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