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 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
 94 {
 95         unsigned long delta;
 96         ktime_t soft, hard, now;
 97 
 98         for (;;) {
 99                 if (hrtimer_active(period_timer))
100                         break;
101 
102                 now = hrtimer_cb_get_time(period_timer);
103                 hrtimer_forward(period_timer, now, period);
104 
105                 soft = hrtimer_get_softexpires(period_timer);
106                 hard = hrtimer_get_expires(period_timer);
107                 delta = ktime_to_ns(ktime_sub(hard, soft));
108                 __hrtimer_start_range_ns(period_timer, soft, delta,
109                                          HRTIMER_MODE_ABS_PINNED, 0);
110         }
111 }
112 
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 
118 void update_rq_clock(struct rq *rq)
119 {
120         s64 delta;
121 
122         if (rq->skip_clock_update > 0)
123                 return;
124 
125         delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126         if (delta < 0)
127                 return;
128         rq->clock += delta;
129         update_rq_clock_task(rq, delta);
130 }
131 
132 /*
133  * Debugging: various feature bits
134  */
135 
136 #define SCHED_FEAT(name, enabled)       \
137         (1UL << __SCHED_FEAT_##name) * enabled |
138 
139 const_debug unsigned int sysctl_sched_features =
140 #include "features.h"
141         0;
142 
143 #undef SCHED_FEAT
144 
145 #ifdef CONFIG_SCHED_DEBUG
146 #define SCHED_FEAT(name, enabled)       \
147         #name ,
148 
149 static const char * const sched_feat_names[] = {
150 #include "features.h"
151 };
152 
153 #undef SCHED_FEAT
154 
155 static int sched_feat_show(struct seq_file *m, void *v)
156 {
157         int i;
158 
159         for (i = 0; i < __SCHED_FEAT_NR; i++) {
160                 if (!(sysctl_sched_features & (1UL << i)))
161                         seq_puts(m, "NO_");
162                 seq_printf(m, "%s ", sched_feat_names[i]);
163         }
164         seq_puts(m, "\n");
165 
166         return 0;
167 }
168 
169 #ifdef HAVE_JUMP_LABEL
170 
171 #define jump_label_key__true  STATIC_KEY_INIT_TRUE
172 #define jump_label_key__false STATIC_KEY_INIT_FALSE
173 
174 #define SCHED_FEAT(name, enabled)       \
175         jump_label_key__##enabled ,
176 
177 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
178 #include "features.h"
179 };
180 
181 #undef SCHED_FEAT
182 
183 static void sched_feat_disable(int i)
184 {
185         if (static_key_enabled(&sched_feat_keys[i]))
186                 static_key_slow_dec(&sched_feat_keys[i]);
187 }
188 
189 static void sched_feat_enable(int i)
190 {
191         if (!static_key_enabled(&sched_feat_keys[i]))
192                 static_key_slow_inc(&sched_feat_keys[i]);
193 }
194 #else
195 static void sched_feat_disable(int i) { };
196 static void sched_feat_enable(int i) { };
197 #endif /* HAVE_JUMP_LABEL */
198 
199 static int sched_feat_set(char *cmp)
200 {
201         int i;
202         int neg = 0;
203 
204         if (strncmp(cmp, "NO_", 3) == 0) {
205                 neg = 1;
206                 cmp += 3;
207         }
208 
209         for (i = 0; i < __SCHED_FEAT_NR; i++) {
210                 if (strcmp(cmp, sched_feat_names[i]) == 0) {
211                         if (neg) {
212                                 sysctl_sched_features &= ~(1UL << i);
213                                 sched_feat_disable(i);
214                         } else {
215                                 sysctl_sched_features |= (1UL << i);
216                                 sched_feat_enable(i);
217                         }
218                         break;
219                 }
220         }
221 
222         return i;
223 }
224 
225 static ssize_t
226 sched_feat_write(struct file *filp, const char __user *ubuf,
227                 size_t cnt, loff_t *ppos)
228 {
229         char buf[64];
230         char *cmp;
231         int i;
232         struct inode *inode;
233 
234         if (cnt > 63)
235                 cnt = 63;
236 
237         if (copy_from_user(&buf, ubuf, cnt))
238                 return -EFAULT;
239 
240         buf[cnt] = 0;
241         cmp = strstrip(buf);
242 
243         /* Ensure the static_key remains in a consistent state */
244         inode = file_inode(filp);
245         mutex_lock(&inode->i_mutex);
246         i = sched_feat_set(cmp);
247         mutex_unlock(&inode->i_mutex);
248         if (i == __SCHED_FEAT_NR)
249                 return -EINVAL;
250 
251         *ppos += cnt;
252 
253         return cnt;
254 }
255 
256 static int sched_feat_open(struct inode *inode, struct file *filp)
257 {
258         return single_open(filp, sched_feat_show, NULL);
259 }
260 
261 static const struct file_operations sched_feat_fops = {
262         .open           = sched_feat_open,
263         .write          = sched_feat_write,
264         .read           = seq_read,
265         .llseek         = seq_lseek,
266         .release        = single_release,
267 };
268 
269 static __init int sched_init_debug(void)
270 {
271         debugfs_create_file("sched_features", 0644, NULL, NULL,
272                         &sched_feat_fops);
273 
274         return 0;
275 }
276 late_initcall(sched_init_debug);
277 #endif /* CONFIG_SCHED_DEBUG */
278 
279 /*
280  * Number of tasks to iterate in a single balance run.
281  * Limited because this is done with IRQs disabled.
282  */
283 const_debug unsigned int sysctl_sched_nr_migrate = 32;
284 
285 /*
286  * period over which we average the RT time consumption, measured
287  * in ms.
288  *
289  * default: 1s
290  */
291 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
292 
293 /*
294  * period over which we measure -rt task cpu usage in us.
295  * default: 1s
296  */
297 unsigned int sysctl_sched_rt_period = 1000000;
298 
299 __read_mostly int scheduler_running;
300 
301 /*
302  * part of the period that we allow rt tasks to run in us.
303  * default: 0.95s
304  */
305 int sysctl_sched_rt_runtime = 950000;
306 
307 /*
308  * __task_rq_lock - lock the rq @p resides on.
309  */
310 static inline struct rq *__task_rq_lock(struct task_struct *p)
311         __acquires(rq->lock)
312 {
313         struct rq *rq;
314 
315         lockdep_assert_held(&p->pi_lock);
316 
317         for (;;) {
318                 rq = task_rq(p);
319                 raw_spin_lock(&rq->lock);
320                 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
321                         return rq;
322                 raw_spin_unlock(&rq->lock);
323 
324                 while (unlikely(task_on_rq_migrating(p)))
325                         cpu_relax();
326         }
327 }
328 
329 /*
330  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
331  */
332 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
333         __acquires(p->pi_lock)
334         __acquires(rq->lock)
335 {
336         struct rq *rq;
337 
338         for (;;) {
339                 raw_spin_lock_irqsave(&p->pi_lock, *flags);
340                 rq = task_rq(p);
341                 raw_spin_lock(&rq->lock);
342                 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
343                         return rq;
344                 raw_spin_unlock(&rq->lock);
345                 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
346 
347                 while (unlikely(task_on_rq_migrating(p)))
348                         cpu_relax();
349         }
350 }
351 
352 static void __task_rq_unlock(struct rq *rq)
353         __releases(rq->lock)
354 {
355         raw_spin_unlock(&rq->lock);
356 }
357 
358 static inline void
359 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
360         __releases(rq->lock)
361         __releases(p->pi_lock)
362 {
363         raw_spin_unlock(&rq->lock);
364         raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
365 }
366 
367 /*
368  * this_rq_lock - lock this runqueue and disable interrupts.
369  */
370 static struct rq *this_rq_lock(void)
371         __acquires(rq->lock)
372 {
373         struct rq *rq;
374 
375         local_irq_disable();
376         rq = this_rq();
377         raw_spin_lock(&rq->lock);
378 
379         return rq;
380 }
381 
382 #ifdef CONFIG_SCHED_HRTICK
383 /*
384  * Use HR-timers to deliver accurate preemption points.
385  */
386 
387 static void hrtick_clear(struct rq *rq)
388 {
389         if (hrtimer_active(&rq->hrtick_timer))
390                 hrtimer_cancel(&rq->hrtick_timer);
391 }
392 
393 /*
394  * High-resolution timer tick.
395  * Runs from hardirq context with interrupts disabled.
396  */
397 static enum hrtimer_restart hrtick(struct hrtimer *timer)
398 {
399         struct rq *rq = container_of(timer, struct rq, hrtick_timer);
400 
401         WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
402 
403         raw_spin_lock(&rq->lock);
404         update_rq_clock(rq);
405         rq->curr->sched_class->task_tick(rq, rq->curr, 1);
406         raw_spin_unlock(&rq->lock);
407 
408         return HRTIMER_NORESTART;
409 }
410 
411 #ifdef CONFIG_SMP
412 
413 static int __hrtick_restart(struct rq *rq)
414 {
415         struct hrtimer *timer = &rq->hrtick_timer;
416         ktime_t time = hrtimer_get_softexpires(timer);
417 
418         return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
419 }
420 
421 /*
422  * called from hardirq (IPI) context
423  */
424 static void __hrtick_start(void *arg)
425 {
426         struct rq *rq = arg;
427 
428         raw_spin_lock(&rq->lock);
429         __hrtick_restart(rq);
430         rq->hrtick_csd_pending = 0;
431         raw_spin_unlock(&rq->lock);
432 }
433 
434 /*
435  * Called to set the hrtick timer state.
436  *
437  * called with rq->lock held and irqs disabled
438  */
439 void hrtick_start(struct rq *rq, u64 delay)
440 {
441         struct hrtimer *timer = &rq->hrtick_timer;
442         ktime_t time;
443         s64 delta;
444 
445         /*
446          * Don't schedule slices shorter than 10000ns, that just
447          * doesn't make sense and can cause timer DoS.
448          */
449         delta = max_t(s64, delay, 10000LL);
450         time = ktime_add_ns(timer->base->get_time(), delta);
451 
452         hrtimer_set_expires(timer, time);
453 
454         if (rq == this_rq()) {
455                 __hrtick_restart(rq);
456         } else if (!rq->hrtick_csd_pending) {
457                 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
458                 rq->hrtick_csd_pending = 1;
459         }
460 }
461 
462 static int
463 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
464 {
465         int cpu = (int)(long)hcpu;
466 
467         switch (action) {
468         case CPU_UP_CANCELED:
469         case CPU_UP_CANCELED_FROZEN:
470         case CPU_DOWN_PREPARE:
471         case CPU_DOWN_PREPARE_FROZEN:
472         case CPU_DEAD:
473         case CPU_DEAD_FROZEN:
474                 hrtick_clear(cpu_rq(cpu));
475                 return NOTIFY_OK;
476         }
477 
478         return NOTIFY_DONE;
479 }
480 
481 static __init void init_hrtick(void)
482 {
483         hotcpu_notifier(hotplug_hrtick, 0);
484 }
485 #else
486 /*
487  * Called to set the hrtick timer state.
488  *
489  * called with rq->lock held and irqs disabled
490  */
491 void hrtick_start(struct rq *rq, u64 delay)
492 {
493         __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
494                         HRTIMER_MODE_REL_PINNED, 0);
495 }
496 
497 static inline void init_hrtick(void)
498 {
499 }
500 #endif /* CONFIG_SMP */
501 
502 static void init_rq_hrtick(struct rq *rq)
503 {
504 #ifdef CONFIG_SMP
505         rq->hrtick_csd_pending = 0;
506 
507         rq->hrtick_csd.flags = 0;
508         rq->hrtick_csd.func = __hrtick_start;
509         rq->hrtick_csd.info = rq;
510 #endif
511 
512         hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
513         rq->hrtick_timer.function = hrtick;
514 }
515 #else   /* CONFIG_SCHED_HRTICK */
516 static inline void hrtick_clear(struct rq *rq)
517 {
518 }
519 
520 static inline void init_rq_hrtick(struct rq *rq)
521 {
522 }
523 
524 static inline void init_hrtick(void)
525 {
526 }
527 #endif  /* CONFIG_SCHED_HRTICK */
528 
529 /*
530  * cmpxchg based fetch_or, macro so it works for different integer types
531  */
532 #define fetch_or(ptr, val)                                              \
533 ({      typeof(*(ptr)) __old, __val = *(ptr);                           \
534         for (;;) {                                                      \
535                 __old = cmpxchg((ptr), __val, __val | (val));           \
536                 if (__old == __val)                                     \
537                         break;                                          \
538                 __val = __old;                                          \
539         }                                                               \
540         __old;                                                          \
541 })
542 
543 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
544 /*
545  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
546  * this avoids any races wrt polling state changes and thereby avoids
547  * spurious IPIs.
548  */
549 static bool set_nr_and_not_polling(struct task_struct *p)
550 {
551         struct thread_info *ti = task_thread_info(p);
552         return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
553 }
554 
555 /*
556  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
557  *
558  * If this returns true, then the idle task promises to call
559  * sched_ttwu_pending() and reschedule soon.
560  */
561 static bool set_nr_if_polling(struct task_struct *p)
562 {
563         struct thread_info *ti = task_thread_info(p);
564         typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
565 
566         for (;;) {
567                 if (!(val & _TIF_POLLING_NRFLAG))
568                         return false;
569                 if (val & _TIF_NEED_RESCHED)
570                         return true;
571                 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
572                 if (old == val)
573                         break;
574                 val = old;
575         }
576         return true;
577 }
578 
579 #else
580 static bool set_nr_and_not_polling(struct task_struct *p)
581 {
582         set_tsk_need_resched(p);
583         return true;
584 }
585 
586 #ifdef CONFIG_SMP
587 static bool set_nr_if_polling(struct task_struct *p)
588 {
589         return false;
590 }
591 #endif
592 #endif
593 
594 /*
595  * resched_curr - mark rq's current task 'to be rescheduled now'.
596  *
597  * On UP this means the setting of the need_resched flag, on SMP it
598  * might also involve a cross-CPU call to trigger the scheduler on
599  * the target CPU.
600  */
601 void resched_curr(struct rq *rq)
602 {
603         struct task_struct *curr = rq->curr;
604         int cpu;
605 
606         lockdep_assert_held(&rq->lock);
607 
608         if (test_tsk_need_resched(curr))
609                 return;
610 
611         cpu = cpu_of(rq);
612 
613         if (cpu == smp_processor_id()) {
614                 set_tsk_need_resched(curr);
615                 set_preempt_need_resched();
616                 return;
617         }
618 
619         if (set_nr_and_not_polling(curr))
620                 smp_send_reschedule(cpu);
621         else
622                 trace_sched_wake_idle_without_ipi(cpu);
623 }
624 
625 void resched_cpu(int cpu)
626 {
627         struct rq *rq = cpu_rq(cpu);
628         unsigned long flags;
629 
630         if (!raw_spin_trylock_irqsave(&rq->lock, flags))
631                 return;
632         resched_curr(rq);
633         raw_spin_unlock_irqrestore(&rq->lock, flags);
634 }
635 
636 #ifdef CONFIG_SMP
637 #ifdef CONFIG_NO_HZ_COMMON
638 /*
639  * In the semi idle case, use the nearest busy cpu for migrating timers
640  * from an idle cpu.  This is good for power-savings.
641  *
642  * We don't do similar optimization for completely idle system, as
643  * selecting an idle cpu will add more delays to the timers than intended
644  * (as that cpu's timer base may not be uptodate wrt jiffies etc).
645  */
646 int get_nohz_timer_target(int pinned)
647 {
648         int cpu = smp_processor_id();
649         int i;
650         struct sched_domain *sd;
651 
652         if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
653                 return cpu;
654 
655         rcu_read_lock();
656         for_each_domain(cpu, sd) {
657                 for_each_cpu(i, sched_domain_span(sd)) {
658                         if (!idle_cpu(i)) {
659                                 cpu = i;
660                                 goto unlock;
661                         }
662                 }
663         }
664 unlock:
665         rcu_read_unlock();
666         return cpu;
667 }
668 /*
669  * When add_timer_on() enqueues a timer into the timer wheel of an
670  * idle CPU then this timer might expire before the next timer event
671  * which is scheduled to wake up that CPU. In case of a completely
672  * idle system the next event might even be infinite time into the
673  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
674  * leaves the inner idle loop so the newly added timer is taken into
675  * account when the CPU goes back to idle and evaluates the timer
676  * wheel for the next timer event.
677  */
678 static void wake_up_idle_cpu(int cpu)
679 {
680         struct rq *rq = cpu_rq(cpu);
681 
682         if (cpu == smp_processor_id())
683                 return;
684 
685         if (set_nr_and_not_polling(rq->idle))
686                 smp_send_reschedule(cpu);
687         else
688                 trace_sched_wake_idle_without_ipi(cpu);
689 }
690 
691 static bool wake_up_full_nohz_cpu(int cpu)
692 {
693         /*
694          * We just need the target to call irq_exit() and re-evaluate
695          * the next tick. The nohz full kick at least implies that.
696          * If needed we can still optimize that later with an
697          * empty IRQ.
698          */
699         if (tick_nohz_full_cpu(cpu)) {
700                 if (cpu != smp_processor_id() ||
701                     tick_nohz_tick_stopped())
702                         tick_nohz_full_kick_cpu(cpu);
703                 return true;
704         }
705 
706         return false;
707 }
708 
709 void wake_up_nohz_cpu(int cpu)
710 {
711         if (!wake_up_full_nohz_cpu(cpu))
712                 wake_up_idle_cpu(cpu);
713 }
714 
715 static inline bool got_nohz_idle_kick(void)
716 {
717         int cpu = smp_processor_id();
718 
719         if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
720                 return false;
721 
722         if (idle_cpu(cpu) && !need_resched())
723                 return true;
724 
725         /*
726          * We can't run Idle Load Balance on this CPU for this time so we
727          * cancel it and clear NOHZ_BALANCE_KICK
728          */
729         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
730         return false;
731 }
732 
733 #else /* CONFIG_NO_HZ_COMMON */
734 
735 static inline bool got_nohz_idle_kick(void)
736 {
737         return false;
738 }
739 
740 #endif /* CONFIG_NO_HZ_COMMON */
741 
742 #ifdef CONFIG_NO_HZ_FULL
743 bool sched_can_stop_tick(void)
744 {
745         /*
746          * More than one running task need preemption.
747          * nr_running update is assumed to be visible
748          * after IPI is sent from wakers.
749          */
750         if (this_rq()->nr_running > 1)
751                 return false;
752 
753         return true;
754 }
755 #endif /* CONFIG_NO_HZ_FULL */
756 
757 void sched_avg_update(struct rq *rq)
758 {
759         s64 period = sched_avg_period();
760 
761         while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
762                 /*
763                  * Inline assembly required to prevent the compiler
764                  * optimising this loop into a divmod call.
765                  * See __iter_div_u64_rem() for another example of this.
766                  */
767                 asm("" : "+rm" (rq->age_stamp));
768                 rq->age_stamp += period;
769                 rq->rt_avg /= 2;
770         }
771 }
772 
773 #endif /* CONFIG_SMP */
774 
775 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
776                         (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
777 /*
778  * Iterate task_group tree rooted at *from, calling @down when first entering a
779  * node and @up when leaving it for the final time.
780  *
781  * Caller must hold rcu_lock or sufficient equivalent.
782  */
783 int walk_tg_tree_from(struct task_group *from,
784                              tg_visitor down, tg_visitor up, void *data)
785 {
786         struct task_group *parent, *child;
787         int ret;
788 
789         parent = from;
790 
791 down:
792         ret = (*down)(parent, data);
793         if (ret)
794                 goto out;
795         list_for_each_entry_rcu(child, &parent->children, siblings) {
796                 parent = child;
797                 goto down;
798 
799 up:
800                 continue;
801         }
802         ret = (*up)(parent, data);
803         if (ret || parent == from)
804                 goto out;
805 
806         child = parent;
807         parent = parent->parent;
808         if (parent)
809                 goto up;
810 out:
811         return ret;
812 }
813 
814 int tg_nop(struct task_group *tg, void *data)
815 {
816         return 0;
817 }
818 #endif
819 
820 static void set_load_weight(struct task_struct *p)
821 {
822         int prio = p->static_prio - MAX_RT_PRIO;
823         struct load_weight *load = &p->se.load;
824 
825         /*
826          * SCHED_IDLE tasks get minimal weight:
827          */
828         if (p->policy == SCHED_IDLE) {
829                 load->weight = scale_load(WEIGHT_IDLEPRIO);
830                 load->inv_weight = WMULT_IDLEPRIO;
831                 return;
832         }
833 
834         load->weight = scale_load(prio_to_weight[prio]);
835         load->inv_weight = prio_to_wmult[prio];
836 }
837 
838 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
839 {
840         update_rq_clock(rq);
841         sched_info_queued(rq, p);
842         p->sched_class->enqueue_task(rq, p, flags);
843 }
844 
845 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
846 {
847         update_rq_clock(rq);
848         sched_info_dequeued(rq, p);
849         p->sched_class->dequeue_task(rq, p, flags);
850 }
851 
852 void activate_task(struct rq *rq, struct task_struct *p, int flags)
853 {
854         if (task_contributes_to_load(p))
855                 rq->nr_uninterruptible--;
856 
857         enqueue_task(rq, p, flags);
858 }
859 
860 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
861 {
862         if (task_contributes_to_load(p))
863                 rq->nr_uninterruptible++;
864 
865         dequeue_task(rq, p, flags);
866 }
867 
868 static void update_rq_clock_task(struct rq *rq, s64 delta)
869 {
870 /*
871  * In theory, the compile should just see 0 here, and optimize out the call
872  * to sched_rt_avg_update. But I don't trust it...
873  */
874 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
875         s64 steal = 0, irq_delta = 0;
876 #endif
877 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
878         irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
879 
880         /*
881          * Since irq_time is only updated on {soft,}irq_exit, we might run into
882          * this case when a previous update_rq_clock() happened inside a
883          * {soft,}irq region.
884          *
885          * When this happens, we stop ->clock_task and only update the
886          * prev_irq_time stamp to account for the part that fit, so that a next
887          * update will consume the rest. This ensures ->clock_task is
888          * monotonic.
889          *
890          * It does however cause some slight miss-attribution of {soft,}irq
891          * time, a more accurate solution would be to update the irq_time using
892          * the current rq->clock timestamp, except that would require using
893          * atomic ops.
894          */
895         if (irq_delta > delta)
896                 irq_delta = delta;
897 
898         rq->prev_irq_time += irq_delta;
899         delta -= irq_delta;
900 #endif
901 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
902         if (static_key_false((&paravirt_steal_rq_enabled))) {
903                 steal = paravirt_steal_clock(cpu_of(rq));
904                 steal -= rq->prev_steal_time_rq;
905 
906                 if (unlikely(steal > delta))
907                         steal = delta;
908 
909                 rq->prev_steal_time_rq += steal;
910                 delta -= steal;
911         }
912 #endif
913 
914         rq->clock_task += delta;
915 
916 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
917         if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
918                 sched_rt_avg_update(rq, irq_delta + steal);
919 #endif
920 }
921 
922 void sched_set_stop_task(int cpu, struct task_struct *stop)
923 {
924         struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
925         struct task_struct *old_stop = cpu_rq(cpu)->stop;
926 
927         if (stop) {
928                 /*
929                  * Make it appear like a SCHED_FIFO task, its something
930                  * userspace knows about and won't get confused about.
931                  *
932                  * Also, it will make PI more or less work without too
933                  * much confusion -- but then, stop work should not
934                  * rely on PI working anyway.
935                  */
936                 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
937 
938                 stop->sched_class = &stop_sched_class;
939         }
940 
941         cpu_rq(cpu)->stop = stop;
942 
943         if (old_stop) {
944                 /*
945                  * Reset it back to a normal scheduling class so that
946                  * it can die in pieces.
947                  */
948                 old_stop->sched_class = &rt_sched_class;
949         }
950 }
951 
952 /*
953  * __normal_prio - return the priority that is based on the static prio
954  */
955 static inline int __normal_prio(struct task_struct *p)
956 {
957         return p->static_prio;
958 }
959 
960 /*
961  * Calculate the expected normal priority: i.e. priority
962  * without taking RT-inheritance into account. Might be
963  * boosted by interactivity modifiers. Changes upon fork,
964  * setprio syscalls, and whenever the interactivity
965  * estimator recalculates.
966  */
967 static inline int normal_prio(struct task_struct *p)
968 {
969         int prio;
970 
971         if (task_has_dl_policy(p))
972                 prio = MAX_DL_PRIO-1;
973         else if (task_has_rt_policy(p))
974                 prio = MAX_RT_PRIO-1 - p->rt_priority;
975         else
976                 prio = __normal_prio(p);
977         return prio;
978 }
979 
980 /*
981  * Calculate the current priority, i.e. the priority
982  * taken into account by the scheduler. This value might
983  * be boosted by RT tasks, or might be boosted by
984  * interactivity modifiers. Will be RT if the task got
985  * RT-boosted. If not then it returns p->normal_prio.
986  */
987 static int effective_prio(struct task_struct *p)
988 {
989         p->normal_prio = normal_prio(p);
990         /*
991          * If we are RT tasks or we were boosted to RT priority,
992          * keep the priority unchanged. Otherwise, update priority
993          * to the normal priority:
994          */
995         if (!rt_prio(p->prio))
996                 return p->normal_prio;
997         return p->prio;
998 }
999 
1000 /**
1001  * task_curr - is this task currently executing on a CPU?
1002  * @p: the task in question.
1003  *
1004  * Return: 1 if the task is currently executing. 0 otherwise.
1005  */
1006 inline int task_curr(const struct task_struct *p)
1007 {
1008         return cpu_curr(task_cpu(p)) == p;
1009 }
1010 
1011 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1012                                        const struct sched_class *prev_class,
1013                                        int oldprio)
1014 {
1015         if (prev_class != p->sched_class) {
1016                 if (prev_class->switched_from)
1017                         prev_class->switched_from(rq, p);
1018                 p->sched_class->switched_to(rq, p);
1019         } else if (oldprio != p->prio || dl_task(p))
1020                 p->sched_class->prio_changed(rq, p, oldprio);
1021 }
1022 
1023 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1024 {
1025         const struct sched_class *class;
1026 
1027         if (p->sched_class == rq->curr->sched_class) {
1028                 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1029         } else {
1030                 for_each_class(class) {
1031                         if (class == rq->curr->sched_class)
1032                                 break;
1033                         if (class == p->sched_class) {
1034                                 resched_curr(rq);
1035                                 break;
1036                         }
1037                 }
1038         }
1039 
1040         /*
1041          * A queue event has occurred, and we're going to schedule.  In
1042          * this case, we can save a useless back to back clock update.
1043          */
1044         if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1045                 rq->skip_clock_update = 1;
1046 }
1047 
1048 #ifdef CONFIG_SMP
1049 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1050 {
1051 #ifdef CONFIG_SCHED_DEBUG
1052         /*
1053          * We should never call set_task_cpu() on a blocked task,
1054          * ttwu() will sort out the placement.
1055          */
1056         WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1057                         !(task_preempt_count(p) & PREEMPT_ACTIVE));
1058 
1059 #ifdef CONFIG_LOCKDEP
1060         /*
1061          * The caller should hold either p->pi_lock or rq->lock, when changing
1062          * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1063          *
1064          * sched_move_task() holds both and thus holding either pins the cgroup,
1065          * see task_group().
1066          *
1067          * Furthermore, all task_rq users should acquire both locks, see
1068          * task_rq_lock().
1069          */
1070         WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1071                                       lockdep_is_held(&task_rq(p)->lock)));
1072 #endif
1073 #endif
1074 
1075         trace_sched_migrate_task(p, new_cpu);
1076 
1077         if (task_cpu(p) != new_cpu) {
1078                 if (p->sched_class->migrate_task_rq)
1079                         p->sched_class->migrate_task_rq(p, new_cpu);
1080                 p->se.nr_migrations++;
1081                 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1082         }
1083 
1084         __set_task_cpu(p, new_cpu);
1085 }
1086 
1087 static void __migrate_swap_task(struct task_struct *p, int cpu)
1088 {
1089         if (task_on_rq_queued(p)) {
1090                 struct rq *src_rq, *dst_rq;
1091 
1092                 src_rq = task_rq(p);
1093                 dst_rq = cpu_rq(cpu);
1094 
1095                 deactivate_task(src_rq, p, 0);
1096                 set_task_cpu(p, cpu);
1097                 activate_task(dst_rq, p, 0);
1098                 check_preempt_curr(dst_rq, p, 0);
1099         } else {
1100                 /*
1101                  * Task isn't running anymore; make it appear like we migrated
1102                  * it before it went to sleep. This means on wakeup we make the
1103                  * previous cpu our targer instead of where it really is.
1104                  */
1105                 p->wake_cpu = cpu;
1106         }
1107 }
1108 
1109 struct migration_swap_arg {
1110         struct task_struct *src_task, *dst_task;
1111         int src_cpu, dst_cpu;
1112 };
1113 
1114 static int migrate_swap_stop(void *data)
1115 {
1116         struct migration_swap_arg *arg = data;
1117         struct rq *src_rq, *dst_rq;
1118         int ret = -EAGAIN;
1119 
1120         src_rq = cpu_rq(arg->src_cpu);
1121         dst_rq = cpu_rq(arg->dst_cpu);
1122 
1123         double_raw_lock(&arg->src_task->pi_lock,
1124                         &arg->dst_task->pi_lock);
1125         double_rq_lock(src_rq, dst_rq);
1126         if (task_cpu(arg->dst_task) != arg->dst_cpu)
1127                 goto unlock;
1128 
1129         if (task_cpu(arg->src_task) != arg->src_cpu)
1130                 goto unlock;
1131 
1132         if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1133                 goto unlock;
1134 
1135         if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1136                 goto unlock;
1137 
1138         __migrate_swap_task(arg->src_task, arg->dst_cpu);
1139         __migrate_swap_task(arg->dst_task, arg->src_cpu);
1140 
1141         ret = 0;
1142 
1143 unlock:
1144         double_rq_unlock(src_rq, dst_rq);
1145         raw_spin_unlock(&arg->dst_task->pi_lock);
1146         raw_spin_unlock(&arg->src_task->pi_lock);
1147 
1148         return ret;
1149 }
1150 
1151 /*
1152  * Cross migrate two tasks
1153  */
1154 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1155 {
1156         struct migration_swap_arg arg;
1157         int ret = -EINVAL;
1158 
1159         arg = (struct migration_swap_arg){
1160                 .src_task = cur,
1161                 .src_cpu = task_cpu(cur),
1162                 .dst_task = p,
1163                 .dst_cpu = task_cpu(p),
1164         };
1165 
1166         if (arg.src_cpu == arg.dst_cpu)
1167                 goto out;
1168 
1169         /*
1170          * These three tests are all lockless; this is OK since all of them
1171          * will be re-checked with proper locks held further down the line.
1172          */
1173         if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1174                 goto out;
1175 
1176         if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1177                 goto out;
1178 
1179         if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1180                 goto out;
1181 
1182         trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1183         ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1184 
1185 out:
1186         return ret;
1187 }
1188 
1189 struct migration_arg {
1190         struct task_struct *task;
1191         int dest_cpu;
1192 };
1193 
1194 static int migration_cpu_stop(void *data);
1195 
1196 /*
1197  * wait_task_inactive - wait for a thread to unschedule.
1198  *
1199  * If @match_state is nonzero, it's the @p->state value just checked and
1200  * not expected to change.  If it changes, i.e. @p might have woken up,
1201  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1202  * we return a positive number (its total switch count).  If a second call
1203  * a short while later returns the same number, the caller can be sure that
1204  * @p has remained unscheduled the whole time.
1205  *
1206  * The caller must ensure that the task *will* unschedule sometime soon,
1207  * else this function might spin for a *long* time. This function can't
1208  * be called with interrupts off, or it may introduce deadlock with
1209  * smp_call_function() if an IPI is sent by the same process we are
1210  * waiting to become inactive.
1211  */
1212 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1213 {
1214         unsigned long flags;
1215         int running, queued;
1216         unsigned long ncsw;
1217         struct rq *rq;
1218 
1219         for (;;) {
1220                 /*
1221                  * We do the initial early heuristics without holding
1222                  * any task-queue locks at all. We'll only try to get
1223                  * the runqueue lock when things look like they will
1224                  * work out!
1225                  */
1226                 rq = task_rq(p);
1227 
1228                 /*
1229                  * If the task is actively running on another CPU
1230                  * still, just relax and busy-wait without holding
1231                  * any locks.
1232                  *
1233                  * NOTE! Since we don't hold any locks, it's not
1234                  * even sure that "rq" stays as the right runqueue!
1235                  * But we don't care, since "task_running()" will
1236                  * return false if the runqueue has changed and p
1237                  * is actually now running somewhere else!
1238                  */
1239                 while (task_running(rq, p)) {
1240                         if (match_state && unlikely(p->state != match_state))
1241                                 return 0;
1242                         cpu_relax();
1243                 }
1244 
1245                 /*
1246                  * Ok, time to look more closely! We need the rq
1247                  * lock now, to be *sure*. If we're wrong, we'll
1248                  * just go back and repeat.
1249                  */
1250                 rq = task_rq_lock(p, &flags);
1251                 trace_sched_wait_task(p);
1252                 running = task_running(rq, p);
1253                 queued = task_on_rq_queued(p);
1254                 ncsw = 0;
1255                 if (!match_state || p->state == match_state)
1256                         ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1257                 task_rq_unlock(rq, p, &flags);
1258 
1259                 /*
1260                  * If it changed from the expected state, bail out now.
1261                  */
1262                 if (unlikely(!ncsw))
1263                         break;
1264 
1265                 /*
1266                  * Was it really running after all now that we
1267                  * checked with the proper locks actually held?
1268                  *
1269                  * Oops. Go back and try again..
1270                  */
1271                 if (unlikely(running)) {
1272                         cpu_relax();
1273                         continue;
1274                 }
1275 
1276                 /*
1277                  * It's not enough that it's not actively running,
1278                  * it must be off the runqueue _entirely_, and not
1279                  * preempted!
1280                  *
1281                  * So if it was still runnable (but just not actively
1282                  * running right now), it's preempted, and we should
1283                  * yield - it could be a while.
1284                  */
1285                 if (unlikely(queued)) {
1286                         ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1287 
1288                         set_current_state(TASK_UNINTERRUPTIBLE);
1289                         schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1290                         continue;
1291                 }
1292 
1293                 /*
1294                  * Ahh, all good. It wasn't running, and it wasn't
1295                  * runnable, which means that it will never become
1296                  * running in the future either. We're all done!
1297                  */
1298                 break;
1299         }
1300 
1301         return ncsw;
1302 }
1303 
1304 /***
1305  * kick_process - kick a running thread to enter/exit the kernel
1306  * @p: the to-be-kicked thread
1307  *
1308  * Cause a process which is running on another CPU to enter
1309  * kernel-mode, without any delay. (to get signals handled.)
1310  *
1311  * NOTE: this function doesn't have to take the runqueue lock,
1312  * because all it wants to ensure is that the remote task enters
1313  * the kernel. If the IPI races and the task has been migrated
1314  * to another CPU then no harm is done and the purpose has been
1315  * achieved as well.
1316  */
1317 void kick_process(struct task_struct *p)
1318 {
1319         int cpu;
1320 
1321         preempt_disable();
1322         cpu = task_cpu(p);
1323         if ((cpu != smp_processor_id()) && task_curr(p))
1324                 smp_send_reschedule(cpu);
1325         preempt_enable();
1326 }
1327 EXPORT_SYMBOL_GPL(kick_process);
1328 #endif /* CONFIG_SMP */
1329 
1330 #ifdef CONFIG_SMP
1331 /*
1332  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1333  */
1334 static int select_fallback_rq(int cpu, struct task_struct *p)
1335 {
1336         int nid = cpu_to_node(cpu);
1337         const struct cpumask *nodemask = NULL;
1338         enum { cpuset, possible, fail } state = cpuset;
1339         int dest_cpu;
1340 
1341         /*
1342          * If the node that the cpu is on has been offlined, cpu_to_node()
1343          * will return -1. There is no cpu on the node, and we should
1344          * select the cpu on the other node.
1345          */
1346         if (nid != -1) {
1347                 nodemask = cpumask_of_node(nid);
1348 
1349                 /* Look for allowed, online CPU in same node. */
1350                 for_each_cpu(dest_cpu, nodemask) {
1351                         if (!cpu_online(dest_cpu))
1352                                 continue;
1353                         if (!cpu_active(dest_cpu))
1354                                 continue;
1355                         if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1356                                 return dest_cpu;
1357                 }
1358         }
1359 
1360         for (;;) {
1361                 /* Any allowed, online CPU? */
1362                 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1363                         if (!cpu_online(dest_cpu))
1364                                 continue;
1365                         if (!cpu_active(dest_cpu))
1366                                 continue;
1367                         goto out;
1368                 }
1369 
1370                 switch (state) {
1371                 case cpuset:
1372                         /* No more Mr. Nice Guy. */
1373                         cpuset_cpus_allowed_fallback(p);
1374                         state = possible;
1375                         break;
1376 
1377                 case possible:
1378                         do_set_cpus_allowed(p, cpu_possible_mask);
1379                         state = fail;
1380                         break;
1381 
1382                 case fail:
1383                         BUG();
1384                         break;
1385                 }
1386         }
1387 
1388 out:
1389         if (state != cpuset) {
1390                 /*
1391                  * Don't tell them about moving exiting tasks or
1392                  * kernel threads (both mm NULL), since they never
1393                  * leave kernel.
1394                  */
1395                 if (p->mm && printk_ratelimit()) {
1396                         printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1397                                         task_pid_nr(p), p->comm, cpu);
1398                 }
1399         }
1400 
1401         return dest_cpu;
1402 }
1403 
1404 /*
1405  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1406  */
1407 static inline
1408 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1409 {
1410         cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1411 
1412         /*
1413          * In order not to call set_task_cpu() on a blocking task we need
1414          * to rely on ttwu() to place the task on a valid ->cpus_allowed
1415          * cpu.
1416          *
1417          * Since this is common to all placement strategies, this lives here.
1418          *
1419          * [ this allows ->select_task() to simply return task_cpu(p) and
1420          *   not worry about this generic constraint ]
1421          */
1422         if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1423                      !cpu_online(cpu)))
1424                 cpu = select_fallback_rq(task_cpu(p), p);
1425 
1426         return cpu;
1427 }
1428 
1429 static void update_avg(u64 *avg, u64 sample)
1430 {
1431         s64 diff = sample - *avg;
1432         *avg += diff >> 3;
1433 }
1434 #endif
1435 
1436 static void
1437 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1438 {
1439 #ifdef CONFIG_SCHEDSTATS
1440         struct rq *rq = this_rq();
1441 
1442 #ifdef CONFIG_SMP
1443         int this_cpu = smp_processor_id();
1444 
1445         if (cpu == this_cpu) {
1446                 schedstat_inc(rq, ttwu_local);
1447                 schedstat_inc(p, se.statistics.nr_wakeups_local);
1448         } else {
1449                 struct sched_domain *sd;
1450 
1451                 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1452                 rcu_read_lock();
1453                 for_each_domain(this_cpu, sd) {
1454                         if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1455                                 schedstat_inc(sd, ttwu_wake_remote);
1456                                 break;
1457                         }
1458                 }
1459                 rcu_read_unlock();
1460         }
1461 
1462         if (wake_flags & WF_MIGRATED)
1463                 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1464 
1465 #endif /* CONFIG_SMP */
1466 
1467         schedstat_inc(rq, ttwu_count);
1468         schedstat_inc(p, se.statistics.nr_wakeups);
1469 
1470         if (wake_flags & WF_SYNC)
1471                 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1472 
1473 #endif /* CONFIG_SCHEDSTATS */
1474 }
1475 
1476 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1477 {
1478         activate_task(rq, p, en_flags);
1479         p->on_rq = TASK_ON_RQ_QUEUED;
1480 
1481         /* if a worker is waking up, notify workqueue */
1482         if (p->flags & PF_WQ_WORKER)
1483                 wq_worker_waking_up(p, cpu_of(rq));
1484 }
1485 
1486 /*
1487  * Mark the task runnable and perform wakeup-preemption.
1488  */
1489 static void
1490 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1491 {
1492         check_preempt_curr(rq, p, wake_flags);
1493         trace_sched_wakeup(p, true);
1494 
1495         p->state = TASK_RUNNING;
1496 #ifdef CONFIG_SMP
1497         if (p->sched_class->task_woken)
1498                 p->sched_class->task_woken(rq, p);
1499 
1500         if (rq->idle_stamp) {
1501                 u64 delta = rq_clock(rq) - rq->idle_stamp;
1502                 u64 max = 2*rq->max_idle_balance_cost;
1503 
1504                 update_avg(&rq->avg_idle, delta);
1505 
1506                 if (rq->avg_idle > max)
1507                         rq->avg_idle = max;
1508 
1509                 rq->idle_stamp = 0;
1510         }
1511 #endif
1512 }
1513 
1514 static void
1515 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1516 {
1517 #ifdef CONFIG_SMP
1518         if (p->sched_contributes_to_load)
1519                 rq->nr_uninterruptible--;
1520 #endif
1521 
1522         ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1523         ttwu_do_wakeup(rq, p, wake_flags);
1524 }
1525 
1526 /*
1527  * Called in case the task @p isn't fully descheduled from its runqueue,
1528  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1529  * since all we need to do is flip p->state to TASK_RUNNING, since
1530  * the task is still ->on_rq.
1531  */
1532 static int ttwu_remote(struct task_struct *p, int wake_flags)
1533 {
1534         struct rq *rq;
1535         int ret = 0;
1536 
1537         rq = __task_rq_lock(p);
1538         if (task_on_rq_queued(p)) {
1539                 /* check_preempt_curr() may use rq clock */
1540                 update_rq_clock(rq);
1541                 ttwu_do_wakeup(rq, p, wake_flags);
1542                 ret = 1;
1543         }
1544         __task_rq_unlock(rq);
1545 
1546         return ret;
1547 }
1548 
1549 #ifdef CONFIG_SMP
1550 void sched_ttwu_pending(void)
1551 {
1552         struct rq *rq = this_rq();
1553         struct llist_node *llist = llist_del_all(&rq->wake_list);
1554         struct task_struct *p;
1555         unsigned long flags;
1556 
1557         if (!llist)
1558                 return;
1559 
1560         raw_spin_lock_irqsave(&rq->lock, flags);
1561 
1562         while (llist) {
1563                 p = llist_entry(llist, struct task_struct, wake_entry);
1564                 llist = llist_next(llist);
1565                 ttwu_do_activate(rq, p, 0);
1566         }
1567 
1568         raw_spin_unlock_irqrestore(&rq->lock, flags);
1569 }
1570 
1571 void scheduler_ipi(void)
1572 {
1573         /*
1574          * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1575          * TIF_NEED_RESCHED remotely (for the first time) will also send
1576          * this IPI.
1577          */
1578         preempt_fold_need_resched();
1579 
1580         if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1581                 return;
1582 
1583         /*
1584          * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1585          * traditionally all their work was done from the interrupt return
1586          * path. Now that we actually do some work, we need to make sure
1587          * we do call them.
1588          *
1589          * Some archs already do call them, luckily irq_enter/exit nest
1590          * properly.
1591          *
1592          * Arguably we should visit all archs and update all handlers,
1593          * however a fair share of IPIs are still resched only so this would
1594          * somewhat pessimize the simple resched case.
1595          */
1596         irq_enter();
1597         sched_ttwu_pending();
1598 
1599         /*
1600          * Check if someone kicked us for doing the nohz idle load balance.
1601          */
1602         if (unlikely(got_nohz_idle_kick())) {
1603                 this_rq()->idle_balance = 1;
1604                 raise_softirq_irqoff(SCHED_SOFTIRQ);
1605         }
1606         irq_exit();
1607 }
1608 
1609 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1610 {
1611         struct rq *rq = cpu_rq(cpu);
1612 
1613         if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1614                 if (!set_nr_if_polling(rq->idle))
1615                         smp_send_reschedule(cpu);
1616                 else
1617                         trace_sched_wake_idle_without_ipi(cpu);
1618         }
1619 }
1620 
1621 void wake_up_if_idle(int cpu)
1622 {
1623         struct rq *rq = cpu_rq(cpu);
1624         unsigned long flags;
1625 
1626         if (!is_idle_task(rq->curr))
1627                 return;
1628 
1629         if (set_nr_if_polling(rq->idle)) {
1630                 trace_sched_wake_idle_without_ipi(cpu);
1631         } else {
1632                 raw_spin_lock_irqsave(&rq->lock, flags);
1633                 if (is_idle_task(rq->curr))
1634                         smp_send_reschedule(cpu);
1635                 /* Else cpu is not in idle, do nothing here */
1636                 raw_spin_unlock_irqrestore(&rq->lock, flags);
1637         }
1638 }
1639 
1640 bool cpus_share_cache(int this_cpu, int that_cpu)
1641 {
1642         return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1643 }
1644 #endif /* CONFIG_SMP */
1645 
1646 static void ttwu_queue(struct task_struct *p, int cpu)
1647 {
1648         struct rq *rq = cpu_rq(cpu);
1649 
1650 #if defined(CONFIG_SMP)
1651         if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1652                 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1653                 ttwu_queue_remote(p, cpu);
1654                 return;
1655         }
1656 #endif
1657 
1658         raw_spin_lock(&rq->lock);
1659         ttwu_do_activate(rq, p, 0);
1660         raw_spin_unlock(&rq->lock);
1661 }
1662 
1663 /**
1664  * try_to_wake_up - wake up a thread
1665  * @p: the thread to be awakened
1666  * @state: the mask of task states that can be woken
1667  * @wake_flags: wake modifier flags (WF_*)
1668  *
1669  * Put it on the run-queue if it's not already there. The "current"
1670  * thread is always on the run-queue (except when the actual
1671  * re-schedule is in progress), and as such you're allowed to do
1672  * the simpler "current->state = TASK_RUNNING" to mark yourself
1673  * runnable without the overhead of this.
1674  *
1675  * Return: %true if @p was woken up, %false if it was already running.
1676  * or @state didn't match @p's state.
1677  */
1678 static int
1679 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1680 {
1681         unsigned long flags;
1682         int cpu, success = 0;
1683 
1684         /*
1685          * If we are going to wake up a thread waiting for CONDITION we
1686          * need to ensure that CONDITION=1 done by the caller can not be
1687          * reordered with p->state check below. This pairs with mb() in
1688          * set_current_state() the waiting thread does.
1689          */
1690         smp_mb__before_spinlock();
1691         raw_spin_lock_irqsave(&p->pi_lock, flags);
1692         if (!(p->state & state))
1693                 goto out;
1694 
1695         success = 1; /* we're going to change ->state */
1696         cpu = task_cpu(p);
1697 
1698         if (p->on_rq && ttwu_remote(p, wake_flags))
1699                 goto stat;
1700 
1701 #ifdef CONFIG_SMP
1702         /*
1703          * If the owning (remote) cpu is still in the middle of schedule() with
1704          * this task as prev, wait until its done referencing the task.
1705          */
1706         while (p->on_cpu)
1707                 cpu_relax();
1708         /*
1709          * Pairs with the smp_wmb() in finish_lock_switch().
1710          */
1711         smp_rmb();
1712 
1713         p->sched_contributes_to_load = !!task_contributes_to_load(p);
1714         p->state = TASK_WAKING;
1715 
1716         if (p->sched_class->task_waking)
1717                 p->sched_class->task_waking(p);
1718 
1719         cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1720         if (task_cpu(p) != cpu) {
1721                 wake_flags |= WF_MIGRATED;
1722                 set_task_cpu(p, cpu);
1723         }
1724 #endif /* CONFIG_SMP */
1725 
1726         ttwu_queue(p, cpu);
1727 stat:
1728         ttwu_stat(p, cpu, wake_flags);
1729 out:
1730         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1731 
1732         return success;
1733 }
1734 
1735 /**
1736  * try_to_wake_up_local - try to wake up a local task with rq lock held
1737  * @p: the thread to be awakened
1738  *
1739  * Put @p on the run-queue if it's not already there. The caller must
1740  * ensure that this_rq() is locked, @p is bound to this_rq() and not
1741  * the current task.
1742  */
1743 static void try_to_wake_up_local(struct task_struct *p)
1744 {
1745         struct rq *rq = task_rq(p);
1746 
1747         if (WARN_ON_ONCE(rq != this_rq()) ||
1748             WARN_ON_ONCE(p == current))
1749                 return;
1750 
1751         lockdep_assert_held(&rq->lock);
1752 
1753         if (!raw_spin_trylock(&p->pi_lock)) {
1754                 raw_spin_unlock(&rq->lock);
1755                 raw_spin_lock(&p->pi_lock);
1756                 raw_spin_lock(&rq->lock);
1757         }
1758 
1759         if (!(p->state & TASK_NORMAL))
1760                 goto out;
1761 
1762         if (!task_on_rq_queued(p))
1763                 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1764 
1765         ttwu_do_wakeup(rq, p, 0);
1766         ttwu_stat(p, smp_processor_id(), 0);
1767 out:
1768         raw_spin_unlock(&p->pi_lock);
1769 }
1770 
1771 /**
1772  * wake_up_process - Wake up a specific process
1773  * @p: The process to be woken up.
1774  *
1775  * Attempt to wake up the nominated process and move it to the set of runnable
1776  * processes.
1777  *
1778  * Return: 1 if the process was woken up, 0 if it was already running.
1779  *
1780  * It may be assumed that this function implies a write memory barrier before
1781  * changing the task state if and only if any tasks are woken up.
1782  */
1783 int wake_up_process(struct task_struct *p)
1784 {
1785         WARN_ON(task_is_stopped_or_traced(p));
1786         return try_to_wake_up(p, TASK_NORMAL, 0);
1787 }
1788 EXPORT_SYMBOL(wake_up_process);
1789 
1790 int wake_up_state(struct task_struct *p, unsigned int state)
1791 {
1792         return try_to_wake_up(p, state, 0);
1793 }
1794 
1795 /*
1796  * This function clears the sched_dl_entity static params.
1797  */
1798 void __dl_clear_params(struct task_struct *p)
1799 {
1800         struct sched_dl_entity *dl_se = &p->dl;
1801 
1802         dl_se->dl_runtime = 0;
1803         dl_se->dl_deadline = 0;
1804         dl_se->dl_period = 0;
1805         dl_se->flags = 0;
1806         dl_se->dl_bw = 0;
1807 }
1808 
1809 /*
1810  * Perform scheduler related setup for a newly forked process p.
1811  * p is forked by current.
1812  *
1813  * __sched_fork() is basic setup used by init_idle() too:
1814  */
1815 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1816 {
1817         p->on_rq                        = 0;
1818 
1819         p->se.on_rq                     = 0;
1820         p->se.exec_start                = 0;
1821         p->se.sum_exec_runtime          = 0;
1822         p->se.prev_sum_exec_runtime     = 0;
1823         p->se.nr_migrations             = 0;
1824         p->se.vruntime                  = 0;
1825         INIT_LIST_HEAD(&p->se.group_node);
1826 
1827 #ifdef CONFIG_SCHEDSTATS
1828         memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1829 #endif
1830 
1831         RB_CLEAR_NODE(&p->dl.rb_node);
1832         hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1833         __dl_clear_params(p);
1834 
1835         INIT_LIST_HEAD(&p->rt.run_list);
1836 
1837 #ifdef CONFIG_PREEMPT_NOTIFIERS
1838         INIT_HLIST_HEAD(&p->preempt_notifiers);
1839 #endif
1840 
1841 #ifdef CONFIG_NUMA_BALANCING
1842         if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1843                 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1844                 p->mm->numa_scan_seq = 0;
1845         }
1846 
1847         if (clone_flags & CLONE_VM)
1848                 p->numa_preferred_nid = current->numa_preferred_nid;
1849         else
1850                 p->numa_preferred_nid = -1;
1851 
1852         p->node_stamp = 0ULL;
1853         p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1854         p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1855         p->numa_work.next = &p->numa_work;
1856         p->numa_faults_memory = NULL;
1857         p->numa_faults_buffer_memory = NULL;
1858         p->last_task_numa_placement = 0;
1859         p->last_sum_exec_runtime = 0;
1860 
1861         INIT_LIST_HEAD(&p->numa_entry);
1862         p->numa_group = NULL;
1863 #endif /* CONFIG_NUMA_BALANCING */
1864 }
1865 
1866 #ifdef CONFIG_NUMA_BALANCING
1867 #ifdef CONFIG_SCHED_DEBUG
1868 void set_numabalancing_state(bool enabled)
1869 {
1870         if (enabled)
1871                 sched_feat_set("NUMA");
1872         else
1873                 sched_feat_set("NO_NUMA");
1874 }
1875 #else
1876 __read_mostly bool numabalancing_enabled;
1877 
1878 void set_numabalancing_state(bool enabled)
1879 {
1880         numabalancing_enabled = enabled;
1881 }
1882 #endif /* CONFIG_SCHED_DEBUG */
1883 
1884 #ifdef CONFIG_PROC_SYSCTL
1885 int sysctl_numa_balancing(struct ctl_table *table, int write,
1886                          void __user *buffer, size_t *lenp, loff_t *ppos)
1887 {
1888         struct ctl_table t;
1889         int err;
1890         int state = numabalancing_enabled;
1891 
1892         if (write && !capable(CAP_SYS_ADMIN))
1893                 return -EPERM;
1894 
1895         t = *table;
1896         t.data = &state;
1897         err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1898         if (err < 0)
1899                 return err;
1900         if (write)
1901                 set_numabalancing_state(state);
1902         return err;
1903 }
1904 #endif
1905 #endif
1906 
1907 /*
1908  * fork()/clone()-time setup:
1909  */
1910 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1911 {
1912         unsigned long flags;
1913         int cpu = get_cpu();
1914 
1915         __sched_fork(clone_flags, p);
1916         /*
1917          * We mark the process as running here. This guarantees that
1918          * nobody will actually run it, and a signal or other external
1919          * event cannot wake it up and insert it on the runqueue either.
1920          */
1921         p->state = TASK_RUNNING;
1922 
1923         /*
1924          * Make sure we do not leak PI boosting priority to the child.
1925          */
1926         p->prio = current->normal_prio;
1927 
1928         /*
1929          * Revert to default priority/policy on fork if requested.
1930          */
1931         if (unlikely(p->sched_reset_on_fork)) {
1932                 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1933                         p->policy = SCHED_NORMAL;
1934                         p->static_prio = NICE_TO_PRIO(0);
1935                         p->rt_priority = 0;
1936                 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1937                         p->static_prio = NICE_TO_PRIO(0);
1938 
1939                 p->prio = p->normal_prio = __normal_prio(p);
1940                 set_load_weight(p);
1941 
1942                 /*
1943                  * We don't need the reset flag anymore after the fork. It has
1944                  * fulfilled its duty:
1945                  */
1946                 p->sched_reset_on_fork = 0;
1947         }
1948 
1949         if (dl_prio(p->prio)) {
1950                 put_cpu();
1951                 return -EAGAIN;
1952         } else if (rt_prio(p->prio)) {
1953                 p->sched_class = &rt_sched_class;
1954         } else {
1955                 p->sched_class = &fair_sched_class;
1956         }
1957 
1958         if (p->sched_class->task_fork)
1959                 p->sched_class->task_fork(p);
1960 
1961         /*
1962          * The child is not yet in the pid-hash so no cgroup attach races,
1963          * and the cgroup is pinned to this child due to cgroup_fork()
1964          * is ran before sched_fork().
1965          *
1966          * Silence PROVE_RCU.
1967          */
1968         raw_spin_lock_irqsave(&p->pi_lock, flags);
1969         set_task_cpu(p, cpu);
1970         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1971 
1972 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1973         if (likely(sched_info_on()))
1974                 memset(&p->sched_info, 0, sizeof(p->sched_info));
1975 #endif
1976 #if defined(CONFIG_SMP)
1977         p->on_cpu = 0;
1978 #endif
1979         init_task_preempt_count(p);
1980 #ifdef CONFIG_SMP
1981         plist_node_init(&p->pushable_tasks, MAX_PRIO);
1982         RB_CLEAR_NODE(&p->pushable_dl_tasks);
1983 #endif
1984 
1985         put_cpu();
1986         return 0;
1987 }
1988 
1989 unsigned long to_ratio(u64 period, u64 runtime)
1990 {
1991         if (runtime == RUNTIME_INF)
1992                 return 1ULL << 20;
1993 
1994         /*
1995          * Doing this here saves a lot of checks in all
1996          * the calling paths, and returning zero seems
1997          * safe for them anyway.
1998          */
1999         if (period == 0)
2000                 return 0;
2001 
2002         return div64_u64(runtime << 20, period);
2003 }
2004 
2005 #ifdef CONFIG_SMP
2006 inline struct dl_bw *dl_bw_of(int i)
2007 {
2008         rcu_lockdep_assert(rcu_read_lock_sched_held(),
2009                            "sched RCU must be held");
2010         return &cpu_rq(i)->rd->dl_bw;
2011 }
2012 
2013 static inline int dl_bw_cpus(int i)
2014 {
2015         struct root_domain *rd = cpu_rq(i)->rd;
2016         int cpus = 0;
2017 
2018         rcu_lockdep_assert(rcu_read_lock_sched_held(),
2019                            "sched RCU must be held");
2020         for_each_cpu_and(i, rd->span, cpu_active_mask)
2021                 cpus++;
2022 
2023         return cpus;
2024 }
2025 #else
2026 inline struct dl_bw *dl_bw_of(int i)
2027 {
2028         return &cpu_rq(i)->dl.dl_bw;
2029 }
2030 
2031 static inline int dl_bw_cpus(int i)
2032 {
2033         return 1;
2034 }
2035 #endif
2036 
2037 static inline
2038 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
2039 {
2040         dl_b->total_bw -= tsk_bw;
2041 }
2042 
2043 static inline
2044 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2045 {
2046         dl_b->total_bw += tsk_bw;
2047 }
2048 
2049 static inline
2050 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2051 {
2052         return dl_b->bw != -1 &&
2053                dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2054 }
2055 
2056 /*
2057  * We must be sure that accepting a new task (or allowing changing the
2058  * parameters of an existing one) is consistent with the bandwidth
2059  * constraints. If yes, this function also accordingly updates the currently
2060  * allocated bandwidth to reflect the new situation.
2061  *
2062  * This function is called while holding p's rq->lock.
2063  */
2064 static int dl_overflow(struct task_struct *p, int policy,
2065                        const struct sched_attr *attr)
2066 {
2067 
2068         struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2069         u64 period = attr->sched_period ?: attr->sched_deadline;
2070         u64 runtime = attr->sched_runtime;
2071         u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2072         int cpus, err = -1;
2073 
2074         if (new_bw == p->dl.dl_bw)
2075                 return 0;
2076 
2077         /*
2078          * Either if a task, enters, leave, or stays -deadline but changes
2079          * its parameters, we may need to update accordingly the total
2080          * allocated bandwidth of the container.
2081          */
2082         raw_spin_lock(&dl_b->lock);
2083         cpus = dl_bw_cpus(task_cpu(p));
2084         if (dl_policy(policy) && !task_has_dl_policy(p) &&
2085             !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2086                 __dl_add(dl_b, new_bw);
2087                 err = 0;
2088         } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2089                    !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2090                 __dl_clear(dl_b, p->dl.dl_bw);
2091                 __dl_add(dl_b, new_bw);
2092                 err = 0;
2093         } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2094                 __dl_clear(dl_b, p->dl.dl_bw);
2095                 err = 0;
2096         }
2097         raw_spin_unlock(&dl_b->lock);
2098 
2099         return err;
2100 }
2101 
2102 extern void init_dl_bw(struct dl_bw *dl_b);
2103 
2104 /*
2105  * wake_up_new_task - wake up a newly created task for the first time.
2106  *
2107  * This function will do some initial scheduler statistics housekeeping
2108  * that must be done for every newly created context, then puts the task
2109  * on the runqueue and wakes it.
2110  */
2111 void wake_up_new_task(struct task_struct *p)
2112 {
2113         unsigned long flags;
2114         struct rq *rq;
2115 
2116         raw_spin_lock_irqsave(&p->pi_lock, flags);
2117 #ifdef CONFIG_SMP
2118         /*
2119          * Fork balancing, do it here and not earlier because:
2120          *  - cpus_allowed can change in the fork path
2121          *  - any previously selected cpu might disappear through hotplug
2122          */
2123         set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2124 #endif
2125 
2126         /* Initialize new task's runnable average */
2127         init_task_runnable_average(p);
2128         rq = __task_rq_lock(p);
2129         activate_task(rq, p, 0);
2130         p->on_rq = TASK_ON_RQ_QUEUED;
2131         trace_sched_wakeup_new(p, true);
2132         check_preempt_curr(rq, p, WF_FORK);
2133 #ifdef CONFIG_SMP
2134         if (p->sched_class->task_woken)
2135                 p->sched_class->task_woken(rq, p);
2136 #endif
2137         task_rq_unlock(rq, p, &flags);
2138 }
2139 
2140 #ifdef CONFIG_PREEMPT_NOTIFIERS
2141 
2142 /**
2143  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2144  * @notifier: notifier struct to register
2145  */
2146 void preempt_notifier_register(struct preempt_notifier *notifier)
2147 {
2148         hlist_add_head(&notifier->link, &current->preempt_notifiers);
2149 }
2150 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2151 
2152 /**
2153  * preempt_notifier_unregister - no longer interested in preemption notifications
2154  * @notifier: notifier struct to unregister
2155  *
2156  * This is safe to call from within a preemption notifier.
2157  */
2158 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2159 {
2160         hlist_del(&notifier->link);
2161 }
2162 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2163 
2164 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2165 {
2166         struct preempt_notifier *notifier;
2167 
2168         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2169                 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2170 }
2171 
2172 static void
2173 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2174                                  struct task_struct *next)
2175 {
2176         struct preempt_notifier *notifier;
2177 
2178         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2179                 notifier->ops->sched_out(notifier, next);
2180 }
2181 
2182 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2183 
2184 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2185 {
2186 }
2187 
2188 static void
2189 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2190                                  struct task_struct *next)
2191 {
2192 }
2193 
2194 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2195 
2196 /**
2197  * prepare_task_switch - prepare to switch tasks
2198  * @rq: the runqueue preparing to switch
2199  * @prev: the current task that is being switched out
2200  * @next: the task we are going to switch to.
2201  *
2202  * This is called with the rq lock held and interrupts off. It must
2203  * be paired with a subsequent finish_task_switch after the context
2204  * switch.
2205  *
2206  * prepare_task_switch sets up locking and calls architecture specific
2207  * hooks.
2208  */
2209 static inline void
2210 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2211                     struct task_struct *next)
2212 {
2213         trace_sched_switch(prev, next);
2214         sched_info_switch(rq, prev, next);
2215         perf_event_task_sched_out(prev, next);
2216         fire_sched_out_preempt_notifiers(prev, next);
2217         prepare_lock_switch(rq, next);
2218         prepare_arch_switch(next);
2219 }
2220 
2221 /**
2222  * finish_task_switch - clean up after a task-switch
2223  * @rq: runqueue associated with task-switch
2224  * @prev: the thread we just switched away from.
2225  *
2226  * finish_task_switch must be called after the context switch, paired
2227  * with a prepare_task_switch call before the context switch.
2228  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2229  * and do any other architecture-specific cleanup actions.
2230  *
2231  * Note that we may have delayed dropping an mm in context_switch(). If
2232  * so, we finish that here outside of the runqueue lock. (Doing it
2233  * with the lock held can cause deadlocks; see schedule() for
2234  * details.)
2235  */
2236 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2237         __releases(rq->lock)
2238 {
2239         struct mm_struct *mm = rq->prev_mm;
2240         long prev_state;
2241 
2242         rq->prev_mm = NULL;
2243 
2244         /*
2245          * A task struct has one reference for the use as "current".
2246          * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2247          * schedule one last time. The schedule call will never return, and
2248          * the scheduled task must drop that reference.
2249          * The test for TASK_DEAD must occur while the runqueue locks are
2250          * still held, otherwise prev could be scheduled on another cpu, die
2251          * there before we look at prev->state, and then the reference would
2252          * be dropped twice.
2253          *              Manfred Spraul <manfred@colorfullife.com>
2254          */
2255         prev_state = prev->state;
2256         vtime_task_switch(prev);
2257         finish_arch_switch(prev);
2258         perf_event_task_sched_in(prev, current);
2259         finish_lock_switch(rq, prev);
2260         finish_arch_post_lock_switch();
2261 
2262         fire_sched_in_preempt_notifiers(current);
2263         if (mm)
2264                 mmdrop(mm);
2265         if (unlikely(prev_state == TASK_DEAD)) {
2266                 if (prev->sched_class->task_dead)
2267                         prev->sched_class->task_dead(prev);
2268 
2269                 /*
2270                  * Remove function-return probe instances associated with this
2271                  * task and put them back on the free list.
2272                  */
2273                 kprobe_flush_task(prev);
2274                 put_task_struct(prev);
2275         }
2276 
2277         tick_nohz_task_switch(current);
2278 }
2279 
2280 #ifdef CONFIG_SMP
2281 
2282 /* rq->lock is NOT held, but preemption is disabled */
2283 static inline void post_schedule(struct rq *rq)
2284 {
2285         if (rq->post_schedule) {
2286                 unsigned long flags;
2287 
2288                 raw_spin_lock_irqsave(&rq->lock, flags);
2289                 if (rq->curr->sched_class->post_schedule)
2290                         rq->curr->sched_class->post_schedule(rq);
2291                 raw_spin_unlock_irqrestore(&rq->lock, flags);
2292 
2293                 rq->post_schedule = 0;
2294         }
2295 }
2296 
2297 #else
2298 
2299 static inline void post_schedule(struct rq *rq)
2300 {
2301 }
2302 
2303 #endif
2304 
2305 /**
2306  * schedule_tail - first thing a freshly forked thread must call.
2307  * @prev: the thread we just switched away from.
2308  */
2309 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2310         __releases(rq->lock)
2311 {
2312         struct rq *rq = this_rq();
2313 
2314         finish_task_switch(rq, prev);
2315 
2316         /*
2317          * FIXME: do we need to worry about rq being invalidated by the
2318          * task_switch?
2319          */
2320         post_schedule(rq);
2321 
2322         if (current->set_child_tid)
2323                 put_user(task_pid_vnr(current), current->set_child_tid);
2324 }
2325 
2326 /*
2327  * context_switch - switch to the new MM and the new
2328  * thread's register state.
2329  */
2330 static inline void
2331 context_switch(struct rq *rq, struct task_struct *prev,
2332                struct task_struct *next)
2333 {
2334         struct mm_struct *mm, *oldmm;
2335 
2336         prepare_task_switch(rq, prev, next);
2337 
2338         mm = next->mm;
2339         oldmm = prev->active_mm;
2340         /*
2341          * For paravirt, this is coupled with an exit in switch_to to
2342          * combine the page table reload and the switch backend into
2343          * one hypercall.
2344          */
2345         arch_start_context_switch(prev);
2346 
2347         if (!mm) {
2348                 next->active_mm = oldmm;
2349                 atomic_inc(&oldmm->mm_count);
2350                 enter_lazy_tlb(oldmm, next);
2351         } else
2352                 switch_mm(oldmm, mm, next);
2353 
2354         if (!prev->mm) {
2355                 prev->active_mm = NULL;
2356                 rq->prev_mm = oldmm;
2357         }
2358         /*
2359          * Since the runqueue lock will be released by the next
2360          * task (which is an invalid locking op but in the case
2361          * of the scheduler it's an obvious special-case), so we
2362          * do an early lockdep release here:
2363          */
2364         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2365 
2366         context_tracking_task_switch(prev, next);
2367         /* Here we just switch the register state and the stack. */
2368         switch_to(prev, next, prev);
2369 
2370         barrier();
2371         /*
2372          * this_rq must be evaluated again because prev may have moved
2373          * CPUs since it called schedule(), thus the 'rq' on its stack
2374          * frame will be invalid.
2375          */
2376         finish_task_switch(this_rq(), prev);
2377 }
2378 
2379 /*
2380  * nr_running and nr_context_switches:
2381  *
2382  * externally visible scheduler statistics: current number of runnable
2383  * threads, total number of context switches performed since bootup.
2384  */
2385 unsigned long nr_running(void)
2386 {
2387         unsigned long i, sum = 0;
2388 
2389         for_each_online_cpu(i)
2390                 sum += cpu_rq(i)->nr_running;
2391 
2392         return sum;
2393 }
2394 
2395 /*
2396  * Check if only the current task is running on the cpu.
2397  */
2398 bool single_task_running(void)
2399 {
2400         if (cpu_rq(smp_processor_id())->nr_running == 1)
2401                 return true;
2402         else
2403                 return false;
2404 }
2405 EXPORT_SYMBOL(single_task_running);
2406 
2407 unsigned long long nr_context_switches(void)
2408 {
2409         int i;
2410         unsigned long long sum = 0;
2411 
2412         for_each_possible_cpu(i)
2413                 sum += cpu_rq(i)->nr_switches;
2414 
2415         return sum;
2416 }
2417 
2418 unsigned long nr_iowait(void)
2419 {
2420         unsigned long i, sum = 0;
2421 
2422         for_each_possible_cpu(i)
2423                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2424 
2425         return sum;
2426 }
2427 
2428 unsigned long nr_iowait_cpu(int cpu)
2429 {
2430         struct rq *this = cpu_rq(cpu);
2431         return atomic_read(&this->nr_iowait);
2432 }
2433 
2434 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2435 {
2436         struct rq *this = this_rq();
2437         *nr_waiters = atomic_read(&this->nr_iowait);
2438         *load = this->cpu_load[0];
2439 }
2440 
2441 #ifdef CONFIG_SMP
2442 
2443 /*
2444  * sched_exec - execve() is a valuable balancing opportunity, because at
2445  * this point the task has the smallest effective memory and cache footprint.
2446  */
2447 void sched_exec(void)
2448 {
2449         struct task_struct *p = current;
2450         unsigned long flags;
2451         int dest_cpu;
2452 
2453         raw_spin_lock_irqsave(&p->pi_lock, flags);
2454         dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2455         if (dest_cpu == smp_processor_id())
2456                 goto unlock;
2457 
2458         if (likely(cpu_active(dest_cpu))) {
2459                 struct migration_arg arg = { p, dest_cpu };
2460 
2461                 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2462                 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2463                 return;
2464         }
2465 unlock:
2466         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2467 }
2468 
2469 #endif
2470 
2471 DEFINE_PER_CPU(struct kernel_stat, kstat);
2472 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2473 
2474 EXPORT_PER_CPU_SYMBOL(kstat);
2475 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2476 
2477 /*
2478  * Return accounted runtime for the task.
2479  * In case the task is currently running, return the runtime plus current's
2480  * pending runtime that have not been accounted yet.
2481  */
2482 unsigned long long task_sched_runtime(struct task_struct *p)
2483 {
2484         unsigned long flags;
2485         struct rq *rq;
2486         u64 ns;
2487 
2488 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2489         /*
2490          * 64-bit doesn't need locks to atomically read a 64bit value.
2491          * So we have a optimization chance when the task's delta_exec is 0.
2492          * Reading ->on_cpu is racy, but this is ok.
2493          *
2494          * If we race with it leaving cpu, we'll take a lock. So we're correct.
2495          * If we race with it entering cpu, unaccounted time is 0. This is
2496          * indistinguishable from the read occurring a few cycles earlier.
2497          * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2498          * been accounted, so we're correct here as well.
2499          */
2500         if (!p->on_cpu || !task_on_rq_queued(p))
2501                 return p->se.sum_exec_runtime;
2502 #endif
2503 
2504         rq = task_rq_lock(p, &flags);
2505         /*
2506          * Must be ->curr _and_ ->on_rq.  If dequeued, we would
2507          * project cycles that may never be accounted to this
2508          * thread, breaking clock_gettime().
2509          */
2510         if (task_current(rq, p) && task_on_rq_queued(p)) {
2511                 update_rq_clock(rq);
2512                 p->sched_class->update_curr(rq);
2513         }
2514         ns = p->se.sum_exec_runtime;
2515         task_rq_unlock(rq, p, &flags);
2516 
2517         return ns;
2518 }
2519 
2520 /*
2521  * This function gets called by the timer code, with HZ frequency.
2522  * We call it with interrupts disabled.
2523  */
2524 void scheduler_tick(void)
2525 {
2526         int cpu = smp_processor_id();
2527         struct rq *rq = cpu_rq(cpu);
2528         struct task_struct *curr = rq->curr;
2529 
2530         sched_clock_tick();
2531 
2532         raw_spin_lock(&rq->lock);
2533         update_rq_clock(rq);
2534         curr->sched_class->task_tick(rq, curr, 0);
2535         update_cpu_load_active(rq);
2536         raw_spin_unlock(&rq->lock);
2537 
2538         perf_event_task_tick();
2539 
2540 #ifdef CONFIG_SMP
2541         rq->idle_balance = idle_cpu(cpu);
2542         trigger_load_balance(rq);
2543 #endif
2544         rq_last_tick_reset(rq);
2545 }
2546 
2547 #ifdef CONFIG_NO_HZ_FULL
2548 /**
2549  * scheduler_tick_max_deferment
2550  *
2551  * Keep at least one tick per second when a single
2552  * active task is running because the scheduler doesn't
2553  * yet completely support full dynticks environment.
2554  *
2555  * This makes sure that uptime, CFS vruntime, load
2556  * balancing, etc... continue to move forward, even
2557  * with a very low granularity.
2558  *
2559  * Return: Maximum deferment in nanoseconds.
2560  */
2561 u64 scheduler_tick_max_deferment(void)
2562 {
2563         struct rq *rq = this_rq();
2564         unsigned long next, now = ACCESS_ONCE(jiffies);
2565 
2566         next = rq->last_sched_tick + HZ;
2567 
2568         if (time_before_eq(next, now))
2569                 return 0;
2570 
2571         return jiffies_to_nsecs(next - now);
2572 }
2573 #endif
2574 
2575 notrace unsigned long get_parent_ip(unsigned long addr)
2576 {
2577         if (in_lock_functions(addr)) {
2578                 addr = CALLER_ADDR2;
2579                 if (in_lock_functions(addr))
2580                         addr = CALLER_ADDR3;
2581         }
2582         return addr;
2583 }
2584 
2585 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2586                                 defined(CONFIG_PREEMPT_TRACER))
2587 
2588 void preempt_count_add(int val)
2589 {
2590 #ifdef CONFIG_DEBUG_PREEMPT
2591         /*
2592          * Underflow?
2593          */
2594         if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2595                 return;
2596 #endif
2597         __preempt_count_add(val);
2598 #ifdef CONFIG_DEBUG_PREEMPT
2599         /*
2600          * Spinlock count overflowing soon?
2601          */
2602         DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2603                                 PREEMPT_MASK - 10);
2604 #endif
2605         if (preempt_count() == val) {
2606                 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2607 #ifdef CONFIG_DEBUG_PREEMPT
2608                 current->preempt_disable_ip = ip;
2609 #endif
2610                 trace_preempt_off(CALLER_ADDR0, ip);
2611         }
2612 }
2613 EXPORT_SYMBOL(preempt_count_add);
2614 NOKPROBE_SYMBOL(preempt_count_add);
2615 
2616 void preempt_count_sub(int val)
2617 {
2618 #ifdef CONFIG_DEBUG_PREEMPT
2619         /*
2620          * Underflow?
2621          */
2622         if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2623                 return;
2624         /*
2625          * Is the spinlock portion underflowing?
2626          */
2627         if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2628                         !(preempt_count() & PREEMPT_MASK)))
2629                 return;
2630 #endif
2631 
2632         if (preempt_count() == val)
2633                 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2634         __preempt_count_sub(val);
2635 }
2636 EXPORT_SYMBOL(preempt_count_sub);
2637 NOKPROBE_SYMBOL(preempt_count_sub);
2638 
2639 #endif
2640 
2641 /*
2642  * Print scheduling while atomic bug:
2643  */
2644 static noinline void __schedule_bug(struct task_struct *prev)
2645 {
2646         if (oops_in_progress)
2647                 return;
2648 
2649         printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2650                 prev->comm, prev->pid, preempt_count());
2651 
2652         debug_show_held_locks(prev);
2653         print_modules();
2654         if (irqs_disabled())
2655                 print_irqtrace_events(prev);
2656 #ifdef CONFIG_DEBUG_PREEMPT
2657         if (in_atomic_preempt_off()) {
2658                 pr_err("Preemption disabled at:");
2659                 print_ip_sym(current->preempt_disable_ip);
2660                 pr_cont("\n");
2661         }
2662 #endif
2663         dump_stack();
2664         add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2665 }
2666 
2667 /*
2668  * Various schedule()-time debugging checks and statistics:
2669  */
2670 static inline void schedule_debug(struct task_struct *prev)
2671 {
2672 #ifdef CONFIG_SCHED_STACK_END_CHECK
2673         BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2674 #endif
2675         /*
2676          * Test if we are atomic. Since do_exit() needs to call into
2677          * schedule() atomically, we ignore that path. Otherwise whine
2678          * if we are scheduling when we should not.
2679          */
2680         if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2681                 __schedule_bug(prev);
2682         rcu_sleep_check();
2683 
2684         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2685 
2686         schedstat_inc(this_rq(), sched_count);
2687 }
2688 
2689 /*
2690  * Pick up the highest-prio task:
2691  */
2692 static inline struct task_struct *
2693 pick_next_task(struct rq *rq, struct task_struct *prev)
2694 {
2695         const struct sched_class *class = &fair_sched_class;
2696         struct task_struct *p;
2697 
2698         /*
2699          * Optimization: we know that if all tasks are in
2700          * the fair class we can call that function directly:
2701          */
2702         if (likely(prev->sched_class == class &&
2703                    rq->nr_running == rq->cfs.h_nr_running)) {
2704                 p = fair_sched_class.pick_next_task(rq, prev);
2705                 if (unlikely(p == RETRY_TASK))
2706                         goto again;
2707 
2708                 /* assumes fair_sched_class->next == idle_sched_class */
2709                 if (unlikely(!p))
2710                         p = idle_sched_class.pick_next_task(rq, prev);
2711 
2712                 return p;
2713         }
2714 
2715 again:
2716         for_each_class(class) {
2717                 p = class->pick_next_task(rq, prev);
2718                 if (p) {
2719                         if (unlikely(p == RETRY_TASK))
2720                                 goto again;
2721                         return p;
2722                 }
2723         }
2724 
2725         BUG(); /* the idle class will always have a runnable task */
2726 }
2727 
2728 /*
2729  * __schedule() is the main scheduler function.
2730  *
2731  * The main means of driving the scheduler and thus entering this function are:
2732  *
2733  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2734  *
2735  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2736  *      paths. For example, see arch/x86/entry_64.S.
2737  *
2738  *      To drive preemption between tasks, the scheduler sets the flag in timer
2739  *      interrupt handler scheduler_tick().
2740  *
2741  *   3. Wakeups don't really cause entry into schedule(). They add a
2742  *      task to the run-queue and that's it.
2743  *
2744  *      Now, if the new task added to the run-queue preempts the current
2745  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2746  *      called on the nearest possible occasion:
2747  *
2748  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2749  *
2750  *         - in syscall or exception context, at the next outmost
2751  *           preempt_enable(). (this might be as soon as the wake_up()'s
2752  *           spin_unlock()!)
2753  *
2754  *         - in IRQ context, return from interrupt-handler to
2755  *           preemptible context
2756  *
2757  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2758  *         then at the next:
2759  *
2760  *          - cond_resched() call
2761  *          - explicit schedule() call
2762  *          - return from syscall or exception to user-space
2763  *          - return from interrupt-handler to user-space
2764  */
2765 static void __sched __schedule(void)
2766 {
2767         struct task_struct *prev, *next;
2768         unsigned long *switch_count;
2769         struct rq *rq;
2770         int cpu;
2771 
2772 need_resched:
2773         preempt_disable();
2774         cpu = smp_processor_id();
2775         rq = cpu_rq(cpu);
2776         rcu_note_context_switch(cpu);
2777         prev = rq->curr;
2778 
2779         schedule_debug(prev);
2780 
2781         if (sched_feat(HRTICK))
2782                 hrtick_clear(rq);
2783 
2784         /*
2785          * Make sure that signal_pending_state()->signal_pending() below
2786          * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2787          * done by the caller to avoid the race with signal_wake_up().
2788          */
2789         smp_mb__before_spinlock();
2790         raw_spin_lock_irq(&rq->lock);
2791 
2792         switch_count = &prev->nivcsw;
2793         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2794                 if (unlikely(signal_pending_state(prev->state, prev))) {
2795                         prev->state = TASK_RUNNING;
2796                 } else {
2797                         deactivate_task(rq, prev, DEQUEUE_SLEEP);
2798                         prev->on_rq = 0;
2799 
2800                         /*
2801                          * If a worker went to sleep, notify and ask workqueue
2802                          * whether it wants to wake up a task to maintain
2803                          * concurrency.
2804                          */
2805                         if (prev->flags & PF_WQ_WORKER) {
2806                                 struct task_struct *to_wakeup;
2807 
2808                                 to_wakeup = wq_worker_sleeping(prev, cpu);
2809                                 if (to_wakeup)
2810                                         try_to_wake_up_local(to_wakeup);
2811                         }
2812                 }
2813                 switch_count = &prev->nvcsw;
2814         }
2815 
2816         if (task_on_rq_queued(prev) || rq->skip_clock_update < 0)
2817                 update_rq_clock(rq);
2818 
2819         next = pick_next_task(rq, prev);
2820         clear_tsk_need_resched(prev);
2821         clear_preempt_need_resched();
2822         rq->skip_clock_update = 0;
2823 
2824         if (likely(prev != next)) {
2825                 rq->nr_switches++;
2826                 rq->curr = next;
2827                 ++*switch_count;
2828 
2829                 context_switch(rq, prev, next); /* unlocks the rq */
2830                 /*
2831                  * The context switch have flipped the stack from under us
2832                  * and restored the local variables which were saved when
2833                  * this task called schedule() in the past. prev == current
2834                  * is still correct, but it can be moved to another cpu/rq.
2835                  */
2836                 cpu = smp_processor_id();
2837                 rq = cpu_rq(cpu);
2838         } else
2839                 raw_spin_unlock_irq(&rq->lock);
2840 
2841         post_schedule(rq);
2842 
2843         sched_preempt_enable_no_resched();
2844         if (need_resched())
2845                 goto need_resched;
2846 }
2847 
2848 static inline void sched_submit_work(struct task_struct *tsk)
2849 {
2850         if (!tsk->state || tsk_is_pi_blocked(tsk))
2851                 return;
2852         /*
2853          * If we are going to sleep and we have plugged IO queued,
2854          * make sure to submit it to avoid deadlocks.
2855          */
2856         if (blk_needs_flush_plug(tsk))
2857                 blk_schedule_flush_plug(tsk);
2858 }
2859 
2860 asmlinkage __visible void __sched schedule(void)
2861 {
2862         struct task_struct *tsk = current;
2863 
2864         sched_submit_work(tsk);
2865         __schedule();
2866 }
2867 EXPORT_SYMBOL(schedule);
2868 
2869 #ifdef CONFIG_CONTEXT_TRACKING
2870 asmlinkage __visible void __sched schedule_user(void)
2871 {
2872         /*
2873          * If we come here after a random call to set_need_resched(),
2874          * or we have been woken up remotely but the IPI has not yet arrived,
2875          * we haven't yet exited the RCU idle mode. Do it here manually until
2876          * we find a better solution.
2877          *
2878          * NB: There are buggy callers of this function.  Ideally we
2879          * should warn if prev_state != IN_USER, but that will trigger
2880          * too frequently to make sense yet.
2881          */
2882         enum ctx_state prev_state = exception_enter();
2883         schedule();
2884         exception_exit(prev_state);
2885 }
2886 #endif
2887 
2888 /**
2889  * schedule_preempt_disabled - called with preemption disabled
2890  *
2891  * Returns with preemption disabled. Note: preempt_count must be 1
2892  */
2893 void __sched schedule_preempt_disabled(void)
2894 {
2895         sched_preempt_enable_no_resched();
2896         schedule();
2897         preempt_disable();
2898 }
2899 
2900 #ifdef CONFIG_PREEMPT
2901 /*
2902  * this is the entry point to schedule() from in-kernel preemption
2903  * off of preempt_enable. Kernel preemptions off return from interrupt
2904  * occur there and call schedule directly.
2905  */
2906 asmlinkage __visible void __sched notrace preempt_schedule(void)
2907 {
2908         /*
2909          * If there is a non-zero preempt_count or interrupts are disabled,
2910          * we do not want to preempt the current task. Just return..
2911          */
2912         if (likely(!preemptible()))
2913                 return;
2914 
2915         do {
2916                 __preempt_count_add(PREEMPT_ACTIVE);
2917                 __schedule();
2918                 __preempt_count_sub(PREEMPT_ACTIVE);
2919 
2920                 /*
2921                  * Check again in case we missed a preemption opportunity
2922                  * between schedule and now.
2923                  */
2924                 barrier();
2925         } while (need_resched());
2926 }
2927 NOKPROBE_SYMBOL(preempt_schedule);
2928 EXPORT_SYMBOL(preempt_schedule);
2929 
2930 #ifdef CONFIG_CONTEXT_TRACKING
2931 /**
2932  * preempt_schedule_context - preempt_schedule called by tracing
2933  *
2934  * The tracing infrastructure uses preempt_enable_notrace to prevent
2935  * recursion and tracing preempt enabling caused by the tracing
2936  * infrastructure itself. But as tracing can happen in areas coming
2937  * from userspace or just about to enter userspace, a preempt enable
2938  * can occur before user_exit() is called. This will cause the scheduler
2939  * to be called when the system is still in usermode.
2940  *
2941  * To prevent this, the preempt_enable_notrace will use this function
2942  * instead of preempt_schedule() to exit user context if needed before
2943  * calling the scheduler.
2944  */
2945 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2946 {
2947         enum ctx_state prev_ctx;
2948 
2949         if (likely(!preemptible()))
2950                 return;
2951 
2952         do {
2953                 __preempt_count_add(PREEMPT_ACTIVE);
2954                 /*
2955                  * Needs preempt disabled in case user_exit() is traced
2956                  * and the tracer calls preempt_enable_notrace() causing
2957                  * an infinite recursion.
2958                  */
2959                 prev_ctx = exception_enter();
2960                 __schedule();
2961                 exception_exit(prev_ctx);
2962 
2963                 __preempt_count_sub(PREEMPT_ACTIVE);
2964                 barrier();
2965         } while (need_resched());
2966 }
2967 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2968 #endif /* CONFIG_CONTEXT_TRACKING */
2969 
2970 #endif /* CONFIG_PREEMPT */
2971 
2972 /*
2973  * this is the entry point to schedule() from kernel preemption
2974  * off of irq context.
2975  * Note, that this is called and return with irqs disabled. This will
2976  * protect us against recursive calling from irq.
2977  */
2978 asmlinkage __visible void __sched preempt_schedule_irq(void)
2979 {
2980         enum ctx_state prev_state;
2981 
2982         /* Catch callers which need to be fixed */
2983         BUG_ON(preempt_count() || !irqs_disabled());
2984 
2985         prev_state = exception_enter();
2986 
2987         do {
2988                 __preempt_count_add(PREEMPT_ACTIVE);
2989                 local_irq_enable();
2990                 __schedule();
2991                 local_irq_disable();
2992                 __preempt_count_sub(PREEMPT_ACTIVE);
2993 
2994                 /*
2995                  * Check again in case we missed a preemption opportunity
2996                  * between schedule and now.
2997                  */
2998                 barrier();
2999         } while (need_resched());
3000 
3001         exception_exit(prev_state);
3002 }
3003 
3004 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3005                           void *key)
3006 {
3007         return try_to_wake_up(curr->private, mode, wake_flags);
3008 }
3009 EXPORT_SYMBOL(default_wake_function);
3010 
3011 #ifdef CONFIG_RT_MUTEXES
3012 
3013 /*
3014  * rt_mutex_setprio - set the current priority of a task
3015  * @p: task
3016  * @prio: prio value (kernel-internal form)
3017  *
3018  * This function changes the 'effective' priority of a task. It does
3019  * not touch ->normal_prio like __setscheduler().
3020  *
3021  * Used by the rt_mutex code to implement priority inheritance
3022  * logic. Call site only calls if the priority of the task changed.
3023  */
3024 void rt_mutex_setprio(struct task_struct *p, int prio)
3025 {
3026         int oldprio, queued, running, enqueue_flag = 0;
3027         struct rq *rq;
3028         const struct sched_class *prev_class;
3029 
3030         BUG_ON(prio > MAX_PRIO);
3031 
3032         rq = __task_rq_lock(p);
3033 
3034         /*
3035          * Idle task boosting is a nono in general. There is one
3036          * exception, when PREEMPT_RT and NOHZ is active:
3037          *
3038          * The idle task calls get_next_timer_interrupt() and holds
3039          * the timer wheel base->lock on the CPU and another CPU wants
3040          * to access the timer (probably to cancel it). We can safely
3041          * ignore the boosting request, as the idle CPU runs this code
3042          * with interrupts disabled and will complete the lock
3043          * protected section without being interrupted. So there is no
3044          * real need to boost.
3045          */
3046         if (unlikely(p == rq->idle)) {
3047                 WARN_ON(p != rq->curr);
3048                 WARN_ON(p->pi_blocked_on);
3049                 goto out_unlock;
3050         }
3051 
3052         trace_sched_pi_setprio(p, prio);
3053         oldprio = p->prio;
3054         prev_class = p->sched_class;
3055         queued = task_on_rq_queued(p);
3056         running = task_current(rq, p);
3057         if (queued)
3058                 dequeue_task(rq, p, 0);
3059         if (running)
3060                 put_prev_task(rq, p);
3061 
3062         /*
3063          * Boosting condition are:
3064          * 1. -rt task is running and holds mutex A
3065          *      --> -dl task blocks on mutex A
3066          *
3067          * 2. -dl task is running and holds mutex A
3068          *      --> -dl task blocks on mutex A and could preempt the
3069          *          running task
3070          */
3071         if (dl_prio(prio)) {
3072                 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3073                 if (!dl_prio(p->normal_prio) ||
3074                     (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3075                         p->dl.dl_boosted = 1;
3076                         p->dl.dl_throttled = 0;
3077                         enqueue_flag = ENQUEUE_REPLENISH;
3078                 } else
3079                         p->dl.dl_boosted = 0;
3080                 p->sched_class = &dl_sched_class;
3081         } else if (rt_prio(prio)) {
3082                 if (dl_prio(oldprio))
3083                         p->dl.dl_boosted = 0;
3084                 if (oldprio < prio)
3085                         enqueue_flag = ENQUEUE_HEAD;
3086                 p->sched_class = &rt_sched_class;
3087         } else {
3088                 if (dl_prio(oldprio))
3089                         p->dl.dl_boosted = 0;
3090                 p->sched_class = &fair_sched_class;
3091         }
3092 
3093         p->prio = prio;
3094 
3095         if (running)
3096                 p->sched_class->set_curr_task(rq);
3097         if (queued)
3098                 enqueue_task(rq, p, enqueue_flag);
3099 
3100         check_class_changed(rq, p, prev_class, oldprio);
3101 out_unlock:
3102         __task_rq_unlock(rq);
3103 }
3104 #endif
3105 
3106 void set_user_nice(struct task_struct *p, long nice)
3107 {
3108         int old_prio, delta, queued;
3109         unsigned long flags;
3110         struct rq *rq;
3111 
3112         if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3113                 return;
3114         /*
3115          * We have to be careful, if called from sys_setpriority(),
3116          * the task might be in the middle of scheduling on another CPU.
3117          */
3118         rq = task_rq_lock(p, &flags);
3119         /*
3120          * The RT priorities are set via sched_setscheduler(), but we still
3121          * allow the 'normal' nice value to be set - but as expected
3122          * it wont have any effect on scheduling until the task is
3123          * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3124          */
3125         if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3126                 p->static_prio = NICE_TO_PRIO(nice);
3127                 goto out_unlock;
3128         }
3129         queued = task_on_rq_queued(p);
3130         if (queued)
3131                 dequeue_task(rq, p, 0);
3132 
3133         p->static_prio = NICE_TO_PRIO(nice);
3134         set_load_weight(p);
3135         old_prio = p->prio;
3136         p->prio = effective_prio(p);
3137         delta = p->prio - old_prio;
3138 
3139         if (queued) {
3140                 enqueue_task(rq, p, 0);
3141                 /*
3142                  * If the task increased its priority or is running and
3143                  * lowered its priority, then reschedule its CPU:
3144                  */
3145                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3146                         resched_curr(rq);
3147         }
3148 out_unlock:
3149         task_rq_unlock(rq, p, &flags);
3150 }
3151 EXPORT_SYMBOL(set_user_nice);
3152 
3153 /*
3154  * can_nice - check if a task can reduce its nice value
3155  * @p: task
3156  * @nice: nice value
3157  */
3158 int can_nice(const struct task_struct *p, const int nice)
3159 {
3160         /* convert nice value [19,-20] to rlimit style value [1,40] */
3161         int nice_rlim = nice_to_rlimit(nice);
3162 
3163         return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3164                 capable(CAP_SYS_NICE));
3165 }
3166 
3167 #ifdef __ARCH_WANT_SYS_NICE
3168 
3169 /*
3170  * sys_nice - change the priority of the current process.
3171  * @increment: priority increment
3172  *
3173  * sys_setpriority is a more generic, but much slower function that
3174  * does similar things.
3175  */
3176 SYSCALL_DEFINE1(nice, int, increment)
3177 {
3178         long nice, retval;
3179 
3180         /*
3181          * Setpriority might change our priority at the same moment.
3182          * We don't have to worry. Conceptually one call occurs first
3183          * and we have a single winner.
3184          */
3185         increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3186         nice = task_nice(current) + increment;
3187 
3188         nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3189         if (increment < 0 && !can_nice(current, nice))
3190                 return -EPERM;
3191 
3192         retval = security_task_setnice(current, nice);
3193         if (retval)
3194                 return retval;
3195 
3196         set_user_nice(current, nice);
3197         return 0;
3198 }
3199 
3200 #endif
3201 
3202 /**
3203  * task_prio - return the priority value of a given task.
3204  * @p: the task in question.
3205  *
3206  * Return: The priority value as seen by users in /proc.
3207  * RT tasks are offset by -200. Normal tasks are centered
3208  * around 0, value goes from -16 to +15.
3209  */
3210 int task_prio(const struct task_struct *p)
3211 {
3212         return p->prio - MAX_RT_PRIO;
3213 }
3214 
3215 /**
3216  * idle_cpu - is a given cpu idle currently?
3217  * @cpu: the processor in question.
3218  *
3219  * Return: 1 if the CPU is currently idle. 0 otherwise.
3220  */
3221 int idle_cpu(int cpu)
3222 {
3223         struct rq *rq = cpu_rq(cpu);
3224 
3225         if (rq->curr != rq->idle)
3226                 return 0;
3227 
3228         if (rq->nr_running)
3229                 return 0;
3230 
3231 #ifdef CONFIG_SMP
3232         if (!llist_empty(&rq->wake_list))
3233                 return 0;
3234 #endif
3235 
3236         return 1;
3237 }
3238 
3239 /**
3240  * idle_task - return the idle task for a given cpu.
3241  * @cpu: the processor in question.
3242  *
3243  * Return: The idle task for the cpu @cpu.
3244  */
3245 struct task_struct *idle_task(int cpu)
3246 {
3247         return cpu_rq(cpu)->idle;
3248 }
3249 
3250 /**
3251  * find_process_by_pid - find a process with a matching PID value.
3252  * @pid: the pid in question.
3253  *
3254  * The task of @pid, if found. %NULL otherwise.
3255  */
3256 static struct task_struct *find_process_by_pid(pid_t pid)
3257 {
3258         return pid ? find_task_by_vpid(pid) : current;
3259 }
3260 
3261 /*
3262  * This function initializes the sched_dl_entity of a newly becoming
3263  * SCHED_DEADLINE task.
3264  *
3265  * Only the static values are considered here, the actual runtime and the
3266  * absolute deadline will be properly calculated when the task is enqueued
3267  * for the first time with its new policy.
3268  */
3269 static void
3270 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3271 {
3272         struct sched_dl_entity *dl_se = &p->dl;
3273 
3274         init_dl_task_timer(dl_se);
3275         dl_se->dl_runtime = attr->sched_runtime;
3276         dl_se->dl_deadline = attr->sched_deadline;
3277         dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3278         dl_se->flags = attr->sched_flags;
3279         dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3280         dl_se->dl_throttled = 0;
3281         dl_se->dl_new = 1;
3282         dl_se->dl_yielded = 0;
3283 }
3284 
3285 /*
3286  * sched_setparam() passes in -1 for its policy, to let the functions
3287  * it calls know not to change it.
3288  */
3289 #define SETPARAM_POLICY -1
3290 
3291 static void __setscheduler_params(struct task_struct *p,
3292                 const struct sched_attr *attr)
3293 {
3294         int policy = attr->sched_policy;
3295 
3296         if (policy == SETPARAM_POLICY)
3297                 policy = p->policy;
3298 
3299         p->policy = policy;
3300 
3301         if (dl_policy(policy))
3302                 __setparam_dl(p, attr);
3303         else if (fair_policy(policy))
3304                 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3305 
3306         /*
3307          * __sched_setscheduler() ensures attr->sched_priority == 0 when
3308          * !rt_policy. Always setting this ensures that things like
3309          * getparam()/getattr() don't report silly values for !rt tasks.
3310          */
3311         p->rt_priority = attr->sched_priority;
3312         p->normal_prio = normal_prio(p);
3313         set_load_weight(p);
3314 }
3315 
3316 /* Actually do priority change: must hold pi & rq lock. */
3317 static void __setscheduler(struct rq *rq, struct task_struct *p,
3318                            const struct sched_attr *attr)
3319 {
3320         __setscheduler_params(p, attr);
3321 
3322         /*
3323          * If we get here, there was no pi waiters boosting the
3324          * task. It is safe to use the normal prio.
3325          */
3326         p->prio = normal_prio(p);
3327 
3328         if (dl_prio(p->prio))
3329                 p->sched_class = &dl_sched_class;
3330         else if (rt_prio(p->prio))
3331                 p->sched_class = &rt_sched_class;
3332         else
3333                 p->sched_class = &fair_sched_class;
3334 }
3335 
3336 static void
3337 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3338 {
3339         struct sched_dl_entity *dl_se = &p->dl;
3340 
3341         attr->sched_priority = p->rt_priority;
3342         attr->sched_runtime = dl_se->dl_runtime;
3343         attr->sched_deadline = dl_se->dl_deadline;
3344         attr->sched_period = dl_se->dl_period;
3345         attr->sched_flags = dl_se->flags;
3346 }
3347 
3348 /*
3349  * This function validates the new parameters of a -deadline task.
3350  * We ask for the deadline not being zero, and greater or equal
3351  * than the runtime, as well as the period of being zero or
3352  * greater than deadline. Furthermore, we have to be sure that
3353  * user parameters are above the internal resolution of 1us (we
3354  * check sched_runtime only since it is always the smaller one) and
3355  * below 2^63 ns (we have to check both sched_deadline and
3356  * sched_period, as the latter can be zero).
3357  */
3358 static bool
3359 __checkparam_dl(const struct sched_attr *attr)
3360 {
3361         /* deadline != 0 */
3362         if (attr->sched_deadline == 0)
3363                 return false;
3364 
3365         /*
3366          * Since we truncate DL_SCALE bits, make sure we're at least
3367          * that big.
3368          */
3369         if (attr->sched_runtime < (1ULL << DL_SCALE))
3370                 return false;
3371 
3372         /*
3373          * Since we use the MSB for wrap-around and sign issues, make
3374          * sure it's not set (mind that period can be equal to zero).
3375          */
3376         if (attr->sched_deadline & (1ULL << 63) ||
3377             attr->sched_period & (1ULL << 63))
3378                 return false;
3379 
3380         /* runtime <= deadline <= period (if period != 0) */
3381         if ((attr->sched_period != 0 &&
3382              attr->sched_period < attr->sched_deadline) ||
3383             attr->sched_deadline < attr->sched_runtime)
3384                 return false;
3385 
3386         return true;
3387 }
3388 
3389 /*
3390  * check the target process has a UID that matches the current process's
3391  */
3392 static bool check_same_owner(struct task_struct *p)
3393 {
3394         const struct cred *cred = current_cred(), *pcred;
3395         bool match;
3396 
3397         rcu_read_lock();
3398         pcred = __task_cred(p);
3399         match = (uid_eq(cred->euid, pcred->euid) ||
3400                  uid_eq(cred->euid, pcred->uid));
3401         rcu_read_unlock();
3402         return match;
3403 }
3404 
3405 static int __sched_setscheduler(struct task_struct *p,
3406                                 const struct sched_attr *attr,
3407                                 bool user)
3408 {
3409         int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3410                       MAX_RT_PRIO - 1 - attr->sched_priority;
3411         int retval, oldprio, oldpolicy = -1, queued, running;
3412         int policy = attr->sched_policy;
3413         unsigned long flags;
3414         const struct sched_class *prev_class;
3415         struct rq *rq;
3416         int reset_on_fork;
3417 
3418         /* may grab non-irq protected spin_locks */
3419         BUG_ON(in_interrupt());
3420 recheck:
3421         /* double check policy once rq lock held */
3422         if (policy < 0) {
3423                 reset_on_fork = p->sched_reset_on_fork;
3424                 policy = oldpolicy = p->policy;
3425         } else {
3426                 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3427 
3428                 if (policy != SCHED_DEADLINE &&
3429                                 policy != SCHED_FIFO && policy != SCHED_RR &&
3430                                 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3431                                 policy != SCHED_IDLE)
3432                         return -EINVAL;
3433         }
3434 
3435         if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3436                 return -EINVAL;
3437 
3438         /*
3439          * Valid priorities for SCHED_FIFO and SCHED_RR are
3440          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3441          * SCHED_BATCH and SCHED_IDLE is 0.
3442          */
3443         if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3444             (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3445                 return -EINVAL;
3446         if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3447             (rt_policy(policy) != (attr->sched_priority != 0)))
3448                 return -EINVAL;
3449 
3450         /*
3451          * Allow unprivileged RT tasks to decrease priority:
3452          */
3453         if (user && !capable(CAP_SYS_NICE)) {
3454                 if (fair_policy(policy)) {
3455                         if (attr->sched_nice < task_nice(p) &&
3456                             !can_nice(p, attr->sched_nice))
3457                                 return -EPERM;
3458                 }
3459 
3460                 if (rt_policy(policy)) {
3461                         unsigned long rlim_rtprio =
3462                                         task_rlimit(p, RLIMIT_RTPRIO);
3463 
3464                         /* can't set/change the rt policy */
3465                         if (policy != p->policy && !rlim_rtprio)
3466                                 return -EPERM;
3467 
3468                         /* can't increase priority */
3469                         if (attr->sched_priority > p->rt_priority &&
3470                             attr->sched_priority > rlim_rtprio)
3471                                 return -EPERM;
3472                 }
3473 
3474                  /*
3475                   * Can't set/change SCHED_DEADLINE policy at all for now
3476                   * (safest behavior); in the future we would like to allow
3477                   * unprivileged DL tasks to increase their relative deadline
3478                   * or reduce their runtime (both ways reducing utilization)
3479                   */
3480                 if (dl_policy(policy))
3481                         return -EPERM;
3482 
3483                 /*
3484                  * Treat SCHED_IDLE as nice 20. Only allow a switch to
3485                  * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3486                  */
3487                 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3488                         if (!can_nice(p, task_nice(p)))
3489                                 return -EPERM;
3490                 }
3491 
3492                 /* can't change other user's priorities */
3493                 if (!check_same_owner(p))
3494                         return -EPERM;
3495 
3496                 /* Normal users shall not reset the sched_reset_on_fork flag */
3497                 if (p->sched_reset_on_fork && !reset_on_fork)
3498                         return -EPERM;
3499         }
3500 
3501         if (user) {
3502                 retval = security_task_setscheduler(p);
3503                 if (retval)
3504                         return retval;
3505         }
3506 
3507         /*
3508          * make sure no PI-waiters arrive (or leave) while we are
3509          * changing the priority of the task:
3510          *
3511          * To be able to change p->policy safely, the appropriate
3512          * runqueue lock must be held.
3513          */
3514         rq = task_rq_lock(p, &flags);
3515 
3516         /*
3517          * Changing the policy of the stop threads its a very bad idea
3518          */
3519         if (p == rq->stop) {
3520                 task_rq_unlock(rq, p, &flags);
3521                 return -EINVAL;
3522         }
3523 
3524         /*
3525          * If not changing anything there's no need to proceed further,
3526          * but store a possible modification of reset_on_fork.
3527          */
3528         if (unlikely(policy == p->policy)) {
3529                 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3530                         goto change;
3531                 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3532                         goto change;
3533                 if (dl_policy(policy))
3534                         goto change;
3535 
3536                 p->sched_reset_on_fork = reset_on_fork;
3537                 task_rq_unlock(rq, p, &flags);
3538                 return 0;
3539         }
3540 change:
3541 
3542         if (user) {
3543 #ifdef CONFIG_RT_GROUP_SCHED
3544                 /*
3545                  * Do not allow realtime tasks into groups that have no runtime
3546                  * assigned.
3547                  */
3548                 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3549                                 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3550                                 !task_group_is_autogroup(task_group(p))) {
3551                         task_rq_unlock(rq, p, &flags);
3552                         return -EPERM;
3553                 }
3554 #endif
3555 #ifdef CONFIG_SMP
3556                 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3557                         cpumask_t *span = rq->rd->span;
3558 
3559                         /*
3560                          * Don't allow tasks with an affinity mask smaller than
3561                          * the entire root_domain to become SCHED_DEADLINE. We
3562                          * will also fail if there's no bandwidth available.
3563                          */
3564                         if (!cpumask_subset(span, &p->cpus_allowed) ||
3565                             rq->rd->dl_bw.bw == 0) {
3566                                 task_rq_unlock(rq, p, &flags);
3567                                 return -EPERM;
3568                         }
3569                 }
3570 #endif
3571         }
3572 
3573         /* recheck policy now with rq lock held */
3574         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3575                 policy = oldpolicy = -1;
3576                 task_rq_unlock(rq, p, &flags);
3577                 goto recheck;
3578         }
3579 
3580         /*
3581          * If setscheduling to SCHED_DEADLINE (or changing the parameters
3582          * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3583          * is available.
3584          */
3585         if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3586                 task_rq_unlock(rq, p, &flags);
3587                 return -EBUSY;
3588         }
3589 
3590         p->sched_reset_on_fork = reset_on_fork;
3591         oldprio = p->prio;
3592 
3593         /*
3594          * Special case for priority boosted tasks.
3595          *
3596          * If the new priority is lower or equal (user space view)
3597          * than the current (boosted) priority, we just store the new
3598          * normal parameters and do not touch the scheduler class and
3599          * the runqueue. This will be done when the task deboost
3600          * itself.
3601          */
3602         if (rt_mutex_check_prio(p, newprio)) {
3603                 __setscheduler_params(p, attr);
3604                 task_rq_unlock(rq, p, &flags);
3605                 return 0;
3606         }
3607 
3608         queued = task_on_rq_queued(p);
3609         running = task_current(rq, p);
3610         if (queued)
3611                 dequeue_task(rq, p, 0);
3612         if (running)
3613                 put_prev_task(rq, p);
3614 
3615         prev_class = p->sched_class;
3616         __setscheduler(rq, p, attr);
3617 
3618         if (running)
3619                 p->sched_class->set_curr_task(rq);
3620         if (queued) {
3621                 /*
3622                  * We enqueue to tail when the priority of a task is
3623                  * increased (user space view).
3624                  */
3625                 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3626         }
3627 
3628         check_class_changed(rq, p, prev_class, oldprio);
3629         task_rq_unlock(rq, p, &flags);
3630 
3631         rt_mutex_adjust_pi(p);
3632 
3633         return 0;
3634 }
3635 
3636 static int _sched_setscheduler(struct task_struct *p, int policy,
3637                                const struct sched_param *param, bool check)
3638 {
3639         struct sched_attr attr = {
3640                 .sched_policy   = policy,
3641                 .sched_priority = param->sched_priority,
3642                 .sched_nice     = PRIO_TO_NICE(p->static_prio),
3643         };
3644 
3645         /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3646         if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3647                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3648                 policy &= ~SCHED_RESET_ON_FORK;
3649                 attr.sched_policy = policy;
3650         }
3651 
3652         return __sched_setscheduler(p, &attr, check);
3653 }
3654 /**
3655  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3656  * @p: the task in question.
3657  * @policy: new policy.
3658  * @param: structure containing the new RT priority.
3659  *
3660  * Return: 0 on success. An error code otherwise.
3661  *
3662  * NOTE that the task may be already dead.
3663  */
3664 int sched_setscheduler(struct task_struct *p, int policy,
3665                        const struct sched_param *param)
3666 {
3667         return _sched_setscheduler(p, policy, param, true);
3668 }
3669 EXPORT_SYMBOL_GPL(sched_setscheduler);
3670 
3671 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3672 {
3673         return __sched_setscheduler(p, attr, true);
3674 }
3675 EXPORT_SYMBOL_GPL(sched_setattr);
3676 
3677 /**
3678  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3679  * @p: the task in question.
3680  * @policy: new policy.
3681  * @param: structure containing the new RT priority.
3682  *
3683  * Just like sched_setscheduler, only don't bother checking if the
3684  * current context has permission.  For example, this is needed in
3685  * stop_machine(): we create temporary high priority worker threads,
3686  * but our caller might not have that capability.
3687  *
3688  * Return: 0 on success. An error code otherwise.
3689  */
3690 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3691                                const struct sched_param *param)
3692 {
3693         return _sched_setscheduler(p, policy, param, false);
3694 }
3695 
3696 static int
3697 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3698 {
3699         struct sched_param lparam;
3700         struct task_struct *p;
3701         int retval;
3702 
3703         if (!param || pid < 0)
3704                 return -EINVAL;
3705         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3706                 return -EFAULT;
3707 
3708         rcu_read_lock();
3709         retval = -ESRCH;
3710         p = find_process_by_pid(pid);
3711         if (p != NULL)
3712                 retval = sched_setscheduler(p, policy, &lparam);
3713         rcu_read_unlock();
3714 
3715         return retval;
3716 }
3717 
3718 /*
3719  * Mimics kernel/events/core.c perf_copy_attr().
3720  */
3721 static int sched_copy_attr(struct sched_attr __user *uattr,
3722                            struct sched_attr *attr)
3723 {
3724         u32 size;
3725         int ret;
3726 
3727         if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3728                 return -EFAULT;
3729 
3730         /*
3731          * zero the full structure, so that a short copy will be nice.
3732          */
3733         memset(attr, 0, sizeof(*attr));
3734 
3735         ret = get_user(size, &uattr->size);
3736         if (ret)
3737                 return ret;
3738 
3739         if (size > PAGE_SIZE)   /* silly large */
3740                 goto err_size;
3741 
3742         if (!size)              /* abi compat */
3743                 size = SCHED_ATTR_SIZE_VER0;
3744 
3745         if (size < SCHED_ATTR_SIZE_VER0)
3746                 goto err_size;
3747 
3748         /*
3749          * If we're handed a bigger struct than we know of,
3750          * ensure all the unknown bits are 0 - i.e. new
3751          * user-space does not rely on any kernel feature
3752          * extensions we dont know about yet.
3753          */
3754         if (size > sizeof(*attr)) {
3755                 unsigned char __user *addr;
3756                 unsigned char __user *end;
3757                 unsigned char val;
3758 
3759                 addr = (void __user *)uattr + sizeof(*attr);
3760                 end  = (void __user *)uattr + size;
3761 
3762                 for (; addr < end; addr++) {
3763                         ret = get_user(val, addr);
3764                         if (ret)
3765                                 return ret;
3766                         if (val)
3767                                 goto err_size;
3768                 }
3769                 size = sizeof(*attr);
3770         }
3771 
3772         ret = copy_from_user(attr, uattr, size);
3773         if (ret)
3774                 return -EFAULT;
3775 
3776         /*
3777          * XXX: do we want to be lenient like existing syscalls; or do we want
3778          * to be strict and return an error on out-of-bounds values?
3779          */
3780         attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3781 
3782         return 0;
3783 
3784 err_size:
3785         put_user(sizeof(*attr), &uattr->size);
3786         return -E2BIG;
3787 }
3788 
3789 /**
3790  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3791  * @pid: the pid in question.
3792  * @policy: new policy.
3793  * @param: structure containing the new RT priority.
3794  *
3795  * Return: 0 on success. An error code otherwise.
3796  */
3797 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3798                 struct sched_param __user *, param)
3799 {
3800         /* negative values for policy are not valid */
3801         if (policy < 0)
3802                 return -EINVAL;
3803 
3804         return do_sched_setscheduler(pid, policy, param);
3805 }
3806 
3807 /**
3808  * sys_sched_setparam - set/change the RT priority of a thread
3809  * @pid: the pid in question.
3810  * @param: structure containing the new RT priority.
3811  *
3812  * Return: 0 on success. An error code otherwise.
3813  */
3814 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3815 {
3816         return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3817 }
3818 
3819 /**
3820  * sys_sched_setattr - same as above, but with extended sched_attr
3821  * @pid: the pid in question.
3822  * @uattr: structure containing the extended parameters.
3823  * @flags: for future extension.
3824  */
3825 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3826                                unsigned int, flags)
3827 {
3828         struct sched_attr attr;
3829         struct task_struct *p;
3830         int retval;
3831 
3832         if (!uattr || pid < 0 || flags)
3833                 return -EINVAL;
3834 
3835         retval = sched_copy_attr(uattr, &attr);
3836         if (retval)
3837                 return retval;
3838 
3839         if ((int)attr.sched_policy < 0)
3840                 return -EINVAL;
3841 
3842         rcu_read_lock();
3843         retval = -ESRCH;
3844         p = find_process_by_pid(pid);
3845         if (p != NULL)
3846                 retval = sched_setattr(p, &attr);
3847         rcu_read_unlock();
3848 
3849         return retval;
3850 }
3851 
3852 /**
3853  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3854  * @pid: the pid in question.
3855  *
3856  * Return: On success, the policy of the thread. Otherwise, a negative error
3857  * code.
3858  */
3859 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3860 {
3861         struct task_struct *p;
3862         int retval;
3863 
3864         if (pid < 0)
3865                 return -EINVAL;
3866 
3867         retval = -ESRCH;
3868         rcu_read_lock();
3869         p = find_process_by_pid(pid);
3870         if (p) {
3871                 retval = security_task_getscheduler(p);
3872                 if (!retval)
3873                         retval = p->policy
3874                                 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3875         }
3876         rcu_read_unlock();
3877         return retval;
3878 }
3879 
3880 /**
3881  * sys_sched_getparam - get the RT priority of a thread
3882  * @pid: the pid in question.
3883  * @param: structure containing the RT priority.
3884  *
3885  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3886  * code.
3887  */
3888 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3889 {
3890         struct sched_param lp = { .sched_priority = 0 };
3891         struct task_struct *p;
3892         int retval;
3893 
3894         if (!param || pid < 0)
3895                 return -EINVAL;
3896 
3897         rcu_read_lock();
3898         p = find_process_by_pid(pid);
3899         retval = -ESRCH;
3900         if (!p)
3901                 goto out_unlock;
3902 
3903         retval = security_task_getscheduler(p);
3904         if (retval)
3905                 goto out_unlock;
3906 
3907         if (task_has_rt_policy(p))
3908                 lp.sched_priority = p->rt_priority;
3909         rcu_read_unlock();
3910 
3911         /*
3912          * This one might sleep, we cannot do it with a spinlock held ...
3913          */
3914         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3915 
3916         return retval;
3917 
3918 out_unlock:
3919         rcu_read_unlock();
3920         return retval;
3921 }
3922 
3923 static int sched_read_attr(struct sched_attr __user *uattr,
3924                            struct sched_attr *attr,
3925                            unsigned int usize)
3926 {
3927         int ret;
3928 
3929         if (!access_ok(VERIFY_WRITE, uattr, usize))
3930                 return -EFAULT;
3931 
3932         /*
3933          * If we're handed a smaller struct than we know of,
3934          * ensure all the unknown bits are 0 - i.e. old
3935          * user-space does not get uncomplete information.
3936          */
3937         if (usize < sizeof(*attr)) {
3938                 unsigned char *addr;
3939                 unsigned char *end;
3940 
3941                 addr = (void *)attr + usize;
3942                 end  = (void *)attr + sizeof(*attr);
3943 
3944                 for (; addr < end; addr++) {
3945                         if (*addr)
3946                                 return -EFBIG;
3947                 }
3948 
3949                 attr->size = usize;
3950         }
3951 
3952         ret = copy_to_user(uattr, attr, attr->size);
3953         if (ret)
3954                 return -EFAULT;
3955 
3956         return 0;
3957 }
3958 
3959 /**
3960  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3961  * @pid: the pid in question.
3962  * @uattr: structure containing the extended parameters.
3963  * @size: sizeof(attr) for fwd/bwd comp.
3964  * @flags: for future extension.
3965  */
3966 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3967                 unsigned int, size, unsigned int, flags)
3968 {
3969         struct sched_attr attr = {
3970                 .size = sizeof(struct sched_attr),
3971         };
3972         struct task_struct *p;
3973         int retval;
3974 
3975         if (!uattr || pid < 0 || size > PAGE_SIZE ||
3976             size < SCHED_ATTR_SIZE_VER0 || flags)
3977                 return -EINVAL;
3978 
3979         rcu_read_lock();
3980         p = find_process_by_pid(pid);
3981         retval = -ESRCH;
3982         if (!p)
3983                 goto out_unlock;
3984 
3985         retval = security_task_getscheduler(p);
3986         if (retval)
3987                 goto out_unlock;
3988 
3989         attr.sched_policy = p->policy;
3990         if (p->sched_reset_on_fork)
3991                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3992         if (task_has_dl_policy(p))
3993                 __getparam_dl(p, &attr);
3994         else if (task_has_rt_policy(p))
3995                 attr.sched_priority = p->rt_priority;
3996         else
3997                 attr.sched_nice = task_nice(p);
3998 
3999         rcu_read_unlock();
4000 
4001         retval = sched_read_attr(uattr, &attr, size);
4002         return retval;
4003 
4004 out_unlock:
4005         rcu_read_unlock();
4006         return retval;
4007 }
4008 
4009 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4010 {
4011         cpumask_var_t cpus_allowed, new_mask;
4012         struct task_struct *p;
4013         int retval;
4014 
4015         rcu_read_lock();
4016 
4017         p = find_process_by_pid(pid);
4018         if (!p) {
4019                 rcu_read_unlock();
4020                 return -ESRCH;
4021         }
4022 
4023         /* Prevent p going away */
4024         get_task_struct(p);
4025         rcu_read_unlock();
4026 
4027         if (p->flags & PF_NO_SETAFFINITY) {
4028                 retval = -EINVAL;
4029                 goto out_put_task;
4030         }
4031         if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4032                 retval = -ENOMEM;
4033                 goto out_put_task;
4034         }
4035         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4036                 retval = -ENOMEM;
4037                 goto out_free_cpus_allowed;
4038         }
4039         retval = -EPERM;
4040         if (!check_same_owner(p)) {
4041                 rcu_read_lock();
4042                 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4043                         rcu_read_unlock();
4044                         goto out_free_new_mask;
4045                 }
4046                 rcu_read_unlock();
4047         }
4048 
4049         retval = security_task_setscheduler(p);
4050         if (retval)
4051                 goto out_free_new_mask;
4052 
4053 
4054         cpuset_cpus_allowed(p, cpus_allowed);
4055         cpumask_and(new_mask, in_mask, cpus_allowed);
4056 
4057         /*
4058          * Since bandwidth control happens on root_domain basis,
4059          * if admission test is enabled, we only admit -deadline
4060          * tasks allowed to run on all the CPUs in the task's
4061          * root_domain.
4062          */
4063 #ifdef CONFIG_SMP
4064         if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4065                 rcu_read_lock();
4066                 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4067                         retval = -EBUSY;
4068                         rcu_read_unlock();
4069                         goto out_free_new_mask;
4070                 }
4071                 rcu_read_unlock();
4072         }
4073 #endif
4074 again:
4075         retval = set_cpus_allowed_ptr(p, new_mask);
4076 
4077         if (!retval) {
4078                 cpuset_cpus_allowed(p, cpus_allowed);
4079                 if (!cpumask_subset(new_mask, cpus_allowed)) {
4080                         /*
4081                          * We must have raced with a concurrent cpuset
4082                          * update. Just reset the cpus_allowed to the
4083                          * cpuset's cpus_allowed
4084                          */
4085                         cpumask_copy(new_mask, cpus_allowed);
4086                         goto again;
4087                 }
4088         }
4089 out_free_new_mask:
4090         free_cpumask_var(new_mask);
4091 out_free_cpus_allowed:
4092         free_cpumask_var(cpus_allowed);
4093 out_put_task:
4094         put_task_struct(p);
4095         return retval;
4096 }
4097 
4098 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4099                              struct cpumask *new_mask)
4100 {
4101         if (len < cpumask_size())
4102                 cpumask_clear(new_mask);
4103         else if (len > cpumask_size())
4104                 len = cpumask_size();
4105 
4106         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4107 }
4108 
4109 /**
4110  * sys_sched_setaffinity - set the cpu affinity of a process
4111  * @pid: pid of the process
4112  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4113  * @user_mask_ptr: user-space pointer to the new cpu mask
4114  *
4115  * Return: 0 on success. An error code otherwise.
4116  */
4117 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4118                 unsigned long __user *, user_mask_ptr)
4119 {
4120         cpumask_var_t new_mask;
4121         int retval;
4122 
4123         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4124                 return -ENOMEM;
4125 
4126         retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4127         if (retval == 0)
4128                 retval = sched_setaffinity(pid, new_mask);
4129         free_cpumask_var(new_mask);
4130         return retval;
4131 }
4132 
4133 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4134 {
4135         struct task_struct *p;
4136         unsigned long flags;
4137         int retval;
4138 
4139         rcu_read_lock();
4140 
4141         retval = -ESRCH;
4142         p = find_process_by_pid(pid);
4143         if (!p)
4144                 goto out_unlock;
4145 
4146         retval = security_task_getscheduler(p);
4147         if (retval)
4148                 goto out_unlock;
4149 
4150         raw_spin_lock_irqsave(&p->pi_lock, flags);
4151         cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4152         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4153 
4154 out_unlock:
4155         rcu_read_unlock();
4156 
4157         return retval;
4158 }
4159 
4160 /**
4161  * sys_sched_getaffinity - get the cpu affinity of a process
4162  * @pid: pid of the process
4163  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4164  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4165  *
4166  * Return: 0 on success. An error code otherwise.
4167  */
4168 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4169                 unsigned long __user *, user_mask_ptr)
4170 {
4171         int ret;
4172         cpumask_var_t mask;
4173 
4174         if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4175                 return -EINVAL;
4176         if (len & (sizeof(unsigned long)-1))
4177                 return -EINVAL;
4178 
4179         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4180                 return -ENOMEM;
4181 
4182         ret = sched_getaffinity(pid, mask);
4183         if (ret == 0) {
4184                 size_t retlen = min_t(size_t, len, cpumask_size());
4185 
4186                 if (copy_to_user(user_mask_ptr, mask, retlen))
4187                         ret = -EFAULT;
4188                 else
4189                         ret = retlen;
4190         }
4191         free_cpumask_var(mask);
4192 
4193         return ret;
4194 }
4195 
4196 /**
4197  * sys_sched_yield - yield the current processor to other threads.
4198  *
4199  * This function yields the current CPU to other tasks. If there are no
4200  * other threads running on this CPU then this function will return.
4201  *
4202  * Return: 0.
4203  */
4204 SYSCALL_DEFINE0(sched_yield)
4205 {
4206         struct rq *rq = this_rq_lock();
4207 
4208         schedstat_inc(rq, yld_count);
4209         current->sched_class->yield_task(rq);
4210 
4211         /*
4212          * Since we are going to call schedule() anyway, there's
4213          * no need to preempt or enable interrupts:
4214          */
4215         __release(rq->lock);
4216         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4217         do_raw_spin_unlock(&rq->lock);
4218         sched_preempt_enable_no_resched();
4219 
4220         schedule();
4221 
4222         return 0;
4223 }
4224 
4225 static void __cond_resched(void)
4226 {
4227         __preempt_count_add(PREEMPT_ACTIVE);
4228         __schedule();
4229         __preempt_count_sub(PREEMPT_ACTIVE);
4230 }
4231 
4232 int __sched _cond_resched(void)
4233 {
4234         if (should_resched()) {
4235                 __cond_resched();
4236                 return 1;
4237         }
4238         return 0;
4239 }
4240 EXPORT_SYMBOL(_cond_resched);
4241 
4242 /*
4243  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4244  * call schedule, and on return reacquire the lock.
4245  *
4246  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4247  * operations here to prevent schedule() from being called twice (once via
4248  * spin_unlock(), once by hand).
4249  */
4250 int __cond_resched_lock(spinlock_t *lock)
4251 {
4252         int resched = should_resched();
4253         int ret = 0;
4254 
4255         lockdep_assert_held(lock);
4256 
4257         if (spin_needbreak(lock) || resched) {
4258                 spin_unlock(lock);
4259                 if (resched)
4260                         __cond_resched();
4261                 else
4262                         cpu_relax();
4263                 ret = 1;
4264                 spin_lock(lock);
4265         }
4266         return ret;
4267 }
4268 EXPORT_SYMBOL(__cond_resched_lock);
4269 
4270 int __sched __cond_resched_softirq(void)
4271 {
4272         BUG_ON(!in_softirq());
4273 
4274         if (should_resched()) {
4275                 local_bh_enable();
4276                 __cond_resched();
4277                 local_bh_disable();
4278                 return 1;
4279         }
4280         return 0;
4281 }
4282 EXPORT_SYMBOL(__cond_resched_softirq);
4283 
4284 /**
4285  * yield - yield the current processor to other threads.
4286  *
4287  * Do not ever use this function, there's a 99% chance you're doing it wrong.
4288  *
4289  * The scheduler is at all times free to pick the calling task as the most
4290  * eligible task to run, if removing the yield() call from your code breaks
4291  * it, its already broken.
4292  *
4293  * Typical broken usage is:
4294  *
4295  * while (!event)
4296  *      yield();
4297  *
4298  * where one assumes that yield() will let 'the other' process run that will
4299  * make event true. If the current task is a SCHED_FIFO task that will never
4300  * happen. Never use yield() as a progress guarantee!!
4301  *
4302  * If you want to use yield() to wait for something, use wait_event().
4303  * If you want to use yield() to be 'nice' for others, use cond_resched().
4304  * If you still want to use yield(), do not!
4305  */
4306 void __sched yield(void)
4307 {
4308         set_current_state(TASK_RUNNING);
4309         sys_sched_yield();
4310 }
4311 EXPORT_SYMBOL(yield);
4312 
4313 /**
4314  * yield_to - yield the current processor to another thread in
4315  * your thread group, or accelerate that thread toward the
4316  * processor it's on.
4317  * @p: target task
4318  * @preempt: whether task preemption is allowed or not
4319  *
4320  * It's the caller's job to ensure that the target task struct
4321  * can't go away on us before we can do any checks.
4322  *
4323  * Return:
4324  *      true (>0) if we indeed boosted the target task.
4325  *      false (0) if we failed to boost the target.
4326  *      -ESRCH if there's no task to yield to.
4327  */
4328 int __sched yield_to(struct task_struct *p, bool preempt)
4329 {
4330         struct task_struct *curr = current;
4331         struct rq *rq, *p_rq;
4332         unsigned long flags;
4333         int yielded = 0;
4334 
4335         local_irq_save(flags);
4336         rq = this_rq();
4337 
4338 again:
4339         p_rq = task_rq(p);
4340         /*
4341          * If we're the only runnable task on the rq and target rq also
4342          * has only one task, there's absolutely no point in yielding.
4343          */
4344         if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4345                 yielded = -ESRCH;
4346                 goto out_irq;
4347         }
4348 
4349         double_rq_lock(rq, p_rq);
4350         if (task_rq(p) != p_rq) {
4351                 double_rq_unlock(rq, p_rq);
4352                 goto again;
4353         }
4354 
4355         if (!curr->sched_class->yield_to_task)
4356                 goto out_unlock;
4357 
4358         if (curr->sched_class != p->sched_class)
4359                 goto out_unlock;
4360 
4361         if (task_running(p_rq, p) || p->state)
4362                 goto out_unlock;
4363 
4364         yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4365         if (yielded) {
4366                 schedstat_inc(rq, yld_count);
4367                 /*
4368                  * Make p's CPU reschedule; pick_next_entity takes care of
4369                  * fairness.
4370                  */
4371                 if (preempt && rq != p_rq)
4372                         resched_curr(p_rq);
4373         }
4374 
4375 out_unlock:
4376         double_rq_unlock(rq, p_rq);
4377 out_irq:
4378         local_irq_restore(flags);
4379 
4380         if (yielded > 0)
4381                 schedule();
4382 
4383         return yielded;
4384 }
4385 EXPORT_SYMBOL_GPL(yield_to);
4386 
4387 /*
4388  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4389  * that process accounting knows that this is a task in IO wait state.
4390  */
4391 void __sched io_schedule(void)
4392 {
4393         struct rq *rq = raw_rq();
4394 
4395         delayacct_blkio_start();
4396         atomic_inc(&rq->nr_iowait);
4397         blk_flush_plug(current);
4398         current->in_iowait = 1;
4399         schedule();
4400         current->in_iowait = 0;
4401         atomic_dec(&rq->nr_iowait);
4402         delayacct_blkio_end();
4403 }
4404 EXPORT_SYMBOL(io_schedule);
4405 
4406 long __sched io_schedule_timeout(long timeout)
4407 {
4408         struct rq *rq = raw_rq();
4409         long ret;
4410 
4411         delayacct_blkio_start();
4412         atomic_inc(&rq->nr_iowait);
4413         blk_flush_plug(current);
4414         current->in_iowait = 1;
4415         ret = schedule_timeout(timeout);
4416         current->in_iowait = 0;
4417         atomic_dec(&rq->nr_iowait);
4418         delayacct_blkio_end();
4419         return ret;
4420 }
4421 
4422 /**
4423  * sys_sched_get_priority_max - return maximum RT priority.
4424  * @policy: scheduling class.
4425  *
4426  * Return: On success, this syscall returns the maximum
4427  * rt_priority that can be used by a given scheduling class.
4428  * On failure, a negative error code is returned.
4429  */
4430 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4431 {
4432         int ret = -EINVAL;
4433 
4434         switch (policy) {
4435         case SCHED_FIFO:
4436         case SCHED_RR:
4437                 ret = MAX_USER_RT_PRIO-1;
4438                 break;
4439         case SCHED_DEADLINE:
4440         case SCHED_NORMAL:
4441         case SCHED_BATCH:
4442         case SCHED_IDLE:
4443                 ret = 0;
4444                 break;
4445         }
4446         return ret;
4447 }
4448 
4449 /**
4450  * sys_sched_get_priority_min - return minimum RT priority.
4451  * @policy: scheduling class.
4452  *
4453  * Return: On success, this syscall returns the minimum
4454  * rt_priority that can be used by a given scheduling class.
4455  * On failure, a negative error code is returned.
4456  */
4457 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4458 {
4459         int ret = -EINVAL;
4460 
4461         switch (policy) {
4462         case SCHED_FIFO:
4463         case SCHED_RR:
4464                 ret = 1;
4465                 break;
4466         case SCHED_DEADLINE:
4467         case SCHED_NORMAL:
4468         case SCHED_BATCH:
4469         case SCHED_IDLE:
4470                 ret = 0;
4471         }
4472         return ret;
4473 }
4474 
4475 /**
4476  * sys_sched_rr_get_interval - return the default timeslice of a process.
4477  * @pid: pid of the process.
4478  * @interval: userspace pointer to the timeslice value.
4479  *
4480  * this syscall writes the default timeslice value of a given process
4481  * into the user-space timespec buffer. A value of '' means infinity.
4482  *
4483  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4484  * an error code.
4485  */
4486 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4487                 struct timespec __user *, interval)
4488 {
4489         struct task_struct *p;
4490         unsigned int time_slice;
4491         unsigned long flags;
4492         struct rq *rq;
4493         int retval;
4494         struct timespec t;
4495 
4496         if (pid < 0)
4497                 return -EINVAL;
4498 
4499         retval = -ESRCH;
4500         rcu_read_lock();
4501         p = find_process_by_pid(pid);
4502         if (!p)
4503                 goto out_unlock;
4504 
4505         retval = security_task_getscheduler(p);
4506         if (retval)
4507                 goto out_unlock;
4508 
4509         rq = task_rq_lock(p, &flags);
4510         time_slice = 0;
4511         if (p->sched_class->get_rr_interval)
4512                 time_slice = p->sched_class->get_rr_interval(rq, p);
4513         task_rq_unlock(rq, p, &flags);
4514 
4515         rcu_read_unlock();
4516         jiffies_to_timespec(time_slice, &t);
4517         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4518         return retval;
4519 
4520 out_unlock:
4521         rcu_read_unlock();
4522         return retval;
4523 }
4524 
4525 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4526 
4527 void sched_show_task(struct task_struct *p)
4528 {
4529         unsigned long free = 0;
4530         int ppid;
4531         unsigned state;
4532 
4533         state = p->state ? __ffs(p->state) + 1 : 0;
4534         printk(KERN_INFO "%-15.15s %c", p->comm,
4535                 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4536 #if BITS_PER_LONG == 32
4537         if (state == TASK_RUNNING)
4538                 printk(KERN_CONT " running  ");
4539         else
4540                 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4541 #else
4542         if (state == TASK_RUNNING)
4543                 printk(KERN_CONT "  running task    ");
4544         else
4545                 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4546 #endif
4547 #ifdef CONFIG_DEBUG_STACK_USAGE
4548         free = stack_not_used(p);
4549 #endif
4550         rcu_read_lock();
4551         ppid = task_pid_nr(rcu_dereference(p->real_parent));
4552         rcu_read_unlock();
4553         printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4554                 task_pid_nr(p), ppid,
4555                 (unsigned long)task_thread_info(p)->flags);
4556 
4557         print_worker_info(KERN_INFO, p);
4558         show_stack(p, NULL);
4559 }
4560 
4561 void show_state_filter(unsigned long state_filter)
4562 {
4563         struct task_struct *g, *p;
4564 
4565 #if BITS_PER_LONG == 32
4566         printk(KERN_INFO
4567                 "  task                PC stack   pid father\n");
4568 #else
4569         printk(KERN_INFO
4570                 "  task                        PC stack   pid father\n");
4571 #endif
4572         rcu_read_lock();
4573         for_each_process_thread(g, p) {
4574                 /*
4575                  * reset the NMI-timeout, listing all files on a slow
4576                  * console might take a lot of time:
4577                  */
4578                 touch_nmi_watchdog();
4579                 if (!state_filter || (p->state & state_filter))
4580                         sched_show_task(p);
4581         }
4582 
4583         touch_all_softlockup_watchdogs();
4584 
4585 #ifdef CONFIG_SCHED_DEBUG
4586         sysrq_sched_debug_show();
4587 #endif
4588         rcu_read_unlock();
4589         /*
4590          * Only show locks if all tasks are dumped:
4591          */
4592         if (!state_filter)
4593                 debug_show_all_locks();
4594 }
4595 
4596 void init_idle_bootup_task(struct task_struct *idle)
4597 {
4598         idle->sched_class = &idle_sched_class;
4599 }
4600 
4601 /**
4602  * init_idle - set up an idle thread for a given CPU
4603  * @idle: task in question
4604  * @cpu: cpu the idle task belongs to
4605  *
4606  * NOTE: this function does not set the idle thread's NEED_RESCHED
4607  * flag, to make booting more robust.
4608  */
4609 void init_idle(struct task_struct *idle, int cpu)
4610 {
4611         struct rq *rq = cpu_rq(cpu);
4612         unsigned long flags;
4613 
4614         raw_spin_lock_irqsave(&rq->lock, flags);
4615 
4616         __sched_fork(0, idle);
4617         idle->state = TASK_RUNNING;
4618         idle->se.exec_start = sched_clock();
4619 
4620         do_set_cpus_allowed(idle, cpumask_of(cpu));
4621         /*
4622          * We're having a chicken and egg problem, even though we are
4623          * holding rq->lock, the cpu isn't yet set to this cpu so the
4624          * lockdep check in task_group() will fail.
4625          *
4626          * Similar case to sched_fork(). / Alternatively we could
4627          * use task_rq_lock() here and obtain the other rq->lock.
4628          *
4629          * Silence PROVE_RCU
4630          */
4631         rcu_read_lock();
4632         __set_task_cpu(idle, cpu);
4633         rcu_read_unlock();
4634 
4635         rq->curr = rq->idle = idle;
4636         idle->on_rq = TASK_ON_RQ_QUEUED;
4637 #if defined(CONFIG_SMP)
4638         idle->on_cpu = 1;
4639 #endif
4640         raw_spin_unlock_irqrestore(&rq->lock, flags);
4641 
4642         /* Set the preempt count _outside_ the spinlocks! */
4643         init_idle_preempt_count(idle, cpu);
4644 
4645         /*
4646          * The idle tasks have their own, simple scheduling class:
4647          */
4648         idle->sched_class = &idle_sched_class;
4649         ftrace_graph_init_idle_task(idle, cpu);
4650         vtime_init_idle(idle, cpu);
4651 #if defined(CONFIG_SMP)
4652         sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4653 #endif
4654 }
4655 
4656 #ifdef CONFIG_SMP
4657 /*
4658  * move_queued_task - move a queued task to new rq.
4659  *
4660  * Returns (locked) new rq. Old rq's lock is released.
4661  */
4662 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4663 {
4664         struct rq *rq = task_rq(p);
4665 
4666         lockdep_assert_held(&rq->lock);
4667 
4668         dequeue_task(rq, p, 0);
4669         p->on_rq = TASK_ON_RQ_MIGRATING;
4670         set_task_cpu(p, new_cpu);
4671         raw_spin_unlock(&rq->lock);
4672 
4673         rq = cpu_rq(new_cpu);
4674 
4675         raw_spin_lock(&rq->lock);
4676         BUG_ON(task_cpu(p) != new_cpu);
4677         p->on_rq = TASK_ON_RQ_QUEUED;
4678         enqueue_task(rq, p, 0);
4679         check_preempt_curr(rq, p, 0);
4680 
4681         return rq;
4682 }
4683 
4684 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4685 {
4686         if (p->sched_class && p->sched_class->set_cpus_allowed)
4687                 p->sched_class->set_cpus_allowed(p, new_mask);
4688 
4689         cpumask_copy(&p->cpus_allowed, new_mask);
4690         p->nr_cpus_allowed = cpumask_weight(new_mask);
4691 }
4692 
4693 /*
4694  * This is how migration works:
4695  *
4696  * 1) we invoke migration_cpu_stop() on the target CPU using
4697  *    stop_one_cpu().
4698  * 2) stopper starts to run (implicitly forcing the migrated thread
4699  *    off the CPU)
4700  * 3) it checks whether the migrated task is still in the wrong runqueue.
4701  * 4) if it's in the wrong runqueue then the migration thread removes
4702  *    it and puts it into the right queue.
4703  * 5) stopper completes and stop_one_cpu() returns and the migration
4704  *    is done.
4705  */
4706 
4707 /*
4708  * Change a given task's CPU affinity. Migrate the thread to a
4709  * proper CPU and schedule it away if the CPU it's executing on
4710  * is removed from the allowed bitmask.
4711  *
4712  * NOTE: the caller must have a valid reference to the task, the
4713  * task must not exit() & deallocate itself prematurely. The
4714  * call is not atomic; no spinlocks may be held.
4715  */
4716 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4717 {
4718         unsigned long flags;
4719         struct rq *rq;
4720         unsigned int dest_cpu;
4721         int ret = 0;
4722 
4723         rq = task_rq_lock(p, &flags);
4724 
4725         if (cpumask_equal(&p->cpus_allowed, new_mask))
4726                 goto out;
4727 
4728         if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4729                 ret = -EINVAL;
4730                 goto out;
4731         }
4732 
4733         do_set_cpus_allowed(p, new_mask);
4734 
4735         /* Can the task run on the task's current CPU? If so, we're done */
4736         if (cpumask_test_cpu(task_cpu(p), new_mask))
4737                 goto out;
4738 
4739         dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4740         if (task_running(rq, p) || p->state == TASK_WAKING) {
4741                 struct migration_arg arg = { p, dest_cpu };
4742                 /* Need help from migration thread: drop lock and wait. */
4743                 task_rq_unlock(rq, p, &flags);
4744                 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4745                 tlb_migrate_finish(p->mm);
4746                 return 0;
4747         } else if (task_on_rq_queued(p))
4748                 rq = move_queued_task(p, dest_cpu);
4749 out:
4750         task_rq_unlock(rq, p, &flags);
4751 
4752         return ret;
4753 }
4754 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4755 
4756 /*
4757  * Move (not current) task off this cpu, onto dest cpu. We're doing
4758  * this because either it can't run here any more (set_cpus_allowed()
4759  * away from this CPU, or CPU going down), or because we're
4760  * attempting to rebalance this task on exec (sched_exec).
4761  *
4762  * So we race with normal scheduler movements, but that's OK, as long
4763  * as the task is no longer on this CPU.
4764  *
4765  * Returns non-zero if task was successfully migrated.
4766  */
4767 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4768 {
4769         struct rq *rq;
4770         int ret = 0;
4771 
4772         if (unlikely(!cpu_active(dest_cpu)))
4773                 return ret;
4774 
4775         rq = cpu_rq(src_cpu);
4776 
4777         raw_spin_lock(&p->pi_lock);
4778         raw_spin_lock(&rq->lock);
4779         /* Already moved. */
4780         if (task_cpu(p) != src_cpu)
4781                 goto done;
4782 
4783         /* Affinity changed (again). */
4784         if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4785                 goto fail;
4786 
4787         /*
4788          * If we're not on a rq, the next wake-up will ensure we're
4789          * placed properly.
4790          */
4791         if (task_on_rq_queued(p))
4792                 rq = move_queued_task(p, dest_cpu);
4793 done:
4794         ret = 1;
4795 fail:
4796         raw_spin_unlock(&rq->lock);
4797         raw_spin_unlock(&p->pi_lock);
4798         return ret;
4799 }
4800 
4801 #ifdef CONFIG_NUMA_BALANCING
4802 /* Migrate current task p to target_cpu */
4803 int migrate_task_to(struct task_struct *p, int target_cpu)
4804 {
4805         struct migration_arg arg = { p, target_cpu };
4806         int curr_cpu = task_cpu(p);
4807 
4808         if (curr_cpu == target_cpu)
4809                 return 0;
4810 
4811         if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4812                 return -EINVAL;
4813 
4814         /* TODO: This is not properly updating schedstats */
4815 
4816         trace_sched_move_numa(p, curr_cpu, target_cpu);
4817         return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4818 }
4819 
4820 /*
4821  * Requeue a task on a given node and accurately track the number of NUMA
4822  * tasks on the runqueues
4823  */
4824 void sched_setnuma(struct task_struct *p, int nid)
4825 {
4826         struct rq *rq;
4827         unsigned long flags;
4828         bool queued, running;
4829 
4830         rq = task_rq_lock(p, &flags);
4831         queued = task_on_rq_queued(p);
4832         running = task_current(rq, p);
4833 
4834         if (queued)
4835                 dequeue_task(rq, p, 0);
4836         if (running)
4837                 put_prev_task(rq, p);
4838 
4839         p->numa_preferred_nid = nid;
4840 
4841         if (running)
4842                 p->sched_class->set_curr_task(rq);
4843         if (queued)
4844                 enqueue_task(rq, p, 0);
4845         task_rq_unlock(rq, p, &flags);
4846 }
4847 #endif
4848 
4849 /*
4850  * migration_cpu_stop - this will be executed by a highprio stopper thread
4851  * and performs thread migration by bumping thread off CPU then
4852  * 'pushing' onto another runqueue.
4853  */
4854 static int migration_cpu_stop(void *data)
4855 {
4856         struct migration_arg *arg = data;
4857 
4858         /*
4859          * The original target cpu might have gone down and we might
4860          * be on another cpu but it doesn't matter.
4861          */
4862         local_irq_disable();
4863         /*
4864          * We need to explicitly wake pending tasks before running
4865          * __migrate_task() such that we will not miss enforcing cpus_allowed
4866          * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4867          */
4868         sched_ttwu_pending();
4869         __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4870         local_irq_enable();
4871         return 0;
4872 }
4873 
4874 #ifdef CONFIG_HOTPLUG_CPU
4875 
4876 /*
4877  * Ensures that the idle task is using init_mm right before its cpu goes
4878  * offline.
4879  */
4880 void idle_task_exit(void)
4881 {
4882         struct mm_struct *mm = current->active_mm;
4883 
4884         BUG_ON(cpu_online(smp_processor_id()));
4885 
4886         if (mm != &init_mm) {
4887                 switch_mm(mm, &init_mm, current);
4888                 finish_arch_post_lock_switch();
4889         }
4890         mmdrop(mm);
4891 }
4892 
4893 /*
4894  * Since this CPU is going 'away' for a while, fold any nr_active delta
4895  * we might have. Assumes we're called after migrate_tasks() so that the
4896  * nr_active count is stable.
4897  *
4898  * Also see the comment "Global load-average calculations".
4899  */
4900 static void calc_load_migrate(struct rq *rq)
4901 {
4902         long delta = calc_load_fold_active(rq);
4903         if (delta)
4904                 atomic_long_add(delta, &calc_load_tasks);
4905 }
4906 
4907 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4908 {
4909 }
4910 
4911 static const struct sched_class fake_sched_class = {
4912         .put_prev_task = put_prev_task_fake,
4913 };
4914 
4915 static struct task_struct fake_task = {
4916         /*
4917          * Avoid pull_{rt,dl}_task()
4918          */
4919         .prio = MAX_PRIO + 1,
4920         .sched_class = &fake_sched_class,
4921 };
4922 
4923 /*
4924  * Migrate all tasks from the rq, sleeping tasks will be migrated by
4925  * try_to_wake_up()->select_task_rq().
4926  *
4927  * Called with rq->lock held even though we'er in stop_machine() and
4928  * there's no concurrency possible, we hold the required locks anyway
4929  * because of lock validation efforts.
4930  */
4931 static void migrate_tasks(unsigned int dead_cpu)
4932 {
4933         struct rq *rq = cpu_rq(dead_cpu);
4934         struct task_struct *next, *stop = rq->stop;
4935         int dest_cpu;
4936 
4937         /*
4938          * Fudge the rq selection such that the below task selection loop
4939          * doesn't get stuck on the currently eligible stop task.
4940          *
4941          * We're currently inside stop_machine() and the rq is either stuck
4942          * in the stop_machine_cpu_stop() loop, or we're executing this code,
4943          * either way we should never end up calling schedule() until we're
4944          * done here.
4945          */
4946         rq->stop = NULL;
4947 
4948         /*
4949          * put_prev_task() and pick_next_task() sched
4950          * class method both need to have an up-to-date
4951          * value of rq->clock[_task]
4952          */
4953         update_rq_clock(rq);
4954 
4955         for ( ; ; ) {
4956                 /*
4957                  * There's this thread running, bail when that's the only
4958                  * remaining thread.
4959                  */
4960                 if (rq->nr_running == 1)
4961                         break;
4962 
4963                 next = pick_next_task(rq, &fake_task);
4964                 BUG_ON(!next);
4965                 next->sched_class->put_prev_task(rq, next);
4966 
4967                 /* Find suitable destination for @next, with force if needed. */
4968                 dest_cpu = select_fallback_rq(dead_cpu, next);
4969                 raw_spin_unlock(&rq->lock);
4970 
4971                 __migrate_task(next, dead_cpu, dest_cpu);
4972 
4973                 raw_spin_lock(&rq->lock);
4974         }
4975 
4976         rq->stop = stop;
4977 }
4978 
4979 #endif /* CONFIG_HOTPLUG_CPU */
4980 
4981 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4982 
4983 static struct ctl_table sd_ctl_dir[] = {
4984         {
4985                 .procname       = "sched_domain",
4986                 .mode           = 0555,
4987         },
4988         {}
4989 };
4990 
4991 static struct ctl_table sd_ctl_root[] = {
4992         {
4993                 .procname       = "kernel",
4994                 .mode           = 0555,
4995                 .child          = sd_ctl_dir,
4996         },
4997         {}
4998 };
4999 
5000 static struct ctl_table *sd_alloc_ctl_entry(int n)
5001 {
5002         struct ctl_table *entry =
5003                 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5004 
5005         return entry;
5006 }
5007 
5008 static void sd_free_ctl_entry(struct ctl_table **tablep)
5009 {
5010         struct ctl_table *entry;
5011 
5012         /*
5013          * In the intermediate directories, both the child directory and
5014          * procname are dynamically allocated and could fail but the mode
5015          * will always be set. In the lowest directory the names are
5016          * static strings and all have proc handlers.
5017          */
5018         for (entry = *tablep; entry->mode; entry++) {
5019                 if (entry->child)
5020                         sd_free_ctl_entry(&entry->child);
5021                 if (entry->proc_handler == NULL)
5022                         kfree(entry->procname);
5023         }
5024 
5025         kfree(*tablep);
5026         *tablep = NULL;
5027 }
5028 
5029 static int min_load_idx = 0;
5030 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5031 
5032 static void
5033 set_table_entry(struct ctl_table *entry,
5034                 const char *procname, void *data, int maxlen,
5035                 umode_t mode, proc_handler *proc_handler,
5036                 bool load_idx)
5037 {
5038         entry->procname = procname;
5039         entry->data = data;
5040         entry->maxlen = maxlen;
5041         entry->mode = mode;
5042         entry->proc_handler = proc_handler;
5043 
5044         if (load_idx) {
5045                 entry->extra1 = &min_load_idx;
5046                 entry->extra2 = &max_load_idx;
5047         }
5048 }
5049 
5050 static struct ctl_table *
5051 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5052 {
5053         struct ctl_table *table = sd_alloc_ctl_entry(14);
5054 
5055         if (table == NULL)
5056                 return NULL;
5057 
5058         set_table_entry(&table[0], "min_interval", &sd->min_interval,
5059                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5060         set_table_entry(&table[1], "max_interval", &sd->max_interval,
5061                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5062         set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5063                 sizeof(int), 0644, proc_dointvec_minmax, true);
5064         set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5065                 sizeof(int), 0644, proc_dointvec_minmax, true);
5066         set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5067                 sizeof(int), 0644, proc_dointvec_minmax, true);
5068         set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5069                 sizeof(int), 0644, proc_dointvec_minmax, true);
5070         set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5071                 sizeof(int), 0644, proc_dointvec_minmax, true);
5072         set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5073                 sizeof(int), 0644, proc_dointvec_minmax, false);
5074         set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5075                 sizeof(int), 0644, proc_dointvec_minmax, false);
5076         set_table_entry(&table[9], "cache_nice_tries",
5077                 &sd->cache_nice_tries,
5078                 sizeof(int), 0644, proc_dointvec_minmax, false);
5079         set_table_entry(&table[10], "flags", &sd->flags,
5080                 sizeof(int), 0644, proc_dointvec_minmax, false);
5081         set_table_entry(&table[11], "max_newidle_lb_cost",
5082                 &sd->max_newidle_lb_cost,
5083                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5084         set_table_entry(&table[12], "name", sd->name,
5085                 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5086         /* &table[13] is terminator */
5087 
5088         return table;
5089 }
5090 
5091 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5092 {
5093         struct ctl_table *entry, *table;
5094         struct sched_domain *sd;
5095         int domain_num = 0, i;
5096         char buf[32];
5097 
5098         for_each_domain(cpu, sd)
5099                 domain_num++;
5100         entry = table = sd_alloc_ctl_entry(domain_num + 1);
5101         if (table == NULL)
5102                 return NULL;
5103 
5104         i = 0;
5105         for_each_domain(cpu, sd) {
5106                 snprintf(buf, 32, "domain%d", i);
5107                 entry->procname = kstrdup(buf, GFP_KERNEL);
5108                 entry->mode = 0555;
5109                 entry->child = sd_alloc_ctl_domain_table(sd);
5110                 entry++;
5111                 i++;
5112         }
5113         return table;
5114 }
5115 
5116 static struct ctl_table_header *sd_sysctl_header;
5117 static void register_sched_domain_sysctl(void)
5118 {
5119         int i, cpu_num = num_possible_cpus();
5120         struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5121         char buf[32];
5122 
5123         WARN_ON(sd_ctl_dir[0].child);
5124         sd_ctl_dir[0].child = entry;
5125 
5126         if (entry == NULL)
5127                 return;
5128 
5129         for_each_possible_cpu(i) {
5130                 snprintf(buf, 32, "cpu%d", i);
5131                 entry->procname = kstrdup(buf, GFP_KERNEL);
5132                 entry->mode = 0555;
5133                 entry->child = sd_alloc_ctl_cpu_table(i);
5134                 entry++;
5135         }
5136 
5137         WARN_ON(sd_sysctl_header);
5138         sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5139 }
5140 
5141 /* may be called multiple times per register */
5142 static void unregister_sched_domain_sysctl(void)
5143 {
5144         if (sd_sysctl_header)
5145                 unregister_sysctl_table(sd_sysctl_header);
5146         sd_sysctl_header = NULL;
5147         if (sd_ctl_dir[0].child)
5148                 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5149 }
5150 #else
5151 static void register_sched_domain_sysctl(void)
5152 {
5153 }
5154 static void unregister_sched_domain_sysctl(void)
5155 {
5156 }
5157 #endif
5158 
5159 static void set_rq_online(struct rq *rq)
5160 {
5161         if (!rq->online) {
5162                 const struct sched_class *class;
5163 
5164                 cpumask_set_cpu(rq->cpu, rq->rd->online);
5165                 rq->online = 1;
5166 
5167                 for_each_class(class) {
5168                         if (class->rq_online)
5169                                 class->rq_online(rq);
5170                 }
5171         }
5172 }
5173 
5174 static void set_rq_offline(struct rq *rq)
5175 {
5176         if (rq->online) {
5177                 const struct sched_class *class;
5178 
5179                 for_each_class(class) {
5180                         if (class->rq_offline)
5181                                 class->rq_offline(rq);
5182                 }
5183 
5184                 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5185                 rq->online = 0;
5186         }
5187 }
5188 
5189 /*
5190  * migration_call - callback that gets triggered when a CPU is added.
5191  * Here we can start up the necessary migration thread for the new CPU.
5192  */
5193 static int
5194 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5195 {
5196         int cpu = (long)hcpu;
5197         unsigned long flags;
5198         struct rq *rq = cpu_rq(cpu);
5199 
5200         switch (action & ~CPU_TASKS_FROZEN) {
5201 
5202         case CPU_UP_PREPARE:
5203                 rq->calc_load_update = calc_load_update;
5204                 break;
5205 
5206         case CPU_ONLINE:
5207                 /* Update our root-domain */
5208                 raw_spin_lock_irqsave(&rq->lock, flags);
5209                 if (rq->rd) {
5210                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5211 
5212                         set_rq_online(rq);
5213                 }
5214                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5215                 break;
5216 
5217 #ifdef CONFIG_HOTPLUG_CPU
5218         case CPU_DYING:
5219                 sched_ttwu_pending();
5220                 /* Update our root-domain */
5221                 raw_spin_lock_irqsave(&rq->lock, flags);
5222                 if (rq->rd) {
5223                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5224                         set_rq_offline(rq);
5225                 }
5226                 migrate_tasks(cpu);
5227                 BUG_ON(rq->nr_running != 1); /* the migration thread */
5228                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5229                 break;
5230 
5231         case CPU_DEAD:
5232                 calc_load_migrate(rq);
5233                 break;
5234 #endif
5235         }
5236 
5237         update_max_interval();
5238 
5239         return NOTIFY_OK;
5240 }
5241 
5242 /*
5243  * Register at high priority so that task migration (migrate_all_tasks)
5244  * happens before everything else.  This has to be lower priority than
5245  * the notifier in the perf_event subsystem, though.
5246  */
5247 static struct notifier_block migration_notifier = {
5248         .notifier_call = migration_call,
5249         .priority = CPU_PRI_MIGRATION,
5250 };
5251 
5252 static void __cpuinit set_cpu_rq_start_time(void)
5253 {
5254         int cpu = smp_processor_id();
5255         struct rq *rq = cpu_rq(cpu);
5256         rq->age_stamp = sched_clock_cpu(cpu);
5257 }
5258 
5259 static int sched_cpu_active(struct notifier_block *nfb,
5260                                       unsigned long action, void *hcpu)
5261 {
5262         switch (action & ~CPU_TASKS_FROZEN) {
5263         case CPU_STARTING:
5264                 set_cpu_rq_start_time();
5265                 return NOTIFY_OK;
5266         case CPU_DOWN_FAILED:
5267                 set_cpu_active((long)hcpu, true);
5268                 return NOTIFY_OK;
5269         default:
5270                 return NOTIFY_DONE;
5271         }
5272 }
5273 
5274 static int sched_cpu_inactive(struct notifier_block *nfb,
5275                                         unsigned long action, void *hcpu)
5276 {
5277         unsigned long flags;
5278         long cpu = (long)hcpu;
5279         struct dl_bw *dl_b;
5280 
5281         switch (action & ~CPU_TASKS_FROZEN) {
5282         case CPU_DOWN_PREPARE:
5283                 set_cpu_active(cpu, false);
5284 
5285                 /* explicitly allow suspend */
5286                 if (!(action & CPU_TASKS_FROZEN)) {
5287                         bool overflow;
5288                         int cpus;
5289 
5290                         rcu_read_lock_sched();
5291                         dl_b = dl_bw_of(cpu);
5292 
5293                         raw_spin_lock_irqsave(&dl_b->lock, flags);
5294                         cpus = dl_bw_cpus(cpu);
5295                         overflow = __dl_overflow(dl_b, cpus, 0, 0);
5296                         raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5297 
5298                         rcu_read_unlock_sched();
5299 
5300                         if (overflow)
5301                                 return notifier_from_errno(-EBUSY);
5302                 }
5303                 return NOTIFY_OK;
5304         }
5305 
5306         return NOTIFY_DONE;
5307 }
5308 
5309 static int __init migration_init(void)
5310 {
5311         void *cpu = (void *)(long)smp_processor_id();
5312         int err;
5313 
5314         /* Initialize migration for the boot CPU */
5315         err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5316         BUG_ON(err == NOTIFY_BAD);
5317         migration_call(&migration_notifier, CPU_ONLINE, cpu);
5318         register_cpu_notifier(&migration_notifier);
5319 
5320         /* Register cpu active notifiers */
5321         cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5322         cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5323 
5324         return 0;
5325 }
5326 early_initcall(migration_init);
5327 #endif
5328 
5329 #ifdef CONFIG_SMP
5330 
5331 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5332 
5333 #ifdef CONFIG_SCHED_DEBUG
5334 
5335 static __read_mostly int sched_debug_enabled;
5336 
5337 static int __init sched_debug_setup(char *str)
5338 {
5339         sched_debug_enabled = 1;
5340 
5341         return 0;
5342 }
5343 early_param("sched_debug", sched_debug_setup);
5344 
5345 static inline bool sched_debug(void)
5346 {
5347         return sched_debug_enabled;
5348 }
5349 
5350 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5351                                   struct cpumask *groupmask)
5352 {
5353         struct sched_group *group = sd->groups;
5354         char str[256];
5355 
5356         cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5357         cpumask_clear(groupmask);
5358 
5359         printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5360 
5361         if (!(sd->flags & SD_LOAD_BALANCE)) {
5362                 printk("does not load-balance\n");
5363                 if (sd->parent)
5364                         printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5365                                         " has parent");
5366                 return -1;
5367         }
5368 
5369         printk(KERN_CONT "span %s level %s\n", str, sd->name);
5370 
5371         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5372                 printk(KERN_ERR "ERROR: domain->span does not contain "
5373                                 "CPU%d\n", cpu);
5374         }
5375         if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5376                 printk(KERN_ERR "ERROR: domain->groups does not contain"
5377                                 " CPU%d\n", cpu);
5378         }
5379 
5380         printk(KERN_DEBUG "%*s groups:", level + 1, "");
5381         do {
5382                 if (!group) {
5383                         printk("\n");
5384                         printk(KERN_ERR "ERROR: group is NULL\n");
5385                         break;
5386                 }
5387 
5388                 /*
5389                  * Even though we initialize ->capacity to something semi-sane,
5390                  * we leave capacity_orig unset. This allows us to detect if
5391                  * domain iteration is still funny without causing /0 traps.
5392                  */
5393                 if (!group->sgc->capacity_orig) {
5394                         printk(KERN_CONT "\n");
5395                         printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5396                         break;
5397                 }
5398 
5399                 if (!cpumask_weight(sched_group_cpus(group))) {
5400                         printk(KERN_CONT "\n");
5401                         printk(KERN_ERR "ERROR: empty group\n");
5402                         break;
5403                 }
5404 
5405                 if (!(sd->flags & SD_OVERLAP) &&
5406                     cpumask_intersects(groupmask, sched_group_cpus(group))) {
5407                         printk(KERN_CONT "\n");
5408                         printk(KERN_ERR "ERROR: repeated CPUs\n");
5409                         break;
5410                 }
5411 
5412                 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5413 
5414                 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5415 
5416                 printk(KERN_CONT " %s", str);
5417                 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5418                         printk(KERN_CONT " (cpu_capacity = %d)",
5419                                 group->sgc->capacity);
5420                 }
5421 
5422                 group = group->next;
5423         } while (group != sd->groups);
5424         printk(KERN_CONT "\n");
5425 
5426         if (!cpumask_equal(sched_domain_span(sd), groupmask))
5427                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5428 
5429         if (sd->parent &&
5430             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5431                 printk(KERN_ERR "ERROR: parent span is not a superset "
5432                         "of domain->span\n");
5433         return 0;
5434 }
5435 
5436 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5437 {
5438         int level = 0;
5439 
5440         if (!sched_debug_enabled)
5441                 return;
5442 
5443         if (!sd) {
5444                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5445                 return;
5446         }
5447 
5448         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5449 
5450         for (;;) {
5451                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5452                         break;
5453                 level++;
5454                 sd = sd->parent;
5455                 if (!sd)
5456                         break;
5457         }
5458 }
5459 #else /* !CONFIG_SCHED_DEBUG */
5460 # define sched_domain_debug(sd, cpu) do { } while (0)
5461 static inline bool sched_debug(void)
5462 {
5463         return false;
5464 }
5465 #endif /* CONFIG_SCHED_DEBUG */
5466 
5467 static int sd_degenerate(struct sched_domain *sd)
5468 {
5469         if (cpumask_weight(sched_domain_span(sd)) == 1)
5470                 return 1;
5471 
5472         /* Following flags need at least 2 groups */
5473         if (sd->flags & (SD_LOAD_BALANCE |
5474                          SD_BALANCE_NEWIDLE |
5475                          SD_BALANCE_FORK |
5476                          SD_BALANCE_EXEC |
5477                          SD_SHARE_CPUCAPACITY |
5478                          SD_SHARE_PKG_RESOURCES |
5479                          SD_SHARE_POWERDOMAIN)) {
5480                 if (sd->groups != sd->groups->next)
5481                         return 0;
5482         }
5483 
5484         /* Following flags don't use groups */
5485         if (sd->flags & (SD_WAKE_AFFINE))
5486                 return 0;
5487 
5488         return 1;
5489 }
5490 
5491 static int
5492 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5493 {
5494         unsigned long cflags = sd->flags, pflags = parent->flags;
5495 
5496         if (sd_degenerate(parent))
5497                 return 1;
5498 
5499         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5500                 return 0;
5501 
5502         /* Flags needing groups don't count if only 1 group in parent */
5503         if (parent->groups == parent->groups->next) {
5504                 pflags &= ~(SD_LOAD_BALANCE |
5505                                 SD_BALANCE_NEWIDLE |
5506                                 SD_BALANCE_FORK |
5507                                 SD_BALANCE_EXEC |
5508                                 SD_SHARE_CPUCAPACITY |
5509                                 SD_SHARE_PKG_RESOURCES |
5510                                 SD_PREFER_SIBLING |
5511                                 SD_SHARE_POWERDOMAIN);
5512                 if (nr_node_ids == 1)
5513                         pflags &= ~SD_SERIALIZE;
5514         }
5515         if (~cflags & pflags)
5516                 return 0;
5517 
5518         return 1;
5519 }
5520 
5521 static void free_rootdomain(struct rcu_head *rcu)
5522 {
5523         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5524 
5525         cpupri_cleanup(&rd->cpupri);
5526         cpudl_cleanup(&rd->cpudl);
5527         free_cpumask_var(rd->dlo_mask);
5528         free_cpumask_var(rd->rto_mask);
5529         free_cpumask_var(rd->online);
5530         free_cpumask_var(rd->span);
5531         kfree(rd);
5532 }
5533 
5534 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5535 {
5536         struct root_domain *old_rd = NULL;
5537         unsigned long flags;
5538 
5539         raw_spin_lock_irqsave(&rq->lock, flags);
5540 
5541         if (rq->rd) {
5542                 old_rd = rq->rd;
5543 
5544                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5545                         set_rq_offline(rq);
5546 
5547                 cpumask_clear_cpu(rq->cpu, old_rd->span);
5548 
5549                 /*
5550                  * If we dont want to free the old_rd yet then
5551                  * set old_rd to NULL to skip the freeing later
5552                  * in this function:
5553                  */
5554                 if (!atomic_dec_and_test(&old_rd->refcount))
5555                         old_rd = NULL;
5556         }
5557 
5558         atomic_inc(&rd->refcount);
5559         rq->rd = rd;
5560 
5561         cpumask_set_cpu(rq->cpu, rd->span);
5562         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5563                 set_rq_online(rq);
5564 
5565         raw_spin_unlock_irqrestore(&rq->lock, flags);
5566 
5567         if (old_rd)
5568                 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5569 }
5570 
5571 static int init_rootdomain(struct root_domain *rd)
5572 {
5573         memset(rd, 0, sizeof(*rd));
5574 
5575         if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5576                 goto out;
5577         if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5578                 goto free_span;
5579         if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5580                 goto free_online;
5581         if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5582                 goto free_dlo_mask;
5583 
5584         init_dl_bw(&rd->dl_bw);
5585         if (cpudl_init(&rd->cpudl) != 0)
5586                 goto free_dlo_mask;
5587 
5588         if (cpupri_init(&rd->cpupri) != 0)
5589                 goto free_rto_mask;
5590         return 0;
5591 
5592 free_rto_mask:
5593         free_cpumask_var(rd->rto_mask);
5594 free_dlo_mask:
5595         free_cpumask_var(rd->dlo_mask);
5596 free_online:
5597         free_cpumask_var(rd->online);
5598 free_span:
5599         free_cpumask_var(rd->span);
5600 out:
5601         return -ENOMEM;
5602 }
5603 
5604 /*
5605  * By default the system creates a single root-domain with all cpus as
5606  * members (mimicking the global state we have today).
5607  */
5608 struct root_domain def_root_domain;
5609 
5610 static void init_defrootdomain(void)
5611 {
5612         init_rootdomain(&def_root_domain);
5613 
5614         atomic_set(&def_root_domain.refcount, 1);
5615 }
5616 
5617 static struct root_domain *alloc_rootdomain(void)
5618 {
5619         struct root_domain *rd;
5620 
5621         rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5622         if (!rd)
5623                 return NULL;
5624 
5625         if (init_rootdomain(rd) != 0) {
5626                 kfree(rd);
5627                 return NULL;
5628         }
5629 
5630         return rd;
5631 }
5632 
5633 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5634 {
5635         struct sched_group *tmp, *first;
5636 
5637         if (!sg)
5638                 return;
5639 
5640         first = sg;
5641         do {
5642                 tmp = sg->next;
5643 
5644                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5645                         kfree(sg->sgc);
5646 
5647                 kfree(sg);
5648                 sg = tmp;
5649         } while (sg != first);
5650 }
5651 
5652 static void free_sched_domain(struct rcu_head *rcu)
5653 {
5654         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5655 
5656         /*
5657          * If its an overlapping domain it has private groups, iterate and
5658          * nuke them all.
5659          */
5660         if (sd->flags & SD_OVERLAP) {
5661                 free_sched_groups(sd->groups, 1);
5662         } else if (atomic_dec_and_test(&sd->groups->ref)) {
5663                 kfree(sd->groups->sgc);
5664                 kfree(sd->groups);
5665         }
5666         kfree(sd);
5667 }
5668 
5669 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5670 {
5671         call_rcu(&sd->rcu, free_sched_domain);
5672 }
5673 
5674 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5675 {
5676         for (; sd; sd = sd->parent)
5677                 destroy_sched_domain(sd, cpu);
5678 }
5679 
5680 /*
5681  * Keep a special pointer to the highest sched_domain that has
5682  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5683  * allows us to avoid some pointer chasing select_idle_sibling().
5684  *
5685  * Also keep a unique ID per domain (we use the first cpu number in
5686  * the cpumask of the domain), this allows us to quickly tell if
5687  * two cpus are in the same cache domain, see cpus_share_cache().
5688  */
5689 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5690 DEFINE_PER_CPU(int, sd_llc_size);
5691 DEFINE_PER_CPU(int, sd_llc_id);
5692 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5693 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5694 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5695 
5696 static void update_top_cache_domain(int cpu)
5697 {
5698         struct sched_domain *sd;
5699         struct sched_domain *busy_sd = NULL;
5700         int id = cpu;
5701         int size = 1;
5702 
5703         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5704         if (sd) {
5705                 id = cpumask_first(sched_domain_span(sd));
5706                 size = cpumask_weight(sched_domain_span(sd));
5707                 busy_sd = sd->parent; /* sd_busy */
5708         }
5709         rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5710 
5711         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5712         per_cpu(sd_llc_size, cpu) = size;
5713         per_cpu(sd_llc_id, cpu) = id;
5714 
5715         sd = lowest_flag_domain(cpu, SD_NUMA);
5716         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5717 
5718         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5719         rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5720 }
5721 
5722 /*
5723  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5724  * hold the hotplug lock.
5725  */
5726 static void
5727 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5728 {
5729         struct rq *rq = cpu_rq(cpu);
5730         struct sched_domain *tmp;
5731 
5732         /* Remove the sched domains which do not contribute to scheduling. */
5733         for (tmp = sd; tmp; ) {
5734                 struct sched_domain *parent = tmp->parent;
5735                 if (!parent)
5736                         break;
5737 
5738                 if (sd_parent_degenerate(tmp, parent)) {
5739                         tmp->parent = parent->parent;
5740                         if (parent->parent)
5741                                 parent->parent->child = tmp;
5742                         /*
5743                          * Transfer SD_PREFER_SIBLING down in case of a
5744                          * degenerate parent; the spans match for this
5745                          * so the property transfers.
5746                          */
5747                         if (parent->flags & SD_PREFER_SIBLING)
5748                                 tmp->flags |= SD_PREFER_SIBLING;
5749                         destroy_sched_domain(parent, cpu);
5750                 } else
5751                         tmp = tmp->parent;
5752         }
5753 
5754         if (sd && sd_degenerate(sd)) {
5755                 tmp = sd;
5756                 sd = sd->parent;
5757                 destroy_sched_domain(tmp, cpu);
5758                 if (sd)
5759                         sd->child = NULL;
5760         }
5761 
5762         sched_domain_debug(sd, cpu);
5763 
5764         rq_attach_root(rq, rd);
5765         tmp = rq->sd;
5766         rcu_assign_pointer(rq->sd, sd);
5767         destroy_sched_domains(tmp, cpu);
5768 
5769         update_top_cache_domain(cpu);
5770 }
5771 
5772 /* cpus with isolated domains */
5773 static cpumask_var_t cpu_isolated_map;
5774 
5775 /* Setup the mask of cpus configured for isolated domains */
5776 static int __init isolated_cpu_setup(char *str)
5777 {
5778         alloc_bootmem_cpumask_var(&cpu_isolated_map);
5779         cpulist_parse(str, cpu_isolated_map);
5780         return 1;
5781 }
5782 
5783 __setup("isolcpus=", isolated_cpu_setup);
5784 
5785 struct s_data {
5786         struct sched_domain ** __percpu sd;
5787         struct root_domain      *rd;
5788 };
5789 
5790 enum s_alloc {
5791         sa_rootdomain,
5792         sa_sd,
5793         sa_sd_storage,
5794         sa_none,
5795 };
5796 
5797 /*
5798  * Build an iteration mask that can exclude certain CPUs from the upwards
5799  * domain traversal.
5800  *
5801  * Asymmetric node setups can result in situations where the domain tree is of
5802  * unequal depth, make sure to skip domains that already cover the entire
5803  * range.
5804  *
5805  * In that case build_sched_domains() will have terminated the iteration early
5806  * and our sibling sd spans will be empty. Domains should always include the
5807  * cpu they're built on, so check that.
5808  *
5809  */
5810 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5811 {
5812         const struct cpumask *span = sched_domain_span(sd);
5813         struct sd_data *sdd = sd->private;
5814         struct sched_domain *sibling;
5815         int i;
5816 
5817         for_each_cpu(i, span) {
5818                 sibling = *per_cpu_ptr(sdd->sd, i);
5819                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5820                         continue;
5821 
5822                 cpumask_set_cpu(i, sched_group_mask(sg));
5823         }
5824 }
5825 
5826 /*
5827  * Return the canonical balance cpu for this group, this is the first cpu
5828  * of this group that's also in the iteration mask.
5829  */
5830 int group_balance_cpu(struct sched_group *sg)
5831 {
5832         return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5833 }
5834 
5835 static int
5836 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5837 {
5838         struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5839         const struct cpumask *span = sched_domain_span(sd);
5840         struct cpumask *covered = sched_domains_tmpmask;
5841         struct sd_data *sdd = sd->private;
5842         struct sched_domain *sibling;
5843         int i;
5844 
5845         cpumask_clear(covered);
5846 
5847         for_each_cpu(i, span) {
5848                 struct cpumask *sg_span;
5849 
5850                 if (cpumask_test_cpu(i, covered))
5851                         continue;
5852 
5853                 sibling = *per_cpu_ptr(sdd->sd, i);
5854 
5855                 /* See the comment near build_group_mask(). */
5856                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5857                         continue;
5858 
5859                 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5860                                 GFP_KERNEL, cpu_to_node(cpu));
5861 
5862                 if (!sg)
5863                         goto fail;
5864 
5865                 sg_span = sched_group_cpus(sg);
5866                 if (sibling->child)
5867                         cpumask_copy(sg_span, sched_domain_span(sibling->child));
5868                 else
5869                         cpumask_set_cpu(i, sg_span);
5870 
5871                 cpumask_or(covered, covered, sg_span);
5872 
5873                 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5874                 if (atomic_inc_return(&sg->sgc->ref) == 1)
5875                         build_group_mask(sd, sg);
5876 
5877                 /*
5878                  * Initialize sgc->capacity such that even if we mess up the
5879                  * domains and no possible iteration will get us here, we won't
5880                  * die on a /0 trap.
5881                  */
5882                 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5883                 sg->sgc->capacity_orig = sg->sgc->capacity;
5884 
5885                 /*
5886                  * Make sure the first group of this domain contains the
5887                  * canonical balance cpu. Otherwise the sched_domain iteration
5888                  * breaks. See update_sg_lb_stats().
5889                  */
5890                 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5891                     group_balance_cpu(sg) == cpu)
5892                         groups = sg;
5893 
5894                 if (!first)
5895                         first = sg;
5896                 if (last)
5897                         last->next = sg;
5898                 last = sg;
5899                 last->next = first;
5900         }
5901         sd->groups = groups;
5902 
5903         return 0;
5904 
5905 fail:
5906         free_sched_groups(first, 0);
5907 
5908         return -ENOMEM;
5909 }
5910 
5911 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5912 {
5913         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5914         struct sched_domain *child = sd->child;
5915 
5916         if (child)
5917                 cpu = cpumask_first(sched_domain_span(child));
5918 
5919         if (sg) {
5920                 *sg = *per_cpu_ptr(sdd->sg, cpu);
5921                 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5922                 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5923         }
5924 
5925         return cpu;
5926 }
5927 
5928 /*
5929  * build_sched_groups will build a circular linked list of the groups
5930  * covered by the given span, and will set each group's ->cpumask correctly,
5931  * and ->cpu_capacity to 0.
5932  *
5933  * Assumes the sched_domain tree is fully constructed
5934  */
5935 static int
5936 build_sched_groups(struct sched_domain *sd, int cpu)
5937 {
5938         struct sched_group *first = NULL, *last = NULL;
5939         struct sd_data *sdd = sd->private;
5940         const struct cpumask *span = sched_domain_span(sd);
5941         struct cpumask *covered;
5942         int i;
5943 
5944         get_group(cpu, sdd, &sd->groups);
5945         atomic_inc(&sd->groups->ref);
5946 
5947         if (cpu != cpumask_first(span))
5948                 return 0;
5949 
5950         lockdep_assert_held(&sched_domains_mutex);
5951         covered = sched_domains_tmpmask;
5952 
5953         cpumask_clear(covered);
5954 
5955         for_each_cpu(i, span) {
5956                 struct sched_group *sg;
5957                 int group, j;
5958 
5959                 if (cpumask_test_cpu(i, covered))
5960                         continue;
5961 
5962                 group = get_group(i, sdd, &sg);
5963                 cpumask_setall(sched_group_mask(sg));
5964 
5965                 for_each_cpu(j, span) {
5966                         if (get_group(j, sdd, NULL) != group)
5967                                 continue;
5968 
5969                         cpumask_set_cpu(j, covered);
5970                         cpumask_set_cpu(j, sched_group_cpus(sg));
5971                 }
5972 
5973                 if (!first)
5974                         first = sg;
5975                 if (last)
5976                         last->next = sg;
5977                 last = sg;
5978         }
5979         last->next = first;
5980 
5981         return 0;
5982 }
5983 
5984 /*
5985  * Initialize sched groups cpu_capacity.
5986  *
5987  * cpu_capacity indicates the capacity of sched group, which is used while
5988  * distributing the load between different sched groups in a sched domain.
5989  * Typically cpu_capacity for all the groups in a sched domain will be same
5990  * unless there are asymmetries in the topology. If there are asymmetries,
5991  * group having more cpu_capacity will pickup more load compared to the
5992  * group having less cpu_capacity.
5993  */
5994 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
5995 {
5996         struct sched_group *sg = sd->groups;
5997 
5998         WARN_ON(!sg);
5999 
6000         do {
6001                 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6002                 sg = sg->next;
6003         } while (sg != sd->groups);
6004 
6005         if (cpu != group_balance_cpu(sg))
6006                 return;
6007 
6008         update_group_capacity(sd, cpu);
6009         atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6010 }
6011 
6012 /*
6013  * Initializers for schedule domains
6014  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6015  */
6016 
6017 static int default_relax_domain_level = -1;
6018 int sched_domain_level_max;
6019 
6020 static int __init setup_relax_domain_level(char *str)
6021 {
6022         if (kstrtoint(str, 0, &default_relax_domain_level))
6023                 pr_warn("Unable to set relax_domain_level\n");
6024 
6025         return 1;
6026 }
6027 __setup("relax_domain_level=", setup_relax_domain_level);
6028 
6029 static void set_domain_attribute(struct sched_domain *sd,
6030                                  struct sched_domain_attr *attr)
6031 {
6032         int request;
6033 
6034         if (!attr || attr->relax_domain_level < 0) {
6035                 if (default_relax_domain_level < 0)
6036                         return;
6037                 else
6038                         request = default_relax_domain_level;
6039         } else
6040                 request = attr->relax_domain_level;
6041         if (request < sd->level) {
6042                 /* turn off idle balance on this domain */
6043                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6044         } else {
6045                 /* turn on idle balance on this domain */
6046                 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6047         }
6048 }
6049 
6050 static void __sdt_free(const struct cpumask *cpu_map);
6051 static int __sdt_alloc(const struct cpumask *cpu_map);
6052 
6053 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6054                                  const struct cpumask *cpu_map)
6055 {
6056         switch (what) {
6057         case sa_rootdomain:
6058                 if (!atomic_read(&d->rd->refcount))
6059                         free_rootdomain(&d->rd->rcu); /* fall through */
6060         case sa_sd:
6061                 free_percpu(d->sd); /* fall through */
6062         case sa_sd_storage:
6063                 __sdt_free(cpu_map); /* fall through */
6064         case sa_none:
6065                 break;
6066         }
6067 }
6068 
6069 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6070                                                    const struct cpumask *cpu_map)
6071 {
6072         memset(d, 0, sizeof(*d));
6073 
6074         if (__sdt_alloc(cpu_map))
6075                 return sa_sd_storage;
6076         d->sd = alloc_percpu(struct sched_domain *);
6077         if (!d->sd)
6078                 return sa_sd_storage;
6079         d->rd = alloc_rootdomain();
6080         if (!d->rd)
6081                 return sa_sd;
6082         return sa_rootdomain;
6083 }
6084 
6085 /*
6086  * NULL the sd_data elements we've used to build the sched_domain and
6087  * sched_group structure so that the subsequent __free_domain_allocs()
6088  * will not free the data we're using.
6089  */
6090 static void claim_allocations(int cpu, struct sched_domain *sd)
6091 {
6092         struct sd_data *sdd = sd->private;
6093 
6094         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6095         *per_cpu_ptr(sdd->sd, cpu) = NULL;
6096 
6097         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6098                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6099 
6100         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6101                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6102 }
6103 
6104 #ifdef CONFIG_NUMA
6105 static int sched_domains_numa_levels;
6106 static int *sched_domains_numa_distance;
6107 static struct cpumask ***sched_domains_numa_masks;
6108 static int sched_domains_curr_level;
6109 #endif
6110 
6111 /*
6112  * SD_flags allowed in topology descriptions.
6113  *
6114  * SD_SHARE_CPUCAPACITY      - describes SMT topologies
6115  * SD_SHARE_PKG_RESOURCES - describes shared caches
6116  * SD_NUMA                - describes NUMA topologies
6117  * SD_SHARE_POWERDOMAIN   - describes shared power domain
6118  *
6119  * Odd one out:
6120  * SD_ASYM_PACKING        - describes SMT quirks
6121  */
6122 #define TOPOLOGY_SD_FLAGS               \
6123         (SD_SHARE_CPUCAPACITY |         \
6124          SD_SHARE_PKG_RESOURCES |       \
6125          SD_NUMA |                      \
6126          SD_ASYM_PACKING |              \
6127          SD_SHARE_POWERDOMAIN)
6128 
6129 static struct sched_domain *
6130 sd_init(struct sched_domain_topology_level *tl, int cpu)
6131 {
6132         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6133         int sd_weight, sd_flags = 0;
6134 
6135 #ifdef CONFIG_NUMA
6136         /*
6137          * Ugly hack to pass state to sd_numa_mask()...
6138          */
6139         sched_domains_curr_level = tl->numa_level;
6140 #endif
6141 
6142         sd_weight = cpumask_weight(tl->mask(cpu));
6143 
6144         if (tl->sd_flags)
6145                 sd_flags = (*tl->sd_flags)();
6146         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6147                         "wrong sd_flags in topology description\n"))
6148                 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6149 
6150         *sd = (struct sched_domain){
6151                 .min_interval           = sd_weight,
6152                 .max_interval           = 2*sd_weight,
6153                 .busy_factor            = 32,
6154                 .imbalance_pct          = 125,
6155 
6156                 .cache_nice_tries       = 0,
6157                 .busy_idx               = 0,
6158                 .idle_idx               = 0,
6159                 .newidle_idx            = 0,
6160                 .wake_idx               = 0,
6161                 .forkexec_idx           = 0,
6162 
6163                 .flags                  = 1*SD_LOAD_BALANCE
6164                                         | 1*SD_BALANCE_NEWIDLE
6165                                         | 1*SD_BALANCE_EXEC
6166                                         | 1*SD_BALANCE_FORK
6167                                         | 0*SD_BALANCE_WAKE
6168                                         | 1*SD_WAKE_AFFINE
6169                                         | 0*SD_SHARE_CPUCAPACITY
6170                                         | 0*SD_SHARE_PKG_RESOURCES
6171                                         | 0*SD_SERIALIZE
6172                                         | 0*SD_PREFER_SIBLING
6173                                         | 0*SD_NUMA
6174                                         | sd_flags
6175                                         ,
6176 
6177                 .last_balance           = jiffies,
6178                 .balance_interval       = sd_weight,
6179                 .smt_gain               = 0,
6180                 .max_newidle_lb_cost    = 0,
6181                 .next_decay_max_lb_cost = jiffies,
6182 #ifdef CONFIG_SCHED_DEBUG
6183                 .name                   = tl->name,
6184 #endif
6185         };
6186 
6187         /*
6188          * Convert topological properties into behaviour.
6189          */
6190 
6191         if (sd->flags & SD_SHARE_CPUCAPACITY) {
6192                 sd->imbalance_pct = 110;
6193                 sd->smt_gain = 1178; /* ~15% */
6194 
6195         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6196                 sd->imbalance_pct = 117;
6197                 sd->cache_nice_tries = 1;
6198                 sd->busy_idx = 2;
6199 
6200 #ifdef CONFIG_NUMA
6201         } else if (sd->flags & SD_NUMA) {
6202                 sd->cache_nice_tries = 2;
6203                 sd->busy_idx = 3;
6204                 sd->idle_idx = 2;
6205 
6206                 sd->flags |= SD_SERIALIZE;
6207                 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6208                         sd->flags &= ~(SD_BALANCE_EXEC |
6209                                        SD_BALANCE_FORK |
6210                                        SD_WAKE_AFFINE);
6211                 }
6212 
6213 #endif
6214         } else {
6215                 sd->flags |= SD_PREFER_SIBLING;
6216                 sd->cache_nice_tries = 1;
6217                 sd->busy_idx = 2;
6218                 sd->idle_idx = 1;
6219         }
6220 
6221         sd->private = &tl->data;
6222 
6223         return sd;
6224 }
6225 
6226 /*
6227  * Topology list, bottom-up.
6228  */
6229 static struct sched_domain_topology_level default_topology[] = {
6230 #ifdef CONFIG_SCHED_SMT
6231         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6232 #endif
6233 #ifdef CONFIG_SCHED_MC
6234         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6235 #endif
6236         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6237         { NULL, },
6238 };
6239 
6240 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6241 
6242 #define for_each_sd_topology(tl)                        \
6243         for (tl = sched_domain_topology; tl->mask; tl++)
6244 
6245 void set_sched_topology(struct sched_domain_topology_level *tl)
6246 {
6247         sched_domain_topology = tl;
6248 }
6249 
6250 #ifdef CONFIG_NUMA
6251 
6252 static const struct cpumask *sd_numa_mask(int cpu)
6253 {
6254         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6255 }
6256 
6257 static void sched_numa_warn(const char *str)
6258 {
6259         static int done = false;
6260         int i,j;
6261 
6262         if (done)
6263                 return;
6264 
6265         done = true;
6266 
6267         printk(KERN_WARNING "ERROR: %s\n\n", str);
6268 
6269         for (i = 0; i < nr_node_ids; i++) {
6270                 printk(KERN_WARNING "  ");
6271                 for (j = 0; j < nr_node_ids; j++)
6272                         printk(KERN_CONT "%02d ", node_distance(i,j));
6273                 printk(KERN_CONT "\n");
6274         }
6275         printk(KERN_WARNING "\n");
6276 }
6277 
6278 static bool find_numa_distance(int distance)
6279 {
6280         int i;
6281 
6282         if (distance == node_distance(0, 0))
6283                 return true;
6284 
6285         for (i = 0; i < sched_domains_numa_levels; i++) {
6286                 if (sched_domains_numa_distance[i] == distance)
6287                         return true;
6288         }
6289 
6290         return false;
6291 }
6292 
6293 static void sched_init_numa(void)
6294 {
6295         int next_distance, curr_distance = node_distance(0, 0);
6296         struct sched_domain_topology_level *tl;
6297         int level = 0;
6298         int i, j, k;
6299 
6300         sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6301         if (!sched_domains_numa_distance)
6302                 return;
6303 
6304         /*
6305          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6306          * unique distances in the node_distance() table.
6307          *
6308          * Assumes node_distance(0,j) includes all distances in
6309          * node_distance(i,j) in order to avoid cubic time.
6310          */
6311         next_distance = curr_distance;
6312         for (i = 0; i < nr_node_ids; i++) {
6313                 for (j = 0; j < nr_node_ids; j++) {
6314                         for (k = 0; k < nr_node_ids; k++) {
6315                                 int distance = node_distance(i, k);
6316 
6317                                 if (distance > curr_distance &&
6318                                     (distance < next_distance ||
6319                                      next_distance == curr_distance))
6320                                         next_distance = distance;
6321 
6322                                 /*
6323                                  * While not a strong assumption it would be nice to know
6324                                  * about cases where if node A is connected to B, B is not
6325                                  * equally connected to A.
6326                                  */
6327                                 if (sched_debug() && node_distance(k, i) != distance)
6328                                         sched_numa_warn("Node-distance not symmetric");
6329 
6330                                 if (sched_debug() && i && !find_numa_distance(distance))
6331                                         sched_numa_warn("Node-0 not representative");
6332                         }
6333                         if (next_distance != curr_distance) {
6334                                 sched_domains_numa_distance[level++] = next_distance;
6335                                 sched_domains_numa_levels = level;
6336                                 curr_distance = next_distance;
6337                         } else break;
6338                 }
6339 
6340                 /*
6341                  * In case of sched_debug() we verify the above assumption.
6342                  */
6343                 if (!sched_debug())
6344                         break;
6345         }
6346 
6347         if (!level)
6348                 return;
6349 
6350         /*
6351          * 'level' contains the number of unique distances, excluding the
6352          * identity distance node_distance(i,i).
6353          *
6354          * The sched_domains_numa_distance[] array includes the actual distance
6355          * numbers.
6356          */
6357 
6358         /*
6359          * Here, we should temporarily reset sched_domains_numa_levels to 0.
6360          * If it fails to allocate memory for array sched_domains_numa_masks[][],
6361          * the array will contain less then 'level' members. This could be
6362          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6363          * in other functions.
6364          *
6365          * We reset it to 'level' at the end of this function.
6366          */
6367         sched_domains_numa_levels = 0;
6368 
6369         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6370         if (!sched_domains_numa_masks)
6371                 return;
6372 
6373         /*
6374          * Now for each level, construct a mask per node which contains all
6375          * cpus of nodes that are that many hops away from us.
6376          */
6377         for (i = 0; i < level; i++) {
6378                 sched_domains_numa_masks[i] =
6379                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6380                 if (!sched_domains_numa_masks[i])
6381                         return;
6382 
6383                 for (j = 0; j < nr_node_ids; j++) {
6384                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6385                         if (!mask)
6386                                 return;
6387 
6388                         sched_domains_numa_masks[i][j] = mask;
6389 
6390                         for (k = 0; k < nr_node_ids; k++) {
6391                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
6392                                         continue;
6393 
6394                                 cpumask_or(mask, mask, cpumask_of_node(k));
6395                         }
6396                 }
6397         }
6398 
6399         /* Compute default topology size */
6400         for (i = 0; sched_domain_topology[i].mask; i++);
6401 
6402         tl = kzalloc((i + level + 1) *
6403                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6404         if (!tl)
6405                 return;
6406 
6407         /*
6408          * Copy the default topology bits..
6409          */
6410         for (i = 0; sched_domain_topology[i].mask; i++)
6411                 tl[i] = sched_domain_topology[i];
6412 
6413         /*
6414          * .. and append 'j' levels of NUMA goodness.
6415          */
6416         for (j = 0; j < level; i++, j++) {
6417                 tl[i] = (struct sched_domain_topology_level){
6418                         .mask = sd_numa_mask,
6419                         .sd_flags = cpu_numa_flags,
6420                         .flags = SDTL_OVERLAP,
6421                         .numa_level = j,
6422                         SD_INIT_NAME(NUMA)
6423                 };
6424         }
6425 
6426         sched_domain_topology = tl;
6427 
6428         sched_domains_numa_levels = level;
6429 }
6430 
6431 static void sched_domains_numa_masks_set(int cpu)
6432 {
6433         int i, j;
6434         int node = cpu_to_node(cpu);
6435 
6436         for (i = 0; i < sched_domains_numa_levels; i++) {
6437                 for (j = 0; j < nr_node_ids; j++) {
6438                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
6439                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6440                 }
6441         }
6442 }
6443 
6444 static void sched_domains_numa_masks_clear(int cpu)
6445 {
6446         int i, j;
6447         for (i = 0; i < sched_domains_numa_levels; i++) {
6448                 for (j = 0; j < nr_node_ids; j++)
6449                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6450         }
6451 }
6452 
6453 /*
6454  * Update sched_domains_numa_masks[level][node] array when new cpus
6455  * are onlined.
6456  */
6457 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6458                                            unsigned long action,
6459                                            void *hcpu)
6460 {
6461         int cpu = (long)hcpu;
6462 
6463         switch (action & ~CPU_TASKS_FROZEN) {
6464         case CPU_ONLINE:
6465                 sched_domains_numa_masks_set(cpu);
6466                 break;
6467 
6468         case CPU_DEAD:
6469                 sched_domains_numa_masks_clear(cpu);
6470                 break;
6471 
6472         default:
6473                 return NOTIFY_DONE;
6474         }
6475 
6476         return NOTIFY_OK;
6477 }
6478 #else
6479 static inline void sched_init_numa(void)
6480 {
6481 }
6482 
6483 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6484                                            unsigned long action,
6485                                            void *hcpu)
6486 {
6487         return 0;
6488 }
6489 #endif /* CONFIG_NUMA */
6490 
6491 static int __sdt_alloc(const struct cpumask *cpu_map)
6492 {
6493         struct sched_domain_topology_level *tl;
6494         int j;
6495 
6496         for_each_sd_topology(tl) {
6497                 struct sd_data *sdd = &tl->data;
6498 
6499                 sdd->sd = alloc_percpu(struct sched_domain *);
6500                 if (!sdd->sd)
6501                         return -ENOMEM;
6502 
6503                 sdd->sg = alloc_percpu(struct sched_group *);
6504                 if (!sdd->sg)
6505                         return -ENOMEM;
6506 
6507                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6508                 if (!sdd->sgc)
6509                         return -ENOMEM;
6510 
6511                 for_each_cpu(j, cpu_map) {
6512                         struct sched_domain *sd;
6513                         struct sched_group *sg;
6514                         struct sched_group_capacity *sgc;
6515 
6516                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6517                                         GFP_KERNEL, cpu_to_node(j));
6518                         if (!sd)
6519                                 return -ENOMEM;
6520 
6521                         *per_cpu_ptr(sdd->sd, j) = sd;
6522 
6523                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6524                                         GFP_KERNEL, cpu_to_node(j));
6525                         if (!sg)
6526                                 return -ENOMEM;
6527 
6528                         sg->next = sg;
6529 
6530                         *per_cpu_ptr(sdd->sg, j) = sg;
6531 
6532                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6533                                         GFP_KERNEL, cpu_to_node(j));
6534                         if (!sgc)
6535                                 return -ENOMEM;
6536 
6537                         *per_cpu_ptr(sdd->sgc, j) = sgc;
6538                 }
6539         }
6540 
6541         return 0;
6542 }
6543 
6544 static void __sdt_free(const struct cpumask *cpu_map)
6545 {
6546         struct sched_domain_topology_level *tl;
6547         int j;
6548 
6549         for_each_sd_topology(tl) {
6550                 struct sd_data *sdd = &tl->data;
6551 
6552                 for_each_cpu(j, cpu_map) {
6553                         struct sched_domain *sd;
6554 
6555                         if (sdd->sd) {
6556                                 sd = *per_cpu_ptr(sdd->sd, j);
6557                                 if (sd && (sd->flags & SD_OVERLAP))
6558                                         free_sched_groups(sd->groups, 0);
6559                                 kfree(*per_cpu_ptr(sdd->sd, j));
6560                         }
6561 
6562                         if (sdd->sg)
6563                                 kfree(*per_cpu_ptr(sdd->sg, j));
6564                         if (sdd->sgc)
6565                                 kfree(*per_cpu_ptr(sdd->sgc, j));
6566                 }
6567                 free_percpu(sdd->sd);
6568                 sdd->sd = NULL;
6569                 free_percpu(sdd->sg);
6570                 sdd->sg = NULL;
6571                 free_percpu(sdd->sgc);
6572                 sdd->sgc = NULL;
6573         }
6574 }
6575 
6576 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6577                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6578                 struct sched_domain *child, int cpu)
6579 {
6580         struct sched_domain *sd = sd_init(tl, cpu);
6581         if (!sd)
6582                 return child;
6583 
6584         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6585         if (child) {
6586                 sd->level = child->level + 1;
6587                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6588                 child->parent = sd;
6589                 sd->child = child;
6590 
6591                 if (!cpumask_subset(sched_domain_span(child),
6592                                     sched_domain_span(sd))) {
6593                         pr_err("BUG: arch topology borken\n");
6594 #ifdef CONFIG_SCHED_DEBUG
6595                         pr_err("     the %s domain not a subset of the %s domain\n",
6596                                         child->name, sd->name);
6597 #endif
6598                         /* Fixup, ensure @sd has at least @child cpus. */
6599                         cpumask_or(sched_domain_span(sd),
6600                                    sched_domain_span(sd),
6601                                    sched_domain_span(child));
6602                 }
6603 
6604         }
6605         set_domain_attribute(sd, attr);
6606 
6607         return sd;
6608 }
6609 
6610 /*
6611  * Build sched domains for a given set of cpus and attach the sched domains
6612  * to the individual cpus
6613  */
6614 static int build_sched_domains(const struct cpumask *cpu_map,
6615                                struct sched_domain_attr *attr)
6616 {
6617         enum s_alloc alloc_state;
6618         struct sched_domain *sd;
6619         struct s_data d;
6620         int i, ret = -ENOMEM;
6621 
6622         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6623         if (alloc_state != sa_rootdomain)
6624                 goto error;
6625 
6626         /* Set up domains for cpus specified by the cpu_map. */
6627         for_each_cpu(i, cpu_map) {
6628                 struct sched_domain_topology_level *tl;
6629 
6630                 sd = NULL;
6631                 for_each_sd_topology(tl) {
6632                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6633                         if (tl == sched_domain_topology)
6634                                 *per_cpu_ptr(d.sd, i) = sd;
6635                         if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6636                                 sd->flags |= SD_OVERLAP;
6637                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6638                                 break;
6639                 }
6640         }
6641 
6642         /* Build the groups for the domains */
6643         for_each_cpu(i, cpu_map) {
6644                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6645                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
6646                         if (sd->flags & SD_OVERLAP) {
6647                                 if (build_overlap_sched_groups(sd, i))
6648                                         goto error;
6649                         } else {
6650                                 if (build_sched_groups(sd, i))
6651                                         goto error;
6652                         }
6653                 }
6654         }
6655 
6656         /* Calculate CPU capacity for physical packages and nodes */
6657         for (i = nr_cpumask_bits-1; i >= 0; i--) {
6658                 if (!cpumask_test_cpu(i, cpu_map))
6659                         continue;
6660 
6661                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6662                         claim_allocations(i, sd);
6663                         init_sched_groups_capacity(i, sd);
6664                 }
6665         }
6666 
6667         /* Attach the domains */
6668         rcu_read_lock();
6669         for_each_cpu(i, cpu_map) {
6670                 sd = *per_cpu_ptr(d.sd, i);
6671                 cpu_attach_domain(sd, d.rd, i);
6672         }
6673         rcu_read_unlock();
6674 
6675         ret = 0;
6676 error:
6677         __free_domain_allocs(&d, alloc_state, cpu_map);
6678         return ret;
6679 }
6680 
6681 static cpumask_var_t *doms_cur; /* current sched domains */
6682 static int ndoms_cur;           /* number of sched domains in 'doms_cur' */
6683 static struct sched_domain_attr *dattr_cur;
6684                                 /* attribues of custom domains in 'doms_cur' */
6685 
6686 /*
6687  * Special case: If a kmalloc of a doms_cur partition (array of
6688  * cpumask) fails, then fallback to a single sched domain,
6689  * as determined by the single cpumask fallback_doms.
6690  */
6691 static cpumask_var_t fallback_doms;
6692 
6693 /*
6694  * arch_update_cpu_topology lets virtualized architectures update the
6695  * cpu core maps. It is supposed to return 1 if the topology changed
6696  * or 0 if it stayed the same.
6697  */
6698 int __weak arch_update_cpu_topology(void)
6699 {
6700         return 0;
6701 }
6702 
6703 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6704 {
6705         int i;
6706         cpumask_var_t *doms;
6707 
6708         doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6709         if (!doms)
6710                 return NULL;
6711         for (i = 0; i < ndoms; i++) {
6712                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6713                         free_sched_domains(doms, i);
6714                         return NULL;
6715                 }
6716         }
6717         return doms;
6718 }
6719 
6720 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6721 {
6722         unsigned int i;
6723         for (i = 0; i < ndoms; i++)
6724                 free_cpumask_var(doms[i]);
6725         kfree(doms);
6726 }
6727 
6728 /*
6729  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6730  * For now this just excludes isolated cpus, but could be used to
6731  * exclude other special cases in the future.
6732  */
6733 static int init_sched_domains(const struct cpumask *cpu_map)
6734 {
6735         int err;
6736 
6737         arch_update_cpu_topology();
6738         ndoms_cur = 1;
6739         doms_cur = alloc_sched_domains(ndoms_cur);
6740         if (!doms_cur)
6741                 doms_cur = &fallback_doms;
6742         cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6743         err = build_sched_domains(doms_cur[0], NULL);
6744         register_sched_domain_sysctl();
6745 
6746         return err;
6747 }
6748 
6749 /*
6750  * Detach sched domains from a group of cpus specified in cpu_map
6751  * These cpus will now be attached to the NULL domain
6752  */
6753 static void detach_destroy_domains(const struct cpumask *cpu_map)
6754 {
6755         int i;
6756 
6757         rcu_read_lock();
6758         for_each_cpu(i, cpu_map)
6759                 cpu_attach_domain(NULL, &def_root_domain, i);
6760         rcu_read_unlock();
6761 }
6762 
6763 /* handle null as "default" */
6764 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6765                         struct sched_domain_attr *new, int idx_new)
6766 {
6767         struct sched_domain_attr tmp;
6768 
6769         /* fast path */
6770         if (!new && !cur)
6771                 return 1;
6772 
6773         tmp = SD_ATTR_INIT;
6774         return !memcmp(cur ? (cur + idx_cur) : &tmp,
6775                         new ? (new + idx_new) : &tmp,
6776                         sizeof(struct sched_domain_attr));
6777 }
6778 
6779 /*
6780  * Partition sched domains as specified by the 'ndoms_new'
6781  * cpumasks in the array doms_new[] of cpumasks. This compares
6782  * doms_new[] to the current sched domain partitioning, doms_cur[].
6783  * It destroys each deleted domain and builds each new domain.
6784  *
6785  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6786  * The masks don't intersect (don't overlap.) We should setup one
6787  * sched domain for each mask. CPUs not in any of the cpumasks will
6788  * not be load balanced. If the same cpumask appears both in the
6789  * current 'doms_cur' domains and in the new 'doms_new', we can leave
6790  * it as it is.
6791  *
6792  * The passed in 'doms_new' should be allocated using
6793  * alloc_sched_domains.  This routine takes ownership of it and will
6794  * free_sched_domains it when done with it. If the caller failed the
6795  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6796  * and partition_sched_domains() will fallback to the single partition
6797  * 'fallback_doms', it also forces the domains to be rebuilt.
6798  *
6799  * If doms_new == NULL it will be replaced with cpu_online_mask.
6800  * ndoms_new == 0 is a special case for destroying existing domains,
6801  * and it will not create the default domain.
6802  *
6803  * Call with hotplug lock held
6804  */
6805 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6806                              struct sched_domain_attr *dattr_new)
6807 {
6808         int i, j, n;
6809         int new_topology;
6810 
6811         mutex_lock(&sched_domains_mutex);
6812 
6813         /* always unregister in case we don't destroy any domains */
6814         unregister_sched_domain_sysctl();
6815 
6816         /* Let architecture update cpu core mappings. */
6817         new_topology = arch_update_cpu_topology();
6818 
6819         n = doms_new ? ndoms_new : 0;
6820 
6821         /* Destroy deleted domains */
6822         for (i = 0; i < ndoms_cur; i++) {
6823                 for (j = 0; j < n && !new_topology; j++) {
6824                         if (cpumask_equal(doms_cur[i], doms_new[j])
6825                             && dattrs_equal(dattr_cur, i, dattr_new, j))
6826                                 goto match1;
6827                 }
6828                 /* no match - a current sched domain not in new doms_new[] */
6829                 detach_destroy_domains(doms_cur[i]);
6830 match1:
6831                 ;
6832         }
6833 
6834         n = ndoms_cur;
6835         if (doms_new == NULL) {
6836                 n = 0;
6837                 doms_new = &fallback_doms;
6838                 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6839                 WARN_ON_ONCE(dattr_new);
6840         }
6841 
6842         /* Build new domains */
6843         for (i = 0; i < ndoms_new; i++) {
6844                 for (j = 0; j < n && !new_topology; j++) {
6845                         if (cpumask_equal(doms_new[i], doms_cur[j])
6846                             && dattrs_equal(dattr_new, i, dattr_cur, j))
6847                                 goto match2;
6848                 }
6849                 /* no match - add a new doms_new */
6850                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6851 match2:
6852                 ;
6853         }
6854 
6855         /* Remember the new sched domains */
6856         if (doms_cur != &fallback_doms)
6857                 free_sched_domains(doms_cur, ndoms_cur);
6858         kfree(dattr_cur);       /* kfree(NULL) is safe */
6859         doms_cur = doms_new;
6860         dattr_cur = dattr_new;
6861         ndoms_cur = ndoms_new;
6862 
6863         register_sched_domain_sysctl();
6864 
6865         mutex_unlock(&sched_domains_mutex);
6866 }
6867 
6868 static int num_cpus_frozen;     /* used to mark begin/end of suspend/resume */
6869 
6870 /*
6871  * Update cpusets according to cpu_active mask.  If cpusets are
6872  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6873  * around partition_sched_domains().
6874  *
6875  * If we come here as part of a suspend/resume, don't touch cpusets because we
6876  * want to restore it back to its original state upon resume anyway.
6877  */
6878 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6879                              void *hcpu)
6880 {
6881         switch (action) {
6882         case CPU_ONLINE_FROZEN:
6883         case CPU_DOWN_FAILED_FROZEN:
6884 
6885                 /*
6886                  * num_cpus_frozen tracks how many CPUs are involved in suspend
6887                  * resume sequence. As long as this is not the last online
6888                  * operation in the resume sequence, just build a single sched
6889                  * domain, ignoring cpusets.
6890                  */
6891                 num_cpus_frozen--;
6892                 if (likely(num_cpus_frozen)) {
6893                         partition_sched_domains(1, NULL, NULL);
6894                         break;
6895                 }
6896 
6897                 /*
6898                  * This is the last CPU online operation. So fall through and
6899                  * restore the original sched domains by considering the
6900                  * cpuset configurations.
6901                  */
6902 
6903         case CPU_ONLINE:
6904         case CPU_DOWN_FAILED:
6905                 cpuset_update_active_cpus(true);
6906                 break;
6907         default:
6908                 return NOTIFY_DONE;
6909         }
6910         return NOTIFY_OK;
6911 }
6912 
6913 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6914                                void *hcpu)
6915 {
6916         switch (action) {
6917         case CPU_DOWN_PREPARE:
6918                 cpuset_update_active_cpus(false);
6919                 break;
6920         case CPU_DOWN_PREPARE_FROZEN:
6921                 num_cpus_frozen++;
6922                 partition_sched_domains(1, NULL, NULL);
6923                 break;
6924         default:
6925                 return NOTIFY_DONE;
6926         }
6927         return NOTIFY_OK;
6928 }
6929 
6930 void __init sched_init_smp(void)
6931 {
6932         cpumask_var_t non_isolated_cpus;
6933 
6934         alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6935         alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6936 
6937         sched_init_numa();
6938 
6939         /*
6940          * There's no userspace yet to cause hotplug operations; hence all the
6941          * cpu masks are stable and all blatant races in the below code cannot
6942          * happen.
6943          */
6944         mutex_lock(&sched_domains_mutex);
6945         init_sched_domains(cpu_active_mask);
6946         cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6947         if (cpumask_empty(non_isolated_cpus))
6948                 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6949         mutex_unlock(&sched_domains_mutex);
6950 
6951         hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6952         hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6953         hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6954 
6955         init_hrtick();
6956 
6957         /* Move init over to a non-isolated CPU */
6958         if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6959                 BUG();
6960         sched_init_granularity();
6961         free_cpumask_var(non_isolated_cpus);
6962 
6963         init_sched_rt_class();
6964         init_sched_dl_class();
6965 }
6966 #else
6967 void __init sched_init_smp(void)
6968 {
6969         sched_init_granularity();
6970 }
6971 #endif /* CONFIG_SMP */
6972 
6973 const_debug unsigned int sysctl_timer_migration = 1;
6974 
6975 int in_sched_functions(unsigned long addr)
6976 {
6977         return in_lock_functions(addr) ||
6978                 (addr >= (unsigned long)__sched_text_start
6979                 && addr < (unsigned long)__sched_text_end);
6980 }
6981 
6982 #ifdef CONFIG_CGROUP_SCHED
6983 /*
6984  * Default task group.
6985  * Every task in system belongs to this group at bootup.
6986  */
6987 struct task_group root_task_group;
6988 LIST_HEAD(task_groups);
6989 #endif
6990 
6991 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6992 
6993 void __init sched_init(void)
6994 {
6995         int i, j;
6996         unsigned long alloc_size = 0, ptr;
6997 
6998 #ifdef CONFIG_FAIR_GROUP_SCHED
6999         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7000 #endif
7001 #ifdef CONFIG_RT_GROUP_SCHED
7002         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7003 #endif
7004 #ifdef CONFIG_CPUMASK_OFFSTACK
7005         alloc_size += num_possible_cpus() * cpumask_size();
7006 #endif
7007         if (alloc_size) {
7008                 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7009 
7010 #ifdef CONFIG_FAIR_GROUP_SCHED
7011                 root_task_group.se = (struct sched_entity **)ptr;
7012                 ptr += nr_cpu_ids * sizeof(void **);
7013 
7014                 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7015                 ptr += nr_cpu_ids * sizeof(void **);
7016 
7017 #endif /* CONFIG_FAIR_GROUP_SCHED */
7018 #ifdef CONFIG_RT_GROUP_SCHED
7019                 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7020                 ptr += nr_cpu_ids * sizeof(void **);
7021 
7022                 root_task_group.rt_rq = (struct rt_rq **)ptr;
7023                 ptr += nr_cpu_ids * sizeof(void **);
7024 
7025 #endif /* CONFIG_RT_GROUP_SCHED */
7026 #ifdef CONFIG_CPUMASK_OFFSTACK
7027                 for_each_possible_cpu(i) {
7028                         per_cpu(load_balance_mask, i) = (void *)ptr;
7029                         ptr += cpumask_size();
7030                 }
7031 #endif /* CONFIG_CPUMASK_OFFSTACK */
7032         }
7033 
7034         init_rt_bandwidth(&def_rt_bandwidth,
7035                         global_rt_period(), global_rt_runtime());
7036         init_dl_bandwidth(&def_dl_bandwidth,
7037                         global_rt_period(), global_rt_runtime());
7038 
7039 #ifdef CONFIG_SMP
7040         init_defrootdomain();
7041 #endif
7042 
7043 #ifdef CONFIG_RT_GROUP_SCHED
7044         init_rt_bandwidth(&root_task_group.rt_bandwidth,
7045                         global_rt_period(), global_rt_runtime());
7046 #endif /* CONFIG_RT_GROUP_SCHED */
7047 
7048 #ifdef CONFIG_CGROUP_SCHED
7049         list_add(&root_task_group.list, &task_groups);
7050         INIT_LIST_HEAD(&root_task_group.children);
7051         INIT_LIST_HEAD(&root_task_group.siblings);
7052         autogroup_init(&init_task);
7053 
7054 #endif /* CONFIG_CGROUP_SCHED */
7055 
7056         for_each_possible_cpu(i) {
7057                 struct rq *rq;
7058 
7059                 rq = cpu_rq(i);
7060                 raw_spin_lock_init(&rq->lock);
7061                 rq->nr_running = 0;
7062                 rq->calc_load_active = 0;
7063                 rq->calc_load_update = jiffies + LOAD_FREQ;
7064                 init_cfs_rq(&rq->cfs);
7065                 init_rt_rq(&rq->rt, rq);
7066                 init_dl_rq(&rq->dl, rq);
7067 #ifdef CONFIG_FAIR_GROUP_SCHED
7068                 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7069                 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7070                 /*
7071                  * How much cpu bandwidth does root_task_group get?
7072                  *
7073                  * In case of task-groups formed thr' the cgroup filesystem, it
7074                  * gets 100% of the cpu resources in the system. This overall
7075                  * system cpu resource is divided among the tasks of
7076                  * root_task_group and its child task-groups in a fair manner,
7077                  * based on each entity's (task or task-group's) weight
7078                  * (se->load.weight).
7079                  *
7080                  * In other words, if root_task_group has 10 tasks of weight
7081                  * 1024) and two child groups A0 and A1 (of weight 1024 each),
7082                  * then A0's share of the cpu resource is:
7083                  *
7084                  *      A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7085                  *
7086                  * We achieve this by letting root_task_group's tasks sit
7087                  * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7088                  */
7089                 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7090                 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7091 #endif /* CONFIG_FAIR_GROUP_SCHED */
7092 
7093                 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7094 #ifdef CONFIG_RT_GROUP_SCHED
7095                 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7096 #endif
7097 
7098                 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7099                         rq->cpu_load[j] = 0;
7100 
7101                 rq->last_load_update_tick = jiffies;
7102 
7103 #ifdef CONFIG_SMP
7104                 rq->sd = NULL;
7105                 rq->rd = NULL;
7106                 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7107                 rq->post_schedule = 0;
7108                 rq->active_balance = 0;
7109                 rq->next_balance = jiffies;
7110                 rq->push_cpu = 0;
7111                 rq->cpu = i;
7112                 rq->online = 0;
7113                 rq->idle_stamp = 0;
7114                 rq->avg_idle = 2*sysctl_sched_migration_cost;
7115                 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7116 
7117                 INIT_LIST_HEAD(&rq->cfs_tasks);
7118 
7119                 rq_attach_root(rq, &def_root_domain);
7120 #ifdef CONFIG_NO_HZ_COMMON
7121                 rq->nohz_flags = 0;
7122 #endif
7123 #ifdef CONFIG_NO_HZ_FULL
7124                 rq->last_sched_tick = 0;
7125 #endif
7126 #endif
7127                 init_rq_hrtick(rq);
7128                 atomic_set(&rq->nr_iowait, 0);
7129         }
7130 
7131         set_load_weight(&init_task);
7132 
7133 #ifdef CONFIG_PREEMPT_NOTIFIERS
7134         INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7135 #endif
7136 
7137         /*
7138          * The boot idle thread does lazy MMU switching as well:
7139          */
7140         atomic_inc(&init_mm.mm_count);
7141         enter_lazy_tlb(&init_mm, current);
7142 
7143         /*
7144          * Make us the idle thread. Technically, schedule() should not be
7145          * called from this thread, however somewhere below it might be,
7146          * but because we are the idle thread, we just pick up running again
7147          * when this runqueue becomes "idle".
7148          */
7149         init_idle(current, smp_processor_id());
7150 
7151         calc_load_update = jiffies + LOAD_FREQ;
7152 
7153         /*
7154          * During early bootup we pretend to be a normal task:
7155          */
7156         current->sched_class = &fair_sched_class;
7157 
7158 #ifdef CONFIG_SMP
7159         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7160         /* May be allocated at isolcpus cmdline parse time */
7161         if (cpu_isolated_map == NULL)
7162                 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7163         idle_thread_set_boot_cpu();
7164         set_cpu_rq_start_time();
7165 #endif
7166         init_sched_fair_class();
7167 
7168         scheduler_running = 1;
7169 }
7170 
7171 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7172 static inline int preempt_count_equals(int preempt_offset)
7173 {
7174         int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7175 
7176         return (nested == preempt_offset);
7177 }
7178 
7179 void __might_sleep(const char *file, int line, int preempt_offset)
7180 {
7181         static unsigned long prev_jiffy;        /* ratelimiting */
7182 
7183         rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7184         if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7185              !is_idle_task(current)) ||
7186             system_state != SYSTEM_RUNNING || oops_in_progress)
7187                 return;
7188         if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7189                 return;
7190         prev_jiffy = jiffies;
7191 
7192         printk(KERN_ERR
7193                 "BUG: sleeping function called from invalid context at %s:%d\n",
7194                         file, line);
7195         printk(KERN_ERR
7196                 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7197                         in_atomic(), irqs_disabled(),
7198                         current->pid, current->comm);
7199 
7200         debug_show_held_locks(current);
7201         if (irqs_disabled())
7202                 print_irqtrace_events(current);
7203 #ifdef CONFIG_DEBUG_PREEMPT
7204         if (!preempt_count_equals(preempt_offset)) {
7205                 pr_err("Preemption disabled at:");
7206                 print_ip_sym(current->preempt_disable_ip);
7207                 pr_cont("\n");
7208         }
7209 #endif
7210         dump_stack();
7211 }
7212 EXPORT_SYMBOL(__might_sleep);
7213 #endif
7214 
7215 #ifdef CONFIG_MAGIC_SYSRQ
7216 static void normalize_task(struct rq *rq, struct task_struct *p)
7217 {
7218         const struct sched_class *prev_class = p->sched_class;
7219         struct sched_attr attr = {
7220                 .sched_policy = SCHED_NORMAL,
7221         };
7222         int old_prio = p->prio;
7223         int queued;
7224 
7225         queued = task_on_rq_queued(p);
7226         if (queued)
7227                 dequeue_task(rq, p, 0);
7228         __setscheduler(rq, p, &attr);
7229         if (queued) {
7230                 enqueue_task(rq, p, 0);
7231                 resched_curr(rq);
7232         }
7233 
7234         check_class_changed(rq, p, prev_class, old_prio);
7235 }
7236 
7237 void normalize_rt_tasks(void)
7238 {
7239         struct task_struct *g, *p;
7240         unsigned long flags;
7241         struct rq *rq;
7242 
7243         read_lock(&tasklist_lock);
7244         for_each_process_thread(g, p) {
7245                 /*
7246                  * Only normalize user tasks:
7247                  */
7248                 if (p->flags & PF_KTHREAD)
7249                         continue;
7250 
7251                 p->se.exec_start                = 0;
7252 #ifdef CONFIG_SCHEDSTATS
7253                 p->se.statistics.wait_start     = 0;
7254                 p->se.statistics.sleep_start    = 0;
7255                 p->se.statistics.block_start    = 0;
7256 #endif
7257 
7258                 if (!dl_task(p) && !rt_task(p)) {
7259                         /*
7260                          * Renice negative nice level userspace
7261                          * tasks back to 0:
7262                          */
7263                         if (task_nice(p) < 0)
7264                                 set_user_nice(p, 0);
7265                         continue;
7266                 }
7267 
7268                 rq = task_rq_lock(p, &flags);
7269                 normalize_task(rq, p);
7270                 task_rq_unlock(rq, p, &flags);
7271         }
7272         read_unlock(&tasklist_lock);
7273 }
7274 
7275 #endif /* CONFIG_MAGIC_SYSRQ */
7276 
7277 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7278 /*
7279  * These functions are only useful for the IA64 MCA handling, or kdb.
7280  *
7281  * They can only be called when the whole system has been
7282  * stopped - every CPU needs to be quiescent, and no scheduling
7283  * activity can take place. Using them for anything else would
7284  * be a serious bug, and as a result, they aren't even visible
7285  * under any other configuration.
7286  */
7287 
7288 /**
7289  * curr_task - return the current task for a given cpu.
7290  * @cpu: the processor in question.
7291  *
7292  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7293  *
7294  * Return: The current task for @cpu.
7295  */
7296 struct task_struct *curr_task(int cpu)
7297 {
7298         return cpu_curr(cpu);
7299 }
7300 
7301 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7302 
7303 #ifdef CONFIG_IA64
7304 /**
7305  * set_curr_task - set the current task for a given cpu.
7306  * @cpu: the processor in question.
7307  * @p: the task pointer to set.
7308  *
7309  * Description: This function must only be used when non-maskable interrupts
7310  * are serviced on a separate stack. It allows the architecture to switch the
7311  * notion of the current task on a cpu in a non-blocking manner. This function
7312  * must be called with all CPU's synchronized, and interrupts disabled, the
7313  * and caller must save the original value of the current task (see
7314  * curr_task() above) and restore that value before reenabling interrupts and
7315  * re-starting the system.
7316  *
7317  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7318  */
7319 void set_curr_task(int cpu, struct task_struct *p)
7320 {
7321         cpu_curr(cpu) = p;
7322 }
7323 
7324 #endif
7325 
7326 #ifdef CONFIG_CGROUP_SCHED
7327 /* task_group_lock serializes the addition/removal of task groups */
7328 static DEFINE_SPINLOCK(task_group_lock);
7329 
7330 static void free_sched_group(struct task_group *tg)
7331 {
7332         free_fair_sched_group(tg);
7333         free_rt_sched_group(tg);
7334         autogroup_free(tg);
7335         kfree(tg);
7336 }
7337 
7338 /* allocate runqueue etc for a new task group */
7339 struct task_group *sched_create_group(struct task_group *parent)
7340 {
7341         struct task_group *tg;
7342 
7343         tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7344         if (!tg)
7345                 return ERR_PTR(-ENOMEM);
7346 
7347         if (!alloc_fair_sched_group(tg, parent))
7348                 goto err;
7349 
7350         if (!alloc_rt_sched_group(tg, parent))
7351                 goto err;
7352 
7353         return tg;
7354 
7355 err:
7356         free_sched_group(tg);
7357         return ERR_PTR(-ENOMEM);
7358 }
7359 
7360 void sched_online_group(struct task_group *tg, struct task_group *parent)
7361 {
7362         unsigned long flags;
7363 
7364         spin_lock_irqsave(&task_group_lock, flags);
7365         list_add_rcu(&tg->list, &task_groups);
7366 
7367         WARN_ON(!parent); /* root should already exist */
7368 
7369         tg->parent = parent;
7370         INIT_LIST_HEAD(&tg->children);
7371         list_add_rcu(&tg->siblings, &parent->children);
7372         spin_unlock_irqrestore(&task_group_lock, flags);
7373 }
7374 
7375 /* rcu callback to free various structures associated with a task group */
7376 static void free_sched_group_rcu(struct rcu_head *rhp)
7377 {
7378         /* now it should be safe to free those cfs_rqs */
7379         free_sched_group(container_of(rhp, struct task_group, rcu));
7380 }
7381 
7382 /* Destroy runqueue etc associated with a task group */
7383 void sched_destroy_group(struct task_group *tg)
7384 {
7385         /* wait for possible concurrent references to cfs_rqs complete */
7386         call_rcu(&tg->rcu, free_sched_group_rcu);
7387 }
7388 
7389 void sched_offline_group(struct task_group *tg)
7390 {
7391         unsigned long flags;
7392         int i;
7393 
7394         /* end participation in shares distribution */
7395         for_each_possible_cpu(i)
7396                 unregister_fair_sched_group(tg, i);
7397 
7398         spin_lock_irqsave(&task_group_lock, flags);
7399         list_del_rcu(&tg->list);
7400         list_del_rcu(&tg->siblings);
7401         spin_unlock_irqrestore(&task_group_lock, flags);
7402 }
7403 
7404 /* change task's runqueue when it moves between groups.
7405  *      The caller of this function should have put the task in its new group