cpu(sg))
6164 return;
6165
6166 update_group_capacity(sd, cpu);
6167 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6168}
6169
6170/*
6171 * Initializers for schedule domains
6172 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6173 */
6174
6175static int default_relax_domain_level = -1;
6176int sched_domain_level_max;
6177
6178static int __init setup_relax_domain_level(char *str)
6179{
6180 if (kstrtoint(str, 0, &default_relax_domain_level))
6181 pr_warn("Unable to set relax_domain_level\n");
6182
6183 return 1;
6184}
6185__setup("relax_domain_level=", setup_relax_domain_level);
6186
6187static void set_domain_attribute(struct sched_domain *sd,
6188 struct sched_domain_attr *attr)
6189{
6190 int request;
6191
6192 if (!attr || attr->relax_domain_level < 0) {
6193 if (default_relax_domain_level < 0)
6194 return;
6195 else
6196 request = default_relax_domain_level;
6197 } else
6198 request = attr->relax_domain_level;
6199 if (request < sd->level) {
6200 /* turn off idle balance on this domain */
6201 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6202 } else {
6203 /* turn on idle balance on this domain */
6204 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6205 }
6206}
6207
6208static void __sdt_free(const struct cpumask *cpu_map);
6209static int __sdt_alloc(const struct cpumask *cpu_map);
6210
6211static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6212 const struct cpumask *cpu_map)
6213{
6214 switch (what) {
6215 case sa_rootdomain:
6216 if (!atomic_read(&d->rd->refcount))
6217 free_rootdomain(&d->rd->rcu); /* fall through */
6218 case sa_sd:
6219 free_percpu(d->sd); /* fall through */
6220 case sa_sd_storage:
6221 __sdt_free(cpu_map); /* fall through */
6222 case sa_none:
6223 break;
6224 }
6225}
6226
6227static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6228 const struct cpumask *cpu_map)
6229{
6230 memset(d, 0, sizeof(*d));
6231
6232 if (__sdt_alloc(cpu_map))
6233 return sa_sd_storage;
6234 d->sd = alloc_percpu(struct sched_domain *);
6235 if (!d->sd)
6236 return sa_sd_storage;
6237 d->rd = alloc_rootdomain();
6238 if (!d->rd)
6239 return sa_sd;
6240 return sa_rootdomain;
6241}
6242
6243/*
6244 * NULL the sd_data elements we've used to build the sched_domain and
6245 * sched_group structure so that the subsequent __free_domain_allocs()
6246 * will not free the data we're using.
6247 */
6248static void claim_allocations(int cpu, struct sched_domain *sd)
6249{
6250 struct sd_data *sdd = sd->private;
6251
6252 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6253 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6254
6255 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6256 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6257
6258 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6259 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6260}
6261
6262#ifdef CONFIG_NUMA
6263static int sched_domains_numa_levels;
6264enum numa_topology_type sched_numa_topology_type;
6265static int *sched_domains_numa_distance;
6266int sched_max_numa_distance;
6267static struct cpumask ***sched_domains_numa_masks;
6268static int sched_domains_curr_level;
6269#endif
6270
6271/*
6272 * SD_flags allowed in topology descriptions.
6273 *
6274 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6275 * SD_SHARE_PKG_RESOURCES - describes shared caches
6276 * SD_NUMA - describes NUMA topologies
6277 * SD_SHARE_POWERDOMAIN - describes shared power domain
6278 *
6279 * Odd one out:
6280 * SD_ASYM_PACKING - describes SMT quirks
6281 */
6282#define TOPOLOGY_SD_FLAGS \
6283 (SD_SHARE_CPUCAPACITY | \
6284 SD_SHARE_PKG_RESOURCES | \
6285 SD_NUMA | \
6286 SD_ASYM_PACKING | \
6287 SD_SHARE_POWERDOMAIN)
6288
6289static struct sched_domain *
6290sd_init(struct sched_domain_topology_level *tl, int cpu)
6291{
6292 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6293 int sd_weight, sd_flags = 0;
6294
6295#ifdef CONFIG_NUMA
6296 /*
6297 * Ugly hack to pass state to sd_numa_mask()...
6298 */
6299 sched_domains_curr_level = tl->numa_level;
6300#endif
6301
6302 sd_weight = cpumask_weight(tl->mask(cpu));
6303
6304 if (tl->sd_flags)
6305 sd_flags = (*tl->sd_flags)();
6306 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6307 "wrong sd_flags in topology description\n"))
6308 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6309
6310 *sd = (struct sched_domain){
6311 .min_interval = sd_weight,
6312 .max_interval = 2*sd_weight,
6313 .busy_factor = 32,
6314 .imbalance_pct = 125,
6315
6316 .cache_nice_tries = 0,
6317 .busy_idx = 0,
6318 .idle_idx = 0,
6319 .newidle_idx = 0,
6320 .wake_idx = 0,
6321 .forkexec_idx = 0,
6322
6323 .flags = 1*SD_LOAD_BALANCE
6324 | 1*SD_BALANCE_NEWIDLE
6325 | 1*SD_BALANCE_EXEC
6326 | 1*SD_BALANCE_FORK
6327 | 0*SD_BALANCE_WAKE
6328 | 1*SD_WAKE_AFFINE
6329 | 0*SD_SHARE_CPUCAPACITY
6330 | 0*SD_SHARE_PKG_RESOURCES
6331 | 0*SD_SERIALIZE
6332 | 0*SD_PREFER_SIBLING
6333 | 0*SD_NUMA
6334 | sd_flags
6335 ,
6336
6337 .last_balance = jiffies,
6338 .balance_interval = sd_weight,
6339 .smt_gain = 0,
6340 .max_newidle_lb_cost = 0,
6341 .next_decay_max_lb_cost = jiffies,
6342#ifdef CONFIG_SCHED_DEBUG
6343 .name = tl->name,
6344#endif
6345 };
6346
6347 /*
6348 * Convert topological properties into behaviour.
6349 */
6350
6351 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6352 sd->flags |= SD_PREFER_SIBLING;
6353 sd->imbalance_pct = 110;
6354 sd->smt_gain = 1178; /* ~15% */
6355
6356 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6357 sd->imbalance_pct = 117;
6358 sd->cache_nice_tries = 1;
6359 sd->busy_idx = 2;
6360
6361#ifdef CONFIG_NUMA
6362 } else if (sd->flags & SD_NUMA) {
6363 sd->cache_nice_tries = 2;
6364 sd->busy_idx = 3;
6365 sd->idle_idx = 2;
6366
6367 sd->flags |= SD_SERIALIZE;
6368 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6369 sd->flags &= ~(SD_BALANCE_EXEC |
6370 SD_BALANCE_FORK |
6371 SD_WAKE_AFFINE);
6372 }
6373
6374#endif
6375 } else {
6376 sd->flags |= SD_PREFER_SIBLING;
6377 sd->cache_nice_tries = 1;
6378 sd->busy_idx = 2;
6379 sd->idle_idx = 1;
6380 }
6381
6382 sd->private = &tl->data;
6383
6384 return sd;
6385}
6386
6387/*
6388 * Topology list, bottom-up.
6389 */
6390static struct sched_domain_topology_level default_topology[] = {
6391#ifdef CONFIG_SCHED_SMT
6392 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6393#endif
6394#ifdef CONFIG_SCHED_MC
6395 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6396#endif
6397 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6398 { NULL, },
6399};
6400
6401static struct sched_domain_topology_level *sched_domain_topology =
6402 default_topology;
6403
6404#define for_each_sd_topology(tl) \
6405 for (tl = sched_domain_topology; tl->mask; tl++)
6406
6407void set_sched_topology(struct sched_domain_topology_level *tl)
6408{
6409 sched_domain_topology = tl;
6410}
6411
6412#ifdef CONFIG_NUMA
6413
6414static const struct cpumask *sd_numa_mask(int cpu)
6415{
6416 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6417}
6418
6419static void sched_numa_warn(const char *str)
6420{
6421 static int done = false;
6422 int i,j;
6423
6424 if (done)
6425 return;
6426
6427 done = true;
6428
6429 printk(KERN_WARNING "ERROR: %s\n\n", str);
6430
6431 for (i = 0; i < nr_node_ids; i++) {
6432 printk(KERN_WARNING " ");
6433 for (j = 0; j < nr_node_ids; j++)
6434 printk(KERN_CONT "%02d ", node_distance(i,j));
6435 printk(KERN_CONT "\n");
6436 }
6437 printk(KERN_WARNING "\n");
6438}
6439
6440bool find_numa_distance(int distance)
6441{
6442 int i;
6443
6444 if (distance == node_distance(0, 0))
6445 return true;
6446
6447 for (i = 0; i < sched_domains_numa_levels; i++) {
6448 if (sched_domains_numa_distance[i] == distance)
6449 return true;
6450 }
6451
6452 return false;
6453}
6454
6455/*
6456 * A system can have three types of NUMA topology:
6457 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6458 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6459 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6460 *
6461 * The difference between a glueless mesh topology and a backplane
6462 * topology lies in whether communication between not directly
6463 * connected nodes goes through intermediary nodes (where programs
6464 * could run), or through backplane controllers. This affects
6465 * placement of programs.
6466 *
6467 * The type of topology can be discerned with the following tests:
6468 * - If the maximum distance between any nodes is 1 hop, the system
6469 * is directly connected.
6470 * - If for two nodes A and B, located N > 1 hops away from each other,
6471 * there is an intermediary node C, which is < N hops away from both
6472 * nodes A and B, the system is a glueless mesh.
6473 */
6474static void init_numa_topology_type(void)
6475{
6476 int a, b, c, n;
6477
6478 n = sched_max_numa_distance;
6479
6480 if (sched_domains_numa_levels <= 1) {
6481 sched_numa_topology_type = NUMA_DIRECT;
6482 return;
6483 }
6484
6485 for_each_online_node(a) {
6486 for_each_online_node(b) {
6487 /* Find two nodes furthest removed from each other. */
6488 if (node_distance(a, b) < n)
6489 continue;
6490
6491 /* Is there an intermediary node between a and b? */
6492 for_each_online_node(c) {
6493 if (node_distance(a, c) < n &&
6494 node_distance(b, c) < n) {
6495 sched_numa_topology_type =
6496 NUMA_GLUELESS_MESH;
6497 return;
6498 }
6499 }
6500
6501 sched_numa_topology_type = NUMA_BACKPLANE;
6502 return;
6503 }
6504 }
6505}
6506
6507static void sched_init_numa(void)
6508{
6509 int next_distance, curr_distance = node_distance(0, 0);
6510 struct sched_domain_topology_level *tl;
6511 int level = 0;
6512 int i, j, k;
6513
6514 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6515 if (!sched_domains_numa_distance)
6516 return;
6517
6518 /*
6519 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6520 * unique distances in the node_distance() table.
6521 *
6522 * Assumes node_distance(0,j) includes all distances in
6523 * node_distance(i,j) in order to avoid cubic time.
6524 */
6525 next_distance = curr_distance;
6526 for (i = 0; i < nr_node_ids; i++) {
6527 for (j = 0; j < nr_node_ids; j++) {
6528 for (k = 0; k < nr_node_ids; k++) {
6529 int distance = node_distance(i, k);
6530
6531 if (distance > curr_distance &&
6532 (distance < next_distance ||
6533 next_distance == curr_distance))
6534 next_distance = distance;
6535
6536 /*
6537 * While not a strong assumption it would be nice to know
6538 * about cases where if node A is connected to B, B is not
6539 * equally connected to A.
6540 */
6541 if (sched_debug() && node_distance(k, i) != distance)
6542 sched_numa_warn("Node-distance not symmetric");
6543
6544 if (sched_debug() && i && !find_numa_distance(distance))
6545 sched_numa_warn("Node-0 not representative");
6546 }
6547 if (next_distance != curr_distance) {
6548 sched_domains_numa_distance[level++] = next_distance;
6549 sched_domains_numa_levels = level;
6550 curr_distance = next_distance;
6551 } else break;
6552 }
6553
6554 /*
6555 * In case of sched_debug() we verify the above assumption.
6556 */
6557 if (!sched_debug())
6558 break;
6559 }
6560
6561 if (!level)
6562 return;
6563
6564 /*
6565 * 'level' contains the number of unique distances, excluding the
6566 * identity distance node_distance(i,i).
6567 *
6568 * The sched_domains_numa_distance[] array includes the actual distance
6569 * numbers.
6570 */
6571
6572 /*
6573 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6574 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6575 * the array will contain less then 'level' members. This could be
6576 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6577 * in other functions.
6578 *
6579 * We reset it to 'level' at the end of this function.
6580 */
6581 sched_domains_numa_levels = 0;
6582
6583 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6584 if (!sched_domains_numa_masks)
6585 return;
6586
6587 /*
6588 * Now for each level, construct a mask per node which contains all
6589 * cpus of nodes that are that many hops away from us.
6590 */
6591 for (i = 0; i < level; i++) {
6592 sched_domains_numa_masks[i] =
6593 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6594 if (!sched_domains_numa_masks[i])
6595 return;
6596
6597 for (j = 0; j < nr_node_ids; j++) {
6598 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6599 if (!mask)
6600 return;
6601
6602 sched_domains_numa_masks[i][j] = mask;
6603
6604 for_each_node(k) {
6605 if (node_distance(j, k) > sched_domains_numa_distance[i])
6606 continue;
6607
6608 cpumask_or(mask, mask, cpumask_of_node(k));
6609 }
6610 }
6611 }
6612
6613 /* Compute default topology size */
6614 for (i = 0; sched_domain_topology[i].mask; i++);
6615
6616 tl = kzalloc((i + level + 1) *
6617 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6618 if (!tl)
6619 return;
6620
6621 /*
6622 * Copy the default topology bits..
6623 */
6624 for (i = 0; sched_domain_topology[i].mask; i++)
6625 tl[i] = sched_domain_topology[i];
6626
6627 /*
6628 * .. and append 'j' levels of NUMA goodness.
6629 */
6630 for (j = 0; j < level; i++, j++) {
6631 tl[i] = (struct sched_domain_topology_level){
6632 .mask = sd_numa_mask,
6633 .sd_flags = cpu_numa_flags,
6634 .flags = SDTL_OVERLAP,
6635 .numa_level = j,
6636 SD_INIT_NAME(NUMA)
6637 };
6638 }
6639
6640 sched_domain_topology = tl;
6641
6642 sched_domains_numa_levels = level;
6643 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6644
6645 init_numa_topology_type();
6646}
6647
6648static void sched_domains_numa_masks_set(int cpu)
6649{
6650 int i, j;
6651 int node = cpu_to_node(cpu);
6652
6653 for (i = 0; i < sched_domains_numa_levels; i++) {
6654 for (j = 0; j < nr_node_ids; j++) {
6655 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6656 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6657 }
6658 }
6659}
6660
6661static void sched_domains_numa_masks_clear(int cpu)
6662{
6663 int i, j;
6664 for (i = 0; i < sched_domains_numa_levels; i++) {
6665 for (j = 0; j < nr_node_ids; j++)
6666 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6667 }
6668}
6669
6670/*
6671 * Update sched_domains_numa_masks[level][node] array when new cpus
6672 * are onlined.
6673 */
6674static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6675 unsigned long action,
6676 void *hcpu)
6677{
6678 int cpu = (long)hcpu;
6679
6680 switch (action & ~CPU_TASKS_FROZEN) {
6681 case CPU_ONLINE:
6682 sched_domains_numa_masks_set(cpu);
6683 break;
6684
6685 case CPU_DEAD:
6686 sched_domains_numa_masks_clear(cpu);
6687 break;
6688
6689 default:
6690 return NOTIFY_DONE;
6691 }
6692
6693 return NOTIFY_OK;
6694}
6695#else
6696static inline void sched_init_numa(void)
6697{
6698}
6699
6700static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6701 unsigned long action,
6702 void *hcpu)
6703{
6704 return 0;
6705}
6706#endif /* CONFIG_NUMA */
6707
6708static int __sdt_alloc(const struct cpumask *cpu_map)
6709{
6710 struct sched_domain_topology_level *tl;
6711 int j;
6712
6713 for_each_sd_topology(tl) {
6714 struct sd_data *sdd = &tl->data;
6715
6716 sdd->sd = alloc_percpu(struct sched_domain *);
6717 if (!sdd->sd)
6718 return -ENOMEM;
6719
6720 sdd->sg = alloc_percpu(struct sched_group *);
6721 if (!sdd->sg)
6722 return -ENOMEM;
6723
6724 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6725 if (!sdd->sgc)
6726 return -ENOMEM;
6727
6728 for_each_cpu(j, cpu_map) {
6729 struct sched_domain *sd;
6730 struct sched_group *sg;
6731 struct sched_group_capacity *sgc;
6732
6733 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6734 GFP_KERNEL, cpu_to_node(j));
6735 if (!sd)
6736 return -ENOMEM;
6737
6738 *per_cpu_ptr(sdd->sd, j) = sd;
6739
6740 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6741 GFP_KERNEL, cpu_to_node(j));
6742 if (!sg)
6743 return -ENOMEM;
6744
6745 sg->next = sg;
6746
6747 *per_cpu_ptr(sdd->sg, j) = sg;
6748
6749 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6750 GFP_KERNEL, cpu_to_node(j));
6751 if (!sgc)
6752 return -ENOMEM;
6753
6754 *per_cpu_ptr(sdd->sgc, j) = sgc;
6755 }
6756 }
6757
6758 return 0;
6759}
6760
6761static void __sdt_free(const struct cpumask *cpu_map)
6762{
6763 struct sched_domain_topology_level *tl;
6764 int j;
6765
6766 for_each_sd_topology(tl) {
6767 struct sd_data *sdd = &tl->data;
6768
6769 for_each_cpu(j, cpu_map) {
6770 struct sched_domain *sd;
6771
6772 if (sdd->sd) {
6773 sd = *per_cpu_ptr(sdd->sd, j);
6774 if (sd && (sd->flags & SD_OVERLAP))
6775 free_sched_groups(sd->groups, 0);
6776 kfree(*per_cpu_ptr(sdd->sd, j));
6777 }
6778
6779 if (sdd->sg)
6780 kfree(*per_cpu_ptr(sdd->sg, j));
6781 if (sdd->sgc)
6782 kfree(*per_cpu_ptr(sdd->sgc, j));
6783 }
6784 free_percpu(sdd->sd);
6785 sdd->sd = NULL;
6786 free_percpu(sdd->sg);
6787 sdd->sg = NULL;
6788 free_percpu(sdd->sgc);
6789 sdd->sgc = NULL;
6790 }
6791}
6792
6793struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6794 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6795 struct sched_domain *child, int cpu)
6796{
6797 struct sched_domain *sd = sd_init(tl, cpu);
6798 if (!sd)
6799 return child;
6800
6801 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6802 if (child) {
6803 sd->level = child->level + 1;
6804 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6805 child->parent = sd;
6806 sd->child = child;
6807
6808 if (!cpumask_subset(sched_domain_span(child),
6809 sched_domain_span(sd))) {
6810 pr_err("BUG: arch topology borken\n");
6811#ifdef CONFIG_SCHED_DEBUG
6812 pr_err(" the %s domain not a subset of the %s domain\n",
6813 child->name, sd->name);
6814#endif
6815 /* Fixup, ensure @sd has at least @child cpus. */
6816 cpumask_or(sched_domain_span(sd),
6817 sched_domain_span(sd),
6818 sched_domain_span(child));
6819 }
6820
6821 }
6822 set_domain_attribute(sd, attr);
6823
6824 return sd;
6825}
6826
6827/*
6828 * Build sched domains for a given set of cpus and attach the sched domains
6829 * to the individual cpus
6830 */
6831static int build_sched_domains(const struct cpumask *cpu_map,
6832 struct sched_domain_attr *attr)
6833{
6834 enum s_alloc alloc_state;
6835 struct sched_domain *sd;
6836 struct s_data d;
6837 int i, ret = -ENOMEM;
6838
6839 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6840 if (alloc_state != sa_rootdomain)
6841 goto error;
6842
6843 /* Set up domains for cpus specified by the cpu_map. */
6844 for_each_cpu(i, cpu_map) {
6845 struct sched_domain_topology_level *tl;
6846
6847 sd = NULL;
6848 for_each_sd_topology(tl) {
6849 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6850 if (tl == sched_domain_topology)
6851 *per_cpu_ptr(d.sd, i) = sd;
6852 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6853 sd->flags |= SD_OVERLAP;
6854 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6855 break;
6856 }
6857 }
6858
6859 /* Build the groups for the domains */
6860 for_each_cpu(i, cpu_map) {
6861 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6862 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6863 if (sd->flags & SD_OVERLAP) {
6864 if (build_overlap_sched_groups(sd, i))
6865 goto error;
6866 } else {
6867 if (build_sched_groups(sd, i))
6868 goto error;
6869 }
6870 }
6871 }
6872
6873 /* Calculate CPU capacity for physical packages and nodes */
6874 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6875 if (!cpumask_test_cpu(i, cpu_map))
6876 continue;
6877
6878 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6879 claim_allocations(i, sd);
6880 init_sched_groups_capacity(i, sd);
6881 }
6882 }
6883
6884 /* Attach the domains */
6885 rcu_read_lock();
6886 for_each_cpu(i, cpu_map) {
6887 sd = *per_cpu_ptr(d.sd, i);
6888 cpu_attach_domain(sd, d.rd, i);
6889 }
6890 rcu_read_unlock();
6891
6892 ret = 0;
6893error:
6894 __free_domain_allocs(&d, alloc_state, cpu_map);
6895 return ret;
6896}
6897
6898static cpumask_var_t *doms_cur; /* current sched domains */
6899static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6900static struct sched_domain_attr *dattr_cur;
6901 /* attribues of custom domains in 'doms_cur' */
6902
6903/*
6904 * Special case: If a kmalloc of a doms_cur partition (array of
6905 * cpumask) fails, then fallback to a single sched domain,
6906 * as determined by the single cpumask fallback_doms.
6907 */
6908static cpumask_var_t fallback_doms;
6909
6910/*
6911 * arch_update_cpu_topology lets virtualized architectures update the
6912 * cpu core maps. It is supposed to return 1 if the topology changed
6913 * or 0 if it stayed the same.
6914 */
6915int __weak arch_update_cpu_topology(void)
6916{
6917 return 0;
6918}
6919
6920cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6921{
6922 int i;
6923 cpumask_var_t *doms;
6924
6925 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6926 if (!doms)
6927 return NULL;
6928 for (i = 0; i < ndoms; i++) {
6929 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6930 free_sched_domains(doms, i);
6931 return NULL;
6932 }
6933 }
6934 return doms;
6935}
6936
6937void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6938{
6939 unsigned int i;
6940 for (i = 0; i < ndoms; i++)
6941 free_cpumask_var(doms[i]);
6942 kfree(doms);
6943}
6944
6945/*
6946 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6947 * For now this just excludes isolated cpus, but could be used to
6948 * exclude other special cases in the future.
6949 */
6950static int init_sched_domains(const struct cpumask *cpu_map)
6951{
6952 int err;
6953
6954 arch_update_cpu_topology();
6955 ndoms_cur = 1;
6956 doms_cur = alloc_sched_domains(ndoms_cur);
6957 if (!doms_cur)
6958 doms_cur = &fallback_doms;
6959 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6960 err = build_sched_domains(doms_cur[0], NULL);
6961 register_sched_domain_sysctl();
6962
6963 return err;
6964}
6965
6966/*
6967 * Detach sched domains from a group of cpus specified in cpu_map
6968 * These cpus will now be attached to the NULL domain
6969 */
6970static void detach_destroy_domains(const struct cpumask *cpu_map)
6971{
6972 int i;
6973
6974 rcu_read_lock();
6975 for_each_cpu(i, cpu_map)
6976 cpu_attach_domain(NULL, &def_root_domain, i);
6977 rcu_read_unlock();
6978}
6979
6980/* handle null as "default" */
6981static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6982 struct sched_domain_attr *new, int idx_new)
6983{
6984 struct sched_domain_attr tmp;
6985
6986 /* fast path */
6987 if (!new && !cur)
6988 return 1;
6989
6990 tmp = SD_ATTR_INIT;
6991 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6992 new ? (new + idx_new) : &tmp,
6993 sizeof(struct sched_domain_attr));
6994}
6995
6996/*
6997 * Partition sched domains as specified by the 'ndoms_new'
6998 * cpumasks in the array doms_new[] of cpumasks. This compares
6999 * doms_new[] to the current sched domain partitioning, doms_cur[].
7000 * It destroys each deleted domain and builds each new domain.
7001 *
7002 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7003 * The masks don't intersect (don't overlap.) We should setup one
7004 * sched domain for each mask. CPUs not in any of the cpumasks will
7005 * not be load balanced. If the same cpumask appears both in the
7006 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7007 * it as it is.
7008 *
7009 * The passed in 'doms_new' should be allocated using
7010 * alloc_sched_domains. This routine takes ownership of it and will
7011 * free_sched_domains it when done with it. If the caller failed the
7012 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7013 * and partition_sched_domains() will fallback to the single partition
7014 * 'fallback_doms', it also forces the domains to be rebuilt.
7015 *
7016 * If doms_new == NULL it will be replaced with cpu_online_mask.
7017 * ndoms_new == 0 is a special case for destroying existing domains,
7018 * and it will not create the default domain.
7019 *
7020 * Call with hotplug lock held
7021 */
7022void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7023 struct sched_domain_attr *dattr_new)
7024{
7025 int i, j, n;
7026 int new_topology;
7027
7028 mutex_lock(&sched_domains_mutex);
7029
7030 /* always unregister in case we don't destroy any domains */
7031 unregister_sched_domain_sysctl();
7032
7033 /* Let architecture update cpu core mappings. */
7034 new_topology = arch_update_cpu_topology();
7035
7036 n = doms_new ? ndoms_new : 0;
7037
7038 /* Destroy deleted domains */
7039 for (i = 0; i < ndoms_cur; i++) {
7040 for (j = 0; j < n && !new_topology; j++) {
7041 if (cpumask_equal(doms_cur[i], doms_new[j])
7042 && dattrs_equal(dattr_cur, i, dattr_new, j))
7043 goto match1;
7044 }
7045 /* no match - a current sched domain not in new doms_new[] */
7046 detach_destroy_domains(doms_cur[i]);
7047match1:
7048 ;
7049 }
7050
7051 n = ndoms_cur;
7052 if (doms_new == NULL) {
7053 n = 0;
7054 doms_new = &fallback_doms;
7055 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7056 WARN_ON_ONCE(dattr_new);
7057 }
7058
7059 /* Build new domains */
7060 for (i = 0; i < ndoms_new; i++) {
7061 for (j = 0; j < n && !new_topology; j++) {
7062 if (cpumask_equal(doms_new[i], doms_cur[j])
7063 && dattrs_equal(dattr_new, i, dattr_cur, j))
7064 goto match2;
7065 }
7066 /* no match - add a new doms_new */
7067 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7068match2:
7069 ;
7070 }
7071
7072 /* Remember the new sched domains */
7073 if (doms_cur != &fallback_doms)
7074 free_sched_domains(doms_cur, ndoms_cur);
7075 kfree(dattr_cur); /* kfree(NULL) is safe */
7076 doms_cur = doms_new;
7077 dattr_cur = dattr_new;
7078 ndoms_cur = ndoms_new;
7079
7080 register_sched_domain_sysctl();
7081
7082 mutex_unlock(&sched_domains_mutex);
7083}
7084
7085static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7086
7087/*
7088 * Update cpusets according to cpu_active mask. If cpusets are
7089 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7090 * around partition_sched_domains().
7091 *
7092 * If we come here as part of a suspend/resume, don't touch cpusets because we
7093 * want to restore it back to its original state upon resume anyway.
7094 */
7095static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7096 void *hcpu)
7097{
7098 switch (action) {
7099 case CPU_ONLINE_FROZEN:
7100 case CPU_DOWN_FAILED_FROZEN:
7101
7102 /*
7103 * num_cpus_frozen tracks how many CPUs are involved in suspend
7104 * resume sequence. As long as this is not the last online
7105 * operation in the resume sequence, just build a single sched
7106 * domain, ignoring cpusets.
7107 */
7108 num_cpus_frozen--;
7109 if (likely(num_cpus_frozen)) {
7110 partition_sched_domains(1, NULL, NULL);
7111 break;
7112 }
7113
7114 /*
7115 * This is the last CPU online operation. So fall through and
7116 * restore the original sched domains by considering the
7117 * cpuset configurations.
7118 */
7119
7120 case CPU_ONLINE:
7121 cpuset_update_active_cpus(true);
7122 break;
7123 default:
7124 return NOTIFY_DONE;
7125 }
7126 return NOTIFY_OK;
7127}
7128
7129static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7130 void *hcpu)
7131{
7132 unsigned long flags;
7133 long cpu = (long)hcpu;
7134 struct dl_bw *dl_b;
7135 bool overflow;
7136 int cpus;
7137
7138 switch (action) {
7139 case CPU_DOWN_PREPARE:
7140 rcu_read_lock_sched();
7141 dl_b = dl_bw_of(cpu);
7142
7143 raw_spin_lock_irqsave(&dl_b->lock, flags);
7144 cpus = dl_bw_cpus(cpu);
7145 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7146 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7147
7148 rcu_read_unlock_sched();
7149
7150 if (overflow)
7151 return notifier_from_errno(-EBUSY);
7152 cpuset_update_active_cpus(false);
7153 break;
7154 case CPU_DOWN_PREPARE_FROZEN:
7155 num_cpus_frozen++;
7156 partition_sched_domains(1, NULL, NULL);
7157 break;
7158 default:
7159 return NOTIFY_DONE;
7160 }
7161 return NOTIFY_OK;
7162}
7163
7164void __init sched_init_smp(void)
7165{
7166 cpumask_var_t non_isolated_cpus;
7167
7168 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7169 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7170
7171 sched_init_numa();
7172
7173 /*
7174 * There's no userspace yet to cause hotplug operations; hence all the
7175 * cpu masks are stable and all blatant races in the below code cannot
7176 * happen.
7177 */
7178 mutex_lock(&sched_domains_mutex);
7179 init_sched_domains(cpu_active_mask);
7180 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7181 if (cpumask_empty(non_isolated_cpus))
7182 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7183 mutex_unlock(&sched_domains_mutex);
7184
7185 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7186 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7187 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7188
7189 init_hrtick();
7190
7191 /* Move init over to a non-isolated CPU */
7192 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7193 BUG();
7194 sched_init_granularity();
7195 free_cpumask_var(non_isolated_cpus);
7196
7197 init_sched_rt_class();
7198 init_sched_dl_class();
7199}
7200#else
7201void __init sched_init_smp(void)
7202{
7203 sched_init_granularity();
7204}
7205#endif /* CONFIG_SMP */
7206
7207int in_sched_functions(unsigned long addr)
7208{
7209 return in_lock_functions(addr) ||
7210 (addr >= (unsigned long)__sched_text_start
7211 && addr < (unsigned long)__sched_text_end);
7212}
7213
7214#ifdef CONFIG_CGROUP_SCHED
7215/*
7216 * Default task group.
7217 * Every task in system belongs to this group at bootup.
7218 */
7219struct task_group root_task_group;
7220LIST_HEAD(task_groups);
7221
7222/* Cacheline aligned slab cache for task_group */
7223static struct kmem_cache *task_group_cache __read_mostly;
7224#endif
7225
7226DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7227
7228void __init sched_init(void)
7229{
7230 int i, j;
7231 unsigned long alloc_size = 0, ptr;
7232
7233#ifdef CONFIG_FAIR_GROUP_SCHED
7234 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7235#endif
7236#ifdef CONFIG_RT_GROUP_SCHED
7237 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7238#endif
7239 if (alloc_size) {
7240 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7241
7242#ifdef CONFIG_FAIR_GROUP_SCHED
7243 root_task_group.se = (struct sched_entity **)ptr;
7244 ptr += nr_cpu_ids * sizeof(void **);
7245
7246 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7247 ptr += nr_cpu_ids * sizeof(void **);
7248
7249#endif /* CONFIG_FAIR_GROUP_SCHED */
7250#ifdef CONFIG_RT_GROUP_SCHED
7251 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7252 ptr += nr_cpu_ids * sizeof(void **);
7253
7254 root_task_group.rt_rq = (struct rt_rq **)ptr;
7255 ptr += nr_cpu_ids * sizeof(void **);
7256
7257#endif /* CONFIG_RT_GROUP_SCHED */
7258 }
7259#ifdef CONFIG_CPUMASK_OFFSTACK
7260 for_each_possible_cpu(i) {
7261 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7262 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7263 }
7264#endif /* CONFIG_CPUMASK_OFFSTACK */
7265
7266 init_rt_bandwidth(&def_rt_bandwidth,
7267 global_rt_period(), global_rt_runtime());
7268 init_dl_bandwidth(&def_dl_bandwidth,
7269 global_rt_period(), global_rt_runtime());
7270
7271#ifdef CONFIG_SMP
7272 init_defrootdomain();
7273#endif
7274
7275#ifdef CONFIG_RT_GROUP_SCHED
7276 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7277 global_rt_period(), global_rt_runtime());
7278#endif /* CONFIG_RT_GROUP_SCHED */
7279
7280#ifdef CONFIG_CGROUP_SCHED
7281 task_group_cache = KMEM_CACHE(task_group, 0);
7282
7283 list_add(&root_task_group.list, &task_groups);
7284 INIT_LIST_HEAD(&root_task_group.children);
7285 INIT_LIST_HEAD(&root_task_group.siblings);
7286 autogroup_init(&init_task);
7287#endif /* CONFIG_CGROUP_SCHED */
7288
7289 for_each_possible_cpu(i) {
7290 struct rq *rq;
7291
7292 rq = cpu_rq(i);
7293 raw_spin_lock_init(&rq->lock);
7294 rq->nr_running = 0;
7295 rq->calc_load_active = 0;
7296 rq->calc_load_update = jiffies + LOAD_FREQ;
7297 init_cfs_rq(&rq->cfs);
7298 init_rt_rq(&rq->rt);
7299 init_dl_rq(&rq->dl);
7300#ifdef CONFIG_FAIR_GROUP_SCHED
7301 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7302 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7303 /*
7304 * How much cpu bandwidth does root_task_group get?
7305 *
7306 * In case of task-groups formed thr' the cgroup filesystem, it
7307 * gets 100% of the cpu resources in the system. This overall
7308 * system cpu resource is divided among the tasks of
7309 * root_task_group and its child task-groups in a fair manner,
7310 * based on each entity's (task or task-group's) weight
7311 * (se->load.weight).
7312 *
7313 * In other words, if root_task_group has 10 tasks of weight
7314 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7315 * then A0's share of the cpu resource is:
7316 *
7317 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7318 *
7319 * We achieve this by letting root_task_group's tasks sit
7320 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7321 */
7322 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7323 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7324#endif /* CONFIG_FAIR_GROUP_SCHED */
7325
7326 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7327#ifdef CONFIG_RT_GROUP_SCHED
7328 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7329#endif
7330
7331 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7332 rq->cpu_load[j] = 0;
7333
7334 rq->last_load_update_tick = jiffies;
7335
7336#ifdef CONFIG_SMP
7337 rq->sd = NULL;
7338 rq->rd = NULL;
7339 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7340 rq->balance_callback = NULL;
7341 rq->active_balance = 0;
7342 rq->next_balance = jiffies;
7343 rq->push_cpu = 0;
7344 rq->cpu = i;
7345 rq->online = 0;
7346 rq->idle_stamp = 0;
7347 rq->avg_idle = 2*sysctl_sched_migration_cost;
7348 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7349
7350 INIT_LIST_HEAD(&rq->cfs_tasks);
7351
7352 rq_attach_root(rq, &def_root_domain);
7353#ifdef CONFIG_NO_HZ_COMMON
7354 rq->nohz_flags = 0;
7355#endif
7356#ifdef CONFIG_NO_HZ_FULL
7357 rq->last_sched_tick = 0;
7358#endif
7359#endif
7360 init_rq_hrtick(rq);
7361 atomic_set(&rq->nr_iowait, 0);
7362 }
7363
7364 set_load_weight(&init_task);
7365
7366#ifdef CONFIG_PREEMPT_NOTIFIERS
7367 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7368#endif
7369
7370 /*
7371 * The boot idle thread does lazy MMU switching as well:
7372 */
7373 atomic_inc(&init_mm.mm_count);
7374 enter_lazy_tlb(&init_mm, current);
7375
7376 /*
7377 * During early bootup we pretend to be a normal task:
7378 */
7379 current->sched_class = &fair_sched_class;
7380
7381 /*
7382 * Make us the idle thread. Technically, schedule() should not be
7383 * called from this thread, however somewhere below it might be,
7384 * but because we are the idle thread, we just pick up running again
7385 * when this runqueue becomes "idle".
7386 */
7387 init_idle(current, smp_processor_id());
7388
7389 calc_load_update = jiffies + LOAD_FREQ;
7390
7391#ifdef CONFIG_SMP
7392 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7393 /* May be allocated at isolcpus cmdline parse time */
7394 if (cpu_isolated_map == NULL)
7395 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7396 idle_thread_set_boot_cpu();
7397 set_cpu_rq_start_time();
7398#endif
7399 init_sched_fair_class();
7400
7401 scheduler_running = 1;
7402}
7403
7404#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7405static inline int preempt_count_equals(int preempt_offset)
7406{
7407 int nested = preempt_count() + rcu_preempt_depth();
7408
7409 return (nested == preempt_offset);
7410}
7411
7412void __might_sleep(const char *file, int line, int preempt_offset)
7413{
7414 /*
7415 * Blocking primitives will set (and therefore destroy) current->state,
7416 * since we will exit with TASK_RUNNING make sure we enter with it,
7417 * otherwise we will destroy state.
7418 */
7419 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7420 "do not call blocking ops when !TASK_RUNNING; "
7421 "state=%lx set at [<%p>] %pS\n",
7422 current->state,
7423 (void *)current->task_state_change,
7424 (void *)current->task_state_change);
7425
7426 ___might_sleep(file, line, preempt_offset);
7427}
7428EXPORT_SYMBOL(__might_sleep);
7429
7430void ___might_sleep(const char *file, int line, int preempt_offset)
7431{
7432 static unsigned long prev_jiffy; /* ratelimiting */
7433
7434 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7435 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7436 !is_idle_task(current)) ||
7437 system_state != SYSTEM_RUNNING || oops_in_progress)
7438 return;
7439 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7440 return;
7441 prev_jiffy = jiffies;
7442
7443 printk(KERN_ERR
7444 "BUG: sleeping function called from invalid context at %s:%d\n",
7445 file, line);
7446 printk(KERN_ERR
7447 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7448 in_atomic(), irqs_disabled(),
7449 current->pid, current->comm);
7450
7451 if (task_stack_end_corrupted(current))
7452 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7453
7454 debug_show_held_locks(current);
7455 if (irqs_disabled())
7456 print_irqtrace_events(current);
7457#ifdef CONFIG_DEBUG_PREEMPT
7458 if (!preempt_count_equals(preempt_offset)) {
7459 pr_err("Preemption disabled at:");
7460 print_ip_sym(current->preempt_disable_ip);
7461 pr_cont("\n");
7462 }
7463#endif
7464 dump_stack();
7465}
7466EXPORT_SYMBOL(___might_sleep);
7467#endif
7468
7469#ifdef CONFIG_MAGIC_SYSRQ
7470void normalize_rt_tasks(void)
7471{
7472 struct task_struct *g, *p;
7473 struct sched_attr attr = {
7474 .sched_policy = SCHED_NORMAL,
7475 };
7476
7477 read_lock(&tasklist_lock);
7478 for_each_process_thread(g, p) {
7479 /*
7480 * Only normalize user tasks:
7481 */
7482 if (p->flags & PF_KTHREAD)
7483 continue;
7484
7485 p->se.exec_start = 0;
7486#ifdef CONFIG_SCHEDSTATS
7487 p->se.statistics.wait_start = 0;
7488 p->se.statistics.sleep_start = 0;
7489 p->se.statistics.block_start = 0;
7490#endif
7491
7492 if (!dl_task(p) && !rt_task(p)) {
7493 /*
7494 * Renice negative nice level userspace
7495 * tasks back to 0:
7496 */
7497 if (task_nice(p) < 0)
7498 set_user_nice(p, 0);
7499 continue;
7500 }
7501
7502 __sched_setscheduler(p, &attr, false, false);
7503 }
7504 read_unlock(&tasklist_lock);
7505}
7506
7507#endif /* CONFIG_MAGIC_SYSRQ */
7508
7509#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7510/*
7511 * These functions are only useful for the IA64 MCA handling, or kdb.
7512 *
7513 * They can only be called when the whole system has been
7514 * stopped - every CPU needs to be quiescent, and no scheduling
7515 * activity can take place. Using them for anything else would
7516 * be a serious bug, and as a result, they aren't even visible
7517 * under any other configuration.
7518 */
7519
7520/**
7521 * curr_task - return the current task for a given cpu.
7522 * @cpu: the processor in question.
7523 *
7524 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7525 *
7526 * Return: The current task for @cpu.
7527 */
7528struct task_struct *curr_task(int cpu)
7529{
7530 return cpu_curr(cpu);
7531}
7532
7533#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7534
7535#ifdef CONFIG_IA64
7536/**
7537 * set_curr_task - set the current task for a given cpu.
7538 * @cpu: the processor in question.
7539 * @p: the task pointer to set.
7540 *
7541 * Description: This function must only be used when non-maskable interrupts
7542 * are serviced on a separate stack. It allows the architecture to switch the
7543 * notion of the current task on a cpu in a non-blocking manner. This function
7544 * must be called with all CPU's synchronized, and interrupts disabled, the
7545 * and caller must save the original value of the current task (see
7546 * curr_task() above) and restore that value before reenabling interrupts and
7547 * re-starting the system.
7548 *
7549 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7550 */
7551void set_curr_task(int cpu, struct task_struct *p)
7552{
7553 cpu_curr(cpu) = p;
7554}
7555
7556#endif
7557
7558#ifdef CONFIG_CGROUP_SCHED
7559/* task_group_lock serializes the addition/removal of task groups */
7560static DEFINE_SPINLOCK(task_group_lock);
7561
7562static void sched_free_group(struct task_group *tg)
7563{
7564 free_fair_sched_group(tg);
7565 free_rt_sched_group(tg);
7566 autogroup_free(tg);
7567 kmem_cache_free(task_group_cache, tg);
7568}
7569
7570/* allocate runqueue etc for a new task group */
7571struct task_group *sched_create_group(struct task_group *parent)
7572{
7573 struct task_group *tg;
7574
7575 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7576 if (!tg)
7577 return ERR_PTR(-ENOMEM);
7578
7579 if (!alloc_fair_sched_group(tg, parent))
7580 goto err;
7581
7582 if (!alloc_rt_sched_group(tg, parent))
7583 goto err;
7584
7585 return tg;
7586
7587err:
7588 sched_free_group(tg);
7589 return ERR_PTR(-ENOMEM);
7590}
7591
7592void sched_online_group(struct task_group *tg, struct task_group *parent)
7593{
7594 unsigned long flags;
7595
7596 spin_lock_irqsave(&task_group_lock, flags);
7597 list_add_rcu(&tg->list, &task_groups);
7598
7599 WARN_ON(!parent); /* root should already exist */
7600
7601 tg->parent = parent;
7602 INIT_LIST_HEAD(&tg->children);
7603 list_add_rcu(&tg->siblings, &parent->children);
7604 spin_unlock_irqrestore(&task_group_lock, flags);
7605}
7606
7607/* rcu callback to free various structures associated with a task group */
7608static void sched_free_group_rcu(struct rcu_head *rhp)
7609{
7610 /* now it should be safe to free those cfs_rqs */
7611 sched_free_group(container_of(rhp, struct task_group, rcu));
7612}
7613
7614void sched_destroy_group(struct task_group *tg)
7615{
7616 /* wait for possible concurrent references to cfs_rqs complete */
7617 call_rcu(&tg->rcu, sched_free_group_rcu);
7618}
7619
7620void sched_offline_group(struct task_group *tg)
7621{
7622 unsigned long flags;
7623
7624 /* end participation in shares distribution */
7625 unregister_fair_sched_group(tg);
7626
7627 spin_lock_irqsave(&task_group_lock, flags);
7628 list_del_rcu(&tg->list);
7629 list_del_rcu(&tg->siblings);
7630 spin_unlock_irqrestore(&task_group_lock, flags);
7631}
7632
7633/* change task's runqueue when it moves between groups.
7634 * The caller of this function should have put the task in its new group
7635 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7636 * reflect its new group.
7637 */
7638void sched_move_task(struct task_struct *tsk)
7639{
7640 struct task_group *tg;
7641 int queued, running;
7642 unsigned long flags;
7643 struct rq *rq;
7644
7645 rq = task_rq_lock(tsk, &flags);
7646
7647 running = task_current(rq, tsk);
7648 queued = task_on_rq_queued(tsk);
7649
7650 if (queued)
7651 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7652 if (unlikely(running))
7653 put_prev_task(rq, tsk);
7654
7655 /*
7656 * All callers are synchronized by task_rq_lock(); we do not use RCU
7657 * which is pointless here. Thus, we pass "true" to task_css_check()
7658 * to prevent lockdep warnings.
7659 */
7660 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7661 struct task_group, css);
7662 tg = autogroup_task_group(tsk, tg);
7663 tsk->sched_task_group = tg;
7664
7665#ifdef CONFIG_FAIR_GROUP_SCHED
7666 if (tsk->sched_class->task_move_group)
7667 tsk->sched_class->task_move_group(tsk);
7668 else
7669#endif
7670 set_task_rq(tsk, task_cpu(tsk));
7671
7672 if (unlikely(running))
7673 tsk->sched_class->set_curr_task(rq);
7674 if (queued)
7675 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7676
7677 task_rq_unlock(rq, tsk, &flags);
7678}
7679#endif /* CONFIG_CGROUP_SCHED */
7680
7681#ifdef CONFIG_RT_GROUP_SCHED
7682/*
7683 * Ensure that the real time constraints are schedulable.
7684 */
7685static DEFINE_MUTEX(rt_constraints_mutex);
7686
7687/* Must be called with tasklist_lock held */
7688static inline int tg_has_rt_tasks(struct task_group *tg)
7689{
7690 struct task_struct *g, *p;
7691
7692 /*
7693 * Autogroups do not have RT tasks; see autogroup_create().
7694 */
7695 if (task_group_is_autogroup(tg))
7696 return 0;
7697
7698 for_each_process_thread(g, p) {
7699 if (rt_task(p) && task_group(p) == tg)
7700 return 1;
7701 }
7702
7703 return 0;
7704}
7705
7706struct rt_schedulable_data {
7707 struct task_group *tg;
7708 u64 rt_period;
7709 u64 rt_runtime;
7710};
7711
7712static int tg_rt_schedulable(struct task_group *tg, void *data)
7713{
7714 struct rt_schedulable_data *d = data;
7715 struct task_group *child;
7716 unsigned long total, sum = 0;
7717 u64 period, runtime;
7718
7719 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7720 runtime = tg->rt_bandwidth.rt_runtime;
7721
7722 if (tg == d->tg) {
7723 period = d->rt_period;
7724 runtime = d->rt_runtime;
7725 }
7726
7727 /*
7728 * Cannot have more runtime than the period.
7729 */
7730 if (runtime > period && runtime != RUNTIME_INF)
7731 return -EINVAL;
7732
7733 /*
7734 * Ensure we don't starve existing RT tasks.
7735 */
7736 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7737 return -EBUSY;
7738
7739 total = to_ratio(period, runtime);
7740
7741 /*
7742 * Nobody can have more than the global setting allows.
7743 */
7744 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7745 return -EINVAL;
7746
7747 /*
7748 * The sum of our children's runtime should not exceed our own.
7749 */
7750 list_for_each_entry_rcu(child, &tg->children, siblings) {
7751 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7752 runtime = child->rt_bandwidth.rt_runtime;
7753
7754 if (child == d->tg) {
7755 period = d->rt_period;
7756 runtime = d->rt_runtime;
7757 }
7758
7759 sum += to_ratio(period, runtime);
7760 }
7761
7762 if (sum > total)
7763 return -EINVAL;
7764
7765 return 0;
7766}
7767
7768static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7769{
7770 int ret;
7771
7772 struct rt_schedulable_data data = {
7773 .tg = tg,
7774 .rt_period = period,
7775 .rt_runtime = runtime,
7776 };
7777
7778 rcu_read_lock();
7779 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7780 rcu_read_unlock();
7781
7782 return ret;
7783}
7784
7785static int tg_set_rt_bandwidth(struct task_group *tg,
7786 u64 rt_period, u64 rt_runtime)
7787{
7788 int i, err = 0;
7789
7790 /*
7791 * Disallowing the root group RT runtime is BAD, it would disallow the
7792 * kernel creating (and or operating) RT threads.
7793 */
7794 if (tg == &root_task_group && rt_runtime == 0)
7795 return -EINVAL;
7796
7797 /* No period doesn't make any sense. */
7798 if (rt_period == 0)
7799 return -EINVAL;
7800
7801 mutex_lock(&rt_constraints_mutex);
7802 read_lock(&tasklist_lock);
7803 err = __rt_schedulable(tg, rt_period, rt_runtime);
7804 if (err)
7805 goto unlock;
7806
7807 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7808 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7809 tg->rt_bandwidth.rt_runtime = rt_runtime;
7810
7811 for_each_possible_cpu(i) {
7812 struct rt_rq *rt_rq = tg->rt_rq[i];
7813
7814 raw_spin_lock(&rt_rq->rt_runtime_lock);
7815 rt_rq->rt_runtime = rt_runtime;
7816 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7817 }
7818 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7819unlock:
7820 read_unlock(&tasklist_lock);
7821 mutex_unlock(&rt_constraints_mutex);
7822
7823 return err;
7824}
7825
7826static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7827{
7828 u64 rt_runtime, rt_period;
7829
7830 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7831 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7832 if (rt_runtime_us < 0)
7833 rt_runtime = RUNTIME_INF;
7834
7835 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7836}
7837
7838static long sched_group_rt_runtime(struct task_group *tg)
7839{
7840 u64 rt_runtime_us;
7841
7842 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7843 return -1;
7844
7845 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7846 do_div(rt_runtime_us, NSEC_PER_USEC);
7847 return rt_runtime_us;
7848}
7849
7850static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7851{
7852 u64 rt_runtime, rt_period;
7853
7854 rt_period = rt_period_us * NSEC_PER_USEC;
7855 rt_runtime = tg->rt_bandwidth.rt_runtime;
7856
7857 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7858}
7859
7860static long sched_group_rt_period(struct task_group *tg)
7861{
7862 u64 rt_period_us;
7863
7864 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7865 do_div(rt_period_us, NSEC_PER_USEC);
7866 return rt_period_us;
7867}
7868#endif /* CONFIG_RT_GROUP_SCHED */
7869
7870#ifdef CONFIG_RT_GROUP_SCHED
7871static int sched_rt_global_constraints(void)
7872{
7873 int ret = 0;
7874
7875 mutex_lock(&rt_constraints_mutex);
7876 read_lock(&tasklist_lock);
7877 ret = __rt_schedulable(NULL, 0, 0);
7878 read_unlock(&tasklist_lock);
7879 mutex_unlock(&rt_constraints_mutex);
7880
7881 return ret;
7882}
7883
7884static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7885{
7886 /* Don't accept realtime tasks when there is no way for them to run */
7887 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7888 return 0;
7889
7890 return 1;
7891}
7892
7893#else /* !CONFIG_RT_GROUP_SCHED */
7894static int sched_rt_global_constraints(void)
7895{
7896 unsigned long flags;
7897 int i, ret = 0;
7898
7899 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7900 for_each_possible_cpu(i) {
7901 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7902
7903 raw_spin_lock(&rt_rq->rt_runtime_lock);
7904 rt_rq->rt_runtime = global_rt_runtime();
7905 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7906 }
7907 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7908
7909 return ret;
7910}
7911#endif /* CONFIG_RT_GROUP_SCHED */
7912
7913static int sched_dl_global_validate(void)
7914{
7915 u64 runtime = global_rt_runtime();
7916 u64 period = global_rt_period();
7917 u64 new_bw = to_ratio(period, runtime);
7918 struct dl_bw *dl_b;
7919 int cpu, ret = 0;
7920 unsigned long flags;
7921
7922 /*
7923 * Here we want to check the bandwidth not being set to some
7924 * value smaller than the currently allocated bandwidth in
7925 * any of the root_domains.
7926 *
7927 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7928 * cycling on root_domains... Discussion on different/better
7929 * solutions is welcome!
7930 */
7931 for_each_possible_cpu(cpu) {
7932 rcu_read_lock_sched();
7933 dl_b = dl_bw_of(cpu);
7934
7935 raw_spin_lock_irqsave(&dl_b->lock, flags);
7936 if (new_bw < dl_b->total_bw)
7937 ret = -EBUSY;
7938 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7939
7940 rcu_read_unlock_sched();
7941
7942 if (ret)
7943 break;
7944 }
7945
7946 return ret;
7947}
7948
7949static void sched_dl_do_global(void)
7950{
7951 u64 new_bw = -1;
7952 struct dl_bw *dl_b;
7953 int cpu;
7954 unsigned long flags;
7955
7956 def_dl_bandwidth.dl_period = global_rt_period();
7957 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7958
7959 if (global_rt_runtime() != RUNTIME_INF)
7960 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7961
7962 /*
7963 * FIXME: As above...
7964 */
7965 for_each_possible_cpu(cpu) {
7966 rcu_read_lock_sched();
7967 dl_b = dl_bw_of(cpu);
7968
7969 raw_spin_lock_irqsave(&dl_b->lock, flags);
7970 dl_b->bw = new_bw;
7971 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7972
7973 rcu_read_unlock_sched();
7974 }
7975}
7976
7977static int sched_rt_global_validate(void)
7978{
7979 if (sysctl_sched_rt_period <= 0)
7980 return -EINVAL;
7981
7982 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7983 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7984 return -EINVAL;
7985
7986 return 0;
7987}
7988
7989static void sched_rt_do_global(void)
7990{
7991 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7992 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7993}
7994
7995int sched_rt_handler(struct ctl_table *table, int write,
7996 void __user *buffer, size_t *lenp,
7997 loff_t *ppos)
7998{
7999 int old_period, old_runtime;
8000 static DEFINE_MUTEX(mutex);
8001 int ret;
8002
8003 mutex_lock(&mutex);
8004 old_period = sysctl_sched_rt_period;
8005 old_runtime = sysctl_sched_rt_runtime;
8006
8007 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8008
8009 if (!ret && write) {
8010 ret = sched_rt_global_validate();
8011 if (ret)
8012 goto undo;
8013
8014 ret = sched_dl_global_validate();
8015 if (ret)
8016 goto undo;
8017
8018 ret = sched_rt_global_constraints();
8019 if (ret)
8020 goto undo;
8021
8022 sched_rt_do_global();
8023 sched_dl_do_global();
8024 }
8025 if (0) {
8026undo:
8027 sysctl_sched_rt_period = old_period;
8028 sysctl_sched_rt_runtime = old_runtime;
8029 }
8030 mutex_unlock(&mutex);
8031
8032 return ret;
8033}
8034
8035int sched_rr_handler(struct ctl_table *table, int write,
8036 void __user *buffer, size_t *lenp,
8037 loff_t *ppos)
8038{
8039 int ret;
8040 static DEFINE_MUTEX(mutex);
8041
8042 mutex_lock(&mutex);
8043 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8044 /* make sure that internally we keep jiffies */
8045 /* also, writing zero resets timeslice to default */
8046 if (!ret && write) {
8047 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8048 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8049 }
8050 mutex_unlock(&mutex);
8051 return ret;
8052}
8053
8054#ifdef CONFIG_CGROUP_SCHED
8055
8056static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8057{
8058 return css ? container_of(css, struct task_group, css) : NULL;
8059}
8060
8061static struct cgroup_subsys_state *
8062cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8063{
8064 struct task_group *parent = css_tg(parent_css);
8065 struct task_group *tg;
8066
8067 if (!parent) {
8068 /* This is early initialization for the top cgroup */
8069 return &root_task_group.css;
8070 }
8071
8072 tg = sched_create_group(parent);
8073 if (IS_ERR(tg))
8074 return ERR_PTR(-ENOMEM);
8075
8076 sched_online_group(tg, parent);
8077
8078 return &tg->css;
8079}
8080
8081static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8082{
8083 struct task_group *tg = css_tg(css);
8084
8085 sched_offline_group(tg);
8086}
8087
8088static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8089{
8090 struct task_group *tg = css_tg(css);
8091
8092 /*
8093 * Relies on the RCU grace period between css_released() and this.
8094 */
8095 sched_free_group(tg);
8096}
8097
8098static void cpu_cgroup_fork(struct task_struct *task)
8099{
8100 sched_move_task(task);
8101}
8102
8103static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8104{
8105 struct task_struct *task;
8106 struct cgroup_subsys_state *css;
8107
8108 cgroup_taskset_for_each(task, css, tset) {
8109#ifdef CONFIG_RT_GROUP_SCHED
8110 if (!sched_rt_can_attach(css_tg(css), task))
8111 return -EINVAL;
8112#else
8113 /* We don't support RT-tasks being in separate groups */
8114 if (task->sched_class != &fair_sched_class)
8115 return -EINVAL;
8116#endif
8117 }
8118 return 0;
8119}
8120
8121static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8122{
8123 struct task_struct *task;
8124 struct cgroup_subsys_state *css;
8125
8126 cgroup_taskset_for_each(task, css, tset)
8127 sched_move_task(task);
8128}
8129
8130#ifdef CONFIG_FAIR_GROUP_SCHED
8131static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8132 struct cftype *cftype, u64 shareval)
8133{
8134 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8135}
8136
8137static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8138 struct cftype *cft)
8139{
8140 struct task_group *tg = css_tg(css);
8141
8142 return (u64) scale_load_down(tg->shares);
8143}
8144
8145#ifdef CONFIG_CFS_BANDWIDTH
8146static DEFINE_MUTEX(cfs_constraints_mutex);
8147
8148const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8149const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8150
8151static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8152
8153static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8154{
8155 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8156 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8157
8158 if (tg == &root_task_group)
8159 return -EINVAL;
8160
8161 /*
8162 * Ensure we have at some amount of bandwidth every period. This is
8163 * to prevent reaching a state of large arrears when throttled via
8164 * entity_tick() resulting in prolonged exit starvation.
8165 */
8166 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8167 return -EINVAL;
8168
8169 /*
8170 * Likewise, bound things on the otherside by preventing insane quota
8171 * periods. This also allows us to normalize in computing quota
8172 * feasibility.
8173 */
8174 if (period > max_cfs_quota_period)
8175 return -EINVAL;
8176
8177 /*
8178 * Prevent race between setting of cfs_rq->runtime_enabled and
8179 * unthrottle_offline_cfs_rqs().
8180 */
8181 get_online_cpus();
8182 mutex_lock(&cfs_constraints_mutex);
8183 ret = __cfs_schedulable(tg, period, quota);
8184 if (ret)
8185 goto out_unlock;
8186
8187 runtime_enabled = quota != RUNTIME_INF;
8188 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8189 /*
8190 * If we need to toggle cfs_bandwidth_used, off->on must occur
8191 * before making related changes, and on->off must occur afterwards
8192 */
8193 if (runtime_enabled && !runtime_was_enabled)
8194 cfs_bandwidth_usage_inc();
8195 raw_spin_lock_irq(&cfs_b->lock);
8196 cfs_b->period = ns_to_ktime(period);
8197 cfs_b->quota = quota;
8198
8199 __refill_cfs_bandwidth_runtime(cfs_b);
8200 /* restart the period timer (if active) to handle new period expiry */
8201 if (runtime_enabled)
8202 start_cfs_bandwidth(cfs_b);
8203 raw_spin_unlock_irq(&cfs_b->lock);
8204
8205 for_each_online_cpu(i) {
8206 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8207 struct rq *rq = cfs_rq->rq;
8208
8209 raw_spin_lock_irq(&rq->lock);
8210 cfs_rq->runtime_enabled = runtime_enabled;
8211 cfs_rq->runtime_remaining = 0;
8212
8213 if (cfs_rq->throttled)
8214 unthrottle_cfs_rq(cfs_rq);
8215 raw_spin_unlock_irq(&rq->lock);
8216 }
8217 if (runtime_was_enabled && !runtime_enabled)
8218 cfs_bandwidth_usage_dec();
8219out_unlock:
8220 mutex_unlock(&cfs_constraints_mutex);
8221 put_online_cpus();
8222
8223 return ret;
8224}
8225
8226int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8227{
8228 u64 quota, period;
8229
8230 period = ktime_to_ns(tg->cfs_bandwidth.period);
8231 if (cfs_quota_us < 0)
8232 quota = RUNTIME_INF;
8233 else
8234 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8235
8236 return tg_set_cfs_bandwidth(tg, period, quota);
8237}
8238
8239long tg_get_cfs_quota(struct task_group *tg)
8240{
8241 u64 quota_us;
8242
8243 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8244 return -1;
8245
8246 quota_us = tg->cfs_bandwidth.quota;
8247 do_div(quota_us, NSEC_PER_USEC);
8248
8249 return quota_us;
8250}
8251
8252int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8253{
8254 u64 quota, period;
8255
8256 period = (u64)cfs_period_us * NSEC_PER_USEC;
8257 quota = tg->cfs_bandwidth.quota;
8258
8259 return tg_set_cfs_bandwidth(tg, period, quota);
8260}
8261
8262long tg_get_cfs_period(struct task_group *tg)
8263{
8264 u64 cfs_period_us;
8265
8266 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8267 do_div(cfs_period_us, NSEC_PER_USEC);
8268
8269 return cfs_period_us;
8270}
8271
8272static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8273 struct cftype *cft)
8274{
8275 return tg_get_cfs_quota(css_tg(css));
8276}
8277
8278static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8279 struct cftype *cftype, s64 cfs_quota_us)
8280{
8281 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8282}
8283
8284static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8285 struct cftype *cft)
8286{
8287 return tg_get_cfs_period(css_tg(css));
8288}
8289
8290static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8291 struct cftype *cftype, u64 cfs_period_us)
8292{
8293 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8294}
8295
8296struct cfs_schedulable_data {
8297 struct task_group *tg;
8298 u64 period, quota;
8299};
8300
8301/*
8302 * normalize group quota/period to be quota/max_period
8303 * note: units are usecs
8304 */
8305static u64 normalize_cfs_quota(struct task_group *tg,
8306 struct cfs_schedulable_data *d)
8307{
8308 u64 quota, period;
8309
8310 if (tg == d->tg) {
8311 period = d->period;
8312 quota = d->quota;
8313 } else {
8314 period = tg_get_cfs_period(tg);
8315 quota = tg_get_cfs_quota(tg);
8316 }
8317
8318 /* note: these should typically be equivalent */
8319 if (quota == RUNTIME_INF || quota == -1)
8320 return RUNTIME_INF;
8321
8322 return to_ratio(period, quota);
8323}
8324
8325static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8326{
8327 struct cfs_schedulable_data *d = data;
8328 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8329 s64 quota = 0, parent_quota = -1;
8330
8331 if (!tg->parent) {
8332 quota = RUNTIME_INF;
8333 } else {
8334 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8335
8336 quota = normalize_cfs_quota(tg, d);
8337 parent_quota = parent_b->hierarchical_quota;
8338
8339 /*
8340 * ensure max(child_quota) <= parent_quota, inherit when no
8341 * limit is set
8342 */
8343 if (quota == RUNTIME_INF)
8344 quota = parent_quota;
8345 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8346 return -EINVAL;
8347 }
8348 cfs_b->hierarchical_quota = quota;
8349
8350 return 0;
8351}
8352
8353static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8354{
8355 int ret;
8356 struct cfs_schedulable_data data = {
8357 .tg = tg,
8358 .period = period,
8359 .quota = quota,
8360 };
8361
8362 if (quota != RUNTIME_INF) {
8363 do_div(data.period, NSEC_PER_USEC);
8364 do_div(data.quota, NSEC_PER_USEC);
8365 }
8366
8367 rcu_read_lock();
8368 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8369 rcu_read_unlock();
8370
8371 return ret;
8372}
8373
8374static int cpu_stats_show(struct seq_file *sf, void *v)
8375{
8376 struct task_group *tg = css_tg(seq_css(sf));
8377 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8378
8379 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8380 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8381 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8382
8383 return 0;
8384}
8385#endif /* CONFIG_CFS_BANDWIDTH */
8386#endif /* CONFIG_FAIR_GROUP_SCHED */
8387
8388#ifdef CONFIG_RT_GROUP_SCHED
8389static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8390 struct cftype *cft, s64 val)
8391{
8392 return sched_group_set_rt_runtime(css_tg(css), val);
8393}
8394
8395static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8396 struct cftype *cft)
8397{
8398 return sched_group_rt_runtime(css_tg(css));
8399}
8400
8401static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8402 struct cftype *cftype, u64 rt_period_us)
8403{
8404 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8405}
8406
8407static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8408 struct cftype *cft)
8409{
8410 return sched_group_rt_period(css_tg(css));
8411}
8412#endif /* CONFIG_RT_GROUP_SCHED */
8413
8414static struct cftype cpu_files[] = {
8415#ifdef CONFIG_FAIR_GROUP_SCHED
8416 {
8417 .name = "shares",
8418 .read_u64 = cpu_shares_read_u64,
8419 .write_u64 = cpu_shares_write_u64,
8420 },
8421#endif
8422#ifdef CONFIG_CFS_BANDWIDTH
8423 {
8424 .name = "cfs_quota_us",
8425 .read_s64 = cpu_cfs_quota_read_s64,
8426 .write_s64 = cpu_cfs_quota_write_s64,
8427 },
8428 {
8429 .name = "cfs_period_us",
8430 .read_u64 = cpu_cfs_period_read_u64,
8431 .write_u64 = cpu_cfs_period_write_u64,
8432 },
8433 {
8434 .name = "stat",
8435 .seq_show = cpu_stats_show,
8436 },
8437#endif
8438#ifdef CONFIG_RT_GROUP_SCHED
8439 {
8440 .name = "rt_runtime_us",
8441 .read_s64 = cpu_rt_runtime_read,
8442 .write_s64 = cpu_rt_runtime_write,
8443 },
8444 {
8445 .name = "rt_period_us",
8446 .read_u64 = cpu_rt_period_read_uint,
8447 .write_u64 = cpu_rt_period_write_uint,
8448 },
8449#endif
8450 { } /* terminate */
8451};
8452
8453struct cgroup_subsys cpu_cgrp_subsys = {
8454 .css_alloc = cpu_cgroup_css_alloc,
8455 .css_released = cpu_cgroup_css_released,
8456 .css_free = cpu_cgroup_css_free,
8457 .fork = cpu_cgroup_fork,
8458 .can_attach = cpu_cgroup_can_attach,
8459 .attach = cpu_cgroup_attach,
8460 .legacy_cftypes = cpu_files,
8461 .early_init = true,
8462};
8463
8464#endif /* CONFIG_CGROUP_SCHED */
8465
8466void dump_cpu_task(int cpu)
8467{
8468 pr_info("Task dump for CPU %d:\n", cpu);
8469 sched_show_task(cpu_curr(cpu));
8470}
8471
8472/*
8473 * Nice levels are multiplicative, with a gentle 10% change for every
8474 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8475 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8476 * that remained on nice 0.
8477 *
8478 * The "10% effect" is relative and cumulative: from _any_ nice level,
8479 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8480 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8481 * If a task goes up by ~10% and another task goes down by ~10% then
8482 * the relative distance between them is ~25%.)
8483 */
8484const int sched_prio_to_weight[40] = {
8485 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8486 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8487 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8488 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8489 /* 0 */ 1024, 820, 655, 526, 423,
8490 /* 5 */ 335, 272, 215, 172, 137,
8491 /* 10 */ 110, 87, 70, 56, 45,
8492 /* 15 */ 36, 29, 23, 18, 15,
8493};
8494
8495/*
8496 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8497 *
8498 * In cases where the weight does not change often, we can use the
8499 * precalculated inverse to speed up arithmetics by turning divisions
8500 * into multiplications:
8501 */
8502const u32 sched_prio_to_wmult[40] = {
8503 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8504 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8505 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8506 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8507 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8508 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8509 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8510 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8511};