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 DEFINE_MUTEX(sched_domains_mutex);
 94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
 95 
 96 static void update_rq_clock_task(struct rq *rq, s64 delta);
 97 
 98 void update_rq_clock(struct rq *rq)
 99 {
100         s64 delta;
101 
102         lockdep_assert_held(&rq->lock);
103 
104         if (rq->clock_skip_update & RQCF_ACT_SKIP)
105                 return;
106 
107         delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
108         if (delta < 0)
109                 return;
110         rq->clock += delta;
111         update_rq_clock_task(rq, delta);
112 }
113 
114 /*
115  * Debugging: various feature bits
116  */
117 
118 #define SCHED_FEAT(name, enabled)       \
119         (1UL << __SCHED_FEAT_##name) * enabled |
120 
121 const_debug unsigned int sysctl_sched_features =
122 #include "features.h"
123         0;
124 
125 #undef SCHED_FEAT
126 
127 #ifdef CONFIG_SCHED_DEBUG
128 #define SCHED_FEAT(name, enabled)       \
129         #name ,
130 
131 static const char * const sched_feat_names[] = {
132 #include "features.h"
133 };
134 
135 #undef SCHED_FEAT
136 
137 static int sched_feat_show(struct seq_file *m, void *v)
138 {
139         int i;
140 
141         for (i = 0; i < __SCHED_FEAT_NR; i++) {
142                 if (!(sysctl_sched_features & (1UL << i)))
143                         seq_puts(m, "NO_");
144                 seq_printf(m, "%s ", sched_feat_names[i]);
145         }
146         seq_puts(m, "\n");
147 
148         return 0;
149 }
150 
151 #ifdef HAVE_JUMP_LABEL
152 
153 #define jump_label_key__true  STATIC_KEY_INIT_TRUE
154 #define jump_label_key__false STATIC_KEY_INIT_FALSE
155 
156 #define SCHED_FEAT(name, enabled)       \
157         jump_label_key__##enabled ,
158 
159 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
160 #include "features.h"
161 };
162 
163 #undef SCHED_FEAT
164 
165 static void sched_feat_disable(int i)
166 {
167         static_key_disable(&sched_feat_keys[i]);
168 }
169 
170 static void sched_feat_enable(int i)
171 {
172         static_key_enable(&sched_feat_keys[i]);
173 }
174 #else
175 static void sched_feat_disable(int i) { };
176 static void sched_feat_enable(int i) { };
177 #endif /* HAVE_JUMP_LABEL */
178 
179 static int sched_feat_set(char *cmp)
180 {
181         int i;
182         int neg = 0;
183 
184         if (strncmp(cmp, "NO_", 3) == 0) {
185                 neg = 1;
186                 cmp += 3;
187         }
188 
189         for (i = 0; i < __SCHED_FEAT_NR; i++) {
190                 if (strcmp(cmp, sched_feat_names[i]) == 0) {
191                         if (neg) {
192                                 sysctl_sched_features &= ~(1UL << i);
193                                 sched_feat_disable(i);
194                         } else {
195                                 sysctl_sched_features |= (1UL << i);
196                                 sched_feat_enable(i);
197                         }
198                         break;
199                 }
200         }
201 
202         return i;
203 }
204 
205 static ssize_t
206 sched_feat_write(struct file *filp, const char __user *ubuf,
207                 size_t cnt, loff_t *ppos)
208 {
209         char buf[64];
210         char *cmp;
211         int i;
212         struct inode *inode;
213 
214         if (cnt > 63)
215                 cnt = 63;
216 
217         if (copy_from_user(&buf, ubuf, cnt))
218                 return -EFAULT;
219 
220         buf[cnt] = 0;
221         cmp = strstrip(buf);
222 
223         /* Ensure the static_key remains in a consistent state */
224         inode = file_inode(filp);
225         mutex_lock(&inode->i_mutex);
226         i = sched_feat_set(cmp);
227         mutex_unlock(&inode->i_mutex);
228         if (i == __SCHED_FEAT_NR)
229                 return -EINVAL;
230 
231         *ppos += cnt;
232 
233         return cnt;
234 }
235 
236 static int sched_feat_open(struct inode *inode, struct file *filp)
237 {
238         return single_open(filp, sched_feat_show, NULL);
239 }
240 
241 static const struct file_operations sched_feat_fops = {
242         .open           = sched_feat_open,
243         .write          = sched_feat_write,
244         .read           = seq_read,
245         .llseek         = seq_lseek,
246         .release        = single_release,
247 };
248 
249 static __init int sched_init_debug(void)
250 {
251         debugfs_create_file("sched_features", 0644, NULL, NULL,
252                         &sched_feat_fops);
253 
254         return 0;
255 }
256 late_initcall(sched_init_debug);
257 #endif /* CONFIG_SCHED_DEBUG */
258 
259 /*
260  * Number of tasks to iterate in a single balance run.
261  * Limited because this is done with IRQs disabled.
262  */
263 const_debug unsigned int sysctl_sched_nr_migrate = 32;
264 
265 /*
266  * period over which we average the RT time consumption, measured
267  * in ms.
268  *
269  * default: 1s
270  */
271 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
272 
273 /*
274  * period over which we measure -rt task cpu usage in us.
275  * default: 1s
276  */
277 unsigned int sysctl_sched_rt_period = 1000000;
278 
279 __read_mostly int scheduler_running;
280 
281 /*
282  * part of the period that we allow rt tasks to run in us.
283  * default: 0.95s
284  */
285 int sysctl_sched_rt_runtime = 950000;
286 
287 /* cpus with isolated domains */
288 cpumask_var_t cpu_isolated_map;
289 
290 /*
291  * this_rq_lock - lock this runqueue and disable interrupts.
292  */
293 static struct rq *this_rq_lock(void)
294         __acquires(rq->lock)
295 {
296         struct rq *rq;
297 
298         local_irq_disable();
299         rq = this_rq();
300         raw_spin_lock(&rq->lock);
301 
302         return rq;
303 }
304 
305 #ifdef CONFIG_SCHED_HRTICK
306 /*
307  * Use HR-timers to deliver accurate preemption points.
308  */
309 
310 static void hrtick_clear(struct rq *rq)
311 {
312         if (hrtimer_active(&rq->hrtick_timer))
313                 hrtimer_cancel(&rq->hrtick_timer);
314 }
315 
316 /*
317  * High-resolution timer tick.
318  * Runs from hardirq context with interrupts disabled.
319  */
320 static enum hrtimer_restart hrtick(struct hrtimer *timer)
321 {
322         struct rq *rq = container_of(timer, struct rq, hrtick_timer);
323 
324         WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
325 
326         raw_spin_lock(&rq->lock);
327         update_rq_clock(rq);
328         rq->curr->sched_class->task_tick(rq, rq->curr, 1);
329         raw_spin_unlock(&rq->lock);
330 
331         return HRTIMER_NORESTART;
332 }
333 
334 #ifdef CONFIG_SMP
335 
336 static void __hrtick_restart(struct rq *rq)
337 {
338         struct hrtimer *timer = &rq->hrtick_timer;
339 
340         hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
341 }
342 
343 /*
344  * called from hardirq (IPI) context
345  */
346 static void __hrtick_start(void *arg)
347 {
348         struct rq *rq = arg;
349 
350         raw_spin_lock(&rq->lock);
351         __hrtick_restart(rq);
352         rq->hrtick_csd_pending = 0;
353         raw_spin_unlock(&rq->lock);
354 }
355 
356 /*
357  * Called to set the hrtick timer state.
358  *
359  * called with rq->lock held and irqs disabled
360  */
361 void hrtick_start(struct rq *rq, u64 delay)
362 {
363         struct hrtimer *timer = &rq->hrtick_timer;
364         ktime_t time;
365         s64 delta;
366 
367         /*
368          * Don't schedule slices shorter than 10000ns, that just
369          * doesn't make sense and can cause timer DoS.
370          */
371         delta = max_t(s64, delay, 10000LL);
372         time = ktime_add_ns(timer->base->get_time(), delta);
373 
374         hrtimer_set_expires(timer, time);
375 
376         if (rq == this_rq()) {
377                 __hrtick_restart(rq);
378         } else if (!rq->hrtick_csd_pending) {
379                 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
380                 rq->hrtick_csd_pending = 1;
381         }
382 }
383 
384 static int
385 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
386 {
387         int cpu = (int)(long)hcpu;
388 
389         switch (action) {
390         case CPU_UP_CANCELED:
391         case CPU_UP_CANCELED_FROZEN:
392         case CPU_DOWN_PREPARE:
393         case CPU_DOWN_PREPARE_FROZEN:
394         case CPU_DEAD:
395         case CPU_DEAD_FROZEN:
396                 hrtick_clear(cpu_rq(cpu));
397                 return NOTIFY_OK;
398         }
399 
400         return NOTIFY_DONE;
401 }
402 
403 static __init void init_hrtick(void)
404 {
405         hotcpu_notifier(hotplug_hrtick, 0);
406 }
407 #else
408 /*
409  * Called to set the hrtick timer state.
410  *
411  * called with rq->lock held and irqs disabled
412  */
413 void hrtick_start(struct rq *rq, u64 delay)
414 {
415         /*
416          * Don't schedule slices shorter than 10000ns, that just
417          * doesn't make sense. Rely on vruntime for fairness.
418          */
419         delay = max_t(u64, delay, 10000LL);
420         hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
421                       HRTIMER_MODE_REL_PINNED);
422 }
423 
424 static inline void init_hrtick(void)
425 {
426 }
427 #endif /* CONFIG_SMP */
428 
429 static void init_rq_hrtick(struct rq *rq)
430 {
431 #ifdef CONFIG_SMP
432         rq->hrtick_csd_pending = 0;
433 
434         rq->hrtick_csd.flags = 0;
435         rq->hrtick_csd.func = __hrtick_start;
436         rq->hrtick_csd.info = rq;
437 #endif
438 
439         hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
440         rq->hrtick_timer.function = hrtick;
441 }
442 #else   /* CONFIG_SCHED_HRTICK */
443 static inline void hrtick_clear(struct rq *rq)
444 {
445 }
446 
447 static inline void init_rq_hrtick(struct rq *rq)
448 {
449 }
450 
451 static inline void init_hrtick(void)
452 {
453 }
454 #endif  /* CONFIG_SCHED_HRTICK */
455 
456 /*
457  * cmpxchg based fetch_or, macro so it works for different integer types
458  */
459 #define fetch_or(ptr, val)                                              \
460 ({      typeof(*(ptr)) __old, __val = *(ptr);                           \
461         for (;;) {                                                      \
462                 __old = cmpxchg((ptr), __val, __val | (val));           \
463                 if (__old == __val)                                     \
464                         break;                                          \
465                 __val = __old;                                          \
466         }                                                               \
467         __old;                                                          \
468 })
469 
470 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
471 /*
472  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
473  * this avoids any races wrt polling state changes and thereby avoids
474  * spurious IPIs.
475  */
476 static bool set_nr_and_not_polling(struct task_struct *p)
477 {
478         struct thread_info *ti = task_thread_info(p);
479         return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
480 }
481 
482 /*
483  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
484  *
485  * If this returns true, then the idle task promises to call
486  * sched_ttwu_pending() and reschedule soon.
487  */
488 static bool set_nr_if_polling(struct task_struct *p)
489 {
490         struct thread_info *ti = task_thread_info(p);
491         typeof(ti->flags) old, val = READ_ONCE(ti->flags);
492 
493         for (;;) {
494                 if (!(val & _TIF_POLLING_NRFLAG))
495                         return false;
496                 if (val & _TIF_NEED_RESCHED)
497                         return true;
498                 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
499                 if (old == val)
500                         break;
501                 val = old;
502         }
503         return true;
504 }
505 
506 #else
507 static bool set_nr_and_not_polling(struct task_struct *p)
508 {
509         set_tsk_need_resched(p);
510         return true;
511 }
512 
513 #ifdef CONFIG_SMP
514 static bool set_nr_if_polling(struct task_struct *p)
515 {
516         return false;
517 }
518 #endif
519 #endif
520 
521 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
522 {
523         struct wake_q_node *node = &task->wake_q;
524 
525         /*
526          * Atomically grab the task, if ->wake_q is !nil already it means
527          * its already queued (either by us or someone else) and will get the
528          * wakeup due to that.
529          *
530          * This cmpxchg() implies a full barrier, which pairs with the write
531          * barrier implied by the wakeup in wake_up_list().
532          */
533         if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
534                 return;
535 
536         get_task_struct(task);
537 
538         /*
539          * The head is context local, there can be no concurrency.
540          */
541         *head->lastp = node;
542         head->lastp = &node->next;
543 }
544 
545 void wake_up_q(struct wake_q_head *head)
546 {
547         struct wake_q_node *node = head->first;
548 
549         while (node != WAKE_Q_TAIL) {
550                 struct task_struct *task;
551 
552                 task = container_of(node, struct task_struct, wake_q);
553                 BUG_ON(!task);
554                 /* task can safely be re-inserted now */
555                 node = node->next;
556                 task->wake_q.next = NULL;
557 
558                 /*
559                  * wake_up_process() implies a wmb() to pair with the queueing
560                  * in wake_q_add() so as not to miss wakeups.
561                  */
562                 wake_up_process(task);
563                 put_task_struct(task);
564         }
565 }
566 
567 /*
568  * resched_curr - mark rq's current task 'to be rescheduled now'.
569  *
570  * On UP this means the setting of the need_resched flag, on SMP it
571  * might also involve a cross-CPU call to trigger the scheduler on
572  * the target CPU.
573  */
574 void resched_curr(struct rq *rq)
575 {
576         struct task_struct *curr = rq->curr;
577         int cpu;
578 
579         lockdep_assert_held(&rq->lock);
580 
581         if (test_tsk_need_resched(curr))
582                 return;
583 
584         cpu = cpu_of(rq);
585 
586         if (cpu == smp_processor_id()) {
587                 set_tsk_need_resched(curr);
588                 set_preempt_need_resched();
589                 return;
590         }
591 
592         if (set_nr_and_not_polling(curr))
593                 smp_send_reschedule(cpu);
594         else
595                 trace_sched_wake_idle_without_ipi(cpu);
596 }
597 
598 void resched_cpu(int cpu)
599 {
600         struct rq *rq = cpu_rq(cpu);
601         unsigned long flags;
602 
603         if (!raw_spin_trylock_irqsave(&rq->lock, flags))
604                 return;
605         resched_curr(rq);
606         raw_spin_unlock_irqrestore(&rq->lock, flags);
607 }
608 
609 #ifdef CONFIG_SMP
610 #ifdef CONFIG_NO_HZ_COMMON
611 /*
612  * In the semi idle case, use the nearest busy cpu for migrating timers
613  * from an idle cpu.  This is good for power-savings.
614  *
615  * We don't do similar optimization for completely idle system, as
616  * selecting an idle cpu will add more delays to the timers than intended
617  * (as that cpu's timer base may not be uptodate wrt jiffies etc).
618  */
619 int get_nohz_timer_target(void)
620 {
621         int i, cpu = smp_processor_id();
622         struct sched_domain *sd;
623 
624         if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
625                 return cpu;
626 
627         rcu_read_lock();
628         for_each_domain(cpu, sd) {
629                 for_each_cpu(i, sched_domain_span(sd)) {
630                         if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
631                                 cpu = i;
632                                 goto unlock;
633                         }
634                 }
635         }
636 
637         if (!is_housekeeping_cpu(cpu))
638                 cpu = housekeeping_any_cpu();
639 unlock:
640         rcu_read_unlock();
641         return cpu;
642 }
643 /*
644  * When add_timer_on() enqueues a timer into the timer wheel of an
645  * idle CPU then this timer might expire before the next timer event
646  * which is scheduled to wake up that CPU. In case of a completely
647  * idle system the next event might even be infinite time into the
648  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
649  * leaves the inner idle loop so the newly added timer is taken into
650  * account when the CPU goes back to idle and evaluates the timer
651  * wheel for the next timer event.
652  */
653 static void wake_up_idle_cpu(int cpu)
654 {
655         struct rq *rq = cpu_rq(cpu);
656 
657         if (cpu == smp_processor_id())
658                 return;
659 
660         if (set_nr_and_not_polling(rq->idle))
661                 smp_send_reschedule(cpu);
662         else
663                 trace_sched_wake_idle_without_ipi(cpu);
664 }
665 
666 static bool wake_up_full_nohz_cpu(int cpu)
667 {
668         /*
669          * We just need the target to call irq_exit() and re-evaluate
670          * the next tick. The nohz full kick at least implies that.
671          * If needed we can still optimize that later with an
672          * empty IRQ.
673          */
674         if (tick_nohz_full_cpu(cpu)) {
675                 if (cpu != smp_processor_id() ||
676                     tick_nohz_tick_stopped())
677                         tick_nohz_full_kick_cpu(cpu);
678                 return true;
679         }
680 
681         return false;
682 }
683 
684 void wake_up_nohz_cpu(int cpu)
685 {
686         if (!wake_up_full_nohz_cpu(cpu))
687                 wake_up_idle_cpu(cpu);
688 }
689 
690 static inline bool got_nohz_idle_kick(void)
691 {
692         int cpu = smp_processor_id();
693 
694         if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
695                 return false;
696 
697         if (idle_cpu(cpu) && !need_resched())
698                 return true;
699 
700         /*
701          * We can't run Idle Load Balance on this CPU for this time so we
702          * cancel it and clear NOHZ_BALANCE_KICK
703          */
704         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
705         return false;
706 }
707 
708 #else /* CONFIG_NO_HZ_COMMON */
709 
710 static inline bool got_nohz_idle_kick(void)
711 {
712         return false;
713 }
714 
715 #endif /* CONFIG_NO_HZ_COMMON */
716 
717 #ifdef CONFIG_NO_HZ_FULL
718 bool sched_can_stop_tick(void)
719 {
720         /*
721          * FIFO realtime policy runs the highest priority task. Other runnable
722          * tasks are of a lower priority. The scheduler tick does nothing.
723          */
724         if (current->policy == SCHED_FIFO)
725                 return true;
726 
727         /*
728          * Round-robin realtime tasks time slice with other tasks at the same
729          * realtime priority. Is this task the only one at this priority?
730          */
731         if (current->policy == SCHED_RR) {
732                 struct sched_rt_entity *rt_se = &current->rt;
733 
734                 return rt_se->run_list.prev == rt_se->run_list.next;
735         }
736 
737         /*
738          * More than one running task need preemption.
739          * nr_running update is assumed to be visible
740          * after IPI is sent from wakers.
741          */
742         if (this_rq()->nr_running > 1)
743                 return false;
744 
745         return true;
746 }
747 #endif /* CONFIG_NO_HZ_FULL */
748 
749 void sched_avg_update(struct rq *rq)
750 {
751         s64 period = sched_avg_period();
752 
753         while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
754                 /*
755                  * Inline assembly required to prevent the compiler
756                  * optimising this loop into a divmod call.
757                  * See __iter_div_u64_rem() for another example of this.
758                  */
759                 asm("" : "+rm" (rq->age_stamp));
760                 rq->age_stamp += period;
761                 rq->rt_avg /= 2;
762         }
763 }
764 
765 #endif /* CONFIG_SMP */
766 
767 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
768                         (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
769 /*
770  * Iterate task_group tree rooted at *from, calling @down when first entering a
771  * node and @up when leaving it for the final time.
772  *
773  * Caller must hold rcu_lock or sufficient equivalent.
774  */
775 int walk_tg_tree_from(struct task_group *from,
776                              tg_visitor down, tg_visitor up, void *data)
777 {
778         struct task_group *parent, *child;
779         int ret;
780 
781         parent = from;
782 
783 down:
784         ret = (*down)(parent, data);
785         if (ret)
786                 goto out;
787         list_for_each_entry_rcu(child, &parent->children, siblings) {
788                 parent = child;
789                 goto down;
790 
791 up:
792                 continue;
793         }
794         ret = (*up)(parent, data);
795         if (ret || parent == from)
796                 goto out;
797 
798         child = parent;
799         parent = parent->parent;
800         if (parent)
801                 goto up;
802 out:
803         return ret;
804 }
805 
806 int tg_nop(struct task_group *tg, void *data)
807 {
808         return 0;
809 }
810 #endif
811 
812 static void set_load_weight(struct task_struct *p)
813 {
814         int prio = p->static_prio - MAX_RT_PRIO;
815         struct load_weight *load = &p->se.load;
816 
817         /*
818          * SCHED_IDLE tasks get minimal weight:
819          */
820         if (p->policy == SCHED_IDLE) {
821                 load->weight = scale_load(WEIGHT_IDLEPRIO);
822                 load->inv_weight = WMULT_IDLEPRIO;
823                 return;
824         }
825 
826         load->weight = scale_load(prio_to_weight[prio]);
827         load->inv_weight = prio_to_wmult[prio];
828 }
829 
830 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
831 {
832         update_rq_clock(rq);
833         sched_info_queued(rq, p);
834         p->sched_class->enqueue_task(rq, p, flags);
835 }
836 
837 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
838 {
839         update_rq_clock(rq);
840         sched_info_dequeued(rq, p);
841         p->sched_class->dequeue_task(rq, p, flags);
842 }
843 
844 void activate_task(struct rq *rq, struct task_struct *p, int flags)
845 {
846         if (task_contributes_to_load(p))
847                 rq->nr_uninterruptible--;
848 
849         enqueue_task(rq, p, flags);
850 }
851 
852 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
853 {
854         if (task_contributes_to_load(p))
855                 rq->nr_uninterruptible++;
856 
857         dequeue_task(rq, p, flags);
858 }
859 
860 static void update_rq_clock_task(struct rq *rq, s64 delta)
861 {
862 /*
863  * In theory, the compile should just see 0 here, and optimize out the call
864  * to sched_rt_avg_update. But I don't trust it...
865  */
866 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
867         s64 steal = 0, irq_delta = 0;
868 #endif
869 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
870         irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
871 
872         /*
873          * Since irq_time is only updated on {soft,}irq_exit, we might run into
874          * this case when a previous update_rq_clock() happened inside a
875          * {soft,}irq region.
876          *
877          * When this happens, we stop ->clock_task and only update the
878          * prev_irq_time stamp to account for the part that fit, so that a next
879          * update will consume the rest. This ensures ->clock_task is
880          * monotonic.
881          *
882          * It does however cause some slight miss-attribution of {soft,}irq
883          * time, a more accurate solution would be to update the irq_time using
884          * the current rq->clock timestamp, except that would require using
885          * atomic ops.
886          */
887         if (irq_delta > delta)
888                 irq_delta = delta;
889 
890         rq->prev_irq_time += irq_delta;
891         delta -= irq_delta;
892 #endif
893 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
894         if (static_key_false((&paravirt_steal_rq_enabled))) {
895                 steal = paravirt_steal_clock(cpu_of(rq));
896                 steal -= rq->prev_steal_time_rq;
897 
898                 if (unlikely(steal > delta))
899                         steal = delta;
900 
901                 rq->prev_steal_time_rq += steal;
902                 delta -= steal;
903         }
904 #endif
905 
906         rq->clock_task += delta;
907 
908 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
909         if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
910                 sched_rt_avg_update(rq, irq_delta + steal);
911 #endif
912 }
913 
914 void sched_set_stop_task(int cpu, struct task_struct *stop)
915 {
916         struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
917         struct task_struct *old_stop = cpu_rq(cpu)->stop;
918 
919         if (stop) {
920                 /*
921                  * Make it appear like a SCHED_FIFO task, its something
922                  * userspace knows about and won't get confused about.
923                  *
924                  * Also, it will make PI more or less work without too
925                  * much confusion -- but then, stop work should not
926                  * rely on PI working anyway.
927                  */
928                 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
929 
930                 stop->sched_class = &stop_sched_class;
931         }
932 
933         cpu_rq(cpu)->stop = stop;
934 
935         if (old_stop) {
936                 /*
937                  * Reset it back to a normal scheduling class so that
938                  * it can die in pieces.
939                  */
940                 old_stop->sched_class = &rt_sched_class;
941         }
942 }
943 
944 /*
945  * __normal_prio - return the priority that is based on the static prio
946  */
947 static inline int __normal_prio(struct task_struct *p)
948 {
949         return p->static_prio;
950 }
951 
952 /*
953  * Calculate the expected normal priority: i.e. priority
954  * without taking RT-inheritance into account. Might be
955  * boosted by interactivity modifiers. Changes upon fork,
956  * setprio syscalls, and whenever the interactivity
957  * estimator recalculates.
958  */
959 static inline int normal_prio(struct task_struct *p)
960 {
961         int prio;
962 
963         if (task_has_dl_policy(p))
964                 prio = MAX_DL_PRIO-1;
965         else if (task_has_rt_policy(p))
966                 prio = MAX_RT_PRIO-1 - p->rt_priority;
967         else
968                 prio = __normal_prio(p);
969         return prio;
970 }
971 
972 /*
973  * Calculate the current priority, i.e. the priority
974  * taken into account by the scheduler. This value might
975  * be boosted by RT tasks, or might be boosted by
976  * interactivity modifiers. Will be RT if the task got
977  * RT-boosted. If not then it returns p->normal_prio.
978  */
979 static int effective_prio(struct task_struct *p)
980 {
981         p->normal_prio = normal_prio(p);
982         /*
983          * If we are RT tasks or we were boosted to RT priority,
984          * keep the priority unchanged. Otherwise, update priority
985          * to the normal priority:
986          */
987         if (!rt_prio(p->prio))
988                 return p->normal_prio;
989         return p->prio;
990 }
991 
992 /**
993  * task_curr - is this task currently executing on a CPU?
994  * @p: the task in question.
995  *
996  * Return: 1 if the task is currently executing. 0 otherwise.
997  */
998 inline int task_curr(const struct task_struct *p)
999 {
1000         return cpu_curr(task_cpu(p)) == p;
1001 }
1002 
1003 /*
1004  * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1005  * use the balance_callback list if you want balancing.
1006  *
1007  * this means any call to check_class_changed() must be followed by a call to
1008  * balance_callback().
1009  */
1010 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1011                                        const struct sched_class *prev_class,
1012                                        int oldprio)
1013 {
1014         if (prev_class != p->sched_class) {
1015                 if (prev_class->switched_from)
1016                         prev_class->switched_from(rq, p);
1017 
1018                 p->sched_class->switched_to(rq, p);
1019         } else if (oldprio != p->prio || dl_task(p))
1020                 p->sched_class->prio_changed(rq, p, oldprio);
1021 }
1022 
1023 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1024 {
1025         const struct sched_class *class;
1026 
1027         if (p->sched_class == rq->curr->sched_class) {
1028                 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1029         } else {
1030                 for_each_class(class) {
1031                         if (class == rq->curr->sched_class)
1032                                 break;
1033                         if (class == p->sched_class) {
1034                                 resched_curr(rq);
1035                                 break;
1036                         }
1037                 }
1038         }
1039 
1040         /*
1041          * A queue event has occurred, and we're going to schedule.  In
1042          * this case, we can save a useless back to back clock update.
1043          */
1044         if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1045                 rq_clock_skip_update(rq, true);
1046 }
1047 
1048 #ifdef CONFIG_SMP
1049 /*
1050  * This is how migration works:
1051  *
1052  * 1) we invoke migration_cpu_stop() on the target CPU using
1053  *    stop_one_cpu().
1054  * 2) stopper starts to run (implicitly forcing the migrated thread
1055  *    off the CPU)
1056  * 3) it checks whether the migrated task is still in the wrong runqueue.
1057  * 4) if it's in the wrong runqueue then the migration thread removes
1058  *    it and puts it into the right queue.
1059  * 5) stopper completes and stop_one_cpu() returns and the migration
1060  *    is done.
1061  */
1062 
1063 /*
1064  * move_queued_task - move a queued task to new rq.
1065  *
1066  * Returns (locked) new rq. Old rq's lock is released.
1067  */
1068 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1069 {
1070         lockdep_assert_held(&rq->lock);
1071 
1072         dequeue_task(rq, p, 0);
1073         p->on_rq = TASK_ON_RQ_MIGRATING;
1074         set_task_cpu(p, new_cpu);
1075         raw_spin_unlock(&rq->lock);
1076 
1077         rq = cpu_rq(new_cpu);
1078 
1079         raw_spin_lock(&rq->lock);
1080         BUG_ON(task_cpu(p) != new_cpu);
1081         p->on_rq = TASK_ON_RQ_QUEUED;
1082         enqueue_task(rq, p, 0);
1083         check_preempt_curr(rq, p, 0);
1084 
1085         return rq;
1086 }
1087 
1088 struct migration_arg {
1089         struct task_struct *task;
1090         int dest_cpu;
1091 };
1092 
1093 /*
1094  * Move (not current) task off this cpu, onto dest cpu. We're doing
1095  * this because either it can't run here any more (set_cpus_allowed()
1096  * away from this CPU, or CPU going down), or because we're
1097  * attempting to rebalance this task on exec (sched_exec).
1098  *
1099  * So we race with normal scheduler movements, but that's OK, as long
1100  * as the task is no longer on this CPU.
1101  */
1102 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1103 {
1104         if (unlikely(!cpu_active(dest_cpu)))
1105                 return rq;
1106 
1107         /* Affinity changed (again). */
1108         if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1109                 return rq;
1110 
1111         rq = move_queued_task(rq, p, dest_cpu);
1112 
1113         return rq;
1114 }
1115 
1116 /*
1117  * migration_cpu_stop - this will be executed by a highprio stopper thread
1118  * and performs thread migration by bumping thread off CPU then
1119  * 'pushing' onto another runqueue.
1120  */
1121 static int migration_cpu_stop(void *data)
1122 {
1123         struct migration_arg *arg = data;
1124         struct task_struct *p = arg->task;
1125         struct rq *rq = this_rq();
1126 
1127         /*
1128          * The original target cpu might have gone down and we might
1129          * be on another cpu but it doesn't matter.
1130          */
1131         local_irq_disable();
1132         /*
1133          * We need to explicitly wake pending tasks before running
1134          * __migrate_task() such that we will not miss enforcing cpus_allowed
1135          * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1136          */
1137         sched_ttwu_pending();
1138 
1139         raw_spin_lock(&p->pi_lock);
1140         raw_spin_lock(&rq->lock);
1141         /*
1142          * If task_rq(p) != rq, it cannot be migrated here, because we're
1143          * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1144          * we're holding p->pi_lock.
1145          */
1146         if (task_rq(p) == rq && task_on_rq_queued(p))
1147                 rq = __migrate_task(rq, p, arg->dest_cpu);
1148         raw_spin_unlock(&rq->lock);
1149         raw_spin_unlock(&p->pi_lock);
1150 
1151         local_irq_enable();
1152         return 0;
1153 }
1154 
1155 /*
1156  * sched_class::set_cpus_allowed must do the below, but is not required to
1157  * actually call this function.
1158  */
1159 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1160 {
1161         cpumask_copy(&p->cpus_allowed, new_mask);
1162         p->nr_cpus_allowed = cpumask_weight(new_mask);
1163 }
1164 
1165 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1166 {
1167         struct rq *rq = task_rq(p);
1168         bool queued, running;
1169 
1170         lockdep_assert_held(&p->pi_lock);
1171 
1172         queued = task_on_rq_queued(p);
1173         running = task_current(rq, p);
1174 
1175         if (queued) {
1176                 /*
1177                  * Because __kthread_bind() calls this on blocked tasks without
1178                  * holding rq->lock.
1179                  */
1180                 lockdep_assert_held(&rq->lock);
1181                 dequeue_task(rq, p, 0);
1182         }
1183         if (running)
1184                 put_prev_task(rq, p);
1185 
1186         p->sched_class->set_cpus_allowed(p, new_mask);
1187 
1188         if (running)
1189                 p->sched_class->set_curr_task(rq);
1190         if (queued)
1191                 enqueue_task(rq, p, 0);
1192 }
1193 
1194 /*
1195  * Change a given task's CPU affinity. Migrate the thread to a
1196  * proper CPU and schedule it away if the CPU it's executing on
1197  * is removed from the allowed bitmask.
1198  *
1199  * NOTE: the caller must have a valid reference to the task, the
1200  * task must not exit() & deallocate itself prematurely. The
1201  * call is not atomic; no spinlocks may be held.
1202  */
1203 static int __set_cpus_allowed_ptr(struct task_struct *p,
1204                                   const struct cpumask *new_mask, bool check)
1205 {
1206         unsigned long flags;
1207         struct rq *rq;
1208         unsigned int dest_cpu;
1209         int ret = 0;
1210 
1211         rq = task_rq_lock(p, &flags);
1212 
1213         /*
1214          * Must re-check here, to close a race against __kthread_bind(),
1215          * sched_setaffinity() is not guaranteed to observe the flag.
1216          */
1217         if (check && (p->flags & PF_NO_SETAFFINITY)) {
1218                 ret = -EINVAL;
1219                 goto out;
1220         }
1221 
1222         if (cpumask_equal(&p->cpus_allowed, new_mask))
1223                 goto out;
1224 
1225         if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1226                 ret = -EINVAL;
1227                 goto out;
1228         }
1229 
1230         do_set_cpus_allowed(p, new_mask);
1231 
1232         /* Can the task run on the task's current CPU? If so, we're done */
1233         if (cpumask_test_cpu(task_cpu(p), new_mask))
1234                 goto out;
1235 
1236         dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1237         if (task_running(rq, p) || p->state == TASK_WAKING) {
1238                 struct migration_arg arg = { p, dest_cpu };
1239                 /* Need help from migration thread: drop lock and wait. */
1240                 task_rq_unlock(rq, p, &flags);
1241                 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1242                 tlb_migrate_finish(p->mm);
1243                 return 0;
1244         } else if (task_on_rq_queued(p)) {
1245                 /*
1246                  * OK, since we're going to drop the lock immediately
1247                  * afterwards anyway.
1248                  */
1249                 lockdep_unpin_lock(&rq->lock);
1250                 rq = move_queued_task(rq, p, dest_cpu);
1251                 lockdep_pin_lock(&rq->lock);
1252         }
1253 out:
1254         task_rq_unlock(rq, p, &flags);
1255 
1256         return ret;
1257 }
1258 
1259 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1260 {
1261         return __set_cpus_allowed_ptr(p, new_mask, false);
1262 }
1263 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1264 
1265 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1266 {
1267 #ifdef CONFIG_SCHED_DEBUG
1268         /*
1269          * We should never call set_task_cpu() on a blocked task,
1270          * ttwu() will sort out the placement.
1271          */
1272         WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1273                         !p->on_rq);
1274 
1275 #ifdef CONFIG_LOCKDEP
1276         /*
1277          * The caller should hold either p->pi_lock or rq->lock, when changing
1278          * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1279          *
1280          * sched_move_task() holds both and thus holding either pins the cgroup,
1281          * see task_group().
1282          *
1283          * Furthermore, all task_rq users should acquire both locks, see
1284          * task_rq_lock().
1285          */
1286         WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1287                                       lockdep_is_held(&task_rq(p)->lock)));
1288 #endif
1289 #endif
1290 
1291         trace_sched_migrate_task(p, new_cpu);
1292 
1293         if (task_cpu(p) != new_cpu) {
1294                 if (p->sched_class->migrate_task_rq)
1295                         p->sched_class->migrate_task_rq(p, new_cpu);
1296                 p->se.nr_migrations++;
1297                 perf_event_task_migrate(p);
1298         }
1299 
1300         __set_task_cpu(p, new_cpu);
1301 }
1302 
1303 static void __migrate_swap_task(struct task_struct *p, int cpu)
1304 {
1305         if (task_on_rq_queued(p)) {
1306                 struct rq *src_rq, *dst_rq;
1307 
1308                 src_rq = task_rq(p);
1309                 dst_rq = cpu_rq(cpu);
1310 
1311                 deactivate_task(src_rq, p, 0);
1312                 set_task_cpu(p, cpu);
1313                 activate_task(dst_rq, p, 0);
1314                 check_preempt_curr(dst_rq, p, 0);
1315         } else {
1316                 /*
1317                  * Task isn't running anymore; make it appear like we migrated
1318                  * it before it went to sleep. This means on wakeup we make the
1319                  * previous cpu our targer instead of where it really is.
1320                  */
1321                 p->wake_cpu = cpu;
1322         }
1323 }
1324 
1325 struct migration_swap_arg {
1326         struct task_struct *src_task, *dst_task;
1327         int src_cpu, dst_cpu;
1328 };
1329 
1330 static int migrate_swap_stop(void *data)
1331 {
1332         struct migration_swap_arg *arg = data;
1333         struct rq *src_rq, *dst_rq;
1334         int ret = -EAGAIN;
1335 
1336         src_rq = cpu_rq(arg->src_cpu);
1337         dst_rq = cpu_rq(arg->dst_cpu);
1338 
1339         double_raw_lock(&arg->src_task->pi_lock,
1340                         &arg->dst_task->pi_lock);
1341         double_rq_lock(src_rq, dst_rq);
1342         if (task_cpu(arg->dst_task) != arg->dst_cpu)
1343                 goto unlock;
1344 
1345         if (task_cpu(arg->src_task) != arg->src_cpu)
1346                 goto unlock;
1347 
1348         if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1349                 goto unlock;
1350 
1351         if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1352                 goto unlock;
1353 
1354         __migrate_swap_task(arg->src_task, arg->dst_cpu);
1355         __migrate_swap_task(arg->dst_task, arg->src_cpu);
1356 
1357         ret = 0;
1358 
1359 unlock:
1360         double_rq_unlock(src_rq, dst_rq);
1361         raw_spin_unlock(&arg->dst_task->pi_lock);
1362         raw_spin_unlock(&arg->src_task->pi_lock);
1363 
1364         return ret;
1365 }
1366 
1367 /*
1368  * Cross migrate two tasks
1369  */
1370 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1371 {
1372         struct migration_swap_arg arg;
1373         int ret = -EINVAL;
1374 
1375         arg = (struct migration_swap_arg){
1376                 .src_task = cur,
1377                 .src_cpu = task_cpu(cur),
1378                 .dst_task = p,
1379                 .dst_cpu = task_cpu(p),
1380         };
1381 
1382         if (arg.src_cpu == arg.dst_cpu)
1383                 goto out;
1384 
1385         /*
1386          * These three tests are all lockless; this is OK since all of them
1387          * will be re-checked with proper locks held further down the line.
1388          */
1389         if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1390                 goto out;
1391 
1392         if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1393                 goto out;
1394 
1395         if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1396                 goto out;
1397 
1398         trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1399         ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1400 
1401 out:
1402         return ret;
1403 }
1404 
1405 /*
1406  * wait_task_inactive - wait for a thread to unschedule.
1407  *
1408  * If @match_state is nonzero, it's the @p->state value just checked and
1409  * not expected to change.  If it changes, i.e. @p might have woken up,
1410  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1411  * we return a positive number (its total switch count).  If a second call
1412  * a short while later returns the same number, the caller can be sure that
1413  * @p has remained unscheduled the whole time.
1414  *
1415  * The caller must ensure that the task *will* unschedule sometime soon,
1416  * else this function might spin for a *long* time. This function can't
1417  * be called with interrupts off, or it may introduce deadlock with
1418  * smp_call_function() if an IPI is sent by the same process we are
1419  * waiting to become inactive.
1420  */
1421 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1422 {
1423         unsigned long flags;
1424         int running, queued;
1425         unsigned long ncsw;
1426         struct rq *rq;
1427 
1428         for (;;) {
1429                 /*
1430                  * We do the initial early heuristics without holding
1431                  * any task-queue locks at all. We'll only try to get
1432                  * the runqueue lock when things look like they will
1433                  * work out!
1434                  */
1435                 rq = task_rq(p);
1436 
1437                 /*
1438                  * If the task is actively running on another CPU
1439                  * still, just relax and busy-wait without holding
1440                  * any locks.
1441                  *
1442                  * NOTE! Since we don't hold any locks, it's not
1443                  * even sure that "rq" stays as the right runqueue!
1444                  * But we don't care, since "task_running()" will
1445                  * return false if the runqueue has changed and p
1446                  * is actually now running somewhere else!
1447                  */
1448                 while (task_running(rq, p)) {
1449                         if (match_state && unlikely(p->state != match_state))
1450                                 return 0;
1451                         cpu_relax();
1452                 }
1453 
1454                 /*
1455                  * Ok, time to look more closely! We need the rq
1456                  * lock now, to be *sure*. If we're wrong, we'll
1457                  * just go back and repeat.
1458                  */
1459                 rq = task_rq_lock(p, &flags);
1460                 trace_sched_wait_task(p);
1461                 running = task_running(rq, p);
1462                 queued = task_on_rq_queued(p);
1463                 ncsw = 0;
1464                 if (!match_state || p->state == match_state)
1465                         ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1466                 task_rq_unlock(rq, p, &flags);
1467 
1468                 /*
1469                  * If it changed from the expected state, bail out now.
1470                  */
1471                 if (unlikely(!ncsw))
1472                         break;
1473 
1474                 /*
1475                  * Was it really running after all now that we
1476                  * checked with the proper locks actually held?
1477                  *
1478                  * Oops. Go back and try again..
1479                  */
1480                 if (unlikely(running)) {
1481                         cpu_relax();
1482                         continue;
1483                 }
1484 
1485                 /*
1486                  * It's not enough that it's not actively running,
1487                  * it must be off the runqueue _entirely_, and not
1488                  * preempted!
1489                  *
1490                  * So if it was still runnable (but just not actively
1491                  * running right now), it's preempted, and we should
1492                  * yield - it could be a while.
1493                  */
1494                 if (unlikely(queued)) {
1495                         ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1496 
1497                         set_current_state(TASK_UNINTERRUPTIBLE);
1498                         schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1499                         continue;
1500                 }
1501 
1502                 /*
1503                  * Ahh, all good. It wasn't running, and it wasn't
1504                  * runnable, which means that it will never become
1505                  * running in the future either. We're all done!
1506                  */
1507                 break;
1508         }
1509 
1510         return ncsw;
1511 }
1512 
1513 /***
1514  * kick_process - kick a running thread to enter/exit the kernel
1515  * @p: the to-be-kicked thread
1516  *
1517  * Cause a process which is running on another CPU to enter
1518  * kernel-mode, without any delay. (to get signals handled.)
1519  *
1520  * NOTE: this function doesn't have to take the runqueue lock,
1521  * because all it wants to ensure is that the remote task enters
1522  * the kernel. If the IPI races and the task has been migrated
1523  * to another CPU then no harm is done and the purpose has been
1524  * achieved as well.
1525  */
1526 void kick_process(struct task_struct *p)
1527 {
1528         int cpu;
1529 
1530         preempt_disable();
1531         cpu = task_cpu(p);
1532         if ((cpu != smp_processor_id()) && task_curr(p))
1533                 smp_send_reschedule(cpu);
1534         preempt_enable();
1535 }
1536 EXPORT_SYMBOL_GPL(kick_process);
1537 
1538 /*
1539  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1540  */
1541 static int select_fallback_rq(int cpu, struct task_struct *p)
1542 {
1543         int nid = cpu_to_node(cpu);
1544         const struct cpumask *nodemask = NULL;
1545         enum { cpuset, possible, fail } state = cpuset;
1546         int dest_cpu;
1547 
1548         /*
1549          * If the node that the cpu is on has been offlined, cpu_to_node()
1550          * will return -1. There is no cpu on the node, and we should
1551          * select the cpu on the other node.
1552          */
1553         if (nid != -1) {
1554                 nodemask = cpumask_of_node(nid);
1555 
1556                 /* Look for allowed, online CPU in same node. */
1557                 for_each_cpu(dest_cpu, nodemask) {
1558                         if (!cpu_online(dest_cpu))
1559                                 continue;
1560                         if (!cpu_active(dest_cpu))
1561                                 continue;
1562                         if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1563                                 return dest_cpu;
1564                 }
1565         }
1566 
1567         for (;;) {
1568                 /* Any allowed, online CPU? */
1569                 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1570                         if (!cpu_online(dest_cpu))
1571                                 continue;
1572                         if (!cpu_active(dest_cpu))
1573                                 continue;
1574                         goto out;
1575                 }
1576 
1577                 switch (state) {
1578                 case cpuset:
1579                         /* No more Mr. Nice Guy. */
1580                         cpuset_cpus_allowed_fallback(p);
1581                         state = possible;
1582                         break;
1583 
1584                 case possible:
1585                         do_set_cpus_allowed(p, cpu_possible_mask);
1586                         state = fail;
1587                         break;
1588 
1589                 case fail:
1590                         BUG();
1591                         break;
1592                 }
1593         }
1594 
1595 out:
1596         if (state != cpuset) {
1597                 /*
1598                  * Don't tell them about moving exiting tasks or
1599                  * kernel threads (both mm NULL), since they never
1600                  * leave kernel.
1601                  */
1602                 if (p->mm && printk_ratelimit()) {
1603                         printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1604                                         task_pid_nr(p), p->comm, cpu);
1605                 }
1606         }
1607 
1608         return dest_cpu;
1609 }
1610 
1611 /*
1612  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1613  */
1614 static inline
1615 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1616 {
1617         lockdep_assert_held(&p->pi_lock);
1618 
1619         if (p->nr_cpus_allowed > 1)
1620                 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1621 
1622         /*
1623          * In order not to call set_task_cpu() on a blocking task we need
1624          * to rely on ttwu() to place the task on a valid ->cpus_allowed
1625          * cpu.
1626          *
1627          * Since this is common to all placement strategies, this lives here.
1628          *
1629          * [ this allows ->select_task() to simply return task_cpu(p) and
1630          *   not worry about this generic constraint ]
1631          */
1632         if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1633                      !cpu_online(cpu)))
1634                 cpu = select_fallback_rq(task_cpu(p), p);
1635 
1636         return cpu;
1637 }
1638 
1639 static void update_avg(u64 *avg, u64 sample)
1640 {
1641         s64 diff = sample - *avg;
1642         *avg += diff >> 3;
1643 }
1644 
1645 #else
1646 
1647 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1648                                          const struct cpumask *new_mask, bool check)
1649 {
1650         return set_cpus_allowed_ptr(p, new_mask);
1651 }
1652 
1653 #endif /* CONFIG_SMP */
1654 
1655 static void
1656 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1657 {
1658 #ifdef CONFIG_SCHEDSTATS
1659         struct rq *rq = this_rq();
1660 
1661 #ifdef CONFIG_SMP
1662         int this_cpu = smp_processor_id();
1663 
1664         if (cpu == this_cpu) {
1665                 schedstat_inc(rq, ttwu_local);
1666                 schedstat_inc(p, se.statistics.nr_wakeups_local);
1667         } else {
1668                 struct sched_domain *sd;
1669 
1670                 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1671                 rcu_read_lock();
1672                 for_each_domain(this_cpu, sd) {
1673                         if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1674                                 schedstat_inc(sd, ttwu_wake_remote);
1675                                 break;
1676                         }
1677                 }
1678                 rcu_read_unlock();
1679         }
1680 
1681         if (wake_flags & WF_MIGRATED)
1682                 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1683 
1684 #endif /* CONFIG_SMP */
1685 
1686         schedstat_inc(rq, ttwu_count);
1687         schedstat_inc(p, se.statistics.nr_wakeups);
1688 
1689         if (wake_flags & WF_SYNC)
1690                 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1691 
1692 #endif /* CONFIG_SCHEDSTATS */
1693 }
1694 
1695 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1696 {
1697         activate_task(rq, p, en_flags);
1698         p->on_rq = TASK_ON_RQ_QUEUED;
1699 
1700         /* if a worker is waking up, notify workqueue */
1701         if (p->flags & PF_WQ_WORKER)
1702                 wq_worker_waking_up(p, cpu_of(rq));
1703 }
1704 
1705 /*
1706  * Mark the task runnable and perform wakeup-preemption.
1707  */
1708 static void
1709 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1710 {
1711         check_preempt_curr(rq, p, wake_flags);
1712         p->state = TASK_RUNNING;
1713         trace_sched_wakeup(p);
1714 
1715 #ifdef CONFIG_SMP
1716         if (p->sched_class->task_woken) {
1717                 /*
1718                  * Our task @p is fully woken up and running; so its safe to
1719                  * drop the rq->lock, hereafter rq is only used for statistics.
1720                  */
1721                 lockdep_unpin_lock(&rq->lock);
1722                 p->sched_class->task_woken(rq, p);
1723                 lockdep_pin_lock(&rq->lock);
1724         }
1725 
1726         if (rq->idle_stamp) {
1727                 u64 delta = rq_clock(rq) - rq->idle_stamp;
1728                 u64 max = 2*rq->max_idle_balance_cost;
1729 
1730                 update_avg(&rq->avg_idle, delta);
1731 
1732                 if (rq->avg_idle > max)
1733                         rq->avg_idle = max;
1734 
1735                 rq->idle_stamp = 0;
1736         }
1737 #endif
1738 }
1739 
1740 static void
1741 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1742 {
1743         lockdep_assert_held(&rq->lock);
1744 
1745 #ifdef CONFIG_SMP
1746         if (p->sched_contributes_to_load)
1747                 rq->nr_uninterruptible--;
1748 #endif
1749 
1750         ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1751         ttwu_do_wakeup(rq, p, wake_flags);
1752 }
1753 
1754 /*
1755  * Called in case the task @p isn't fully descheduled from its runqueue,
1756  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1757  * since all we need to do is flip p->state to TASK_RUNNING, since
1758  * the task is still ->on_rq.
1759  */
1760 static int ttwu_remote(struct task_struct *p, int wake_flags)
1761 {
1762         struct rq *rq;
1763         int ret = 0;
1764 
1765         rq = __task_rq_lock(p);
1766         if (task_on_rq_queued(p)) {
1767                 /* check_preempt_curr() may use rq clock */
1768                 update_rq_clock(rq);
1769                 ttwu_do_wakeup(rq, p, wake_flags);
1770                 ret = 1;
1771         }
1772         __task_rq_unlock(rq);
1773 
1774         return ret;
1775 }
1776 
1777 #ifdef CONFIG_SMP
1778 void sched_ttwu_pending(void)
1779 {
1780         struct rq *rq = this_rq();
1781         struct llist_node *llist = llist_del_all(&rq->wake_list);
1782         struct task_struct *p;
1783         unsigned long flags;
1784 
1785         if (!llist)
1786                 return;
1787 
1788         raw_spin_lock_irqsave(&rq->lock, flags);
1789         lockdep_pin_lock(&rq->lock);
1790 
1791         while (llist) {
1792                 p = llist_entry(llist, struct task_struct, wake_entry);
1793                 llist = llist_next(llist);
1794                 ttwu_do_activate(rq, p, 0);
1795         }
1796 
1797         lockdep_unpin_lock(&rq->lock);
1798         raw_spin_unlock_irqrestore(&rq->lock, flags);
1799 }
1800 
1801 void scheduler_ipi(void)
1802 {
1803         /*
1804          * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1805          * TIF_NEED_RESCHED remotely (for the first time) will also send
1806          * this IPI.
1807          */
1808         preempt_fold_need_resched();
1809 
1810         if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1811                 return;
1812 
1813         /*
1814          * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1815          * traditionally all their work was done from the interrupt return
1816          * path. Now that we actually do some work, we need to make sure
1817          * we do call them.
1818          *
1819          * Some archs already do call them, luckily irq_enter/exit nest
1820          * properly.
1821          *
1822          * Arguably we should visit all archs and update all handlers,
1823          * however a fair share of IPIs are still resched only so this would
1824          * somewhat pessimize the simple resched case.
1825          */
1826         irq_enter();
1827         sched_ttwu_pending();
1828 
1829         /*
1830          * Check if someone kicked us for doing the nohz idle load balance.
1831          */
1832         if (unlikely(got_nohz_idle_kick())) {
1833                 this_rq()->idle_balance = 1;
1834                 raise_softirq_irqoff(SCHED_SOFTIRQ);
1835         }
1836         irq_exit();
1837 }
1838 
1839 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1840 {
1841         struct rq *rq = cpu_rq(cpu);
1842 
1843         if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1844                 if (!set_nr_if_polling(rq->idle))
1845                         smp_send_reschedule(cpu);
1846                 else
1847                         trace_sched_wake_idle_without_ipi(cpu);
1848         }
1849 }
1850 
1851 void wake_up_if_idle(int cpu)
1852 {
1853         struct rq *rq = cpu_rq(cpu);
1854         unsigned long flags;
1855 
1856         rcu_read_lock();
1857 
1858         if (!is_idle_task(rcu_dereference(rq->curr)))
1859                 goto out;
1860 
1861         if (set_nr_if_polling(rq->idle)) {
1862                 trace_sched_wake_idle_without_ipi(cpu);
1863         } else {
1864                 raw_spin_lock_irqsave(&rq->lock, flags);
1865                 if (is_idle_task(rq->curr))
1866                         smp_send_reschedule(cpu);
1867                 /* Else cpu is not in idle, do nothing here */
1868                 raw_spin_unlock_irqrestore(&rq->lock, flags);
1869         }
1870 
1871 out:
1872         rcu_read_unlock();
1873 }
1874 
1875 bool cpus_share_cache(int this_cpu, int that_cpu)
1876 {
1877         return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1878 }
1879 #endif /* CONFIG_SMP */
1880 
1881 static void ttwu_queue(struct task_struct *p, int cpu)
1882 {
1883         struct rq *rq = cpu_rq(cpu);
1884 
1885 #if defined(CONFIG_SMP)
1886         if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1887                 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1888                 ttwu_queue_remote(p, cpu);
1889                 return;
1890         }
1891 #endif
1892 
1893         raw_spin_lock(&rq->lock);
1894         lockdep_pin_lock(&rq->lock);
1895         ttwu_do_activate(rq, p, 0);
1896         lockdep_unpin_lock(&rq->lock);
1897         raw_spin_unlock(&rq->lock);
1898 }
1899 
1900 /**
1901  * try_to_wake_up - wake up a thread
1902  * @p: the thread to be awakened
1903  * @state: the mask of task states that can be woken
1904  * @wake_flags: wake modifier flags (WF_*)
1905  *
1906  * Put it on the run-queue if it's not already there. The "current"
1907  * thread is always on the run-queue (except when the actual
1908  * re-schedule is in progress), and as such you're allowed to do
1909  * the simpler "current->state = TASK_RUNNING" to mark yourself
1910  * runnable without the overhead of this.
1911  *
1912  * Return: %true if @p was woken up, %false if it was already running.
1913  * or @state didn't match @p's state.
1914  */
1915 static int
1916 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1917 {
1918         unsigned long flags;
1919         int cpu, success = 0;
1920 
1921         /*
1922          * If we are going to wake up a thread waiting for CONDITION we
1923          * need to ensure that CONDITION=1 done by the caller can not be
1924          * reordered with p->state check below. This pairs with mb() in
1925          * set_current_state() the waiting thread does.
1926          */
1927         smp_mb__before_spinlock();
1928         raw_spin_lock_irqsave(&p->pi_lock, flags);
1929         if (!(p->state & state))
1930                 goto out;
1931 
1932         trace_sched_waking(p);
1933 
1934         success = 1; /* we're going to change ->state */
1935         cpu = task_cpu(p);
1936 
1937         if (p->on_rq && ttwu_remote(p, wake_flags))
1938                 goto stat;
1939 
1940 #ifdef CONFIG_SMP
1941         /*
1942          * If the owning (remote) cpu is still in the middle of schedule() with
1943          * this task as prev, wait until its done referencing the task.
1944          */
1945         while (p->on_cpu)
1946                 cpu_relax();
1947         /*
1948          * Pairs with the smp_wmb() in finish_lock_switch().
1949          */
1950         smp_rmb();
1951 
1952         p->sched_contributes_to_load = !!task_contributes_to_load(p);
1953         p->state = TASK_WAKING;
1954 
1955         if (p->sched_class->task_waking)
1956                 p->sched_class->task_waking(p);
1957 
1958         cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1959         if (task_cpu(p) != cpu) {
1960                 wake_flags |= WF_MIGRATED;
1961                 set_task_cpu(p, cpu);
1962         }
1963 #endif /* CONFIG_SMP */
1964 
1965         ttwu_queue(p, cpu);
1966 stat:
1967         ttwu_stat(p, cpu, wake_flags);
1968 out:
1969         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1970 
1971         return success;
1972 }
1973 
1974 /**
1975  * try_to_wake_up_local - try to wake up a local task with rq lock held
1976  * @p: the thread to be awakened
1977  *
1978  * Put @p on the run-queue if it's not already there. The caller must
1979  * ensure that this_rq() is locked, @p is bound to this_rq() and not
1980  * the current task.
1981  */
1982 static void try_to_wake_up_local(struct task_struct *p)
1983 {
1984         struct rq *rq = task_rq(p);
1985 
1986         if (WARN_ON_ONCE(rq != this_rq()) ||
1987             WARN_ON_ONCE(p == current))
1988                 return;
1989 
1990         lockdep_assert_held(&rq->lock);
1991 
1992         if (!raw_spin_trylock(&p->pi_lock)) {
1993                 /*
1994                  * This is OK, because current is on_cpu, which avoids it being
1995                  * picked for load-balance and preemption/IRQs are still
1996                  * disabled avoiding further scheduler activity on it and we've
1997                  * not yet picked a replacement task.
1998                  */
1999                 lockdep_unpin_lock(&rq->lock);
2000                 raw_spin_unlock(&rq->lock);
2001                 raw_spin_lock(&p->pi_lock);
2002                 raw_spin_lock(&rq->lock);
2003                 lockdep_pin_lock(&rq->lock);
2004         }
2005 
2006         if (!(p->state & TASK_NORMAL))
2007                 goto out;
2008 
2009         trace_sched_waking(p);
2010 
2011         if (!task_on_rq_queued(p))
2012                 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2013 
2014         ttwu_do_wakeup(rq, p, 0);
2015         ttwu_stat(p, smp_processor_id(), 0);
2016 out:
2017         raw_spin_unlock(&p->pi_lock);
2018 }
2019 
2020 /**
2021  * wake_up_process - Wake up a specific process
2022  * @p: The process to be woken up.
2023  *
2024  * Attempt to wake up the nominated process and move it to the set of runnable
2025  * processes.
2026  *
2027  * Return: 1 if the process was woken up, 0 if it was already running.
2028  *
2029  * It may be assumed that this function implies a write memory barrier before
2030  * changing the task state if and only if any tasks are woken up.
2031  */
2032 int wake_up_process(struct task_struct *p)
2033 {
2034         WARN_ON(task_is_stopped_or_traced(p));
2035         return try_to_wake_up(p, TASK_NORMAL, 0);
2036 }
2037 EXPORT_SYMBOL(wake_up_process);
2038 
2039 int wake_up_state(struct task_struct *p, unsigned int state)
2040 {
2041         return try_to_wake_up(p, state, 0);
2042 }
2043 
2044 /*
2045  * This function clears the sched_dl_entity static params.
2046  */
2047 void __dl_clear_params(struct task_struct *p)
2048 {
2049         struct sched_dl_entity *dl_se = &p->dl;
2050 
2051         dl_se->dl_runtime = 0;
2052         dl_se->dl_deadline = 0;
2053         dl_se->dl_period = 0;
2054         dl_se->flags = 0;
2055         dl_se->dl_bw = 0;
2056 
2057         dl_se->dl_throttled = 0;
2058         dl_se->dl_new = 1;
2059         dl_se->dl_yielded = 0;
2060 }
2061 
2062 /*
2063  * Perform scheduler related setup for a newly forked process p.
2064  * p is forked by current.
2065  *
2066  * __sched_fork() is basic setup used by init_idle() too:
2067  */
2068 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2069 {
2070         p->on_rq                        = 0;
2071 
2072         p->se.on_rq                     = 0;
2073         p->se.exec_start                = 0;
2074         p->se.sum_exec_runtime          = 0;
2075         p->se.prev_sum_exec_runtime     = 0;
2076         p->se.nr_migrations             = 0;
2077         p->se.vruntime                  = 0;
2078         INIT_LIST_HEAD(&p->se.group_node);
2079 
2080 #ifdef CONFIG_SCHEDSTATS
2081         memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2082 #endif
2083 
2084         RB_CLEAR_NODE(&p->dl.rb_node);
2085         init_dl_task_timer(&p->dl);
2086         __dl_clear_params(p);
2087 
2088         INIT_LIST_HEAD(&p->rt.run_list);
2089 
2090 #ifdef CONFIG_PREEMPT_NOTIFIERS
2091         INIT_HLIST_HEAD(&p->preempt_notifiers);
2092 #endif
2093 
2094 #ifdef CONFIG_NUMA_BALANCING
2095         if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2096                 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2097                 p->mm->numa_scan_seq = 0;
2098         }
2099 
2100         if (clone_flags & CLONE_VM)
2101                 p->numa_preferred_nid = current->numa_preferred_nid;
2102         else
2103                 p->numa_preferred_nid = -1;
2104 
2105         p->node_stamp = 0ULL;
2106         p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2107         p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2108         p->numa_work.next = &p->numa_work;
2109         p->numa_faults = NULL;
2110         p->last_task_numa_placement = 0;
2111         p->last_sum_exec_runtime = 0;
2112 
2113         p->numa_group = NULL;
2114 #endif /* CONFIG_NUMA_BALANCING */
2115 }
2116 
2117 #ifdef CONFIG_NUMA_BALANCING
2118 #ifdef CONFIG_SCHED_DEBUG
2119 void set_numabalancing_state(bool enabled)
2120 {
2121         if (enabled)
2122                 sched_feat_set("NUMA");
2123         else
2124                 sched_feat_set("NO_NUMA");
2125 }
2126 #else
2127 __read_mostly bool numabalancing_enabled;
2128 
2129 void set_numabalancing_state(bool enabled)
2130 {
2131         numabalancing_enabled = enabled;
2132 }
2133 #endif /* CONFIG_SCHED_DEBUG */
2134 
2135 #ifdef CONFIG_PROC_SYSCTL
2136 int sysctl_numa_balancing(struct ctl_table *table, int write,
2137                          void __user *buffer, size_t *lenp, loff_t *ppos)
2138 {
2139         struct ctl_table t;
2140         int err;
2141         int state = numabalancing_enabled;
2142 
2143         if (write && !capable(CAP_SYS_ADMIN))
2144                 return -EPERM;
2145 
2146         t = *table;
2147         t.data = &state;
2148         err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2149         if (err < 0)
2150                 return err;
2151         if (write)
2152                 set_numabalancing_state(state);
2153         return err;
2154 }
2155 #endif
2156 #endif
2157 
2158 /*
2159  * fork()/clone()-time setup:
2160  */
2161 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2162 {
2163         unsigned long flags;
2164         int cpu = get_cpu();
2165 
2166         __sched_fork(clone_flags, p);
2167         /*
2168          * We mark the process as running here. This guarantees that
2169          * nobody will actually run it, and a signal or other external
2170          * event cannot wake it up and insert it on the runqueue either.
2171          */
2172         p->state = TASK_RUNNING;
2173 
2174         /*
2175          * Make sure we do not leak PI boosting priority to the child.
2176          */
2177         p->prio = current->normal_prio;
2178 
2179         /*
2180          * Revert to default priority/policy on fork if requested.
2181          */
2182         if (unlikely(p->sched_reset_on_fork)) {
2183                 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2184                         p->policy = SCHED_NORMAL;
2185                         p->static_prio = NICE_TO_PRIO(0);
2186                         p->rt_priority = 0;
2187                 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2188                         p->static_prio = NICE_TO_PRIO(0);
2189 
2190                 p->prio = p->normal_prio = __normal_prio(p);
2191                 set_load_weight(p);
2192 
2193                 /*
2194                  * We don't need the reset flag anymore after the fork. It has
2195                  * fulfilled its duty:
2196                  */
2197                 p->sched_reset_on_fork = 0;
2198         }
2199 
2200         if (dl_prio(p->prio)) {
2201                 put_cpu();
2202                 return -EAGAIN;
2203         } else if (rt_prio(p->prio)) {
2204                 p->sched_class = &rt_sched_class;
2205         } else {
2206                 p->sched_class = &fair_sched_class;
2207         }
2208 
2209         if (p->sched_class->task_fork)
2210                 p->sched_class->task_fork(p);
2211 
2212         /*
2213          * The child is not yet in the pid-hash so no cgroup attach races,
2214          * and the cgroup is pinned to this child due to cgroup_fork()
2215          * is ran before sched_fork().
2216          *
2217          * Silence PROVE_RCU.
2218          */
2219         raw_spin_lock_irqsave(&p->pi_lock, flags);
2220         set_task_cpu(p, cpu);
2221         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2222 
2223 #ifdef CONFIG_SCHED_INFO
2224         if (likely(sched_info_on()))
2225                 memset(&p->sched_info, 0, sizeof(p->sched_info));
2226 #endif
2227 #if defined(CONFIG_SMP)
2228         p->on_cpu = 0;
2229 #endif
2230         init_task_preempt_count(p);
2231 #ifdef CONFIG_SMP
2232         plist_node_init(&p->pushable_tasks, MAX_PRIO);
2233         RB_CLEAR_NODE(&p->pushable_dl_tasks);
2234 #endif
2235 
2236         put_cpu();
2237         return 0;
2238 }
2239 
2240 unsigned long to_ratio(u64 period, u64 runtime)
2241 {
2242         if (runtime == RUNTIME_INF)
2243                 return 1ULL << 20;
2244 
2245         /*
2246          * Doing this here saves a lot of checks in all
2247          * the calling paths, and returning zero seems
2248          * safe for them anyway.
2249          */
2250         if (period == 0)
2251                 return 0;
2252 
2253         return div64_u64(runtime << 20, period);
2254 }
2255 
2256 #ifdef CONFIG_SMP
2257 inline struct dl_bw *dl_bw_of(int i)
2258 {
2259         RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2260                          "sched RCU must be held");
2261         return &cpu_rq(i)->rd->dl_bw;
2262 }
2263 
2264 static inline int dl_bw_cpus(int i)
2265 {
2266         struct root_domain *rd = cpu_rq(i)->rd;
2267         int cpus = 0;
2268 
2269         RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2270                          "sched RCU must be held");
2271         for_each_cpu_and(i, rd->span, cpu_active_mask)
2272                 cpus++;
2273 
2274         return cpus;
2275 }
2276 #else
2277 inline struct dl_bw *dl_bw_of(int i)
2278 {
2279         return &cpu_rq(i)->dl.dl_bw;
2280 }
2281 
2282 static inline int dl_bw_cpus(int i)
2283 {
2284         return 1;
2285 }
2286 #endif
2287 
2288 /*
2289  * We must be sure that accepting a new task (or allowing changing the
2290  * parameters of an existing one) is consistent with the bandwidth
2291  * constraints. If yes, this function also accordingly updates the currently
2292  * allocated bandwidth to reflect the new situation.
2293  *
2294  * This function is called while holding p's rq->lock.
2295  *
2296  * XXX we should delay bw change until the task's 0-lag point, see
2297  * __setparam_dl().
2298  */
2299 static int dl_overflow(struct task_struct *p, int policy,
2300                        const struct sched_attr *attr)
2301 {
2302 
2303         struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2304         u64 period = attr->sched_period ?: attr->sched_deadline;
2305         u64 runtime = attr->sched_runtime;
2306         u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2307         int cpus, err = -1;
2308 
2309         if (new_bw == p->dl.dl_bw)
2310                 return 0;
2311 
2312         /*
2313          * Either if a task, enters, leave, or stays -deadline but changes
2314          * its parameters, we may need to update accordingly the total
2315          * allocated bandwidth of the container.
2316          */
2317         raw_spin_lock(&dl_b->lock);
2318         cpus = dl_bw_cpus(task_cpu(p));
2319         if (dl_policy(policy) && !task_has_dl_policy(p) &&
2320             !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2321                 __dl_add(dl_b, new_bw);
2322                 err = 0;
2323         } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2324                    !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2325                 __dl_clear(dl_b, p->dl.dl_bw);
2326                 __dl_add(dl_b, new_bw);
2327                 err = 0;
2328         } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2329                 __dl_clear(dl_b, p->dl.dl_bw);
2330                 err = 0;
2331         }
2332         raw_spin_unlock(&dl_b->lock);
2333 
2334         return err;
2335 }
2336 
2337 extern void init_dl_bw(struct dl_bw *dl_b);
2338 
2339 /*
2340  * wake_up_new_task - wake up a newly created task for the first time.
2341  *
2342  * This function will do some initial scheduler statistics housekeeping
2343  * that must be done for every newly created context, then puts the task
2344  * on the runqueue and wakes it.
2345  */
2346 void wake_up_new_task(struct task_struct *p)
2347 {
2348         unsigned long flags;
2349         struct rq *rq;
2350 
2351         raw_spin_lock_irqsave(&p->pi_lock, flags);
2352 #ifdef CONFIG_SMP
2353         /*
2354          * Fork balancing, do it here and not earlier because:
2355          *  - cpus_allowed can change in the fork path
2356          *  - any previously selected cpu might disappear through hotplug
2357          */
2358         set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2359 #endif
2360 
2361         /* Initialize new task's runnable average */
2362         init_entity_runnable_average(&p->se);
2363         rq = __task_rq_lock(p);
2364         activate_task(rq, p, 0);
2365         p->on_rq = TASK_ON_RQ_QUEUED;
2366         trace_sched_wakeup_new(p);
2367         check_preempt_curr(rq, p, WF_FORK);
2368 #ifdef CONFIG_SMP
2369         if (p->sched_class->task_woken) {
2370                 /*
2371                  * Nothing relies on rq->lock after this, so its fine to
2372                  * drop it.
2373                  */
2374                 lockdep_unpin_lock(&rq->lock);
2375                 p->sched_class->task_woken(rq, p);
2376                 lockdep_pin_lock(&rq->lock);
2377         }
2378 #endif
2379         task_rq_unlock(rq, p, &flags);
2380 }
2381 
2382 #ifdef CONFIG_PREEMPT_NOTIFIERS
2383 
2384 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2385 
2386 void preempt_notifier_inc(void)
2387 {
2388         static_key_slow_inc(&preempt_notifier_key);
2389 }
2390 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2391 
2392 void preempt_notifier_dec(void)
2393 {
2394         static_key_slow_dec(&preempt_notifier_key);
2395 }
2396 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2397 
2398 /**
2399  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2400  * @notifier: notifier struct to register
2401  */
2402 void preempt_notifier_register(struct preempt_notifier *notifier)
2403 {
2404         if (!static_key_false(&preempt_notifier_key))
2405                 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2406 
2407         hlist_add_head(&notifier->link, &current->preempt_notifiers);
2408 }
2409 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2410 
2411 /**
2412  * preempt_notifier_unregister - no longer interested in preemption notifications
2413  * @notifier: notifier struct to unregister
2414  *
2415  * This is *not* safe to call from within a preemption notifier.
2416  */
2417 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2418 {
2419         hlist_del(&notifier->link);
2420 }
2421 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2422 
2423 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2424 {
2425         struct preempt_notifier *notifier;
2426 
2427         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2428                 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2429 }
2430 
2431 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2432 {
2433         if (static_key_false(&preempt_notifier_key))
2434                 __fire_sched_in_preempt_notifiers(curr);
2435 }
2436 
2437 static void
2438 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2439                                    struct task_struct *next)
2440 {
2441         struct preempt_notifier *notifier;
2442 
2443         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2444                 notifier->ops->sched_out(notifier, next);
2445 }
2446 
2447 static __always_inline void
2448 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2449                                  struct task_struct *next)
2450 {
2451         if (static_key_false(&preempt_notifier_key))
2452                 __fire_sched_out_preempt_notifiers(curr, next);
2453 }
2454 
2455 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2456 
2457 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2458 {
2459 }
2460 
2461 static inline void
2462 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2463                                  struct task_struct *next)
2464 {
2465 }
2466 
2467 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2468 
2469 /**
2470  * prepare_task_switch - prepare to switch tasks
2471  * @rq: the runqueue preparing to switch
2472  * @prev: the current task that is being switched out
2473  * @next: the task we are going to switch to.
2474  *
2475  * This is called with the rq lock held and interrupts off. It must
2476  * be paired with a subsequent finish_task_switch after the context
2477  * switch.
2478  *
2479  * prepare_task_switch sets up locking and calls architecture specific
2480  * hooks.
2481  */
2482 static inline void
2483 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2484                     struct task_struct *next)
2485 {
2486         trace_sched_switch(prev, next);
2487         sched_info_switch(rq, prev, next);
2488         perf_event_task_sched_out(prev, next);
2489         fire_sched_out_preempt_notifiers(prev, next);
2490         prepare_lock_switch(rq, next);
2491         prepare_arch_switch(next);
2492 }
2493 
2494 /**
2495  * finish_task_switch - clean up after a task-switch
2496  * @prev: the thread we just switched away from.
2497  *
2498  * finish_task_switch must be called after the context switch, paired
2499  * with a prepare_task_switch call before the context switch.
2500  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2501  * and do any other architecture-specific cleanup actions.
2502  *
2503  * Note that we may have delayed dropping an mm in context_switch(). If
2504  * so, we finish that here outside of the runqueue lock. (Doing it
2505  * with the lock held can cause deadlocks; see schedule() for
2506  * details.)
2507  *
2508  * The context switch have flipped the stack from under us and restored the
2509  * local variables which were saved when this task called schedule() in the
2510  * past. prev == current is still correct but we need to recalculate this_rq
2511  * because prev may have moved to another CPU.
2512  */
2513 static struct rq *finish_task_switch(struct task_struct *prev)
2514         __releases(rq->lock)
2515 {
2516         struct rq *rq = this_rq();
2517         struct mm_struct *mm = rq->prev_mm;
2518         long prev_state;
2519 
2520         rq->prev_mm = NULL;
2521 
2522         /*
2523          * A task struct has one reference for the use as "current".
2524          * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2525          * schedule one last time. The schedule call will never return, and
2526          * the scheduled task must drop that reference.
2527          *
2528          * We must observe prev->state before clearing prev->on_cpu (in
2529          * finish_lock_switch), otherwise a concurrent wakeup can get prev
2530          * running on another CPU and we could rave with its RUNNING -> DEAD
2531          * transition, resulting in a double drop.
2532          */
2533         prev_state = prev->state;
2534         vtime_task_switch(prev);
2535         perf_event_task_sched_in(prev, current);
2536         finish_lock_switch(rq, prev);
2537         finish_arch_post_lock_switch();
2538 
2539         fire_sched_in_preempt_notifiers(current);
2540         if (mm)
2541                 mmdrop(mm);
2542         if (unlikely(prev_state == TASK_DEAD)) {
2543                 if (prev->sched_class->task_dead)
2544                         prev->sched_class->task_dead(prev);
2545 
2546                 /*
2547                  * Remove function-return probe instances associated with this
2548                  * task and put them back on the free list.
2549                  */
2550                 kprobe_flush_task(prev);
2551                 put_task_struct(prev);
2552         }
2553 
2554         tick_nohz_task_switch();
2555         return rq;
2556 }
2557 
2558 #ifdef CONFIG_SMP
2559 
2560 /* rq->lock is NOT held, but preemption is disabled */
2561 static void __balance_callback(struct rq *rq)
2562 {
2563         struct callback_head *head, *next;
2564         void (*func)(struct rq *rq);
2565         unsigned long flags;
2566 
2567         raw_spin_lock_irqsave(&rq->lock, flags);
2568         head = rq->balance_callback;
2569         rq->balance_callback = NULL;
2570         while (head) {
2571                 func = (void (*)(struct rq *))head->func;
2572                 next = head->next;
2573                 head->next = NULL;
2574                 head = next;
2575 
2576                 func(rq);
2577         }
2578         raw_spin_unlock_irqrestore(&rq->lock, flags);
2579 }
2580 
2581 static inline void balance_callback(struct rq *rq)
2582 {
2583         if (unlikely(rq->balance_callback))
2584                 __balance_callback(rq);
2585 }
2586 
2587 #else
2588 
2589 static inline void balance_callback(struct rq *rq)
2590 {
2591 }
2592 
2593 #endif
2594 
2595 /**
2596  * schedule_tail - first thing a freshly forked thread must call.
2597  * @prev: the thread we just switched away from.
2598  */
2599 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2600         __releases(rq->lock)
2601 {
2602         struct rq *rq;
2603 
2604         /* finish_task_switch() drops rq->lock and enables preemtion */
2605         preempt_disable();
2606         rq = finish_task_switch(prev);
2607         balance_callback(rq);
2608         preempt_enable();
2609 
2610         if (current->set_child_tid)
2611                 put_user(task_pid_vnr(current), current->set_child_tid);
2612 }
2613 
2614 /*
2615  * context_switch - switch to the new MM and the new thread's register state.
2616  */
2617 static inline struct rq *
2618 context_switch(struct rq *rq, struct task_struct *prev,
2619                struct task_struct *next)
2620 {
2621         struct mm_struct *mm, *oldmm;
2622 
2623         prepare_task_switch(rq, prev, next);
2624 
2625         mm = next->mm;
2626         oldmm = prev->active_mm;
2627         /*
2628          * For paravirt, this is coupled with an exit in switch_to to
2629          * combine the page table reload and the switch backend into
2630          * one hypercall.
2631          */
2632         arch_start_context_switch(prev);
2633 
2634         if (!mm) {
2635                 next->active_mm = oldmm;
2636                 atomic_inc(&oldmm->mm_count);
2637                 enter_lazy_tlb(oldmm, next);
2638         } else
2639                 switch_mm(oldmm, mm, next);
2640 
2641         if (!prev->mm) {
2642                 prev->active_mm = NULL;
2643                 rq->prev_mm = oldmm;
2644         }
2645         /*
2646          * Since the runqueue lock will be released by the next
2647          * task (which is an invalid locking op but in the case
2648          * of the scheduler it's an obvious special-case), so we
2649          * do an early lockdep release here:
2650          */
2651         lockdep_unpin_lock(&rq->lock);
2652         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2653 
2654         /* Here we just switch the register state and the stack. */
2655         switch_to(prev, next, prev);
2656         barrier();
2657 
2658         return finish_task_switch(prev);
2659 }
2660 
2661 /*
2662  * nr_running and nr_context_switches:
2663  *
2664  * externally visible scheduler statistics: current number of runnable
2665  * threads, total number of context switches performed since bootup.
2666  */
2667 unsigned long nr_running(void)
2668 {
2669         unsigned long i, sum = 0;
2670 
2671         for_each_online_cpu(i)
2672                 sum += cpu_rq(i)->nr_running;
2673 
2674         return sum;
2675 }
2676 
2677 /*
2678  * Check if only the current task is running on the cpu.
2679  *
2680  * Caution: this function does not check that the caller has disabled
2681  * preemption, thus the result might have a time-of-check-to-time-of-use
2682  * race.  The caller is responsible to use it correctly, for example:
2683  *
2684  * - from a non-preemptable section (of course)
2685  *
2686  * - from a thread that is bound to a single CPU
2687  *
2688  * - in a loop with very short iterations (e.g. a polling loop)
2689  */
2690 bool single_task_running(void)
2691 {
2692         return raw_rq()->nr_running == 1;
2693 }
2694 EXPORT_SYMBOL(single_task_running);
2695 
2696 unsigned long long nr_context_switches(void)
2697 {
2698         int i;
2699         unsigned long long sum = 0;
2700 
2701         for_each_possible_cpu(i)
2702                 sum += cpu_rq(i)->nr_switches;
2703 
2704         return sum;
2705 }
2706 
2707 unsigned long nr_iowait(void)
2708 {
2709         unsigned long i, sum = 0;
2710 
2711         for_each_possible_cpu(i)
2712                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2713 
2714         return sum;
2715 }
2716 
2717 unsigned long nr_iowait_cpu(int cpu)
2718 {
2719         struct rq *this = cpu_rq(cpu);
2720         return atomic_read(&this->nr_iowait);
2721 }
2722 
2723 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2724 {
2725         struct rq *rq = this_rq();
2726         *nr_waiters = atomic_read(&rq->nr_iowait);
2727         *load = rq->load.weight;
2728 }
2729 
2730 #ifdef CONFIG_SMP
2731 
2732 /*
2733  * sched_exec - execve() is a valuable balancing opportunity, because at
2734  * this point the task has the smallest effective memory and cache footprint.
2735  */
2736 void sched_exec(void)
2737 {
2738         struct task_struct *p = current;
2739         unsigned long flags;
2740         int dest_cpu;
2741 
2742         raw_spin_lock_irqsave(&p->pi_lock, flags);
2743         dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2744         if (dest_cpu == smp_processor_id())
2745                 goto unlock;
2746 
2747         if (likely(cpu_active(dest_cpu))) {
2748                 struct migration_arg arg = { p, dest_cpu };
2749 
2750                 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2751                 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2752                 return;
2753         }
2754 unlock:
2755         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2756 }
2757 
2758 #endif
2759 
2760 DEFINE_PER_CPU(struct kernel_stat, kstat);
2761 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2762 
2763 EXPORT_PER_CPU_SYMBOL(kstat);
2764 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2765 
2766 /*
2767  * Return accounted runtime for the task.
2768  * In case the task is currently running, return the runtime plus current's
2769  * pending runtime that have not been accounted yet.
2770  */
2771 unsigned long long task_sched_runtime(struct task_struct *p)
2772 {
2773         unsigned long flags;
2774         struct rq *rq;
2775         u64 ns;
2776 
2777 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2778         /*
2779          * 64-bit doesn't need locks to atomically read a 64bit value.
2780          * So we have a optimization chance when the task's delta_exec is 0.
2781          * Reading ->on_cpu is racy, but this is ok.
2782          *
2783          * If we race with it leaving cpu, we'll take a lock. So we're correct.
2784          * If we race with it entering cpu, unaccounted time is 0. This is
2785          * indistinguishable from the read occurring a few cycles earlier.
2786          * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2787          * been accounted, so we're correct here as well.
2788          */
2789         if (!p->on_cpu || !task_on_rq_queued(p))
2790                 return p->se.sum_exec_runtime;
2791 #endif
2792 
2793         rq = task_rq_lock(p, &flags);
2794         /*
2795          * Must be ->curr _and_ ->on_rq.  If dequeued, we would
2796          * project cycles that may never be accounted to this
2797          * thread, breaking clock_gettime().
2798          */
2799         if (task_current(rq, p) && task_on_rq_queued(p)) {
2800                 update_rq_clock(rq);
2801                 p->sched_class->update_curr(rq);
2802         }
2803         ns = p->se.sum_exec_runtime;
2804         task_rq_unlock(rq, p, &flags);
2805 
2806         return ns;
2807 }
2808 
2809 /*
2810  * This function gets called by the timer code, with HZ frequency.
2811  * We call it with interrupts disabled.
2812  */
2813 void scheduler_tick(void)
2814 {
2815         int cpu = smp_processor_id();
2816         struct rq *rq = cpu_rq(cpu);
2817         struct task_struct *curr = rq->curr;
2818 
2819         sched_clock_tick();
2820 
2821         raw_spin_lock(&rq->lock);
2822         update_rq_clock(rq);
2823         curr->sched_class->task_tick(rq, curr, 0);
2824         update_cpu_load_active(rq);
2825         calc_global_load_tick(rq);
2826         raw_spin_unlock(&rq->lock);
2827 
2828         perf_event_task_tick();
2829 
2830 #ifdef CONFIG_SMP
2831         rq->idle_balance = idle_cpu(cpu);
2832         trigger_load_balance(rq);
2833 #endif
2834         rq_last_tick_reset(rq);
2835 }
2836 
2837 #ifdef CONFIG_NO_HZ_FULL
2838 /**
2839  * scheduler_tick_max_deferment
2840  *
2841  * Keep at least one tick per second when a single
2842  * active task is running because the scheduler doesn't
2843  * yet completely support full dynticks environment.
2844  *
2845  * This makes sure that uptime, CFS vruntime, load
2846  * balancing, etc... continue to move forward, even
2847  * with a very low granularity.
2848  *
2849  * Return: Maximum deferment in nanoseconds.
2850  */
2851 u64 scheduler_tick_max_deferment(void)
2852 {
2853         struct rq *rq = this_rq();
2854         unsigned long next, now = READ_ONCE(jiffies);
2855 
2856         next = rq->last_sched_tick + HZ;
2857 
2858         if (time_before_eq(next, now))
2859                 return 0;
2860 
2861         return jiffies_to_nsecs(next - now);
2862 }
2863 #endif
2864 
2865 notrace unsigned long get_parent_ip(unsigned long addr)
2866 {
2867         if (in_lock_functions(addr)) {
2868                 addr = CALLER_ADDR2;
2869                 if (in_lock_functions(addr))
2870                         addr = CALLER_ADDR3;
2871         }
2872         return addr;
2873 }
2874 
2875 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2876                                 defined(CONFIG_PREEMPT_TRACER))
2877 
2878 void preempt_count_add(int val)
2879 {
2880 #ifdef CONFIG_DEBUG_PREEMPT
2881         /*
2882          * Underflow?
2883          */
2884         if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2885                 return;
2886 #endif
2887         __preempt_count_add(val);
2888 #ifdef CONFIG_DEBUG_PREEMPT
2889         /*
2890          * Spinlock count overflowing soon?
2891          */
2892         DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2893                                 PREEMPT_MASK - 10);
2894 #endif
2895         if (preempt_count() == val) {
2896                 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2897 #ifdef CONFIG_DEBUG_PREEMPT
2898                 current->preempt_disable_ip = ip;
2899 #endif
2900                 trace_preempt_off(CALLER_ADDR0, ip);
2901         }
2902 }
2903 EXPORT_SYMBOL(preempt_count_add);
2904 NOKPROBE_SYMBOL(preempt_count_add);
2905 
2906 void preempt_count_sub(int val)
2907 {
2908 #ifdef CONFIG_DEBUG_PREEMPT
2909         /*
2910          * Underflow?
2911          */
2912         if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2913                 return;
2914         /*
2915          * Is the spinlock portion underflowing?
2916          */
2917         if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2918                         !(preempt_count() & PREEMPT_MASK)))
2919                 return;
2920 #endif
2921 
2922         if (preempt_count() == val)
2923                 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2924         __preempt_count_sub(val);
2925 }
2926 EXPORT_SYMBOL(preempt_count_sub);
2927 NOKPROBE_SYMBOL(preempt_count_sub);
2928 
2929 #endif
2930 
2931 /*
2932  * Print scheduling while atomic bug:
2933  */
2934 static noinline void __schedule_bug(struct task_struct *prev)
2935 {
2936         if (oops_in_progress)
2937                 return;
2938 
2939         printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2940                 prev->comm, prev->pid, preempt_count());
2941 
2942         debug_show_held_locks(prev);
2943         print_modules();
2944         if (irqs_disabled())
2945                 print_irqtrace_events(prev);
2946 #ifdef CONFIG_DEBUG_PREEMPT
2947         if (in_atomic_preempt_off()) {
2948                 pr_err("Preemption disabled at:");
2949                 print_ip_sym(current->preempt_disable_ip);
2950                 pr_cont("\n");
2951         }
2952 #endif
2953         dump_stack();
2954         add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2955 }
2956 
2957 /*
2958  * Various schedule()-time debugging checks and statistics:
2959  */
2960 static inline void schedule_debug(struct task_struct *prev)
2961 {
2962 #ifdef CONFIG_SCHED_STACK_END_CHECK
2963         BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2964 #endif
2965         /*
2966          * Test if we are atomic. Since do_exit() needs to call into
2967          * schedule() atomically, we ignore that path. Otherwise whine
2968          * if we are scheduling when we should not.
2969          */
2970         if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2971                 __schedule_bug(prev);
2972         rcu_sleep_check();
2973 
2974         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2975 
2976         schedstat_inc(this_rq(), sched_count);
2977 }
2978 
2979 /*
2980  * Pick up the highest-prio task:
2981  */
2982 static inline struct task_struct *
2983 pick_next_task(struct rq *rq, struct task_struct *prev)
2984 {
2985         const struct sched_class *class = &fair_sched_class;
2986         struct task_struct *p;
2987 
2988         /*
2989          * Optimization: we know that if all tasks are in
2990          * the fair class we can call that function directly:
2991          */
2992         if (likely(prev->sched_class == class &&
2993                    rq->nr_running == rq->cfs.h_nr_running)) {
2994                 p = fair_sched_class.pick_next_task(rq, prev);
2995                 if (unlikely(p == RETRY_TASK))
2996                         goto again;
2997 
2998                 /* assumes fair_sched_class->next == idle_sched_class */
2999                 if (unlikely(!p))
3000                         p = idle_sched_class.pick_next_task(rq, prev);
3001 
3002                 return p;
3003         }
3004 
3005 again:
3006         for_each_class(class) {
3007                 p = class->pick_next_task(rq, prev);
3008                 if (p) {
3009                         if (unlikely(p == RETRY_TASK))
3010                                 goto again;
3011                         return p;
3012                 }
3013         }
3014 
3015         BUG(); /* the idle class will always have a runnable task */
3016 }
3017 
3018 /*
3019  * __schedule() is the main scheduler function.
3020  *
3021  * The main means of driving the scheduler and thus entering this function are:
3022  *
3023  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3024  *
3025  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3026  *      paths. For example, see arch/x86/entry_64.S.
3027  *
3028  *      To drive preemption between tasks, the scheduler sets the flag in timer
3029  *      interrupt handler scheduler_tick().
3030  *
3031  *   3. Wakeups don't really cause entry into schedule(). They add a
3032  *      task to the run-queue and that's it.
3033  *
3034  *      Now, if the new task added to the run-queue preempts the current
3035  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3036  *      called on the nearest possible occasion:
3037  *
3038  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
3039  *
3040  *         - in syscall or exception context, at the next outmost
3041  *           preempt_enable(). (this might be as soon as the wake_up()'s
3042  *           spin_unlock()!)
3043  *
3044  *         - in IRQ context, return from interrupt-handler to
3045  *           preemptible context
3046  *
3047  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3048  *         then at the next:
3049  *
3050  *          - cond_resched() call
3051  *          - explicit schedule() call
3052  *          - return from syscall or exception to user-space
3053  *          - return from interrupt-handler to user-space
3054  *
3055  * WARNING: must be called with preemption disabled!
3056  */
3057 static void __sched __schedule(void)
3058 {
3059         struct task_struct *prev, *next;
3060         unsigned long *switch_count;
3061         struct rq *rq;
3062         int cpu;
3063 
3064         cpu = smp_processor_id();
3065         rq = cpu_rq(cpu);
3066         rcu_note_context_switch();
3067         prev = rq->curr;
3068 
3069         schedule_debug(prev);
3070 
3071         if (sched_feat(HRTICK))
3072                 hrtick_clear(rq);
3073 
3074         /*
3075          * Make sure that signal_pending_state()->signal_pending() below
3076          * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3077          * done by the caller to avoid the race with signal_wake_up().
3078          */
3079         smp_mb__before_spinlock();
3080         raw_spin_lock_irq(&rq->lock);
3081         lockdep_pin_lock(&rq->lock);
3082 
3083         rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3084 
3085         switch_count = &prev->nivcsw;
3086         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3087                 if (unlikely(signal_pending_state(prev->state, prev))) {
3088                         prev->state = TASK_RUNNING;
3089                 } else {
3090                         deactivate_task(rq, prev, DEQUEUE_SLEEP);
3091                         prev->on_rq = 0;
3092 
3093                         /*
3094                          * If a worker went to sleep, notify and ask workqueue
3095                          * whether it wants to wake up a task to maintain
3096                          * concurrency.
3097                          */
3098                         if (prev->flags & PF_WQ_WORKER) {
3099                                 struct task_struct *to_wakeup;
3100 
3101                                 to_wakeup = wq_worker_sleeping(prev, cpu);
3102                                 if (to_wakeup)
3103                                         try_to_wake_up_local(to_wakeup);
3104                         }
3105                 }
3106                 switch_count = &prev->nvcsw;
3107         }
3108 
3109         if (task_on_rq_queued(prev))
3110                 update_rq_clock(rq);
3111 
3112         next = pick_next_task(rq, prev);
3113         clear_tsk_need_resched(prev);
3114         clear_preempt_need_resched();
3115         rq->clock_skip_update = 0;
3116 
3117         if (likely(prev != next)) {
3118                 rq->nr_switches++;
3119                 rq->curr = next;
3120                 ++*switch_count;
3121 
3122                 rq = context_switch(rq, prev, next); /* unlocks the rq */
3123                 cpu = cpu_of(rq);
3124         } else {
3125                 lockdep_unpin_lock(&rq->lock);
3126                 raw_spin_unlock_irq(&rq->lock);
3127         }
3128 
3129         balance_callback(rq);
3130 }
3131 
3132 static inline void sched_submit_work(struct task_struct *tsk)
3133 {
3134         if (!tsk->state || tsk_is_pi_blocked(tsk))
3135                 return;
3136         /*
3137          * If we are going to sleep and we have plugged IO queued,
3138          * make sure to submit it to avoid deadlocks.
3139          */
3140         if (blk_needs_flush_plug(tsk))
3141                 blk_schedule_flush_plug(tsk);
3142 }
3143 
3144 asmlinkage __visible void __sched schedule(void)
3145 {
3146         struct task_struct *tsk = current;
3147 
3148         sched_submit_work(tsk);
3149         do {
3150                 preempt_disable();
3151                 __schedule();
3152                 sched_preempt_enable_no_resched();
3153         } while (need_resched());
3154 }
3155 EXPORT_SYMBOL(schedule);
3156 
3157 #ifdef CONFIG_CONTEXT_TRACKING
3158 asmlinkage __visible void __sched schedule_user(void)
3159 {
3160         /*
3161          * If we come here after a random call to set_need_resched(),
3162          * or we have been woken up remotely but the IPI has not yet arrived,
3163          * we haven't yet exited the RCU idle mode. Do it here manually until
3164          * we find a better solution.
3165          *
3166          * NB: There are buggy callers of this function.  Ideally we
3167          * should warn if prev_state != CONTEXT_USER, but that will trigger
3168          * too frequently to make sense yet.
3169          */
3170         enum ctx_state prev_state = exception_enter();
3171         schedule();
3172         exception_exit(prev_state);
3173 }
3174 #endif
3175 
3176 /**
3177  * schedule_preempt_disabled - called with preemption disabled
3178  *
3179  * Returns with preemption disabled. Note: preempt_count must be 1
3180  */
3181 void __sched schedule_preempt_disabled(void)
3182 {
3183         sched_preempt_enable_no_resched();
3184         schedule();
3185         preempt_disable();
3186 }
3187 
3188 static void __sched notrace preempt_schedule_common(void)
3189 {
3190         do {
3191                 preempt_active_enter();
3192                 __schedule();
3193                 preempt_active_exit();
3194 
3195                 /*
3196                  * Check again in case we missed a preemption opportunity
3197                  * between schedule and now.
3198                  */
3199         } while (need_resched());
3200 }
3201 
3202 #ifdef CONFIG_PREEMPT
3203 /*
3204  * this is the entry point to schedule() from in-kernel preemption
3205  * off of preempt_enable. Kernel preemptions off return from interrupt
3206  * occur there and call schedule directly.
3207  */
3208 asmlinkage __visible void __sched notrace preempt_schedule(void)
3209 {
3210         /*
3211          * If there is a non-zero preempt_count or interrupts are disabled,
3212          * we do not want to preempt the current task. Just return..
3213          */
3214         if (likely(!preemptible()))
3215                 return;
3216 
3217         preempt_schedule_common();
3218 }
3219 NOKPROBE_SYMBOL(preempt_schedule);
3220 EXPORT_SYMBOL(preempt_schedule);
3221 
3222 /**
3223  * preempt_schedule_notrace - preempt_schedule called by tracing
3224  *
3225  * The tracing infrastructure uses preempt_enable_notrace to prevent
3226  * recursion and tracing preempt enabling caused by the tracing
3227  * infrastructure itself. But as tracing can happen in areas coming
3228  * from userspace or just about to enter userspace, a preempt enable
3229  * can occur before user_exit() is called. This will cause the scheduler
3230  * to be called when the system is still in usermode.
3231  *
3232  * To prevent this, the preempt_enable_notrace will use this function
3233  * instead of preempt_schedule() to exit user context if needed before
3234  * calling the scheduler.
3235  */
3236 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3237 {
3238         enum ctx_state prev_ctx;
3239 
3240         if (likely(!preemptible()))
3241                 return;
3242 
3243         do {
3244                 /*
3245                  * Use raw __prempt_count() ops that don't call function.
3246                  * We can't call functions before disabling preemption which
3247                  * disarm preemption tracing recursions.
3248                  */
3249                 __preempt_count_add(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
3250                 barrier();
3251                 /*
3252                  * Needs preempt disabled in case user_exit() is traced
3253                  * and the tracer calls preempt_enable_notrace() causing
3254                  * an infinite recursion.
3255                  */
3256                 prev_ctx = exception_enter();
3257                 __schedule();
3258                 exception_exit(prev_ctx);
3259 
3260                 barrier();
3261                 __preempt_count_sub(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
3262         } while (need_resched());
3263 }
3264 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3265 
3266 #endif /* CONFIG_PREEMPT */
3267 
3268 /*
3269  * this is the entry point to schedule() from kernel preemption
3270  * off of irq context.
3271  * Note, that this is called and return with irqs disabled. This will
3272  * protect us against recursive calling from irq.
3273  */
3274 asmlinkage __visible void __sched preempt_schedule_irq(void)
3275 {
3276         enum ctx_state prev_state;
3277 
3278         /* Catch callers which need to be fixed */
3279         BUG_ON(preempt_count() || !irqs_disabled());
3280 
3281         prev_state = exception_enter();
3282 
3283         do {
3284                 preempt_active_enter();
3285                 local_irq_enable();
3286                 __schedule();
3287                 local_irq_disable();
3288                 preempt_active_exit();
3289         } while (need_resched());
3290 
3291         exception_exit(prev_state);
3292 }
3293 
3294 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3295                           void *key)
3296 {
3297         return try_to_wake_up(curr->private, mode, wake_flags);
3298 }
3299 EXPORT_SYMBOL(default_wake_function);
3300 
3301 #ifdef CONFIG_RT_MUTEXES
3302 
3303 /*
3304  * rt_mutex_setprio - set the current priority of a task
3305  * @p: task
3306  * @prio: prio value (kernel-internal form)
3307  *
3308  * This function changes the 'effective' priority of a task. It does
3309  * not touch ->normal_prio like __setscheduler().
3310  *
3311  * Used by the rt_mutex code to implement priority inheritance
3312  * logic. Call site only calls if the priority of the task changed.
3313  */
3314 void rt_mutex_setprio(struct task_struct *p, int prio)
3315 {
3316         int oldprio, queued, running, enqueue_flag = 0;
3317         struct rq *rq;
3318         const struct sched_class *prev_class;
3319 
3320         BUG_ON(prio > MAX_PRIO);
3321 
3322         rq = __task_rq_lock(p);
3323 
3324         /*
3325          * Idle task boosting is a nono in general. There is one
3326          * exception, when PREEMPT_RT and NOHZ is active:
3327          *
3328          * The idle task calls get_next_timer_interrupt() and holds
3329          * the timer wheel base->lock on the CPU and another CPU wants
3330          * to access the timer (probably to cancel it). We can safely
3331          * ignore the boosting request, as the idle CPU runs this code
3332          * with interrupts disabled and will complete the lock
3333          * protected section without being interrupted. So there is no
3334          * real need to boost.
3335          */
3336         if (unlikely(p == rq->idle)) {
3337                 WARN_ON(p != rq->curr);
3338                 WARN_ON(p->pi_blocked_on);
3339                 goto out_unlock;
3340         }
3341 
3342         trace_sched_pi_setprio(p, prio);
3343         oldprio = p->prio;
3344         prev_class = p->sched_class;
3345         queued = task_on_rq_queued(p);
3346         running = task_current(rq, p);
3347         if (queued)
3348                 dequeue_task(rq, p, 0);
3349         if (running)
3350                 put_prev_task(rq, p);
3351 
3352         /*
3353          * Boosting condition are:
3354          * 1. -rt task is running and holds mutex A
3355          *      --> -dl task blocks on mutex A
3356          *
3357          * 2. -dl task is running and holds mutex A
3358          *      --> -dl task blocks on mutex A and could preempt the
3359          *          running task
3360          */
3361         if (dl_prio(prio)) {
3362                 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3363                 if (!dl_prio(p->normal_prio) ||
3364                     (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3365                         p->dl.dl_boosted = 1;
3366                         enqueue_flag = ENQUEUE_REPLENISH;
3367                 } else
3368                         p->dl.dl_boosted = 0;
3369                 p->sched_class = &dl_sched_class;
3370         } else if (rt_prio(prio)) {
3371                 if (dl_prio(oldprio))
3372                         p->dl.dl_boosted = 0;
3373                 if (oldprio < prio)
3374                         enqueue_flag = ENQUEUE_HEAD;
3375                 p->sched_class = &rt_sched_class;
3376         } else {
3377                 if (dl_prio(oldprio))
3378                         p->dl.dl_boosted = 0;
3379                 if (rt_prio(oldprio))
3380                         p->rt.timeout = 0;
3381                 p->sched_class = &fair_sched_class;
3382         }
3383 
3384         p->prio = prio;
3385 
3386         if (running)
3387                 p->sched_class->set_curr_task(rq);
3388         if (queued)
3389                 enqueue_task(rq, p, enqueue_flag);
3390 
3391         check_class_changed(rq, p, prev_class, oldprio);
3392 out_unlock:
3393         preempt_disable(); /* avoid rq from going away on us */
3394         __task_rq_unlock(rq);
3395 
3396         balance_callback(rq);
3397         preempt_enable();
3398 }
3399 #endif
3400 
3401 void set_user_nice(struct task_struct *p, long nice)
3402 {
3403         int old_prio, delta, queued;
3404         unsigned long flags;
3405         struct rq *rq;
3406 
3407         if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3408                 return;
3409         /*
3410          * We have to be careful, if called from sys_setpriority(),
3411          * the task might be in the middle of scheduling on another CPU.
3412          */
3413         rq = task_rq_lock(p, &flags);
3414         /*
3415          * The RT priorities are set via sched_setscheduler(), but we still
3416          * allow the 'normal' nice value to be set - but as expected
3417          * it wont have any effect on scheduling until the task is
3418          * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3419          */
3420         if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3421                 p->static_prio = NICE_TO_PRIO(nice);
3422                 goto out_unlock;
3423         }
3424         queued = task_on_rq_queued(p);
3425         if (queued)
3426                 dequeue_task(rq, p, 0);
3427 
3428         p->static_prio = NICE_TO_PRIO(nice);
3429         set_load_weight(p);
3430         old_prio = p->prio;
3431         p->prio = effective_prio(p);
3432         delta = p->prio - old_prio;
3433 
3434         if (queued) {
3435                 enqueue_task(rq, p, 0);
3436                 /*
3437                  * If the task increased its priority or is running and
3438                  * lowered its priority, then reschedule its CPU:
3439                  */
3440                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3441                         resched_curr(rq);
3442         }
3443 out_unlock:
3444         task_rq_unlock(rq, p, &flags);
3445 }
3446 EXPORT_SYMBOL(set_user_nice);
3447 
3448 /*
3449  * can_nice - check if a task can reduce its nice value
3450  * @p: task
3451  * @nice: nice value
3452  */
3453 int can_nice(const struct task_struct *p, const int nice)
3454 {
3455         /* convert nice value [19,-20] to rlimit style value [1,40] */
3456         int nice_rlim = nice_to_rlimit(nice);
3457 
3458         return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3459                 capable(CAP_SYS_NICE));
3460 }
3461 
3462 #ifdef __ARCH_WANT_SYS_NICE
3463 
3464 /*
3465  * sys_nice - change the priority of the current process.
3466  * @increment: priority increment
3467  *
3468  * sys_setpriority is a more generic, but much slower function that
3469  * does similar things.
3470  */
3471 SYSCALL_DEFINE1(nice, int, increment)
3472 {
3473         long nice, retval;
3474 
3475         /*
3476          * Setpriority might change our priority at the same moment.
3477          * We don't have to worry. Conceptually one call occurs first
3478          * and we have a single winner.
3479          */
3480         increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3481         nice = task_nice(current) + increment;
3482 
3483         nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3484         if (increment < 0 && !can_nice(current, nice))
3485                 return -EPERM;
3486 
3487         retval = security_task_setnice(current, nice);
3488         if (retval)
3489                 return retval;
3490 
3491         set_user_nice(current, nice);
3492         return 0;
3493 }
3494 
3495 #endif
3496 
3497 /**
3498  * task_prio - return the priority value of a given task.
3499  * @p: the task in question.
3500  *
3501  * Return: The priority value as seen by users in /proc.
3502  * RT tasks are offset by -200. Normal tasks are centered
3503  * around 0, value goes from -16 to +15.
3504  */
3505 int task_prio(const struct task_struct *p)
3506 {
3507         return p->prio - MAX_RT_PRIO;
3508 }
3509 
3510 /**
3511  * idle_cpu - is a given cpu idle currently?
3512  * @cpu: the processor in question.
3513  *
3514  * Return: 1 if the CPU is currently idle. 0 otherwise.
3515  */
3516 int idle_cpu(int cpu)
3517 {
3518         struct rq *rq = cpu_rq(cpu);
3519 
3520         if (rq->curr != rq->idle)
3521                 return 0;
3522 
3523         if (rq->nr_running)
3524                 return 0;
3525 
3526 #ifdef CONFIG_SMP
3527         if (!llist_empty(&rq->wake_list))
3528                 return 0;
3529 #endif
3530 
3531         return 1;
3532 }
3533 
3534 /**
3535  * idle_task - return the idle task for a given cpu.
3536  * @cpu: the processor in question.
3537  *
3538  * Return: The idle task for the cpu @cpu.
3539  */
3540 struct task_struct *idle_task(int cpu)
3541 {
3542         return cpu_rq(cpu)->idle;
3543 }
3544 
3545 /**
3546  * find_process_by_pid - find a process with a matching PID value.
3547  * @pid: the pid in question.
3548  *
3549  * The task of @pid, if found. %NULL otherwise.
3550  */
3551 static struct task_struct *find_process_by_pid(pid_t pid)
3552 {
3553         return pid ? find_task_by_vpid(pid) : current;
3554 }
3555 
3556 /*
3557  * This function initializes the sched_dl_entity of a newly becoming
3558  * SCHED_DEADLINE task.
3559  *
3560  * Only the static values are considered here, the actual runtime and the
3561  * absolute deadline will be properly calculated when the task is enqueued
3562  * for the first time with its new policy.
3563  */
3564 static void
3565 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3566 {
3567         struct sched_dl_entity *dl_se = &p->dl;
3568 
3569         dl_se->dl_runtime = attr->sched_runtime;
3570         dl_se->dl_deadline = attr->sched_deadline;
3571         dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3572         dl_se->flags = attr->sched_flags;
3573         dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3574 
3575         /*
3576          * Changing the parameters of a task is 'tricky' and we're not doing
3577          * the correct thing -- also see task_dead_dl() and switched_from_dl().
3578          *
3579          * What we SHOULD do is delay the bandwidth release until the 0-lag
3580          * point. This would include retaining the task_struct until that time
3581          * and change dl_overflow() to not immediately decrement the current
3582          * amount.
3583          *
3584          * Instead we retain the current runtime/deadline and let the new
3585          * parameters take effect after the current reservation period lapses.
3586          * This is safe (albeit pessimistic) because the 0-lag point is always
3587          * before the current scheduling deadline.
3588          *
3589          * We can still have temporary overloads because we do not delay the
3590          * change in bandwidth until that time; so admission control is
3591          * not on the safe side. It does however guarantee tasks will never
3592          * consume more than promised.
3593          */
3594 }
3595 
3596 /*
3597  * sched_setparam() passes in -1 for its policy, to let the functions
3598  * it calls know not to change it.
3599  */
3600 #define SETPARAM_POLICY -1
3601 
3602 static void __setscheduler_params(struct task_struct *p,
3603                 const struct sched_attr *attr)
3604 {
3605         int policy = attr->sched_policy;
3606 
3607         if (policy == SETPARAM_POLICY)
3608                 policy = p->policy;
3609 
3610         p->policy = policy;
3611 
3612         if (dl_policy(policy))
3613                 __setparam_dl(p, attr);
3614         else if (fair_policy(policy))
3615                 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3616 
3617         /*
3618          * __sched_setscheduler() ensures attr->sched_priority == 0 when
3619          * !rt_policy. Always setting this ensures that things like
3620          * getparam()/getattr() don't report silly values for !rt tasks.
3621          */
3622         p->rt_priority = attr->sched_priority;
3623         p->normal_prio = normal_prio(p);
3624         set_load_weight(p);
3625 }
3626 
3627 /* Actually do priority change: must hold pi & rq lock. */
3628 static void __setscheduler(struct rq *rq, struct task_struct *p,
3629                            const struct sched_attr *attr, bool keep_boost)
3630 {
3631         __setscheduler_params(p, attr);
3632 
3633         /*
3634          * Keep a potential priority boosting if called from
3635          * sched_setscheduler().
3636          */
3637         if (keep_boost)
3638                 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3639         else
3640                 p->prio = normal_prio(p);
3641 
3642         if (dl_prio(p->prio))
3643                 p->sched_class = &dl_sched_class;
3644         else if (rt_prio(p->prio))
3645                 p->sched_class = &rt_sched_class;
3646         else
3647                 p->sched_class = &fair_sched_class;
3648 }
3649 
3650 static void
3651 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3652 {
3653         struct sched_dl_entity *dl_se = &p->dl;
3654 
3655         attr->sched_priority = p->rt_priority;
3656         attr->sched_runtime = dl_se->dl_runtime;
3657         attr->sched_deadline = dl_se->dl_deadline;
3658         attr->sched_period = dl_se->dl_period;
3659         attr->sched_flags = dl_se->flags;
3660 }
3661 
3662 /*
3663  * This function validates the new parameters of a -deadline task.
3664  * We ask for the deadline not being zero, and greater or equal
3665  * than the runtime, as well as the period of being zero or
3666  * greater than deadline. Furthermore, we have to be sure that
3667  * user parameters are above the internal resolution of 1us (we
3668  * check sched_runtime only since it is always the smaller one) and
3669  * below 2^63 ns (we have to check both sched_deadline and
3670  * sched_period, as the latter can be zero).
3671  */
3672 static bool
3673 __checkparam_dl(const struct sched_attr *attr)
3674 {
3675         /* deadline != 0 */
3676         if (attr->sched_deadline == 0)
3677                 return false;
3678 
3679         /*
3680          * Since we truncate DL_SCALE bits, make sure we're at least
3681          * that big.
3682          */
3683         if (attr->sched_runtime < (1ULL << DL_SCALE))
3684                 return false;
3685 
3686         /*
3687          * Since we use the MSB for wrap-around and sign issues, make
3688          * sure it's not set (mind that period can be equal to zero).
3689          */
3690         if (attr->sched_deadline & (1ULL << 63) ||
3691             attr->sched_period & (1ULL << 63))
3692                 return false;
3693 
3694         /* runtime <= deadline <= period (if period != 0) */
3695         if ((attr->sched_period != 0 &&
3696              attr->sched_period < attr->sched_deadline) ||
3697             attr->sched_deadline < attr->sched_runtime)
3698                 return false;
3699 
3700         return true;
3701 }
3702 
3703 /*
3704  * check the target process has a UID that matches the current process's
3705  */
3706 static bool check_same_owner(struct task_struct *p)
3707 {
3708         const struct cred *cred = current_cred(), *pcred;
3709         bool match;
3710 
3711         rcu_read_lock();
3712         pcred = __task_cred(p);
3713         match = (uid_eq(cred->euid, pcred->euid) ||
3714                  uid_eq(cred->euid, pcred->uid));
3715         rcu_read_unlock();
3716         return match;
3717 }
3718 
3719 static bool dl_param_changed(struct task_struct *p,
3720                 const struct sched_attr *attr)
3721 {
3722         struct sched_dl_entity *dl_se = &p->dl;
3723 
3724         if (dl_se->dl_runtime != attr->sched_runtime ||
3725                 dl_se->dl_deadline != attr->sched_deadline ||
3726                 dl_se->dl_period != attr->sched_period ||
3727                 dl_se->flags != attr->sched_flags)
3728                 return true;
3729 
3730         return false;
3731 }
3732 
3733 static int __sched_setscheduler(struct task_struct *p,
3734                                 const struct sched_attr *attr,
3735                                 bool user, bool pi)
3736 {
3737         int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3738                       MAX_RT_PRIO - 1 - attr->sched_priority;
3739         int retval, oldprio, oldpolicy = -1, queued, running;
3740         int new_effective_prio, policy = attr->sched_policy;
3741         unsigned long flags;
3742         const struct sched_class *prev_class;
3743         struct rq *rq;
3744         int reset_on_fork;
3745 
3746         /* may grab non-irq protected spin_locks */
3747         BUG_ON(in_interrupt());
3748 recheck:
3749         /* double check policy once rq lock held */
3750         if (policy < 0) {
3751                 reset_on_fork = p->sched_reset_on_fork;
3752                 policy = oldpolicy = p->policy;
3753         } else {
3754                 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3755 
3756                 if (policy != SCHED_DEADLINE &&
3757                                 policy != SCHED_FIFO && policy != SCHED_RR &&
3758                                 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3759                                 policy != SCHED_IDLE)
3760                         return -EINVAL;
3761         }
3762 
3763         if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3764                 return -EINVAL;
3765 
3766         /*
3767          * Valid priorities for SCHED_FIFO and SCHED_RR are
3768          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3769          * SCHED_BATCH and SCHED_IDLE is 0.
3770          */
3771         if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3772             (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3773                 return -EINVAL;
3774         if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3775             (rt_policy(policy) != (attr->sched_priority != 0)))
3776                 return -EINVAL;
3777 
3778         /*
3779          * Allow unprivileged RT tasks to decrease priority:
3780          */
3781         if (user && !capable(CAP_SYS_NICE)) {
3782                 if (fair_policy(policy)) {
3783                         if (attr->sched_nice < task_nice(p) &&
3784                             !can_nice(p, attr->sched_nice))
3785                                 return -EPERM;
3786                 }
3787 
3788                 if (rt_policy(policy)) {
3789                         unsigned long rlim_rtprio =
3790                                         task_rlimit(p, RLIMIT_RTPRIO);
3791 
3792                         /* can't set/change the rt policy */
3793                         if (policy != p->policy && !rlim_rtprio)
3794                                 return -EPERM;
3795 
3796                         /* can't increase priority */
3797                         if (attr->sched_priority > p->rt_priority &&
3798                             attr->sched_priority > rlim_rtprio)
3799                                 return -EPERM;
3800                 }
3801 
3802                  /*
3803                   * Can't set/change SCHED_DEADLINE policy at all for now
3804                   * (safest behavior); in the future we would like to allow
3805                   * unprivileged DL tasks to increase their relative deadline
3806                   * or reduce their runtime (both ways reducing utilization)
3807                   */
3808                 if (dl_policy(policy))
3809                         return -EPERM;
3810 
3811                 /*
3812                  * Treat SCHED_IDLE as nice 20. Only allow a switch to
3813                  * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3814                  */
3815                 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3816                         if (!can_nice(p, task_nice(p)))
3817                                 return -EPERM;
3818                 }
3819 
3820                 /* can't change other user's priorities */
3821                 if (!check_same_owner(p))
3822                         return -EPERM;
3823 
3824                 /* Normal users shall not reset the sched_reset_on_fork flag */
3825                 if (p->sched_reset_on_fork && !reset_on_fork)
3826                         return -EPERM;
3827         }
3828 
3829         if (user) {
3830                 retval = security_task_setscheduler(p);
3831                 if (retval)
3832                         return retval;
3833         }
3834 
3835         /*
3836          * make sure no PI-waiters arrive (or leave) while we are
3837          * changing the priority of the task:
3838          *
3839          * To be able to change p->policy safely, the appropriate
3840          * runqueue lock must be held.
3841          */
3842         rq = task_rq_lock(p, &flags);
3843 
3844         /*
3845          * Changing the policy of the stop threads its a very bad idea
3846          */
3847         if (p == rq->stop) {
3848                 task_rq_unlock(rq, p, &flags);
3849                 return -EINVAL;
3850         }
3851 
3852         /*
3853          * If not changing anything there's no need to proceed further,
3854          * but store a possible modification of reset_on_fork.
3855          */
3856         if (unlikely(policy == p->policy)) {
3857                 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3858                         goto change;
3859                 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3860                         goto change;
3861                 if (dl_policy(policy) && dl_param_changed(p, attr))
3862                         goto change;
3863 
3864                 p->sched_reset_on_fork = reset_on_fork;
3865                 task_rq_unlock(rq, p, &flags);
3866                 return 0;
3867         }
3868 change:
3869 
3870         if (user) {
3871 #ifdef CONFIG_RT_GROUP_SCHED
3872                 /*
3873                  * Do not allow realtime tasks into groups that have no runtime
3874                  * assigned.
3875                  */
3876                 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3877                                 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3878                                 !task_group_is_autogroup(task_group(p))) {
3879                         task_rq_unlock(rq, p, &flags);
3880                         return -EPERM;
3881                 }
3882 #endif
3883 #ifdef CONFIG_SMP
3884                 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3885                         cpumask_t *span = rq->rd->span;
3886 
3887                         /*
3888                          * Don't allow tasks with an affinity mask smaller than
3889                          * the entire root_domain to become SCHED_DEADLINE. We
3890                          * will also fail if there's no bandwidth available.
3891                          */
3892                         if (!cpumask_subset(span, &p->cpus_allowed) ||
3893                             rq->rd->dl_bw.bw == 0) {
3894                                 task_rq_unlock(rq, p, &flags);
3895                                 return -EPERM;
3896                         }
3897                 }
3898 #endif
3899         }
3900 
3901         /* recheck policy now with rq lock held */
3902         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3903                 policy = oldpolicy = -1;
3904                 task_rq_unlock(rq, p, &flags);
3905                 goto recheck;
3906         }
3907 
3908         /*
3909          * If setscheduling to SCHED_DEADLINE (or changing the parameters
3910          * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3911          * is available.
3912          */
3913         if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3914                 task_rq_unlock(rq, p, &flags);
3915                 return -EBUSY;
3916         }
3917 
3918         p->sched_reset_on_fork = reset_on_fork;
3919         oldprio = p->prio;
3920 
3921         if (pi) {
3922                 /*
3923                  * Take priority boosted tasks into account. If the new
3924                  * effective priority is unchanged, we just store the new
3925                  * normal parameters and do not touch the scheduler class and
3926                  * the runqueue. This will be done when the task deboost
3927                  * itself.
3928                  */
3929                 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3930                 if (new_effective_prio == oldprio) {
3931                         __setscheduler_params(p, attr);
3932                         task_rq_unlock(rq, p, &flags);
3933                         return 0;
3934                 }
3935         }
3936 
3937         queued = task_on_rq_queued(p);
3938         running = task_current(rq, p);
3939         if (queued)
3940                 dequeue_task(rq, p, 0);
3941         if (running)
3942                 put_prev_task(rq, p);
3943 
3944         prev_class = p->sched_class;
3945         __setscheduler(rq, p, attr, pi);
3946 
3947         if (running)
3948                 p->sched_class->set_curr_task(rq);
3949         if (queued) {
3950                 /*
3951                  * We enqueue to tail when the priority of a task is
3952                  * increased (user space view).
3953                  */
3954                 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3955         }
3956 
3957         check_class_changed(rq, p, prev_class, oldprio);
3958         preempt_disable(); /* avoid rq from going away on us */
3959         task_rq_unlock(rq, p, &flags);
3960 
3961         if (pi)
3962                 rt_mutex_adjust_pi(p);
3963 
3964         /*
3965          * Run balance callbacks after we've adjusted the PI chain.
3966          */
3967         balance_callback(rq);
3968         preempt_enable();
3969 
3970         return 0;
3971 }
3972 
3973 static int _sched_setscheduler(struct task_struct *p, int policy,
3974                                const struct sched_param *param, bool check)
3975 {
3976         struct sched_attr attr = {
3977                 .sched_policy   = policy,
3978                 .sched_priority = param->sched_priority,
3979                 .sched_nice     = PRIO_TO_NICE(p->static_prio),
3980         };
3981 
3982         /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3983         if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3984                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3985                 policy &= ~SCHED_RESET_ON_FORK;
3986                 attr.sched_policy = policy;
3987         }
3988 
3989         return __sched_setscheduler(p, &attr, check, true);
3990 }
3991 /**
3992  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3993  * @p: the task in question.
3994  * @policy: new policy.
3995  * @param: structure containing the new RT priority.
3996  *
3997  * Return: 0 on success. An error code otherwise.
3998  *
3999  * NOTE that the task may be already dead.
4000  */
4001 int sched_setscheduler(struct task_struct *p, int policy,
4002                        const struct sched_param *param)
4003 {
4004         return _sched_setscheduler(p, policy, param, true);
4005 }
4006 EXPORT_SYMBOL_GPL(sched_setscheduler);
4007 
4008 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4009 {
4010         return __sched_setscheduler(p, attr, true, true);
4011 }
4012 EXPORT_SYMBOL_GPL(sched_setattr);
4013 
4014 /**
4015  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4016  * @p: the task in question.
4017  * @policy: new policy.
4018  * @param: structure containing the new RT priority.
4019  *
4020  * Just like sched_setscheduler, only don't bother checking if the
4021  * current context has permission.  For example, this is needed in
4022  * stop_machine(): we create temporary high priority worker threads,
4023  * but our caller might not have that capability.
4024  *
4025  * Return: 0 on success. An error code otherwise.
4026  */
4027 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4028                                const struct sched_param *param)
4029 {
4030         return _sched_setscheduler(p, policy, param, false);
4031 }
4032 
4033 static int
4034 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4035 {
4036         struct sched_param lparam;
4037         struct task_struct *p;
4038         int retval;
4039 
4040         if (!param || pid < 0)
4041                 return -EINVAL;
4042         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4043                 return -EFAULT;
4044 
4045         rcu_read_lock();
4046         retval = -ESRCH;
4047         p = find_process_by_pid(pid);
4048         if (p != NULL)
4049                 retval = sched_setscheduler(p, policy, &lparam);
4050         rcu_read_unlock();
4051 
4052         return retval;
4053 }
4054 
4055 /*
4056  * Mimics kernel/events/core.c perf_copy_attr().
4057  */
4058 static int sched_copy_attr(struct sched_attr __user *uattr,
4059                            struct sched_attr *attr)
4060 {
4061         u32 size;
4062         int ret;
4063 
4064         if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4065                 return -EFAULT;
4066 
4067         /*
4068          * zero the full structure, so that a short copy will be nice.
4069          */
4070         memset(attr, 0, sizeof(*attr));
4071 
4072         ret = get_user(size, &uattr->size);
4073         if (ret)
4074                 return ret;
4075 
4076         if (size > PAGE_SIZE)   /* silly large */
4077                 goto err_size;
4078 
4079         if (!size)              /* abi compat */
4080                 size = SCHED_ATTR_SIZE_VER0;
4081 
4082         if (size < SCHED_ATTR_SIZE_VER0)
4083                 goto err_size;
4084 
4085         /*
4086          * If we're handed a bigger struct than we know of,
4087          * ensure all the unknown bits are 0 - i.e. new
4088          * user-space does not rely on any kernel feature
4089          * extensions we dont know about yet.
4090          */
4091         if (size > sizeof(*attr)) {
4092                 unsigned char __user *addr;
4093                 unsigned char __user *end;
4094                 unsigned char val;
4095 
4096                 addr = (void __user *)uattr + sizeof(*attr);
4097                 end  = (void __user *)uattr + size;
4098 
4099                 for (; addr < end; addr++) {
4100                         ret = get_user(val, addr);
4101                         if (ret)
4102                                 return ret;
4103                         if (val)
4104                                 goto err_size;
4105                 }
4106                 size = sizeof(*attr);
4107         }
4108 
4109         ret = copy_from_user(attr, uattr, size);
4110         if (ret)
4111                 return -EFAULT;
4112 
4113         /*
4114          * XXX: do we want to be lenient like existing syscalls; or do we want
4115          * to be strict and return an error on out-of-bounds values?
4116          */
4117         attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4118 
4119         return 0;
4120 
4121 err_size:
4122         put_user(sizeof(*attr), &uattr->size);
4123         return -E2BIG;
4124 }
4125 
4126 /**
4127  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4128  * @pid: the pid in question.
4129  * @policy: new policy.
4130  * @param: structure containing the new RT priority.
4131  *
4132  * Return: 0 on success. An error code otherwise.
4133  */
4134 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4135                 struct sched_param __user *, param)
4136 {
4137         /* negative values for policy are not valid */
4138         if (policy < 0)
4139                 return -EINVAL;
4140 
4141         return do_sched_setscheduler(pid, policy, param);
4142 }
4143 
4144 /**
4145  * sys_sched_setparam - set/change the RT priority of a thread
4146  * @pid: the pid in question.
4147  * @param: structure containing the new RT priority.
4148  *
4149  * Return: 0 on success. An error code otherwise.
4150  */
4151 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4152 {
4153         return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4154 }
4155 
4156 /**
4157  * sys_sched_setattr - same as above, but with extended sched_attr
4158  * @pid: the pid in question.
4159  * @uattr: structure containing the extended parameters.
4160  * @flags: for future extension.
4161  */
4162 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4163                                unsigned int, flags)
4164 {
4165         struct sched_attr attr;
4166         struct task_struct *p;
4167         int retval;
4168 
4169         if (!uattr || pid < 0 || flags)
4170                 return -EINVAL;
4171 
4172         retval = sched_copy_attr(uattr, &attr);
4173         if (retval)
4174                 return retval;
4175 
4176         if ((int)attr.sched_policy < 0)
4177                 return -EINVAL;
4178 
4179         rcu_read_lock();
4180         retval = -ESRCH;
4181         p = find_process_by_pid(pid);
4182         if (p != NULL)
4183                 retval = sched_setattr(p, &attr);
4184         rcu_read_unlock();
4185 
4186         return retval;
4187 }
4188 
4189 /**
4190  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4191  * @pid: the pid in question.
4192  *
4193  * Return: On success, the policy of the thread. Otherwise, a negative error
4194  * code.
4195  */
4196 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4197 {
4198         struct task_struct *p;
4199         int retval;
4200 
4201         if (pid < 0)
4202                 return -EINVAL;
4203 
4204         retval = -ESRCH;
4205         rcu_read_lock();
4206         p = find_process_by_pid(pid);
4207         if (p) {
4208                 retval = security_task_getscheduler(p);
4209                 if (!retval)
4210                         retval = p->policy
4211                                 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4212         }
4213         rcu_read_unlock();
4214         return retval;
4215 }
4216 
4217 /**
4218  * sys_sched_getparam - get the RT priority of a thread
4219  * @pid: the pid in question.
4220  * @param: structure containing the RT priority.
4221  *
4222  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4223  * code.
4224  */
4225 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4226 {
4227         struct sched_param lp = { .sched_priority = 0 };
4228         struct task_struct *p;
4229         int retval;
4230 
4231         if (!param || pid < 0)
4232                 return -EINVAL;
4233 
4234         rcu_read_lock();
4235         p = find_process_by_pid(pid);
4236         retval = -ESRCH;
4237         if (!p)
4238                 goto out_unlock;
4239 
4240         retval = security_task_getscheduler(p);
4241         if (retval)
4242                 goto out_unlock;
4243 
4244         if (task_has_rt_policy(p))
4245                 lp.sched_priority = p->rt_priority;
4246         rcu_read_unlock();
4247 
4248         /*
4249          * This one might sleep, we cannot do it with a spinlock held ...
4250          */
4251         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4252 
4253         return retval;
4254 
4255 out_unlock:
4256         rcu_read_unlock();
4257         return retval;
4258 }
4259 
4260 static int sched_read_attr(struct sched_attr __user *uattr,
4261                            struct sched_attr *attr,
4262                            unsigned int usize)
4263 {
4264         int ret;
4265 
4266         if (!access_ok(VERIFY_WRITE, uattr, usize))
4267                 return -EFAULT;
4268 
4269         /*
4270          * If we're handed a smaller struct than we know of,
4271          * ensure all the unknown bits are 0 - i.e. old
4272          * user-space does not get uncomplete information.
4273          */
4274         if (usize < sizeof(*attr)) {
4275                 unsigned char *addr;
4276                 unsigned char *end;
4277 
4278                 addr = (void *)attr + usize;
4279                 end  = (void *)attr + sizeof(*attr);
4280 
4281                 for (; addr < end; addr++) {
4282                         if (*addr)
4283                                 return -EFBIG;
4284                 }
4285 
4286                 attr->size = usize;
4287         }
4288 
4289         ret = copy_to_user(uattr, attr, attr->size);
4290         if (ret)
4291                 return -EFAULT;
4292 
4293         return 0;
4294 }
4295 
4296 /**
4297  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4298  * @pid: the pid in question.
4299  * @uattr: structure containing the extended parameters.
4300  * @size: sizeof(attr) for fwd/bwd comp.
4301  * @flags: for future extension.
4302  */
4303 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4304                 unsigned int, size, unsigned int, flags)
4305 {
4306         struct sched_attr attr = {
4307                 .size = sizeof(struct sched_attr),
4308         };
4309         struct task_struct *p;
4310         int retval;
4311 
4312         if (!uattr || pid < 0 || size > PAGE_SIZE ||
4313             size < SCHED_ATTR_SIZE_VER0 || flags)
4314                 return -EINVAL;
4315 
4316         rcu_read_lock();
4317         p = find_process_by_pid(pid);
4318         retval = -ESRCH;
4319         if (!p)
4320                 goto out_unlock;
4321 
4322         retval = security_task_getscheduler(p);
4323         if (retval)
4324                 goto out_unlock;
4325 
4326         attr.sched_policy = p->policy;
4327         if (p->sched_reset_on_fork)
4328                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4329         if (task_has_dl_policy(p))
4330                 __getparam_dl(p, &attr);
4331         else if (task_has_rt_policy(p))
4332                 attr.sched_priority = p->rt_priority;
4333         else
4334                 attr.sched_nice = task_nice(p);
4335 
4336         rcu_read_unlock();
4337 
4338         retval = sched_read_attr(uattr, &attr, size);
4339         return retval;
4340 
4341 out_unlock:
4342         rcu_read_unlock();
4343         return retval;
4344 }
4345 
4346 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4347 {
4348         cpumask_var_t cpus_allowed, new_mask;
4349         struct task_struct *p;
4350         int retval;
4351 
4352         rcu_read_lock();
4353 
4354         p = find_process_by_pid(pid);
4355         if (!p) {
4356                 rcu_read_unlock();
4357                 return -ESRCH;
4358         }
4359 
4360         /* Prevent p going away */
4361         get_task_struct(p);
4362         rcu_read_unlock();
4363 
4364         if (p->flags & PF_NO_SETAFFINITY) {
4365                 retval = -EINVAL;
4366                 goto out_put_task;
4367         }
4368         if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4369                 retval = -ENOMEM;
4370                 goto out_put_task;
4371         }
4372         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4373                 retval = -ENOMEM;
4374                 goto out_free_cpus_allowed;
4375         }
4376         retval = -EPERM;
4377         if (!check_same_owner(p)) {
4378                 rcu_read_lock();
4379                 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4380                         rcu_read_unlock();
4381                         goto out_free_new_mask;
4382                 }
4383                 rcu_read_unlock();
4384         }
4385 
4386         retval = security_task_setscheduler(p);
4387         if (retval)
4388                 goto out_free_new_mask;
4389 
4390 
4391         cpuset_cpus_allowed(p, cpus_allowed);
4392         cpumask_and(new_mask, in_mask, cpus_allowed);
4393 
4394         /*
4395          * Since bandwidth control happens on root_domain basis,
4396          * if admission test is enabled, we only admit -deadline
4397          * tasks allowed to run on all the CPUs in the task's
4398          * root_domain.
4399          */
4400 #ifdef CONFIG_SMP
4401         if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4402                 rcu_read_lock();
4403                 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4404                         retval = -EBUSY;
4405                         rcu_read_unlock();
4406                         goto out_free_new_mask;
4407                 }
4408                 rcu_read_unlock();
4409         }
4410 #endif
4411 again:
4412         retval = __set_cpus_allowed_ptr(p, new_mask, true);
4413 
4414         if (!retval) {
4415                 cpuset_cpus_allowed(p, cpus_allowed);
4416                 if (!cpumask_subset(new_mask, cpus_allowed)) {
4417                         /*
4418                          * We must have raced with a concurrent cpuset
4419                          * update. Just reset the cpus_allowed to the
4420                          * cpuset's cpus_allowed
4421                          */
4422                         cpumask_copy(new_mask, cpus_allowed);
4423                         goto again;
4424                 }
4425         }
4426 out_free_new_mask:
4427         free_cpumask_var(new_mask);
4428 out_free_cpus_allowed:
4429         free_cpumask_var(cpus_allowed);
4430 out_put_task:
4431         put_task_struct(p);
4432         return retval;
4433 }
4434 
4435 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4436                              struct cpumask *new_mask)
4437 {
4438         if (len < cpumask_size())
4439                 cpumask_clear(new_mask);
4440         else if (len > cpumask_size())
4441                 len = cpumask_size();
4442 
4443         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4444 }
4445 
4446 /**
4447  * sys_sched_setaffinity - set the cpu affinity of a process
4448  * @pid: pid of the process
4449  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4450  * @user_mask_ptr: user-space pointer to the new cpu mask
4451  *
4452  * Return: 0 on success. An error code otherwise.
4453  */
4454 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4455                 unsigned long __user *, user_mask_ptr)
4456 {
4457         cpumask_var_t new_mask;
4458         int retval;
4459 
4460         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4461                 return -ENOMEM;
4462 
4463         retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4464         if (retval == 0)
4465                 retval = sched_setaffinity(pid, new_mask);
4466         free_cpumask_var(new_mask);
4467         return retval;
4468 }
4469 
4470 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4471 {
4472         struct task_struct *p;
4473         unsigned long flags;
4474         int retval;
4475 
4476         rcu_read_lock();
4477 
4478         retval = -ESRCH;
4479         p = find_process_by_pid(pid);
4480         if (!p)
4481                 goto out_unlock;
4482 
4483         retval = security_task_getscheduler(p);
4484         if (retval)
4485                 goto out_unlock;
4486 
4487         raw_spin_lock_irqsave(&p->pi_lock, flags);
4488         cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4489         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4490 
4491 out_unlock:
4492         rcu_read_unlock();
4493 
4494         return retval;
4495 }
4496 
4497 /**
4498  * sys_sched_getaffinity - get the cpu affinity of a process
4499  * @pid: pid of the process
4500  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4501  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4502  *
4503  * Return: 0 on success. An error code otherwise.
4504  */
4505 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4506                 unsigned long __user *, user_mask_ptr)
4507 {
4508         int ret;
4509         cpumask_var_t mask;
4510 
4511         if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4512                 return -EINVAL;
4513         if (len & (sizeof(unsigned long)-1))
4514                 return -EINVAL;
4515 
4516         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4517                 return -ENOMEM;
4518 
4519         ret = sched_getaffinity(pid, mask);
4520         if (ret == 0) {
4521                 size_t retlen = min_t(size_t, len, cpumask_size());
4522 
4523                 if (copy_to_user(user_mask_ptr, mask, retlen))
4524                         ret = -EFAULT;
4525                 else
4526                         ret = retlen;
4527         }
4528         free_cpumask_var(mask);
4529 
4530         return ret;
4531 }
4532 
4533 /**
4534  * sys_sched_yield - yield the current processor to other threads.
4535  *
4536  * This function yields the current CPU to other tasks. If there are no
4537  * other threads running on this CPU then this function will return.
4538  *
4539  * Return: 0.
4540  */
4541 SYSCALL_DEFINE0(sched_yield)
4542 {
4543         struct rq *rq = this_rq_lock();
4544 
4545         schedstat_inc(rq, yld_count);
4546         current->sched_class->yield_task(rq);
4547 
4548         /*
4549          * Since we are going to call schedule() anyway, there's
4550          * no need to preempt or enable interrupts:
4551          */
4552         __release(rq->lock);
4553         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4554         do_raw_spin_unlock(&rq->lock);
4555         sched_preempt_enable_no_resched();
4556 
4557         schedule();
4558 
4559         return 0;
4560 }
4561 
4562 int __sched _cond_resched(void)
4563 {
4564         if (should_resched(0)) {
4565                 preempt_schedule_common();
4566                 return 1;
4567         }
4568         return 0;
4569 }
4570 EXPORT_SYMBOL(_cond_resched);
4571 
4572 /*
4573  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4574  * call schedule, and on return reacquire the lock.
4575  *
4576  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4577  * operations here to prevent schedule() from being called twice (once via
4578  * spin_unlock(), once by hand).
4579  */
4580 int __cond_resched_lock(spinlock_t *lock)
4581 {
4582         int resched = should_resched(PREEMPT_LOCK_OFFSET);
4583         int ret = 0;
4584 
4585         lockdep_assert_held(lock);
4586 
4587         if (spin_needbreak(lock) || resched) {
4588                 spin_unlock(lock);
4589                 if (resched)
4590                         preempt_schedule_common();
4591                 else
4592                         cpu_relax();
4593                 ret = 1;
4594                 spin_lock(lock);
4595         }
4596         return ret;
4597 }
4598 EXPORT_SYMBOL(__cond_resched_lock);
4599 
4600 int __sched __cond_resched_softirq(void)
4601 {
4602         BUG_ON(!in_softirq());
4603 
4604         if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4605                 local_bh_enable();
4606                 preempt_schedule_common();
4607                 local_bh_disable();
4608                 return 1;
4609         }
4610         return 0;
4611 }
4612 EXPORT_SYMBOL(__cond_resched_softirq);
4613 
4614 /**
4615  * yield - yield the current processor to other threads.
4616  *
4617  * Do not ever use this function, there's a 99% chance you're doing it wrong.
4618  *
4619  * The scheduler is at all times free to pick the calling task as the most
4620  * eligible task to run, if removing the yield() call from your code breaks
4621  * it, its already broken.
4622  *
4623  * Typical broken usage is:
4624  *
4625  * while (!event)
4626  *      yield();
4627  *
4628  * where one assumes that yield() will let 'the other' process run that will
4629  * make event true. If the current task is a SCHED_FIFO task that will never
4630  * happen. Never use yield() as a progress guarantee!!
4631  *
4632  * If you want to use yield() to wait for something, use wait_event().
4633  * If you want to use yield() to be 'nice' for others, use cond_resched().
4634  * If you still want to use yield(), do not!
4635  */
4636 void __sched yield(void)
4637 {
4638         set_current_state(TASK_RUNNING);
4639         sys_sched_yield();
4640 }
4641 EXPORT_SYMBOL(yield);
4642 
4643 /**
4644  * yield_to - yield the current processor to another thread in
4645  * your thread group, or accelerate that thread toward the
4646  * processor it's on.
4647  * @p: target task
4648  * @preempt: whether task preemption is allowed or not
4649  *
4650  * It's the caller's job to ensure that the target task struct
4651  * can't go away on us before we can do any checks.
4652  *
4653  * Return:
4654  *      true (>0) if we indeed boosted the target task.
4655  *      false (0) if we failed to boost the target.
4656  *      -ESRCH if there's no task to yield to.
4657  */
4658 int __sched yield_to(struct task_struct *p, bool preempt)
4659 {
4660         struct task_struct *curr = current;
4661         struct rq *rq, *p_rq;
4662         unsigned long flags;
4663         int yielded = 0;
4664 
4665         local_irq_save(flags);
4666         rq = this_rq();
4667 
4668 again:
4669         p_rq = task_rq(p);
4670         /*
4671          * If we're the only runnable task on the rq and target rq also
4672          * has only one task, there's absolutely no point in yielding.
4673          */
4674         if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4675                 yielded = -ESRCH;
4676                 goto out_irq;
4677         }
4678 
4679         double_rq_lock(rq, p_rq);
4680         if (task_rq(p) != p_rq) {
4681                 double_rq_unlock(rq, p_rq);
4682                 goto again;
4683         }
4684 
4685         if (!curr->sched_class->yield_to_task)
4686                 goto out_unlock;
4687 
4688         if (curr->sched_class != p->sched_class)
4689                 goto out_unlock;
4690 
4691         if (task_running(p_rq, p) || p->state)
4692                 goto out_unlock;
4693 
4694         yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4695         if (yielded) {
4696                 schedstat_inc(rq, yld_count);
4697                 /*
4698                  * Make p's CPU reschedule; pick_next_entity takes care of
4699                  * fairness.
4700                  */
4701                 if (preempt && rq != p_rq)
4702                         resched_curr(p_rq);
4703         }
4704 
4705 out_unlock:
4706         double_rq_unlock(rq, p_rq);
4707 out_irq:
4708         local_irq_restore(flags);
4709 
4710         if (yielded > 0)
4711                 schedule();
4712 
4713         return yielded;
4714 }
4715 EXPORT_SYMBOL_GPL(yield_to);
4716 
4717 /*
4718  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4719  * that process accounting knows that this is a task in IO wait state.
4720  */
4721 long __sched io_schedule_timeout(long timeout)
4722 {
4723         int old_iowait = current->in_iowait;
4724         struct rq *rq;
4725         long ret;
4726 
4727         current->in_iowait = 1;
4728         blk_schedule_flush_plug(current);
4729 
4730         delayacct_blkio_start();
4731         rq = raw_rq();
4732         atomic_inc(&rq->nr_iowait);
4733         ret = schedule_timeout(timeout);
4734         current->in_iowait = old_iowait;
4735         atomic_dec(&rq->nr_iowait);
4736         delayacct_blkio_end();
4737 
4738         return ret;
4739 }
4740 EXPORT_SYMBOL(io_schedule_timeout);
4741 
4742 /**
4743  * sys_sched_get_priority_max - return maximum RT priority.
4744  * @policy: scheduling class.
4745  *
4746  * Return: On success, this syscall returns the maximum
4747  * rt_priority that can be used by a given scheduling class.
4748  * On failure, a negative error code is returned.
4749  */
4750 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4751 {
4752         int ret = -EINVAL;
4753 
4754         switch (policy) {
4755         case SCHED_FIFO:
4756         case SCHED_RR:
4757                 ret = MAX_USER_RT_PRIO-1;
4758                 break;
4759         case SCHED_DEADLINE:
4760         case SCHED_NORMAL:
4761         case SCHED_BATCH:
4762         case SCHED_IDLE:
4763                 ret = 0;
4764                 break;
4765         }
4766         return ret;
4767 }
4768 
4769 /**
4770  * sys_sched_get_priority_min - return minimum RT priority.
4771  * @policy: scheduling class.
4772  *
4773  * Return: On success, this syscall returns the minimum
4774  * rt_priority that can be used by a given scheduling class.
4775  * On failure, a negative error code is returned.
4776  */
4777 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4778 {
4779         int ret = -EINVAL;
4780 
4781         switch (policy) {
4782         case SCHED_FIFO:
4783         case SCHED_RR:
4784                 ret = 1;
4785                 break;
4786         case SCHED_DEADLINE:
4787         case SCHED_NORMAL:
4788         case SCHED_BATCH:
4789         case SCHED_IDLE:
4790                 ret = 0;
4791         }
4792         return ret;
4793 }
4794 
4795 /**
4796  * sys_sched_rr_get_interval - return the default timeslice of a process.
4797  * @pid: pid of the process.
4798  * @interval: userspace pointer to the timeslice value.
4799  *
4800  * this syscall writes the default timeslice value of a given process
4801  * into the user-space timespec buffer. A value of '' means infinity.
4802  *
4803  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4804  * an error code.
4805  */
4806 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4807                 struct timespec __user *, interval)
4808 {
4809         struct task_struct *p;
4810         unsigned int time_slice;
4811         unsigned long flags;
4812         struct rq *rq;
4813         int retval;
4814         struct timespec t;
4815 
4816         if (pid < 0)
4817                 return -EINVAL;
4818 
4819         retval = -ESRCH;
4820         rcu_read_lock();
4821         p = find_process_by_pid(pid);
4822         if (!p)
4823                 goto out_unlock;
4824 
4825         retval = security_task_getscheduler(p);
4826         if (retval)
4827                 goto out_unlock;
4828 
4829         rq = task_rq_lock(p, &flags);
4830         time_slice = 0;
4831         if (p->sched_class->get_rr_interval)
4832                 time_slice = p->sched_class->get_rr_interval(rq, p);
4833         task_rq_unlock(rq, p, &flags);
4834 
4835         rcu_read_unlock();
4836         jiffies_to_timespec(time_slice, &t);
4837         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4838         return retval;
4839 
4840 out_unlock:
4841         rcu_read_unlock();
4842         return retval;
4843 }
4844 
4845 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4846 
4847 void sched_show_task(struct task_struct *p)
4848 {
4849         unsigned long free = 0;
4850         int ppid;
4851         unsigned long state = p->state;
4852 
4853         if (state)
4854                 state = __ffs(state) + 1;
4855         printk(KERN_INFO "%-15.15s %c", p->comm,
4856                 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4857 #if BITS_PER_LONG == 32
4858         if (state == TASK_RUNNING)
4859                 printk(KERN_CONT " running  ");
4860         else
4861                 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4862 #else
4863         if (state == TASK_RUNNING)
4864                 printk(KERN_CONT "  running task    ");
4865         else
4866                 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4867 #endif
4868 #ifdef CONFIG_DEBUG_STACK_USAGE
4869         free = stack_not_used(p);
4870 #endif
4871         ppid = 0;
4872         rcu_read_lock();
4873         if (pid_alive(p))
4874                 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4875         rcu_read_unlock();
4876         printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4877                 task_pid_nr(p), ppid,
4878                 (unsigned long)task_thread_info(p)->flags);
4879 
4880         print_worker_info(KERN_INFO, p);
4881         show_stack(p, NULL);
4882 }
4883 
4884 void show_state_filter(unsigned long state_filter)
4885 {
4886         struct task_struct *g, *p;
4887 
4888 #if BITS_PER_LONG == 32
4889         printk(KERN_INFO
4890                 "  task                PC stack   pid father\n");
4891 #else
4892         printk(KERN_INFO
4893                 "  task                        PC stack   pid father\n");
4894 #endif
4895         rcu_read_lock();
4896         for_each_process_thread(g, p) {
4897                 /*
4898                  * reset the NMI-timeout, listing all files on a slow
4899                  * console might take a lot of time:
4900                  */
4901                 touch_nmi_watchdog();
4902                 if (!state_filter || (p->state & state_filter))
4903                         sched_show_task(p);
4904         }
4905 
4906         touch_all_softlockup_watchdogs();
4907 
4908 #ifdef CONFIG_SCHED_DEBUG
4909         sysrq_sched_debug_show();
4910 #endif
4911         rcu_read_unlock();
4912         /*
4913          * Only show locks if all tasks are dumped:
4914          */
4915         if (!state_filter)
4916                 debug_show_all_locks();
4917 }
4918 
4919 void init_idle_bootup_task(struct task_struct *idle)
4920 {
4921         idle->sched_class = &idle_sched_class;
4922 }
4923 
4924 /**
4925  * init_idle - set up an idle thread for a given CPU
4926  * @idle: task in question
4927  * @cpu: cpu the idle task belongs to
4928  *
4929  * NOTE: this function does not set the idle thread's NEED_RESCHED
4930  * flag, to make booting more robust.
4931  */
4932 void init_idle(struct task_struct *idle, int cpu)
4933 {
4934         struct rq *rq = cpu_rq(cpu);
4935         unsigned long flags;
4936 
4937         raw_spin_lock_irqsave(&idle->pi_lock, flags);
4938         raw_spin_lock(&rq->lock);
4939 
4940         __sched_fork(0, idle);
4941         idle->state = TASK_RUNNING;
4942         idle->se.exec_start = sched_clock();
4943 
4944 #ifdef CONFIG_SMP
4945         /*
4946          * Its possible that init_idle() gets called multiple times on a task,
4947          * in that case do_set_cpus_allowed() will not do the right thing.
4948          *
4949          * And since this is boot we can forgo the serialization.
4950          */
4951         set_cpus_allowed_common(idle, cpumask_of(cpu));
4952 #endif
4953         /*
4954          * We're having a chicken and egg problem, even though we are
4955          * holding rq->lock, the cpu isn't yet set to this cpu so the
4956          * lockdep check in task_group() will fail.
4957          *
4958          * Similar case to sched_fork(). / Alternatively we could
4959          * use task_rq_lock() here and obtain the other rq->lock.
4960          *
4961          * Silence PROVE_RCU
4962          */
4963         rcu_read_lock();
4964         __set_task_cpu(idle, cpu);
4965         rcu_read_unlock();
4966 
4967         rq->curr = rq->idle = idle;
4968         idle->on_rq = TASK_ON_RQ_QUEUED;
4969 #ifdef CONFIG_SMP
4970         idle->on_cpu = 1;
4971 #endif
4972         raw_spin_unlock(&rq->lock);
4973         raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
4974 
4975         /* Set the preempt count _outside_ the spinlocks! */
4976         init_idle_preempt_count(idle, cpu);
4977 
4978         /*
4979          * The idle tasks have their own, simple scheduling class:
4980          */
4981         idle->sched_class = &idle_sched_class;
4982         ftrace_graph_init_idle_task(idle, cpu);
4983         vtime_init_idle(idle, cpu);
4984 #ifdef CONFIG_SMP
4985         sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4986 #endif
4987 }
4988 
4989 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4990                               const struct cpumask *trial)
4991 {
4992         int ret = 1, trial_cpus;
4993         struct dl_bw *cur_dl_b;
4994         unsigned long flags;
4995 
4996         if (!cpumask_weight(cur))
4997                 return ret;
4998 
4999         rcu_read_lock_sched();
5000         cur_dl_b = dl_bw_of(cpumask_any(cur));
5001         trial_cpus = cpumask_weight(trial);
5002 
5003         raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5004         if (cur_dl_b->bw != -1 &&
5005             cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5006                 ret = 0;
5007         raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5008         rcu_read_unlock_sched();
5009 
5010         return ret;
5011 }
5012 
5013 int task_can_attach(struct task_struct *p,
5014                     const struct cpumask *cs_cpus_allowed)
5015 {
5016         int ret = 0;
5017 
5018         /*
5019          * Kthreads which disallow setaffinity shouldn't be moved
5020          * to a new cpuset; we don't want to change their cpu
5021          * affinity and isolating such threads by their set of
5022          * allowed nodes is unnecessary.  Thus, cpusets are not
5023          * applicable for such threads.  This prevents checking for
5024          * success of set_cpus_allowed_ptr() on all attached tasks
5025          * before cpus_allowed may be changed.
5026          */
5027         if (p->flags & PF_NO_SETAFFINITY) {
5028                 ret = -EINVAL;
5029                 goto out;
5030         }
5031 
5032 #ifdef CONFIG_SMP
5033         if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5034                                               cs_cpus_allowed)) {
5035                 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5036                                                         cs_cpus_allowed);
5037                 struct dl_bw *dl_b;
5038                 bool overflow;
5039                 int cpus;
5040                 unsigned long flags;
5041 
5042                 rcu_read_lock_sched();
5043                 dl_b = dl_bw_of(dest_cpu);
5044                 raw_spin_lock_irqsave(&dl_b->lock, flags);
5045                 cpus = dl_bw_cpus(dest_cpu);
5046                 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5047                 if (overflow)
5048                         ret = -EBUSY;
5049                 else {
5050                         /*
5051                          * We reserve space for this task in the destination
5052                          * root_domain, as we can't fail after this point.
5053                          * We will free resources in the source root_domain
5054                          * later on (see set_cpus_allowed_dl()).
5055                          */
5056                         __dl_add(dl_b, p->dl.dl_bw);
5057                 }
5058                 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5059                 rcu_read_unlock_sched();
5060 
5061         }
5062 #endif
5063 out:
5064         return ret;
5065 }
5066 
5067 #ifdef CONFIG_SMP
5068 
5069 #ifdef CONFIG_NUMA_BALANCING
5070 /* Migrate current task p to target_cpu */
5071 int migrate_task_to(struct task_struct *p, int target_cpu)
5072 {
5073         struct migration_arg arg = { p, target_cpu };
5074         int curr_cpu = task_cpu(p);
5075 
5076         if (curr_cpu == target_cpu)
5077                 return 0;
5078 
5079         if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5080                 return -EINVAL;
5081 
5082         /* TODO: This is not properly updating schedstats */
5083 
5084         trace_sched_move_numa(p, curr_cpu, target_cpu);
5085         return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5086 }
5087 
5088 /*
5089  * Requeue a task on a given node and accurately track the number of NUMA
5090  * tasks on the runqueues
5091  */
5092 void sched_setnuma(struct task_struct *p, int nid)
5093 {
5094         struct rq *rq;
5095         unsigned long flags;
5096         bool queued, running;
5097 
5098         rq = task_rq_lock(p, &flags);
5099         queued = task_on_rq_queued(p);
5100         running = task_current(rq, p);
5101 
5102         if (queued)
5103                 dequeue_task(rq, p, 0);
5104         if (running)
5105                 put_prev_task(rq, p);
5106 
5107         p->numa_preferred_nid = nid;
5108 
5109         if (running)
5110                 p->sched_class->set_curr_task(rq);
5111         if (queued)
5112                 enqueue_task(rq, p, 0);
5113         task_rq_unlock(rq, p, &flags);
5114 }
5115 #endif /* CONFIG_NUMA_BALANCING */
5116 
5117 #ifdef CONFIG_HOTPLUG_CPU
5118 /*
5119  * Ensures that the idle task is using init_mm right before its cpu goes
5120  * offline.
5121  */
5122 void idle_task_exit(void)
5123 {
5124         struct mm_struct *mm = current->active_mm;
5125 
5126         BUG_ON(cpu_online(smp_processor_id()));
5127 
5128         if (mm != &init_mm) {
5129                 switch_mm(mm, &init_mm, current);
5130                 finish_arch_post_lock_switch();
5131         }
5132         mmdrop(mm);
5133 }
5134 
5135 /*
5136  * Since this CPU is going 'away' for a while, fold any nr_active delta
5137  * we might have. Assumes we're called after migrate_tasks() so that the
5138  * nr_active count is stable.
5139  *
5140  * Also see the comment "Global load-average calculations".
5141  */
5142 static void calc_load_migrate(struct rq *rq)
5143 {
5144         long delta = calc_load_fold_active(rq);
5145         if (delta)
5146                 atomic_long_add(delta, &calc_load_tasks);
5147 }
5148 
5149 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5150 {
5151 }
5152 
5153 static const struct sched_class fake_sched_class = {
5154         .put_prev_task = put_prev_task_fake,
5155 };
5156 
5157 static struct task_struct fake_task = {
5158         /*
5159          * Avoid pull_{rt,dl}_task()
5160          */
5161         .prio = MAX_PRIO + 1,
5162         .sched_class = &fake_sched_class,
5163 };
5164 
5165 /*
5166  * Migrate all tasks from the rq, sleeping tasks will be migrated by
5167  * try_to_wake_up()->select_task_rq().
5168  *
5169  * Called with rq->lock held even though we'er in stop_machine() and
5170  * there's no concurrency possible, we hold the required locks anyway
5171  * because of lock validation efforts.
5172  */
5173 static void migrate_tasks(struct rq *dead_rq)
5174 {
5175         struct rq *rq = dead_rq;
5176         struct task_struct *next, *stop = rq->stop;
5177         int dest_cpu;
5178 
5179         /*
5180          * Fudge the rq selection such that the below task selection loop
5181          * doesn't get stuck on the currently eligible stop task.
5182          *
5183          * We're currently inside stop_machine() and the rq is either stuck
5184          * in the stop_machine_cpu_stop() loop, or we're executing this code,
5185          * either way we should never end up calling schedule() until we're
5186          * done here.
5187          */
5188         rq->stop = NULL;
5189 
5190         /*
5191          * put_prev_task() and pick_next_task() sched
5192          * class method both need to have an up-to-date
5193          * value of rq->clock[_task]
5194          */
5195         update_rq_clock(rq);
5196 
5197         for (;;) {
5198                 /*
5199                  * There's this thread running, bail when that's the only
5200                  * remaining thread.
5201                  */
5202                 if (rq->nr_running == 1)
5203                         break;
5204 
5205                 /*
5206                  * pick_next_task assumes pinned rq->lock.
5207                  */
5208                 lockdep_pin_lock(&rq->lock);
5209                 next = pick_next_task(rq, &fake_task);
5210                 BUG_ON(!next);
5211                 next->sched_class->put_prev_task(rq, next);
5212 
5213                 /*
5214                  * Rules for changing task_struct::cpus_allowed are holding
5215                  * both pi_lock and rq->lock, such that holding either
5216                  * stabilizes the mask.
5217                  *
5218                  * Drop rq->lock is not quite as disastrous as it usually is
5219                  * because !cpu_active at this point, which means load-balance
5220                  * will not interfere. Also, stop-machine.
5221                  */
5222                 lockdep_unpin_lock(&rq->lock);
5223                 raw_spin_unlock(&rq->lock);
5224                 raw_spin_lock(&next->pi_lock);
5225                 raw_spin_lock(&rq->lock);
5226 
5227                 /*
5228                  * Since we're inside stop-machine, _nothing_ should have
5229                  * changed the task, WARN if weird stuff happened, because in
5230                  * that case the above rq->lock drop is a fail too.
5231                  */
5232                 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5233                         raw_spin_unlock(&next->pi_lock);
5234                         continue;
5235                 }
5236 
5237                 /* Find suitable destination for @next, with force if needed. */
5238                 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5239 
5240                 rq = __migrate_task(rq, next, dest_cpu);
5241                 if (rq != dead_rq) {
5242                         raw_spin_unlock(&rq->lock);
5243                         rq = dead_rq;
5244                         raw_spin_lock(&rq->lock);
5245                 }
5246                 raw_spin_unlock(&next->pi_lock);
5247         }
5248 
5249         rq->stop = stop;
5250 }
5251 #endif /* CONFIG_HOTPLUG_CPU */
5252 
5253 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5254 
5255 static struct ctl_table sd_ctl_dir[] = {
5256         {
5257                 .procname       = "sched_domain",
5258                 .mode           = 0555,
5259         },
5260         {}
5261 };
5262 
5263 static struct ctl_table sd_ctl_root[] = {
5264         {
5265                 .procname       = "kernel",
5266                 .mode           = 0555,
5267                 .child          = sd_ctl_dir,
5268         },
5269         {}
5270 };
5271 
5272 static struct ctl_table *sd_alloc_ctl_entry(int n)
5273 {
5274         struct ctl_table *entry =
5275                 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5276 
5277         return entry;
5278 }
5279 
5280 static void sd_free_ctl_entry(struct ctl_table **tablep)
5281 {
5282         struct ctl_table *entry;
5283 
5284         /*
5285          * In the intermediate directories, both the child directory and
5286          * procname are dynamically allocated and could fail but the mode
5287          * will always be set. In the lowest directory the names are
5288          * static strings and all have proc handlers.
5289          */
5290         for (entry = *tablep; entry->mode; entry++) {
5291                 if (entry->child)
5292                         sd_free_ctl_entry(&entry->child);
5293                 if (entry->proc_handler == NULL)
5294                         kfree(entry->procname);
5295         }
5296 
5297         kfree(*tablep);
5298         *tablep = NULL;
5299 }
5300 
5301 static int min_load_idx = 0;
5302 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5303 
5304 static void
5305 set_table_entry(struct ctl_table *entry,
5306                 const char *procname, void *data, int maxlen,
5307                 umode_t mode, proc_handler *proc_handler,
5308                 bool load_idx)
5309 {
5310         entry->procname = procname;
5311         entry->data = data;
5312         entry->maxlen = maxlen;
5313         entry->mode = mode;
5314         entry->proc_handler = proc_handler;
5315 
5316         if (load_idx) {
5317                 entry->extra1 = &min_load_idx;
5318                 entry->extra2 = &max_load_idx;
5319         }
5320 }
5321 
5322 static struct ctl_table *
5323 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5324 {
5325         struct ctl_table *table = sd_alloc_ctl_entry(14);
5326 
5327         if (table == NULL)
5328                 return NULL;
5329 
5330         set_table_entry(&table[0], "min_interval", &sd->min_interval,
5331                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5332         set_table_entry(&table[1], "max_interval", &sd->max_interval,
5333                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5334         set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5335                 sizeof(int), 0644, proc_dointvec_minmax, true);
5336         set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5337                 sizeof(int), 0644, proc_dointvec_minmax, true);
5338         set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5339                 sizeof(int), 0644, proc_dointvec_minmax, true);
5340         set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5341                 sizeof(int), 0644, proc_dointvec_minmax, true);
5342         set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5343                 sizeof(int), 0644, proc_dointvec_minmax, true);
5344         set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5345                 sizeof(int), 0644, proc_dointvec_minmax, false);
5346         set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5347                 sizeof(int), 0644, proc_dointvec_minmax, false);
5348         set_table_entry(&table[9], "cache_nice_tries",
5349                 &sd->cache_nice_tries,
5350                 sizeof(int), 0644, proc_dointvec_minmax, false);
5351         set_table_entry(&table[10], "flags", &sd->flags,
5352                 sizeof(int), 0644, proc_dointvec_minmax, false);
5353         set_table_entry(&table[11], "max_newidle_lb_cost",
5354                 &sd->max_newidle_lb_cost,
5355                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5356         set_table_entry(&table[12], "name", sd->name,
5357                 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5358         /* &table[13] is terminator */
5359 
5360         return table;
5361 }
5362 
5363 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5364 {
5365         struct ctl_table *entry, *table;
5366         struct sched_domain *sd;
5367         int domain_num = 0, i;
5368         char buf[32];
5369 
5370         for_each_domain(cpu, sd)
5371                 domain_num++;
5372         entry = table = sd_alloc_ctl_entry(domain_num + 1);
5373         if (table == NULL)
5374                 return NULL;
5375 
5376         i = 0;
5377         for_each_domain(cpu, sd) {
5378                 snprintf(buf, 32, "domain%d", i);
5379                 entry->procname = kstrdup(buf, GFP_KERNEL);
5380                 entry->mode = 0555;
5381                 entry->child = sd_alloc_ctl_domain_table(sd);
5382                 entry++;
5383                 i++;
5384         }
5385         return table;
5386 }
5387 
5388 static struct ctl_table_header *sd_sysctl_header;
5389 static void register_sched_domain_sysctl(void)
5390 {
5391         int i, cpu_num = num_possible_cpus();
5392         struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5393         char buf[32];
5394 
5395         WARN_ON(sd_ctl_dir[0].child);
5396         sd_ctl_dir[0].child = entry;
5397 
5398         if (entry == NULL)
5399                 return;
5400 
5401         for_each_possible_cpu(i) {
5402                 snprintf(buf, 32, "cpu%d", i);
5403                 entry->procname = kstrdup(buf, GFP_KERNEL);
5404                 entry->mode = 0555;
5405                 entry->child = sd_alloc_ctl_cpu_table(i);
5406                 entry++;
5407         }
5408 
5409         WARN_ON(sd_sysctl_header);
5410         sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5411 }
5412 
5413 /* may be called multiple times per register */
5414 static void unregister_sched_domain_sysctl(void)
5415 {
5416         unregister_sysctl_table(sd_sysctl_header);
5417         sd_sysctl_header = NULL;
5418         if (sd_ctl_dir[0].child)
5419                 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5420 }
5421 #else
5422 static void register_sched_domain_sysctl(void)
5423 {
5424 }
5425 static void unregister_sched_domain_sysctl(void)
5426 {
5427 }
5428 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5429 
5430 static void set_rq_online(struct rq *rq)
5431 {
5432         if (!rq->online) {
5433                 const struct sched_class *class;
5434 
5435                 cpumask_set_cpu(rq->cpu, rq->rd->online);
5436                 rq->online = 1;
5437 
5438                 for_each_class(class) {
5439                         if (class->rq_online)
5440                                 class->rq_online(rq);
5441                 }
5442         }
5443 }
5444 
5445 static void set_rq_offline(struct rq *rq)
5446 {
5447         if (rq->online) {
5448                 const struct sched_class *class;
5449 
5450                 for_each_class(class) {
5451                         if (class->rq_offline)
5452                                 class->rq_offline(rq);
5453                 }
5454 
5455                 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5456                 rq->online = 0;
5457         }
5458 }
5459 
5460 /*
5461  * migration_call - callback that gets triggered when a CPU is added.
5462  * Here we can start up the necessary migration thread for the new CPU.
5463  */
5464 static int
5465 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5466 {
5467         int cpu = (long)hcpu;
5468         unsigned long flags;
5469         struct rq *rq = cpu_rq(cpu);
5470 
5471         switch (action & ~CPU_TASKS_FROZEN) {
5472 
5473         case CPU_UP_PREPARE:
5474                 rq->calc_load_update = calc_load_update;
5475                 break;
5476 
5477         case CPU_ONLINE:
5478                 /* Update our root-domain */
5479                 raw_spin_lock_irqsave(&rq->lock, flags);
5480                 if (rq->rd) {
5481                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5482 
5483                         set_rq_online(rq);
5484                 }
5485                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5486                 break;
5487 
5488 #ifdef CONFIG_HOTPLUG_CPU
5489         case CPU_DYING:
5490                 sched_ttwu_pending();
5491                 /* Update our root-domain */
5492                 raw_spin_lock_irqsave(&rq->lock, flags);
5493                 if (rq->rd) {
5494                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5495                         set_rq_offline(rq);
5496                 }
5497                 migrate_tasks(rq);
5498                 BUG_ON(rq->nr_running != 1); /* the migration thread */
5499                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5500                 break;
5501 
5502         case CPU_DEAD:
5503                 calc_load_migrate(rq);
5504                 break;
5505 #endif
5506         }
5507 
5508         update_max_interval();
5509 
5510         return NOTIFY_OK;
5511 }
5512 
5513 /*
5514  * Register at high priority so that task migration (migrate_all_tasks)
5515  * happens before everything else.  This has to be lower priority than
5516  * the notifier in the perf_event subsystem, though.
5517  */
5518 static struct notifier_block migration_notifier = {
5519         .notifier_call = migration_call,
5520         .priority = CPU_PRI_MIGRATION,
5521 };
5522 
5523 static void set_cpu_rq_start_time(void)
5524 {
5525         int cpu = smp_processor_id();
5526         struct rq *rq = cpu_rq(cpu);
5527         rq->age_stamp = sched_clock_cpu(cpu);
5528 }
5529 
5530 static int sched_cpu_active(struct notifier_block *nfb,
5531                                       unsigned long action, void *hcpu)
5532 {
5533         switch (action & ~CPU_TASKS_FROZEN) {
5534         case CPU_STARTING:
5535                 set_cpu_rq_start_time();
5536                 return NOTIFY_OK;
5537         case CPU_ONLINE:
5538                 /*
5539                  * At this point a starting CPU has marked itself as online via
5540                  * set_cpu_online(). But it might not yet have marked itself
5541                  * as active, which is essential from here on.
5542                  *
5543                  * Thus, fall-through and help the starting CPU along.
5544                  */
5545         case CPU_DOWN_FAILED:
5546                 set_cpu_active((long)hcpu, true);
5547                 return NOTIFY_OK;
5548         default:
5549                 return NOTIFY_DONE;
5550         }
5551 }
5552 
5553 static int sched_cpu_inactive(struct notifier_block *nfb,
5554                                         unsigned long action, void *hcpu)
5555 {
5556         switch (action & ~CPU_TASKS_FROZEN) {
5557         case CPU_DOWN_PREPARE:
5558                 set_cpu_active((long)hcpu, false);
5559                 return NOTIFY_OK;
5560         default:
5561                 return NOTIFY_DONE;
5562         }
5563 }
5564 
5565 static int __init migration_init(void)
5566 {
5567         void *cpu = (void *)(long)smp_processor_id();
5568         int err;
5569 
5570         /* Initialize migration for the boot CPU */
5571         err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5572         BUG_ON(err == NOTIFY_BAD);
5573         migration_call(&migration_notifier, CPU_ONLINE, cpu);
5574         register_cpu_notifier(&migration_notifier);
5575 
5576         /* Register cpu active notifiers */
5577         cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5578         cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5579 
5580         return 0;
5581 }
5582 early_initcall(migration_init);
5583 
5584 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5585 
5586 #ifdef CONFIG_SCHED_DEBUG
5587 
5588 static __read_mostly int sched_debug_enabled;
5589 
5590 static int __init sched_debug_setup(char *str)
5591 {
5592         sched_debug_enabled = 1;
5593 
5594         return 0;
5595 }
5596 early_param("sched_debug", sched_debug_setup);
5597 
5598 static inline bool sched_debug(void)
5599 {
5600         return sched_debug_enabled;
5601 }
5602 
5603 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5604                                   struct cpumask *groupmask)
5605 {
5606         struct sched_group *group = sd->groups;
5607 
5608         cpumask_clear(groupmask);
5609 
5610         printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5611 
5612         if (!(sd->flags & SD_LOAD_BALANCE)) {
5613                 printk("does not load-balance\n");
5614                 if (sd->parent)
5615                         printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5616                                         " has parent");
5617                 return -1;
5618         }
5619 
5620         printk(KERN_CONT "span %*pbl level %s\n",
5621                cpumask_pr_args(sched_domain_span(sd)), sd->name);
5622 
5623         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5624                 printk(KERN_ERR "ERROR: domain->span does not contain "
5625                                 "CPU%d\n", cpu);
5626         }
5627         if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5628                 printk(KERN_ERR "ERROR: domain->groups does not contain"
5629                                 " CPU%d\n", cpu);
5630         }
5631 
5632         printk(KERN_DEBUG "%*s groups:", level + 1, "");
5633         do {
5634                 if (!group) {
5635                         printk("\n");
5636                         printk(KERN_ERR "ERROR: group is NULL\n");
5637                         break;
5638                 }
5639 
5640                 if (!cpumask_weight(sched_group_cpus(group))) {
5641                         printk(KERN_CONT "\n");
5642                         printk(KERN_ERR "ERROR: empty group\n");
5643                         break;
5644                 }
5645 
5646                 if (!(sd->flags & SD_OVERLAP) &&
5647                     cpumask_intersects(groupmask, sched_group_cpus(group))) {
5648                         printk(KERN_CONT "\n");
5649                         printk(KERN_ERR "ERROR: repeated CPUs\n");
5650                         break;
5651                 }
5652 
5653                 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5654 
5655                 printk(KERN_CONT " %*pbl",
5656                        cpumask_pr_args(sched_group_cpus(group)));
5657                 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5658                         printk(KERN_CONT " (cpu_capacity = %d)",
5659                                 group->sgc->capacity);
5660                 }
5661 
5662                 group = group->next;
5663         } while (group != sd->groups);
5664         printk(KERN_CONT "\n");
5665 
5666         if (!cpumask_equal(sched_domain_span(sd), groupmask))
5667                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5668 
5669         if (sd->parent &&
5670             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5671                 printk(KERN_ERR "ERROR: parent span is not a superset "
5672                         "of domain->span\n");
5673         return 0;
5674 }
5675 
5676 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5677 {
5678         int level = 0;
5679 
5680         if (!sched_debug_enabled)
5681                 return;
5682 
5683         if (!sd) {
5684                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5685                 return;
5686         }
5687 
5688         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5689 
5690         for (;;) {
5691                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5692                         break;
5693                 level++;
5694                 sd = sd->parent;
5695                 if (!sd)
5696                         break;
5697         }
5698 }
5699 #else /* !CONFIG_SCHED_DEBUG */
5700 # define sched_domain_debug(sd, cpu) do { } while (0)
5701 static inline bool sched_debug(void)
5702 {
5703         return false;
5704 }
5705 #endif /* CONFIG_SCHED_DEBUG */
5706 
5707 static int sd_degenerate(struct sched_domain *sd)
5708 {
5709         if (cpumask_weight(sched_domain_span(sd)) == 1)
5710                 return 1;
5711 
5712         /* Following flags need at least 2 groups */
5713         if (sd->flags & (SD_LOAD_BALANCE |
5714                          SD_BALANCE_NEWIDLE |
5715                          SD_BALANCE_FORK |
5716                          SD_BALANCE_EXEC |
5717                          SD_SHARE_CPUCAPACITY |
5718                          SD_SHARE_PKG_RESOURCES |
5719                          SD_SHARE_POWERDOMAIN)) {
5720                 if (sd->groups != sd->groups->next)
5721                         return 0;
5722         }
5723 
5724         /* Following flags don't use groups */
5725         if (sd->flags & (SD_WAKE_AFFINE))
5726                 return 0;
5727 
5728         return 1;
5729 }
5730 
5731 static int
5732 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5733 {
5734         unsigned long cflags = sd->flags, pflags = parent->flags;
5735 
5736         if (sd_degenerate(parent))
5737                 return 1;
5738 
5739         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5740                 return 0;
5741 
5742         /* Flags needing groups don't count if only 1 group in parent */
5743         if (parent->groups == parent->groups->next) {
5744                 pflags &= ~(SD_LOAD_BALANCE |
5745                                 SD_BALANCE_NEWIDLE |
5746                                 SD_BALANCE_FORK |
5747                                 SD_BALANCE_EXEC |
5748                                 SD_SHARE_CPUCAPACITY |
5749                                 SD_SHARE_PKG_RESOURCES |
5750                                 SD_PREFER_SIBLING |
5751                                 SD_SHARE_POWERDOMAIN);
5752                 if (nr_node_ids == 1)
5753                         pflags &= ~SD_SERIALIZE;
5754         }
5755         if (~cflags & pflags)
5756                 return 0;
5757 
5758         return 1;
5759 }
5760 
5761 static void free_rootdomain(struct rcu_head *rcu)
5762 {
5763         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5764 
5765         cpupri_cleanup(&rd->cpupri);
5766         cpudl_cleanup(&rd->cpudl);
5767         free_cpumask_var(rd->dlo_mask);
5768         free_cpumask_var(rd->rto_mask);
5769         free_cpumask_var(rd->online);
5770         free_cpumask_var(rd->span);
5771         kfree(rd);
5772 }
5773 
5774 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5775 {
5776         struct root_domain *old_rd = NULL;
5777         unsigned long flags;
5778 
5779         raw_spin_lock_irqsave(&rq->lock, flags);
5780 
5781         if (rq->rd) {
5782                 old_rd = rq->rd;
5783 
5784                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5785                         set_rq_offline(rq);
5786 
5787                 cpumask_clear_cpu(rq->cpu, old_rd->span);
5788 
5789                 /*
5790                  * If we dont want to free the old_rd yet then
5791                  * set old_rd to NULL to skip the freeing later
5792                  * in this function:
5793                  */
5794                 if (!atomic_dec_and_test(&old_rd->refcount))
5795                         old_rd = NULL;
5796         }
5797 
5798         atomic_inc(&rd->refcount);
5799         rq->rd = rd;
5800 
5801         cpumask_set_cpu(rq->cpu, rd->span);
5802         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5803                 set_rq_online(rq);
5804 
5805         raw_spin_unlock_irqrestore(&rq->lock, flags);
5806 
5807         if (old_rd)
5808                 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5809 }
5810 
5811 static int init_rootdomain(struct root_domain *rd)
5812 {
5813         memset(rd, 0, sizeof(*rd));
5814 
5815         if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5816                 goto out;
5817         if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5818                 goto free_span;
5819         if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5820                 goto free_online;
5821         if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5822                 goto free_dlo_mask;
5823 
5824         init_dl_bw(&rd->dl_bw);
5825         if (cpudl_init(&rd->cpudl) != 0)
5826                 goto free_dlo_mask;
5827 
5828         if (cpupri_init(&rd->cpupri) != 0)
5829                 goto free_rto_mask;
5830         return 0;
5831 
5832 free_rto_mask:
5833         free_cpumask_var(rd->rto_mask);
5834 free_dlo_mask:
5835         free_cpumask_var(rd->dlo_mask);
5836 free_online:
5837         free_cpumask_var(rd->online);
5838 free_span:
5839         free_cpumask_var(rd->span);
5840 out:
5841         return -ENOMEM;
5842 }
5843 
5844 /*
5845  * By default the system creates a single root-domain with all cpus as
5846  * members (mimicking the global state we have today).
5847  */
5848 struct root_domain def_root_domain;
5849 
5850 static void init_defrootdomain(void)
5851 {
5852         init_rootdomain(&def_root_domain);
5853 
5854         atomic_set(&def_root_domain.refcount, 1);
5855 }
5856 
5857 static struct root_domain *alloc_rootdomain(void)
5858 {
5859         struct root_domain *rd;
5860 
5861         rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5862         if (!rd)
5863                 return NULL;
5864 
5865         if (init_rootdomain(rd) != 0) {
5866                 kfree(rd);
5867                 return NULL;
5868         }
5869 
5870         return rd;
5871 }
5872 
5873 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5874 {
5875         struct sched_group *tmp, *first;
5876 
5877         if (!sg)
5878                 return;
5879 
5880         first = sg;
5881         do {
5882                 tmp = sg->next;
5883 
5884                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5885                         kfree(sg->sgc);
5886 
5887                 kfree(sg);
5888                 sg = tmp;
5889         } while (sg != first);
5890 }
5891 
5892 static void free_sched_domain(struct rcu_head *rcu)
5893 {
5894         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5895 
5896         /*
5897          * If its an overlapping domain it has private groups, iterate and
5898          * nuke them all.
5899          */
5900         if (sd->flags & SD_OVERLAP) {
5901                 free_sched_groups(sd->groups, 1);
5902         } else if (atomic_dec_and_test(&sd->groups->ref)) {
5903                 kfree(sd->groups->sgc);
5904                 kfree(sd->groups);
5905         }
5906         kfree(sd);
5907 }
5908 
5909 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5910 {
5911         call_rcu(&sd->rcu, free_sched_domain);
5912 }
5913 
5914 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5915 {
5916         for (; sd; sd = sd->parent)
5917                 destroy_sched_domain(sd, cpu);
5918 }
5919 
5920 /*
5921  * Keep a special pointer to the highest sched_domain that has
5922  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5923  * allows us to avoid some pointer chasing select_idle_sibling().
5924  *
5925  * Also keep a unique ID per domain (we use the first cpu number in
5926  * the cpumask of the domain), this allows us to quickly tell if
5927  * two cpus are in the same cache domain, see cpus_share_cache().
5928  */
5929 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5930 DEFINE_PER_CPU(int, sd_llc_size);
5931 DEFINE_PER_CPU(int, sd_llc_id);
5932 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5933 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5934 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5935 
5936 static void update_top_cache_domain(int cpu)
5937 {
5938         struct sched_domain *sd;
5939         struct sched_domain *busy_sd = NULL;
5940         int id = cpu;
5941         int size = 1;
5942 
5943         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5944         if (sd) {
5945                 id = cpumask_first(sched_domain_span(sd));
5946                 size = cpumask_weight(sched_domain_span(sd));
5947                 busy_sd = sd->parent; /* sd_busy */
5948         }
5949         rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5950 
5951         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5952         per_cpu(sd_llc_size, cpu) = size;
5953         per_cpu(sd_llc_id, cpu) = id;
5954 
5955         sd = lowest_flag_domain(cpu, SD_NUMA);
5956         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5957 
5958         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5959         rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5960 }
5961 
5962 /*
5963  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5964  * hold the hotplug lock.
5965  */
5966 static void
5967 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5968 {
5969         struct rq *rq = cpu_rq(cpu);
5970         struct sched_domain *tmp;
5971 
5972         /* Remove the sched domains which do not contribute to scheduling. */
5973         for (tmp = sd; tmp; ) {
5974                 struct sched_domain *parent = tmp->parent;
5975                 if (!parent)
5976                         break;
5977 
5978                 if (sd_parent_degenerate(tmp, parent)) {
5979                         tmp->parent = parent->parent;
5980                         if (parent->parent)
5981                                 parent->parent->child = tmp;
5982                         /*
5983                          * Transfer SD_PREFER_SIBLING down in case of a
5984                          * degenerate parent; the spans match for this
5985                          * so the property transfers.
5986                          */
5987                         if (parent->flags & SD_PREFER_SIBLING)
5988                                 tmp->flags |= SD_PREFER_SIBLING;
5989                         destroy_sched_domain(parent, cpu);
5990                 } else
5991                         tmp = tmp->parent;
5992         }
5993 
5994         if (sd && sd_degenerate(sd)) {
5995                 tmp = sd;
5996                 sd = sd->parent;
5997                 destroy_sched_domain(tmp, cpu);
5998                 if (sd)
5999                         sd->child = NULL;
6000         }
6001 
6002         sched_domain_debug(sd, cpu);
6003 
6004         rq_attach_root(rq, rd);
6005         tmp = rq->sd;
6006         rcu_assign_pointer(rq->sd, sd);
6007         destroy_sched_domains(tmp, cpu);
6008 
6009         update_top_cache_domain(cpu);
6010 }
6011 
6012 /* Setup the mask of cpus configured for isolated domains */
6013 static int __init isolated_cpu_setup(char *str)
6014 {
6015         alloc_bootmem_cpumask_var(&cpu_isolated_map);
6016         cpulist_parse(str, cpu_isolated_map);
6017         return 1;
6018 }
6019 
6020 __setup("isolcpus=", isolated_cpu_setup);
6021 
6022 struct s_data {
6023         struct sched_domain ** __percpu sd;
6024         struct root_domain      *rd;
6025 };
6026 
6027 enum s_alloc {
6028         sa_rootdomain,
6029         sa_sd,
6030         sa_sd_storage,
6031         sa_none,
6032 };
6033 
6034 /*
6035  * Build an iteration mask that can exclude certain CPUs from the upwards
6036  * domain traversal.
6037  *
6038  * Asymmetric node setups can result in situations where the domain tree is of
6039  * unequal depth, make sure to skip domains that already cover the entire
6040  * range.
6041  *
6042  * In that case build_sched_domains() will have terminated the iteration early
6043  * and our sibling sd spans will be empty. Domains should always include the
6044  * cpu they're built on, so check that.
6045  *
6046  */
6047 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6048 {
6049         const struct cpumask *span = sched_domain_span(sd);
6050         struct sd_data *sdd = sd->private;
6051         struct sched_domain *sibling;
6052         int i;
6053 
6054         for_each_cpu(i, span) {
6055                 sibling = *per_cpu_ptr(sdd->sd, i);
6056                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6057                         continue;
6058 
6059                 cpumask_set_cpu(i, sched_group_mask(sg));
6060         }
6061 }
6062 
6063 /*
6064  * Return the canonical balance cpu for this group, this is the first cpu
6065  * of this group that's also in the iteration mask.
6066  */
6067 int group_balance_cpu(struct sched_group *sg)
6068 {
6069         return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6070 }
6071 
6072 static int
6073 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6074 {
6075         struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6076         const struct cpumask *span = sched_domain_span(sd);
6077         struct cpumask *covered = sched_domains_tmpmask;
6078         struct sd_data *sdd = sd->private;
6079         struct sched_domain *sibling;
6080         int i;
6081 
6082         cpumask_clear(covered);
6083 
6084         for_each_cpu(i, span) {
6085                 struct cpumask *sg_span;
6086 
6087                 if (cpumask_test_cpu(i, covered))
6088                         continue;
6089 
6090                 sibling = *per_cpu_ptr(sdd->sd, i);
6091 
6092                 /* See the comment near build_group_mask(). */
6093                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6094                         continue;
6095 
6096                 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6097                                 GFP_KERNEL, cpu_to_node(cpu));
6098 
6099                 if (!sg)
6100                         goto fail;
6101 
6102                 sg_span = sched_group_cpus(sg);
6103                 if (sibling->child)
6104                         cpumask_copy(sg_span, sched_domain_span(sibling->child));
6105                 else
6106                         cpumask_set_cpu(i, sg_span);
6107 
6108                 cpumask_or(covered, covered, sg_span);
6109 
6110                 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6111                 if (atomic_inc_return(&sg->sgc->ref) == 1)
6112                         build_group_mask(sd, sg);
6113 
6114                 /*
6115                  * Initialize sgc->capacity such that even if we mess up the
6116                  * domains and no possible iteration will get us here, we won't
6117                  * die on a /0 trap.
6118                  */
6119                 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6120 
6121                 /*
6122                  * Make sure the first group of this domain contains the
6123                  * canonical balance cpu. Otherwise the sched_domain iteration
6124                  * breaks. See update_sg_lb_stats().
6125                  */
6126                 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6127                     group_balance_cpu(sg) == cpu)
6128                         groups = sg;
6129 
6130                 if (!first)
6131                         first = sg;
6132                 if (last)
6133                         last->next = sg;
6134                 last = sg;
6135                 last->next = first;
6136         }
6137         sd->groups = groups;
6138 
6139         return 0;
6140 
6141 fail:
6142         free_sched_groups(first, 0);
6143 
6144         return -ENOMEM;
6145 }
6146 
6147 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6148 {
6149         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6150         struct sched_domain *child = sd->child;
6151 
6152         if (child)
6153                 cpu = cpumask_first(sched_domain_span(child));
6154 
6155         if (sg) {
6156                 *sg = *per_cpu_ptr(sdd->sg, cpu);
6157                 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6158                 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6159         }
6160 
6161         return cpu;
6162 }
6163 
6164 /*
6165  * build_sched_groups will build a circular linked list of the groups
6166  * covered by the given span, and will set each group's ->cpumask correctly,
6167  * and ->cpu_capacity to 0.
6168  *
6169  * Assumes the sched_domain tree is fully constructed
6170  */
6171 static int
6172 build_sched_groups(struct sched_domain *sd, int cpu)
6173 {
6174         struct sched_group *first = NULL, *last = NULL;
6175         struct sd_data *sdd = sd->private;
6176         const struct cpumask *span = sched_domain_span(sd);
6177         struct cpumask *covered;
6178         int i;
6179 
6180         get_group(cpu, sdd, &sd->groups);
6181         atomic_inc(&sd->groups->ref);
6182 
6183         if (cpu != cpumask_first(span))
6184                 return 0;
6185 
6186         lockdep_assert_held(&sched_domains_mutex);
6187         covered = sched_domains_tmpmask;
6188 
6189         cpumask_clear(covered);
6190 
6191         for_each_cpu(i, span) {
6192                 struct sched_group *sg;
6193                 int group, j;
6194 
6195                 if (cpumask_test_cpu(i, covered))
6196                         continue;
6197 
6198                 group = get_group(i, sdd, &sg);
6199                 cpumask_setall(sched_group_mask(sg));
6200 
6201                 for_each_cpu(j, span) {
6202                         if (get_group(j, sdd, NULL) != group)
6203                                 continue;
6204 
6205                         cpumask_set_cpu(j, covered);
6206                         cpumask_set_cpu(j, sched_group_cpus(sg));
6207                 }
6208 
6209                 if (!first)
6210                         first = sg;
6211                 if (last)
6212                         last->next = sg;
6213                 last = sg;
6214         }
6215         last->next = first;
6216 
6217         return 0;
6218 }
6219 
6220 /*
6221  * Initialize sched groups cpu_capacity.
6222  *
6223  * cpu_capacity indicates the capacity of sched group, which is used while
6224  * distributing the load between different sched groups in a sched domain.
6225  * Typically cpu_capacity for all the groups in a sched domain will be same
6226  * unless there are asymmetries in the topology. If there are asymmetries,
6227  * group having more cpu_capacity will pickup more load compared to the
6228  * group having less cpu_capacity.
6229  */
6230 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6231 {
6232         struct sched_group *sg = sd->groups;
6233 
6234         WARN_ON(!sg);
6235 
6236         do {
6237                 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6238                 sg = sg->next;
6239         } while (sg != sd->groups);
6240 
6241         if (cpu != group_balance_cpu(sg))
6242                 return;
6243 
6244         update_group_capacity(sd, cpu);
6245         atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6246 }
6247 
6248 /*
6249  * Initializers for schedule domains
6250  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6251  */
6252 
6253 static int default_relax_domain_level = -1;
6254 int sched_domain_level_max;
6255 
6256 static int __init setup_relax_domain_level(char *str)
6257 {
6258         if (kstrtoint(str, 0, &default_relax_domain_level))
6259                 pr_warn("Unable to set relax_domain_level\n");
6260 
6261         return 1;
6262 }
6263 __setup("relax_domain_level=", setup_relax_domain_level);
6264 
6265 static void set_domain_attribute(struct sched_domain *sd,
6266                                  struct sched_domain_attr *attr)
6267 {
6268         int request;
6269 
6270         if (!attr || attr->relax_domain_level < 0) {
6271                 if (default_relax_domain_level < 0)
6272                         return;
6273                 else
6274                         request = default_relax_domain_level;
6275         } else
6276                 request = attr->relax_domain_level;
6277         if (request < sd->level) {
6278                 /* turn off idle balance on this domain */
6279                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6280         } else {
6281                 /* turn on idle balance on this domain */
6282                 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6283         }
6284 }
6285 
6286 static void __sdt_free(const struct cpumask *cpu_map);
6287 static int __sdt_alloc(const struct cpumask *cpu_map);
6288 
6289 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6290                                  const struct cpumask *cpu_map)
6291 {
6292         switch (what) {
6293         case sa_rootdomain:
6294                 if (!atomic_read(&d->rd->refcount))
6295                         free_rootdomain(&d->rd->rcu); /* fall through */
6296         case sa_sd:
6297                 free_percpu(d->sd); /* fall through */
6298         case sa_sd_storage:
6299                 __sdt_free(cpu_map); /* fall through */
6300         case sa_none:
6301                 break;
6302         }
6303 }
6304 
6305 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6306                                                    const struct cpumask *cpu_map)
6307 {
6308         memset(d, 0, sizeof(*d));
6309 
6310         if (__sdt_alloc(cpu_map))
6311                 return sa_sd_storage;
6312         d->sd = alloc_percpu(struct sched_domain *);
6313         if (!d->sd)
6314                 return sa_sd_storage;
6315         d->rd = alloc_rootdomain();
6316         if (!d->rd)
6317                 return sa_sd;
6318         return sa_rootdomain;
6319 }
6320 
6321 /*
6322  * NULL the sd_data elements we've used to build the sched_domain and
6323  * sched_group structure so that the subsequent __free_domain_allocs()
6324  * will not free the data we're using.
6325  */
6326 static void claim_allocations(int cpu, struct sched_domain *sd)
6327 {
6328         struct sd_data *sdd = sd->private;
6329 
6330         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6331         *per_cpu_ptr(sdd->sd, cpu) = NULL;
6332 
6333         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6334                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6335 
6336         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6337                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6338 }
6339 
6340 #ifdef CONFIG_NUMA
6341 static int sched_domains_numa_levels;
6342 enum numa_topology_type sched_numa_topology_type;
6343 static int *sched_domains_numa_distance;
6344 int sched_max_numa_distance;
6345 static struct cpumask ***sched_domains_numa_masks;
6346 static int sched_domains_curr_level;
6347 #endif
6348 
6349 /*
6350  * SD_flags allowed in topology descriptions.
6351  *
6352  * SD_SHARE_CPUCAPACITY      - describes SMT topologies
6353  * SD_SHARE_PKG_RESOURCES - describes shared caches
6354  * SD_NUMA                - describes NUMA topologies
6355  * SD_SHARE_POWERDOMAIN   - describes shared power domain
6356  *
6357  * Odd one out:
6358  * SD_ASYM_PACKING        - describes SMT quirks
6359  */
6360 #define TOPOLOGY_SD_FLAGS               \
6361         (SD_SHARE_CPUCAPACITY |         \
6362          SD_SHARE_PKG_RESOURCES |       \
6363          SD_NUMA |                      \
6364          SD_ASYM_PACKING |              \
6365          SD_SHARE_POWERDOMAIN)
6366 
6367 static struct sched_domain *
6368 sd_init(struct sched_domain_topology_level *tl, int cpu)
6369 {
6370         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6371         int sd_weight, sd_flags = 0;
6372 
6373 #ifdef CONFIG_NUMA
6374         /*
6375          * Ugly hack to pass state to sd_numa_mask()...
6376          */
6377         sched_domains_curr_level = tl->numa_level;
6378 #endif
6379 
6380         sd_weight = cpumask_weight(tl->mask(cpu));
6381 
6382         if (tl->sd_flags)
6383                 sd_flags = (*tl->sd_flags)();
6384         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6385                         "wrong sd_flags in topology description\n"))
6386                 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6387 
6388         *sd = (struct sched_domain){
6389                 .min_interval           = sd_weight,
6390                 .max_interval           = 2*sd_weight,
6391                 .busy_factor            = 32,
6392                 .imbalance_pct          = 125,
6393 
6394                 .cache_nice_tries       = 0,
6395                 .busy_idx               = 0,
6396                 .idle_idx               = 0,
6397                 .newidle_idx            = 0,
6398                 .wake_idx               = 0,
6399                 .forkexec_idx           = 0,
6400 
6401                 .flags                  = 1*SD_LOAD_BALANCE
6402                                         | 1*SD_BALANCE_NEWIDLE
6403                                         | 1*SD_BALANCE_EXEC
6404                                         | 1*SD_BALANCE_FORK
6405                                         | 0*SD_BALANCE_WAKE
6406                                         | 1*SD_WAKE_AFFINE
6407                                         | 0*SD_SHARE_CPUCAPACITY
6408                                         | 0*SD_SHARE_PKG_RESOURCES
6409                                         | 0*SD_SERIALIZE
6410                                         | 0*SD_PREFER_SIBLING
6411                                         | 0*SD_NUMA
6412                                         | sd_flags
6413                                         ,
6414 
6415                 .last_balance           = jiffies,
6416                 .balance_interval       = sd_weight,
6417                 .smt_gain               = 0,
6418                 .max_newidle_lb_cost    = 0,
6419                 .next_decay_max_lb_cost = jiffies,
6420 #ifdef CONFIG_SCHED_DEBUG
6421                 .name                   = tl->name,
6422 #endif
6423         };
6424 
6425         /*
6426          * Convert topological properties into behaviour.
6427          */
6428 
6429         if (sd->flags & SD_SHARE_CPUCAPACITY) {
6430                 sd->flags |= SD_PREFER_SIBLING;
6431                 sd->imbalance_pct = 110;
6432                 sd->smt_gain = 1178; /* ~15% */
6433 
6434         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6435                 sd->imbalance_pct = 117;
6436                 sd->cache_nice_tries = 1;
6437                 sd->busy_idx = 2;
6438 
6439 #ifdef CONFIG_NUMA
6440         } else if (sd->flags & SD_NUMA) {
6441                 sd->cache_nice_tries = 2;
6442                 sd->busy_idx = 3;
6443                 sd->idle_idx = 2;
6444 
6445                 sd->flags |= SD_SERIALIZE;
6446                 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6447                         sd->flags &= ~(SD_BALANCE_EXEC |
6448                                        SD_BALANCE_FORK |
6449                                        SD_WAKE_AFFINE);
6450                 }
6451 
6452 #endif
6453         } else {
6454                 sd->flags |= SD_PREFER_SIBLING;
6455                 sd->cache_nice_tries = 1;
6456                 sd->busy_idx = 2;
6457                 sd->idle_idx = 1;
6458         }
6459 
6460         sd->private = &tl->data;
6461 
6462         return sd;
6463 }
6464 
6465 /*
6466  * Topology list, bottom-up.
6467  */
6468 static struct sched_domain_topology_level default_topology[] = {
6469 #ifdef CONFIG_SCHED_SMT
6470         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6471 #endif
6472 #ifdef CONFIG_SCHED_MC
6473         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6474 #endif
6475         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6476         { NULL, },
6477 };
6478 
6479 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6480 
6481 #define for_each_sd_topology(tl)                        \
6482         for (tl = sched_domain_topology; tl->mask; tl++)
6483 
6484 void set_sched_topology(struct sched_domain_topology_level *tl)
6485 {
6486         sched_domain_topology = tl;
6487 }
6488 
6489 #ifdef CONFIG_NUMA
6490 
6491 static const struct cpumask *sd_numa_mask(int cpu)
6492 {
6493         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6494 }
6495 
6496 static void sched_numa_warn(const char *str)
6497 {
6498         static int done = false;
6499         int i,j;
6500 
6501         if (done)
6502                 return;
6503 
6504         done = true;
6505 
6506         printk(KERN_WARNING "ERROR: %s\n\n", str);
6507 
6508         for (i = 0; i < nr_node_ids; i++) {
6509                 printk(KERN_WARNING "  ");
6510                 for (j = 0; j < nr_node_ids; j++)
6511                         printk(KERN_CONT "%02d ", node_distance(i,j));
6512                 printk(KERN_CONT "\n");
6513         }
6514         printk(KERN_WARNING "\n");
6515 }
6516 
6517 bool find_numa_distance(int distance)
6518 {
6519         int i;
6520 
6521         if (distance == node_distance(0, 0))
6522                 return true;
6523 
6524         for (i = 0; i < sched_domains_numa_levels; i++) {
6525                 if (sched_domains_numa_distance[i] == distance)
6526                         return true;
6527         }
6528 
6529         return false;
6530 }
6531 
6532 /*
6533  * A system can have three types of NUMA topology:
6534  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6535  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6536  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6537  *
6538  * The difference between a glueless mesh topology and a backplane
6539  * topology lies in whether communication between not directly
6540  * connected nodes goes through intermediary nodes (where programs
6541  * could run), or through backplane controllers. This affects
6542  * placement of programs.
6543  *
6544  * The type of topology can be discerned with the following tests:
6545  * - If the maximum distance between any nodes is 1 hop, the system
6546  *   is directly connected.
6547  * - If for two nodes A and B, located N > 1 hops away from each other,
6548  *   there is an intermediary node C, which is < N hops away from both
6549  *   nodes A and B, the system is a glueless mesh.
6550  */
6551 static void init_numa_topology_type(void)
6552 {
6553         int a, b, c, n;
6554 
6555         n = sched_max_numa_distance;
6556 
6557         if (sched_domains_numa_levels <= 1) {
6558                 sched_numa_topology_type = NUMA_DIRECT;
6559                 return;
6560         }
6561 
6562         for_each_online_node(a) {
6563                 for_each_online_node(b) {
6564                         /* Find two nodes furthest removed from each other. */
6565                         if (node_distance(a, b) < n)
6566                                 continue;
6567 
6568                         /* Is there an intermediary node between a and b? */
6569                         for_each_online_node(c) {
6570                                 if (node_distance(a, c) < n &&
6571                                     node_distance(b, c) < n) {
6572                                         sched_numa_topology_type =
6573                                                         NUMA_GLUELESS_MESH;
6574                                         return;
6575                                 }
6576                         }
6577 
6578                         sched_numa_topology_type = NUMA_BACKPLANE;
6579                         return;
6580                 }
6581         }
6582 }
6583 
6584 static void sched_init_numa(void)
6585 {
6586         int next_distance, curr_distance = node_distance(0, 0);
6587         struct sched_domain_topology_level *tl;
6588         int level = 0;
6589         int i, j, k;
6590 
6591         sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6592         if (!sched_domains_numa_distance)
6593                 return;
6594 
6595         /*
6596          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6597          * unique distances in the node_distance() table.
6598          *
6599          * Assumes node_distance(0,j) includes all distances in
6600          * node_distance(i,j) in order to avoid cubic time.
6601          */
6602         next_distance = curr_distance;
6603         for (i = 0; i < nr_node_ids; i++) {
6604                 for (j = 0; j < nr_node_ids; j++) {
6605                         for (k = 0; k < nr_node_ids; k++) {
6606                                 int distance = node_distance(i, k);
6607 
6608                                 if (distance > curr_distance &&
6609                                     (distance < next_distance ||
6610                                      next_distance == curr_distance))
6611                                         next_distance = distance;
6612 
6613                                 /*
6614                                  * While not a strong assumption it would be nice to know
6615                                  * about cases where if node A is connected to B, B is not
6616                                  * equally connected to A.
6617                                  */
6618                                 if (sched_debug() && node_distance(k, i) != distance)
6619                                         sched_numa_warn("Node-distance not symmetric");
6620 
6621                                 if (sched_debug() && i && !find_numa_distance(distance))
6622                                         sched_numa_warn("Node-0 not representative");
6623                         }
6624                         if (next_distance != curr_distance) {
6625                                 sched_domains_numa_distance[level++] = next_distance;
6626                                 sched_domains_numa_levels = level;
6627                                 curr_distance = next_distance;
6628                         } else break;
6629                 }
6630 
6631                 /*
6632                  * In case of sched_debug() we verify the above assumption.
6633                  */
6634                 if (!sched_debug())
6635                         break;
6636         }
6637 
6638         if (!level)
6639                 return;
6640 
6641         /*
6642          * 'level' contains the number of unique distances, excluding the
6643          * identity distance node_distance(i,i).
6644          *
6645          * The sched_domains_numa_distance[] array includes the actual distance
6646          * numbers.
6647          */
6648 
6649         /*
6650          * Here, we should temporarily reset sched_domains_numa_levels to 0.
6651          * If it fails to allocate memory for array sched_domains_numa_masks[][],
6652          * the array will contain less then 'level' members. This could be
6653          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6654          * in other functions.
6655          *
6656          * We reset it to 'level' at the end of this function.
6657          */
6658         sched_domains_numa_levels = 0;
6659 
6660         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6661         if (!sched_domains_numa_masks)
6662                 return;
6663 
6664         /*
6665          * Now for each level, construct a mask per node which contains all
6666          * cpus of nodes that are that many hops away from us.
6667          */
6668         for (i = 0; i < level; i++) {
6669                 sched_domains_numa_masks[i] =
6670                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6671                 if (!sched_domains_numa_masks[i])
6672                         return;
6673 
6674                 for (j = 0; j < nr_node_ids; j++) {
6675                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6676                         if (!mask)
6677                                 return;
6678 
6679                         sched_domains_numa_masks[i][j] = mask;
6680 
6681                         for (k = 0; k < nr_node_ids; k++) {
6682                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
6683                                         continue;
6684 
6685                                 cpumask_or(mask, mask, cpumask_of_node(k));
6686                         }
6687                 }
6688         }
6689 
6690         /* Compute default topology size */
6691         for (i = 0; sched_domain_topology[i].mask; i++);
6692 
6693         tl = kzalloc((i + level + 1) *
6694                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6695         if (!tl)
6696                 return;
6697 
6698         /*
6699          * Copy the default topology bits..
6700          */
6701         for (i = 0; sched_domain_topology[i].mask; i++)
6702                 tl[i] = sched_domain_topology[i];
6703 
6704         /*
6705          * .. and append 'j' levels of NUMA goodness.
6706          */
6707         for (j = 0; j < level; i++, j++) {
6708                 tl[i] = (struct sched_domain_topology_level){
6709                         .mask = sd_numa_mask,
6710                         .sd_flags = cpu_numa_flags,
6711                         .flags = SDTL_OVERLAP,
6712                         .numa_level = j,
6713                         SD_INIT_NAME(NUMA)
6714                 };
6715         }
6716 
6717         sched_domain_topology = tl;
6718 
6719         sched_domains_numa_levels = level;
6720         sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6721 
6722         init_numa_topology_type();
6723 }
6724 
6725 static void sched_domains_numa_masks_set(int cpu)
6726 {
6727         int i, j;
6728         int node = cpu_to_node(cpu);
6729 
6730         for (i = 0; i < sched_domains_numa_levels; i++) {
6731                 for (j = 0; j < nr_node_ids; j++) {
6732                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
6733                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6734                 }
6735         }
6736 }
6737 
6738 static void sched_domains_numa_masks_clear(int cpu)
6739 {
6740         int i, j;
6741         for (i = 0; i < sched_domains_numa_levels; i++) {
6742                 for (j = 0; j < nr_node_ids; j++)
6743                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6744         }
6745 }
6746 
6747 /*
6748  * Update sched_domains_numa_masks[level][node] array when new cpus
6749  * are onlined.
6750  */
6751 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6752                                            unsigned long action,
6753                                            void *hcpu)
6754 {
6755         int cpu = (long)hcpu;
6756 
6757         switch (action & ~CPU_TASKS_FROZEN) {
6758         case CPU_ONLINE:
6759                 sched_domains_numa_masks_set(cpu);
6760                 break;
6761 
6762         case CPU_DEAD:
6763                 sched_domains_numa_masks_clear(cpu);
6764                 break;
6765 
6766         default:
6767                 return NOTIFY_DONE;
6768         }
6769 
6770         return NOTIFY_OK;
6771 }
6772 #else
6773 static inline void sched_init_numa(void)
6774 {
6775 }
6776 
6777 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6778                                            unsigned long action,
6779                                            void *hcpu)
6780 {
6781         return 0;
6782 }
6783 #endif /* CONFIG_NUMA */
6784 
6785 static int __sdt_alloc(const struct cpumask *cpu_map)
6786 {
6787         struct sched_domain_topology_level *tl;
6788         int j;
6789 
6790         for_each_sd_topology(tl) {
6791                 struct sd_data *sdd = &tl->data;
6792 
6793                 sdd->sd = alloc_percpu(struct sched_domain *);
6794                 if (!sdd->sd)
6795                         return -ENOMEM;
6796 
6797                 sdd->sg = alloc_percpu(struct sched_group *);
6798                 if (!sdd->sg)
6799                         return -ENOMEM;
6800 
6801                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6802                 if (!sdd->sgc)
6803                         return -ENOMEM;
6804 
6805                 for_each_cpu(j, cpu_map) {
6806                         struct sched_domain *sd;
6807                         struct sched_group *sg;
6808                         struct sched_group_capacity *sgc;
6809 
6810                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6811                                         GFP_KERNEL, cpu_to_node(j));
6812                         if (!sd)
6813                                 return -ENOMEM;
6814 
6815                         *per_cpu_ptr(sdd->sd, j) = sd;
6816 
6817                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6818                                         GFP_KERNEL, cpu_to_node(j));
6819                         if (!sg)
6820                                 return -ENOMEM;
6821 
6822                         sg->next = sg;
6823 
6824                         *per_cpu_ptr(sdd->sg, j) = sg;
6825 
6826                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6827                                         GFP_KERNEL, cpu_to_node(j));
6828                         if (!sgc)
6829                                 return -ENOMEM;
6830 
6831                         *per_cpu_ptr(sdd->sgc, j) = sgc;
6832                 }
6833         }
6834 
6835         return 0;
6836 }
6837 
6838 static void __sdt_free(const struct cpumask *cpu_map)
6839 {
6840         struct sched_domain_topology_level *tl;
6841         int j;
6842 
6843         for_each_sd_topology(tl) {
6844                 struct sd_data *sdd = &tl->data;
6845 
6846                 for_each_cpu(j, cpu_map) {
6847                         struct sched_domain *sd;
6848 
6849                         if (sdd->sd) {
6850                                 sd = *per_cpu_ptr(sdd->sd, j);
6851                                 if (sd && (sd->flags & SD_OVERLAP))
6852                                         free_sched_groups(sd->groups, 0);
6853                                 kfree(*per_cpu_ptr(sdd->sd, j));
6854                         }
6855 
6856                         if (sdd->sg)
6857                                 kfree(*per_cpu_ptr(sdd->sg, j));
6858                         if (sdd->sgc)
6859                                 kfree(*per_cpu_ptr(sdd->sgc, j));
6860                 }
6861                 free_percpu(sdd->sd);
6862                 sdd->sd = NULL;
6863                 free_percpu(sdd->sg);
6864                 sdd->sg = NULL;
6865                 free_percpu(sdd->sgc);
6866                 sdd->sgc = NULL;
6867         }
6868 }
6869 
6870 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6871                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6872                 struct sched_domain *child, int cpu)
6873 {
6874         struct sched_domain *sd = sd_init(tl, cpu);
6875         if (!sd)
6876                 return child;
6877 
6878         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6879         if (child) {
6880                 sd->level = child->level + 1;
6881                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6882                 child->parent = sd;
6883                 sd->child = child;
6884 
6885                 if (!cpumask_subset(sched_domain_span(child),
6886                                     sched_domain_span(sd))) {
6887                         pr_err("BUG: arch topology borken\n");
6888 #ifdef CONFIG_SCHED_DEBUG
6889                         pr_err("     the %s domain not a subset of the %s domain\n",
6890                                         child->name, sd->name);
6891 #endif
6892                         /* Fixup, ensure @sd has at least @child cpus. */
6893                         cpumask_or(sched_domain_span(sd),
6894                                    sched_domain_span(sd),
6895                                    sched_domain_span(child));
6896                 }
6897 
6898         }
6899         set_domain_attribute(sd, attr);
6900 
6901         return sd;
6902 }
6903 
6904 /*
6905  * Build sched domains for a given set of cpus and attach the sched domains
6906  * to the individual cpus
6907  */
6908 static int build_sched_domains(const struct cpumask *cpu_map,
6909                                struct sched_domain_attr *attr)
6910 {
6911         enum s_alloc alloc_state;
6912         struct sched_domain *sd;
6913         struct s_data d;
6914         int i, ret = -ENOMEM;
6915 
6916         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6917         if (alloc_state != sa_rootdomain)
6918                 goto error;
6919 
6920         /* Set up domains for cpus specified by the cpu_map. */
6921         for_each_cpu(i, cpu_map) {
6922                 struct sched_domain_topology_level *tl;
6923 
6924                 sd = NULL;
6925                 for_each_sd_topology(tl) {
6926                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6927                         if (tl == sched_domain_topology)
6928                                 *per_cpu_ptr(d.sd, i) = sd;
6929                         if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6930                                 sd->flags |= SD_OVERLAP;
6931                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6932                                 break;
6933                 }
6934         }
6935 
6936         /* Build the groups for the domains */
6937         for_each_cpu(i, cpu_map) {
6938                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6939                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
6940                         if (sd->flags & SD_OVERLAP) {
6941                                 if (build_overlap_sched_groups(sd, i))
6942                                         goto error;
6943                         } else {
6944                                 if (build_sched_groups(sd, i))
6945                                         goto error;
6946                         }
6947                 }
6948         }
6949 
6950         /* Calculate CPU capacity for physical packages and nodes */
6951         for (i = nr_cpumask_bits-1; i >= 0; i--) {
6952                 if (!cpumask_test_cpu(i, cpu_map))
6953                         continue;
6954 
6955                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6956                         claim_allocations(i, sd);
6957                         init_sched_groups_capacity(i, sd);
6958                 }
6959         }
6960 
6961         /* Attach the domains */
6962         rcu_read_lock();
6963         for_each_cpu(i, cpu_map) {
6964                 sd = *per_cpu_ptr(d.sd, i);
6965                 cpu_attach_domain(sd, d.rd, i);
6966         }
6967         rcu_read_unlock();
6968 
6969         ret = 0;
6970 error:
6971         __free_domain_allocs(&d, alloc_state, cpu_map);
6972         return ret;
6973 }
6974 
6975 static cpumask_var_t *doms_cur; /* current sched domains */
6976 static int ndoms_cur;           /* number of sched domains in 'doms_cur' */
6977 static struct sched_domain_attr *dattr_cur;
6978                                 /* attribues of custom domains in 'doms_cur' */
6979 
6980 /*
6981  * Special case: If a kmalloc of a doms_cur partition (array of
6982  * cpumask) fails, then fallback to a single sched domain,
6983  * as determined by the single cpumask fallback_doms.
6984  */
6985 static cpumask_var_t fallback_doms;
6986 
6987 /*
6988  * arch_update_cpu_topology lets virtualized architectures update the
6989  * cpu core maps. It is supposed to return 1 if the topology changed
6990  * or 0 if it stayed the same.
6991  */
6992 int __weak arch_update_cpu_topology(void)
6993 {
6994         return 0;
6995 }
6996 
6997 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6998 {
6999         int i;
7000         cpumask_var_t *doms;
7001 
7002         doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7003         if (!doms)
7004                 return NULL;
7005         for (i = 0; i < ndoms; i++) {
7006                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7007                         free_sched_domains(doms, i);
7008                         return NULL;
7009                 }
7010         }
7011         return doms;
7012 }
7013 
7014 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7015 {
7016         unsigned int i;
7017         for (i = 0; i < ndoms; i++)
7018                 free_cpumask_var(doms[i]);
7019         kfree(doms);
7020 }
7021 
7022 /*
7023  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7024  * For now this just excludes isolated cpus, but could be used to
7025  * exclude other special cases in the future.
7026  */
7027 static int init_sched_domains(const struct cpumask *cpu_map)
7028 {
7029         int err;
7030 
7031         arch_update_cpu_topology();
7032         ndoms_cur = 1;
7033         doms_cur = alloc_sched_domains(ndoms_cur);
7034         if (!doms_cur)
7035                 doms_cur = &fallback_doms;
7036         cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7037         err = build_sched_domains(doms_cur[0], NULL);
7038         register_sched_domain_sysctl();
7039 
7040         return err;
7041 }
7042 
7043 /*
7044  * Detach sched domains from a group of cpus specified in cpu_map
7045  * These cpus will now be attached to the NULL domain
7046  */
7047 static void detach_destroy_domains(const struct cpumask *cpu_map)
7048 {
7049         int i;
7050 
7051         rcu_read_lock();
7052         for_each_cpu(i, cpu_map)
7053                 cpu_attach_domain(NULL, &def_root_domain, i);
7054         rcu_read_unlock();
7055 }
7056 
7057 /* handle null as "default" */
7058 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7059                         struct sched_domain_attr *new, int idx_new)
7060 {
7061         struct sched_domain_attr tmp;
7062 
7063         /* fast path */
7064         if (!new && !cur)
7065                 return 1;
7066 
7067         tmp = SD_ATTR_INIT;
7068         return !memcmp(cur ? (cur + idx_cur) : &tmp,
7069                         new ? (new + idx_new) : &tmp,
7070                         sizeof(struct sched_domain_attr));
7071 }
7072 
7073 /*
7074  * Partition sched domains as specified by the 'ndoms_new'
7075  * cpumasks in the array doms_new[] of cpumasks. This compares
7076  * doms_new[] to the current sched domain partitioning, doms_cur[].
7077  * It destroys each deleted domain and builds each new domain.
7078  *
7079  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7080  * The masks don't intersect (don't overlap.) We should setup one
7081  * sched domain for each mask. CPUs not in any of the cpumasks will
7082  * not be load balanced. If the same cpumask appears both in the
7083  * current 'doms_cur' domains and in the new 'doms_new', we can leave
7084  * it as it is.
7085  *
7086  * The passed in 'doms_new' should be allocated using
7087  * alloc_sched_domains.  This routine takes ownership of it and will
7088  * free_sched_domains it when done with it. If the caller failed the
7089  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7090  * and partition_sched_domains() will fallback to the single partition
7091  * 'fallback_doms', it also forces the domains to be rebuilt.
7092  *
7093  * If doms_new == NULL it will be replaced with cpu_online_mask.
7094  * ndoms_new == 0 is a special case for destroying existing domains,
7095  * and it will not create the default domain.
7096  *
7097  * Call with hotplug lock held
7098  */
7099 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7100                              struct sched_domain_attr *dattr_new)
7101 {
7102         int i, j, n;
7103         int new_topology;
7104 
7105         mutex_lock(&sched_domains_mutex);
7106 
7107         /* always unregister in case we don't destroy any domains */
7108         unregister_sched_domain_sysctl();
7109 
7110         /* Let architecture update cpu core mappings. */
7111         new_topology = arch_update_cpu_topology();
7112 
7113         n = doms_new ? ndoms_new : 0;
7114 
7115         /* Destroy deleted domains */
7116         for (i = 0; i < ndoms_cur; i++) {
7117                 for (j = 0; j < n && !new_topology; j++) {
7118                         if (cpumask_equal(doms_cur[i], doms_new[j])
7119                             && dattrs_equal(dattr_cur, i, dattr_new, j))
7120                                 goto match1;
7121                 }
7122                 /* no match - a current sched domain not in new doms_new[] */
7123                 detach_destroy_domains(doms_cur[i]);
7124 match1:
7125                 ;
7126         }
7127 
7128         n = ndoms_cur;
7129         if (doms_new == NULL) {
7130                 n = 0;
7131                 doms_new = &fallback_doms;
7132                 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7133                 WARN_ON_ONCE(dattr_new);
7134         }
7135 
7136         /* Build new domains */
7137         for (i = 0; i < ndoms_new; i++) {
7138                 for (j = 0; j < n && !new_topology; j++) {
7139                         if (cpumask_equal(doms_new[i], doms_cur[j])
7140                             && dattrs_equal(dattr_new, i, dattr_cur, j))
7141                                 goto match2;
7142                 }
7143                 /* no match - add a new doms_new */
7144                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7145 match2:
7146                 ;
7147         }
7148 
7149         /* Remember the new sched domains */
7150         if (doms_cur != &fallback_doms)
7151                 free_sched_domains(doms_cur, ndoms_cur);
7152         kfree(dattr_cur);       /* kfree(NULL) is safe */
7153         doms_cur = doms_new;
7154         dattr_cur = dattr_new;
7155         ndoms_cur = ndoms_new;
7156 
7157         register_sched_domain_sysctl();
7158 
7159         mutex_unlock(&sched_domains_mutex);
7160 }
7161 
7162 static int num_cpus_frozen;     /* used to mark begin/end of suspend/resume */
7163 
7164 /*
7165  * Update cpusets according to cpu_active mask.  If cpusets are
7166  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7167  * around partition_sched_domains().
7168  *
7169  * If we come here as part of a suspend/resume, don't touch cpusets because we
7170  * want to restore it back to its original state upon resume anyway.
7171  */
7172 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7173                              void *hcpu)
7174 {
7175         switch (action) {
7176         case CPU_ONLINE_FROZEN:
7177         case CPU_DOWN_FAILED_FROZEN:
7178 
7179                 /*
7180                  * num_cpus_frozen tracks how many CPUs are involved in suspend
7181                  * resume sequence. As long as this is not the last online
7182                  * operation in the resume sequence, just build a single sched
7183                  * domain, ignoring cpusets.
7184                  */
7185                 num_cpus_frozen--;
7186                 if (likely(num_cpus_frozen)) {
7187                         partition_sched_domains(1, NULL, NULL);
7188                         break;
7189                 }
7190 
7191                 /*
7192                  * This is the last CPU online operation. So fall through and
7193                  * restore the original sched domains by considering the
7194                  * cpuset configurations.
7195                  */
7196 
7197         case CPU_ONLINE:
7198                 cpuset_update_active_cpus(true);
7199                 break;
7200         default:
7201                 return NOTIFY_DONE;
7202         }
7203         return NOTIFY_OK;
7204 }
7205 
7206 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7207                                void *hcpu)
7208 {
7209         unsigned long flags;
7210         long cpu = (long)hcpu;
7211         struct dl_bw *dl_b;
7212         bool overflow;
7213         int cpus;
7214 
7215         switch (action) {
7216         case CPU_DOWN_PREPARE:
7217                 rcu_read_lock_sched();
7218                 dl_b = dl_bw_of(cpu);
7219 
7220                 raw_spin_lock_irqsave(&dl_b->lock, flags);
7221                 cpus = dl_bw_cpus(cpu);
7222                 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7223                 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7224 
7225                 rcu_read_unlock_sched();
7226 
7227                 if (overflow)
7228                         return notifier_from_errno(-EBUSY);
7229                 cpuset_update_active_cpus(false);
7230                 break;
7231         case CPU_DOWN_PREPARE_FROZEN:
7232                 num_cpus_frozen++;
7233                 partition_sched_domains(1, NULL, NULL);
7234                 break;
7235         default:
7236                 return NOTIFY_DONE;
7237         }
7238         return NOTIFY_OK;
7239 }
7240 
7241 void __init sched_init_smp(void)
7242 {
7243         cpumask_var_t non_isolated_cpus;
7244 
7245         alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7246         alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7247 
7248         sched_init_numa();
7249 
7250         /*
7251          * There's no userspace yet to cause hotplug operations; hence all the
7252          * cpu masks are stable and all blatant races in the below code cannot
7253          * happen.
7254          */
7255         mutex_lock(&sched_domains_mutex);
7256         init_sched_domains(cpu_active_mask);
7257         cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7258         if (cpumask_empty(non_isolated_cpus))
7259                 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7260         mutex_unlock(&sched_domains_mutex);
7261 
7262         hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7263         hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7264         hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7265 
7266         init_hrtick();
7267 
7268         /* Move init over to a non-isolated CPU */
7269         if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7270                 BUG();
7271         sched_init_granularity();
7272         free_cpumask_var(non_isolated_cpus);
7273 
7274         init_sched_rt_class();
7275         init_sched_dl_class();
7276 }
7277 #else
7278 void __init sched_init_smp(void)
7279 {
7280         sched_init_granularity();
7281 }
7282 #endif /* CONFIG_SMP */
7283 
7284 int in_sched_functions(unsigned long addr)
7285 {
7286         return in_lock_functions(addr) ||
7287                 (addr >= (unsigned long)__sched_text_start
7288                 && addr < (unsigned long)__sched_text_end);
7289 }
7290 
7291 #ifdef CONFIG_CGROUP_SCHED
7292 /*
7293  * Default task group.
7294  * Every task in system belongs to this group at bootup.
7295  */
7296 struct task_group root_task_group;
7297 LIST_HEAD(task_groups);
7298 #endif
7299 
7300 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7301 
7302 void __init sched_init(void)
7303 {
7304         int i, j;
7305         unsigned long alloc_size = 0, ptr;
7306 
7307 #ifdef CONFIG_FAIR_GROUP_SCHED
7308         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7309 #endif
7310 #ifdef CONFIG_RT_GROUP_SCHED
7311         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7312 #endif
7313         if (alloc_size) {
7314                 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7315 
7316 #ifdef CONFIG_FAIR_GROUP_SCHED
7317                 root_task_group.se = (struct sched_entity **)ptr;
7318                 ptr += nr_cpu_ids * sizeof(void **);
7319 
7320                 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7321                 ptr += nr_cpu_ids * sizeof(void **);
7322 
7323 #endif /* CONFIG_FAIR_GROUP_SCHED */
7324 #ifdef CONFIG_RT_GROUP_SCHED
7325                 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7326                 ptr += nr_cpu_ids * sizeof(void **);
7327 
7328                 root_task_group.rt_rq = (struct rt_rq **)ptr;
7329                 ptr += nr_cpu_ids * sizeof(void **);
7330 
7331 #endif /* CONFIG_RT_GROUP_SCHED */
7332         }
7333 #ifdef CONFIG_CPUMASK_OFFSTACK
7334         for_each_possible_cpu(i) {
7335                 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7336                         cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7337         }
7338 #endif /* CONFIG_CPUMASK_OFFSTACK */
7339 
7340         init_rt_bandwidth(&def_rt_bandwidth,
7341                         global_rt_period(), global_rt_runtime());
7342         init_dl_bandwidth(&def_dl_bandwidth,
7343                         global_rt_period(), global_rt_runtime());
7344 
7345 #ifdef CONFIG_SMP
7346         init_defrootdomain();
7347 #endif
7348 
7349 #ifdef CONFIG_RT_GROUP_SCHED
7350         init_rt_bandwidth(&root_task_group.rt_bandwidth,
7351                         global_rt_period(), global_rt_runtime());
7352 #endif /* CONFIG_RT_GROUP_SCHED */
7353 
7354 #ifdef CONFIG_CGROUP_SCHED
7355         list_add(&root_task_group.list, &task_groups);
7356         INIT_LIST_HEAD(&root_task_group.children);
7357         INIT_LIST_HEAD(&root_task_group.siblings);
7358         autogroup_init(&init_task);
7359 
7360 #endif /* CONFIG_CGROUP_SCHED */
7361 
7362         for_each_possible_cpu(i) {
7363                 struct rq *rq;
7364 
7365                 rq = cpu_rq(i);
7366                 raw_spin_lock_init(&rq->lock);
7367                 rq->nr_running = 0;
7368                 rq->calc_load_active = 0;
7369                 rq->calc_load_update = jiffies + LOAD_FREQ;
7370                 init_cfs_rq(&rq->cfs);
7371                 init_rt_rq(&rq->rt);
7372                 init_dl_rq(&rq->dl);
7373 #ifdef CONFIG_FAIR_GROUP_SCHED
7374                 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7375                 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7376                 /*
7377                  * How much cpu bandwidth does root_task_group get?
7378                  *
7379                  * In case of task-groups formed thr' the cgroup filesystem, it
7380                  * gets 100% of the cpu resources in the system. This overall
7381                  * system cpu resource is divided among the tasks of
7382                  * root_task_group and its child task-groups in a fair manner,
7383                  * based on each entity's (task or task-group's) weight
7384                  * (se->load.weight).
7385                  *
7386                  * In other words, if root_task_group has 10 tasks of weight
7387                  * 1024) and two child groups A0 and A1 (of weight 1024 each),
7388                  * then A0's share of the cpu resource is:
7389                  *
7390                  *      A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7391                  *
7392                  * We achieve this by letting root_task_group's tasks sit
7393                  * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7394                  */
7395                 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7396                 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7397 #endif /* CONFIG_FAIR_GROUP_SCHED */
7398 
7399                 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7400 #ifdef CONFIG_RT_GROUP_SCHED
7401                 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7402 #endif
7403 
7404                 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7405                         rq->cpu_load[j] = 0;
7406 
7407                 rq->last_load_update_tick = jiffies;
7408 
7409 #ifdef CONFIG_SMP
7410                 rq->sd = NULL;
7411                 rq->rd = NULL;
7412                 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7413                 rq->balance_callback = NULL;
7414                 rq->active_balance = 0;
7415                 rq->next_balance = jiffies;
7416                 rq->push_cpu = 0;
7417                 rq->cpu = i;
7418                 rq->online = 0;
7419                 rq->idle_stamp = 0;
7420                 rq->avg_idle = 2*sysctl_sched_migration_cost;
7421                 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7422 
7423                 INIT_LIST_HEAD(&rq->cfs_tasks);
7424 
7425                 rq_attach_root(rq, &def_root_domain);
7426 #ifdef CONFIG_NO_HZ_COMMON
7427                 rq->nohz_flags = 0;
7428 #endif
7429 #ifdef CONFIG_NO_HZ_FULL
7430                 rq->last_sched_tick = 0;
7431 #endif
7432 #endif
7433                 init_rq_hrtick(rq);
7434                 atomic_set(&rq->nr_iowait, 0);
7435         }
7436 
7437         set_load_weight(&init_task);
7438 
7439 #ifdef CONFIG_PREEMPT_NOTIFIERS
7440         INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7441 #endif
7442 
7443         /*
7444          * The boot idle thread does lazy MMU switching as well:
7445          */
7446         atomic_inc(&init_mm.mm_count);
7447         enter_lazy_tlb(&init_mm, current);
7448 
7449         /*
7450          * During early bootup we pretend to be a normal task:
7451          */
7452         current->sched_class = &fair_sched_class;
7453 
7454         /*
7455          * Make us the idle thread. Technically, schedule() should not be
7456          * called from this thread, however somewhere below it might be,
7457          * but because we are the idle thread, we just pick up running again
7458          * when this runqueue becomes "idle".
7459          */
7460         init_idle(current, smp_processor_id());
7461 
7462         calc_load_update = jiffies + LOAD_FREQ;
7463 
7464 #ifdef CONFIG_SMP
7465         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7466         /* May be allocated at isolcpus cmdline parse time */
7467         if (cpu_isolated_map