Version:  2.0.40 2.2.26 2.4.37 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10

Linux/kernel/sched/core.c

  1 /*
  2  *  kernel/sched/core.c
  3  *
  4  *  Kernel scheduler and related syscalls
  5  *
  6  *  Copyright (C) 1991-2002  Linus Torvalds
  7  *
  8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
  9  *              make semaphores SMP safe
 10  *  1998-11-19  Implemented schedule_timeout() and related stuff
 11  *              by Andrea Arcangeli
 12  *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
 13  *              hybrid priority-list and round-robin design with
 14  *              an array-switch method of distributing timeslices
 15  *              and per-CPU runqueues.  Cleanups and useful suggestions
 16  *              by Davide Libenzi, preemptible kernel bits by Robert Love.
 17  *  2003-09-03  Interactivity tuning by Con Kolivas.
 18  *  2004-04-02  Scheduler domains code by Nick Piggin
 19  *  2007-04-15  Work begun on replacing all interactivity tuning with a
 20  *              fair scheduling design by Con Kolivas.
 21  *  2007-05-05  Load balancing (smp-nice) and other improvements
 22  *              by Peter Williams
 23  *  2007-05-06  Interactivity improvements to CFS by Mike Galbraith
 24  *  2007-07-01  Group scheduling enhancements by Srivatsa Vaddagiri
 25  *  2007-11-29  RT balancing improvements by Steven Rostedt, Gregory Haskins,
 26  *              Thomas Gleixner, Mike Kravetz
 27  */
 28 
 29 #include <linux/mm.h>
 30 #include <linux/module.h>
 31 #include <linux/nmi.h>
 32 #include <linux/init.h>
 33 #include <linux/uaccess.h>
 34 #include <linux/highmem.h>
 35 #include <asm/mmu_context.h>
 36 #include <linux/interrupt.h>
 37 #include <linux/capability.h>
 38 #include <linux/completion.h>
 39 #include <linux/kernel_stat.h>
 40 #include <linux/debug_locks.h>
 41 #include <linux/perf_event.h>
 42 #include <linux/security.h>
 43 #include <linux/notifier.h>
 44 #include <linux/profile.h>
 45 #include <linux/freezer.h>
 46 #include <linux/vmalloc.h>
 47 #include <linux/blkdev.h>
 48 #include <linux/delay.h>
 49 #include <linux/pid_namespace.h>
 50 #include <linux/smp.h>
 51 #include <linux/threads.h>
 52 #include <linux/timer.h>
 53 #include <linux/rcupdate.h>
 54 #include <linux/cpu.h>
 55 #include <linux/cpuset.h>
 56 #include <linux/percpu.h>
 57 #include <linux/proc_fs.h>
 58 #include <linux/seq_file.h>
 59 #include <linux/sysctl.h>
 60 #include <linux/syscalls.h>
 61 #include <linux/times.h>
 62 #include <linux/tsacct_kern.h>
 63 #include <linux/kprobes.h>
 64 #include <linux/delayacct.h>
 65 #include <linux/unistd.h>
 66 #include <linux/pagemap.h>
 67 #include <linux/hrtimer.h>
 68 #include <linux/tick.h>
 69 #include <linux/debugfs.h>
 70 #include <linux/ctype.h>
 71 #include <linux/ftrace.h>
 72 #include <linux/slab.h>
 73 #include <linux/init_task.h>
 74 #include <linux/binfmts.h>
 75 #include <linux/context_tracking.h>
 76 #include <linux/compiler.h>
 77 
 78 #include <asm/switch_to.h>
 79 #include <asm/tlb.h>
 80 #include <asm/irq_regs.h>
 81 #include <asm/mutex.h>
 82 #ifdef CONFIG_PARAVIRT
 83 #include <asm/paravirt.h>
 84 #endif
 85 
 86 #include "sched.h"
 87 #include "../workqueue_internal.h"
 88 #include "../smpboot.h"
 89 
 90 #define CREATE_TRACE_POINTS
 91 #include <trace/events/sched.h>
 92 
 93 #ifdef smp_mb__before_atomic
 94 void __smp_mb__before_atomic(void)
 95 {
 96         smp_mb__before_atomic();
 97 }
 98 EXPORT_SYMBOL(__smp_mb__before_atomic);
 99 #endif
100 
101 #ifdef smp_mb__after_atomic
102 void __smp_mb__after_atomic(void)
103 {
104         smp_mb__after_atomic();
105 }
106 EXPORT_SYMBOL(__smp_mb__after_atomic);
107 #endif
108 
109 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
110 {
111         unsigned long delta;
112         ktime_t soft, hard, now;
113 
114         for (;;) {
115                 if (hrtimer_active(period_timer))
116                         break;
117 
118                 now = hrtimer_cb_get_time(period_timer);
119                 hrtimer_forward(period_timer, now, period);
120 
121                 soft = hrtimer_get_softexpires(period_timer);
122                 hard = hrtimer_get_expires(period_timer);
123                 delta = ktime_to_ns(ktime_sub(hard, soft));
124                 __hrtimer_start_range_ns(period_timer, soft, delta,
125                                          HRTIMER_MODE_ABS_PINNED, 0);
126         }
127 }
128 
129 DEFINE_MUTEX(sched_domains_mutex);
130 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
131 
132 static void update_rq_clock_task(struct rq *rq, s64 delta);
133 
134 void update_rq_clock(struct rq *rq)
135 {
136         s64 delta;
137 
138         if (rq->skip_clock_update > 0)
139                 return;
140 
141         delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
142         rq->clock += delta;
143         update_rq_clock_task(rq, delta);
144 }
145 
146 /*
147  * Debugging: various feature bits
148  */
149 
150 #define SCHED_FEAT(name, enabled)       \
151         (1UL << __SCHED_FEAT_##name) * enabled |
152 
153 const_debug unsigned int sysctl_sched_features =
154 #include "features.h"
155         0;
156 
157 #undef SCHED_FEAT
158 
159 #ifdef CONFIG_SCHED_DEBUG
160 #define SCHED_FEAT(name, enabled)       \
161         #name ,
162 
163 static const char * const sched_feat_names[] = {
164 #include "features.h"
165 };
166 
167 #undef SCHED_FEAT
168 
169 static int sched_feat_show(struct seq_file *m, void *v)
170 {
171         int i;
172 
173         for (i = 0; i < __SCHED_FEAT_NR; i++) {
174                 if (!(sysctl_sched_features & (1UL << i)))
175                         seq_puts(m, "NO_");
176                 seq_printf(m, "%s ", sched_feat_names[i]);
177         }
178         seq_puts(m, "\n");
179 
180         return 0;
181 }
182 
183 #ifdef HAVE_JUMP_LABEL
184 
185 #define jump_label_key__true  STATIC_KEY_INIT_TRUE
186 #define jump_label_key__false STATIC_KEY_INIT_FALSE
187 
188 #define SCHED_FEAT(name, enabled)       \
189         jump_label_key__##enabled ,
190 
191 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
192 #include "features.h"
193 };
194 
195 #undef SCHED_FEAT
196 
197 static void sched_feat_disable(int i)
198 {
199         if (static_key_enabled(&sched_feat_keys[i]))
200                 static_key_slow_dec(&sched_feat_keys[i]);
201 }
202 
203 static void sched_feat_enable(int i)
204 {
205         if (!static_key_enabled(&sched_feat_keys[i]))
206                 static_key_slow_inc(&sched_feat_keys[i]);
207 }
208 #else
209 static void sched_feat_disable(int i) { };
210 static void sched_feat_enable(int i) { };
211 #endif /* HAVE_JUMP_LABEL */
212 
213 static int sched_feat_set(char *cmp)
214 {
215         int i;
216         int neg = 0;
217 
218         if (strncmp(cmp, "NO_", 3) == 0) {
219                 neg = 1;
220                 cmp += 3;
221         }
222 
223         for (i = 0; i < __SCHED_FEAT_NR; i++) {
224                 if (strcmp(cmp, sched_feat_names[i]) == 0) {
225                         if (neg) {
226                                 sysctl_sched_features &= ~(1UL << i);
227                                 sched_feat_disable(i);
228                         } else {
229                                 sysctl_sched_features |= (1UL << i);
230                                 sched_feat_enable(i);
231                         }
232                         break;
233                 }
234         }
235 
236         return i;
237 }
238 
239 static ssize_t
240 sched_feat_write(struct file *filp, const char __user *ubuf,
241                 size_t cnt, loff_t *ppos)
242 {
243         char buf[64];
244         char *cmp;
245         int i;
246 
247         if (cnt > 63)
248                 cnt = 63;
249 
250         if (copy_from_user(&buf, ubuf, cnt))
251                 return -EFAULT;
252 
253         buf[cnt] = 0;
254         cmp = strstrip(buf);
255 
256         i = sched_feat_set(cmp);
257         if (i == __SCHED_FEAT_NR)
258                 return -EINVAL;
259 
260         *ppos += cnt;
261 
262         return cnt;
263 }
264 
265 static int sched_feat_open(struct inode *inode, struct file *filp)
266 {
267         return single_open(filp, sched_feat_show, NULL);
268 }
269 
270 static const struct file_operations sched_feat_fops = {
271         .open           = sched_feat_open,
272         .write          = sched_feat_write,
273         .read           = seq_read,
274         .llseek         = seq_lseek,
275         .release        = single_release,
276 };
277 
278 static __init int sched_init_debug(void)
279 {
280         debugfs_create_file("sched_features", 0644, NULL, NULL,
281                         &sched_feat_fops);
282 
283         return 0;
284 }
285 late_initcall(sched_init_debug);
286 #endif /* CONFIG_SCHED_DEBUG */
287 
288 /*
289  * Number of tasks to iterate in a single balance run.
290  * Limited because this is done with IRQs disabled.
291  */
292 const_debug unsigned int sysctl_sched_nr_migrate = 32;
293 
294 /*
295  * period over which we average the RT time consumption, measured
296  * in ms.
297  *
298  * default: 1s
299  */
300 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
301 
302 /*
303  * period over which we measure -rt task cpu usage in us.
304  * default: 1s
305  */
306 unsigned int sysctl_sched_rt_period = 1000000;
307 
308 __read_mostly int scheduler_running;
309 
310 /*
311  * part of the period that we allow rt tasks to run in us.
312  * default: 0.95s
313  */
314 int sysctl_sched_rt_runtime = 950000;
315 
316 /*
317  * __task_rq_lock - lock the rq @p resides on.
318  */
319 static inline struct rq *__task_rq_lock(struct task_struct *p)
320         __acquires(rq->lock)
321 {
322         struct rq *rq;
323 
324         lockdep_assert_held(&p->pi_lock);
325 
326         for (;;) {
327                 rq = task_rq(p);
328                 raw_spin_lock(&rq->lock);
329                 if (likely(rq == task_rq(p)))
330                         return rq;
331                 raw_spin_unlock(&rq->lock);
332         }
333 }
334 
335 /*
336  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
337  */
338 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
339         __acquires(p->pi_lock)
340         __acquires(rq->lock)
341 {
342         struct rq *rq;
343 
344         for (;;) {
345                 raw_spin_lock_irqsave(&p->pi_lock, *flags);
346                 rq = task_rq(p);
347                 raw_spin_lock(&rq->lock);
348                 if (likely(rq == task_rq(p)))
349                         return rq;
350                 raw_spin_unlock(&rq->lock);
351                 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
352         }
353 }
354 
355 static void __task_rq_unlock(struct rq *rq)
356         __releases(rq->lock)
357 {
358         raw_spin_unlock(&rq->lock);
359 }
360 
361 static inline void
362 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
363         __releases(rq->lock)
364         __releases(p->pi_lock)
365 {
366         raw_spin_unlock(&rq->lock);
367         raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
368 }
369 
370 /*
371  * this_rq_lock - lock this runqueue and disable interrupts.
372  */
373 static struct rq *this_rq_lock(void)
374         __acquires(rq->lock)
375 {
376         struct rq *rq;
377 
378         local_irq_disable();
379         rq = this_rq();
380         raw_spin_lock(&rq->lock);
381 
382         return rq;
383 }
384 
385 #ifdef CONFIG_SCHED_HRTICK
386 /*
387  * Use HR-timers to deliver accurate preemption points.
388  */
389 
390 static void hrtick_clear(struct rq *rq)
391 {
392         if (hrtimer_active(&rq->hrtick_timer))
393                 hrtimer_cancel(&rq->hrtick_timer);
394 }
395 
396 /*
397  * High-resolution timer tick.
398  * Runs from hardirq context with interrupts disabled.
399  */
400 static enum hrtimer_restart hrtick(struct hrtimer *timer)
401 {
402         struct rq *rq = container_of(timer, struct rq, hrtick_timer);
403 
404         WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
405 
406         raw_spin_lock(&rq->lock);
407         update_rq_clock(rq);
408         rq->curr->sched_class->task_tick(rq, rq->curr, 1);
409         raw_spin_unlock(&rq->lock);
410 
411         return HRTIMER_NORESTART;
412 }
413 
414 #ifdef CONFIG_SMP
415 
416 static int __hrtick_restart(struct rq *rq)
417 {
418         struct hrtimer *timer = &rq->hrtick_timer;
419         ktime_t time = hrtimer_get_softexpires(timer);
420 
421         return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
422 }
423 
424 /*
425  * called from hardirq (IPI) context
426  */
427 static void __hrtick_start(void *arg)
428 {
429         struct rq *rq = arg;
430 
431         raw_spin_lock(&rq->lock);
432         __hrtick_restart(rq);
433         rq->hrtick_csd_pending = 0;
434         raw_spin_unlock(&rq->lock);
435 }
436 
437 /*
438  * Called to set the hrtick timer state.
439  *
440  * called with rq->lock held and irqs disabled
441  */
442 void hrtick_start(struct rq *rq, u64 delay)
443 {
444         struct hrtimer *timer = &rq->hrtick_timer;
445         ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
446 
447         hrtimer_set_expires(timer, time);
448 
449         if (rq == this_rq()) {
450                 __hrtick_restart(rq);
451         } else if (!rq->hrtick_csd_pending) {
452                 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
453                 rq->hrtick_csd_pending = 1;
454         }
455 }
456 
457 static int
458 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
459 {
460         int cpu = (int)(long)hcpu;
461 
462         switch (action) {
463         case CPU_UP_CANCELED:
464         case CPU_UP_CANCELED_FROZEN:
465         case CPU_DOWN_PREPARE:
466         case CPU_DOWN_PREPARE_FROZEN:
467         case CPU_DEAD:
468         case CPU_DEAD_FROZEN:
469                 hrtick_clear(cpu_rq(cpu));
470                 return NOTIFY_OK;
471         }
472 
473         return NOTIFY_DONE;
474 }
475 
476 static __init void init_hrtick(void)
477 {
478         hotcpu_notifier(hotplug_hrtick, 0);
479 }
480 #else
481 /*
482  * Called to set the hrtick timer state.
483  *
484  * called with rq->lock held and irqs disabled
485  */
486 void hrtick_start(struct rq *rq, u64 delay)
487 {
488         __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
489                         HRTIMER_MODE_REL_PINNED, 0);
490 }
491 
492 static inline void init_hrtick(void)
493 {
494 }
495 #endif /* CONFIG_SMP */
496 
497 static void init_rq_hrtick(struct rq *rq)
498 {
499 #ifdef CONFIG_SMP
500         rq->hrtick_csd_pending = 0;
501 
502         rq->hrtick_csd.flags = 0;
503         rq->hrtick_csd.func = __hrtick_start;
504         rq->hrtick_csd.info = rq;
505 #endif
506 
507         hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
508         rq->hrtick_timer.function = hrtick;
509 }
510 #else   /* CONFIG_SCHED_HRTICK */
511 static inline void hrtick_clear(struct rq *rq)
512 {
513 }
514 
515 static inline void init_rq_hrtick(struct rq *rq)
516 {
517 }
518 
519 static inline void init_hrtick(void)
520 {
521 }
522 #endif  /* CONFIG_SCHED_HRTICK */
523 
524 /*
525  * cmpxchg based fetch_or, macro so it works for different integer types
526  */
527 #define fetch_or(ptr, val)                                              \
528 ({      typeof(*(ptr)) __old, __val = *(ptr);                           \
529         for (;;) {                                                      \
530                 __old = cmpxchg((ptr), __val, __val | (val));           \
531                 if (__old == __val)                                     \
532                         break;                                          \
533                 __val = __old;                                          \
534         }                                                               \
535         __old;                                                          \
536 })
537 
538 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
539 /*
540  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
541  * this avoids any races wrt polling state changes and thereby avoids
542  * spurious IPIs.
543  */
544 static bool set_nr_and_not_polling(struct task_struct *p)
545 {
546         struct thread_info *ti = task_thread_info(p);
547         return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
548 }
549 
550 /*
551  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
552  *
553  * If this returns true, then the idle task promises to call
554  * sched_ttwu_pending() and reschedule soon.
555  */
556 static bool set_nr_if_polling(struct task_struct *p)
557 {
558         struct thread_info *ti = task_thread_info(p);
559         typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
560 
561         for (;;) {
562                 if (!(val & _TIF_POLLING_NRFLAG))
563                         return false;
564                 if (val & _TIF_NEED_RESCHED)
565                         return true;
566                 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
567                 if (old == val)
568                         break;
569                 val = old;
570         }
571         return true;
572 }
573 
574 #else
575 static bool set_nr_and_not_polling(struct task_struct *p)
576 {
577         set_tsk_need_resched(p);
578         return true;
579 }
580 
581 #ifdef CONFIG_SMP
582 static bool set_nr_if_polling(struct task_struct *p)
583 {
584         return false;
585 }
586 #endif
587 #endif
588 
589 /*
590  * resched_task - mark a task 'to be rescheduled now'.
591  *
592  * On UP this means the setting of the need_resched flag, on SMP it
593  * might also involve a cross-CPU call to trigger the scheduler on
594  * the target CPU.
595  */
596 void resched_task(struct task_struct *p)
597 {
598         int cpu;
599 
600         lockdep_assert_held(&task_rq(p)->lock);
601 
602         if (test_tsk_need_resched(p))
603                 return;
604 
605         cpu = task_cpu(p);
606 
607         if (cpu == smp_processor_id()) {
608                 set_tsk_need_resched(p);
609                 set_preempt_need_resched();
610                 return;
611         }
612 
613         if (set_nr_and_not_polling(p))
614                 smp_send_reschedule(cpu);
615         else
616                 trace_sched_wake_idle_without_ipi(cpu);
617 }
618 
619 void resched_cpu(int cpu)
620 {
621         struct rq *rq = cpu_rq(cpu);
622         unsigned long flags;
623 
624         if (!raw_spin_trylock_irqsave(&rq->lock, flags))
625                 return;
626         resched_task(cpu_curr(cpu));
627         raw_spin_unlock_irqrestore(&rq->lock, flags);
628 }
629 
630 #ifdef CONFIG_SMP
631 #ifdef CONFIG_NO_HZ_COMMON
632 /*
633  * In the semi idle case, use the nearest busy cpu for migrating timers
634  * from an idle cpu.  This is good for power-savings.
635  *
636  * We don't do similar optimization for completely idle system, as
637  * selecting an idle cpu will add more delays to the timers than intended
638  * (as that cpu's timer base may not be uptodate wrt jiffies etc).
639  */
640 int get_nohz_timer_target(int pinned)
641 {
642         int cpu = smp_processor_id();
643         int i;
644         struct sched_domain *sd;
645 
646         if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
647                 return cpu;
648 
649         rcu_read_lock();
650         for_each_domain(cpu, sd) {
651                 for_each_cpu(i, sched_domain_span(sd)) {
652                         if (!idle_cpu(i)) {
653                                 cpu = i;
654                                 goto unlock;
655                         }
656                 }
657         }
658 unlock:
659         rcu_read_unlock();
660         return cpu;
661 }
662 /*
663  * When add_timer_on() enqueues a timer into the timer wheel of an
664  * idle CPU then this timer might expire before the next timer event
665  * which is scheduled to wake up that CPU. In case of a completely
666  * idle system the next event might even be infinite time into the
667  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
668  * leaves the inner idle loop so the newly added timer is taken into
669  * account when the CPU goes back to idle and evaluates the timer
670  * wheel for the next timer event.
671  */
672 static void wake_up_idle_cpu(int cpu)
673 {
674         struct rq *rq = cpu_rq(cpu);
675 
676         if (cpu == smp_processor_id())
677                 return;
678 
679         if (set_nr_and_not_polling(rq->idle))
680                 smp_send_reschedule(cpu);
681         else
682                 trace_sched_wake_idle_without_ipi(cpu);
683 }
684 
685 static bool wake_up_full_nohz_cpu(int cpu)
686 {
687         if (tick_nohz_full_cpu(cpu)) {
688                 if (cpu != smp_processor_id() ||
689                     tick_nohz_tick_stopped())
690                         smp_send_reschedule(cpu);
691                 return true;
692         }
693 
694         return false;
695 }
696 
697 void wake_up_nohz_cpu(int cpu)
698 {
699         if (!wake_up_full_nohz_cpu(cpu))
700                 wake_up_idle_cpu(cpu);
701 }
702 
703 static inline bool got_nohz_idle_kick(void)
704 {
705         int cpu = smp_processor_id();
706 
707         if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
708                 return false;
709 
710         if (idle_cpu(cpu) && !need_resched())
711                 return true;
712 
713         /*
714          * We can't run Idle Load Balance on this CPU for this time so we
715          * cancel it and clear NOHZ_BALANCE_KICK
716          */
717         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
718         return false;
719 }
720 
721 #else /* CONFIG_NO_HZ_COMMON */
722 
723 static inline bool got_nohz_idle_kick(void)
724 {
725         return false;
726 }
727 
728 #endif /* CONFIG_NO_HZ_COMMON */
729 
730 #ifdef CONFIG_NO_HZ_FULL
731 bool sched_can_stop_tick(void)
732 {
733        struct rq *rq;
734 
735        rq = this_rq();
736 
737        /* Make sure rq->nr_running update is visible after the IPI */
738        smp_rmb();
739 
740        /* More than one running task need preemption */
741        if (rq->nr_running > 1)
742                return false;
743 
744        return true;
745 }
746 #endif /* CONFIG_NO_HZ_FULL */
747 
748 void sched_avg_update(struct rq *rq)
749 {
750         s64 period = sched_avg_period();
751 
752         while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
753                 /*
754                  * Inline assembly required to prevent the compiler
755                  * optimising this loop into a divmod call.
756                  * See __iter_div_u64_rem() for another example of this.
757                  */
758                 asm("" : "+rm" (rq->age_stamp));
759                 rq->age_stamp += period;
760                 rq->rt_avg /= 2;
761         }
762 }
763 
764 #endif /* CONFIG_SMP */
765 
766 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
767                         (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
768 /*
769  * Iterate task_group tree rooted at *from, calling @down when first entering a
770  * node and @up when leaving it for the final time.
771  *
772  * Caller must hold rcu_lock or sufficient equivalent.
773  */
774 int walk_tg_tree_from(struct task_group *from,
775                              tg_visitor down, tg_visitor up, void *data)
776 {
777         struct task_group *parent, *child;
778         int ret;
779 
780         parent = from;
781 
782 down:
783         ret = (*down)(parent, data);
784         if (ret)
785                 goto out;
786         list_for_each_entry_rcu(child, &parent->children, siblings) {
787                 parent = child;
788                 goto down;
789 
790 up:
791                 continue;
792         }
793         ret = (*up)(parent, data);
794         if (ret || parent == from)
795                 goto out;
796 
797         child = parent;
798         parent = parent->parent;
799         if (parent)
800                 goto up;
801 out:
802         return ret;
803 }
804 
805 int tg_nop(struct task_group *tg, void *data)
806 {
807         return 0;
808 }
809 #endif
810 
811 static void set_load_weight(struct task_struct *p)
812 {
813         int prio = p->static_prio - MAX_RT_PRIO;
814         struct load_weight *load = &p->se.load;
815 
816         /*
817          * SCHED_IDLE tasks get minimal weight:
818          */
819         if (p->policy == SCHED_IDLE) {
820                 load->weight = scale_load(WEIGHT_IDLEPRIO);
821                 load->inv_weight = WMULT_IDLEPRIO;
822                 return;
823         }
824 
825         load->weight = scale_load(prio_to_weight[prio]);
826         load->inv_weight = prio_to_wmult[prio];
827 }
828 
829 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
830 {
831         update_rq_clock(rq);
832         sched_info_queued(rq, p);
833         p->sched_class->enqueue_task(rq, p, flags);
834 }
835 
836 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
837 {
838         update_rq_clock(rq);
839         sched_info_dequeued(rq, p);
840         p->sched_class->dequeue_task(rq, p, flags);
841 }
842 
843 void activate_task(struct rq *rq, struct task_struct *p, int flags)
844 {
845         if (task_contributes_to_load(p))
846                 rq->nr_uninterruptible--;
847 
848         enqueue_task(rq, p, flags);
849 }
850 
851 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
852 {
853         if (task_contributes_to_load(p))
854                 rq->nr_uninterruptible++;
855 
856         dequeue_task(rq, p, flags);
857 }
858 
859 static void update_rq_clock_task(struct rq *rq, s64 delta)
860 {
861 /*
862  * In theory, the compile should just see 0 here, and optimize out the call
863  * to sched_rt_avg_update. But I don't trust it...
864  */
865 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
866         s64 steal = 0, irq_delta = 0;
867 #endif
868 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
869         irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
870 
871         /*
872          * Since irq_time is only updated on {soft,}irq_exit, we might run into
873          * this case when a previous update_rq_clock() happened inside a
874          * {soft,}irq region.
875          *
876          * When this happens, we stop ->clock_task and only update the
877          * prev_irq_time stamp to account for the part that fit, so that a next
878          * update will consume the rest. This ensures ->clock_task is
879          * monotonic.
880          *
881          * It does however cause some slight miss-attribution of {soft,}irq
882          * time, a more accurate solution would be to update the irq_time using
883          * the current rq->clock timestamp, except that would require using
884          * atomic ops.
885          */
886         if (irq_delta > delta)
887                 irq_delta = delta;
888 
889         rq->prev_irq_time += irq_delta;
890         delta -= irq_delta;
891 #endif
892 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
893         if (static_key_false((&paravirt_steal_rq_enabled))) {
894                 steal = paravirt_steal_clock(cpu_of(rq));
895                 steal -= rq->prev_steal_time_rq;
896 
897                 if (unlikely(steal > delta))
898                         steal = delta;
899 
900                 rq->prev_steal_time_rq += steal;
901                 delta -= steal;
902         }
903 #endif
904 
905         rq->clock_task += delta;
906 
907 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
908         if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
909                 sched_rt_avg_update(rq, irq_delta + steal);
910 #endif
911 }
912 
913 void sched_set_stop_task(int cpu, struct task_struct *stop)
914 {
915         struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
916         struct task_struct *old_stop = cpu_rq(cpu)->stop;
917 
918         if (stop) {
919                 /*
920                  * Make it appear like a SCHED_FIFO task, its something
921                  * userspace knows about and won't get confused about.
922                  *
923                  * Also, it will make PI more or less work without too
924                  * much confusion -- but then, stop work should not
925                  * rely on PI working anyway.
926                  */
927                 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
928 
929                 stop->sched_class = &stop_sched_class;
930         }
931 
932         cpu_rq(cpu)->stop = stop;
933 
934         if (old_stop) {
935                 /*
936                  * Reset it back to a normal scheduling class so that
937                  * it can die in pieces.
938                  */
939                 old_stop->sched_class = &rt_sched_class;
940         }
941 }
942 
943 /*
944  * __normal_prio - return the priority that is based on the static prio
945  */
946 static inline int __normal_prio(struct task_struct *p)
947 {
948         return p->static_prio;
949 }
950 
951 /*
952  * Calculate the expected normal priority: i.e. priority
953  * without taking RT-inheritance into account. Might be
954  * boosted by interactivity modifiers. Changes upon fork,
955  * setprio syscalls, and whenever the interactivity
956  * estimator recalculates.
957  */
958 static inline int normal_prio(struct task_struct *p)
959 {
960         int prio;
961 
962         if (task_has_dl_policy(p))
963                 prio = MAX_DL_PRIO-1;
964         else if (task_has_rt_policy(p))
965                 prio = MAX_RT_PRIO-1 - p->rt_priority;
966         else
967                 prio = __normal_prio(p);
968         return prio;
969 }
970 
971 /*
972  * Calculate the current priority, i.e. the priority
973  * taken into account by the scheduler. This value might
974  * be boosted by RT tasks, or might be boosted by
975  * interactivity modifiers. Will be RT if the task got
976  * RT-boosted. If not then it returns p->normal_prio.
977  */
978 static int effective_prio(struct task_struct *p)
979 {
980         p->normal_prio = normal_prio(p);
981         /*
982          * If we are RT tasks or we were boosted to RT priority,
983          * keep the priority unchanged. Otherwise, update priority
984          * to the normal priority:
985          */
986         if (!rt_prio(p->prio))
987                 return p->normal_prio;
988         return p->prio;
989 }
990 
991 /**
992  * task_curr - is this task currently executing on a CPU?
993  * @p: the task in question.
994  *
995  * Return: 1 if the task is currently executing. 0 otherwise.
996  */
997 inline int task_curr(const struct task_struct *p)
998 {
999         return cpu_curr(task_cpu(p)) == p;
1000 }
1001 
1002 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1003                                        const struct sched_class *prev_class,
1004                                        int oldprio)
1005 {
1006         if (prev_class != p->sched_class) {
1007                 if (prev_class->switched_from)
1008                         prev_class->switched_from(rq, p);
1009                 p->sched_class->switched_to(rq, p);
1010         } else if (oldprio != p->prio || dl_task(p))
1011                 p->sched_class->prio_changed(rq, p, oldprio);
1012 }
1013 
1014 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1015 {
1016         const struct sched_class *class;
1017 
1018         if (p->sched_class == rq->curr->sched_class) {
1019                 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1020         } else {
1021                 for_each_class(class) {
1022                         if (class == rq->curr->sched_class)
1023                                 break;
1024                         if (class == p->sched_class) {
1025                                 resched_task(rq->curr);
1026                                 break;
1027                         }
1028                 }
1029         }
1030 
1031         /*
1032          * A queue event has occurred, and we're going to schedule.  In
1033          * this case, we can save a useless back to back clock update.
1034          */
1035         if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1036                 rq->skip_clock_update = 1;
1037 }
1038 
1039 #ifdef CONFIG_SMP
1040 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1041 {
1042 #ifdef CONFIG_SCHED_DEBUG
1043         /*
1044          * We should never call set_task_cpu() on a blocked task,
1045          * ttwu() will sort out the placement.
1046          */
1047         WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1048                         !(task_preempt_count(p) & PREEMPT_ACTIVE));
1049 
1050 #ifdef CONFIG_LOCKDEP
1051         /*
1052          * The caller should hold either p->pi_lock or rq->lock, when changing
1053          * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1054          *
1055          * sched_move_task() holds both and thus holding either pins the cgroup,
1056          * see task_group().
1057          *
1058          * Furthermore, all task_rq users should acquire both locks, see
1059          * task_rq_lock().
1060          */
1061         WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1062                                       lockdep_is_held(&task_rq(p)->lock)));
1063 #endif
1064 #endif
1065 
1066         trace_sched_migrate_task(p, new_cpu);
1067 
1068         if (task_cpu(p) != new_cpu) {
1069                 if (p->sched_class->migrate_task_rq)
1070                         p->sched_class->migrate_task_rq(p, new_cpu);
1071                 p->se.nr_migrations++;
1072                 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1073         }
1074 
1075         __set_task_cpu(p, new_cpu);
1076 }
1077 
1078 static void __migrate_swap_task(struct task_struct *p, int cpu)
1079 {
1080         if (p->on_rq) {
1081                 struct rq *src_rq, *dst_rq;
1082 
1083                 src_rq = task_rq(p);
1084                 dst_rq = cpu_rq(cpu);
1085 
1086                 deactivate_task(src_rq, p, 0);
1087                 set_task_cpu(p, cpu);
1088                 activate_task(dst_rq, p, 0);
1089                 check_preempt_curr(dst_rq, p, 0);
1090         } else {
1091                 /*
1092                  * Task isn't running anymore; make it appear like we migrated
1093                  * it before it went to sleep. This means on wakeup we make the
1094                  * previous cpu our targer instead of where it really is.
1095                  */
1096                 p->wake_cpu = cpu;
1097         }
1098 }
1099 
1100 struct migration_swap_arg {
1101         struct task_struct *src_task, *dst_task;
1102         int src_cpu, dst_cpu;
1103 };
1104 
1105 static int migrate_swap_stop(void *data)
1106 {
1107         struct migration_swap_arg *arg = data;
1108         struct rq *src_rq, *dst_rq;
1109         int ret = -EAGAIN;
1110 
1111         src_rq = cpu_rq(arg->src_cpu);
1112         dst_rq = cpu_rq(arg->dst_cpu);
1113 
1114         double_raw_lock(&arg->src_task->pi_lock,
1115                         &arg->dst_task->pi_lock);
1116         double_rq_lock(src_rq, dst_rq);
1117         if (task_cpu(arg->dst_task) != arg->dst_cpu)
1118                 goto unlock;
1119 
1120         if (task_cpu(arg->src_task) != arg->src_cpu)
1121                 goto unlock;
1122 
1123         if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1124                 goto unlock;
1125 
1126         if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1127                 goto unlock;
1128 
1129         __migrate_swap_task(arg->src_task, arg->dst_cpu);
1130         __migrate_swap_task(arg->dst_task, arg->src_cpu);
1131 
1132         ret = 0;
1133 
1134 unlock:
1135         double_rq_unlock(src_rq, dst_rq);
1136         raw_spin_unlock(&arg->dst_task->pi_lock);
1137         raw_spin_unlock(&arg->src_task->pi_lock);
1138 
1139         return ret;
1140 }
1141 
1142 /*
1143  * Cross migrate two tasks
1144  */
1145 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1146 {
1147         struct migration_swap_arg arg;
1148         int ret = -EINVAL;
1149 
1150         arg = (struct migration_swap_arg){
1151                 .src_task = cur,
1152                 .src_cpu = task_cpu(cur),
1153                 .dst_task = p,
1154                 .dst_cpu = task_cpu(p),
1155         };
1156 
1157         if (arg.src_cpu == arg.dst_cpu)
1158                 goto out;
1159 
1160         /*
1161          * These three tests are all lockless; this is OK since all of them
1162          * will be re-checked with proper locks held further down the line.
1163          */
1164         if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1165                 goto out;
1166 
1167         if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1168                 goto out;
1169 
1170         if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1171                 goto out;
1172 
1173         trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1174         ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1175 
1176 out:
1177         return ret;
1178 }
1179 
1180 struct migration_arg {
1181         struct task_struct *task;
1182         int dest_cpu;
1183 };
1184 
1185 static int migration_cpu_stop(void *data);
1186 
1187 /*
1188  * wait_task_inactive - wait for a thread to unschedule.
1189  *
1190  * If @match_state is nonzero, it's the @p->state value just checked and
1191  * not expected to change.  If it changes, i.e. @p might have woken up,
1192  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1193  * we return a positive number (its total switch count).  If a second call
1194  * a short while later returns the same number, the caller can be sure that
1195  * @p has remained unscheduled the whole time.
1196  *
1197  * The caller must ensure that the task *will* unschedule sometime soon,
1198  * else this function might spin for a *long* time. This function can't
1199  * be called with interrupts off, or it may introduce deadlock with
1200  * smp_call_function() if an IPI is sent by the same process we are
1201  * waiting to become inactive.
1202  */
1203 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1204 {
1205         unsigned long flags;
1206         int running, on_rq;
1207         unsigned long ncsw;
1208         struct rq *rq;
1209 
1210         for (;;) {
1211                 /*
1212                  * We do the initial early heuristics without holding
1213                  * any task-queue locks at all. We'll only try to get
1214                  * the runqueue lock when things look like they will
1215                  * work out!
1216                  */
1217                 rq = task_rq(p);
1218 
1219                 /*
1220                  * If the task is actively running on another CPU
1221                  * still, just relax and busy-wait without holding
1222                  * any locks.
1223                  *
1224                  * NOTE! Since we don't hold any locks, it's not
1225                  * even sure that "rq" stays as the right runqueue!
1226                  * But we don't care, since "task_running()" will
1227                  * return false if the runqueue has changed and p
1228                  * is actually now running somewhere else!
1229                  */
1230                 while (task_running(rq, p)) {
1231                         if (match_state && unlikely(p->state != match_state))
1232                                 return 0;
1233                         cpu_relax();
1234                 }
1235 
1236                 /*
1237                  * Ok, time to look more closely! We need the rq
1238                  * lock now, to be *sure*. If we're wrong, we'll
1239                  * just go back and repeat.
1240                  */
1241                 rq = task_rq_lock(p, &flags);
1242                 trace_sched_wait_task(p);
1243                 running = task_running(rq, p);
1244                 on_rq = p->on_rq;
1245                 ncsw = 0;
1246                 if (!match_state || p->state == match_state)
1247                         ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1248                 task_rq_unlock(rq, p, &flags);
1249 
1250                 /*
1251                  * If it changed from the expected state, bail out now.
1252                  */
1253                 if (unlikely(!ncsw))
1254                         break;
1255 
1256                 /*
1257                  * Was it really running after all now that we
1258                  * checked with the proper locks actually held?
1259                  *
1260                  * Oops. Go back and try again..
1261                  */
1262                 if (unlikely(running)) {
1263                         cpu_relax();
1264                         continue;
1265                 }
1266 
1267                 /*
1268                  * It's not enough that it's not actively running,
1269                  * it must be off the runqueue _entirely_, and not
1270                  * preempted!
1271                  *
1272                  * So if it was still runnable (but just not actively
1273                  * running right now), it's preempted, and we should
1274                  * yield - it could be a while.
1275                  */
1276                 if (unlikely(on_rq)) {
1277                         ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1278 
1279                         set_current_state(TASK_UNINTERRUPTIBLE);
1280                         schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1281                         continue;
1282                 }
1283 
1284                 /*
1285                  * Ahh, all good. It wasn't running, and it wasn't
1286                  * runnable, which means that it will never become
1287                  * running in the future either. We're all done!
1288                  */
1289                 break;
1290         }
1291 
1292         return ncsw;
1293 }
1294 
1295 /***
1296  * kick_process - kick a running thread to enter/exit the kernel
1297  * @p: the to-be-kicked thread
1298  *
1299  * Cause a process which is running on another CPU to enter
1300  * kernel-mode, without any delay. (to get signals handled.)
1301  *
1302  * NOTE: this function doesn't have to take the runqueue lock,
1303  * because all it wants to ensure is that the remote task enters
1304  * the kernel. If the IPI races and the task has been migrated
1305  * to another CPU then no harm is done and the purpose has been
1306  * achieved as well.
1307  */
1308 void kick_process(struct task_struct *p)
1309 {
1310         int cpu;
1311 
1312         preempt_disable();
1313         cpu = task_cpu(p);
1314         if ((cpu != smp_processor_id()) && task_curr(p))
1315                 smp_send_reschedule(cpu);
1316         preempt_enable();
1317 }
1318 EXPORT_SYMBOL_GPL(kick_process);
1319 #endif /* CONFIG_SMP */
1320 
1321 #ifdef CONFIG_SMP
1322 /*
1323  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1324  */
1325 static int select_fallback_rq(int cpu, struct task_struct *p)
1326 {
1327         int nid = cpu_to_node(cpu);
1328         const struct cpumask *nodemask = NULL;
1329         enum { cpuset, possible, fail } state = cpuset;
1330         int dest_cpu;
1331 
1332         /*
1333          * If the node that the cpu is on has been offlined, cpu_to_node()
1334          * will return -1. There is no cpu on the node, and we should
1335          * select the cpu on the other node.
1336          */
1337         if (nid != -1) {
1338                 nodemask = cpumask_of_node(nid);
1339 
1340                 /* Look for allowed, online CPU in same node. */
1341                 for_each_cpu(dest_cpu, nodemask) {
1342                         if (!cpu_online(dest_cpu))
1343                                 continue;
1344                         if (!cpu_active(dest_cpu))
1345                                 continue;
1346                         if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1347                                 return dest_cpu;
1348                 }
1349         }
1350 
1351         for (;;) {
1352                 /* Any allowed, online CPU? */
1353                 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1354                         if (!cpu_online(dest_cpu))
1355                                 continue;
1356                         if (!cpu_active(dest_cpu))
1357                                 continue;
1358                         goto out;
1359                 }
1360 
1361                 switch (state) {
1362                 case cpuset:
1363                         /* No more Mr. Nice Guy. */
1364                         cpuset_cpus_allowed_fallback(p);
1365                         state = possible;
1366                         break;
1367 
1368                 case possible:
1369                         do_set_cpus_allowed(p, cpu_possible_mask);
1370                         state = fail;
1371                         break;
1372 
1373                 case fail:
1374                         BUG();
1375                         break;
1376                 }
1377         }
1378 
1379 out:
1380         if (state != cpuset) {
1381                 /*
1382                  * Don't tell them about moving exiting tasks or
1383                  * kernel threads (both mm NULL), since they never
1384                  * leave kernel.
1385                  */
1386                 if (p->mm && printk_ratelimit()) {
1387                         printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1388                                         task_pid_nr(p), p->comm, cpu);
1389                 }
1390         }
1391 
1392         return dest_cpu;
1393 }
1394 
1395 /*
1396  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1397  */
1398 static inline
1399 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1400 {
1401         cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1402 
1403         /*
1404          * In order not to call set_task_cpu() on a blocking task we need
1405          * to rely on ttwu() to place the task on a valid ->cpus_allowed
1406          * cpu.
1407          *
1408          * Since this is common to all placement strategies, this lives here.
1409          *
1410          * [ this allows ->select_task() to simply return task_cpu(p) and
1411          *   not worry about this generic constraint ]
1412          */
1413         if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1414                      !cpu_online(cpu)))
1415                 cpu = select_fallback_rq(task_cpu(p), p);
1416 
1417         return cpu;
1418 }
1419 
1420 static void update_avg(u64 *avg, u64 sample)
1421 {
1422         s64 diff = sample - *avg;
1423         *avg += diff >> 3;
1424 }
1425 #endif
1426 
1427 static void
1428 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1429 {
1430 #ifdef CONFIG_SCHEDSTATS
1431         struct rq *rq = this_rq();
1432 
1433 #ifdef CONFIG_SMP
1434         int this_cpu = smp_processor_id();
1435 
1436         if (cpu == this_cpu) {
1437                 schedstat_inc(rq, ttwu_local);
1438                 schedstat_inc(p, se.statistics.nr_wakeups_local);
1439         } else {
1440                 struct sched_domain *sd;
1441 
1442                 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1443                 rcu_read_lock();
1444                 for_each_domain(this_cpu, sd) {
1445                         if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1446                                 schedstat_inc(sd, ttwu_wake_remote);
1447                                 break;
1448                         }
1449                 }
1450                 rcu_read_unlock();
1451         }
1452 
1453         if (wake_flags & WF_MIGRATED)
1454                 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1455 
1456 #endif /* CONFIG_SMP */
1457 
1458         schedstat_inc(rq, ttwu_count);
1459         schedstat_inc(p, se.statistics.nr_wakeups);
1460 
1461         if (wake_flags & WF_SYNC)
1462                 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1463 
1464 #endif /* CONFIG_SCHEDSTATS */
1465 }
1466 
1467 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1468 {
1469         activate_task(rq, p, en_flags);
1470         p->on_rq = 1;
1471 
1472         /* if a worker is waking up, notify workqueue */
1473         if (p->flags & PF_WQ_WORKER)
1474                 wq_worker_waking_up(p, cpu_of(rq));
1475 }
1476 
1477 /*
1478  * Mark the task runnable and perform wakeup-preemption.
1479  */
1480 static void
1481 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1482 {
1483         check_preempt_curr(rq, p, wake_flags);
1484         trace_sched_wakeup(p, true);
1485 
1486         p->state = TASK_RUNNING;
1487 #ifdef CONFIG_SMP
1488         if (p->sched_class->task_woken)
1489                 p->sched_class->task_woken(rq, p);
1490 
1491         if (rq->idle_stamp) {
1492                 u64 delta = rq_clock(rq) - rq->idle_stamp;
1493                 u64 max = 2*rq->max_idle_balance_cost;
1494 
1495                 update_avg(&rq->avg_idle, delta);
1496 
1497                 if (rq->avg_idle > max)
1498                         rq->avg_idle = max;
1499 
1500                 rq->idle_stamp = 0;
1501         }
1502 #endif
1503 }
1504 
1505 static void
1506 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1507 {
1508 #ifdef CONFIG_SMP
1509         if (p->sched_contributes_to_load)
1510                 rq->nr_uninterruptible--;
1511 #endif
1512 
1513         ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1514         ttwu_do_wakeup(rq, p, wake_flags);
1515 }
1516 
1517 /*
1518  * Called in case the task @p isn't fully descheduled from its runqueue,
1519  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1520  * since all we need to do is flip p->state to TASK_RUNNING, since
1521  * the task is still ->on_rq.
1522  */
1523 static int ttwu_remote(struct task_struct *p, int wake_flags)
1524 {
1525         struct rq *rq;
1526         int ret = 0;
1527 
1528         rq = __task_rq_lock(p);
1529         if (p->on_rq) {
1530                 /* check_preempt_curr() may use rq clock */
1531                 update_rq_clock(rq);
1532                 ttwu_do_wakeup(rq, p, wake_flags);
1533                 ret = 1;
1534         }
1535         __task_rq_unlock(rq);
1536 
1537         return ret;
1538 }
1539 
1540 #ifdef CONFIG_SMP
1541 void sched_ttwu_pending(void)
1542 {
1543         struct rq *rq = this_rq();
1544         struct llist_node *llist = llist_del_all(&rq->wake_list);
1545         struct task_struct *p;
1546         unsigned long flags;
1547 
1548         if (!llist)
1549                 return;
1550 
1551         raw_spin_lock_irqsave(&rq->lock, flags);
1552 
1553         while (llist) {
1554                 p = llist_entry(llist, struct task_struct, wake_entry);
1555                 llist = llist_next(llist);
1556                 ttwu_do_activate(rq, p, 0);
1557         }
1558 
1559         raw_spin_unlock_irqrestore(&rq->lock, flags);
1560 }
1561 
1562 void scheduler_ipi(void)
1563 {
1564         /*
1565          * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1566          * TIF_NEED_RESCHED remotely (for the first time) will also send
1567          * this IPI.
1568          */
1569         preempt_fold_need_resched();
1570 
1571         if (llist_empty(&this_rq()->wake_list)
1572                         && !tick_nohz_full_cpu(smp_processor_id())
1573                         && !got_nohz_idle_kick())
1574                 return;
1575 
1576         /*
1577          * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1578          * traditionally all their work was done from the interrupt return
1579          * path. Now that we actually do some work, we need to make sure
1580          * we do call them.
1581          *
1582          * Some archs already do call them, luckily irq_enter/exit nest
1583          * properly.
1584          *
1585          * Arguably we should visit all archs and update all handlers,
1586          * however a fair share of IPIs are still resched only so this would
1587          * somewhat pessimize the simple resched case.
1588          */
1589         irq_enter();
1590         tick_nohz_full_check();
1591         sched_ttwu_pending();
1592 
1593         /*
1594          * Check if someone kicked us for doing the nohz idle load balance.
1595          */
1596         if (unlikely(got_nohz_idle_kick())) {
1597                 this_rq()->idle_balance = 1;
1598                 raise_softirq_irqoff(SCHED_SOFTIRQ);
1599         }
1600         irq_exit();
1601 }
1602 
1603 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1604 {
1605         struct rq *rq = cpu_rq(cpu);
1606 
1607         if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1608                 if (!set_nr_if_polling(rq->idle))
1609                         smp_send_reschedule(cpu);
1610                 else
1611                         trace_sched_wake_idle_without_ipi(cpu);
1612         }
1613 }
1614 
1615 bool cpus_share_cache(int this_cpu, int that_cpu)
1616 {
1617         return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1618 }
1619 #endif /* CONFIG_SMP */
1620 
1621 static void ttwu_queue(struct task_struct *p, int cpu)
1622 {
1623         struct rq *rq = cpu_rq(cpu);
1624 
1625 #if defined(CONFIG_SMP)
1626         if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1627                 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1628                 ttwu_queue_remote(p, cpu);
1629                 return;
1630         }
1631 #endif
1632 
1633         raw_spin_lock(&rq->lock);
1634         ttwu_do_activate(rq, p, 0);
1635         raw_spin_unlock(&rq->lock);
1636 }
1637 
1638 /**
1639  * try_to_wake_up - wake up a thread
1640  * @p: the thread to be awakened
1641  * @state: the mask of task states that can be woken
1642  * @wake_flags: wake modifier flags (WF_*)
1643  *
1644  * Put it on the run-queue if it's not already there. The "current"
1645  * thread is always on the run-queue (except when the actual
1646  * re-schedule is in progress), and as such you're allowed to do
1647  * the simpler "current->state = TASK_RUNNING" to mark yourself
1648  * runnable without the overhead of this.
1649  *
1650  * Return: %true if @p was woken up, %false if it was already running.
1651  * or @state didn't match @p's state.
1652  */
1653 static int
1654 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1655 {
1656         unsigned long flags;
1657         int cpu, success = 0;
1658 
1659         /*
1660          * If we are going to wake up a thread waiting for CONDITION we
1661          * need to ensure that CONDITION=1 done by the caller can not be
1662          * reordered with p->state check below. This pairs with mb() in
1663          * set_current_state() the waiting thread does.
1664          */
1665         smp_mb__before_spinlock();
1666         raw_spin_lock_irqsave(&p->pi_lock, flags);
1667         if (!(p->state & state))
1668                 goto out;
1669 
1670         success = 1; /* we're going to change ->state */
1671         cpu = task_cpu(p);
1672 
1673         if (p->on_rq && ttwu_remote(p, wake_flags))
1674                 goto stat;
1675 
1676 #ifdef CONFIG_SMP
1677         /*
1678          * If the owning (remote) cpu is still in the middle of schedule() with
1679          * this task as prev, wait until its done referencing the task.
1680          */
1681         while (p->on_cpu)
1682                 cpu_relax();
1683         /*
1684          * Pairs with the smp_wmb() in finish_lock_switch().
1685          */
1686         smp_rmb();
1687 
1688         p->sched_contributes_to_load = !!task_contributes_to_load(p);
1689         p->state = TASK_WAKING;
1690 
1691         if (p->sched_class->task_waking)
1692                 p->sched_class->task_waking(p);
1693 
1694         cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1695         if (task_cpu(p) != cpu) {
1696                 wake_flags |= WF_MIGRATED;
1697                 set_task_cpu(p, cpu);
1698         }
1699 #endif /* CONFIG_SMP */
1700 
1701         ttwu_queue(p, cpu);
1702 stat:
1703         ttwu_stat(p, cpu, wake_flags);
1704 out:
1705         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1706 
1707         return success;
1708 }
1709 
1710 /**
1711  * try_to_wake_up_local - try to wake up a local task with rq lock held
1712  * @p: the thread to be awakened
1713  *
1714  * Put @p on the run-queue if it's not already there. The caller must
1715  * ensure that this_rq() is locked, @p is bound to this_rq() and not
1716  * the current task.
1717  */
1718 static void try_to_wake_up_local(struct task_struct *p)
1719 {
1720         struct rq *rq = task_rq(p);
1721 
1722         if (WARN_ON_ONCE(rq != this_rq()) ||
1723             WARN_ON_ONCE(p == current))
1724                 return;
1725 
1726         lockdep_assert_held(&rq->lock);
1727 
1728         if (!raw_spin_trylock(&p->pi_lock)) {
1729                 raw_spin_unlock(&rq->lock);
1730                 raw_spin_lock(&p->pi_lock);
1731                 raw_spin_lock(&rq->lock);
1732         }
1733 
1734         if (!(p->state & TASK_NORMAL))
1735                 goto out;
1736 
1737         if (!p->on_rq)
1738                 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1739 
1740         ttwu_do_wakeup(rq, p, 0);
1741         ttwu_stat(p, smp_processor_id(), 0);
1742 out:
1743         raw_spin_unlock(&p->pi_lock);
1744 }
1745 
1746 /**
1747  * wake_up_process - Wake up a specific process
1748  * @p: The process to be woken up.
1749  *
1750  * Attempt to wake up the nominated process and move it to the set of runnable
1751  * processes.
1752  *
1753  * Return: 1 if the process was woken up, 0 if it was already running.
1754  *
1755  * It may be assumed that this function implies a write memory barrier before
1756  * changing the task state if and only if any tasks are woken up.
1757  */
1758 int wake_up_process(struct task_struct *p)
1759 {
1760         WARN_ON(task_is_stopped_or_traced(p));
1761         return try_to_wake_up(p, TASK_NORMAL, 0);
1762 }
1763 EXPORT_SYMBOL(wake_up_process);
1764 
1765 int wake_up_state(struct task_struct *p, unsigned int state)
1766 {
1767         return try_to_wake_up(p, state, 0);
1768 }
1769 
1770 /*
1771  * Perform scheduler related setup for a newly forked process p.
1772  * p is forked by current.
1773  *
1774  * __sched_fork() is basic setup used by init_idle() too:
1775  */
1776 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1777 {
1778         p->on_rq                        = 0;
1779 
1780         p->se.on_rq                     = 0;
1781         p->se.exec_start                = 0;
1782         p->se.sum_exec_runtime          = 0;
1783         p->se.prev_sum_exec_runtime     = 0;
1784         p->se.nr_migrations             = 0;
1785         p->se.vruntime                  = 0;
1786         INIT_LIST_HEAD(&p->se.group_node);
1787 
1788 #ifdef CONFIG_SCHEDSTATS
1789         memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1790 #endif
1791 
1792         RB_CLEAR_NODE(&p->dl.rb_node);
1793         hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1794         p->dl.dl_runtime = p->dl.runtime = 0;
1795         p->dl.dl_deadline = p->dl.deadline = 0;
1796         p->dl.dl_period = 0;
1797         p->dl.flags = 0;
1798 
1799         INIT_LIST_HEAD(&p->rt.run_list);
1800 
1801 #ifdef CONFIG_PREEMPT_NOTIFIERS
1802         INIT_HLIST_HEAD(&p->preempt_notifiers);
1803 #endif
1804 
1805 #ifdef CONFIG_NUMA_BALANCING
1806         if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1807                 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1808                 p->mm->numa_scan_seq = 0;
1809         }
1810 
1811         if (clone_flags & CLONE_VM)
1812                 p->numa_preferred_nid = current->numa_preferred_nid;
1813         else
1814                 p->numa_preferred_nid = -1;
1815 
1816         p->node_stamp = 0ULL;
1817         p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1818         p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1819         p->numa_work.next = &p->numa_work;
1820         p->numa_faults_memory = NULL;
1821         p->numa_faults_buffer_memory = NULL;
1822         p->last_task_numa_placement = 0;
1823         p->last_sum_exec_runtime = 0;
1824 
1825         INIT_LIST_HEAD(&p->numa_entry);
1826         p->numa_group = NULL;
1827 #endif /* CONFIG_NUMA_BALANCING */
1828 }
1829 
1830 #ifdef CONFIG_NUMA_BALANCING
1831 #ifdef CONFIG_SCHED_DEBUG
1832 void set_numabalancing_state(bool enabled)
1833 {
1834         if (enabled)
1835                 sched_feat_set("NUMA");
1836         else
1837                 sched_feat_set("NO_NUMA");
1838 }
1839 #else
1840 __read_mostly bool numabalancing_enabled;
1841 
1842 void set_numabalancing_state(bool enabled)
1843 {
1844         numabalancing_enabled = enabled;
1845 }
1846 #endif /* CONFIG_SCHED_DEBUG */
1847 
1848 #ifdef CONFIG_PROC_SYSCTL
1849 int sysctl_numa_balancing(struct ctl_table *table, int write,
1850                          void __user *buffer, size_t *lenp, loff_t *ppos)
1851 {
1852         struct ctl_table t;
1853         int err;
1854         int state = numabalancing_enabled;
1855 
1856         if (write && !capable(CAP_SYS_ADMIN))
1857                 return -EPERM;
1858 
1859         t = *table;
1860         t.data = &state;
1861         err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1862         if (err < 0)
1863                 return err;
1864         if (write)
1865                 set_numabalancing_state(state);
1866         return err;
1867 }
1868 #endif
1869 #endif
1870 
1871 /*
1872  * fork()/clone()-time setup:
1873  */
1874 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1875 {
1876         unsigned long flags;
1877         int cpu = get_cpu();
1878 
1879         __sched_fork(clone_flags, p);
1880         /*
1881          * We mark the process as running here. This guarantees that
1882          * nobody will actually run it, and a signal or other external
1883          * event cannot wake it up and insert it on the runqueue either.
1884          */
1885         p->state = TASK_RUNNING;
1886 
1887         /*
1888          * Make sure we do not leak PI boosting priority to the child.
1889          */
1890         p->prio = current->normal_prio;
1891 
1892         /*
1893          * Revert to default priority/policy on fork if requested.
1894          */
1895         if (unlikely(p->sched_reset_on_fork)) {
1896                 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1897                         p->policy = SCHED_NORMAL;
1898                         p->static_prio = NICE_TO_PRIO(0);
1899                         p->rt_priority = 0;
1900                 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1901                         p->static_prio = NICE_TO_PRIO(0);
1902 
1903                 p->prio = p->normal_prio = __normal_prio(p);
1904                 set_load_weight(p);
1905 
1906                 /*
1907                  * We don't need the reset flag anymore after the fork. It has
1908                  * fulfilled its duty:
1909                  */
1910                 p->sched_reset_on_fork = 0;
1911         }
1912 
1913         if (dl_prio(p->prio)) {
1914                 put_cpu();
1915                 return -EAGAIN;
1916         } else if (rt_prio(p->prio)) {
1917                 p->sched_class = &rt_sched_class;
1918         } else {
1919                 p->sched_class = &fair_sched_class;
1920         }
1921 
1922         if (p->sched_class->task_fork)
1923                 p->sched_class->task_fork(p);
1924 
1925         /*
1926          * The child is not yet in the pid-hash so no cgroup attach races,
1927          * and the cgroup is pinned to this child due to cgroup_fork()
1928          * is ran before sched_fork().
1929          *
1930          * Silence PROVE_RCU.
1931          */
1932         raw_spin_lock_irqsave(&p->pi_lock, flags);
1933         set_task_cpu(p, cpu);
1934         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1935 
1936 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1937         if (likely(sched_info_on()))
1938                 memset(&p->sched_info, 0, sizeof(p->sched_info));
1939 #endif
1940 #if defined(CONFIG_SMP)
1941         p->on_cpu = 0;
1942 #endif
1943         init_task_preempt_count(p);
1944 #ifdef CONFIG_SMP
1945         plist_node_init(&p->pushable_tasks, MAX_PRIO);
1946         RB_CLEAR_NODE(&p->pushable_dl_tasks);
1947 #endif
1948 
1949         put_cpu();
1950         return 0;
1951 }
1952 
1953 unsigned long to_ratio(u64 period, u64 runtime)
1954 {
1955         if (runtime == RUNTIME_INF)
1956                 return 1ULL << 20;
1957 
1958         /*
1959          * Doing this here saves a lot of checks in all
1960          * the calling paths, and returning zero seems
1961          * safe for them anyway.
1962          */
1963         if (period == 0)
1964                 return 0;
1965 
1966         return div64_u64(runtime << 20, period);
1967 }
1968 
1969 #ifdef CONFIG_SMP
1970 inline struct dl_bw *dl_bw_of(int i)
1971 {
1972         return &cpu_rq(i)->rd->dl_bw;
1973 }
1974 
1975 static inline int dl_bw_cpus(int i)
1976 {
1977         struct root_domain *rd = cpu_rq(i)->rd;
1978         int cpus = 0;
1979 
1980         for_each_cpu_and(i, rd->span, cpu_active_mask)
1981                 cpus++;
1982 
1983         return cpus;
1984 }
1985 #else
1986 inline struct dl_bw *dl_bw_of(int i)
1987 {
1988         return &cpu_rq(i)->dl.dl_bw;
1989 }
1990 
1991 static inline int dl_bw_cpus(int i)
1992 {
1993         return 1;
1994 }
1995 #endif
1996 
1997 static inline
1998 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1999 {
2000         dl_b->total_bw -= tsk_bw;
2001 }
2002 
2003 static inline
2004 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2005 {
2006         dl_b->total_bw += tsk_bw;
2007 }
2008 
2009 static inline
2010 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2011 {
2012         return dl_b->bw != -1 &&
2013                dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2014 }
2015 
2016 /*
2017  * We must be sure that accepting a new task (or allowing changing the
2018  * parameters of an existing one) is consistent with the bandwidth
2019  * constraints. If yes, this function also accordingly updates the currently
2020  * allocated bandwidth to reflect the new situation.
2021  *
2022  * This function is called while holding p's rq->lock.
2023  */
2024 static int dl_overflow(struct task_struct *p, int policy,
2025                        const struct sched_attr *attr)
2026 {
2027 
2028         struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2029         u64 period = attr->sched_period ?: attr->sched_deadline;
2030         u64 runtime = attr->sched_runtime;
2031         u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2032         int cpus, err = -1;
2033 
2034         if (new_bw == p->dl.dl_bw)
2035                 return 0;
2036 
2037         /*
2038          * Either if a task, enters, leave, or stays -deadline but changes
2039          * its parameters, we may need to update accordingly the total
2040          * allocated bandwidth of the container.
2041          */
2042         raw_spin_lock(&dl_b->lock);
2043         cpus = dl_bw_cpus(task_cpu(p));
2044         if (dl_policy(policy) && !task_has_dl_policy(p) &&
2045             !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2046                 __dl_add(dl_b, new_bw);
2047                 err = 0;
2048         } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2049                    !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2050                 __dl_clear(dl_b, p->dl.dl_bw);
2051                 __dl_add(dl_b, new_bw);
2052                 err = 0;
2053         } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2054                 __dl_clear(dl_b, p->dl.dl_bw);
2055                 err = 0;
2056         }
2057         raw_spin_unlock(&dl_b->lock);
2058 
2059         return err;
2060 }
2061 
2062 extern void init_dl_bw(struct dl_bw *dl_b);
2063 
2064 /*
2065  * wake_up_new_task - wake up a newly created task for the first time.
2066  *
2067  * This function will do some initial scheduler statistics housekeeping
2068  * that must be done for every newly created context, then puts the task
2069  * on the runqueue and wakes it.
2070  */
2071 void wake_up_new_task(struct task_struct *p)
2072 {
2073         unsigned long flags;
2074         struct rq *rq;
2075 
2076         raw_spin_lock_irqsave(&p->pi_lock, flags);
2077 #ifdef CONFIG_SMP
2078         /*
2079          * Fork balancing, do it here and not earlier because:
2080          *  - cpus_allowed can change in the fork path
2081          *  - any previously selected cpu might disappear through hotplug
2082          */
2083         set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2084 #endif
2085 
2086         /* Initialize new task's runnable average */
2087         init_task_runnable_average(p);
2088         rq = __task_rq_lock(p);
2089         activate_task(rq, p, 0);
2090         p->on_rq = 1;
2091         trace_sched_wakeup_new(p, true);
2092         check_preempt_curr(rq, p, WF_FORK);
2093 #ifdef CONFIG_SMP
2094         if (p->sched_class->task_woken)
2095                 p->sched_class->task_woken(rq, p);
2096 #endif
2097         task_rq_unlock(rq, p, &flags);
2098 }
2099 
2100 #ifdef CONFIG_PREEMPT_NOTIFIERS
2101 
2102 /**
2103  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2104  * @notifier: notifier struct to register
2105  */
2106 void preempt_notifier_register(struct preempt_notifier *notifier)
2107 {
2108         hlist_add_head(&notifier->link, &current->preempt_notifiers);
2109 }
2110 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2111 
2112 /**
2113  * preempt_notifier_unregister - no longer interested in preemption notifications
2114  * @notifier: notifier struct to unregister
2115  *
2116  * This is safe to call from within a preemption notifier.
2117  */
2118 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2119 {
2120         hlist_del(&notifier->link);
2121 }
2122 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2123 
2124 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2125 {
2126         struct preempt_notifier *notifier;
2127 
2128         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2129                 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2130 }
2131 
2132 static void
2133 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2134                                  struct task_struct *next)
2135 {
2136         struct preempt_notifier *notifier;
2137 
2138         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2139                 notifier->ops->sched_out(notifier, next);
2140 }
2141 
2142 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2143 
2144 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2145 {
2146 }
2147 
2148 static void
2149 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2150                                  struct task_struct *next)
2151 {
2152 }
2153 
2154 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2155 
2156 /**
2157  * prepare_task_switch - prepare to switch tasks
2158  * @rq: the runqueue preparing to switch
2159  * @prev: the current task that is being switched out
2160  * @next: the task we are going to switch to.
2161  *
2162  * This is called with the rq lock held and interrupts off. It must
2163  * be paired with a subsequent finish_task_switch after the context
2164  * switch.
2165  *
2166  * prepare_task_switch sets up locking and calls architecture specific
2167  * hooks.
2168  */
2169 static inline void
2170 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2171                     struct task_struct *next)
2172 {
2173         trace_sched_switch(prev, next);
2174         sched_info_switch(rq, prev, next);
2175         perf_event_task_sched_out(prev, next);
2176         fire_sched_out_preempt_notifiers(prev, next);
2177         prepare_lock_switch(rq, next);
2178         prepare_arch_switch(next);
2179 }
2180 
2181 /**
2182  * finish_task_switch - clean up after a task-switch
2183  * @rq: runqueue associated with task-switch
2184  * @prev: the thread we just switched away from.
2185  *
2186  * finish_task_switch must be called after the context switch, paired
2187  * with a prepare_task_switch call before the context switch.
2188  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2189  * and do any other architecture-specific cleanup actions.
2190  *
2191  * Note that we may have delayed dropping an mm in context_switch(). If
2192  * so, we finish that here outside of the runqueue lock. (Doing it
2193  * with the lock held can cause deadlocks; see schedule() for
2194  * details.)
2195  */
2196 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2197         __releases(rq->lock)
2198 {
2199         struct mm_struct *mm = rq->prev_mm;
2200         long prev_state;
2201 
2202         rq->prev_mm = NULL;
2203 
2204         /*
2205          * A task struct has one reference for the use as "current".
2206          * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2207          * schedule one last time. The schedule call will never return, and
2208          * the scheduled task must drop that reference.
2209          * The test for TASK_DEAD must occur while the runqueue locks are
2210          * still held, otherwise prev could be scheduled on another cpu, die
2211          * there before we look at prev->state, and then the reference would
2212          * be dropped twice.
2213          *              Manfred Spraul <manfred@colorfullife.com>
2214          */
2215         prev_state = prev->state;
2216         vtime_task_switch(prev);
2217         finish_arch_switch(prev);
2218         perf_event_task_sched_in(prev, current);
2219         finish_lock_switch(rq, prev);
2220         finish_arch_post_lock_switch();
2221 
2222         fire_sched_in_preempt_notifiers(current);
2223         if (mm)
2224                 mmdrop(mm);
2225         if (unlikely(prev_state == TASK_DEAD)) {
2226                 if (prev->sched_class->task_dead)
2227                         prev->sched_class->task_dead(prev);
2228 
2229                 /*
2230                  * Remove function-return probe instances associated with this
2231                  * task and put them back on the free list.
2232                  */
2233                 kprobe_flush_task(prev);
2234                 put_task_struct(prev);
2235         }
2236 
2237         tick_nohz_task_switch(current);
2238 }
2239 
2240 #ifdef CONFIG_SMP
2241 
2242 /* rq->lock is NOT held, but preemption is disabled */
2243 static inline void post_schedule(struct rq *rq)
2244 {
2245         if (rq->post_schedule) {
2246                 unsigned long flags;
2247 
2248                 raw_spin_lock_irqsave(&rq->lock, flags);
2249                 if (rq->curr->sched_class->post_schedule)
2250                         rq->curr->sched_class->post_schedule(rq);
2251                 raw_spin_unlock_irqrestore(&rq->lock, flags);
2252 
2253                 rq->post_schedule = 0;
2254         }
2255 }
2256 
2257 #else
2258 
2259 static inline void post_schedule(struct rq *rq)
2260 {
2261 }
2262 
2263 #endif
2264 
2265 /**
2266  * schedule_tail - first thing a freshly forked thread must call.
2267  * @prev: the thread we just switched away from.
2268  */
2269 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2270         __releases(rq->lock)
2271 {
2272         struct rq *rq = this_rq();
2273 
2274         finish_task_switch(rq, prev);
2275 
2276         /*
2277          * FIXME: do we need to worry about rq being invalidated by the
2278          * task_switch?
2279          */
2280         post_schedule(rq);
2281 
2282 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2283         /* In this case, finish_task_switch does not reenable preemption */
2284         preempt_enable();
2285 #endif
2286         if (current->set_child_tid)
2287                 put_user(task_pid_vnr(current), current->set_child_tid);
2288 }
2289 
2290 /*
2291  * context_switch - switch to the new MM and the new
2292  * thread's register state.
2293  */
2294 static inline void
2295 context_switch(struct rq *rq, struct task_struct *prev,
2296                struct task_struct *next)
2297 {
2298         struct mm_struct *mm, *oldmm;
2299 
2300         prepare_task_switch(rq, prev, next);
2301 
2302         mm = next->mm;
2303         oldmm = prev->active_mm;
2304         /*
2305          * For paravirt, this is coupled with an exit in switch_to to
2306          * combine the page table reload and the switch backend into
2307          * one hypercall.
2308          */
2309         arch_start_context_switch(prev);
2310 
2311         if (!mm) {
2312                 next->active_mm = oldmm;
2313                 atomic_inc(&oldmm->mm_count);
2314                 enter_lazy_tlb(oldmm, next);
2315         } else
2316                 switch_mm(oldmm, mm, next);
2317 
2318         if (!prev->mm) {
2319                 prev->active_mm = NULL;
2320                 rq->prev_mm = oldmm;
2321         }
2322         /*
2323          * Since the runqueue lock will be released by the next
2324          * task (which is an invalid locking op but in the case
2325          * of the scheduler it's an obvious special-case), so we
2326          * do an early lockdep release here:
2327          */
2328 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2329         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2330 #endif
2331 
2332         context_tracking_task_switch(prev, next);
2333         /* Here we just switch the register state and the stack. */
2334         switch_to(prev, next, prev);
2335 
2336         barrier();
2337         /*
2338          * this_rq must be evaluated again because prev may have moved
2339          * CPUs since it called schedule(), thus the 'rq' on its stack
2340          * frame will be invalid.
2341          */
2342         finish_task_switch(this_rq(), prev);
2343 }
2344 
2345 /*
2346  * nr_running and nr_context_switches:
2347  *
2348  * externally visible scheduler statistics: current number of runnable
2349  * threads, total number of context switches performed since bootup.
2350  */
2351 unsigned long nr_running(void)
2352 {
2353         unsigned long i, sum = 0;
2354 
2355         for_each_online_cpu(i)
2356                 sum += cpu_rq(i)->nr_running;
2357 
2358         return sum;
2359 }
2360 
2361 unsigned long long nr_context_switches(void)
2362 {
2363         int i;
2364         unsigned long long sum = 0;
2365 
2366         for_each_possible_cpu(i)
2367                 sum += cpu_rq(i)->nr_switches;
2368 
2369         return sum;
2370 }
2371 
2372 unsigned long nr_iowait(void)
2373 {
2374         unsigned long i, sum = 0;
2375 
2376         for_each_possible_cpu(i)
2377                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2378 
2379         return sum;
2380 }
2381 
2382 unsigned long nr_iowait_cpu(int cpu)
2383 {
2384         struct rq *this = cpu_rq(cpu);
2385         return atomic_read(&this->nr_iowait);
2386 }
2387 
2388 #ifdef CONFIG_SMP
2389 
2390 /*
2391  * sched_exec - execve() is a valuable balancing opportunity, because at
2392  * this point the task has the smallest effective memory and cache footprint.
2393  */
2394 void sched_exec(void)
2395 {
2396         struct task_struct *p = current;
2397         unsigned long flags;
2398         int dest_cpu;
2399 
2400         raw_spin_lock_irqsave(&p->pi_lock, flags);
2401         dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2402         if (dest_cpu == smp_processor_id())
2403                 goto unlock;
2404 
2405         if (likely(cpu_active(dest_cpu))) {
2406                 struct migration_arg arg = { p, dest_cpu };
2407 
2408                 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2409                 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2410                 return;
2411         }
2412 unlock:
2413         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2414 }
2415 
2416 #endif
2417 
2418 DEFINE_PER_CPU(struct kernel_stat, kstat);
2419 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2420 
2421 EXPORT_PER_CPU_SYMBOL(kstat);
2422 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2423 
2424 /*
2425  * Return any ns on the sched_clock that have not yet been accounted in
2426  * @p in case that task is currently running.
2427  *
2428  * Called with task_rq_lock() held on @rq.
2429  */
2430 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2431 {
2432         u64 ns = 0;
2433 
2434         if (task_current(rq, p)) {
2435                 update_rq_clock(rq);
2436                 ns = rq_clock_task(rq) - p->se.exec_start;
2437                 if ((s64)ns < 0)
2438                         ns = 0;
2439         }
2440 
2441         return ns;
2442 }
2443 
2444 unsigned long long task_delta_exec(struct task_struct *p)
2445 {
2446         unsigned long flags;
2447         struct rq *rq;
2448         u64 ns = 0;
2449 
2450         rq = task_rq_lock(p, &flags);
2451         ns = do_task_delta_exec(p, rq);
2452         task_rq_unlock(rq, p, &flags);
2453 
2454         return ns;
2455 }
2456 
2457 /*
2458  * Return accounted runtime for the task.
2459  * In case the task is currently running, return the runtime plus current's
2460  * pending runtime that have not been accounted yet.
2461  */
2462 unsigned long long task_sched_runtime(struct task_struct *p)
2463 {
2464         unsigned long flags;
2465         struct rq *rq;
2466         u64 ns = 0;
2467 
2468 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2469         /*
2470          * 64-bit doesn't need locks to atomically read a 64bit value.
2471          * So we have a optimization chance when the task's delta_exec is 0.
2472          * Reading ->on_cpu is racy, but this is ok.
2473          *
2474          * If we race with it leaving cpu, we'll take a lock. So we're correct.
2475          * If we race with it entering cpu, unaccounted time is 0. This is
2476          * indistinguishable from the read occurring a few cycles earlier.
2477          */
2478         if (!p->on_cpu)
2479                 return p->se.sum_exec_runtime;
2480 #endif
2481 
2482         rq = task_rq_lock(p, &flags);
2483         ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2484         task_rq_unlock(rq, p, &flags);
2485 
2486         return ns;
2487 }
2488 
2489 /*
2490  * This function gets called by the timer code, with HZ frequency.
2491  * We call it with interrupts disabled.
2492  */
2493 void scheduler_tick(void)
2494 {
2495         int cpu = smp_processor_id();
2496         struct rq *rq = cpu_rq(cpu);
2497         struct task_struct *curr = rq->curr;
2498 
2499         sched_clock_tick();
2500 
2501         raw_spin_lock(&rq->lock);
2502         update_rq_clock(rq);
2503         curr->sched_class->task_tick(rq, curr, 0);
2504         update_cpu_load_active(rq);
2505         raw_spin_unlock(&rq->lock);
2506 
2507         perf_event_task_tick();
2508 
2509 #ifdef CONFIG_SMP
2510         rq->idle_balance = idle_cpu(cpu);
2511         trigger_load_balance(rq);
2512 #endif
2513         rq_last_tick_reset(rq);
2514 }
2515 
2516 #ifdef CONFIG_NO_HZ_FULL
2517 /**
2518  * scheduler_tick_max_deferment
2519  *
2520  * Keep at least one tick per second when a single
2521  * active task is running because the scheduler doesn't
2522  * yet completely support full dynticks environment.
2523  *
2524  * This makes sure that uptime, CFS vruntime, load
2525  * balancing, etc... continue to move forward, even
2526  * with a very low granularity.
2527  *
2528  * Return: Maximum deferment in nanoseconds.
2529  */
2530 u64 scheduler_tick_max_deferment(void)
2531 {
2532         struct rq *rq = this_rq();
2533         unsigned long next, now = ACCESS_ONCE(jiffies);
2534 
2535         next = rq->last_sched_tick + HZ;
2536 
2537         if (time_before_eq(next, now))
2538                 return 0;
2539 
2540         return jiffies_to_nsecs(next - now);
2541 }
2542 #endif
2543 
2544 notrace unsigned long get_parent_ip(unsigned long addr)
2545 {
2546         if (in_lock_functions(addr)) {
2547                 addr = CALLER_ADDR2;
2548                 if (in_lock_functions(addr))
2549                         addr = CALLER_ADDR3;
2550         }
2551         return addr;
2552 }
2553 
2554 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2555                                 defined(CONFIG_PREEMPT_TRACER))
2556 
2557 void preempt_count_add(int val)
2558 {
2559 #ifdef CONFIG_DEBUG_PREEMPT
2560         /*
2561          * Underflow?
2562          */
2563         if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2564                 return;
2565 #endif
2566         __preempt_count_add(val);
2567 #ifdef CONFIG_DEBUG_PREEMPT
2568         /*
2569          * Spinlock count overflowing soon?
2570          */
2571         DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2572                                 PREEMPT_MASK - 10);
2573 #endif
2574         if (preempt_count() == val) {
2575                 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2576 #ifdef CONFIG_DEBUG_PREEMPT
2577                 current->preempt_disable_ip = ip;
2578 #endif
2579                 trace_preempt_off(CALLER_ADDR0, ip);
2580         }
2581 }
2582 EXPORT_SYMBOL(preempt_count_add);
2583 NOKPROBE_SYMBOL(preempt_count_add);
2584 
2585 void preempt_count_sub(int val)
2586 {
2587 #ifdef CONFIG_DEBUG_PREEMPT
2588         /*
2589          * Underflow?
2590          */
2591         if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2592                 return;
2593         /*
2594          * Is the spinlock portion underflowing?
2595          */
2596         if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2597                         !(preempt_count() & PREEMPT_MASK)))
2598                 return;
2599 #endif
2600 
2601         if (preempt_count() == val)
2602                 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2603         __preempt_count_sub(val);
2604 }
2605 EXPORT_SYMBOL(preempt_count_sub);
2606 NOKPROBE_SYMBOL(preempt_count_sub);
2607 
2608 #endif
2609 
2610 /*
2611  * Print scheduling while atomic bug:
2612  */
2613 static noinline void __schedule_bug(struct task_struct *prev)
2614 {
2615         if (oops_in_progress)
2616                 return;
2617 
2618         printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2619                 prev->comm, prev->pid, preempt_count());
2620 
2621         debug_show_held_locks(prev);
2622         print_modules();
2623         if (irqs_disabled())
2624                 print_irqtrace_events(prev);
2625 #ifdef CONFIG_DEBUG_PREEMPT
2626         if (in_atomic_preempt_off()) {
2627                 pr_err("Preemption disabled at:");
2628                 print_ip_sym(current->preempt_disable_ip);
2629                 pr_cont("\n");
2630         }
2631 #endif
2632         dump_stack();
2633         add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2634 }
2635 
2636 /*
2637  * Various schedule()-time debugging checks and statistics:
2638  */
2639 static inline void schedule_debug(struct task_struct *prev)
2640 {
2641         /*
2642          * Test if we are atomic. Since do_exit() needs to call into
2643          * schedule() atomically, we ignore that path. Otherwise whine
2644          * if we are scheduling when we should not.
2645          */
2646         if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2647                 __schedule_bug(prev);
2648         rcu_sleep_check();
2649 
2650         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2651 
2652         schedstat_inc(this_rq(), sched_count);
2653 }
2654 
2655 /*
2656  * Pick up the highest-prio task:
2657  */
2658 static inline struct task_struct *
2659 pick_next_task(struct rq *rq, struct task_struct *prev)
2660 {
2661         const struct sched_class *class = &fair_sched_class;
2662         struct task_struct *p;
2663 
2664         /*
2665          * Optimization: we know that if all tasks are in
2666          * the fair class we can call that function directly:
2667          */
2668         if (likely(prev->sched_class == class &&
2669                    rq->nr_running == rq->cfs.h_nr_running)) {
2670                 p = fair_sched_class.pick_next_task(rq, prev);
2671                 if (unlikely(p == RETRY_TASK))
2672                         goto again;
2673 
2674                 /* assumes fair_sched_class->next == idle_sched_class */
2675                 if (unlikely(!p))
2676                         p = idle_sched_class.pick_next_task(rq, prev);
2677 
2678                 return p;
2679         }
2680 
2681 again:
2682         for_each_class(class) {
2683                 p = class->pick_next_task(rq, prev);
2684                 if (p) {
2685                         if (unlikely(p == RETRY_TASK))
2686                                 goto again;
2687                         return p;
2688                 }
2689         }
2690 
2691         BUG(); /* the idle class will always have a runnable task */
2692 }
2693 
2694 /*
2695  * __schedule() is the main scheduler function.
2696  *
2697  * The main means of driving the scheduler and thus entering this function are:
2698  *
2699  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2700  *
2701  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2702  *      paths. For example, see arch/x86/entry_64.S.
2703  *
2704  *      To drive preemption between tasks, the scheduler sets the flag in timer
2705  *      interrupt handler scheduler_tick().
2706  *
2707  *   3. Wakeups don't really cause entry into schedule(). They add a
2708  *      task to the run-queue and that's it.
2709  *
2710  *      Now, if the new task added to the run-queue preempts the current
2711  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2712  *      called on the nearest possible occasion:
2713  *
2714  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2715  *
2716  *         - in syscall or exception context, at the next outmost
2717  *           preempt_enable(). (this might be as soon as the wake_up()'s
2718  *           spin_unlock()!)
2719  *
2720  *         - in IRQ context, return from interrupt-handler to
2721  *           preemptible context
2722  *
2723  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2724  *         then at the next:
2725  *
2726  *          - cond_resched() call
2727  *          - explicit schedule() call
2728  *          - return from syscall or exception to user-space
2729  *          - return from interrupt-handler to user-space
2730  */
2731 static void __sched __schedule(void)
2732 {
2733         struct task_struct *prev, *next;
2734         unsigned long *switch_count;
2735         struct rq *rq;
2736         int cpu;
2737 
2738 need_resched:
2739         preempt_disable();
2740         cpu = smp_processor_id();
2741         rq = cpu_rq(cpu);
2742         rcu_note_context_switch(cpu);
2743         prev = rq->curr;
2744 
2745         schedule_debug(prev);
2746 
2747         if (sched_feat(HRTICK))
2748                 hrtick_clear(rq);
2749 
2750         /*
2751          * Make sure that signal_pending_state()->signal_pending() below
2752          * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2753          * done by the caller to avoid the race with signal_wake_up().
2754          */
2755         smp_mb__before_spinlock();
2756         raw_spin_lock_irq(&rq->lock);
2757 
2758         switch_count = &prev->nivcsw;
2759         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2760                 if (unlikely(signal_pending_state(prev->state, prev))) {
2761                         prev->state = TASK_RUNNING;
2762                 } else {
2763                         deactivate_task(rq, prev, DEQUEUE_SLEEP);
2764                         prev->on_rq = 0;
2765 
2766                         /*
2767                          * If a worker went to sleep, notify and ask workqueue
2768                          * whether it wants to wake up a task to maintain
2769                          * concurrency.
2770                          */
2771                         if (prev->flags & PF_WQ_WORKER) {
2772                                 struct task_struct *to_wakeup;
2773 
2774                                 to_wakeup = wq_worker_sleeping(prev, cpu);
2775                                 if (to_wakeup)
2776                                         try_to_wake_up_local(to_wakeup);
2777                         }
2778                 }
2779                 switch_count = &prev->nvcsw;
2780         }
2781 
2782         if (prev->on_rq || rq->skip_clock_update < 0)
2783                 update_rq_clock(rq);
2784 
2785         next = pick_next_task(rq, prev);
2786         clear_tsk_need_resched(prev);
2787         clear_preempt_need_resched();
2788         rq->skip_clock_update = 0;
2789 
2790         if (likely(prev != next)) {
2791                 rq->nr_switches++;
2792                 rq->curr = next;
2793                 ++*switch_count;
2794 
2795                 context_switch(rq, prev, next); /* unlocks the rq */
2796                 /*
2797                  * The context switch have flipped the stack from under us
2798                  * and restored the local variables which were saved when
2799                  * this task called schedule() in the past. prev == current
2800                  * is still correct, but it can be moved to another cpu/rq.
2801                  */
2802                 cpu = smp_processor_id();
2803                 rq = cpu_rq(cpu);
2804         } else
2805                 raw_spin_unlock_irq(&rq->lock);
2806 
2807         post_schedule(rq);
2808 
2809         sched_preempt_enable_no_resched();
2810         if (need_resched())
2811                 goto need_resched;
2812 }
2813 
2814 static inline void sched_submit_work(struct task_struct *tsk)
2815 {
2816         if (!tsk->state || tsk_is_pi_blocked(tsk))
2817                 return;
2818         /*
2819          * If we are going to sleep and we have plugged IO queued,
2820          * make sure to submit it to avoid deadlocks.
2821          */
2822         if (blk_needs_flush_plug(tsk))
2823                 blk_schedule_flush_plug(tsk);
2824 }
2825 
2826 asmlinkage __visible void __sched schedule(void)
2827 {
2828         struct task_struct *tsk = current;
2829 
2830         sched_submit_work(tsk);
2831         __schedule();
2832 }
2833 EXPORT_SYMBOL(schedule);
2834 
2835 #ifdef CONFIG_CONTEXT_TRACKING
2836 asmlinkage __visible void __sched schedule_user(void)
2837 {
2838         /*
2839          * If we come here after a random call to set_need_resched(),
2840          * or we have been woken up remotely but the IPI has not yet arrived,
2841          * we haven't yet exited the RCU idle mode. Do it here manually until
2842          * we find a better solution.
2843          */
2844         user_exit();
2845         schedule();
2846         user_enter();
2847 }
2848 #endif
2849 
2850 /**
2851  * schedule_preempt_disabled - called with preemption disabled
2852  *
2853  * Returns with preemption disabled. Note: preempt_count must be 1
2854  */
2855 void __sched schedule_preempt_disabled(void)
2856 {
2857         sched_preempt_enable_no_resched();
2858         schedule();
2859         preempt_disable();
2860 }
2861 
2862 #ifdef CONFIG_PREEMPT
2863 /*
2864  * this is the entry point to schedule() from in-kernel preemption
2865  * off of preempt_enable. Kernel preemptions off return from interrupt
2866  * occur there and call schedule directly.
2867  */
2868 asmlinkage __visible void __sched notrace preempt_schedule(void)
2869 {
2870         /*
2871          * If there is a non-zero preempt_count or interrupts are disabled,
2872          * we do not want to preempt the current task. Just return..
2873          */
2874         if (likely(!preemptible()))
2875                 return;
2876 
2877         do {
2878                 __preempt_count_add(PREEMPT_ACTIVE);
2879                 __schedule();
2880                 __preempt_count_sub(PREEMPT_ACTIVE);
2881 
2882                 /*
2883                  * Check again in case we missed a preemption opportunity
2884                  * between schedule and now.
2885                  */
2886                 barrier();
2887         } while (need_resched());
2888 }
2889 NOKPROBE_SYMBOL(preempt_schedule);
2890 EXPORT_SYMBOL(preempt_schedule);
2891 #endif /* CONFIG_PREEMPT */
2892 
2893 /*
2894  * this is the entry point to schedule() from kernel preemption
2895  * off of irq context.
2896  * Note, that this is called and return with irqs disabled. This will
2897  * protect us against recursive calling from irq.
2898  */
2899 asmlinkage __visible void __sched preempt_schedule_irq(void)
2900 {
2901         enum ctx_state prev_state;
2902 
2903         /* Catch callers which need to be fixed */
2904         BUG_ON(preempt_count() || !irqs_disabled());
2905 
2906         prev_state = exception_enter();
2907 
2908         do {
2909                 __preempt_count_add(PREEMPT_ACTIVE);
2910                 local_irq_enable();
2911                 __schedule();
2912                 local_irq_disable();
2913                 __preempt_count_sub(PREEMPT_ACTIVE);
2914 
2915                 /*
2916                  * Check again in case we missed a preemption opportunity
2917                  * between schedule and now.
2918                  */
2919                 barrier();
2920         } while (need_resched());
2921 
2922         exception_exit(prev_state);
2923 }
2924 
2925 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2926                           void *key)
2927 {
2928         return try_to_wake_up(curr->private, mode, wake_flags);
2929 }
2930 EXPORT_SYMBOL(default_wake_function);
2931 
2932 #ifdef CONFIG_RT_MUTEXES
2933 
2934 /*
2935  * rt_mutex_setprio - set the current priority of a task
2936  * @p: task
2937  * @prio: prio value (kernel-internal form)
2938  *
2939  * This function changes the 'effective' priority of a task. It does
2940  * not touch ->normal_prio like __setscheduler().
2941  *
2942  * Used by the rt_mutex code to implement priority inheritance
2943  * logic. Call site only calls if the priority of the task changed.
2944  */
2945 void rt_mutex_setprio(struct task_struct *p, int prio)
2946 {
2947         int oldprio, on_rq, running, enqueue_flag = 0;
2948         struct rq *rq;
2949         const struct sched_class *prev_class;
2950 
2951         BUG_ON(prio > MAX_PRIO);
2952 
2953         rq = __task_rq_lock(p);
2954 
2955         /*
2956          * Idle task boosting is a nono in general. There is one
2957          * exception, when PREEMPT_RT and NOHZ is active:
2958          *
2959          * The idle task calls get_next_timer_interrupt() and holds
2960          * the timer wheel base->lock on the CPU and another CPU wants
2961          * to access the timer (probably to cancel it). We can safely
2962          * ignore the boosting request, as the idle CPU runs this code
2963          * with interrupts disabled and will complete the lock
2964          * protected section without being interrupted. So there is no
2965          * real need to boost.
2966          */
2967         if (unlikely(p == rq->idle)) {
2968                 WARN_ON(p != rq->curr);
2969                 WARN_ON(p->pi_blocked_on);
2970                 goto out_unlock;
2971         }
2972 
2973         trace_sched_pi_setprio(p, prio);
2974         p->pi_top_task = rt_mutex_get_top_task(p);
2975         oldprio = p->prio;
2976         prev_class = p->sched_class;
2977         on_rq = p->on_rq;
2978         running = task_current(rq, p);
2979         if (on_rq)
2980                 dequeue_task(rq, p, 0);
2981         if (running)
2982                 p->sched_class->put_prev_task(rq, p);
2983 
2984         /*
2985          * Boosting condition are:
2986          * 1. -rt task is running and holds mutex A
2987          *      --> -dl task blocks on mutex A
2988          *
2989          * 2. -dl task is running and holds mutex A
2990          *      --> -dl task blocks on mutex A and could preempt the
2991          *          running task
2992          */
2993         if (dl_prio(prio)) {
2994                 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2995                         dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2996                         p->dl.dl_boosted = 1;
2997                         p->dl.dl_throttled = 0;
2998                         enqueue_flag = ENQUEUE_REPLENISH;
2999                 } else
3000                         p->dl.dl_boosted = 0;
3001                 p->sched_class = &dl_sched_class;
3002         } else if (rt_prio(prio)) {
3003                 if (dl_prio(oldprio))
3004                         p->dl.dl_boosted = 0;
3005                 if (oldprio < prio)
3006                         enqueue_flag = ENQUEUE_HEAD;
3007                 p->sched_class = &rt_sched_class;
3008         } else {
3009                 if (dl_prio(oldprio))
3010                         p->dl.dl_boosted = 0;
3011                 p->sched_class = &fair_sched_class;
3012         }
3013 
3014         p->prio = prio;
3015 
3016         if (running)
3017                 p->sched_class->set_curr_task(rq);
3018         if (on_rq)
3019                 enqueue_task(rq, p, enqueue_flag);
3020 
3021         check_class_changed(rq, p, prev_class, oldprio);
3022 out_unlock:
3023         __task_rq_unlock(rq);
3024 }
3025 #endif
3026 
3027 void set_user_nice(struct task_struct *p, long nice)
3028 {
3029         int old_prio, delta, on_rq;
3030         unsigned long flags;
3031         struct rq *rq;
3032 
3033         if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3034                 return;
3035         /*
3036          * We have to be careful, if called from sys_setpriority(),
3037          * the task might be in the middle of scheduling on another CPU.
3038          */
3039         rq = task_rq_lock(p, &flags);
3040         /*
3041          * The RT priorities are set via sched_setscheduler(), but we still
3042          * allow the 'normal' nice value to be set - but as expected
3043          * it wont have any effect on scheduling until the task is
3044          * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3045          */
3046         if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3047                 p->static_prio = NICE_TO_PRIO(nice);
3048                 goto out_unlock;
3049         }
3050         on_rq = p->on_rq;
3051         if (on_rq)
3052                 dequeue_task(rq, p, 0);
3053 
3054         p->static_prio = NICE_TO_PRIO(nice);
3055         set_load_weight(p);
3056         old_prio = p->prio;
3057         p->prio = effective_prio(p);
3058         delta = p->prio - old_prio;
3059 
3060         if (on_rq) {
3061                 enqueue_task(rq, p, 0);
3062                 /*
3063                  * If the task increased its priority or is running and
3064                  * lowered its priority, then reschedule its CPU:
3065                  */
3066                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3067                         resched_task(rq->curr);
3068         }
3069 out_unlock:
3070         task_rq_unlock(rq, p, &flags);
3071 }
3072 EXPORT_SYMBOL(set_user_nice);
3073 
3074 /*
3075  * can_nice - check if a task can reduce its nice value
3076  * @p: task
3077  * @nice: nice value
3078  */
3079 int can_nice(const struct task_struct *p, const int nice)
3080 {
3081         /* convert nice value [19,-20] to rlimit style value [1,40] */
3082         int nice_rlim = nice_to_rlimit(nice);
3083 
3084         return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3085                 capable(CAP_SYS_NICE));
3086 }
3087 
3088 #ifdef __ARCH_WANT_SYS_NICE
3089 
3090 /*
3091  * sys_nice - change the priority of the current process.
3092  * @increment: priority increment
3093  *
3094  * sys_setpriority is a more generic, but much slower function that
3095  * does similar things.
3096  */
3097 SYSCALL_DEFINE1(nice, int, increment)
3098 {
3099         long nice, retval;
3100 
3101         /*
3102          * Setpriority might change our priority at the same moment.
3103          * We don't have to worry. Conceptually one call occurs first
3104          * and we have a single winner.
3105          */
3106         increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3107         nice = task_nice(current) + increment;
3108 
3109         nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3110         if (increment < 0 && !can_nice(current, nice))
3111                 return -EPERM;
3112 
3113         retval = security_task_setnice(current, nice);
3114         if (retval)
3115                 return retval;
3116 
3117         set_user_nice(current, nice);
3118         return 0;
3119 }
3120 
3121 #endif
3122 
3123 /**
3124  * task_prio - return the priority value of a given task.
3125  * @p: the task in question.
3126  *
3127  * Return: The priority value as seen by users in /proc.
3128  * RT tasks are offset by -200. Normal tasks are centered
3129  * around 0, value goes from -16 to +15.
3130  */
3131 int task_prio(const struct task_struct *p)
3132 {
3133         return p->prio - MAX_RT_PRIO;
3134 }
3135 
3136 /**
3137  * idle_cpu - is a given cpu idle currently?
3138  * @cpu: the processor in question.
3139  *
3140  * Return: 1 if the CPU is currently idle. 0 otherwise.
3141  */
3142 int idle_cpu(int cpu)
3143 {
3144         struct rq *rq = cpu_rq(cpu);
3145 
3146         if (rq->curr != rq->idle)
3147                 return 0;
3148 
3149         if (rq->nr_running)
3150                 return 0;
3151 
3152 #ifdef CONFIG_SMP
3153         if (!llist_empty(&rq->wake_list))
3154                 return 0;
3155 #endif
3156 
3157         return 1;
3158 }
3159 
3160 /**
3161  * idle_task - return the idle task for a given cpu.
3162  * @cpu: the processor in question.
3163  *
3164  * Return: The idle task for the cpu @cpu.
3165  */
3166 struct task_struct *idle_task(int cpu)
3167 {
3168         return cpu_rq(cpu)->idle;
3169 }
3170 
3171 /**
3172  * find_process_by_pid - find a process with a matching PID value.
3173  * @pid: the pid in question.
3174  *
3175  * The task of @pid, if found. %NULL otherwise.
3176  */
3177 static struct task_struct *find_process_by_pid(pid_t pid)
3178 {
3179         return pid ? find_task_by_vpid(pid) : current;
3180 }
3181 
3182 /*
3183  * This function initializes the sched_dl_entity of a newly becoming
3184  * SCHED_DEADLINE task.
3185  *
3186  * Only the static values are considered here, the actual runtime and the
3187  * absolute deadline will be properly calculated when the task is enqueued
3188  * for the first time with its new policy.
3189  */
3190 static void
3191 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3192 {
3193         struct sched_dl_entity *dl_se = &p->dl;
3194 
3195         init_dl_task_timer(dl_se);
3196         dl_se->dl_runtime = attr->sched_runtime;
3197         dl_se->dl_deadline = attr->sched_deadline;
3198         dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3199         dl_se->flags = attr->sched_flags;
3200         dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3201         dl_se->dl_throttled = 0;
3202         dl_se->dl_new = 1;
3203         dl_se->dl_yielded = 0;
3204 }
3205 
3206 static void __setscheduler_params(struct task_struct *p,
3207                 const struct sched_attr *attr)
3208 {
3209         int policy = attr->sched_policy;
3210 
3211         if (policy == -1) /* setparam */
3212                 policy = p->policy;
3213 
3214         p->policy = policy;
3215 
3216         if (dl_policy(policy))
3217                 __setparam_dl(p, attr);
3218         else if (fair_policy(policy))
3219                 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3220 
3221         /*
3222          * __sched_setscheduler() ensures attr->sched_priority == 0 when
3223          * !rt_policy. Always setting this ensures that things like
3224          * getparam()/getattr() don't report silly values for !rt tasks.
3225          */
3226         p->rt_priority = attr->sched_priority;
3227         p->normal_prio = normal_prio(p);
3228         set_load_weight(p);
3229 }
3230 
3231 /* Actually do priority change: must hold pi & rq lock. */
3232 static void __setscheduler(struct rq *rq, struct task_struct *p,
3233                            const struct sched_attr *attr)
3234 {
3235         __setscheduler_params(p, attr);
3236 
3237         /*
3238          * If we get here, there was no pi waiters boosting the
3239          * task. It is safe to use the normal prio.
3240          */
3241         p->prio = normal_prio(p);
3242 
3243         if (dl_prio(p->prio))
3244                 p->sched_class = &dl_sched_class;
3245         else if (rt_prio(p->prio))
3246                 p->sched_class = &rt_sched_class;
3247         else
3248                 p->sched_class = &fair_sched_class;
3249 }
3250 
3251 static void
3252 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3253 {
3254         struct sched_dl_entity *dl_se = &p->dl;
3255 
3256         attr->sched_priority = p->rt_priority;
3257         attr->sched_runtime = dl_se->dl_runtime;
3258         attr->sched_deadline = dl_se->dl_deadline;
3259         attr->sched_period = dl_se->dl_period;
3260         attr->sched_flags = dl_se->flags;
3261 }
3262 
3263 /*
3264  * This function validates the new parameters of a -deadline task.
3265  * We ask for the deadline not being zero, and greater or equal
3266  * than the runtime, as well as the period of being zero or
3267  * greater than deadline. Furthermore, we have to be sure that
3268  * user parameters are above the internal resolution of 1us (we
3269  * check sched_runtime only since it is always the smaller one) and
3270  * below 2^63 ns (we have to check both sched_deadline and
3271  * sched_period, as the latter can be zero).
3272  */
3273 static bool
3274 __checkparam_dl(const struct sched_attr *attr)
3275 {
3276         /* deadline != 0 */
3277         if (attr->sched_deadline == 0)
3278                 return false;
3279 
3280         /*
3281          * Since we truncate DL_SCALE bits, make sure we're at least
3282          * that big.
3283          */
3284         if (attr->sched_runtime < (1ULL << DL_SCALE))
3285                 return false;
3286 
3287         /*
3288          * Since we use the MSB for wrap-around and sign issues, make
3289          * sure it's not set (mind that period can be equal to zero).
3290          */
3291         if (attr->sched_deadline & (1ULL << 63) ||
3292             attr->sched_period & (1ULL << 63))
3293                 return false;
3294 
3295         /* runtime <= deadline <= period (if period != 0) */
3296         if ((attr->sched_period != 0 &&
3297              attr->sched_period < attr->sched_deadline) ||
3298             attr->sched_deadline < attr->sched_runtime)
3299                 return false;
3300 
3301         return true;
3302 }
3303 
3304 /*
3305  * check the target process has a UID that matches the current process's
3306  */
3307 static bool check_same_owner(struct task_struct *p)
3308 {
3309         const struct cred *cred = current_cred(), *pcred;
3310         bool match;
3311 
3312         rcu_read_lock();
3313         pcred = __task_cred(p);
3314         match = (uid_eq(cred->euid, pcred->euid) ||
3315                  uid_eq(cred->euid, pcred->uid));
3316         rcu_read_unlock();
3317         return match;
3318 }
3319 
3320 static int __sched_setscheduler(struct task_struct *p,
3321                                 const struct sched_attr *attr,
3322                                 bool user)
3323 {
3324         int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3325                       MAX_RT_PRIO - 1 - attr->sched_priority;
3326         int retval, oldprio, oldpolicy = -1, on_rq, running;
3327         int policy = attr->sched_policy;
3328         unsigned long flags;
3329         const struct sched_class *prev_class;
3330         struct rq *rq;
3331         int reset_on_fork;
3332 
3333         /* may grab non-irq protected spin_locks */
3334         BUG_ON(in_interrupt());
3335 recheck:
3336         /* double check policy once rq lock held */
3337         if (policy < 0) {
3338                 reset_on_fork = p->sched_reset_on_fork;
3339                 policy = oldpolicy = p->policy;
3340         } else {
3341                 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3342 
3343                 if (policy != SCHED_DEADLINE &&
3344                                 policy != SCHED_FIFO && policy != SCHED_RR &&
3345                                 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3346                                 policy != SCHED_IDLE)
3347                         return -EINVAL;
3348         }
3349 
3350         if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3351                 return -EINVAL;
3352 
3353         /*
3354          * Valid priorities for SCHED_FIFO and SCHED_RR are
3355          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3356          * SCHED_BATCH and SCHED_IDLE is 0.
3357          */
3358         if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3359             (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3360                 return -EINVAL;
3361         if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3362             (rt_policy(policy) != (attr->sched_priority != 0)))
3363                 return -EINVAL;
3364 
3365         /*
3366          * Allow unprivileged RT tasks to decrease priority:
3367          */
3368         if (user && !capable(CAP_SYS_NICE)) {
3369                 if (fair_policy(policy)) {
3370                         if (attr->sched_nice < task_nice(p) &&
3371                             !can_nice(p, attr->sched_nice))
3372                                 return -EPERM;
3373                 }
3374 
3375                 if (rt_policy(policy)) {
3376                         unsigned long rlim_rtprio =
3377                                         task_rlimit(p, RLIMIT_RTPRIO);
3378 
3379                         /* can't set/change the rt policy */
3380                         if (policy != p->policy && !rlim_rtprio)
3381                                 return -EPERM;
3382 
3383                         /* can't increase priority */
3384                         if (attr->sched_priority > p->rt_priority &&
3385                             attr->sched_priority > rlim_rtprio)
3386                                 return -EPERM;
3387                 }
3388 
3389                  /*
3390                   * Can't set/change SCHED_DEADLINE policy at all for now
3391                   * (safest behavior); in the future we would like to allow
3392                   * unprivileged DL tasks to increase their relative deadline
3393                   * or reduce their runtime (both ways reducing utilization)
3394                   */
3395                 if (dl_policy(policy))
3396                         return -EPERM;
3397 
3398                 /*
3399                  * Treat SCHED_IDLE as nice 20. Only allow a switch to
3400                  * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3401                  */
3402                 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3403                         if (!can_nice(p, task_nice(p)))
3404                                 return -EPERM;
3405                 }
3406 
3407                 /* can't change other user's priorities */
3408                 if (!check_same_owner(p))
3409                         return -EPERM;
3410 
3411                 /* Normal users shall not reset the sched_reset_on_fork flag */
3412                 if (p->sched_reset_on_fork && !reset_on_fork)
3413                         return -EPERM;
3414         }
3415 
3416         if (user) {
3417                 retval = security_task_setscheduler(p);
3418                 if (retval)
3419                         return retval;
3420         }
3421 
3422         /*
3423          * make sure no PI-waiters arrive (or leave) while we are
3424          * changing the priority of the task:
3425          *
3426          * To be able to change p->policy safely, the appropriate
3427          * runqueue lock must be held.
3428          */
3429         rq = task_rq_lock(p, &flags);
3430 
3431         /*
3432          * Changing the policy of the stop threads its a very bad idea
3433          */
3434         if (p == rq->stop) {
3435                 task_rq_unlock(rq, p, &flags);
3436                 return -EINVAL;
3437         }
3438 
3439         /*
3440          * If not changing anything there's no need to proceed further,
3441          * but store a possible modification of reset_on_fork.
3442          */
3443         if (unlikely(policy == p->policy)) {
3444                 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3445                         goto change;
3446                 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3447                         goto change;
3448                 if (dl_policy(policy))
3449                         goto change;
3450 
3451                 p->sched_reset_on_fork = reset_on_fork;
3452                 task_rq_unlock(rq, p, &flags);
3453                 return 0;
3454         }
3455 change:
3456 
3457         if (user) {
3458 #ifdef CONFIG_RT_GROUP_SCHED
3459                 /*
3460                  * Do not allow realtime tasks into groups that have no runtime
3461                  * assigned.
3462                  */
3463                 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3464                                 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3465                                 !task_group_is_autogroup(task_group(p))) {
3466                         task_rq_unlock(rq, p, &flags);
3467                         return -EPERM;
3468                 }
3469 #endif
3470 #ifdef CONFIG_SMP
3471                 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3472                         cpumask_t *span = rq->rd->span;
3473 
3474                         /*
3475                          * Don't allow tasks with an affinity mask smaller than
3476                          * the entire root_domain to become SCHED_DEADLINE. We
3477                          * will also fail if there's no bandwidth available.
3478                          */
3479                         if (!cpumask_subset(span, &p->cpus_allowed) ||
3480                             rq->rd->dl_bw.bw == 0) {
3481                                 task_rq_unlock(rq, p, &flags);
3482                                 return -EPERM;
3483                         }
3484                 }
3485 #endif
3486         }
3487 
3488         /* recheck policy now with rq lock held */
3489         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3490                 policy = oldpolicy = -1;
3491                 task_rq_unlock(rq, p, &flags);
3492                 goto recheck;
3493         }
3494 
3495         /*
3496          * If setscheduling to SCHED_DEADLINE (or changing the parameters
3497          * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3498          * is available.
3499          */
3500         if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3501                 task_rq_unlock(rq, p, &flags);
3502                 return -EBUSY;
3503         }
3504 
3505         p->sched_reset_on_fork = reset_on_fork;
3506         oldprio = p->prio;
3507 
3508         /*
3509          * Special case for priority boosted tasks.
3510          *
3511          * If the new priority is lower or equal (user space view)
3512          * than the current (boosted) priority, we just store the new
3513          * normal parameters and do not touch the scheduler class and
3514          * the runqueue. This will be done when the task deboost
3515          * itself.
3516          */
3517         if (rt_mutex_check_prio(p, newprio)) {
3518                 __setscheduler_params(p, attr);
3519                 task_rq_unlock(rq, p, &flags);
3520                 return 0;
3521         }
3522 
3523         on_rq = p->on_rq;
3524         running = task_current(rq, p);
3525         if (on_rq)
3526                 dequeue_task(rq, p, 0);
3527         if (running)
3528                 p->sched_class->put_prev_task(rq, p);
3529 
3530         prev_class = p->sched_class;
3531         __setscheduler(rq, p, attr);
3532 
3533         if (running)
3534                 p->sched_class->set_curr_task(rq);
3535         if (on_rq) {
3536                 /*
3537                  * We enqueue to tail when the priority of a task is
3538                  * increased (user space view).
3539                  */
3540                 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3541         }
3542 
3543         check_class_changed(rq, p, prev_class, oldprio);
3544         task_rq_unlock(rq, p, &flags);
3545 
3546         rt_mutex_adjust_pi(p);
3547 
3548         return 0;
3549 }
3550 
3551 static int _sched_setscheduler(struct task_struct *p, int policy,
3552                                const struct sched_param *param, bool check)
3553 {
3554         struct sched_attr attr = {
3555                 .sched_policy   = policy,
3556                 .sched_priority = param->sched_priority,
3557                 .sched_nice     = PRIO_TO_NICE(p->static_prio),
3558         };
3559 
3560         /*
3561          * Fixup the legacy SCHED_RESET_ON_FORK hack
3562          */
3563         if (policy & SCHED_RESET_ON_FORK) {
3564                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3565                 policy &= ~SCHED_RESET_ON_FORK;
3566                 attr.sched_policy = policy;
3567         }
3568 
3569         return __sched_setscheduler(p, &attr, check);
3570 }
3571 /**
3572  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3573  * @p: the task in question.
3574  * @policy: new policy.
3575  * @param: structure containing the new RT priority.
3576  *
3577  * Return: 0 on success. An error code otherwise.
3578  *
3579  * NOTE that the task may be already dead.
3580  */
3581 int sched_setscheduler(struct task_struct *p, int policy,
3582                        const struct sched_param *param)
3583 {
3584         return _sched_setscheduler(p, policy, param, true);
3585 }
3586 EXPORT_SYMBOL_GPL(sched_setscheduler);
3587 
3588 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3589 {
3590         return __sched_setscheduler(p, attr, true);
3591 }
3592 EXPORT_SYMBOL_GPL(sched_setattr);
3593 
3594 /**
3595  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3596  * @p: the task in question.
3597  * @policy: new policy.
3598  * @param: structure containing the new RT priority.
3599  *
3600  * Just like sched_setscheduler, only don't bother checking if the
3601  * current context has permission.  For example, this is needed in
3602  * stop_machine(): we create temporary high priority worker threads,
3603  * but our caller might not have that capability.
3604  *
3605  * Return: 0 on success. An error code otherwise.
3606  */
3607 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3608                                const struct sched_param *param)
3609 {
3610         return _sched_setscheduler(p, policy, param, false);
3611 }
3612 
3613 static int
3614 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3615 {
3616         struct sched_param lparam;
3617         struct task_struct *p;
3618         int retval;
3619 
3620         if (!param || pid < 0)
3621                 return -EINVAL;
3622         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3623                 return -EFAULT;
3624 
3625         rcu_read_lock();
3626         retval = -ESRCH;
3627         p = find_process_by_pid(pid);
3628         if (p != NULL)
3629                 retval = sched_setscheduler(p, policy, &lparam);
3630         rcu_read_unlock();
3631 
3632         return retval;
3633 }
3634 
3635 /*
3636  * Mimics kernel/events/core.c perf_copy_attr().
3637  */
3638 static int sched_copy_attr(struct sched_attr __user *uattr,
3639                            struct sched_attr *attr)
3640 {
3641         u32 size;
3642         int ret;
3643 
3644         if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3645                 return -EFAULT;
3646 
3647         /*
3648          * zero the full structure, so that a short copy will be nice.
3649          */
3650         memset(attr, 0, sizeof(*attr));
3651 
3652         ret = get_user(size, &uattr->size);
3653         if (ret)
3654                 return ret;
3655 
3656         if (size > PAGE_SIZE)   /* silly large */
3657                 goto err_size;
3658 
3659         if (!size)              /* abi compat */
3660                 size = SCHED_ATTR_SIZE_VER0;
3661 
3662         if (size < SCHED_ATTR_SIZE_VER0)
3663                 goto err_size;
3664 
3665         /*
3666          * If we're handed a bigger struct than we know of,
3667          * ensure all the unknown bits are 0 - i.e. new
3668          * user-space does not rely on any kernel feature
3669          * extensions we dont know about yet.
3670          */
3671         if (size > sizeof(*attr)) {
3672                 unsigned char __user *addr;
3673                 unsigned char __user *end;
3674                 unsigned char val;
3675 
3676                 addr = (void __user *)uattr + sizeof(*attr);
3677                 end  = (void __user *)uattr + size;
3678 
3679                 for (; addr < end; addr++) {
3680                         ret = get_user(val, addr);
3681                         if (ret)
3682                                 return ret;
3683                         if (val)
3684                                 goto err_size;
3685                 }
3686                 size = sizeof(*attr);
3687         }
3688 
3689         ret = copy_from_user(attr, uattr, size);
3690         if (ret)
3691                 return -EFAULT;
3692 
3693         /*
3694          * XXX: do we want to be lenient like existing syscalls; or do we want
3695          * to be strict and return an error on out-of-bounds values?
3696          */
3697         attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3698 
3699         return 0;
3700 
3701 err_size:
3702         put_user(sizeof(*attr), &uattr->size);
3703         return -E2BIG;
3704 }
3705 
3706 /**
3707  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3708  * @pid: the pid in question.
3709  * @policy: new policy.
3710  * @param: structure containing the new RT priority.
3711  *
3712  * Return: 0 on success. An error code otherwise.
3713  */
3714 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3715                 struct sched_param __user *, param)
3716 {
3717         /* negative values for policy are not valid */
3718         if (policy < 0)
3719                 return -EINVAL;
3720 
3721         return do_sched_setscheduler(pid, policy, param);
3722 }
3723 
3724 /**
3725  * sys_sched_setparam - set/change the RT priority of a thread
3726  * @pid: the pid in question.
3727  * @param: structure containing the new RT priority.
3728  *
3729  * Return: 0 on success. An error code otherwise.
3730  */
3731 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3732 {
3733         return do_sched_setscheduler(pid, -1, param);
3734 }
3735 
3736 /**
3737  * sys_sched_setattr - same as above, but with extended sched_attr
3738  * @pid: the pid in question.
3739  * @uattr: structure containing the extended parameters.
3740  * @flags: for future extension.
3741  */
3742 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3743                                unsigned int, flags)
3744 {
3745         struct sched_attr attr;
3746         struct task_struct *p;
3747         int retval;
3748 
3749         if (!uattr || pid < 0 || flags)
3750                 return -EINVAL;
3751 
3752         retval = sched_copy_attr(uattr, &attr);
3753         if (retval)
3754                 return retval;
3755 
3756         if ((int)attr.sched_policy < 0)
3757                 return -EINVAL;
3758 
3759         rcu_read_lock();
3760         retval = -ESRCH;
3761         p = find_process_by_pid(pid);
3762         if (p != NULL)
3763                 retval = sched_setattr(p, &attr);
3764         rcu_read_unlock();
3765 
3766         return retval;
3767 }
3768 
3769 /**
3770  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3771  * @pid: the pid in question.
3772  *
3773  * Return: On success, the policy of the thread. Otherwise, a negative error
3774  * code.
3775  */
3776 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3777 {
3778         struct task_struct *p;
3779         int retval;
3780 
3781         if (pid < 0)
3782                 return -EINVAL;
3783 
3784         retval = -ESRCH;
3785         rcu_read_lock();
3786         p = find_process_by_pid(pid);
3787         if (p) {
3788                 retval = security_task_getscheduler(p);
3789                 if (!retval)
3790                         retval = p->policy
3791                                 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3792         }
3793         rcu_read_unlock();
3794         return retval;
3795 }
3796 
3797 /**
3798  * sys_sched_getparam - get the RT priority of a thread
3799  * @pid: the pid in question.
3800  * @param: structure containing the RT priority.
3801  *
3802  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3803  * code.
3804  */
3805 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3806 {
3807         struct sched_param lp = { .sched_priority = 0 };
3808         struct task_struct *p;
3809         int retval;
3810 
3811         if (!param || pid < 0)
3812                 return -EINVAL;
3813 
3814         rcu_read_lock();
3815         p = find_process_by_pid(pid);
3816         retval = -ESRCH;
3817         if (!p)
3818                 goto out_unlock;
3819 
3820         retval = security_task_getscheduler(p);
3821         if (retval)
3822                 goto out_unlock;
3823 
3824         if (task_has_rt_policy(p))
3825                 lp.sched_priority = p->rt_priority;
3826         rcu_read_unlock();
3827 
3828         /*
3829          * This one might sleep, we cannot do it with a spinlock held ...
3830          */
3831         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3832 
3833         return retval;
3834 
3835 out_unlock:
3836         rcu_read_unlock();
3837         return retval;
3838 }
3839 
3840 static int sched_read_attr(struct sched_attr __user *uattr,
3841                            struct sched_attr *attr,
3842                            unsigned int usize)
3843 {
3844         int ret;
3845 
3846         if (!access_ok(VERIFY_WRITE, uattr, usize))
3847                 return -EFAULT;
3848 
3849         /*
3850          * If we're handed a smaller struct than we know of,
3851          * ensure all the unknown bits are 0 - i.e. old
3852          * user-space does not get uncomplete information.
3853          */
3854         if (usize < sizeof(*attr)) {
3855                 unsigned char *addr;
3856                 unsigned char *end;
3857 
3858                 addr = (void *)attr + usize;
3859                 end  = (void *)attr + sizeof(*attr);
3860 
3861                 for (; addr < end; addr++) {
3862                         if (*addr)
3863                                 return -EFBIG;
3864                 }
3865 
3866                 attr->size = usize;
3867         }
3868 
3869         ret = copy_to_user(uattr, attr, attr->size);
3870         if (ret)
3871                 return -EFAULT;
3872 
3873         return 0;
3874 }
3875 
3876 /**
3877  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3878  * @pid: the pid in question.
3879  * @uattr: structure containing the extended parameters.
3880  * @size: sizeof(attr) for fwd/bwd comp.
3881  * @flags: for future extension.
3882  */
3883 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3884                 unsigned int, size, unsigned int, flags)
3885 {
3886         struct sched_attr attr = {
3887                 .size = sizeof(struct sched_attr),
3888         };
3889         struct task_struct *p;
3890         int retval;
3891 
3892         if (!uattr || pid < 0 || size > PAGE_SIZE ||
3893             size < SCHED_ATTR_SIZE_VER0 || flags)
3894                 return -EINVAL;
3895 
3896         rcu_read_lock();
3897         p = find_process_by_pid(pid);
3898         retval = -ESRCH;
3899         if (!p)
3900                 goto out_unlock;
3901 
3902         retval = security_task_getscheduler(p);
3903         if (retval)
3904                 goto out_unlock;
3905 
3906         attr.sched_policy = p->policy;
3907         if (p->sched_reset_on_fork)
3908                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3909         if (task_has_dl_policy(p))
3910                 __getparam_dl(p, &attr);
3911         else if (task_has_rt_policy(p))
3912                 attr.sched_priority = p->rt_priority;
3913         else
3914                 attr.sched_nice = task_nice(p);
3915 
3916         rcu_read_unlock();
3917 
3918         retval = sched_read_attr(uattr, &attr, size);
3919         return retval;
3920 
3921 out_unlock:
3922         rcu_read_unlock();
3923         return retval;
3924 }
3925 
3926 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3927 {
3928         cpumask_var_t cpus_allowed, new_mask;
3929         struct task_struct *p;
3930         int retval;
3931 
3932         rcu_read_lock();
3933 
3934         p = find_process_by_pid(pid);
3935         if (!p) {
3936                 rcu_read_unlock();
3937                 return -ESRCH;
3938         }
3939 
3940         /* Prevent p going away */
3941         get_task_struct(p);
3942         rcu_read_unlock();
3943 
3944         if (p->flags & PF_NO_SETAFFINITY) {
3945                 retval = -EINVAL;
3946                 goto out_put_task;
3947         }
3948         if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3949                 retval = -ENOMEM;
3950                 goto out_put_task;
3951         }
3952         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3953                 retval = -ENOMEM;
3954                 goto out_free_cpus_allowed;
3955         }
3956         retval = -EPERM;
3957         if (!check_same_owner(p)) {
3958                 rcu_read_lock();
3959                 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3960                         rcu_read_unlock();
3961                         goto out_unlock;
3962                 }
3963                 rcu_read_unlock();
3964         }
3965 
3966         retval = security_task_setscheduler(p);
3967         if (retval)
3968                 goto out_unlock;
3969 
3970 
3971         cpuset_cpus_allowed(p, cpus_allowed);
3972         cpumask_and(new_mask, in_mask, cpus_allowed);
3973 
3974         /*
3975          * Since bandwidth control happens on root_domain basis,
3976          * if admission test is enabled, we only admit -deadline
3977          * tasks allowed to run on all the CPUs in the task's
3978          * root_domain.
3979          */
3980 #ifdef CONFIG_SMP
3981         if (task_has_dl_policy(p)) {
3982                 const struct cpumask *span = task_rq(p)->rd->span;
3983 
3984                 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3985                         retval = -EBUSY;
3986                         goto out_unlock;
3987                 }
3988         }
3989 #endif
3990 again:
3991         retval = set_cpus_allowed_ptr(p, new_mask);
3992 
3993         if (!retval) {
3994                 cpuset_cpus_allowed(p, cpus_allowed);
3995                 if (!cpumask_subset(new_mask, cpus_allowed)) {
3996                         /*
3997                          * We must have raced with a concurrent cpuset
3998                          * update. Just reset the cpus_allowed to the
3999                          * cpuset's cpus_allowed
4000                          */
4001                         cpumask_copy(new_mask, cpus_allowed);
4002                         goto again;
4003                 }
4004         }
4005 out_unlock:
4006         free_cpumask_var(new_mask);
4007 out_free_cpus_allowed:
4008         free_cpumask_var(cpus_allowed);
4009 out_put_task:
4010         put_task_struct(p);
4011         return retval;
4012 }
4013 
4014 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4015                              struct cpumask *new_mask)
4016 {
4017         if (len < cpumask_size())
4018                 cpumask_clear(new_mask);
4019         else if (len > cpumask_size())
4020                 len = cpumask_size();
4021 
4022         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4023 }
4024 
4025 /**
4026  * sys_sched_setaffinity - set the cpu affinity of a process
4027  * @pid: pid of the process
4028  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4029  * @user_mask_ptr: user-space pointer to the new cpu mask
4030  *
4031  * Return: 0 on success. An error code otherwise.
4032  */
4033 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4034                 unsigned long __user *, user_mask_ptr)
4035 {
4036         cpumask_var_t new_mask;
4037         int retval;
4038 
4039         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4040                 return -ENOMEM;
4041 
4042         retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4043         if (retval == 0)
4044                 retval = sched_setaffinity(pid, new_mask);
4045         free_cpumask_var(new_mask);
4046         return retval;
4047 }
4048 
4049 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4050 {
4051         struct task_struct *p;
4052         unsigned long flags;
4053         int retval;
4054 
4055         rcu_read_lock();
4056 
4057         retval = -ESRCH;
4058         p = find_process_by_pid(pid);
4059         if (!p)
4060                 goto out_unlock;
4061 
4062         retval = security_task_getscheduler(p);
4063         if (retval)
4064                 goto out_unlock;
4065 
4066         raw_spin_lock_irqsave(&p->pi_lock, flags);
4067         cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4068         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4069 
4070 out_unlock:
4071         rcu_read_unlock();
4072 
4073         return retval;
4074 }
4075 
4076 /**
4077  * sys_sched_getaffinity - get the cpu affinity of a process
4078  * @pid: pid of the process
4079  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4080  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4081  *
4082  * Return: 0 on success. An error code otherwise.
4083  */
4084 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4085                 unsigned long __user *, user_mask_ptr)
4086 {
4087         int ret;
4088         cpumask_var_t mask;
4089 
4090         if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4091                 return -EINVAL;
4092         if (len & (sizeof(unsigned long)-1))
4093                 return -EINVAL;
4094 
4095         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4096                 return -ENOMEM;
4097 
4098         ret = sched_getaffinity(pid, mask);
4099         if (ret == 0) {
4100                 size_t retlen = min_t(size_t, len, cpumask_size());
4101 
4102                 if (copy_to_user(user_mask_ptr, mask, retlen))
4103                         ret = -EFAULT;
4104                 else
4105                         ret = retlen;
4106         }
4107         free_cpumask_var(mask);
4108 
4109         return ret;
4110 }
4111 
4112 /**
4113  * sys_sched_yield - yield the current processor to other threads.
4114  *
4115  * This function yields the current CPU to other tasks. If there are no
4116  * other threads running on this CPU then this function will return.
4117  *
4118  * Return: 0.
4119  */
4120 SYSCALL_DEFINE0(sched_yield)
4121 {
4122         struct rq *rq = this_rq_lock();
4123 
4124         schedstat_inc(rq, yld_count);
4125         current->sched_class->yield_task(rq);
4126 
4127         /*
4128          * Since we are going to call schedule() anyway, there's
4129          * no need to preempt or enable interrupts:
4130          */
4131         __release(rq->lock);
4132         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4133         do_raw_spin_unlock(&rq->lock);
4134         sched_preempt_enable_no_resched();
4135 
4136         schedule();
4137 
4138         return 0;
4139 }
4140 
4141 static void __cond_resched(void)
4142 {
4143         __preempt_count_add(PREEMPT_ACTIVE);
4144         __schedule();
4145         __preempt_count_sub(PREEMPT_ACTIVE);
4146 }
4147 
4148 int __sched _cond_resched(void)
4149 {
4150         if (should_resched()) {
4151                 __cond_resched();
4152                 return 1;
4153         }
4154         return 0;
4155 }
4156 EXPORT_SYMBOL(_cond_resched);
4157 
4158 /*
4159  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4160  * call schedule, and on return reacquire the lock.
4161  *
4162  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4163  * operations here to prevent schedule() from being called twice (once via
4164  * spin_unlock(), once by hand).
4165  */
4166 int __cond_resched_lock(spinlock_t *lock)
4167 {
4168         int resched = should_resched();
4169         int ret = 0;
4170 
4171         lockdep_assert_held(lock);
4172 
4173         if (spin_needbreak(lock) || resched) {
4174                 spin_unlock(lock);
4175                 if (resched)
4176                         __cond_resched();
4177                 else
4178                         cpu_relax();
4179                 ret = 1;
4180                 spin_lock(lock);
4181         }
4182         return ret;
4183 }
4184 EXPORT_SYMBOL(__cond_resched_lock);
4185 
4186 int __sched __cond_resched_softirq(void)
4187 {
4188         BUG_ON(!in_softirq());
4189 
4190         if (should_resched()) {
4191                 local_bh_enable();
4192                 __cond_resched();
4193                 local_bh_disable();
4194                 return 1;
4195         }
4196         return 0;
4197 }
4198 EXPORT_SYMBOL(__cond_resched_softirq);
4199 
4200 /**
4201  * yield - yield the current processor to other threads.
4202  *
4203  * Do not ever use this function, there's a 99% chance you're doing it wrong.
4204  *
4205  * The scheduler is at all times free to pick the calling task as the most
4206  * eligible task to run, if removing the yield() call from your code breaks
4207  * it, its already broken.
4208  *
4209  * Typical broken usage is:
4210  *
4211  * while (!event)
4212  *      yield();
4213  *
4214  * where one assumes that yield() will let 'the other' process run that will
4215  * make event true. If the current task is a SCHED_FIFO task that will never
4216  * happen. Never use yield() as a progress guarantee!!
4217  *
4218  * If you want to use yield() to wait for something, use wait_event().
4219  * If you want to use yield() to be 'nice' for others, use cond_resched().
4220  * If you still want to use yield(), do not!
4221  */
4222 void __sched yield(void)
4223 {
4224         set_current_state(TASK_RUNNING);
4225         sys_sched_yield();
4226 }
4227 EXPORT_SYMBOL(yield);
4228 
4229 /**
4230  * yield_to - yield the current processor to another thread in
4231  * your thread group, or accelerate that thread toward the
4232  * processor it's on.
4233  * @p: target task
4234  * @preempt: whether task preemption is allowed or not
4235  *
4236  * It's the caller's job to ensure that the target task struct
4237  * can't go away on us before we can do any checks.
4238  *
4239  * Return:
4240  *      true (>0) if we indeed boosted the target task.
4241  *      false (0) if we failed to boost the target.
4242  *      -ESRCH if there's no task to yield to.
4243  */
4244 int __sched yield_to(struct task_struct *p, bool preempt)
4245 {
4246         struct task_struct *curr = current;
4247         struct rq *rq, *p_rq;
4248         unsigned long flags;
4249         int yielded = 0;
4250 
4251         local_irq_save(flags);
4252         rq = this_rq();
4253 
4254 again:
4255         p_rq = task_rq(p);
4256         /*
4257          * If we're the only runnable task on the rq and target rq also
4258          * has only one task, there's absolutely no point in yielding.
4259          */
4260         if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4261                 yielded = -ESRCH;
4262                 goto out_irq;
4263         }
4264 
4265         double_rq_lock(rq, p_rq);
4266         if (task_rq(p) != p_rq) {
4267                 double_rq_unlock(rq, p_rq);
4268                 goto again;
4269         }
4270 
4271         if (!curr->sched_class->yield_to_task)
4272                 goto out_unlock;
4273 
4274         if (curr->sched_class != p->sched_class)
4275                 goto out_unlock;
4276 
4277         if (task_running(p_rq, p) || p->state)
4278                 goto out_unlock;
4279 
4280         yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4281         if (yielded) {
4282                 schedstat_inc(rq, yld_count);
4283                 /*
4284                  * Make p's CPU reschedule; pick_next_entity takes care of
4285                  * fairness.
4286                  */
4287                 if (preempt && rq != p_rq)
4288                         resched_task(p_rq->curr);
4289         }
4290 
4291 out_unlock:
4292         double_rq_unlock(rq, p_rq);
4293 out_irq:
4294         local_irq_restore(flags);
4295 
4296         if (yielded > 0)
4297                 schedule();
4298 
4299         return yielded;
4300 }
4301 EXPORT_SYMBOL_GPL(yield_to);
4302 
4303 /*
4304  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4305  * that process accounting knows that this is a task in IO wait state.
4306  */
4307 void __sched io_schedule(void)
4308 {
4309         struct rq *rq = raw_rq();
4310 
4311         delayacct_blkio_start();
4312         atomic_inc(&rq->nr_iowait);
4313         blk_flush_plug(current);
4314         current->in_iowait = 1;
4315         schedule();
4316         current->in_iowait = 0;
4317         atomic_dec(&rq->nr_iowait);
4318         delayacct_blkio_end();
4319 }
4320 EXPORT_SYMBOL(io_schedule);
4321 
4322 long __sched io_schedule_timeout(long timeout)
4323 {
4324         struct rq *rq = raw_rq();
4325         long ret;
4326 
4327         delayacct_blkio_start();
4328         atomic_inc(&rq->nr_iowait);
4329         blk_flush_plug(current);
4330         current->in_iowait = 1;
4331         ret = schedule_timeout(timeout);
4332         current->in_iowait = 0;
4333         atomic_dec(&rq->nr_iowait);
4334         delayacct_blkio_end();
4335         return ret;
4336 }
4337 
4338 /**
4339  * sys_sched_get_priority_max - return maximum RT priority.
4340  * @policy: scheduling class.
4341  *
4342  * Return: On success, this syscall returns the maximum
4343  * rt_priority that can be used by a given scheduling class.
4344  * On failure, a negative error code is returned.
4345  */
4346 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4347 {
4348         int ret = -EINVAL;
4349 
4350         switch (policy) {
4351         case SCHED_FIFO:
4352         case SCHED_RR:
4353                 ret = MAX_USER_RT_PRIO-1;
4354                 break;
4355         case SCHED_DEADLINE:
4356         case SCHED_NORMAL:
4357         case SCHED_BATCH:
4358         case SCHED_IDLE:
4359                 ret = 0;
4360                 break;
4361         }
4362         return ret;
4363 }
4364 
4365 /**
4366  * sys_sched_get_priority_min - return minimum RT priority.
4367  * @policy: scheduling class.
4368  *
4369  * Return: On success, this syscall returns the minimum
4370  * rt_priority that can be used by a given scheduling class.
4371  * On failure, a negative error code is returned.
4372  */
4373 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4374 {
4375         int ret = -EINVAL;
4376 
4377         switch (policy) {
4378         case SCHED_FIFO:
4379         case SCHED_RR:
4380                 ret = 1;
4381                 break;
4382         case SCHED_DEADLINE:
4383         case SCHED_NORMAL:
4384         case SCHED_BATCH:
4385         case SCHED_IDLE:
4386                 ret = 0;
4387         }
4388         return ret;
4389 }
4390 
4391 /**
4392  * sys_sched_rr_get_interval - return the default timeslice of a process.
4393  * @pid: pid of the process.
4394  * @interval: userspace pointer to the timeslice value.
4395  *
4396  * this syscall writes the default timeslice value of a given process
4397  * into the user-space timespec buffer. A value of '' means infinity.
4398  *
4399  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4400  * an error code.
4401  */
4402 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4403                 struct timespec __user *, interval)
4404 {
4405         struct task_struct *p;
4406         unsigned int time_slice;
4407         unsigned long flags;
4408         struct rq *rq;
4409         int retval;
4410         struct timespec t;
4411 
4412         if (pid < 0)
4413                 return -EINVAL;
4414 
4415         retval = -ESRCH;
4416         rcu_read_lock();
4417         p = find_process_by_pid(pid);
4418         if (!p)
4419                 goto out_unlock;
4420 
4421         retval = security_task_getscheduler(p);
4422         if (retval)
4423                 goto out_unlock;
4424 
4425         rq = task_rq_lock(p, &flags);
4426         time_slice = 0;
4427         if (p->sched_class->get_rr_interval)
4428                 time_slice = p->sched_class->get_rr_interval(rq, p);
4429         task_rq_unlock(rq, p, &flags);
4430 
4431         rcu_read_unlock();
4432         jiffies_to_timespec(time_slice, &t);
4433         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4434         return retval;
4435 
4436 out_unlock:
4437         rcu_read_unlock();
4438         return retval;
4439 }
4440 
4441 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4442 
4443 void sched_show_task(struct task_struct *p)
4444 {
4445         unsigned long free = 0;
4446         int ppid;
4447         unsigned state;
4448 
4449         state = p->state ? __ffs(p->state) + 1 : 0;
4450         printk(KERN_INFO "%-15.15s %c", p->comm,
4451                 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4452 #if BITS_PER_LONG == 32
4453         if (state == TASK_RUNNING)
4454                 printk(KERN_CONT " running  ");
4455         else
4456                 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4457 #else
4458         if (state == TASK_RUNNING)
4459                 printk(KERN_CONT "  running task    ");
4460         else
4461                 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4462 #endif
4463 #ifdef CONFIG_DEBUG_STACK_USAGE
4464         free = stack_not_used(p);
4465 #endif
4466         rcu_read_lock();
4467         ppid = task_pid_nr(rcu_dereference(p->real_parent));
4468         rcu_read_unlock();
4469         printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4470                 task_pid_nr(p), ppid,
4471                 (unsigned long)task_thread_info(p)->flags);
4472 
4473         print_worker_info(KERN_INFO, p);
4474         show_stack(p, NULL);
4475 }
4476 
4477 void show_state_filter(unsigned long state_filter)
4478 {
4479         struct task_struct *g, *p;
4480 
4481 #if BITS_PER_LONG == 32
4482         printk(KERN_INFO
4483                 "  task                PC stack   pid father\n");
4484 #else
4485         printk(KERN_INFO
4486                 "  task                        PC stack   pid father\n");
4487 #endif
4488         rcu_read_lock();
4489         do_each_thread(g, p) {
4490                 /*
4491                  * reset the NMI-timeout, listing all files on a slow
4492                  * console might take a lot of time:
4493                  */
4494                 touch_nmi_watchdog();
4495                 if (!state_filter || (p->state & state_filter))
4496                         sched_show_task(p);
4497         } while_each_thread(g, p);
4498 
4499         touch_all_softlockup_watchdogs();
4500 
4501 #ifdef CONFIG_SCHED_DEBUG
4502         sysrq_sched_debug_show();
4503 #endif
4504         rcu_read_unlock();
4505         /*
4506          * Only show locks if all tasks are dumped:
4507          */
4508         if (!state_filter)
4509                 debug_show_all_locks();
4510 }
4511 
4512 void init_idle_bootup_task(struct task_struct *idle)
4513 {
4514         idle->sched_class = &idle_sched_class;
4515 }
4516 
4517 /**
4518  * init_idle - set up an idle thread for a given CPU
4519  * @idle: task in question
4520  * @cpu: cpu the idle task belongs to
4521  *
4522  * NOTE: this function does not set the idle thread's NEED_RESCHED
4523  * flag, to make booting more robust.
4524  */
4525 void init_idle(struct task_struct *idle, int cpu)
4526 {
4527         struct rq *rq = cpu_rq(cpu);
4528         unsigned long flags;
4529 
4530         raw_spin_lock_irqsave(&rq->lock, flags);
4531 
4532         __sched_fork(0, idle);
4533         idle->state = TASK_RUNNING;
4534         idle->se.exec_start = sched_clock();
4535 
4536         do_set_cpus_allowed(idle, cpumask_of(cpu));
4537         /*
4538          * We're having a chicken and egg problem, even though we are
4539          * holding rq->lock, the cpu isn't yet set to this cpu so the
4540          * lockdep check in task_group() will fail.
4541          *
4542          * Similar case to sched_fork(). / Alternatively we could
4543          * use task_rq_lock() here and obtain the other rq->lock.
4544          *
4545          * Silence PROVE_RCU
4546          */
4547         rcu_read_lock();
4548         __set_task_cpu(idle, cpu);
4549         rcu_read_unlock();
4550 
4551         rq->curr = rq->idle = idle;
4552         idle->on_rq = 1;
4553 #if defined(CONFIG_SMP)
4554         idle->on_cpu = 1;
4555 #endif
4556         raw_spin_unlock_irqrestore(&rq->lock, flags);
4557 
4558         /* Set the preempt count _outside_ the spinlocks! */
4559         init_idle_preempt_count(idle, cpu);
4560 
4561         /*
4562          * The idle tasks have their own, simple scheduling class:
4563          */
4564         idle->sched_class = &idle_sched_class;
4565         ftrace_graph_init_idle_task(idle, cpu);
4566         vtime_init_idle(idle, cpu);
4567 #if defined(CONFIG_SMP)
4568         sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4569 #endif
4570 }
4571 
4572 #ifdef CONFIG_SMP
4573 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4574 {
4575         if (p->sched_class && p->sched_class->set_cpus_allowed)
4576                 p->sched_class->set_cpus_allowed(p, new_mask);
4577 
4578         cpumask_copy(&p->cpus_allowed, new_mask);
4579         p->nr_cpus_allowed = cpumask_weight(new_mask);
4580 }
4581 
4582 /*
4583  * This is how migration works:
4584  *
4585  * 1) we invoke migration_cpu_stop() on the target CPU using
4586  *    stop_one_cpu().
4587  * 2) stopper starts to run (implicitly forcing the migrated thread
4588  *    off the CPU)
4589  * 3) it checks whether the migrated task is still in the wrong runqueue.
4590  * 4) if it's in the wrong runqueue then the migration thread removes
4591  *    it and puts it into the right queue.
4592  * 5) stopper completes and stop_one_cpu() returns and the migration
4593  *    is done.
4594  */
4595 
4596 /*
4597  * Change a given task's CPU affinity. Migrate the thread to a
4598  * proper CPU and schedule it away if the CPU it's executing on
4599  * is removed from the allowed bitmask.
4600  *
4601  * NOTE: the caller must have a valid reference to the task, the
4602  * task must not exit() & deallocate itself prematurely. The
4603  * call is not atomic; no spinlocks may be held.
4604  */
4605 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4606 {
4607         unsigned long flags;
4608         struct rq *rq;
4609         unsigned int dest_cpu;
4610         int ret = 0;
4611 
4612         rq = task_rq_lock(p, &flags);
4613 
4614         if (cpumask_equal(&p->cpus_allowed, new_mask))
4615                 goto out;
4616 
4617         if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4618                 ret = -EINVAL;
4619                 goto out;
4620         }
4621 
4622         do_set_cpus_allowed(p, new_mask);
4623 
4624         /* Can the task run on the task's current CPU? If so, we're done */
4625         if (cpumask_test_cpu(task_cpu(p), new_mask))
4626                 goto out;
4627 
4628         dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4629         if (p->on_rq) {
4630                 struct migration_arg arg = { p, dest_cpu };
4631                 /* Need help from migration thread: drop lock and wait. */
4632                 task_rq_unlock(rq, p, &flags);
4633                 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4634                 tlb_migrate_finish(p->mm);
4635                 return 0;
4636         }
4637 out:
4638         task_rq_unlock(rq, p, &flags);
4639 
4640         return ret;
4641 }
4642 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4643 
4644 /*
4645  * Move (not current) task off this cpu, onto dest cpu. We're doing
4646  * this because either it can't run here any more (set_cpus_allowed()
4647  * away from this CPU, or CPU going down), or because we're
4648  * attempting to rebalance this task on exec (sched_exec).
4649  *
4650  * So we race with normal scheduler movements, but that's OK, as long
4651  * as the task is no longer on this CPU.
4652  *
4653  * Returns non-zero if task was successfully migrated.
4654  */
4655 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4656 {
4657         struct rq *rq_dest, *rq_src;
4658         int ret = 0;
4659 
4660         if (unlikely(!cpu_active(dest_cpu)))
4661                 return ret;
4662 
4663         rq_src = cpu_rq(src_cpu);
4664         rq_dest = cpu_rq(dest_cpu);
4665 
4666         raw_spin_lock(&p->pi_lock);
4667         double_rq_lock(rq_src, rq_dest);
4668         /* Already moved. */
4669         if (task_cpu(p) != src_cpu)
4670                 goto done;
4671         /* Affinity changed (again). */
4672         if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4673                 goto fail;
4674 
4675         /*
4676          * If we're not on a rq, the next wake-up will ensure we're
4677          * placed properly.
4678          */
4679         if (p->on_rq) {
4680                 dequeue_task(rq_src, p, 0);
4681                 set_task_cpu(p, dest_cpu);
4682                 enqueue_task(rq_dest, p, 0);
4683                 check_preempt_curr(rq_dest, p, 0);
4684         }
4685 done:
4686         ret = 1;
4687 fail:
4688         double_rq_unlock(rq_src, rq_dest);
4689         raw_spin_unlock(&p->pi_lock);
4690         return ret;
4691 }
4692 
4693 #ifdef CONFIG_NUMA_BALANCING
4694 /* Migrate current task p to target_cpu */
4695 int migrate_task_to(struct task_struct *p, int target_cpu)
4696 {
4697         struct migration_arg arg = { p, target_cpu };
4698         int curr_cpu = task_cpu(p);
4699 
4700         if (curr_cpu == target_cpu)
4701                 return 0;
4702 
4703         if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4704                 return -EINVAL;
4705 
4706         /* TODO: This is not properly updating schedstats */
4707 
4708         trace_sched_move_numa(p, curr_cpu, target_cpu);
4709         return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4710 }
4711 
4712 /*
4713  * Requeue a task on a given node and accurately track the number of NUMA
4714  * tasks on the runqueues
4715  */
4716 void sched_setnuma(struct task_struct *p, int nid)
4717 {
4718         struct rq *rq;
4719         unsigned long flags;
4720         bool on_rq, running;
4721 
4722         rq = task_rq_lock(p, &flags);
4723         on_rq = p->on_rq;
4724         running = task_current(rq, p);
4725 
4726         if (on_rq)
4727                 dequeue_task(rq, p, 0);
4728         if (running)
4729                 p->sched_class->put_prev_task(rq, p);
4730 
4731         p->numa_preferred_nid = nid;
4732 
4733         if (running)
4734                 p->sched_class->set_curr_task(rq);
4735         if (on_rq)
4736                 enqueue_task(rq, p, 0);
4737         task_rq_unlock(rq, p, &flags);
4738 }
4739 #endif
4740 
4741 /*
4742  * migration_cpu_stop - this will be executed by a highprio stopper thread
4743  * and performs thread migration by bumping thread off CPU then
4744  * 'pushing' onto another runqueue.
4745  */
4746 static int migration_cpu_stop(void *data)
4747 {
4748         struct migration_arg *arg = data;
4749 
4750         /*
4751          * The original target cpu might have gone down and we might
4752          * be on another cpu but it doesn't matter.
4753          */
4754         local_irq_disable();
4755         __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4756         local_irq_enable();
4757         return 0;
4758 }
4759 
4760 #ifdef CONFIG_HOTPLUG_CPU
4761 
4762 /*
4763  * Ensures that the idle task is using init_mm right before its cpu goes
4764  * offline.
4765  */
4766 void idle_task_exit(void)
4767 {
4768         struct mm_struct *mm = current->active_mm;
4769 
4770         BUG_ON(cpu_online(smp_processor_id()));
4771 
4772         if (mm != &init_mm) {
4773                 switch_mm(mm, &init_mm, current);
4774                 finish_arch_post_lock_switch();
4775         }
4776         mmdrop(mm);
4777 }
4778 
4779 /*
4780  * Since this CPU is going 'away' for a while, fold any nr_active delta
4781  * we might have. Assumes we're called after migrate_tasks() so that the
4782  * nr_active count is stable.
4783  *
4784  * Also see the comment "Global load-average calculations".
4785  */
4786 static void calc_load_migrate(struct rq *rq)
4787 {
4788         long delta = calc_load_fold_active(rq);
4789         if (delta)
4790                 atomic_long_add(delta, &calc_load_tasks);
4791 }
4792 
4793 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4794 {
4795 }
4796 
4797 static const struct sched_class fake_sched_class = {
4798         .put_prev_task = put_prev_task_fake,
4799 };
4800 
4801 static struct task_struct fake_task = {
4802         /*
4803          * Avoid pull_{rt,dl}_task()
4804          */
4805         .prio = MAX_PRIO + 1,
4806         .sched_class = &fake_sched_class,
4807 };
4808 
4809 /*
4810  * Migrate all tasks from the rq, sleeping tasks will be migrated by
4811  * try_to_wake_up()->select_task_rq().
4812  *
4813  * Called with rq->lock held even though we'er in stop_machine() and
4814  * there's no concurrency possible, we hold the required locks anyway
4815  * because of lock validation efforts.
4816  */
4817 static void migrate_tasks(unsigned int dead_cpu)
4818 {
4819         struct rq *rq = cpu_rq(dead_cpu);
4820         struct task_struct *next, *stop = rq->stop;
4821         int dest_cpu;
4822 
4823         /*
4824          * Fudge the rq selection such that the below task selection loop
4825          * doesn't get stuck on the currently eligible stop task.
4826          *
4827          * We're currently inside stop_machine() and the rq is either stuck
4828          * in the stop_machine_cpu_stop() loop, or we're executing this code,
4829          * either way we should never end up calling schedule() until we're
4830          * done here.
4831          */
4832         rq->stop = NULL;
4833 
4834         /*
4835          * put_prev_task() and pick_next_task() sched
4836          * class method both need to have an up-to-date
4837          * value of rq->clock[_task]
4838          */
4839         update_rq_clock(rq);
4840 
4841         for ( ; ; ) {
4842                 /*
4843                  * There's this thread running, bail when that's the only
4844                  * remaining thread.
4845                  */
4846                 if (rq->nr_running == 1)
4847                         break;
4848 
4849                 next = pick_next_task(rq, &fake_task);
4850                 BUG_ON(!next);
4851                 next->sched_class->put_prev_task(rq, next);
4852 
4853                 /* Find suitable destination for @next, with force if needed. */
4854                 dest_cpu = select_fallback_rq(dead_cpu, next);
4855                 raw_spin_unlock(&rq->lock);
4856 
4857                 __migrate_task(next, dead_cpu, dest_cpu);
4858 
4859                 raw_spin_lock(&rq->lock);
4860         }
4861 
4862         rq->stop = stop;
4863 }
4864 
4865 #endif /* CONFIG_HOTPLUG_CPU */
4866 
4867 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4868 
4869 static struct ctl_table sd_ctl_dir[] = {
4870         {
4871                 .procname       = "sched_domain",
4872                 .mode           = 0555,
4873         },
4874         {}
4875 };
4876 
4877 static struct ctl_table sd_ctl_root[] = {
4878         {
4879                 .procname       = "kernel",
4880                 .mode           = 0555,
4881                 .child          = sd_ctl_dir,
4882         },
4883         {}
4884 };
4885 
4886 static struct ctl_table *sd_alloc_ctl_entry(int n)
4887 {
4888         struct ctl_table *entry =
4889                 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4890 
4891         return entry;
4892 }
4893 
4894 static void sd_free_ctl_entry(struct ctl_table **tablep)
4895 {
4896         struct ctl_table *entry;
4897 
4898         /*
4899          * In the intermediate directories, both the child directory and
4900          * procname are dynamically allocated and could fail but the mode
4901          * will always be set. In the lowest directory the names are
4902          * static strings and all have proc handlers.
4903          */
4904         for (entry = *tablep; entry->mode; entry++) {
4905                 if (entry->child)
4906                         sd_free_ctl_entry(&entry->child);
4907                 if (entry->proc_handler == NULL)
4908                         kfree(entry->procname);
4909         }
4910 
4911         kfree(*tablep);
4912         *tablep = NULL;
4913 }
4914 
4915 static int min_load_idx = 0;
4916 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4917 
4918 static void
4919 set_table_entry(struct ctl_table *entry,
4920                 const char *procname, void *data, int maxlen,
4921                 umode_t mode, proc_handler *proc_handler,
4922                 bool load_idx)
4923 {
4924         entry->procname = procname;
4925         entry->data = data;
4926         entry->maxlen = maxlen;
4927         entry->mode = mode;
4928         entry->proc_handler = proc_handler;
4929 
4930         if (load_idx) {
4931                 entry->extra1 = &min_load_idx;
4932                 entry->extra2 = &max_load_idx;
4933         }
4934 }
4935 
4936 static struct ctl_table *
4937 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4938 {
4939         struct ctl_table *table = sd_alloc_ctl_entry(14);
4940 
4941         if (table == NULL)
4942                 return NULL;
4943 
4944         set_table_entry(&table[0], "min_interval", &sd->min_interval,
4945                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4946         set_table_entry(&table[1], "max_interval", &sd->max_interval,
4947                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4948         set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4949                 sizeof(int), 0644, proc_dointvec_minmax, true);
4950         set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4951                 sizeof(int), 0644, proc_dointvec_minmax, true);
4952         set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4953                 sizeof(int), 0644, proc_dointvec_minmax, true);
4954         set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4955                 sizeof(int), 0644, proc_dointvec_minmax, true);
4956         set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4957                 sizeof(int), 0644, proc_dointvec_minmax, true);
4958         set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4959                 sizeof(int), 0644, proc_dointvec_minmax, false);
4960         set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4961                 sizeof(int), 0644, proc_dointvec_minmax, false);
4962         set_table_entry(&table[9], "cache_nice_tries",
4963                 &sd->cache_nice_tries,
4964                 sizeof(int), 0644, proc_dointvec_minmax, false);
4965         set_table_entry(&table[10], "flags", &sd->flags,
4966                 sizeof(int), 0644, proc_dointvec_minmax, false);
4967         set_table_entry(&table[11], "max_newidle_lb_cost",
4968                 &sd->max_newidle_lb_cost,
4969                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4970         set_table_entry(&table[12], "name", sd->name,
4971                 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4972         /* &table[13] is terminator */
4973 
4974         return table;
4975 }
4976 
4977 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4978 {
4979         struct ctl_table *entry, *table;
4980         struct sched_domain *sd;
4981         int domain_num = 0, i;
4982         char buf[32];
4983 
4984         for_each_domain(cpu, sd)
4985                 domain_num++;
4986         entry = table = sd_alloc_ctl_entry(domain_num + 1);
4987         if (table == NULL)
4988                 return NULL;
4989 
4990         i = 0;
4991         for_each_domain(cpu, sd) {
4992                 snprintf(buf, 32, "domain%d", i);
4993                 entry->procname = kstrdup(buf, GFP_KERNEL);
4994                 entry->mode = 0555;
4995                 entry->child = sd_alloc_ctl_domain_table(sd);
4996                 entry++;
4997                 i++;
4998         }
4999         return table;
5000 }
5001 
5002 static struct ctl_table_header *sd_sysctl_header;
5003 static void register_sched_domain_sysctl(void)
5004 {
5005         int i, cpu_num = num_possible_cpus();
5006         struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5007         char buf[32];
5008 
5009         WARN_ON(sd_ctl_dir[0].child);
5010         sd_ctl_dir[0].child = entry;
5011 
5012         if (entry == NULL)
5013                 return;
5014 
5015         for_each_possible_cpu(i) {
5016                 snprintf(buf, 32, "cpu%d", i);
5017                 entry->procname = kstrdup(buf, GFP_KERNEL);
5018                 entry->mode = 0555;
5019                 entry->child = sd_alloc_ctl_cpu_table(i);
5020                 entry++;
5021         }
5022 
5023         WARN_ON(sd_sysctl_header);
5024         sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5025 }
5026 
5027 /* may be called multiple times per register */
5028 static void unregister_sched_domain_sysctl(void)
5029 {
5030         if (sd_sysctl_header)
5031                 unregister_sysctl_table(sd_sysctl_header);
5032         sd_sysctl_header = NULL;
5033         if (sd_ctl_dir[0].child)
5034                 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5035 }
5036 #else
5037 static void register_sched_domain_sysctl(void)
5038 {
5039 }
5040 static void unregister_sched_domain_sysctl(void)
5041 {
5042 }
5043 #endif
5044 
5045 static void set_rq_online(struct rq *rq)
5046 {
5047         if (!rq->online) {
5048                 const struct sched_class *class;
5049 
5050                 cpumask_set_cpu(rq->cpu, rq->rd->online);
5051                 rq->online = 1;
5052 
5053                 for_each_class(class) {
5054                         if (class->rq_online)
5055                                 class->rq_online(rq);
5056                 }
5057         }
5058 }
5059 
5060 static void set_rq_offline(struct rq *rq)
5061 {
5062         if (rq->online) {
5063                 const struct sched_class *class;
5064 
5065                 for_each_class(class) {
5066                         if (class->rq_offline)
5067                                 class->rq_offline(rq);
5068                 }
5069 
5070                 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5071                 rq->online = 0;
5072         }
5073 }
5074 
5075 /*
5076  * migration_call - callback that gets triggered when a CPU is added.
5077  * Here we can start up the necessary migration thread for the new CPU.
5078  */
5079 static int
5080 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5081 {
5082         int cpu = (long)hcpu;
5083         unsigned long flags;
5084         struct rq *rq = cpu_rq(cpu);
5085 
5086         switch (action & ~CPU_TASKS_FROZEN) {
5087 
5088         case CPU_UP_PREPARE:
5089                 rq->calc_load_update = calc_load_update;
5090                 break;
5091 
5092         case CPU_ONLINE:
5093                 /* Update our root-domain */
5094                 raw_spin_lock_irqsave(&rq->lock, flags);
5095                 if (rq->rd) {
5096                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5097 
5098                         set_rq_online(rq);
5099                 }
5100                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5101                 break;
5102 
5103 #ifdef CONFIG_HOTPLUG_CPU
5104         case CPU_DYING:
5105                 sched_ttwu_pending();
5106                 /* Update our root-domain */
5107                 raw_spin_lock_irqsave(&rq->lock, flags);
5108                 if (rq->rd) {
5109                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5110                         set_rq_offline(rq);
5111                 }
5112                 migrate_tasks(cpu);
5113                 BUG_ON(rq->nr_running != 1); /* the migration thread */
5114                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5115                 break;
5116 
5117         case CPU_DEAD:
5118                 calc_load_migrate(rq);
5119                 break;
5120 #endif
5121         }
5122 
5123         update_max_interval();
5124 
5125         return NOTIFY_OK;
5126 }
5127 
5128 /*
5129  * Register at high priority so that task migration (migrate_all_tasks)
5130  * happens before everything else.  This has to be lower priority than
5131  * the notifier in the perf_event subsystem, though.
5132  */
5133 static struct notifier_block migration_notifier = {
5134         .notifier_call = migration_call,
5135         .priority = CPU_PRI_MIGRATION,
5136 };
5137 
5138 static void __cpuinit set_cpu_rq_start_time(void)
5139 {
5140         int cpu = smp_processor_id();
5141         struct rq *rq = cpu_rq(cpu);
5142         rq->age_stamp = sched_clock_cpu(cpu);
5143 }
5144 
5145 static int sched_cpu_active(struct notifier_block *nfb,
5146                                       unsigned long action, void *hcpu)
5147 {
5148         switch (action & ~CPU_TASKS_FROZEN) {
5149         case CPU_STARTING:
5150                 set_cpu_rq_start_time();
5151                 return NOTIFY_OK;
5152         case CPU_DOWN_FAILED:
5153                 set_cpu_active((long)hcpu, true);
5154                 return NOTIFY_OK;
5155         default:
5156                 return NOTIFY_DONE;
5157         }
5158 }
5159 
5160 static int sched_cpu_inactive(struct notifier_block *nfb,
5161                                         unsigned long action, void *hcpu)
5162 {
5163         unsigned long flags;
5164         long cpu = (long)hcpu;
5165 
5166         switch (action & ~CPU_TASKS_FROZEN) {
5167         case CPU_DOWN_PREPARE:
5168                 set_cpu_active(cpu, false);
5169 
5170                 /* explicitly allow suspend */
5171                 if (!(action & CPU_TASKS_FROZEN)) {
5172                         struct dl_bw *dl_b = dl_bw_of(cpu);
5173                         bool overflow;
5174                         int cpus;
5175 
5176                         raw_spin_lock_irqsave(&dl_b->lock, flags);
5177                         cpus = dl_bw_cpus(cpu);
5178                         overflow = __dl_overflow(dl_b, cpus, 0, 0);
5179                         raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5180 
5181                         if (overflow)
5182                                 return notifier_from_errno(-EBUSY);
5183                 }
5184                 return NOTIFY_OK;
5185         }
5186 
5187         return NOTIFY_DONE;
5188 }
5189 
5190 static int __init migration_init(void)
5191 {
5192         void *cpu = (void *)(long)smp_processor_id();
5193         int err;
5194 
5195         /* Initialize migration for the boot CPU */
5196         err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5197         BUG_ON(err == NOTIFY_BAD);
5198         migration_call(&migration_notifier, CPU_ONLINE, cpu);
5199         register_cpu_notifier(&migration_notifier);
5200 
5201         /* Register cpu active notifiers */
5202         cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5203         cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5204 
5205         return 0;
5206 }
5207 early_initcall(migration_init);
5208 #endif
5209 
5210 #ifdef CONFIG_SMP
5211 
5212 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5213 
5214 #ifdef CONFIG_SCHED_DEBUG
5215 
5216 static __read_mostly int sched_debug_enabled;
5217 
5218 static int __init sched_debug_setup(char *str)
5219 {
5220         sched_debug_enabled = 1;
5221 
5222         return 0;
5223 }
5224 early_param("sched_debug", sched_debug_setup);
5225 
5226 static inline bool sched_debug(void)
5227 {
5228         return sched_debug_enabled;
5229 }
5230 
5231 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5232                                   struct cpumask *groupmask)
5233 {
5234         struct sched_group *group = sd->groups;
5235         char str[256];
5236 
5237         cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5238         cpumask_clear(groupmask);
5239 
5240         printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5241 
5242         if (!(sd->flags & SD_LOAD_BALANCE)) {
5243                 printk("does not load-balance\n");
5244                 if (sd->parent)
5245                         printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5246                                         " has parent");
5247                 return -1;
5248         }
5249 
5250         printk(KERN_CONT "span %s level %s\n", str, sd->name);
5251 
5252         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5253                 printk(KERN_ERR "ERROR: domain->span does not contain "
5254                                 "CPU%d\n", cpu);
5255         }
5256         if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5257                 printk(KERN_ERR "ERROR: domain->groups does not contain"
5258                                 " CPU%d\n", cpu);
5259         }
5260 
5261         printk(KERN_DEBUG "%*s groups:", level + 1, "");
5262         do {
5263                 if (!group) {
5264                         printk("\n");
5265                         printk(KERN_ERR "ERROR: group is NULL\n");
5266                         break;
5267                 }
5268 
5269                 /*
5270                  * Even though we initialize ->capacity to something semi-sane,
5271                  * we leave capacity_orig unset. This allows us to detect if
5272                  * domain iteration is still funny without causing /0 traps.
5273                  */
5274                 if (!group->sgc->capacity_orig) {
5275                         printk(KERN_CONT "\n");
5276                         printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5277                         break;
5278                 }
5279 
5280                 if (!cpumask_weight(sched_group_cpus(group))) {
5281                         printk(KERN_CONT "\n");
5282                         printk(KERN_ERR "ERROR: empty group\n");
5283                         break;
5284                 }
5285 
5286                 if (!(sd->flags & SD_OVERLAP) &&
5287                     cpumask_intersects(groupmask, sched_group_cpus(group))) {
5288                         printk(KERN_CONT "\n");
5289                         printk(KERN_ERR "ERROR: repeated CPUs\n");
5290                         break;
5291                 }
5292 
5293                 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5294 
5295                 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5296 
5297                 printk(KERN_CONT " %s", str);
5298                 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5299                         printk(KERN_CONT " (cpu_capacity = %d)",
5300                                 group->sgc->capacity);
5301                 }
5302 
5303                 group = group->next;
5304         } while (group != sd->groups);
5305         printk(KERN_CONT "\n");
5306 
5307         if (!cpumask_equal(sched_domain_span(sd), groupmask))
5308                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5309 
5310         if (sd->parent &&
5311             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5312                 printk(KERN_ERR "ERROR: parent span is not a superset "
5313                         "of domain->span\n");
5314         return 0;
5315 }
5316 
5317 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5318 {
5319         int level = 0;
5320 
5321         if (!sched_debug_enabled)
5322                 return;
5323 
5324         if (!sd) {
5325                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5326                 return;
5327         }
5328 
5329         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5330 
5331         for (;;) {
5332                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5333                         break;
5334                 level++;
5335                 sd = sd->parent;
5336                 if (!sd)
5337                         break;
5338         }
5339 }
5340 #else /* !CONFIG_SCHED_DEBUG */
5341 # define sched_domain_debug(sd, cpu) do { } while (0)
5342 static inline bool sched_debug(void)
5343 {
5344         return false;
5345 }
5346 #endif /* CONFIG_SCHED_DEBUG */
5347 
5348 static int sd_degenerate(struct sched_domain *sd)
5349 {
5350         if (cpumask_weight(sched_domain_span(sd)) == 1)
5351                 return 1;
5352 
5353         /* Following flags need at least 2 groups */
5354         if (sd->flags & (SD_LOAD_BALANCE |
5355                          SD_BALANCE_NEWIDLE |
5356                          SD_BALANCE_FORK |
5357                          SD_BALANCE_EXEC |
5358                          SD_SHARE_CPUCAPACITY |
5359                          SD_SHARE_PKG_RESOURCES |
5360                          SD_SHARE_POWERDOMAIN)) {
5361                 if (sd->groups != sd->groups->next)
5362                         return 0;
5363         }
5364 
5365         /* Following flags don't use groups */
5366         if (sd->flags & (SD_WAKE_AFFINE))
5367                 return 0;
5368 
5369         return 1;
5370 }
5371 
5372 static int
5373 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5374 {
5375         unsigned long cflags = sd->flags, pflags = parent->flags;
5376 
5377         if (sd_degenerate(parent))
5378                 return 1;
5379 
5380         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5381                 return 0;
5382 
5383         /* Flags needing groups don't count if only 1 group in parent */
5384         if (parent->groups == parent->groups->next) {
5385                 pflags &= ~(SD_LOAD_BALANCE |
5386                                 SD_BALANCE_NEWIDLE |
5387                                 SD_BALANCE_FORK |
5388                                 SD_BALANCE_EXEC |
5389                                 SD_SHARE_CPUCAPACITY |
5390                                 SD_SHARE_PKG_RESOURCES |
5391                                 SD_PREFER_SIBLING |
5392                                 SD_SHARE_POWERDOMAIN);
5393                 if (nr_node_ids == 1)
5394                         pflags &= ~SD_SERIALIZE;
5395         }
5396         if (~cflags & pflags)
5397                 return 0;
5398 
5399         return 1;
5400 }
5401 
5402 static void free_rootdomain(struct rcu_head *rcu)
5403 {
5404         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5405 
5406         cpupri_cleanup(&rd->cpupri);
5407         cpudl_cleanup(&rd->cpudl);
5408         free_cpumask_var(rd->dlo_mask);
5409         free_cpumask_var(rd->rto_mask);
5410         free_cpumask_var(rd->online);
5411         free_cpumask_var(rd->span);
5412         kfree(rd);
5413 }
5414 
5415 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5416 {
5417         struct root_domain *old_rd = NULL;
5418         unsigned long flags;
5419 
5420         raw_spin_lock_irqsave(&rq->lock, flags);
5421 
5422         if (rq->rd) {
5423                 old_rd = rq->rd;
5424 
5425                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5426                         set_rq_offline(rq);
5427 
5428                 cpumask_clear_cpu(rq->cpu, old_rd->span);
5429 
5430                 /*
5431                  * If we dont want to free the old_rd yet then
5432                  * set old_rd to NULL to skip the freeing later
5433                  * in this function:
5434                  */
5435                 if (!atomic_dec_and_test(&old_rd->refcount))
5436                         old_rd = NULL;
5437         }
5438 
5439         atomic_inc(&rd->refcount);
5440         rq->rd = rd;
5441 
5442         cpumask_set_cpu(rq->cpu, rd->span);
5443         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5444                 set_rq_online(rq);
5445 
5446         raw_spin_unlock_irqrestore(&rq->lock, flags);
5447 
5448         if (old_rd)
5449                 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5450 }
5451 
5452 static int init_rootdomain(struct root_domain *rd)
5453 {
5454         memset(rd, 0, sizeof(*rd));
5455 
5456         if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5457                 goto out;
5458         if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5459                 goto free_span;
5460         if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5461                 goto free_online;
5462         if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5463                 goto free_dlo_mask;
5464 
5465         init_dl_bw(&rd->dl_bw);
5466         if (cpudl_init(&rd->cpudl) != 0)
5467                 goto free_dlo_mask;
5468 
5469         if (cpupri_init(&rd->cpupri) != 0)
5470                 goto free_rto_mask;
5471         return 0;
5472 
5473 free_rto_mask:
5474         free_cpumask_var(rd->rto_mask);
5475 free_dlo_mask:
5476         free_cpumask_var(rd->dlo_mask);
5477 free_online:
5478         free_cpumask_var(rd->online);
5479 free_span:
5480         free_cpumask_var(rd->span);
5481 out:
5482         return -ENOMEM;
5483 }
5484 
5485 /*
5486  * By default the system creates a single root-domain with all cpus as
5487  * members (mimicking the global state we have today).
5488  */
5489 struct root_domain def_root_domain;
5490 
5491 static void init_defrootdomain(void)
5492 {
5493         init_rootdomain(&def_root_domain);
5494 
5495         atomic_set(&def_root_domain.refcount, 1);
5496 }
5497 
5498 static struct root_domain *alloc_rootdomain(void)
5499 {
5500         struct root_domain *rd;
5501 
5502         rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5503         if (!rd)
5504                 return NULL;
5505 
5506         if (init_rootdomain(rd) != 0) {
5507                 kfree(rd);
5508                 return NULL;
5509         }
5510 
5511         return rd;
5512 }
5513 
5514 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5515 {
5516         struct sched_group *tmp, *first;
5517 
5518         if (!sg)
5519                 return;
5520 
5521         first = sg;
5522         do {
5523                 tmp = sg->next;
5524 
5525                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5526                         kfree(sg->sgc);
5527 
5528                 kfree(sg);
5529                 sg = tmp;
5530         } while (sg != first);
5531 }
5532 
5533 static void free_sched_domain(struct rcu_head *rcu)
5534 {
5535         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5536 
5537         /*
5538          * If its an overlapping domain it has private groups, iterate and
5539          * nuke them all.
5540          */
5541         if (sd->flags & SD_OVERLAP) {
5542                 free_sched_groups(sd->groups, 1);
5543         } else if (atomic_dec_and_test(&sd->groups->ref)) {
5544                 kfree(sd->groups->sgc);
5545                 kfree(sd->groups);
5546         }
5547         kfree(sd);
5548 }
5549 
5550 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5551 {
5552         call_rcu(&sd->rcu, free_sched_domain);
5553 }
5554 
5555 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5556 {
5557         for (; sd; sd = sd->parent)
5558                 destroy_sched_domain(sd, cpu);
5559 }
5560 
5561 /*
5562  * Keep a special pointer to the highest sched_domain that has
5563  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5564  * allows us to avoid some pointer chasing select_idle_sibling().
5565  *
5566  * Also keep a unique ID per domain (we use the first cpu number in
5567  * the cpumask of the domain), this allows us to quickly tell if
5568  * two cpus are in the same cache domain, see cpus_share_cache().
5569  */
5570 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5571 DEFINE_PER_CPU(int, sd_llc_size);
5572 DEFINE_PER_CPU(int, sd_llc_id);
5573 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5574 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5575 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5576 
5577 static void update_top_cache_domain(int cpu)
5578 {
5579         struct sched_domain *sd;
5580         struct sched_domain *busy_sd = NULL;
5581         int id = cpu;
5582         int size = 1;
5583 
5584         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5585         if (sd) {
5586                 id = cpumask_first(sched_domain_span(sd));
5587                 size = cpumask_weight(sched_domain_span(sd));
5588                 busy_sd = sd->parent; /* sd_busy */
5589         }
5590         rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5591 
5592         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5593         per_cpu(sd_llc_size, cpu) = size;
5594         per_cpu(sd_llc_id, cpu) = id;
5595 
5596         sd = lowest_flag_domain(cpu, SD_NUMA);
5597         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5598 
5599         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5600         rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5601 }
5602 
5603 /*
5604  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5605  * hold the hotplug lock.
5606  */
5607 static void
5608 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5609 {
5610         struct rq *rq = cpu_rq(cpu);
5611         struct sched_domain *tmp;
5612 
5613         /* Remove the sched domains which do not contribute to scheduling. */
5614         for (tmp = sd; tmp; ) {
5615                 struct sched_domain *parent = tmp->parent;
5616                 if (!parent)
5617                         break;
5618 
5619                 if (sd_parent_degenerate(tmp, parent)) {
5620                         tmp->parent = parent->parent;
5621                         if (parent->parent)
5622                                 parent->parent->child = tmp;
5623                         /*
5624                          * Transfer SD_PREFER_SIBLING down in case of a
5625                          * degenerate parent; the spans match for this
5626                          * so the property transfers.
5627                          */
5628                         if (parent->flags & SD_PREFER_SIBLING)
5629                                 tmp->flags |= SD_PREFER_SIBLING;
5630                         destroy_sched_domain(parent, cpu);
5631                 } else
5632                         tmp = tmp->parent;
5633         }
5634 
5635         if (sd && sd_degenerate(sd)) {
5636                 tmp = sd;
5637                 sd = sd->parent;
5638                 destroy_sched_domain(tmp, cpu);
5639                 if (sd)
5640                         sd->child = NULL;
5641         }
5642 
5643         sched_domain_debug(sd, cpu);
5644 
5645         rq_attach_root(rq, rd);
5646         tmp = rq->sd;
5647         rcu_assign_pointer(rq->sd, sd);
5648         destroy_sched_domains(tmp, cpu);
5649 
5650         update_top_cache_domain(cpu);
5651 }
5652 
5653 /* cpus with isolated domains */
5654 static cpumask_var_t cpu_isolated_map;
5655 
5656 /* Setup the mask of cpus configured for isolated domains */
5657 static int __init isolated_cpu_setup(char *str)
5658 {
5659         alloc_bootmem_cpumask_var(&cpu_isolated_map);
5660         cpulist_parse(str, cpu_isolated_map);
5661         return 1;
5662 }
5663 
5664 __setup("isolcpus=", isolated_cpu_setup);
5665 
5666 struct s_data {
5667         struct sched_domain ** __percpu sd;
5668         struct root_domain      *rd;
5669 };
5670 
5671 enum s_alloc {
5672         sa_rootdomain,
5673         sa_sd,
5674         sa_sd_storage,
5675         sa_none,
5676 };
5677 
5678 /*
5679  * Build an iteration mask that can exclude certain CPUs from the upwards
5680  * domain traversal.
5681  *
5682  * Asymmetric node setups can result in situations where the domain tree is of
5683  * unequal depth, make sure to skip domains that already cover the entire
5684  * range.
5685  *
5686  * In that case build_sched_domains() will have terminated the iteration early
5687  * and our sibling sd spans will be empty. Domains should always include the
5688  * cpu they're built on, so check that.
5689  *
5690  */
5691 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5692 {
5693         const struct cpumask *span = sched_domain_span(sd);
5694         struct sd_data *sdd = sd->private;
5695         struct sched_domain *sibling;
5696         int i;
5697 
5698         for_each_cpu(i, span) {
5699                 sibling = *per_cpu_ptr(sdd->sd, i);
5700                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5701                         continue;
5702 
5703                 cpumask_set_cpu(i, sched_group_mask(sg));
5704         }
5705 }
5706 
5707 /*
5708  * Return the canonical balance cpu for this group, this is the first cpu
5709  * of this group that's also in the iteration mask.
5710  */
5711 int group_balance_cpu(struct sched_group *sg)
5712 {
5713         return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5714 }
5715 
5716 static int
5717 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5718 {
5719         struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5720         const struct cpumask *span = sched_domain_span(sd);
5721         struct cpumask *covered = sched_domains_tmpmask;
5722         struct sd_data *sdd = sd->private;
5723         struct sched_domain *child;
5724         int i;
5725 
5726         cpumask_clear(covered);
5727 
5728         for_each_cpu(i, span) {
5729                 struct cpumask *sg_span;
5730 
5731                 if (cpumask_test_cpu(i, covered))
5732                         continue;
5733 
5734                 child = *per_cpu_ptr(sdd->sd, i);
5735 
5736                 /* See the comment near build_group_mask(). */
5737                 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5738                         continue;
5739 
5740                 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5741                                 GFP_KERNEL, cpu_to_node(cpu));
5742 
5743                 if (!sg)
5744                         goto fail;
5745 
5746                 sg_span = sched_group_cpus(sg);
5747                 if (child->child) {
5748                         child = child->child;
5749                         cpumask_copy(sg_span, sched_domain_span(child));
5750                 } else
5751                         cpumask_set_cpu(i, sg_span);
5752 
5753                 cpumask_or(covered, covered, sg_span);
5754 
5755                 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5756                 if (atomic_inc_return(&sg->sgc->ref) == 1)
5757                         build_group_mask(sd, sg);
5758 
5759                 /*
5760                  * Initialize sgc->capacity such that even if we mess up the
5761                  * domains and no possible iteration will get us here, we won't
5762                  * die on a /0 trap.
5763                  */
5764                 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5765                 sg->sgc->capacity_orig = sg->sgc->capacity;
5766 
5767                 /*
5768                  * Make sure the first group of this domain contains the
5769                  * canonical balance cpu. Otherwise the sched_domain iteration
5770                  * breaks. See update_sg_lb_stats().
5771                  */
5772                 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5773                     group_balance_cpu(sg) == cpu)
5774                         groups = sg;
5775 
5776                 if (!first)
5777                         first = sg;
5778                 if (last)
5779                         last->next = sg;
5780                 last = sg;
5781                 last->next = first;
5782         }
5783         sd->groups = groups;
5784 
5785         return 0;
5786 
5787 fail:
5788         free_sched_groups(first, 0);
5789 
5790         return -ENOMEM;
5791 }
5792 
5793 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5794 {
5795         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5796         struct sched_domain *child = sd->child;
5797 
5798         if (child)
5799                 cpu = cpumask_first(sched_domain_span(child));
5800 
5801         if (sg) {
5802                 *sg = *per_cpu_ptr(sdd->sg, cpu);
5803                 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5804                 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5805         }
5806 
5807         return cpu;
5808 }
5809 
5810 /*
5811  * build_sched_groups will build a circular linked list of the groups
5812  * covered by the given span, and will set each group's ->cpumask correctly,
5813  * and ->cpu_capacity to 0.
5814  *
5815  * Assumes the sched_domain tree is fully constructed
5816  */
5817 static int
5818 build_sched_groups(struct sched_domain *sd, int cpu)
5819 {
5820         struct sched_group *first = NULL, *last = NULL;
5821         struct sd_data *sdd = sd->private;
5822         const struct cpumask *span = sched_domain_span(sd);
5823         struct cpumask *covered;
5824         int i;
5825 
5826         get_group(cpu, sdd, &sd->groups);
5827         atomic_inc(&sd->groups->ref);
5828 
5829         if (cpu != cpumask_first(span))
5830                 return 0;
5831 
5832         lockdep_assert_held(&sched_domains_mutex);
5833         covered = sched_domains_tmpmask;
5834 
5835         cpumask_clear(covered);
5836 
5837         for_each_cpu(i, span) {
5838                 struct sched_group *sg;
5839                 int group, j;
5840 
5841                 if (cpumask_test_cpu(i, covered))
5842                         continue;
5843 
5844                 group = get_group(i, sdd, &sg);
5845                 cpumask_setall(sched_group_mask(sg));
5846 
5847                 for_each_cpu(j, span) {
5848                         if (get_group(j, sdd, NULL) != group)
5849                                 continue;
5850 
5851                         cpumask_set_cpu(j, covered);
5852                         cpumask_set_cpu(j, sched_group_cpus(sg));
5853                 }
5854 
5855                 if (!first)
5856                         first = sg;
5857                 if (last)
5858                         last->next = sg;
5859                 last = sg;
5860         }
5861         last->next = first;
5862 
5863         return 0;
5864 }
5865 
5866 /*
5867  * Initialize sched groups cpu_capacity.
5868  *
5869  * cpu_capacity indicates the capacity of sched group, which is used while
5870  * distributing the load between different sched groups in a sched domain.
5871  * Typically cpu_capacity for all the groups in a sched domain will be same
5872  * unless there are asymmetries in the topology. If there are asymmetries,
5873  * group having more cpu_capacity will pickup more load compared to the
5874  * group having less cpu_capacity.
5875  */
5876 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
5877 {
5878         struct sched_group *sg = sd->groups;
5879 
5880         WARN_ON(!sg);
5881 
5882         do {
5883                 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5884                 sg = sg->next;
5885         } while (sg != sd->groups);
5886 
5887         if (cpu != group_balance_cpu(sg))
5888                 return;
5889 
5890         update_group_capacity(sd, cpu);
5891         atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
5892 }
5893 
5894 /*
5895  * Initializers for schedule domains
5896  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5897  */
5898 
5899 static int default_relax_domain_level = -1;
5900 int sched_domain_level_max;
5901 
5902 static int __init setup_relax_domain_level(char *str)
5903 {
5904         if (kstrtoint(str, 0, &default_relax_domain_level))
5905                 pr_warn("Unable to set relax_domain_level\n");
5906 
5907         return 1;
5908 }
5909 __setup("relax_domain_level=", setup_relax_domain_level);
5910 
5911 static void set_domain_attribute(struct sched_domain *sd,
5912                                  struct sched_domain_attr *attr)
5913 {
5914         int request;
5915 
5916         if (!attr || attr->relax_domain_level < 0) {
5917                 if (default_relax_domain_level < 0)
5918                         return;
5919                 else
5920                         request = default_relax_domain_level;
5921         } else
5922                 request = attr->relax_domain_level;
5923         if (request < sd->level) {
5924                 /* turn off idle balance on this domain */
5925                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5926         } else {
5927                 /* turn on idle balance on this domain */
5928                 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5929         }
5930 }
5931 
5932 static void __sdt_free(const struct cpumask *cpu_map);
5933 static int __sdt_alloc(const struct cpumask *cpu_map);
5934 
5935 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5936                                  const struct cpumask *cpu_map)
5937 {
5938         switch (what) {
5939         case sa_rootdomain:
5940                 if (!atomic_read(&d->rd->refcount))
5941                         free_rootdomain(&d->rd->rcu); /* fall through */
5942         case sa_sd:
5943                 free_percpu(d->sd); /* fall through */
5944         case sa_sd_storage:
5945                 __sdt_free(cpu_map); /* fall through */
5946         case sa_none:
5947                 break;
5948         }
5949 }
5950 
5951 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5952                                                    const struct cpumask *cpu_map)
5953 {
5954         memset(d, 0, sizeof(*d));
5955 
5956         if (__sdt_alloc(cpu_map))
5957                 return sa_sd_storage;
5958         d->sd = alloc_percpu(struct sched_domain *);
5959         if (!d->sd)
5960                 return sa_sd_storage;
5961         d->rd = alloc_rootdomain();
5962         if (!d->rd)
5963                 return sa_sd;
5964         return sa_rootdomain;
5965 }
5966 
5967 /*
5968  * NULL the sd_data elements we've used to build the sched_domain and
5969  * sched_group structure so that the subsequent __free_domain_allocs()
5970  * will not free the data we're using.
5971  */
5972 static void claim_allocations(int cpu, struct sched_domain *sd)
5973 {
5974         struct sd_data *sdd = sd->private;
5975 
5976         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5977         *per_cpu_ptr(sdd->sd, cpu) = NULL;
5978 
5979         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5980                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5981 
5982         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
5983                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
5984 }
5985 
5986 #ifdef CONFIG_NUMA
5987 static int sched_domains_numa_levels;
5988 static int *sched_domains_numa_distance;
5989 static struct cpumask ***sched_domains_numa_masks;
5990 static int sched_domains_curr_level;
5991 #endif
5992 
5993 /*
5994  * SD_flags allowed in topology descriptions.
5995  *
5996  * SD_SHARE_CPUCAPACITY      - describes SMT topologies
5997  * SD_SHARE_PKG_RESOURCES - describes shared caches
5998  * SD_NUMA                - describes NUMA topologies
5999  * SD_SHARE_POWERDOMAIN   - describes shared power domain
6000  *
6001  * Odd one out:
6002  * SD_ASYM_PACKING        - describes SMT quirks
6003  */
6004 #define TOPOLOGY_SD_FLAGS               \
6005         (SD_SHARE_CPUCAPACITY |         \
6006          SD_SHARE_PKG_RESOURCES |       \
6007          SD_NUMA |                      \
6008          SD_ASYM_PACKING |              \
6009          SD_SHARE_POWERDOMAIN)
6010 
6011 static struct sched_domain *
6012 sd_init(struct sched_domain_topology_level *tl, int cpu)
6013 {
6014         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6015         int sd_weight, sd_flags = 0;
6016 
6017 #ifdef CONFIG_NUMA
6018         /*
6019          * Ugly hack to pass state to sd_numa_mask()...
6020          */
6021         sched_domains_curr_level = tl->numa_level;
6022 #endif
6023 
6024         sd_weight = cpumask_weight(tl->mask(cpu));
6025 
6026         if (tl->sd_flags)
6027                 sd_flags = (*tl->sd_flags)();
6028         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6029                         "wrong sd_flags in topology description\n"))
6030                 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6031 
6032         *sd = (struct sched_domain){
6033                 .min_interval           = sd_weight,
6034                 .max_interval           = 2*sd_weight,
6035                 .busy_factor            = 32,
6036                 .imbalance_pct          = 125,
6037 
6038                 .cache_nice_tries       = 0,
6039                 .busy_idx               = 0,
6040                 .idle_idx               = 0,
6041                 .newidle_idx            = 0,
6042                 .wake_idx               = 0,
6043                 .forkexec_idx           = 0,
6044 
6045                 .flags                  = 1*SD_LOAD_BALANCE
6046                                         | 1*SD_BALANCE_NEWIDLE
6047                                         | 1*SD_BALANCE_EXEC
6048                                         | 1*SD_BALANCE_FORK
6049                                         | 0*SD_BALANCE_WAKE
6050                                         | 1*SD_WAKE_AFFINE
6051                                         | 0*SD_SHARE_CPUCAPACITY
6052                                         | 0*SD_SHARE_PKG_RESOURCES
6053                                         | 0*SD_SERIALIZE
6054                                         | 0*SD_PREFER_SIBLING
6055                                         | 0*SD_NUMA
6056                                         | sd_flags
6057                                         ,
6058 
6059                 .last_balance           = jiffies,
6060                 .balance_interval       = sd_weight,
6061                 .smt_gain               = 0,
6062                 .max_newidle_lb_cost    = 0,
6063                 .next_decay_max_lb_cost = jiffies,
6064 #ifdef CONFIG_SCHED_DEBUG
6065                 .name                   = tl->name,
6066 #endif
6067         };
6068 
6069         /*
6070          * Convert topological properties into behaviour.
6071          */
6072 
6073         if (sd->flags & SD_SHARE_CPUCAPACITY) {
6074                 sd->imbalance_pct = 110;
6075                 sd->smt_gain = 1178; /* ~15% */
6076 
6077         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6078                 sd->imbalance_pct = 117;
6079                 sd->cache_nice_tries = 1;
6080                 sd->busy_idx = 2;
6081 
6082 #ifdef CONFIG_NUMA
6083         } else if (sd->flags & SD_NUMA) {
6084                 sd->cache_nice_tries = 2;
6085                 sd->busy_idx = 3;
6086                 sd->idle_idx = 2;
6087 
6088                 sd->flags |= SD_SERIALIZE;
6089                 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6090                         sd->flags &= ~(SD_BALANCE_EXEC |
6091                                        SD_BALANCE_FORK |
6092                                        SD_WAKE_AFFINE);
6093                 }
6094 
6095 #endif
6096         } else {
6097                 sd->flags |= SD_PREFER_SIBLING;
6098                 sd->cache_nice_tries = 1;
6099                 sd->busy_idx = 2;
6100                 sd->idle_idx = 1;
6101         }
6102 
6103         sd->private = &tl->data;
6104 
6105         return sd;
6106 }
6107 
6108 /*
6109  * Topology list, bottom-up.
6110  */
6111 static struct sched_domain_topology_level default_topology[] = {
6112 #ifdef CONFIG_SCHED_SMT
6113         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6114 #endif
6115 #ifdef CONFIG_SCHED_MC
6116         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6117 #endif
6118         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6119         { NULL, },
6120 };
6121 
6122 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6123 
6124 #define for_each_sd_topology(tl)                        \
6125         for (tl = sched_domain_topology; tl->mask; tl++)
6126 
6127 void set_sched_topology(struct sched_domain_topology_level *tl)
6128 {
6129         sched_domain_topology = tl;
6130 }
6131 
6132 #ifdef CONFIG_NUMA
6133 
6134 static const struct cpumask *sd_numa_mask(int cpu)
6135 {
6136         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6137 }
6138 
6139 static void sched_numa_warn(const char *str)
6140 {
6141         static int done = false;
6142         int i,j;
6143 
6144         if (done)
6145                 return;
6146 
6147         done = true;
6148 
6149         printk(KERN_WARNING "ERROR: %s\n\n", str);
6150 
6151         for (i = 0; i < nr_node_ids; i++) {
6152                 printk(KERN_WARNING "  ");
6153                 for (j = 0; j < nr_node_ids; j++)
6154                         printk(KERN_CONT "%02d ", node_distance(i,j));
6155                 printk(KERN_CONT "\n");
6156         }
6157         printk(KERN_WARNING "\n");
6158 }
6159 
6160 static bool find_numa_distance(int distance)
6161 {
6162         int i;
6163 
6164         if (distance == node_distance(0, 0))
6165                 return true;
6166 
6167         for (i = 0; i < sched_domains_numa_levels; i++) {
6168                 if (sched_domains_numa_distance[i] == distance)
6169                         return true;
6170         }
6171 
6172         return false;
6173 }
6174 
6175 static void sched_init_numa(void)
6176 {
6177         int next_distance, curr_distance = node_distance(0, 0);
6178         struct sched_domain_topology_level *tl;
6179         int level = 0;
6180         int i, j, k;
6181 
6182         sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6183         if (!sched_domains_numa_distance)
6184                 return;
6185 
6186         /*
6187          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6188          * unique distances in the node_distance() table.
6189          *
6190          * Assumes node_distance(0,j) includes all distances in
6191          * node_distance(i,j) in order to avoid cubic time.
6192          */
6193         next_distance = curr_distance;
6194         for (i = 0; i < nr_node_ids; i++) {
6195                 for (j = 0; j < nr_node_ids; j++) {
6196                         for (k = 0; k < nr_node_ids; k++) {
6197                                 int distance = node_distance(i, k);
6198 
6199                                 if (distance > curr_distance &&
6200                                     (distance < next_distance ||
6201                                      next_distance == curr_distance))
6202                                         next_distance = distance;
6203 
6204                                 /*
6205                                  * While not a strong assumption it would be nice to know
6206                                  * about cases where if node A is connected to B, B is not
6207                                  * equally connected to A.
6208                                  */
6209                                 if (sched_debug() && node_distance(k, i) != distance)
6210                                         sched_numa_warn("Node-distance not symmetric");
6211 
6212                                 if (sched_debug() && i && !find_numa_distance(distance))
6213                                         sched_numa_warn("Node-0 not representative");
6214                         }
6215                         if (next_distance != curr_distance) {
6216                                 sched_domains_numa_distance[level++] = next_distance;
6217                                 sched_domains_numa_levels = level;
6218                                 curr_distance = next_distance;
6219                         } else break;
6220                 }
6221 
6222                 /*
6223                  * In case of sched_debug() we verify the above assumption.
6224                  */
6225                 if (!sched_debug())
6226                         break;
6227         }
6228         /*
6229          * 'level' contains the number of unique distances, excluding the
6230          * identity distance node_distance(i,i).
6231          *
6232          * The sched_domains_numa_distance[] array includes the actual distance
6233          * numbers.
6234          */
6235 
6236         /*
6237          * Here, we should temporarily reset sched_domains_numa_levels to 0.
6238          * If it fails to allocate memory for array sched_domains_numa_masks[][],
6239          * the array will contain less then 'level' members. This could be
6240          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6241          * in other functions.
6242          *
6243          * We reset it to 'level' at the end of this function.
6244          */
6245         sched_domains_numa_levels = 0;
6246 
6247         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6248         if (!sched_domains_numa_masks)
6249                 return;
6250 
6251         /*
6252          * Now for each level, construct a mask per node which contains all
6253          * cpus of nodes that are that many hops away from us.
6254          */
6255         for (i = 0; i < level; i++) {
6256                 sched_domains_numa_masks[i] =
6257                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6258                 if (!sched_domains_numa_masks[i])
6259                         return;
6260 
6261                 for (j = 0; j < nr_node_ids; j++) {
6262                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6263                         if (!mask)
6264                                 return;
6265 
6266                         sched_domains_numa_masks[i][j] = mask;
6267 
6268                         for (k = 0; k < nr_node_ids; k++) {
6269                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
6270                                         continue;
6271 
6272                                 cpumask_or(mask, mask, cpumask_of_node(k));
6273                         }
6274                 }
6275         }
6276 
6277         /* Compute default topology size */
6278         for (i = 0; sched_domain_topology[i].mask; i++);
6279 
6280         tl = kzalloc((i + level + 1) *
6281                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6282         if (!tl)
6283                 return;
6284 
6285         /*
6286          * Copy the default topology bits..
6287          */
6288         for (i = 0; sched_domain_topology[i].mask; i++)
6289                 tl[i] = sched_domain_topology[i];
6290 
6291         /*
6292          * .. and append 'j' levels of NUMA goodness.
6293          */
6294         for (j = 0; j < level; i++, j++) {
6295                 tl[i] = (struct sched_domain_topology_level){
6296                         .mask = sd_numa_mask,
6297                         .sd_flags = cpu_numa_flags,
6298                         .flags = SDTL_OVERLAP,
6299                         .numa_level = j,
6300                         SD_INIT_NAME(NUMA)
6301                 };
6302         }
6303 
6304         sched_domain_topology = tl;
6305 
6306         sched_domains_numa_levels = level;
6307 }
6308 
6309 static void sched_domains_numa_masks_set(int cpu)
6310 {
6311         int i, j;
6312         int node = cpu_to_node(cpu);
6313 
6314         for (i = 0; i < sched_domains_numa_levels; i++) {
6315                 for (j = 0; j < nr_node_ids; j++) {
6316                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
6317                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6318                 }
6319         }
6320 }
6321 
6322 static void sched_domains_numa_masks_clear(int cpu)
6323 {
6324         int i, j;
6325         for (i = 0; i < sched_domains_numa_levels; i++) {
6326                 for (j = 0; j < nr_node_ids; j++)
6327                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6328         }
6329 }
6330 
6331 /*
6332  * Update sched_domains_numa_masks[level][node] array when new cpus
6333  * are onlined.
6334  */
6335 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6336                                            unsigned long action,
6337                                            void *hcpu)
6338 {
6339         int cpu = (long)hcpu;
6340 
6341         switch (action & ~CPU_TASKS_FROZEN) {
6342         case CPU_ONLINE:
6343                 sched_domains_numa_masks_set(cpu);
6344                 break;
6345 
6346         case CPU_DEAD:
6347                 sched_domains_numa_masks_clear(cpu);
6348                 break;
6349 
6350         default:
6351                 return NOTIFY_DONE;
6352         }
6353 
6354         return NOTIFY_OK;
6355 }
6356 #else
6357 static inline void sched_init_numa(void)
6358 {
6359 }
6360 
6361 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6362                                            unsigned long action,
6363                                            void *hcpu)
6364 {
6365         return 0;
6366 }
6367 #endif /* CONFIG_NUMA */
6368 
6369 static int __sdt_alloc(const struct cpumask *cpu_map)
6370 {
6371         struct sched_domain_topology_level *tl;
6372         int j;
6373 
6374         for_each_sd_topology(tl) {
6375                 struct sd_data *sdd = &tl->data;
6376 
6377                 sdd->sd = alloc_percpu(struct sched_domain *);
6378                 if (!sdd->sd)
6379                         return -ENOMEM;
6380 
6381                 sdd->sg = alloc_percpu(struct sched_group *);
6382                 if (!sdd->sg)
6383                         return -ENOMEM;
6384 
6385                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6386                 if (!sdd->sgc)
6387                         return -ENOMEM;
6388 
6389                 for_each_cpu(j, cpu_map) {
6390                         struct sched_domain *sd;
6391                         struct sched_group *sg;
6392                         struct sched_group_capacity *sgc;
6393 
6394                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6395                                         GFP_KERNEL, cpu_to_node(j));
6396                         if (!sd)
6397                                 return -ENOMEM;
6398 
6399                         *per_cpu_ptr(sdd->sd, j) = sd;
6400 
6401                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6402                                         GFP_KERNEL, cpu_to_node(j));
6403                         if (!sg)
6404                                 return -ENOMEM;
6405 
6406                         sg->next = sg;
6407 
6408                         *per_cpu_ptr(sdd->sg, j) = sg;
6409 
6410                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6411                                         GFP_KERNEL, cpu_to_node(j));
6412                         if (!sgc)
6413                                 return -ENOMEM;
6414 
6415                         *per_cpu_ptr(sdd->sgc, j) = sgc;
6416                 }
6417         }
6418 
6419         return 0;
6420 }
6421 
6422 static void __sdt_free(const struct cpumask *cpu_map)
6423 {
6424         struct sched_domain_topology_level *tl;
6425         int j;
6426 
6427         for_each_sd_topology(tl) {
6428                 struct sd_data *sdd = &tl->data;
6429 
6430                 for_each_cpu(j, cpu_map) {
6431                         struct sched_domain *sd;
6432 
6433                         if (sdd->sd) {
6434                                 sd = *per_cpu_ptr(sdd->sd, j);
6435                                 if (sd && (sd->flags & SD_OVERLAP))
6436                                         free_sched_groups(sd->groups, 0);
6437                                 kfree(*per_cpu_ptr(sdd->sd, j));
6438                         }
6439 
6440                         if (sdd->sg)
6441                                 kfree(*per_cpu_ptr(sdd->sg, j));
6442                         if (sdd->sgc)
6443                                 kfree(*per_cpu_ptr(sdd->sgc, j));
6444                 }
6445                 free_percpu(sdd->sd);
6446                 sdd->sd = NULL;
6447                 free_percpu(sdd->sg);
6448                 sdd->sg = NULL;
6449                 free_percpu(sdd->sgc);
6450                 sdd->sgc = NULL;
6451         }
6452 }
6453 
6454 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6455                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6456                 struct sched_domain *child, int cpu)
6457 {
6458         struct sched_domain *sd = sd_init(tl, cpu);
6459         if (!sd)
6460                 return child;
6461 
6462         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6463         if (child) {
6464                 sd->level = child->level + 1;
6465                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6466                 child->parent = sd;
6467                 sd->child = child;
6468         }
6469         set_domain_attribute(sd, attr);
6470 
6471         return sd;
6472 }
6473 
6474 /*
6475  * Build sched domains for a given set of cpus and attach the sched domains
6476  * to the individual cpus
6477  */
6478 static int build_sched_domains(const struct cpumask *cpu_map,
6479                                struct sched_domain_attr *attr)
6480 {
6481         enum s_alloc alloc_state;
6482         struct sched_domain *sd;
6483         struct s_data d;
6484         int i, ret = -ENOMEM;
6485 
6486         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6487         if (alloc_state != sa_rootdomain)
6488                 goto error;
6489 
6490         /* Set up domains for cpus specified by the cpu_map. */
6491         for_each_cpu(i, cpu_map) {
6492                 struct sched_domain_topology_level *tl;
6493 
6494                 sd = NULL;
6495                 for_each_sd_topology(tl) {
6496                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6497                         if (tl == sched_domain_topology)
6498                                 *per_cpu_ptr(d.sd, i) = sd;
6499                         if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6500                                 sd->flags |= SD_OVERLAP;
6501                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6502                                 break;
6503                 }
6504         }
6505 
6506         /* Build the groups for the domains */
6507         for_each_cpu(i, cpu_map) {
6508                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6509                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
6510                         if (sd->flags & SD_OVERLAP) {
6511                                 if (build_overlap_sched_groups(sd, i))
6512                                         goto error;
6513                         } else {
6514                                 if (build_sched_groups(sd, i))
6515                                         goto error;
6516                         }
6517                 }
6518         }
6519 
6520         /* Calculate CPU capacity for physical packages and nodes */
6521         for (i = nr_cpumask_bits-1; i >= 0; i--) {
6522                 if (!cpumask_test_cpu(i, cpu_map))
6523                         continue;
6524 
6525                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6526                         claim_allocations(i, sd);
6527                         init_sched_groups_capacity(i, sd);
6528                 }
6529         }
6530 
6531         /* Attach the domains */
6532         rcu_read_lock();
6533         for_each_cpu(i, cpu_map) {
6534                 sd = *per_cpu_ptr(d.sd, i);
6535                 cpu_attach_domain(sd, d.rd, i);
6536         }
6537         rcu_read_unlock();
6538 
6539         ret = 0;
6540 error:
6541         __free_domain_allocs(&d, alloc_state, cpu_map);
6542         return ret;
6543 }
6544 
6545 static cpumask_var_t *doms_cur; /* current sched domains */
6546 static int ndoms_cur;           /* number of sched domains in 'doms_cur' */
6547 static struct sched_domain_attr *dattr_cur;
6548                                 /* attribues of custom domains in 'doms_cur' */
6549 
6550 /*
6551  * Special case: If a kmalloc of a doms_cur partition (array of
6552  * cpumask) fails, then fallback to a single sched domain,
6553  * as determined by the single cpumask fallback_doms.
6554  */
6555 static cpumask_var_t fallback_doms;
6556 
6557 /*
6558  * arch_update_cpu_topology lets virtualized architectures update the
6559  * cpu core maps. It is supposed to return 1 if the topology changed
6560  * or 0 if it stayed the same.
6561  */
6562 int __weak arch_update_cpu_topology(void)
6563 {
6564         return 0;
6565 }
6566 
6567 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6568 {
6569         int i;
6570         cpumask_var_t *doms;
6571 
6572         doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6573         if (!doms)
6574                 return NULL;
6575         for (i = 0; i < ndoms; i++) {
6576                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6577                         free_sched_domains(doms, i);
6578                         return NULL;
6579                 }
6580         }
6581         return doms;
6582 }
6583 
6584 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6585 {
6586         unsigned int i;
6587         for (i = 0; i < ndoms; i++)
6588                 free_cpumask_var(doms[i]);
6589         kfree(doms);
6590 }
6591 
6592 /*
6593  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6594  * For now this just excludes isolated cpus, but could be used to
6595  * exclude other special cases in the future.
6596  */
6597 static int init_sched_domains(const struct cpumask *cpu_map)
6598 {
6599         int err;
6600 
6601         arch_update_cpu_topology();
6602         ndoms_cur = 1;
6603         doms_cur = alloc_sched_domains(ndoms_cur);
6604         if (!doms_cur)
6605                 doms_cur = &fallback_doms;
6606         cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6607         err = build_sched_domains(doms_cur[0], NULL);
6608         register_sched_domain_sysctl();
6609 
6610         return err;
6611 }
6612 
6613 /*
6614  * Detach sched domains from a group of cpus specified in cpu_map
6615  * These cpus will now be attached to the NULL domain
6616  */
6617 static void detach_destroy_domains(const struct cpumask *cpu_map)
6618 {
6619         int i;
6620 
6621         rcu_read_lock();
6622         for_each_cpu(i, cpu_map)
6623                 cpu_attach_domain(NULL, &def_root_domain, i);
6624         rcu_read_unlock();
6625 }
6626 
6627 /* handle null as "default" */
6628 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6629                         struct sched_domain_attr *new, int idx_new)
6630 {
6631         struct sched_domain_attr tmp;
6632 
6633         /* fast path */
6634         if (!new && !cur)
6635                 return 1;
6636 
6637         tmp = SD_ATTR_INIT;
6638         return !memcmp(cur ? (cur + idx_cur) : &tmp,
6639                         new ? (new + idx_new) : &tmp,
6640                         sizeof(struct sched_domain_attr));
6641 }
6642 
6643 /*
6644  * Partition sched domains as specified by the 'ndoms_new'
6645  * cpumasks in the array doms_new[] of cpumasks. This compares
6646  * doms_new[] to the current sched domain partitioning, doms_cur[].
6647  * It destroys each deleted domain and builds each new domain.
6648  *
6649  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6650  * The masks don't intersect (don't overlap.) We should setup one
6651  * sched domain for each mask. CPUs not in any of the cpumasks will
6652  * not be load balanced. If the same cpumask appears both in the
6653  * current 'doms_cur' domains and in the new 'doms_new', we can leave
6654  * it as it is.
6655  *
6656  * The passed in 'doms_new' should be allocated using
6657  * alloc_sched_domains.  This routine takes ownership of it and will
6658  * free_sched_domains it when done with it. If the caller failed the
6659  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6660  * and partition_sched_domains() will fallback to the single partition
6661  * 'fallback_doms', it also forces the domains to be rebuilt.
6662  *
6663  * If doms_new == NULL it will be replaced with cpu_online_mask.
6664  * ndoms_new == 0 is a special case for destroying existing domains,
6665  * and it will not create the default domain.
6666  *
6667  * Call with hotplug lock held
6668  */
6669 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6670                              struct sched_domain_attr *dattr_new)
6671 {
6672         int i, j, n;
6673         int new_topology;
6674 
6675         mutex_lock(&sched_domains_mutex);
6676 
6677         /* always unregister in case we don't destroy any domains */
6678         unregister_sched_domain_sysctl();
6679 
6680         /* Let architecture update cpu core mappings. */
6681         new_topology = arch_update_cpu_topology();
6682 
6683         n = doms_new ? ndoms_new : 0;
6684 
6685         /* Destroy deleted domains */
6686         for (i = 0; i < ndoms_cur; i++) {
6687                 for (j = 0; j < n && !new_topology; j++) {
6688                         if (cpumask_equal(doms_cur[i], doms_new[j])
6689                             && dattrs_equal(dattr_cur, i, dattr_new, j))
6690                                 goto match1;
6691                 }
6692                 /* no match - a current sched domain not in new doms_new[] */
6693                 detach_destroy_domains(doms_cur[i]);
6694 match1:
6695                 ;
6696         }
6697 
6698         n = ndoms_cur;
6699         if (doms_new == NULL) {
6700                 n = 0;
6701                 doms_new = &fallback_doms;
6702                 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6703                 WARN_ON_ONCE(dattr_new);
6704         }
6705 
6706         /* Build new domains */
6707         for (i = 0; i < ndoms_new; i++) {
6708                 for (j = 0; j < n && !new_topology; j++) {
6709                         if (cpumask_equal(doms_new[i], doms_cur[j])
6710                             && dattrs_equal(dattr_new, i, dattr_cur, j))
6711                                 goto match2;
6712                 }
6713                 /* no match - add a new doms_new */
6714                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6715 match2:
6716                 ;
6717         }
6718 
6719         /* Remember the new sched domains */
6720         if (doms_cur != &fallback_doms)
6721                 free_sched_domains(doms_cur, ndoms_cur);
6722         kfree(dattr_cur);       /* kfree(NULL) is safe */
6723         doms_cur = doms_new;
6724         dattr_cur = dattr_new;
6725         ndoms_cur = ndoms_new;
6726 
6727         register_sched_domain_sysctl();
6728 
6729         mutex_unlock(&sched_domains_mutex);
6730 }
6731 
6732 static int num_cpus_frozen;     /* used to mark begin/end of suspend/resume */
6733 
6734 /*
6735  * Update cpusets according to cpu_active mask.  If cpusets are
6736  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6737  * around partition_sched_domains().
6738  *
6739  * If we come here as part of a suspend/resume, don't touch cpusets because we
6740  * want to restore it back to its original state upon resume anyway.
6741  */
6742 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6743                              void *hcpu)
6744 {
6745         switch (action) {
6746         case CPU_ONLINE_FROZEN:
6747         case CPU_DOWN_FAILED_FROZEN:
6748 
6749                 /*
6750                  * num_cpus_frozen tracks how many CPUs are involved in suspend
6751                  * resume sequence. As long as this is not the last online
6752                  * operation in the resume sequence, just build a single sched
6753                  * domain, ignoring cpusets.
6754                  */
6755                 num_cpus_frozen--;
6756                 if (likely(num_cpus_frozen)) {
6757                         partition_sched_domains(1, NULL, NULL);
6758                         break;
6759                 }
6760 
6761                 /*
6762                  * This is the last CPU online operation. So fall through and
6763                  * restore the original sched domains by considering the
6764                  * cpuset configurations.
6765                  */
6766 
6767         case CPU_ONLINE:
6768         case CPU_DOWN_FAILED:
6769                 cpuset_update_active_cpus(true);
6770                 break;
6771         default:
6772                 return NOTIFY_DONE;
6773         }
6774         return NOTIFY_OK;
6775 }
6776 
6777 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6778                                void *hcpu)
6779 {
6780         switch (action) {
6781         case CPU_DOWN_PREPARE:
6782                 cpuset_update_active_cpus(false);
6783                 break;
6784         case CPU_DOWN_PREPARE_FROZEN:
6785                 num_cpus_frozen++;
6786                 partition_sched_domains(1, NULL, NULL);
6787                 break;
6788         default:
6789                 return NOTIFY_DONE;
6790         }
6791         return NOTIFY_OK;
6792 }
6793 
6794 void __init sched_init_smp(void)
6795 {
6796         cpumask_var_t non_isolated_cpus;
6797 
6798         alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6799         alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6800 
6801         sched_init_numa();
6802 
6803         /*
6804          * There's no userspace yet to cause hotplug operations; hence all the
6805          * cpu masks are stable and all blatant races in the below code cannot
6806          * happen.
6807          */
6808         mutex_lock(&sched_domains_mutex);
6809         init_sched_domains(cpu_active_mask);
6810         cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6811         if (cpumask_empty(non_isolated_cpus))
6812                 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6813         mutex_unlock(&sched_domains_mutex);
6814 
6815         hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6816         hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6817         hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6818 
6819         init_hrtick();
6820 
6821         /* Move init over to a non-isolated CPU */
6822         if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6823                 BUG();
6824         sched_init_granularity();
6825         free_cpumask_var(non_isolated_cpus);
6826 
6827         init_sched_rt_class();
6828         init_sched_dl_class();
6829 }
6830 #else
6831 void __init sched_init_smp(void)
6832 {
6833         sched_init_granularity();
6834 }
6835 #endif /* CONFIG_SMP */
6836 
6837 const_debug unsigned int sysctl_timer_migration = 1;
6838 
6839 int in_sched_functions(unsigned long addr)
6840 {
6841         return in_lock_functions(addr) ||
6842                 (addr >= (unsigned long)__sched_text_start
6843                 && addr < (unsigned long)__sched_text_end);
6844 }
6845 
6846 #ifdef CONFIG_CGROUP_SCHED
6847 /*
6848  * Default task group.
6849  * Every task in system belongs to this group at bootup.
6850  */
6851 struct task_group root_task_group;
6852 LIST_HEAD(task_groups);
6853 #endif
6854 
6855 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6856 
6857 void __init sched_init(void)
6858 {
6859         int i, j;
6860         unsigned long alloc_size = 0, ptr;
6861 
6862 #ifdef CONFIG_FAIR_GROUP_SCHED
6863         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6864 #endif
6865 #ifdef CONFIG_RT_GROUP_SCHED
6866         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6867 #endif
6868 #ifdef CONFIG_CPUMASK_OFFSTACK
6869         alloc_size += num_possible_cpus() * cpumask_size();
6870 #endif
6871         if (alloc_size) {
6872                 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6873 
6874 #ifdef CONFIG_FAIR_GROUP_SCHED
6875                 root_task_group.se = (struct sched_entity **)ptr;
6876                 ptr += nr_cpu_ids * sizeof(void **);
6877 
6878                 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6879                 ptr += nr_cpu_ids * sizeof(void **);
6880 
6881 #endif /* CONFIG_FAIR_GROUP_SCHED */
6882 #ifdef CONFIG_RT_GROUP_SCHED
6883                 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6884                 ptr += nr_cpu_ids * sizeof(void **);
6885 
6886                 root_task_group.rt_rq = (struct rt_rq **)ptr;
6887                 ptr += nr_cpu_ids * sizeof(void **);
6888 
6889 #endif /* CONFIG_RT_GROUP_SCHED */
6890 #ifdef CONFIG_CPUMASK_OFFSTACK
6891                 for_each_possible_cpu(i) {
6892                         per_cpu(load_balance_mask, i) = (void *)ptr;
6893                         ptr += cpumask_size();
6894                 }
6895 #endif /* CONFIG_CPUMASK_OFFSTACK */
6896         }
6897 
6898         init_rt_bandwidth(&def_rt_bandwidth,
6899                         global_rt_period(), global_rt_runtime());
6900         init_dl_bandwidth(&def_dl_bandwidth,
6901                         global_rt_period(), global_rt_runtime());
6902 
6903 #ifdef CONFIG_SMP
6904         init_defrootdomain();
6905 #endif
6906 
6907 #ifdef CONFIG_RT_GROUP_SCHED
6908         init_rt_bandwidth(&root_task_group.rt_bandwidth,
6909                         global_rt_period(), global_rt_runtime());
6910 #endif /* CONFIG_RT_GROUP_SCHED */
6911 
6912 #ifdef CONFIG_CGROUP_SCHED
6913         list_add(&root_task_group.list, &task_groups);
6914         INIT_LIST_HEAD(&root_task_group.children);
6915         INIT_LIST_HEAD(&root_task_group.siblings);
6916         autogroup_init(&init_task);
6917 
6918 #endif /* CONFIG_CGROUP_SCHED */
6919 
6920         for_each_possible_cpu(i) {
6921                 struct rq *rq;
6922 
6923                 rq = cpu_rq(i);
6924                 raw_spin_lock_init(&rq->lock);
6925                 rq->nr_running = 0;
6926                 rq->calc_load_active = 0;
6927                 rq->calc_load_update = jiffies + LOAD_FREQ;
6928                 init_cfs_rq(&rq->cfs);
6929                 init_rt_rq(&rq->rt, rq);
6930                 init_dl_rq(&rq->dl, rq);
6931 #ifdef CONFIG_FAIR_GROUP_SCHED
6932                 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6933                 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6934                 /*
6935                  * How much cpu bandwidth does root_task_group get?
6936                  *
6937                  * In case of task-groups formed thr' the cgroup filesystem, it
6938                  * gets 100% of the cpu resources in the system. This overall
6939                  * system cpu resource is divided among the tasks of
6940                  * root_task_group and its child task-groups in a fair manner,
6941                  * based on each entity's (task or task-group's) weight
6942                  * (se->load.weight).
6943                  *
6944                  * In other words, if root_task_group has 10 tasks of weight
6945                  * 1024) and two child groups A0 and A1 (of weight 1024 each),
6946                  * then A0's share of the cpu resource is:
6947                  *
6948                  *      A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6949                  *
6950                  * We achieve this by letting root_task_group's tasks sit
6951                  * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6952                  */
6953                 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6954                 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6955 #endif /* CONFIG_FAIR_GROUP_SCHED */
6956 
6957                 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6958 #ifdef CONFIG_RT_GROUP_SCHED
6959                 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6960 #endif
6961 
6962                 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6963                         rq->cpu_load[j] = 0;
6964 
6965                 rq->last_load_update_tick = jiffies;
6966 
6967 #ifdef CONFIG_SMP
6968                 rq->sd = NULL;
6969                 rq->rd = NULL;
6970                 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
6971                 rq->post_schedule = 0;
6972                 rq->active_balance = 0;
6973                 rq->next_balance = jiffies;
6974                 rq->push_cpu = 0;
6975                 rq->cpu = i;
6976                 rq->online = 0;
6977                 rq->idle_stamp = 0;
6978                 rq->avg_idle = 2*sysctl_sched_migration_cost;
6979                 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6980 
6981                 INIT_LIST_HEAD(&rq->cfs_tasks);
6982 
6983                 rq_attach_root(rq, &def_root_domain);
6984 #ifdef CONFIG_NO_HZ_COMMON
6985                 rq->nohz_flags = 0;
6986 #endif
6987 #ifdef CONFIG_NO_HZ_FULL
6988                 rq->last_sched_tick = 0;
6989 #endif
6990 #endif
6991                 init_rq_hrtick(rq);
6992                 atomic_set(&rq->nr_iowait, 0);
6993         }
6994 
6995         set_load_weight(&init_task);
6996 
6997 #ifdef CONFIG_PREEMPT_NOTIFIERS
6998         INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6999 #endif
7000 
7001         /*
7002          * The boot idle thread does lazy MMU switching as well:
7003          */
7004         atomic_inc(&init_mm.mm_count);
7005         enter_lazy_tlb(&init_mm, current);
7006 
7007         /*
7008          * Make us the idle thread. Technically, schedule() should not be
7009          * called from this thread, however somewhere below it might be,
7010          * but because we are the idle thread, we just pick up running again
7011          * when this runqueue becomes "idle".
7012          */
7013         init_idle(current, smp_processor_id());
7014 
7015         calc_load_update = jiffies + LOAD_FREQ;
7016 
7017         /*
7018          * During early bootup we pretend to be a normal task:
7019          */
7020         current->sched_class = &fair_sched_class;
7021 
7022 #ifdef CONFIG_SMP
7023         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7024         /* May be allocated at isolcpus cmdline parse time */
7025         if (cpu_isolated_map == NULL)
7026                 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7027         idle_thread_set_boot_cpu();
7028         set_cpu_rq_start_time();
7029 #endif
7030         init_sched_fair_class();
7031 
7032         scheduler_running = 1;
7033 }
7034 
7035 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7036 static inline int preempt_count_equals(int preempt_offset)
7037 {
7038         int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7039 
7040         return (nested == preempt_offset);
7041 }
7042 
7043 void __might_sleep(const char *file, int line, int preempt_offset)
7044 {
7045         static unsigned long prev_jiffy;        /* ratelimiting */
7046 
7047         rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7048         if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7049              !is_idle_task(current)) ||
7050             system_state != SYSTEM_RUNNING || oops_in_progress)
7051                 return;
7052         if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7053                 return;
7054         prev_jiffy = jiffies;
7055 
7056         printk(KERN_ERR
7057                 "BUG: sleeping function called from invalid context at %s:%d\n",
7058                         file, line);
7059         printk(KERN_ERR
7060                 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7061                         in_atomic(), irqs_disabled(),
7062                         current->pid, current->comm);
7063 
7064         debug_show_held_locks(current);
7065         if (irqs_disabled())
7066                 print_irqtrace_events(current);
7067 #ifdef CONFIG_DEBUG_PREEMPT
7068         if (!preempt_count_equals(preempt_offset)) {
7069                 pr_err("Preemption disabled at:");
7070                 print_ip_sym(current->preempt_disable_ip);
7071                 pr_cont("\n");
7072         }
7073 #endif
7074         dump_stack();
7075 }
7076 EXPORT_SYMBOL(__might_sleep);
7077 #endif
7078 
7079 #ifdef CONFIG_MAGIC_SYSRQ
7080 static void normalize_task(struct rq *rq, struct task_struct *p)
7081 {
7082         const struct sched_class *prev_class = p->sched_class;
7083         struct sched_attr attr = {
7084                 .sched_policy = SCHED_NORMAL,
7085         };
7086         int old_prio = p->prio;
7087         int on_rq;
7088 
7089         on_rq = p->on_rq;
7090         if (on_rq)
7091                 dequeue_task(rq, p, 0);
7092         __setscheduler(rq, p, &attr);
7093         if (on_rq) {
7094                 enqueue_task(rq, p, 0);
7095                 resched_task(rq->curr);
7096         }
7097 
7098         check_class_changed(rq, p, prev_class, old_prio);
7099 }
7100 
7101 void normalize_rt_tasks(void)
7102 {
7103         struct task_struct *g, *p;
7104         unsigned long flags;
7105         struct rq *rq;
7106 
7107         read_lock_irqsave(&tasklist_lock, flags);
7108         do_each_thread(g, p) {
7109                 /*
7110                  * Only normalize user tasks:
7111                  */
7112                 if (!p->mm)
7113                         continue;
7114 
7115                 p->se.exec_start                = 0;
7116 #ifdef CONFIG_SCHEDSTATS
7117                 p->se.statistics.wait_start     = 0;
7118                 p->se.statistics.sleep_start    = 0;
7119                 p->se.statistics.block_start    = 0;
7120 #endif
7121 
7122                 if (!dl_task(p) && !rt_task(p)) {
7123                         /*
7124                          * Renice negative nice level userspace
7125                          * tasks back to 0:
7126                          */
7127                         if (task_nice(p) < 0 && p->mm)
7128                                 set_user_nice(p, 0);
7129                         continue;
7130                 }
7131 
7132                 raw_spin_lock(&p->pi_lock);
7133                 rq = __task_rq_lock(p);
7134 
7135                 normalize_task(rq, p);
7136 
7137                 __task_rq_unlock(rq);
7138                 raw_spin_unlock(&p->pi_lock);
7139         } while_each_thread(g, p);
7140 
7141         read_unlock_irqrestore(&tasklist_lock, flags);
7142 }
7143 
7144 #endif /* CONFIG_MAGIC_SYSRQ */
7145 
7146 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7147 /*
7148  * These functions are only useful for the IA64 MCA handling, or kdb.
7149  *
7150  * They can only be called when the whole system has been
7151  * stopped - every CPU needs to be quiescent, and no scheduling
7152  * activity can take place. Using them for anything else would
7153  * be a serious bug, and as a result, they aren't even visible
7154  * under any other configuration.
7155  */
7156 
7157 /**
7158  * curr_task - return the current task for a given cpu.
7159  * @cpu: the processor in question.
7160  *
7161  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7162  *
7163  * Return: The current task for @cpu.
7164  */
7165 struct task_struct *curr_task(int cpu)
7166 {
7167         return cpu_curr(cpu);
7168 }
7169 
7170 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7171 
7172 #ifdef CONFIG_IA64
7173 /**
7174  * set_curr_task - set the current task for a given cpu.
7175  * @cpu: the processor in question.
7176  * @p: the task pointer to set.
7177  *
7178  * Description: This function must only be used when non-maskable interrupts
7179  * are serviced on a separate stack. It allows the architecture to switch the
7180  * notion of the current task on a cpu in a non-blocking manner. This function
7181  * must be called with all CPU's synchronized, and interrupts disabled, the
7182  * and caller must save the original value of the current task (see
7183  * curr_task() above) and restore that value before reenabling interrupts and
7184  * re-starting the system.
7185  *
7186  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7187  */
7188 void set_curr_task(int cpu, struct task_struct *p)
7189 {
7190         cpu_curr(cpu) = p;
7191 }
7192 
7193 #endif
7194 
7195 #ifdef CONFIG_CGROUP_SCHED
7196 /* task_group_lock serializes the addition/removal of task groups */
7197 static DEFINE_SPINLOCK(task_group_lock);
7198 
7199 static void free_sched_group(struct task_group *tg)
7200 {
7201         free_fair_sched_group(tg);
7202         free_rt_sched_group(tg);
7203         autogroup_free(tg);
7204         kfree(tg);
7205 }
7206 
7207 /* allocate runqueue etc for a new task group */
7208 struct task_group *sched_create_group(struct task_group *parent)
7209 {
7210         struct task_group *tg;
7211 
7212         tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7213         if (!tg)
7214                 return ERR_PTR(-ENOMEM);
7215 
7216         if (!alloc_fair_sched_group(tg, parent))
7217                 goto err;
7218 
7219         if (!alloc_rt_sched_group(tg, parent))
7220                 goto err;
7221 
7222         return tg;
7223 
7224 err:
7225         free_sched_group(tg);
7226         return ERR_PTR(-ENOMEM);
7227 }
7228 
7229 void sched_online_group(struct task_group *tg, struct task_group *parent)
7230 {
7231         unsigned long flags;
7232 
7233         spin_lock_irqsave(&task_group_lock, flags);
7234         list_add_rcu(&tg->list, &task_groups);
7235 
7236         WARN_ON(!parent); /* root should already exist */
7237 
7238         tg->parent = parent;
7239         INIT_LIST_HEAD(&tg->children);
7240         list_add_rcu(&tg->siblings, &parent->children);
7241         spin_unlock_irqrestore(&task_group_lock, flags);
7242 }
7243 
7244 /* rcu callback to free various structures associated with a task group */
7245 static void free_sched_group_rcu(struct rcu_head *rhp)
7246 {
7247         /* now it should be safe to free those cfs_rqs */
7248         free_sched_group(container_of(rhp, struct task_group, rcu));
7249 }
7250 
7251 /* Destroy runqueue etc associated with a task group */
7252 void sched_destroy_group(struct task_group *tg)
7253 {
7254         /* wait for possible concurrent references to cfs_rqs complete */
7255         call_rcu(&tg->rcu, free_sched_group_rcu);
7256 }
7257 
7258 void sched_offline_group(struct task_group *tg)
7259 {
7260         unsigned long flags;
7261         int i;
7262 
7263         /* end participation in shares distribution */
7264         for_each_possible_cpu(i)
7265                 unregister_fair_sched_group(tg, i);
7266 
7267         spin_lock_irqsave(&task_group_lock, flags);
7268         list_del_rcu(&tg->list);
7269         list_del_rcu(&tg->siblings);
7270         spin_unlock_irqrestore(&task_group_lock, flags);
7271 }
7272 
7273 /* change task's runqueue when it moves between groups.
7274  *      The caller of this function should have put the task in its new group
7275  *      by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7276  *      reflect its new group.
7277  */
7278 void sched_move_task(struct task_struct *tsk)
7279 {
7280         struct task_group *tg;
7281         int on_rq, running;
7282         unsigned long flags;
7283         struct rq *rq;
7284 
7285         rq = task_rq_lock(tsk, &flags);
7286 
7287         running = task_current(rq, tsk);
7288         on_rq = tsk->on_rq;
7289 
7290         if (on_rq)
7291                 dequeue_task(rq, tsk, 0);
7292         if (unlikely(running))
7293                 tsk->sched_class->put_prev_task(rq, tsk);
7294 
7295         tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7296                                 lockdep_is_held(&tsk->sighand->siglock)),
7297                           struct task_group, css);
7298         tg = autogroup_task_group(tsk, tg);
7299         tsk->sched_task_group = tg;
7300 
7301 #ifdef CONFIG_FAIR_GROUP_SCHED
7302         if (tsk->sched_class->task_move_group)
7303                 tsk->sched_class->task_move_group(tsk, on_rq);
7304         else
7305 #endif
7306                 set_task_rq(tsk, task_cpu(tsk));
7307 
7308         if (unlikely(running))
7309                 tsk->sched_class->set_curr_task(rq);
7310         if (on_rq)
7311                 enqueue_task(rq, tsk, 0);
7312 
7313         task_rq_unlock(rq, tsk, &flags);
7314 }
7315 #endif /* CONFIG_CGROUP_SCHED */
7316 
7317 #ifdef CONFIG_RT_GROUP_SCHED
7318 /*
7319  * Ensure that the real time constraints are schedulable.
7320  */
7321 static DEFINE_MUTEX(rt_constraints_mutex);
7322 
7323 /* Must be called with tasklist_lock held */
7324 static inline int tg_has_rt_tasks(struct task_group *tg)
7325 {
7326         struct task_struct *g, *p;
7327 
7328         do_each_thread(g, p) {
7329                 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7330                         return 1;
7331         } while_each_thread(g, p);
7332 
7333         return 0;
7334 }
7335 
7336 struct rt_schedulable_data {
7337         struct task_group *tg;
7338         u64 rt_period;
7339         u64 rt_runtime;
7340 };
7341 
7342 static int tg_rt_schedulable(struct task_group *tg, void *data)
7343 {
7344         struct rt_schedulable_data *d = data;
7345         struct task_group *child;
7346         unsigned long total, sum = 0;
7347         u64 period, runtime;
7348 
7349         period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7350         runtime = tg->rt_bandwidth.rt_runtime;
7351 
7352         if (tg == d->tg) {
7353                 period = d->rt_period;
7354                 runtime = d->rt_runtime;
7355         }
7356 
7357         /*
7358          * Cannot have more runtime than the period.
7359          */
7360         if (runtime > period && runtime != RUNTIME_INF)
7361                 return -EINVAL;
7362 
7363         /*
7364          * Ensure we don't starve existing RT tasks.
7365          */
7366         if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7367                 return -EBUSY;
7368 
7369         total = to_ratio(period, runtime);
7370 
7371         /*
7372          * Nobody can have more than the global setting allows.
7373          */
7374         if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7375                 return -EINVAL;
7376 
7377         /*
7378          * The sum of our children's runtime should not exceed our own.
7379          */
7380         list_for_each_entry_rcu(child, &tg->children, siblings) {
7381                 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7382                 runtime = child->rt_bandwidth.rt_runtime;
7383 
7384                 if (child == d->tg) {
7385                         period = d->rt_period;
7386                         runtime = d->rt_runtime;
7387                 }
7388 
7389                 sum += to_ratio(period, runtime);
7390         }
7391 
7392         if (sum > total)
7393                 return -EINVAL;
7394 
7395         return 0;
7396 }
7397 
7398 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7399 {
7400         int ret;
7401 
7402         struct rt_schedulable_data data = {
7403                 .tg = tg,
7404                 .rt_period = period,
7405                 .rt_runtime = runtime,
7406         };
7407 
7408         rcu_read_lock();
7409