~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/sched/core.c

Version: ~ [ linux-5.8 ] ~ [ linux-5.7.12 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.55 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.136 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.191 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.232 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.232 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  *  kernel/sched/core.c
  3  *
  4  *  Kernel scheduler and related syscalls
  5  *
  6  *  Copyright (C) 1991-2002  Linus Torvalds
  7  *
  8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
  9  *              make semaphores SMP safe
 10  *  1998-11-19  Implemented schedule_timeout() and related stuff
 11  *              by Andrea Arcangeli
 12  *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
 13  *              hybrid priority-list and round-robin design with
 14  *              an array-switch method of distributing timeslices
 15  *              and per-CPU runqueues.  Cleanups and useful suggestions
 16  *              by Davide Libenzi, preemptible kernel bits by Robert Love.
 17  *  2003-09-03  Interactivity tuning by Con Kolivas.
 18  *  2004-04-02  Scheduler domains code by Nick Piggin
 19  *  2007-04-15  Work begun on replacing all interactivity tuning with a
 20  *              fair scheduling design by Con Kolivas.
 21  *  2007-05-05  Load balancing (smp-nice) and other improvements
 22  *              by Peter Williams
 23  *  2007-05-06  Interactivity improvements to CFS by Mike Galbraith
 24  *  2007-07-01  Group scheduling enhancements by Srivatsa Vaddagiri
 25  *  2007-11-29  RT balancing improvements by Steven Rostedt, Gregory Haskins,
 26  *              Thomas Gleixner, Mike Kravetz
 27  */
 28 
 29 #include <linux/mm.h>
 30 #include <linux/module.h>
 31 #include <linux/nmi.h>
 32 #include <linux/init.h>
 33 #include <linux/uaccess.h>
 34 #include <linux/highmem.h>
 35 #include <asm/mmu_context.h>
 36 #include <linux/interrupt.h>
 37 #include <linux/capability.h>
 38 #include <linux/completion.h>
 39 #include <linux/kernel_stat.h>
 40 #include <linux/debug_locks.h>
 41 #include <linux/perf_event.h>
 42 #include <linux/security.h>
 43 #include <linux/notifier.h>
 44 #include <linux/profile.h>
 45 #include <linux/freezer.h>
 46 #include <linux/vmalloc.h>
 47 #include <linux/blkdev.h>
 48 #include <linux/delay.h>
 49 #include <linux/pid_namespace.h>
 50 #include <linux/smp.h>
 51 #include <linux/threads.h>
 52 #include <linux/timer.h>
 53 #include <linux/rcupdate.h>
 54 #include <linux/cpu.h>
 55 #include <linux/cpuset.h>
 56 #include <linux/percpu.h>
 57 #include <linux/proc_fs.h>
 58 #include <linux/seq_file.h>
 59 #include <linux/sysctl.h>
 60 #include <linux/syscalls.h>
 61 #include <linux/times.h>
 62 #include <linux/tsacct_kern.h>
 63 #include <linux/kprobes.h>
 64 #include <linux/delayacct.h>
 65 #include <linux/unistd.h>
 66 #include <linux/pagemap.h>
 67 #include <linux/hrtimer.h>
 68 #include <linux/tick.h>
 69 #include <linux/debugfs.h>
 70 #include <linux/ctype.h>
 71 #include <linux/ftrace.h>
 72 #include <linux/slab.h>
 73 #include <linux/init_task.h>
 74 #include <linux/binfmts.h>
 75 #include <linux/context_tracking.h>
 76 #include <linux/compiler.h>
 77 
 78 #include <asm/switch_to.h>
 79 #include <asm/tlb.h>
 80 #include <asm/irq_regs.h>
 81 #include <asm/mutex.h>
 82 #ifdef CONFIG_PARAVIRT
 83 #include <asm/paravirt.h>
 84 #endif
 85 
 86 #include "sched.h"
 87 #include "../workqueue_internal.h"
 88 #include "../smpboot.h"
 89 
 90 #define CREATE_TRACE_POINTS
 91 #include <trace/events/sched.h>
 92 
 93 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
 94 {
 95         unsigned long delta;
 96         ktime_t soft, hard, now;
 97 
 98         for (;;) {
 99                 if (hrtimer_active(period_timer))
100                         break;
101 
102                 now = hrtimer_cb_get_time(period_timer);
103                 hrtimer_forward(period_timer, now, period);
104 
105                 soft = hrtimer_get_softexpires(period_timer);
106                 hard = hrtimer_get_expires(period_timer);
107                 delta = ktime_to_ns(ktime_sub(hard, soft));
108                 __hrtimer_start_range_ns(period_timer, soft, delta,
109                                          HRTIMER_MODE_ABS_PINNED, 0);
110         }
111 }
112 
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 
118 void update_rq_clock(struct rq *rq)
119 {
120         s64 delta;
121 
122         if (rq->skip_clock_update > 0)
123                 return;
124 
125         delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126         rq->clock += delta;
127         update_rq_clock_task(rq, delta);
128 }
129 
130 /*
131  * Debugging: various feature bits
132  */
133 
134 #define SCHED_FEAT(name, enabled)       \
135         (1UL << __SCHED_FEAT_##name) * enabled |
136 
137 const_debug unsigned int sysctl_sched_features =
138 #include "features.h"
139         0;
140 
141 #undef SCHED_FEAT
142 
143 #ifdef CONFIG_SCHED_DEBUG
144 #define SCHED_FEAT(name, enabled)       \
145         #name ,
146 
147 static const char * const sched_feat_names[] = {
148 #include "features.h"
149 };
150 
151 #undef SCHED_FEAT
152 
153 static int sched_feat_show(struct seq_file *m, void *v)
154 {
155         int i;
156 
157         for (i = 0; i < __SCHED_FEAT_NR; i++) {
158                 if (!(sysctl_sched_features & (1UL << i)))
159                         seq_puts(m, "NO_");
160                 seq_printf(m, "%s ", sched_feat_names[i]);
161         }
162         seq_puts(m, "\n");
163 
164         return 0;
165 }
166 
167 #ifdef HAVE_JUMP_LABEL
168 
169 #define jump_label_key__true  STATIC_KEY_INIT_TRUE
170 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 
172 #define SCHED_FEAT(name, enabled)       \
173         jump_label_key__##enabled ,
174 
175 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
176 #include "features.h"
177 };
178 
179 #undef SCHED_FEAT
180 
181 static void sched_feat_disable(int i)
182 {
183         if (static_key_enabled(&sched_feat_keys[i]))
184                 static_key_slow_dec(&sched_feat_keys[i]);
185 }
186 
187 static void sched_feat_enable(int i)
188 {
189         if (!static_key_enabled(&sched_feat_keys[i]))
190                 static_key_slow_inc(&sched_feat_keys[i]);
191 }
192 #else
193 static void sched_feat_disable(int i) { };
194 static void sched_feat_enable(int i) { };
195 #endif /* HAVE_JUMP_LABEL */
196 
197 static int sched_feat_set(char *cmp)
198 {
199         int i;
200         int neg = 0;
201 
202         if (strncmp(cmp, "NO_", 3) == 0) {
203                 neg = 1;
204                 cmp += 3;
205         }
206 
207         for (i = 0; i < __SCHED_FEAT_NR; i++) {
208                 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209                         if (neg) {
210                                 sysctl_sched_features &= ~(1UL << i);
211                                 sched_feat_disable(i);
212                         } else {
213                                 sysctl_sched_features |= (1UL << i);
214                                 sched_feat_enable(i);
215                         }
216                         break;
217                 }
218         }
219 
220         return i;
221 }
222 
223 static ssize_t
224 sched_feat_write(struct file *filp, const char __user *ubuf,
225                 size_t cnt, loff_t *ppos)
226 {
227         char buf[64];
228         char *cmp;
229         int i;
230 
231         if (cnt > 63)
232                 cnt = 63;
233 
234         if (copy_from_user(&buf, ubuf, cnt))
235                 return -EFAULT;
236 
237         buf[cnt] = 0;
238         cmp = strstrip(buf);
239 
240         i = sched_feat_set(cmp);
241         if (i == __SCHED_FEAT_NR)
242                 return -EINVAL;
243 
244         *ppos += cnt;
245 
246         return cnt;
247 }
248 
249 static int sched_feat_open(struct inode *inode, struct file *filp)
250 {
251         return single_open(filp, sched_feat_show, NULL);
252 }
253 
254 static const struct file_operations sched_feat_fops = {
255         .open           = sched_feat_open,
256         .write          = sched_feat_write,
257         .read           = seq_read,
258         .llseek         = seq_lseek,
259         .release        = single_release,
260 };
261 
262 static __init int sched_init_debug(void)
263 {
264         debugfs_create_file("sched_features", 0644, NULL, NULL,
265                         &sched_feat_fops);
266 
267         return 0;
268 }
269 late_initcall(sched_init_debug);
270 #endif /* CONFIG_SCHED_DEBUG */
271 
272 /*
273  * Number of tasks to iterate in a single balance run.
274  * Limited because this is done with IRQs disabled.
275  */
276 const_debug unsigned int sysctl_sched_nr_migrate = 32;
277 
278 /*
279  * period over which we average the RT time consumption, measured
280  * in ms.
281  *
282  * default: 1s
283  */
284 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
285 
286 /*
287  * period over which we measure -rt task cpu usage in us.
288  * default: 1s
289  */
290 unsigned int sysctl_sched_rt_period = 1000000;
291 
292 __read_mostly int scheduler_running;
293 
294 /*
295  * part of the period that we allow rt tasks to run in us.
296  * default: 0.95s
297  */
298 int sysctl_sched_rt_runtime = 950000;
299 
300 /*
301  * __task_rq_lock - lock the rq @p resides on.
302  */
303 static inline struct rq *__task_rq_lock(struct task_struct *p)
304         __acquires(rq->lock)
305 {
306         struct rq *rq;
307 
308         lockdep_assert_held(&p->pi_lock);
309 
310         for (;;) {
311                 rq = task_rq(p);
312                 raw_spin_lock(&rq->lock);
313                 if (likely(rq == task_rq(p)))
314                         return rq;
315                 raw_spin_unlock(&rq->lock);
316         }
317 }
318 
319 /*
320  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
321  */
322 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
323         __acquires(p->pi_lock)
324         __acquires(rq->lock)
325 {
326         struct rq *rq;
327 
328         for (;;) {
329                 raw_spin_lock_irqsave(&p->pi_lock, *flags);
330                 rq = task_rq(p);
331                 raw_spin_lock(&rq->lock);
332                 if (likely(rq == task_rq(p)))
333                         return rq;
334                 raw_spin_unlock(&rq->lock);
335                 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
336         }
337 }
338 
339 static void __task_rq_unlock(struct rq *rq)
340         __releases(rq->lock)
341 {
342         raw_spin_unlock(&rq->lock);
343 }
344 
345 static inline void
346 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
347         __releases(rq->lock)
348         __releases(p->pi_lock)
349 {
350         raw_spin_unlock(&rq->lock);
351         raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
352 }
353 
354 /*
355  * this_rq_lock - lock this runqueue and disable interrupts.
356  */
357 static struct rq *this_rq_lock(void)
358         __acquires(rq->lock)
359 {
360         struct rq *rq;
361 
362         local_irq_disable();
363         rq = this_rq();
364         raw_spin_lock(&rq->lock);
365 
366         return rq;
367 }
368 
369 #ifdef CONFIG_SCHED_HRTICK
370 /*
371  * Use HR-timers to deliver accurate preemption points.
372  */
373 
374 static void hrtick_clear(struct rq *rq)
375 {
376         if (hrtimer_active(&rq->hrtick_timer))
377                 hrtimer_cancel(&rq->hrtick_timer);
378 }
379 
380 /*
381  * High-resolution timer tick.
382  * Runs from hardirq context with interrupts disabled.
383  */
384 static enum hrtimer_restart hrtick(struct hrtimer *timer)
385 {
386         struct rq *rq = container_of(timer, struct rq, hrtick_timer);
387 
388         WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
389 
390         raw_spin_lock(&rq->lock);
391         update_rq_clock(rq);
392         rq->curr->sched_class->task_tick(rq, rq->curr, 1);
393         raw_spin_unlock(&rq->lock);
394 
395         return HRTIMER_NORESTART;
396 }
397 
398 #ifdef CONFIG_SMP
399 
400 static int __hrtick_restart(struct rq *rq)
401 {
402         struct hrtimer *timer = &rq->hrtick_timer;
403         ktime_t time = hrtimer_get_softexpires(timer);
404 
405         return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
406 }
407 
408 /*
409  * called from hardirq (IPI) context
410  */
411 static void __hrtick_start(void *arg)
412 {
413         struct rq *rq = arg;
414 
415         raw_spin_lock(&rq->lock);
416         __hrtick_restart(rq);
417         rq->hrtick_csd_pending = 0;
418         raw_spin_unlock(&rq->lock);
419 }
420 
421 /*
422  * Called to set the hrtick timer state.
423  *
424  * called with rq->lock held and irqs disabled
425  */
426 void hrtick_start(struct rq *rq, u64 delay)
427 {
428         struct hrtimer *timer = &rq->hrtick_timer;
429         ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 
431         hrtimer_set_expires(timer, time);
432 
433         if (rq == this_rq()) {
434                 __hrtick_restart(rq);
435         } else if (!rq->hrtick_csd_pending) {
436                 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
437                 rq->hrtick_csd_pending = 1;
438         }
439 }
440 
441 static int
442 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443 {
444         int cpu = (int)(long)hcpu;
445 
446         switch (action) {
447         case CPU_UP_CANCELED:
448         case CPU_UP_CANCELED_FROZEN:
449         case CPU_DOWN_PREPARE:
450         case CPU_DOWN_PREPARE_FROZEN:
451         case CPU_DEAD:
452         case CPU_DEAD_FROZEN:
453                 hrtick_clear(cpu_rq(cpu));
454                 return NOTIFY_OK;
455         }
456 
457         return NOTIFY_DONE;
458 }
459 
460 static __init void init_hrtick(void)
461 {
462         hotcpu_notifier(hotplug_hrtick, 0);
463 }
464 #else
465 /*
466  * Called to set the hrtick timer state.
467  *
468  * called with rq->lock held and irqs disabled
469  */
470 void hrtick_start(struct rq *rq, u64 delay)
471 {
472         __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
473                         HRTIMER_MODE_REL_PINNED, 0);
474 }
475 
476 static inline void init_hrtick(void)
477 {
478 }
479 #endif /* CONFIG_SMP */
480 
481 static void init_rq_hrtick(struct rq *rq)
482 {
483 #ifdef CONFIG_SMP
484         rq->hrtick_csd_pending = 0;
485 
486         rq->hrtick_csd.flags = 0;
487         rq->hrtick_csd.func = __hrtick_start;
488         rq->hrtick_csd.info = rq;
489 #endif
490 
491         hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
492         rq->hrtick_timer.function = hrtick;
493 }
494 #else   /* CONFIG_SCHED_HRTICK */
495 static inline void hrtick_clear(struct rq *rq)
496 {
497 }
498 
499 static inline void init_rq_hrtick(struct rq *rq)
500 {
501 }
502 
503 static inline void init_hrtick(void)
504 {
505 }
506 #endif  /* CONFIG_SCHED_HRTICK */
507 
508 /*
509  * resched_task - mark a task 'to be rescheduled now'.
510  *
511  * On UP this means the setting of the need_resched flag, on SMP it
512  * might also involve a cross-CPU call to trigger the scheduler on
513  * the target CPU.
514  */
515 void resched_task(struct task_struct *p)
516 {
517         int cpu;
518 
519         lockdep_assert_held(&task_rq(p)->lock);
520 
521         if (test_tsk_need_resched(p))
522                 return;
523 
524         set_tsk_need_resched(p);
525 
526         cpu = task_cpu(p);
527         if (cpu == smp_processor_id()) {
528                 set_preempt_need_resched();
529                 return;
530         }
531 
532         /* NEED_RESCHED must be visible before we test polling */
533         smp_mb();
534         if (!tsk_is_polling(p))
535                 smp_send_reschedule(cpu);
536 }
537 
538 void resched_cpu(int cpu)
539 {
540         struct rq *rq = cpu_rq(cpu);
541         unsigned long flags;
542 
543         if (!raw_spin_trylock_irqsave(&rq->lock, flags))
544                 return;
545         resched_task(cpu_curr(cpu));
546         raw_spin_unlock_irqrestore(&rq->lock, flags);
547 }
548 
549 #ifdef CONFIG_SMP
550 #ifdef CONFIG_NO_HZ_COMMON
551 /*
552  * In the semi idle case, use the nearest busy cpu for migrating timers
553  * from an idle cpu.  This is good for power-savings.
554  *
555  * We don't do similar optimization for completely idle system, as
556  * selecting an idle cpu will add more delays to the timers than intended
557  * (as that cpu's timer base may not be uptodate wrt jiffies etc).
558  */
559 int get_nohz_timer_target(int pinned)
560 {
561         int cpu = smp_processor_id();
562         int i;
563         struct sched_domain *sd;
564 
565         if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
566                 return cpu;
567 
568         rcu_read_lock();
569         for_each_domain(cpu, sd) {
570                 for_each_cpu(i, sched_domain_span(sd)) {
571                         if (!idle_cpu(i)) {
572                                 cpu = i;
573                                 goto unlock;
574                         }
575                 }
576         }
577 unlock:
578         rcu_read_unlock();
579         return cpu;
580 }
581 /*
582  * When add_timer_on() enqueues a timer into the timer wheel of an
583  * idle CPU then this timer might expire before the next timer event
584  * which is scheduled to wake up that CPU. In case of a completely
585  * idle system the next event might even be infinite time into the
586  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
587  * leaves the inner idle loop so the newly added timer is taken into
588  * account when the CPU goes back to idle and evaluates the timer
589  * wheel for the next timer event.
590  */
591 static void wake_up_idle_cpu(int cpu)
592 {
593         struct rq *rq = cpu_rq(cpu);
594 
595         if (cpu == smp_processor_id())
596                 return;
597 
598         /*
599          * This is safe, as this function is called with the timer
600          * wheel base lock of (cpu) held. When the CPU is on the way
601          * to idle and has not yet set rq->curr to idle then it will
602          * be serialized on the timer wheel base lock and take the new
603          * timer into account automatically.
604          */
605         if (rq->curr != rq->idle)
606                 return;
607 
608         /*
609          * We can set TIF_RESCHED on the idle task of the other CPU
610          * lockless. The worst case is that the other CPU runs the
611          * idle task through an additional NOOP schedule()
612          */
613         set_tsk_need_resched(rq->idle);
614 
615         /* NEED_RESCHED must be visible before we test polling */
616         smp_mb();
617         if (!tsk_is_polling(rq->idle))
618                 smp_send_reschedule(cpu);
619 }
620 
621 static bool wake_up_full_nohz_cpu(int cpu)
622 {
623         if (tick_nohz_full_cpu(cpu)) {
624                 if (cpu != smp_processor_id() ||
625                     tick_nohz_tick_stopped())
626                         smp_send_reschedule(cpu);
627                 return true;
628         }
629 
630         return false;
631 }
632 
633 void wake_up_nohz_cpu(int cpu)
634 {
635         if (!wake_up_full_nohz_cpu(cpu))
636                 wake_up_idle_cpu(cpu);
637 }
638 
639 static inline bool got_nohz_idle_kick(void)
640 {
641         int cpu = smp_processor_id();
642 
643         if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
644                 return false;
645 
646         if (idle_cpu(cpu) && !need_resched())
647                 return true;
648 
649         /*
650          * We can't run Idle Load Balance on this CPU for this time so we
651          * cancel it and clear NOHZ_BALANCE_KICK
652          */
653         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
654         return false;
655 }
656 
657 #else /* CONFIG_NO_HZ_COMMON */
658 
659 static inline bool got_nohz_idle_kick(void)
660 {
661         return false;
662 }
663 
664 #endif /* CONFIG_NO_HZ_COMMON */
665 
666 #ifdef CONFIG_NO_HZ_FULL
667 bool sched_can_stop_tick(void)
668 {
669        struct rq *rq;
670 
671        rq = this_rq();
672 
673        /* Make sure rq->nr_running update is visible after the IPI */
674        smp_rmb();
675 
676        /* More than one running task need preemption */
677        if (rq->nr_running > 1)
678                return false;
679 
680        return true;
681 }
682 #endif /* CONFIG_NO_HZ_FULL */
683 
684 void sched_avg_update(struct rq *rq)
685 {
686         s64 period = sched_avg_period();
687 
688         while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
689                 /*
690                  * Inline assembly required to prevent the compiler
691                  * optimising this loop into a divmod call.
692                  * See __iter_div_u64_rem() for another example of this.
693                  */
694                 asm("" : "+rm" (rq->age_stamp));
695                 rq->age_stamp += period;
696                 rq->rt_avg /= 2;
697         }
698 }
699 
700 #endif /* CONFIG_SMP */
701 
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703                         (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
704 /*
705  * Iterate task_group tree rooted at *from, calling @down when first entering a
706  * node and @up when leaving it for the final time.
707  *
708  * Caller must hold rcu_lock or sufficient equivalent.
709  */
710 int walk_tg_tree_from(struct task_group *from,
711                              tg_visitor down, tg_visitor up, void *data)
712 {
713         struct task_group *parent, *child;
714         int ret;
715 
716         parent = from;
717 
718 down:
719         ret = (*down)(parent, data);
720         if (ret)
721                 goto out;
722         list_for_each_entry_rcu(child, &parent->children, siblings) {
723                 parent = child;
724                 goto down;
725 
726 up:
727                 continue;
728         }
729         ret = (*up)(parent, data);
730         if (ret || parent == from)
731                 goto out;
732 
733         child = parent;
734         parent = parent->parent;
735         if (parent)
736                 goto up;
737 out:
738         return ret;
739 }
740 
741 int tg_nop(struct task_group *tg, void *data)
742 {
743         return 0;
744 }
745 #endif
746 
747 static void set_load_weight(struct task_struct *p)
748 {
749         int prio = p->static_prio - MAX_RT_PRIO;
750         struct load_weight *load = &p->se.load;
751 
752         /*
753          * SCHED_IDLE tasks get minimal weight:
754          */
755         if (p->policy == SCHED_IDLE) {
756                 load->weight = scale_load(WEIGHT_IDLEPRIO);
757                 load->inv_weight = WMULT_IDLEPRIO;
758                 return;
759         }
760 
761         load->weight = scale_load(prio_to_weight[prio]);
762         load->inv_weight = prio_to_wmult[prio];
763 }
764 
765 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
766 {
767         update_rq_clock(rq);
768         sched_info_queued(rq, p);
769         p->sched_class->enqueue_task(rq, p, flags);
770 }
771 
772 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
773 {
774         update_rq_clock(rq);
775         sched_info_dequeued(rq, p);
776         p->sched_class->dequeue_task(rq, p, flags);
777 }
778 
779 void activate_task(struct rq *rq, struct task_struct *p, int flags)
780 {
781         if (task_contributes_to_load(p))
782                 rq->nr_uninterruptible--;
783 
784         enqueue_task(rq, p, flags);
785 }
786 
787 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
788 {
789         if (task_contributes_to_load(p))
790                 rq->nr_uninterruptible++;
791 
792         dequeue_task(rq, p, flags);
793 }
794 
795 static void update_rq_clock_task(struct rq *rq, s64 delta)
796 {
797 /*
798  * In theory, the compile should just see 0 here, and optimize out the call
799  * to sched_rt_avg_update. But I don't trust it...
800  */
801 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802         s64 steal = 0, irq_delta = 0;
803 #endif
804 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
805         irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
806 
807         /*
808          * Since irq_time is only updated on {soft,}irq_exit, we might run into
809          * this case when a previous update_rq_clock() happened inside a
810          * {soft,}irq region.
811          *
812          * When this happens, we stop ->clock_task and only update the
813          * prev_irq_time stamp to account for the part that fit, so that a next
814          * update will consume the rest. This ensures ->clock_task is
815          * monotonic.
816          *
817          * It does however cause some slight miss-attribution of {soft,}irq
818          * time, a more accurate solution would be to update the irq_time using
819          * the current rq->clock timestamp, except that would require using
820          * atomic ops.
821          */
822         if (irq_delta > delta)
823                 irq_delta = delta;
824 
825         rq->prev_irq_time += irq_delta;
826         delta -= irq_delta;
827 #endif
828 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
829         if (static_key_false((&paravirt_steal_rq_enabled))) {
830                 steal = paravirt_steal_clock(cpu_of(rq));
831                 steal -= rq->prev_steal_time_rq;
832 
833                 if (unlikely(steal > delta))
834                         steal = delta;
835 
836                 rq->prev_steal_time_rq += steal;
837                 delta -= steal;
838         }
839 #endif
840 
841         rq->clock_task += delta;
842 
843 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
844         if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
845                 sched_rt_avg_update(rq, irq_delta + steal);
846 #endif
847 }
848 
849 void sched_set_stop_task(int cpu, struct task_struct *stop)
850 {
851         struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
852         struct task_struct *old_stop = cpu_rq(cpu)->stop;
853 
854         if (stop) {
855                 /*
856                  * Make it appear like a SCHED_FIFO task, its something
857                  * userspace knows about and won't get confused about.
858                  *
859                  * Also, it will make PI more or less work without too
860                  * much confusion -- but then, stop work should not
861                  * rely on PI working anyway.
862                  */
863                 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
864 
865                 stop->sched_class = &stop_sched_class;
866         }
867 
868         cpu_rq(cpu)->stop = stop;
869 
870         if (old_stop) {
871                 /*
872                  * Reset it back to a normal scheduling class so that
873                  * it can die in pieces.
874                  */
875                 old_stop->sched_class = &rt_sched_class;
876         }
877 }
878 
879 /*
880  * __normal_prio - return the priority that is based on the static prio
881  */
882 static inline int __normal_prio(struct task_struct *p)
883 {
884         return p->static_prio;
885 }
886 
887 /*
888  * Calculate the expected normal priority: i.e. priority
889  * without taking RT-inheritance into account. Might be
890  * boosted by interactivity modifiers. Changes upon fork,
891  * setprio syscalls, and whenever the interactivity
892  * estimator recalculates.
893  */
894 static inline int normal_prio(struct task_struct *p)
895 {
896         int prio;
897 
898         if (task_has_dl_policy(p))
899                 prio = MAX_DL_PRIO-1;
900         else if (task_has_rt_policy(p))
901                 prio = MAX_RT_PRIO-1 - p->rt_priority;
902         else
903                 prio = __normal_prio(p);
904         return prio;
905 }
906 
907 /*
908  * Calculate the current priority, i.e. the priority
909  * taken into account by the scheduler. This value might
910  * be boosted by RT tasks, or might be boosted by
911  * interactivity modifiers. Will be RT if the task got
912  * RT-boosted. If not then it returns p->normal_prio.
913  */
914 static int effective_prio(struct task_struct *p)
915 {
916         p->normal_prio = normal_prio(p);
917         /*
918          * If we are RT tasks or we were boosted to RT priority,
919          * keep the priority unchanged. Otherwise, update priority
920          * to the normal priority:
921          */
922         if (!rt_prio(p->prio))
923                 return p->normal_prio;
924         return p->prio;
925 }
926 
927 /**
928  * task_curr - is this task currently executing on a CPU?
929  * @p: the task in question.
930  *
931  * Return: 1 if the task is currently executing. 0 otherwise.
932  */
933 inline int task_curr(const struct task_struct *p)
934 {
935         return cpu_curr(task_cpu(p)) == p;
936 }
937 
938 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
939                                        const struct sched_class *prev_class,
940                                        int oldprio)
941 {
942         if (prev_class != p->sched_class) {
943                 if (prev_class->switched_from)
944                         prev_class->switched_from(rq, p);
945                 p->sched_class->switched_to(rq, p);
946         } else if (oldprio != p->prio || dl_task(p))
947                 p->sched_class->prio_changed(rq, p, oldprio);
948 }
949 
950 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
951 {
952         const struct sched_class *class;
953 
954         if (p->sched_class == rq->curr->sched_class) {
955                 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
956         } else {
957                 for_each_class(class) {
958                         if (class == rq->curr->sched_class)
959                                 break;
960                         if (class == p->sched_class) {
961                                 resched_task(rq->curr);
962                                 break;
963                         }
964                 }
965         }
966 
967         /*
968          * A queue event has occurred, and we're going to schedule.  In
969          * this case, we can save a useless back to back clock update.
970          */
971         if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
972                 rq->skip_clock_update = 1;
973 }
974 
975 #ifdef CONFIG_SMP
976 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
977 {
978 #ifdef CONFIG_SCHED_DEBUG
979         /*
980          * We should never call set_task_cpu() on a blocked task,
981          * ttwu() will sort out the placement.
982          */
983         WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
984                         !(task_preempt_count(p) & PREEMPT_ACTIVE));
985 
986 #ifdef CONFIG_LOCKDEP
987         /*
988          * The caller should hold either p->pi_lock or rq->lock, when changing
989          * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
990          *
991          * sched_move_task() holds both and thus holding either pins the cgroup,
992          * see task_group().
993          *
994          * Furthermore, all task_rq users should acquire both locks, see
995          * task_rq_lock().
996          */
997         WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
998                                       lockdep_is_held(&task_rq(p)->lock)));
999 #endif
1000 #endif
1001 
1002         trace_sched_migrate_task(p, new_cpu);
1003 
1004         if (task_cpu(p) != new_cpu) {
1005                 if (p->sched_class->migrate_task_rq)
1006                         p->sched_class->migrate_task_rq(p, new_cpu);
1007                 p->se.nr_migrations++;
1008                 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1009         }
1010 
1011         __set_task_cpu(p, new_cpu);
1012 }
1013 
1014 static void __migrate_swap_task(struct task_struct *p, int cpu)
1015 {
1016         if (p->on_rq) {
1017                 struct rq *src_rq, *dst_rq;
1018 
1019                 src_rq = task_rq(p);
1020                 dst_rq = cpu_rq(cpu);
1021 
1022                 deactivate_task(src_rq, p, 0);
1023                 set_task_cpu(p, cpu);
1024                 activate_task(dst_rq, p, 0);
1025                 check_preempt_curr(dst_rq, p, 0);
1026         } else {
1027                 /*
1028                  * Task isn't running anymore; make it appear like we migrated
1029                  * it before it went to sleep. This means on wakeup we make the
1030                  * previous cpu our targer instead of where it really is.
1031                  */
1032                 p->wake_cpu = cpu;
1033         }
1034 }
1035 
1036 struct migration_swap_arg {
1037         struct task_struct *src_task, *dst_task;
1038         int src_cpu, dst_cpu;
1039 };
1040 
1041 static int migrate_swap_stop(void *data)
1042 {
1043         struct migration_swap_arg *arg = data;
1044         struct rq *src_rq, *dst_rq;
1045         int ret = -EAGAIN;
1046 
1047         src_rq = cpu_rq(arg->src_cpu);
1048         dst_rq = cpu_rq(arg->dst_cpu);
1049 
1050         double_raw_lock(&arg->src_task->pi_lock,
1051                         &arg->dst_task->pi_lock);
1052         double_rq_lock(src_rq, dst_rq);
1053         if (task_cpu(arg->dst_task) != arg->dst_cpu)
1054                 goto unlock;
1055 
1056         if (task_cpu(arg->src_task) != arg->src_cpu)
1057                 goto unlock;
1058 
1059         if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1060                 goto unlock;
1061 
1062         if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1063                 goto unlock;
1064 
1065         __migrate_swap_task(arg->src_task, arg->dst_cpu);
1066         __migrate_swap_task(arg->dst_task, arg->src_cpu);
1067 
1068         ret = 0;
1069 
1070 unlock:
1071         double_rq_unlock(src_rq, dst_rq);
1072         raw_spin_unlock(&arg->dst_task->pi_lock);
1073         raw_spin_unlock(&arg->src_task->pi_lock);
1074 
1075         return ret;
1076 }
1077 
1078 /*
1079  * Cross migrate two tasks
1080  */
1081 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1082 {
1083         struct migration_swap_arg arg;
1084         int ret = -EINVAL;
1085 
1086         arg = (struct migration_swap_arg){
1087                 .src_task = cur,
1088                 .src_cpu = task_cpu(cur),
1089                 .dst_task = p,
1090                 .dst_cpu = task_cpu(p),
1091         };
1092 
1093         if (arg.src_cpu == arg.dst_cpu)
1094                 goto out;
1095 
1096         /*
1097          * These three tests are all lockless; this is OK since all of them
1098          * will be re-checked with proper locks held further down the line.
1099          */
1100         if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1101                 goto out;
1102 
1103         if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1104                 goto out;
1105 
1106         if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1107                 goto out;
1108 
1109         trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1110         ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1111 
1112 out:
1113         return ret;
1114 }
1115 
1116 struct migration_arg {
1117         struct task_struct *task;
1118         int dest_cpu;
1119 };
1120 
1121 static int migration_cpu_stop(void *data);
1122 
1123 /*
1124  * wait_task_inactive - wait for a thread to unschedule.
1125  *
1126  * If @match_state is nonzero, it's the @p->state value just checked and
1127  * not expected to change.  If it changes, i.e. @p might have woken up,
1128  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1129  * we return a positive number (its total switch count).  If a second call
1130  * a short while later returns the same number, the caller can be sure that
1131  * @p has remained unscheduled the whole time.
1132  *
1133  * The caller must ensure that the task *will* unschedule sometime soon,
1134  * else this function might spin for a *long* time. This function can't
1135  * be called with interrupts off, or it may introduce deadlock with
1136  * smp_call_function() if an IPI is sent by the same process we are
1137  * waiting to become inactive.
1138  */
1139 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1140 {
1141         unsigned long flags;
1142         int running, on_rq;
1143         unsigned long ncsw;
1144         struct rq *rq;
1145 
1146         for (;;) {
1147                 /*
1148                  * We do the initial early heuristics without holding
1149                  * any task-queue locks at all. We'll only try to get
1150                  * the runqueue lock when things look like they will
1151                  * work out!
1152                  */
1153                 rq = task_rq(p);
1154 
1155                 /*
1156                  * If the task is actively running on another CPU
1157                  * still, just relax and busy-wait without holding
1158                  * any locks.
1159                  *
1160                  * NOTE! Since we don't hold any locks, it's not
1161                  * even sure that "rq" stays as the right runqueue!
1162                  * But we don't care, since "task_running()" will
1163                  * return false if the runqueue has changed and p
1164                  * is actually now running somewhere else!
1165                  */
1166                 while (task_running(rq, p)) {
1167                         if (match_state && unlikely(p->state != match_state))
1168                                 return 0;
1169                         cpu_relax();
1170                 }
1171 
1172                 /*
1173                  * Ok, time to look more closely! We need the rq
1174                  * lock now, to be *sure*. If we're wrong, we'll
1175                  * just go back and repeat.
1176                  */
1177                 rq = task_rq_lock(p, &flags);
1178                 trace_sched_wait_task(p);
1179                 running = task_running(rq, p);
1180                 on_rq = p->on_rq;
1181                 ncsw = 0;
1182                 if (!match_state || p->state == match_state)
1183                         ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1184                 task_rq_unlock(rq, p, &flags);
1185 
1186                 /*
1187                  * If it changed from the expected state, bail out now.
1188                  */
1189                 if (unlikely(!ncsw))
1190                         break;
1191 
1192                 /*
1193                  * Was it really running after all now that we
1194                  * checked with the proper locks actually held?
1195                  *
1196                  * Oops. Go back and try again..
1197                  */
1198                 if (unlikely(running)) {
1199                         cpu_relax();
1200                         continue;
1201                 }
1202 
1203                 /*
1204                  * It's not enough that it's not actively running,
1205                  * it must be off the runqueue _entirely_, and not
1206                  * preempted!
1207                  *
1208                  * So if it was still runnable (but just not actively
1209                  * running right now), it's preempted, and we should
1210                  * yield - it could be a while.
1211                  */
1212                 if (unlikely(on_rq)) {
1213                         ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1214 
1215                         set_current_state(TASK_UNINTERRUPTIBLE);
1216                         schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1217                         continue;
1218                 }
1219 
1220                 /*
1221                  * Ahh, all good. It wasn't running, and it wasn't
1222                  * runnable, which means that it will never become
1223                  * running in the future either. We're all done!
1224                  */
1225                 break;
1226         }
1227 
1228         return ncsw;
1229 }
1230 
1231 /***
1232  * kick_process - kick a running thread to enter/exit the kernel
1233  * @p: the to-be-kicked thread
1234  *
1235  * Cause a process which is running on another CPU to enter
1236  * kernel-mode, without any delay. (to get signals handled.)
1237  *
1238  * NOTE: this function doesn't have to take the runqueue lock,
1239  * because all it wants to ensure is that the remote task enters
1240  * the kernel. If the IPI races and the task has been migrated
1241  * to another CPU then no harm is done and the purpose has been
1242  * achieved as well.
1243  */
1244 void kick_process(struct task_struct *p)
1245 {
1246         int cpu;
1247 
1248         preempt_disable();
1249         cpu = task_cpu(p);
1250         if ((cpu != smp_processor_id()) && task_curr(p))
1251                 smp_send_reschedule(cpu);
1252         preempt_enable();
1253 }
1254 EXPORT_SYMBOL_GPL(kick_process);
1255 #endif /* CONFIG_SMP */
1256 
1257 #ifdef CONFIG_SMP
1258 /*
1259  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1260  */
1261 static int select_fallback_rq(int cpu, struct task_struct *p)
1262 {
1263         int nid = cpu_to_node(cpu);
1264         const struct cpumask *nodemask = NULL;
1265         enum { cpuset, possible, fail } state = cpuset;
1266         int dest_cpu;
1267 
1268         /*
1269          * If the node that the cpu is on has been offlined, cpu_to_node()
1270          * will return -1. There is no cpu on the node, and we should
1271          * select the cpu on the other node.
1272          */
1273         if (nid != -1) {
1274                 nodemask = cpumask_of_node(nid);
1275 
1276                 /* Look for allowed, online CPU in same node. */
1277                 for_each_cpu(dest_cpu, nodemask) {
1278                         if (!cpu_online(dest_cpu))
1279                                 continue;
1280                         if (!cpu_active(dest_cpu))
1281                                 continue;
1282                         if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1283                                 return dest_cpu;
1284                 }
1285         }
1286 
1287         for (;;) {
1288                 /* Any allowed, online CPU? */
1289                 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1290                         if (!cpu_online(dest_cpu))
1291                                 continue;
1292                         if (!cpu_active(dest_cpu))
1293                                 continue;
1294                         goto out;
1295                 }
1296 
1297                 switch (state) {
1298                 case cpuset:
1299                         /* No more Mr. Nice Guy. */
1300                         cpuset_cpus_allowed_fallback(p);
1301                         state = possible;
1302                         break;
1303 
1304                 case possible:
1305                         do_set_cpus_allowed(p, cpu_possible_mask);
1306                         state = fail;
1307                         break;
1308 
1309                 case fail:
1310                         BUG();
1311                         break;
1312                 }
1313         }
1314 
1315 out:
1316         if (state != cpuset) {
1317                 /*
1318                  * Don't tell them about moving exiting tasks or
1319                  * kernel threads (both mm NULL), since they never
1320                  * leave kernel.
1321                  */
1322                 if (p->mm && printk_ratelimit()) {
1323                         printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1324                                         task_pid_nr(p), p->comm, cpu);
1325                 }
1326         }
1327 
1328         return dest_cpu;
1329 }
1330 
1331 /*
1332  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1333  */
1334 static inline
1335 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1336 {
1337         cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1338 
1339         /*
1340          * In order not to call set_task_cpu() on a blocking task we need
1341          * to rely on ttwu() to place the task on a valid ->cpus_allowed
1342          * cpu.
1343          *
1344          * Since this is common to all placement strategies, this lives here.
1345          *
1346          * [ this allows ->select_task() to simply return task_cpu(p) and
1347          *   not worry about this generic constraint ]
1348          */
1349         if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1350                      !cpu_online(cpu)))
1351                 cpu = select_fallback_rq(task_cpu(p), p);
1352 
1353         return cpu;
1354 }
1355 
1356 static void update_avg(u64 *avg, u64 sample)
1357 {
1358         s64 diff = sample - *avg;
1359         *avg += diff >> 3;
1360 }
1361 #endif
1362 
1363 static void
1364 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1365 {
1366 #ifdef CONFIG_SCHEDSTATS
1367         struct rq *rq = this_rq();
1368 
1369 #ifdef CONFIG_SMP
1370         int this_cpu = smp_processor_id();
1371 
1372         if (cpu == this_cpu) {
1373                 schedstat_inc(rq, ttwu_local);
1374                 schedstat_inc(p, se.statistics.nr_wakeups_local);
1375         } else {
1376                 struct sched_domain *sd;
1377 
1378                 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1379                 rcu_read_lock();
1380                 for_each_domain(this_cpu, sd) {
1381                         if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1382                                 schedstat_inc(sd, ttwu_wake_remote);
1383                                 break;
1384                         }
1385                 }
1386                 rcu_read_unlock();
1387         }
1388 
1389         if (wake_flags & WF_MIGRATED)
1390                 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1391 
1392 #endif /* CONFIG_SMP */
1393 
1394         schedstat_inc(rq, ttwu_count);
1395         schedstat_inc(p, se.statistics.nr_wakeups);
1396 
1397         if (wake_flags & WF_SYNC)
1398                 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1399 
1400 #endif /* CONFIG_SCHEDSTATS */
1401 }
1402 
1403 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1404 {
1405         activate_task(rq, p, en_flags);
1406         p->on_rq = 1;
1407 
1408         /* if a worker is waking up, notify workqueue */
1409         if (p->flags & PF_WQ_WORKER)
1410                 wq_worker_waking_up(p, cpu_of(rq));
1411 }
1412 
1413 /*
1414  * Mark the task runnable and perform wakeup-preemption.
1415  */
1416 static void
1417 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1418 {
1419         check_preempt_curr(rq, p, wake_flags);
1420         trace_sched_wakeup(p, true);
1421 
1422         p->state = TASK_RUNNING;
1423 #ifdef CONFIG_SMP
1424         if (p->sched_class->task_woken)
1425                 p->sched_class->task_woken(rq, p);
1426 
1427         if (rq->idle_stamp) {
1428                 u64 delta = rq_clock(rq) - rq->idle_stamp;
1429                 u64 max = 2*rq->max_idle_balance_cost;
1430 
1431                 update_avg(&rq->avg_idle, delta);
1432 
1433                 if (rq->avg_idle > max)
1434                         rq->avg_idle = max;
1435 
1436                 rq->idle_stamp = 0;
1437         }
1438 #endif
1439 }
1440 
1441 static void
1442 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1443 {
1444 #ifdef CONFIG_SMP
1445         if (p->sched_contributes_to_load)
1446                 rq->nr_uninterruptible--;
1447 #endif
1448 
1449         ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1450         ttwu_do_wakeup(rq, p, wake_flags);
1451 }
1452 
1453 /*
1454  * Called in case the task @p isn't fully descheduled from its runqueue,
1455  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1456  * since all we need to do is flip p->state to TASK_RUNNING, since
1457  * the task is still ->on_rq.
1458  */
1459 static int ttwu_remote(struct task_struct *p, int wake_flags)
1460 {
1461         struct rq *rq;
1462         int ret = 0;
1463 
1464         rq = __task_rq_lock(p);
1465         if (p->on_rq) {
1466                 /* check_preempt_curr() may use rq clock */
1467                 update_rq_clock(rq);
1468                 ttwu_do_wakeup(rq, p, wake_flags);
1469                 ret = 1;
1470         }
1471         __task_rq_unlock(rq);
1472 
1473         return ret;
1474 }
1475 
1476 #ifdef CONFIG_SMP
1477 static void sched_ttwu_pending(void)
1478 {
1479         struct rq *rq = this_rq();
1480         struct llist_node *llist = llist_del_all(&rq->wake_list);
1481         struct task_struct *p;
1482 
1483         raw_spin_lock(&rq->lock);
1484 
1485         while (llist) {
1486                 p = llist_entry(llist, struct task_struct, wake_entry);
1487                 llist = llist_next(llist);
1488                 ttwu_do_activate(rq, p, 0);
1489         }
1490 
1491         raw_spin_unlock(&rq->lock);
1492 }
1493 
1494 void scheduler_ipi(void)
1495 {
1496         /*
1497          * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1498          * TIF_NEED_RESCHED remotely (for the first time) will also send
1499          * this IPI.
1500          */
1501         preempt_fold_need_resched();
1502 
1503         if (llist_empty(&this_rq()->wake_list)
1504                         && !tick_nohz_full_cpu(smp_processor_id())
1505                         && !got_nohz_idle_kick())
1506                 return;
1507 
1508         /*
1509          * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1510          * traditionally all their work was done from the interrupt return
1511          * path. Now that we actually do some work, we need to make sure
1512          * we do call them.
1513          *
1514          * Some archs already do call them, luckily irq_enter/exit nest
1515          * properly.
1516          *
1517          * Arguably we should visit all archs and update all handlers,
1518          * however a fair share of IPIs are still resched only so this would
1519          * somewhat pessimize the simple resched case.
1520          */
1521         irq_enter();
1522         tick_nohz_full_check();
1523         sched_ttwu_pending();
1524 
1525         /*
1526          * Check if someone kicked us for doing the nohz idle load balance.
1527          */
1528         if (unlikely(got_nohz_idle_kick())) {
1529                 this_rq()->idle_balance = 1;
1530                 raise_softirq_irqoff(SCHED_SOFTIRQ);
1531         }
1532         irq_exit();
1533 }
1534 
1535 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1536 {
1537         if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1538                 smp_send_reschedule(cpu);
1539 }
1540 
1541 bool cpus_share_cache(int this_cpu, int that_cpu)
1542 {
1543         return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1544 }
1545 #endif /* CONFIG_SMP */
1546 
1547 static void ttwu_queue(struct task_struct *p, int cpu)
1548 {
1549         struct rq *rq = cpu_rq(cpu);
1550 
1551 #if defined(CONFIG_SMP)
1552         if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1553                 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1554                 ttwu_queue_remote(p, cpu);
1555                 return;
1556         }
1557 #endif
1558 
1559         raw_spin_lock(&rq->lock);
1560         ttwu_do_activate(rq, p, 0);
1561         raw_spin_unlock(&rq->lock);
1562 }
1563 
1564 /**
1565  * try_to_wake_up - wake up a thread
1566  * @p: the thread to be awakened
1567  * @state: the mask of task states that can be woken
1568  * @wake_flags: wake modifier flags (WF_*)
1569  *
1570  * Put it on the run-queue if it's not already there. The "current"
1571  * thread is always on the run-queue (except when the actual
1572  * re-schedule is in progress), and as such you're allowed to do
1573  * the simpler "current->state = TASK_RUNNING" to mark yourself
1574  * runnable without the overhead of this.
1575  *
1576  * Return: %true if @p was woken up, %false if it was already running.
1577  * or @state didn't match @p's state.
1578  */
1579 static int
1580 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1581 {
1582         unsigned long flags;
1583         int cpu, success = 0;
1584 
1585         /*
1586          * If we are going to wake up a thread waiting for CONDITION we
1587          * need to ensure that CONDITION=1 done by the caller can not be
1588          * reordered with p->state check below. This pairs with mb() in
1589          * set_current_state() the waiting thread does.
1590          */
1591         smp_mb__before_spinlock();
1592         raw_spin_lock_irqsave(&p->pi_lock, flags);
1593         if (!(p->state & state))
1594                 goto out;
1595 
1596         success = 1; /* we're going to change ->state */
1597         cpu = task_cpu(p);
1598 
1599         if (p->on_rq && ttwu_remote(p, wake_flags))
1600                 goto stat;
1601 
1602 #ifdef CONFIG_SMP
1603         /*
1604          * If the owning (remote) cpu is still in the middle of schedule() with
1605          * this task as prev, wait until its done referencing the task.
1606          */
1607         while (p->on_cpu)
1608                 cpu_relax();
1609         /*
1610          * Pairs with the smp_wmb() in finish_lock_switch().
1611          */
1612         smp_rmb();
1613 
1614         p->sched_contributes_to_load = !!task_contributes_to_load(p);
1615         p->state = TASK_WAKING;
1616 
1617         if (p->sched_class->task_waking)
1618                 p->sched_class->task_waking(p);
1619 
1620         cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1621         if (task_cpu(p) != cpu) {
1622                 wake_flags |= WF_MIGRATED;
1623                 set_task_cpu(p, cpu);
1624         }
1625 #endif /* CONFIG_SMP */
1626 
1627         ttwu_queue(p, cpu);
1628 stat:
1629         ttwu_stat(p, cpu, wake_flags);
1630 out:
1631         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1632 
1633         return success;
1634 }
1635 
1636 /**
1637  * try_to_wake_up_local - try to wake up a local task with rq lock held
1638  * @p: the thread to be awakened
1639  *
1640  * Put @p on the run-queue if it's not already there. The caller must
1641  * ensure that this_rq() is locked, @p is bound to this_rq() and not
1642  * the current task.
1643  */
1644 static void try_to_wake_up_local(struct task_struct *p)
1645 {
1646         struct rq *rq = task_rq(p);
1647 
1648         if (WARN_ON_ONCE(rq != this_rq()) ||
1649             WARN_ON_ONCE(p == current))
1650                 return;
1651 
1652         lockdep_assert_held(&rq->lock);
1653 
1654         if (!raw_spin_trylock(&p->pi_lock)) {
1655                 raw_spin_unlock(&rq->lock);
1656                 raw_spin_lock(&p->pi_lock);
1657                 raw_spin_lock(&rq->lock);
1658         }
1659 
1660         if (!(p->state & TASK_NORMAL))
1661                 goto out;
1662 
1663         if (!p->on_rq)
1664                 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1665 
1666         ttwu_do_wakeup(rq, p, 0);
1667         ttwu_stat(p, smp_processor_id(), 0);
1668 out:
1669         raw_spin_unlock(&p->pi_lock);
1670 }
1671 
1672 /**
1673  * wake_up_process - Wake up a specific process
1674  * @p: The process to be woken up.
1675  *
1676  * Attempt to wake up the nominated process and move it to the set of runnable
1677  * processes.
1678  *
1679  * Return: 1 if the process was woken up, 0 if it was already running.
1680  *
1681  * It may be assumed that this function implies a write memory barrier before
1682  * changing the task state if and only if any tasks are woken up.
1683  */
1684 int wake_up_process(struct task_struct *p)
1685 {
1686         WARN_ON(task_is_stopped_or_traced(p));
1687         return try_to_wake_up(p, TASK_NORMAL, 0);
1688 }
1689 EXPORT_SYMBOL(wake_up_process);
1690 
1691 int wake_up_state(struct task_struct *p, unsigned int state)
1692 {
1693         return try_to_wake_up(p, state, 0);
1694 }
1695 
1696 /*
1697  * Perform scheduler related setup for a newly forked process p.
1698  * p is forked by current.
1699  *
1700  * __sched_fork() is basic setup used by init_idle() too:
1701  */
1702 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1703 {
1704         p->on_rq                        = 0;
1705 
1706         p->se.on_rq                     = 0;
1707         p->se.exec_start                = 0;
1708         p->se.sum_exec_runtime          = 0;
1709         p->se.prev_sum_exec_runtime     = 0;
1710         p->se.nr_migrations             = 0;
1711         p->se.vruntime                  = 0;
1712         INIT_LIST_HEAD(&p->se.group_node);
1713 
1714 #ifdef CONFIG_SCHEDSTATS
1715         memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1716 #endif
1717 
1718         RB_CLEAR_NODE(&p->dl.rb_node);
1719         hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1720         p->dl.dl_runtime = p->dl.runtime = 0;
1721         p->dl.dl_deadline = p->dl.deadline = 0;
1722         p->dl.dl_period = 0;
1723         p->dl.flags = 0;
1724 
1725         INIT_LIST_HEAD(&p->rt.run_list);
1726 
1727 #ifdef CONFIG_PREEMPT_NOTIFIERS
1728         INIT_HLIST_HEAD(&p->preempt_notifiers);
1729 #endif
1730 
1731 #ifdef CONFIG_NUMA_BALANCING
1732         if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1733                 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1734                 p->mm->numa_scan_seq = 0;
1735         }
1736 
1737         if (clone_flags & CLONE_VM)
1738                 p->numa_preferred_nid = current->numa_preferred_nid;
1739         else
1740                 p->numa_preferred_nid = -1;
1741 
1742         p->node_stamp = 0ULL;
1743         p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1744         p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1745         p->numa_work.next = &p->numa_work;
1746         p->numa_faults_memory = NULL;
1747         p->numa_faults_buffer_memory = NULL;
1748         p->last_task_numa_placement = 0;
1749         p->last_sum_exec_runtime = 0;
1750 
1751         INIT_LIST_HEAD(&p->numa_entry);
1752         p->numa_group = NULL;
1753 #endif /* CONFIG_NUMA_BALANCING */
1754 }
1755 
1756 #ifdef CONFIG_NUMA_BALANCING
1757 #ifdef CONFIG_SCHED_DEBUG
1758 void set_numabalancing_state(bool enabled)
1759 {
1760         if (enabled)
1761                 sched_feat_set("NUMA");
1762         else
1763                 sched_feat_set("NO_NUMA");
1764 }
1765 #else
1766 __read_mostly bool numabalancing_enabled;
1767 
1768 void set_numabalancing_state(bool enabled)
1769 {
1770         numabalancing_enabled = enabled;
1771 }
1772 #endif /* CONFIG_SCHED_DEBUG */
1773 
1774 #ifdef CONFIG_PROC_SYSCTL
1775 int sysctl_numa_balancing(struct ctl_table *table, int write,
1776                          void __user *buffer, size_t *lenp, loff_t *ppos)
1777 {
1778         struct ctl_table t;
1779         int err;
1780         int state = numabalancing_enabled;
1781 
1782         if (write && !capable(CAP_SYS_ADMIN))
1783                 return -EPERM;
1784 
1785         t = *table;
1786         t.data = &state;
1787         err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1788         if (err < 0)
1789                 return err;
1790         if (write)
1791                 set_numabalancing_state(state);
1792         return err;
1793 }
1794 #endif
1795 #endif
1796 
1797 /*
1798  * fork()/clone()-time setup:
1799  */
1800 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1801 {
1802         unsigned long flags;
1803         int cpu = get_cpu();
1804 
1805         __sched_fork(clone_flags, p);
1806         /*
1807          * We mark the process as running here. This guarantees that
1808          * nobody will actually run it, and a signal or other external
1809          * event cannot wake it up and insert it on the runqueue either.
1810          */
1811         p->state = TASK_RUNNING;
1812 
1813         /*
1814          * Make sure we do not leak PI boosting priority to the child.
1815          */
1816         p->prio = current->normal_prio;
1817 
1818         /*
1819          * Revert to default priority/policy on fork if requested.
1820          */
1821         if (unlikely(p->sched_reset_on_fork)) {
1822                 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1823                         p->policy = SCHED_NORMAL;
1824                         p->static_prio = NICE_TO_PRIO(0);
1825                         p->rt_priority = 0;
1826                 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1827                         p->static_prio = NICE_TO_PRIO(0);
1828 
1829                 p->prio = p->normal_prio = __normal_prio(p);
1830                 set_load_weight(p);
1831 
1832                 /*
1833                  * We don't need the reset flag anymore after the fork. It has
1834                  * fulfilled its duty:
1835                  */
1836                 p->sched_reset_on_fork = 0;
1837         }
1838 
1839         if (dl_prio(p->prio)) {
1840                 put_cpu();
1841                 return -EAGAIN;
1842         } else if (rt_prio(p->prio)) {
1843                 p->sched_class = &rt_sched_class;
1844         } else {
1845                 p->sched_class = &fair_sched_class;
1846         }
1847 
1848         if (p->sched_class->task_fork)
1849                 p->sched_class->task_fork(p);
1850 
1851         /*
1852          * The child is not yet in the pid-hash so no cgroup attach races,
1853          * and the cgroup is pinned to this child due to cgroup_fork()
1854          * is ran before sched_fork().
1855          *
1856          * Silence PROVE_RCU.
1857          */
1858         raw_spin_lock_irqsave(&p->pi_lock, flags);
1859         set_task_cpu(p, cpu);
1860         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1861 
1862 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1863         if (likely(sched_info_on()))
1864                 memset(&p->sched_info, 0, sizeof(p->sched_info));
1865 #endif
1866 #if defined(CONFIG_SMP)
1867         p->on_cpu = 0;
1868 #endif
1869         init_task_preempt_count(p);
1870 #ifdef CONFIG_SMP
1871         plist_node_init(&p->pushable_tasks, MAX_PRIO);
1872         RB_CLEAR_NODE(&p->pushable_dl_tasks);
1873 #endif
1874 
1875         put_cpu();
1876         return 0;
1877 }
1878 
1879 unsigned long to_ratio(u64 period, u64 runtime)
1880 {
1881         if (runtime == RUNTIME_INF)
1882                 return 1ULL << 20;
1883 
1884         /*
1885          * Doing this here saves a lot of checks in all
1886          * the calling paths, and returning zero seems
1887          * safe for them anyway.
1888          */
1889         if (period == 0)
1890                 return 0;
1891 
1892         return div64_u64(runtime << 20, period);
1893 }
1894 
1895 #ifdef CONFIG_SMP
1896 inline struct dl_bw *dl_bw_of(int i)
1897 {
1898         return &cpu_rq(i)->rd->dl_bw;
1899 }
1900 
1901 static inline int dl_bw_cpus(int i)
1902 {
1903         struct root_domain *rd = cpu_rq(i)->rd;
1904         int cpus = 0;
1905 
1906         for_each_cpu_and(i, rd->span, cpu_active_mask)
1907                 cpus++;
1908 
1909         return cpus;
1910 }
1911 #else
1912 inline struct dl_bw *dl_bw_of(int i)
1913 {
1914         return &cpu_rq(i)->dl.dl_bw;
1915 }
1916 
1917 static inline int dl_bw_cpus(int i)
1918 {
1919         return 1;
1920 }
1921 #endif
1922 
1923 static inline
1924 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1925 {
1926         dl_b->total_bw -= tsk_bw;
1927 }
1928 
1929 static inline
1930 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1931 {
1932         dl_b->total_bw += tsk_bw;
1933 }
1934 
1935 static inline
1936 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1937 {
1938         return dl_b->bw != -1 &&
1939                dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1940 }
1941 
1942 /*
1943  * We must be sure that accepting a new task (or allowing changing the
1944  * parameters of an existing one) is consistent with the bandwidth
1945  * constraints. If yes, this function also accordingly updates the currently
1946  * allocated bandwidth to reflect the new situation.
1947  *
1948  * This function is called while holding p's rq->lock.
1949  */
1950 static int dl_overflow(struct task_struct *p, int policy,
1951                        const struct sched_attr *attr)
1952 {
1953 
1954         struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1955         u64 period = attr->sched_period ?: attr->sched_deadline;
1956         u64 runtime = attr->sched_runtime;
1957         u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
1958         int cpus, err = -1;
1959 
1960         if (new_bw == p->dl.dl_bw)
1961                 return 0;
1962 
1963         /*
1964          * Either if a task, enters, leave, or stays -deadline but changes
1965          * its parameters, we may need to update accordingly the total
1966          * allocated bandwidth of the container.
1967          */
1968         raw_spin_lock(&dl_b->lock);
1969         cpus = dl_bw_cpus(task_cpu(p));
1970         if (dl_policy(policy) && !task_has_dl_policy(p) &&
1971             !__dl_overflow(dl_b, cpus, 0, new_bw)) {
1972                 __dl_add(dl_b, new_bw);
1973                 err = 0;
1974         } else if (dl_policy(policy) && task_has_dl_policy(p) &&
1975                    !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
1976                 __dl_clear(dl_b, p->dl.dl_bw);
1977                 __dl_add(dl_b, new_bw);
1978                 err = 0;
1979         } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
1980                 __dl_clear(dl_b, p->dl.dl_bw);
1981                 err = 0;
1982         }
1983         raw_spin_unlock(&dl_b->lock);
1984 
1985         return err;
1986 }
1987 
1988 extern void init_dl_bw(struct dl_bw *dl_b);
1989 
1990 /*
1991  * wake_up_new_task - wake up a newly created task for the first time.
1992  *
1993  * This function will do some initial scheduler statistics housekeeping
1994  * that must be done for every newly created context, then puts the task
1995  * on the runqueue and wakes it.
1996  */
1997 void wake_up_new_task(struct task_struct *p)
1998 {
1999         unsigned long flags;
2000         struct rq *rq;
2001 
2002         raw_spin_lock_irqsave(&p->pi_lock, flags);
2003 #ifdef CONFIG_SMP
2004         /*
2005          * Fork balancing, do it here and not earlier because:
2006          *  - cpus_allowed can change in the fork path
2007          *  - any previously selected cpu might disappear through hotplug
2008          */
2009         set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2010 #endif
2011 
2012         /* Initialize new task's runnable average */
2013         init_task_runnable_average(p);
2014         rq = __task_rq_lock(p);
2015         activate_task(rq, p, 0);
2016         p->on_rq = 1;
2017         trace_sched_wakeup_new(p, true);
2018         check_preempt_curr(rq, p, WF_FORK);
2019 #ifdef CONFIG_SMP
2020         if (p->sched_class->task_woken)
2021                 p->sched_class->task_woken(rq, p);
2022 #endif
2023         task_rq_unlock(rq, p, &flags);
2024 }
2025 
2026 #ifdef CONFIG_PREEMPT_NOTIFIERS
2027 
2028 /**
2029  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2030  * @notifier: notifier struct to register
2031  */
2032 void preempt_notifier_register(struct preempt_notifier *notifier)
2033 {
2034         hlist_add_head(&notifier->link, &current->preempt_notifiers);
2035 }
2036 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2037 
2038 /**
2039  * preempt_notifier_unregister - no longer interested in preemption notifications
2040  * @notifier: notifier struct to unregister
2041  *
2042  * This is safe to call from within a preemption notifier.
2043  */
2044 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2045 {
2046         hlist_del(&notifier->link);
2047 }
2048 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2049 
2050 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2051 {
2052         struct preempt_notifier *notifier;
2053 
2054         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2055                 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2056 }
2057 
2058 static void
2059 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2060                                  struct task_struct *next)
2061 {
2062         struct preempt_notifier *notifier;
2063 
2064         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2065                 notifier->ops->sched_out(notifier, next);
2066 }
2067 
2068 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2069 
2070 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2071 {
2072 }
2073 
2074 static void
2075 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2076                                  struct task_struct *next)
2077 {
2078 }
2079 
2080 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2081 
2082 /**
2083  * prepare_task_switch - prepare to switch tasks
2084  * @rq: the runqueue preparing to switch
2085  * @prev: the current task that is being switched out
2086  * @next: the task we are going to switch to.
2087  *
2088  * This is called with the rq lock held and interrupts off. It must
2089  * be paired with a subsequent finish_task_switch after the context
2090  * switch.
2091  *
2092  * prepare_task_switch sets up locking and calls architecture specific
2093  * hooks.
2094  */
2095 static inline void
2096 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2097                     struct task_struct *next)
2098 {
2099         trace_sched_switch(prev, next);
2100         sched_info_switch(rq, prev, next);
2101         perf_event_task_sched_out(prev, next);
2102         fire_sched_out_preempt_notifiers(prev, next);
2103         prepare_lock_switch(rq, next);
2104         prepare_arch_switch(next);
2105 }
2106 
2107 /**
2108  * finish_task_switch - clean up after a task-switch
2109  * @rq: runqueue associated with task-switch
2110  * @prev: the thread we just switched away from.
2111  *
2112  * finish_task_switch must be called after the context switch, paired
2113  * with a prepare_task_switch call before the context switch.
2114  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2115  * and do any other architecture-specific cleanup actions.
2116  *
2117  * Note that we may have delayed dropping an mm in context_switch(). If
2118  * so, we finish that here outside of the runqueue lock. (Doing it
2119  * with the lock held can cause deadlocks; see schedule() for
2120  * details.)
2121  */
2122 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2123         __releases(rq->lock)
2124 {
2125         struct mm_struct *mm = rq->prev_mm;
2126         long prev_state;
2127 
2128         rq->prev_mm = NULL;
2129 
2130         /*
2131          * A task struct has one reference for the use as "current".
2132          * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2133          * schedule one last time. The schedule call will never return, and
2134          * the scheduled task must drop that reference.
2135          * The test for TASK_DEAD must occur while the runqueue locks are
2136          * still held, otherwise prev could be scheduled on another cpu, die
2137          * there before we look at prev->state, and then the reference would
2138          * be dropped twice.
2139          *              Manfred Spraul <manfred@colorfullife.com>
2140          */
2141         prev_state = prev->state;
2142         vtime_task_switch(prev);
2143         finish_arch_switch(prev);
2144         perf_event_task_sched_in(prev, current);
2145         finish_lock_switch(rq, prev);
2146         finish_arch_post_lock_switch();
2147 
2148         fire_sched_in_preempt_notifiers(current);
2149         if (mm)
2150                 mmdrop(mm);
2151         if (unlikely(prev_state == TASK_DEAD)) {
2152                 if (prev->sched_class->task_dead)
2153                         prev->sched_class->task_dead(prev);
2154 
2155                 /*
2156                  * Remove function-return probe instances associated with this
2157                  * task and put them back on the free list.
2158                  */
2159                 kprobe_flush_task(prev);
2160                 put_task_struct(prev);
2161         }
2162 
2163         tick_nohz_task_switch(current);
2164 }
2165 
2166 #ifdef CONFIG_SMP
2167 
2168 /* rq->lock is NOT held, but preemption is disabled */
2169 static inline void post_schedule(struct rq *rq)
2170 {
2171         if (rq->post_schedule) {
2172                 unsigned long flags;
2173 
2174                 raw_spin_lock_irqsave(&rq->lock, flags);
2175                 if (rq->curr->sched_class->post_schedule)
2176                         rq->curr->sched_class->post_schedule(rq);
2177                 raw_spin_unlock_irqrestore(&rq->lock, flags);
2178 
2179                 rq->post_schedule = 0;
2180         }
2181 }
2182 
2183 #else
2184 
2185 static inline void post_schedule(struct rq *rq)
2186 {
2187 }
2188 
2189 #endif
2190 
2191 /**
2192  * schedule_tail - first thing a freshly forked thread must call.
2193  * @prev: the thread we just switched away from.
2194  */
2195 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2196         __releases(rq->lock)
2197 {
2198         struct rq *rq = this_rq();
2199 
2200         finish_task_switch(rq, prev);
2201 
2202         /*
2203          * FIXME: do we need to worry about rq being invalidated by the
2204          * task_switch?
2205          */
2206         post_schedule(rq);
2207 
2208 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2209         /* In this case, finish_task_switch does not reenable preemption */
2210         preempt_enable();
2211 #endif
2212         if (current->set_child_tid)
2213                 put_user(task_pid_vnr(current), current->set_child_tid);
2214 }
2215 
2216 /*
2217  * context_switch - switch to the new MM and the new
2218  * thread's register state.
2219  */
2220 static inline void
2221 context_switch(struct rq *rq, struct task_struct *prev,
2222                struct task_struct *next)
2223 {
2224         struct mm_struct *mm, *oldmm;
2225 
2226         prepare_task_switch(rq, prev, next);
2227 
2228         mm = next->mm;
2229         oldmm = prev->active_mm;
2230         /*
2231          * For paravirt, this is coupled with an exit in switch_to to
2232          * combine the page table reload and the switch backend into
2233          * one hypercall.
2234          */
2235         arch_start_context_switch(prev);
2236 
2237         if (!mm) {
2238                 next->active_mm = oldmm;
2239                 atomic_inc(&oldmm->mm_count);
2240                 enter_lazy_tlb(oldmm, next);
2241         } else
2242                 switch_mm(oldmm, mm, next);
2243 
2244         if (!prev->mm) {
2245                 prev->active_mm = NULL;
2246                 rq->prev_mm = oldmm;
2247         }
2248         /*
2249          * Since the runqueue lock will be released by the next
2250          * task (which is an invalid locking op but in the case
2251          * of the scheduler it's an obvious special-case), so we
2252          * do an early lockdep release here:
2253          */
2254 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2255         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2256 #endif
2257 
2258         context_tracking_task_switch(prev, next);
2259         /* Here we just switch the register state and the stack. */
2260         switch_to(prev, next, prev);
2261 
2262         barrier();
2263         /*
2264          * this_rq must be evaluated again because prev may have moved
2265          * CPUs since it called schedule(), thus the 'rq' on its stack
2266          * frame will be invalid.
2267          */
2268         finish_task_switch(this_rq(), prev);
2269 }
2270 
2271 /*
2272  * nr_running and nr_context_switches:
2273  *
2274  * externally visible scheduler statistics: current number of runnable
2275  * threads, total number of context switches performed since bootup.
2276  */
2277 unsigned long nr_running(void)
2278 {
2279         unsigned long i, sum = 0;
2280 
2281         for_each_online_cpu(i)
2282                 sum += cpu_rq(i)->nr_running;
2283 
2284         return sum;
2285 }
2286 
2287 unsigned long long nr_context_switches(void)
2288 {
2289         int i;
2290         unsigned long long sum = 0;
2291 
2292         for_each_possible_cpu(i)
2293                 sum += cpu_rq(i)->nr_switches;
2294 
2295         return sum;
2296 }
2297 
2298 unsigned long nr_iowait(void)
2299 {
2300         unsigned long i, sum = 0;
2301 
2302         for_each_possible_cpu(i)
2303                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2304 
2305         return sum;
2306 }
2307 
2308 unsigned long nr_iowait_cpu(int cpu)
2309 {
2310         struct rq *this = cpu_rq(cpu);
2311         return atomic_read(&this->nr_iowait);
2312 }
2313 
2314 #ifdef CONFIG_SMP
2315 
2316 /*
2317  * sched_exec - execve() is a valuable balancing opportunity, because at
2318  * this point the task has the smallest effective memory and cache footprint.
2319  */
2320 void sched_exec(void)
2321 {
2322         struct task_struct *p = current;
2323         unsigned long flags;
2324         int dest_cpu;
2325 
2326         raw_spin_lock_irqsave(&p->pi_lock, flags);
2327         dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2328         if (dest_cpu == smp_processor_id())
2329                 goto unlock;
2330 
2331         if (likely(cpu_active(dest_cpu))) {
2332                 struct migration_arg arg = { p, dest_cpu };
2333 
2334                 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2335                 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2336                 return;
2337         }
2338 unlock:
2339         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2340 }
2341 
2342 #endif
2343 
2344 DEFINE_PER_CPU(struct kernel_stat, kstat);
2345 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2346 
2347 EXPORT_PER_CPU_SYMBOL(kstat);
2348 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2349 
2350 /*
2351  * Return any ns on the sched_clock that have not yet been accounted in
2352  * @p in case that task is currently running.
2353  *
2354  * Called with task_rq_lock() held on @rq.
2355  */
2356 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2357 {
2358         u64 ns = 0;
2359 
2360         if (task_current(rq, p)) {
2361                 update_rq_clock(rq);
2362                 ns = rq_clock_task(rq) - p->se.exec_start;
2363                 if ((s64)ns < 0)
2364                         ns = 0;
2365         }
2366 
2367         return ns;
2368 }
2369 
2370 unsigned long long task_delta_exec(struct task_struct *p)
2371 {
2372         unsigned long flags;
2373         struct rq *rq;
2374         u64 ns = 0;
2375 
2376         rq = task_rq_lock(p, &flags);
2377         ns = do_task_delta_exec(p, rq);
2378         task_rq_unlock(rq, p, &flags);
2379 
2380         return ns;
2381 }
2382 
2383 /*
2384  * Return accounted runtime for the task.
2385  * In case the task is currently running, return the runtime plus current's
2386  * pending runtime that have not been accounted yet.
2387  */
2388 unsigned long long task_sched_runtime(struct task_struct *p)
2389 {
2390         unsigned long flags;
2391         struct rq *rq;
2392         u64 ns = 0;
2393 
2394 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2395         /*
2396          * 64-bit doesn't need locks to atomically read a 64bit value.
2397          * So we have a optimization chance when the task's delta_exec is 0.
2398          * Reading ->on_cpu is racy, but this is ok.
2399          *
2400          * If we race with it leaving cpu, we'll take a lock. So we're correct.
2401          * If we race with it entering cpu, unaccounted time is 0. This is
2402          * indistinguishable from the read occurring a few cycles earlier.
2403          */
2404         if (!p->on_cpu)
2405                 return p->se.sum_exec_runtime;
2406 #endif
2407 
2408         rq = task_rq_lock(p, &flags);
2409         ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2410         task_rq_unlock(rq, p, &flags);
2411 
2412         return ns;
2413 }
2414 
2415 /*
2416  * This function gets called by the timer code, with HZ frequency.
2417  * We call it with interrupts disabled.
2418  */
2419 void scheduler_tick(void)
2420 {
2421         int cpu = smp_processor_id();
2422         struct rq *rq = cpu_rq(cpu);
2423         struct task_struct *curr = rq->curr;
2424 
2425         sched_clock_tick();
2426 
2427         raw_spin_lock(&rq->lock);
2428         update_rq_clock(rq);
2429         curr->sched_class->task_tick(rq, curr, 0);
2430         update_cpu_load_active(rq);
2431         raw_spin_unlock(&rq->lock);
2432 
2433         perf_event_task_tick();
2434 
2435 #ifdef CONFIG_SMP
2436         rq->idle_balance = idle_cpu(cpu);
2437         trigger_load_balance(rq);
2438 #endif
2439         rq_last_tick_reset(rq);
2440 }
2441 
2442 #ifdef CONFIG_NO_HZ_FULL
2443 /**
2444  * scheduler_tick_max_deferment
2445  *
2446  * Keep at least one tick per second when a single
2447  * active task is running because the scheduler doesn't
2448  * yet completely support full dynticks environment.
2449  *
2450  * This makes sure that uptime, CFS vruntime, load
2451  * balancing, etc... continue to move forward, even
2452  * with a very low granularity.
2453  *
2454  * Return: Maximum deferment in nanoseconds.
2455  */
2456 u64 scheduler_tick_max_deferment(void)
2457 {
2458         struct rq *rq = this_rq();
2459         unsigned long next, now = ACCESS_ONCE(jiffies);
2460 
2461         next = rq->last_sched_tick + HZ;
2462 
2463         if (time_before_eq(next, now))
2464                 return 0;
2465 
2466         return jiffies_to_nsecs(next - now);
2467 }
2468 #endif
2469 
2470 notrace unsigned long get_parent_ip(unsigned long addr)
2471 {
2472         if (in_lock_functions(addr)) {
2473                 addr = CALLER_ADDR2;
2474                 if (in_lock_functions(addr))
2475                         addr = CALLER_ADDR3;
2476         }
2477         return addr;
2478 }
2479 
2480 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2481                                 defined(CONFIG_PREEMPT_TRACER))
2482 
2483 void __kprobes preempt_count_add(int val)
2484 {
2485 #ifdef CONFIG_DEBUG_PREEMPT
2486         /*
2487          * Underflow?
2488          */
2489         if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2490                 return;
2491 #endif
2492         __preempt_count_add(val);
2493 #ifdef CONFIG_DEBUG_PREEMPT
2494         /*
2495          * Spinlock count overflowing soon?
2496          */
2497         DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2498                                 PREEMPT_MASK - 10);
2499 #endif
2500         if (preempt_count() == val) {
2501                 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2502 #ifdef CONFIG_DEBUG_PREEMPT
2503                 current->preempt_disable_ip = ip;
2504 #endif
2505                 trace_preempt_off(CALLER_ADDR0, ip);
2506         }
2507 }
2508 EXPORT_SYMBOL(preempt_count_add);
2509 
2510 void __kprobes preempt_count_sub(int val)
2511 {
2512 #ifdef CONFIG_DEBUG_PREEMPT
2513         /*
2514          * Underflow?
2515          */
2516         if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2517                 return;
2518         /*
2519          * Is the spinlock portion underflowing?
2520          */
2521         if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2522                         !(preempt_count() & PREEMPT_MASK)))
2523                 return;
2524 #endif
2525 
2526         if (preempt_count() == val)
2527                 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2528         __preempt_count_sub(val);
2529 }
2530 EXPORT_SYMBOL(preempt_count_sub);
2531 
2532 #endif
2533 
2534 /*
2535  * Print scheduling while atomic bug:
2536  */
2537 static noinline void __schedule_bug(struct task_struct *prev)
2538 {
2539         if (oops_in_progress)
2540                 return;
2541 
2542         printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2543                 prev->comm, prev->pid, preempt_count());
2544 
2545         debug_show_held_locks(prev);
2546         print_modules();
2547         if (irqs_disabled())
2548                 print_irqtrace_events(prev);
2549 #ifdef CONFIG_DEBUG_PREEMPT
2550         if (in_atomic_preempt_off()) {
2551                 pr_err("Preemption disabled at:");
2552                 print_ip_sym(current->preempt_disable_ip);
2553                 pr_cont("\n");
2554         }
2555 #endif
2556         dump_stack();
2557         add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2558 }
2559 
2560 /*
2561  * Various schedule()-time debugging checks and statistics:
2562  */
2563 static inline void schedule_debug(struct task_struct *prev)
2564 {
2565         /*
2566          * Test if we are atomic. Since do_exit() needs to call into
2567          * schedule() atomically, we ignore that path. Otherwise whine
2568          * if we are scheduling when we should not.
2569          */
2570         if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2571                 __schedule_bug(prev);
2572         rcu_sleep_check();
2573 
2574         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2575 
2576         schedstat_inc(this_rq(), sched_count);
2577 }
2578 
2579 /*
2580  * Pick up the highest-prio task:
2581  */
2582 static inline struct task_struct *
2583 pick_next_task(struct rq *rq, struct task_struct *prev)
2584 {
2585         const struct sched_class *class = &fair_sched_class;
2586         struct task_struct *p;
2587 
2588         /*
2589          * Optimization: we know that if all tasks are in
2590          * the fair class we can call that function directly:
2591          */
2592         if (likely(prev->sched_class == class &&
2593                    rq->nr_running == rq->cfs.h_nr_running)) {
2594                 p = fair_sched_class.pick_next_task(rq, prev);
2595                 if (unlikely(p == RETRY_TASK))
2596                         goto again;
2597 
2598                 /* assumes fair_sched_class->next == idle_sched_class */
2599                 if (unlikely(!p))
2600                         p = idle_sched_class.pick_next_task(rq, prev);
2601 
2602                 return p;
2603         }
2604 
2605 again:
2606         for_each_class(class) {
2607                 p = class->pick_next_task(rq, prev);
2608                 if (p) {
2609                         if (unlikely(p == RETRY_TASK))
2610                                 goto again;
2611                         return p;
2612                 }
2613         }
2614 
2615         BUG(); /* the idle class will always have a runnable task */
2616 }
2617 
2618 /*
2619  * __schedule() is the main scheduler function.
2620  *
2621  * The main means of driving the scheduler and thus entering this function are:
2622  *
2623  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2624  *
2625  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2626  *      paths. For example, see arch/x86/entry_64.S.
2627  *
2628  *      To drive preemption between tasks, the scheduler sets the flag in timer
2629  *      interrupt handler scheduler_tick().
2630  *
2631  *   3. Wakeups don't really cause entry into schedule(). They add a
2632  *      task to the run-queue and that's it.
2633  *
2634  *      Now, if the new task added to the run-queue preempts the current
2635  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2636  *      called on the nearest possible occasion:
2637  *
2638  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2639  *
2640  *         - in syscall or exception context, at the next outmost
2641  *           preempt_enable(). (this might be as soon as the wake_up()'s
2642  *           spin_unlock()!)
2643  *
2644  *         - in IRQ context, return from interrupt-handler to
2645  *           preemptible context
2646  *
2647  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2648  *         then at the next:
2649  *
2650  *          - cond_resched() call
2651  *          - explicit schedule() call
2652  *          - return from syscall or exception to user-space
2653  *          - return from interrupt-handler to user-space
2654  */
2655 static void __sched __schedule(void)
2656 {
2657         struct task_struct *prev, *next;
2658         unsigned long *switch_count;
2659         struct rq *rq;
2660         int cpu;
2661 
2662 need_resched:
2663         preempt_disable();
2664         cpu = smp_processor_id();
2665         rq = cpu_rq(cpu);
2666         rcu_note_context_switch(cpu);
2667         prev = rq->curr;
2668 
2669         schedule_debug(prev);
2670 
2671         if (sched_feat(HRTICK))
2672                 hrtick_clear(rq);
2673 
2674         /*
2675          * Make sure that signal_pending_state()->signal_pending() below
2676          * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2677          * done by the caller to avoid the race with signal_wake_up().
2678          */
2679         smp_mb__before_spinlock();
2680         raw_spin_lock_irq(&rq->lock);
2681 
2682         switch_count = &prev->nivcsw;
2683         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2684                 if (unlikely(signal_pending_state(prev->state, prev))) {
2685                         prev->state = TASK_RUNNING;
2686                 } else {
2687                         deactivate_task(rq, prev, DEQUEUE_SLEEP);
2688                         prev->on_rq = 0;
2689 
2690                         /*
2691                          * If a worker went to sleep, notify and ask workqueue
2692                          * whether it wants to wake up a task to maintain
2693                          * concurrency.
2694                          */
2695                         if (prev->flags & PF_WQ_WORKER) {
2696                                 struct task_struct *to_wakeup;
2697 
2698                                 to_wakeup = wq_worker_sleeping(prev, cpu);
2699                                 if (to_wakeup)
2700                                         try_to_wake_up_local(to_wakeup);
2701                         }
2702                 }
2703                 switch_count = &prev->nvcsw;
2704         }
2705 
2706         if (prev->on_rq || rq->skip_clock_update < 0)
2707                 update_rq_clock(rq);
2708 
2709         next = pick_next_task(rq, prev);
2710         clear_tsk_need_resched(prev);
2711         clear_preempt_need_resched();
2712         rq->skip_clock_update = 0;
2713 
2714         if (likely(prev != next)) {
2715                 rq->nr_switches++;
2716                 rq->curr = next;
2717                 ++*switch_count;
2718 
2719                 context_switch(rq, prev, next); /* unlocks the rq */
2720                 /*
2721                  * The context switch have flipped the stack from under us
2722                  * and restored the local variables which were saved when
2723                  * this task called schedule() in the past. prev == current
2724                  * is still correct, but it can be moved to another cpu/rq.
2725                  */
2726                 cpu = smp_processor_id();
2727                 rq = cpu_rq(cpu);
2728         } else
2729                 raw_spin_unlock_irq(&rq->lock);
2730 
2731         post_schedule(rq);
2732 
2733         sched_preempt_enable_no_resched();
2734         if (need_resched())
2735                 goto need_resched;
2736 }
2737 
2738 static inline void sched_submit_work(struct task_struct *tsk)
2739 {
2740         if (!tsk->state || tsk_is_pi_blocked(tsk))
2741                 return;
2742         /*
2743          * If we are going to sleep and we have plugged IO queued,
2744          * make sure to submit it to avoid deadlocks.
2745          */
2746         if (blk_needs_flush_plug(tsk))
2747                 blk_schedule_flush_plug(tsk);
2748 }
2749 
2750 asmlinkage __visible void __sched schedule(void)
2751 {
2752         struct task_struct *tsk = current;
2753 
2754         sched_submit_work(tsk);
2755         __schedule();
2756 }
2757 EXPORT_SYMBOL(schedule);
2758 
2759 #ifdef CONFIG_CONTEXT_TRACKING
2760 asmlinkage __visible void __sched schedule_user(void)
2761 {
2762         /*
2763          * If we come here after a random call to set_need_resched(),
2764          * or we have been woken up remotely but the IPI has not yet arrived,
2765          * we haven't yet exited the RCU idle mode. Do it here manually until
2766          * we find a better solution.
2767          */
2768         user_exit();
2769         schedule();
2770         user_enter();
2771 }
2772 #endif
2773 
2774 /**
2775  * schedule_preempt_disabled - called with preemption disabled
2776  *
2777  * Returns with preemption disabled. Note: preempt_count must be 1
2778  */
2779 void __sched schedule_preempt_disabled(void)
2780 {
2781         sched_preempt_enable_no_resched();
2782         schedule();
2783         preempt_disable();
2784 }
2785 
2786 #ifdef CONFIG_PREEMPT
2787 /*
2788  * this is the entry point to schedule() from in-kernel preemption
2789  * off of preempt_enable. Kernel preemptions off return from interrupt
2790  * occur there and call schedule directly.
2791  */
2792 asmlinkage __visible void __sched notrace preempt_schedule(void)
2793 {
2794         /*
2795          * If there is a non-zero preempt_count or interrupts are disabled,
2796          * we do not want to preempt the current task. Just return..
2797          */
2798         if (likely(!preemptible()))
2799                 return;
2800 
2801         do {
2802                 __preempt_count_add(PREEMPT_ACTIVE);
2803                 __schedule();
2804                 __preempt_count_sub(PREEMPT_ACTIVE);
2805 
2806                 /*
2807                  * Check again in case we missed a preemption opportunity
2808                  * between schedule and now.
2809                  */
2810                 barrier();
2811         } while (need_resched());
2812 }
2813 EXPORT_SYMBOL(preempt_schedule);
2814 #endif /* CONFIG_PREEMPT */
2815 
2816 /*
2817  * this is the entry point to schedule() from kernel preemption
2818  * off of irq context.
2819  * Note, that this is called and return with irqs disabled. This will
2820  * protect us against recursive calling from irq.
2821  */
2822 asmlinkage __visible void __sched preempt_schedule_irq(void)
2823 {
2824         enum ctx_state prev_state;
2825 
2826         /* Catch callers which need to be fixed */
2827         BUG_ON(preempt_count() || !irqs_disabled());
2828 
2829         prev_state = exception_enter();
2830 
2831         do {
2832                 __preempt_count_add(PREEMPT_ACTIVE);
2833                 local_irq_enable();
2834                 __schedule();
2835                 local_irq_disable();
2836                 __preempt_count_sub(PREEMPT_ACTIVE);
2837 
2838                 /*
2839                  * Check again in case we missed a preemption opportunity
2840                  * between schedule and now.
2841                  */
2842                 barrier();
2843         } while (need_resched());
2844 
2845         exception_exit(prev_state);
2846 }
2847 
2848 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2849                           void *key)
2850 {
2851         return try_to_wake_up(curr->private, mode, wake_flags);
2852 }
2853 EXPORT_SYMBOL(default_wake_function);
2854 
2855 #ifdef CONFIG_RT_MUTEXES
2856 
2857 /*
2858  * rt_mutex_setprio - set the current priority of a task
2859  * @p: task
2860  * @prio: prio value (kernel-internal form)
2861  *
2862  * This function changes the 'effective' priority of a task. It does
2863  * not touch ->normal_prio like __setscheduler().
2864  *
2865  * Used by the rt_mutex code to implement priority inheritance
2866  * logic. Call site only calls if the priority of the task changed.
2867  */
2868 void rt_mutex_setprio(struct task_struct *p, int prio)
2869 {
2870         int oldprio, on_rq, running, enqueue_flag = 0;
2871         struct rq *rq;
2872         const struct sched_class *prev_class;
2873 
2874         BUG_ON(prio > MAX_PRIO);
2875 
2876         rq = __task_rq_lock(p);
2877 
2878         /*
2879          * Idle task boosting is a nono in general. There is one
2880          * exception, when PREEMPT_RT and NOHZ is active:
2881          *
2882          * The idle task calls get_next_timer_interrupt() and holds
2883          * the timer wheel base->lock on the CPU and another CPU wants
2884          * to access the timer (probably to cancel it). We can safely
2885          * ignore the boosting request, as the idle CPU runs this code
2886          * with interrupts disabled and will complete the lock
2887          * protected section without being interrupted. So there is no
2888          * real need to boost.
2889          */
2890         if (unlikely(p == rq->idle)) {
2891                 WARN_ON(p != rq->curr);
2892                 WARN_ON(p->pi_blocked_on);
2893                 goto out_unlock;
2894         }
2895 
2896         trace_sched_pi_setprio(p, prio);
2897         p->pi_top_task = rt_mutex_get_top_task(p);
2898         oldprio = p->prio;
2899         prev_class = p->sched_class;
2900         on_rq = p->on_rq;
2901         running = task_current(rq, p);
2902         if (on_rq)
2903                 dequeue_task(rq, p, 0);
2904         if (running)
2905                 p->sched_class->put_prev_task(rq, p);
2906 
2907         /*
2908          * Boosting condition are:
2909          * 1. -rt task is running and holds mutex A
2910          *      --> -dl task blocks on mutex A
2911          *
2912          * 2. -dl task is running and holds mutex A
2913          *      --> -dl task blocks on mutex A and could preempt the
2914          *          running task
2915          */
2916         if (dl_prio(prio)) {
2917                 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2918                         dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2919                         p->dl.dl_boosted = 1;
2920                         p->dl.dl_throttled = 0;
2921                         enqueue_flag = ENQUEUE_REPLENISH;
2922                 } else
2923                         p->dl.dl_boosted = 0;
2924                 p->sched_class = &dl_sched_class;
2925         } else if (rt_prio(prio)) {
2926                 if (dl_prio(oldprio))
2927                         p->dl.dl_boosted = 0;
2928                 if (oldprio < prio)
2929                         enqueue_flag = ENQUEUE_HEAD;
2930                 p->sched_class = &rt_sched_class;
2931         } else {
2932                 if (dl_prio(oldprio))
2933                         p->dl.dl_boosted = 0;
2934                 p->sched_class = &fair_sched_class;
2935         }
2936 
2937         p->prio = prio;
2938 
2939         if (running)
2940                 p->sched_class->set_curr_task(rq);
2941         if (on_rq)
2942                 enqueue_task(rq, p, enqueue_flag);
2943 
2944         check_class_changed(rq, p, prev_class, oldprio);
2945 out_unlock:
2946         __task_rq_unlock(rq);
2947 }
2948 #endif
2949 
2950 void set_user_nice(struct task_struct *p, long nice)
2951 {
2952         int old_prio, delta, on_rq;
2953         unsigned long flags;
2954         struct rq *rq;
2955 
2956         if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
2957                 return;
2958         /*
2959          * We have to be careful, if called from sys_setpriority(),
2960          * the task might be in the middle of scheduling on another CPU.
2961          */
2962         rq = task_rq_lock(p, &flags);
2963         /*
2964          * The RT priorities are set via sched_setscheduler(), but we still
2965          * allow the 'normal' nice value to be set - but as expected
2966          * it wont have any effect on scheduling until the task is
2967          * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
2968          */
2969         if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2970                 p->static_prio = NICE_TO_PRIO(nice);
2971                 goto out_unlock;
2972         }
2973         on_rq = p->on_rq;
2974         if (on_rq)
2975                 dequeue_task(rq, p, 0);
2976 
2977         p->static_prio = NICE_TO_PRIO(nice);
2978         set_load_weight(p);
2979         old_prio = p->prio;
2980         p->prio = effective_prio(p);
2981         delta = p->prio - old_prio;
2982 
2983         if (on_rq) {
2984                 enqueue_task(rq, p, 0);
2985                 /*
2986                  * If the task increased its priority or is running and
2987                  * lowered its priority, then reschedule its CPU:
2988                  */
2989                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
2990                         resched_task(rq->curr);
2991         }
2992 out_unlock:
2993         task_rq_unlock(rq, p, &flags);
2994 }
2995 EXPORT_SYMBOL(set_user_nice);
2996 
2997 /*
2998  * can_nice - check if a task can reduce its nice value
2999  * @p: task
3000  * @nice: nice value
3001  */
3002 int can_nice(const struct task_struct *p, const int nice)
3003 {
3004         /* convert nice value [19,-20] to rlimit style value [1,40] */
3005         int nice_rlim = 20 - nice;
3006 
3007         return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3008                 capable(CAP_SYS_NICE));
3009 }
3010 
3011 #ifdef __ARCH_WANT_SYS_NICE
3012 
3013 /*
3014  * sys_nice - change the priority of the current process.
3015  * @increment: priority increment
3016  *
3017  * sys_setpriority is a more generic, but much slower function that
3018  * does similar things.
3019  */
3020 SYSCALL_DEFINE1(nice, int, increment)
3021 {
3022         long nice, retval;
3023         if (!ccs_capable(CCS_SYS_NICE))
3024                 return -EPERM;
3025 
3026         /*
3027          * Setpriority might change our priority at the same moment.
3028          * We don't have to worry. Conceptually one call occurs first
3029          * and we have a single winner.
3030          */
3031         if (increment < -40)
3032                 increment = -40;
3033         if (increment > 40)
3034                 increment = 40;
3035 
3036         nice = task_nice(current) + increment;
3037         if (nice < MIN_NICE)
3038                 nice = MIN_NICE;
3039         if (nice > MAX_NICE)
3040                 nice = MAX_NICE;
3041 
3042         if (increment < 0 && !can_nice(current, nice))
3043                 return -EPERM;
3044 
3045         retval = security_task_setnice(current, nice);
3046         if (retval)
3047                 return retval;
3048 
3049         set_user_nice(current, nice);
3050         return 0;
3051 }
3052 
3053 #endif
3054 
3055 /**
3056  * task_prio - return the priority value of a given task.
3057  * @p: the task in question.
3058  *
3059  * Return: The priority value as seen by users in /proc.
3060  * RT tasks are offset by -200. Normal tasks are centered
3061  * around 0, value goes from -16 to +15.
3062  */
3063 int task_prio(const struct task_struct *p)
3064 {
3065         return p->prio - MAX_RT_PRIO;
3066 }
3067 
3068 /**
3069  * idle_cpu - is a given cpu idle currently?
3070  * @cpu: the processor in question.
3071  *
3072  * Return: 1 if the CPU is currently idle. 0 otherwise.
3073  */
3074 int idle_cpu(int cpu)
3075 {
3076         struct rq *rq = cpu_rq(cpu);
3077 
3078         if (rq->curr != rq->idle)
3079                 return 0;
3080 
3081         if (rq->nr_running)
3082                 return 0;
3083 
3084 #ifdef CONFIG_SMP
3085         if (!llist_empty(&rq->wake_list))
3086                 return 0;
3087 #endif
3088 
3089         return 1;
3090 }
3091 
3092 /**
3093  * idle_task - return the idle task for a given cpu.
3094  * @cpu: the processor in question.
3095  *
3096  * Return: The idle task for the cpu @cpu.
3097  */
3098 struct task_struct *idle_task(int cpu)
3099 {
3100         return cpu_rq(cpu)->idle;
3101 }
3102 
3103 /**
3104  * find_process_by_pid - find a process with a matching PID value.
3105  * @pid: the pid in question.
3106  *
3107  * The task of @pid, if found. %NULL otherwise.
3108  */
3109 static struct task_struct *find_process_by_pid(pid_t pid)
3110 {
3111         return pid ? find_task_by_vpid(pid) : current;
3112 }
3113 
3114 /*
3115  * This function initializes the sched_dl_entity of a newly becoming
3116  * SCHED_DEADLINE task.
3117  *
3118  * Only the static values are considered here, the actual runtime and the
3119  * absolute deadline will be properly calculated when the task is enqueued
3120  * for the first time with its new policy.
3121  */
3122 static void
3123 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3124 {
3125         struct sched_dl_entity *dl_se = &p->dl;
3126 
3127         init_dl_task_timer(dl_se);
3128         dl_se->dl_runtime = attr->sched_runtime;
3129         dl_se->dl_deadline = attr->sched_deadline;
3130         dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3131         dl_se->flags = attr->sched_flags;
3132         dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3133         dl_se->dl_throttled = 0;
3134         dl_se->dl_new = 1;
3135         dl_se->dl_yielded = 0;
3136 }
3137 
3138 static void __setscheduler_params(struct task_struct *p,
3139                 const struct sched_attr *attr)
3140 {
3141         int policy = attr->sched_policy;
3142 
3143         if (policy == -1) /* setparam */
3144                 policy = p->policy;
3145 
3146         p->policy = policy;
3147 
3148         if (dl_policy(policy))
3149                 __setparam_dl(p, attr);
3150         else if (fair_policy(policy))
3151                 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3152 
3153         /*
3154          * __sched_setscheduler() ensures attr->sched_priority == 0 when
3155          * !rt_policy. Always setting this ensures that things like
3156          * getparam()/getattr() don't report silly values for !rt tasks.
3157          */
3158         p->rt_priority = attr->sched_priority;
3159         p->normal_prio = normal_prio(p);
3160         set_load_weight(p);
3161 }
3162 
3163 /* Actually do priority change: must hold pi & rq lock. */
3164 static void __setscheduler(struct rq *rq, struct task_struct *p,
3165                            const struct sched_attr *attr)
3166 {
3167         __setscheduler_params(p, attr);
3168 
3169         /*
3170          * If we get here, there was no pi waiters boosting the
3171          * task. It is safe to use the normal prio.
3172          */
3173         p->prio = normal_prio(p);
3174 
3175         if (dl_prio(p->prio))
3176                 p->sched_class = &dl_sched_class;
3177         else if (rt_prio(p->prio))
3178                 p->sched_class = &rt_sched_class;
3179         else
3180                 p->sched_class = &fair_sched_class;
3181 }
3182 
3183 static void
3184 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3185 {
3186         struct sched_dl_entity *dl_se = &p->dl;
3187 
3188         attr->sched_priority = p->rt_priority;
3189         attr->sched_runtime = dl_se->dl_runtime;
3190         attr->sched_deadline = dl_se->dl_deadline;
3191         attr->sched_period = dl_se->dl_period;
3192         attr->sched_flags = dl_se->flags;
3193 }
3194 
3195 /*
3196  * This function validates the new parameters of a -deadline task.
3197  * We ask for the deadline not being zero, and greater or equal
3198  * than the runtime, as well as the period of being zero or
3199  * greater than deadline. Furthermore, we have to be sure that
3200  * user parameters are above the internal resolution of 1us (we
3201  * check sched_runtime only since it is always the smaller one) and
3202  * below 2^63 ns (we have to check both sched_deadline and
3203  * sched_period, as the latter can be zero).
3204  */
3205 static bool
3206 __checkparam_dl(const struct sched_attr *attr)
3207 {
3208         /* deadline != 0 */
3209         if (attr->sched_deadline == 0)
3210                 return false;
3211 
3212         /*
3213          * Since we truncate DL_SCALE bits, make sure we're at least
3214          * that big.
3215          */
3216         if (attr->sched_runtime < (1ULL << DL_SCALE))
3217                 return false;
3218 
3219         /*
3220          * Since we use the MSB for wrap-around and sign issues, make
3221          * sure it's not set (mind that period can be equal to zero).
3222          */
3223         if (attr->sched_deadline & (1ULL << 63) ||
3224             attr->sched_period & (1ULL << 63))
3225                 return false;
3226 
3227         /* runtime <= deadline <= period (if period != 0) */
3228         if ((attr->sched_period != 0 &&
3229              attr->sched_period < attr->sched_deadline) ||
3230             attr->sched_deadline < attr->sched_runtime)
3231                 return false;
3232 
3233         return true;
3234 }
3235 
3236 /*
3237  * check the target process has a UID that matches the current process's
3238  */
3239 static bool check_same_owner(struct task_struct *p)
3240 {
3241         const struct cred *cred = current_cred(), *pcred;
3242         bool match;
3243 
3244         rcu_read_lock();
3245         pcred = __task_cred(p);
3246         match = (uid_eq(cred->euid, pcred->euid) ||
3247                  uid_eq(cred->euid, pcred->uid));
3248         rcu_read_unlock();
3249         return match;
3250 }
3251 
3252 static int __sched_setscheduler(struct task_struct *p,
3253                                 const struct sched_attr *attr,
3254                                 bool user)
3255 {
3256         int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3257                       MAX_RT_PRIO - 1 - attr->sched_priority;
3258         int retval, oldprio, oldpolicy = -1, on_rq, running;
3259         int policy = attr->sched_policy;
3260         unsigned long flags;
3261         const struct sched_class *prev_class;
3262         struct rq *rq;
3263         int reset_on_fork;
3264 
3265         /* may grab non-irq protected spin_locks */
3266         BUG_ON(in_interrupt());
3267 recheck:
3268         /* double check policy once rq lock held */
3269         if (policy < 0) {
3270                 reset_on_fork = p->sched_reset_on_fork;
3271                 policy = oldpolicy = p->policy;
3272         } else {
3273                 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3274 
3275                 if (policy != SCHED_DEADLINE &&
3276                                 policy != SCHED_FIFO && policy != SCHED_RR &&
3277                                 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3278                                 policy != SCHED_IDLE)
3279                         return -EINVAL;
3280         }
3281 
3282         if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3283                 return -EINVAL;
3284 
3285         /*
3286          * Valid priorities for SCHED_FIFO and SCHED_RR are
3287          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3288          * SCHED_BATCH and SCHED_IDLE is 0.
3289          */
3290         if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3291             (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3292                 return -EINVAL;
3293         if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3294             (rt_policy(policy) != (attr->sched_priority != 0)))
3295                 return -EINVAL;
3296 
3297         /*
3298          * Allow unprivileged RT tasks to decrease priority:
3299          */
3300         if (user && !capable(CAP_SYS_NICE)) {
3301                 if (fair_policy(policy)) {
3302                         if (attr->sched_nice < task_nice(p) &&
3303                             !can_nice(p, attr->sched_nice))
3304                                 return -EPERM;
3305                 }
3306 
3307                 if (rt_policy(policy)) {
3308                         unsigned long rlim_rtprio =
3309                                         task_rlimit(p, RLIMIT_RTPRIO);
3310 
3311                         /* can't set/change the rt policy */
3312                         if (policy != p->policy && !rlim_rtprio)
3313                                 return -EPERM;
3314 
3315                         /* can't increase priority */
3316                         if (attr->sched_priority > p->rt_priority &&
3317                             attr->sched_priority > rlim_rtprio)
3318                                 return -EPERM;
3319                 }
3320 
3321                  /*
3322                   * Can't set/change SCHED_DEADLINE policy at all for now
3323                   * (safest behavior); in the future we would like to allow
3324                   * unprivileged DL tasks to increase their relative deadline
3325                   * or reduce their runtime (both ways reducing utilization)
3326                   */
3327                 if (dl_policy(policy))
3328                         return -EPERM;
3329 
3330                 /*
3331                  * Treat SCHED_IDLE as nice 20. Only allow a switch to
3332                  * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3333                  */
3334                 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3335                         if (!can_nice(p, task_nice(p)))
3336                                 return -EPERM;
3337                 }
3338 
3339                 /* can't change other user's priorities */
3340                 if (!check_same_owner(p))
3341                         return -EPERM;
3342 
3343                 /* Normal users shall not reset the sched_reset_on_fork flag */
3344                 if (p->sched_reset_on_fork && !reset_on_fork)
3345                         return -EPERM;
3346         }
3347 
3348         if (user) {
3349                 retval = security_task_setscheduler(p);
3350                 if (retval)
3351                         return retval;
3352         }
3353 
3354         /*
3355          * make sure no PI-waiters arrive (or leave) while we are
3356          * changing the priority of the task:
3357          *
3358          * To be able to change p->policy safely, the appropriate
3359          * runqueue lock must be held.
3360          */
3361         rq = task_rq_lock(p, &flags);
3362 
3363         /*
3364          * Changing the policy of the stop threads its a very bad idea
3365          */
3366         if (p == rq->stop) {
3367                 task_rq_unlock(rq, p, &flags);
3368                 return -EINVAL;
3369         }
3370 
3371         /*
3372          * If not changing anything there's no need to proceed further,
3373          * but store a possible modification of reset_on_fork.
3374          */
3375         if (unlikely(policy == p->policy)) {
3376                 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3377                         goto change;
3378                 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3379                         goto change;
3380                 if (dl_policy(policy))
3381                         goto change;
3382 
3383                 p->sched_reset_on_fork = reset_on_fork;
3384                 task_rq_unlock(rq, p, &flags);
3385                 return 0;
3386         }
3387 change:
3388 
3389         if (user) {
3390 #ifdef CONFIG_RT_GROUP_SCHED
3391                 /*
3392                  * Do not allow realtime tasks into groups that have no runtime
3393                  * assigned.
3394                  */
3395                 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3396                                 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3397                                 !task_group_is_autogroup(task_group(p))) {
3398                         task_rq_unlock(rq, p, &flags);
3399                         return -EPERM;
3400                 }
3401 #endif
3402 #ifdef CONFIG_SMP
3403                 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3404                         cpumask_t *span = rq->rd->span;
3405 
3406                         /*
3407                          * Don't allow tasks with an affinity mask smaller than
3408                          * the entire root_domain to become SCHED_DEADLINE. We
3409                          * will also fail if there's no bandwidth available.
3410                          */
3411                         if (!cpumask_subset(span, &p->cpus_allowed) ||
3412                             rq->rd->dl_bw.bw == 0) {
3413                                 task_rq_unlock(rq, p, &flags);
3414                                 return -EPERM;
3415                         }
3416                 }
3417 #endif
3418         }
3419 
3420         /* recheck policy now with rq lock held */
3421         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3422                 policy = oldpolicy = -1;
3423                 task_rq_unlock(rq, p, &flags);
3424                 goto recheck;
3425         }
3426 
3427         /*
3428          * If setscheduling to SCHED_DEADLINE (or changing the parameters
3429          * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3430          * is available.
3431          */
3432         if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3433                 task_rq_unlock(rq, p, &flags);
3434                 return -EBUSY;
3435         }
3436 
3437         p->sched_reset_on_fork = reset_on_fork;
3438         oldprio = p->prio;
3439 
3440         /*
3441          * Special case for priority boosted tasks.
3442          *
3443          * If the new priority is lower or equal (user space view)
3444          * than the current (boosted) priority, we just store the new
3445          * normal parameters and do not touch the scheduler class and
3446          * the runqueue. This will be done when the task deboost
3447          * itself.
3448          */
3449         if (rt_mutex_check_prio(p, newprio)) {
3450                 __setscheduler_params(p, attr);
3451                 task_rq_unlock(rq, p, &flags);
3452                 return 0;
3453         }
3454 
3455         on_rq = p->on_rq;
3456         running = task_current(rq, p);
3457         if (on_rq)
3458                 dequeue_task(rq, p, 0);
3459         if (running)
3460                 p->sched_class->put_prev_task(rq, p);
3461 
3462         prev_class = p->sched_class;
3463         __setscheduler(rq, p, attr);
3464 
3465         if (running)
3466                 p->sched_class->set_curr_task(rq);
3467         if (on_rq) {
3468                 /*
3469                  * We enqueue to tail when the priority of a task is
3470                  * increased (user space view).
3471                  */
3472                 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3473         }
3474 
3475         check_class_changed(rq, p, prev_class, oldprio);
3476         task_rq_unlock(rq, p, &flags);
3477 
3478         rt_mutex_adjust_pi(p);
3479 
3480         return 0;
3481 }
3482 
3483 static int _sched_setscheduler(struct task_struct *p, int policy,
3484                                const struct sched_param *param, bool check)
3485 {
3486         struct sched_attr attr = {
3487                 .sched_policy   = policy,
3488                 .sched_priority = param->sched_priority,
3489                 .sched_nice     = PRIO_TO_NICE(p->static_prio),
3490         };
3491 
3492         /*
3493          * Fixup the legacy SCHED_RESET_ON_FORK hack
3494          */
3495         if (policy & SCHED_RESET_ON_FORK) {
3496                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3497                 policy &= ~SCHED_RESET_ON_FORK;
3498                 attr.sched_policy = policy;
3499         }
3500 
3501         return __sched_setscheduler(p, &attr, check);
3502 }
3503 /**
3504  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3505  * @p: the task in question.
3506  * @policy: new policy.
3507  * @param: structure containing the new RT priority.
3508  *
3509  * Return: 0 on success. An error code otherwise.
3510  *
3511  * NOTE that the task may be already dead.
3512  */
3513 int sched_setscheduler(struct task_struct *p, int policy,
3514                        const struct sched_param *param)
3515 {
3516         return _sched_setscheduler(p, policy, param, true);
3517 }
3518 EXPORT_SYMBOL_GPL(sched_setscheduler);
3519 
3520 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3521 {
3522         return __sched_setscheduler(p, attr, true);
3523 }
3524 EXPORT_SYMBOL_GPL(sched_setattr);
3525 
3526 /**
3527  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3528  * @p: the task in question.
3529  * @policy: new policy.
3530  * @param: structure containing the new RT priority.
3531  *
3532  * Just like sched_setscheduler, only don't bother checking if the
3533  * current context has permission.  For example, this is needed in
3534  * stop_machine(): we create temporary high priority worker threads,
3535  * but our caller might not have that capability.
3536  *
3537  * Return: 0 on success. An error code otherwise.
3538  */
3539 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3540                                const struct sched_param *param)
3541 {
3542         return _sched_setscheduler(p, policy, param, false);
3543 }
3544 
3545 static int
3546 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3547 {
3548         struct sched_param lparam;
3549         struct task_struct *p;
3550         int retval;
3551 
3552         if (!param || pid < 0)
3553                 return -EINVAL;
3554         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3555                 return -EFAULT;
3556 
3557         rcu_read_lock();
3558         retval = -ESRCH;
3559         p = find_process_by_pid(pid);
3560         if (p != NULL)
3561                 retval = sched_setscheduler(p, policy, &lparam);
3562         rcu_read_unlock();
3563 
3564         return retval;
3565 }
3566 
3567 /*
3568  * Mimics kernel/events/core.c perf_copy_attr().
3569  */
3570 static int sched_copy_attr(struct sched_attr __user *uattr,
3571                            struct sched_attr *attr)
3572 {
3573         u32 size;
3574         int ret;
3575 
3576         if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3577                 return -EFAULT;
3578 
3579         /*
3580          * zero the full structure, so that a short copy will be nice.
3581          */
3582         memset(attr, 0, sizeof(*attr));
3583 
3584         ret = get_user(size, &uattr->size);
3585         if (ret)
3586                 return ret;
3587 
3588         if (size > PAGE_SIZE)   /* silly large */
3589                 goto err_size;
3590 
3591         if (!size)              /* abi compat */
3592                 size = SCHED_ATTR_SIZE_VER0;
3593 
3594         if (size < SCHED_ATTR_SIZE_VER0)
3595                 goto err_size;
3596 
3597         /*
3598          * If we're handed a bigger struct than we know of,
3599          * ensure all the unknown bits are 0 - i.e. new
3600          * user-space does not rely on any kernel feature
3601          * extensions we dont know about yet.
3602          */
3603         if (size > sizeof(*attr)) {
3604                 unsigned char __user *addr;
3605                 unsigned char __user *end;
3606                 unsigned char val;
3607 
3608                 addr = (void __user *)uattr + sizeof(*attr);
3609                 end  = (void __user *)uattr + size;
3610 
3611                 for (; addr < end; addr++) {
3612                         ret = get_user(val, addr);
3613                         if (ret)
3614                                 return ret;
3615                         if (val)
3616                                 goto err_size;
3617                 }
3618                 size = sizeof(*attr);
3619         }
3620 
3621         ret = copy_from_user(attr, uattr, size);
3622         if (ret)
3623                 return -EFAULT;
3624 
3625         /*
3626          * XXX: do we want to be lenient like existing syscalls; or do we want
3627          * to be strict and return an error on out-of-bounds values?
3628          */
3629         attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3630 
3631 out:
3632         return ret;
3633 
3634 err_size:
3635         put_user(sizeof(*attr), &uattr->size);
3636         ret = -E2BIG;
3637         goto out;
3638 }
3639 
3640 /**
3641  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3642  * @pid: the pid in question.
3643  * @policy: new policy.
3644  * @param: structure containing the new RT priority.
3645  *
3646  * Return: 0 on success. An error code otherwise.
3647  */
3648 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3649                 struct sched_param __user *, param)
3650 {
3651         /* negative values for policy are not valid */
3652         if (policy < 0)
3653                 return -EINVAL;
3654 
3655         return do_sched_setscheduler(pid, policy, param);
3656 }
3657 
3658 /**
3659  * sys_sched_setparam - set/change the RT priority of a thread
3660  * @pid: the pid in question.
3661  * @param: structure containing the new RT priority.
3662  *
3663  * Return: 0 on success. An error code otherwise.
3664  */
3665 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3666 {
3667         return do_sched_setscheduler(pid, -1, param);
3668 }
3669 
3670 /**
3671  * sys_sched_setattr - same as above, but with extended sched_attr
3672  * @pid: the pid in question.
3673  * @uattr: structure containing the extended parameters.
3674  * @flags: for future extension.
3675  */
3676 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3677                                unsigned int, flags)
3678 {
3679         struct sched_attr attr;
3680         struct task_struct *p;
3681         int retval;
3682 
3683         if (!uattr || pid < 0 || flags)
3684                 return -EINVAL;
3685 
3686         retval = sched_copy_attr(uattr, &attr);
3687         if (retval)
3688                 return retval;
3689 
3690         if ((int)attr.sched_policy < 0)
3691                 return -EINVAL;
3692 
3693         rcu_read_lock();
3694         retval = -ESRCH;
3695         p = find_process_by_pid(pid);
3696         if (p != NULL)
3697                 retval = sched_setattr(p, &attr);
3698         rcu_read_unlock();
3699 
3700         return retval;
3701 }
3702 
3703 /**
3704  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3705  * @pid: the pid in question.
3706  *
3707  * Return: On success, the policy of the thread. Otherwise, a negative error
3708  * code.
3709  */
3710 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3711 {
3712         struct task_struct *p;
3713         int retval;
3714 
3715         if (pid < 0)
3716                 return -EINVAL;
3717 
3718         retval = -ESRCH;
3719         rcu_read_lock();
3720         p = find_process_by_pid(pid);
3721         if (p) {
3722                 retval = security_task_getscheduler(p);
3723                 if (!retval)
3724                         retval = p->policy
3725                                 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3726         }
3727         rcu_read_unlock();
3728         return retval;
3729 }
3730 
3731 /**
3732  * sys_sched_getparam - get the RT priority of a thread
3733  * @pid: the pid in question.
3734  * @param: structure containing the RT priority.
3735  *
3736  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3737  * code.
3738  */
3739 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3740 {
3741         struct sched_param lp = { .sched_priority = 0 };
3742         struct task_struct *p;
3743         int retval;
3744 
3745         if (!param || pid < 0)
3746                 return -EINVAL;
3747 
3748         rcu_read_lock();
3749         p = find_process_by_pid(pid);
3750         retval = -ESRCH;
3751         if (!p)
3752                 goto out_unlock;
3753 
3754         retval = security_task_getscheduler(p);
3755         if (retval)
3756                 goto out_unlock;
3757 
3758         if (task_has_rt_policy(p))
3759                 lp.sched_priority = p->rt_priority;
3760         rcu_read_unlock();
3761 
3762         /*
3763          * This one might sleep, we cannot do it with a spinlock held ...
3764          */
3765         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3766 
3767         return retval;
3768 
3769 out_unlock:
3770         rcu_read_unlock();
3771         return retval;
3772 }
3773 
3774 static int sched_read_attr(struct sched_attr __user *uattr,
3775                            struct sched_attr *attr,
3776                            unsigned int usize)
3777 {
3778         int ret;
3779 
3780         if (!access_ok(VERIFY_WRITE, uattr, usize))
3781                 return -EFAULT;
3782 
3783         /*
3784          * If we're handed a smaller struct than we know of,
3785          * ensure all the unknown bits are 0 - i.e. old
3786          * user-space does not get uncomplete information.
3787          */
3788         if (usize < sizeof(*attr)) {
3789                 unsigned char *addr;
3790                 unsigned char *end;
3791 
3792                 addr = (void *)attr + usize;
3793                 end  = (void *)attr + sizeof(*attr);
3794 
3795                 for (; addr < end; addr++) {
3796                         if (*addr)
3797                                 goto err_size;
3798                 }
3799 
3800                 attr->size = usize;
3801         }
3802 
3803         ret = copy_to_user(uattr, attr, attr->size);
3804         if (ret)
3805                 return -EFAULT;
3806 
3807 out:
3808         return ret;
3809 
3810 err_size:
3811         ret = -E2BIG;
3812         goto out;
3813 }
3814 
3815 /**
3816  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3817  * @pid: the pid in question.
3818  * @uattr: structure containing the extended parameters.
3819  * @size: sizeof(attr) for fwd/bwd comp.
3820  * @flags: for future extension.
3821  */
3822 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3823                 unsigned int, size, unsigned int, flags)
3824 {
3825         struct sched_attr attr = {
3826                 .size = sizeof(struct sched_attr),
3827         };
3828         struct task_struct *p;
3829         int retval;
3830 
3831         if (!uattr || pid < 0 || size > PAGE_SIZE ||
3832             size < SCHED_ATTR_SIZE_VER0 || flags)
3833                 return -EINVAL;
3834 
3835         rcu_read_lock();
3836         p = find_process_by_pid(pid);
3837         retval = -ESRCH;
3838         if (!p)
3839                 goto out_unlock;
3840 
3841         retval = security_task_getscheduler(p);
3842         if (retval)
3843                 goto out_unlock;
3844 
3845         attr.sched_policy = p->policy;
3846         if (p->sched_reset_on_fork)
3847                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3848         if (task_has_dl_policy(p))
3849                 __getparam_dl(p, &attr);
3850         else if (task_has_rt_policy(p))
3851                 attr.sched_priority = p->rt_priority;
3852         else
3853                 attr.sched_nice = task_nice(p);
3854 
3855         rcu_read_unlock();
3856 
3857         retval = sched_read_attr(uattr, &attr, size);
3858         return retval;
3859 
3860 out_unlock:
3861         rcu_read_unlock();
3862         return retval;
3863 }
3864 
3865 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3866 {
3867         cpumask_var_t cpus_allowed, new_mask;
3868         struct task_struct *p;
3869         int retval;
3870 
3871         rcu_read_lock();
3872 
3873         p = find_process_by_pid(pid);
3874         if (!p) {
3875                 rcu_read_unlock();
3876                 return -ESRCH;
3877         }
3878 
3879         /* Prevent p going away */
3880         get_task_struct(p);
3881         rcu_read_unlock();
3882 
3883         if (p->flags & PF_NO_SETAFFINITY) {
3884                 retval = -EINVAL;
3885                 goto out_put_task;
3886         }
3887         if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3888                 retval = -ENOMEM;
3889                 goto out_put_task;
3890         }
3891         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3892                 retval = -ENOMEM;
3893                 goto out_free_cpus_allowed;
3894         }
3895         retval = -EPERM;
3896         if (!check_same_owner(p)) {
3897                 rcu_read_lock();
3898                 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3899                         rcu_read_unlock();
3900                         goto out_unlock;
3901                 }
3902                 rcu_read_unlock();
3903         }
3904 
3905         retval = security_task_setscheduler(p);
3906         if (retval)
3907                 goto out_unlock;
3908 
3909 
3910         cpuset_cpus_allowed(p, cpus_allowed);
3911         cpumask_and(new_mask, in_mask, cpus_allowed);
3912 
3913         /*
3914          * Since bandwidth control happens on root_domain basis,
3915          * if admission test is enabled, we only admit -deadline
3916          * tasks allowed to run on all the CPUs in the task's
3917          * root_domain.
3918          */
3919 #ifdef CONFIG_SMP
3920         if (task_has_dl_policy(p)) {
3921                 const struct cpumask *span = task_rq(p)->rd->span;
3922 
3923                 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3924                         retval = -EBUSY;
3925                         goto out_unlock;
3926                 }
3927         }
3928 #endif
3929 again:
3930         retval = set_cpus_allowed_ptr(p, new_mask);
3931 
3932         if (!retval) {
3933                 cpuset_cpus_allowed(p, cpus_allowed);
3934                 if (!cpumask_subset(new_mask, cpus_allowed)) {
3935                         /*
3936                          * We must have raced with a concurrent cpuset
3937                          * update. Just reset the cpus_allowed to the
3938                          * cpuset's cpus_allowed
3939                          */
3940                         cpumask_copy(new_mask, cpus_allowed);
3941                         goto again;
3942                 }
3943         }
3944 out_unlock:
3945         free_cpumask_var(new_mask);
3946 out_free_cpus_allowed:
3947         free_cpumask_var(cpus_allowed);
3948 out_put_task:
3949         put_task_struct(p);
3950         return retval;
3951 }
3952 
3953 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3954                              struct cpumask *new_mask)
3955 {
3956         if (len < cpumask_size())
3957                 cpumask_clear(new_mask);
3958         else if (len > cpumask_size())
3959                 len = cpumask_size();
3960 
3961         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3962 }
3963 
3964 /**
3965  * sys_sched_setaffinity - set the cpu affinity of a process
3966  * @pid: pid of the process
3967  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3968  * @user_mask_ptr: user-space pointer to the new cpu mask
3969  *
3970  * Return: 0 on success. An error code otherwise.
3971  */
3972 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3973                 unsigned long __user *, user_mask_ptr)
3974 {
3975         cpumask_var_t new_mask;
3976         int retval;
3977 
3978         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3979                 return -ENOMEM;
3980 
3981         retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3982         if (retval == 0)
3983                 retval = sched_setaffinity(pid, new_mask);
3984         free_cpumask_var(new_mask);
3985         return retval;
3986 }
3987 
3988 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3989 {
3990         struct task_struct *p;
3991         unsigned long flags;
3992         int retval;
3993 
3994         rcu_read_lock();
3995 
3996         retval = -ESRCH;
3997         p = find_process_by_pid(pid);
3998         if (!p)
3999                 goto out_unlock;
4000 
4001         retval = security_task_getscheduler(p);
4002         if (retval)
4003                 goto out_unlock;
4004 
4005         raw_spin_lock_irqsave(&p->pi_lock, flags);
4006         cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4007         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4008 
4009 out_unlock:
4010         rcu_read_unlock();
4011 
4012         return retval;
4013 }
4014 
4015 /**
4016  * sys_sched_getaffinity - get the cpu affinity of a process
4017  * @pid: pid of the process
4018  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4019  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4020  *
4021  * Return: 0 on success. An error code otherwise.
4022  */
4023 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4024                 unsigned long __user *, user_mask_ptr)
4025 {
4026         int ret;
4027         cpumask_var_t mask;
4028 
4029         if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4030                 return -EINVAL;
4031         if (len & (sizeof(unsigned long)-1))
4032                 return -EINVAL;
4033 
4034         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4035                 return -ENOMEM;
4036 
4037         ret = sched_getaffinity(pid, mask);
4038         if (ret == 0) {
4039                 size_t retlen = min_t(size_t, len, cpumask_size());
4040 
4041                 if (copy_to_user(user_mask_ptr, mask, retlen))
4042                         ret = -EFAULT;
4043                 else
4044                         ret = retlen;
4045         }
4046         free_cpumask_var(mask);
4047 
4048         return ret;
4049 }
4050 
4051 /**
4052  * sys_sched_yield - yield the current processor to other threads.
4053  *
4054  * This function yields the current CPU to other tasks. If there are no
4055  * other threads running on this CPU then this function will return.
4056  *
4057  * Return: 0.
4058  */
4059 SYSCALL_DEFINE0(sched_yield)
4060 {
4061         struct rq *rq = this_rq_lock();
4062 
4063         schedstat_inc(rq, yld_count);
4064         current->sched_class->yield_task(rq);
4065 
4066         /*
4067          * Since we are going to call schedule() anyway, there's
4068          * no need to preempt or enable interrupts:
4069          */
4070         __release(rq->lock);
4071         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4072         do_raw_spin_unlock(&rq->lock);
4073         sched_preempt_enable_no_resched();
4074 
4075         schedule();
4076 
4077         return 0;
4078 }
4079 
4080 static void __cond_resched(void)
4081 {
4082         __preempt_count_add(PREEMPT_ACTIVE);
4083         __schedule();
4084         __preempt_count_sub(PREEMPT_ACTIVE);
4085 }
4086 
4087 int __sched _cond_resched(void)
4088 {
4089         if (should_resched()) {
4090                 __cond_resched();
4091                 return 1;
4092         }
4093         return 0;
4094 }
4095 EXPORT_SYMBOL(_cond_resched);
4096 
4097 /*
4098  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4099  * call schedule, and on return reacquire the lock.
4100  *
4101  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4102  * operations here to prevent schedule() from being called twice (once via
4103  * spin_unlock(), once by hand).
4104  */
4105 int __cond_resched_lock(spinlock_t *lock)
4106 {
4107         int resched = should_resched();
4108         int ret = 0;
4109 
4110         lockdep_assert_held(lock);
4111 
4112         if (spin_needbreak(lock) || resched) {
4113                 spin_unlock(lock);
4114                 if (resched)
4115                         __cond_resched();
4116                 else
4117                         cpu_relax();
4118                 ret = 1;
4119                 spin_lock(lock);
4120         }
4121         return ret;
4122 }
4123 EXPORT_SYMBOL(__cond_resched_lock);
4124 
4125 int __sched __cond_resched_softirq(void)
4126 {
4127         BUG_ON(!in_softirq());
4128 
4129         if (should_resched()) {
4130                 local_bh_enable();
4131                 __cond_resched();
4132                 local_bh_disable();
4133                 return 1;
4134         }
4135         return 0;
4136 }
4137 EXPORT_SYMBOL(__cond_resched_softirq);
4138 
4139 /**
4140  * yield - yield the current processor to other threads.
4141  *
4142  * Do not ever use this function, there's a 99% chance you're doing it wrong.
4143  *
4144  * The scheduler is at all times free to pick the calling task as the most
4145  * eligible task to run, if removing the yield() call from your code breaks
4146  * it, its already broken.
4147  *
4148  * Typical broken usage is:
4149  *
4150  * while (!event)
4151  *      yield();
4152  *
4153  * where one assumes that yield() will let 'the other' process run that will
4154  * make event true. If the current task is a SCHED_FIFO task that will never
4155  * happen. Never use yield() as a progress guarantee!!
4156  *
4157  * If you want to use yield() to wait for something, use wait_event().
4158  * If you want to use yield() to be 'nice' for others, use cond_resched().
4159  * If you still want to use yield(), do not!
4160  */
4161 void __sched yield(void)
4162 {
4163         set_current_state(TASK_RUNNING);
4164         sys_sched_yield();
4165 }
4166 EXPORT_SYMBOL(yield);
4167 
4168 /**
4169  * yield_to - yield the current processor to another thread in
4170  * your thread group, or accelerate that thread toward the
4171  * processor it's on.
4172  * @p: target task
4173  * @preempt: whether task preemption is allowed or not
4174  *
4175  * It's the caller's job to ensure that the target task struct
4176  * can't go away on us before we can do any checks.
4177  *
4178  * Return:
4179  *      true (>0) if we indeed boosted the target task.
4180  *      false (0) if we failed to boost the target.
4181  *      -ESRCH if there's no task to yield to.
4182  */
4183 bool __sched yield_to(struct task_struct *p, bool preempt)
4184 {
4185         struct task_struct *curr = current;
4186         struct rq *rq, *p_rq;
4187         unsigned long flags;
4188         int yielded = 0;
4189 
4190         local_irq_save(flags);
4191         rq = this_rq();
4192 
4193 again:
4194         p_rq = task_rq(p);
4195         /*
4196          * If we're the only runnable task on the rq and target rq also
4197          * has only one task, there's absolutely no point in yielding.
4198          */
4199         if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4200                 yielded = -ESRCH;
4201                 goto out_irq;
4202         }
4203 
4204         double_rq_lock(rq, p_rq);
4205         if (task_rq(p) != p_rq) {
4206                 double_rq_unlock(rq, p_rq);
4207                 goto again;
4208         }
4209 
4210         if (!curr->sched_class->yield_to_task)
4211                 goto out_unlock;
4212 
4213         if (curr->sched_class != p->sched_class)
4214                 goto out_unlock;
4215 
4216         if (task_running(p_rq, p) || p->state)
4217                 goto out_unlock;
4218 
4219         yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4220         if (yielded) {
4221                 schedstat_inc(rq, yld_count);
4222                 /*
4223                  * Make p's CPU reschedule; pick_next_entity takes care of
4224                  * fairness.
4225                  */
4226                 if (preempt && rq != p_rq)
4227                         resched_task(p_rq->curr);
4228         }
4229 
4230 out_unlock:
4231         double_rq_unlock(rq, p_rq);
4232 out_irq:
4233         local_irq_restore(flags);
4234 
4235         if (yielded > 0)
4236                 schedule();
4237 
4238         return yielded;
4239 }
4240 EXPORT_SYMBOL_GPL(yield_to);
4241 
4242 /*
4243  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4244  * that process accounting knows that this is a task in IO wait state.
4245  */
4246 void __sched io_schedule(void)
4247 {
4248         struct rq *rq = raw_rq();
4249 
4250         delayacct_blkio_start();
4251         atomic_inc(&rq->nr_iowait);
4252         blk_flush_plug(current);
4253         current->in_iowait = 1;
4254         schedule();
4255         current->in_iowait = 0;
4256         atomic_dec(&rq->nr_iowait);
4257         delayacct_blkio_end();
4258 }
4259 EXPORT_SYMBOL(io_schedule);
4260 
4261 long __sched io_schedule_timeout(long timeout)
4262 {
4263         struct rq *rq = raw_rq();
4264         long ret;
4265 
4266         delayacct_blkio_start();
4267         atomic_inc(&rq->nr_iowait);
4268         blk_flush_plug(current);
4269         current->in_iowait = 1;
4270         ret = schedule_timeout(timeout);
4271         current->in_iowait = 0;
4272         atomic_dec(&rq->nr_iowait);
4273         delayacct_blkio_end();
4274         return ret;
4275 }
4276 
4277 /**
4278  * sys_sched_get_priority_max - return maximum RT priority.
4279  * @policy: scheduling class.
4280  *
4281  * Return: On success, this syscall returns the maximum
4282  * rt_priority that can be used by a given scheduling class.
4283  * On failure, a negative error code is returned.
4284  */
4285 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4286 {
4287         int ret = -EINVAL;
4288 
4289         switch (policy) {
4290         case SCHED_FIFO:
4291         case SCHED_RR:
4292                 ret = MAX_USER_RT_PRIO-1;
4293                 break;
4294         case SCHED_DEADLINE:
4295         case SCHED_NORMAL:
4296         case SCHED_BATCH:
4297         case SCHED_IDLE:
4298                 ret = 0;
4299                 break;
4300         }
4301         return ret;
4302 }
4303 
4304 /**
4305  * sys_sched_get_priority_min - return minimum RT priority.
4306  * @policy: scheduling class.
4307  *
4308  * Return: On success, this syscall returns the minimum
4309  * rt_priority that can be used by a given scheduling class.
4310  * On failure, a negative error code is returned.
4311  */
4312 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4313 {
4314         int ret = -EINVAL;
4315 
4316         switch (policy) {
4317         case SCHED_FIFO:
4318         case SCHED_RR:
4319                 ret = 1;
4320                 break;
4321         case SCHED_DEADLINE:
4322         case SCHED_NORMAL:
4323         case SCHED_BATCH:
4324         case SCHED_IDLE:
4325                 ret = 0;
4326         }
4327         return ret;
4328 }
4329 
4330 /**
4331  * sys_sched_rr_get_interval - return the default timeslice of a process.
4332  * @pid: pid of the process.
4333  * @interval: userspace pointer to the timeslice value.
4334  *
4335  * this syscall writes the default timeslice value of a given process
4336  * into the user-space timespec buffer. A value of '' means infinity.
4337  *
4338  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4339  * an error code.
4340  */
4341 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4342                 struct timespec __user *, interval)
4343 {
4344         struct task_struct *p;
4345         unsigned int time_slice;
4346         unsigned long flags;
4347         struct rq *rq;
4348         int retval;
4349         struct timespec t;
4350 
4351         if (pid < 0)
4352                 return -EINVAL;
4353 
4354         retval = -ESRCH;
4355         rcu_read_lock();
4356         p = find_process_by_pid(pid);
4357         if (!p)
4358                 goto out_unlock;
4359 
4360         retval = security_task_getscheduler(p);
4361         if (retval)
4362                 goto out_unlock;
4363 
4364         rq = task_rq_lock(p, &flags);
4365         time_slice = 0;
4366         if (p->sched_class->get_rr_interval)
4367                 time_slice = p->sched_class->get_rr_interval(rq, p);
4368         task_rq_unlock(rq, p, &flags);
4369 
4370         rcu_read_unlock();
4371         jiffies_to_timespec(time_slice, &t);
4372         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4373         return retval;
4374 
4375 out_unlock:
4376         rcu_read_unlock();
4377         return retval;
4378 }
4379 
4380 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4381 
4382 void sched_show_task(struct task_struct *p)
4383 {
4384         unsigned long free = 0;
4385         int ppid;
4386         unsigned state;
4387 
4388         state = p->state ? __ffs(p->state) + 1 : 0;
4389         printk(KERN_INFO "%-15.15s %c", p->comm,
4390                 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4391 #if BITS_PER_LONG == 32
4392         if (state == TASK_RUNNING)
4393                 printk(KERN_CONT " running  ");
4394         else
4395                 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4396 #else
4397         if (state == TASK_RUNNING)
4398                 printk(KERN_CONT "  running task    ");
4399         else
4400                 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4401 #endif
4402 #ifdef CONFIG_DEBUG_STACK_USAGE
4403         free = stack_not_used(p);
4404 #endif
4405         rcu_read_lock();
4406         ppid = task_pid_nr(rcu_dereference(p->real_parent));
4407         rcu_read_unlock();
4408         printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4409                 task_pid_nr(p), ppid,
4410                 (unsigned long)task_thread_info(p)->flags);
4411 
4412         print_worker_info(KERN_INFO, p);
4413         show_stack(p, NULL);
4414 }
4415 
4416 void show_state_filter(unsigned long state_filter)
4417 {
4418         struct task_struct *g, *p;
4419 
4420 #if BITS_PER_LONG == 32
4421         printk(KERN_INFO
4422                 "  task                PC stack   pid father\n");
4423 #else
4424         printk(KERN_INFO
4425                 "  task                        PC stack   pid father\n");
4426 #endif
4427         rcu_read_lock();
4428         do_each_thread(g, p) {
4429                 /*
4430                  * reset the NMI-timeout, listing all files on a slow
4431                  * console might take a lot of time:
4432                  */
4433                 touch_nmi_watchdog();
4434                 if (!state_filter || (p->state & state_filter))
4435                         sched_show_task(p);
4436         } while_each_thread(g, p);
4437 
4438         touch_all_softlockup_watchdogs();
4439 
4440 #ifdef CONFIG_SCHED_DEBUG
4441         sysrq_sched_debug_show();
4442 #endif
4443         rcu_read_unlock();
4444         /*
4445          * Only show locks if all tasks are dumped:
4446          */
4447         if (!state_filter)
4448                 debug_show_all_locks();
4449 }
4450 
4451 void init_idle_bootup_task(struct task_struct *idle)
4452 {
4453         idle->sched_class = &idle_sched_class;
4454 }
4455 
4456 /**
4457  * init_idle - set up an idle thread for a given CPU
4458  * @idle: task in question
4459  * @cpu: cpu the idle task belongs to
4460  *
4461  * NOTE: this function does not set the idle thread's NEED_RESCHED
4462  * flag, to make booting more robust.
4463  */
4464 void init_idle(struct task_struct *idle, int cpu)
4465 {
4466         struct rq *rq = cpu_rq(cpu);
4467         unsigned long flags;
4468 
4469         raw_spin_lock_irqsave(&rq->lock, flags);
4470 
4471         __sched_fork(0, idle);
4472         idle->state = TASK_RUNNING;
4473         idle->se.exec_start = sched_clock();
4474 
4475         do_set_cpus_allowed(idle, cpumask_of(cpu));
4476         /*
4477          * We're having a chicken and egg problem, even though we are
4478          * holding rq->lock, the cpu isn't yet set to this cpu so the
4479          * lockdep check in task_group() will fail.
4480          *
4481          * Similar case to sched_fork(). / Alternatively we could
4482          * use task_rq_lock() here and obtain the other rq->lock.
4483          *
4484          * Silence PROVE_RCU
4485          */
4486         rcu_read_lock();
4487         __set_task_cpu(idle, cpu);
4488         rcu_read_unlock();
4489 
4490         rq->curr = rq->idle = idle;
4491         idle->on_rq = 1;
4492 #if defined(CONFIG_SMP)
4493         idle->on_cpu = 1;
4494 #endif
4495         raw_spin_unlock_irqrestore(&rq->lock, flags);
4496 
4497         /* Set the preempt count _outside_ the spinlocks! */
4498         init_idle_preempt_count(idle, cpu);
4499 
4500         /*
4501          * The idle tasks have their own, simple scheduling class:
4502          */
4503         idle->sched_class = &idle_sched_class;
4504         ftrace_graph_init_idle_task(idle, cpu);
4505         vtime_init_idle(idle, cpu);
4506 #if defined(CONFIG_SMP)
4507         sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4508 #endif
4509 }
4510 
4511 #ifdef CONFIG_SMP
4512 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4513 {
4514         if (p->sched_class && p->sched_class->set_cpus_allowed)
4515                 p->sched_class->set_cpus_allowed(p, new_mask);
4516 
4517         cpumask_copy(&p->cpus_allowed, new_mask);
4518         p->nr_cpus_allowed = cpumask_weight(new_mask);
4519 }
4520 
4521 /*
4522  * This is how migration works:
4523  *
4524  * 1) we invoke migration_cpu_stop() on the target CPU using
4525  *    stop_one_cpu().
4526  * 2) stopper starts to run (implicitly forcing the migrated thread
4527  *    off the CPU)
4528  * 3) it checks whether the migrated task is still in the wrong runqueue.
4529  * 4) if it's in the wrong runqueue then the migration thread removes
4530  *    it and puts it into the right queue.
4531  * 5) stopper completes and stop_one_cpu() returns and the migration
4532  *    is done.
4533  */
4534 
4535 /*
4536  * Change a given task's CPU affinity. Migrate the thread to a
4537  * proper CPU and schedule it away if the CPU it's executing on
4538  * is removed from the allowed bitmask.
4539  *
4540  * NOTE: the caller must have a valid reference to the task, the
4541  * task must not exit() & deallocate itself prematurely. The
4542  * call is not atomic; no spinlocks may be held.
4543  */
4544 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4545 {
4546         unsigned long flags;
4547         struct rq *rq;
4548         unsigned int dest_cpu;
4549         int ret = 0;
4550 
4551         rq = task_rq_lock(p, &flags);
4552 
4553         if (cpumask_equal(&p->cpus_allowed, new_mask))
4554                 goto out;
4555 
4556         if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4557                 ret = -EINVAL;
4558                 goto out;
4559         }
4560 
4561         do_set_cpus_allowed(p, new_mask);
4562 
4563         /* Can the task run on the task's current CPU? If so, we're done */
4564         if (cpumask_test_cpu(task_cpu(p), new_mask))
4565                 goto out;
4566 
4567         dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4568         if (p->on_rq) {
4569                 struct migration_arg arg = { p, dest_cpu };
4570                 /* Need help from migration thread: drop lock and wait. */
4571                 task_rq_unlock(rq, p, &flags);
4572                 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4573                 tlb_migrate_finish(p->mm);
4574                 return 0;
4575         }
4576 out:
4577         task_rq_unlock(rq, p, &flags);
4578 
4579         return ret;
4580 }
4581 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4582 
4583 /*
4584  * Move (not current) task off this cpu, onto dest cpu. We're doing
4585  * this because either it can't run here any more (set_cpus_allowed()
4586  * away from this CPU, or CPU going down), or because we're
4587  * attempting to rebalance this task on exec (sched_exec).
4588  *
4589  * So we race with normal scheduler movements, but that's OK, as long
4590  * as the task is no longer on this CPU.
4591  *
4592  * Returns non-zero if task was successfully migrated.
4593  */
4594 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4595 {
4596         struct rq *rq_dest, *rq_src;
4597         int ret = 0;
4598 
4599         if (unlikely(!cpu_active(dest_cpu)))
4600                 return ret;
4601 
4602         rq_src = cpu_rq(src_cpu);
4603         rq_dest = cpu_rq(dest_cpu);
4604 
4605         raw_spin_lock(&p->pi_lock);
4606         double_rq_lock(rq_src, rq_dest);
4607         /* Already moved. */
4608         if (task_cpu(p) != src_cpu)
4609                 goto done;
4610         /* Affinity changed (again). */
4611         if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4612                 goto fail;
4613 
4614         /*
4615          * If we're not on a rq, the next wake-up will ensure we're
4616          * placed properly.
4617          */
4618         if (p->on_rq) {
4619                 dequeue_task(rq_src, p, 0);
4620                 set_task_cpu(p, dest_cpu);
4621                 enqueue_task(rq_dest, p, 0);
4622                 check_preempt_curr(rq_dest, p, 0);
4623         }
4624 done:
4625         ret = 1;
4626 fail:
4627         double_rq_unlock(rq_src, rq_dest);
4628         raw_spin_unlock(&p->pi_lock);
4629         return ret;
4630 }
4631 
4632 #ifdef CONFIG_NUMA_BALANCING
4633 /* Migrate current task p to target_cpu */
4634 int migrate_task_to(struct task_struct *p, int target_cpu)
4635 {
4636         struct migration_arg arg = { p, target_cpu };
4637         int curr_cpu = task_cpu(p);
4638 
4639         if (curr_cpu == target_cpu)
4640                 return 0;
4641 
4642         if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4643                 return -EINVAL;
4644 
4645         /* TODO: This is not properly updating schedstats */
4646 
4647         trace_sched_move_numa(p, curr_cpu, target_cpu);
4648         return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4649 }
4650 
4651 /*
4652  * Requeue a task on a given node and accurately track the number of NUMA
4653  * tasks on the runqueues
4654  */
4655 void sched_setnuma(struct task_struct *p, int nid)
4656 {
4657         struct rq *rq;
4658         unsigned long flags;
4659         bool on_rq, running;
4660 
4661         rq = task_rq_lock(p, &flags);
4662         on_rq = p->on_rq;
4663         running = task_current(rq, p);
4664 
4665         if (on_rq)
4666                 dequeue_task(rq, p, 0);
4667         if (running)
4668                 p->sched_class->put_prev_task(rq, p);
4669 
4670         p->numa_preferred_nid = nid;
4671 
4672         if (running)
4673                 p->sched_class->set_curr_task(rq);
4674         if (on_rq)
4675                 enqueue_task(rq, p, 0);
4676         task_rq_unlock(rq, p, &flags);
4677 }
4678 #endif
4679 
4680 /*
4681  * migration_cpu_stop - this will be executed by a highprio stopper thread
4682  * and performs thread migration by bumping thread off CPU then
4683  * 'pushing' onto another runqueue.
4684  */
4685 static int migration_cpu_stop(void *data)
4686 {
4687         struct migration_arg *arg = data;
4688 
4689         /*
4690          * The original target cpu might have gone down and we might
4691          * be on another cpu but it doesn't matter.
4692          */
4693         local_irq_disable();
4694         __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4695         local_irq_enable();
4696         return 0;
4697 }
4698 
4699 #ifdef CONFIG_HOTPLUG_CPU
4700 
4701 /*
4702  * Ensures that the idle task is using init_mm right before its cpu goes
4703  * offline.
4704  */
4705 void idle_task_exit(void)
4706 {
4707         struct mm_struct *mm = current->active_mm;
4708 
4709         BUG_ON(cpu_online(smp_processor_id()));
4710 
4711         if (mm != &init_mm) {
4712                 switch_mm(mm, &init_mm, current);
4713                 finish_arch_post_lock_switch();
4714         }
4715         mmdrop(mm);
4716 }
4717 
4718 /*
4719  * Since this CPU is going 'away' for a while, fold any nr_active delta
4720  * we might have. Assumes we're called after migrate_tasks() so that the
4721  * nr_active count is stable.
4722  *
4723  * Also see the comment "Global load-average calculations".
4724  */
4725 static void calc_load_migrate(struct rq *rq)
4726 {
4727         long delta = calc_load_fold_active(rq);
4728         if (delta)
4729                 atomic_long_add(delta, &calc_load_tasks);
4730 }
4731 
4732 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4733 {
4734 }
4735 
4736 static const struct sched_class fake_sched_class = {
4737         .put_prev_task = put_prev_task_fake,
4738 };
4739 
4740 static struct task_struct fake_task = {
4741         /*
4742          * Avoid pull_{rt,dl}_task()
4743          */
4744         .prio = MAX_PRIO + 1,
4745         .sched_class = &fake_sched_class,
4746 };
4747 
4748 /*
4749  * Migrate all tasks from the rq, sleeping tasks will be migrated by
4750  * try_to_wake_up()->select_task_rq().
4751  *
4752  * Called with rq->lock held even though we'er in stop_machine() and
4753  * there's no concurrency possible, we hold the required locks anyway
4754  * because of lock validation efforts.
4755  */
4756 static void migrate_tasks(unsigned int dead_cpu)
4757 {
4758         struct rq *rq = cpu_rq(dead_cpu);
4759         struct task_struct *next, *stop = rq->stop;
4760         int dest_cpu;
4761 
4762         /*
4763          * Fudge the rq selection such that the below task selection loop
4764          * doesn't get stuck on the currently eligible stop task.
4765          *
4766          * We're currently inside stop_machine() and the rq is either stuck
4767          * in the stop_machine_cpu_stop() loop, or we're executing this code,
4768          * either way we should never end up calling schedule() until we're
4769          * done here.
4770          */
4771         rq->stop = NULL;
4772 
4773         /*
4774          * put_prev_task() and pick_next_task() sched
4775          * class method both need to have an up-to-date
4776          * value of rq->clock[_task]
4777          */
4778         update_rq_clock(rq);
4779 
4780         for ( ; ; ) {
4781                 /*
4782                  * There's this thread running, bail when that's the only
4783                  * remaining thread.
4784                  */
4785                 if (rq->nr_running == 1)
4786                         break;
4787 
4788                 next = pick_next_task(rq, &fake_task);
4789                 BUG_ON(!next);
4790                 next->sched_class->put_prev_task(rq, next);
4791 
4792                 /* Find suitable destination for @next, with force if needed. */
4793                 dest_cpu = select_fallback_rq(dead_cpu, next);
4794                 raw_spin_unlock(&rq->lock);
4795 
4796                 __migrate_task(next, dead_cpu, dest_cpu);
4797 
4798                 raw_spin_lock(&rq->lock);
4799         }
4800 
4801         rq->stop = stop;
4802 }
4803 
4804 #endif /* CONFIG_HOTPLUG_CPU */
4805 
4806 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4807 
4808 static struct ctl_table sd_ctl_dir[] = {
4809         {
4810                 .procname       = "sched_domain",
4811                 .mode           = 0555,
4812         },
4813         {}
4814 };
4815 
4816 static struct ctl_table sd_ctl_root[] = {
4817         {
4818                 .procname       = "kernel",
4819                 .mode           = 0555,
4820                 .child          = sd_ctl_dir,
4821         },
4822         {}
4823 };
4824 
4825 static struct ctl_table *sd_alloc_ctl_entry(int n)
4826 {
4827         struct ctl_table *entry =
4828                 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4829 
4830         return entry;
4831 }
4832 
4833 static void sd_free_ctl_entry(struct ctl_table **tablep)
4834 {
4835         struct ctl_table *entry;
4836 
4837         /*
4838          * In the intermediate directories, both the child directory and
4839          * procname are dynamically allocated and could fail but the mode
4840          * will always be set. In the lowest directory the names are
4841          * static strings and all have proc handlers.
4842          */
4843         for (entry = *tablep; entry->mode; entry++) {
4844                 if (entry->child)
4845                         sd_free_ctl_entry(&entry->child);
4846                 if (entry->proc_handler == NULL)
4847                         kfree(entry->procname);
4848         }
4849 
4850         kfree(*tablep);
4851         *tablep = NULL;
4852 }
4853 
4854 static int min_load_idx = 0;
4855 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4856 
4857 static void
4858 set_table_entry(struct ctl_table *entry,
4859                 const char *procname, void *data, int maxlen,
4860                 umode_t mode, proc_handler *proc_handler,
4861                 bool load_idx)
4862 {
4863         entry->procname = procname;
4864         entry->data = data;
4865         entry->maxlen = maxlen;
4866         entry->mode = mode;
4867         entry->proc_handler = proc_handler;
4868 
4869         if (load_idx) {
4870                 entry->extra1 = &min_load_idx;
4871                 entry->extra2 = &max_load_idx;
4872         }
4873 }
4874 
4875 static struct ctl_table *
4876 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4877 {
4878         struct ctl_table *table = sd_alloc_ctl_entry(14);
4879 
4880         if (table == NULL)
4881                 return NULL;
4882 
4883         set_table_entry(&table[0], "min_interval", &sd->min_interval,
4884                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4885         set_table_entry(&table[1], "max_interval", &sd->max_interval,
4886                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4887         set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4888                 sizeof(int), 0644, proc_dointvec_minmax, true);
4889         set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4890                 sizeof(int), 0644, proc_dointvec_minmax, true);
4891         set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4892                 sizeof(int), 0644, proc_dointvec_minmax, true);
4893         set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4894                 sizeof(int), 0644, proc_dointvec_minmax, true);
4895         set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4896                 sizeof(int), 0644, proc_dointvec_minmax, true);
4897         set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4898                 sizeof(int), 0644, proc_dointvec_minmax, false);
4899         set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4900                 sizeof(int), 0644, proc_dointvec_minmax, false);
4901         set_table_entry(&table[9], "cache_nice_tries",
4902                 &sd->cache_nice_tries,
4903                 sizeof(int), 0644, proc_dointvec_minmax, false);
4904         set_table_entry(&table[10], "flags", &sd->flags,
4905                 sizeof(int), 0644, proc_dointvec_minmax, false);
4906         set_table_entry(&table[11], "max_newidle_lb_cost",
4907                 &sd->max_newidle_lb_cost,
4908                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4909         set_table_entry(&table[12], "name", sd->name,
4910                 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4911         /* &table[13] is terminator */
4912 
4913         return table;
4914 }
4915 
4916 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4917 {
4918         struct ctl_table *entry, *table;
4919         struct sched_domain *sd;
4920         int domain_num = 0, i;
4921         char buf[32];
4922 
4923         for_each_domain(cpu, sd)
4924                 domain_num++;
4925         entry = table = sd_alloc_ctl_entry(domain_num + 1);
4926         if (table == NULL)
4927                 return NULL;
4928 
4929         i = 0;
4930         for_each_domain(cpu, sd) {
4931                 snprintf(buf, 32, "domain%d", i);
4932                 entry->procname = kstrdup(buf, GFP_KERNEL);
4933                 entry->mode = 0555;
4934                 entry->child = sd_alloc_ctl_domain_table(sd);
4935                 entry++;
4936                 i++;
4937         }
4938         return table;
4939 }
4940 
4941 static struct ctl_table_header *sd_sysctl_header;
4942 static void register_sched_domain_sysctl(void)
4943 {
4944         int i, cpu_num = num_possible_cpus();
4945         struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4946         char buf[32];
4947 
4948         WARN_ON(sd_ctl_dir[0].child);
4949         sd_ctl_dir[0].child = entry;
4950 
4951         if (entry == NULL)
4952                 return;
4953 
4954         for_each_possible_cpu(i) {
4955                 snprintf(buf, 32, "cpu%d", i);
4956                 entry->procname = kstrdup(buf, GFP_KERNEL);
4957                 entry->mode = 0555;
4958                 entry->child = sd_alloc_ctl_cpu_table(i);
4959                 entry++;
4960         }
4961 
4962         WARN_ON(sd_sysctl_header);
4963         sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4964 }
4965 
4966 /* may be called multiple times per register */
4967 static void unregister_sched_domain_sysctl(void)
4968 {
4969         if (sd_sysctl_header)
4970                 unregister_sysctl_table(sd_sysctl_header);
4971         sd_sysctl_header = NULL;
4972         if (sd_ctl_dir[0].child)
4973                 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4974 }
4975 #else
4976 static void register_sched_domain_sysctl(void)
4977 {
4978 }
4979 static void unregister_sched_domain_sysctl(void)
4980 {
4981 }
4982 #endif
4983 
4984 static void set_rq_online(struct rq *rq)
4985 {
4986         if (!rq->online) {
4987                 const struct sched_class *class;
4988 
4989                 cpumask_set_cpu(rq->cpu, rq->rd->online);
4990                 rq->online = 1;
4991 
4992                 for_each_class(class) {
4993                         if (class->rq_online)
4994                                 class->rq_online(rq);
4995                 }
4996         }
4997 }
4998 
4999 static void set_rq_offline(struct rq *rq)
5000 {
5001         if (rq->online) {
5002                 const struct sched_class *class;
5003 
5004                 for_each_class(class) {
5005                         if (class->rq_offline)
5006                                 class->rq_offline(rq);
5007                 }
5008 
5009                 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5010                 rq->online = 0;
5011         }
5012 }
5013 
5014 /*
5015  * migration_call - callback that gets triggered when a CPU is added.
5016  * Here we can start up the necessary migration thread for the new CPU.
5017  */
5018 static int
5019 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5020 {
5021         int cpu = (long)hcpu;
5022         unsigned long flags;
5023         struct rq *rq = cpu_rq(cpu);
5024 
5025         switch (action & ~CPU_TASKS_FROZEN) {
5026 
5027         case CPU_UP_PREPARE:
5028                 rq->calc_load_update = calc_load_update;
5029                 break;
5030 
5031         case CPU_ONLINE:
5032                 /* Update our root-domain */
5033                 raw_spin_lock_irqsave(&rq->lock, flags);
5034                 if (rq->rd) {
5035                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5036 
5037                         set_rq_online(rq);
5038                 }
5039                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5040                 break;
5041 
5042 #ifdef CONFIG_HOTPLUG_CPU
5043         case CPU_DYING:
5044                 sched_ttwu_pending();
5045                 /* Update our root-domain */
5046                 raw_spin_lock_irqsave(&rq->lock, flags);
5047                 if (rq->rd) {
5048                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5049                         set_rq_offline(rq);
5050                 }
5051                 migrate_tasks(cpu);
5052                 BUG_ON(rq->nr_running != 1); /* the migration thread */
5053                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5054                 break;
5055 
5056         case CPU_DEAD:
5057                 calc_load_migrate(rq);
5058                 break;
5059 #endif
5060         }
5061 
5062         update_max_interval();
5063 
5064         return NOTIFY_OK;
5065 }
5066 
5067 /*
5068  * Register at high priority so that task migration (migrate_all_tasks)
5069  * happens before everything else.  This has to be lower priority than
5070  * the notifier in the perf_event subsystem, though.
5071  */
5072 static struct notifier_block migration_notifier = {
5073         .notifier_call = migration_call,
5074         .priority = CPU_PRI_MIGRATION,
5075 };
5076 
5077 static int sched_cpu_active(struct notifier_block *nfb,
5078                                       unsigned long action, void *hcpu)
5079 {
5080         switch (action & ~CPU_TASKS_FROZEN) {
5081         case CPU_DOWN_FAILED:
5082                 set_cpu_active((long)hcpu, true);
5083                 return NOTIFY_OK;
5084         default:
5085                 return NOTIFY_DONE;
5086         }
5087 }
5088 
5089 static int sched_cpu_inactive(struct notifier_block *nfb,
5090                                         unsigned long action, void *hcpu)
5091 {
5092         unsigned long flags;
5093         long cpu = (long)hcpu;
5094 
5095         switch (action & ~CPU_TASKS_FROZEN) {
5096         case CPU_DOWN_PREPARE:
5097                 set_cpu_active(cpu, false);
5098 
5099                 /* explicitly allow suspend */
5100                 if (!(action & CPU_TASKS_FROZEN)) {
5101                         struct dl_bw *dl_b = dl_bw_of(cpu);
5102                         bool overflow;
5103                         int cpus;
5104 
5105                         raw_spin_lock_irqsave(&dl_b->lock, flags);
5106                         cpus = dl_bw_cpus(cpu);
5107                         overflow = __dl_overflow(dl_b, cpus, 0, 0);
5108                         raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5109 
5110                         if (overflow)
5111                                 return notifier_from_errno(-EBUSY);
5112                 }
5113                 return NOTIFY_OK;
5114         }
5115 
5116         return NOTIFY_DONE;
5117 }
5118 
5119 static int __init migration_init(void)
5120 {
5121         void *cpu = (void *)(long)smp_processor_id();
5122         int err;
5123 
5124         /* Initialize migration for the boot CPU */
5125         err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5126         BUG_ON(err == NOTIFY_BAD);
5127         migration_call(&migration_notifier, CPU_ONLINE, cpu);
5128         register_cpu_notifier(&migration_notifier);
5129 
5130         /* Register cpu active notifiers */
5131         cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5132         cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5133 
5134         return 0;
5135 }
5136 early_initcall(migration_init);
5137 #endif
5138 
5139 #ifdef CONFIG_SMP
5140 
5141 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5142 
5143 #ifdef CONFIG_SCHED_DEBUG
5144 
5145 static __read_mostly int sched_debug_enabled;
5146 
5147 static int __init sched_debug_setup(char *str)
5148 {
5149         sched_debug_enabled = 1;
5150 
5151         return 0;
5152 }
5153 early_param("sched_debug", sched_debug_setup);
5154 
5155 static inline bool sched_debug(void)
5156 {
5157         return sched_debug_enabled;
5158 }
5159 
5160 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5161                                   struct cpumask *groupmask)
5162 {
5163         struct sched_group *group = sd->groups;
5164         char str[256];
5165 
5166         cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5167         cpumask_clear(groupmask);
5168 
5169         printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5170 
5171         if (!(sd->flags & SD_LOAD_BALANCE)) {
5172                 printk("does not load-balance\n");
5173                 if (sd->parent)
5174                         printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5175                                         " has parent");
5176                 return -1;
5177         }
5178 
5179         printk(KERN_CONT "span %s level %s\n", str, sd->name);
5180 
5181         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5182                 printk(KERN_ERR "ERROR: domain->span does not contain "
5183                                 "CPU%d\n", cpu);
5184         }
5185         if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5186                 printk(KERN_ERR "ERROR: domain->groups does not contain"
5187                                 " CPU%d\n", cpu);
5188         }
5189 
5190         printk(KERN_DEBUG "%*s groups:", level + 1, "");
5191         do {
5192                 if (!group) {
5193                         printk("\n");
5194                         printk(KERN_ERR "ERROR: group is NULL\n");
5195                         break;
5196                 }
5197 
5198                 /*
5199                  * Even though we initialize ->power to something semi-sane,
5200                  * we leave power_orig unset. This allows us to detect if
5201                  * domain iteration is still funny without causing /0 traps.
5202                  */
5203                 if (!group->sgp->power_orig) {
5204                         printk(KERN_CONT "\n");
5205                         printk(KERN_ERR "ERROR: domain->cpu_power not "
5206                                         "set\n");
5207                         break;
5208                 }
5209 
5210                 if (!cpumask_weight(sched_group_cpus(group))) {
5211                         printk(KERN_CONT "\n");
5212                         printk(KERN_ERR "ERROR: empty group\n");
5213                         break;
5214                 }
5215 
5216                 if (!(sd->flags & SD_OVERLAP) &&
5217                     cpumask_intersects(groupmask, sched_group_cpus(group))) {
5218                         printk(KERN_CONT "\n");
5219                         printk(KERN_ERR "ERROR: repeated CPUs\n");
5220                         break;
5221                 }
5222 
5223                 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5224 
5225                 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5226 
5227                 printk(KERN_CONT " %s", str);
5228                 if (group->sgp->power != SCHED_POWER_SCALE) {
5229                         printk(KERN_CONT " (cpu_power = %d)",
5230                                 group->sgp->power);
5231                 }
5232 
5233                 group = group->next;
5234         } while (group != sd->groups);
5235         printk(KERN_CONT "\n");
5236 
5237         if (!cpumask_equal(sched_domain_span(sd), groupmask))
5238                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5239 
5240         if (sd->parent &&
5241             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5242                 printk(KERN_ERR "ERROR: parent span is not a superset "
5243                         "of domain->span\n");
5244         return 0;
5245 }
5246 
5247 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5248 {
5249         int level = 0;
5250 
5251         if (!sched_debug_enabled)
5252                 return;
5253 
5254         if (!sd) {
5255                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5256                 return;
5257         }
5258 
5259         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5260 
5261         for (;;) {
5262                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5263                         break;
5264                 level++;
5265                 sd = sd->parent;
5266                 if (!sd)
5267                         break;
5268         }
5269 }
5270 #else /* !CONFIG_SCHED_DEBUG */
5271 # define sched_domain_debug(sd, cpu) do { } while (0)
5272 static inline bool sched_debug(void)
5273 {
5274         return false;
5275 }
5276 #endif /* CONFIG_SCHED_DEBUG */
5277 
5278 static int sd_degenerate(struct sched_domain *sd)
5279 {
5280         if (cpumask_weight(sched_domain_span(sd)) == 1)
5281                 return 1;
5282 
5283         /* Following flags need at least 2 groups */
5284         if (sd->flags & (SD_LOAD_BALANCE |
5285                          SD_BALANCE_NEWIDLE |
5286                          SD_BALANCE_FORK |
5287                          SD_BALANCE_EXEC |
5288                          SD_SHARE_CPUPOWER |
5289                          SD_SHARE_PKG_RESOURCES)) {
5290                 if (sd->groups != sd->groups->next)
5291                         return 0;
5292         }
5293 
5294         /* Following flags don't use groups */
5295         if (sd->flags & (SD_WAKE_AFFINE))
5296                 return 0;
5297 
5298         return 1;
5299 }
5300 
5301 static int
5302 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5303 {
5304         unsigned long cflags = sd->flags, pflags = parent->flags;
5305 
5306         if (sd_degenerate(parent))
5307                 return 1;
5308 
5309         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5310                 return 0;
5311 
5312         /* Flags needing groups don't count if only 1 group in parent */
5313         if (parent->groups == parent->groups->next) {
5314                 pflags &= ~(SD_LOAD_BALANCE |
5315                                 SD_BALANCE_NEWIDLE |
5316                                 SD_BALANCE_FORK |
5317                                 SD_BALANCE_EXEC |
5318                                 SD_SHARE_CPUPOWER |
5319                                 SD_SHARE_PKG_RESOURCES |
5320                                 SD_PREFER_SIBLING);
5321                 if (nr_node_ids == 1)
5322                         pflags &= ~SD_SERIALIZE;
5323         }
5324         if (~cflags & pflags)
5325                 return 0;
5326 
5327         return 1;
5328 }
5329 
5330 static void free_rootdomain(struct rcu_head *rcu)
5331 {
5332         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5333 
5334         cpupri_cleanup(&rd->cpupri);
5335         cpudl_cleanup(&rd->cpudl);
5336         free_cpumask_var(rd->dlo_mask);
5337         free_cpumask_var(rd->rto_mask);
5338         free_cpumask_var(rd->online);
5339         free_cpumask_var(rd->span);
5340         kfree(rd);
5341 }
5342 
5343 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5344 {
5345         struct root_domain *old_rd = NULL;
5346         unsigned long flags;
5347 
5348         raw_spin_lock_irqsave(&rq->lock, flags);
5349 
5350         if (rq->rd) {
5351                 old_rd = rq->rd;
5352 
5353                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5354                         set_rq_offline(rq);
5355 
5356                 cpumask_clear_cpu(rq->cpu, old_rd->span);
5357 
5358                 /*
5359                  * If we dont want to free the old_rd yet then
5360                  * set old_rd to NULL to skip the freeing later
5361                  * in this function:
5362                  */
5363                 if (!atomic_dec_and_test(&old_rd->refcount))
5364                         old_rd = NULL;
5365         }
5366 
5367         atomic_inc(&rd->refcount);
5368         rq->rd = rd;
5369 
5370         cpumask_set_cpu(rq->cpu, rd->span);
5371         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5372                 set_rq_online(rq);
5373 
5374         raw_spin_unlock_irqrestore(&rq->lock, flags);
5375 
5376         if (old_rd)
5377                 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5378 }
5379 
5380 static int init_rootdomain(struct root_domain *rd)
5381 {
5382         memset(rd, 0, sizeof(*rd));
5383 
5384         if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5385                 goto out;
5386         if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5387                 goto free_span;
5388         if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5389                 goto free_online;
5390         if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5391                 goto free_dlo_mask;
5392 
5393         init_dl_bw(&rd->dl_bw);
5394         if (cpudl_init(&rd->cpudl) != 0)
5395                 goto free_dlo_mask;
5396 
5397         if (cpupri_init(&rd->cpupri) != 0)
5398                 goto free_rto_mask;
5399         return 0;
5400 
5401 free_rto_mask:
5402         free_cpumask_var(rd->rto_mask);
5403 free_dlo_mask:
5404         free_cpumask_var(rd->dlo_mask);
5405 free_online:
5406         free_cpumask_var(rd->online);
5407 free_span:
5408         free_cpumask_var(rd->span);
5409 out:
5410         return -ENOMEM;
5411 }
5412 
5413 /*
5414  * By default the system creates a single root-domain with all cpus as
5415  * members (mimicking the global state we have today).
5416  */
5417 struct root_domain def_root_domain;
5418 
5419 static void init_defrootdomain(void)
5420 {
5421         init_rootdomain(&def_root_domain);
5422 
5423         atomic_set(&def_root_domain.refcount, 1);
5424 }
5425 
5426 static struct root_domain *alloc_rootdomain(void)
5427 {
5428         struct root_domain *rd;
5429 
5430         rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5431         if (!rd)
5432                 return NULL;
5433 
5434         if (init_rootdomain(rd) != 0) {
5435                 kfree(rd);
5436                 return NULL;
5437         }
5438 
5439         return rd;
5440 }
5441 
5442 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5443 {
5444         struct sched_group *tmp, *first;
5445 
5446         if (!sg)
5447                 return;
5448 
5449         first = sg;
5450         do {
5451                 tmp = sg->next;
5452 
5453                 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5454                         kfree(sg->sgp);
5455 
5456                 kfree(sg);
5457                 sg = tmp;
5458         } while (sg != first);
5459 }
5460 
5461 static void free_sched_domain(struct rcu_head *rcu)
5462 {
5463         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5464 
5465         /*
5466          * If its an overlapping domain it has private groups, iterate and
5467          * nuke them all.
5468          */
5469         if (sd->flags & SD_OVERLAP) {
5470                 free_sched_groups(sd->groups, 1);
5471         } else if (atomic_dec_and_test(&sd->groups->ref)) {
5472                 kfree(sd->groups->sgp);
5473                 kfree(sd->groups);
5474         }
5475         kfree(sd);
5476 }
5477 
5478 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5479 {
5480         call_rcu(&sd->rcu, free_sched_domain);
5481 }
5482 
5483 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5484 {
5485         for (; sd; sd = sd->parent)
5486                 destroy_sched_domain(sd, cpu);
5487 }
5488 
5489 /*
5490  * Keep a special pointer to the highest sched_domain that has
5491  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5492  * allows us to avoid some pointer chasing select_idle_sibling().
5493  *
5494  * Also keep a unique ID per domain (we use the first cpu number in
5495  * the cpumask of the domain), this allows us to quickly tell if
5496  * two cpus are in the same cache domain, see cpus_share_cache().
5497  */
5498 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5499 DEFINE_PER_CPU(int, sd_llc_size);
5500 DEFINE_PER_CPU(int, sd_llc_id);
5501 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5502 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5503 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5504 
5505 static void update_top_cache_domain(int cpu)
5506 {
5507         struct sched_domain *sd;
5508         struct sched_domain *busy_sd = NULL;
5509         int id = cpu;
5510         int size = 1;
5511 
5512         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5513         if (sd) {
5514                 id = cpumask_first(sched_domain_span(sd));
5515                 size = cpumask_weight(sched_domain_span(sd));
5516                 busy_sd = sd->parent; /* sd_busy */
5517         }
5518         rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5519 
5520         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5521         per_cpu(sd_llc_size, cpu) = size;
5522         per_cpu(sd_llc_id, cpu) = id;
5523 
5524         sd = lowest_flag_domain(cpu, SD_NUMA);
5525         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5526 
5527         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5528         rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5529 }
5530 
5531 /*
5532  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5533  * hold the hotplug lock.
5534  */
5535 static void
5536 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5537 {
5538         struct rq *rq = cpu_rq(cpu);
5539         struct sched_domain *tmp;
5540 
5541         /* Remove the sched domains which do not contribute to scheduling. */
5542         for (tmp = sd; tmp; ) {
5543                 struct sched_domain *parent = tmp->parent;
5544                 if (!parent)
5545                         break;
5546 
5547                 if (sd_parent_degenerate(tmp, parent)) {
5548                         tmp->parent = parent->parent;
5549                         if (parent->parent)
5550                                 parent->parent->child = tmp;
5551                         /*
5552                          * Transfer SD_PREFER_SIBLING down in case of a
5553                          * degenerate parent; the spans match for this
5554                          * so the property transfers.
5555                          */
5556                         if (parent->flags & SD_PREFER_SIBLING)
5557                                 tmp->flags |= SD_PREFER_SIBLING;
5558                         destroy_sched_domain(parent, cpu);
5559                 } else
5560                         tmp = tmp->parent;
5561         }
5562 
5563         if (sd && sd_degenerate(sd)) {
5564                 tmp = sd;
5565                 sd = sd->parent;
5566                 destroy_sched_domain(tmp, cpu);
5567                 if (sd)
5568                         sd->child = NULL;
5569         }
5570 
5571         sched_domain_debug(sd, cpu);
5572 
5573         rq_attach_root(rq, rd);
5574         tmp = rq->sd;
5575         rcu_assign_pointer(rq->sd, sd);
5576         destroy_sched_domains(tmp, cpu);
5577 
5578         update_top_cache_domain(cpu);
5579 }
5580 
5581 /* cpus with isolated domains */
5582 static cpumask_var_t cpu_isolated_map;
5583 
5584 /* Setup the mask of cpus configured for isolated domains */
5585 static int __init isolated_cpu_setup(char *str)
5586 {
5587         alloc_bootmem_cpumask_var(&cpu_isolated_map);
5588         cpulist_parse(str, cpu_isolated_map);
5589         return 1;
5590 }
5591 
5592 __setup("isolcpus=", isolated_cpu_setup);
5593 
5594 static const struct cpumask *cpu_cpu_mask(int cpu)
5595 {
5596         return cpumask_of_node(cpu_to_node(cpu));
5597 }
5598 
5599 struct sd_data {
5600         struct sched_domain **__percpu sd;
5601         struct sched_group **__percpu sg;
5602         struct sched_group_power **__percpu sgp;
5603 };
5604 
5605 struct s_data {
5606         struct sched_domain ** __percpu sd;
5607         struct root_domain      *rd;
5608 };
5609 
5610 enum s_alloc {
5611         sa_rootdomain,
5612         sa_sd,
5613         sa_sd_storage,
5614         sa_none,
5615 };
5616 
5617 struct sched_domain_topology_level;
5618 
5619 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5620 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5621 
5622 #define SDTL_OVERLAP    0x01
5623 
5624 struct sched_domain_topology_level {
5625         sched_domain_init_f init;
5626         sched_domain_mask_f mask;
5627         int                 flags;
5628         int                 numa_level;
5629         struct sd_data      data;
5630 };
5631 
5632 /*
5633  * Build an iteration mask that can exclude certain CPUs from the upwards
5634  * domain traversal.
5635  *
5636  * Asymmetric node setups can result in situations where the domain tree is of
5637  * unequal depth, make sure to skip domains that already cover the entire
5638  * range.
5639  *
5640  * In that case build_sched_domains() will have terminated the iteration early
5641  * and our sibling sd spans will be empty. Domains should always include the
5642  * cpu they're built on, so check that.
5643  *
5644  */
5645 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5646 {
5647         const struct cpumask *span = sched_domain_span(sd);
5648         struct sd_data *sdd = sd->private;
5649         struct sched_domain *sibling;
5650         int i;
5651 
5652         for_each_cpu(i, span) {
5653                 sibling = *per_cpu_ptr(sdd->sd, i);
5654                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5655                         continue;
5656 
5657                 cpumask_set_cpu(i, sched_group_mask(sg));
5658         }
5659 }
5660 
5661 /*
5662  * Return the canonical balance cpu for this group, this is the first cpu
5663  * of this group that's also in the iteration mask.
5664  */
5665 int group_balance_cpu(struct sched_group *sg)
5666 {
5667         return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5668 }
5669 
5670 static int
5671 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5672 {
5673         struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5674         const struct cpumask *span = sched_domain_span(sd);
5675         struct cpumask *covered = sched_domains_tmpmask;
5676         struct sd_data *sdd = sd->private;
5677         struct sched_domain *child;
5678         int i;
5679 
5680         cpumask_clear(covered);
5681 
5682         for_each_cpu(i, span) {
5683                 struct cpumask *sg_span;
5684 
5685                 if (cpumask_test_cpu(i, covered))
5686                         continue;
5687 
5688                 child = *per_cpu_ptr(sdd->sd, i);
5689 
5690                 /* See the comment near build_group_mask(). */
5691                 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5692                         continue;
5693 
5694                 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5695                                 GFP_KERNEL, cpu_to_node(cpu));
5696 
5697                 if (!sg)
5698                         goto fail;
5699 
5700                 sg_span = sched_group_cpus(sg);
5701                 if (child->child) {
5702                         child = child->child;
5703                         cpumask_copy(sg_span, sched_domain_span(child));
5704                 } else
5705                         cpumask_set_cpu(i, sg_span);
5706 
5707                 cpumask_or(covered, covered, sg_span);
5708 
5709                 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5710                 if (atomic_inc_return(&sg->sgp->ref) == 1)
5711                         build_group_mask(sd, sg);
5712 
5713                 /*
5714                  * Initialize sgp->power such that even if we mess up the
5715                  * domains and no possible iteration will get us here, we won't
5716                  * die on a /0 trap.
5717                  */
5718                 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5719                 sg->sgp->power_orig = sg->sgp->power;
5720 
5721                 /*
5722                  * Make sure the first group of this domain contains the
5723                  * canonical balance cpu. Otherwise the sched_domain iteration
5724                  * breaks. See update_sg_lb_stats().
5725                  */
5726                 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5727                     group_balance_cpu(sg) == cpu)
5728                         groups = sg;
5729 
5730                 if (!first)
5731                         first = sg;
5732                 if (last)
5733                         last->next = sg;
5734                 last = sg;
5735                 last->next = first;
5736         }
5737         sd->groups = groups;
5738 
5739         return 0;
5740 
5741 fail:
5742         free_sched_groups(first, 0);
5743 
5744         return -ENOMEM;
5745 }
5746 
5747 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5748 {
5749         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5750         struct sched_domain *child = sd->child;
5751 
5752         if (child)
5753                 cpu = cpumask_first(sched_domain_span(child));
5754 
5755         if (sg) {
5756                 *sg = *per_cpu_ptr(sdd->sg, cpu);
5757                 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5758                 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5759         }
5760 
5761         return cpu;
5762 }
5763 
5764 /*
5765  * build_sched_groups will build a circular linked list of the groups
5766  * covered by the given span, and will set each group's ->cpumask correctly,
5767  * and ->cpu_power to 0.
5768  *
5769  * Assumes the sched_domain tree is fully constructed
5770  */
5771 static int
5772 build_sched_groups(struct sched_domain *sd, int cpu)
5773 {
5774         struct sched_group *first = NULL, *last = NULL;
5775         struct sd_data *sdd = sd->private;
5776         const struct cpumask *span = sched_domain_span(sd);
5777         struct cpumask *covered;
5778         int i;
5779 
5780         get_group(cpu, sdd, &sd->groups);
5781         atomic_inc(&sd->groups->ref);
5782 
5783         if (cpu != cpumask_first(span))
5784                 return 0;
5785 
5786         lockdep_assert_held(&sched_domains_mutex);
5787         covered = sched_domains_tmpmask;
5788 
5789         cpumask_clear(covered);
5790 
5791         for_each_cpu(i, span) {
5792                 struct sched_group *sg;
5793                 int group, j;
5794 
5795                 if (cpumask_test_cpu(i, covered))
5796                         continue;
5797 
5798                 group = get_group(i, sdd, &sg);
5799                 cpumask_clear(sched_group_cpus(sg));
5800                 sg->sgp->power = 0;
5801                 cpumask_setall(sched_group_mask(sg));
5802 
5803                 for_each_cpu(j, span) {
5804                         if (get_group(j, sdd, NULL) != group)
5805                                 continue;
5806 
5807                         cpumask_set_cpu(j, covered);
5808                         cpumask_set_cpu(j, sched_group_cpus(sg));
5809                 }
5810 
5811                 if (!first)
5812                         first = sg;
5813                 if (last)
5814                         last->next = sg;
5815                 last = sg;
5816         }
5817         last->next = first;
5818 
5819         return 0;
5820 }
5821 
5822 /*
5823  * Initialize sched groups cpu_power.
5824  *
5825  * cpu_power indicates the capacity of sched group, which is used while
5826  * distributing the load between different sched groups in a sched domain.
5827  * Typically cpu_power for all the groups in a sched domain will be same unless
5828  * there are asymmetries in the topology. If there are asymmetries, group
5829  * having more cpu_power will pickup more load compared to the group having
5830  * less cpu_power.
5831  */
5832 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5833 {
5834         struct sched_group *sg = sd->groups;
5835 
5836         WARN_ON(!sg);
5837 
5838         do {
5839                 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5840                 sg = sg->next;
5841         } while (sg != sd->groups);
5842 
5843         if (cpu != group_balance_cpu(sg))
5844                 return;
5845 
5846         update_group_power(sd, cpu);
5847         atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5848 }
5849 
5850 int __weak arch_sd_sibling_asym_packing(void)
5851 {
5852        return 0*SD_ASYM_PACKING;
5853 }
5854 
5855 /*
5856  * Initializers for schedule domains
5857  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5858  */
5859 
5860 #ifdef CONFIG_SCHED_DEBUG
5861 # define SD_INIT_NAME(sd, type)         sd->name = #type
5862 #else
5863 # define SD_INIT_NAME(sd, type)         do { } while (0)
5864 #endif
5865 
5866 #define SD_INIT_FUNC(type)                                              \
5867 static noinline struct sched_domain *                                   \
5868 sd_init_##type(struct sched_domain_topology_level *tl, int cpu)         \
5869 {                                                                       \
5870         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);       \
5871         *sd = SD_##type##_INIT;                                         \
5872         SD_INIT_NAME(sd, type);                                         \
5873         sd->private = &tl->data;                                        \
5874         return sd;                                                      \
5875 }
5876 
5877 SD_INIT_FUNC(CPU)
5878 #ifdef CONFIG_SCHED_SMT
5879  SD_INIT_FUNC(SIBLING)
5880 #endif
5881 #ifdef CONFIG_SCHED_MC
5882  SD_INIT_FUNC(MC)
5883 #endif
5884 #ifdef CONFIG_SCHED_BOOK
5885  SD_INIT_FUNC(BOOK)
5886 #endif
5887 
5888 static int default_relax_domain_level = -1;
5889 int sched_domain_level_max;
5890 
5891 static int __init setup_relax_domain_level(char *str)
5892 {
5893         if (kstrtoint(str, 0, &default_relax_domain_level))
5894                 pr_warn("Unable to set relax_domain_level\n");
5895 
5896         return 1;
5897 }
5898 __setup("relax_domain_level=", setup_relax_domain_level);
5899 
5900 static void set_domain_attribute(struct sched_domain *sd,
5901                                  struct sched_domain_attr *attr)
5902 {
5903         int request;
5904 
5905         if (!attr || attr->relax_domain_level < 0) {
5906                 if (default_relax_domain_level < 0)
5907                         return;
5908                 else
5909                         request = default_relax_domain_level;
5910         } else
5911                 request = attr->relax_domain_level;
5912         if (request < sd->level) {
5913                 /* turn off idle balance on this domain */
5914                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5915         } else {
5916                 /* turn on idle balance on this domain */
5917                 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5918         }
5919 }
5920 
5921 static void __sdt_free(const struct cpumask *cpu_map);
5922 static int __sdt_alloc(const struct cpumask *cpu_map);
5923 
5924 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5925                                  const struct cpumask *cpu_map)
5926 {
5927         switch (what) {
5928         case sa_rootdomain:
5929                 if (!atomic_read(&d->rd->refcount))
5930                         free_rootdomain(&d->rd->rcu); /* fall through */
5931         case sa_sd:
5932                 free_percpu(d->sd); /* fall through */
5933         case sa_sd_storage:
5934                 __sdt_free(cpu_map); /* fall through */
5935         case sa_none:
5936                 break;
5937         }
5938 }
5939 
5940 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5941                                                    const struct cpumask *cpu_map)
5942 {
5943         memset(d, 0, sizeof(*d));
5944 
5945         if (__sdt_alloc(cpu_map))
5946                 return sa_sd_storage;
5947         d->sd = alloc_percpu(struct sched_domain *);
5948         if (!d->sd)
5949                 return sa_sd_storage;
5950         d->rd = alloc_rootdomain();
5951         if (!d->rd)
5952                 return sa_sd;
5953         return sa_rootdomain;
5954 }
5955 
5956 /*
5957  * NULL the sd_data elements we've used to build the sched_domain and
5958  * sched_group structure so that the subsequent __free_domain_allocs()
5959  * will not free the data we're using.
5960  */
5961 static void claim_allocations(int cpu, struct sched_domain *sd)
5962 {
5963         struct sd_data *sdd = sd->private;
5964 
5965         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5966         *per_cpu_ptr(sdd->sd, cpu) = NULL;
5967 
5968         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5969                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5970 
5971         if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5972                 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5973 }
5974 
5975 #ifdef CONFIG_SCHED_SMT
5976 static const struct cpumask *cpu_smt_mask(int cpu)
5977 {
5978         return topology_thread_cpumask(cpu);
5979 }
5980 #endif
5981 
5982 /*
5983  * Topology list, bottom-up.
5984  */
5985 static struct sched_domain_topology_level default_topology[] = {
5986 #ifdef CONFIG_SCHED_SMT
5987         { sd_init_SIBLING, cpu_smt_mask, },
5988 #endif
5989 #ifdef CONFIG_SCHED_MC
5990         { sd_init_MC, cpu_coregroup_mask, },
5991 #endif
5992 #ifdef CONFIG_SCHED_BOOK
5993         { sd_init_BOOK, cpu_book_mask, },
5994 #endif
5995         { sd_init_CPU, cpu_cpu_mask, },
5996         { NULL, },
5997 };
5998 
5999 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6000 
6001 #define for_each_sd_topology(tl)                        \
6002         for (tl = sched_domain_topology; tl->init; tl++)
6003 
6004 #ifdef CONFIG_NUMA
6005 
6006 static int sched_domains_numa_levels;
6007 static int *sched_domains_numa_distance;
6008 static struct cpumask ***sched_domains_numa_masks;
6009 static int sched_domains_curr_level;
6010 
6011 static inline int sd_local_flags(int level)
6012 {
6013         if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6014                 return 0;
6015 
6016         return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6017 }
6018 
6019 static struct sched_domain *
6020 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6021 {
6022         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6023         int level = tl->numa_level;
6024         int sd_weight = cpumask_weight(
6025                         sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6026 
6027         *sd = (struct sched_domain){
6028                 .min_interval           = sd_weight,
6029                 .max_interval           = 2*sd_weight,
6030                 .busy_factor            = 32,
6031                 .imbalance_pct          = 125,
6032                 .cache_nice_tries       = 2,
6033                 .busy_idx               = 3,
6034                 .idle_idx               = 2,
6035                 .newidle_idx            = 0,
6036                 .wake_idx               = 0,
6037                 .forkexec_idx           = 0,
6038 
6039                 .flags                  = 1*SD_LOAD_BALANCE
6040                                         | 1*SD_BALANCE_NEWIDLE
6041                                         | 0*SD_BALANCE_EXEC
6042                                         | 0*SD_BALANCE_FORK
6043                                         | 0*SD_BALANCE_WAKE
6044                                         | 0*SD_WAKE_AFFINE
6045                                         | 0*SD_SHARE_CPUPOWER
6046                                         | 0*SD_SHARE_PKG_RESOURCES
6047                                         | 1*SD_SERIALIZE
6048                                         | 0*SD_PREFER_SIBLING
6049                                         | 1*SD_NUMA
6050                                         | sd_local_flags(level)
6051                                         ,
6052                 .last_balance           = jiffies,
6053                 .balance_interval       = sd_weight,
6054                 .max_newidle_lb_cost    = 0,
6055                 .next_decay_max_lb_cost = jiffies,
6056         };
6057         SD_INIT_NAME(sd, NUMA);
6058         sd->private = &tl->data;
6059 
6060         /*
6061          * Ugly hack to pass state to sd_numa_mask()...
6062          */
6063         sched_domains_curr_level = tl->numa_level;
6064 
6065         return sd;
6066 }
6067 
6068 static const struct cpumask *sd_numa_mask(int cpu)
6069 {
6070         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6071 }
6072 
6073 static void sched_numa_warn(const char *str)
6074 {
6075         static int done = false;
6076         int i,j;
6077 
6078         if (done)
6079                 return;
6080 
6081         done = true;
6082 
6083         printk(KERN_WARNING "ERROR: %s\n\n", str);
6084 
6085         for (i = 0; i < nr_node_ids; i++) {
6086                 printk(KERN_WARNING "  ");
6087                 for (j = 0; j < nr_node_ids; j++)
6088                         printk(KERN_CONT "%02d ", node_distance(i,j));
6089                 printk(KERN_CONT "\n");
6090         }
6091         printk(KERN_WARNING "\n");
6092 }
6093 
6094 static bool find_numa_distance(int distance)
6095 {
6096         int i;
6097 
6098         if (distance == node_distance(0, 0))
6099                 return true;
6100 
6101         for (i = 0; i < sched_domains_numa_levels; i++) {
6102                 if (sched_domains_numa_distance[i] == distance)
6103                         return true;
6104         }
6105 
6106         return false;
6107 }
6108 
6109 static void sched_init_numa(void)
6110 {
6111         int next_distance, curr_distance = node_distance(0, 0);
6112         struct sched_domain_topology_level *tl;
6113         int level = 0;
6114         int i, j, k;
6115 
6116         sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6117         if (!sched_domains_numa_distance)
6118                 return;
6119 
6120         /*
6121          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6122          * unique distances in the node_distance() table.
6123          *
6124          * Assumes node_distance(0,j) includes all distances in
6125          * node_distance(i,j) in order to avoid cubic time.
6126          */
6127         next_distance = curr_distance;
6128         for (i = 0; i < nr_node_ids; i++) {
6129                 for (j = 0; j < nr_node_ids; j++) {
6130                         for (k = 0; k < nr_node_ids; k++) {
6131                                 int distance = node_distance(i, k);
6132 
6133                                 if (distance > curr_distance &&
6134                                     (distance < next_distance ||
6135                                      next_distance == curr_distance))
6136                                         next_distance = distance;
6137 
6138                                 /*
6139                                  * While not a strong assumption it would be nice to know
6140                                  * about cases where if node A is connected to B, B is not
6141                                  * equally connected to A.
6142                                  */
6143                                 if (sched_debug() && node_distance(k, i) != distance)
6144                                         sched_numa_warn("Node-distance not symmetric");
6145 
6146                                 if (sched_debug() && i && !find_numa_distance(distance))
6147                                         sched_numa_warn("Node-0 not representative");
6148                         }
6149                         if (next_distance != curr_distance) {
6150                                 sched_domains_numa_distance[level++] = next_distance;
6151                                 sched_domains_numa_levels = level;
6152                                 curr_distance = next_distance;
6153                         } else break;
6154                 }
6155 
6156                 /*
6157                  * In case of sched_debug() we verify the above assumption.
6158                  */
6159                 if (!sched_debug())
6160                         break;
6161         }
6162         /*
6163          * 'level' contains the number of unique distances, excluding the
6164          * identity distance node_distance(i,i).
6165          *
6166          * The sched_domains_numa_distance[] array includes the actual distance
6167          * numbers.
6168          */
6169 
6170         /*
6171          * Here, we should temporarily reset sched_domains_numa_levels to 0.
6172          * If it fails to allocate memory for array sched_domains_numa_masks[][],
6173          * the array will contain less then 'level' members. This could be
6174          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6175          * in other functions.
6176          *
6177          * We reset it to 'level' at the end of this function.
6178          */
6179         sched_domains_numa_levels = 0;
6180 
6181         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6182         if (!sched_domains_numa_masks)
6183                 return;
6184 
6185         /*
6186          * Now for each level, construct a mask per node which contains all
6187          * cpus of nodes that are that many hops away from us.
6188          */
6189         for (i = 0; i < level; i++) {
6190                 sched_domains_numa_masks[i] =
6191                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6192                 if (!sched_domains_numa_masks[i])
6193                         return;
6194 
6195                 for (j = 0; j < nr_node_ids; j++) {
6196                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6197                         if (!mask)
6198                                 return;
6199 
6200                         sched_domains_numa_masks[i][j] = mask;
6201 
6202                         for (k = 0; k < nr_node_ids; k++) {
6203                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
6204                                         continue;
6205 
6206                                 cpumask_or(mask, mask, cpumask_of_node(k));
6207                         }
6208                 }
6209         }
6210 
6211         tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6212                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6213         if (!tl)
6214                 return;
6215 
6216         /*
6217          * Copy the default topology bits..
6218          */
6219         for (i = 0; default_topology[i].init; i++)
6220                 tl[i] = default_topology[i];
6221 
6222         /*
6223          * .. and append 'j' levels of NUMA goodness.
6224          */
6225         for (j = 0; j < level; i++, j++) {
6226                 tl[i] = (struct sched_domain_topology_level){
6227                         .init = sd_numa_init,
6228                         .mask = sd_numa_mask,
6229                         .flags = SDTL_OVERLAP,
6230                         .numa_level = j,
6231                 };
6232         }
6233 
6234         sched_domain_topology = tl;
6235 
6236         sched_domains_numa_levels = level;
6237 }
6238 
6239 static void sched_domains_numa_masks_set(int cpu)
6240 {
6241         int i, j;
6242         int node = cpu_to_node(cpu);
6243 
6244         for (i = 0; i < sched_domains_numa_levels; i++) {
6245                 for (j = 0; j < nr_node_ids; j++) {
6246                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
6247                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6248                 }
6249         }
6250 }
6251 
6252 static void sched_domains_numa_masks_clear(int cpu)
6253 {
6254         int i, j;
6255         for (i = 0; i < sched_domains_numa_levels; i++) {
6256                 for (j = 0; j < nr_node_ids; j++)
6257                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6258         }
6259 }
6260 
6261 /*
6262  * Update sched_domains_numa_masks[level][node] array when new cpus
6263  * are onlined.
6264  */
6265 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6266                                            unsigned long action,
6267                                            void *hcpu)
6268 {
6269         int cpu = (long)hcpu;
6270 
6271         switch (action & ~CPU_TASKS_FROZEN) {
6272         case CPU_ONLINE:
6273                 sched_domains_numa_masks_set(cpu);
6274                 break;
6275 
6276         case CPU_DEAD:
6277                 sched_domains_numa_masks_clear(cpu);
6278                 break;
6279 
6280         default:
6281                 return NOTIFY_DONE;
6282         }
6283 
6284         return NOTIFY_OK;
6285 }
6286 #else
6287 static inline void sched_init_numa(void)
6288 {
6289 }
6290 
6291 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6292                                            unsigned long action,
6293                                            void *hcpu)
6294 {
6295         return 0;
6296 }
6297 #endif /* CONFIG_NUMA */
6298 
6299 static int __sdt_alloc(const struct cpumask *cpu_map)
6300 {
6301         struct sched_domain_topology_level *tl;
6302         int j;
6303 
6304         for_each_sd_topology(tl) {
6305                 struct sd_data *sdd = &tl->data;
6306 
6307                 sdd->sd = alloc_percpu(struct sched_domain *);
6308                 if (!sdd->sd)
6309                         return -ENOMEM;
6310 
6311                 sdd->sg = alloc_percpu(struct sched_group *);
6312                 if (!sdd->sg)
6313                         return -ENOMEM;
6314 
6315                 sdd->sgp = alloc_percpu(struct sched_group_power *);
6316                 if (!sdd->sgp)
6317                         return -ENOMEM;
6318 
6319                 for_each_cpu(j, cpu_map) {
6320                         struct sched_domain *sd;
6321                         struct sched_group *sg;
6322                         struct sched_group_power *sgp;
6323 
6324                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6325                                         GFP_KERNEL, cpu_to_node(j));
6326                         if (!sd)
6327                                 return -ENOMEM;
6328 
6329                         *per_cpu_ptr(sdd->sd, j) = sd;
6330 
6331                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6332                                         GFP_KERNEL, cpu_to_node(j));
6333                         if (!sg)
6334                                 return -ENOMEM;
6335 
6336                         sg->next = sg;
6337 
6338                         *per_cpu_ptr(sdd->sg, j) = sg;
6339 
6340                         sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6341                                         GFP_KERNEL, cpu_to_node(j));
6342                         if (!sgp)
6343                                 return -ENOMEM;
6344 
6345                         *per_cpu_ptr(sdd->sgp, j) = sgp;
6346                 }
6347         }
6348 
6349         return 0;
6350 }
6351 
6352 static void __sdt_free(const struct cpumask *cpu_map)
6353 {
6354         struct sched_domain_topology_level *tl;
6355         int j;
6356 
6357         for_each_sd_topology(tl) {
6358                 struct sd_data *sdd = &tl->data;
6359 
6360                 for_each_cpu(j, cpu_map) {
6361                         struct sched_domain *sd;
6362 
6363                         if (sdd->sd) {
6364                                 sd = *per_cpu_ptr(sdd->sd, j);
6365                                 if (sd && (sd->flags & SD_OVERLAP))
6366                                         free_sched_groups(sd->groups, 0);
6367                                 kfree(*per_cpu_ptr(sdd->sd, j));
6368                         }
6369 
6370                         if (sdd->sg)
6371                                 kfree(*per_cpu_ptr(sdd->sg, j));
6372                         if (sdd->sgp)
6373                                 kfree(*per_cpu_ptr(sdd->sgp, j));
6374                 }
6375                 free_percpu(sdd->sd);
6376                 sdd->sd = NULL;
6377                 free_percpu(sdd->sg);
6378                 sdd->sg = NULL;
6379                 free_percpu(sdd->sgp);
6380                 sdd->sgp = NULL;
6381         }
6382 }
6383 
6384 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6385                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6386                 struct sched_domain *child, int cpu)
6387 {
6388         struct sched_domain *sd = tl->init(tl, cpu);
6389         if (!sd)
6390                 return child;
6391 
6392         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6393         if (child) {
6394                 sd->level = child->level + 1;
6395                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6396                 child->parent = sd;
6397                 sd->child = child;
6398         }
6399         set_domain_attribute(sd, attr);
6400 
6401         return sd;
6402 }
6403 
6404 /*
6405  * Build sched domains for a given set of cpus and attach the sched domains
6406  * to the individual cpus
6407  */
6408 static int build_sched_domains(const struct cpumask *cpu_map,
6409                                struct sched_domain_attr *attr)
6410 {
6411         enum s_alloc alloc_state;
6412         struct sched_domain *sd;
6413         struct s_data d;
6414         int i, ret = -ENOMEM;
6415 
6416         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6417         if (alloc_state != sa_rootdomain)
6418                 goto error;
6419 
6420         /* Set up domains for cpus specified by the cpu_map. */
6421         for_each_cpu(i, cpu_map) {
6422                 struct sched_domain_topology_level *tl;
6423 
6424                 sd = NULL;
6425                 for_each_sd_topology(tl) {
6426                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6427                         if (tl == sched_domain_topology)
6428                                 *per_cpu_ptr(d.sd, i) = sd;
6429                         if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6430                                 sd->flags |= SD_OVERLAP;
6431                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6432                                 break;
6433                 }
6434         }
6435 
6436         /* Build the groups for the domains */
6437         for_each_cpu(i, cpu_map) {
6438                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6439                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
6440                         if (sd->flags & SD_OVERLAP) {
6441                                 if (build_overlap_sched_groups(sd, i))
6442                                         goto error;
6443                         } else {
6444                                 if (build_sched_groups(sd, i))
6445                                         goto error;
6446                         }
6447                 }
6448         }
6449 
6450         /* Calculate CPU power for physical packages and nodes */
6451         for (i = nr_cpumask_bits-1; i >= 0; i--) {
6452                 if (!cpumask_test_cpu(i, cpu_map))
6453                         continue;
6454 
6455                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6456                         claim_allocations(i, sd);
6457                         init_sched_groups_power(i, sd);
6458                 }
6459         }
6460 
6461         /* Attach the domains */
6462         rcu_read_lock();
6463         for_each_cpu(i, cpu_map) {
6464                 sd = *per_cpu_ptr(d.sd, i);
6465                 cpu_attach_domain(sd, d.rd, i);
6466         }
6467         rcu_read_unlock();
6468 
6469         ret = 0;
6470 error:
6471         __free_domain_allocs(&d, alloc_state, cpu_map);
6472         return ret;
6473 }
6474 
6475 static cpumask_var_t *doms_cur; /* current sched domains */
6476 static int ndoms_cur;           /* number of sched domains in 'doms_cur' */
6477 static struct sched_domain_attr *dattr_cur;
6478                                 /* attribues of custom domains in 'doms_cur' */
6479 
6480 /*
6481  * Special case: If a kmalloc of a doms_cur partition (array of
6482  * cpumask) fails, then fallback to a single sched domain,
6483  * as determined by the single cpumask fallback_doms.
6484  */
6485 static cpumask_var_t fallback_doms;
6486 
6487 /*
6488  * arch_update_cpu_topology lets virtualized architectures update the
6489  * cpu core maps. It is supposed to return 1 if the topology changed
6490  * or 0 if it stayed the same.
6491  */
6492 int __weak arch_update_cpu_topology(void)
6493 {
6494         return 0;
6495 }
6496 
6497 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6498 {
6499         int i;
6500         cpumask_var_t *doms;
6501 
6502         doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6503         if (!doms)
6504                 return NULL;
6505         for (i = 0; i < ndoms; i++) {
6506                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6507                         free_sched_domains(doms, i);
6508                         return NULL;
6509                 }
6510         }
6511         return doms;
6512 }
6513 
6514 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6515 {
6516         unsigned int i;
6517         for (i = 0; i < ndoms; i++)
6518                 free_cpumask_var(doms[i]);
6519         kfree(doms);
6520 }
6521 
6522 /*
6523  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6524  * For now this just excludes isolated cpus, but could be used to
6525  * exclude other special cases in the future.
6526  */
6527 static int init_sched_domains(const struct cpumask *cpu_map)
6528 {
6529         int err;
6530 
6531         arch_update_cpu_topology();
6532         ndoms_cur = 1;
6533         doms_cur = alloc_sched_domains(ndoms_cur);
6534         if (!doms_cur)
6535                 doms_cur = &fallback_doms;
6536         cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6537         err = build_sched_domains(doms_cur[0], NULL);
6538         register_sched_domain_sysctl();
6539 
6540         return err;
6541 }
6542 
6543 /*
6544  * Detach sched domains from a group of cpus specified in cpu_map
6545  * These cpus will now be attached to the NULL domain
6546  */
6547 static void detach_destroy_domains(const struct cpumask *cpu_map)
6548 {
6549         int i;
6550 
6551         rcu_read_lock();
6552         for_each_cpu(i, cpu_map)
6553                 cpu_attach_domain(NULL, &def_root_domain, i);
6554         rcu_read_unlock();
6555 }
6556 
6557 /* handle null as "default" */
6558 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6559                         struct sched_domain_attr *new, int idx_new)
6560 {
6561         struct sched_domain_attr tmp;
6562 
6563         /* fast path */
6564         if (!new && !cur)
6565                 return 1;
6566 
6567         tmp = SD_ATTR_INIT;
6568         return !memcmp(cur ? (cur + idx_cur) : &tmp,
6569                         new ? (new + idx_new) : &tmp,
6570                         sizeof(struct sched_domain_attr));
6571 }
6572 
6573 /*
6574  * Partition sched domains as specified by the 'ndoms_new'
6575  * cpumasks in the array doms_new[] of cpumasks. This compares
6576  * doms_new[] to the current sched domain partitioning, doms_cur[].
6577  * It destroys each deleted domain and builds each new domain.
6578  *
6579  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6580  * The masks don't intersect (don't overlap.) We should setup one
6581  * sched domain for each mask. CPUs not in any of the cpumasks will
6582  * not be load balanced. If the same cpumask appears both in the
6583  * current 'doms_cur' domains and in the new 'doms_new', we can leave
6584  * it as it is.
6585  *
6586  * The passed in 'doms_new' should be allocated using
6587  * alloc_sched_domains.  This routine takes ownership of it and will
6588  * free_sched_domains it when done with it. If the caller failed the
6589  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6590  * and partition_sched_domains() will fallback to the single partition
6591  * 'fallback_doms', it also forces the domains to be rebuilt.
6592  *
6593  * If doms_new == NULL it will be replaced with cpu_online_mask.
6594  * ndoms_new == 0 is a special case for destroying existing domains,
6595  * and it will not create the default domain.
6596  *
6597  * Call with hotplug lock held
6598  */
6599 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6600                              struct sched_domain_attr *dattr_new)
6601 {
6602         int i, j, n;
6603         int new_topology;
6604 
6605         mutex_lock(&sched_domains_mutex);
6606 
6607         /* always unregister in case we don't destroy any domains */
6608         unregister_sched_domain_sysctl();
6609 
6610         /* Let architecture update cpu core mappings. */
6611         new_topology = arch_update_cpu_topology();
6612 
6613         n = doms_new ? ndoms_new : 0;
6614 
6615         /* Destroy deleted domains */
6616         for (i = 0; i < ndoms_cur; i++) {
6617                 for (j = 0; j < n && !new_topology; j++) {
6618                         if (cpumask_equal(doms_cur[i], doms_new[j])
6619                             && dattrs_equal(dattr_cur, i, dattr_new, j))
6620                                 goto match1;
6621                 }
6622                 /* no match - a current sched domain not in new doms_new[] */
6623                 detach_destroy_domains(doms_cur[i]);
6624 match1:
6625                 ;
6626         }
6627 
6628         n = ndoms_cur;
6629         if (doms_new == NULL) {
6630                 n = 0;
6631                 doms_new = &fallback_doms;
6632                 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6633                 WARN_ON_ONCE(dattr_new);
6634         }
6635 
6636         /* Build new domains */
6637         for (i = 0; i < ndoms_new; i++) {
6638                 for (j = 0; j < n && !new_topology; j++) {
6639                         if (cpumask_equal(doms_new[i], doms_cur[j])
6640                             && dattrs_equal(dattr_new, i, dattr_cur, j))
6641                                 goto match2;
6642                 }
6643                 /* no match - add a new doms_new */
6644                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6645 match2:
6646                 ;
6647         }
6648 
6649         /* Remember the new sched domains */
6650         if (doms_cur != &fallback_doms)
6651                 free_sched_domains(doms_cur, ndoms_cur);
6652         kfree(dattr_cur);       /* kfree(NULL) is safe */
6653         doms_cur = doms_new;
6654         dattr_cur = dattr_new;
6655         ndoms_cur = ndoms_new;
6656 
6657         register_sched_domain_sysctl();
6658 
6659         mutex_unlock(&sched_domains_mutex);
6660 }
6661 
6662 static int num_cpus_frozen;     /* used to mark begin/end of suspend/resume */
6663 
6664 /*
6665  * Update cpusets according to cpu_active mask.  If cpusets are
6666  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6667  * around partition_sched_domains().
6668  *
6669  * If we come here as part of a suspend/resume, don't touch cpusets because we
6670  * want to restore it back to its original state upon resume anyway.
6671  */
6672 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6673                              void *hcpu)
6674 {
6675         switch (action) {
6676         case CPU_ONLINE_FROZEN:
6677         case CPU_DOWN_FAILED_FROZEN:
6678 
6679                 /*
6680                  * num_cpus_frozen tracks how many CPUs are involved in suspend
6681                  * resume sequence. As long as this is not the last online
6682                  * operation in the resume sequence, just build a single sched
6683                  * domain, ignoring cpusets.
6684                  */
6685                 num_cpus_frozen--;
6686                 if (likely(num_cpus_frozen)) {
6687                         partition_sched_domains(1, NULL, NULL);
6688                         break;
6689                 }
6690 
6691                 /*
6692                  * This is the last CPU online operation. So fall through and
6693                  * restore the original sched domains by considering the
6694                  * cpuset configurations.
6695                  */
6696 
6697         case CPU_ONLINE:
6698         case CPU_DOWN_FAILED:
6699                 cpuset_update_active_cpus(true);
6700                 break;
6701         default:
6702                 return NOTIFY_DONE;
6703         }
6704         return NOTIFY_OK;
6705 }
6706 
6707 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6708                                void *hcpu)
6709 {
6710         switch (action) {
6711         case CPU_DOWN_PREPARE:
6712                 cpuset_update_active_cpus(false);
6713                 break;
6714         case CPU_DOWN_PREPARE_FROZEN:
6715                 num_cpus_frozen++;
6716                 partition_sched_domains(1, NULL, NULL);
6717                 break;
6718         default:
6719                 return NOTIFY_DONE;
6720         }
6721         return NOTIFY_OK;
6722 }
6723 
6724 void __init sched_init_smp(void)
6725 {
6726         cpumask_var_t non_isolated_cpus;
6727 
6728         alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6729         alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6730 
6731         sched_init_numa();
6732 
6733         /*
6734          * There's no userspace yet to cause hotplug operations; hence all the
6735          * cpu masks are stable and all blatant races in the below code cannot
6736          * happen.
6737          */
6738         mutex_lock(&sched_domains_mutex);
6739         init_sched_domains(cpu_active_mask);
6740         cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6741         if (cpumask_empty(non_isolated_cpus))
6742                 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6743         mutex_unlock(&sched_domains_mutex);
6744 
6745         hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6746         hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6747         hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6748 
6749         init_hrtick();
6750 
6751         /* Move init over to a non-isolated CPU */
6752         if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6753                 BUG();
6754         sched_init_granularity();
6755         free_cpumask_var(non_isolated_cpus);
6756 
6757         init_sched_rt_class();
6758         init_sched_dl_class();
6759 }
6760 #else
6761 void __init sched_init_smp(void)
6762 {
6763         sched_init_granularity();
6764 }
6765 #endif /* CONFIG_SMP */
6766 
6767 const_debug unsigned int sysctl_timer_migration = 1;
6768 
6769 int in_sched_functions(unsigned long addr)
6770 {
6771         return in_lock_functions(addr) ||
6772                 (addr >= (unsigned long)__sched_text_start
6773                 && addr < (unsigned long)__sched_text_end);
6774 }
6775 
6776 #ifdef CONFIG_CGROUP_SCHED
6777 /*
6778  * Default task group.
6779  * Every task in system belongs to this group at bootup.
6780  */
6781 struct task_group root_task_group;
6782 LIST_HEAD(task_groups);
6783 #endif
6784 
6785 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6786 
6787 void __init sched_init(void)
6788 {
6789         int i, j;
6790         unsigned long alloc_size = 0, ptr;
6791 
6792 #ifdef CONFIG_FAIR_GROUP_SCHED
6793         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6794 #endif
6795 #ifdef CONFIG_RT_GROUP_SCHED
6796         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6797 #endif
6798 #ifdef CONFIG_CPUMASK_OFFSTACK
6799         alloc_size += num_possible_cpus() * cpumask_size();
6800 #endif
6801         if (alloc_size) {
6802                 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6803 
6804 #ifdef CONFIG_FAIR_GROUP_SCHED
6805                 root_task_group.se = (struct sched_entity **)ptr;
6806                 ptr += nr_cpu_ids * sizeof(void **);
6807 
6808                 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6809                 ptr += nr_cpu_ids * sizeof(void **);
6810 
6811 #endif /* CONFIG_FAIR_GROUP_SCHED */
6812 #ifdef CONFIG_RT_GROUP_SCHED
6813                 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6814                 ptr += nr_cpu_ids * sizeof(void **);
6815 
6816                 root_task_group.rt_rq = (struct rt_rq **)ptr;
6817                 ptr += nr_cpu_ids * sizeof(void **);
6818 
6819 #endif /* CONFIG_RT_GROUP_SCHED */
6820 #ifdef CONFIG_CPUMASK_OFFSTACK
6821                 for_each_possible_cpu(i) {
6822                         per_cpu(load_balance_mask, i) = (void *)ptr;
6823                         ptr += cpumask_size