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

Linux/kernel/sched/core.c

  1 /*
  2  *  kernel/sched/core.c
  3  *
  4  *  Kernel scheduler and related syscalls
  5  *
  6  *  Copyright (C) 1991-2002  Linus Torvalds
  7  *
  8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
  9  *              make semaphores SMP safe
 10  *  1998-11-19  Implemented schedule_timeout() and related stuff
 11  *              by Andrea Arcangeli
 12  *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
 13  *              hybrid priority-list and round-robin design with
 14  *              an array-switch method of distributing timeslices
 15  *              and per-CPU runqueues.  Cleanups and useful suggestions
 16  *              by Davide Libenzi, preemptible kernel bits by Robert Love.
 17  *  2003-09-03  Interactivity tuning by Con Kolivas.
 18  *  2004-04-02  Scheduler domains code by Nick Piggin
 19  *  2007-04-15  Work begun on replacing all interactivity tuning with a
 20  *              fair scheduling design by Con Kolivas.
 21  *  2007-05-05  Load balancing (smp-nice) and other improvements
 22  *              by Peter Williams
 23  *  2007-05-06  Interactivity improvements to CFS by Mike Galbraith
 24  *  2007-07-01  Group scheduling enhancements by Srivatsa Vaddagiri
 25  *  2007-11-29  RT balancing improvements by Steven Rostedt, Gregory Haskins,
 26  *              Thomas Gleixner, Mike Kravetz
 27  */
 28 
 29 #include <linux/mm.h>
 30 #include <linux/module.h>
 31 #include <linux/nmi.h>
 32 #include <linux/init.h>
 33 #include <linux/uaccess.h>
 34 #include <linux/highmem.h>
 35 #include <asm/mmu_context.h>
 36 #include <linux/interrupt.h>
 37 #include <linux/capability.h>
 38 #include <linux/completion.h>
 39 #include <linux/kernel_stat.h>
 40 #include <linux/debug_locks.h>
 41 #include <linux/perf_event.h>
 42 #include <linux/security.h>
 43 #include <linux/notifier.h>
 44 #include <linux/profile.h>
 45 #include <linux/freezer.h>
 46 #include <linux/vmalloc.h>
 47 #include <linux/blkdev.h>
 48 #include <linux/delay.h>
 49 #include <linux/pid_namespace.h>
 50 #include <linux/smp.h>
 51 #include <linux/threads.h>
 52 #include <linux/timer.h>
 53 #include <linux/rcupdate.h>
 54 #include <linux/cpu.h>
 55 #include <linux/cpuset.h>
 56 #include <linux/percpu.h>
 57 #include <linux/proc_fs.h>
 58 #include <linux/seq_file.h>
 59 #include <linux/sysctl.h>
 60 #include <linux/syscalls.h>
 61 #include <linux/times.h>
 62 #include <linux/tsacct_kern.h>
 63 #include <linux/kprobes.h>
 64 #include <linux/delayacct.h>
 65 #include <linux/unistd.h>
 66 #include <linux/pagemap.h>
 67 #include <linux/hrtimer.h>
 68 #include <linux/tick.h>
 69 #include <linux/debugfs.h>
 70 #include <linux/ctype.h>
 71 #include <linux/ftrace.h>
 72 #include <linux/slab.h>
 73 #include <linux/init_task.h>
 74 #include <linux/binfmts.h>
 75 #include <linux/context_tracking.h>
 76 
 77 #include <asm/switch_to.h>
 78 #include <asm/tlb.h>
 79 #include <asm/irq_regs.h>
 80 #include <asm/mutex.h>
 81 #ifdef CONFIG_PARAVIRT
 82 #include <asm/paravirt.h>
 83 #endif
 84 
 85 #include "sched.h"
 86 #include "../workqueue_internal.h"
 87 #include "../smpboot.h"
 88 
 89 #define CREATE_TRACE_POINTS
 90 #include <trace/events/sched.h>
 91 
 92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
 93 {
 94         unsigned long delta;
 95         ktime_t soft, hard, now;
 96 
 97         for (;;) {
 98                 if (hrtimer_active(period_timer))
 99                         break;
100 
101                 now = hrtimer_cb_get_time(period_timer);
102                 hrtimer_forward(period_timer, now, period);
103 
104                 soft = hrtimer_get_softexpires(period_timer);
105                 hard = hrtimer_get_expires(period_timer);
106                 delta = ktime_to_ns(ktime_sub(hard, soft));
107                 __hrtimer_start_range_ns(period_timer, soft, delta,
108                                          HRTIMER_MODE_ABS_PINNED, 0);
109         }
110 }
111 
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
114 
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
116 
117 void update_rq_clock(struct rq *rq)
118 {
119         s64 delta;
120 
121         if (rq->skip_clock_update > 0)
122                 return;
123 
124         delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
125         rq->clock += delta;
126         update_rq_clock_task(rq, delta);
127 }
128 
129 /*
130  * Debugging: various feature bits
131  */
132 
133 #define SCHED_FEAT(name, enabled)       \
134         (1UL << __SCHED_FEAT_##name) * enabled |
135 
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
138         0;
139 
140 #undef SCHED_FEAT
141 
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled)       \
144         #name ,
145 
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
148 };
149 
150 #undef SCHED_FEAT
151 
152 static int sched_feat_show(struct seq_file *m, void *v)
153 {
154         int i;
155 
156         for (i = 0; i < __SCHED_FEAT_NR; i++) {
157                 if (!(sysctl_sched_features & (1UL << i)))
158                         seq_puts(m, "NO_");
159                 seq_printf(m, "%s ", sched_feat_names[i]);
160         }
161         seq_puts(m, "\n");
162 
163         return 0;
164 }
165 
166 #ifdef HAVE_JUMP_LABEL
167 
168 #define jump_label_key__true  STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
170 
171 #define SCHED_FEAT(name, enabled)       \
172         jump_label_key__##enabled ,
173 
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
176 };
177 
178 #undef SCHED_FEAT
179 
180 static void sched_feat_disable(int i)
181 {
182         if (static_key_enabled(&sched_feat_keys[i]))
183                 static_key_slow_dec(&sched_feat_keys[i]);
184 }
185 
186 static void sched_feat_enable(int i)
187 {
188         if (!static_key_enabled(&sched_feat_keys[i]))
189                 static_key_slow_inc(&sched_feat_keys[i]);
190 }
191 #else
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
195 
196 static int sched_feat_set(char *cmp)
197 {
198         int i;
199         int neg = 0;
200 
201         if (strncmp(cmp, "NO_", 3) == 0) {
202                 neg = 1;
203                 cmp += 3;
204         }
205 
206         for (i = 0; i < __SCHED_FEAT_NR; i++) {
207                 if (strcmp(cmp, sched_feat_names[i]) == 0) {
208                         if (neg) {
209                                 sysctl_sched_features &= ~(1UL << i);
210                                 sched_feat_disable(i);
211                         } else {
212                                 sysctl_sched_features |= (1UL << i);
213                                 sched_feat_enable(i);
214                         }
215                         break;
216                 }
217         }
218 
219         return i;
220 }
221 
222 static ssize_t
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224                 size_t cnt, loff_t *ppos)
225 {
226         char buf[64];
227         char *cmp;
228         int i;
229 
230         if (cnt > 63)
231                 cnt = 63;
232 
233         if (copy_from_user(&buf, ubuf, cnt))
234                 return -EFAULT;
235 
236         buf[cnt] = 0;
237         cmp = strstrip(buf);
238 
239         i = sched_feat_set(cmp);
240         if (i == __SCHED_FEAT_NR)
241                 return -EINVAL;
242 
243         *ppos += cnt;
244 
245         return cnt;
246 }
247 
248 static int sched_feat_open(struct inode *inode, struct file *filp)
249 {
250         return single_open(filp, sched_feat_show, NULL);
251 }
252 
253 static const struct file_operations sched_feat_fops = {
254         .open           = sched_feat_open,
255         .write          = sched_feat_write,
256         .read           = seq_read,
257         .llseek         = seq_lseek,
258         .release        = single_release,
259 };
260 
261 static __init int sched_init_debug(void)
262 {
263         debugfs_create_file("sched_features", 0644, NULL, NULL,
264                         &sched_feat_fops);
265 
266         return 0;
267 }
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
270 
271 /*
272  * Number of tasks to iterate in a single balance run.
273  * Limited because this is done with IRQs disabled.
274  */
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
276 
277 /*
278  * period over which we average the RT time consumption, measured
279  * in ms.
280  *
281  * default: 1s
282  */
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
284 
285 /*
286  * period over which we measure -rt task cpu usage in us.
287  * default: 1s
288  */
289 unsigned int sysctl_sched_rt_period = 1000000;
290 
291 __read_mostly int scheduler_running;
292 
293 /*
294  * part of the period that we allow rt tasks to run in us.
295  * default: 0.95s
296  */
297 int sysctl_sched_rt_runtime = 950000;
298 
299 /*
300  * __task_rq_lock - lock the rq @p resides on.
301  */
302 static inline struct rq *__task_rq_lock(struct task_struct *p)
303         __acquires(rq->lock)
304 {
305         struct rq *rq;
306 
307         lockdep_assert_held(&p->pi_lock);
308 
309         for (;;) {
310                 rq = task_rq(p);
311                 raw_spin_lock(&rq->lock);
312                 if (likely(rq == task_rq(p)))
313                         return rq;
314                 raw_spin_unlock(&rq->lock);
315         }
316 }
317 
318 /*
319  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
320  */
321 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
322         __acquires(p->pi_lock)
323         __acquires(rq->lock)
324 {
325         struct rq *rq;
326 
327         for (;;) {
328                 raw_spin_lock_irqsave(&p->pi_lock, *flags);
329                 rq = task_rq(p);
330                 raw_spin_lock(&rq->lock);
331                 if (likely(rq == task_rq(p)))
332                         return rq;
333                 raw_spin_unlock(&rq->lock);
334                 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
335         }
336 }
337 
338 static void __task_rq_unlock(struct rq *rq)
339         __releases(rq->lock)
340 {
341         raw_spin_unlock(&rq->lock);
342 }
343 
344 static inline void
345 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
346         __releases(rq->lock)
347         __releases(p->pi_lock)
348 {
349         raw_spin_unlock(&rq->lock);
350         raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
351 }
352 
353 /*
354  * this_rq_lock - lock this runqueue and disable interrupts.
355  */
356 static struct rq *this_rq_lock(void)
357         __acquires(rq->lock)
358 {
359         struct rq *rq;
360 
361         local_irq_disable();
362         rq = this_rq();
363         raw_spin_lock(&rq->lock);
364 
365         return rq;
366 }
367 
368 #ifdef CONFIG_SCHED_HRTICK
369 /*
370  * Use HR-timers to deliver accurate preemption points.
371  */
372 
373 static void hrtick_clear(struct rq *rq)
374 {
375         if (hrtimer_active(&rq->hrtick_timer))
376                 hrtimer_cancel(&rq->hrtick_timer);
377 }
378 
379 /*
380  * High-resolution timer tick.
381  * Runs from hardirq context with interrupts disabled.
382  */
383 static enum hrtimer_restart hrtick(struct hrtimer *timer)
384 {
385         struct rq *rq = container_of(timer, struct rq, hrtick_timer);
386 
387         WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
388 
389         raw_spin_lock(&rq->lock);
390         update_rq_clock(rq);
391         rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392         raw_spin_unlock(&rq->lock);
393 
394         return HRTIMER_NORESTART;
395 }
396 
397 #ifdef CONFIG_SMP
398 
399 static int __hrtick_restart(struct rq *rq)
400 {
401         struct hrtimer *timer = &rq->hrtick_timer;
402         ktime_t time = hrtimer_get_softexpires(timer);
403 
404         return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
405 }
406 
407 /*
408  * called from hardirq (IPI) context
409  */
410 static void __hrtick_start(void *arg)
411 {
412         struct rq *rq = arg;
413 
414         raw_spin_lock(&rq->lock);
415         __hrtick_restart(rq);
416         rq->hrtick_csd_pending = 0;
417         raw_spin_unlock(&rq->lock);
418 }
419 
420 /*
421  * Called to set the hrtick timer state.
422  *
423  * called with rq->lock held and irqs disabled
424  */
425 void hrtick_start(struct rq *rq, u64 delay)
426 {
427         struct hrtimer *timer = &rq->hrtick_timer;
428         ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
429 
430         hrtimer_set_expires(timer, time);
431 
432         if (rq == this_rq()) {
433                 __hrtick_restart(rq);
434         } else if (!rq->hrtick_csd_pending) {
435                 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
436                 rq->hrtick_csd_pending = 1;
437         }
438 }
439 
440 static int
441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
442 {
443         int cpu = (int)(long)hcpu;
444 
445         switch (action) {
446         case CPU_UP_CANCELED:
447         case CPU_UP_CANCELED_FROZEN:
448         case CPU_DOWN_PREPARE:
449         case CPU_DOWN_PREPARE_FROZEN:
450         case CPU_DEAD:
451         case CPU_DEAD_FROZEN:
452                 hrtick_clear(cpu_rq(cpu));
453                 return NOTIFY_OK;
454         }
455 
456         return NOTIFY_DONE;
457 }
458 
459 static __init void init_hrtick(void)
460 {
461         hotcpu_notifier(hotplug_hrtick, 0);
462 }
463 #else
464 /*
465  * Called to set the hrtick timer state.
466  *
467  * called with rq->lock held and irqs disabled
468  */
469 void hrtick_start(struct rq *rq, u64 delay)
470 {
471         __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
472                         HRTIMER_MODE_REL_PINNED, 0);
473 }
474 
475 static inline void init_hrtick(void)
476 {
477 }
478 #endif /* CONFIG_SMP */
479 
480 static void init_rq_hrtick(struct rq *rq)
481 {
482 #ifdef CONFIG_SMP
483         rq->hrtick_csd_pending = 0;
484 
485         rq->hrtick_csd.flags = 0;
486         rq->hrtick_csd.func = __hrtick_start;
487         rq->hrtick_csd.info = rq;
488 #endif
489 
490         hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
491         rq->hrtick_timer.function = hrtick;
492 }
493 #else   /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq *rq)
495 {
496 }
497 
498 static inline void init_rq_hrtick(struct rq *rq)
499 {
500 }
501 
502 static inline void init_hrtick(void)
503 {
504 }
505 #endif  /* CONFIG_SCHED_HRTICK */
506 
507 /*
508  * resched_task - mark a task 'to be rescheduled now'.
509  *
510  * On UP this means the setting of the need_resched flag, on SMP it
511  * might also involve a cross-CPU call to trigger the scheduler on
512  * the target CPU.
513  */
514 void resched_task(struct task_struct *p)
515 {
516         int cpu;
517 
518         lockdep_assert_held(&task_rq(p)->lock);
519 
520         if (test_tsk_need_resched(p))
521                 return;
522 
523         set_tsk_need_resched(p);
524 
525         cpu = task_cpu(p);
526         if (cpu == smp_processor_id()) {
527                 set_preempt_need_resched();
528                 return;
529         }
530 
531         /* NEED_RESCHED must be visible before we test polling */
532         smp_mb();
533         if (!tsk_is_polling(p))
534                 smp_send_reschedule(cpu);
535 }
536 
537 void resched_cpu(int cpu)
538 {
539         struct rq *rq = cpu_rq(cpu);
540         unsigned long flags;
541 
542         if (!raw_spin_trylock_irqsave(&rq->lock, flags))
543                 return;
544         resched_task(cpu_curr(cpu));
545         raw_spin_unlock_irqrestore(&rq->lock, flags);
546 }
547 
548 #ifdef CONFIG_SMP
549 #ifdef CONFIG_NO_HZ_COMMON
550 /*
551  * In the semi idle case, use the nearest busy cpu for migrating timers
552  * from an idle cpu.  This is good for power-savings.
553  *
554  * We don't do similar optimization for completely idle system, as
555  * selecting an idle cpu will add more delays to the timers than intended
556  * (as that cpu's timer base may not be uptodate wrt jiffies etc).
557  */
558 int get_nohz_timer_target(void)
559 {
560         int cpu = smp_processor_id();
561         int i;
562         struct sched_domain *sd;
563 
564         rcu_read_lock();
565         for_each_domain(cpu, sd) {
566                 for_each_cpu(i, sched_domain_span(sd)) {
567                         if (!idle_cpu(i)) {
568                                 cpu = i;
569                                 goto unlock;
570                         }
571                 }
572         }
573 unlock:
574         rcu_read_unlock();
575         return cpu;
576 }
577 /*
578  * When add_timer_on() enqueues a timer into the timer wheel of an
579  * idle CPU then this timer might expire before the next timer event
580  * which is scheduled to wake up that CPU. In case of a completely
581  * idle system the next event might even be infinite time into the
582  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
583  * leaves the inner idle loop so the newly added timer is taken into
584  * account when the CPU goes back to idle and evaluates the timer
585  * wheel for the next timer event.
586  */
587 static void wake_up_idle_cpu(int cpu)
588 {
589         struct rq *rq = cpu_rq(cpu);
590 
591         if (cpu == smp_processor_id())
592                 return;
593 
594         /*
595          * This is safe, as this function is called with the timer
596          * wheel base lock of (cpu) held. When the CPU is on the way
597          * to idle and has not yet set rq->curr to idle then it will
598          * be serialized on the timer wheel base lock and take the new
599          * timer into account automatically.
600          */
601         if (rq->curr != rq->idle)
602                 return;
603 
604         /*
605          * We can set TIF_RESCHED on the idle task of the other CPU
606          * lockless. The worst case is that the other CPU runs the
607          * idle task through an additional NOOP schedule()
608          */
609         set_tsk_need_resched(rq->idle);
610 
611         /* NEED_RESCHED must be visible before we test polling */
612         smp_mb();
613         if (!tsk_is_polling(rq->idle))
614                 smp_send_reschedule(cpu);
615 }
616 
617 static bool wake_up_full_nohz_cpu(int cpu)
618 {
619         if (tick_nohz_full_cpu(cpu)) {
620                 if (cpu != smp_processor_id() ||
621                     tick_nohz_tick_stopped())
622                         smp_send_reschedule(cpu);
623                 return true;
624         }
625 
626         return false;
627 }
628 
629 void wake_up_nohz_cpu(int cpu)
630 {
631         if (!wake_up_full_nohz_cpu(cpu))
632                 wake_up_idle_cpu(cpu);
633 }
634 
635 static inline bool got_nohz_idle_kick(void)
636 {
637         int cpu = smp_processor_id();
638 
639         if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
640                 return false;
641 
642         if (idle_cpu(cpu) && !need_resched())
643                 return true;
644 
645         /*
646          * We can't run Idle Load Balance on this CPU for this time so we
647          * cancel it and clear NOHZ_BALANCE_KICK
648          */
649         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
650         return false;
651 }
652 
653 #else /* CONFIG_NO_HZ_COMMON */
654 
655 static inline bool got_nohz_idle_kick(void)
656 {
657         return false;
658 }
659 
660 #endif /* CONFIG_NO_HZ_COMMON */
661 
662 #ifdef CONFIG_NO_HZ_FULL
663 bool sched_can_stop_tick(void)
664 {
665        struct rq *rq;
666 
667        rq = this_rq();
668 
669        /* Make sure rq->nr_running update is visible after the IPI */
670        smp_rmb();
671 
672        /* More than one running task need preemption */
673        if (rq->nr_running > 1)
674                return false;
675 
676        return true;
677 }
678 #endif /* CONFIG_NO_HZ_FULL */
679 
680 void sched_avg_update(struct rq *rq)
681 {
682         s64 period = sched_avg_period();
683 
684         while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
685                 /*
686                  * Inline assembly required to prevent the compiler
687                  * optimising this loop into a divmod call.
688                  * See __iter_div_u64_rem() for another example of this.
689                  */
690                 asm("" : "+rm" (rq->age_stamp));
691                 rq->age_stamp += period;
692                 rq->rt_avg /= 2;
693         }
694 }
695 
696 #endif /* CONFIG_SMP */
697 
698 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
699                         (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
700 /*
701  * Iterate task_group tree rooted at *from, calling @down when first entering a
702  * node and @up when leaving it for the final time.
703  *
704  * Caller must hold rcu_lock or sufficient equivalent.
705  */
706 int walk_tg_tree_from(struct task_group *from,
707                              tg_visitor down, tg_visitor up, void *data)
708 {
709         struct task_group *parent, *child;
710         int ret;
711 
712         parent = from;
713 
714 down:
715         ret = (*down)(parent, data);
716         if (ret)
717                 goto out;
718         list_for_each_entry_rcu(child, &parent->children, siblings) {
719                 parent = child;
720                 goto down;
721 
722 up:
723                 continue;
724         }
725         ret = (*up)(parent, data);
726         if (ret || parent == from)
727                 goto out;
728 
729         child = parent;
730         parent = parent->parent;
731         if (parent)
732                 goto up;
733 out:
734         return ret;
735 }
736 
737 int tg_nop(struct task_group *tg, void *data)
738 {
739         return 0;
740 }
741 #endif
742 
743 static void set_load_weight(struct task_struct *p)
744 {
745         int prio = p->static_prio - MAX_RT_PRIO;
746         struct load_weight *load = &p->se.load;
747 
748         /*
749          * SCHED_IDLE tasks get minimal weight:
750          */
751         if (p->policy == SCHED_IDLE) {
752                 load->weight = scale_load(WEIGHT_IDLEPRIO);
753                 load->inv_weight = WMULT_IDLEPRIO;
754                 return;
755         }
756 
757         load->weight = scale_load(prio_to_weight[prio]);
758         load->inv_weight = prio_to_wmult[prio];
759 }
760 
761 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
762 {
763         update_rq_clock(rq);
764         sched_info_queued(rq, p);
765         p->sched_class->enqueue_task(rq, p, flags);
766 }
767 
768 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
769 {
770         update_rq_clock(rq);
771         sched_info_dequeued(rq, p);
772         p->sched_class->dequeue_task(rq, p, flags);
773 }
774 
775 void activate_task(struct rq *rq, struct task_struct *p, int flags)
776 {
777         if (task_contributes_to_load(p))
778                 rq->nr_uninterruptible--;
779 
780         enqueue_task(rq, p, flags);
781 }
782 
783 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
784 {
785         if (task_contributes_to_load(p))
786                 rq->nr_uninterruptible++;
787 
788         dequeue_task(rq, p, flags);
789 }
790 
791 static void update_rq_clock_task(struct rq *rq, s64 delta)
792 {
793 /*
794  * In theory, the compile should just see 0 here, and optimize out the call
795  * to sched_rt_avg_update. But I don't trust it...
796  */
797 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
798         s64 steal = 0, irq_delta = 0;
799 #endif
800 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
801         irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
802 
803         /*
804          * Since irq_time is only updated on {soft,}irq_exit, we might run into
805          * this case when a previous update_rq_clock() happened inside a
806          * {soft,}irq region.
807          *
808          * When this happens, we stop ->clock_task and only update the
809          * prev_irq_time stamp to account for the part that fit, so that a next
810          * update will consume the rest. This ensures ->clock_task is
811          * monotonic.
812          *
813          * It does however cause some slight miss-attribution of {soft,}irq
814          * time, a more accurate solution would be to update the irq_time using
815          * the current rq->clock timestamp, except that would require using
816          * atomic ops.
817          */
818         if (irq_delta > delta)
819                 irq_delta = delta;
820 
821         rq->prev_irq_time += irq_delta;
822         delta -= irq_delta;
823 #endif
824 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
825         if (static_key_false((&paravirt_steal_rq_enabled))) {
826                 u64 st;
827 
828                 steal = paravirt_steal_clock(cpu_of(rq));
829                 steal -= rq->prev_steal_time_rq;
830 
831                 if (unlikely(steal > delta))
832                         steal = delta;
833 
834                 st = steal_ticks(steal);
835                 steal = st * TICK_NSEC;
836 
837                 rq->prev_steal_time_rq += steal;
838 
839                 delta -= steal;
840         }
841 #endif
842 
843         rq->clock_task += delta;
844 
845 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
846         if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
847                 sched_rt_avg_update(rq, irq_delta + steal);
848 #endif
849 }
850 
851 void sched_set_stop_task(int cpu, struct task_struct *stop)
852 {
853         struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
854         struct task_struct *old_stop = cpu_rq(cpu)->stop;
855 
856         if (stop) {
857                 /*
858                  * Make it appear like a SCHED_FIFO task, its something
859                  * userspace knows about and won't get confused about.
860                  *
861                  * Also, it will make PI more or less work without too
862                  * much confusion -- but then, stop work should not
863                  * rely on PI working anyway.
864                  */
865                 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
866 
867                 stop->sched_class = &stop_sched_class;
868         }
869 
870         cpu_rq(cpu)->stop = stop;
871 
872         if (old_stop) {
873                 /*
874                  * Reset it back to a normal scheduling class so that
875                  * it can die in pieces.
876                  */
877                 old_stop->sched_class = &rt_sched_class;
878         }
879 }
880 
881 /*
882  * __normal_prio - return the priority that is based on the static prio
883  */
884 static inline int __normal_prio(struct task_struct *p)
885 {
886         return p->static_prio;
887 }
888 
889 /*
890  * Calculate the expected normal priority: i.e. priority
891  * without taking RT-inheritance into account. Might be
892  * boosted by interactivity modifiers. Changes upon fork,
893  * setprio syscalls, and whenever the interactivity
894  * estimator recalculates.
895  */
896 static inline int normal_prio(struct task_struct *p)
897 {
898         int prio;
899 
900         if (task_has_dl_policy(p))
901                 prio = MAX_DL_PRIO-1;
902         else if (task_has_rt_policy(p))
903                 prio = MAX_RT_PRIO-1 - p->rt_priority;
904         else
905                 prio = __normal_prio(p);
906         return prio;
907 }
908 
909 /*
910  * Calculate the current priority, i.e. the priority
911  * taken into account by the scheduler. This value might
912  * be boosted by RT tasks, or might be boosted by
913  * interactivity modifiers. Will be RT if the task got
914  * RT-boosted. If not then it returns p->normal_prio.
915  */
916 static int effective_prio(struct task_struct *p)
917 {
918         p->normal_prio = normal_prio(p);
919         /*
920          * If we are RT tasks or we were boosted to RT priority,
921          * keep the priority unchanged. Otherwise, update priority
922          * to the normal priority:
923          */
924         if (!rt_prio(p->prio))
925                 return p->normal_prio;
926         return p->prio;
927 }
928 
929 /**
930  * task_curr - is this task currently executing on a CPU?
931  * @p: the task in question.
932  *
933  * Return: 1 if the task is currently executing. 0 otherwise.
934  */
935 inline int task_curr(const struct task_struct *p)
936 {
937         return cpu_curr(task_cpu(p)) == p;
938 }
939 
940 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
941                                        const struct sched_class *prev_class,
942                                        int oldprio)
943 {
944         if (prev_class != p->sched_class) {
945                 if (prev_class->switched_from)
946                         prev_class->switched_from(rq, p);
947                 p->sched_class->switched_to(rq, p);
948         } else if (oldprio != p->prio || dl_task(p))
949                 p->sched_class->prio_changed(rq, p, oldprio);
950 }
951 
952 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
953 {
954         const struct sched_class *class;
955 
956         if (p->sched_class == rq->curr->sched_class) {
957                 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
958         } else {
959                 for_each_class(class) {
960                         if (class == rq->curr->sched_class)
961                                 break;
962                         if (class == p->sched_class) {
963                                 resched_task(rq->curr);
964                                 break;
965                         }
966                 }
967         }
968 
969         /*
970          * A queue event has occurred, and we're going to schedule.  In
971          * this case, we can save a useless back to back clock update.
972          */
973         if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
974                 rq->skip_clock_update = 1;
975 }
976 
977 #ifdef CONFIG_SMP
978 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
979 {
980 #ifdef CONFIG_SCHED_DEBUG
981         /*
982          * We should never call set_task_cpu() on a blocked task,
983          * ttwu() will sort out the placement.
984          */
985         WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
986                         !(task_preempt_count(p) & PREEMPT_ACTIVE));
987 
988 #ifdef CONFIG_LOCKDEP
989         /*
990          * The caller should hold either p->pi_lock or rq->lock, when changing
991          * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
992          *
993          * sched_move_task() holds both and thus holding either pins the cgroup,
994          * see task_group().
995          *
996          * Furthermore, all task_rq users should acquire both locks, see
997          * task_rq_lock().
998          */
999         WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1000                                       lockdep_is_held(&task_rq(p)->lock)));
1001 #endif
1002 #endif
1003 
1004         trace_sched_migrate_task(p, new_cpu);
1005 
1006         if (task_cpu(p) != new_cpu) {
1007                 if (p->sched_class->migrate_task_rq)
1008                         p->sched_class->migrate_task_rq(p, new_cpu);
1009                 p->se.nr_migrations++;
1010                 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1011         }
1012 
1013         __set_task_cpu(p, new_cpu);
1014 }
1015 
1016 static void __migrate_swap_task(struct task_struct *p, int cpu)
1017 {
1018         if (p->on_rq) {
1019                 struct rq *src_rq, *dst_rq;
1020 
1021                 src_rq = task_rq(p);
1022                 dst_rq = cpu_rq(cpu);
1023 
1024                 deactivate_task(src_rq, p, 0);
1025                 set_task_cpu(p, cpu);
1026                 activate_task(dst_rq, p, 0);
1027                 check_preempt_curr(dst_rq, p, 0);
1028         } else {
1029                 /*
1030                  * Task isn't running anymore; make it appear like we migrated
1031                  * it before it went to sleep. This means on wakeup we make the
1032                  * previous cpu our targer instead of where it really is.
1033                  */
1034                 p->wake_cpu = cpu;
1035         }
1036 }
1037 
1038 struct migration_swap_arg {
1039         struct task_struct *src_task, *dst_task;
1040         int src_cpu, dst_cpu;
1041 };
1042 
1043 static int migrate_swap_stop(void *data)
1044 {
1045         struct migration_swap_arg *arg = data;
1046         struct rq *src_rq, *dst_rq;
1047         int ret = -EAGAIN;
1048 
1049         src_rq = cpu_rq(arg->src_cpu);
1050         dst_rq = cpu_rq(arg->dst_cpu);
1051 
1052         double_raw_lock(&arg->src_task->pi_lock,
1053                         &arg->dst_task->pi_lock);
1054         double_rq_lock(src_rq, dst_rq);
1055         if (task_cpu(arg->dst_task) != arg->dst_cpu)
1056                 goto unlock;
1057 
1058         if (task_cpu(arg->src_task) != arg->src_cpu)
1059                 goto unlock;
1060 
1061         if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1062                 goto unlock;
1063 
1064         if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1065                 goto unlock;
1066 
1067         __migrate_swap_task(arg->src_task, arg->dst_cpu);
1068         __migrate_swap_task(arg->dst_task, arg->src_cpu);
1069 
1070         ret = 0;
1071 
1072 unlock:
1073         double_rq_unlock(src_rq, dst_rq);
1074         raw_spin_unlock(&arg->dst_task->pi_lock);
1075         raw_spin_unlock(&arg->src_task->pi_lock);
1076 
1077         return ret;
1078 }
1079 
1080 /*
1081  * Cross migrate two tasks
1082  */
1083 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1084 {
1085         struct migration_swap_arg arg;
1086         int ret = -EINVAL;
1087 
1088         arg = (struct migration_swap_arg){
1089                 .src_task = cur,
1090                 .src_cpu = task_cpu(cur),
1091                 .dst_task = p,
1092                 .dst_cpu = task_cpu(p),
1093         };
1094 
1095         if (arg.src_cpu == arg.dst_cpu)
1096                 goto out;
1097 
1098         /*
1099          * These three tests are all lockless; this is OK since all of them
1100          * will be re-checked with proper locks held further down the line.
1101          */
1102         if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1103                 goto out;
1104 
1105         if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1106                 goto out;
1107 
1108         if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1109                 goto out;
1110 
1111         trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1112         ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1113 
1114 out:
1115         return ret;
1116 }
1117 
1118 struct migration_arg {
1119         struct task_struct *task;
1120         int dest_cpu;
1121 };
1122 
1123 static int migration_cpu_stop(void *data);
1124 
1125 /*
1126  * wait_task_inactive - wait for a thread to unschedule.
1127  *
1128  * If @match_state is nonzero, it's the @p->state value just checked and
1129  * not expected to change.  If it changes, i.e. @p might have woken up,
1130  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1131  * we return a positive number (its total switch count).  If a second call
1132  * a short while later returns the same number, the caller can be sure that
1133  * @p has remained unscheduled the whole time.
1134  *
1135  * The caller must ensure that the task *will* unschedule sometime soon,
1136  * else this function might spin for a *long* time. This function can't
1137  * be called with interrupts off, or it may introduce deadlock with
1138  * smp_call_function() if an IPI is sent by the same process we are
1139  * waiting to become inactive.
1140  */
1141 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1142 {
1143         unsigned long flags;
1144         int running, on_rq;
1145         unsigned long ncsw;
1146         struct rq *rq;
1147 
1148         for (;;) {
1149                 /*
1150                  * We do the initial early heuristics without holding
1151                  * any task-queue locks at all. We'll only try to get
1152                  * the runqueue lock when things look like they will
1153                  * work out!
1154                  */
1155                 rq = task_rq(p);
1156 
1157                 /*
1158                  * If the task is actively running on another CPU
1159                  * still, just relax and busy-wait without holding
1160                  * any locks.
1161                  *
1162                  * NOTE! Since we don't hold any locks, it's not
1163                  * even sure that "rq" stays as the right runqueue!
1164                  * But we don't care, since "task_running()" will
1165                  * return false if the runqueue has changed and p
1166                  * is actually now running somewhere else!
1167                  */
1168                 while (task_running(rq, p)) {
1169                         if (match_state && unlikely(p->state != match_state))
1170                                 return 0;
1171                         cpu_relax();
1172                 }
1173 
1174                 /*
1175                  * Ok, time to look more closely! We need the rq
1176                  * lock now, to be *sure*. If we're wrong, we'll
1177                  * just go back and repeat.
1178                  */
1179                 rq = task_rq_lock(p, &flags);
1180                 trace_sched_wait_task(p);
1181                 running = task_running(rq, p);
1182                 on_rq = p->on_rq;
1183                 ncsw = 0;
1184                 if (!match_state || p->state == match_state)
1185                         ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1186                 task_rq_unlock(rq, p, &flags);
1187 
1188                 /*
1189                  * If it changed from the expected state, bail out now.
1190                  */
1191                 if (unlikely(!ncsw))
1192                         break;
1193 
1194                 /*
1195                  * Was it really running after all now that we
1196                  * checked with the proper locks actually held?
1197                  *
1198                  * Oops. Go back and try again..
1199                  */
1200                 if (unlikely(running)) {
1201                         cpu_relax();
1202                         continue;
1203                 }
1204 
1205                 /*
1206                  * It's not enough that it's not actively running,
1207                  * it must be off the runqueue _entirely_, and not
1208                  * preempted!
1209                  *
1210                  * So if it was still runnable (but just not actively
1211                  * running right now), it's preempted, and we should
1212                  * yield - it could be a while.
1213                  */
1214                 if (unlikely(on_rq)) {
1215                         ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1216 
1217                         set_current_state(TASK_UNINTERRUPTIBLE);
1218                         schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1219                         continue;
1220                 }
1221 
1222                 /*
1223                  * Ahh, all good. It wasn't running, and it wasn't
1224                  * runnable, which means that it will never become
1225                  * running in the future either. We're all done!
1226                  */
1227                 break;
1228         }
1229 
1230         return ncsw;
1231 }
1232 
1233 /***
1234  * kick_process - kick a running thread to enter/exit the kernel
1235  * @p: the to-be-kicked thread
1236  *
1237  * Cause a process which is running on another CPU to enter
1238  * kernel-mode, without any delay. (to get signals handled.)
1239  *
1240  * NOTE: this function doesn't have to take the runqueue lock,
1241  * because all it wants to ensure is that the remote task enters
1242  * the kernel. If the IPI races and the task has been migrated
1243  * to another CPU then no harm is done and the purpose has been
1244  * achieved as well.
1245  */
1246 void kick_process(struct task_struct *p)
1247 {
1248         int cpu;
1249 
1250         preempt_disable();
1251         cpu = task_cpu(p);
1252         if ((cpu != smp_processor_id()) && task_curr(p))
1253                 smp_send_reschedule(cpu);
1254         preempt_enable();
1255 }
1256 EXPORT_SYMBOL_GPL(kick_process);
1257 #endif /* CONFIG_SMP */
1258 
1259 #ifdef CONFIG_SMP
1260 /*
1261  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1262  */
1263 static int select_fallback_rq(int cpu, struct task_struct *p)
1264 {
1265         int nid = cpu_to_node(cpu);
1266         const struct cpumask *nodemask = NULL;
1267         enum { cpuset, possible, fail } state = cpuset;
1268         int dest_cpu;
1269 
1270         /*
1271          * If the node that the cpu is on has been offlined, cpu_to_node()
1272          * will return -1. There is no cpu on the node, and we should
1273          * select the cpu on the other node.
1274          */
1275         if (nid != -1) {
1276                 nodemask = cpumask_of_node(nid);
1277 
1278                 /* Look for allowed, online CPU in same node. */
1279                 for_each_cpu(dest_cpu, nodemask) {
1280                         if (!cpu_online(dest_cpu))
1281                                 continue;
1282                         if (!cpu_active(dest_cpu))
1283                                 continue;
1284                         if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1285                                 return dest_cpu;
1286                 }
1287         }
1288 
1289         for (;;) {
1290                 /* Any allowed, online CPU? */
1291                 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1292                         if (!cpu_online(dest_cpu))
1293                                 continue;
1294                         if (!cpu_active(dest_cpu))
1295                                 continue;
1296                         goto out;
1297                 }
1298 
1299                 switch (state) {
1300                 case cpuset:
1301                         /* No more Mr. Nice Guy. */
1302                         cpuset_cpus_allowed_fallback(p);
1303                         state = possible;
1304                         break;
1305 
1306                 case possible:
1307                         do_set_cpus_allowed(p, cpu_possible_mask);
1308                         state = fail;
1309                         break;
1310 
1311                 case fail:
1312                         BUG();
1313                         break;
1314                 }
1315         }
1316 
1317 out:
1318         if (state != cpuset) {
1319                 /*
1320                  * Don't tell them about moving exiting tasks or
1321                  * kernel threads (both mm NULL), since they never
1322                  * leave kernel.
1323                  */
1324                 if (p->mm && printk_ratelimit()) {
1325                         printk_sched("process %d (%s) no longer affine to cpu%d\n",
1326                                         task_pid_nr(p), p->comm, cpu);
1327                 }
1328         }
1329 
1330         return dest_cpu;
1331 }
1332 
1333 /*
1334  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1335  */
1336 static inline
1337 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1338 {
1339         cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1340 
1341         /*
1342          * In order not to call set_task_cpu() on a blocking task we need
1343          * to rely on ttwu() to place the task on a valid ->cpus_allowed
1344          * cpu.
1345          *
1346          * Since this is common to all placement strategies, this lives here.
1347          *
1348          * [ this allows ->select_task() to simply return task_cpu(p) and
1349          *   not worry about this generic constraint ]
1350          */
1351         if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1352                      !cpu_online(cpu)))
1353                 cpu = select_fallback_rq(task_cpu(p), p);
1354 
1355         return cpu;
1356 }
1357 
1358 static void update_avg(u64 *avg, u64 sample)
1359 {
1360         s64 diff = sample - *avg;
1361         *avg += diff >> 3;
1362 }
1363 #endif
1364 
1365 static void
1366 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1367 {
1368 #ifdef CONFIG_SCHEDSTATS
1369         struct rq *rq = this_rq();
1370 
1371 #ifdef CONFIG_SMP
1372         int this_cpu = smp_processor_id();
1373 
1374         if (cpu == this_cpu) {
1375                 schedstat_inc(rq, ttwu_local);
1376                 schedstat_inc(p, se.statistics.nr_wakeups_local);
1377         } else {
1378                 struct sched_domain *sd;
1379 
1380                 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1381                 rcu_read_lock();
1382                 for_each_domain(this_cpu, sd) {
1383                         if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1384                                 schedstat_inc(sd, ttwu_wake_remote);
1385                                 break;
1386                         }
1387                 }
1388                 rcu_read_unlock();
1389         }
1390 
1391         if (wake_flags & WF_MIGRATED)
1392                 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1393 
1394 #endif /* CONFIG_SMP */
1395 
1396         schedstat_inc(rq, ttwu_count);
1397         schedstat_inc(p, se.statistics.nr_wakeups);
1398 
1399         if (wake_flags & WF_SYNC)
1400                 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1401 
1402 #endif /* CONFIG_SCHEDSTATS */
1403 }
1404 
1405 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1406 {
1407         activate_task(rq, p, en_flags);
1408         p->on_rq = 1;
1409 
1410         /* if a worker is waking up, notify workqueue */
1411         if (p->flags & PF_WQ_WORKER)
1412                 wq_worker_waking_up(p, cpu_of(rq));
1413 }
1414 
1415 /*
1416  * Mark the task runnable and perform wakeup-preemption.
1417  */
1418 static void
1419 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1420 {
1421         check_preempt_curr(rq, p, wake_flags);
1422         trace_sched_wakeup(p, true);
1423 
1424         p->state = TASK_RUNNING;
1425 #ifdef CONFIG_SMP
1426         if (p->sched_class->task_woken)
1427                 p->sched_class->task_woken(rq, p);
1428 
1429         if (rq->idle_stamp) {
1430                 u64 delta = rq_clock(rq) - rq->idle_stamp;
1431                 u64 max = 2*rq->max_idle_balance_cost;
1432 
1433                 update_avg(&rq->avg_idle, delta);
1434 
1435                 if (rq->avg_idle > max)
1436                         rq->avg_idle = max;
1437 
1438                 rq->idle_stamp = 0;
1439         }
1440 #endif
1441 }
1442 
1443 static void
1444 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1445 {
1446 #ifdef CONFIG_SMP
1447         if (p->sched_contributes_to_load)
1448                 rq->nr_uninterruptible--;
1449 #endif
1450 
1451         ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1452         ttwu_do_wakeup(rq, p, wake_flags);
1453 }
1454 
1455 /*
1456  * Called in case the task @p isn't fully descheduled from its runqueue,
1457  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1458  * since all we need to do is flip p->state to TASK_RUNNING, since
1459  * the task is still ->on_rq.
1460  */
1461 static int ttwu_remote(struct task_struct *p, int wake_flags)
1462 {
1463         struct rq *rq;
1464         int ret = 0;
1465 
1466         rq = __task_rq_lock(p);
1467         if (p->on_rq) {
1468                 /* check_preempt_curr() may use rq clock */
1469                 update_rq_clock(rq);
1470                 ttwu_do_wakeup(rq, p, wake_flags);
1471                 ret = 1;
1472         }
1473         __task_rq_unlock(rq);
1474 
1475         return ret;
1476 }
1477 
1478 #ifdef CONFIG_SMP
1479 static void sched_ttwu_pending(void)
1480 {
1481         struct rq *rq = this_rq();
1482         struct llist_node *llist = llist_del_all(&rq->wake_list);
1483         struct task_struct *p;
1484 
1485         raw_spin_lock(&rq->lock);
1486 
1487         while (llist) {
1488                 p = llist_entry(llist, struct task_struct, wake_entry);
1489                 llist = llist_next(llist);
1490                 ttwu_do_activate(rq, p, 0);
1491         }
1492 
1493         raw_spin_unlock(&rq->lock);
1494 }
1495 
1496 void scheduler_ipi(void)
1497 {
1498         /*
1499          * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1500          * TIF_NEED_RESCHED remotely (for the first time) will also send
1501          * this IPI.
1502          */
1503         preempt_fold_need_resched();
1504 
1505         if (llist_empty(&this_rq()->wake_list)
1506                         && !tick_nohz_full_cpu(smp_processor_id())
1507                         && !got_nohz_idle_kick())
1508                 return;
1509 
1510         /*
1511          * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1512          * traditionally all their work was done from the interrupt return
1513          * path. Now that we actually do some work, we need to make sure
1514          * we do call them.
1515          *
1516          * Some archs already do call them, luckily irq_enter/exit nest
1517          * properly.
1518          *
1519          * Arguably we should visit all archs and update all handlers,
1520          * however a fair share of IPIs are still resched only so this would
1521          * somewhat pessimize the simple resched case.
1522          */
1523         irq_enter();
1524         tick_nohz_full_check();
1525         sched_ttwu_pending();
1526 
1527         /*
1528          * Check if someone kicked us for doing the nohz idle load balance.
1529          */
1530         if (unlikely(got_nohz_idle_kick())) {
1531                 this_rq()->idle_balance = 1;
1532                 raise_softirq_irqoff(SCHED_SOFTIRQ);
1533         }
1534         irq_exit();
1535 }
1536 
1537 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1538 {
1539         if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1540                 smp_send_reschedule(cpu);
1541 }
1542 
1543 bool cpus_share_cache(int this_cpu, int that_cpu)
1544 {
1545         return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1546 }
1547 #endif /* CONFIG_SMP */
1548 
1549 static void ttwu_queue(struct task_struct *p, int cpu)
1550 {
1551         struct rq *rq = cpu_rq(cpu);
1552 
1553 #if defined(CONFIG_SMP)
1554         if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1555                 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1556                 ttwu_queue_remote(p, cpu);
1557                 return;
1558         }
1559 #endif
1560 
1561         raw_spin_lock(&rq->lock);
1562         ttwu_do_activate(rq, p, 0);
1563         raw_spin_unlock(&rq->lock);
1564 }
1565 
1566 /**
1567  * try_to_wake_up - wake up a thread
1568  * @p: the thread to be awakened
1569  * @state: the mask of task states that can be woken
1570  * @wake_flags: wake modifier flags (WF_*)
1571  *
1572  * Put it on the run-queue if it's not already there. The "current"
1573  * thread is always on the run-queue (except when the actual
1574  * re-schedule is in progress), and as such you're allowed to do
1575  * the simpler "current->state = TASK_RUNNING" to mark yourself
1576  * runnable without the overhead of this.
1577  *
1578  * Return: %true if @p was woken up, %false if it was already running.
1579  * or @state didn't match @p's state.
1580  */
1581 static int
1582 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1583 {
1584         unsigned long flags;
1585         int cpu, success = 0;
1586 
1587         /*
1588          * If we are going to wake up a thread waiting for CONDITION we
1589          * need to ensure that CONDITION=1 done by the caller can not be
1590          * reordered with p->state check below. This pairs with mb() in
1591          * set_current_state() the waiting thread does.
1592          */
1593         smp_mb__before_spinlock();
1594         raw_spin_lock_irqsave(&p->pi_lock, flags);
1595         if (!(p->state & state))
1596                 goto out;
1597 
1598         success = 1; /* we're going to change ->state */
1599         cpu = task_cpu(p);
1600 
1601         if (p->on_rq && ttwu_remote(p, wake_flags))
1602                 goto stat;
1603 
1604 #ifdef CONFIG_SMP
1605         /*
1606          * If the owning (remote) cpu is still in the middle of schedule() with
1607          * this task as prev, wait until its done referencing the task.
1608          */
1609         while (p->on_cpu)
1610                 cpu_relax();
1611         /*
1612          * Pairs with the smp_wmb() in finish_lock_switch().
1613          */
1614         smp_rmb();
1615 
1616         p->sched_contributes_to_load = !!task_contributes_to_load(p);
1617         p->state = TASK_WAKING;
1618 
1619         if (p->sched_class->task_waking)
1620                 p->sched_class->task_waking(p);
1621 
1622         cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1623         if (task_cpu(p) != cpu) {
1624                 wake_flags |= WF_MIGRATED;
1625                 set_task_cpu(p, cpu);
1626         }
1627 #endif /* CONFIG_SMP */
1628 
1629         ttwu_queue(p, cpu);
1630 stat:
1631         ttwu_stat(p, cpu, wake_flags);
1632 out:
1633         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1634 
1635         return success;
1636 }
1637 
1638 /**
1639  * try_to_wake_up_local - try to wake up a local task with rq lock held
1640  * @p: the thread to be awakened
1641  *
1642  * Put @p on the run-queue if it's not already there. The caller must
1643  * ensure that this_rq() is locked, @p is bound to this_rq() and not
1644  * the current task.
1645  */
1646 static void try_to_wake_up_local(struct task_struct *p)
1647 {
1648         struct rq *rq = task_rq(p);
1649 
1650         if (WARN_ON_ONCE(rq != this_rq()) ||
1651             WARN_ON_ONCE(p == current))
1652                 return;
1653 
1654         lockdep_assert_held(&rq->lock);
1655 
1656         if (!raw_spin_trylock(&p->pi_lock)) {
1657                 raw_spin_unlock(&rq->lock);
1658                 raw_spin_lock(&p->pi_lock);
1659                 raw_spin_lock(&rq->lock);
1660         }
1661 
1662         if (!(p->state & TASK_NORMAL))
1663                 goto out;
1664 
1665         if (!p->on_rq)
1666                 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1667 
1668         ttwu_do_wakeup(rq, p, 0);
1669         ttwu_stat(p, smp_processor_id(), 0);
1670 out:
1671         raw_spin_unlock(&p->pi_lock);
1672 }
1673 
1674 /**
1675  * wake_up_process - Wake up a specific process
1676  * @p: The process to be woken up.
1677  *
1678  * Attempt to wake up the nominated process and move it to the set of runnable
1679  * processes.
1680  *
1681  * Return: 1 if the process was woken up, 0 if it was already running.
1682  *
1683  * It may be assumed that this function implies a write memory barrier before
1684  * changing the task state if and only if any tasks are woken up.
1685  */
1686 int wake_up_process(struct task_struct *p)
1687 {
1688         WARN_ON(task_is_stopped_or_traced(p));
1689         return try_to_wake_up(p, TASK_NORMAL, 0);
1690 }
1691 EXPORT_SYMBOL(wake_up_process);
1692 
1693 int wake_up_state(struct task_struct *p, unsigned int state)
1694 {
1695         return try_to_wake_up(p, state, 0);
1696 }
1697 
1698 /*
1699  * Perform scheduler related setup for a newly forked process p.
1700  * p is forked by current.
1701  *
1702  * __sched_fork() is basic setup used by init_idle() too:
1703  */
1704 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1705 {
1706         p->on_rq                        = 0;
1707 
1708         p->se.on_rq                     = 0;
1709         p->se.exec_start                = 0;
1710         p->se.sum_exec_runtime          = 0;
1711         p->se.prev_sum_exec_runtime     = 0;
1712         p->se.nr_migrations             = 0;
1713         p->se.vruntime                  = 0;
1714         INIT_LIST_HEAD(&p->se.group_node);
1715 
1716 #ifdef CONFIG_SCHEDSTATS
1717         memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1718 #endif
1719 
1720         RB_CLEAR_NODE(&p->dl.rb_node);
1721         hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1722         p->dl.dl_runtime = p->dl.runtime = 0;
1723         p->dl.dl_deadline = p->dl.deadline = 0;
1724         p->dl.dl_period = 0;
1725         p->dl.flags = 0;
1726 
1727         INIT_LIST_HEAD(&p->rt.run_list);
1728 
1729 #ifdef CONFIG_PREEMPT_NOTIFIERS
1730         INIT_HLIST_HEAD(&p->preempt_notifiers);
1731 #endif
1732 
1733 #ifdef CONFIG_NUMA_BALANCING
1734         if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1735                 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1736                 p->mm->numa_scan_seq = 0;
1737         }
1738 
1739         if (clone_flags & CLONE_VM)
1740                 p->numa_preferred_nid = current->numa_preferred_nid;
1741         else
1742                 p->numa_preferred_nid = -1;
1743 
1744         p->node_stamp = 0ULL;
1745         p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1746         p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1747         p->numa_work.next = &p->numa_work;
1748         p->numa_faults = NULL;
1749         p->numa_faults_buffer = NULL;
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                 task_numa_free(prev);
2153 
2154                 if (prev->sched_class->task_dead)
2155                         prev->sched_class->task_dead(prev);
2156 
2157                 /*
2158                  * Remove function-return probe instances associated with this
2159                  * task and put them back on the free list.
2160                  */
2161                 kprobe_flush_task(prev);
2162                 put_task_struct(prev);
2163         }
2164 
2165         tick_nohz_task_switch(current);
2166 }
2167 
2168 #ifdef CONFIG_SMP
2169 
2170 /* assumes rq->lock is held */
2171 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2172 {
2173         if (prev->sched_class->pre_schedule)
2174                 prev->sched_class->pre_schedule(rq, prev);
2175 }
2176 
2177 /* rq->lock is NOT held, but preemption is disabled */
2178 static inline void post_schedule(struct rq *rq)
2179 {
2180         if (rq->post_schedule) {
2181                 unsigned long flags;
2182 
2183                 raw_spin_lock_irqsave(&rq->lock, flags);
2184                 if (rq->curr->sched_class->post_schedule)
2185                         rq->curr->sched_class->post_schedule(rq);
2186                 raw_spin_unlock_irqrestore(&rq->lock, flags);
2187 
2188                 rq->post_schedule = 0;
2189         }
2190 }
2191 
2192 #else
2193 
2194 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2195 {
2196 }
2197 
2198 static inline void post_schedule(struct rq *rq)
2199 {
2200 }
2201 
2202 #endif
2203 
2204 /**
2205  * schedule_tail - first thing a freshly forked thread must call.
2206  * @prev: the thread we just switched away from.
2207  */
2208 asmlinkage void schedule_tail(struct task_struct *prev)
2209         __releases(rq->lock)
2210 {
2211         struct rq *rq = this_rq();
2212 
2213         finish_task_switch(rq, prev);
2214 
2215         /*
2216          * FIXME: do we need to worry about rq being invalidated by the
2217          * task_switch?
2218          */
2219         post_schedule(rq);
2220 
2221 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2222         /* In this case, finish_task_switch does not reenable preemption */
2223         preempt_enable();
2224 #endif
2225         if (current->set_child_tid)
2226                 put_user(task_pid_vnr(current), current->set_child_tid);
2227 }
2228 
2229 /*
2230  * context_switch - switch to the new MM and the new
2231  * thread's register state.
2232  */
2233 static inline void
2234 context_switch(struct rq *rq, struct task_struct *prev,
2235                struct task_struct *next)
2236 {
2237         struct mm_struct *mm, *oldmm;
2238 
2239         prepare_task_switch(rq, prev, next);
2240 
2241         mm = next->mm;
2242         oldmm = prev->active_mm;
2243         /*
2244          * For paravirt, this is coupled with an exit in switch_to to
2245          * combine the page table reload and the switch backend into
2246          * one hypercall.
2247          */
2248         arch_start_context_switch(prev);
2249 
2250         if (!mm) {
2251                 next->active_mm = oldmm;
2252                 atomic_inc(&oldmm->mm_count);
2253                 enter_lazy_tlb(oldmm, next);
2254         } else
2255                 switch_mm(oldmm, mm, next);
2256 
2257         if (!prev->mm) {
2258                 prev->active_mm = NULL;
2259                 rq->prev_mm = oldmm;
2260         }
2261         /*
2262          * Since the runqueue lock will be released by the next
2263          * task (which is an invalid locking op but in the case
2264          * of the scheduler it's an obvious special-case), so we
2265          * do an early lockdep release here:
2266          */
2267 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2268         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2269 #endif
2270 
2271         context_tracking_task_switch(prev, next);
2272         /* Here we just switch the register state and the stack. */
2273         switch_to(prev, next, prev);
2274 
2275         barrier();
2276         /*
2277          * this_rq must be evaluated again because prev may have moved
2278          * CPUs since it called schedule(), thus the 'rq' on its stack
2279          * frame will be invalid.
2280          */
2281         finish_task_switch(this_rq(), prev);
2282 }
2283 
2284 /*
2285  * nr_running and nr_context_switches:
2286  *
2287  * externally visible scheduler statistics: current number of runnable
2288  * threads, total number of context switches performed since bootup.
2289  */
2290 unsigned long nr_running(void)
2291 {
2292         unsigned long i, sum = 0;
2293 
2294         for_each_online_cpu(i)
2295                 sum += cpu_rq(i)->nr_running;
2296 
2297         return sum;
2298 }
2299 
2300 unsigned long long nr_context_switches(void)
2301 {
2302         int i;
2303         unsigned long long sum = 0;
2304 
2305         for_each_possible_cpu(i)
2306                 sum += cpu_rq(i)->nr_switches;
2307 
2308         return sum;
2309 }
2310 
2311 unsigned long nr_iowait(void)
2312 {
2313         unsigned long i, sum = 0;
2314 
2315         for_each_possible_cpu(i)
2316                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2317 
2318         return sum;
2319 }
2320 
2321 unsigned long nr_iowait_cpu(int cpu)
2322 {
2323         struct rq *this = cpu_rq(cpu);
2324         return atomic_read(&this->nr_iowait);
2325 }
2326 
2327 #ifdef CONFIG_SMP
2328 
2329 /*
2330  * sched_exec - execve() is a valuable balancing opportunity, because at
2331  * this point the task has the smallest effective memory and cache footprint.
2332  */
2333 void sched_exec(void)
2334 {
2335         struct task_struct *p = current;
2336         unsigned long flags;
2337         int dest_cpu;
2338 
2339         raw_spin_lock_irqsave(&p->pi_lock, flags);
2340         dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2341         if (dest_cpu == smp_processor_id())
2342                 goto unlock;
2343 
2344         if (likely(cpu_active(dest_cpu))) {
2345                 struct migration_arg arg = { p, dest_cpu };
2346 
2347                 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2348                 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2349                 return;
2350         }
2351 unlock:
2352         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2353 }
2354 
2355 #endif
2356 
2357 DEFINE_PER_CPU(struct kernel_stat, kstat);
2358 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2359 
2360 EXPORT_PER_CPU_SYMBOL(kstat);
2361 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2362 
2363 /*
2364  * Return any ns on the sched_clock that have not yet been accounted in
2365  * @p in case that task is currently running.
2366  *
2367  * Called with task_rq_lock() held on @rq.
2368  */
2369 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2370 {
2371         u64 ns = 0;
2372 
2373         if (task_current(rq, p)) {
2374                 update_rq_clock(rq);
2375                 ns = rq_clock_task(rq) - p->se.exec_start;
2376                 if ((s64)ns < 0)
2377                         ns = 0;
2378         }
2379 
2380         return ns;
2381 }
2382 
2383 unsigned long long task_delta_exec(struct task_struct *p)
2384 {
2385         unsigned long flags;
2386         struct rq *rq;
2387         u64 ns = 0;
2388 
2389         rq = task_rq_lock(p, &flags);
2390         ns = do_task_delta_exec(p, rq);
2391         task_rq_unlock(rq, p, &flags);
2392 
2393         return ns;
2394 }
2395 
2396 /*
2397  * Return accounted runtime for the task.
2398  * In case the task is currently running, return the runtime plus current's
2399  * pending runtime that have not been accounted yet.
2400  */
2401 unsigned long long task_sched_runtime(struct task_struct *p)
2402 {
2403         unsigned long flags;
2404         struct rq *rq;
2405         u64 ns = 0;
2406 
2407 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2408         /*
2409          * 64-bit doesn't need locks to atomically read a 64bit value.
2410          * So we have a optimization chance when the task's delta_exec is 0.
2411          * Reading ->on_cpu is racy, but this is ok.
2412          *
2413          * If we race with it leaving cpu, we'll take a lock. So we're correct.
2414          * If we race with it entering cpu, unaccounted time is 0. This is
2415          * indistinguishable from the read occurring a few cycles earlier.
2416          */
2417         if (!p->on_cpu)
2418                 return p->se.sum_exec_runtime;
2419 #endif
2420 
2421         rq = task_rq_lock(p, &flags);
2422         ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2423         task_rq_unlock(rq, p, &flags);
2424 
2425         return ns;
2426 }
2427 
2428 /*
2429  * This function gets called by the timer code, with HZ frequency.
2430  * We call it with interrupts disabled.
2431  */
2432 void scheduler_tick(void)
2433 {
2434         int cpu = smp_processor_id();
2435         struct rq *rq = cpu_rq(cpu);
2436         struct task_struct *curr = rq->curr;
2437 
2438         sched_clock_tick();
2439 
2440         raw_spin_lock(&rq->lock);
2441         update_rq_clock(rq);
2442         curr->sched_class->task_tick(rq, curr, 0);
2443         update_cpu_load_active(rq);
2444         raw_spin_unlock(&rq->lock);
2445 
2446         perf_event_task_tick();
2447 
2448 #ifdef CONFIG_SMP
2449         rq->idle_balance = idle_cpu(cpu);
2450         trigger_load_balance(rq);
2451 #endif
2452         rq_last_tick_reset(rq);
2453 }
2454 
2455 #ifdef CONFIG_NO_HZ_FULL
2456 /**
2457  * scheduler_tick_max_deferment
2458  *
2459  * Keep at least one tick per second when a single
2460  * active task is running because the scheduler doesn't
2461  * yet completely support full dynticks environment.
2462  *
2463  * This makes sure that uptime, CFS vruntime, load
2464  * balancing, etc... continue to move forward, even
2465  * with a very low granularity.
2466  *
2467  * Return: Maximum deferment in nanoseconds.
2468  */
2469 u64 scheduler_tick_max_deferment(void)
2470 {
2471         struct rq *rq = this_rq();
2472         unsigned long next, now = ACCESS_ONCE(jiffies);
2473 
2474         next = rq->last_sched_tick + HZ;
2475 
2476         if (time_before_eq(next, now))
2477                 return 0;
2478 
2479         return jiffies_to_nsecs(next - now);
2480 }
2481 #endif
2482 
2483 notrace unsigned long get_parent_ip(unsigned long addr)
2484 {
2485         if (in_lock_functions(addr)) {
2486                 addr = CALLER_ADDR2;
2487                 if (in_lock_functions(addr))
2488                         addr = CALLER_ADDR3;
2489         }
2490         return addr;
2491 }
2492 
2493 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2494                                 defined(CONFIG_PREEMPT_TRACER))
2495 
2496 void __kprobes preempt_count_add(int val)
2497 {
2498 #ifdef CONFIG_DEBUG_PREEMPT
2499         /*
2500          * Underflow?
2501          */
2502         if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2503                 return;
2504 #endif
2505         __preempt_count_add(val);
2506 #ifdef CONFIG_DEBUG_PREEMPT
2507         /*
2508          * Spinlock count overflowing soon?
2509          */
2510         DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2511                                 PREEMPT_MASK - 10);
2512 #endif
2513         if (preempt_count() == val)
2514                 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2515 }
2516 EXPORT_SYMBOL(preempt_count_add);
2517 
2518 void __kprobes preempt_count_sub(int val)
2519 {
2520 #ifdef CONFIG_DEBUG_PREEMPT
2521         /*
2522          * Underflow?
2523          */
2524         if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2525                 return;
2526         /*
2527          * Is the spinlock portion underflowing?
2528          */
2529         if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2530                         !(preempt_count() & PREEMPT_MASK)))
2531                 return;
2532 #endif
2533 
2534         if (preempt_count() == val)
2535                 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2536         __preempt_count_sub(val);
2537 }
2538 EXPORT_SYMBOL(preempt_count_sub);
2539 
2540 #endif
2541 
2542 /*
2543  * Print scheduling while atomic bug:
2544  */
2545 static noinline void __schedule_bug(struct task_struct *prev)
2546 {
2547         if (oops_in_progress)
2548                 return;
2549 
2550         printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2551                 prev->comm, prev->pid, preempt_count());
2552 
2553         debug_show_held_locks(prev);
2554         print_modules();
2555         if (irqs_disabled())
2556                 print_irqtrace_events(prev);
2557         dump_stack();
2558         add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2559 }
2560 
2561 /*
2562  * Various schedule()-time debugging checks and statistics:
2563  */
2564 static inline void schedule_debug(struct task_struct *prev)
2565 {
2566         /*
2567          * Test if we are atomic. Since do_exit() needs to call into
2568          * schedule() atomically, we ignore that path. Otherwise whine
2569          * if we are scheduling when we should not.
2570          */
2571         if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2572                 __schedule_bug(prev);
2573         rcu_sleep_check();
2574 
2575         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2576 
2577         schedstat_inc(this_rq(), sched_count);
2578 }
2579 
2580 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2581 {
2582         if (prev->on_rq || rq->skip_clock_update < 0)
2583                 update_rq_clock(rq);
2584         prev->sched_class->put_prev_task(rq, prev);
2585 }
2586 
2587 /*
2588  * Pick up the highest-prio task:
2589  */
2590 static inline struct task_struct *
2591 pick_next_task(struct rq *rq)
2592 {
2593         const struct sched_class *class;
2594         struct task_struct *p;
2595 
2596         /*
2597          * Optimization: we know that if all tasks are in
2598          * the fair class we can call that function directly:
2599          */
2600         if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2601                 p = fair_sched_class.pick_next_task(rq);
2602                 if (likely(p))
2603                         return p;
2604         }
2605 
2606         for_each_class(class) {
2607                 p = class->pick_next_task(rq);
2608                 if (p)
2609                         return p;
2610         }
2611 
2612         BUG(); /* the idle class will always have a runnable task */
2613 }
2614 
2615 /*
2616  * __schedule() is the main scheduler function.
2617  *
2618  * The main means of driving the scheduler and thus entering this function are:
2619  *
2620  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2621  *
2622  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2623  *      paths. For example, see arch/x86/entry_64.S.
2624  *
2625  *      To drive preemption between tasks, the scheduler sets the flag in timer
2626  *      interrupt handler scheduler_tick().
2627  *
2628  *   3. Wakeups don't really cause entry into schedule(). They add a
2629  *      task to the run-queue and that's it.
2630  *
2631  *      Now, if the new task added to the run-queue preempts the current
2632  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2633  *      called on the nearest possible occasion:
2634  *
2635  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2636  *
2637  *         - in syscall or exception context, at the next outmost
2638  *           preempt_enable(). (this might be as soon as the wake_up()'s
2639  *           spin_unlock()!)
2640  *
2641  *         - in IRQ context, return from interrupt-handler to
2642  *           preemptible context
2643  *
2644  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2645  *         then at the next:
2646  *
2647  *          - cond_resched() call
2648  *          - explicit schedule() call
2649  *          - return from syscall or exception to user-space
2650  *          - return from interrupt-handler to user-space
2651  */
2652 static void __sched __schedule(void)
2653 {
2654         struct task_struct *prev, *next;
2655         unsigned long *switch_count;
2656         struct rq *rq;
2657         int cpu;
2658 
2659 need_resched:
2660         preempt_disable();
2661         cpu = smp_processor_id();
2662         rq = cpu_rq(cpu);
2663         rcu_note_context_switch(cpu);
2664         prev = rq->curr;
2665 
2666         schedule_debug(prev);
2667 
2668         if (sched_feat(HRTICK))
2669                 hrtick_clear(rq);
2670 
2671         /*
2672          * Make sure that signal_pending_state()->signal_pending() below
2673          * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2674          * done by the caller to avoid the race with signal_wake_up().
2675          */
2676         smp_mb__before_spinlock();
2677         raw_spin_lock_irq(&rq->lock);
2678 
2679         switch_count = &prev->nivcsw;
2680         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2681                 if (unlikely(signal_pending_state(prev->state, prev))) {
2682                         prev->state = TASK_RUNNING;
2683                 } else {
2684                         deactivate_task(rq, prev, DEQUEUE_SLEEP);
2685                         prev->on_rq = 0;
2686 
2687                         /*
2688                          * If a worker went to sleep, notify and ask workqueue
2689                          * whether it wants to wake up a task to maintain
2690                          * concurrency.
2691                          */
2692                         if (prev->flags & PF_WQ_WORKER) {
2693                                 struct task_struct *to_wakeup;
2694 
2695                                 to_wakeup = wq_worker_sleeping(prev, cpu);
2696                                 if (to_wakeup)
2697                                         try_to_wake_up_local(to_wakeup);
2698                         }
2699                 }
2700                 switch_count = &prev->nvcsw;
2701         }
2702 
2703         pre_schedule(rq, prev);
2704 
2705         if (unlikely(!rq->nr_running))
2706                 idle_balance(cpu, rq);
2707 
2708         put_prev_task(rq, prev);
2709         next = pick_next_task(rq);
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 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 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 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 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 static long __sched
2856 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2857 {
2858         unsigned long flags;
2859         wait_queue_t wait;
2860 
2861         init_waitqueue_entry(&wait, current);
2862 
2863         __set_current_state(state);
2864 
2865         spin_lock_irqsave(&q->lock, flags);
2866         __add_wait_queue(q, &wait);
2867         spin_unlock(&q->lock);
2868         timeout = schedule_timeout(timeout);
2869         spin_lock_irq(&q->lock);
2870         __remove_wait_queue(q, &wait);
2871         spin_unlock_irqrestore(&q->lock, flags);
2872 
2873         return timeout;
2874 }
2875 
2876 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2877 {
2878         sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2879 }
2880 EXPORT_SYMBOL(interruptible_sleep_on);
2881 
2882 long __sched
2883 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2884 {
2885         return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2886 }
2887 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2888 
2889 void __sched sleep_on(wait_queue_head_t *q)
2890 {
2891         sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2892 }
2893 EXPORT_SYMBOL(sleep_on);
2894 
2895 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2896 {
2897         return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2898 }
2899 EXPORT_SYMBOL(sleep_on_timeout);
2900 
2901 #ifdef CONFIG_RT_MUTEXES
2902 
2903 /*
2904  * rt_mutex_setprio - set the current priority of a task
2905  * @p: task
2906  * @prio: prio value (kernel-internal form)
2907  *
2908  * This function changes the 'effective' priority of a task. It does
2909  * not touch ->normal_prio like __setscheduler().
2910  *
2911  * Used by the rt_mutex code to implement priority inheritance logic.
2912  */
2913 void rt_mutex_setprio(struct task_struct *p, int prio)
2914 {
2915         int oldprio, on_rq, running, enqueue_flag = 0;
2916         struct rq *rq;
2917         const struct sched_class *prev_class;
2918 
2919         BUG_ON(prio > MAX_PRIO);
2920 
2921         rq = __task_rq_lock(p);
2922 
2923         /*
2924          * Idle task boosting is a nono in general. There is one
2925          * exception, when PREEMPT_RT and NOHZ is active:
2926          *
2927          * The idle task calls get_next_timer_interrupt() and holds
2928          * the timer wheel base->lock on the CPU and another CPU wants
2929          * to access the timer (probably to cancel it). We can safely
2930          * ignore the boosting request, as the idle CPU runs this code
2931          * with interrupts disabled and will complete the lock
2932          * protected section without being interrupted. So there is no
2933          * real need to boost.
2934          */
2935         if (unlikely(p == rq->idle)) {
2936                 WARN_ON(p != rq->curr);
2937                 WARN_ON(p->pi_blocked_on);
2938                 goto out_unlock;
2939         }
2940 
2941         trace_sched_pi_setprio(p, prio);
2942         p->pi_top_task = rt_mutex_get_top_task(p);
2943         oldprio = p->prio;
2944         prev_class = p->sched_class;
2945         on_rq = p->on_rq;
2946         running = task_current(rq, p);
2947         if (on_rq)
2948                 dequeue_task(rq, p, 0);
2949         if (running)
2950                 p->sched_class->put_prev_task(rq, p);
2951 
2952         /*
2953          * Boosting condition are:
2954          * 1. -rt task is running and holds mutex A
2955          *      --> -dl task blocks on mutex A
2956          *
2957          * 2. -dl task is running and holds mutex A
2958          *      --> -dl task blocks on mutex A and could preempt the
2959          *          running task
2960          */
2961         if (dl_prio(prio)) {
2962                 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2963                         dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2964                         p->dl.dl_boosted = 1;
2965                         p->dl.dl_throttled = 0;
2966                         enqueue_flag = ENQUEUE_REPLENISH;
2967                 } else
2968                         p->dl.dl_boosted = 0;
2969                 p->sched_class = &dl_sched_class;
2970         } else if (rt_prio(prio)) {
2971                 if (dl_prio(oldprio))
2972                         p->dl.dl_boosted = 0;
2973                 if (oldprio < prio)
2974                         enqueue_flag = ENQUEUE_HEAD;
2975                 p->sched_class = &rt_sched_class;
2976         } else {
2977                 if (dl_prio(oldprio))
2978                         p->dl.dl_boosted = 0;
2979                 p->sched_class = &fair_sched_class;
2980         }
2981 
2982         p->prio = prio;
2983 
2984         if (running)
2985                 p->sched_class->set_curr_task(rq);
2986         if (on_rq)
2987                 enqueue_task(rq, p, enqueue_flag);
2988 
2989         check_class_changed(rq, p, prev_class, oldprio);
2990 out_unlock:
2991         __task_rq_unlock(rq);
2992 }
2993 #endif
2994 
2995 void set_user_nice(struct task_struct *p, long nice)
2996 {
2997         int old_prio, delta, on_rq;
2998         unsigned long flags;
2999         struct rq *rq;
3000 
3001         if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3002                 return;
3003         /*
3004          * We have to be careful, if called from sys_setpriority(),
3005          * the task might be in the middle of scheduling on another CPU.
3006          */
3007         rq = task_rq_lock(p, &flags);
3008         /*
3009          * The RT priorities are set via sched_setscheduler(), but we still
3010          * allow the 'normal' nice value to be set - but as expected
3011          * it wont have any effect on scheduling until the task is
3012          * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3013          */
3014         if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3015                 p->static_prio = NICE_TO_PRIO(nice);
3016                 goto out_unlock;
3017         }
3018         on_rq = p->on_rq;
3019         if (on_rq)
3020                 dequeue_task(rq, p, 0);
3021 
3022         p->static_prio = NICE_TO_PRIO(nice);
3023         set_load_weight(p);
3024         old_prio = p->prio;
3025         p->prio = effective_prio(p);
3026         delta = p->prio - old_prio;
3027 
3028         if (on_rq) {
3029                 enqueue_task(rq, p, 0);
3030                 /*
3031                  * If the task increased its priority or is running and
3032                  * lowered its priority, then reschedule its CPU:
3033                  */
3034                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3035                         resched_task(rq->curr);
3036         }
3037 out_unlock:
3038         task_rq_unlock(rq, p, &flags);
3039 }
3040 EXPORT_SYMBOL(set_user_nice);
3041 
3042 /*
3043  * can_nice - check if a task can reduce its nice value
3044  * @p: task
3045  * @nice: nice value
3046  */
3047 int can_nice(const struct task_struct *p, const int nice)
3048 {
3049         /* convert nice value [19,-20] to rlimit style value [1,40] */
3050         int nice_rlim = 20 - nice;
3051 
3052         return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3053                 capable(CAP_SYS_NICE));
3054 }
3055 
3056 #ifdef __ARCH_WANT_SYS_NICE
3057 
3058 /*
3059  * sys_nice - change the priority of the current process.
3060  * @increment: priority increment
3061  *
3062  * sys_setpriority is a more generic, but much slower function that
3063  * does similar things.
3064  */
3065 SYSCALL_DEFINE1(nice, int, increment)
3066 {
3067         long nice, retval;
3068 
3069         /*
3070          * Setpriority might change our priority at the same moment.
3071          * We don't have to worry. Conceptually one call occurs first
3072          * and we have a single winner.
3073          */
3074         if (increment < -40)
3075                 increment = -40;
3076         if (increment > 40)
3077                 increment = 40;
3078 
3079         nice = TASK_NICE(current) + increment;
3080         if (nice < -20)
3081                 nice = -20;
3082         if (nice > 19)
3083                 nice = 19;
3084 
3085         if (increment < 0 && !can_nice(current, nice))
3086                 return -EPERM;
3087 
3088         retval = security_task_setnice(current, nice);
3089         if (retval)
3090                 return retval;
3091 
3092         set_user_nice(current, nice);
3093         return 0;
3094 }
3095 
3096 #endif
3097 
3098 /**
3099  * task_prio - return the priority value of a given task.
3100  * @p: the task in question.
3101  *
3102  * Return: The priority value as seen by users in /proc.
3103  * RT tasks are offset by -200. Normal tasks are centered
3104  * around 0, value goes from -16 to +15.
3105  */
3106 int task_prio(const struct task_struct *p)
3107 {
3108         return p->prio - MAX_RT_PRIO;
3109 }
3110 
3111 /**
3112  * task_nice - return the nice value of a given task.
3113  * @p: the task in question.
3114  *
3115  * Return: The nice value [ -20 ... 0 ... 19 ].
3116  */
3117 int task_nice(const struct task_struct *p)
3118 {
3119         return TASK_NICE(p);
3120 }
3121 EXPORT_SYMBOL(task_nice);
3122 
3123 /**
3124  * idle_cpu - is a given cpu idle currently?
3125  * @cpu: the processor in question.
3126  *
3127  * Return: 1 if the CPU is currently idle. 0 otherwise.
3128  */
3129 int idle_cpu(int cpu)
3130 {
3131         struct rq *rq = cpu_rq(cpu);
3132 
3133         if (rq->curr != rq->idle)
3134                 return 0;
3135 
3136         if (rq->nr_running)
3137                 return 0;
3138 
3139 #ifdef CONFIG_SMP
3140         if (!llist_empty(&rq->wake_list))
3141                 return 0;
3142 #endif
3143 
3144         return 1;
3145 }
3146 
3147 /**
3148  * idle_task - return the idle task for a given cpu.
3149  * @cpu: the processor in question.
3150  *
3151  * Return: The idle task for the cpu @cpu.
3152  */
3153 struct task_struct *idle_task(int cpu)
3154 {
3155         return cpu_rq(cpu)->idle;
3156 }
3157 
3158 /**
3159  * find_process_by_pid - find a process with a matching PID value.
3160  * @pid: the pid in question.
3161  *
3162  * The task of @pid, if found. %NULL otherwise.
3163  */
3164 static struct task_struct *find_process_by_pid(pid_t pid)
3165 {
3166         return pid ? find_task_by_vpid(pid) : current;
3167 }
3168 
3169 /*
3170  * This function initializes the sched_dl_entity of a newly becoming
3171  * SCHED_DEADLINE task.
3172  *
3173  * Only the static values are considered here, the actual runtime and the
3174  * absolute deadline will be properly calculated when the task is enqueued
3175  * for the first time with its new policy.
3176  */
3177 static void
3178 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3179 {
3180         struct sched_dl_entity *dl_se = &p->dl;
3181 
3182         init_dl_task_timer(dl_se);
3183         dl_se->dl_runtime = attr->sched_runtime;
3184         dl_se->dl_deadline = attr->sched_deadline;
3185         dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3186         dl_se->flags = attr->sched_flags;
3187         dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3188         dl_se->dl_throttled = 0;
3189         dl_se->dl_new = 1;
3190 }
3191 
3192 /* Actually do priority change: must hold pi & rq lock. */
3193 static void __setscheduler(struct rq *rq, struct task_struct *p,
3194                            const struct sched_attr *attr)
3195 {
3196         int policy = attr->sched_policy;
3197 
3198         if (policy == -1) /* setparam */
3199                 policy = p->policy;
3200 
3201         p->policy = policy;
3202 
3203         if (dl_policy(policy))
3204                 __setparam_dl(p, attr);
3205         else if (fair_policy(policy))
3206                 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3207 
3208         /*
3209          * __sched_setscheduler() ensures attr->sched_priority == 0 when
3210          * !rt_policy. Always setting this ensures that things like
3211          * getparam()/getattr() don't report silly values for !rt tasks.
3212          */
3213         p->rt_priority = attr->sched_priority;
3214 
3215         p->normal_prio = normal_prio(p);
3216         p->prio = rt_mutex_getprio(p);
3217 
3218         if (dl_prio(p->prio))
3219                 p->sched_class = &dl_sched_class;
3220         else if (rt_prio(p->prio))
3221                 p->sched_class = &rt_sched_class;
3222         else
3223                 p->sched_class = &fair_sched_class;
3224 
3225         set_load_weight(p);
3226 }
3227 
3228 static void
3229 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3230 {
3231         struct sched_dl_entity *dl_se = &p->dl;
3232 
3233         attr->sched_priority = p->rt_priority;
3234         attr->sched_runtime = dl_se->dl_runtime;
3235         attr->sched_deadline = dl_se->dl_deadline;
3236         attr->sched_period = dl_se->dl_period;
3237         attr->sched_flags = dl_se->flags;
3238 }
3239 
3240 /*
3241  * This function validates the new parameters of a -deadline task.
3242  * We ask for the deadline not being zero, and greater or equal
3243  * than the runtime, as well as the period of being zero or
3244  * greater than deadline. Furthermore, we have to be sure that
3245  * user parameters are above the internal resolution (1us); we
3246  * check sched_runtime only since it is always the smaller one.
3247  */
3248 static bool
3249 __checkparam_dl(const struct sched_attr *attr)
3250 {
3251         return attr && attr->sched_deadline != 0 &&
3252                 (attr->sched_period == 0 ||
3253                 (s64)(attr->sched_period   - attr->sched_deadline) >= 0) &&
3254                 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0  &&
3255                 attr->sched_runtime >= (2 << (DL_SCALE - 1));
3256 }
3257 
3258 /*
3259  * check the target process has a UID that matches the current process's
3260  */
3261 static bool check_same_owner(struct task_struct *p)
3262 {
3263         const struct cred *cred = current_cred(), *pcred;
3264         bool match;
3265 
3266         rcu_read_lock();
3267         pcred = __task_cred(p);
3268         match = (uid_eq(cred->euid, pcred->euid) ||
3269                  uid_eq(cred->euid, pcred->uid));
3270         rcu_read_unlock();
3271         return match;
3272 }
3273 
3274 static int __sched_setscheduler(struct task_struct *p,
3275                                 const struct sched_attr *attr,
3276                                 bool user)
3277 {
3278         int retval, oldprio, oldpolicy = -1, on_rq, running;
3279         int policy = attr->sched_policy;
3280         unsigned long flags;
3281         const struct sched_class *prev_class;
3282         struct rq *rq;
3283         int reset_on_fork;
3284 
3285         /* may grab non-irq protected spin_locks */
3286         BUG_ON(in_interrupt());
3287 recheck:
3288         /* double check policy once rq lock held */
3289         if (policy < 0) {
3290                 reset_on_fork = p->sched_reset_on_fork;
3291                 policy = oldpolicy = p->policy;
3292         } else {
3293                 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3294 
3295                 if (policy != SCHED_DEADLINE &&
3296                                 policy != SCHED_FIFO && policy != SCHED_RR &&
3297                                 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3298                                 policy != SCHED_IDLE)
3299                         return -EINVAL;
3300         }
3301 
3302         if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3303                 return -EINVAL;
3304 
3305         /*
3306          * Valid priorities for SCHED_FIFO and SCHED_RR are
3307          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3308          * SCHED_BATCH and SCHED_IDLE is 0.
3309          */
3310         if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3311             (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3312                 return -EINVAL;
3313         if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3314             (rt_policy(policy) != (attr->sched_priority != 0)))
3315                 return -EINVAL;
3316 
3317         /*
3318          * Allow unprivileged RT tasks to decrease priority:
3319          */
3320         if (user && !capable(CAP_SYS_NICE)) {
3321                 if (fair_policy(policy)) {
3322                         if (attr->sched_nice < TASK_NICE(p) &&
3323                             !can_nice(p, attr->sched_nice))
3324                                 return -EPERM;
3325                 }
3326 
3327                 if (rt_policy(policy)) {
3328                         unsigned long rlim_rtprio =
3329                                         task_rlimit(p, RLIMIT_RTPRIO);
3330 
3331                         /* can't set/change the rt policy */
3332                         if (policy != p->policy && !rlim_rtprio)
3333                                 return -EPERM;
3334 
3335                         /* can't increase priority */
3336                         if (attr->sched_priority > p->rt_priority &&
3337                             attr->sched_priority > rlim_rtprio)
3338                                 return -EPERM;
3339                 }
3340 
3341                  /*
3342                   * Can't set/change SCHED_DEADLINE policy at all for now
3343                   * (safest behavior); in the future we would like to allow
3344                   * unprivileged DL tasks to increase their relative deadline
3345                   * or reduce their runtime (both ways reducing utilization)
3346                   */
3347                 if (dl_policy(policy))
3348                         return -EPERM;
3349 
3350                 /*
3351                  * Treat SCHED_IDLE as nice 20. Only allow a switch to
3352                  * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3353                  */
3354                 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3355                         if (!can_nice(p, TASK_NICE(p)))
3356                                 return -EPERM;
3357                 }
3358 
3359                 /* can't change other user's priorities */
3360                 if (!check_same_owner(p))
3361                         return -EPERM;
3362 
3363                 /* Normal users shall not reset the sched_reset_on_fork flag */
3364                 if (p->sched_reset_on_fork && !reset_on_fork)
3365                         return -EPERM;
3366         }
3367 
3368         if (user) {
3369                 retval = security_task_setscheduler(p);
3370                 if (retval)
3371                         return retval;
3372         }
3373 
3374         /*
3375          * make sure no PI-waiters arrive (or leave) while we are
3376          * changing the priority of the task:
3377          *
3378          * To be able to change p->policy safely, the appropriate
3379          * runqueue lock must be held.
3380          */
3381         rq = task_rq_lock(p, &flags);
3382 
3383         /*
3384          * Changing the policy of the stop threads its a very bad idea
3385          */
3386         if (p == rq->stop) {
3387                 task_rq_unlock(rq, p, &flags);
3388                 return -EINVAL;
3389         }
3390 
3391         /*
3392          * If not changing anything there's no need to proceed further:
3393          */
3394         if (unlikely(policy == p->policy)) {
3395                 if (fair_policy(policy) && attr->sched_nice != TASK_NICE(p))
3396                         goto change;
3397                 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3398                         goto change;
3399                 if (dl_policy(policy))
3400                         goto change;
3401 
3402                 task_rq_unlock(rq, p, &flags);
3403                 return 0;
3404         }
3405 change:
3406 
3407         if (user) {
3408 #ifdef CONFIG_RT_GROUP_SCHED
3409                 /*
3410                  * Do not allow realtime tasks into groups that have no runtime
3411                  * assigned.
3412                  */
3413                 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3414                                 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3415                                 !task_group_is_autogroup(task_group(p))) {
3416                         task_rq_unlock(rq, p, &flags);
3417                         return -EPERM;
3418                 }
3419 #endif
3420 #ifdef CONFIG_SMP
3421                 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3422                         cpumask_t *span = rq->rd->span;
3423 
3424                         /*
3425                          * Don't allow tasks with an affinity mask smaller than
3426                          * the entire root_domain to become SCHED_DEADLINE. We
3427                          * will also fail if there's no bandwidth available.
3428                          */
3429                         if (!cpumask_subset(span, &p->cpus_allowed) ||
3430                             rq->rd->dl_bw.bw == 0) {
3431                                 task_rq_unlock(rq, p, &flags);
3432                                 return -EPERM;
3433                         }
3434                 }
3435 #endif
3436         }
3437 
3438         /* recheck policy now with rq lock held */
3439         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3440                 policy = oldpolicy = -1;
3441                 task_rq_unlock(rq, p, &flags);
3442                 goto recheck;
3443         }
3444 
3445         /*
3446          * If setscheduling to SCHED_DEADLINE (or changing the parameters
3447          * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3448          * is available.
3449          */
3450         if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3451                 task_rq_unlock(rq, p, &flags);
3452                 return -EBUSY;
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         p->sched_reset_on_fork = reset_on_fork;
3463 
3464         oldprio = p->prio;
3465         prev_class = p->sched_class;
3466         __setscheduler(rq, p, attr);
3467 
3468         if (running)
3469                 p->sched_class->set_curr_task(rq);
3470         if (on_rq)
3471                 enqueue_task(rq, p, 0);
3472 
3473         check_class_changed(rq, p, prev_class, oldprio);
3474         task_rq_unlock(rq, p, &flags);
3475 
3476         rt_mutex_adjust_pi(p);
3477 
3478         return 0;
3479 }
3480 
3481 static int _sched_setscheduler(struct task_struct *p, int policy,
3482                                const struct sched_param *param, bool check)
3483 {
3484         struct sched_attr attr = {
3485                 .sched_policy   = policy,
3486                 .sched_priority = param->sched_priority,
3487                 .sched_nice     = PRIO_TO_NICE(p->static_prio),
3488         };
3489 
3490         /*
3491          * Fixup the legacy SCHED_RESET_ON_FORK hack
3492          */
3493         if (policy & SCHED_RESET_ON_FORK) {
3494                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3495                 policy &= ~SCHED_RESET_ON_FORK;
3496                 attr.sched_policy = policy;
3497         }
3498 
3499         return __sched_setscheduler(p, &attr, check);
3500 }
3501 /**
3502  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3503  * @p: the task in question.
3504  * @policy: new policy.
3505  * @param: structure containing the new RT priority.
3506  *
3507  * Return: 0 on success. An error code otherwise.
3508  *
3509  * NOTE that the task may be already dead.
3510  */
3511 int sched_setscheduler(struct task_struct *p, int policy,
3512                        const struct sched_param *param)
3513 {
3514         return _sched_setscheduler(p, policy, param, true);
3515 }
3516 EXPORT_SYMBOL_GPL(sched_setscheduler);
3517 
3518 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3519 {
3520         return __sched_setscheduler(p, attr, true);
3521 }
3522 EXPORT_SYMBOL_GPL(sched_setattr);
3523 
3524 /**
3525  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3526  * @p: the task in question.
3527  * @policy: new policy.
3528  * @param: structure containing the new RT priority.
3529  *
3530  * Just like sched_setscheduler, only don't bother checking if the
3531  * current context has permission.  For example, this is needed in
3532  * stop_machine(): we create temporary high priority worker threads,
3533  * but our caller might not have that capability.
3534  *
3535  * Return: 0 on success. An error code otherwise.
3536  */
3537 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3538                                const struct sched_param *param)
3539 {
3540         return _sched_setscheduler(p, policy, param, false);
3541 }
3542 
3543 static int
3544 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3545 {
3546         struct sched_param lparam;
3547         struct task_struct *p;
3548         int retval;
3549 
3550         if (!param || pid < 0)
3551                 return -EINVAL;
3552         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3553                 return -EFAULT;
3554 
3555         rcu_read_lock();
3556         retval = -ESRCH;
3557         p = find_process_by_pid(pid);
3558         if (p != NULL)
3559                 retval = sched_setscheduler(p, policy, &lparam);
3560         rcu_read_unlock();
3561 
3562         return retval;
3563 }
3564 
3565 /*
3566  * Mimics kernel/events/core.c perf_copy_attr().
3567  */
3568 static int sched_copy_attr(struct sched_attr __user *uattr,
3569                            struct sched_attr *attr)
3570 {
3571         u32 size;
3572         int ret;
3573 
3574         if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3575                 return -EFAULT;
3576 
3577         /*
3578          * zero the full structure, so that a short copy will be nice.
3579          */
3580         memset(attr, 0, sizeof(*attr));
3581 
3582         ret = get_user(size, &uattr->size);
3583         if (ret)
3584                 return ret;
3585 
3586         if (size > PAGE_SIZE)   /* silly large */
3587                 goto err_size;
3588 
3589         if (!size)              /* abi compat */
3590                 size = SCHED_ATTR_SIZE_VER0;
3591 
3592         if (size < SCHED_ATTR_SIZE_VER0)
3593                 goto err_size;
3594 
3595         /*
3596          * If we're handed a bigger struct than we know of,
3597          * ensure all the unknown bits are 0 - i.e. new
3598          * user-space does not rely on any kernel feature
3599          * extensions we dont know about yet.
3600          */
3601         if (size > sizeof(*attr)) {
3602                 unsigned char __user *addr;
3603                 unsigned char __user *end;
3604                 unsigned char val;
3605 
3606                 addr = (void __user *)uattr + sizeof(*attr);
3607                 end  = (void __user *)uattr + size;
3608 
3609                 for (; addr < end; addr++) {
3610                         ret = get_user(val, addr);
3611                         if (ret)
3612                                 return ret;
3613                         if (val)
3614                                 goto err_size;
3615                 }
3616                 size = sizeof(*attr);
3617         }
3618 
3619         ret = copy_from_user(attr, uattr, size);
3620         if (ret)
3621                 return -EFAULT;
3622 
3623         /*
3624          * XXX: do we want to be lenient like existing syscalls; or do we want
3625          * to be strict and return an error on out-of-bounds values?
3626          */
3627         attr->sched_nice = clamp(attr->sched_nice, -20, 19);
3628 
3629 out:
3630         return ret;
3631 
3632 err_size:
3633         put_user(sizeof(*attr), &uattr->size);
3634         ret = -E2BIG;
3635         goto out;
3636 }
3637 
3638 /**
3639  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3640  * @pid: the pid in question.
3641  * @policy: new policy.
3642  * @param: structure containing the new RT priority.
3643  *
3644  * Return: 0 on success. An error code otherwise.
3645  */
3646 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3647                 struct sched_param __user *, param)
3648 {
3649         /* negative values for policy are not valid */
3650         if (policy < 0)
3651                 return -EINVAL;
3652 
3653         return do_sched_setscheduler(pid, policy, param);
3654 }
3655 
3656 /**
3657  * sys_sched_setparam - set/change the RT priority of a thread
3658  * @pid: the pid in question.
3659  * @param: structure containing the new RT priority.
3660  *
3661  * Return: 0 on success. An error code otherwise.
3662  */
3663 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3664 {
3665         return do_sched_setscheduler(pid, -1, param);
3666 }
3667 
3668 /**
3669  * sys_sched_setattr - same as above, but with extended sched_attr
3670  * @pid: the pid in question.
3671  * @uattr: structure containing the extended parameters.
3672  */
3673 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3674                                unsigned int, flags)
3675 {
3676         struct sched_attr attr;
3677         struct task_struct *p;
3678         int retval;
3679 
3680         if (!uattr || pid < 0 || flags)
3681                 return -EINVAL;
3682 
3683         if (sched_copy_attr(uattr, &attr))
3684                 return -EFAULT;
3685 
3686         rcu_read_lock();
3687         retval = -ESRCH;
3688         p = find_process_by_pid(pid);
3689         if (p != NULL)
3690                 retval = sched_setattr(p, &attr);
3691         rcu_read_unlock();
3692 
3693         return retval;
3694 }
3695 
3696 /**
3697  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3698  * @pid: the pid in question.
3699  *
3700  * Return: On success, the policy of the thread. Otherwise, a negative error
3701  * code.
3702  */
3703 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3704 {
3705         struct task_struct *p;
3706         int retval;
3707 
3708         if (pid < 0)
3709                 return -EINVAL;
3710 
3711         retval = -ESRCH;
3712         rcu_read_lock();
3713         p = find_process_by_pid(pid);
3714         if (p) {
3715                 retval = security_task_getscheduler(p);
3716                 if (!retval)
3717                         retval = p->policy
3718                                 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3719         }
3720         rcu_read_unlock();
3721         return retval;
3722 }
3723 
3724 /**
3725  * sys_sched_getparam - get the RT priority of a thread
3726  * @pid: the pid in question.
3727  * @param: structure containing the RT priority.
3728  *
3729  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3730  * code.
3731  */
3732 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3733 {
3734         struct sched_param lp;
3735         struct task_struct *p;
3736         int retval;
3737 
3738         if (!param || pid < 0)
3739                 return -EINVAL;
3740 
3741         rcu_read_lock();
3742         p = find_process_by_pid(pid);
3743         retval = -ESRCH;
3744         if (!p)
3745                 goto out_unlock;
3746 
3747         retval = security_task_getscheduler(p);
3748         if (retval)
3749                 goto out_unlock;
3750 
3751         if (task_has_dl_policy(p)) {
3752                 retval = -EINVAL;
3753                 goto out_unlock;
3754         }
3755         lp.sched_priority = p->rt_priority;
3756         rcu_read_unlock();
3757 
3758         /*
3759          * This one might sleep, we cannot do it with a spinlock held ...
3760          */
3761         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3762 
3763         return retval;
3764 
3765 out_unlock:
3766         rcu_read_unlock();
3767         return retval;
3768 }
3769 
3770 static int sched_read_attr(struct sched_attr __user *uattr,
3771                            struct sched_attr *attr,
3772                            unsigned int usize)
3773 {
3774         int ret;
3775 
3776         if (!access_ok(VERIFY_WRITE, uattr, usize))
3777                 return -EFAULT;
3778 
3779         /*
3780          * If we're handed a smaller struct than we know of,
3781          * ensure all the unknown bits are 0 - i.e. old
3782          * user-space does not get uncomplete information.
3783          */
3784         if (usize < sizeof(*attr)) {
3785                 unsigned char *addr;
3786                 unsigned char *end;
3787 
3788                 addr = (void *)attr + usize;
3789                 end  = (void *)attr + sizeof(*attr);
3790 
3791                 for (; addr < end; addr++) {
3792                         if (*addr)
3793                                 goto err_size;
3794                 }
3795 
3796                 attr->size = usize;
3797         }
3798 
3799         ret = copy_to_user(uattr, attr, attr->size);
3800         if (ret)
3801                 return -EFAULT;
3802 
3803 out:
3804         return ret;
3805 
3806 err_size:
3807         ret = -E2BIG;
3808         goto out;
3809 }
3810 
3811 /**
3812  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3813  * @pid: the pid in question.
3814  * @uattr: structure containing the extended parameters.
3815  * @size: sizeof(attr) for fwd/bwd comp.
3816  */
3817 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3818                 unsigned int, size, unsigned int, flags)
3819 {
3820         struct sched_attr attr = {
3821                 .size = sizeof(struct sched_attr),
3822         };
3823         struct task_struct *p;
3824         int retval;
3825 
3826         if (!uattr || pid < 0 || size > PAGE_SIZE ||
3827             size < SCHED_ATTR_SIZE_VER0 || flags)
3828                 return -EINVAL;
3829 
3830         rcu_read_lock();
3831         p = find_process_by_pid(pid);
3832         retval = -ESRCH;
3833         if (!p)
3834                 goto out_unlock;
3835 
3836         retval = security_task_getscheduler(p);
3837         if (retval)
3838                 goto out_unlock;
3839 
3840         attr.sched_policy = p->policy;
3841         if (p->sched_reset_on_fork)
3842                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3843         if (task_has_dl_policy(p))
3844                 __getparam_dl(p, &attr);
3845         else if (task_has_rt_policy(p))
3846                 attr.sched_priority = p->rt_priority;
3847         else
3848                 attr.sched_nice = TASK_NICE(p);
3849 
3850         rcu_read_unlock();
3851 
3852         retval = sched_read_attr(uattr, &attr, size);
3853         return retval;
3854 
3855 out_unlock:
3856         rcu_read_unlock();
3857         return retval;
3858 }
3859 
3860 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3861 {
3862         cpumask_var_t cpus_allowed, new_mask;
3863         struct task_struct *p;
3864         int retval;
3865 
3866         rcu_read_lock();
3867 
3868         p = find_process_by_pid(pid);
3869         if (!p) {
3870                 rcu_read_unlock();
3871                 return -ESRCH;
3872         }
3873 
3874         /* Prevent p going away */
3875         get_task_struct(p);
3876         rcu_read_unlock();
3877 
3878         if (p->flags & PF_NO_SETAFFINITY) {
3879                 retval = -EINVAL;
3880                 goto out_put_task;
3881         }
3882         if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3883                 retval = -ENOMEM;
3884                 goto out_put_task;
3885         }
3886         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3887                 retval = -ENOMEM;
3888                 goto out_free_cpus_allowed;
3889         }
3890         retval = -EPERM;
3891         if (!check_same_owner(p)) {
3892                 rcu_read_lock();
3893                 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3894                         rcu_read_unlock();
3895                         goto out_unlock;
3896                 }
3897                 rcu_read_unlock();
3898         }
3899 
3900         retval = security_task_setscheduler(p);
3901         if (retval)
3902                 goto out_unlock;
3903 
3904 
3905         cpuset_cpus_allowed(p, cpus_allowed);
3906         cpumask_and(new_mask, in_mask, cpus_allowed);
3907 
3908         /*
3909          * Since bandwidth control happens on root_domain basis,
3910          * if admission test is enabled, we only admit -deadline
3911          * tasks allowed to run on all the CPUs in the task's
3912          * root_domain.
3913          */
3914 #ifdef CONFIG_SMP
3915         if (task_has_dl_policy(p)) {
3916                 const struct cpumask *span = task_rq(p)->rd->span;
3917 
3918                 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3919                         retval = -EBUSY;
3920                         goto out_unlock;
3921                 }
3922         }
3923 #endif
3924 again:
3925         retval = set_cpus_allowed_ptr(p, new_mask);
3926 
3927         if (!retval) {
3928                 cpuset_cpus_allowed(p, cpus_allowed);
3929                 if (!cpumask_subset(new_mask, cpus_allowed)) {
3930                         /*
3931                          * We must have raced with a concurrent cpuset
3932                          * update. Just reset the cpus_allowed to the
3933                          * cpuset's cpus_allowed
3934                          */
3935                         cpumask_copy(new_mask, cpus_allowed);
3936                         goto again;
3937                 }
3938         }
3939 out_unlock:
3940         free_cpumask_var(new_mask);
3941 out_free_cpus_allowed:
3942         free_cpumask_var(cpus_allowed);
3943 out_put_task:
3944         put_task_struct(p);
3945         return retval;
3946 }
3947 
3948 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3949                              struct cpumask *new_mask)
3950 {
3951         if (len < cpumask_size())
3952                 cpumask_clear(new_mask);
3953         else if (len > cpumask_size())
3954                 len = cpumask_size();
3955 
3956         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3957 }
3958 
3959 /**
3960  * sys_sched_setaffinity - set the cpu affinity of a process
3961  * @pid: pid of the process
3962  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3963  * @user_mask_ptr: user-space pointer to the new cpu mask
3964  *
3965  * Return: 0 on success. An error code otherwise.
3966  */
3967 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3968                 unsigned long __user *, user_mask_ptr)
3969 {
3970         cpumask_var_t new_mask;
3971         int retval;
3972 
3973         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3974                 return -ENOMEM;
3975 
3976         retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3977         if (retval == 0)
3978                 retval = sched_setaffinity(pid, new_mask);
3979         free_cpumask_var(new_mask);
3980         return retval;
3981 }
3982 
3983 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3984 {
3985         struct task_struct *p;
3986         unsigned long flags;
3987         int retval;
3988 
3989         rcu_read_lock();
3990 
3991         retval = -ESRCH;
3992         p = find_process_by_pid(pid);
3993         if (!p)
3994                 goto out_unlock;
3995 
3996         retval = security_task_getscheduler(p);
3997         if (retval)
3998                 goto out_unlock;
3999 
4000         raw_spin_lock_irqsave(&p->pi_lock, flags);
4001         cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4002         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4003 
4004 out_unlock:
4005         rcu_read_unlock();
4006 
4007         return retval;
4008 }
4009 
4010 /**
4011  * sys_sched_getaffinity - get the cpu affinity of a process
4012  * @pid: pid of the process
4013  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4014  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4015  *
4016  * Return: 0 on success. An error code otherwise.
4017  */
4018 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4019                 unsigned long __user *, user_mask_ptr)
4020 {
4021         int ret;
4022         cpumask_var_t mask;
4023 
4024         if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4025                 return -EINVAL;
4026         if (len & (sizeof(unsigned long)-1))
4027                 return -EINVAL;
4028 
4029         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4030                 return -ENOMEM;
4031 
4032         ret = sched_getaffinity(pid, mask);
4033         if (ret == 0) {
4034                 size_t retlen = min_t(size_t, len, cpumask_size());
4035 
4036                 if (copy_to_user(user_mask_ptr, mask, retlen))
4037                         ret = -EFAULT;
4038                 else
4039                         ret = retlen;
4040         }
4041         free_cpumask_var(mask);
4042 
4043         return ret;
4044 }
4045 
4046 /**
4047  * sys_sched_yield - yield the current processor to other threads.
4048  *
4049  * This function yields the current CPU to other tasks. If there are no
4050  * other threads running on this CPU then this function will return.
4051  *
4052  * Return: 0.
4053  */
4054 SYSCALL_DEFINE0(sched_yield)
4055 {
4056         struct rq *rq = this_rq_lock();
4057 
4058         schedstat_inc(rq, yld_count);
4059         current->sched_class->yield_task(rq);
4060 
4061         /*
4062          * Since we are going to call schedule() anyway, there's
4063          * no need to preempt or enable interrupts:
4064          */
4065         __release(rq->lock);
4066         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4067         do_raw_spin_unlock(&rq->lock);
4068         sched_preempt_enable_no_resched();
4069 
4070         schedule();
4071 
4072         return 0;
4073 }
4074 
4075 static void __cond_resched(void)
4076 {
4077         __preempt_count_add(PREEMPT_ACTIVE);
4078         __schedule();
4079         __preempt_count_sub(PREEMPT_ACTIVE);
4080 }
4081 
4082 int __sched _cond_resched(void)
4083 {
4084         if (should_resched()) {
4085                 __cond_resched();
4086                 return 1;
4087         }
4088         return 0;
4089 }
4090 EXPORT_SYMBOL(_cond_resched);
4091 
4092 /*
4093  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4094  * call schedule, and on return reacquire the lock.
4095  *
4096  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4097  * operations here to prevent schedule() from being called twice (once via
4098  * spin_unlock(), once by hand).
4099  */
4100 int __cond_resched_lock(spinlock_t *lock)
4101 {
4102         int resched = should_resched();
4103         int ret = 0;
4104 
4105         lockdep_assert_held(lock);
4106 
4107         if (spin_needbreak(lock) || resched) {
4108                 spin_unlock(lock);
4109                 if (resched)
4110                         __cond_resched();
4111                 else
4112                         cpu_relax();
4113                 ret = 1;
4114                 spin_lock(lock);
4115         }
4116         return ret;
4117 }
4118 EXPORT_SYMBOL(__cond_resched_lock);
4119 
4120 int __sched __cond_resched_softirq(void)
4121 {
4122         BUG_ON(!in_softirq());
4123 
4124         if (should_resched()) {
4125                 local_bh_enable();
4126                 __cond_resched();
4127                 local_bh_disable();
4128                 return 1;
4129         }
4130         return 0;
4131 }
4132 EXPORT_SYMBOL(__cond_resched_softirq);
4133 
4134 /**
4135  * yield - yield the current processor to other threads.
4136  *
4137  * Do not ever use this function, there's a 99% chance you're doing it wrong.
4138  *
4139  * The scheduler is at all times free to pick the calling task as the most
4140  * eligible task to run, if removing the yield() call from your code breaks
4141  * it, its already broken.
4142  *
4143  * Typical broken usage is:
4144  *
4145  * while (!event)
4146  *      yield();
4147  *
4148  * where one assumes that yield() will let 'the other' process run that will
4149  * make event true. If the current task is a SCHED_FIFO task that will never
4150  * happen. Never use yield() as a progress guarantee!!
4151  *
4152  * If you want to use yield() to wait for something, use wait_event().
4153  * If you want to use yield() to be 'nice' for others, use cond_resched().
4154  * If you still want to use yield(), do not!
4155  */
4156 void __sched yield(void)
4157 {
4158         set_current_state(TASK_RUNNING);
4159         sys_sched_yield();
4160 }
4161 EXPORT_SYMBOL(yield);
4162 
4163 /**
4164  * yield_to - yield the current processor to another thread in
4165  * your thread group, or accelerate that thread toward the
4166  * processor it's on.
4167  * @p: target task
4168  * @preempt: whether task preemption is allowed or not
4169  *
4170  * It's the caller's job to ensure that the target task struct
4171  * can't go away on us before we can do any checks.
4172  *
4173  * Return:
4174  *      true (>0) if we indeed boosted the target task.
4175  *      false (0) if we failed to boost the target.
4176  *      -ESRCH if there's no task to yield to.
4177  */
4178 bool __sched yield_to(struct task_struct *p, bool preempt)
4179 {
4180         struct task_struct *curr = current;
4181         struct rq *rq, *p_rq;
4182         unsigned long flags;
4183         int yielded = 0;
4184 
4185         local_irq_save(flags);
4186         rq = this_rq();
4187 
4188 again:
4189         p_rq = task_rq(p);
4190         /*
4191          * If we're the only runnable task on the rq and target rq also
4192          * has only one task, there's absolutely no point in yielding.
4193          */
4194         if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4195                 yielded = -ESRCH;
4196                 goto out_irq;
4197         }
4198 
4199         double_rq_lock(rq, p_rq);
4200         if (task_rq(p) != p_rq) {
4201                 double_rq_unlock(rq, p_rq);
4202                 goto again;
4203         }
4204 
4205         if (!curr->sched_class->yield_to_task)
4206                 goto out_unlock;
4207 
4208         if (curr->sched_class != p->sched_class)
4209                 goto out_unlock;
4210 
4211         if (task_running(p_rq, p) || p->state)
4212                 goto out_unlock;
4213 
4214         yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4215         if (yielded) {
4216                 schedstat_inc(rq, yld_count);
4217                 /*
4218                  * Make p's CPU reschedule; pick_next_entity takes care of
4219                  * fairness.
4220                  */
4221                 if (preempt && rq != p_rq)
4222                         resched_task(p_rq->curr);
4223         }
4224 
4225 out_unlock:
4226         double_rq_unlock(rq, p_rq);
4227 out_irq:
4228         local_irq_restore(flags);
4229 
4230         if (yielded > 0)
4231                 schedule();
4232 
4233         return yielded;
4234 }
4235 EXPORT_SYMBOL_GPL(yield_to);
4236 
4237 /*
4238  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4239  * that process accounting knows that this is a task in IO wait state.
4240  */
4241 void __sched io_schedule(void)
4242 {
4243         struct rq *rq = raw_rq();
4244 
4245         delayacct_blkio_start();
4246         atomic_inc(&rq->nr_iowait);
4247         blk_flush_plug(current);
4248         current->in_iowait = 1;
4249         schedule();
4250         current->in_iowait = 0;
4251         atomic_dec(&rq->nr_iowait);
4252         delayacct_blkio_end();
4253 }
4254 EXPORT_SYMBOL(io_schedule);
4255 
4256 long __sched io_schedule_timeout(long timeout)
4257 {
4258         struct rq *rq = raw_rq();
4259         long ret;
4260 
4261         delayacct_blkio_start();
4262         atomic_inc(&rq->nr_iowait);
4263         blk_flush_plug(current);
4264         current->in_iowait = 1;
4265         ret = schedule_timeout(timeout);
4266         current->in_iowait = 0;
4267         atomic_dec(&rq->nr_iowait);
4268         delayacct_blkio_end();
4269         return ret;
4270 }
4271 
4272 /**
4273  * sys_sched_get_priority_max - return maximum RT priority.
4274  * @policy: scheduling class.
4275  *
4276  * Return: On success, this syscall returns the maximum
4277  * rt_priority that can be used by a given scheduling class.
4278  * On failure, a negative error code is returned.
4279  */
4280 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4281 {
4282         int ret = -EINVAL;
4283 
4284         switch (policy) {
4285         case SCHED_FIFO:
4286         case SCHED_RR:
4287                 ret = MAX_USER_RT_PRIO-1;
4288                 break;
4289         case SCHED_DEADLINE:
4290         case SCHED_NORMAL:
4291         case SCHED_BATCH:
4292         case SCHED_IDLE:
4293                 ret = 0;
4294                 break;
4295         }
4296         return ret;
4297 }
4298 
4299 /**
4300  * sys_sched_get_priority_min - return minimum RT priority.
4301  * @policy: scheduling class.
4302  *
4303  * Return: On success, this syscall returns the minimum
4304  * rt_priority that can be used by a given scheduling class.
4305  * On failure, a negative error code is returned.
4306  */
4307 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4308 {
4309         int ret = -EINVAL;
4310 
4311         switch (policy) {
4312         case SCHED_FIFO:
4313         case SCHED_RR:
4314                 ret = 1;
4315                 break;
4316         case SCHED_DEADLINE:
4317         case SCHED_NORMAL:
4318         case SCHED_BATCH:
4319         case SCHED_IDLE:
4320                 ret = 0;
4321         }
4322         return ret;
4323 }
4324 
4325 /**
4326  * sys_sched_rr_get_interval - return the default timeslice of a process.
4327  * @pid: pid of the process.
4328  * @interval: userspace pointer to the timeslice value.
4329  *
4330  * this syscall writes the default timeslice value of a given process
4331  * into the user-space timespec buffer. A value of '' means infinity.
4332  *
4333  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4334  * an error code.
4335  */
4336 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4337                 struct timespec __user *, interval)
4338 {
4339         struct task_struct *p;
4340         unsigned int time_slice;
4341         unsigned long flags;
4342         struct rq *rq;
4343         int retval;
4344         struct timespec t;
4345 
4346         if (pid < 0)
4347                 return -EINVAL;
4348 
4349         retval = -ESRCH;
4350         rcu_read_lock();
4351         p = find_process_by_pid(pid);
4352         if (!p)
4353                 goto out_unlock;
4354 
4355         retval = security_task_getscheduler(p);
4356         if (retval)
4357                 goto out_unlock;
4358 
4359         rq = task_rq_lock(p, &flags);
4360         time_slice = 0;
4361         if (p->sched_class->get_rr_interval)
4362                 time_slice = p->sched_class->get_rr_interval(rq, p);
4363         task_rq_unlock(rq, p, &flags);
4364 
4365         rcu_read_unlock();
4366         jiffies_to_timespec(time_slice, &t);
4367         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4368         return retval;
4369 
4370 out_unlock:
4371         rcu_read_unlock();
4372         return retval;
4373 }
4374 
4375 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4376 
4377 void sched_show_task(struct task_struct *p)
4378 {
4379         unsigned long free = 0;
4380         int ppid;
4381         unsigned state;
4382 
4383         state = p->state ? __ffs(p->state) + 1 : 0;
4384         printk(KERN_INFO "%-15.15s %c", p->comm,
4385                 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4386 #if BITS_PER_LONG == 32
4387         if (state == TASK_RUNNING)
4388                 printk(KERN_CONT " running  ");
4389         else
4390                 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4391 #else
4392         if (state == TASK_RUNNING)
4393                 printk(KERN_CONT "  running task    ");
4394         else
4395                 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4396 #endif
4397 #ifdef CONFIG_DEBUG_STACK_USAGE
4398         free = stack_not_used(p);
4399 #endif
4400         rcu_read_lock();
4401         ppid = task_pid_nr(rcu_dereference(p->real_parent));
4402         rcu_read_unlock();
4403         printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4404                 task_pid_nr(p), ppid,
4405                 (unsigned long)task_thread_info(p)->flags);
4406 
4407         print_worker_info(KERN_INFO, p);
4408         show_stack(p, NULL);
4409 }
4410 
4411 void show_state_filter(unsigned long state_filter)
4412 {
4413         struct task_struct *g, *p;
4414 
4415 #if BITS_PER_LONG == 32
4416         printk(KERN_INFO
4417                 "  task                PC stack   pid father\n");
4418 #else
4419         printk(KERN_INFO
4420                 "  task                        PC stack   pid father\n");
4421 #endif
4422         rcu_read_lock();
4423         do_each_thread(g, p) {
4424                 /*
4425                  * reset the NMI-timeout, listing all files on a slow
4426                  * console might take a lot of time:
4427                  */
4428                 touch_nmi_watchdog();
4429                 if (!state_filter || (p->state & state_filter))
4430                         sched_show_task(p);
4431         } while_each_thread(g, p);
4432 
4433         touch_all_softlockup_watchdogs();
4434 
4435 #ifdef CONFIG_SCHED_DEBUG
4436         sysrq_sched_debug_show();
4437 #endif
4438         rcu_read_unlock();
4439         /*
4440          * Only show locks if all tasks are dumped:
4441          */
4442         if (!state_filter)
4443                 debug_show_all_locks();
4444 }
4445 
4446 void init_idle_bootup_task(struct task_struct *idle)
4447 {
4448         idle->sched_class = &idle_sched_class;
4449 }
4450 
4451 /**
4452  * init_idle - set up an idle thread for a given CPU
4453  * @idle: task in question
4454  * @cpu: cpu the idle task belongs to
4455  *
4456  * NOTE: this function does not set the idle thread's NEED_RESCHED
4457  * flag, to make booting more robust.
4458  */
4459 void init_idle(struct task_struct *idle, int cpu)
4460 {
4461         struct rq *rq = cpu_rq(cpu);
4462         unsigned long flags;
4463 
4464         raw_spin_lock_irqsave(&rq->lock, flags);
4465 
4466         __sched_fork(0, idle);
4467         idle->state = TASK_RUNNING;
4468         idle->se.exec_start = sched_clock();
4469 
4470         do_set_cpus_allowed(idle, cpumask_of(cpu));
4471         /*
4472          * We're having a chicken and egg problem, even though we are
4473          * holding rq->lock, the cpu isn't yet set to this cpu so the
4474          * lockdep check in task_group() will fail.
4475          *
4476          * Similar case to sched_fork(). / Alternatively we could
4477          * use task_rq_lock() here and obtain the other rq->lock.
4478          *
4479          * Silence PROVE_RCU
4480          */
4481         rcu_read_lock();
4482         __set_task_cpu(idle, cpu);
4483         rcu_read_unlock();
4484 
4485         rq->curr = rq->idle = idle;
4486 #if defined(CONFIG_SMP)
4487         idle->on_cpu = 1;
4488 #endif
4489         raw_spin_unlock_irqrestore(&rq->lock, flags);
4490 
4491         /* Set the preempt count _outside_ the spinlocks! */
4492         init_idle_preempt_count(idle, cpu);
4493 
4494         /*
4495          * The idle tasks have their own, simple scheduling class:
4496          */
4497         idle->sched_class = &idle_sched_class;
4498         ftrace_graph_init_idle_task(idle, cpu);
4499         vtime_init_idle(idle, cpu);
4500 #if defined(CONFIG_SMP)
4501         sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4502 #endif
4503 }
4504 
4505 #ifdef CONFIG_SMP
4506 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4507 {
4508         if (p->sched_class && p->sched_class->set_cpus_allowed)
4509                 p->sched_class->set_cpus_allowed(p, new_mask);
4510 
4511         cpumask_copy(&p->cpus_allowed, new_mask);
4512         p->nr_cpus_allowed = cpumask_weight(new_mask);
4513 }
4514 
4515 /*
4516  * This is how migration works:
4517  *
4518  * 1) we invoke migration_cpu_stop() on the target CPU using
4519  *    stop_one_cpu().
4520  * 2) stopper starts to run (implicitly forcing the migrated thread
4521  *    off the CPU)
4522  * 3) it checks whether the migrated task is still in the wrong runqueue.
4523  * 4) if it's in the wrong runqueue then the migration thread removes
4524  *    it and puts it into the right queue.
4525  * 5) stopper completes and stop_one_cpu() returns and the migration
4526  *    is done.
4527  */
4528 
4529 /*
4530  * Change a given task's CPU affinity. Migrate the thread to a
4531  * proper CPU and schedule it away if the CPU it's executing on
4532  * is removed from the allowed bitmask.
4533  *
4534  * NOTE: the caller must have a valid reference to the task, the
4535  * task must not exit() & deallocate itself prematurely. The
4536  * call is not atomic; no spinlocks may be held.
4537  */
4538 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4539 {
4540         unsigned long flags;
4541         struct rq *rq;
4542         unsigned int dest_cpu;
4543         int ret = 0;
4544 
4545         rq = task_rq_lock(p, &flags);
4546 
4547         if (cpumask_equal(&p->cpus_allowed, new_mask))
4548                 goto out;
4549 
4550         if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4551                 ret = -EINVAL;
4552                 goto out;
4553         }
4554 
4555         do_set_cpus_allowed(p, new_mask);
4556 
4557         /* Can the task run on the task's current CPU? If so, we're done */
4558         if (cpumask_test_cpu(task_cpu(p), new_mask))
4559                 goto out;
4560 
4561         dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4562         if (p->on_rq) {
4563                 struct migration_arg arg = { p, dest_cpu };
4564                 /* Need help from migration thread: drop lock and wait. */
4565                 task_rq_unlock(rq, p, &flags);
4566                 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4567                 tlb_migrate_finish(p->mm);
4568                 return 0;
4569         }
4570 out:
4571         task_rq_unlock(rq, p, &flags);
4572 
4573         return ret;
4574 }
4575 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4576 
4577 /*
4578  * Move (not current) task off this cpu, onto dest cpu. We're doing
4579  * this because either it can't run here any more (set_cpus_allowed()
4580  * away from this CPU, or CPU going down), or because we're
4581  * attempting to rebalance this task on exec (sched_exec).
4582  *
4583  * So we race with normal scheduler movements, but that's OK, as long
4584  * as the task is no longer on this CPU.
4585  *
4586  * Returns non-zero if task was successfully migrated.
4587  */
4588 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4589 {
4590         struct rq *rq_dest, *rq_src;
4591         int ret = 0;
4592 
4593         if (unlikely(!cpu_active(dest_cpu)))
4594                 return ret;
4595 
4596         rq_src = cpu_rq(src_cpu);
4597         rq_dest = cpu_rq(dest_cpu);
4598 
4599         raw_spin_lock(&p->pi_lock);
4600         double_rq_lock(rq_src, rq_dest);
4601         /* Already moved. */
4602         if (task_cpu(p) != src_cpu)
4603                 goto done;
4604         /* Affinity changed (again). */
4605         if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4606                 goto fail;
4607 
4608         /*
4609          * If we're not on a rq, the next wake-up will ensure we're
4610          * placed properly.
4611          */
4612         if (p->on_rq) {
4613                 dequeue_task(rq_src, p, 0);
4614                 set_task_cpu(p, dest_cpu);
4615                 enqueue_task(rq_dest, p, 0);
4616                 check_preempt_curr(rq_dest, p, 0);
4617         }
4618 done:
4619         ret = 1;
4620 fail:
4621         double_rq_unlock(rq_src, rq_dest);
4622         raw_spin_unlock(&p->pi_lock);
4623         return ret;
4624 }
4625 
4626 #ifdef CONFIG_NUMA_BALANCING
4627 /* Migrate current task p to target_cpu */
4628 int migrate_task_to(struct task_struct *p, int target_cpu)
4629 {
4630         struct migration_arg arg = { p, target_cpu };
4631         int curr_cpu = task_cpu(p);
4632 
4633         if (curr_cpu == target_cpu)
4634                 return 0;
4635 
4636         if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4637                 return -EINVAL;
4638 
4639         /* TODO: This is not properly updating schedstats */
4640 
4641         trace_sched_move_numa(p, curr_cpu, target_cpu);
4642         return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4643 }
4644 
4645 /*
4646  * Requeue a task on a given node and accurately track the number of NUMA
4647  * tasks on the runqueues
4648  */
4649 void sched_setnuma(struct task_struct *p, int nid)
4650 {
4651         struct rq *rq;
4652         unsigned long flags;
4653         bool on_rq, running;
4654 
4655         rq = task_rq_lock(p, &flags);
4656         on_rq = p->on_rq;
4657         running = task_current(rq, p);
4658 
4659         if (on_rq)
4660                 dequeue_task(rq, p, 0);
4661         if (running)
4662                 p->sched_class->put_prev_task(rq, p);
4663 
4664         p->numa_preferred_nid = nid;
4665 
4666         if (running)
4667                 p->sched_class->set_curr_task(rq);
4668         if (on_rq)
4669                 enqueue_task(rq, p, 0);
4670         task_rq_unlock(rq, p, &flags);
4671 }
4672 #endif
4673 
4674 /*
4675  * migration_cpu_stop - this will be executed by a highprio stopper thread
4676  * and performs thread migration by bumping thread off CPU then
4677  * 'pushing' onto another runqueue.
4678  */
4679 static int migration_cpu_stop(void *data)
4680 {
4681         struct migration_arg *arg = data;
4682 
4683         /*
4684          * The original target cpu might have gone down and we might
4685          * be on another cpu but it doesn't matter.
4686          */
4687         local_irq_disable();
4688         __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4689         local_irq_enable();
4690         return 0;
4691 }
4692 
4693 #ifdef CONFIG_HOTPLUG_CPU
4694 
4695 /*
4696  * Ensures that the idle task is using init_mm right before its cpu goes
4697  * offline.
4698  */
4699 void idle_task_exit(void)
4700 {
4701         struct mm_struct *mm = current->active_mm;
4702 
4703         BUG_ON(cpu_online(smp_processor_id()));
4704 
4705         if (mm != &init_mm)
4706                 switch_mm(mm, &init_mm, current);
4707         mmdrop(mm);
4708 }
4709 
4710 /*
4711  * Since this CPU is going 'away' for a while, fold any nr_active delta
4712  * we might have. Assumes we're called after migrate_tasks() so that the
4713  * nr_active count is stable.
4714  *
4715  * Also see the comment "Global load-average calculations".
4716  */
4717 static void calc_load_migrate(struct rq *rq)
4718 {
4719         long delta = calc_load_fold_active(rq);
4720         if (delta)
4721                 atomic_long_add(delta, &calc_load_tasks);
4722 }
4723 
4724 /*
4725  * Migrate all tasks from the rq, sleeping tasks will be migrated by
4726  * try_to_wake_up()->select_task_rq().
4727  *
4728  * Called with rq->lock held even though we'er in stop_machine() and
4729  * there's no concurrency possible, we hold the required locks anyway
4730  * because of lock validation efforts.
4731  */
4732 static void migrate_tasks(unsigned int dead_cpu)
4733 {
4734         struct rq *rq = cpu_rq(dead_cpu);
4735         struct task_struct *next, *stop = rq->stop;
4736         int dest_cpu;
4737 
4738         /*
4739          * Fudge the rq selection such that the below task selection loop
4740          * doesn't get stuck on the currently eligible stop task.
4741          *
4742          * We're currently inside stop_machine() and the rq is either stuck
4743          * in the stop_machine_cpu_stop() loop, or we're executing this code,
4744          * either way we should never end up calling schedule() until we're
4745          * done here.
4746          */
4747         rq->stop = NULL;
4748 
4749         /*
4750          * put_prev_task() and pick_next_task() sched
4751          * class method both need to have an up-to-date
4752          * value of rq->clock[_task]
4753          */
4754         update_rq_clock(rq);
4755 
4756         for ( ; ; ) {
4757                 /*
4758                  * There's this thread running, bail when that's the only
4759                  * remaining thread.
4760                  */
4761                 if (rq->nr_running == 1)
4762                         break;
4763 
4764                 next = pick_next_task(rq);
4765                 BUG_ON(!next);
4766                 next->sched_class->put_prev_task(rq, next);
4767 
4768                 /* Find suitable destination for @next, with force if needed. */
4769                 dest_cpu = select_fallback_rq(dead_cpu, next);
4770                 raw_spin_unlock(&rq->lock);
4771 
4772                 __migrate_task(next, dead_cpu, dest_cpu);
4773 
4774                 raw_spin_lock(&rq->lock);
4775         }
4776 
4777         rq->stop = stop;
4778 }
4779 
4780 #endif /* CONFIG_HOTPLUG_CPU */
4781 
4782 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4783 
4784 static struct ctl_table sd_ctl_dir[] = {
4785         {
4786                 .procname       = "sched_domain",
4787                 .mode           = 0555,
4788         },
4789         {}
4790 };
4791 
4792 static struct ctl_table sd_ctl_root[] = {
4793         {
4794                 .procname       = "kernel",
4795                 .mode           = 0555,
4796                 .child          = sd_ctl_dir,
4797         },
4798         {}
4799 };
4800 
4801 static struct ctl_table *sd_alloc_ctl_entry(int n)
4802 {
4803         struct ctl_table *entry =
4804                 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4805 
4806         return entry;
4807 }
4808 
4809 static void sd_free_ctl_entry(struct ctl_table **tablep)
4810 {
4811         struct ctl_table *entry;
4812 
4813         /*
4814          * In the intermediate directories, both the child directory and
4815          * procname are dynamically allocated and could fail but the mode
4816          * will always be set. In the lowest directory the names are
4817          * static strings and all have proc handlers.
4818          */
4819         for (entry = *tablep; entry->mode; entry++) {
4820                 if (entry->child)
4821                         sd_free_ctl_entry(&entry->child);
4822                 if (entry->proc_handler == NULL)
4823                         kfree(entry->procname);
4824         }
4825 
4826         kfree(*tablep);
4827         *tablep = NULL;
4828 }
4829 
4830 static int min_load_idx = 0;
4831 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4832 
4833 static void
4834 set_table_entry(struct ctl_table *entry,
4835                 const char *procname, void *data, int maxlen,
4836                 umode_t mode, proc_handler *proc_handler,
4837                 bool load_idx)
4838 {
4839         entry->procname = procname;
4840         entry->data = data;
4841         entry->maxlen = maxlen;
4842         entry->mode = mode;
4843         entry->proc_handler = proc_handler;
4844 
4845         if (load_idx) {
4846                 entry->extra1 = &min_load_idx;
4847                 entry->extra2 = &max_load_idx;
4848         }
4849 }
4850 
4851 static struct ctl_table *
4852 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4853 {
4854         struct ctl_table *table = sd_alloc_ctl_entry(13);
4855 
4856         if (table == NULL)
4857                 return NULL;
4858 
4859         set_table_entry(&table[0], "min_interval", &sd->min_interval,
4860                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4861         set_table_entry(&table[1], "max_interval", &sd->max_interval,
4862                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4863         set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4864                 sizeof(int), 0644, proc_dointvec_minmax, true);
4865         set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4866                 sizeof(int), 0644, proc_dointvec_minmax, true);
4867         set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4868                 sizeof(int), 0644, proc_dointvec_minmax, true);
4869         set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4870                 sizeof(int), 0644, proc_dointvec_minmax, true);
4871         set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4872                 sizeof(int), 0644, proc_dointvec_minmax, true);
4873         set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4874                 sizeof(int), 0644, proc_dointvec_minmax, false);
4875         set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4876                 sizeof(int), 0644, proc_dointvec_minmax, false);
4877         set_table_entry(&table[9], "cache_nice_tries",
4878                 &sd->cache_nice_tries,
4879                 sizeof(int), 0644, proc_dointvec_minmax, false);
4880         set_table_entry(&table[10], "flags", &sd->flags,
4881                 sizeof(int), 0644, proc_dointvec_minmax, false);
4882         set_table_entry(&table[11], "name", sd->name,
4883                 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4884         /* &table[12] is terminator */
4885 
4886         return table;
4887 }
4888 
4889 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4890 {
4891         struct ctl_table *entry, *table;
4892         struct sched_domain *sd;
4893         int domain_num = 0, i;
4894         char buf[32];
4895 
4896         for_each_domain(cpu, sd)
4897                 domain_num++;
4898         entry = table = sd_alloc_ctl_entry(domain_num + 1);
4899         if (table == NULL)
4900                 return NULL;
4901 
4902         i = 0;
4903         for_each_domain(cpu, sd) {
4904                 snprintf(buf, 32, "domain%d", i);
4905                 entry->procname = kstrdup(buf, GFP_KERNEL);
4906                 entry->mode = 0555;
4907                 entry->child = sd_alloc_ctl_domain_table(sd);
4908                 entry++;
4909                 i++;
4910         }
4911         return table;
4912 }
4913 
4914 static struct ctl_table_header *sd_sysctl_header;
4915 static void register_sched_domain_sysctl(void)
4916 {
4917         int i, cpu_num = num_possible_cpus();
4918         struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4919         char buf[32];
4920 
4921         WARN_ON(sd_ctl_dir[0].child);
4922         sd_ctl_dir[0].child = entry;
4923 
4924         if (entry == NULL)
4925                 return;
4926 
4927         for_each_possible_cpu(i) {
4928                 snprintf(buf, 32, "cpu%d", i);
4929                 entry->procname = kstrdup(buf, GFP_KERNEL);
4930                 entry->mode = 0555;
4931                 entry->child = sd_alloc_ctl_cpu_table(i);
4932                 entry++;
4933         }
4934 
4935         WARN_ON(sd_sysctl_header);
4936         sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4937 }
4938 
4939 /* may be called multiple times per register */
4940 static void unregister_sched_domain_sysctl(void)
4941 {
4942         if (sd_sysctl_header)
4943                 unregister_sysctl_table(sd_sysctl_header);
4944         sd_sysctl_header = NULL;
4945         if (sd_ctl_dir[0].child)
4946                 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4947 }
4948 #else
4949 static void register_sched_domain_sysctl(void)
4950 {
4951 }
4952 static void unregister_sched_domain_sysctl(void)
4953 {
4954 }
4955 #endif
4956 
4957 static void set_rq_online(struct rq *rq)
4958 {
4959         if (!rq->online) {
4960                 const struct sched_class *class;
4961 
4962                 cpumask_set_cpu(rq->cpu, rq->rd->online);
4963                 rq->online = 1;
4964 
4965                 for_each_class(class) {
4966                         if (class->rq_online)
4967                                 class->rq_online(rq);
4968                 }
4969         }
4970 }
4971 
4972 static void set_rq_offline(struct rq *rq)
4973 {
4974         if (rq->online) {
4975                 const struct sched_class *class;
4976 
4977                 for_each_class(class) {
4978                         if (class->rq_offline)
4979                                 class->rq_offline(rq);
4980                 }
4981 
4982                 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4983                 rq->online = 0;
4984         }
4985 }
4986 
4987 /*
4988  * migration_call - callback that gets triggered when a CPU is added.
4989  * Here we can start up the necessary migration thread for the new CPU.
4990  */
4991 static int
4992 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4993 {
4994         int cpu = (long)hcpu;
4995         unsigned long flags;
4996         struct rq *rq = cpu_rq(cpu);
4997 
4998         switch (action & ~CPU_TASKS_FROZEN) {
4999 
5000         case CPU_UP_PREPARE:
5001                 rq->calc_load_update = calc_load_update;
5002                 break;
5003 
5004         case CPU_ONLINE:
5005                 /* Update our root-domain */
5006                 raw_spin_lock_irqsave(&rq->lock, flags);
5007                 if (rq->rd) {
5008                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5009 
5010                         set_rq_online(rq);
5011                 }
5012                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5013                 break;
5014 
5015 #ifdef CONFIG_HOTPLUG_CPU
5016         case CPU_DYING:
5017                 sched_ttwu_pending();
5018                 /* Update our root-domain */
5019                 raw_spin_lock_irqsave(&rq->lock, flags);
5020                 if (rq->rd) {
5021                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5022                         set_rq_offline(rq);
5023                 }
5024                 migrate_tasks(cpu);
5025                 BUG_ON(rq->nr_running != 1); /* the migration thread */
5026                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5027                 break;
5028 
5029         case CPU_DEAD:
5030                 calc_load_migrate(rq);
5031                 break;
5032 #endif
5033         }
5034 
5035         update_max_interval();
5036 
5037         return NOTIFY_OK;
5038 }
5039 
5040 /*
5041  * Register at high priority so that task migration (migrate_all_tasks)
5042  * happens before everything else.  This has to be lower priority than
5043  * the notifier in the perf_event subsystem, though.
5044  */
5045 static struct notifier_block migration_notifier = {
5046         .notifier_call = migration_call,
5047         .priority = CPU_PRI_MIGRATION,
5048 };
5049 
5050 static int sched_cpu_active(struct notifier_block *nfb,
5051                                       unsigned long action, void *hcpu)
5052 {
5053         switch (action & ~CPU_TASKS_FROZEN) {
5054         case CPU_STARTING:
5055         case CPU_DOWN_FAILED:
5056                 set_cpu_active((long)hcpu, true);
5057                 return NOTIFY_OK;
5058         default:
5059                 return NOTIFY_DONE;
5060         }
5061 }
5062 
5063 static int sched_cpu_inactive(struct notifier_block *nfb,
5064                                         unsigned long action, void *hcpu)
5065 {
5066         unsigned long flags;
5067         long cpu = (long)hcpu;
5068 
5069         switch (action & ~CPU_TASKS_FROZEN) {
5070         case CPU_DOWN_PREPARE:
5071                 set_cpu_active(cpu, false);
5072 
5073                 /* explicitly allow suspend */
5074                 if (!(action & CPU_TASKS_FROZEN)) {
5075                         struct dl_bw *dl_b = dl_bw_of(cpu);
5076                         bool overflow;
5077                         int cpus;
5078 
5079                         raw_spin_lock_irqsave(&dl_b->lock, flags);
5080                         cpus = dl_bw_cpus(cpu);
5081                         overflow = __dl_overflow(dl_b, cpus, 0, 0);
5082                         raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5083 
5084                         if (overflow)
5085                                 return notifier_from_errno(-EBUSY);
5086                 }
5087                 return NOTIFY_OK;
5088         }
5089 
5090         return NOTIFY_DONE;
5091 }
5092 
5093 static int __init migration_init(void)
5094 {
5095         void *cpu = (void *)(long)smp_processor_id();
5096         int err;
5097 
5098         /* Initialize migration for the boot CPU */
5099         err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5100         BUG_ON(err == NOTIFY_BAD);
5101         migration_call(&migration_notifier, CPU_ONLINE, cpu);
5102         register_cpu_notifier(&migration_notifier);
5103 
5104         /* Register cpu active notifiers */
5105         cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5106         cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5107 
5108         return 0;
5109 }
5110 early_initcall(migration_init);
5111 #endif
5112 
5113 #ifdef CONFIG_SMP
5114 
5115 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5116 
5117 #ifdef CONFIG_SCHED_DEBUG
5118 
5119 static __read_mostly int sched_debug_enabled;
5120 
5121 static int __init sched_debug_setup(char *str)
5122 {
5123         sched_debug_enabled = 1;
5124 
5125         return 0;
5126 }
5127 early_param("sched_debug", sched_debug_setup);
5128 
5129 static inline bool sched_debug(void)
5130 {
5131         return sched_debug_enabled;
5132 }
5133 
5134 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5135                                   struct cpumask *groupmask)
5136 {
5137         struct sched_group *group = sd->groups;
5138         char str[256];
5139 
5140         cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5141         cpumask_clear(groupmask);
5142 
5143         printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5144 
5145         if (!(sd->flags & SD_LOAD_BALANCE)) {
5146                 printk("does not load-balance\n");
5147                 if (sd->parent)
5148                         printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5149                                         " has parent");
5150                 return -1;
5151         }
5152 
5153         printk(KERN_CONT "span %s level %s\n", str, sd->name);
5154 
5155         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5156                 printk(KERN_ERR "ERROR: domain->span does not contain "
5157                                 "CPU%d\n", cpu);
5158         }
5159         if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5160                 printk(KERN_ERR "ERROR: domain->groups does not contain"
5161                                 " CPU%d\n", cpu);
5162         }
5163 
5164         printk(KERN_DEBUG "%*s groups:", level + 1, "");
5165         do {
5166                 if (!group) {
5167                         printk("\n");
5168                         printk(KERN_ERR "ERROR: group is NULL\n");
5169                         break;
5170                 }
5171 
5172                 /*
5173                  * Even though we initialize ->power to something semi-sane,
5174                  * we leave power_orig unset. This allows us to detect if
5175                  * domain iteration is still funny without causing /0 traps.
5176                  */
5177                 if (!group->sgp->power_orig) {
5178                         printk(KERN_CONT "\n");
5179                         printk(KERN_ERR "ERROR: domain->cpu_power not "
5180                                         "set\n");
5181                         break;
5182                 }
5183 
5184                 if (!cpumask_weight(sched_group_cpus(group))) {
5185                         printk(KERN_CONT "\n");
5186                         printk(KERN_ERR "ERROR: empty group\n");
5187                         break;
5188                 }
5189 
5190                 if (!(sd->flags & SD_OVERLAP) &&
5191                     cpumask_intersects(groupmask, sched_group_cpus(group))) {
5192                         printk(KERN_CONT "\n");
5193                         printk(KERN_ERR "ERROR: repeated CPUs\n");
5194                         break;
5195                 }
5196 
5197                 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5198 
5199                 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5200 
5201                 printk(KERN_CONT " %s", str);
5202                 if (group->sgp->power != SCHED_POWER_SCALE) {
5203                         printk(KERN_CONT " (cpu_power = %d)",
5204                                 group->sgp->power);
5205                 }
5206 
5207                 group = group->next;
5208         } while (group != sd->groups);
5209         printk(KERN_CONT "\n");
5210 
5211         if (!cpumask_equal(sched_domain_span(sd), groupmask))
5212                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5213 
5214         if (sd->parent &&
5215             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5216                 printk(KERN_ERR "ERROR: parent span is not a superset "
5217                         "of domain->span\n");
5218         return 0;
5219 }
5220 
5221 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5222 {
5223         int level = 0;
5224 
5225         if (!sched_debug_enabled)
5226                 return;
5227 
5228         if (!sd) {
5229                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5230                 return;
5231         }
5232 
5233         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5234 
5235         for (;;) {
5236                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5237                         break;
5238                 level++;
5239                 sd = sd->parent;
5240                 if (!sd)
5241                         break;
5242         }
5243 }
5244 #else /* !CONFIG_SCHED_DEBUG */
5245 # define sched_domain_debug(sd, cpu) do { } while (0)
5246 static inline bool sched_debug(void)
5247 {
5248         return false;
5249 }
5250 #endif /* CONFIG_SCHED_DEBUG */
5251 
5252 static int sd_degenerate(struct sched_domain *sd)
5253 {
5254         if (cpumask_weight(sched_domain_span(sd)) == 1)
5255                 return 1;
5256 
5257         /* Following flags need at least 2 groups */
5258         if (sd->flags & (SD_LOAD_BALANCE |
5259                          SD_BALANCE_NEWIDLE |
5260                          SD_BALANCE_FORK |
5261                          SD_BALANCE_EXEC |
5262                          SD_SHARE_CPUPOWER |
5263                          SD_SHARE_PKG_RESOURCES)) {
5264                 if (sd->groups != sd->groups->next)
5265                         return 0;
5266         }
5267 
5268         /* Following flags don't use groups */
5269         if (sd->flags & (SD_WAKE_AFFINE))
5270                 return 0;
5271 
5272         return 1;
5273 }
5274 
5275 static int
5276 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5277 {
5278         unsigned long cflags = sd->flags, pflags = parent->flags;
5279 
5280         if (sd_degenerate(parent))
5281                 return 1;
5282 
5283         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5284                 return 0;
5285 
5286         /* Flags needing groups don't count if only 1 group in parent */
5287         if (parent->groups == parent->groups->next) {
5288                 pflags &= ~(SD_LOAD_BALANCE |
5289                                 SD_BALANCE_NEWIDLE |
5290                                 SD_BALANCE_FORK |
5291                                 SD_BALANCE_EXEC |
5292                                 SD_SHARE_CPUPOWER |
5293                                 SD_SHARE_PKG_RESOURCES |
5294                                 SD_PREFER_SIBLING);
5295                 if (nr_node_ids == 1)
5296                         pflags &= ~SD_SERIALIZE;
5297         }
5298         if (~cflags & pflags)
5299                 return 0;
5300 
5301         return 1;
5302 }
5303 
5304 static void free_rootdomain(struct rcu_head *rcu)
5305 {
5306         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5307 
5308         cpupri_cleanup(&rd->cpupri);
5309         cpudl_cleanup(&rd->cpudl);
5310         free_cpumask_var(rd->dlo_mask);
5311         free_cpumask_var(rd->rto_mask);
5312         free_cpumask_var(rd->online);
5313         free_cpumask_var(rd->span);
5314         kfree(rd);
5315 }
5316 
5317 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5318 {
5319         struct root_domain *old_rd = NULL;
5320         unsigned long flags;
5321 
5322         raw_spin_lock_irqsave(&rq->lock, flags);
5323 
5324         if (rq->rd) {
5325                 old_rd = rq->rd;
5326 
5327                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5328                         set_rq_offline(rq);
5329 
5330                 cpumask_clear_cpu(rq->cpu, old_rd->span);
5331 
5332                 /*
5333                  * If we dont want to free the old_rd yet then
5334                  * set old_rd to NULL to skip the freeing later
5335                  * in this function:
5336                  */
5337                 if (!atomic_dec_and_test(&old_rd->refcount))
5338                         old_rd = NULL;
5339         }
5340 
5341         atomic_inc(&rd->refcount);
5342         rq->rd = rd;
5343 
5344         cpumask_set_cpu(rq->cpu, rd->span);
5345         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5346                 set_rq_online(rq);
5347 
5348         raw_spin_unlock_irqrestore(&rq->lock, flags);
5349 
5350         if (old_rd)
5351                 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5352 }
5353 
5354 static int init_rootdomain(struct root_domain *rd)
5355 {
5356         memset(rd, 0, sizeof(*rd));
5357 
5358         if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5359                 goto out;
5360         if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5361                 goto free_span;
5362         if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5363                 goto free_online;
5364         if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5365                 goto free_dlo_mask;
5366 
5367         init_dl_bw(&rd->dl_bw);
5368         if (cpudl_init(&rd->cpudl) != 0)
5369                 goto free_dlo_mask;
5370 
5371         if (cpupri_init(&rd->cpupri) != 0)
5372                 goto free_rto_mask;
5373         return 0;
5374 
5375 free_rto_mask:
5376         free_cpumask_var(rd->rto_mask);
5377 free_dlo_mask:
5378         free_cpumask_var(rd->dlo_mask);
5379 free_online:
5380         free_cpumask_var(rd->online);
5381 free_span:
5382         free_cpumask_var(rd->span);
5383 out:
5384         return -ENOMEM;
5385 }
5386 
5387 /*
5388  * By default the system creates a single root-domain with all cpus as
5389  * members (mimicking the global state we have today).
5390  */
5391 struct root_domain def_root_domain;
5392 
5393 static void init_defrootdomain(void)
5394 {
5395         init_rootdomain(&def_root_domain);
5396 
5397         atomic_set(&def_root_domain.refcount, 1);
5398 }
5399 
5400 static struct root_domain *alloc_rootdomain(void)
5401 {
5402         struct root_domain *rd;
5403 
5404         rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5405         if (!rd)
5406                 return NULL;
5407 
5408         if (init_rootdomain(rd) != 0) {
5409                 kfree(rd);
5410                 return NULL;
5411         }
5412 
5413         return rd;
5414 }
5415 
5416 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5417 {
5418         struct sched_group *tmp, *first;
5419 
5420         if (!sg)
5421                 return;
5422 
5423         first = sg;
5424         do {
5425                 tmp = sg->next;
5426 
5427                 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5428                         kfree(sg->sgp);
5429 
5430                 kfree(sg);
5431                 sg = tmp;
5432         } while (sg != first);
5433 }
5434 
5435 static void free_sched_domain(struct rcu_head *rcu)
5436 {
5437         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5438 
5439         /*
5440          * If its an overlapping domain it has private groups, iterate and
5441          * nuke them all.
5442          */
5443         if (sd->flags & SD_OVERLAP) {
5444                 free_sched_groups(sd->groups, 1);
5445         } else if (atomic_dec_and_test(&sd->groups->ref)) {
5446                 kfree(sd->groups->sgp);
5447                 kfree(sd->groups);
5448         }
5449         kfree(sd);
5450 }
5451 
5452 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5453 {
5454         call_rcu(&sd->rcu, free_sched_domain);
5455 }
5456 
5457 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5458 {
5459         for (; sd; sd = sd->parent)
5460                 destroy_sched_domain(sd, cpu);
5461 }
5462 
5463 /*
5464  * Keep a special pointer to the highest sched_domain that has
5465  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5466  * allows us to avoid some pointer chasing select_idle_sibling().
5467  *
5468  * Also keep a unique ID per domain (we use the first cpu number in
5469  * the cpumask of the domain), this allows us to quickly tell if
5470  * two cpus are in the same cache domain, see cpus_share_cache().
5471  */
5472 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5473 DEFINE_PER_CPU(int, sd_llc_size);
5474 DEFINE_PER_CPU(int, sd_llc_id);
5475 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5476 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5477 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5478 
5479 static void update_top_cache_domain(int cpu)
5480 {
5481         struct sched_domain *sd;
5482         struct sched_domain *busy_sd = NULL;
5483         int id = cpu;
5484         int size = 1;
5485 
5486         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5487         if (sd) {
5488                 id = cpumask_first(sched_domain_span(sd));
5489                 size = cpumask_weight(sched_domain_span(sd));
5490                 busy_sd = sd->parent; /* sd_busy */
5491         }
5492         rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5493 
5494         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5495         per_cpu(sd_llc_size, cpu) = size;
5496         per_cpu(sd_llc_id, cpu) = id;
5497 
5498         sd = lowest_flag_domain(cpu, SD_NUMA);
5499         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5500 
5501         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5502         rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5503 }
5504 
5505 /*
5506  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5507  * hold the hotplug lock.
5508  */
5509 static void
5510 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5511 {
5512         struct rq *rq = cpu_rq(cpu);
5513         struct sched_domain *tmp;
5514 
5515         /* Remove the sched domains which do not contribute to scheduling. */
5516         for (tmp = sd; tmp; ) {
5517                 struct sched_domain *parent = tmp->parent;
5518                 if (!parent)
5519                         break;
5520 
5521                 if (sd_parent_degenerate(tmp, parent)) {
5522                         tmp->parent = parent->parent;
5523                         if (parent->parent)
5524                                 parent->parent->child = tmp;
5525                         /*
5526                          * Transfer SD_PREFER_SIBLING down in case of a
5527                          * degenerate parent; the spans match for this
5528                          * so the property transfers.
5529                          */
5530                         if (parent->flags & SD_PREFER_SIBLING)
5531                                 tmp->flags |= SD_PREFER_SIBLING;
5532                         destroy_sched_domain(parent, cpu);
5533                 } else
5534                         tmp = tmp->parent;
5535         }
5536 
5537         if (sd && sd_degenerate(sd)) {
5538                 tmp = sd;
5539                 sd = sd->parent;
5540                 destroy_sched_domain(tmp, cpu);
5541                 if (sd)
5542                         sd->child = NULL;
5543         }
5544 
5545         sched_domain_debug(sd, cpu);
5546 
5547         rq_attach_root(rq, rd);
5548         tmp = rq->sd;
5549         rcu_assign_pointer(rq->sd, sd);
5550         destroy_sched_domains(tmp, cpu);
5551 
5552         update_top_cache_domain(cpu);
5553 }
5554 
5555 /* cpus with isolated domains */
5556 static cpumask_var_t cpu_isolated_map;
5557 
5558 /* Setup the mask of cpus configured for isolated domains */
5559 static int __init isolated_cpu_setup(char *str)
5560 {
5561         alloc_bootmem_cpumask_var(&cpu_isolated_map);
5562         cpulist_parse(str, cpu_isolated_map);
5563         return 1;
5564 }
5565 
5566 __setup("isolcpus=", isolated_cpu_setup);
5567 
5568 static const struct cpumask *cpu_cpu_mask(int cpu)
5569 {
5570         return cpumask_of_node(cpu_to_node(cpu));
5571 }
5572 
5573 struct sd_data {
5574         struct sched_domain **__percpu sd;
5575         struct sched_group **__percpu sg;
5576         struct sched_group_power **__percpu sgp;
5577 };
5578 
5579 struct s_data {
5580         struct sched_domain ** __percpu sd;
5581         struct root_domain      *rd;
5582 };
5583 
5584 enum s_alloc {
5585         sa_rootdomain,
5586         sa_sd,
5587         sa_sd_storage,
5588         sa_none,
5589 };
5590 
5591 struct sched_domain_topology_level;
5592 
5593 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5594 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5595 
5596 #define SDTL_OVERLAP    0x01
5597 
5598 struct sched_domain_topology_level {
5599         sched_domain_init_f init;
5600         sched_domain_mask_f mask;
5601         int                 flags;
5602         int                 numa_level;
5603         struct sd_data      data;
5604 };
5605 
5606 /*
5607  * Build an iteration mask that can exclude certain CPUs from the upwards
5608  * domain traversal.
5609  *
5610  * Asymmetric node setups can result in situations where the domain tree is of
5611  * unequal depth, make sure to skip domains that already cover the entire
5612  * range.
5613  *
5614  * In that case build_sched_domains() will have terminated the iteration early
5615  * and our sibling sd spans will be empty. Domains should always include the
5616  * cpu they're built on, so check that.
5617  *
5618  */
5619 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5620 {
5621         const struct cpumask *span = sched_domain_span(sd);
5622         struct sd_data *sdd = sd->private;
5623         struct sched_domain *sibling;
5624         int i;
5625 
5626         for_each_cpu(i, span) {
5627                 sibling = *per_cpu_ptr(sdd->sd, i);
5628                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5629                         continue;
5630 
5631                 cpumask_set_cpu(i, sched_group_mask(sg));
5632         }
5633 }
5634 
5635 /*
5636  * Return the canonical balance cpu for this group, this is the first cpu
5637  * of this group that's also in the iteration mask.
5638  */
5639 int group_balance_cpu(struct sched_group *sg)
5640 {
5641         return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5642 }
5643 
5644 static int
5645 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5646 {
5647         struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5648         const struct cpumask *span = sched_domain_span(sd);
5649         struct cpumask *covered = sched_domains_tmpmask;
5650         struct sd_data *sdd = sd->private;
5651         struct sched_domain *child;
5652         int i;
5653 
5654         cpumask_clear(covered);
5655 
5656         for_each_cpu(i, span) {
5657                 struct cpumask *sg_span;
5658 
5659                 if (cpumask_test_cpu(i, covered))
5660                         continue;
5661 
5662                 child = *per_cpu_ptr(sdd->sd, i);
5663 
5664                 /* See the comment near build_group_mask(). */
5665                 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5666                         continue;
5667 
5668                 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5669                                 GFP_KERNEL, cpu_to_node(cpu));
5670 
5671                 if (!sg)
5672                         goto fail;
5673 
5674                 sg_span = sched_group_cpus(sg);
5675                 if (child->child) {
5676                         child = child->child;
5677                         cpumask_copy(sg_span, sched_domain_span(child));
5678                 } else
5679                         cpumask_set_cpu(i, sg_span);
5680 
5681                 cpumask_or(covered, covered, sg_span);
5682 
5683                 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5684                 if (atomic_inc_return(&sg->sgp->ref) == 1)
5685                         build_group_mask(sd, sg);
5686 
5687                 /*
5688                  * Initialize sgp->power such that even if we mess up the
5689                  * domains and no possible iteration will get us here, we won't
5690                  * die on a /0 trap.
5691                  */
5692                 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5693                 sg->sgp->power_orig = sg->sgp->power;
5694 
5695                 /*
5696                  * Make sure the first group of this domain contains the
5697                  * canonical balance cpu. Otherwise the sched_domain iteration
5698                  * breaks. See update_sg_lb_stats().
5699                  */
5700                 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5701                     group_balance_cpu(sg) == cpu)
5702                         groups = sg;
5703 
5704                 if (!first)
5705                         first = sg;
5706                 if (last)
5707                         last->next = sg;
5708                 last = sg;
5709                 last->next = first;
5710         }
5711         sd->groups = groups;
5712 
5713         return 0;
5714 
5715 fail:
5716         free_sched_groups(first, 0);
5717 
5718         return -ENOMEM;
5719 }
5720 
5721 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5722 {
5723         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5724         struct sched_domain *child = sd->child;
5725 
5726         if (child)
5727                 cpu = cpumask_first(sched_domain_span(child));
5728 
5729         if (sg) {
5730                 *sg = *per_cpu_ptr(sdd->sg, cpu);
5731                 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5732                 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5733         }
5734 
5735         return cpu;
5736 }
5737 
5738 /*
5739  * build_sched_groups will build a circular linked list of the groups
5740  * covered by the given span, and will set each group's ->cpumask correctly,
5741  * and ->cpu_power to 0.
5742  *
5743  * Assumes the sched_domain tree is fully constructed
5744  */
5745 static int
5746 build_sched_groups(struct sched_domain *sd, int cpu)
5747 {
5748         struct sched_group *first = NULL, *last = NULL;
5749         struct sd_data *sdd = sd->private;
5750         const struct cpumask *span = sched_domain_span(sd);
5751         struct cpumask *covered;
5752         int i;
5753 
5754         get_group(cpu, sdd, &sd->groups);
5755         atomic_inc(&sd->groups->ref);
5756 
5757         if (cpu != cpumask_first(span))
5758                 return 0;
5759 
5760         lockdep_assert_held(&sched_domains_mutex);
5761         covered = sched_domains_tmpmask;
5762 
5763         cpumask_clear(covered);
5764 
5765         for_each_cpu(i, span) {
5766                 struct sched_group *sg;
5767                 int group, j;
5768 
5769                 if (cpumask_test_cpu(i, covered))
5770                         continue;
5771 
5772                 group = get_group(i, sdd, &sg);
5773                 cpumask_clear(sched_group_cpus(sg));
5774                 sg->sgp->power = 0;
5775                 cpumask_setall(sched_group_mask(sg));
5776 
5777                 for_each_cpu(j, span) {
5778                         if (get_group(j, sdd, NULL) != group)
5779                                 continue;
5780 
5781                         cpumask_set_cpu(j, covered);
5782                         cpumask_set_cpu(j, sched_group_cpus(sg));
5783                 }
5784 
5785                 if (!first)
5786                         first = sg;
5787                 if (last)
5788                         last->next = sg;
5789                 last = sg;
5790         }
5791         last->next = first;
5792 
5793         return 0;
5794 }
5795 
5796 /*
5797  * Initialize sched groups cpu_power.
5798  *
5799  * cpu_power indicates the capacity of sched group, which is used while
5800  * distributing the load between different sched groups in a sched domain.
5801  * Typically cpu_power for all the groups in a sched domain will be same unless
5802  * there are asymmetries in the topology. If there are asymmetries, group
5803  * having more cpu_power will pickup more load compared to the group having
5804  * less cpu_power.
5805  */
5806 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5807 {
5808         struct sched_group *sg = sd->groups;
5809 
5810         WARN_ON(!sg);
5811 
5812         do {
5813                 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5814                 sg = sg->next;
5815         } while (sg != sd->groups);
5816 
5817         if (cpu != group_balance_cpu(sg))
5818                 return;
5819 
5820         update_group_power(sd, cpu);
5821         atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5822 }
5823 
5824 int __weak arch_sd_sibling_asym_packing(void)
5825 {
5826        return 0*SD_ASYM_PACKING;
5827 }
5828 
5829 /*
5830  * Initializers for schedule domains
5831  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5832  */
5833 
5834 #ifdef CONFIG_SCHED_DEBUG
5835 # define SD_INIT_NAME(sd, type)         sd->name = #type
5836 #else
5837 # define SD_INIT_NAME(sd, type)         do { } while (0)
5838 #endif
5839 
5840 #define SD_INIT_FUNC(type)                                              \
5841 static noinline struct sched_domain *                                   \
5842 sd_init_##type(struct sched_domain_topology_level *tl, int cpu)         \
5843 {                                                                       \
5844         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);       \
5845         *sd = SD_##type##_INIT;                                         \
5846         SD_INIT_NAME(sd, type);                                         \
5847         sd->private = &tl->data;                                        \
5848         return sd;                                                      \
5849 }
5850 
5851 SD_INIT_FUNC(CPU)
5852 #ifdef CONFIG_SCHED_SMT
5853  SD_INIT_FUNC(SIBLING)
5854 #endif
5855 #ifdef CONFIG_SCHED_MC
5856  SD_INIT_FUNC(MC)
5857 #endif
5858 #ifdef CONFIG_SCHED_BOOK
5859  SD_INIT_FUNC(BOOK)
5860 #endif
5861 
5862 static int default_relax_domain_level = -1;
5863 int sched_domain_level_max;
5864 
5865 static int __init setup_relax_domain_level(char *str)
5866 {
5867         if (kstrtoint(str, 0, &default_relax_domain_level))
5868                 pr_warn("Unable to set relax_domain_level\n");
5869 
5870         return 1;
5871 }
5872 __setup("relax_domain_level=", setup_relax_domain_level);
5873 
5874 static void set_domain_attribute(struct sched_domain *sd,
5875                                  struct sched_domain_attr *attr)
5876 {
5877         int request;
5878 
5879         if (!attr || attr->relax_domain_level < 0) {
5880                 if (default_relax_domain_level < 0)
5881                         return;
5882                 else
5883                         request = default_relax_domain_level;
5884         } else
5885                 request = attr->relax_domain_level;
5886         if (request < sd->level) {
5887                 /* turn off idle balance on this domain */
5888                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5889         } else {
5890                 /* turn on idle balance on this domain */
5891                 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5892         }
5893 }
5894 
5895 static void __sdt_free(const struct cpumask *cpu_map);
5896 static int __sdt_alloc(const struct cpumask *cpu_map);
5897 
5898 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5899                                  const struct cpumask *cpu_map)
5900 {
5901         switch (what) {
5902         case sa_rootdomain:
5903                 if (!atomic_read(&d->rd->refcount))
5904                         free_rootdomain(&d->rd->rcu); /* fall through */
5905         case sa_sd:
5906                 free_percpu(d->sd); /* fall through */
5907         case sa_sd_storage:
5908                 __sdt_free(cpu_map); /* fall through */
5909         case sa_none:
5910                 break;
5911         }
5912 }
5913 
5914 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5915                                                    const struct cpumask *cpu_map)
5916 {
5917         memset(d, 0, sizeof(*d));
5918 
5919         if (__sdt_alloc(cpu_map))
5920                 return sa_sd_storage;
5921         d->sd = alloc_percpu(struct sched_domain *);
5922         if (!d->sd)
5923                 return sa_sd_storage;
5924         d->rd = alloc_rootdomain();
5925         if (!d->rd)
5926                 return sa_sd;
5927         return sa_rootdomain;
5928 }
5929 
5930 /*
5931  * NULL the sd_data elements we've used to build the sched_domain and
5932  * sched_group structure so that the subsequent __free_domain_allocs()
5933  * will not free the data we're using.
5934  */
5935 static void claim_allocations(int cpu, struct sched_domain *sd)
5936 {
5937         struct sd_data *sdd = sd->private;
5938 
5939         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5940         *per_cpu_ptr(sdd->sd, cpu) = NULL;
5941 
5942         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5943                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5944 
5945         if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5946                 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5947 }
5948 
5949 #ifdef CONFIG_SCHED_SMT
5950 static const struct cpumask *cpu_smt_mask(int cpu)
5951 {
5952         return topology_thread_cpumask(cpu);
5953 }
5954 #endif
5955 
5956 /*
5957  * Topology list, bottom-up.
5958  */
5959 static struct sched_domain_topology_level default_topology[] = {
5960 #ifdef CONFIG_SCHED_SMT
5961         { sd_init_SIBLING, cpu_smt_mask, },
5962 #endif
5963 #ifdef CONFIG_SCHED_MC
5964         { sd_init_MC, cpu_coregroup_mask, },
5965 #endif
5966 #ifdef CONFIG_SCHED_BOOK
5967         { sd_init_BOOK, cpu_book_mask, },
5968 #endif
5969         { sd_init_CPU, cpu_cpu_mask, },
5970         { NULL, },
5971 };
5972 
5973 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5974 
5975 #define for_each_sd_topology(tl)                        \
5976         for (tl = sched_domain_topology; tl->init; tl++)
5977 
5978 #ifdef CONFIG_NUMA
5979 
5980 static int sched_domains_numa_levels;
5981 static int *sched_domains_numa_distance;
5982 static struct cpumask ***sched_domains_numa_masks;
5983 static int sched_domains_curr_level;
5984 
5985 static inline int sd_local_flags(int level)
5986 {
5987         if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5988                 return 0;
5989 
5990         return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5991 }
5992 
5993 static struct sched_domain *
5994 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5995 {
5996         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5997         int level = tl->numa_level;
5998         int sd_weight = cpumask_weight(
5999                         sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6000 
6001         *sd = (struct sched_domain){
6002                 .min_interval           = sd_weight,
6003                 .max_interval           = 2*sd_weight,
6004                 .busy_factor            = 32,
6005                 .imbalance_pct          = 125,
6006                 .cache_nice_tries       = 2,
6007                 .busy_idx               = 3,
6008                 .idle_idx               = 2,
6009                 .newidle_idx            = 0,
6010                 .wake_idx               = 0,
6011                 .forkexec_idx           = 0,
6012 
6013                 .flags                  = 1*SD_LOAD_BALANCE
6014                                         | 1*SD_BALANCE_NEWIDLE
6015                                         | 0*SD_BALANCE_EXEC
6016                                         | 0*SD_BALANCE_FORK
6017                                         | 0*SD_BALANCE_WAKE
6018                                         | 0*SD_WAKE_AFFINE
6019                                         | 0*SD_SHARE_CPUPOWER
6020                                         | 0*SD_SHARE_PKG_RESOURCES
6021                                         | 1*SD_SERIALIZE
6022                                         | 0*SD_PREFER_SIBLING
6023                                         | 1*SD_NUMA
6024                                         | sd_local_flags(level)
6025                                         ,
6026                 .last_balance           = jiffies,
6027                 .balance_interval       = sd_weight,
6028         };
6029         SD_INIT_NAME(sd, NUMA);
6030         sd->private = &tl->data;
6031 
6032         /*
6033          * Ugly hack to pass state to sd_numa_mask()...
6034          */
6035         sched_domains_curr_level = tl->numa_level;
6036 
6037         return sd;
6038 }
6039 
6040 static const struct cpumask *sd_numa_mask(int cpu)
6041 {
6042         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6043 }
6044 
6045 static void sched_numa_warn(const char *str)
6046 {
6047         static int done = false;
6048         int i,j;
6049 
6050         if (done)
6051                 return;
6052 
6053         done = true;
6054 
6055         printk(KERN_WARNING "ERROR: %s\n\n", str);
6056 
6057         for (i = 0; i < nr_node_ids; i++) {
6058                 printk(KERN_WARNING "  ");
6059                 for (j = 0; j < nr_node_ids; j++)
6060                         printk(KERN_CONT "%02d ", node_distance(i,j));
6061                 printk(KERN_CONT "\n");
6062         }
6063         printk(KERN_WARNING "\n");
6064 }
6065 
6066 static bool find_numa_distance(int distance)
6067 {
6068         int i;
6069 
6070         if (distance == node_distance(0, 0))
6071                 return true;
6072 
6073         for (i = 0; i < sched_domains_numa_levels; i++) {
6074                 if (sched_domains_numa_distance[i] == distance)
6075                         return true;
6076         }
6077 
6078         return false;
6079 }
6080 
6081 static void sched_init_numa(void)
6082 {
6083         int next_distance, curr_distance = node_distance(0, 0);
6084         struct sched_domain_topology_level *tl;
6085         int level = 0;
6086         int i, j, k;
6087 
6088         sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6089         if (!sched_domains_numa_distance)
6090                 return;
6091 
6092         /*
6093          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6094          * unique distances in the node_distance() table.
6095          *
6096          * Assumes node_distance(0,j) includes all distances in
6097          * node_distance(i,j) in order to avoid cubic time.
6098          */
6099         next_distance = curr_distance;
6100         for (i = 0; i < nr_node_ids; i++) {
6101                 for (j = 0; j < nr_node_ids; j++) {
6102                         for (k = 0; k < nr_node_ids; k++) {
6103                                 int distance = node_distance(i, k);
6104 
6105                                 if (distance > curr_distance &&
6106                                     (distance < next_distance ||
6107                                      next_distance == curr_distance))
6108                                         next_distance = distance;
6109 
6110                                 /*
6111                                  * While not a strong assumption it would be nice to know
6112                                  * about cases where if node A is connected to B, B is not
6113                                  * equally connected to A.
6114                                  */
6115                                 if (sched_debug() && node_distance(k, i) != distance)
6116                                         sched_numa_warn("Node-distance not symmetric");
6117 
6118                                 if (sched_debug() && i && !find_numa_distance(distance))
6119                                         sched_numa_warn("Node-0 not representative");
6120                         }
6121                         if (next_distance != curr_distance) {
6122                                 sched_domains_numa_distance[level++] = next_distance;
6123                                 sched_domains_numa_levels = level;
6124                                 curr_distance = next_distance;
6125                         } else break;
6126                 }
6127 
6128                 /*
6129                  * In case of sched_debug() we verify the above assumption.
6130                  */
6131                 if (!sched_debug())
6132                         break;
6133         }
6134         /*
6135          * 'level' contains the number of unique distances, excluding the
6136          * identity distance node_distance(i,i).
6137          *
6138          * The sched_domains_numa_distance[] array includes the actual distance
6139          * numbers.
6140          */
6141 
6142         /*
6143          * Here, we should temporarily reset sched_domains_numa_levels to 0.
6144          * If it fails to allocate memory for array sched_domains_numa_masks[][],
6145          * the array will contain less then 'level' members. This could be
6146          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6147          * in other functions.
6148          *
6149          * We reset it to 'level' at the end of this function.
6150          */
6151         sched_domains_numa_levels = 0;
6152 
6153         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6154         if (!sched_domains_numa_masks)
6155                 return;
6156 
6157         /*
6158          * Now for each level, construct a mask per node which contains all
6159          * cpus of nodes that are that many hops away from us.
6160          */
6161         for (i = 0; i < level; i++) {
6162                 sched_domains_numa_masks[i] =
6163                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6164                 if (!sched_domains_numa_masks[i])
6165                         return;
6166 
6167                 for (j = 0; j < nr_node_ids; j++) {
6168                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6169                         if (!mask)
6170                                 return;
6171 
6172                         sched_domains_numa_masks[i][j] = mask;
6173 
6174                         for (k = 0; k < nr_node_ids; k++) {
6175                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
6176                                         continue;
6177 
6178                                 cpumask_or(mask, mask, cpumask_of_node(k));
6179                         }
6180                 }
6181         }
6182 
6183         tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6184                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6185         if (!tl)
6186                 return;
6187 
6188         /*
6189          * Copy the default topology bits..
6190          */
6191         for (i = 0; default_topology[i].init; i++)
6192                 tl[i] = default_topology[i];
6193 
6194         /*
6195          * .. and append 'j' levels of NUMA goodness.
6196          */
6197         for (j = 0; j < level; i++, j++) {
6198                 tl[i] = (struct sched_domain_topology_level){
6199                         .init = sd_numa_init,
6200                         .mask = sd_numa_mask,
6201                         .flags = SDTL_OVERLAP,
6202                         .numa_level = j,
6203                 };
6204         }
6205 
6206         sched_domain_topology = tl;
6207 
6208         sched_domains_numa_levels = level;
6209 }
6210 
6211 static void sched_domains_numa_masks_set(int cpu)
6212 {
6213         int i, j;
6214         int node = cpu_to_node(cpu);
6215 
6216         for (i = 0; i < sched_domains_numa_levels; i++) {
6217                 for (j = 0; j < nr_node_ids; j++) {
6218                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
6219                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6220                 }
6221         }
6222 }
6223 
6224 static void sched_domains_numa_masks_clear(int cpu)
6225 {
6226         int i, j;
6227         for (i = 0; i < sched_domains_numa_levels; i++) {
6228                 for (j = 0; j < nr_node_ids; j++)
6229                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6230         }
6231 }
6232 
6233 /*
6234  * Update sched_domains_numa_masks[level][node] array when new cpus
6235  * are onlined.
6236  */
6237 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6238                                            unsigned long action,
6239                                            void *hcpu)
6240 {
6241         int cpu = (long)hcpu;
6242 
6243         switch (action & ~CPU_TASKS_FROZEN) {
6244         case CPU_ONLINE:
6245                 sched_domains_numa_masks_set(cpu);
6246                 break;
6247 
6248         case CPU_DEAD:
6249                 sched_domains_numa_masks_clear(cpu);
6250                 break;
6251 
6252         default:
6253                 return NOTIFY_DONE;
6254         }
6255 
6256         return NOTIFY_OK;
6257 }
6258 #else
6259 static inline void sched_init_numa(void)
6260 {
6261 }
6262 
6263 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6264                                            unsigned long action,
6265                                            void *hcpu)
6266 {
6267         return 0;
6268 }
6269 #endif /* CONFIG_NUMA */
6270 
6271 static int __sdt_alloc(const struct cpumask *cpu_map)
6272 {
6273         struct sched_domain_topology_level *tl;
6274         int j;
6275 
6276         for_each_sd_topology(tl) {
6277                 struct sd_data *sdd = &tl->data;
6278 
6279                 sdd->sd = alloc_percpu(struct sched_domain *);
6280                 if (!sdd->sd)
6281                         return -ENOMEM;
6282 
6283                 sdd->sg = alloc_percpu(struct sched_group *);
6284                 if (!sdd->sg)
6285                         return -ENOMEM;
6286 
6287                 sdd->sgp = alloc_percpu(struct sched_group_power *);
6288                 if (!sdd->sgp)
6289                         return -ENOMEM;
6290 
6291                 for_each_cpu(j, cpu_map) {
6292                         struct sched_domain *sd;
6293                         struct sched_group *sg;
6294                         struct sched_group_power *sgp;
6295 
6296                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6297                                         GFP_KERNEL, cpu_to_node(j));
6298                         if (!sd)
6299                                 return -ENOMEM;
6300 
6301                         *per_cpu_ptr(sdd->sd, j) = sd;
6302 
6303                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6304                                         GFP_KERNEL, cpu_to_node(j));
6305                         if (!sg)
6306                                 return -ENOMEM;
6307 
6308                         sg->next = sg;
6309 
6310                         *per_cpu_ptr(sdd->sg, j) = sg;
6311 
6312                         sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6313                                         GFP_KERNEL, cpu_to_node(j));
6314                         if (!sgp)
6315                                 return -ENOMEM;
6316 
6317                         *per_cpu_ptr(sdd->sgp, j) = sgp;
6318                 }
6319         }
6320 
6321         return 0;
6322 }
6323 
6324 static void __sdt_free(const struct cpumask *cpu_map)
6325 {
6326         struct sched_domain_topology_level *tl;
6327         int j;
6328 
6329         for_each_sd_topology(tl) {
6330                 struct sd_data *sdd = &tl->data;
6331 
6332                 for_each_cpu(j, cpu_map) {
6333                         struct sched_domain *sd;
6334 
6335                         if (sdd->sd) {
6336                                 sd = *per_cpu_ptr(sdd->sd, j);
6337                                 if (sd && (sd->flags & SD_OVERLAP))
6338                                         free_sched_groups(sd->groups, 0);
6339                                 kfree(*per_cpu_ptr(sdd->sd, j));
6340                         }
6341 
6342                         if (sdd->sg)
6343                                 kfree(*per_cpu_ptr(sdd->sg, j));
6344                         if (sdd->sgp)
6345                                 kfree(*per_cpu_ptr(sdd->sgp, j));
6346                 }
6347                 free_percpu(sdd->sd);
6348                 sdd->sd = NULL;
6349                 free_percpu(sdd->sg);
6350                 sdd->sg = NULL;
6351                 free_percpu(sdd->sgp);
6352                 sdd->sgp = NULL;
6353         }
6354 }
6355 
6356 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6357                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6358                 struct sched_domain *child, int cpu)
6359 {
6360         struct sched_domain *sd = tl->init(tl, cpu);
6361         if (!sd)
6362                 return child;
6363 
6364         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6365         if (child) {
6366                 sd->level = child->level + 1;
6367                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6368                 child->parent = sd;
6369                 sd->child = child;
6370         }
6371         set_domain_attribute(sd, attr);
6372 
6373         return sd;
6374 }
6375 
6376 /*
6377  * Build sched domains for a given set of cpus and attach the sched domains
6378  * to the individual cpus
6379  */
6380 static int build_sched_domains(const struct cpumask *cpu_map,
6381                                struct sched_domain_attr *attr)
6382 {
6383         enum s_alloc alloc_state;
6384         struct sched_domain *sd;
6385         struct s_data d;
6386         int i, ret = -ENOMEM;
6387 
6388         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6389         if (alloc_state != sa_rootdomain)
6390                 goto error;
6391 
6392         /* Set up domains for cpus specified by the cpu_map. */
6393         for_each_cpu(i, cpu_map) {
6394                 struct sched_domain_topology_level *tl;
6395 
6396                 sd = NULL;
6397                 for_each_sd_topology(tl) {
6398                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6399                         if (tl == sched_domain_topology)
6400                                 *per_cpu_ptr(d.sd, i) = sd;
6401                         if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6402                                 sd->flags |= SD_OVERLAP;
6403                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6404                                 break;
6405                 }
6406         }
6407 
6408         /* Build the groups for the domains */
6409         for_each_cpu(i, cpu_map) {
6410                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6411                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
6412                         if (sd->flags & SD_OVERLAP) {
6413                                 if (build_overlap_sched_groups(sd, i))
6414                                         goto error;
6415                         } else {
6416                                 if (build_sched_groups(sd, i))
6417                                         goto error;
6418                         }
6419                 }
6420         }
6421 
6422         /* Calculate CPU power for physical packages and nodes */
6423         for (i = nr_cpumask_bits-1; i >= 0; i--) {
6424                 if (!cpumask_test_cpu(i, cpu_map))
6425                         continue;
6426 
6427                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6428                         claim_allocations(i, sd);
6429                         init_sched_groups_power(i, sd);
6430                 }
6431         }
6432 
6433         /* Attach the domains */
6434         rcu_read_lock();
6435         for_each_cpu(i, cpu_map) {
6436                 sd = *per_cpu_ptr(d.sd, i);
6437                 cpu_attach_domain(sd, d.rd, i);
6438         }
6439         rcu_read_unlock();
6440 
6441         ret = 0;
6442 error:
6443         __free_domain_allocs(&d, alloc_state, cpu_map);
6444         return ret;
6445 }
6446 
6447 static cpumask_var_t *doms_cur; /* current sched domains */
6448 static int ndoms_cur;           /* number of sched domains in 'doms_cur' */
6449 static struct sched_domain_attr *dattr_cur;
6450                                 /* attribues of custom domains in 'doms_cur' */
6451 
6452 /*
6453  * Special case: If a kmalloc of a doms_cur partition (array of
6454  * cpumask) fails, then fallback to a single sched domain,
6455  * as determined by the single cpumask fallback_doms.
6456  */
6457 static cpumask_var_t fallback_doms;
6458 
6459 /*
6460  * arch_update_cpu_topology lets virtualized architectures update the
6461  * cpu core maps. It is supposed to return 1 if the topology changed
6462  * or 0 if it stayed the same.
6463  */
6464 int __attribute__((weak)) arch_update_cpu_topology(void)
6465 {
6466         return 0;
6467 }
6468 
6469 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6470 {
6471         int i;
6472         cpumask_var_t *doms;
6473 
6474         doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6475         if (!doms)
6476                 return NULL;
6477         for (i = 0; i < ndoms; i++) {
6478                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6479                         free_sched_domains(doms, i);
6480                         return NULL;
6481                 }
6482         }
6483         return doms;
6484 }
6485 
6486 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6487 {
6488         unsigned int i;
6489         for (i = 0; i < ndoms; i++)
6490                 free_cpumask_var(doms[i]);
6491         kfree(doms);
6492 }
6493 
6494 /*
6495  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6496  * For now this just excludes isolated cpus, but could be used to
6497  * exclude other special cases in the future.
6498  */
6499 static int init_sched_domains(const struct cpumask *cpu_map)
6500 {
6501         int err;
6502 
6503         arch_update_cpu_topology();
6504         ndoms_cur = 1;
6505         doms_cur = alloc_sched_domains(ndoms_cur);
6506         if (!doms_cur)
6507                 doms_cur = &fallback_doms;
6508         cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6509         err = build_sched_domains(doms_cur[0], NULL);
6510         register_sched_domain_sysctl();
6511 
6512         return err;
6513 }
6514 
6515 /*
6516  * Detach sched domains from a group of cpus specified in cpu_map
6517  * These cpus will now be attached to the NULL domain
6518  */
6519 static void detach_destroy_domains(const struct cpumask *cpu_map)
6520 {
6521         int i;
6522 
6523         rcu_read_lock();
6524         for_each_cpu(i, cpu_map)
6525                 cpu_attach_domain(NULL, &def_root_domain, i);
6526         rcu_read_unlock();
6527 }
6528 
6529 /* handle null as "default" */
6530 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6531                         struct sched_domain_attr *new, int idx_new)
6532 {
6533         struct sched_domain_attr tmp;
6534 
6535         /* fast path */
6536         if (!new && !cur)
6537                 return 1;
6538 
6539         tmp = SD_ATTR_INIT;
6540         return !memcmp(cur ? (cur + idx_cur) : &tmp,
6541                         new ? (new + idx_new) : &tmp,
6542                         sizeof(struct sched_domain_attr));
6543 }
6544 
6545 /*
6546  * Partition sched domains as specified by the 'ndoms_new'
6547  * cpumasks in the array doms_new[] of cpumasks. This compares
6548  * doms_new[] to the current sched domain partitioning, doms_cur[].
6549  * It destroys each deleted domain and builds each new domain.
6550  *
6551  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6552  * The masks don't intersect (don't overlap.) We should setup one
6553  * sched domain for each mask. CPUs not in any of the cpumasks will
6554  * not be load balanced. If the same cpumask appears both in the
6555  * current 'doms_cur' domains and in the new 'doms_new', we can leave
6556  * it as it is.
6557  *
6558  * The passed in 'doms_new' should be allocated using
6559  * alloc_sched_domains.  This routine takes ownership of it and will
6560  * free_sched_domains it when done with it. If the caller failed the
6561  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6562  * and partition_sched_domains() will fallback to the single partition
6563  * 'fallback_doms', it also forces the domains to be rebuilt.
6564  *
6565  * If doms_new == NULL it will be replaced with cpu_online_mask.
6566  * ndoms_new == 0 is a special case for destroying existing domains,
6567  * and it will not create the default domain.
6568  *
6569  * Call with hotplug lock held
6570  */
6571 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6572                              struct sched_domain_attr *dattr_new)
6573 {
6574         int i, j, n;
6575         int new_topology;
6576 
6577         mutex_lock(&sched_domains_mutex);
6578 
6579         /* always unregister in case we don't destroy any domains */
6580         unregister_sched_domain_sysctl();
6581 
6582         /* Let architecture update cpu core mappings. */
6583         new_topology = arch_update_cpu_topology();
6584 
6585         n = doms_new ? ndoms_new : 0;
6586 
6587         /* Destroy deleted domains */
6588         for (i = 0; i < ndoms_cur; i++) {
6589                 for (j = 0; j < n && !new_topology; j++) {
6590                         if (cpumask_equal(doms_cur[i], doms_new[j])
6591                             && dattrs_equal(dattr_cur, i, dattr_new, j))
6592                                 goto match1;
6593                 }
6594                 /* no match - a current sched domain not in new doms_new[] */
6595                 detach_destroy_domains(doms_cur[i]);
6596 match1:
6597                 ;
6598         }
6599 
6600         n = ndoms_cur;
6601         if (doms_new == NULL) {
6602                 n = 0;
6603                 doms_new = &fallback_doms;
6604                 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6605                 WARN_ON_ONCE(dattr_new);
6606         }
6607 
6608         /* Build new domains */
6609         for (i = 0; i < ndoms_new; i++) {
6610                 for (j = 0; j < n && !new_topology; j++) {
6611                         if (cpumask_equal(doms_new[i], doms_cur[j])
6612                             && dattrs_equal(dattr_new, i, dattr_cur, j))
6613                                 goto match2;
6614                 }
6615                 /* no match - add a new doms_new */
6616                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6617 match2:
6618                 ;
6619         }
6620 
6621         /* Remember the new sched domains */
6622         if (doms_cur != &fallback_doms)
6623                 free_sched_domains(doms_cur, ndoms_cur);
6624         kfree(dattr_cur);       /* kfree(NULL) is safe */
6625         doms_cur = doms_new;
6626         dattr_cur = dattr_new;
6627         ndoms_cur = ndoms_new;
6628 
6629         register_sched_domain_sysctl();
6630 
6631         mutex_unlock(&sched_domains_mutex);
6632 }
6633 
6634 static int num_cpus_frozen;     /* used to mark begin/end of suspend/resume */
6635 
6636 /*
6637  * Update cpusets according to cpu_active mask.  If cpusets are
6638  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6639  * around partition_sched_domains().
6640  *
6641  * If we come here as part of a suspend/resume, don't touch cpusets because we
6642  * want to restore it back to its original state upon resume anyway.
6643  */
6644 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6645                              void *hcpu)
6646 {
6647         switch (action) {
6648         case CPU_ONLINE_FROZEN:
6649         case CPU_DOWN_FAILED_FROZEN:
6650 
6651                 /*
6652                  * num_cpus_frozen tracks how many CPUs are involved in suspend
6653                  * resume sequence. As long as this is not the last online
6654                  * operation in the resume sequence, just build a single sched
6655                  * domain, ignoring cpusets.
6656                  */
6657                 num_cpus_frozen--;
6658                 if (likely(num_cpus_frozen)) {
6659                         partition_sched_domains(1, NULL, NULL);
6660                         break;
6661                 }
6662 
6663                 /*
6664                  * This is the last CPU online operation. So fall through and
6665                  * restore the original sched domains by considering the
6666                  * cpuset configurations.
6667                  */
6668 
6669         case CPU_ONLINE:
6670         case CPU_DOWN_FAILED:
6671                 cpuset_update_active_cpus(true);
6672                 break;
6673         default:
6674                 return NOTIFY_DONE;
6675         }
6676         return NOTIFY_OK;
6677 }
6678 
6679 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6680                                void *hcpu)
6681 {
6682         switch (action) {
6683         case CPU_DOWN_PREPARE:
6684                 cpuset_update_active_cpus(false);
6685                 break;
6686         case CPU_DOWN_PREPARE_FROZEN:
6687                 num_cpus_frozen++;
6688                 partition_sched_domains(1, NULL, NULL);
6689                 break;
6690         default:
6691                 return NOTIFY_DONE;
6692         }
6693         return NOTIFY_OK;
6694 }
6695 
6696 void __init sched_init_smp(void)
6697 {
6698         cpumask_var_t non_isolated_cpus;
6699 
6700         alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6701         alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6702 
6703         sched_init_numa();
6704 
6705         /*
6706          * There's no userspace yet to cause hotplug operations; hence all the
6707          * cpu masks are stable and all blatant races in the below code cannot
6708          * happen.
6709          */
6710         mutex_lock(&sched_domains_mutex);
6711         init_sched_domains(cpu_active_mask);
6712         cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6713         if (cpumask_empty(non_isolated_cpus))
6714                 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6715         mutex_unlock(&sched_domains_mutex);
6716 
6717         hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6718         hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6719         hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6720 
6721         init_hrtick();
6722 
6723         /* Move init over to a non-isolated CPU */
6724         if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6725                 BUG();
6726         sched_init_granularity();
6727         free_cpumask_var(non_isolated_cpus);
6728 
6729         init_sched_rt_class();
6730         init_sched_dl_class();
6731 }
6732 #else
6733 void __init sched_init_smp(void)
6734 {
6735         sched_init_granularity();
6736 }
6737 #endif /* CONFIG_SMP */
6738 
6739 const_debug unsigned int sysctl_timer_migration = 1;
6740 
6741 int in_sched_functions(unsigned long addr)
6742 {
6743         return in_lock_functions(addr) ||
6744                 (addr >= (unsigned long)__sched_text_start
6745                 && addr < (unsigned long)__sched_text_end);
6746 }
6747 
6748 #ifdef CONFIG_CGROUP_SCHED
6749 /*
6750  * Default task group.
6751  * Every task in system belongs to this group at bootup.
6752  */
6753 struct task_group root_task_group;
6754 LIST_HEAD(task_groups);
6755 #endif
6756 
6757 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6758 
6759 void __init sched_init(void)
6760 {
6761         int i, j;
6762         unsigned long alloc_size = 0, ptr;
6763 
6764 #ifdef CONFIG_FAIR_GROUP_SCHED
6765         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6766 #endif
6767 #ifdef CONFIG_RT_GROUP_SCHED
6768         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6769 #endif
6770 #ifdef CONFIG_CPUMASK_OFFSTACK
6771         alloc_size += num_possible_cpus() * cpumask_size();
6772 #endif
6773         if (alloc_size) {
6774                 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6775 
6776 #ifdef CONFIG_FAIR_GROUP_SCHED
6777                 root_task_group.se = (struct sched_entity **)ptr;
6778                 ptr += nr_cpu_ids * sizeof(void **);
6779 
6780                 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6781                 ptr += nr_cpu_ids * sizeof(void **);
6782 
6783 #endif /* CONFIG_FAIR_GROUP_SCHED */
6784 #ifdef CONFIG_RT_GROUP_SCHED
6785                 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6786                 ptr += nr_cpu_ids * sizeof(void **);
6787 
6788                 root_task_group.rt_rq = (struct rt_rq **)ptr;
6789                 ptr += nr_cpu_ids * sizeof(void **);
6790 
6791 #endif /* CONFIG_RT_GROUP_SCHED */
6792 #ifdef CONFIG_CPUMASK_OFFSTACK
6793                 for_each_possible_cpu(i) {
6794                         per_cpu(load_balance_mask, i) = (void *)ptr;
6795                         ptr += cpumask_size();
6796                 }
6797 #endif /* CONFIG_CPUMASK_OFFSTACK */
6798         }
6799 
6800         init_rt_bandwidth(&def_rt_bandwidth,
6801                         global_rt_period(), global_rt_runtime());
6802         init_dl_bandwidth(&def_dl_bandwidth,
6803                         global_rt_period(), global_rt_runtime());
6804 
6805 #ifdef CONFIG_SMP
6806         init_defrootdomain();
6807 #endif
6808 
6809 #ifdef CONFIG_RT_GROUP_SCHED
6810         init_rt_bandwidth(&root_task_group.rt_bandwidth,
6811                         global_rt_period(), global_rt_runtime());
6812 #endif /* CONFIG_RT_GROUP_SCHED */
6813 
6814 #ifdef CONFIG_CGROUP_SCHED
6815         list_add(&root_task_group.list, &task_groups);
6816         INIT_LIST_HEAD(&root_task_group.children);
6817         INIT_LIST_HEAD(&root_task_group.siblings);
6818         autogroup_init(&init_task);
6819 
6820 #endif /* CONFIG_CGROUP_SCHED */
6821 
6822         for_each_possible_cpu(i) {
6823                 struct rq *rq;
6824 
6825                 rq = cpu_rq(i);
6826                 raw_spin_lock_init(&rq->lock);
6827                 rq->nr_running = 0;
6828                 rq->calc_load_active = 0;
6829                 rq->calc_load_update = jiffies + LOAD_FREQ;
6830                 init_cfs_rq(&rq->cfs);
6831                 init_rt_rq(&rq->rt, rq);
6832                 init_dl_rq(&rq->dl, rq);
6833 #ifdef CONFIG_FAIR_GROUP_SCHED
6834                 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6835                 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6836                 /*
6837                  * How much cpu bandwidth does root_task_group get?
6838                  *
6839                  * In case of task-groups formed thr' the cgroup filesystem, it
6840                  * gets 100% of the cpu resources in the system. This overall
6841                  * system cpu resource is divided among the tasks of
6842                  * root_task_group and its child task-groups in a fair manner,
6843                  * based on each entity's (task or task-group's) weight
6844                  * (se->load.weight).
6845                  *
6846                  * In other words, if root_task_group has 10 tasks of weight
6847                  * 1024) and two child groups A0 and A1 (of weight 1024 each),
6848                  * then A0's share of the cpu resource is:
6849                  *
6850                  *      A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6851                  *
6852                  * We achieve this by letting root_task_group's tasks sit
6853                  * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6854                  */
6855                 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6856                 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6857 #endif /* CONFIG_FAIR_GROUP_SCHED */
6858 
6859                 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6860 #ifdef CONFIG_RT_GROUP_SCHED
6861                 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6862                 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6863 #endif
6864 
6865                 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6866                         rq->cpu_load[j] = 0;
6867 
6868                 rq->last_load_update_tick = jiffies;
6869 
6870 #ifdef CONFIG_SMP
6871                 rq->sd = NULL;
6872                 rq->rd = NULL;
6873                 rq->cpu_power = SCHED_POWER_SCALE;
6874                 rq->post_schedule = 0;
6875                 rq->active_balance = 0;
6876                 rq->next_balance = jiffies;
6877                 rq->push_cpu = 0;
6878                 rq->cpu = i;
6879                 rq->online = 0;
6880                 rq->idle_stamp = 0;
6881                 rq->avg_idle = 2*sysctl_sched_migration_cost;
6882                 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6883 
6884                 INIT_LIST_HEAD(&rq->cfs_tasks);
6885 
6886                 rq_attach_root(rq, &def_root_domain);
6887 #ifdef CONFIG_NO_HZ_COMMON
6888                 rq->nohz_flags = 0;
6889 #endif
6890 #ifdef CONFIG_NO_HZ_FULL
6891                 rq->last_sched_tick = 0;
6892 #endif
6893 #endif
6894                 init_rq_hrtick(rq);
6895                 atomic_set(&rq->nr_iowait, 0);
6896         }
6897 
6898         set_load_weight(&init_task);
6899 
6900 #ifdef CONFIG_PREEMPT_NOTIFIERS
6901         INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6902 #endif
6903 
6904         /*
6905          * The boot idle thread does lazy MMU switching as well:
6906          */
6907         atomic_inc(&init_mm.mm_count);
6908         enter_lazy_tlb(&init_mm, current);
6909 
6910         /*
6911          * Make us the idle thread. Technically, schedule() should not be
6912          * called from this thread, however somewhere below it might be,
6913          * but because we are the idle thread, we just pick up running again
6914          * when this runqueue becomes "idle".
6915          */
6916         init_idle(current, smp_processor_id());
6917 
6918         calc_load_update = jiffies + LOAD_FREQ;
6919 
6920         /*
6921          * During early bootup we pretend to be a normal task:
6922          */
6923         current->sched_class = &fair_sched_class;
6924 
6925 #ifdef CONFIG_SMP
6926         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6927         /* May be allocated at isolcpus cmdline parse time */
6928         if (cpu_isolated_map == NULL)
6929                 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6930         idle_thread_set_boot_cpu();
6931 #endif
6932         init_sched_fair_class();
6933 
6934         scheduler_running = 1;
6935 }
6936 
6937 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6938 static inline int preempt_count_equals(int preempt_offset)
6939 {
6940         int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6941 
6942         return (nested == preempt_offset);
6943 }
6944 
6945 void __might_sleep(const char *file, int line, int preempt_offset)
6946 {
6947         static unsigned long prev_jiffy;        /* ratelimiting */
6948 
6949         rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6950         if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6951             system_state != SYSTEM_RUNNING || oops_in_progress)
6952                 return;
6953         if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6954                 return;
6955         prev_jiffy = jiffies;
6956 
6957         printk(KERN_ERR
6958                 "BUG: sleeping function called from invalid context at %s:%d\n",
6959                         file, line);
6960         printk(KERN_ERR
6961                 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6962                         in_atomic(), irqs_disabled(),
6963                         current->pid, current->comm);
6964 
6965         debug_show_held_locks(current);
6966         if (irqs_disabled())
6967                 print_irqtrace_events(current);
6968         dump_stack();
6969 }
6970 EXPORT_SYMBOL(__might_sleep);
6971 #endif
6972 
6973 #ifdef CONFIG_MAGIC_SYSRQ
6974 static void normalize_task(struct rq *rq, struct task_struct *p)
6975 {
6976         const struct sched_class *prev_class = p->sched_class;
6977         struct sched_attr attr = {
6978                 .sched_policy = SCHED_NORMAL,
6979         };
6980         int old_prio = p->prio;
6981         int on_rq;
6982 
6983         on_rq = p->on_rq;
6984         if (on_rq)
6985                 dequeue_task(rq, p, 0);
6986         __setscheduler(rq, p, &attr);
6987         if (on_rq) {
6988                 enqueue_task(rq, p, 0);
6989                 resched_task(rq->curr);
6990         }
6991 
6992         check_class_changed(rq, p, prev_class, old_prio);
6993 }
6994 
6995 void normalize_rt_tasks(void)
6996 {
6997         struct task_struct *g, *p;
6998         unsigned long flags;
6999         struct rq *rq;
7000 
7001         read_lock_irqsave(&tasklist_lock, flags);
7002         do_each_thread(g, p) {
7003                 /*
7004                  * Only normalize user tasks:
7005                  */
7006                 if (!p->mm)
7007                         continue;
7008 
7009                 p->se.exec_start                = 0;
7010 #ifdef CONFIG_SCHEDSTATS
7011                 p->se.statistics.wait_start     = 0;
7012                 p->se.statistics.sleep_start    = 0;
7013                 p->se.statistics.block_start    = 0;
7014 #endif
7015 
7016                 if (!dl_task(p) && !rt_task(p)) {
7017                         /*
7018                          * Renice negative nice level userspace
7019                          * tasks back to 0:
7020                          */
7021                         if (TASK_NICE(p) < 0 && p->mm)
7022                                 set_user_nice(p, 0);
7023                         continue;
7024                 }
7025 
7026                 raw_spin_lock(&p->pi_lock);
7027                 rq = __task_rq_lock(p);
7028 
7029                 normalize_task(rq, p);
7030 
7031                 __task_rq_unlock(rq);
7032                 raw_spin_unlock(&p->pi_lock);
7033         } while_each_thread(g, p);
7034 
7035         read_unlock_irqrestore(&tasklist_lock, flags);
7036 }
7037 
7038 #endif /* CONFIG_MAGIC_SYSRQ */
7039 
7040 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7041 /*
7042  * These functions are only useful for the IA64 MCA handling, or kdb.
7043  *
7044  * They can only be called when the whole system has been
7045  * stopped - every CPU needs to be quiescent, and no scheduling
7046  * activity can take place. Using them for anything else would
7047  * be a serious bug, and as a result, they aren't even visible
7048  * under any other configuration.
7049  */
7050 
7051 /**
7052  * curr_task - return the current task for a given cpu.
7053  * @cpu: the processor in question.
7054  *
7055  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7056  *
7057  * Return: The current task for @cpu.
7058  */
7059 struct task_struct *curr_task(int cpu)
7060 {
7061         return cpu_curr(cpu);
7062 }
7063 
7064 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7065 
7066 #ifdef CONFIG_IA64
7067 /**
7068  * set_curr_task - set the current task for a given cpu.
7069  * @cpu: the processor in question.
7070  * @p: the task pointer to set.
7071  *
7072  * Description: This function must only be used when non-maskable interrupts
7073  * are serviced on a separate stack. It allows the architecture to switch the
7074  * notion of the current task on a cpu in a non-blocking manner. This function
7075  * must be called with all CPU's synchronized, and interrupts disabled, the
7076  * and caller must save the original value of the current task (see
7077  * curr_task() above) and restore that value before reenabling interrupts and
7078  * re-starting the system.
7079  *
7080  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7081  */
7082 void set_curr_task(int cpu, struct task_struct *p)
7083 {
7084         cpu_curr(cpu) = p;
7085 }
7086 
7087 #endif
7088 
7089 #ifdef CONFIG_CGROUP_SCHED
7090 /* task_group_lock serializes the addition/removal of task groups */
7091 static DEFINE_SPINLOCK(task_group_lock);
7092 
7093 static void free_sched_group(struct task_group *tg)
7094 {
7095         free_fair_sched_group(tg);
7096         free_rt_sched_group(tg);
7097         autogroup_free(tg);
7098         kfree(tg);
7099 }
7100 
7101 /* allocate runqueue etc for a new task group */
7102 struct task_group *sched_create_group(struct task_group *parent)
7103 {
7104         struct task_group *tg;
7105 
7106         tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7107         if (!tg)
7108                 return ERR_PTR(-ENOMEM);
7109 
7110         if (!alloc_fair_sched_group(tg, parent))
7111                 goto err;
7112 
7113         if (!alloc_rt_sched_group(tg, parent))
7114                 goto err;
7115 
7116         return tg;
7117 
7118 err:
7119         free_sched_group(tg);
7120         return ERR_PTR(-ENOMEM);
7121 }
7122 
7123 void sched_online_group(struct task_group *tg, struct task_group *parent)
7124 {
7125         unsigned long flags;
7126 
7127         spin_lock_irqsave(&task_group_lock, flags);
7128         list_add_rcu(&tg->list, &task_groups);
7129 
7130         WARN_ON(!parent); /* root should already exist */
7131 
7132         tg->parent = parent;
7133         INIT_LIST_HEAD(&tg->children);
7134         list_add_rcu(&tg->siblings, &parent->children);
7135         spin_unlock_irqrestore(&task_group_lock, flags);
7136 }
7137 
7138 /* rcu callback to free various structures associated with a task group */
7139 static void free_sched_group_rcu(struct rcu_head *rhp)
7140 {
7141         /* now it should be safe to free those cfs_rqs */
7142         free_sched_group(container_of(rhp, struct task_group, rcu));
7143 }
7144 
7145 /* Destroy runqueue etc associated with a task group */
7146 void sched_destroy_group(struct task_group *tg)
7147 {
7148         /* wait for possible concurrent references to cfs_rqs complete */
7149         call_rcu(&tg->rcu, free_sched_group_rcu);
7150 }
7151 
7152 void sched_offline_group(struct task_group *tg)
7153 {
7154         unsigned long flags;
7155         int i;
7156 
7157         /* end participation in shares distribution */
7158         for_each_possible_cpu(i)
7159                 unregister_fair_sched_group(tg, i);
7160 
7161         spin_lock_irqsave(&task_group_lock, flags);
7162         list_del_rcu(&tg->list);
7163         list_del_rcu(&tg->siblings);
7164         spin_unlock_irqrestore(&task_group_lock, flags);
7165 }
7166 
7167 /* change task's runqueue when it moves between groups.
7168  *      The caller of this function should have put the task in its new group
7169  *      by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7170  *      reflect its new group.
7171  */
7172 void sched_move_task(struct task_struct *tsk)
7173 {
7174         struct task_group *tg;
7175         int on_rq, running;
7176         unsigned long flags;
7177         struct rq *rq;
7178 
7179         rq = task_rq_lock(tsk, &flags);
7180 
7181         running = task_current(rq, tsk);
7182         on_rq = tsk->on_rq;
7183 
7184         if (on_rq)
7185                 dequeue_task(rq, tsk, 0);
7186         if (unlikely(running))
7187                 tsk->sched_class->put_prev_task(rq, tsk);
7188 
7189         tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
7190                                 lockdep_is_held(&tsk->sighand->siglock)),
7191                           struct task_group, css);
7192         tg = autogroup_task_group(tsk, tg);
7193         tsk->sched_task_group = tg;
7194 
7195 #ifdef CONFIG_FAIR_GROUP_SCHED
7196         if (tsk->sched_class->task_move_group)
7197                 tsk->sched_class->task_move_group(tsk, on_rq);
7198         else
7199 #endif
7200                 set_task_rq(tsk, task_cpu(tsk));
7201 
7202         if (unlikely(running))
7203                 tsk->sched_class->set_curr_task(rq);
7204         if (on_rq)
7205                 enqueue_task(rq, tsk, 0);
7206 
7207         task_rq_unlock(rq, tsk, &flags);
7208 }
7209 #endif /* CONFIG_CGROUP_SCHED */
7210 
7211 #ifdef CONFIG_RT_GROUP_SCHED
7212 /*
7213  * Ensure that the real time constraints are schedulable.
7214  */
7215 static DEFINE_MUTEX(rt_constraints_mutex);
7216 
7217 /* Must be called with tasklist_lock held */
7218 static inline int tg_has_rt_tasks(struct task_group *tg)
7219 {
7220         struct task_struct *g, *p;
7221 
7222         do_each_thread(g, p) {
7223                 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7224                         return 1;
7225         } while_each_thread(g, p);
7226 
7227         return 0;
7228 }
7229 
7230 struct rt_schedulable_data {
7231         struct task_group *tg;
7232         u64 rt_period;
7233         u64 rt_runtime;
7234 };
7235 
7236 static int tg_rt_schedulable(struct task_group *tg, void *data)
7237 {
7238         struct rt_schedulable_data *d = data;
7239         struct task_group *child;
7240         unsigned long total, sum = 0;
7241         u64 period, runtime;
7242 
7243         period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7244         runtime = tg->rt_bandwidth.rt_runtime;
7245 
7246         if (tg == d->tg) {
7247                 period = d->rt_period;
7248                 runtime = d->rt_runtime;
7249         }
7250 
7251         /*
7252          * Cannot have more runtime than the period.
7253          */
7254         if (runtime > period && runtime != RUNTIME_INF)
7255                 return -EINVAL;
7256 
7257         /*
7258          * Ensure we don't starve existing RT tasks.
7259          */
7260         if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7261                 return -EBUSY;
7262 
7263         total = to_ratio(period, runtime);
7264 
7265         /*
7266          * Nobody can have more than the global setting allows.
7267          */
7268         if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7269                 return -EINVAL;
7270 
7271         /*
7272          * The sum of our children's runtime should not exceed our own.
7273          */
7274         list_for_each_entry_rcu(child, &tg->children, siblings) {
7275                 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7276                 runtime = child->rt_bandwidth.rt_runtime;
7277 
7278                 if (child == d->tg) {
7279                         period = d->rt_period;
7280                         runtime = d->rt_runtime;
7281                 }
7282 
7283                 sum += to_ratio(period, runtime);
7284         }
7285 
7286         if (sum > total)
7287                 return -EINVAL;
7288 
7289         return 0;
7290 }
7291 
7292 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7293 {
7294         int ret;
7295 
7296         struct rt_schedulable_data data = {
7297                 .tg = tg,
7298                 .rt_period = period,
7299                 .rt_runtime = runtime,
7300         };
7301 
7302         rcu_read_lock();
7303         ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7304         rcu_read_unlock();
7305 
7306         return ret;
7307 }
7308 
7309 static int tg_set_rt_bandwidth(struct task_group *tg,
7310                 u64 rt_period, u64 rt_runtime)
7311 {
7312         int i, err = 0;
7313 
7314         mutex_lock(&rt_constraints_mutex);
7315         read_lock(&tasklist_lock);
7316         err = __rt_schedulable(tg, rt_period, rt_runtime);
7317         if (err)
7318                 goto unlock;
7319 
7320         raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7321         tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7322         tg->rt_bandwidth.rt_runtime = rt_runtime;
7323 
7324         for_each_possible_cpu(i) {
7325                 struct rt_rq *rt_rq = tg->rt_rq[i];
7326 
7327                 raw_spin_lock(&rt_rq->rt_runtime_lock);
7328                 rt_rq->rt_runtime = rt_runtime;
7329                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7330         }
7331         raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7332 unlock:
7333         read_unlock(&tasklist_lock);
7334         mutex_unlock(&rt_constraints_mutex);
7335 
7336         return err;
7337 }
7338 
7339 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7340 {
7341         u64 rt_runtime, rt_period;
7342 
7343         rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7344         rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7345         if (rt_runtime_us < 0)
7346                 rt_runtime = RUNTIME_INF;
7347 
7348         return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7349 }
7350 
7351 static long sched_group_rt_runtime(struct task_group *tg)
7352 {
7353         u64 rt_runtime_us;
7354 
7355         if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7356                 return -1;
7357 
7358         rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7359         do_div(rt_runtime_us, NSEC_PER_USEC);
7360         return rt_runtime_us;
7361 }
7362 
7363 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7364 {
7365         u64 rt_runtime, rt_period;
7366 
7367         rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7368         rt_runtime = tg->rt_bandwidth.rt_runtime;
7369 
7370         if (rt_period == 0)
7371                 return -EINVAL;
7372 
7373         return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7374 }
7375 
7376 static long sched_group_rt_period(struct task_group *tg)
7377 {
7378         u64 rt_period_us;
7379 
7380         rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7381         do_div(rt_period_us, NSEC_PER_USEC);
7382         return rt_period_us;
7383 }
7384 #endif /* CONFIG_RT_GROUP_SCHED */
7385 
7386 #ifdef CONFIG_RT_GROUP_SCHED
7387 static int sched_rt_global_constraints(void)
7388 {
7389         int ret = 0;
7390 
7391         mutex_lock(&rt_constraints_mutex);
7392         read_lock(&tasklist_lock);
7393         ret = __rt_schedulable(NULL, 0, 0);
7394         read_unlock(&tasklist_lock);
7395         mutex_unlock(&rt_constraints_mutex);
7396 
7397         return ret;
7398 }
7399 
7400 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7401 {
7402         /* Don't accept realtime tasks when there is no way for them to run */
7403         if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7404                 return 0;
7405 
7406         return 1;
7407 }
7408 
7409 #else /* !CONFIG_RT_GROUP_SCHED */
7410 static int sched_rt_global_constraints(void)
7411 {
7412         unsigned long flags;
7413         int i, ret = 0;
7414 
7415         raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7416