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