Version:  2.0.40 2.2.26 2.4.37 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10

Linux/kernel/sched/core.c

  1 /*
  2  *  kernel/sched/core.c
  3  *
  4  *  Kernel scheduler and related syscalls
  5  *
  6  *  Copyright (C) 1991-2002  Linus Torvalds
  7  *
  8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
  9  *              make semaphores SMP safe
 10  *  1998-11-19  Implemented schedule_timeout() and related stuff
 11  *              by Andrea Arcangeli
 12  *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
 13  *              hybrid priority-list and round-robin design with
 14  *              an array-switch method of distributing timeslices
 15  *              and per-CPU runqueues.  Cleanups and useful suggestions
 16  *              by Davide Libenzi, preemptible kernel bits by Robert Love.
 17  *  2003-09-03  Interactivity tuning by Con Kolivas.
 18  *  2004-04-02  Scheduler domains code by Nick Piggin
 19  *  2007-04-15  Work begun on replacing all interactivity tuning with a
 20  *              fair scheduling design by Con Kolivas.
 21  *  2007-05-05  Load balancing (smp-nice) and other improvements
 22  *              by Peter Williams
 23  *  2007-05-06  Interactivity improvements to CFS by Mike Galbraith
 24  *  2007-07-01  Group scheduling enhancements by Srivatsa Vaddagiri
 25  *  2007-11-29  RT balancing improvements by Steven Rostedt, Gregory Haskins,
 26  *              Thomas Gleixner, Mike Kravetz
 27  */
 28 
 29 #include <linux/mm.h>
 30 #include <linux/module.h>
 31 #include <linux/nmi.h>
 32 #include <linux/init.h>
 33 #include <linux/uaccess.h>
 34 #include <linux/highmem.h>
 35 #include <asm/mmu_context.h>
 36 #include <linux/interrupt.h>
 37 #include <linux/capability.h>
 38 #include <linux/completion.h>
 39 #include <linux/kernel_stat.h>
 40 #include <linux/debug_locks.h>
 41 #include <linux/perf_event.h>
 42 #include <linux/security.h>
 43 #include <linux/notifier.h>
 44 #include <linux/profile.h>
 45 #include <linux/freezer.h>
 46 #include <linux/vmalloc.h>
 47 #include <linux/blkdev.h>
 48 #include <linux/delay.h>
 49 #include <linux/pid_namespace.h>
 50 #include <linux/smp.h>
 51 #include <linux/threads.h>
 52 #include <linux/timer.h>
 53 #include <linux/rcupdate.h>
 54 #include <linux/cpu.h>
 55 #include <linux/cpuset.h>
 56 #include <linux/percpu.h>
 57 #include <linux/proc_fs.h>
 58 #include <linux/seq_file.h>
 59 #include <linux/sysctl.h>
 60 #include <linux/syscalls.h>
 61 #include <linux/times.h>
 62 #include <linux/tsacct_kern.h>
 63 #include <linux/kprobes.h>
 64 #include <linux/delayacct.h>
 65 #include <linux/unistd.h>
 66 #include <linux/pagemap.h>
 67 #include <linux/hrtimer.h>
 68 #include <linux/tick.h>
 69 #include <linux/debugfs.h>
 70 #include <linux/ctype.h>
 71 #include <linux/ftrace.h>
 72 #include <linux/slab.h>
 73 #include <linux/init_task.h>
 74 #include <linux/binfmts.h>
 75 #include <linux/context_tracking.h>
 76 #include <linux/compiler.h>
 77 
 78 #include <asm/switch_to.h>
 79 #include <asm/tlb.h>
 80 #include <asm/irq_regs.h>
 81 #include <asm/mutex.h>
 82 #ifdef CONFIG_PARAVIRT
 83 #include <asm/paravirt.h>
 84 #endif
 85 
 86 #include "sched.h"
 87 #include "../workqueue_internal.h"
 88 #include "../smpboot.h"
 89 
 90 #define CREATE_TRACE_POINTS
 91 #include <trace/events/sched.h>
 92 
 93 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
 94 {
 95         unsigned long delta;
 96         ktime_t soft, hard, now;
 97 
 98         for (;;) {
 99                 if (hrtimer_active(period_timer))
100                         break;
101 
102                 now = hrtimer_cb_get_time(period_timer);
103                 hrtimer_forward(period_timer, now, period);
104 
105                 soft = hrtimer_get_softexpires(period_timer);
106                 hard = hrtimer_get_expires(period_timer);
107                 delta = ktime_to_ns(ktime_sub(hard, soft));
108                 __hrtimer_start_range_ns(period_timer, soft, delta,
109                                          HRTIMER_MODE_ABS_PINNED, 0);
110         }
111 }
112 
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 
118 void update_rq_clock(struct rq *rq)
119 {
120         s64 delta;
121 
122         lockdep_assert_held(&rq->lock);
123 
124         if (rq->clock_skip_update & RQCF_ACT_SKIP)
125                 return;
126 
127         delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
128         if (delta < 0)
129                 return;
130         rq->clock += delta;
131         update_rq_clock_task(rq, delta);
132 }
133 
134 /*
135  * Debugging: various feature bits
136  */
137 
138 #define SCHED_FEAT(name, enabled)       \
139         (1UL << __SCHED_FEAT_##name) * enabled |
140 
141 const_debug unsigned int sysctl_sched_features =
142 #include "features.h"
143         0;
144 
145 #undef SCHED_FEAT
146 
147 #ifdef CONFIG_SCHED_DEBUG
148 #define SCHED_FEAT(name, enabled)       \
149         #name ,
150 
151 static const char * const sched_feat_names[] = {
152 #include "features.h"
153 };
154 
155 #undef SCHED_FEAT
156 
157 static int sched_feat_show(struct seq_file *m, void *v)
158 {
159         int i;
160 
161         for (i = 0; i < __SCHED_FEAT_NR; i++) {
162                 if (!(sysctl_sched_features & (1UL << i)))
163                         seq_puts(m, "NO_");
164                 seq_printf(m, "%s ", sched_feat_names[i]);
165         }
166         seq_puts(m, "\n");
167 
168         return 0;
169 }
170 
171 #ifdef HAVE_JUMP_LABEL
172 
173 #define jump_label_key__true  STATIC_KEY_INIT_TRUE
174 #define jump_label_key__false STATIC_KEY_INIT_FALSE
175 
176 #define SCHED_FEAT(name, enabled)       \
177         jump_label_key__##enabled ,
178 
179 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
180 #include "features.h"
181 };
182 
183 #undef SCHED_FEAT
184 
185 static void sched_feat_disable(int i)
186 {
187         if (static_key_enabled(&sched_feat_keys[i]))
188                 static_key_slow_dec(&sched_feat_keys[i]);
189 }
190 
191 static void sched_feat_enable(int i)
192 {
193         if (!static_key_enabled(&sched_feat_keys[i]))
194                 static_key_slow_inc(&sched_feat_keys[i]);
195 }
196 #else
197 static void sched_feat_disable(int i) { };
198 static void sched_feat_enable(int i) { };
199 #endif /* HAVE_JUMP_LABEL */
200 
201 static int sched_feat_set(char *cmp)
202 {
203         int i;
204         int neg = 0;
205 
206         if (strncmp(cmp, "NO_", 3) == 0) {
207                 neg = 1;
208                 cmp += 3;
209         }
210 
211         for (i = 0; i < __SCHED_FEAT_NR; i++) {
212                 if (strcmp(cmp, sched_feat_names[i]) == 0) {
213                         if (neg) {
214                                 sysctl_sched_features &= ~(1UL << i);
215                                 sched_feat_disable(i);
216                         } else {
217                                 sysctl_sched_features |= (1UL << i);
218                                 sched_feat_enable(i);
219                         }
220                         break;
221                 }
222         }
223 
224         return i;
225 }
226 
227 static ssize_t
228 sched_feat_write(struct file *filp, const char __user *ubuf,
229                 size_t cnt, loff_t *ppos)
230 {
231         char buf[64];
232         char *cmp;
233         int i;
234         struct inode *inode;
235 
236         if (cnt > 63)
237                 cnt = 63;
238 
239         if (copy_from_user(&buf, ubuf, cnt))
240                 return -EFAULT;
241 
242         buf[cnt] = 0;
243         cmp = strstrip(buf);
244 
245         /* Ensure the static_key remains in a consistent state */
246         inode = file_inode(filp);
247         mutex_lock(&inode->i_mutex);
248         i = sched_feat_set(cmp);
249         mutex_unlock(&inode->i_mutex);
250         if (i == __SCHED_FEAT_NR)
251                 return -EINVAL;
252 
253         *ppos += cnt;
254 
255         return cnt;
256 }
257 
258 static int sched_feat_open(struct inode *inode, struct file *filp)
259 {
260         return single_open(filp, sched_feat_show, NULL);
261 }
262 
263 static const struct file_operations sched_feat_fops = {
264         .open           = sched_feat_open,
265         .write          = sched_feat_write,
266         .read           = seq_read,
267         .llseek         = seq_lseek,
268         .release        = single_release,
269 };
270 
271 static __init int sched_init_debug(void)
272 {
273         debugfs_create_file("sched_features", 0644, NULL, NULL,
274                         &sched_feat_fops);
275 
276         return 0;
277 }
278 late_initcall(sched_init_debug);
279 #endif /* CONFIG_SCHED_DEBUG */
280 
281 /*
282  * Number of tasks to iterate in a single balance run.
283  * Limited because this is done with IRQs disabled.
284  */
285 const_debug unsigned int sysctl_sched_nr_migrate = 32;
286 
287 /*
288  * period over which we average the RT time consumption, measured
289  * in ms.
290  *
291  * default: 1s
292  */
293 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
294 
295 /*
296  * period over which we measure -rt task cpu usage in us.
297  * default: 1s
298  */
299 unsigned int sysctl_sched_rt_period = 1000000;
300 
301 __read_mostly int scheduler_running;
302 
303 /*
304  * part of the period that we allow rt tasks to run in us.
305  * default: 0.95s
306  */
307 int sysctl_sched_rt_runtime = 950000;
308 
309 /*
310  * this_rq_lock - lock this runqueue and disable interrupts.
311  */
312 static struct rq *this_rq_lock(void)
313         __acquires(rq->lock)
314 {
315         struct rq *rq;
316 
317         local_irq_disable();
318         rq = this_rq();
319         raw_spin_lock(&rq->lock);
320 
321         return rq;
322 }
323 
324 #ifdef CONFIG_SCHED_HRTICK
325 /*
326  * Use HR-timers to deliver accurate preemption points.
327  */
328 
329 static void hrtick_clear(struct rq *rq)
330 {
331         if (hrtimer_active(&rq->hrtick_timer))
332                 hrtimer_cancel(&rq->hrtick_timer);
333 }
334 
335 /*
336  * High-resolution timer tick.
337  * Runs from hardirq context with interrupts disabled.
338  */
339 static enum hrtimer_restart hrtick(struct hrtimer *timer)
340 {
341         struct rq *rq = container_of(timer, struct rq, hrtick_timer);
342 
343         WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
344 
345         raw_spin_lock(&rq->lock);
346         update_rq_clock(rq);
347         rq->curr->sched_class->task_tick(rq, rq->curr, 1);
348         raw_spin_unlock(&rq->lock);
349 
350         return HRTIMER_NORESTART;
351 }
352 
353 #ifdef CONFIG_SMP
354 
355 static int __hrtick_restart(struct rq *rq)
356 {
357         struct hrtimer *timer = &rq->hrtick_timer;
358         ktime_t time = hrtimer_get_softexpires(timer);
359 
360         return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
361 }
362 
363 /*
364  * called from hardirq (IPI) context
365  */
366 static void __hrtick_start(void *arg)
367 {
368         struct rq *rq = arg;
369 
370         raw_spin_lock(&rq->lock);
371         __hrtick_restart(rq);
372         rq->hrtick_csd_pending = 0;
373         raw_spin_unlock(&rq->lock);
374 }
375 
376 /*
377  * Called to set the hrtick timer state.
378  *
379  * called with rq->lock held and irqs disabled
380  */
381 void hrtick_start(struct rq *rq, u64 delay)
382 {
383         struct hrtimer *timer = &rq->hrtick_timer;
384         ktime_t time;
385         s64 delta;
386 
387         /*
388          * Don't schedule slices shorter than 10000ns, that just
389          * doesn't make sense and can cause timer DoS.
390          */
391         delta = max_t(s64, delay, 10000LL);
392         time = ktime_add_ns(timer->base->get_time(), delta);
393 
394         hrtimer_set_expires(timer, time);
395 
396         if (rq == this_rq()) {
397                 __hrtick_restart(rq);
398         } else if (!rq->hrtick_csd_pending) {
399                 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
400                 rq->hrtick_csd_pending = 1;
401         }
402 }
403 
404 static int
405 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
406 {
407         int cpu = (int)(long)hcpu;
408 
409         switch (action) {
410         case CPU_UP_CANCELED:
411         case CPU_UP_CANCELED_FROZEN:
412         case CPU_DOWN_PREPARE:
413         case CPU_DOWN_PREPARE_FROZEN:
414         case CPU_DEAD:
415         case CPU_DEAD_FROZEN:
416                 hrtick_clear(cpu_rq(cpu));
417                 return NOTIFY_OK;
418         }
419 
420         return NOTIFY_DONE;
421 }
422 
423 static __init void init_hrtick(void)
424 {
425         hotcpu_notifier(hotplug_hrtick, 0);
426 }
427 #else
428 /*
429  * Called to set the hrtick timer state.
430  *
431  * called with rq->lock held and irqs disabled
432  */
433 void hrtick_start(struct rq *rq, u64 delay)
434 {
435         /*
436          * Don't schedule slices shorter than 10000ns, that just
437          * doesn't make sense. Rely on vruntime for fairness.
438          */
439         delay = max_t(u64, delay, 10000LL);
440         __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
441                         HRTIMER_MODE_REL_PINNED, 0);
442 }
443 
444 static inline void init_hrtick(void)
445 {
446 }
447 #endif /* CONFIG_SMP */
448 
449 static void init_rq_hrtick(struct rq *rq)
450 {
451 #ifdef CONFIG_SMP
452         rq->hrtick_csd_pending = 0;
453 
454         rq->hrtick_csd.flags = 0;
455         rq->hrtick_csd.func = __hrtick_start;
456         rq->hrtick_csd.info = rq;
457 #endif
458 
459         hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
460         rq->hrtick_timer.function = hrtick;
461 }
462 #else   /* CONFIG_SCHED_HRTICK */
463 static inline void hrtick_clear(struct rq *rq)
464 {
465 }
466 
467 static inline void init_rq_hrtick(struct rq *rq)
468 {
469 }
470 
471 static inline void init_hrtick(void)
472 {
473 }
474 #endif  /* CONFIG_SCHED_HRTICK */
475 
476 /*
477  * cmpxchg based fetch_or, macro so it works for different integer types
478  */
479 #define fetch_or(ptr, val)                                              \
480 ({      typeof(*(ptr)) __old, __val = *(ptr);                           \
481         for (;;) {                                                      \
482                 __old = cmpxchg((ptr), __val, __val | (val));           \
483                 if (__old == __val)                                     \
484                         break;                                          \
485                 __val = __old;                                          \
486         }                                                               \
487         __old;                                                          \
488 })
489 
490 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
491 /*
492  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
493  * this avoids any races wrt polling state changes and thereby avoids
494  * spurious IPIs.
495  */
496 static bool set_nr_and_not_polling(struct task_struct *p)
497 {
498         struct thread_info *ti = task_thread_info(p);
499         return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
500 }
501 
502 /*
503  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
504  *
505  * If this returns true, then the idle task promises to call
506  * sched_ttwu_pending() and reschedule soon.
507  */
508 static bool set_nr_if_polling(struct task_struct *p)
509 {
510         struct thread_info *ti = task_thread_info(p);
511         typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
512 
513         for (;;) {
514                 if (!(val & _TIF_POLLING_NRFLAG))
515                         return false;
516                 if (val & _TIF_NEED_RESCHED)
517                         return true;
518                 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
519                 if (old == val)
520                         break;
521                 val = old;
522         }
523         return true;
524 }
525 
526 #else
527 static bool set_nr_and_not_polling(struct task_struct *p)
528 {
529         set_tsk_need_resched(p);
530         return true;
531 }
532 
533 #ifdef CONFIG_SMP
534 static bool set_nr_if_polling(struct task_struct *p)
535 {
536         return false;
537 }
538 #endif
539 #endif
540 
541 /*
542  * resched_curr - mark rq's current task 'to be rescheduled now'.
543  *
544  * On UP this means the setting of the need_resched flag, on SMP it
545  * might also involve a cross-CPU call to trigger the scheduler on
546  * the target CPU.
547  */
548 void resched_curr(struct rq *rq)
549 {
550         struct task_struct *curr = rq->curr;
551         int cpu;
552 
553         lockdep_assert_held(&rq->lock);
554 
555         if (test_tsk_need_resched(curr))
556                 return;
557 
558         cpu = cpu_of(rq);
559 
560         if (cpu == smp_processor_id()) {
561                 set_tsk_need_resched(curr);
562                 set_preempt_need_resched();
563                 return;
564         }
565 
566         if (set_nr_and_not_polling(curr))
567                 smp_send_reschedule(cpu);
568         else
569                 trace_sched_wake_idle_without_ipi(cpu);
570 }
571 
572 void resched_cpu(int cpu)
573 {
574         struct rq *rq = cpu_rq(cpu);
575         unsigned long flags;
576 
577         if (!raw_spin_trylock_irqsave(&rq->lock, flags))
578                 return;
579         resched_curr(rq);
580         raw_spin_unlock_irqrestore(&rq->lock, flags);
581 }
582 
583 #ifdef CONFIG_SMP
584 #ifdef CONFIG_NO_HZ_COMMON
585 /*
586  * In the semi idle case, use the nearest busy cpu for migrating timers
587  * from an idle cpu.  This is good for power-savings.
588  *
589  * We don't do similar optimization for completely idle system, as
590  * selecting an idle cpu will add more delays to the timers than intended
591  * (as that cpu's timer base may not be uptodate wrt jiffies etc).
592  */
593 int get_nohz_timer_target(int pinned)
594 {
595         int cpu = smp_processor_id();
596         int i;
597         struct sched_domain *sd;
598 
599         if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
600                 return cpu;
601 
602         rcu_read_lock();
603         for_each_domain(cpu, sd) {
604                 for_each_cpu(i, sched_domain_span(sd)) {
605                         if (!idle_cpu(i)) {
606                                 cpu = i;
607                                 goto unlock;
608                         }
609                 }
610         }
611 unlock:
612         rcu_read_unlock();
613         return cpu;
614 }
615 /*
616  * When add_timer_on() enqueues a timer into the timer wheel of an
617  * idle CPU then this timer might expire before the next timer event
618  * which is scheduled to wake up that CPU. In case of a completely
619  * idle system the next event might even be infinite time into the
620  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
621  * leaves the inner idle loop so the newly added timer is taken into
622  * account when the CPU goes back to idle and evaluates the timer
623  * wheel for the next timer event.
624  */
625 static void wake_up_idle_cpu(int cpu)
626 {
627         struct rq *rq = cpu_rq(cpu);
628 
629         if (cpu == smp_processor_id())
630                 return;
631 
632         if (set_nr_and_not_polling(rq->idle))
633                 smp_send_reschedule(cpu);
634         else
635                 trace_sched_wake_idle_without_ipi(cpu);
636 }
637 
638 static bool wake_up_full_nohz_cpu(int cpu)
639 {
640         /*
641          * We just need the target to call irq_exit() and re-evaluate
642          * the next tick. The nohz full kick at least implies that.
643          * If needed we can still optimize that later with an
644          * empty IRQ.
645          */
646         if (tick_nohz_full_cpu(cpu)) {
647                 if (cpu != smp_processor_id() ||
648                     tick_nohz_tick_stopped())
649                         tick_nohz_full_kick_cpu(cpu);
650                 return true;
651         }
652 
653         return false;
654 }
655 
656 void wake_up_nohz_cpu(int cpu)
657 {
658         if (!wake_up_full_nohz_cpu(cpu))
659                 wake_up_idle_cpu(cpu);
660 }
661 
662 static inline bool got_nohz_idle_kick(void)
663 {
664         int cpu = smp_processor_id();
665 
666         if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
667                 return false;
668 
669         if (idle_cpu(cpu) && !need_resched())
670                 return true;
671 
672         /*
673          * We can't run Idle Load Balance on this CPU for this time so we
674          * cancel it and clear NOHZ_BALANCE_KICK
675          */
676         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
677         return false;
678 }
679 
680 #else /* CONFIG_NO_HZ_COMMON */
681 
682 static inline bool got_nohz_idle_kick(void)
683 {
684         return false;
685 }
686 
687 #endif /* CONFIG_NO_HZ_COMMON */
688 
689 #ifdef CONFIG_NO_HZ_FULL
690 bool sched_can_stop_tick(void)
691 {
692         /*
693          * More than one running task need preemption.
694          * nr_running update is assumed to be visible
695          * after IPI is sent from wakers.
696          */
697         if (this_rq()->nr_running > 1)
698                 return false;
699 
700         return true;
701 }
702 #endif /* CONFIG_NO_HZ_FULL */
703 
704 void sched_avg_update(struct rq *rq)
705 {
706         s64 period = sched_avg_period();
707 
708         while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
709                 /*
710                  * Inline assembly required to prevent the compiler
711                  * optimising this loop into a divmod call.
712                  * See __iter_div_u64_rem() for another example of this.
713                  */
714                 asm("" : "+rm" (rq->age_stamp));
715                 rq->age_stamp += period;
716                 rq->rt_avg /= 2;
717         }
718 }
719 
720 #endif /* CONFIG_SMP */
721 
722 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
723                         (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
724 /*
725  * Iterate task_group tree rooted at *from, calling @down when first entering a
726  * node and @up when leaving it for the final time.
727  *
728  * Caller must hold rcu_lock or sufficient equivalent.
729  */
730 int walk_tg_tree_from(struct task_group *from,
731                              tg_visitor down, tg_visitor up, void *data)
732 {
733         struct task_group *parent, *child;
734         int ret;
735 
736         parent = from;
737 
738 down:
739         ret = (*down)(parent, data);
740         if (ret)
741                 goto out;
742         list_for_each_entry_rcu(child, &parent->children, siblings) {
743                 parent = child;
744                 goto down;
745 
746 up:
747                 continue;
748         }
749         ret = (*up)(parent, data);
750         if (ret || parent == from)
751                 goto out;
752 
753         child = parent;
754         parent = parent->parent;
755         if (parent)
756                 goto up;
757 out:
758         return ret;
759 }
760 
761 int tg_nop(struct task_group *tg, void *data)
762 {
763         return 0;
764 }
765 #endif
766 
767 static void set_load_weight(struct task_struct *p)
768 {
769         int prio = p->static_prio - MAX_RT_PRIO;
770         struct load_weight *load = &p->se.load;
771 
772         /*
773          * SCHED_IDLE tasks get minimal weight:
774          */
775         if (p->policy == SCHED_IDLE) {
776                 load->weight = scale_load(WEIGHT_IDLEPRIO);
777                 load->inv_weight = WMULT_IDLEPRIO;
778                 return;
779         }
780 
781         load->weight = scale_load(prio_to_weight[prio]);
782         load->inv_weight = prio_to_wmult[prio];
783 }
784 
785 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
786 {
787         update_rq_clock(rq);
788         sched_info_queued(rq, p);
789         p->sched_class->enqueue_task(rq, p, flags);
790 }
791 
792 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
793 {
794         update_rq_clock(rq);
795         sched_info_dequeued(rq, p);
796         p->sched_class->dequeue_task(rq, p, flags);
797 }
798 
799 void activate_task(struct rq *rq, struct task_struct *p, int flags)
800 {
801         if (task_contributes_to_load(p))
802                 rq->nr_uninterruptible--;
803 
804         enqueue_task(rq, p, flags);
805 }
806 
807 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
808 {
809         if (task_contributes_to_load(p))
810                 rq->nr_uninterruptible++;
811 
812         dequeue_task(rq, p, flags);
813 }
814 
815 static void update_rq_clock_task(struct rq *rq, s64 delta)
816 {
817 /*
818  * In theory, the compile should just see 0 here, and optimize out the call
819  * to sched_rt_avg_update. But I don't trust it...
820  */
821 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
822         s64 steal = 0, irq_delta = 0;
823 #endif
824 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
825         irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
826 
827         /*
828          * Since irq_time is only updated on {soft,}irq_exit, we might run into
829          * this case when a previous update_rq_clock() happened inside a
830          * {soft,}irq region.
831          *
832          * When this happens, we stop ->clock_task and only update the
833          * prev_irq_time stamp to account for the part that fit, so that a next
834          * update will consume the rest. This ensures ->clock_task is
835          * monotonic.
836          *
837          * It does however cause some slight miss-attribution of {soft,}irq
838          * time, a more accurate solution would be to update the irq_time using
839          * the current rq->clock timestamp, except that would require using
840          * atomic ops.
841          */
842         if (irq_delta > delta)
843                 irq_delta = delta;
844 
845         rq->prev_irq_time += irq_delta;
846         delta -= irq_delta;
847 #endif
848 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
849         if (static_key_false((&paravirt_steal_rq_enabled))) {
850                 steal = paravirt_steal_clock(cpu_of(rq));
851                 steal -= rq->prev_steal_time_rq;
852 
853                 if (unlikely(steal > delta))
854                         steal = delta;
855 
856                 rq->prev_steal_time_rq += steal;
857                 delta -= steal;
858         }
859 #endif
860 
861         rq->clock_task += delta;
862 
863 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
864         if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
865                 sched_rt_avg_update(rq, irq_delta + steal);
866 #endif
867 }
868 
869 void sched_set_stop_task(int cpu, struct task_struct *stop)
870 {
871         struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
872         struct task_struct *old_stop = cpu_rq(cpu)->stop;
873 
874         if (stop) {
875                 /*
876                  * Make it appear like a SCHED_FIFO task, its something
877                  * userspace knows about and won't get confused about.
878                  *
879                  * Also, it will make PI more or less work without too
880                  * much confusion -- but then, stop work should not
881                  * rely on PI working anyway.
882                  */
883                 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
884 
885                 stop->sched_class = &stop_sched_class;
886         }
887 
888         cpu_rq(cpu)->stop = stop;
889 
890         if (old_stop) {
891                 /*
892                  * Reset it back to a normal scheduling class so that
893                  * it can die in pieces.
894                  */
895                 old_stop->sched_class = &rt_sched_class;
896         }
897 }
898 
899 /*
900  * __normal_prio - return the priority that is based on the static prio
901  */
902 static inline int __normal_prio(struct task_struct *p)
903 {
904         return p->static_prio;
905 }
906 
907 /*
908  * Calculate the expected normal priority: i.e. priority
909  * without taking RT-inheritance into account. Might be
910  * boosted by interactivity modifiers. Changes upon fork,
911  * setprio syscalls, and whenever the interactivity
912  * estimator recalculates.
913  */
914 static inline int normal_prio(struct task_struct *p)
915 {
916         int prio;
917 
918         if (task_has_dl_policy(p))
919                 prio = MAX_DL_PRIO-1;
920         else if (task_has_rt_policy(p))
921                 prio = MAX_RT_PRIO-1 - p->rt_priority;
922         else
923                 prio = __normal_prio(p);
924         return prio;
925 }
926 
927 /*
928  * Calculate the current priority, i.e. the priority
929  * taken into account by the scheduler. This value might
930  * be boosted by RT tasks, or might be boosted by
931  * interactivity modifiers. Will be RT if the task got
932  * RT-boosted. If not then it returns p->normal_prio.
933  */
934 static int effective_prio(struct task_struct *p)
935 {
936         p->normal_prio = normal_prio(p);
937         /*
938          * If we are RT tasks or we were boosted to RT priority,
939          * keep the priority unchanged. Otherwise, update priority
940          * to the normal priority:
941          */
942         if (!rt_prio(p->prio))
943                 return p->normal_prio;
944         return p->prio;
945 }
946 
947 /**
948  * task_curr - is this task currently executing on a CPU?
949  * @p: the task in question.
950  *
951  * Return: 1 if the task is currently executing. 0 otherwise.
952  */
953 inline int task_curr(const struct task_struct *p)
954 {
955         return cpu_curr(task_cpu(p)) == p;
956 }
957 
958 /*
959  * Can drop rq->lock because from sched_class::switched_from() methods drop it.
960  */
961 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
962                                        const struct sched_class *prev_class,
963                                        int oldprio)
964 {
965         if (prev_class != p->sched_class) {
966                 if (prev_class->switched_from)
967                         prev_class->switched_from(rq, p);
968                 /* Possble rq->lock 'hole'.  */
969                 p->sched_class->switched_to(rq, p);
970         } else if (oldprio != p->prio || dl_task(p))
971                 p->sched_class->prio_changed(rq, p, oldprio);
972 }
973 
974 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
975 {
976         const struct sched_class *class;
977 
978         if (p->sched_class == rq->curr->sched_class) {
979                 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
980         } else {
981                 for_each_class(class) {
982                         if (class == rq->curr->sched_class)
983                                 break;
984                         if (class == p->sched_class) {
985                                 resched_curr(rq);
986                                 break;
987                         }
988                 }
989         }
990 
991         /*
992          * A queue event has occurred, and we're going to schedule.  In
993          * this case, we can save a useless back to back clock update.
994          */
995         if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
996                 rq_clock_skip_update(rq, true);
997 }
998 
999 #ifdef CONFIG_SMP
1000 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1001 {
1002 #ifdef CONFIG_SCHED_DEBUG
1003         /*
1004          * We should never call set_task_cpu() on a blocked task,
1005          * ttwu() will sort out the placement.
1006          */
1007         WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1008                         !p->on_rq);
1009 
1010 #ifdef CONFIG_LOCKDEP
1011         /*
1012          * The caller should hold either p->pi_lock or rq->lock, when changing
1013          * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1014          *
1015          * sched_move_task() holds both and thus holding either pins the cgroup,
1016          * see task_group().
1017          *
1018          * Furthermore, all task_rq users should acquire both locks, see
1019          * task_rq_lock().
1020          */
1021         WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1022                                       lockdep_is_held(&task_rq(p)->lock)));
1023 #endif
1024 #endif
1025 
1026         trace_sched_migrate_task(p, new_cpu);
1027 
1028         if (task_cpu(p) != new_cpu) {
1029                 if (p->sched_class->migrate_task_rq)
1030                         p->sched_class->migrate_task_rq(p, new_cpu);
1031                 p->se.nr_migrations++;
1032                 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1033         }
1034 
1035         __set_task_cpu(p, new_cpu);
1036 }
1037 
1038 static void __migrate_swap_task(struct task_struct *p, int cpu)
1039 {
1040         if (task_on_rq_queued(p)) {
1041                 struct rq *src_rq, *dst_rq;
1042 
1043                 src_rq = task_rq(p);
1044                 dst_rq = cpu_rq(cpu);
1045 
1046                 deactivate_task(src_rq, p, 0);
1047                 set_task_cpu(p, cpu);
1048                 activate_task(dst_rq, p, 0);
1049                 check_preempt_curr(dst_rq, p, 0);
1050         } else {
1051                 /*
1052                  * Task isn't running anymore; make it appear like we migrated
1053                  * it before it went to sleep. This means on wakeup we make the
1054                  * previous cpu our targer instead of where it really is.
1055                  */
1056                 p->wake_cpu = cpu;
1057         }
1058 }
1059 
1060 struct migration_swap_arg {
1061         struct task_struct *src_task, *dst_task;
1062         int src_cpu, dst_cpu;
1063 };
1064 
1065 static int migrate_swap_stop(void *data)
1066 {
1067         struct migration_swap_arg *arg = data;
1068         struct rq *src_rq, *dst_rq;
1069         int ret = -EAGAIN;
1070 
1071         src_rq = cpu_rq(arg->src_cpu);
1072         dst_rq = cpu_rq(arg->dst_cpu);
1073 
1074         double_raw_lock(&arg->src_task->pi_lock,
1075                         &arg->dst_task->pi_lock);
1076         double_rq_lock(src_rq, dst_rq);
1077         if (task_cpu(arg->dst_task) != arg->dst_cpu)
1078                 goto unlock;
1079 
1080         if (task_cpu(arg->src_task) != arg->src_cpu)
1081                 goto unlock;
1082 
1083         if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1084                 goto unlock;
1085 
1086         if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1087                 goto unlock;
1088 
1089         __migrate_swap_task(arg->src_task, arg->dst_cpu);
1090         __migrate_swap_task(arg->dst_task, arg->src_cpu);
1091 
1092         ret = 0;
1093 
1094 unlock:
1095         double_rq_unlock(src_rq, dst_rq);
1096         raw_spin_unlock(&arg->dst_task->pi_lock);
1097         raw_spin_unlock(&arg->src_task->pi_lock);
1098 
1099         return ret;
1100 }
1101 
1102 /*
1103  * Cross migrate two tasks
1104  */
1105 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1106 {
1107         struct migration_swap_arg arg;
1108         int ret = -EINVAL;
1109 
1110         arg = (struct migration_swap_arg){
1111                 .src_task = cur,
1112                 .src_cpu = task_cpu(cur),
1113                 .dst_task = p,
1114                 .dst_cpu = task_cpu(p),
1115         };
1116 
1117         if (arg.src_cpu == arg.dst_cpu)
1118                 goto out;
1119 
1120         /*
1121          * These three tests are all lockless; this is OK since all of them
1122          * will be re-checked with proper locks held further down the line.
1123          */
1124         if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1125                 goto out;
1126 
1127         if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1128                 goto out;
1129 
1130         if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1131                 goto out;
1132 
1133         trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1134         ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1135 
1136 out:
1137         return ret;
1138 }
1139 
1140 struct migration_arg {
1141         struct task_struct *task;
1142         int dest_cpu;
1143 };
1144 
1145 static int migration_cpu_stop(void *data);
1146 
1147 /*
1148  * wait_task_inactive - wait for a thread to unschedule.
1149  *
1150  * If @match_state is nonzero, it's the @p->state value just checked and
1151  * not expected to change.  If it changes, i.e. @p might have woken up,
1152  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1153  * we return a positive number (its total switch count).  If a second call
1154  * a short while later returns the same number, the caller can be sure that
1155  * @p has remained unscheduled the whole time.
1156  *
1157  * The caller must ensure that the task *will* unschedule sometime soon,
1158  * else this function might spin for a *long* time. This function can't
1159  * be called with interrupts off, or it may introduce deadlock with
1160  * smp_call_function() if an IPI is sent by the same process we are
1161  * waiting to become inactive.
1162  */
1163 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1164 {
1165         unsigned long flags;
1166         int running, queued;
1167         unsigned long ncsw;
1168         struct rq *rq;
1169 
1170         for (;;) {
1171                 /*
1172                  * We do the initial early heuristics without holding
1173                  * any task-queue locks at all. We'll only try to get
1174                  * the runqueue lock when things look like they will
1175                  * work out!
1176                  */
1177                 rq = task_rq(p);
1178 
1179                 /*
1180                  * If the task is actively running on another CPU
1181                  * still, just relax and busy-wait without holding
1182                  * any locks.
1183                  *
1184                  * NOTE! Since we don't hold any locks, it's not
1185                  * even sure that "rq" stays as the right runqueue!
1186                  * But we don't care, since "task_running()" will
1187                  * return false if the runqueue has changed and p
1188                  * is actually now running somewhere else!
1189                  */
1190                 while (task_running(rq, p)) {
1191                         if (match_state && unlikely(p->state != match_state))
1192                                 return 0;
1193                         cpu_relax();
1194                 }
1195 
1196                 /*
1197                  * Ok, time to look more closely! We need the rq
1198                  * lock now, to be *sure*. If we're wrong, we'll
1199                  * just go back and repeat.
1200                  */
1201                 rq = task_rq_lock(p, &flags);
1202                 trace_sched_wait_task(p);
1203                 running = task_running(rq, p);
1204                 queued = task_on_rq_queued(p);
1205                 ncsw = 0;
1206                 if (!match_state || p->state == match_state)
1207                         ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1208                 task_rq_unlock(rq, p, &flags);
1209 
1210                 /*
1211                  * If it changed from the expected state, bail out now.
1212                  */
1213                 if (unlikely(!ncsw))
1214                         break;
1215 
1216                 /*
1217                  * Was it really running after all now that we
1218                  * checked with the proper locks actually held?
1219                  *
1220                  * Oops. Go back and try again..
1221                  */
1222                 if (unlikely(running)) {
1223                         cpu_relax();
1224                         continue;
1225                 }
1226 
1227                 /*
1228                  * It's not enough that it's not actively running,
1229                  * it must be off the runqueue _entirely_, and not
1230                  * preempted!
1231                  *
1232                  * So if it was still runnable (but just not actively
1233                  * running right now), it's preempted, and we should
1234                  * yield - it could be a while.
1235                  */
1236                 if (unlikely(queued)) {
1237                         ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1238 
1239                         set_current_state(TASK_UNINTERRUPTIBLE);
1240                         schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1241                         continue;
1242                 }
1243 
1244                 /*
1245                  * Ahh, all good. It wasn't running, and it wasn't
1246                  * runnable, which means that it will never become
1247                  * running in the future either. We're all done!
1248                  */
1249                 break;
1250         }
1251 
1252         return ncsw;
1253 }
1254 
1255 /***
1256  * kick_process - kick a running thread to enter/exit the kernel
1257  * @p: the to-be-kicked thread
1258  *
1259  * Cause a process which is running on another CPU to enter
1260  * kernel-mode, without any delay. (to get signals handled.)
1261  *
1262  * NOTE: this function doesn't have to take the runqueue lock,
1263  * because all it wants to ensure is that the remote task enters
1264  * the kernel. If the IPI races and the task has been migrated
1265  * to another CPU then no harm is done and the purpose has been
1266  * achieved as well.
1267  */
1268 void kick_process(struct task_struct *p)
1269 {
1270         int cpu;
1271 
1272         preempt_disable();
1273         cpu = task_cpu(p);
1274         if ((cpu != smp_processor_id()) && task_curr(p))
1275                 smp_send_reschedule(cpu);
1276         preempt_enable();
1277 }
1278 EXPORT_SYMBOL_GPL(kick_process);
1279 #endif /* CONFIG_SMP */
1280 
1281 #ifdef CONFIG_SMP
1282 /*
1283  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1284  */
1285 static int select_fallback_rq(int cpu, struct task_struct *p)
1286 {
1287         int nid = cpu_to_node(cpu);
1288         const struct cpumask *nodemask = NULL;
1289         enum { cpuset, possible, fail } state = cpuset;
1290         int dest_cpu;
1291 
1292         /*
1293          * If the node that the cpu is on has been offlined, cpu_to_node()
1294          * will return -1. There is no cpu on the node, and we should
1295          * select the cpu on the other node.
1296          */
1297         if (nid != -1) {
1298                 nodemask = cpumask_of_node(nid);
1299 
1300                 /* Look for allowed, online CPU in same node. */
1301                 for_each_cpu(dest_cpu, nodemask) {
1302                         if (!cpu_online(dest_cpu))
1303                                 continue;
1304                         if (!cpu_active(dest_cpu))
1305                                 continue;
1306                         if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1307                                 return dest_cpu;
1308                 }
1309         }
1310 
1311         for (;;) {
1312                 /* Any allowed, online CPU? */
1313                 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1314                         if (!cpu_online(dest_cpu))
1315                                 continue;
1316                         if (!cpu_active(dest_cpu))
1317                                 continue;
1318                         goto out;
1319                 }
1320 
1321                 switch (state) {
1322                 case cpuset:
1323                         /* No more Mr. Nice Guy. */
1324                         cpuset_cpus_allowed_fallback(p);
1325                         state = possible;
1326                         break;
1327 
1328                 case possible:
1329                         do_set_cpus_allowed(p, cpu_possible_mask);
1330                         state = fail;
1331                         break;
1332 
1333                 case fail:
1334                         BUG();
1335                         break;
1336                 }
1337         }
1338 
1339 out:
1340         if (state != cpuset) {
1341                 /*
1342                  * Don't tell them about moving exiting tasks or
1343                  * kernel threads (both mm NULL), since they never
1344                  * leave kernel.
1345                  */
1346                 if (p->mm && printk_ratelimit()) {
1347                         printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1348                                         task_pid_nr(p), p->comm, cpu);
1349                 }
1350         }
1351 
1352         return dest_cpu;
1353 }
1354 
1355 /*
1356  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1357  */
1358 static inline
1359 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1360 {
1361         if (p->nr_cpus_allowed > 1)
1362                 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1363 
1364         /*
1365          * In order not to call set_task_cpu() on a blocking task we need
1366          * to rely on ttwu() to place the task on a valid ->cpus_allowed
1367          * cpu.
1368          *
1369          * Since this is common to all placement strategies, this lives here.
1370          *
1371          * [ this allows ->select_task() to simply return task_cpu(p) and
1372          *   not worry about this generic constraint ]
1373          */
1374         if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1375                      !cpu_online(cpu)))
1376                 cpu = select_fallback_rq(task_cpu(p), p);
1377 
1378         return cpu;
1379 }
1380 
1381 static void update_avg(u64 *avg, u64 sample)
1382 {
1383         s64 diff = sample - *avg;
1384         *avg += diff >> 3;
1385 }
1386 #endif
1387 
1388 static void
1389 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1390 {
1391 #ifdef CONFIG_SCHEDSTATS
1392         struct rq *rq = this_rq();
1393 
1394 #ifdef CONFIG_SMP
1395         int this_cpu = smp_processor_id();
1396 
1397         if (cpu == this_cpu) {
1398                 schedstat_inc(rq, ttwu_local);
1399                 schedstat_inc(p, se.statistics.nr_wakeups_local);
1400         } else {
1401                 struct sched_domain *sd;
1402 
1403                 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1404                 rcu_read_lock();
1405                 for_each_domain(this_cpu, sd) {
1406                         if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1407                                 schedstat_inc(sd, ttwu_wake_remote);
1408                                 break;
1409                         }
1410                 }
1411                 rcu_read_unlock();
1412         }
1413 
1414         if (wake_flags & WF_MIGRATED)
1415                 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1416 
1417 #endif /* CONFIG_SMP */
1418 
1419         schedstat_inc(rq, ttwu_count);
1420         schedstat_inc(p, se.statistics.nr_wakeups);
1421 
1422         if (wake_flags & WF_SYNC)
1423                 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1424 
1425 #endif /* CONFIG_SCHEDSTATS */
1426 }
1427 
1428 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1429 {
1430         activate_task(rq, p, en_flags);
1431         p->on_rq = TASK_ON_RQ_QUEUED;
1432 
1433         /* if a worker is waking up, notify workqueue */
1434         if (p->flags & PF_WQ_WORKER)
1435                 wq_worker_waking_up(p, cpu_of(rq));
1436 }
1437 
1438 /*
1439  * Mark the task runnable and perform wakeup-preemption.
1440  */
1441 static void
1442 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1443 {
1444         check_preempt_curr(rq, p, wake_flags);
1445         trace_sched_wakeup(p, true);
1446 
1447         p->state = TASK_RUNNING;
1448 #ifdef CONFIG_SMP
1449         if (p->sched_class->task_woken)
1450                 p->sched_class->task_woken(rq, p);
1451 
1452         if (rq->idle_stamp) {
1453                 u64 delta = rq_clock(rq) - rq->idle_stamp;
1454                 u64 max = 2*rq->max_idle_balance_cost;
1455 
1456                 update_avg(&rq->avg_idle, delta);
1457 
1458                 if (rq->avg_idle > max)
1459                         rq->avg_idle = max;
1460 
1461                 rq->idle_stamp = 0;
1462         }
1463 #endif
1464 }
1465 
1466 static void
1467 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1468 {
1469 #ifdef CONFIG_SMP
1470         if (p->sched_contributes_to_load)
1471                 rq->nr_uninterruptible--;
1472 #endif
1473 
1474         ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1475         ttwu_do_wakeup(rq, p, wake_flags);
1476 }
1477 
1478 /*
1479  * Called in case the task @p isn't fully descheduled from its runqueue,
1480  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1481  * since all we need to do is flip p->state to TASK_RUNNING, since
1482  * the task is still ->on_rq.
1483  */
1484 static int ttwu_remote(struct task_struct *p, int wake_flags)
1485 {
1486         struct rq *rq;
1487         int ret = 0;
1488 
1489         rq = __task_rq_lock(p);
1490         if (task_on_rq_queued(p)) {
1491                 /* check_preempt_curr() may use rq clock */
1492                 update_rq_clock(rq);
1493                 ttwu_do_wakeup(rq, p, wake_flags);
1494                 ret = 1;
1495         }
1496         __task_rq_unlock(rq);
1497 
1498         return ret;
1499 }
1500 
1501 #ifdef CONFIG_SMP
1502 void sched_ttwu_pending(void)
1503 {
1504         struct rq *rq = this_rq();
1505         struct llist_node *llist = llist_del_all(&rq->wake_list);
1506         struct task_struct *p;
1507         unsigned long flags;
1508 
1509         if (!llist)
1510                 return;
1511 
1512         raw_spin_lock_irqsave(&rq->lock, flags);
1513 
1514         while (llist) {
1515                 p = llist_entry(llist, struct task_struct, wake_entry);
1516                 llist = llist_next(llist);
1517                 ttwu_do_activate(rq, p, 0);
1518         }
1519 
1520         raw_spin_unlock_irqrestore(&rq->lock, flags);
1521 }
1522 
1523 void scheduler_ipi(void)
1524 {
1525         /*
1526          * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1527          * TIF_NEED_RESCHED remotely (for the first time) will also send
1528          * this IPI.
1529          */
1530         preempt_fold_need_resched();
1531 
1532         if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1533                 return;
1534 
1535         /*
1536          * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1537          * traditionally all their work was done from the interrupt return
1538          * path. Now that we actually do some work, we need to make sure
1539          * we do call them.
1540          *
1541          * Some archs already do call them, luckily irq_enter/exit nest
1542          * properly.
1543          *
1544          * Arguably we should visit all archs and update all handlers,
1545          * however a fair share of IPIs are still resched only so this would
1546          * somewhat pessimize the simple resched case.
1547          */
1548         irq_enter();
1549         sched_ttwu_pending();
1550 
1551         /*
1552          * Check if someone kicked us for doing the nohz idle load balance.
1553          */
1554         if (unlikely(got_nohz_idle_kick())) {
1555                 this_rq()->idle_balance = 1;
1556                 raise_softirq_irqoff(SCHED_SOFTIRQ);
1557         }
1558         irq_exit();
1559 }
1560 
1561 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1562 {
1563         struct rq *rq = cpu_rq(cpu);
1564 
1565         if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1566                 if (!set_nr_if_polling(rq->idle))
1567                         smp_send_reschedule(cpu);
1568                 else
1569                         trace_sched_wake_idle_without_ipi(cpu);
1570         }
1571 }
1572 
1573 void wake_up_if_idle(int cpu)
1574 {
1575         struct rq *rq = cpu_rq(cpu);
1576         unsigned long flags;
1577 
1578         rcu_read_lock();
1579 
1580         if (!is_idle_task(rcu_dereference(rq->curr)))
1581                 goto out;
1582 
1583         if (set_nr_if_polling(rq->idle)) {
1584                 trace_sched_wake_idle_without_ipi(cpu);
1585         } else {
1586                 raw_spin_lock_irqsave(&rq->lock, flags);
1587                 if (is_idle_task(rq->curr))
1588                         smp_send_reschedule(cpu);
1589                 /* Else cpu is not in idle, do nothing here */
1590                 raw_spin_unlock_irqrestore(&rq->lock, flags);
1591         }
1592 
1593 out:
1594         rcu_read_unlock();
1595 }
1596 
1597 bool cpus_share_cache(int this_cpu, int that_cpu)
1598 {
1599         return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1600 }
1601 #endif /* CONFIG_SMP */
1602 
1603 static void ttwu_queue(struct task_struct *p, int cpu)
1604 {
1605         struct rq *rq = cpu_rq(cpu);
1606 
1607 #if defined(CONFIG_SMP)
1608         if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1609                 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1610                 ttwu_queue_remote(p, cpu);
1611                 return;
1612         }
1613 #endif
1614 
1615         raw_spin_lock(&rq->lock);
1616         ttwu_do_activate(rq, p, 0);
1617         raw_spin_unlock(&rq->lock);
1618 }
1619 
1620 /**
1621  * try_to_wake_up - wake up a thread
1622  * @p: the thread to be awakened
1623  * @state: the mask of task states that can be woken
1624  * @wake_flags: wake modifier flags (WF_*)
1625  *
1626  * Put it on the run-queue if it's not already there. The "current"
1627  * thread is always on the run-queue (except when the actual
1628  * re-schedule is in progress), and as such you're allowed to do
1629  * the simpler "current->state = TASK_RUNNING" to mark yourself
1630  * runnable without the overhead of this.
1631  *
1632  * Return: %true if @p was woken up, %false if it was already running.
1633  * or @state didn't match @p's state.
1634  */
1635 static int
1636 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1637 {
1638         unsigned long flags;
1639         int cpu, success = 0;
1640 
1641         /*
1642          * If we are going to wake up a thread waiting for CONDITION we
1643          * need to ensure that CONDITION=1 done by the caller can not be
1644          * reordered with p->state check below. This pairs with mb() in
1645          * set_current_state() the waiting thread does.
1646          */
1647         smp_mb__before_spinlock();
1648         raw_spin_lock_irqsave(&p->pi_lock, flags);
1649         if (!(p->state & state))
1650                 goto out;
1651 
1652         success = 1; /* we're going to change ->state */
1653         cpu = task_cpu(p);
1654 
1655         if (p->on_rq && ttwu_remote(p, wake_flags))
1656                 goto stat;
1657 
1658 #ifdef CONFIG_SMP
1659         /*
1660          * If the owning (remote) cpu is still in the middle of schedule() with
1661          * this task as prev, wait until its done referencing the task.
1662          */
1663         while (p->on_cpu)
1664                 cpu_relax();
1665         /*
1666          * Pairs with the smp_wmb() in finish_lock_switch().
1667          */
1668         smp_rmb();
1669 
1670         p->sched_contributes_to_load = !!task_contributes_to_load(p);
1671         p->state = TASK_WAKING;
1672 
1673         if (p->sched_class->task_waking)
1674                 p->sched_class->task_waking(p);
1675 
1676         cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1677         if (task_cpu(p) != cpu) {
1678                 wake_flags |= WF_MIGRATED;
1679                 set_task_cpu(p, cpu);
1680         }
1681 #endif /* CONFIG_SMP */
1682 
1683         ttwu_queue(p, cpu);
1684 stat:
1685         ttwu_stat(p, cpu, wake_flags);
1686 out:
1687         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1688 
1689         return success;
1690 }
1691 
1692 /**
1693  * try_to_wake_up_local - try to wake up a local task with rq lock held
1694  * @p: the thread to be awakened
1695  *
1696  * Put @p on the run-queue if it's not already there. The caller must
1697  * ensure that this_rq() is locked, @p is bound to this_rq() and not
1698  * the current task.
1699  */
1700 static void try_to_wake_up_local(struct task_struct *p)
1701 {
1702         struct rq *rq = task_rq(p);
1703 
1704         if (WARN_ON_ONCE(rq != this_rq()) ||
1705             WARN_ON_ONCE(p == current))
1706                 return;
1707 
1708         lockdep_assert_held(&rq->lock);
1709 
1710         if (!raw_spin_trylock(&p->pi_lock)) {
1711                 raw_spin_unlock(&rq->lock);
1712                 raw_spin_lock(&p->pi_lock);
1713                 raw_spin_lock(&rq->lock);
1714         }
1715 
1716         if (!(p->state & TASK_NORMAL))
1717                 goto out;
1718 
1719         if (!task_on_rq_queued(p))
1720                 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1721 
1722         ttwu_do_wakeup(rq, p, 0);
1723         ttwu_stat(p, smp_processor_id(), 0);
1724 out:
1725         raw_spin_unlock(&p->pi_lock);
1726 }
1727 
1728 /**
1729  * wake_up_process - Wake up a specific process
1730  * @p: The process to be woken up.
1731  *
1732  * Attempt to wake up the nominated process and move it to the set of runnable
1733  * processes.
1734  *
1735  * Return: 1 if the process was woken up, 0 if it was already running.
1736  *
1737  * It may be assumed that this function implies a write memory barrier before
1738  * changing the task state if and only if any tasks are woken up.
1739  */
1740 int wake_up_process(struct task_struct *p)
1741 {
1742         WARN_ON(task_is_stopped_or_traced(p));
1743         return try_to_wake_up(p, TASK_NORMAL, 0);
1744 }
1745 EXPORT_SYMBOL(wake_up_process);
1746 
1747 int wake_up_state(struct task_struct *p, unsigned int state)
1748 {
1749         return try_to_wake_up(p, state, 0);
1750 }
1751 
1752 /*
1753  * This function clears the sched_dl_entity static params.
1754  */
1755 void __dl_clear_params(struct task_struct *p)
1756 {
1757         struct sched_dl_entity *dl_se = &p->dl;
1758 
1759         dl_se->dl_runtime = 0;
1760         dl_se->dl_deadline = 0;
1761         dl_se->dl_period = 0;
1762         dl_se->flags = 0;
1763         dl_se->dl_bw = 0;
1764 
1765         dl_se->dl_throttled = 0;
1766         dl_se->dl_new = 1;
1767         dl_se->dl_yielded = 0;
1768 }
1769 
1770 /*
1771  * Perform scheduler related setup for a newly forked process p.
1772  * p is forked by current.
1773  *
1774  * __sched_fork() is basic setup used by init_idle() too:
1775  */
1776 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1777 {
1778         p->on_rq                        = 0;
1779 
1780         p->se.on_rq                     = 0;
1781         p->se.exec_start                = 0;
1782         p->se.sum_exec_runtime          = 0;
1783         p->se.prev_sum_exec_runtime     = 0;
1784         p->se.nr_migrations             = 0;
1785         p->se.vruntime                  = 0;
1786 #ifdef CONFIG_SMP
1787         p->se.avg.decay_count           = 0;
1788 #endif
1789         INIT_LIST_HEAD(&p->se.group_node);
1790 
1791 #ifdef CONFIG_SCHEDSTATS
1792         memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1793 #endif
1794 
1795         RB_CLEAR_NODE(&p->dl.rb_node);
1796         init_dl_task_timer(&p->dl);
1797         __dl_clear_params(p);
1798 
1799         INIT_LIST_HEAD(&p->rt.run_list);
1800 
1801 #ifdef CONFIG_PREEMPT_NOTIFIERS
1802         INIT_HLIST_HEAD(&p->preempt_notifiers);
1803 #endif
1804 
1805 #ifdef CONFIG_NUMA_BALANCING
1806         if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1807                 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1808                 p->mm->numa_scan_seq = 0;
1809         }
1810 
1811         if (clone_flags & CLONE_VM)
1812                 p->numa_preferred_nid = current->numa_preferred_nid;
1813         else
1814                 p->numa_preferred_nid = -1;
1815 
1816         p->node_stamp = 0ULL;
1817         p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1818         p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1819         p->numa_work.next = &p->numa_work;
1820         p->numa_faults = NULL;
1821         p->last_task_numa_placement = 0;
1822         p->last_sum_exec_runtime = 0;
1823 
1824         p->numa_group = NULL;
1825 #endif /* CONFIG_NUMA_BALANCING */
1826 }
1827 
1828 #ifdef CONFIG_NUMA_BALANCING
1829 #ifdef CONFIG_SCHED_DEBUG
1830 void set_numabalancing_state(bool enabled)
1831 {
1832         if (enabled)
1833                 sched_feat_set("NUMA");
1834         else
1835                 sched_feat_set("NO_NUMA");
1836 }
1837 #else
1838 __read_mostly bool numabalancing_enabled;
1839 
1840 void set_numabalancing_state(bool enabled)
1841 {
1842         numabalancing_enabled = enabled;
1843 }
1844 #endif /* CONFIG_SCHED_DEBUG */
1845 
1846 #ifdef CONFIG_PROC_SYSCTL
1847 int sysctl_numa_balancing(struct ctl_table *table, int write,
1848                          void __user *buffer, size_t *lenp, loff_t *ppos)
1849 {
1850         struct ctl_table t;
1851         int err;
1852         int state = numabalancing_enabled;
1853 
1854         if (write && !capable(CAP_SYS_ADMIN))
1855                 return -EPERM;
1856 
1857         t = *table;
1858         t.data = &state;
1859         err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1860         if (err < 0)
1861                 return err;
1862         if (write)
1863                 set_numabalancing_state(state);
1864         return err;
1865 }
1866 #endif
1867 #endif
1868 
1869 /*
1870  * fork()/clone()-time setup:
1871  */
1872 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1873 {
1874         unsigned long flags;
1875         int cpu = get_cpu();
1876 
1877         __sched_fork(clone_flags, p);
1878         /*
1879          * We mark the process as running here. This guarantees that
1880          * nobody will actually run it, and a signal or other external
1881          * event cannot wake it up and insert it on the runqueue either.
1882          */
1883         p->state = TASK_RUNNING;
1884 
1885         /*
1886          * Make sure we do not leak PI boosting priority to the child.
1887          */
1888         p->prio = current->normal_prio;
1889 
1890         /*
1891          * Revert to default priority/policy on fork if requested.
1892          */
1893         if (unlikely(p->sched_reset_on_fork)) {
1894                 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1895                         p->policy = SCHED_NORMAL;
1896                         p->static_prio = NICE_TO_PRIO(0);
1897                         p->rt_priority = 0;
1898                 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1899                         p->static_prio = NICE_TO_PRIO(0);
1900 
1901                 p->prio = p->normal_prio = __normal_prio(p);
1902                 set_load_weight(p);
1903 
1904                 /*
1905                  * We don't need the reset flag anymore after the fork. It has
1906                  * fulfilled its duty:
1907                  */
1908                 p->sched_reset_on_fork = 0;
1909         }
1910 
1911         if (dl_prio(p->prio)) {
1912                 put_cpu();
1913                 return -EAGAIN;
1914         } else if (rt_prio(p->prio)) {
1915                 p->sched_class = &rt_sched_class;
1916         } else {
1917                 p->sched_class = &fair_sched_class;
1918         }
1919 
1920         if (p->sched_class->task_fork)
1921                 p->sched_class->task_fork(p);
1922 
1923         /*
1924          * The child is not yet in the pid-hash so no cgroup attach races,
1925          * and the cgroup is pinned to this child due to cgroup_fork()
1926          * is ran before sched_fork().
1927          *
1928          * Silence PROVE_RCU.
1929          */
1930         raw_spin_lock_irqsave(&p->pi_lock, flags);
1931         set_task_cpu(p, cpu);
1932         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1933 
1934 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1935         if (likely(sched_info_on()))
1936                 memset(&p->sched_info, 0, sizeof(p->sched_info));
1937 #endif
1938 #if defined(CONFIG_SMP)
1939         p->on_cpu = 0;
1940 #endif
1941         init_task_preempt_count(p);
1942 #ifdef CONFIG_SMP
1943         plist_node_init(&p->pushable_tasks, MAX_PRIO);
1944         RB_CLEAR_NODE(&p->pushable_dl_tasks);
1945 #endif
1946 
1947         put_cpu();
1948         return 0;
1949 }
1950 
1951 unsigned long to_ratio(u64 period, u64 runtime)
1952 {
1953         if (runtime == RUNTIME_INF)
1954                 return 1ULL << 20;
1955 
1956         /*
1957          * Doing this here saves a lot of checks in all
1958          * the calling paths, and returning zero seems
1959          * safe for them anyway.
1960          */
1961         if (period == 0)
1962                 return 0;
1963 
1964         return div64_u64(runtime << 20, period);
1965 }
1966 
1967 #ifdef CONFIG_SMP
1968 inline struct dl_bw *dl_bw_of(int i)
1969 {
1970         rcu_lockdep_assert(rcu_read_lock_sched_held(),
1971                            "sched RCU must be held");
1972         return &cpu_rq(i)->rd->dl_bw;
1973 }
1974 
1975 static inline int dl_bw_cpus(int i)
1976 {
1977         struct root_domain *rd = cpu_rq(i)->rd;
1978         int cpus = 0;
1979 
1980         rcu_lockdep_assert(rcu_read_lock_sched_held(),
1981                            "sched RCU must be held");
1982         for_each_cpu_and(i, rd->span, cpu_active_mask)
1983                 cpus++;
1984 
1985         return cpus;
1986 }
1987 #else
1988 inline struct dl_bw *dl_bw_of(int i)
1989 {
1990         return &cpu_rq(i)->dl.dl_bw;
1991 }
1992 
1993 static inline int dl_bw_cpus(int i)
1994 {
1995         return 1;
1996 }
1997 #endif
1998 
1999 /*
2000  * We must be sure that accepting a new task (or allowing changing the
2001  * parameters of an existing one) is consistent with the bandwidth
2002  * constraints. If yes, this function also accordingly updates the currently
2003  * allocated bandwidth to reflect the new situation.
2004  *
2005  * This function is called while holding p's rq->lock.
2006  *
2007  * XXX we should delay bw change until the task's 0-lag point, see
2008  * __setparam_dl().
2009  */
2010 static int dl_overflow(struct task_struct *p, int policy,
2011                        const struct sched_attr *attr)
2012 {
2013 
2014         struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2015         u64 period = attr->sched_period ?: attr->sched_deadline;
2016         u64 runtime = attr->sched_runtime;
2017         u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2018         int cpus, err = -1;
2019 
2020         if (new_bw == p->dl.dl_bw)
2021                 return 0;
2022 
2023         /*
2024          * Either if a task, enters, leave, or stays -deadline but changes
2025          * its parameters, we may need to update accordingly the total
2026          * allocated bandwidth of the container.
2027          */
2028         raw_spin_lock(&dl_b->lock);
2029         cpus = dl_bw_cpus(task_cpu(p));
2030         if (dl_policy(policy) && !task_has_dl_policy(p) &&
2031             !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2032                 __dl_add(dl_b, new_bw);
2033                 err = 0;
2034         } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2035                    !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2036                 __dl_clear(dl_b, p->dl.dl_bw);
2037                 __dl_add(dl_b, new_bw);
2038                 err = 0;
2039         } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2040                 __dl_clear(dl_b, p->dl.dl_bw);
2041                 err = 0;
2042         }
2043         raw_spin_unlock(&dl_b->lock);
2044 
2045         return err;
2046 }
2047 
2048 extern void init_dl_bw(struct dl_bw *dl_b);
2049 
2050 /*
2051  * wake_up_new_task - wake up a newly created task for the first time.
2052  *
2053  * This function will do some initial scheduler statistics housekeeping
2054  * that must be done for every newly created context, then puts the task
2055  * on the runqueue and wakes it.
2056  */
2057 void wake_up_new_task(struct task_struct *p)
2058 {
2059         unsigned long flags;
2060         struct rq *rq;
2061 
2062         raw_spin_lock_irqsave(&p->pi_lock, flags);
2063 #ifdef CONFIG_SMP
2064         /*
2065          * Fork balancing, do it here and not earlier because:
2066          *  - cpus_allowed can change in the fork path
2067          *  - any previously selected cpu might disappear through hotplug
2068          */
2069         set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2070 #endif
2071 
2072         /* Initialize new task's runnable average */
2073         init_task_runnable_average(p);
2074         rq = __task_rq_lock(p);
2075         activate_task(rq, p, 0);
2076         p->on_rq = TASK_ON_RQ_QUEUED;
2077         trace_sched_wakeup_new(p, true);
2078         check_preempt_curr(rq, p, WF_FORK);
2079 #ifdef CONFIG_SMP
2080         if (p->sched_class->task_woken)
2081                 p->sched_class->task_woken(rq, p);
2082 #endif
2083         task_rq_unlock(rq, p, &flags);
2084 }
2085 
2086 #ifdef CONFIG_PREEMPT_NOTIFIERS
2087 
2088 /**
2089  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2090  * @notifier: notifier struct to register
2091  */
2092 void preempt_notifier_register(struct preempt_notifier *notifier)
2093 {
2094         hlist_add_head(&notifier->link, &current->preempt_notifiers);
2095 }
2096 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2097 
2098 /**
2099  * preempt_notifier_unregister - no longer interested in preemption notifications
2100  * @notifier: notifier struct to unregister
2101  *
2102  * This is safe to call from within a preemption notifier.
2103  */
2104 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2105 {
2106         hlist_del(&notifier->link);
2107 }
2108 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2109 
2110 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2111 {
2112         struct preempt_notifier *notifier;
2113 
2114         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2115                 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2116 }
2117 
2118 static void
2119 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2120                                  struct task_struct *next)
2121 {
2122         struct preempt_notifier *notifier;
2123 
2124         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2125                 notifier->ops->sched_out(notifier, next);
2126 }
2127 
2128 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2129 
2130 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2131 {
2132 }
2133 
2134 static void
2135 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2136                                  struct task_struct *next)
2137 {
2138 }
2139 
2140 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2141 
2142 /**
2143  * prepare_task_switch - prepare to switch tasks
2144  * @rq: the runqueue preparing to switch
2145  * @prev: the current task that is being switched out
2146  * @next: the task we are going to switch to.
2147  *
2148  * This is called with the rq lock held and interrupts off. It must
2149  * be paired with a subsequent finish_task_switch after the context
2150  * switch.
2151  *
2152  * prepare_task_switch sets up locking and calls architecture specific
2153  * hooks.
2154  */
2155 static inline void
2156 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2157                     struct task_struct *next)
2158 {
2159         trace_sched_switch(prev, next);
2160         sched_info_switch(rq, prev, next);
2161         perf_event_task_sched_out(prev, next);
2162         fire_sched_out_preempt_notifiers(prev, next);
2163         prepare_lock_switch(rq, next);
2164         prepare_arch_switch(next);
2165 }
2166 
2167 /**
2168  * finish_task_switch - clean up after a task-switch
2169  * @prev: the thread we just switched away from.
2170  *
2171  * finish_task_switch must be called after the context switch, paired
2172  * with a prepare_task_switch call before the context switch.
2173  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2174  * and do any other architecture-specific cleanup actions.
2175  *
2176  * Note that we may have delayed dropping an mm in context_switch(). If
2177  * so, we finish that here outside of the runqueue lock. (Doing it
2178  * with the lock held can cause deadlocks; see schedule() for
2179  * details.)
2180  *
2181  * The context switch have flipped the stack from under us and restored the
2182  * local variables which were saved when this task called schedule() in the
2183  * past. prev == current is still correct but we need to recalculate this_rq
2184  * because prev may have moved to another CPU.
2185  */
2186 static struct rq *finish_task_switch(struct task_struct *prev)
2187         __releases(rq->lock)
2188 {
2189         struct rq *rq = this_rq();
2190         struct mm_struct *mm = rq->prev_mm;
2191         long prev_state;
2192 
2193         rq->prev_mm = NULL;
2194 
2195         /*
2196          * A task struct has one reference for the use as "current".
2197          * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2198          * schedule one last time. The schedule call will never return, and
2199          * the scheduled task must drop that reference.
2200          * The test for TASK_DEAD must occur while the runqueue locks are
2201          * still held, otherwise prev could be scheduled on another cpu, die
2202          * there before we look at prev->state, and then the reference would
2203          * be dropped twice.
2204          *              Manfred Spraul <manfred@colorfullife.com>
2205          */
2206         prev_state = prev->state;
2207         vtime_task_switch(prev);
2208         finish_arch_switch(prev);
2209         perf_event_task_sched_in(prev, current);
2210         finish_lock_switch(rq, prev);
2211         finish_arch_post_lock_switch();
2212 
2213         fire_sched_in_preempt_notifiers(current);
2214         if (mm)
2215                 mmdrop(mm);
2216         if (unlikely(prev_state == TASK_DEAD)) {
2217                 if (prev->sched_class->task_dead)
2218                         prev->sched_class->task_dead(prev);
2219 
2220                 /*
2221                  * Remove function-return probe instances associated with this
2222                  * task and put them back on the free list.
2223                  */
2224                 kprobe_flush_task(prev);
2225                 put_task_struct(prev);
2226         }
2227 
2228         tick_nohz_task_switch(current);
2229         return rq;
2230 }
2231 
2232 #ifdef CONFIG_SMP
2233 
2234 /* rq->lock is NOT held, but preemption is disabled */
2235 static inline void post_schedule(struct rq *rq)
2236 {
2237         if (rq->post_schedule) {
2238                 unsigned long flags;
2239 
2240                 raw_spin_lock_irqsave(&rq->lock, flags);
2241                 if (rq->curr->sched_class->post_schedule)
2242                         rq->curr->sched_class->post_schedule(rq);
2243                 raw_spin_unlock_irqrestore(&rq->lock, flags);
2244 
2245                 rq->post_schedule = 0;
2246         }
2247 }
2248 
2249 #else
2250 
2251 static inline void post_schedule(struct rq *rq)
2252 {
2253 }
2254 
2255 #endif
2256 
2257 /**
2258  * schedule_tail - first thing a freshly forked thread must call.
2259  * @prev: the thread we just switched away from.
2260  */
2261 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2262         __releases(rq->lock)
2263 {
2264         struct rq *rq;
2265 
2266         /* finish_task_switch() drops rq->lock and enables preemtion */
2267         preempt_disable();
2268         rq = finish_task_switch(prev);
2269         post_schedule(rq);
2270         preempt_enable();
2271 
2272         if (current->set_child_tid)
2273                 put_user(task_pid_vnr(current), current->set_child_tid);
2274 }
2275 
2276 /*
2277  * context_switch - switch to the new MM and the new thread's register state.
2278  */
2279 static inline struct rq *
2280 context_switch(struct rq *rq, struct task_struct *prev,
2281                struct task_struct *next)
2282 {
2283         struct mm_struct *mm, *oldmm;
2284 
2285         prepare_task_switch(rq, prev, next);
2286 
2287         mm = next->mm;
2288         oldmm = prev->active_mm;
2289         /*
2290          * For paravirt, this is coupled with an exit in switch_to to
2291          * combine the page table reload and the switch backend into
2292          * one hypercall.
2293          */
2294         arch_start_context_switch(prev);
2295 
2296         if (!mm) {
2297                 next->active_mm = oldmm;
2298                 atomic_inc(&oldmm->mm_count);
2299                 enter_lazy_tlb(oldmm, next);
2300         } else
2301                 switch_mm(oldmm, mm, next);
2302 
2303         if (!prev->mm) {
2304                 prev->active_mm = NULL;
2305                 rq->prev_mm = oldmm;
2306         }
2307         /*
2308          * Since the runqueue lock will be released by the next
2309          * task (which is an invalid locking op but in the case
2310          * of the scheduler it's an obvious special-case), so we
2311          * do an early lockdep release here:
2312          */
2313         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2314 
2315         context_tracking_task_switch(prev, next);
2316         /* Here we just switch the register state and the stack. */
2317         switch_to(prev, next, prev);
2318         barrier();
2319 
2320         return finish_task_switch(prev);
2321 }
2322 
2323 /*
2324  * nr_running and nr_context_switches:
2325  *
2326  * externally visible scheduler statistics: current number of runnable
2327  * threads, total number of context switches performed since bootup.
2328  */
2329 unsigned long nr_running(void)
2330 {
2331         unsigned long i, sum = 0;
2332 
2333         for_each_online_cpu(i)
2334                 sum += cpu_rq(i)->nr_running;
2335 
2336         return sum;
2337 }
2338 
2339 /*
2340  * Check if only the current task is running on the cpu.
2341  */
2342 bool single_task_running(void)
2343 {
2344         if (cpu_rq(smp_processor_id())->nr_running == 1)
2345                 return true;
2346         else
2347                 return false;
2348 }
2349 EXPORT_SYMBOL(single_task_running);
2350 
2351 unsigned long long nr_context_switches(void)
2352 {
2353         int i;
2354         unsigned long long sum = 0;
2355 
2356         for_each_possible_cpu(i)
2357                 sum += cpu_rq(i)->nr_switches;
2358 
2359         return sum;
2360 }
2361 
2362 unsigned long nr_iowait(void)
2363 {
2364         unsigned long i, sum = 0;
2365 
2366         for_each_possible_cpu(i)
2367                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2368 
2369         return sum;
2370 }
2371 
2372 unsigned long nr_iowait_cpu(int cpu)
2373 {
2374         struct rq *this = cpu_rq(cpu);
2375         return atomic_read(&this->nr_iowait);
2376 }
2377 
2378 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2379 {
2380         struct rq *this = this_rq();
2381         *nr_waiters = atomic_read(&this->nr_iowait);
2382         *load = this->cpu_load[0];
2383 }
2384 
2385 #ifdef CONFIG_SMP
2386 
2387 /*
2388  * sched_exec - execve() is a valuable balancing opportunity, because at
2389  * this point the task has the smallest effective memory and cache footprint.
2390  */
2391 void sched_exec(void)
2392 {
2393         struct task_struct *p = current;
2394         unsigned long flags;
2395         int dest_cpu;
2396 
2397         raw_spin_lock_irqsave(&p->pi_lock, flags);
2398         dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2399         if (dest_cpu == smp_processor_id())
2400                 goto unlock;
2401 
2402         if (likely(cpu_active(dest_cpu))) {
2403                 struct migration_arg arg = { p, dest_cpu };
2404 
2405                 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2406                 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2407                 return;
2408         }
2409 unlock:
2410         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2411 }
2412 
2413 #endif
2414 
2415 DEFINE_PER_CPU(struct kernel_stat, kstat);
2416 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2417 
2418 EXPORT_PER_CPU_SYMBOL(kstat);
2419 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2420 
2421 /*
2422  * Return accounted runtime for the task.
2423  * In case the task is currently running, return the runtime plus current's
2424  * pending runtime that have not been accounted yet.
2425  */
2426 unsigned long long task_sched_runtime(struct task_struct *p)
2427 {
2428         unsigned long flags;
2429         struct rq *rq;
2430         u64 ns;
2431 
2432 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2433         /*
2434          * 64-bit doesn't need locks to atomically read a 64bit value.
2435          * So we have a optimization chance when the task's delta_exec is 0.
2436          * Reading ->on_cpu is racy, but this is ok.
2437          *
2438          * If we race with it leaving cpu, we'll take a lock. So we're correct.
2439          * If we race with it entering cpu, unaccounted time is 0. This is
2440          * indistinguishable from the read occurring a few cycles earlier.
2441          * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2442          * been accounted, so we're correct here as well.
2443          */
2444         if (!p->on_cpu || !task_on_rq_queued(p))
2445                 return p->se.sum_exec_runtime;
2446 #endif
2447 
2448         rq = task_rq_lock(p, &flags);
2449         /*
2450          * Must be ->curr _and_ ->on_rq.  If dequeued, we would
2451          * project cycles that may never be accounted to this
2452          * thread, breaking clock_gettime().
2453          */
2454         if (task_current(rq, p) && task_on_rq_queued(p)) {
2455                 update_rq_clock(rq);
2456                 p->sched_class->update_curr(rq);
2457         }
2458         ns = p->se.sum_exec_runtime;
2459         task_rq_unlock(rq, p, &flags);
2460 
2461         return ns;
2462 }
2463 
2464 /*
2465  * This function gets called by the timer code, with HZ frequency.
2466  * We call it with interrupts disabled.
2467  */
2468 void scheduler_tick(void)
2469 {
2470         int cpu = smp_processor_id();
2471         struct rq *rq = cpu_rq(cpu);
2472         struct task_struct *curr = rq->curr;
2473 
2474         sched_clock_tick();
2475 
2476         raw_spin_lock(&rq->lock);
2477         update_rq_clock(rq);
2478         curr->sched_class->task_tick(rq, curr, 0);
2479         update_cpu_load_active(rq);
2480         raw_spin_unlock(&rq->lock);
2481 
2482         perf_event_task_tick();
2483 
2484 #ifdef CONFIG_SMP
2485         rq->idle_balance = idle_cpu(cpu);
2486         trigger_load_balance(rq);
2487 #endif
2488         rq_last_tick_reset(rq);
2489 }
2490 
2491 #ifdef CONFIG_NO_HZ_FULL
2492 /**
2493  * scheduler_tick_max_deferment
2494  *
2495  * Keep at least one tick per second when a single
2496  * active task is running because the scheduler doesn't
2497  * yet completely support full dynticks environment.
2498  *
2499  * This makes sure that uptime, CFS vruntime, load
2500  * balancing, etc... continue to move forward, even
2501  * with a very low granularity.
2502  *
2503  * Return: Maximum deferment in nanoseconds.
2504  */
2505 u64 scheduler_tick_max_deferment(void)
2506 {
2507         struct rq *rq = this_rq();
2508         unsigned long next, now = ACCESS_ONCE(jiffies);
2509 
2510         next = rq->last_sched_tick + HZ;
2511 
2512         if (time_before_eq(next, now))
2513                 return 0;
2514 
2515         return jiffies_to_nsecs(next - now);
2516 }
2517 #endif
2518 
2519 notrace unsigned long get_parent_ip(unsigned long addr)
2520 {
2521         if (in_lock_functions(addr)) {
2522                 addr = CALLER_ADDR2;
2523                 if (in_lock_functions(addr))
2524                         addr = CALLER_ADDR3;
2525         }
2526         return addr;
2527 }
2528 
2529 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2530                                 defined(CONFIG_PREEMPT_TRACER))
2531 
2532 void preempt_count_add(int val)
2533 {
2534 #ifdef CONFIG_DEBUG_PREEMPT
2535         /*
2536          * Underflow?
2537          */
2538         if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2539                 return;
2540 #endif
2541         __preempt_count_add(val);
2542 #ifdef CONFIG_DEBUG_PREEMPT
2543         /*
2544          * Spinlock count overflowing soon?
2545          */
2546         DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2547                                 PREEMPT_MASK - 10);
2548 #endif
2549         if (preempt_count() == val) {
2550                 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2551 #ifdef CONFIG_DEBUG_PREEMPT
2552                 current->preempt_disable_ip = ip;
2553 #endif
2554                 trace_preempt_off(CALLER_ADDR0, ip);
2555         }
2556 }
2557 EXPORT_SYMBOL(preempt_count_add);
2558 NOKPROBE_SYMBOL(preempt_count_add);
2559 
2560 void preempt_count_sub(int val)
2561 {
2562 #ifdef CONFIG_DEBUG_PREEMPT
2563         /*
2564          * Underflow?
2565          */
2566         if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2567                 return;
2568         /*
2569          * Is the spinlock portion underflowing?
2570          */
2571         if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2572                         !(preempt_count() & PREEMPT_MASK)))
2573                 return;
2574 #endif
2575 
2576         if (preempt_count() == val)
2577                 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2578         __preempt_count_sub(val);
2579 }
2580 EXPORT_SYMBOL(preempt_count_sub);
2581 NOKPROBE_SYMBOL(preempt_count_sub);
2582 
2583 #endif
2584 
2585 /*
2586  * Print scheduling while atomic bug:
2587  */
2588 static noinline void __schedule_bug(struct task_struct *prev)
2589 {
2590         if (oops_in_progress)
2591                 return;
2592 
2593         printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2594                 prev->comm, prev->pid, preempt_count());
2595 
2596         debug_show_held_locks(prev);
2597         print_modules();
2598         if (irqs_disabled())
2599                 print_irqtrace_events(prev);
2600 #ifdef CONFIG_DEBUG_PREEMPT
2601         if (in_atomic_preempt_off()) {
2602                 pr_err("Preemption disabled at:");
2603                 print_ip_sym(current->preempt_disable_ip);
2604                 pr_cont("\n");
2605         }
2606 #endif
2607         dump_stack();
2608         add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2609 }
2610 
2611 /*
2612  * Various schedule()-time debugging checks and statistics:
2613  */
2614 static inline void schedule_debug(struct task_struct *prev)
2615 {
2616 #ifdef CONFIG_SCHED_STACK_END_CHECK
2617         BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2618 #endif
2619         /*
2620          * Test if we are atomic. Since do_exit() needs to call into
2621          * schedule() atomically, we ignore that path. Otherwise whine
2622          * if we are scheduling when we should not.
2623          */
2624         if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2625                 __schedule_bug(prev);
2626         rcu_sleep_check();
2627 
2628         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2629 
2630         schedstat_inc(this_rq(), sched_count);
2631 }
2632 
2633 /*
2634  * Pick up the highest-prio task:
2635  */
2636 static inline struct task_struct *
2637 pick_next_task(struct rq *rq, struct task_struct *prev)
2638 {
2639         const struct sched_class *class = &fair_sched_class;
2640         struct task_struct *p;
2641 
2642         /*
2643          * Optimization: we know that if all tasks are in
2644          * the fair class we can call that function directly:
2645          */
2646         if (likely(prev->sched_class == class &&
2647                    rq->nr_running == rq->cfs.h_nr_running)) {
2648                 p = fair_sched_class.pick_next_task(rq, prev);
2649                 if (unlikely(p == RETRY_TASK))
2650                         goto again;
2651 
2652                 /* assumes fair_sched_class->next == idle_sched_class */
2653                 if (unlikely(!p))
2654                         p = idle_sched_class.pick_next_task(rq, prev);
2655 
2656                 return p;
2657         }
2658 
2659 again:
2660         for_each_class(class) {
2661                 p = class->pick_next_task(rq, prev);
2662                 if (p) {
2663                         if (unlikely(p == RETRY_TASK))
2664                                 goto again;
2665                         return p;
2666                 }
2667         }
2668 
2669         BUG(); /* the idle class will always have a runnable task */
2670 }
2671 
2672 /*
2673  * __schedule() is the main scheduler function.
2674  *
2675  * The main means of driving the scheduler and thus entering this function are:
2676  *
2677  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2678  *
2679  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2680  *      paths. For example, see arch/x86/entry_64.S.
2681  *
2682  *      To drive preemption between tasks, the scheduler sets the flag in timer
2683  *      interrupt handler scheduler_tick().
2684  *
2685  *   3. Wakeups don't really cause entry into schedule(). They add a
2686  *      task to the run-queue and that's it.
2687  *
2688  *      Now, if the new task added to the run-queue preempts the current
2689  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2690  *      called on the nearest possible occasion:
2691  *
2692  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2693  *
2694  *         - in syscall or exception context, at the next outmost
2695  *           preempt_enable(). (this might be as soon as the wake_up()'s
2696  *           spin_unlock()!)
2697  *
2698  *         - in IRQ context, return from interrupt-handler to
2699  *           preemptible context
2700  *
2701  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2702  *         then at the next:
2703  *
2704  *          - cond_resched() call
2705  *          - explicit schedule() call
2706  *          - return from syscall or exception to user-space
2707  *          - return from interrupt-handler to user-space
2708  *
2709  * WARNING: all callers must re-check need_resched() afterward and reschedule
2710  * accordingly in case an event triggered the need for rescheduling (such as
2711  * an interrupt waking up a task) while preemption was disabled in __schedule().
2712  */
2713 static void __sched __schedule(void)
2714 {
2715         struct task_struct *prev, *next;
2716         unsigned long *switch_count;
2717         struct rq *rq;
2718         int cpu;
2719 
2720         preempt_disable();
2721         cpu = smp_processor_id();
2722         rq = cpu_rq(cpu);
2723         rcu_note_context_switch();
2724         prev = rq->curr;
2725 
2726         schedule_debug(prev);
2727 
2728         if (sched_feat(HRTICK))
2729                 hrtick_clear(rq);
2730 
2731         /*
2732          * Make sure that signal_pending_state()->signal_pending() below
2733          * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2734          * done by the caller to avoid the race with signal_wake_up().
2735          */
2736         smp_mb__before_spinlock();
2737         raw_spin_lock_irq(&rq->lock);
2738 
2739         rq->clock_skip_update <<= 1; /* promote REQ to ACT */
2740 
2741         switch_count = &prev->nivcsw;
2742         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2743                 if (unlikely(signal_pending_state(prev->state, prev))) {
2744                         prev->state = TASK_RUNNING;
2745                 } else {
2746                         deactivate_task(rq, prev, DEQUEUE_SLEEP);
2747                         prev->on_rq = 0;
2748 
2749                         /*
2750                          * If a worker went to sleep, notify and ask workqueue
2751                          * whether it wants to wake up a task to maintain
2752                          * concurrency.
2753                          */
2754                         if (prev->flags & PF_WQ_WORKER) {
2755                                 struct task_struct *to_wakeup;
2756 
2757                                 to_wakeup = wq_worker_sleeping(prev, cpu);
2758                                 if (to_wakeup)
2759                                         try_to_wake_up_local(to_wakeup);
2760                         }
2761                 }
2762                 switch_count = &prev->nvcsw;
2763         }
2764 
2765         if (task_on_rq_queued(prev))
2766                 update_rq_clock(rq);
2767 
2768         next = pick_next_task(rq, prev);
2769         clear_tsk_need_resched(prev);
2770         clear_preempt_need_resched();
2771         rq->clock_skip_update = 0;
2772 
2773         if (likely(prev != next)) {
2774                 rq->nr_switches++;
2775                 rq->curr = next;
2776                 ++*switch_count;
2777 
2778                 rq = context_switch(rq, prev, next); /* unlocks the rq */
2779                 cpu = cpu_of(rq);
2780         } else
2781                 raw_spin_unlock_irq(&rq->lock);
2782 
2783         post_schedule(rq);
2784 
2785         sched_preempt_enable_no_resched();
2786 }
2787 
2788 static inline void sched_submit_work(struct task_struct *tsk)
2789 {
2790         if (!tsk->state || tsk_is_pi_blocked(tsk))
2791                 return;
2792         /*
2793          * If we are going to sleep and we have plugged IO queued,
2794          * make sure to submit it to avoid deadlocks.
2795          */
2796         if (blk_needs_flush_plug(tsk))
2797                 blk_schedule_flush_plug(tsk);
2798 }
2799 
2800 asmlinkage __visible void __sched schedule(void)
2801 {
2802         struct task_struct *tsk = current;
2803 
2804         sched_submit_work(tsk);
2805         do {
2806                 __schedule();
2807         } while (need_resched());
2808 }
2809 EXPORT_SYMBOL(schedule);
2810 
2811 #ifdef CONFIG_CONTEXT_TRACKING
2812 asmlinkage __visible void __sched schedule_user(void)
2813 {
2814         /*
2815          * If we come here after a random call to set_need_resched(),
2816          * or we have been woken up remotely but the IPI has not yet arrived,
2817          * we haven't yet exited the RCU idle mode. Do it here manually until
2818          * we find a better solution.
2819          *
2820          * NB: There are buggy callers of this function.  Ideally we
2821          * should warn if prev_state != IN_USER, but that will trigger
2822          * too frequently to make sense yet.
2823          */
2824         enum ctx_state prev_state = exception_enter();
2825         schedule();
2826         exception_exit(prev_state);
2827 }
2828 #endif
2829 
2830 /**
2831  * schedule_preempt_disabled - called with preemption disabled
2832  *
2833  * Returns with preemption disabled. Note: preempt_count must be 1
2834  */
2835 void __sched schedule_preempt_disabled(void)
2836 {
2837         sched_preempt_enable_no_resched();
2838         schedule();
2839         preempt_disable();
2840 }
2841 
2842 static void __sched notrace preempt_schedule_common(void)
2843 {
2844         do {
2845                 __preempt_count_add(PREEMPT_ACTIVE);
2846                 __schedule();
2847                 __preempt_count_sub(PREEMPT_ACTIVE);
2848 
2849                 /*
2850                  * Check again in case we missed a preemption opportunity
2851                  * between schedule and now.
2852                  */
2853                 barrier();
2854         } while (need_resched());
2855 }
2856 
2857 #ifdef CONFIG_PREEMPT
2858 /*
2859  * this is the entry point to schedule() from in-kernel preemption
2860  * off of preempt_enable. Kernel preemptions off return from interrupt
2861  * occur there and call schedule directly.
2862  */
2863 asmlinkage __visible void __sched notrace preempt_schedule(void)
2864 {
2865         /*
2866          * If there is a non-zero preempt_count or interrupts are disabled,
2867          * we do not want to preempt the current task. Just return..
2868          */
2869         if (likely(!preemptible()))
2870                 return;
2871 
2872         preempt_schedule_common();
2873 }
2874 NOKPROBE_SYMBOL(preempt_schedule);
2875 EXPORT_SYMBOL(preempt_schedule);
2876 
2877 #ifdef CONFIG_CONTEXT_TRACKING
2878 /**
2879  * preempt_schedule_context - preempt_schedule called by tracing
2880  *
2881  * The tracing infrastructure uses preempt_enable_notrace to prevent
2882  * recursion and tracing preempt enabling caused by the tracing
2883  * infrastructure itself. But as tracing can happen in areas coming
2884  * from userspace or just about to enter userspace, a preempt enable
2885  * can occur before user_exit() is called. This will cause the scheduler
2886  * to be called when the system is still in usermode.
2887  *
2888  * To prevent this, the preempt_enable_notrace will use this function
2889  * instead of preempt_schedule() to exit user context if needed before
2890  * calling the scheduler.
2891  */
2892 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2893 {
2894         enum ctx_state prev_ctx;
2895 
2896         if (likely(!preemptible()))
2897                 return;
2898 
2899         do {
2900                 __preempt_count_add(PREEMPT_ACTIVE);
2901                 /*
2902                  * Needs preempt disabled in case user_exit() is traced
2903                  * and the tracer calls preempt_enable_notrace() causing
2904                  * an infinite recursion.
2905                  */
2906                 prev_ctx = exception_enter();
2907                 __schedule();
2908                 exception_exit(prev_ctx);
2909 
2910                 __preempt_count_sub(PREEMPT_ACTIVE);
2911                 barrier();
2912         } while (need_resched());
2913 }
2914 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2915 #endif /* CONFIG_CONTEXT_TRACKING */
2916 
2917 #endif /* CONFIG_PREEMPT */
2918 
2919 /*
2920  * this is the entry point to schedule() from kernel preemption
2921  * off of irq context.
2922  * Note, that this is called and return with irqs disabled. This will
2923  * protect us against recursive calling from irq.
2924  */
2925 asmlinkage __visible void __sched preempt_schedule_irq(void)
2926 {
2927         enum ctx_state prev_state;
2928 
2929         /* Catch callers which need to be fixed */
2930         BUG_ON(preempt_count() || !irqs_disabled());
2931 
2932         prev_state = exception_enter();
2933 
2934         do {
2935                 __preempt_count_add(PREEMPT_ACTIVE);
2936                 local_irq_enable();
2937                 __schedule();
2938                 local_irq_disable();
2939                 __preempt_count_sub(PREEMPT_ACTIVE);
2940 
2941                 /*
2942                  * Check again in case we missed a preemption opportunity
2943                  * between schedule and now.
2944                  */
2945                 barrier();
2946         } while (need_resched());
2947 
2948         exception_exit(prev_state);
2949 }
2950 
2951 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2952                           void *key)
2953 {
2954         return try_to_wake_up(curr->private, mode, wake_flags);
2955 }
2956 EXPORT_SYMBOL(default_wake_function);
2957 
2958 #ifdef CONFIG_RT_MUTEXES
2959 
2960 /*
2961  * rt_mutex_setprio - set the current priority of a task
2962  * @p: task
2963  * @prio: prio value (kernel-internal form)
2964  *
2965  * This function changes the 'effective' priority of a task. It does
2966  * not touch ->normal_prio like __setscheduler().
2967  *
2968  * Used by the rt_mutex code to implement priority inheritance
2969  * logic. Call site only calls if the priority of the task changed.
2970  */
2971 void rt_mutex_setprio(struct task_struct *p, int prio)
2972 {
2973         int oldprio, queued, running, enqueue_flag = 0;
2974         struct rq *rq;
2975         const struct sched_class *prev_class;
2976 
2977         BUG_ON(prio > MAX_PRIO);
2978 
2979         rq = __task_rq_lock(p);
2980 
2981         /*
2982          * Idle task boosting is a nono in general. There is one
2983          * exception, when PREEMPT_RT and NOHZ is active:
2984          *
2985          * The idle task calls get_next_timer_interrupt() and holds
2986          * the timer wheel base->lock on the CPU and another CPU wants
2987          * to access the timer (probably to cancel it). We can safely
2988          * ignore the boosting request, as the idle CPU runs this code
2989          * with interrupts disabled and will complete the lock
2990          * protected section without being interrupted. So there is no
2991          * real need to boost.
2992          */
2993         if (unlikely(p == rq->idle)) {
2994                 WARN_ON(p != rq->curr);
2995                 WARN_ON(p->pi_blocked_on);
2996                 goto out_unlock;
2997         }
2998 
2999         trace_sched_pi_setprio(p, prio);
3000         oldprio = p->prio;
3001         prev_class = p->sched_class;
3002         queued = task_on_rq_queued(p);
3003         running = task_current(rq, p);
3004         if (queued)
3005                 dequeue_task(rq, p, 0);
3006         if (running)
3007                 put_prev_task(rq, p);
3008 
3009         /*
3010          * Boosting condition are:
3011          * 1. -rt task is running and holds mutex A
3012          *      --> -dl task blocks on mutex A
3013          *
3014          * 2. -dl task is running and holds mutex A
3015          *      --> -dl task blocks on mutex A and could preempt the
3016          *          running task
3017          */
3018         if (dl_prio(prio)) {
3019                 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3020                 if (!dl_prio(p->normal_prio) ||
3021                     (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3022                         p->dl.dl_boosted = 1;
3023                         p->dl.dl_throttled = 0;
3024                         enqueue_flag = ENQUEUE_REPLENISH;
3025                 } else
3026                         p->dl.dl_boosted = 0;
3027                 p->sched_class = &dl_sched_class;
3028         } else if (rt_prio(prio)) {
3029                 if (dl_prio(oldprio))
3030                         p->dl.dl_boosted = 0;
3031                 if (oldprio < prio)
3032                         enqueue_flag = ENQUEUE_HEAD;
3033                 p->sched_class = &rt_sched_class;
3034         } else {
3035                 if (dl_prio(oldprio))
3036                         p->dl.dl_boosted = 0;
3037                 if (rt_prio(oldprio))
3038                         p->rt.timeout = 0;
3039                 p->sched_class = &fair_sched_class;
3040         }
3041 
3042         p->prio = prio;
3043 
3044         if (running)
3045                 p->sched_class->set_curr_task(rq);
3046         if (queued)
3047                 enqueue_task(rq, p, enqueue_flag);
3048 
3049         check_class_changed(rq, p, prev_class, oldprio);
3050 out_unlock:
3051         __task_rq_unlock(rq);
3052 }
3053 #endif
3054 
3055 void set_user_nice(struct task_struct *p, long nice)
3056 {
3057         int old_prio, delta, queued;
3058         unsigned long flags;
3059         struct rq *rq;
3060 
3061         if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3062                 return;
3063         /*
3064          * We have to be careful, if called from sys_setpriority(),
3065          * the task might be in the middle of scheduling on another CPU.
3066          */
3067         rq = task_rq_lock(p, &flags);
3068         /*
3069          * The RT priorities are set via sched_setscheduler(), but we still
3070          * allow the 'normal' nice value to be set - but as expected
3071          * it wont have any effect on scheduling until the task is
3072          * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3073          */
3074         if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3075                 p->static_prio = NICE_TO_PRIO(nice);
3076                 goto out_unlock;
3077         }
3078         queued = task_on_rq_queued(p);
3079         if (queued)
3080                 dequeue_task(rq, p, 0);
3081 
3082         p->static_prio = NICE_TO_PRIO(nice);
3083         set_load_weight(p);
3084         old_prio = p->prio;
3085         p->prio = effective_prio(p);
3086         delta = p->prio - old_prio;
3087 
3088         if (queued) {
3089                 enqueue_task(rq, p, 0);
3090                 /*
3091                  * If the task increased its priority or is running and
3092                  * lowered its priority, then reschedule its CPU:
3093                  */
3094                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3095                         resched_curr(rq);
3096         }
3097 out_unlock:
3098         task_rq_unlock(rq, p, &flags);
3099 }
3100 EXPORT_SYMBOL(set_user_nice);
3101 
3102 /*
3103  * can_nice - check if a task can reduce its nice value
3104  * @p: task
3105  * @nice: nice value
3106  */
3107 int can_nice(const struct task_struct *p, const int nice)
3108 {
3109         /* convert nice value [19,-20] to rlimit style value [1,40] */
3110         int nice_rlim = nice_to_rlimit(nice);
3111 
3112         return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3113                 capable(CAP_SYS_NICE));
3114 }
3115 
3116 #ifdef __ARCH_WANT_SYS_NICE
3117 
3118 /*
3119  * sys_nice - change the priority of the current process.
3120  * @increment: priority increment
3121  *
3122  * sys_setpriority is a more generic, but much slower function that
3123  * does similar things.
3124  */
3125 SYSCALL_DEFINE1(nice, int, increment)
3126 {
3127         long nice, retval;
3128 
3129         /*
3130          * Setpriority might change our priority at the same moment.
3131          * We don't have to worry. Conceptually one call occurs first
3132          * and we have a single winner.
3133          */
3134         increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3135         nice = task_nice(current) + increment;
3136 
3137         nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3138         if (increment < 0 && !can_nice(current, nice))
3139                 return -EPERM;
3140 
3141         retval = security_task_setnice(current, nice);
3142         if (retval)
3143                 return retval;
3144 
3145         set_user_nice(current, nice);
3146         return 0;
3147 }
3148 
3149 #endif
3150 
3151 /**
3152  * task_prio - return the priority value of a given task.
3153  * @p: the task in question.
3154  *
3155  * Return: The priority value as seen by users in /proc.
3156  * RT tasks are offset by -200. Normal tasks are centered
3157  * around 0, value goes from -16 to +15.
3158  */
3159 int task_prio(const struct task_struct *p)
3160 {
3161         return p->prio - MAX_RT_PRIO;
3162 }
3163 
3164 /**
3165  * idle_cpu - is a given cpu idle currently?
3166  * @cpu: the processor in question.
3167  *
3168  * Return: 1 if the CPU is currently idle. 0 otherwise.
3169  */
3170 int idle_cpu(int cpu)
3171 {
3172         struct rq *rq = cpu_rq(cpu);
3173 
3174         if (rq->curr != rq->idle)
3175                 return 0;
3176 
3177         if (rq->nr_running)
3178                 return 0;
3179 
3180 #ifdef CONFIG_SMP
3181         if (!llist_empty(&rq->wake_list))
3182                 return 0;
3183 #endif
3184 
3185         return 1;
3186 }
3187 
3188 /**
3189  * idle_task - return the idle task for a given cpu.
3190  * @cpu: the processor in question.
3191  *
3192  * Return: The idle task for the cpu @cpu.
3193  */
3194 struct task_struct *idle_task(int cpu)
3195 {
3196         return cpu_rq(cpu)->idle;
3197 }
3198 
3199 /**
3200  * find_process_by_pid - find a process with a matching PID value.
3201  * @pid: the pid in question.
3202  *
3203  * The task of @pid, if found. %NULL otherwise.
3204  */
3205 static struct task_struct *find_process_by_pid(pid_t pid)
3206 {
3207         return pid ? find_task_by_vpid(pid) : current;
3208 }
3209 
3210 /*
3211  * This function initializes the sched_dl_entity of a newly becoming
3212  * SCHED_DEADLINE task.
3213  *
3214  * Only the static values are considered here, the actual runtime and the
3215  * absolute deadline will be properly calculated when the task is enqueued
3216  * for the first time with its new policy.
3217  */
3218 static void
3219 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3220 {
3221         struct sched_dl_entity *dl_se = &p->dl;
3222 
3223         dl_se->dl_runtime = attr->sched_runtime;
3224         dl_se->dl_deadline = attr->sched_deadline;
3225         dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3226         dl_se->flags = attr->sched_flags;
3227         dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3228 
3229         /*
3230          * Changing the parameters of a task is 'tricky' and we're not doing
3231          * the correct thing -- also see task_dead_dl() and switched_from_dl().
3232          *
3233          * What we SHOULD do is delay the bandwidth release until the 0-lag
3234          * point. This would include retaining the task_struct until that time
3235          * and change dl_overflow() to not immediately decrement the current
3236          * amount.
3237          *
3238          * Instead we retain the current runtime/deadline and let the new
3239          * parameters take effect after the current reservation period lapses.
3240          * This is safe (albeit pessimistic) because the 0-lag point is always
3241          * before the current scheduling deadline.
3242          *
3243          * We can still have temporary overloads because we do not delay the
3244          * change in bandwidth until that time; so admission control is
3245          * not on the safe side. It does however guarantee tasks will never
3246          * consume more than promised.
3247          */
3248 }
3249 
3250 /*
3251  * sched_setparam() passes in -1 for its policy, to let the functions
3252  * it calls know not to change it.
3253  */
3254 #define SETPARAM_POLICY -1
3255 
3256 static void __setscheduler_params(struct task_struct *p,
3257                 const struct sched_attr *attr)
3258 {
3259         int policy = attr->sched_policy;
3260 
3261         if (policy == SETPARAM_POLICY)
3262                 policy = p->policy;
3263 
3264         p->policy = policy;
3265 
3266         if (dl_policy(policy))
3267                 __setparam_dl(p, attr);
3268         else if (fair_policy(policy))
3269                 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3270 
3271         /*
3272          * __sched_setscheduler() ensures attr->sched_priority == 0 when
3273          * !rt_policy. Always setting this ensures that things like
3274          * getparam()/getattr() don't report silly values for !rt tasks.
3275          */
3276         p->rt_priority = attr->sched_priority;
3277         p->normal_prio = normal_prio(p);
3278         set_load_weight(p);
3279 }
3280 
3281 /* Actually do priority change: must hold pi & rq lock. */
3282 static void __setscheduler(struct rq *rq, struct task_struct *p,
3283                            const struct sched_attr *attr)
3284 {
3285         __setscheduler_params(p, attr);
3286 
3287         /*
3288          * If we get here, there was no pi waiters boosting the
3289          * task. It is safe to use the normal prio.
3290          */
3291         p->prio = normal_prio(p);
3292 
3293         if (dl_prio(p->prio))
3294                 p->sched_class = &dl_sched_class;
3295         else if (rt_prio(p->prio))
3296                 p->sched_class = &rt_sched_class;
3297         else
3298                 p->sched_class = &fair_sched_class;
3299 }
3300 
3301 static void
3302 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3303 {
3304         struct sched_dl_entity *dl_se = &p->dl;
3305 
3306         attr->sched_priority = p->rt_priority;
3307         attr->sched_runtime = dl_se->dl_runtime;
3308         attr->sched_deadline = dl_se->dl_deadline;
3309         attr->sched_period = dl_se->dl_period;
3310         attr->sched_flags = dl_se->flags;
3311 }
3312 
3313 /*
3314  * This function validates the new parameters of a -deadline task.
3315  * We ask for the deadline not being zero, and greater or equal
3316  * than the runtime, as well as the period of being zero or
3317  * greater than deadline. Furthermore, we have to be sure that
3318  * user parameters are above the internal resolution of 1us (we
3319  * check sched_runtime only since it is always the smaller one) and
3320  * below 2^63 ns (we have to check both sched_deadline and
3321  * sched_period, as the latter can be zero).
3322  */
3323 static bool
3324 __checkparam_dl(const struct sched_attr *attr)
3325 {
3326         /* deadline != 0 */
3327         if (attr->sched_deadline == 0)
3328                 return false;
3329 
3330         /*
3331          * Since we truncate DL_SCALE bits, make sure we're at least
3332          * that big.
3333          */
3334         if (attr->sched_runtime < (1ULL << DL_SCALE))
3335                 return false;
3336 
3337         /*
3338          * Since we use the MSB for wrap-around and sign issues, make
3339          * sure it's not set (mind that period can be equal to zero).
3340          */
3341         if (attr->sched_deadline & (1ULL << 63) ||
3342             attr->sched_period & (1ULL << 63))
3343                 return false;
3344 
3345         /* runtime <= deadline <= period (if period != 0) */
3346         if ((attr->sched_period != 0 &&
3347              attr->sched_period < attr->sched_deadline) ||
3348             attr->sched_deadline < attr->sched_runtime)
3349                 return false;
3350 
3351         return true;
3352 }
3353 
3354 /*
3355  * check the target process has a UID that matches the current process's
3356  */
3357 static bool check_same_owner(struct task_struct *p)
3358 {
3359         const struct cred *cred = current_cred(), *pcred;
3360         bool match;
3361 
3362         rcu_read_lock();
3363         pcred = __task_cred(p);
3364         match = (uid_eq(cred->euid, pcred->euid) ||
3365                  uid_eq(cred->euid, pcred->uid));
3366         rcu_read_unlock();
3367         return match;
3368 }
3369 
3370 static bool dl_param_changed(struct task_struct *p,
3371                 const struct sched_attr *attr)
3372 {
3373         struct sched_dl_entity *dl_se = &p->dl;
3374 
3375         if (dl_se->dl_runtime != attr->sched_runtime ||
3376                 dl_se->dl_deadline != attr->sched_deadline ||
3377                 dl_se->dl_period != attr->sched_period ||
3378                 dl_se->flags != attr->sched_flags)
3379                 return true;
3380 
3381         return false;
3382 }
3383 
3384 static int __sched_setscheduler(struct task_struct *p,
3385                                 const struct sched_attr *attr,
3386                                 bool user)
3387 {
3388         int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3389                       MAX_RT_PRIO - 1 - attr->sched_priority;
3390         int retval, oldprio, oldpolicy = -1, queued, running;
3391         int policy = attr->sched_policy;
3392         unsigned long flags;
3393         const struct sched_class *prev_class;
3394         struct rq *rq;
3395         int reset_on_fork;
3396 
3397         /* may grab non-irq protected spin_locks */
3398         BUG_ON(in_interrupt());
3399 recheck:
3400         /* double check policy once rq lock held */
3401         if (policy < 0) {
3402                 reset_on_fork = p->sched_reset_on_fork;
3403                 policy = oldpolicy = p->policy;
3404         } else {
3405                 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3406 
3407                 if (policy != SCHED_DEADLINE &&
3408                                 policy != SCHED_FIFO && policy != SCHED_RR &&
3409                                 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3410                                 policy != SCHED_IDLE)
3411                         return -EINVAL;
3412         }
3413 
3414         if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3415                 return -EINVAL;
3416 
3417         /*
3418          * Valid priorities for SCHED_FIFO and SCHED_RR are
3419          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3420          * SCHED_BATCH and SCHED_IDLE is 0.
3421          */
3422         if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3423             (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3424                 return -EINVAL;
3425         if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3426             (rt_policy(policy) != (attr->sched_priority != 0)))
3427                 return -EINVAL;
3428 
3429         /*
3430          * Allow unprivileged RT tasks to decrease priority:
3431          */
3432         if (user && !capable(CAP_SYS_NICE)) {
3433                 if (fair_policy(policy)) {
3434                         if (attr->sched_nice < task_nice(p) &&
3435                             !can_nice(p, attr->sched_nice))
3436                                 return -EPERM;
3437                 }
3438 
3439                 if (rt_policy(policy)) {
3440                         unsigned long rlim_rtprio =
3441                                         task_rlimit(p, RLIMIT_RTPRIO);
3442 
3443                         /* can't set/change the rt policy */
3444                         if (policy != p->policy && !rlim_rtprio)
3445                                 return -EPERM;
3446 
3447                         /* can't increase priority */
3448                         if (attr->sched_priority > p->rt_priority &&
3449                             attr->sched_priority > rlim_rtprio)
3450                                 return -EPERM;
3451                 }
3452 
3453                  /*
3454                   * Can't set/change SCHED_DEADLINE policy at all for now
3455                   * (safest behavior); in the future we would like to allow
3456                   * unprivileged DL tasks to increase their relative deadline
3457                   * or reduce their runtime (both ways reducing utilization)
3458                   */
3459                 if (dl_policy(policy))
3460                         return -EPERM;
3461 
3462                 /*
3463                  * Treat SCHED_IDLE as nice 20. Only allow a switch to
3464                  * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3465                  */
3466                 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3467                         if (!can_nice(p, task_nice(p)))
3468                                 return -EPERM;
3469                 }
3470 
3471                 /* can't change other user's priorities */
3472                 if (!check_same_owner(p))
3473                         return -EPERM;
3474 
3475                 /* Normal users shall not reset the sched_reset_on_fork flag */
3476                 if (p->sched_reset_on_fork && !reset_on_fork)
3477                         return -EPERM;
3478         }
3479 
3480         if (user) {
3481                 retval = security_task_setscheduler(p);
3482                 if (retval)
3483                         return retval;
3484         }
3485 
3486         /*
3487          * make sure no PI-waiters arrive (or leave) while we are
3488          * changing the priority of the task:
3489          *
3490          * To be able to change p->policy safely, the appropriate
3491          * runqueue lock must be held.
3492          */
3493         rq = task_rq_lock(p, &flags);
3494 
3495         /*
3496          * Changing the policy of the stop threads its a very bad idea
3497          */
3498         if (p == rq->stop) {
3499                 task_rq_unlock(rq, p, &flags);
3500                 return -EINVAL;
3501         }
3502 
3503         /*
3504          * If not changing anything there's no need to proceed further,
3505          * but store a possible modification of reset_on_fork.
3506          */
3507         if (unlikely(policy == p->policy)) {
3508                 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3509                         goto change;
3510                 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3511                         goto change;
3512                 if (dl_policy(policy) && dl_param_changed(p, attr))
3513                         goto change;
3514 
3515                 p->sched_reset_on_fork = reset_on_fork;
3516                 task_rq_unlock(rq, p, &flags);
3517                 return 0;
3518         }
3519 change:
3520 
3521         if (user) {
3522 #ifdef CONFIG_RT_GROUP_SCHED
3523                 /*
3524                  * Do not allow realtime tasks into groups that have no runtime
3525                  * assigned.
3526                  */
3527                 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3528                                 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3529                                 !task_group_is_autogroup(task_group(p))) {
3530                         task_rq_unlock(rq, p, &flags);
3531                         return -EPERM;
3532                 }
3533 #endif
3534 #ifdef CONFIG_SMP
3535                 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3536                         cpumask_t *span = rq->rd->span;
3537 
3538                         /*
3539                          * Don't allow tasks with an affinity mask smaller than
3540                          * the entire root_domain to become SCHED_DEADLINE. We
3541                          * will also fail if there's no bandwidth available.
3542                          */
3543                         if (!cpumask_subset(span, &p->cpus_allowed) ||
3544                             rq->rd->dl_bw.bw == 0) {
3545                                 task_rq_unlock(rq, p, &flags);
3546                                 return -EPERM;
3547                         }
3548                 }
3549 #endif
3550         }
3551 
3552         /* recheck policy now with rq lock held */
3553         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3554                 policy = oldpolicy = -1;
3555                 task_rq_unlock(rq, p, &flags);
3556                 goto recheck;
3557         }
3558 
3559         /*
3560          * If setscheduling to SCHED_DEADLINE (or changing the parameters
3561          * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3562          * is available.
3563          */
3564         if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3565                 task_rq_unlock(rq, p, &flags);
3566                 return -EBUSY;
3567         }
3568 
3569         p->sched_reset_on_fork = reset_on_fork;
3570         oldprio = p->prio;
3571 
3572         /*
3573          * Special case for priority boosted tasks.
3574          *
3575          * If the new priority is lower or equal (user space view)
3576          * than the current (boosted) priority, we just store the new
3577          * normal parameters and do not touch the scheduler class and
3578          * the runqueue. This will be done when the task deboost
3579          * itself.
3580          */
3581         if (rt_mutex_check_prio(p, newprio)) {
3582                 __setscheduler_params(p, attr);
3583                 task_rq_unlock(rq, p, &flags);
3584                 return 0;
3585         }
3586 
3587         queued = task_on_rq_queued(p);
3588         running = task_current(rq, p);
3589         if (queued)
3590                 dequeue_task(rq, p, 0);
3591         if (running)
3592                 put_prev_task(rq, p);
3593 
3594         prev_class = p->sched_class;
3595         __setscheduler(rq, p, attr);
3596 
3597         if (running)
3598                 p->sched_class->set_curr_task(rq);
3599         if (queued) {
3600                 /*
3601                  * We enqueue to tail when the priority of a task is
3602                  * increased (user space view).
3603                  */
3604                 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3605         }
3606 
3607         check_class_changed(rq, p, prev_class, oldprio);
3608         task_rq_unlock(rq, p, &flags);
3609 
3610         rt_mutex_adjust_pi(p);
3611 
3612         return 0;
3613 }
3614 
3615 static int _sched_setscheduler(struct task_struct *p, int policy,
3616                                const struct sched_param *param, bool check)
3617 {
3618         struct sched_attr attr = {
3619                 .sched_policy   = policy,
3620                 .sched_priority = param->sched_priority,
3621                 .sched_nice     = PRIO_TO_NICE(p->static_prio),
3622         };
3623 
3624         /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3625         if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3626                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3627                 policy &= ~SCHED_RESET_ON_FORK;
3628                 attr.sched_policy = policy;
3629         }
3630 
3631         return __sched_setscheduler(p, &attr, check);
3632 }
3633 /**
3634  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3635  * @p: the task in question.
3636  * @policy: new policy.
3637  * @param: structure containing the new RT priority.
3638  *
3639  * Return: 0 on success. An error code otherwise.
3640  *
3641  * NOTE that the task may be already dead.
3642  */
3643 int sched_setscheduler(struct task_struct *p, int policy,
3644                        const struct sched_param *param)
3645 {
3646         return _sched_setscheduler(p, policy, param, true);
3647 }
3648 EXPORT_SYMBOL_GPL(sched_setscheduler);
3649 
3650 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3651 {
3652         return __sched_setscheduler(p, attr, true);
3653 }
3654 EXPORT_SYMBOL_GPL(sched_setattr);
3655 
3656 /**
3657  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3658  * @p: the task in question.
3659  * @policy: new policy.
3660  * @param: structure containing the new RT priority.
3661  *
3662  * Just like sched_setscheduler, only don't bother checking if the
3663  * current context has permission.  For example, this is needed in
3664  * stop_machine(): we create temporary high priority worker threads,
3665  * but our caller might not have that capability.
3666  *
3667  * Return: 0 on success. An error code otherwise.
3668  */
3669 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3670                                const struct sched_param *param)
3671 {
3672         return _sched_setscheduler(p, policy, param, false);
3673 }
3674 
3675 static int
3676 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3677 {
3678         struct sched_param lparam;
3679         struct task_struct *p;
3680         int retval;
3681 
3682         if (!param || pid < 0)
3683                 return -EINVAL;
3684         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3685                 return -EFAULT;
3686 
3687         rcu_read_lock();
3688         retval = -ESRCH;
3689         p = find_process_by_pid(pid);
3690         if (p != NULL)
3691                 retval = sched_setscheduler(p, policy, &lparam);
3692         rcu_read_unlock();
3693 
3694         return retval;
3695 }
3696 
3697 /*
3698  * Mimics kernel/events/core.c perf_copy_attr().
3699  */
3700 static int sched_copy_attr(struct sched_attr __user *uattr,
3701                            struct sched_attr *attr)
3702 {
3703         u32 size;
3704         int ret;
3705 
3706         if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3707                 return -EFAULT;
3708 
3709         /*
3710          * zero the full structure, so that a short copy will be nice.
3711          */
3712         memset(attr, 0, sizeof(*attr));
3713 
3714         ret = get_user(size, &uattr->size);
3715         if (ret)
3716                 return ret;
3717 
3718         if (size > PAGE_SIZE)   /* silly large */
3719                 goto err_size;
3720 
3721         if (!size)              /* abi compat */
3722                 size = SCHED_ATTR_SIZE_VER0;
3723 
3724         if (size < SCHED_ATTR_SIZE_VER0)
3725                 goto err_size;
3726 
3727         /*
3728          * If we're handed a bigger struct than we know of,
3729          * ensure all the unknown bits are 0 - i.e. new
3730          * user-space does not rely on any kernel feature
3731          * extensions we dont know about yet.
3732          */
3733         if (size > sizeof(*attr)) {
3734                 unsigned char __user *addr;
3735                 unsigned char __user *end;
3736                 unsigned char val;
3737 
3738                 addr = (void __user *)uattr + sizeof(*attr);
3739                 end  = (void __user *)uattr + size;
3740 
3741                 for (; addr < end; addr++) {
3742                         ret = get_user(val, addr);
3743                         if (ret)
3744                                 return ret;
3745                         if (val)
3746                                 goto err_size;
3747                 }
3748                 size = sizeof(*attr);
3749         }
3750 
3751         ret = copy_from_user(attr, uattr, size);
3752         if (ret)
3753                 return -EFAULT;
3754 
3755         /*
3756          * XXX: do we want to be lenient like existing syscalls; or do we want
3757          * to be strict and return an error on out-of-bounds values?
3758          */
3759         attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3760 
3761         return 0;
3762 
3763 err_size:
3764         put_user(sizeof(*attr), &uattr->size);
3765         return -E2BIG;
3766 }
3767 
3768 /**
3769  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3770  * @pid: the pid in question.
3771  * @policy: new policy.
3772  * @param: structure containing the new RT priority.
3773  *
3774  * Return: 0 on success. An error code otherwise.
3775  */
3776 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3777                 struct sched_param __user *, param)
3778 {
3779         /* negative values for policy are not valid */
3780         if (policy < 0)
3781                 return -EINVAL;
3782 
3783         return do_sched_setscheduler(pid, policy, param);
3784 }
3785 
3786 /**
3787  * sys_sched_setparam - set/change the RT priority of a thread
3788  * @pid: the pid in question.
3789  * @param: structure containing the new RT priority.
3790  *
3791  * Return: 0 on success. An error code otherwise.
3792  */
3793 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3794 {
3795         return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3796 }
3797 
3798 /**
3799  * sys_sched_setattr - same as above, but with extended sched_attr
3800  * @pid: the pid in question.
3801  * @uattr: structure containing the extended parameters.
3802  * @flags: for future extension.
3803  */
3804 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3805                                unsigned int, flags)
3806 {
3807         struct sched_attr attr;
3808         struct task_struct *p;
3809         int retval;
3810 
3811         if (!uattr || pid < 0 || flags)
3812                 return -EINVAL;
3813 
3814         retval = sched_copy_attr(uattr, &attr);
3815         if (retval)
3816                 return retval;
3817 
3818         if ((int)attr.sched_policy < 0)
3819                 return -EINVAL;
3820 
3821         rcu_read_lock();
3822         retval = -ESRCH;
3823         p = find_process_by_pid(pid);
3824         if (p != NULL)
3825                 retval = sched_setattr(p, &attr);
3826         rcu_read_unlock();
3827 
3828         return retval;
3829 }
3830 
3831 /**
3832  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3833  * @pid: the pid in question.
3834  *
3835  * Return: On success, the policy of the thread. Otherwise, a negative error
3836  * code.
3837  */
3838 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3839 {
3840         struct task_struct *p;
3841         int retval;
3842 
3843         if (pid < 0)
3844                 return -EINVAL;
3845 
3846         retval = -ESRCH;
3847         rcu_read_lock();
3848         p = find_process_by_pid(pid);
3849         if (p) {
3850                 retval = security_task_getscheduler(p);
3851                 if (!retval)
3852                         retval = p->policy
3853                                 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3854         }
3855         rcu_read_unlock();
3856         return retval;
3857 }
3858 
3859 /**
3860  * sys_sched_getparam - get the RT priority of a thread
3861  * @pid: the pid in question.
3862  * @param: structure containing the RT priority.
3863  *
3864  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3865  * code.
3866  */
3867 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3868 {
3869         struct sched_param lp = { .sched_priority = 0 };
3870         struct task_struct *p;
3871         int retval;
3872 
3873         if (!param || pid < 0)
3874                 return -EINVAL;
3875 
3876         rcu_read_lock();
3877         p = find_process_by_pid(pid);
3878         retval = -ESRCH;
3879         if (!p)
3880                 goto out_unlock;
3881 
3882         retval = security_task_getscheduler(p);
3883         if (retval)
3884                 goto out_unlock;
3885 
3886         if (task_has_rt_policy(p))
3887                 lp.sched_priority = p->rt_priority;
3888         rcu_read_unlock();
3889 
3890         /*
3891          * This one might sleep, we cannot do it with a spinlock held ...
3892          */
3893         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3894 
3895         return retval;
3896 
3897 out_unlock:
3898         rcu_read_unlock();
3899         return retval;
3900 }
3901 
3902 static int sched_read_attr(struct sched_attr __user *uattr,
3903                            struct sched_attr *attr,
3904                            unsigned int usize)
3905 {
3906         int ret;
3907 
3908         if (!access_ok(VERIFY_WRITE, uattr, usize))
3909                 return -EFAULT;
3910 
3911         /*
3912          * If we're handed a smaller struct than we know of,
3913          * ensure all the unknown bits are 0 - i.e. old
3914          * user-space does not get uncomplete information.
3915          */
3916         if (usize < sizeof(*attr)) {
3917                 unsigned char *addr;
3918                 unsigned char *end;
3919 
3920                 addr = (void *)attr + usize;
3921                 end  = (void *)attr + sizeof(*attr);
3922 
3923                 for (; addr < end; addr++) {
3924                         if (*addr)
3925                                 return -EFBIG;
3926                 }
3927 
3928                 attr->size = usize;
3929         }
3930 
3931         ret = copy_to_user(uattr, attr, attr->size);
3932         if (ret)
3933                 return -EFAULT;
3934 
3935         return 0;
3936 }
3937 
3938 /**
3939  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3940  * @pid: the pid in question.
3941  * @uattr: structure containing the extended parameters.
3942  * @size: sizeof(attr) for fwd/bwd comp.
3943  * @flags: for future extension.
3944  */
3945 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3946                 unsigned int, size, unsigned int, flags)
3947 {
3948         struct sched_attr attr = {
3949                 .size = sizeof(struct sched_attr),
3950         };
3951         struct task_struct *p;
3952         int retval;
3953 
3954         if (!uattr || pid < 0 || size > PAGE_SIZE ||
3955             size < SCHED_ATTR_SIZE_VER0 || flags)
3956                 return -EINVAL;
3957 
3958         rcu_read_lock();
3959         p = find_process_by_pid(pid);
3960         retval = -ESRCH;
3961         if (!p)
3962                 goto out_unlock;
3963 
3964         retval = security_task_getscheduler(p);
3965         if (retval)
3966                 goto out_unlock;
3967 
3968         attr.sched_policy = p->policy;
3969         if (p->sched_reset_on_fork)
3970                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3971         if (task_has_dl_policy(p))
3972                 __getparam_dl(p, &attr);
3973         else if (task_has_rt_policy(p))
3974                 attr.sched_priority = p->rt_priority;
3975         else
3976                 attr.sched_nice = task_nice(p);
3977 
3978         rcu_read_unlock();
3979 
3980         retval = sched_read_attr(uattr, &attr, size);
3981         return retval;
3982 
3983 out_unlock:
3984         rcu_read_unlock();
3985         return retval;
3986 }
3987 
3988 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3989 {
3990         cpumask_var_t cpus_allowed, new_mask;
3991         struct task_struct *p;
3992         int retval;
3993 
3994         rcu_read_lock();
3995 
3996         p = find_process_by_pid(pid);
3997         if (!p) {
3998                 rcu_read_unlock();
3999                 return -ESRCH;
4000         }
4001 
4002         /* Prevent p going away */
4003         get_task_struct(p);
4004         rcu_read_unlock();
4005 
4006         if (p->flags & PF_NO_SETAFFINITY) {
4007                 retval = -EINVAL;
4008                 goto out_put_task;
4009         }
4010         if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4011                 retval = -ENOMEM;
4012                 goto out_put_task;
4013         }
4014         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4015                 retval = -ENOMEM;
4016                 goto out_free_cpus_allowed;
4017         }
4018         retval = -EPERM;
4019         if (!check_same_owner(p)) {
4020                 rcu_read_lock();
4021                 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4022                         rcu_read_unlock();
4023                         goto out_free_new_mask;
4024                 }
4025                 rcu_read_unlock();
4026         }
4027 
4028         retval = security_task_setscheduler(p);
4029         if (retval)
4030                 goto out_free_new_mask;
4031 
4032 
4033         cpuset_cpus_allowed(p, cpus_allowed);
4034         cpumask_and(new_mask, in_mask, cpus_allowed);
4035 
4036         /*
4037          * Since bandwidth control happens on root_domain basis,
4038          * if admission test is enabled, we only admit -deadline
4039          * tasks allowed to run on all the CPUs in the task's
4040          * root_domain.
4041          */
4042 #ifdef CONFIG_SMP
4043         if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4044                 rcu_read_lock();
4045                 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4046                         retval = -EBUSY;
4047                         rcu_read_unlock();
4048                         goto out_free_new_mask;
4049                 }
4050                 rcu_read_unlock();
4051         }
4052 #endif
4053 again:
4054         retval = set_cpus_allowed_ptr(p, new_mask);
4055 
4056         if (!retval) {
4057                 cpuset_cpus_allowed(p, cpus_allowed);
4058                 if (!cpumask_subset(new_mask, cpus_allowed)) {
4059                         /*
4060                          * We must have raced with a concurrent cpuset
4061                          * update. Just reset the cpus_allowed to the
4062                          * cpuset's cpus_allowed
4063                          */
4064                         cpumask_copy(new_mask, cpus_allowed);
4065                         goto again;
4066                 }
4067         }
4068 out_free_new_mask:
4069         free_cpumask_var(new_mask);
4070 out_free_cpus_allowed:
4071         free_cpumask_var(cpus_allowed);
4072 out_put_task:
4073         put_task_struct(p);
4074         return retval;
4075 }
4076 
4077 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4078                              struct cpumask *new_mask)
4079 {
4080         if (len < cpumask_size())
4081                 cpumask_clear(new_mask);
4082         else if (len > cpumask_size())
4083                 len = cpumask_size();
4084 
4085         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4086 }
4087 
4088 /**
4089  * sys_sched_setaffinity - set the cpu affinity of a process
4090  * @pid: pid of the process
4091  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4092  * @user_mask_ptr: user-space pointer to the new cpu mask
4093  *
4094  * Return: 0 on success. An error code otherwise.
4095  */
4096 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4097                 unsigned long __user *, user_mask_ptr)
4098 {
4099         cpumask_var_t new_mask;
4100         int retval;
4101 
4102         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4103                 return -ENOMEM;
4104 
4105         retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4106         if (retval == 0)
4107                 retval = sched_setaffinity(pid, new_mask);
4108         free_cpumask_var(new_mask);
4109         return retval;
4110 }
4111 
4112 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4113 {
4114         struct task_struct *p;
4115         unsigned long flags;
4116         int retval;
4117 
4118         rcu_read_lock();
4119 
4120         retval = -ESRCH;
4121         p = find_process_by_pid(pid);
4122         if (!p)
4123                 goto out_unlock;
4124 
4125         retval = security_task_getscheduler(p);
4126         if (retval)
4127                 goto out_unlock;
4128 
4129         raw_spin_lock_irqsave(&p->pi_lock, flags);
4130         cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4131         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4132 
4133 out_unlock:
4134         rcu_read_unlock();
4135 
4136         return retval;
4137 }
4138 
4139 /**
4140  * sys_sched_getaffinity - get the cpu affinity of a process
4141  * @pid: pid of the process
4142  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4143  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4144  *
4145  * Return: 0 on success. An error code otherwise.
4146  */
4147 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4148                 unsigned long __user *, user_mask_ptr)
4149 {
4150         int ret;
4151         cpumask_var_t mask;
4152 
4153         if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4154                 return -EINVAL;
4155         if (len & (sizeof(unsigned long)-1))
4156                 return -EINVAL;
4157 
4158         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4159                 return -ENOMEM;
4160 
4161         ret = sched_getaffinity(pid, mask);
4162         if (ret == 0) {
4163                 size_t retlen = min_t(size_t, len, cpumask_size());
4164 
4165                 if (copy_to_user(user_mask_ptr, mask, retlen))
4166                         ret = -EFAULT;
4167                 else
4168                         ret = retlen;
4169         }
4170         free_cpumask_var(mask);
4171 
4172         return ret;
4173 }
4174 
4175 /**
4176  * sys_sched_yield - yield the current processor to other threads.
4177  *
4178  * This function yields the current CPU to other tasks. If there are no
4179  * other threads running on this CPU then this function will return.
4180  *
4181  * Return: 0.
4182  */
4183 SYSCALL_DEFINE0(sched_yield)
4184 {
4185         struct rq *rq = this_rq_lock();
4186 
4187         schedstat_inc(rq, yld_count);
4188         current->sched_class->yield_task(rq);
4189 
4190         /*
4191          * Since we are going to call schedule() anyway, there's
4192          * no need to preempt or enable interrupts:
4193          */
4194         __release(rq->lock);
4195         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4196         do_raw_spin_unlock(&rq->lock);
4197         sched_preempt_enable_no_resched();
4198 
4199         schedule();
4200 
4201         return 0;
4202 }
4203 
4204 int __sched _cond_resched(void)
4205 {
4206         if (should_resched()) {
4207                 preempt_schedule_common();
4208                 return 1;
4209         }
4210         return 0;
4211 }
4212 EXPORT_SYMBOL(_cond_resched);
4213 
4214 /*
4215  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4216  * call schedule, and on return reacquire the lock.
4217  *
4218  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4219  * operations here to prevent schedule() from being called twice (once via
4220  * spin_unlock(), once by hand).
4221  */
4222 int __cond_resched_lock(spinlock_t *lock)
4223 {
4224         int resched = should_resched();
4225         int ret = 0;
4226 
4227         lockdep_assert_held(lock);
4228 
4229         if (spin_needbreak(lock) || resched) {
4230                 spin_unlock(lock);
4231                 if (resched)
4232                         preempt_schedule_common();
4233                 else
4234                         cpu_relax();
4235                 ret = 1;
4236                 spin_lock(lock);
4237         }
4238         return ret;
4239 }
4240 EXPORT_SYMBOL(__cond_resched_lock);
4241 
4242 int __sched __cond_resched_softirq(void)
4243 {
4244         BUG_ON(!in_softirq());
4245 
4246         if (should_resched()) {
4247                 local_bh_enable();
4248                 preempt_schedule_common();
4249                 local_bh_disable();
4250                 return 1;
4251         }
4252         return 0;
4253 }
4254 EXPORT_SYMBOL(__cond_resched_softirq);
4255 
4256 /**
4257  * yield - yield the current processor to other threads.
4258  *
4259  * Do not ever use this function, there's a 99% chance you're doing it wrong.
4260  *
4261  * The scheduler is at all times free to pick the calling task as the most
4262  * eligible task to run, if removing the yield() call from your code breaks
4263  * it, its already broken.
4264  *
4265  * Typical broken usage is:
4266  *
4267  * while (!event)
4268  *      yield();
4269  *
4270  * where one assumes that yield() will let 'the other' process run that will
4271  * make event true. If the current task is a SCHED_FIFO task that will never
4272  * happen. Never use yield() as a progress guarantee!!
4273  *
4274  * If you want to use yield() to wait for something, use wait_event().
4275  * If you want to use yield() to be 'nice' for others, use cond_resched().
4276  * If you still want to use yield(), do not!
4277  */
4278 void __sched yield(void)
4279 {
4280         set_current_state(TASK_RUNNING);
4281         sys_sched_yield();
4282 }
4283 EXPORT_SYMBOL(yield);
4284 
4285 /**
4286  * yield_to - yield the current processor to another thread in
4287  * your thread group, or accelerate that thread toward the
4288  * processor it's on.
4289  * @p: target task
4290  * @preempt: whether task preemption is allowed or not
4291  *
4292  * It's the caller's job to ensure that the target task struct
4293  * can't go away on us before we can do any checks.
4294  *
4295  * Return:
4296  *      true (>0) if we indeed boosted the target task.
4297  *      false (0) if we failed to boost the target.
4298  *      -ESRCH if there's no task to yield to.
4299  */
4300 int __sched yield_to(struct task_struct *p, bool preempt)
4301 {
4302         struct task_struct *curr = current;
4303         struct rq *rq, *p_rq;
4304         unsigned long flags;
4305         int yielded = 0;
4306 
4307         local_irq_save(flags);
4308         rq = this_rq();
4309 
4310 again:
4311         p_rq = task_rq(p);
4312         /*
4313          * If we're the only runnable task on the rq and target rq also
4314          * has only one task, there's absolutely no point in yielding.
4315          */
4316         if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4317                 yielded = -ESRCH;
4318                 goto out_irq;
4319         }
4320 
4321         double_rq_lock(rq, p_rq);
4322         if (task_rq(p) != p_rq) {
4323                 double_rq_unlock(rq, p_rq);
4324                 goto again;
4325         }
4326 
4327         if (!curr->sched_class->yield_to_task)
4328                 goto out_unlock;
4329 
4330         if (curr->sched_class != p->sched_class)
4331                 goto out_unlock;
4332 
4333         if (task_running(p_rq, p) || p->state)
4334                 goto out_unlock;
4335 
4336         yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4337         if (yielded) {
4338                 schedstat_inc(rq, yld_count);
4339                 /*
4340                  * Make p's CPU reschedule; pick_next_entity takes care of
4341                  * fairness.
4342                  */
4343                 if (preempt && rq != p_rq)
4344                         resched_curr(p_rq);
4345         }
4346 
4347 out_unlock:
4348         double_rq_unlock(rq, p_rq);
4349 out_irq:
4350         local_irq_restore(flags);
4351 
4352         if (yielded > 0)
4353                 schedule();
4354 
4355         return yielded;
4356 }
4357 EXPORT_SYMBOL_GPL(yield_to);
4358 
4359 /*
4360  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4361  * that process accounting knows that this is a task in IO wait state.
4362  */
4363 long __sched io_schedule_timeout(long timeout)
4364 {
4365         int old_iowait = current->in_iowait;
4366         struct rq *rq;
4367         long ret;
4368 
4369         current->in_iowait = 1;
4370         if (old_iowait)
4371                 blk_schedule_flush_plug(current);
4372         else
4373                 blk_flush_plug(current);
4374 
4375         delayacct_blkio_start();
4376         rq = raw_rq();
4377         atomic_inc(&rq->nr_iowait);
4378         ret = schedule_timeout(timeout);
4379         current->in_iowait = old_iowait;
4380         atomic_dec(&rq->nr_iowait);
4381         delayacct_blkio_end();
4382 
4383         return ret;
4384 }
4385 EXPORT_SYMBOL(io_schedule_timeout);
4386 
4387 /**
4388  * sys_sched_get_priority_max - return maximum RT priority.
4389  * @policy: scheduling class.
4390  *
4391  * Return: On success, this syscall returns the maximum
4392  * rt_priority that can be used by a given scheduling class.
4393  * On failure, a negative error code is returned.
4394  */
4395 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4396 {
4397         int ret = -EINVAL;
4398 
4399         switch (policy) {
4400         case SCHED_FIFO:
4401         case SCHED_RR:
4402                 ret = MAX_USER_RT_PRIO-1;
4403                 break;
4404         case SCHED_DEADLINE:
4405         case SCHED_NORMAL:
4406         case SCHED_BATCH:
4407         case SCHED_IDLE:
4408                 ret = 0;
4409                 break;
4410         }
4411         return ret;
4412 }
4413 
4414 /**
4415  * sys_sched_get_priority_min - return minimum RT priority.
4416  * @policy: scheduling class.
4417  *
4418  * Return: On success, this syscall returns the minimum
4419  * rt_priority that can be used by a given scheduling class.
4420  * On failure, a negative error code is returned.
4421  */
4422 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4423 {
4424         int ret = -EINVAL;
4425 
4426         switch (policy) {
4427         case SCHED_FIFO:
4428         case SCHED_RR:
4429                 ret = 1;
4430                 break;
4431         case SCHED_DEADLINE:
4432         case SCHED_NORMAL:
4433         case SCHED_BATCH:
4434         case SCHED_IDLE:
4435                 ret = 0;
4436         }
4437         return ret;
4438 }
4439 
4440 /**
4441  * sys_sched_rr_get_interval - return the default timeslice of a process.
4442  * @pid: pid of the process.
4443  * @interval: userspace pointer to the timeslice value.
4444  *
4445  * this syscall writes the default timeslice value of a given process
4446  * into the user-space timespec buffer. A value of '' means infinity.
4447  *
4448  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4449  * an error code.
4450  */
4451 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4452                 struct timespec __user *, interval)
4453 {
4454         struct task_struct *p;
4455         unsigned int time_slice;
4456         unsigned long flags;
4457         struct rq *rq;
4458         int retval;
4459         struct timespec t;
4460 
4461         if (pid < 0)
4462                 return -EINVAL;
4463 
4464         retval = -ESRCH;
4465         rcu_read_lock();
4466         p = find_process_by_pid(pid);
4467         if (!p)
4468                 goto out_unlock;
4469 
4470         retval = security_task_getscheduler(p);
4471         if (retval)
4472                 goto out_unlock;
4473 
4474         rq = task_rq_lock(p, &flags);
4475         time_slice = 0;
4476         if (p->sched_class->get_rr_interval)
4477                 time_slice = p->sched_class->get_rr_interval(rq, p);
4478         task_rq_unlock(rq, p, &flags);
4479 
4480         rcu_read_unlock();
4481         jiffies_to_timespec(time_slice, &t);
4482         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4483         return retval;
4484 
4485 out_unlock:
4486         rcu_read_unlock();
4487         return retval;
4488 }
4489 
4490 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4491 
4492 void sched_show_task(struct task_struct *p)
4493 {
4494         unsigned long free = 0;
4495         int ppid;
4496         unsigned long state = p->state;
4497 
4498         if (state)
4499                 state = __ffs(state) + 1;
4500         printk(KERN_INFO "%-15.15s %c", p->comm,
4501                 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4502 #if BITS_PER_LONG == 32
4503         if (state == TASK_RUNNING)
4504                 printk(KERN_CONT " running  ");
4505         else
4506                 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4507 #else
4508         if (state == TASK_RUNNING)
4509                 printk(KERN_CONT "  running task    ");
4510         else
4511                 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4512 #endif
4513 #ifdef CONFIG_DEBUG_STACK_USAGE
4514         free = stack_not_used(p);
4515 #endif
4516         ppid = 0;
4517         rcu_read_lock();
4518         if (pid_alive(p))
4519                 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4520         rcu_read_unlock();
4521         printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4522                 task_pid_nr(p), ppid,
4523                 (unsigned long)task_thread_info(p)->flags);
4524 
4525         print_worker_info(KERN_INFO, p);
4526         show_stack(p, NULL);
4527 }
4528 
4529 void show_state_filter(unsigned long state_filter)
4530 {
4531         struct task_struct *g, *p;
4532 
4533 #if BITS_PER_LONG == 32
4534         printk(KERN_INFO
4535                 "  task                PC stack   pid father\n");
4536 #else
4537         printk(KERN_INFO
4538                 "  task                        PC stack   pid father\n");
4539 #endif
4540         rcu_read_lock();
4541         for_each_process_thread(g, p) {
4542                 /*
4543                  * reset the NMI-timeout, listing all files on a slow
4544                  * console might take a lot of time:
4545                  */
4546                 touch_nmi_watchdog();
4547                 if (!state_filter || (p->state & state_filter))
4548                         sched_show_task(p);
4549         }
4550 
4551         touch_all_softlockup_watchdogs();
4552 
4553 #ifdef CONFIG_SCHED_DEBUG
4554         sysrq_sched_debug_show();
4555 #endif
4556         rcu_read_unlock();
4557         /*
4558          * Only show locks if all tasks are dumped:
4559          */
4560         if (!state_filter)
4561                 debug_show_all_locks();
4562 }
4563 
4564 void init_idle_bootup_task(struct task_struct *idle)
4565 {
4566         idle->sched_class = &idle_sched_class;
4567 }
4568 
4569 /**
4570  * init_idle - set up an idle thread for a given CPU
4571  * @idle: task in question
4572  * @cpu: cpu the idle task belongs to
4573  *
4574  * NOTE: this function does not set the idle thread's NEED_RESCHED
4575  * flag, to make booting more robust.
4576  */
4577 void init_idle(struct task_struct *idle, int cpu)
4578 {
4579         struct rq *rq = cpu_rq(cpu);
4580         unsigned long flags;
4581 
4582         raw_spin_lock_irqsave(&rq->lock, flags);
4583 
4584         __sched_fork(0, idle);
4585         idle->state = TASK_RUNNING;
4586         idle->se.exec_start = sched_clock();
4587 
4588         do_set_cpus_allowed(idle, cpumask_of(cpu));
4589         /*
4590          * We're having a chicken and egg problem, even though we are
4591          * holding rq->lock, the cpu isn't yet set to this cpu so the
4592          * lockdep check in task_group() will fail.
4593          *
4594          * Similar case to sched_fork(). / Alternatively we could
4595          * use task_rq_lock() here and obtain the other rq->lock.
4596          *
4597          * Silence PROVE_RCU
4598          */
4599         rcu_read_lock();
4600         __set_task_cpu(idle, cpu);
4601         rcu_read_unlock();
4602 
4603         rq->curr = rq->idle = idle;
4604         idle->on_rq = TASK_ON_RQ_QUEUED;
4605 #if defined(CONFIG_SMP)
4606         idle->on_cpu = 1;
4607 #endif
4608         raw_spin_unlock_irqrestore(&rq->lock, flags);
4609 
4610         /* Set the preempt count _outside_ the spinlocks! */
4611         init_idle_preempt_count(idle, cpu);
4612 
4613         /*
4614          * The idle tasks have their own, simple scheduling class:
4615          */
4616         idle->sched_class = &idle_sched_class;
4617         ftrace_graph_init_idle_task(idle, cpu);
4618         vtime_init_idle(idle, cpu);
4619 #if defined(CONFIG_SMP)
4620         sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4621 #endif
4622 }
4623 
4624 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4625                               const struct cpumask *trial)
4626 {
4627         int ret = 1, trial_cpus;
4628         struct dl_bw *cur_dl_b;
4629         unsigned long flags;
4630 
4631         if (!cpumask_weight(cur))
4632                 return ret;
4633 
4634         rcu_read_lock_sched();
4635         cur_dl_b = dl_bw_of(cpumask_any(cur));
4636         trial_cpus = cpumask_weight(trial);
4637 
4638         raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4639         if (cur_dl_b->bw != -1 &&
4640             cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4641                 ret = 0;
4642         raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4643         rcu_read_unlock_sched();
4644 
4645         return ret;
4646 }
4647 
4648 int task_can_attach(struct task_struct *p,
4649                     const struct cpumask *cs_cpus_allowed)
4650 {
4651         int ret = 0;
4652 
4653         /*
4654          * Kthreads which disallow setaffinity shouldn't be moved
4655          * to a new cpuset; we don't want to change their cpu
4656          * affinity and isolating such threads by their set of
4657          * allowed nodes is unnecessary.  Thus, cpusets are not
4658          * applicable for such threads.  This prevents checking for
4659          * success of set_cpus_allowed_ptr() on all attached tasks
4660          * before cpus_allowed may be changed.
4661          */
4662         if (p->flags & PF_NO_SETAFFINITY) {
4663                 ret = -EINVAL;
4664                 goto out;
4665         }
4666 
4667 #ifdef CONFIG_SMP
4668         if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4669                                               cs_cpus_allowed)) {
4670                 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4671                                                         cs_cpus_allowed);
4672                 struct dl_bw *dl_b;
4673                 bool overflow;
4674                 int cpus;
4675                 unsigned long flags;
4676 
4677                 rcu_read_lock_sched();
4678                 dl_b = dl_bw_of(dest_cpu);
4679                 raw_spin_lock_irqsave(&dl_b->lock, flags);
4680                 cpus = dl_bw_cpus(dest_cpu);
4681                 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4682                 if (overflow)
4683                         ret = -EBUSY;
4684                 else {
4685                         /*
4686                          * We reserve space for this task in the destination
4687                          * root_domain, as we can't fail after this point.
4688                          * We will free resources in the source root_domain
4689                          * later on (see set_cpus_allowed_dl()).
4690                          */
4691                         __dl_add(dl_b, p->dl.dl_bw);
4692                 }
4693                 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4694                 rcu_read_unlock_sched();
4695 
4696         }
4697 #endif
4698 out:
4699         return ret;
4700 }
4701 
4702 #ifdef CONFIG_SMP
4703 /*
4704  * move_queued_task - move a queued task to new rq.
4705  *
4706  * Returns (locked) new rq. Old rq's lock is released.
4707  */
4708 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4709 {
4710         struct rq *rq = task_rq(p);
4711 
4712         lockdep_assert_held(&rq->lock);
4713 
4714         dequeue_task(rq, p, 0);
4715         p->on_rq = TASK_ON_RQ_MIGRATING;
4716         set_task_cpu(p, new_cpu);
4717         raw_spin_unlock(&rq->lock);
4718 
4719         rq = cpu_rq(new_cpu);
4720 
4721         raw_spin_lock(&rq->lock);
4722         BUG_ON(task_cpu(p) != new_cpu);
4723         p->on_rq = TASK_ON_RQ_QUEUED;
4724         enqueue_task(rq, p, 0);
4725         check_preempt_curr(rq, p, 0);
4726 
4727         return rq;
4728 }
4729 
4730 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4731 {
4732         if (p->sched_class->set_cpus_allowed)
4733                 p->sched_class->set_cpus_allowed(p, new_mask);
4734 
4735         cpumask_copy(&p->cpus_allowed, new_mask);
4736         p->nr_cpus_allowed = cpumask_weight(new_mask);
4737 }
4738 
4739 /*
4740  * This is how migration works:
4741  *
4742  * 1) we invoke migration_cpu_stop() on the target CPU using
4743  *    stop_one_cpu().
4744  * 2) stopper starts to run (implicitly forcing the migrated thread
4745  *    off the CPU)
4746  * 3) it checks whether the migrated task is still in the wrong runqueue.
4747  * 4) if it's in the wrong runqueue then the migration thread removes
4748  *    it and puts it into the right queue.
4749  * 5) stopper completes and stop_one_cpu() returns and the migration
4750  *    is done.
4751  */
4752 
4753 /*
4754  * Change a given task's CPU affinity. Migrate the thread to a
4755  * proper CPU and schedule it away if the CPU it's executing on
4756  * is removed from the allowed bitmask.
4757  *
4758  * NOTE: the caller must have a valid reference to the task, the
4759  * task must not exit() & deallocate itself prematurely. The
4760  * call is not atomic; no spinlocks may be held.
4761  */
4762 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4763 {
4764         unsigned long flags;
4765         struct rq *rq;
4766         unsigned int dest_cpu;
4767         int ret = 0;
4768 
4769         rq = task_rq_lock(p, &flags);
4770 
4771         if (cpumask_equal(&p->cpus_allowed, new_mask))
4772                 goto out;
4773 
4774         if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4775                 ret = -EINVAL;
4776                 goto out;
4777         }
4778 
4779         do_set_cpus_allowed(p, new_mask);
4780 
4781         /* Can the task run on the task's current CPU? If so, we're done */
4782         if (cpumask_test_cpu(task_cpu(p), new_mask))
4783                 goto out;
4784 
4785         dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4786         if (task_running(rq, p) || p->state == TASK_WAKING) {
4787                 struct migration_arg arg = { p, dest_cpu };
4788                 /* Need help from migration thread: drop lock and wait. */
4789                 task_rq_unlock(rq, p, &flags);
4790                 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4791                 tlb_migrate_finish(p->mm);
4792                 return 0;
4793         } else if (task_on_rq_queued(p))
4794                 rq = move_queued_task(p, dest_cpu);
4795 out:
4796         task_rq_unlock(rq, p, &flags);
4797 
4798         return ret;
4799 }
4800 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4801 
4802 /*
4803  * Move (not current) task off this cpu, onto dest cpu. We're doing
4804  * this because either it can't run here any more (set_cpus_allowed()
4805  * away from this CPU, or CPU going down), or because we're
4806  * attempting to rebalance this task on exec (sched_exec).
4807  *
4808  * So we race with normal scheduler movements, but that's OK, as long
4809  * as the task is no longer on this CPU.
4810  *
4811  * Returns non-zero if task was successfully migrated.
4812  */
4813 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4814 {
4815         struct rq *rq;
4816         int ret = 0;
4817 
4818         if (unlikely(!cpu_active(dest_cpu)))
4819                 return ret;
4820 
4821         rq = cpu_rq(src_cpu);
4822 
4823         raw_spin_lock(&p->pi_lock);
4824         raw_spin_lock(&rq->lock);
4825         /* Already moved. */
4826         if (task_cpu(p) != src_cpu)
4827                 goto done;
4828 
4829         /* Affinity changed (again). */
4830         if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4831                 goto fail;
4832 
4833         /*
4834          * If we're not on a rq, the next wake-up will ensure we're
4835          * placed properly.
4836          */
4837         if (task_on_rq_queued(p))
4838                 rq = move_queued_task(p, dest_cpu);
4839 done:
4840         ret = 1;
4841 fail:
4842         raw_spin_unlock(&rq->lock);
4843         raw_spin_unlock(&p->pi_lock);
4844         return ret;
4845 }
4846 
4847 #ifdef CONFIG_NUMA_BALANCING
4848 /* Migrate current task p to target_cpu */
4849 int migrate_task_to(struct task_struct *p, int target_cpu)
4850 {
4851         struct migration_arg arg = { p, target_cpu };
4852         int curr_cpu = task_cpu(p);
4853 
4854         if (curr_cpu == target_cpu)
4855                 return 0;
4856 
4857         if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4858                 return -EINVAL;
4859 
4860         /* TODO: This is not properly updating schedstats */
4861 
4862         trace_sched_move_numa(p, curr_cpu, target_cpu);
4863         return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4864 }
4865 
4866 /*
4867  * Requeue a task on a given node and accurately track the number of NUMA
4868  * tasks on the runqueues
4869  */
4870 void sched_setnuma(struct task_struct *p, int nid)
4871 {
4872         struct rq *rq;
4873         unsigned long flags;
4874         bool queued, running;
4875 
4876         rq = task_rq_lock(p, &flags);
4877         queued = task_on_rq_queued(p);
4878         running = task_current(rq, p);
4879 
4880         if (queued)
4881                 dequeue_task(rq, p, 0);
4882         if (running)
4883                 put_prev_task(rq, p);
4884 
4885         p->numa_preferred_nid = nid;
4886 
4887         if (running)
4888                 p->sched_class->set_curr_task(rq);
4889         if (queued)
4890                 enqueue_task(rq, p, 0);
4891         task_rq_unlock(rq, p, &flags);
4892 }
4893 #endif
4894 
4895 /*
4896  * migration_cpu_stop - this will be executed by a highprio stopper thread
4897  * and performs thread migration by bumping thread off CPU then
4898  * 'pushing' onto another runqueue.
4899  */
4900 static int migration_cpu_stop(void *data)
4901 {
4902         struct migration_arg *arg = data;
4903 
4904         /*
4905          * The original target cpu might have gone down and we might
4906          * be on another cpu but it doesn't matter.
4907          */
4908         local_irq_disable();
4909         /*
4910          * We need to explicitly wake pending tasks before running
4911          * __migrate_task() such that we will not miss enforcing cpus_allowed
4912          * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4913          */
4914         sched_ttwu_pending();
4915         __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4916         local_irq_enable();
4917         return 0;
4918 }
4919 
4920 #ifdef CONFIG_HOTPLUG_CPU
4921 
4922 /*
4923  * Ensures that the idle task is using init_mm right before its cpu goes
4924  * offline.
4925  */
4926 void idle_task_exit(void)
4927 {
4928         struct mm_struct *mm = current->active_mm;
4929 
4930         BUG_ON(cpu_online(smp_processor_id()));
4931 
4932         if (mm != &init_mm) {
4933                 switch_mm(mm, &init_mm, current);
4934                 finish_arch_post_lock_switch();
4935         }
4936         mmdrop(mm);
4937 }
4938 
4939 /*
4940  * Since this CPU is going 'away' for a while, fold any nr_active delta
4941  * we might have. Assumes we're called after migrate_tasks() so that the
4942  * nr_active count is stable.
4943  *
4944  * Also see the comment "Global load-average calculations".
4945  */
4946 static void calc_load_migrate(struct rq *rq)
4947 {
4948         long delta = calc_load_fold_active(rq);
4949         if (delta)
4950                 atomic_long_add(delta, &calc_load_tasks);
4951 }
4952 
4953 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4954 {
4955 }
4956 
4957 static const struct sched_class fake_sched_class = {
4958         .put_prev_task = put_prev_task_fake,
4959 };
4960 
4961 static struct task_struct fake_task = {
4962         /*
4963          * Avoid pull_{rt,dl}_task()
4964          */
4965         .prio = MAX_PRIO + 1,
4966         .sched_class = &fake_sched_class,
4967 };
4968 
4969 /*
4970  * Migrate all tasks from the rq, sleeping tasks will be migrated by
4971  * try_to_wake_up()->select_task_rq().
4972  *
4973  * Called with rq->lock held even though we'er in stop_machine() and
4974  * there's no concurrency possible, we hold the required locks anyway
4975  * because of lock validation efforts.
4976  */
4977 static void migrate_tasks(unsigned int dead_cpu)
4978 {
4979         struct rq *rq = cpu_rq(dead_cpu);
4980         struct task_struct *next, *stop = rq->stop;
4981         int dest_cpu;
4982 
4983         /*
4984          * Fudge the rq selection such that the below task selection loop
4985          * doesn't get stuck on the currently eligible stop task.
4986          *
4987          * We're currently inside stop_machine() and the rq is either stuck
4988          * in the stop_machine_cpu_stop() loop, or we're executing this code,
4989          * either way we should never end up calling schedule() until we're
4990          * done here.
4991          */
4992         rq->stop = NULL;
4993 
4994         /*
4995          * put_prev_task() and pick_next_task() sched
4996          * class method both need to have an up-to-date
4997          * value of rq->clock[_task]
4998          */
4999         update_rq_clock(rq);
5000 
5001         for ( ; ; ) {
5002                 /*
5003                  * There's this thread running, bail when that's the only
5004                  * remaining thread.
5005                  */
5006                 if (rq->nr_running == 1)
5007                         break;
5008 
5009                 next = pick_next_task(rq, &fake_task);
5010                 BUG_ON(!next);
5011                 next->sched_class->put_prev_task(rq, next);
5012 
5013                 /* Find suitable destination for @next, with force if needed. */
5014                 dest_cpu = select_fallback_rq(dead_cpu, next);
5015                 raw_spin_unlock(&rq->lock);
5016 
5017                 __migrate_task(next, dead_cpu, dest_cpu);
5018 
5019                 raw_spin_lock(&rq->lock);
5020         }
5021 
5022         rq->stop = stop;
5023 }
5024 
5025 #endif /* CONFIG_HOTPLUG_CPU */
5026 
5027 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5028 
5029 static struct ctl_table sd_ctl_dir[] = {
5030         {
5031                 .procname       = "sched_domain",
5032                 .mode           = 0555,
5033         },
5034         {}
5035 };
5036 
5037 static struct ctl_table sd_ctl_root[] = {
5038         {
5039                 .procname       = "kernel",
5040                 .mode           = 0555,
5041                 .child          = sd_ctl_dir,
5042         },
5043         {}
5044 };
5045 
5046 static struct ctl_table *sd_alloc_ctl_entry(int n)
5047 {
5048         struct ctl_table *entry =
5049                 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5050 
5051         return entry;
5052 }
5053 
5054 static void sd_free_ctl_entry(struct ctl_table **tablep)
5055 {
5056         struct ctl_table *entry;
5057 
5058         /*
5059          * In the intermediate directories, both the child directory and
5060          * procname are dynamically allocated and could fail but the mode
5061          * will always be set. In the lowest directory the names are
5062          * static strings and all have proc handlers.
5063          */
5064         for (entry = *tablep; entry->mode; entry++) {
5065                 if (entry->child)
5066                         sd_free_ctl_entry(&entry->child);
5067                 if (entry->proc_handler == NULL)
5068                         kfree(entry->procname);
5069         }
5070 
5071         kfree(*tablep);
5072         *tablep = NULL;
5073 }
5074 
5075 static int min_load_idx = 0;
5076 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5077 
5078 static void
5079 set_table_entry(struct ctl_table *entry,
5080                 const char *procname, void *data, int maxlen,
5081                 umode_t mode, proc_handler *proc_handler,
5082                 bool load_idx)
5083 {
5084         entry->procname = procname;
5085         entry->data = data;
5086         entry->maxlen = maxlen;
5087         entry->mode = mode;
5088         entry->proc_handler = proc_handler;
5089 
5090         if (load_idx) {
5091                 entry->extra1 = &min_load_idx;
5092                 entry->extra2 = &max_load_idx;
5093         }
5094 }
5095 
5096 static struct ctl_table *
5097 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5098 {
5099         struct ctl_table *table = sd_alloc_ctl_entry(14);
5100 
5101         if (table == NULL)
5102                 return NULL;
5103 
5104         set_table_entry(&table[0], "min_interval", &sd->min_interval,
5105                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5106         set_table_entry(&table[1], "max_interval", &sd->max_interval,
5107                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5108         set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5109                 sizeof(int), 0644, proc_dointvec_minmax, true);
5110         set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5111                 sizeof(int), 0644, proc_dointvec_minmax, true);
5112         set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5113                 sizeof(int), 0644, proc_dointvec_minmax, true);
5114         set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5115                 sizeof(int), 0644, proc_dointvec_minmax, true);
5116         set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5117                 sizeof(int), 0644, proc_dointvec_minmax, true);
5118         set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5119                 sizeof(int), 0644, proc_dointvec_minmax, false);
5120         set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5121                 sizeof(int), 0644, proc_dointvec_minmax, false);
5122         set_table_entry(&table[9], "cache_nice_tries",
5123                 &sd->cache_nice_tries,
5124                 sizeof(int), 0644, proc_dointvec_minmax, false);
5125         set_table_entry(&table[10], "flags", &sd->flags,
5126                 sizeof(int), 0644, proc_dointvec_minmax, false);
5127         set_table_entry(&table[11], "max_newidle_lb_cost",
5128                 &sd->max_newidle_lb_cost,
5129                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5130         set_table_entry(&table[12], "name", sd->name,
5131                 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5132         /* &table[13] is terminator */
5133 
5134         return table;
5135 }
5136 
5137 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5138 {
5139         struct ctl_table *entry, *table;
5140         struct sched_domain *sd;
5141         int domain_num = 0, i;
5142         char buf[32];
5143 
5144         for_each_domain(cpu, sd)
5145                 domain_num++;
5146         entry = table = sd_alloc_ctl_entry(domain_num + 1);
5147         if (table == NULL)
5148                 return NULL;
5149 
5150         i = 0;
5151         for_each_domain(cpu, sd) {
5152                 snprintf(buf, 32, "domain%d", i);
5153                 entry->procname = kstrdup(buf, GFP_KERNEL);
5154                 entry->mode = 0555;
5155                 entry->child = sd_alloc_ctl_domain_table(sd);
5156                 entry++;
5157                 i++;
5158         }
5159         return table;
5160 }
5161 
5162 static struct ctl_table_header *sd_sysctl_header;
5163 static void register_sched_domain_sysctl(void)
5164 {
5165         int i, cpu_num = num_possible_cpus();
5166         struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5167         char buf[32];
5168 
5169         WARN_ON(sd_ctl_dir[0].child);
5170         sd_ctl_dir[0].child = entry;
5171 
5172         if (entry == NULL)
5173                 return;
5174 
5175         for_each_possible_cpu(i) {
5176                 snprintf(buf, 32, "cpu%d", i);
5177                 entry->procname = kstrdup(buf, GFP_KERNEL);
5178                 entry->mode = 0555;
5179                 entry->child = sd_alloc_ctl_cpu_table(i);
5180                 entry++;
5181         }
5182 
5183         WARN_ON(sd_sysctl_header);
5184         sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5185 }
5186 
5187 /* may be called multiple times per register */
5188 static void unregister_sched_domain_sysctl(void)
5189 {
5190         if (sd_sysctl_header)
5191                 unregister_sysctl_table(sd_sysctl_header);
5192         sd_sysctl_header = NULL;
5193         if (sd_ctl_dir[0].child)
5194                 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5195 }
5196 #else
5197 static void register_sched_domain_sysctl(void)
5198 {
5199 }
5200 static void unregister_sched_domain_sysctl(void)
5201 {
5202 }
5203 #endif
5204 
5205 static void set_rq_online(struct rq *rq)
5206 {
5207         if (!rq->online) {
5208                 const struct sched_class *class;
5209 
5210                 cpumask_set_cpu(rq->cpu, rq->rd->online);
5211                 rq->online = 1;
5212 
5213                 for_each_class(class) {
5214                         if (class->rq_online)
5215                                 class->rq_online(rq);
5216                 }
5217         }
5218 }
5219 
5220 static void set_rq_offline(struct rq *rq)
5221 {
5222         if (rq->online) {
5223                 const struct sched_class *class;
5224 
5225                 for_each_class(class) {
5226                         if (class->rq_offline)
5227                                 class->rq_offline(rq);
5228                 }
5229 
5230                 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5231                 rq->online = 0;
5232         }
5233 }
5234 
5235 /*
5236  * migration_call - callback that gets triggered when a CPU is added.
5237  * Here we can start up the necessary migration thread for the new CPU.
5238  */
5239 static int
5240 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5241 {
5242         int cpu = (long)hcpu;
5243         unsigned long flags;
5244         struct rq *rq = cpu_rq(cpu);
5245 
5246         switch (action & ~CPU_TASKS_FROZEN) {
5247 
5248         case CPU_UP_PREPARE:
5249                 rq->calc_load_update = calc_load_update;
5250                 break;
5251 
5252         case CPU_ONLINE:
5253                 /* Update our root-domain */
5254                 raw_spin_lock_irqsave(&rq->lock, flags);
5255                 if (rq->rd) {
5256                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5257 
5258                         set_rq_online(rq);
5259                 }
5260                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5261                 break;
5262 
5263 #ifdef CONFIG_HOTPLUG_CPU
5264         case CPU_DYING:
5265                 sched_ttwu_pending();
5266                 /* Update our root-domain */
5267                 raw_spin_lock_irqsave(&rq->lock, flags);
5268                 if (rq->rd) {
5269                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5270                         set_rq_offline(rq);
5271                 }
5272                 migrate_tasks(cpu);
5273                 BUG_ON(rq->nr_running != 1); /* the migration thread */
5274                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5275                 break;
5276 
5277         case CPU_DEAD:
5278                 calc_load_migrate(rq);
5279                 break;
5280 #endif
5281         }
5282 
5283         update_max_interval();
5284 
5285         return NOTIFY_OK;
5286 }
5287 
5288 /*
5289  * Register at high priority so that task migration (migrate_all_tasks)
5290  * happens before everything else.  This has to be lower priority than
5291  * the notifier in the perf_event subsystem, though.
5292  */
5293 static struct notifier_block migration_notifier = {
5294         .notifier_call = migration_call,
5295         .priority = CPU_PRI_MIGRATION,
5296 };
5297 
5298 static void __cpuinit set_cpu_rq_start_time(void)
5299 {
5300         int cpu = smp_processor_id();
5301         struct rq *rq = cpu_rq(cpu);
5302         rq->age_stamp = sched_clock_cpu(cpu);
5303 }
5304 
5305 static int sched_cpu_active(struct notifier_block *nfb,
5306                                       unsigned long action, void *hcpu)
5307 {
5308         switch (action & ~CPU_TASKS_FROZEN) {
5309         case CPU_STARTING:
5310                 set_cpu_rq_start_time();
5311                 return NOTIFY_OK;
5312         case CPU_DOWN_FAILED:
5313                 set_cpu_active((long)hcpu, true);
5314                 return NOTIFY_OK;
5315         default:
5316                 return NOTIFY_DONE;
5317         }
5318 }
5319 
5320 static int sched_cpu_inactive(struct notifier_block *nfb,
5321                                         unsigned long action, void *hcpu)
5322 {
5323         unsigned long flags;
5324         long cpu = (long)hcpu;
5325         struct dl_bw *dl_b;
5326 
5327         switch (action & ~CPU_TASKS_FROZEN) {
5328         case CPU_DOWN_PREPARE:
5329                 set_cpu_active(cpu, false);
5330 
5331                 /* explicitly allow suspend */
5332                 if (!(action & CPU_TASKS_FROZEN)) {
5333                         bool overflow;
5334                         int cpus;
5335 
5336                         rcu_read_lock_sched();
5337                         dl_b = dl_bw_of(cpu);
5338 
5339                         raw_spin_lock_irqsave(&dl_b->lock, flags);
5340                         cpus = dl_bw_cpus(cpu);
5341                         overflow = __dl_overflow(dl_b, cpus, 0, 0);
5342                         raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5343 
5344                         rcu_read_unlock_sched();
5345 
5346                         if (overflow)
5347                                 return notifier_from_errno(-EBUSY);
5348                 }
5349                 return NOTIFY_OK;
5350         }
5351 
5352         return NOTIFY_DONE;
5353 }
5354 
5355 static int __init migration_init(void)
5356 {
5357         void *cpu = (void *)(long)smp_processor_id();
5358         int err;
5359 
5360         /* Initialize migration for the boot CPU */
5361         err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5362         BUG_ON(err == NOTIFY_BAD);
5363         migration_call(&migration_notifier, CPU_ONLINE, cpu);
5364         register_cpu_notifier(&migration_notifier);
5365 
5366         /* Register cpu active notifiers */
5367         cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5368         cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5369 
5370         return 0;
5371 }
5372 early_initcall(migration_init);
5373 #endif
5374 
5375 #ifdef CONFIG_SMP
5376 
5377 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5378 
5379 #ifdef CONFIG_SCHED_DEBUG
5380 
5381 static __read_mostly int sched_debug_enabled;
5382 
5383 static int __init sched_debug_setup(char *str)
5384 {
5385         sched_debug_enabled = 1;
5386 
5387         return 0;
5388 }
5389 early_param("sched_debug", sched_debug_setup);
5390 
5391 static inline bool sched_debug(void)
5392 {
5393         return sched_debug_enabled;
5394 }
5395 
5396 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5397                                   struct cpumask *groupmask)
5398 {
5399         struct sched_group *group = sd->groups;
5400 
5401         cpumask_clear(groupmask);
5402 
5403         printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5404 
5405         if (!(sd->flags & SD_LOAD_BALANCE)) {
5406                 printk("does not load-balance\n");
5407                 if (sd->parent)
5408                         printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5409                                         " has parent");
5410                 return -1;
5411         }
5412 
5413         printk(KERN_CONT "span %*pbl level %s\n",
5414                cpumask_pr_args(sched_domain_span(sd)), sd->name);
5415 
5416         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5417                 printk(KERN_ERR "ERROR: domain->span does not contain "
5418                                 "CPU%d\n", cpu);
5419         }
5420         if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5421                 printk(KERN_ERR "ERROR: domain->groups does not contain"
5422                                 " CPU%d\n", cpu);
5423         }
5424 
5425         printk(KERN_DEBUG "%*s groups:", level + 1, "");
5426         do {
5427                 if (!group) {
5428                         printk("\n");
5429                         printk(KERN_ERR "ERROR: group is NULL\n");
5430                         break;
5431                 }
5432 
5433                 /*
5434                  * Even though we initialize ->capacity to something semi-sane,
5435                  * we leave capacity_orig unset. This allows us to detect if
5436                  * domain iteration is still funny without causing /0 traps.
5437                  */
5438                 if (!group->sgc->capacity_orig) {
5439                         printk(KERN_CONT "\n");
5440                         printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5441                         break;
5442                 }
5443 
5444                 if (!cpumask_weight(sched_group_cpus(group))) {
5445                         printk(KERN_CONT "\n");
5446                         printk(KERN_ERR "ERROR: empty group\n");
5447                         break;
5448                 }
5449 
5450                 if (!(sd->flags & SD_OVERLAP) &&
5451                     cpumask_intersects(groupmask, sched_group_cpus(group))) {
5452                         printk(KERN_CONT "\n");
5453                         printk(KERN_ERR "ERROR: repeated CPUs\n");
5454                         break;
5455                 }
5456 
5457                 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5458 
5459                 printk(KERN_CONT " %*pbl",
5460                        cpumask_pr_args(sched_group_cpus(group)));
5461                 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5462                         printk(KERN_CONT " (cpu_capacity = %d)",
5463                                 group->sgc->capacity);
5464                 }
5465 
5466                 group = group->next;
5467         } while (group != sd->groups);
5468         printk(KERN_CONT "\n");
5469 
5470         if (!cpumask_equal(sched_domain_span(sd), groupmask))
5471                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5472 
5473         if (sd->parent &&
5474             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5475                 printk(KERN_ERR "ERROR: parent span is not a superset "
5476                         "of domain->span\n");
5477         return 0;
5478 }
5479 
5480 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5481 {
5482         int level = 0;
5483 
5484         if (!sched_debug_enabled)
5485                 return;
5486 
5487         if (!sd) {
5488                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5489                 return;
5490         }
5491 
5492         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5493 
5494         for (;;) {
5495                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5496                         break;
5497                 level++;
5498                 sd = sd->parent;
5499                 if (!sd)
5500                         break;
5501         }
5502 }
5503 #else /* !CONFIG_SCHED_DEBUG */
5504 # define sched_domain_debug(sd, cpu) do { } while (0)
5505 static inline bool sched_debug(void)
5506 {
5507         return false;
5508 }
5509 #endif /* CONFIG_SCHED_DEBUG */
5510 
5511 static int sd_degenerate(struct sched_domain *sd)
5512 {
5513         if (cpumask_weight(sched_domain_span(sd)) == 1)
5514                 return 1;
5515 
5516         /* Following flags need at least 2 groups */
5517         if (sd->flags & (SD_LOAD_BALANCE |
5518                          SD_BALANCE_NEWIDLE |
5519                          SD_BALANCE_FORK |
5520                          SD_BALANCE_EXEC |
5521                          SD_SHARE_CPUCAPACITY |
5522                          SD_SHARE_PKG_RESOURCES |
5523                          SD_SHARE_POWERDOMAIN)) {
5524                 if (sd->groups != sd->groups->next)
5525                         return 0;
5526         }
5527 
5528         /* Following flags don't use groups */
5529         if (sd->flags & (SD_WAKE_AFFINE))
5530                 return 0;
5531 
5532         return 1;
5533 }
5534 
5535 static int
5536 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5537 {
5538         unsigned long cflags = sd->flags, pflags = parent->flags;
5539 
5540         if (sd_degenerate(parent))
5541                 return 1;
5542 
5543         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5544                 return 0;
5545 
5546         /* Flags needing groups don't count if only 1 group in parent */
5547         if (parent->groups == parent->groups->next) {
5548                 pflags &= ~(SD_LOAD_BALANCE |
5549                                 SD_BALANCE_NEWIDLE |
5550                                 SD_BALANCE_FORK |
5551                                 SD_BALANCE_EXEC |
5552                                 SD_SHARE_CPUCAPACITY |
5553                                 SD_SHARE_PKG_RESOURCES |
5554                                 SD_PREFER_SIBLING |
5555                                 SD_SHARE_POWERDOMAIN);
5556                 if (nr_node_ids == 1)
5557                         pflags &= ~SD_SERIALIZE;
5558         }
5559         if (~cflags & pflags)
5560                 return 0;
5561 
5562         return 1;
5563 }
5564 
5565 static void free_rootdomain(struct rcu_head *rcu)
5566 {
5567         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5568 
5569         cpupri_cleanup(&rd->cpupri);
5570         cpudl_cleanup(&rd->cpudl);
5571         free_cpumask_var(rd->dlo_mask);
5572         free_cpumask_var(rd->rto_mask);
5573         free_cpumask_var(rd->online);
5574         free_cpumask_var(rd->span);
5575         kfree(rd);
5576 }
5577 
5578 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5579 {
5580         struct root_domain *old_rd = NULL;
5581         unsigned long flags;
5582 
5583         raw_spin_lock_irqsave(&rq->lock, flags);
5584 
5585         if (rq->rd) {
5586                 old_rd = rq->rd;
5587 
5588                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5589                         set_rq_offline(rq);
5590 
5591                 cpumask_clear_cpu(rq->cpu, old_rd->span);
5592 
5593                 /*
5594                  * If we dont want to free the old_rd yet then
5595                  * set old_rd to NULL to skip the freeing later
5596                  * in this function:
5597                  */
5598                 if (!atomic_dec_and_test(&old_rd->refcount))
5599                         old_rd = NULL;
5600         }
5601 
5602         atomic_inc(&rd->refcount);
5603         rq->rd = rd;
5604 
5605         cpumask_set_cpu(rq->cpu, rd->span);
5606         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5607                 set_rq_online(rq);
5608 
5609         raw_spin_unlock_irqrestore(&rq->lock, flags);
5610 
5611         if (old_rd)
5612                 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5613 }
5614 
5615 static int init_rootdomain(struct root_domain *rd)
5616 {
5617         memset(rd, 0, sizeof(*rd));
5618 
5619         if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5620                 goto out;
5621         if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5622                 goto free_span;
5623         if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5624                 goto free_online;
5625         if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5626                 goto free_dlo_mask;
5627 
5628         init_dl_bw(&rd->dl_bw);
5629         if (cpudl_init(&rd->cpudl) != 0)
5630                 goto free_dlo_mask;
5631 
5632         if (cpupri_init(&rd->cpupri) != 0)
5633                 goto free_rto_mask;
5634         return 0;
5635 
5636 free_rto_mask:
5637         free_cpumask_var(rd->rto_mask);
5638 free_dlo_mask:
5639         free_cpumask_var(rd->dlo_mask);
5640 free_online:
5641         free_cpumask_var(rd->online);
5642 free_span:
5643         free_cpumask_var(rd->span);
5644 out:
5645         return -ENOMEM;
5646 }
5647 
5648 /*
5649  * By default the system creates a single root-domain with all cpus as
5650  * members (mimicking the global state we have today).
5651  */
5652 struct root_domain def_root_domain;
5653 
5654 static void init_defrootdomain(void)
5655 {
5656         init_rootdomain(&def_root_domain);
5657 
5658         atomic_set(&def_root_domain.refcount, 1);
5659 }
5660 
5661 static struct root_domain *alloc_rootdomain(void)
5662 {
5663         struct root_domain *rd;
5664 
5665         rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5666         if (!rd)
5667                 return NULL;
5668 
5669         if (init_rootdomain(rd) != 0) {
5670                 kfree(rd);
5671                 return NULL;
5672         }
5673 
5674         return rd;
5675 }
5676 
5677 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5678 {
5679         struct sched_group *tmp, *first;
5680 
5681         if (!sg)
5682                 return;
5683 
5684         first = sg;
5685         do {
5686                 tmp = sg->next;
5687 
5688                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5689                         kfree(sg->sgc);
5690 
5691                 kfree(sg);
5692                 sg = tmp;
5693         } while (sg != first);
5694 }
5695 
5696 static void free_sched_domain(struct rcu_head *rcu)
5697 {
5698         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5699 
5700         /*
5701          * If its an overlapping domain it has private groups, iterate and
5702          * nuke them all.
5703          */
5704         if (sd->flags & SD_OVERLAP) {
5705                 free_sched_groups(sd->groups, 1);
5706         } else if (atomic_dec_and_test(&sd->groups->ref)) {
5707                 kfree(sd->groups->sgc);
5708                 kfree(sd->groups);
5709         }
5710         kfree(sd);
5711 }
5712 
5713 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5714 {
5715         call_rcu(&sd->rcu, free_sched_domain);
5716 }
5717 
5718 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5719 {
5720         for (; sd; sd = sd->parent)
5721                 destroy_sched_domain(sd, cpu);
5722 }
5723 
5724 /*
5725  * Keep a special pointer to the highest sched_domain that has
5726  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5727  * allows us to avoid some pointer chasing select_idle_sibling().
5728  *
5729  * Also keep a unique ID per domain (we use the first cpu number in
5730  * the cpumask of the domain), this allows us to quickly tell if
5731  * two cpus are in the same cache domain, see cpus_share_cache().
5732  */
5733 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5734 DEFINE_PER_CPU(int, sd_llc_size);
5735 DEFINE_PER_CPU(int, sd_llc_id);
5736 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5737 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5738 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5739 
5740 static void update_top_cache_domain(int cpu)
5741 {
5742         struct sched_domain *sd;
5743         struct sched_domain *busy_sd = NULL;
5744         int id = cpu;
5745         int size = 1;
5746 
5747         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5748         if (sd) {
5749                 id = cpumask_first(sched_domain_span(sd));
5750                 size = cpumask_weight(sched_domain_span(sd));
5751                 busy_sd = sd->parent; /* sd_busy */
5752         }
5753         rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5754 
5755         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5756         per_cpu(sd_llc_size, cpu) = size;
5757         per_cpu(sd_llc_id, cpu) = id;
5758 
5759         sd = lowest_flag_domain(cpu, SD_NUMA);
5760         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5761 
5762         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5763         rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5764 }
5765 
5766 /*
5767  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5768  * hold the hotplug lock.
5769  */
5770 static void
5771 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5772 {
5773         struct rq *rq = cpu_rq(cpu);
5774         struct sched_domain *tmp;
5775 
5776         /* Remove the sched domains which do not contribute to scheduling. */
5777         for (tmp = sd; tmp; ) {
5778                 struct sched_domain *parent = tmp->parent;
5779                 if (!parent)
5780                         break;
5781 
5782                 if (sd_parent_degenerate(tmp, parent)) {
5783                         tmp->parent = parent->parent;
5784                         if (parent->parent)
5785                                 parent->parent->child = tmp;
5786                         /*
5787                          * Transfer SD_PREFER_SIBLING down in case of a
5788                          * degenerate parent; the spans match for this
5789                          * so the property transfers.
5790                          */
5791                         if (parent->flags & SD_PREFER_SIBLING)
5792                                 tmp->flags |= SD_PREFER_SIBLING;
5793                         destroy_sched_domain(parent, cpu);
5794                 } else
5795                         tmp = tmp->parent;
5796         }
5797 
5798         if (sd && sd_degenerate(sd)) {
5799                 tmp = sd;
5800                 sd = sd->parent;
5801                 destroy_sched_domain(tmp, cpu);
5802                 if (sd)
5803                         sd->child = NULL;
5804         }
5805 
5806         sched_domain_debug(sd, cpu);
5807 
5808         rq_attach_root(rq, rd);
5809         tmp = rq->sd;
5810         rcu_assign_pointer(rq->sd, sd);
5811         destroy_sched_domains(tmp, cpu);
5812 
5813         update_top_cache_domain(cpu);
5814 }
5815 
5816 /* cpus with isolated domains */
5817 static cpumask_var_t cpu_isolated_map;
5818 
5819 /* Setup the mask of cpus configured for isolated domains */
5820 static int __init isolated_cpu_setup(char *str)
5821 {
5822         alloc_bootmem_cpumask_var(&cpu_isolated_map);
5823         cpulist_parse(str, cpu_isolated_map);
5824         return 1;
5825 }
5826 
5827 __setup("isolcpus=", isolated_cpu_setup);
5828 
5829 struct s_data {
5830         struct sched_domain ** __percpu sd;
5831         struct root_domain      *rd;
5832 };
5833 
5834 enum s_alloc {
5835         sa_rootdomain,
5836         sa_sd,
5837         sa_sd_storage,
5838         sa_none,
5839 };
5840 
5841 /*
5842  * Build an iteration mask that can exclude certain CPUs from the upwards
5843  * domain traversal.
5844  *
5845  * Asymmetric node setups can result in situations where the domain tree is of
5846  * unequal depth, make sure to skip domains that already cover the entire
5847  * range.
5848  *
5849  * In that case build_sched_domains() will have terminated the iteration early
5850  * and our sibling sd spans will be empty. Domains should always include the
5851  * cpu they're built on, so check that.
5852  *
5853  */
5854 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5855 {
5856         const struct cpumask *span = sched_domain_span(sd);
5857         struct sd_data *sdd = sd->private;
5858         struct sched_domain *sibling;
5859         int i;
5860 
5861         for_each_cpu(i, span) {
5862                 sibling = *per_cpu_ptr(sdd->sd, i);
5863                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5864                         continue;
5865 
5866                 cpumask_set_cpu(i, sched_group_mask(sg));
5867         }
5868 }
5869 
5870 /*
5871  * Return the canonical balance cpu for this group, this is the first cpu
5872  * of this group that's also in the iteration mask.
5873  */
5874 int group_balance_cpu(struct sched_group *sg)
5875 {
5876         return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5877 }
5878 
5879 static int
5880 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5881 {
5882         struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5883         const struct cpumask *span = sched_domain_span(sd);
5884         struct cpumask *covered = sched_domains_tmpmask;
5885         struct sd_data *sdd = sd->private;
5886         struct sched_domain *sibling;
5887         int i;
5888 
5889         cpumask_clear(covered);
5890 
5891         for_each_cpu(i, span) {
5892                 struct cpumask *sg_span;
5893 
5894                 if (cpumask_test_cpu(i, covered))
5895                         continue;
5896 
5897                 sibling = *per_cpu_ptr(sdd->sd, i);
5898 
5899                 /* See the comment near build_group_mask(). */
5900                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5901                         continue;
5902 
5903                 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5904                                 GFP_KERNEL, cpu_to_node(cpu));
5905 
5906                 if (!sg)
5907                         goto fail;
5908 
5909                 sg_span = sched_group_cpus(sg);
5910                 if (sibling->child)
5911                         cpumask_copy(sg_span, sched_domain_span(sibling->child));
5912                 else
5913                         cpumask_set_cpu(i, sg_span);
5914 
5915                 cpumask_or(covered, covered, sg_span);
5916 
5917                 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5918                 if (atomic_inc_return(&sg->sgc->ref) == 1)
5919                         build_group_mask(sd, sg);
5920 
5921                 /*
5922                  * Initialize sgc->capacity such that even if we mess up the
5923                  * domains and no possible iteration will get us here, we won't
5924                  * die on a /0 trap.
5925                  */
5926                 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5927                 sg->sgc->capacity_orig = sg->sgc->capacity;
5928 
5929                 /*
5930                  * Make sure the first group of this domain contains the
5931                  * canonical balance cpu. Otherwise the sched_domain iteration
5932                  * breaks. See update_sg_lb_stats().
5933                  */
5934                 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5935                     group_balance_cpu(sg) == cpu)
5936                         groups = sg;
5937 
5938                 if (!first)
5939                         first = sg;
5940                 if (last)
5941                         last->next = sg;
5942                 last = sg;
5943                 last->next = first;
5944         }
5945         sd->groups = groups;
5946 
5947         return 0;
5948 
5949 fail:
5950         free_sched_groups(first, 0);
5951 
5952         return -ENOMEM;
5953 }
5954 
5955 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5956 {
5957         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5958         struct sched_domain *child = sd->child;
5959 
5960         if (child)
5961                 cpu = cpumask_first(sched_domain_span(child));
5962 
5963         if (sg) {
5964                 *sg = *per_cpu_ptr(sdd->sg, cpu);
5965                 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5966                 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5967         }
5968 
5969         return cpu;
5970 }
5971 
5972 /*
5973  * build_sched_groups will build a circular linked list of the groups
5974  * covered by the given span, and will set each group's ->cpumask correctly,
5975  * and ->cpu_capacity to 0.
5976  *
5977  * Assumes the sched_domain tree is fully constructed
5978  */
5979 static int
5980 build_sched_groups(struct sched_domain *sd, int cpu)
5981 {
5982         struct sched_group *first = NULL, *last = NULL;
5983         struct sd_data *sdd = sd->private;
5984         const struct cpumask *span = sched_domain_span(sd);
5985         struct cpumask *covered;
5986         int i;
5987 
5988         get_group(cpu, sdd, &sd->groups);
5989         atomic_inc(&sd->groups->ref);
5990 
5991         if (cpu != cpumask_first(span))
5992                 return 0;
5993 
5994         lockdep_assert_held(&sched_domains_mutex);
5995         covered = sched_domains_tmpmask;
5996 
5997         cpumask_clear(covered);
5998 
5999         for_each_cpu(i, span) {
6000                 struct sched_group *sg;
6001                 int group, j;
6002 
6003                 if (cpumask_test_cpu(i, covered))
6004                         continue;
6005 
6006                 group = get_group(i, sdd, &sg);
6007                 cpumask_setall(sched_group_mask(sg));
6008 
6009                 for_each_cpu(j, span) {
6010                         if (get_group(j, sdd, NULL) != group)
6011                                 continue;
6012 
6013                         cpumask_set_cpu(j, covered);
6014                         cpumask_set_cpu(j, sched_group_cpus(sg));
6015                 }
6016 
6017                 if (!first)
6018                         first = sg;
6019                 if (last)
6020                         last->next = sg;
6021                 last = sg;
6022         }
6023         last->next = first;
6024 
6025         return 0;
6026 }
6027 
6028 /*
6029  * Initialize sched groups cpu_capacity.
6030  *
6031  * cpu_capacity indicates the capacity of sched group, which is used while
6032  * distributing the load between different sched groups in a sched domain.
6033  * Typically cpu_capacity for all the groups in a sched domain will be same
6034  * unless there are asymmetries in the topology. If there are asymmetries,
6035  * group having more cpu_capacity will pickup more load compared to the
6036  * group having less cpu_capacity.
6037  */
6038 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6039 {
6040         struct sched_group *sg = sd->groups;
6041 
6042         WARN_ON(!sg);
6043 
6044         do {
6045                 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6046                 sg = sg->next;
6047         } while (sg != sd->groups);
6048 
6049         if (cpu != group_balance_cpu(sg))
6050                 return;
6051 
6052         update_group_capacity(sd, cpu);
6053         atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6054 }
6055 
6056 /*
6057  * Initializers for schedule domains
6058  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6059  */
6060 
6061 static int default_relax_domain_level = -1;
6062 int sched_domain_level_max;
6063 
6064 static int __init setup_relax_domain_level(char *str)
6065 {
6066         if (kstrtoint(str, 0, &default_relax_domain_level))
6067                 pr_warn("Unable to set relax_domain_level\n");
6068 
6069         return 1;
6070 }
6071 __setup("relax_domain_level=", setup_relax_domain_level);
6072 
6073 static void set_domain_attribute(struct sched_domain *sd,
6074                                  struct sched_domain_attr *attr)
6075 {
6076         int request;
6077 
6078         if (!attr || attr->relax_domain_level < 0) {
6079                 if (default_relax_domain_level < 0)
6080                         return;
6081                 else
6082                         request = default_relax_domain_level;
6083         } else
6084                 request = attr->relax_domain_level;
6085         if (request < sd->level) {
6086                 /* turn off idle balance on this domain */
6087                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6088         } else {
6089                 /* turn on idle balance on this domain */
6090                 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6091         }
6092 }
6093 
6094 static void __sdt_free(const struct cpumask *cpu_map);
6095 static int __sdt_alloc(const struct cpumask *cpu_map);
6096 
6097 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6098                                  const struct cpumask *cpu_map)
6099 {
6100         switch (what) {
6101         case sa_rootdomain:
6102                 if (!atomic_read(&d->rd->refcount))
6103                         free_rootdomain(&d->rd->rcu); /* fall through */
6104         case sa_sd:
6105                 free_percpu(d->sd); /* fall through */
6106         case sa_sd_storage:
6107                 __sdt_free(cpu_map); /* fall through */
6108         case sa_none:
6109                 break;
6110         }
6111 }
6112 
6113 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6114                                                    const struct cpumask *cpu_map)
6115 {
6116         memset(d, 0, sizeof(*d));
6117 
6118         if (__sdt_alloc(cpu_map))
6119                 return sa_sd_storage;
6120         d->sd = alloc_percpu(struct sched_domain *);
6121         if (!d->sd)
6122                 return sa_sd_storage;
6123         d->rd = alloc_rootdomain();
6124         if (!d->rd)
6125                 return sa_sd;
6126         return sa_rootdomain;
6127 }
6128 
6129 /*
6130  * NULL the sd_data elements we've used to build the sched_domain and
6131  * sched_group structure so that the subsequent __free_domain_allocs()
6132  * will not free the data we're using.
6133  */
6134 static void claim_allocations(int cpu, struct sched_domain *sd)
6135 {
6136         struct sd_data *sdd = sd->private;
6137 
6138         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6139         *per_cpu_ptr(sdd->sd, cpu) = NULL;
6140 
6141         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6142                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6143 
6144         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6145                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6146 }
6147 
6148 #ifdef CONFIG_NUMA
6149 static int sched_domains_numa_levels;
6150 enum numa_topology_type sched_numa_topology_type;
6151 static int *sched_domains_numa_distance;
6152 int sched_max_numa_distance;
6153 static struct cpumask ***sched_domains_numa_masks;
6154 static int sched_domains_curr_level;
6155 #endif
6156 
6157 /*
6158  * SD_flags allowed in topology descriptions.
6159  *
6160  * SD_SHARE_CPUCAPACITY      - describes SMT topologies
6161  * SD_SHARE_PKG_RESOURCES - describes shared caches
6162  * SD_NUMA                - describes NUMA topologies
6163  * SD_SHARE_POWERDOMAIN   - describes shared power domain
6164  *
6165  * Odd one out:
6166  * SD_ASYM_PACKING        - describes SMT quirks
6167  */
6168 #define TOPOLOGY_SD_FLAGS               \
6169         (SD_SHARE_CPUCAPACITY |         \
6170          SD_SHARE_PKG_RESOURCES |       \
6171          SD_NUMA |                      \
6172          SD_ASYM_PACKING |              \
6173          SD_SHARE_POWERDOMAIN)
6174 
6175 static struct sched_domain *
6176 sd_init(struct sched_domain_topology_level *tl, int cpu)
6177 {
6178         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6179         int sd_weight, sd_flags = 0;
6180 
6181 #ifdef CONFIG_NUMA
6182         /*
6183          * Ugly hack to pass state to sd_numa_mask()...
6184          */
6185         sched_domains_curr_level = tl->numa_level;
6186 #endif
6187 
6188         sd_weight = cpumask_weight(tl->mask(cpu));
6189 
6190         if (tl->sd_flags)
6191                 sd_flags = (*tl->sd_flags)();
6192         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6193                         "wrong sd_flags in topology description\n"))
6194                 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6195 
6196         *sd = (struct sched_domain){
6197                 .min_interval           = sd_weight,
6198                 .max_interval           = 2*sd_weight,
6199                 .busy_factor            = 32,
6200                 .imbalance_pct          = 125,
6201 
6202                 .cache_nice_tries       = 0,
6203                 .busy_idx               = 0,
6204                 .idle_idx               = 0,
6205                 .newidle_idx            = 0,
6206                 .wake_idx               = 0,
6207                 .forkexec_idx           = 0,
6208 
6209                 .flags                  = 1*SD_LOAD_BALANCE
6210                                         | 1*SD_BALANCE_NEWIDLE
6211                                         | 1*SD_BALANCE_EXEC
6212                                         | 1*SD_BALANCE_FORK
6213                                         | 0*SD_BALANCE_WAKE
6214                                         | 1*SD_WAKE_AFFINE
6215                                         | 0*SD_SHARE_CPUCAPACITY
6216                                         | 0*SD_SHARE_PKG_RESOURCES
6217                                         | 0*SD_SERIALIZE
6218                                         | 0*SD_PREFER_SIBLING
6219                                         | 0*SD_NUMA
6220                                         | sd_flags
6221                                         ,
6222 
6223                 .last_balance           = jiffies,
6224                 .balance_interval       = sd_weight,
6225                 .smt_gain               = 0,
6226                 .max_newidle_lb_cost    = 0,
6227                 .next_decay_max_lb_cost = jiffies,
6228 #ifdef CONFIG_SCHED_DEBUG
6229                 .name                   = tl->name,
6230 #endif
6231         };
6232 
6233         /*
6234          * Convert topological properties into behaviour.
6235          */
6236 
6237         if (sd->flags & SD_SHARE_CPUCAPACITY) {
6238                 sd->imbalance_pct = 110;
6239                 sd->smt_gain = 1178; /* ~15% */
6240 
6241         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6242                 sd->imbalance_pct = 117;
6243                 sd->cache_nice_tries = 1;
6244                 sd->busy_idx = 2;
6245 
6246 #ifdef CONFIG_NUMA
6247         } else if (sd->flags & SD_NUMA) {
6248                 sd->cache_nice_tries = 2;
6249                 sd->busy_idx = 3;
6250                 sd->idle_idx = 2;
6251 
6252                 sd->flags |= SD_SERIALIZE;
6253                 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6254                         sd->flags &= ~(SD_BALANCE_EXEC |
6255                                        SD_BALANCE_FORK |
6256                                        SD_WAKE_AFFINE);
6257                 }
6258 
6259 #endif
6260         } else {
6261                 sd->flags |= SD_PREFER_SIBLING;
6262                 sd->cache_nice_tries = 1;
6263                 sd->busy_idx = 2;
6264                 sd->idle_idx = 1;
6265         }
6266 
6267         sd->private = &tl->data;
6268 
6269         return sd;
6270 }
6271 
6272 /*
6273  * Topology list, bottom-up.
6274  */
6275 static struct sched_domain_topology_level default_topology[] = {
6276 #ifdef CONFIG_SCHED_SMT
6277         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6278 #endif
6279 #ifdef CONFIG_SCHED_MC
6280         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6281 #endif
6282         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6283         { NULL, },
6284 };
6285 
6286 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6287 
6288 #define for_each_sd_topology(tl)                        \
6289         for (tl = sched_domain_topology; tl->mask; tl++)
6290 
6291 void set_sched_topology(struct sched_domain_topology_level *tl)
6292 {
6293         sched_domain_topology = tl;
6294 }
6295 
6296 #ifdef CONFIG_NUMA
6297 
6298 static const struct cpumask *sd_numa_mask(int cpu)
6299 {
6300         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6301 }
6302 
6303 static void sched_numa_warn(const char *str)
6304 {
6305         static int done = false;
6306         int i,j;
6307 
6308         if (done)
6309                 return;
6310 
6311         done = true;
6312 
6313         printk(KERN_WARNING "ERROR: %s\n\n", str);
6314 
6315         for (i = 0; i < nr_node_ids; i++) {
6316                 printk(KERN_WARNING "  ");
6317                 for (j = 0; j < nr_node_ids; j++)
6318                         printk(KERN_CONT "%02d ", node_distance(i,j));
6319                 printk(KERN_CONT "\n");
6320         }
6321         printk(KERN_WARNING "\n");
6322 }
6323 
6324 bool find_numa_distance(int distance)
6325 {
6326         int i;
6327 
6328         if (distance == node_distance(0, 0))
6329                 return true;
6330 
6331         for (i = 0; i < sched_domains_numa_levels; i++) {
6332                 if (sched_domains_numa_distance[i] == distance)
6333                         return true;
6334         }
6335 
6336         return false;
6337 }
6338 
6339 /*
6340  * A system can have three types of NUMA topology:
6341  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6342  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6343  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6344  *
6345  * The difference between a glueless mesh topology and a backplane
6346  * topology lies in whether communication between not directly
6347  * connected nodes goes through intermediary nodes (where programs
6348  * could run), or through backplane controllers. This affects
6349  * placement of programs.
6350  *
6351  * The type of topology can be discerned with the following tests:
6352  * - If the maximum distance between any nodes is 1 hop, the system
6353  *   is directly connected.
6354  * - If for two nodes A and B, located N > 1 hops away from each other,
6355  *   there is an intermediary node C, which is < N hops away from both
6356  *   nodes A and B, the system is a glueless mesh.
6357  */
6358 static void init_numa_topology_type(void)
6359 {
6360         int a, b, c, n;
6361 
6362         n = sched_max_numa_distance;
6363 
6364         if (n <= 1)
6365                 sched_numa_topology_type = NUMA_DIRECT;
6366 
6367         for_each_online_node(a) {
6368                 for_each_online_node(b) {
6369                         /* Find two nodes furthest removed from each other. */
6370                         if (node_distance(a, b) < n)
6371                                 continue;
6372 
6373                         /* Is there an intermediary node between a and b? */
6374                         for_each_online_node(c) {
6375                                 if (node_distance(a, c) < n &&
6376                                     node_distance(b, c) < n) {
6377                                         sched_numa_topology_type =
6378                                                         NUMA_GLUELESS_MESH;
6379                                         return;
6380                                 }
6381                         }
6382 
6383                         sched_numa_topology_type = NUMA_BACKPLANE;
6384                         return;
6385                 }
6386         }
6387 }
6388 
6389 static void sched_init_numa(void)
6390 {
6391         int next_distance, curr_distance = node_distance(0, 0);
6392         struct sched_domain_topology_level *tl;
6393         int level = 0;
6394         int i, j, k;
6395 
6396         sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6397         if (!sched_domains_numa_distance)
6398                 return;
6399 
6400         /*
6401          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6402          * unique distances in the node_distance() table.
6403          *
6404          * Assumes node_distance(0,j) includes all distances in
6405          * node_distance(i,j) in order to avoid cubic time.
6406          */
6407         next_distance = curr_distance;
6408         for (i = 0; i < nr_node_ids; i++) {
6409                 for (j = 0; j < nr_node_ids; j++) {
6410                         for (k = 0; k < nr_node_ids; k++) {
6411                                 int distance = node_distance(i, k);
6412 
6413                                 if (distance > curr_distance &&
6414                                     (distance < next_distance ||
6415                                      next_distance == curr_distance))
6416                                         next_distance = distance;
6417 
6418                                 /*
6419                                  * While not a strong assumption it would be nice to know
6420                                  * about cases where if node A is connected to B, B is not
6421                                  * equally connected to A.
6422                                  */
6423                                 if (sched_debug() && node_distance(k, i) != distance)
6424                                         sched_numa_warn("Node-distance not symmetric");
6425 
6426                                 if (sched_debug() && i && !find_numa_distance(distance))
6427                                         sched_numa_warn("Node-0 not representative");
6428                         }
6429                         if (next_distance != curr_distance) {
6430                                 sched_domains_numa_distance[level++] = next_distance;
6431                                 sched_domains_numa_levels = level;
6432                                 curr_distance = next_distance;
6433                         } else break;
6434                 }
6435 
6436                 /*
6437                  * In case of sched_debug() we verify the above assumption.
6438                  */
6439                 if (!sched_debug())
6440                         break;
6441         }
6442 
6443         if (!level)
6444                 return;
6445 
6446         /*
6447          * 'level' contains the number of unique distances, excluding the
6448          * identity distance node_distance(i,i).
6449          *
6450          * The sched_domains_numa_distance[] array includes the actual distance
6451          * numbers.
6452          */
6453 
6454         /*
6455          * Here, we should temporarily reset sched_domains_numa_levels to 0.
6456          * If it fails to allocate memory for array sched_domains_numa_masks[][],
6457          * the array will contain less then 'level' members. This could be
6458          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6459          * in other functions.
6460          *
6461          * We reset it to 'level' at the end of this function.
6462          */
6463         sched_domains_numa_levels = 0;
6464 
6465         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6466         if (!sched_domains_numa_masks)
6467                 return;
6468 
6469         /*
6470          * Now for each level, construct a mask per node which contains all
6471          * cpus of nodes that are that many hops away from us.
6472          */
6473         for (i = 0; i < level; i++) {
6474                 sched_domains_numa_masks[i] =
6475                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6476                 if (!sched_domains_numa_masks[i])
6477                         return;
6478 
6479                 for (j = 0; j < nr_node_ids; j++) {
6480                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6481                         if (!mask)
6482                                 return;
6483 
6484                         sched_domains_numa_masks[i][j] = mask;
6485 
6486                         for (k = 0; k < nr_node_ids; k++) {
6487                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
6488                                         continue;
6489 
6490                                 cpumask_or(mask, mask, cpumask_of_node(k));
6491                         }
6492                 }
6493         }
6494 
6495         /* Compute default topology size */
6496         for (i = 0; sched_domain_topology[i].mask; i++);
6497 
6498         tl = kzalloc((i + level + 1) *
6499                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6500         if (!tl)
6501                 return;
6502 
6503         /*
6504          * Copy the default topology bits..
6505          */
6506         for (i = 0; sched_domain_topology[i].mask; i++)
6507                 tl[i] = sched_domain_topology[i];
6508 
6509         /*
6510          * .. and append 'j' levels of NUMA goodness.
6511          */
6512         for (j = 0; j < level; i++, j++) {
6513                 tl[i] = (struct sched_domain_topology_level){
6514                         .mask = sd_numa_mask,
6515                         .sd_flags = cpu_numa_flags,
6516                         .flags = SDTL_OVERLAP,
6517                         .numa_level = j,
6518                         SD_INIT_NAME(NUMA)
6519                 };
6520         }
6521 
6522         sched_domain_topology = tl;
6523 
6524         sched_domains_numa_levels = level;
6525         sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6526 
6527         init_numa_topology_type();
6528 }
6529 
6530 static void sched_domains_numa_masks_set(int cpu)
6531 {
6532         int i, j;
6533         int node = cpu_to_node(cpu);
6534 
6535         for (i = 0; i < sched_domains_numa_levels; i++) {
6536                 for (j = 0; j < nr_node_ids; j++) {
6537                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
6538                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6539                 }
6540         }
6541 }
6542 
6543 static void sched_domains_numa_masks_clear(int cpu)
6544 {
6545         int i, j;
6546         for (i = 0; i < sched_domains_numa_levels; i++) {
6547                 for (j = 0; j < nr_node_ids; j++)
6548                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6549         }
6550 }
6551 
6552 /*
6553  * Update sched_domains_numa_masks[level][node] array when new cpus
6554  * are onlined.
6555  */
6556 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6557                                            unsigned long action,
6558                                            void *hcpu)
6559 {
6560         int cpu = (long)hcpu;
6561 
6562         switch (action & ~CPU_TASKS_FROZEN) {
6563         case CPU_ONLINE:
6564                 sched_domains_numa_masks_set(cpu);
6565                 break;
6566 
6567         case CPU_DEAD:
6568                 sched_domains_numa_masks_clear(cpu);
6569                 break;
6570 
6571         default:
6572                 return NOTIFY_DONE;
6573         }
6574 
6575         return NOTIFY_OK;
6576 }
6577 #else
6578 static inline void sched_init_numa(void)
6579 {
6580 }
6581 
6582 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6583                                            unsigned long action,
6584                                            void *hcpu)
6585 {
6586         return 0;
6587 }
6588 #endif /* CONFIG_NUMA */
6589 
6590 static int __sdt_alloc(const struct cpumask *cpu_map)
6591 {
6592         struct sched_domain_topology_level *tl;
6593         int j;
6594 
6595         for_each_sd_topology(tl) {
6596                 struct sd_data *sdd = &tl->data;
6597 
6598                 sdd->sd = alloc_percpu(struct sched_domain *);
6599                 if (!sdd->sd)
6600                         return -ENOMEM;
6601 
6602                 sdd->sg = alloc_percpu(struct sched_group *);
6603                 if (!sdd->sg)
6604                         return -ENOMEM;
6605 
6606                 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6607                 if (!sdd->sgc)
6608                         return -ENOMEM;
6609 
6610                 for_each_cpu(j, cpu_map) {
6611                         struct sched_domain *sd;
6612                         struct sched_group *sg;
6613                         struct sched_group_capacity *sgc;
6614 
6615                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6616                                         GFP_KERNEL, cpu_to_node(j));
6617                         if (!sd)
6618                                 return -ENOMEM;
6619 
6620                         *per_cpu_ptr(sdd->sd, j) = sd;
6621 
6622                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6623                                         GFP_KERNEL, cpu_to_node(j));
6624                         if (!sg)
6625                                 return -ENOMEM;
6626 
6627                         sg->next = sg;
6628 
6629                         *per_cpu_ptr(sdd->sg, j) = sg;
6630 
6631                         sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6632                                         GFP_KERNEL, cpu_to_node(j));
6633                         if (!sgc)
6634                                 return -ENOMEM;
6635 
6636                         *per_cpu_ptr(sdd->sgc, j) = sgc;
6637                 }
6638         }
6639 
6640         return 0;
6641 }
6642 
6643 static void __sdt_free(const struct cpumask *cpu_map)
6644 {
6645         struct sched_domain_topology_level *tl;
6646         int j;
6647 
6648         for_each_sd_topology(tl) {
6649                 struct sd_data *sdd = &tl->data;
6650 
6651                 for_each_cpu(j, cpu_map) {
6652                         struct sched_domain *sd;
6653 
6654                         if (sdd->sd) {
6655                                 sd = *per_cpu_ptr(sdd->sd, j);
6656                                 if (sd && (sd->flags & SD_OVERLAP))
6657                                         free_sched_groups(sd->groups, 0);
6658                                 kfree(*per_cpu_ptr(sdd->sd, j));
6659                         }
6660 
6661                         if (sdd->sg)
6662                                 kfree(*per_cpu_ptr(sdd->sg, j));
6663                         if (sdd->sgc)
6664                                 kfree(*per_cpu_ptr(sdd->sgc, j));
6665                 }
6666                 free_percpu(sdd->sd);
6667                 sdd->sd = NULL;
6668                 free_percpu(sdd->sg);
6669                 sdd->sg = NULL;
6670                 free_percpu(sdd->sgc);
6671                 sdd->sgc = NULL;
6672         }
6673 }
6674 
6675 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6676                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6677                 struct sched_domain *child, int cpu)
6678 {
6679         struct sched_domain *sd = sd_init(tl, cpu);
6680         if (!sd)
6681                 return child;
6682 
6683         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6684         if (child) {
6685                 sd->level = child->level + 1;
6686                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6687                 child->parent = sd;
6688                 sd->child = child;
6689 
6690                 if (!cpumask_subset(sched_domain_span(child),
6691                                     sched_domain_span(sd))) {
6692                         pr_err("BUG: arch topology borken\n");
6693 #ifdef CONFIG_SCHED_DEBUG
6694                         pr_err("     the %s domain not a subset of the %s domain\n",
6695                                         child->name, sd->name);
6696 #endif
6697                         /* Fixup, ensure @sd has at least @child cpus. */
6698                         cpumask_or(sched_domain_span(sd),
6699                                    sched_domain_span(sd),
6700                                    sched_domain_span(child));
6701                 }
6702 
6703         }
6704         set_domain_attribute(sd, attr);
6705 
6706         return sd;
6707 }
6708 
6709 /*
6710  * Build sched domains for a given set of cpus and attach the sched domains
6711  * to the individual cpus
6712  */
6713 static int build_sched_domains(const struct cpumask *cpu_map,
6714                                struct sched_domain_attr *attr)
6715 {
6716         enum s_alloc alloc_state;
6717         struct sched_domain *sd;
6718         struct s_data d;
6719         int i, ret = -ENOMEM;
6720 
6721         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6722         if (alloc_state != sa_rootdomain)
6723                 goto error;
6724 
6725         /* Set up domains for cpus specified by the cpu_map. */
6726         for_each_cpu(i, cpu_map) {
6727                 struct sched_domain_topology_level *tl;
6728 
6729                 sd = NULL;
6730                 for_each_sd_topology(tl) {
6731                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6732                         if (tl == sched_domain_topology)
6733                                 *per_cpu_ptr(d.sd, i) = sd;
6734                         if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6735                                 sd->flags |= SD_OVERLAP;
6736                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6737                                 break;
6738                 }
6739         }
6740 
6741         /* Build the groups for the domains */
6742         for_each_cpu(i, cpu_map) {
6743                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6744                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
6745                         if (sd->flags & SD_OVERLAP) {
6746                                 if (build_overlap_sched_groups(sd, i))
6747                                         goto error;
6748                         } else {
6749                                 if (build_sched_groups(sd, i))
6750                                         goto error;
6751                         }
6752                 }
6753         }
6754 
6755         /* Calculate CPU capacity for physical packages and nodes */
6756         for (i = nr_cpumask_bits-1; i >= 0; i--) {
6757                 if (!cpumask_test_cpu(i, cpu_map))
6758                         continue;
6759 
6760                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6761                         claim_allocations(i, sd);
6762                         init_sched_groups_capacity(i, sd);
6763                 }
6764         }
6765 
6766         /* Attach the domains */
6767         rcu_read_lock();
6768         for_each_cpu(i, cpu_map) {
6769                 sd = *per_cpu_ptr(d.sd, i);
6770                 cpu_attach_domain(sd, d.rd, i);
6771         }
6772         rcu_read_unlock();
6773 
6774         ret = 0;
6775 error:
6776         __free_domain_allocs(&d, alloc_state, cpu_map);
6777         return ret;
6778 }
6779 
6780 static cpumask_var_t *doms_cur; /* current sched domains */
6781 static int ndoms_cur;           /* number of sched domains in 'doms_cur' */
6782 static struct sched_domain_attr *dattr_cur;
6783                                 /* attribues of custom domains in 'doms_cur' */
6784 
6785 /*
6786  * Special case: If a kmalloc of a doms_cur partition (array of
6787  * cpumask) fails, then fallback to a single sched domain,
6788  * as determined by the single cpumask fallback_doms.
6789  */
6790 static cpumask_var_t fallback_doms;
6791 
6792 /*
6793  * arch_update_cpu_topology lets virtualized architectures update the
6794  * cpu core maps. It is supposed to return 1 if the topology changed
6795  * or 0 if it stayed the same.
6796  */
6797 int __weak arch_update_cpu_topology(void)
6798 {
6799         return 0;
6800 }
6801 
6802 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6803 {
6804         int i;
6805         cpumask_var_t *doms;
6806 
6807         doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6808         if (!doms)
6809                 return NULL;
6810         for (i = 0; i < ndoms; i++) {
6811                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6812                         free_sched_domains(doms, i);
6813                         return NULL;
6814                 }
6815         }
6816         return doms;
6817 }
6818 
6819 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6820 {
6821         unsigned int i;
6822         for (i = 0; i < ndoms; i++)
6823                 free_cpumask_var(doms[i]);
6824         kfree(doms);
6825 }
6826 
6827 /*
6828  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6829  * For now this just excludes isolated cpus, but could be used to
6830  * exclude other special cases in the future.
6831  */
6832 static int init_sched_domains(const struct cpumask *cpu_map)
6833 {
6834         int err;
6835 
6836         arch_update_cpu_topology();
6837         ndoms_cur = 1;
6838         doms_cur = alloc_sched_domains(ndoms_cur);
6839         if (!doms_cur)
6840                 doms_cur = &fallback_doms;
6841         cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6842         err = build_sched_domains(doms_cur[0], NULL);
6843         register_sched_domain_sysctl();
6844 
6845         return err;
6846 }
6847 
6848 /*
6849  * Detach sched domains from a group of cpus specified in cpu_map
6850  * These cpus will now be attached to the NULL domain
6851  */
6852 static void detach_destroy_domains(const struct cpumask *cpu_map)
6853 {
6854         int i;
6855 
6856         rcu_read_lock();
6857         for_each_cpu(i, cpu_map)
6858                 cpu_attach_domain(NULL, &def_root_domain, i);
6859         rcu_read_unlock();
6860 }
6861 
6862 /* handle null as "default" */
6863 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6864                         struct sched_domain_attr *new, int idx_new)
6865 {
6866         struct sched_domain_attr tmp;
6867 
6868         /* fast path */
6869         if (!new && !cur)
6870                 return 1;
6871 
6872         tmp = SD_ATTR_INIT;
6873         return !memcmp(cur ? (cur + idx_cur) : &tmp,
6874                         new ? (new + idx_new) : &tmp,
6875                         sizeof(struct sched_domain_attr));
6876 }
6877 
6878 /*
6879  * Partition sched domains as specified by the 'ndoms_new'
6880  * cpumasks in the array doms_new[] of cpumasks. This compares
6881  * doms_new[] to the current sched domain partitioning, doms_cur[].
6882  * It destroys each deleted domain and builds each new domain.
6883  *
6884  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6885  * The masks don't intersect (don't overlap.) We should setup one
6886  * sched domain for each mask. CPUs not in any of the cpumasks will
6887  * not be load balanced. If the same cpumask appears both in the
6888  * current 'doms_cur' domains and in the new 'doms_new', we can leave
6889  * it as it is.
6890  *
6891  * The passed in 'doms_new' should be allocated using
6892  * alloc_sched_domains.  This routine takes ownership of it and will
6893  * free_sched_domains it when done with it. If the caller failed the
6894  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6895  * and partition_sched_domains() will fallback to the single partition
6896  * 'fallback_doms', it also forces the domains to be rebuilt.
6897  *
6898  * If doms_new == NULL it will be replaced with cpu_online_mask.
6899  * ndoms_new == 0 is a special case for destroying existing domains,
6900  * and it will not create the default domain.
6901  *
6902  * Call with hotplug lock held
6903  */
6904 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6905                              struct sched_domain_attr *dattr_new)
6906 {
6907         int i, j, n;
6908         int new_topology;
6909 
6910         mutex_lock(&sched_domains_mutex);
6911 
6912         /* always unregister in case we don't destroy any domains */
6913         unregister_sched_domain_sysctl();
6914 
6915         /* Let architecture update cpu core mappings. */
6916         new_topology = arch_update_cpu_topology();
6917 
6918         n = doms_new ? ndoms_new : 0;
6919 
6920         /* Destroy deleted domains */
6921         for (i = 0; i < ndoms_cur; i++) {
6922                 for (j = 0; j < n && !new_topology; j++) {
6923                         if (cpumask_equal(doms_cur[i], doms_new[j])
6924                             && dattrs_equal(dattr_cur, i, dattr_new, j))
6925                                 goto match1;
6926                 }
6927                 /* no match - a current sched domain not in new doms_new[] */
6928                 detach_destroy_domains(doms_cur[i]);
6929 match1:
6930                 ;
6931         }
6932 
6933         n = ndoms_cur;
6934         if (doms_new == NULL) {
6935                 n = 0;
6936                 doms_new = &fallback_doms;
6937                 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6938                 WARN_ON_ONCE(dattr_new);
6939         }
6940 
6941         /* Build new domains */
6942         for (i = 0; i < ndoms_new; i++) {
6943                 for (j = 0; j < n && !new_topology; j++) {
6944                         if (cpumask_equal(doms_new[i], doms_cur[j])
6945                             && dattrs_equal(dattr_new, i, dattr_cur, j))
6946                                 goto match2;
6947                 }
6948                 /* no match - add a new doms_new */
6949                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6950 match2:
6951                 ;
6952         }
6953 
6954         /* Remember the new sched domains */
6955         if (doms_cur != &fallback_doms)
6956                 free_sched_domains(doms_cur, ndoms_cur);
6957         kfree(dattr_cur);       /* kfree(NULL) is safe */
6958         doms_cur = doms_new;
6959         dattr_cur = dattr_new;
6960         ndoms_cur = ndoms_new;
6961 
6962         register_sched_domain_sysctl();
6963 
6964         mutex_unlock(&sched_domains_mutex);
6965 }
6966 
6967 static int num_cpus_frozen;     /* used to mark begin/end of suspend/resume */
6968 
6969 /*
6970  * Update cpusets according to cpu_active mask.  If cpusets are
6971  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6972  * around partition_sched_domains().
6973  *
6974  * If we come here as part of a suspend/resume, don't touch cpusets because we
6975  * want to restore it back to its original state upon resume anyway.
6976  */
6977 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6978                              void *hcpu)
6979 {
6980         switch (action) {
6981         case CPU_ONLINE_FROZEN:
6982         case CPU_DOWN_FAILED_FROZEN:
6983 
6984                 /*
6985                  * num_cpus_frozen tracks how many CPUs are involved in suspend
6986                  * resume sequence. As long as this is not the last online
6987                  * operation in the resume sequence, just build a single sched
6988                  * domain, ignoring cpusets.
6989                  */
6990                 num_cpus_frozen--;
6991                 if (likely(num_cpus_frozen)) {
6992                         partition_sched_domains(1, NULL, NULL);
6993                         break;
6994                 }
6995 
6996                 /*
6997                  * This is the last CPU online operation. So fall through and
6998                  * restore the original sched domains by considering the
6999                  * cpuset configurations.
7000                  */
7001 
7002         case CPU_ONLINE:
7003         case CPU_DOWN_FAILED:
7004                 cpuset_update_active_cpus(true);
7005                 break;
7006         default:
7007                 return NOTIFY_DONE;
7008         }
7009         return NOTIFY_OK;
7010 }
7011 
7012 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7013                                void *hcpu)
7014 {
7015         switch (action) {
7016         case CPU_DOWN_PREPARE:
7017                 cpuset_update_active_cpus(false);
7018                 break;
7019         case CPU_DOWN_PREPARE_FROZEN:
7020                 num_cpus_frozen++;
7021                 partition_sched_domains(1, NULL, NULL);
7022                 break;
7023         default:
7024                 return NOTIFY_DONE;
7025         }
7026         return NOTIFY_OK;
7027 }
7028 
7029 void __init sched_init_smp(void)
7030 {
7031         cpumask_var_t non_isolated_cpus;
7032 
7033         alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7034         alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7035 
7036         sched_init_numa();
7037 
7038         /*
7039          * There's no userspace yet to cause hotplug operations; hence all the
7040          * cpu masks are stable and all blatant races in the below code cannot
7041          * happen.
7042          */
7043         mutex_lock(&sched_domains_mutex);
7044         init_sched_domains(cpu_active_mask);
7045         cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7046         if (cpumask_empty(non_isolated_cpus))
7047                 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7048         mutex_unlock(&sched_domains_mutex);
7049 
7050         hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7051         hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7052         hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7053 
7054         init_hrtick();
7055 
7056         /* Move init over to a non-isolated CPU */
7057         if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7058                 BUG();
7059         sched_init_granularity();
7060         free_cpumask_var(non_isolated_cpus);
7061 
7062         init_sched_rt_class();
7063         init_sched_dl_class();
7064 }
7065 #else
7066 void __init sched_init_smp(void)
7067 {
7068         sched_init_granularity();
7069 }
7070 #endif /* CONFIG_SMP */
7071 
7072 const_debug unsigned int sysctl_timer_migration = 1;
7073 
7074 int in_sched_functions(unsigned long addr)
7075 {
7076         return in_lock_functions(addr) ||
7077                 (addr >= (unsigned long)__sched_text_start
7078                 && addr < (unsigned long)__sched_text_end);
7079 }
7080 
7081 #ifdef CONFIG_CGROUP_SCHED
7082 /*
7083  * Default task group.
7084  * Every task in system belongs to this group at bootup.
7085  */
7086 struct task_group root_task_group;
7087 LIST_HEAD(task_groups);
7088 #endif
7089 
7090 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7091 
7092 void __init sched_init(void)
7093 {
7094         int i, j;
7095         unsigned long alloc_size = 0, ptr;
7096 
7097 #ifdef CONFIG_FAIR_GROUP_SCHED
7098         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7099 #endif
7100 #ifdef CONFIG_RT_GROUP_SCHED
7101         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7102 #endif
7103         if (alloc_size) {
7104                 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7105 
7106 #ifdef CONFIG_FAIR_GROUP_SCHED
7107                 root_task_group.se = (struct sched_entity **)ptr;
7108                 ptr += nr_cpu_ids * sizeof(void **);
7109 
7110                 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7111                 ptr += nr_cpu_ids * sizeof(void **);
7112 
7113 #endif /* CONFIG_FAIR_GROUP_SCHED */
7114 #ifdef CONFIG_RT_GROUP_SCHED
7115                 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7116                 ptr += nr_cpu_ids * sizeof(void **);
7117 
7118                 root_task_group.rt_rq = (struct rt_rq **)ptr;
7119                 ptr += nr_cpu_ids * sizeof(void **);
7120 
7121 #endif /* CONFIG_RT_GROUP_SCHED */
7122         }
7123 #ifdef CONFIG_CPUMASK_OFFSTACK
7124         for_each_possible_cpu(i) {
7125                 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7126                         cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7127         }
7128 #endif /* CONFIG_CPUMASK_OFFSTACK */
7129 
7130         init_rt_bandwidth(&def_rt_bandwidth,
7131                         global_rt_period(), global_rt_runtime());
7132         init_dl_bandwidth(&def_dl_bandwidth,
7133                         global_rt_period(), global_rt_runtime());
7134 
7135 #ifdef CONFIG_SMP
7136         init_defrootdomain();
7137 #endif
7138 
7139 #ifdef CONFIG_RT_GROUP_SCHED
7140         init_rt_bandwidth(&root_task_group.rt_bandwidth,
7141                         global_rt_period(), global_rt_runtime());
7142 #endif /* CONFIG_RT_GROUP_SCHED */
7143 
7144 #ifdef CONFIG_CGROUP_SCHED
7145         list_add(&root_task_group.list, &task_groups);
7146         INIT_LIST_HEAD(&root_task_group.children);
7147         INIT_LIST_HEAD(&root_task_group.siblings);
7148         autogroup_init(&init_task);
7149 
7150 #endif /* CONFIG_CGROUP_SCHED */
7151 
7152         for_each_possible_cpu(i) {
7153                 struct rq *rq;
7154 
7155                 rq = cpu_rq(i);
7156                 raw_spin_lock_init(&rq->lock);
7157                 rq->nr_running = 0;
7158                 rq->calc_load_active = 0;
7159                 rq->calc_load_update = jiffies + LOAD_FREQ;
7160                 init_cfs_rq(&rq->cfs);
7161                 init_rt_rq(&rq->rt, rq);
7162                 init_dl_rq(&rq->dl, rq);
7163 #ifdef CONFIG_FAIR_GROUP_SCHED
7164                 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7165                 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7166                 /*
7167                  * How much cpu bandwidth does root_task_group get?
7168                  *
7169                  * In case of task-groups formed thr' the cgroup filesystem, it
7170                  * gets 100% of the cpu resources in the system. This overall
7171                  * system cpu resource is divided among the tasks of
7172                  * root_task_group and its child task-groups in a fair manner,
7173                  * based on each entity's (task or task-group's) weight
7174                  * (se->load.weight).
7175                  *
7176                  * In other words, if root_task_group has 10 tasks of weight
7177                  * 1024) and two child groups A0 and A1 (of weight 1024 each),
7178                  * then A0's share of the cpu resource is:
7179                  *
7180                  *      A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7181                  *
7182                  * We achieve this by letting root_task_group's tasks sit
7183                  * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7184                  */
7185                 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7186                 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7187 #endif /* CONFIG_FAIR_GROUP_SCHED */
7188 
7189                 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7190 #ifdef CONFIG_RT_GROUP_SCHED
7191                 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7192 #endif
7193 
7194                 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7195                         rq->cpu_load[j] = 0;
7196 
7197                 rq->last_load_update_tick = jiffies;
7198 
7199 #ifdef CONFIG_SMP
7200                 rq->sd = NULL;
7201                 rq->rd = NULL;
7202                 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7203                 rq->post_schedule = 0;
7204                 rq->active_balance = 0;
7205                 rq->next_balance = jiffies;
7206                 rq->push_cpu = 0;
7207                 rq->cpu = i;
7208                 rq->online = 0;
7209                 rq->idle_stamp = 0;
7210                 rq->avg_idle = 2*sysctl_sched_migration_cost;
7211                 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7212 
7213                 INIT_LIST_HEAD(&rq->cfs_tasks);
7214 
7215                 rq_attach_root(rq, &def_root_domain);
7216 #ifdef CONFIG_NO_HZ_COMMON
7217                 rq->nohz_flags = 0;
7218 #endif
7219 #ifdef CONFIG_NO_HZ_FULL
7220                 rq->last_sched_tick = 0;
7221 #endif
7222 #endif
7223                 init_rq_hrtick(rq);
7224                 atomic_set(&rq->nr_iowait, 0);
7225         }
7226 
7227         set_load_weight(&init_task);
7228 
7229 #ifdef CONFIG_PREEMPT_NOTIFIERS
7230         INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7231 #endif
7232 
7233         /*
7234          * The boot idle thread does lazy MMU switching as well:
7235          */
7236         atomic_inc(&init_mm.mm_count);
7237         enter_lazy_tlb(&init_mm, current);
7238 
7239         /*
7240          * During early bootup we pretend to be a normal task:
7241          */
7242         current->sched_class = &fair_sched_class;
7243 
7244         /*
7245          * Make us the idle thread. Technically, schedule() should not be
7246          * called from this thread, however somewhere below it might be,
7247          * but because we are the idle thread, we just pick up running again
7248          * when this runqueue becomes "idle".
7249          */
7250         init_idle(current, smp_processor_id());
7251 
7252         calc_load_update = jiffies + LOAD_FREQ;
7253 
7254 #ifdef CONFIG_SMP
7255         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7256         /* May be allocated at isolcpus cmdline parse time */
7257         if (cpu_isolated_map == NULL)
7258                 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7259         idle_thread_set_boot_cpu();
7260         set_cpu_rq_start_time();
7261 #endif
7262         init_sched_fair_class();
7263 
7264         scheduler_running = 1;
7265 }
7266 
7267 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7268 static inline int preempt_count_equals(int preempt_offset)
7269 {
7270         int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7271 
7272         return (nested == preempt_offset);
7273 }
7274 
7275 void __might_sleep(const char *file, int line, int preempt_offset)
7276 {
7277         /*
7278          * Blocking primitives will set (and therefore destroy) current->state,
7279          * since we will exit with TASK_RUNNING make sure we enter with it,
7280          * otherwise we will destroy state.
7281          */
7282         WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7283                         "do not call blocking ops when !TASK_RUNNING; "
7284                         "state=%lx set at [<%p>] %pS\n",
7285                         current->state,
7286                         (void *)current->task_state_change,
7287                         (void *)current->task_state_change);
7288 
7289         ___might_sleep(file, line, preempt_offset);
7290 }
7291 EXPORT_SYMBOL(__might_sleep);
7292 
7293 void ___might_sleep(const char *file, int line, int preempt_offset)
7294 {
7295         static unsigned long prev_jiffy;        /* ratelimiting */
7296 
7297         rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7298         if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7299              !is_idle_task(current)) ||
7300             system_state != SYSTEM_RUNNING || oops_in_progress)
7301                 return;
7302         if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7303                 return;
7304         prev_jiffy = jiffies;
7305 
7306         printk(KERN_ERR
7307                 "BUG: sleeping function called from invalid context at %s:%d\n",
7308                         file, line);
7309         printk(KERN_ERR
7310                 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7311                         in_atomic(), irqs_disabled(),
7312                         current->pid, current->comm);
7313 
7314         if (task_stack_end_corrupted(current))
7315                 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7316 
7317         debug_show_held_locks(current);
7318         if (irqs_disabled())
7319                 print_irqtrace_events(current);
7320 #ifdef CONFIG_DEBUG_PREEMPT
7321         if (!preempt_count_equals(preempt_offset)) {
7322                 pr_err("Preemption disabled at:");
7323                 print_ip_sym(current->preempt_disable_ip);
7324                 pr_cont("\n");
7325         }
7326 #endif
7327         dump_stack();
7328 }
7329 EXPORT_SYMBOL(___might_sleep);
7330 #endif
7331 
7332 #ifdef CONFIG_MAGIC_SYSRQ
7333 static void normalize_task(struct rq *rq, struct task_struct *p)
7334 {
7335         const struct sched_class *prev_class = p->sched_class;
7336         struct sched_attr attr = {
7337                 .sched_policy = SCHED_NORMAL,
7338         };
7339         int old_prio = p->prio;
7340         int queued;
7341 
7342         queued = task_on_rq_queued(p);
7343         if (queued)
7344                 dequeue_task(rq, p, 0);
7345         __setscheduler(rq, p, &attr);
7346         if (queued) {
7347                 enqueue_task(rq, p, 0);
7348                 resched_curr(rq);
7349         }
7350 
7351         check_class_changed(rq, p, prev_class, old_prio);
7352 }
7353 
7354 void normalize_rt_tasks(void)
7355 {
7356         struct task_struct *g, *p;
7357         unsigned long flags;
7358         struct rq *rq;
7359 
7360         read_lock(&tasklist_lock);
7361         for_each_process_thread(g, p) {
7362                 /*
7363                  * Only normalize user tasks:
7364                  */
7365                 if (p->flags & PF_KTHREAD)
7366                         continue;
7367 
7368                 p->se.exec_start                = 0;
7369 #ifdef CONFIG_SCHEDSTATS
7370                 p->se.statistics.wait_start     = 0;
7371                 p->se.statistics.sleep_start    = 0;
7372                 p->se.statistics.block_start    = 0;
7373 #endif
7374 
7375                 if (!dl_task(p) && !rt_task(p)) {
7376                         /*
7377                          * Renice negative nice level userspace
7378                          * tasks back to 0:
7379                          */
7380                         if (task_nice(p) < 0)
7381                                 set_user_nice(p, 0);
7382                         continue;
7383                 }
7384 
7385                 rq = task_rq_lock(p, &flags);
7386                 normalize_task(rq, p);
7387                 task_rq_unlock(rq, p, &flags);
7388         }
7389         read_unlock(&tasklist_lock);
7390 }
7391 
7392 #endif /* CONFIG_MAGIC_SYSRQ */
7393 
7394 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7395 /*
7396  * These functions are only useful for the IA64 MCA handling, or kdb.
7397  *
7398  * They can only be called when the whole system has been
7399  * stopped - every CPU needs to be quiescent, and no scheduling
7400  * activity can take place. Using them for anything else would
7401  * be a serious bug, and as a result, they aren't even visible
7402  * under any other configuration.
7403  */
7404 
7405 /**
7406  * curr_task - return the current task for a given cpu.
7407  * @cpu: the processor in question.
7408  *
7409  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7410  *
7411  * Return: The current task for @cpu.
7412  */
7413 struct task_struct *curr_task(int cpu)
7414 {
7415         return cpu_curr(cpu);
7416 }
7417 
7418 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7419 
7420 #ifdef CONFIG_IA64
7421 /**
7422  * set_curr_task - set the current task for a given cpu.
7423  * @cpu: the processor in question.
7424  * @p: the task pointer to set.
7425  *
7426  * Description: This function must only be used when non-maskable interrupts
7427  * are serviced on a separate stack. It allows the architecture to switch the
7428  * notion of the current task on a cpu in a non-blocking manner. This function
7429  * must be called with all CPU's synchronized, and interrupts disabled, the
7430  * and caller must save the original value of the current task (see
7431  * curr_task() above) and restore that value before reenabling interrupts and
7432  * re-starting the system.
7433  *
7434  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7435  */
7436 void set_curr_task(int cpu, struct task_struct *p)
7437 {
7438         cpu_curr(cpu) = p;
7439 }
7440 
7441 #endif
7442 
7443 #ifdef CONFIG_CGROUP_SCHED
7444 /* task_group_lock serializes the addition/removal of task groups */
7445 static DEFINE_SPINLOCK(task_group_lock);
7446 
7447 static void free_sched_group(struct task_group *tg)
7448 {
7449         free_fair_sched_group(tg);
7450         free_rt_sched_group(tg);
7451         autogroup_free(tg);
7452         kfree(tg);
7453 }
7454 
7455 /* allocate runqueue etc for a new task group */
7456 struct task_group *sched_create_group(struct task_group *parent)
7457 {
7458         struct task_group *tg;
7459 
7460         tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7461         if (!tg)
7462                 return ERR_PTR(-ENOMEM);
7463 
7464         if (!alloc_fair_sched_group(tg, parent))
7465                 goto err;
7466 
7467         if (!alloc_rt_sched_group(tg, parent))
7468                 goto err;
7469 
7470         return tg;
7471 
7472 err:
7473         free_sched_group(tg);
7474         return ERR_PTR(-ENOMEM);
7475 }
7476 
7477 void sched_online_group(struct task_group *tg, struct task_group *parent)
7478 {
7479         unsigned long flags;
7480 
7481         spin_lock_irqsave(&task_group_lock, flags);
7482         list_add_rcu(&tg->list, &task_groups);
7483