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