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

Linux/kernel/sched/core.c

  1 /*
  2  *  kernel/sched/core.c
  3  *
  4  *  Kernel scheduler and related syscalls
  5  *
  6  *  Copyright (C) 1991-2002  Linus Torvalds
  7  *
  8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
  9  *              make semaphores SMP safe
 10  *  1998-11-19  Implemented schedule_timeout() and related stuff
 11  *              by Andrea Arcangeli
 12  *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
 13  *              hybrid priority-list and round-robin design with
 14  *              an array-switch method of distributing timeslices
 15  *              and per-CPU runqueues.  Cleanups and useful suggestions
 16  *              by Davide Libenzi, preemptible kernel bits by Robert Love.
 17  *  2003-09-03  Interactivity tuning by Con Kolivas.
 18  *  2004-04-02  Scheduler domains code by Nick Piggin
 19  *  2007-04-15  Work begun on replacing all interactivity tuning with a
 20  *              fair scheduling design by Con Kolivas.
 21  *  2007-05-05  Load balancing (smp-nice) and other improvements
 22  *              by Peter Williams
 23  *  2007-05-06  Interactivity improvements to CFS by Mike Galbraith
 24  *  2007-07-01  Group scheduling enhancements by Srivatsa Vaddagiri
 25  *  2007-11-29  RT balancing improvements by Steven Rostedt, Gregory Haskins,
 26  *              Thomas Gleixner, Mike Kravetz
 27  */
 28 
 29 #include <linux/mm.h>
 30 #include <linux/module.h>
 31 #include <linux/nmi.h>
 32 #include <linux/init.h>
 33 #include <linux/uaccess.h>
 34 #include <linux/highmem.h>
 35 #include <asm/mmu_context.h>
 36 #include <linux/interrupt.h>
 37 #include <linux/capability.h>
 38 #include <linux/completion.h>
 39 #include <linux/kernel_stat.h>
 40 #include <linux/debug_locks.h>
 41 #include <linux/perf_event.h>
 42 #include <linux/security.h>
 43 #include <linux/notifier.h>
 44 #include <linux/profile.h>
 45 #include <linux/freezer.h>
 46 #include <linux/vmalloc.h>
 47 #include <linux/blkdev.h>
 48 #include <linux/delay.h>
 49 #include <linux/pid_namespace.h>
 50 #include <linux/smp.h>
 51 #include <linux/threads.h>
 52 #include <linux/timer.h>
 53 #include <linux/rcupdate.h>
 54 #include <linux/cpu.h>
 55 #include <linux/cpuset.h>
 56 #include <linux/percpu.h>
 57 #include <linux/proc_fs.h>
 58 #include <linux/seq_file.h>
 59 #include <linux/sysctl.h>
 60 #include <linux/syscalls.h>
 61 #include <linux/times.h>
 62 #include <linux/tsacct_kern.h>
 63 #include <linux/kprobes.h>
 64 #include <linux/delayacct.h>
 65 #include <linux/unistd.h>
 66 #include <linux/pagemap.h>
 67 #include <linux/hrtimer.h>
 68 #include <linux/tick.h>
 69 #include <linux/debugfs.h>
 70 #include <linux/ctype.h>
 71 #include <linux/ftrace.h>
 72 #include <linux/slab.h>
 73 #include <linux/init_task.h>
 74 #include <linux/binfmts.h>
 75 #include <linux/context_tracking.h>
 76 #include <linux/compiler.h>
 77 
 78 #include <asm/switch_to.h>
 79 #include <asm/tlb.h>
 80 #include <asm/irq_regs.h>
 81 #include <asm/mutex.h>
 82 #ifdef CONFIG_PARAVIRT
 83 #include <asm/paravirt.h>
 84 #endif
 85 
 86 #include "sched.h"
 87 #include "../workqueue_internal.h"
 88 #include "../smpboot.h"
 89 
 90 #define CREATE_TRACE_POINTS
 91 #include <trace/events/sched.h>
 92 
 93 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
 94 {
 95         unsigned long delta;
 96         ktime_t soft, hard, now;
 97 
 98         for (;;) {
 99                 if (hrtimer_active(period_timer))
100                         break;
101 
102                 now = hrtimer_cb_get_time(period_timer);
103                 hrtimer_forward(period_timer, now, period);
104 
105                 soft = hrtimer_get_softexpires(period_timer);
106                 hard = hrtimer_get_expires(period_timer);
107                 delta = ktime_to_ns(ktime_sub(hard, soft));
108                 __hrtimer_start_range_ns(period_timer, soft, delta,
109                                          HRTIMER_MODE_ABS_PINNED, 0);
110         }
111 }
112 
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 
118 void update_rq_clock(struct rq *rq)
119 {
120         s64 delta;
121 
122         if (rq->skip_clock_update > 0)
123                 return;
124 
125         delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126         rq->clock += delta;
127         update_rq_clock_task(rq, delta);
128 }
129 
130 /*
131  * Debugging: various feature bits
132  */
133 
134 #define SCHED_FEAT(name, enabled)       \
135         (1UL << __SCHED_FEAT_##name) * enabled |
136 
137 const_debug unsigned int sysctl_sched_features =
138 #include "features.h"
139         0;
140 
141 #undef SCHED_FEAT
142 
143 #ifdef CONFIG_SCHED_DEBUG
144 #define SCHED_FEAT(name, enabled)       \
145         #name ,
146 
147 static const char * const sched_feat_names[] = {
148 #include "features.h"
149 };
150 
151 #undef SCHED_FEAT
152 
153 static int sched_feat_show(struct seq_file *m, void *v)
154 {
155         int i;
156 
157         for (i = 0; i < __SCHED_FEAT_NR; i++) {
158                 if (!(sysctl_sched_features & (1UL << i)))
159                         seq_puts(m, "NO_");
160                 seq_printf(m, "%s ", sched_feat_names[i]);
161         }
162         seq_puts(m, "\n");
163 
164         return 0;
165 }
166 
167 #ifdef HAVE_JUMP_LABEL
168 
169 #define jump_label_key__true  STATIC_KEY_INIT_TRUE
170 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 
172 #define SCHED_FEAT(name, enabled)       \
173         jump_label_key__##enabled ,
174 
175 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
176 #include "features.h"
177 };
178 
179 #undef SCHED_FEAT
180 
181 static void sched_feat_disable(int i)
182 {
183         if (static_key_enabled(&sched_feat_keys[i]))
184                 static_key_slow_dec(&sched_feat_keys[i]);
185 }
186 
187 static void sched_feat_enable(int i)
188 {
189         if (!static_key_enabled(&sched_feat_keys[i]))
190                 static_key_slow_inc(&sched_feat_keys[i]);
191 }
192 #else
193 static void sched_feat_disable(int i) { };
194 static void sched_feat_enable(int i) { };
195 #endif /* HAVE_JUMP_LABEL */
196 
197 static int sched_feat_set(char *cmp)
198 {
199         int i;
200         int neg = 0;
201 
202         if (strncmp(cmp, "NO_", 3) == 0) {
203                 neg = 1;
204                 cmp += 3;
205         }
206 
207         for (i = 0; i < __SCHED_FEAT_NR; i++) {
208                 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209                         if (neg) {
210                                 sysctl_sched_features &= ~(1UL << i);
211                                 sched_feat_disable(i);
212                         } else {
213                                 sysctl_sched_features |= (1UL << i);
214                                 sched_feat_enable(i);
215                         }
216                         break;
217                 }
218         }
219 
220         return i;
221 }
222 
223 static ssize_t
224 sched_feat_write(struct file *filp, const char __user *ubuf,
225                 size_t cnt, loff_t *ppos)
226 {
227         char buf[64];
228         char *cmp;
229         int i;
230 
231         if (cnt > 63)
232                 cnt = 63;
233 
234         if (copy_from_user(&buf, ubuf, cnt))
235                 return -EFAULT;
236 
237         buf[cnt] = 0;
238         cmp = strstrip(buf);
239 
240         i = sched_feat_set(cmp);
241         if (i == __SCHED_FEAT_NR)
242                 return -EINVAL;
243 
244         *ppos += cnt;
245 
246         return cnt;
247 }
248 
249 static int sched_feat_open(struct inode *inode, struct file *filp)
250 {
251         return single_open(filp, sched_feat_show, NULL);
252 }
253 
254 static const struct file_operations sched_feat_fops = {
255         .open           = sched_feat_open,
256         .write          = sched_feat_write,
257         .read           = seq_read,
258         .llseek         = seq_lseek,
259         .release        = single_release,
260 };
261 
262 static __init int sched_init_debug(void)
263 {
264         debugfs_create_file("sched_features", 0644, NULL, NULL,
265                         &sched_feat_fops);
266 
267         return 0;
268 }
269 late_initcall(sched_init_debug);
270 #endif /* CONFIG_SCHED_DEBUG */
271 
272 /*
273  * Number of tasks to iterate in a single balance run.
274  * Limited because this is done with IRQs disabled.
275  */
276 const_debug unsigned int sysctl_sched_nr_migrate = 32;
277 
278 /*
279  * period over which we average the RT time consumption, measured
280  * in ms.
281  *
282  * default: 1s
283  */
284 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
285 
286 /*
287  * period over which we measure -rt task cpu usage in us.
288  * default: 1s
289  */
290 unsigned int sysctl_sched_rt_period = 1000000;
291 
292 __read_mostly int scheduler_running;
293 
294 /*
295  * part of the period that we allow rt tasks to run in us.
296  * default: 0.95s
297  */
298 int sysctl_sched_rt_runtime = 950000;
299 
300 /*
301  * __task_rq_lock - lock the rq @p resides on.
302  */
303 static inline struct rq *__task_rq_lock(struct task_struct *p)
304         __acquires(rq->lock)
305 {
306         struct rq *rq;
307 
308         lockdep_assert_held(&p->pi_lock);
309 
310         for (;;) {
311                 rq = task_rq(p);
312                 raw_spin_lock(&rq->lock);
313                 if (likely(rq == task_rq(p)))
314                         return rq;
315                 raw_spin_unlock(&rq->lock);
316         }
317 }
318 
319 /*
320  * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
321  */
322 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
323         __acquires(p->pi_lock)
324         __acquires(rq->lock)
325 {
326         struct rq *rq;
327 
328         for (;;) {
329                 raw_spin_lock_irqsave(&p->pi_lock, *flags);
330                 rq = task_rq(p);
331                 raw_spin_lock(&rq->lock);
332                 if (likely(rq == task_rq(p)))
333                         return rq;
334                 raw_spin_unlock(&rq->lock);
335                 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
336         }
337 }
338 
339 static void __task_rq_unlock(struct rq *rq)
340         __releases(rq->lock)
341 {
342         raw_spin_unlock(&rq->lock);
343 }
344 
345 static inline void
346 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
347         __releases(rq->lock)
348         __releases(p->pi_lock)
349 {
350         raw_spin_unlock(&rq->lock);
351         raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
352 }
353 
354 /*
355  * this_rq_lock - lock this runqueue and disable interrupts.
356  */
357 static struct rq *this_rq_lock(void)
358         __acquires(rq->lock)
359 {
360         struct rq *rq;
361 
362         local_irq_disable();
363         rq = this_rq();
364         raw_spin_lock(&rq->lock);
365 
366         return rq;
367 }
368 
369 #ifdef CONFIG_SCHED_HRTICK
370 /*
371  * Use HR-timers to deliver accurate preemption points.
372  */
373 
374 static void hrtick_clear(struct rq *rq)
375 {
376         if (hrtimer_active(&rq->hrtick_timer))
377                 hrtimer_cancel(&rq->hrtick_timer);
378 }
379 
380 /*
381  * High-resolution timer tick.
382  * Runs from hardirq context with interrupts disabled.
383  */
384 static enum hrtimer_restart hrtick(struct hrtimer *timer)
385 {
386         struct rq *rq = container_of(timer, struct rq, hrtick_timer);
387 
388         WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
389 
390         raw_spin_lock(&rq->lock);
391         update_rq_clock(rq);
392         rq->curr->sched_class->task_tick(rq, rq->curr, 1);
393         raw_spin_unlock(&rq->lock);
394 
395         return HRTIMER_NORESTART;
396 }
397 
398 #ifdef CONFIG_SMP
399 
400 static int __hrtick_restart(struct rq *rq)
401 {
402         struct hrtimer *timer = &rq->hrtick_timer;
403         ktime_t time = hrtimer_get_softexpires(timer);
404 
405         return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
406 }
407 
408 /*
409  * called from hardirq (IPI) context
410  */
411 static void __hrtick_start(void *arg)
412 {
413         struct rq *rq = arg;
414 
415         raw_spin_lock(&rq->lock);
416         __hrtick_restart(rq);
417         rq->hrtick_csd_pending = 0;
418         raw_spin_unlock(&rq->lock);
419 }
420 
421 /*
422  * Called to set the hrtick timer state.
423  *
424  * called with rq->lock held and irqs disabled
425  */
426 void hrtick_start(struct rq *rq, u64 delay)
427 {
428         struct hrtimer *timer = &rq->hrtick_timer;
429         ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 
431         hrtimer_set_expires(timer, time);
432 
433         if (rq == this_rq()) {
434                 __hrtick_restart(rq);
435         } else if (!rq->hrtick_csd_pending) {
436                 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
437                 rq->hrtick_csd_pending = 1;
438         }
439 }
440 
441 static int
442 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443 {
444         int cpu = (int)(long)hcpu;
445 
446         switch (action) {
447         case CPU_UP_CANCELED:
448         case CPU_UP_CANCELED_FROZEN:
449         case CPU_DOWN_PREPARE:
450         case CPU_DOWN_PREPARE_FROZEN:
451         case CPU_DEAD:
452         case CPU_DEAD_FROZEN:
453                 hrtick_clear(cpu_rq(cpu));
454                 return NOTIFY_OK;
455         }
456 
457         return NOTIFY_DONE;
458 }
459 
460 static __init void init_hrtick(void)
461 {
462         hotcpu_notifier(hotplug_hrtick, 0);
463 }
464 #else
465 /*
466  * Called to set the hrtick timer state.
467  *
468  * called with rq->lock held and irqs disabled
469  */
470 void hrtick_start(struct rq *rq, u64 delay)
471 {
472         __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
473                         HRTIMER_MODE_REL_PINNED, 0);
474 }
475 
476 static inline void init_hrtick(void)
477 {
478 }
479 #endif /* CONFIG_SMP */
480 
481 static void init_rq_hrtick(struct rq *rq)
482 {
483 #ifdef CONFIG_SMP
484         rq->hrtick_csd_pending = 0;
485 
486         rq->hrtick_csd.flags = 0;
487         rq->hrtick_csd.func = __hrtick_start;
488         rq->hrtick_csd.info = rq;
489 #endif
490 
491         hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
492         rq->hrtick_timer.function = hrtick;
493 }
494 #else   /* CONFIG_SCHED_HRTICK */
495 static inline void hrtick_clear(struct rq *rq)
496 {
497 }
498 
499 static inline void init_rq_hrtick(struct rq *rq)
500 {
501 }
502 
503 static inline void init_hrtick(void)
504 {
505 }
506 #endif  /* CONFIG_SCHED_HRTICK */
507 
508 /*
509  * resched_task - mark a task 'to be rescheduled now'.
510  *
511  * On UP this means the setting of the need_resched flag, on SMP it
512  * might also involve a cross-CPU call to trigger the scheduler on
513  * the target CPU.
514  */
515 void resched_task(struct task_struct *p)
516 {
517         int cpu;
518 
519         lockdep_assert_held(&task_rq(p)->lock);
520 
521         if (test_tsk_need_resched(p))
522                 return;
523 
524         set_tsk_need_resched(p);
525 
526         cpu = task_cpu(p);
527         if (cpu == smp_processor_id()) {
528                 set_preempt_need_resched();
529                 return;
530         }
531 
532         /* NEED_RESCHED must be visible before we test polling */
533         smp_mb();
534         if (!tsk_is_polling(p))
535                 smp_send_reschedule(cpu);
536 }
537 
538 void resched_cpu(int cpu)
539 {
540         struct rq *rq = cpu_rq(cpu);
541         unsigned long flags;
542 
543         if (!raw_spin_trylock_irqsave(&rq->lock, flags))
544                 return;
545         resched_task(cpu_curr(cpu));
546         raw_spin_unlock_irqrestore(&rq->lock, flags);
547 }
548 
549 #ifdef CONFIG_SMP
550 #ifdef CONFIG_NO_HZ_COMMON
551 /*
552  * In the semi idle case, use the nearest busy cpu for migrating timers
553  * from an idle cpu.  This is good for power-savings.
554  *
555  * We don't do similar optimization for completely idle system, as
556  * selecting an idle cpu will add more delays to the timers than intended
557  * (as that cpu's timer base may not be uptodate wrt jiffies etc).
558  */
559 int get_nohz_timer_target(int pinned)
560 {
561         int cpu = smp_processor_id();
562         int i;
563         struct sched_domain *sd;
564 
565         if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
566                 return cpu;
567 
568         rcu_read_lock();
569         for_each_domain(cpu, sd) {
570                 for_each_cpu(i, sched_domain_span(sd)) {
571                         if (!idle_cpu(i)) {
572                                 cpu = i;
573                                 goto unlock;
574                         }
575                 }
576         }
577 unlock:
578         rcu_read_unlock();
579         return cpu;
580 }
581 /*
582  * When add_timer_on() enqueues a timer into the timer wheel of an
583  * idle CPU then this timer might expire before the next timer event
584  * which is scheduled to wake up that CPU. In case of a completely
585  * idle system the next event might even be infinite time into the
586  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
587  * leaves the inner idle loop so the newly added timer is taken into
588  * account when the CPU goes back to idle and evaluates the timer
589  * wheel for the next timer event.
590  */
591 static void wake_up_idle_cpu(int cpu)
592 {
593         struct rq *rq = cpu_rq(cpu);
594 
595         if (cpu == smp_processor_id())
596                 return;
597 
598         /*
599          * This is safe, as this function is called with the timer
600          * wheel base lock of (cpu) held. When the CPU is on the way
601          * to idle and has not yet set rq->curr to idle then it will
602          * be serialized on the timer wheel base lock and take the new
603          * timer into account automatically.
604          */
605         if (rq->curr != rq->idle)
606                 return;
607 
608         /*
609          * We can set TIF_RESCHED on the idle task of the other CPU
610          * lockless. The worst case is that the other CPU runs the
611          * idle task through an additional NOOP schedule()
612          */
613         set_tsk_need_resched(rq->idle);
614 
615         /* NEED_RESCHED must be visible before we test polling */
616         smp_mb();
617         if (!tsk_is_polling(rq->idle))
618                 smp_send_reschedule(cpu);
619 }
620 
621 static bool wake_up_full_nohz_cpu(int cpu)
622 {
623         if (tick_nohz_full_cpu(cpu)) {
624                 if (cpu != smp_processor_id() ||
625                     tick_nohz_tick_stopped())
626                         smp_send_reschedule(cpu);
627                 return true;
628         }
629 
630         return false;
631 }
632 
633 void wake_up_nohz_cpu(int cpu)
634 {
635         if (!wake_up_full_nohz_cpu(cpu))
636                 wake_up_idle_cpu(cpu);
637 }
638 
639 static inline bool got_nohz_idle_kick(void)
640 {
641         int cpu = smp_processor_id();
642 
643         if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
644                 return false;
645 
646         if (idle_cpu(cpu) && !need_resched())
647                 return true;
648 
649         /*
650          * We can't run Idle Load Balance on this CPU for this time so we
651          * cancel it and clear NOHZ_BALANCE_KICK
652          */
653         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
654         return false;
655 }
656 
657 #else /* CONFIG_NO_HZ_COMMON */
658 
659 static inline bool got_nohz_idle_kick(void)
660 {
661         return false;
662 }
663 
664 #endif /* CONFIG_NO_HZ_COMMON */
665 
666 #ifdef CONFIG_NO_HZ_FULL
667 bool sched_can_stop_tick(void)
668 {
669        struct rq *rq;
670 
671        rq = this_rq();
672 
673        /* Make sure rq->nr_running update is visible after the IPI */
674        smp_rmb();
675 
676        /* More than one running task need preemption */
677        if (rq->nr_running > 1)
678                return false;
679 
680        return true;
681 }
682 #endif /* CONFIG_NO_HZ_FULL */
683 
684 void sched_avg_update(struct rq *rq)
685 {
686         s64 period = sched_avg_period();
687 
688         while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
689                 /*
690                  * Inline assembly required to prevent the compiler
691                  * optimising this loop into a divmod call.
692                  * See __iter_div_u64_rem() for another example of this.
693                  */
694                 asm("" : "+rm" (rq->age_stamp));
695                 rq->age_stamp += period;
696                 rq->rt_avg /= 2;
697         }
698 }
699 
700 #endif /* CONFIG_SMP */
701 
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703                         (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
704 /*
705  * Iterate task_group tree rooted at *from, calling @down when first entering a
706  * node and @up when leaving it for the final time.
707  *
708  * Caller must hold rcu_lock or sufficient equivalent.
709  */
710 int walk_tg_tree_from(struct task_group *from,
711                              tg_visitor down, tg_visitor up, void *data)
712 {
713         struct task_group *parent, *child;
714         int ret;
715 
716         parent = from;
717 
718 down:
719         ret = (*down)(parent, data);
720         if (ret)
721                 goto out;
722         list_for_each_entry_rcu(child, &parent->children, siblings) {
723                 parent = child;
724                 goto down;
725 
726 up:
727                 continue;
728         }
729         ret = (*up)(parent, data);
730         if (ret || parent == from)
731                 goto out;
732 
733         child = parent;
734         parent = parent->parent;
735         if (parent)
736                 goto up;
737 out:
738         return ret;
739 }
740 
741 int tg_nop(struct task_group *tg, void *data)
742 {
743         return 0;
744 }
745 #endif
746 
747 static void set_load_weight(struct task_struct *p)
748 {
749         int prio = p->static_prio - MAX_RT_PRIO;
750         struct load_weight *load = &p->se.load;
751 
752         /*
753          * SCHED_IDLE tasks get minimal weight:
754          */
755         if (p->policy == SCHED_IDLE) {
756                 load->weight = scale_load(WEIGHT_IDLEPRIO);
757                 load->inv_weight = WMULT_IDLEPRIO;
758                 return;
759         }
760 
761         load->weight = scale_load(prio_to_weight[prio]);
762         load->inv_weight = prio_to_wmult[prio];
763 }
764 
765 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
766 {
767         update_rq_clock(rq);
768         sched_info_queued(rq, p);
769         p->sched_class->enqueue_task(rq, p, flags);
770 }
771 
772 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
773 {
774         update_rq_clock(rq);
775         sched_info_dequeued(rq, p);
776         p->sched_class->dequeue_task(rq, p, flags);
777 }
778 
779 void activate_task(struct rq *rq, struct task_struct *p, int flags)
780 {
781         if (task_contributes_to_load(p))
782                 rq->nr_uninterruptible--;
783 
784         enqueue_task(rq, p, flags);
785 }
786 
787 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
788 {
789         if (task_contributes_to_load(p))
790                 rq->nr_uninterruptible++;
791 
792         dequeue_task(rq, p, flags);
793 }
794 
795 static void update_rq_clock_task(struct rq *rq, s64 delta)
796 {
797 /*
798  * In theory, the compile should just see 0 here, and optimize out the call
799  * to sched_rt_avg_update. But I don't trust it...
800  */
801 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802         s64 steal = 0, irq_delta = 0;
803 #endif
804 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
805         irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
806 
807         /*
808          * Since irq_time is only updated on {soft,}irq_exit, we might run into
809          * this case when a previous update_rq_clock() happened inside a
810          * {soft,}irq region.
811          *
812          * When this happens, we stop ->clock_task and only update the
813          * prev_irq_time stamp to account for the part that fit, so that a next
814          * update will consume the rest. This ensures ->clock_task is
815          * monotonic.
816          *
817          * It does however cause some slight miss-attribution of {soft,}irq
818          * time, a more accurate solution would be to update the irq_time using
819          * the current rq->clock timestamp, except that would require using
820          * atomic ops.
821          */
822         if (irq_delta > delta)
823                 irq_delta = delta;
824 
825         rq->prev_irq_time += irq_delta;
826         delta -= irq_delta;
827 #endif
828 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
829         if (static_key_false((&paravirt_steal_rq_enabled))) {
830                 steal = paravirt_steal_clock(cpu_of(rq));
831                 steal -= rq->prev_steal_time_rq;
832 
833                 if (unlikely(steal > delta))
834                         steal = delta;
835 
836                 rq->prev_steal_time_rq += steal;
837                 delta -= steal;
838         }
839 #endif
840 
841         rq->clock_task += delta;
842 
843 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
844         if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
845                 sched_rt_avg_update(rq, irq_delta + steal);
846 #endif
847 }
848 
849 void sched_set_stop_task(int cpu, struct task_struct *stop)
850 {
851         struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
852         struct task_struct *old_stop = cpu_rq(cpu)->stop;
853 
854         if (stop) {
855                 /*
856                  * Make it appear like a SCHED_FIFO task, its something
857                  * userspace knows about and won't get confused about.
858                  *
859                  * Also, it will make PI more or less work without too
860                  * much confusion -- but then, stop work should not
861                  * rely on PI working anyway.
862                  */
863                 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
864 
865                 stop->sched_class = &stop_sched_class;
866         }
867 
868         cpu_rq(cpu)->stop = stop;
869 
870         if (old_stop) {
871                 /*
872                  * Reset it back to a normal scheduling class so that
873                  * it can die in pieces.
874                  */
875                 old_stop->sched_class = &rt_sched_class;
876         }
877 }
878 
879 /*
880  * __normal_prio - return the priority that is based on the static prio
881  */
882 static inline int __normal_prio(struct task_struct *p)
883 {
884         return p->static_prio;
885 }
886 
887 /*
888  * Calculate the expected normal priority: i.e. priority
889  * without taking RT-inheritance into account. Might be
890  * boosted by interactivity modifiers. Changes upon fork,
891  * setprio syscalls, and whenever the interactivity
892  * estimator recalculates.
893  */
894 static inline int normal_prio(struct task_struct *p)
895 {
896         int prio;
897 
898         if (task_has_dl_policy(p))
899                 prio = MAX_DL_PRIO-1;
900         else if (task_has_rt_policy(p))
901                 prio = MAX_RT_PRIO-1 - p->rt_priority;
902         else
903                 prio = __normal_prio(p);
904         return prio;
905 }
906 
907 /*
908  * Calculate the current priority, i.e. the priority
909  * taken into account by the scheduler. This value might
910  * be boosted by RT tasks, or might be boosted by
911  * interactivity modifiers. Will be RT if the task got
912  * RT-boosted. If not then it returns p->normal_prio.
913  */
914 static int effective_prio(struct task_struct *p)
915 {
916         p->normal_prio = normal_prio(p);
917         /*
918          * If we are RT tasks or we were boosted to RT priority,
919          * keep the priority unchanged. Otherwise, update priority
920          * to the normal priority:
921          */
922         if (!rt_prio(p->prio))
923                 return p->normal_prio;
924         return p->prio;
925 }
926 
927 /**
928  * task_curr - is this task currently executing on a CPU?
929  * @p: the task in question.
930  *
931  * Return: 1 if the task is currently executing. 0 otherwise.
932  */
933 inline int task_curr(const struct task_struct *p)
934 {
935         return cpu_curr(task_cpu(p)) == p;
936 }
937 
938 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
939                                        const struct sched_class *prev_class,
940                                        int oldprio)
941 {
942         if (prev_class != p->sched_class) {
943                 if (prev_class->switched_from)
944                         prev_class->switched_from(rq, p);
945                 p->sched_class->switched_to(rq, p);
946         } else if (oldprio != p->prio || dl_task(p))
947                 p->sched_class->prio_changed(rq, p, oldprio);
948 }
949 
950 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
951 {
952         const struct sched_class *class;
953 
954         if (p->sched_class == rq->curr->sched_class) {
955                 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
956         } else {
957                 for_each_class(class) {
958                         if (class == rq->curr->sched_class)
959                                 break;
960                         if (class == p->sched_class) {
961                                 resched_task(rq->curr);
962                                 break;
963                         }
964                 }
965         }
966 
967         /*
968          * A queue event has occurred, and we're going to schedule.  In
969          * this case, we can save a useless back to back clock update.
970          */
971         if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
972                 rq->skip_clock_update = 1;
973 }
974 
975 #ifdef CONFIG_SMP
976 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
977 {
978 #ifdef CONFIG_SCHED_DEBUG
979         /*
980          * We should never call set_task_cpu() on a blocked task,
981          * ttwu() will sort out the placement.
982          */
983         WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
984                         !(task_preempt_count(p) & PREEMPT_ACTIVE));
985 
986 #ifdef CONFIG_LOCKDEP
987         /*
988          * The caller should hold either p->pi_lock or rq->lock, when changing
989          * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
990          *
991          * sched_move_task() holds both and thus holding either pins the cgroup,
992          * see task_group().
993          *
994          * Furthermore, all task_rq users should acquire both locks, see
995          * task_rq_lock().
996          */
997         WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
998                                       lockdep_is_held(&task_rq(p)->lock)));
999 #endif
1000 #endif
1001 
1002         trace_sched_migrate_task(p, new_cpu);
1003 
1004         if (task_cpu(p) != new_cpu) {
1005                 if (p->sched_class->migrate_task_rq)
1006                         p->sched_class->migrate_task_rq(p, new_cpu);
1007                 p->se.nr_migrations++;
1008                 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1009         }
1010 
1011         __set_task_cpu(p, new_cpu);
1012 }
1013 
1014 static void __migrate_swap_task(struct task_struct *p, int cpu)
1015 {
1016         if (p->on_rq) {
1017                 struct rq *src_rq, *dst_rq;
1018 
1019                 src_rq = task_rq(p);
1020                 dst_rq = cpu_rq(cpu);
1021 
1022                 deactivate_task(src_rq, p, 0);
1023                 set_task_cpu(p, cpu);
1024                 activate_task(dst_rq, p, 0);
1025                 check_preempt_curr(dst_rq, p, 0);
1026         } else {
1027                 /*
1028                  * Task isn't running anymore; make it appear like we migrated
1029                  * it before it went to sleep. This means on wakeup we make the
1030                  * previous cpu our targer instead of where it really is.
1031                  */
1032                 p->wake_cpu = cpu;
1033         }
1034 }
1035 
1036 struct migration_swap_arg {
1037         struct task_struct *src_task, *dst_task;
1038         int src_cpu, dst_cpu;
1039 };
1040 
1041 static int migrate_swap_stop(void *data)
1042 {
1043         struct migration_swap_arg *arg = data;
1044         struct rq *src_rq, *dst_rq;
1045         int ret = -EAGAIN;
1046 
1047         src_rq = cpu_rq(arg->src_cpu);
1048         dst_rq = cpu_rq(arg->dst_cpu);
1049 
1050         double_raw_lock(&arg->src_task->pi_lock,
1051                         &arg->dst_task->pi_lock);
1052         double_rq_lock(src_rq, dst_rq);
1053         if (task_cpu(arg->dst_task) != arg->dst_cpu)
1054                 goto unlock;
1055 
1056         if (task_cpu(arg->src_task) != arg->src_cpu)
1057                 goto unlock;
1058 
1059         if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1060                 goto unlock;
1061 
1062         if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1063                 goto unlock;
1064 
1065         __migrate_swap_task(arg->src_task, arg->dst_cpu);
1066         __migrate_swap_task(arg->dst_task, arg->src_cpu);
1067 
1068         ret = 0;
1069 
1070 unlock:
1071         double_rq_unlock(src_rq, dst_rq);
1072         raw_spin_unlock(&arg->dst_task->pi_lock);
1073         raw_spin_unlock(&arg->src_task->pi_lock);
1074 
1075         return ret;
1076 }
1077 
1078 /*
1079  * Cross migrate two tasks
1080  */
1081 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1082 {
1083         struct migration_swap_arg arg;
1084         int ret = -EINVAL;
1085 
1086         arg = (struct migration_swap_arg){
1087                 .src_task = cur,
1088                 .src_cpu = task_cpu(cur),
1089                 .dst_task = p,
1090                 .dst_cpu = task_cpu(p),
1091         };
1092 
1093         if (arg.src_cpu == arg.dst_cpu)
1094                 goto out;
1095 
1096         /*
1097          * These three tests are all lockless; this is OK since all of them
1098          * will be re-checked with proper locks held further down the line.
1099          */
1100         if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1101                 goto out;
1102 
1103         if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1104                 goto out;
1105 
1106         if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1107                 goto out;
1108 
1109         trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1110         ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1111 
1112 out:
1113         return ret;
1114 }
1115 
1116 struct migration_arg {
1117         struct task_struct *task;
1118         int dest_cpu;
1119 };
1120 
1121 static int migration_cpu_stop(void *data);
1122 
1123 /*
1124  * wait_task_inactive - wait for a thread to unschedule.
1125  *
1126  * If @match_state is nonzero, it's the @p->state value just checked and
1127  * not expected to change.  If it changes, i.e. @p might have woken up,
1128  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1129  * we return a positive number (its total switch count).  If a second call
1130  * a short while later returns the same number, the caller can be sure that
1131  * @p has remained unscheduled the whole time.
1132  *
1133  * The caller must ensure that the task *will* unschedule sometime soon,
1134  * else this function might spin for a *long* time. This function can't
1135  * be called with interrupts off, or it may introduce deadlock with
1136  * smp_call_function() if an IPI is sent by the same process we are
1137  * waiting to become inactive.
1138  */
1139 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1140 {
1141         unsigned long flags;
1142         int running, on_rq;
1143         unsigned long ncsw;
1144         struct rq *rq;
1145 
1146         for (;;) {
1147                 /*
1148                  * We do the initial early heuristics without holding
1149                  * any task-queue locks at all. We'll only try to get
1150                  * the runqueue lock when things look like they will
1151                  * work out!
1152                  */
1153                 rq = task_rq(p);
1154 
1155                 /*
1156                  * If the task is actively running on another CPU
1157                  * still, just relax and busy-wait without holding
1158                  * any locks.
1159                  *
1160                  * NOTE! Since we don't hold any locks, it's not
1161                  * even sure that "rq" stays as the right runqueue!
1162                  * But we don't care, since "task_running()" will
1163                  * return false if the runqueue has changed and p
1164                  * is actually now running somewhere else!
1165                  */
1166                 while (task_running(rq, p)) {
1167                         if (match_state && unlikely(p->state != match_state))
1168                                 return 0;
1169                         cpu_relax();
1170                 }
1171 
1172                 /*
1173                  * Ok, time to look more closely! We need the rq
1174                  * lock now, to be *sure*. If we're wrong, we'll
1175                  * just go back and repeat.
1176                  */
1177                 rq = task_rq_lock(p, &flags);
1178                 trace_sched_wait_task(p);
1179                 running = task_running(rq, p);
1180                 on_rq = p->on_rq;
1181                 ncsw = 0;
1182                 if (!match_state || p->state == match_state)
1183                         ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1184                 task_rq_unlock(rq, p, &flags);
1185 
1186                 /*
1187                  * If it changed from the expected state, bail out now.
1188                  */
1189                 if (unlikely(!ncsw))
1190                         break;
1191 
1192                 /*
1193                  * Was it really running after all now that we
1194                  * checked with the proper locks actually held?
1195                  *
1196                  * Oops. Go back and try again..
1197                  */
1198                 if (unlikely(running)) {
1199                         cpu_relax();
1200                         continue;
1201                 }
1202 
1203                 /*
1204                  * It's not enough that it's not actively running,
1205                  * it must be off the runqueue _entirely_, and not
1206                  * preempted!
1207                  *
1208                  * So if it was still runnable (but just not actively
1209                  * running right now), it's preempted, and we should
1210                  * yield - it could be a while.
1211                  */
1212                 if (unlikely(on_rq)) {
1213                         ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1214 
1215                         set_current_state(TASK_UNINTERRUPTIBLE);
1216                         schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1217                         continue;
1218                 }
1219 
1220                 /*
1221                  * Ahh, all good. It wasn't running, and it wasn't
1222                  * runnable, which means that it will never become
1223                  * running in the future either. We're all done!
1224                  */
1225                 break;
1226         }
1227 
1228         return ncsw;
1229 }
1230 
1231 /***
1232  * kick_process - kick a running thread to enter/exit the kernel
1233  * @p: the to-be-kicked thread
1234  *
1235  * Cause a process which is running on another CPU to enter
1236  * kernel-mode, without any delay. (to get signals handled.)
1237  *
1238  * NOTE: this function doesn't have to take the runqueue lock,
1239  * because all it wants to ensure is that the remote task enters
1240  * the kernel. If the IPI races and the task has been migrated
1241  * to another CPU then no harm is done and the purpose has been
1242  * achieved as well.
1243  */
1244 void kick_process(struct task_struct *p)
1245 {
1246         int cpu;
1247 
1248         preempt_disable();
1249         cpu = task_cpu(p);
1250         if ((cpu != smp_processor_id()) && task_curr(p))
1251                 smp_send_reschedule(cpu);
1252         preempt_enable();
1253 }
1254 EXPORT_SYMBOL_GPL(kick_process);
1255 #endif /* CONFIG_SMP */
1256 
1257 #ifdef CONFIG_SMP
1258 /*
1259  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1260  */
1261 static int select_fallback_rq(int cpu, struct task_struct *p)
1262 {
1263         int nid = cpu_to_node(cpu);
1264         const struct cpumask *nodemask = NULL;
1265         enum { cpuset, possible, fail } state = cpuset;
1266         int dest_cpu;
1267 
1268         /*
1269          * If the node that the cpu is on has been offlined, cpu_to_node()
1270          * will return -1. There is no cpu on the node, and we should
1271          * select the cpu on the other node.
1272          */
1273         if (nid != -1) {
1274                 nodemask = cpumask_of_node(nid);
1275 
1276                 /* Look for allowed, online CPU in same node. */
1277                 for_each_cpu(dest_cpu, nodemask) {
1278                         if (!cpu_online(dest_cpu))
1279                                 continue;
1280                         if (!cpu_active(dest_cpu))
1281                                 continue;
1282                         if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1283                                 return dest_cpu;
1284                 }
1285         }
1286 
1287         for (;;) {
1288                 /* Any allowed, online CPU? */
1289                 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1290                         if (!cpu_online(dest_cpu))
1291                                 continue;
1292                         if (!cpu_active(dest_cpu))
1293                                 continue;
1294                         goto out;
1295                 }
1296 
1297                 switch (state) {
1298                 case cpuset:
1299                         /* No more Mr. Nice Guy. */
1300                         cpuset_cpus_allowed_fallback(p);
1301                         state = possible;
1302                         break;
1303 
1304                 case possible:
1305                         do_set_cpus_allowed(p, cpu_possible_mask);
1306                         state = fail;
1307                         break;
1308 
1309                 case fail:
1310                         BUG();
1311                         break;
1312                 }
1313         }
1314 
1315 out:
1316         if (state != cpuset) {
1317                 /*
1318                  * Don't tell them about moving exiting tasks or
1319                  * kernel threads (both mm NULL), since they never
1320                  * leave kernel.
1321                  */
1322                 if (p->mm && printk_ratelimit()) {
1323                         printk_sched("process %d (%s) no longer affine to cpu%d\n",
1324                                         task_pid_nr(p), p->comm, cpu);
1325                 }
1326         }
1327 
1328         return dest_cpu;
1329 }
1330 
1331 /*
1332  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1333  */
1334 static inline
1335 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1336 {
1337         cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1338 
1339         /*
1340          * In order not to call set_task_cpu() on a blocking task we need
1341          * to rely on ttwu() to place the task on a valid ->cpus_allowed
1342          * cpu.
1343          *
1344          * Since this is common to all placement strategies, this lives here.
1345          *
1346          * [ this allows ->select_task() to simply return task_cpu(p) and
1347          *   not worry about this generic constraint ]
1348          */
1349         if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1350                      !cpu_online(cpu)))
1351                 cpu = select_fallback_rq(task_cpu(p), p);
1352 
1353         return cpu;
1354 }
1355 
1356 static void update_avg(u64 *avg, u64 sample)
1357 {
1358         s64 diff = sample - *avg;
1359         *avg += diff >> 3;
1360 }
1361 #endif
1362 
1363 static void
1364 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1365 {
1366 #ifdef CONFIG_SCHEDSTATS
1367         struct rq *rq = this_rq();
1368 
1369 #ifdef CONFIG_SMP
1370         int this_cpu = smp_processor_id();
1371 
1372         if (cpu == this_cpu) {
1373                 schedstat_inc(rq, ttwu_local);
1374                 schedstat_inc(p, se.statistics.nr_wakeups_local);
1375         } else {
1376                 struct sched_domain *sd;
1377 
1378                 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1379                 rcu_read_lock();
1380                 for_each_domain(this_cpu, sd) {
1381                         if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1382                                 schedstat_inc(sd, ttwu_wake_remote);
1383                                 break;
1384                         }
1385                 }
1386                 rcu_read_unlock();
1387         }
1388 
1389         if (wake_flags & WF_MIGRATED)
1390                 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1391 
1392 #endif /* CONFIG_SMP */
1393 
1394         schedstat_inc(rq, ttwu_count);
1395         schedstat_inc(p, se.statistics.nr_wakeups);
1396 
1397         if (wake_flags & WF_SYNC)
1398                 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1399 
1400 #endif /* CONFIG_SCHEDSTATS */
1401 }
1402 
1403 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1404 {
1405         activate_task(rq, p, en_flags);
1406         p->on_rq = 1;
1407 
1408         /* if a worker is waking up, notify workqueue */
1409         if (p->flags & PF_WQ_WORKER)
1410                 wq_worker_waking_up(p, cpu_of(rq));
1411 }
1412 
1413 /*
1414  * Mark the task runnable and perform wakeup-preemption.
1415  */
1416 static void
1417 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1418 {
1419         check_preempt_curr(rq, p, wake_flags);
1420         trace_sched_wakeup(p, true);
1421 
1422         p->state = TASK_RUNNING;
1423 #ifdef CONFIG_SMP
1424         if (p->sched_class->task_woken)
1425                 p->sched_class->task_woken(rq, p);
1426 
1427         if (rq->idle_stamp) {
1428                 u64 delta = rq_clock(rq) - rq->idle_stamp;
1429                 u64 max = 2*rq->max_idle_balance_cost;
1430 
1431                 update_avg(&rq->avg_idle, delta);
1432 
1433                 if (rq->avg_idle > max)
1434                         rq->avg_idle = max;
1435 
1436                 rq->idle_stamp = 0;
1437         }
1438 #endif
1439 }
1440 
1441 static void
1442 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1443 {
1444 #ifdef CONFIG_SMP
1445         if (p->sched_contributes_to_load)
1446                 rq->nr_uninterruptible--;
1447 #endif
1448 
1449         ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1450         ttwu_do_wakeup(rq, p, wake_flags);
1451 }
1452 
1453 /*
1454  * Called in case the task @p isn't fully descheduled from its runqueue,
1455  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1456  * since all we need to do is flip p->state to TASK_RUNNING, since
1457  * the task is still ->on_rq.
1458  */
1459 static int ttwu_remote(struct task_struct *p, int wake_flags)
1460 {
1461         struct rq *rq;
1462         int ret = 0;
1463 
1464         rq = __task_rq_lock(p);
1465         if (p->on_rq) {
1466                 /* check_preempt_curr() may use rq clock */
1467                 update_rq_clock(rq);
1468                 ttwu_do_wakeup(rq, p, wake_flags);
1469                 ret = 1;
1470         }
1471         __task_rq_unlock(rq);
1472 
1473         return ret;
1474 }
1475 
1476 #ifdef CONFIG_SMP
1477 static void sched_ttwu_pending(void)
1478 {
1479         struct rq *rq = this_rq();
1480         struct llist_node *llist = llist_del_all(&rq->wake_list);
1481         struct task_struct *p;
1482 
1483         raw_spin_lock(&rq->lock);
1484 
1485         while (llist) {
1486                 p = llist_entry(llist, struct task_struct, wake_entry);
1487                 llist = llist_next(llist);
1488                 ttwu_do_activate(rq, p, 0);
1489         }
1490 
1491         raw_spin_unlock(&rq->lock);
1492 }
1493 
1494 void scheduler_ipi(void)
1495 {
1496         /*
1497          * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1498          * TIF_NEED_RESCHED remotely (for the first time) will also send
1499          * this IPI.
1500          */
1501         preempt_fold_need_resched();
1502 
1503         if (llist_empty(&this_rq()->wake_list)
1504                         && !tick_nohz_full_cpu(smp_processor_id())
1505                         && !got_nohz_idle_kick())
1506                 return;
1507 
1508         /*
1509          * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1510          * traditionally all their work was done from the interrupt return
1511          * path. Now that we actually do some work, we need to make sure
1512          * we do call them.
1513          *
1514          * Some archs already do call them, luckily irq_enter/exit nest
1515          * properly.
1516          *
1517          * Arguably we should visit all archs and update all handlers,
1518          * however a fair share of IPIs are still resched only so this would
1519          * somewhat pessimize the simple resched case.
1520          */
1521         irq_enter();
1522         tick_nohz_full_check();
1523         sched_ttwu_pending();
1524 
1525         /*
1526          * Check if someone kicked us for doing the nohz idle load balance.
1527          */
1528         if (unlikely(got_nohz_idle_kick())) {
1529                 this_rq()->idle_balance = 1;
1530                 raise_softirq_irqoff(SCHED_SOFTIRQ);
1531         }
1532         irq_exit();
1533 }
1534 
1535 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1536 {
1537         if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1538                 smp_send_reschedule(cpu);
1539 }
1540 
1541 bool cpus_share_cache(int this_cpu, int that_cpu)
1542 {
1543         return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1544 }
1545 #endif /* CONFIG_SMP */
1546 
1547 static void ttwu_queue(struct task_struct *p, int cpu)
1548 {
1549         struct rq *rq = cpu_rq(cpu);
1550 
1551 #if defined(CONFIG_SMP)
1552         if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1553                 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1554                 ttwu_queue_remote(p, cpu);
1555                 return;
1556         }
1557 #endif
1558 
1559         raw_spin_lock(&rq->lock);
1560         ttwu_do_activate(rq, p, 0);
1561         raw_spin_unlock(&rq->lock);
1562 }
1563 
1564 /**
1565  * try_to_wake_up - wake up a thread
1566  * @p: the thread to be awakened
1567  * @state: the mask of task states that can be woken
1568  * @wake_flags: wake modifier flags (WF_*)
1569  *
1570  * Put it on the run-queue if it's not already there. The "current"
1571  * thread is always on the run-queue (except when the actual
1572  * re-schedule is in progress), and as such you're allowed to do
1573  * the simpler "current->state = TASK_RUNNING" to mark yourself
1574  * runnable without the overhead of this.
1575  *
1576  * Return: %true if @p was woken up, %false if it was already running.
1577  * or @state didn't match @p's state.
1578  */
1579 static int
1580 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1581 {
1582         unsigned long flags;
1583         int cpu, success = 0;
1584 
1585         /*
1586          * If we are going to wake up a thread waiting for CONDITION we
1587          * need to ensure that CONDITION=1 done by the caller can not be
1588          * reordered with p->state check below. This pairs with mb() in
1589          * set_current_state() the waiting thread does.
1590          */
1591         smp_mb__before_spinlock();
1592         raw_spin_lock_irqsave(&p->pi_lock, flags);
1593         if (!(p->state & state))
1594                 goto out;
1595 
1596         success = 1; /* we're going to change ->state */
1597         cpu = task_cpu(p);
1598 
1599         if (p->on_rq && ttwu_remote(p, wake_flags))
1600                 goto stat;
1601 
1602 #ifdef CONFIG_SMP
1603         /*
1604          * If the owning (remote) cpu is still in the middle of schedule() with
1605          * this task as prev, wait until its done referencing the task.
1606          */
1607         while (p->on_cpu)
1608                 cpu_relax();
1609         /*
1610          * Pairs with the smp_wmb() in finish_lock_switch().
1611          */
1612         smp_rmb();
1613 
1614         p->sched_contributes_to_load = !!task_contributes_to_load(p);
1615         p->state = TASK_WAKING;
1616 
1617         if (p->sched_class->task_waking)
1618                 p->sched_class->task_waking(p);
1619 
1620         cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1621         if (task_cpu(p) != cpu) {
1622                 wake_flags |= WF_MIGRATED;
1623                 set_task_cpu(p, cpu);
1624         }
1625 #endif /* CONFIG_SMP */
1626 
1627         ttwu_queue(p, cpu);
1628 stat:
1629         ttwu_stat(p, cpu, wake_flags);
1630 out:
1631         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1632 
1633         return success;
1634 }
1635 
1636 /**
1637  * try_to_wake_up_local - try to wake up a local task with rq lock held
1638  * @p: the thread to be awakened
1639  *
1640  * Put @p on the run-queue if it's not already there. The caller must
1641  * ensure that this_rq() is locked, @p is bound to this_rq() and not
1642  * the current task.
1643  */
1644 static void try_to_wake_up_local(struct task_struct *p)
1645 {
1646         struct rq *rq = task_rq(p);
1647 
1648         if (WARN_ON_ONCE(rq != this_rq()) ||
1649             WARN_ON_ONCE(p == current))
1650                 return;
1651 
1652         lockdep_assert_held(&rq->lock);
1653 
1654         if (!raw_spin_trylock(&p->pi_lock)) {
1655                 raw_spin_unlock(&rq->lock);
1656                 raw_spin_lock(&p->pi_lock);
1657                 raw_spin_lock(&rq->lock);
1658         }
1659 
1660         if (!(p->state & TASK_NORMAL))
1661                 goto out;
1662 
1663         if (!p->on_rq)
1664                 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1665 
1666         ttwu_do_wakeup(rq, p, 0);
1667         ttwu_stat(p, smp_processor_id(), 0);
1668 out:
1669         raw_spin_unlock(&p->pi_lock);
1670 }
1671 
1672 /**
1673  * wake_up_process - Wake up a specific process
1674  * @p: The process to be woken up.
1675  *
1676  * Attempt to wake up the nominated process and move it to the set of runnable
1677  * processes.
1678  *
1679  * Return: 1 if the process was woken up, 0 if it was already running.
1680  *
1681  * It may be assumed that this function implies a write memory barrier before
1682  * changing the task state if and only if any tasks are woken up.
1683  */
1684 int wake_up_process(struct task_struct *p)
1685 {
1686         WARN_ON(task_is_stopped_or_traced(p));
1687         return try_to_wake_up(p, TASK_NORMAL, 0);
1688 }
1689 EXPORT_SYMBOL(wake_up_process);
1690 
1691 int wake_up_state(struct task_struct *p, unsigned int state)
1692 {
1693         return try_to_wake_up(p, state, 0);
1694 }
1695 
1696 /*
1697  * Perform scheduler related setup for a newly forked process p.
1698  * p is forked by current.
1699  *
1700  * __sched_fork() is basic setup used by init_idle() too:
1701  */
1702 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1703 {
1704         p->on_rq                        = 0;
1705 
1706         p->se.on_rq                     = 0;
1707         p->se.exec_start                = 0;
1708         p->se.sum_exec_runtime          = 0;
1709         p->se.prev_sum_exec_runtime     = 0;
1710         p->se.nr_migrations             = 0;
1711         p->se.vruntime                  = 0;
1712         INIT_LIST_HEAD(&p->se.group_node);
1713 
1714 #ifdef CONFIG_SCHEDSTATS
1715         memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1716 #endif
1717 
1718         RB_CLEAR_NODE(&p->dl.rb_node);
1719         hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1720         p->dl.dl_runtime = p->dl.runtime = 0;
1721         p->dl.dl_deadline = p->dl.deadline = 0;
1722         p->dl.dl_period = 0;
1723         p->dl.flags = 0;
1724 
1725         INIT_LIST_HEAD(&p->rt.run_list);
1726 
1727 #ifdef CONFIG_PREEMPT_NOTIFIERS
1728         INIT_HLIST_HEAD(&p->preempt_notifiers);
1729 #endif
1730 
1731 #ifdef CONFIG_NUMA_BALANCING
1732         if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1733                 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1734                 p->mm->numa_scan_seq = 0;
1735         }
1736 
1737         if (clone_flags & CLONE_VM)
1738                 p->numa_preferred_nid = current->numa_preferred_nid;
1739         else
1740                 p->numa_preferred_nid = -1;
1741 
1742         p->node_stamp = 0ULL;
1743         p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1744         p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1745         p->numa_work.next = &p->numa_work;
1746         p->numa_faults_memory = NULL;
1747         p->numa_faults_buffer_memory = NULL;
1748         p->last_task_numa_placement = 0;
1749         p->last_sum_exec_runtime = 0;
1750 
1751         INIT_LIST_HEAD(&p->numa_entry);
1752         p->numa_group = NULL;
1753 #endif /* CONFIG_NUMA_BALANCING */
1754 }
1755 
1756 #ifdef CONFIG_NUMA_BALANCING
1757 #ifdef CONFIG_SCHED_DEBUG
1758 void set_numabalancing_state(bool enabled)
1759 {
1760         if (enabled)
1761                 sched_feat_set("NUMA");
1762         else
1763                 sched_feat_set("NO_NUMA");
1764 }
1765 #else
1766 __read_mostly bool numabalancing_enabled;
1767 
1768 void set_numabalancing_state(bool enabled)
1769 {
1770         numabalancing_enabled = enabled;
1771 }
1772 #endif /* CONFIG_SCHED_DEBUG */
1773 
1774 #ifdef CONFIG_PROC_SYSCTL
1775 int sysctl_numa_balancing(struct ctl_table *table, int write,
1776                          void __user *buffer, size_t *lenp, loff_t *ppos)
1777 {
1778         struct ctl_table t;
1779         int err;
1780         int state = numabalancing_enabled;
1781 
1782         if (write && !capable(CAP_SYS_ADMIN))
1783                 return -EPERM;
1784 
1785         t = *table;
1786         t.data = &state;
1787         err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1788         if (err < 0)
1789                 return err;
1790         if (write)
1791                 set_numabalancing_state(state);
1792         return err;
1793 }
1794 #endif
1795 #endif
1796 
1797 /*
1798  * fork()/clone()-time setup:
1799  */
1800 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1801 {
1802         unsigned long flags;
1803         int cpu = get_cpu();
1804 
1805         __sched_fork(clone_flags, p);
1806         /*
1807          * We mark the process as running here. This guarantees that
1808          * nobody will actually run it, and a signal or other external
1809          * event cannot wake it up and insert it on the runqueue either.
1810          */
1811         p->state = TASK_RUNNING;
1812 
1813         /*
1814          * Make sure we do not leak PI boosting priority to the child.
1815          */
1816         p->prio = current->normal_prio;
1817 
1818         /*
1819          * Revert to default priority/policy on fork if requested.
1820          */
1821         if (unlikely(p->sched_reset_on_fork)) {
1822                 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1823                         p->policy = SCHED_NORMAL;
1824                         p->static_prio = NICE_TO_PRIO(0);
1825                         p->rt_priority = 0;
1826                 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1827                         p->static_prio = NICE_TO_PRIO(0);
1828 
1829                 p->prio = p->normal_prio = __normal_prio(p);
1830                 set_load_weight(p);
1831 
1832                 /*
1833                  * We don't need the reset flag anymore after the fork. It has
1834                  * fulfilled its duty:
1835                  */
1836                 p->sched_reset_on_fork = 0;
1837         }
1838 
1839         if (dl_prio(p->prio)) {
1840                 put_cpu();
1841                 return -EAGAIN;
1842         } else if (rt_prio(p->prio)) {
1843                 p->sched_class = &rt_sched_class;
1844         } else {
1845                 p->sched_class = &fair_sched_class;
1846         }
1847 
1848         if (p->sched_class->task_fork)
1849                 p->sched_class->task_fork(p);
1850 
1851         /*
1852          * The child is not yet in the pid-hash so no cgroup attach races,
1853          * and the cgroup is pinned to this child due to cgroup_fork()
1854          * is ran before sched_fork().
1855          *
1856          * Silence PROVE_RCU.
1857          */
1858         raw_spin_lock_irqsave(&p->pi_lock, flags);
1859         set_task_cpu(p, cpu);
1860         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1861 
1862 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1863         if (likely(sched_info_on()))
1864                 memset(&p->sched_info, 0, sizeof(p->sched_info));
1865 #endif
1866 #if defined(CONFIG_SMP)
1867         p->on_cpu = 0;
1868 #endif
1869         init_task_preempt_count(p);
1870 #ifdef CONFIG_SMP
1871         plist_node_init(&p->pushable_tasks, MAX_PRIO);
1872         RB_CLEAR_NODE(&p->pushable_dl_tasks);
1873 #endif
1874 
1875         put_cpu();
1876         return 0;
1877 }
1878 
1879 unsigned long to_ratio(u64 period, u64 runtime)
1880 {
1881         if (runtime == RUNTIME_INF)
1882                 return 1ULL << 20;
1883 
1884         /*
1885          * Doing this here saves a lot of checks in all
1886          * the calling paths, and returning zero seems
1887          * safe for them anyway.
1888          */
1889         if (period == 0)
1890                 return 0;
1891 
1892         return div64_u64(runtime << 20, period);
1893 }
1894 
1895 #ifdef CONFIG_SMP
1896 inline struct dl_bw *dl_bw_of(int i)
1897 {
1898         return &cpu_rq(i)->rd->dl_bw;
1899 }
1900 
1901 static inline int dl_bw_cpus(int i)
1902 {
1903         struct root_domain *rd = cpu_rq(i)->rd;
1904         int cpus = 0;
1905 
1906         for_each_cpu_and(i, rd->span, cpu_active_mask)
1907                 cpus++;
1908 
1909         return cpus;
1910 }
1911 #else
1912 inline struct dl_bw *dl_bw_of(int i)
1913 {
1914         return &cpu_rq(i)->dl.dl_bw;
1915 }
1916 
1917 static inline int dl_bw_cpus(int i)
1918 {
1919         return 1;
1920 }
1921 #endif
1922 
1923 static inline
1924 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1925 {
1926         dl_b->total_bw -= tsk_bw;
1927 }
1928 
1929 static inline
1930 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1931 {
1932         dl_b->total_bw += tsk_bw;
1933 }
1934 
1935 static inline
1936 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1937 {
1938         return dl_b->bw != -1 &&
1939                dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1940 }
1941 
1942 /*
1943  * We must be sure that accepting a new task (or allowing changing the
1944  * parameters of an existing one) is consistent with the bandwidth
1945  * constraints. If yes, this function also accordingly updates the currently
1946  * allocated bandwidth to reflect the new situation.
1947  *
1948  * This function is called while holding p's rq->lock.
1949  */
1950 static int dl_overflow(struct task_struct *p, int policy,
1951                        const struct sched_attr *attr)
1952 {
1953 
1954         struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1955         u64 period = attr->sched_period ?: attr->sched_deadline;
1956         u64 runtime = attr->sched_runtime;
1957         u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
1958         int cpus, err = -1;
1959 
1960         if (new_bw == p->dl.dl_bw)
1961                 return 0;
1962 
1963         /*
1964          * Either if a task, enters, leave, or stays -deadline but changes
1965          * its parameters, we may need to update accordingly the total
1966          * allocated bandwidth of the container.
1967          */
1968         raw_spin_lock(&dl_b->lock);
1969         cpus = dl_bw_cpus(task_cpu(p));
1970         if (dl_policy(policy) && !task_has_dl_policy(p) &&
1971             !__dl_overflow(dl_b, cpus, 0, new_bw)) {
1972                 __dl_add(dl_b, new_bw);
1973                 err = 0;
1974         } else if (dl_policy(policy) && task_has_dl_policy(p) &&
1975                    !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
1976                 __dl_clear(dl_b, p->dl.dl_bw);
1977                 __dl_add(dl_b, new_bw);
1978                 err = 0;
1979         } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
1980                 __dl_clear(dl_b, p->dl.dl_bw);
1981                 err = 0;
1982         }
1983         raw_spin_unlock(&dl_b->lock);
1984 
1985         return err;
1986 }
1987 
1988 extern void init_dl_bw(struct dl_bw *dl_b);
1989 
1990 /*
1991  * wake_up_new_task - wake up a newly created task for the first time.
1992  *
1993  * This function will do some initial scheduler statistics housekeeping
1994  * that must be done for every newly created context, then puts the task
1995  * on the runqueue and wakes it.
1996  */
1997 void wake_up_new_task(struct task_struct *p)
1998 {
1999         unsigned long flags;
2000         struct rq *rq;
2001 
2002         raw_spin_lock_irqsave(&p->pi_lock, flags);
2003 #ifdef CONFIG_SMP
2004         /*
2005          * Fork balancing, do it here and not earlier because:
2006          *  - cpus_allowed can change in the fork path
2007          *  - any previously selected cpu might disappear through hotplug
2008          */
2009         set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2010 #endif
2011 
2012         /* Initialize new task's runnable average */
2013         init_task_runnable_average(p);
2014         rq = __task_rq_lock(p);
2015         activate_task(rq, p, 0);
2016         p->on_rq = 1;
2017         trace_sched_wakeup_new(p, true);
2018         check_preempt_curr(rq, p, WF_FORK);
2019 #ifdef CONFIG_SMP
2020         if (p->sched_class->task_woken)
2021                 p->sched_class->task_woken(rq, p);
2022 #endif
2023         task_rq_unlock(rq, p, &flags);
2024 }
2025 
2026 #ifdef CONFIG_PREEMPT_NOTIFIERS
2027 
2028 /**
2029  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2030  * @notifier: notifier struct to register
2031  */
2032 void preempt_notifier_register(struct preempt_notifier *notifier)
2033 {
2034         hlist_add_head(&notifier->link, &current->preempt_notifiers);
2035 }
2036 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2037 
2038 /**
2039  * preempt_notifier_unregister - no longer interested in preemption notifications
2040  * @notifier: notifier struct to unregister
2041  *
2042  * This is safe to call from within a preemption notifier.
2043  */
2044 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2045 {
2046         hlist_del(&notifier->link);
2047 }
2048 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2049 
2050 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2051 {
2052         struct preempt_notifier *notifier;
2053 
2054         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2055                 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2056 }
2057 
2058 static void
2059 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2060                                  struct task_struct *next)
2061 {
2062         struct preempt_notifier *notifier;
2063 
2064         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2065                 notifier->ops->sched_out(notifier, next);
2066 }
2067 
2068 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2069 
2070 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2071 {
2072 }
2073 
2074 static void
2075 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2076                                  struct task_struct *next)
2077 {
2078 }
2079 
2080 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2081 
2082 /**
2083  * prepare_task_switch - prepare to switch tasks
2084  * @rq: the runqueue preparing to switch
2085  * @prev: the current task that is being switched out
2086  * @next: the task we are going to switch to.
2087  *
2088  * This is called with the rq lock held and interrupts off. It must
2089  * be paired with a subsequent finish_task_switch after the context
2090  * switch.
2091  *
2092  * prepare_task_switch sets up locking and calls architecture specific
2093  * hooks.
2094  */
2095 static inline void
2096 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2097                     struct task_struct *next)
2098 {
2099         trace_sched_switch(prev, next);
2100         sched_info_switch(rq, prev, next);
2101         perf_event_task_sched_out(prev, next);
2102         fire_sched_out_preempt_notifiers(prev, next);
2103         prepare_lock_switch(rq, next);
2104         prepare_arch_switch(next);
2105 }
2106 
2107 /**
2108  * finish_task_switch - clean up after a task-switch
2109  * @rq: runqueue associated with task-switch
2110  * @prev: the thread we just switched away from.
2111  *
2112  * finish_task_switch must be called after the context switch, paired
2113  * with a prepare_task_switch call before the context switch.
2114  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2115  * and do any other architecture-specific cleanup actions.
2116  *
2117  * Note that we may have delayed dropping an mm in context_switch(). If
2118  * so, we finish that here outside of the runqueue lock. (Doing it
2119  * with the lock held can cause deadlocks; see schedule() for
2120  * details.)
2121  */
2122 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2123         __releases(rq->lock)
2124 {
2125         struct mm_struct *mm = rq->prev_mm;
2126         long prev_state;
2127 
2128         rq->prev_mm = NULL;
2129 
2130         /*
2131          * A task struct has one reference for the use as "current".
2132          * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2133          * schedule one last time. The schedule call will never return, and
2134          * the scheduled task must drop that reference.
2135          * The test for TASK_DEAD must occur while the runqueue locks are
2136          * still held, otherwise prev could be scheduled on another cpu, die
2137          * there before we look at prev->state, and then the reference would
2138          * be dropped twice.
2139          *              Manfred Spraul <manfred@colorfullife.com>
2140          */
2141         prev_state = prev->state;
2142         vtime_task_switch(prev);
2143         finish_arch_switch(prev);
2144         perf_event_task_sched_in(prev, current);
2145         finish_lock_switch(rq, prev);
2146         finish_arch_post_lock_switch();
2147 
2148         fire_sched_in_preempt_notifiers(current);
2149         if (mm)
2150                 mmdrop(mm);
2151         if (unlikely(prev_state == TASK_DEAD)) {
2152                 if (prev->sched_class->task_dead)
2153                         prev->sched_class->task_dead(prev);
2154 
2155                 /*
2156                  * Remove function-return probe instances associated with this
2157                  * task and put them back on the free list.
2158                  */
2159                 kprobe_flush_task(prev);
2160                 put_task_struct(prev);
2161         }
2162 
2163         tick_nohz_task_switch(current);
2164 }
2165 
2166 #ifdef CONFIG_SMP
2167 
2168 /* rq->lock is NOT held, but preemption is disabled */
2169 static inline void post_schedule(struct rq *rq)
2170 {
2171         if (rq->post_schedule) {
2172                 unsigned long flags;
2173 
2174                 raw_spin_lock_irqsave(&rq->lock, flags);
2175                 if (rq->curr->sched_class->post_schedule)
2176                         rq->curr->sched_class->post_schedule(rq);
2177                 raw_spin_unlock_irqrestore(&rq->lock, flags);
2178 
2179                 rq->post_schedule = 0;
2180         }
2181 }
2182 
2183 #else
2184 
2185 static inline void post_schedule(struct rq *rq)
2186 {
2187 }
2188 
2189 #endif
2190 
2191 /**
2192  * schedule_tail - first thing a freshly forked thread must call.
2193  * @prev: the thread we just switched away from.
2194  */
2195 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2196         __releases(rq->lock)
2197 {
2198         struct rq *rq = this_rq();
2199 
2200         finish_task_switch(rq, prev);
2201 
2202         /*
2203          * FIXME: do we need to worry about rq being invalidated by the
2204          * task_switch?
2205          */
2206         post_schedule(rq);
2207 
2208 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2209         /* In this case, finish_task_switch does not reenable preemption */
2210         preempt_enable();
2211 #endif
2212         if (current->set_child_tid)
2213                 put_user(task_pid_vnr(current), current->set_child_tid);
2214 }
2215 
2216 /*
2217  * context_switch - switch to the new MM and the new
2218  * thread's register state.
2219  */
2220 static inline void
2221 context_switch(struct rq *rq, struct task_struct *prev,
2222                struct task_struct *next)
2223 {
2224         struct mm_struct *mm, *oldmm;
2225 
2226         prepare_task_switch(rq, prev, next);
2227 
2228         mm = next->mm;
2229         oldmm = prev->active_mm;
2230         /*
2231          * For paravirt, this is coupled with an exit in switch_to to
2232          * combine the page table reload and the switch backend into
2233          * one hypercall.
2234          */
2235         arch_start_context_switch(prev);
2236 
2237         if (!mm) {
2238                 next->active_mm = oldmm;
2239                 atomic_inc(&oldmm->mm_count);
2240                 enter_lazy_tlb(oldmm, next);
2241         } else
2242                 switch_mm(oldmm, mm, next);
2243 
2244         if (!prev->mm) {
2245                 prev->active_mm = NULL;
2246                 rq->prev_mm = oldmm;
2247         }
2248         /*
2249          * Since the runqueue lock will be released by the next
2250          * task (which is an invalid locking op but in the case
2251          * of the scheduler it's an obvious special-case), so we
2252          * do an early lockdep release here:
2253          */
2254 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2255         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2256 #endif
2257 
2258         context_tracking_task_switch(prev, next);
2259         /* Here we just switch the register state and the stack. */
2260         switch_to(prev, next, prev);
2261 
2262         barrier();
2263         /*
2264          * this_rq must be evaluated again because prev may have moved
2265          * CPUs since it called schedule(), thus the 'rq' on its stack
2266          * frame will be invalid.
2267          */
2268         finish_task_switch(this_rq(), prev);
2269 }
2270 
2271 /*
2272  * nr_running and nr_context_switches:
2273  *
2274  * externally visible scheduler statistics: current number of runnable
2275  * threads, total number of context switches performed since bootup.
2276  */
2277 unsigned long nr_running(void)
2278 {
2279         unsigned long i, sum = 0;
2280 
2281         for_each_online_cpu(i)
2282                 sum += cpu_rq(i)->nr_running;
2283 
2284         return sum;
2285 }
2286 
2287 unsigned long long nr_context_switches(void)
2288 {
2289         int i;
2290         unsigned long long sum = 0;
2291 
2292         for_each_possible_cpu(i)
2293                 sum += cpu_rq(i)->nr_switches;
2294 
2295         return sum;
2296 }
2297 
2298 unsigned long nr_iowait(void)
2299 {
2300         unsigned long i, sum = 0;
2301 
2302         for_each_possible_cpu(i)
2303                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2304 
2305         return sum;
2306 }
2307 
2308 unsigned long nr_iowait_cpu(int cpu)
2309 {
2310         struct rq *this = cpu_rq(cpu);
2311         return atomic_read(&this->nr_iowait);
2312 }
2313 
2314 #ifdef CONFIG_SMP
2315 
2316 /*
2317  * sched_exec - execve() is a valuable balancing opportunity, because at
2318  * this point the task has the smallest effective memory and cache footprint.
2319  */
2320 void sched_exec(void)
2321 {
2322         struct task_struct *p = current;
2323         unsigned long flags;
2324         int dest_cpu;
2325 
2326         raw_spin_lock_irqsave(&p->pi_lock, flags);
2327         dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2328         if (dest_cpu == smp_processor_id())
2329                 goto unlock;
2330 
2331         if (likely(cpu_active(dest_cpu))) {
2332                 struct migration_arg arg = { p, dest_cpu };
2333 
2334                 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2335                 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2336                 return;
2337         }
2338 unlock:
2339         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2340 }
2341 
2342 #endif
2343 
2344 DEFINE_PER_CPU(struct kernel_stat, kstat);
2345 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2346 
2347 EXPORT_PER_CPU_SYMBOL(kstat);
2348 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2349 
2350 /*
2351  * Return any ns on the sched_clock that have not yet been accounted in
2352  * @p in case that task is currently running.
2353  *
2354  * Called with task_rq_lock() held on @rq.
2355  */
2356 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2357 {
2358         u64 ns = 0;
2359 
2360         if (task_current(rq, p)) {
2361                 update_rq_clock(rq);
2362                 ns = rq_clock_task(rq) - p->se.exec_start;
2363                 if ((s64)ns < 0)
2364                         ns = 0;
2365         }
2366 
2367         return ns;
2368 }
2369 
2370 unsigned long long task_delta_exec(struct task_struct *p)
2371 {
2372         unsigned long flags;
2373         struct rq *rq;
2374         u64 ns = 0;
2375 
2376         rq = task_rq_lock(p, &flags);
2377         ns = do_task_delta_exec(p, rq);
2378         task_rq_unlock(rq, p, &flags);
2379 
2380         return ns;
2381 }
2382 
2383 /*
2384  * Return accounted runtime for the task.
2385  * In case the task is currently running, return the runtime plus current's
2386  * pending runtime that have not been accounted yet.
2387  */
2388 unsigned long long task_sched_runtime(struct task_struct *p)
2389 {
2390         unsigned long flags;
2391         struct rq *rq;
2392         u64 ns = 0;
2393 
2394 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2395         /*
2396          * 64-bit doesn't need locks to atomically read a 64bit value.
2397          * So we have a optimization chance when the task's delta_exec is 0.
2398          * Reading ->on_cpu is racy, but this is ok.
2399          *
2400          * If we race with it leaving cpu, we'll take a lock. So we're correct.
2401          * If we race with it entering cpu, unaccounted time is 0. This is
2402          * indistinguishable from the read occurring a few cycles earlier.
2403          */
2404         if (!p->on_cpu)
2405                 return p->se.sum_exec_runtime;
2406 #endif
2407 
2408         rq = task_rq_lock(p, &flags);
2409         ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2410         task_rq_unlock(rq, p, &flags);
2411 
2412         return ns;
2413 }
2414 
2415 /*
2416  * This function gets called by the timer code, with HZ frequency.
2417  * We call it with interrupts disabled.
2418  */
2419 void scheduler_tick(void)
2420 {
2421         int cpu = smp_processor_id();
2422         struct rq *rq = cpu_rq(cpu);
2423         struct task_struct *curr = rq->curr;
2424 
2425         sched_clock_tick();
2426 
2427         raw_spin_lock(&rq->lock);
2428         update_rq_clock(rq);
2429         curr->sched_class->task_tick(rq, curr, 0);
2430         update_cpu_load_active(rq);
2431         raw_spin_unlock(&rq->lock);
2432 
2433         perf_event_task_tick();
2434 
2435 #ifdef CONFIG_SMP
2436         rq->idle_balance = idle_cpu(cpu);
2437         trigger_load_balance(rq);
2438 #endif
2439         rq_last_tick_reset(rq);
2440 }
2441 
2442 #ifdef CONFIG_NO_HZ_FULL
2443 /**
2444  * scheduler_tick_max_deferment
2445  *
2446  * Keep at least one tick per second when a single
2447  * active task is running because the scheduler doesn't
2448  * yet completely support full dynticks environment.
2449  *
2450  * This makes sure that uptime, CFS vruntime, load
2451  * balancing, etc... continue to move forward, even
2452  * with a very low granularity.
2453  *
2454  * Return: Maximum deferment in nanoseconds.
2455  */
2456 u64 scheduler_tick_max_deferment(void)
2457 {
2458         struct rq *rq = this_rq();
2459         unsigned long next, now = ACCESS_ONCE(jiffies);
2460 
2461         next = rq->last_sched_tick + HZ;
2462 
2463         if (time_before_eq(next, now))
2464                 return 0;
2465 
2466         return jiffies_to_nsecs(next - now);
2467 }
2468 #endif
2469 
2470 notrace unsigned long get_parent_ip(unsigned long addr)
2471 {
2472         if (in_lock_functions(addr)) {
2473                 addr = CALLER_ADDR2;
2474                 if (in_lock_functions(addr))
2475                         addr = CALLER_ADDR3;
2476         }
2477         return addr;
2478 }
2479 
2480 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2481                                 defined(CONFIG_PREEMPT_TRACER))
2482 
2483 void __kprobes preempt_count_add(int val)
2484 {
2485 #ifdef CONFIG_DEBUG_PREEMPT
2486         /*
2487          * Underflow?
2488          */
2489         if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2490                 return;
2491 #endif
2492         __preempt_count_add(val);
2493 #ifdef CONFIG_DEBUG_PREEMPT
2494         /*
2495          * Spinlock count overflowing soon?
2496          */
2497         DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2498                                 PREEMPT_MASK - 10);
2499 #endif
2500         if (preempt_count() == val) {
2501                 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2502 #ifdef CONFIG_DEBUG_PREEMPT
2503                 current->preempt_disable_ip = ip;
2504 #endif
2505                 trace_preempt_off(CALLER_ADDR0, ip);
2506         }
2507 }
2508 EXPORT_SYMBOL(preempt_count_add);
2509 
2510 void __kprobes preempt_count_sub(int val)
2511 {
2512 #ifdef CONFIG_DEBUG_PREEMPT
2513         /*
2514          * Underflow?
2515          */
2516         if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2517                 return;
2518         /*
2519          * Is the spinlock portion underflowing?
2520          */
2521         if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2522                         !(preempt_count() & PREEMPT_MASK)))
2523                 return;
2524 #endif
2525 
2526         if (preempt_count() == val)
2527                 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2528         __preempt_count_sub(val);
2529 }
2530 EXPORT_SYMBOL(preempt_count_sub);
2531 
2532 #endif
2533 
2534 /*
2535  * Print scheduling while atomic bug:
2536  */
2537 static noinline void __schedule_bug(struct task_struct *prev)
2538 {
2539         if (oops_in_progress)
2540                 return;
2541 
2542         printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2543                 prev->comm, prev->pid, preempt_count());
2544 
2545         debug_show_held_locks(prev);
2546         print_modules();
2547         if (irqs_disabled())
2548                 print_irqtrace_events(prev);
2549 #ifdef CONFIG_DEBUG_PREEMPT
2550         if (in_atomic_preempt_off()) {
2551                 pr_err("Preemption disabled at:");
2552                 print_ip_sym(current->preempt_disable_ip);
2553                 pr_cont("\n");
2554         }
2555 #endif
2556         dump_stack();
2557         add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2558 }
2559 
2560 /*
2561  * Various schedule()-time debugging checks and statistics:
2562  */
2563 static inline void schedule_debug(struct task_struct *prev)
2564 {
2565         /*
2566          * Test if we are atomic. Since do_exit() needs to call into
2567          * schedule() atomically, we ignore that path. Otherwise whine
2568          * if we are scheduling when we should not.
2569          */
2570         if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2571                 __schedule_bug(prev);
2572         rcu_sleep_check();
2573 
2574         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2575 
2576         schedstat_inc(this_rq(), sched_count);
2577 }
2578 
2579 /*
2580  * Pick up the highest-prio task:
2581  */
2582 static inline struct task_struct *
2583 pick_next_task(struct rq *rq, struct task_struct *prev)
2584 {
2585         const struct sched_class *class = &fair_sched_class;
2586         struct task_struct *p;
2587 
2588         /*
2589          * Optimization: we know that if all tasks are in
2590          * the fair class we can call that function directly:
2591          */
2592         if (likely(prev->sched_class == class &&
2593                    rq->nr_running == rq->cfs.h_nr_running)) {
2594                 p = fair_sched_class.pick_next_task(rq, prev);
2595                 if (unlikely(p == RETRY_TASK))
2596                         goto again;
2597 
2598                 /* assumes fair_sched_class->next == idle_sched_class */
2599                 if (unlikely(!p))
2600                         p = idle_sched_class.pick_next_task(rq, prev);
2601 
2602                 return p;
2603         }
2604 
2605 again:
2606         for_each_class(class) {
2607                 p = class->pick_next_task(rq, prev);
2608                 if (p) {
2609                         if (unlikely(p == RETRY_TASK))
2610                                 goto again;
2611                         return p;
2612                 }
2613         }
2614 
2615         BUG(); /* the idle class will always have a runnable task */
2616 }
2617 
2618 /*
2619  * __schedule() is the main scheduler function.
2620  *
2621  * The main means of driving the scheduler and thus entering this function are:
2622  *
2623  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2624  *
2625  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2626  *      paths. For example, see arch/x86/entry_64.S.
2627  *
2628  *      To drive preemption between tasks, the scheduler sets the flag in timer
2629  *      interrupt handler scheduler_tick().
2630  *
2631  *   3. Wakeups don't really cause entry into schedule(). They add a
2632  *      task to the run-queue and that's it.
2633  *
2634  *      Now, if the new task added to the run-queue preempts the current
2635  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2636  *      called on the nearest possible occasion:
2637  *
2638  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2639  *
2640  *         - in syscall or exception context, at the next outmost
2641  *           preempt_enable(). (this might be as soon as the wake_up()'s
2642  *           spin_unlock()!)
2643  *
2644  *         - in IRQ context, return from interrupt-handler to
2645  *           preemptible context
2646  *
2647  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2648  *         then at the next:
2649  *
2650  *          - cond_resched() call
2651  *          - explicit schedule() call
2652  *          - return from syscall or exception to user-space
2653  *          - return from interrupt-handler to user-space
2654  */
2655 static void __sched __schedule(void)
2656 {
2657         struct task_struct *prev, *next;
2658         unsigned long *switch_count;
2659         struct rq *rq;
2660         int cpu;
2661 
2662 need_resched:
2663         preempt_disable();
2664         cpu = smp_processor_id();
2665         rq = cpu_rq(cpu);
2666         rcu_note_context_switch(cpu);
2667         prev = rq->curr;
2668 
2669         schedule_debug(prev);
2670 
2671         if (sched_feat(HRTICK))
2672                 hrtick_clear(rq);
2673 
2674         /*
2675          * Make sure that signal_pending_state()->signal_pending() below
2676          * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2677          * done by the caller to avoid the race with signal_wake_up().
2678          */
2679         smp_mb__before_spinlock();
2680         raw_spin_lock_irq(&rq->lock);
2681 
2682         switch_count = &prev->nivcsw;
2683         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2684                 if (unlikely(signal_pending_state(prev->state, prev))) {
2685                         prev->state = TASK_RUNNING;
2686                 } else {
2687                         deactivate_task(rq, prev, DEQUEUE_SLEEP);
2688                         prev->on_rq = 0;
2689 
2690                         /*
2691                          * If a worker went to sleep, notify and ask workqueue
2692                          * whether it wants to wake up a task to maintain
2693                          * concurrency.
2694                          */
2695                         if (prev->flags & PF_WQ_WORKER) {
2696                                 struct task_struct *to_wakeup;
2697 
2698                                 to_wakeup = wq_worker_sleeping(prev, cpu);
2699                                 if (to_wakeup)
2700                                         try_to_wake_up_local(to_wakeup);
2701                         }
2702                 }
2703                 switch_count = &prev->nvcsw;
2704         }
2705 
2706         if (prev->on_rq || rq->skip_clock_update < 0)
2707                 update_rq_clock(rq);
2708 
2709         next = pick_next_task(rq, prev);
2710         clear_tsk_need_resched(prev);
2711         clear_preempt_need_resched();
2712         rq->skip_clock_update = 0;
2713 
2714         if (likely(prev != next)) {
2715                 rq->nr_switches++;
2716                 rq->curr = next;
2717                 ++*switch_count;
2718 
2719                 context_switch(rq, prev, next); /* unlocks the rq */
2720                 /*
2721                  * The context switch have flipped the stack from under us
2722                  * and restored the local variables which were saved when
2723                  * this task called schedule() in the past. prev == current
2724                  * is still correct, but it can be moved to another cpu/rq.
2725                  */
2726                 cpu = smp_processor_id();
2727                 rq = cpu_rq(cpu);
2728         } else
2729                 raw_spin_unlock_irq(&rq->lock);
2730 
2731         post_schedule(rq);
2732 
2733         sched_preempt_enable_no_resched();
2734         if (need_resched())
2735                 goto need_resched;
2736 }
2737 
2738 static inline void sched_submit_work(struct task_struct *tsk)
2739 {
2740         if (!tsk->state || tsk_is_pi_blocked(tsk))
2741                 return;
2742         /*
2743          * If we are going to sleep and we have plugged IO queued,
2744          * make sure to submit it to avoid deadlocks.
2745          */
2746         if (blk_needs_flush_plug(tsk))
2747                 blk_schedule_flush_plug(tsk);
2748 }
2749 
2750 asmlinkage __visible void __sched schedule(void)
2751 {
2752         struct task_struct *tsk = current;
2753 
2754         sched_submit_work(tsk);
2755         __schedule();
2756 }
2757 EXPORT_SYMBOL(schedule);
2758 
2759 #ifdef CONFIG_CONTEXT_TRACKING
2760 asmlinkage __visible void __sched schedule_user(void)
2761 {
2762         /*
2763          * If we come here after a random call to set_need_resched(),
2764          * or we have been woken up remotely but the IPI has not yet arrived,
2765          * we haven't yet exited the RCU idle mode. Do it here manually until
2766          * we find a better solution.
2767          */
2768         user_exit();
2769         schedule();
2770         user_enter();
2771 }
2772 #endif
2773 
2774 /**
2775  * schedule_preempt_disabled - called with preemption disabled
2776  *
2777  * Returns with preemption disabled. Note: preempt_count must be 1
2778  */
2779 void __sched schedule_preempt_disabled(void)
2780 {
2781         sched_preempt_enable_no_resched();
2782         schedule();
2783         preempt_disable();
2784 }
2785 
2786 #ifdef CONFIG_PREEMPT
2787 /*
2788  * this is the entry point to schedule() from in-kernel preemption
2789  * off of preempt_enable. Kernel preemptions off return from interrupt
2790  * occur there and call schedule directly.
2791  */
2792 asmlinkage __visible void __sched notrace preempt_schedule(void)
2793 {
2794         /*
2795          * If there is a non-zero preempt_count or interrupts are disabled,
2796          * we do not want to preempt the current task. Just return..
2797          */
2798         if (likely(!preemptible()))
2799                 return;
2800 
2801         do {
2802                 __preempt_count_add(PREEMPT_ACTIVE);
2803                 __schedule();
2804                 __preempt_count_sub(PREEMPT_ACTIVE);
2805 
2806                 /*
2807                  * Check again in case we missed a preemption opportunity
2808                  * between schedule and now.
2809                  */
2810                 barrier();
2811         } while (need_resched());
2812 }
2813 EXPORT_SYMBOL(preempt_schedule);
2814 #endif /* CONFIG_PREEMPT */
2815 
2816 /*
2817  * this is the entry point to schedule() from kernel preemption
2818  * off of irq context.
2819  * Note, that this is called and return with irqs disabled. This will
2820  * protect us against recursive calling from irq.
2821  */
2822 asmlinkage __visible void __sched preempt_schedule_irq(void)
2823 {
2824         enum ctx_state prev_state;
2825 
2826         /* Catch callers which need to be fixed */
2827         BUG_ON(preempt_count() || !irqs_disabled());
2828 
2829         prev_state = exception_enter();
2830 
2831         do {
2832                 __preempt_count_add(PREEMPT_ACTIVE);
2833                 local_irq_enable();
2834                 __schedule();
2835                 local_irq_disable();
2836                 __preempt_count_sub(PREEMPT_ACTIVE);
2837 
2838                 /*
2839                  * Check again in case we missed a preemption opportunity
2840                  * between schedule and now.
2841                  */
2842                 barrier();
2843         } while (need_resched());
2844 
2845         exception_exit(prev_state);
2846 }
2847 
2848 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2849                           void *key)
2850 {
2851         return try_to_wake_up(curr->private, mode, wake_flags);
2852 }
2853 EXPORT_SYMBOL(default_wake_function);
2854 
2855 #ifdef CONFIG_RT_MUTEXES
2856 
2857 /*
2858  * rt_mutex_setprio - set the current priority of a task
2859  * @p: task
2860  * @prio: prio value (kernel-internal form)
2861  *
2862  * This function changes the 'effective' priority of a task. It does
2863  * not touch ->normal_prio like __setscheduler().
2864  *
2865  * Used by the rt_mutex code to implement priority inheritance
2866  * logic. Call site only calls if the priority of the task changed.
2867  */
2868 void rt_mutex_setprio(struct task_struct *p, int prio)
2869 {
2870         int oldprio, on_rq, running, enqueue_flag = 0;
2871         struct rq *rq;
2872         const struct sched_class *prev_class;
2873 
2874         BUG_ON(prio > MAX_PRIO);
2875 
2876         rq = __task_rq_lock(p);
2877 
2878         /*
2879          * Idle task boosting is a nono in general. There is one
2880          * exception, when PREEMPT_RT and NOHZ is active:
2881          *
2882          * The idle task calls get_next_timer_interrupt() and holds
2883          * the timer wheel base->lock on the CPU and another CPU wants
2884          * to access the timer (probably to cancel it). We can safely
2885          * ignore the boosting request, as the idle CPU runs this code
2886          * with interrupts disabled and will complete the lock
2887          * protected section without being interrupted. So there is no
2888          * real need to boost.
2889          */
2890         if (unlikely(p == rq->idle)) {
2891                 WARN_ON(p != rq->curr);
2892                 WARN_ON(p->pi_blocked_on);
2893                 goto out_unlock;
2894         }
2895 
2896         trace_sched_pi_setprio(p, prio);
2897         p->pi_top_task = rt_mutex_get_top_task(p);
2898         oldprio = p->prio;
2899         prev_class = p->sched_class;
2900         on_rq = p->on_rq;
2901         running = task_current(rq, p);
2902         if (on_rq)
2903                 dequeue_task(rq, p, 0);
2904         if (running)
2905                 p->sched_class->put_prev_task(rq, p);
2906 
2907         /*
2908          * Boosting condition are:
2909          * 1. -rt task is running and holds mutex A
2910          *      --> -dl task blocks on mutex A
2911          *
2912          * 2. -dl task is running and holds mutex A
2913          *      --> -dl task blocks on mutex A and could preempt the
2914          *          running task
2915          */
2916         if (dl_prio(prio)) {
2917                 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2918                         dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2919                         p->dl.dl_boosted = 1;
2920                         p->dl.dl_throttled = 0;
2921                         enqueue_flag = ENQUEUE_REPLENISH;
2922                 } else
2923                         p->dl.dl_boosted = 0;
2924                 p->sched_class = &dl_sched_class;
2925         } else if (rt_prio(prio)) {
2926                 if (dl_prio(oldprio))
2927                         p->dl.dl_boosted = 0;
2928                 if (oldprio < prio)
2929                         enqueue_flag = ENQUEUE_HEAD;
2930                 p->sched_class = &rt_sched_class;
2931         } else {
2932                 if (dl_prio(oldprio))
2933                         p->dl.dl_boosted = 0;
2934                 p->sched_class = &fair_sched_class;
2935         }
2936 
2937         p->prio = prio;
2938 
2939         if (running)
2940                 p->sched_class->set_curr_task(rq);
2941         if (on_rq)
2942                 enqueue_task(rq, p, enqueue_flag);
2943 
2944         check_class_changed(rq, p, prev_class, oldprio);
2945 out_unlock:
2946         __task_rq_unlock(rq);
2947 }
2948 #endif
2949 
2950 void set_user_nice(struct task_struct *p, long nice)
2951 {
2952         int old_prio, delta, on_rq;
2953         unsigned long flags;
2954         struct rq *rq;
2955 
2956         if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
2957                 return;
2958         /*
2959          * We have to be careful, if called from sys_setpriority(),
2960          * the task might be in the middle of scheduling on another CPU.
2961          */
2962         rq = task_rq_lock(p, &flags);
2963         /*
2964          * The RT priorities are set via sched_setscheduler(), but we still
2965          * allow the 'normal' nice value to be set - but as expected
2966          * it wont have any effect on scheduling until the task is
2967          * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
2968          */
2969         if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2970                 p->static_prio = NICE_TO_PRIO(nice);
2971                 goto out_unlock;
2972         }
2973         on_rq = p->on_rq;
2974         if (on_rq)
2975                 dequeue_task(rq, p, 0);
2976 
2977         p->static_prio = NICE_TO_PRIO(nice);
2978         set_load_weight(p);
2979         old_prio = p->prio;
2980         p->prio = effective_prio(p);
2981         delta = p->prio - old_prio;
2982 
2983         if (on_rq) {
2984                 enqueue_task(rq, p, 0);
2985                 /*
2986                  * If the task increased its priority or is running and
2987                  * lowered its priority, then reschedule its CPU:
2988                  */
2989                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
2990                         resched_task(rq->curr);
2991         }
2992 out_unlock:
2993         task_rq_unlock(rq, p, &flags);
2994 }
2995 EXPORT_SYMBOL(set_user_nice);
2996 
2997 /*
2998  * can_nice - check if a task can reduce its nice value
2999  * @p: task
3000  * @nice: nice value
3001  */
3002 int can_nice(const struct task_struct *p, const int nice)
3003 {
3004         /* convert nice value [19,-20] to rlimit style value [1,40] */
3005         int nice_rlim = 20 - nice;
3006 
3007         return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3008                 capable(CAP_SYS_NICE));
3009 }
3010 
3011 #ifdef __ARCH_WANT_SYS_NICE
3012 
3013 /*
3014  * sys_nice - change the priority of the current process.
3015  * @increment: priority increment
3016  *
3017  * sys_setpriority is a more generic, but much slower function that
3018  * does similar things.
3019  */
3020 SYSCALL_DEFINE1(nice, int, increment)
3021 {
3022         long nice, retval;
3023 
3024         /*
3025          * Setpriority might change our priority at the same moment.
3026          * We don't have to worry. Conceptually one call occurs first
3027          * and we have a single winner.
3028          */
3029         if (increment < -40)
3030                 increment = -40;
3031         if (increment > 40)
3032                 increment = 40;
3033 
3034         nice = task_nice(current) + increment;
3035         if (nice < MIN_NICE)
3036                 nice = MIN_NICE;
3037         if (nice > MAX_NICE)
3038                 nice = MAX_NICE;
3039 
3040         if (increment < 0 && !can_nice(current, nice))
3041                 return -EPERM;
3042 
3043         retval = security_task_setnice(current, nice);
3044         if (retval)
3045                 return retval;
3046 
3047         set_user_nice(current, nice);
3048         return 0;
3049 }
3050 
3051 #endif
3052 
3053 /**
3054  * task_prio - return the priority value of a given task.
3055  * @p: the task in question.
3056  *
3057  * Return: The priority value as seen by users in /proc.
3058  * RT tasks are offset by -200. Normal tasks are centered
3059  * around 0, value goes from -16 to +15.
3060  */
3061 int task_prio(const struct task_struct *p)
3062 {
3063         return p->prio - MAX_RT_PRIO;
3064 }
3065 
3066 /**
3067  * idle_cpu - is a given cpu idle currently?
3068  * @cpu: the processor in question.
3069  *
3070  * Return: 1 if the CPU is currently idle. 0 otherwise.
3071  */
3072 int idle_cpu(int cpu)
3073 {
3074         struct rq *rq = cpu_rq(cpu);
3075 
3076         if (rq->curr != rq->idle)
3077                 return 0;
3078 
3079         if (rq->nr_running)
3080                 return 0;
3081 
3082 #ifdef CONFIG_SMP
3083         if (!llist_empty(&rq->wake_list))
3084                 return 0;
3085 #endif
3086 
3087         return 1;
3088 }
3089 
3090 /**
3091  * idle_task - return the idle task for a given cpu.
3092  * @cpu: the processor in question.
3093  *
3094  * Return: The idle task for the cpu @cpu.
3095  */
3096 struct task_struct *idle_task(int cpu)
3097 {
3098         return cpu_rq(cpu)->idle;
3099 }
3100 
3101 /**
3102  * find_process_by_pid - find a process with a matching PID value.
3103  * @pid: the pid in question.
3104  *
3105  * The task of @pid, if found. %NULL otherwise.
3106  */
3107 static struct task_struct *find_process_by_pid(pid_t pid)
3108 {
3109         return pid ? find_task_by_vpid(pid) : current;
3110 }
3111 
3112 /*
3113  * This function initializes the sched_dl_entity of a newly becoming
3114  * SCHED_DEADLINE task.
3115  *
3116  * Only the static values are considered here, the actual runtime and the
3117  * absolute deadline will be properly calculated when the task is enqueued
3118  * for the first time with its new policy.
3119  */
3120 static void
3121 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3122 {
3123         struct sched_dl_entity *dl_se = &p->dl;
3124 
3125         init_dl_task_timer(dl_se);
3126         dl_se->dl_runtime = attr->sched_runtime;
3127         dl_se->dl_deadline = attr->sched_deadline;
3128         dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3129         dl_se->flags = attr->sched_flags;
3130         dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3131         dl_se->dl_throttled = 0;
3132         dl_se->dl_new = 1;
3133         dl_se->dl_yielded = 0;
3134 }
3135 
3136 static void __setscheduler_params(struct task_struct *p,
3137                 const struct sched_attr *attr)
3138 {
3139         int policy = attr->sched_policy;
3140 
3141         if (policy == -1) /* setparam */
3142                 policy = p->policy;
3143 
3144         p->policy = policy;
3145 
3146         if (dl_policy(policy))
3147                 __setparam_dl(p, attr);
3148         else if (fair_policy(policy))
3149                 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3150 
3151         /*
3152          * __sched_setscheduler() ensures attr->sched_priority == 0 when
3153          * !rt_policy. Always setting this ensures that things like
3154          * getparam()/getattr() don't report silly values for !rt tasks.
3155          */
3156         p->rt_priority = attr->sched_priority;
3157         p->normal_prio = normal_prio(p);
3158         set_load_weight(p);
3159 }
3160 
3161 /* Actually do priority change: must hold pi & rq lock. */
3162 static void __setscheduler(struct rq *rq, struct task_struct *p,
3163                            const struct sched_attr *attr)
3164 {
3165         __setscheduler_params(p, attr);
3166 
3167         /*
3168          * If we get here, there was no pi waiters boosting the
3169          * task. It is safe to use the normal prio.
3170          */
3171         p->prio = normal_prio(p);
3172 
3173         if (dl_prio(p->prio))
3174                 p->sched_class = &dl_sched_class;
3175         else if (rt_prio(p->prio))
3176                 p->sched_class = &rt_sched_class;
3177         else
3178                 p->sched_class = &fair_sched_class;
3179 }
3180 
3181 static void
3182 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3183 {
3184         struct sched_dl_entity *dl_se = &p->dl;
3185 
3186         attr->sched_priority = p->rt_priority;
3187         attr->sched_runtime = dl_se->dl_runtime;
3188         attr->sched_deadline = dl_se->dl_deadline;
3189         attr->sched_period = dl_se->dl_period;
3190         attr->sched_flags = dl_se->flags;
3191 }
3192 
3193 /*
3194  * This function validates the new parameters of a -deadline task.
3195  * We ask for the deadline not being zero, and greater or equal
3196  * than the runtime, as well as the period of being zero or
3197  * greater than deadline. Furthermore, we have to be sure that
3198  * user parameters are above the internal resolution of 1us (we
3199  * check sched_runtime only since it is always the smaller one) and
3200  * below 2^63 ns (we have to check both sched_deadline and
3201  * sched_period, as the latter can be zero).
3202  */
3203 static bool
3204 __checkparam_dl(const struct sched_attr *attr)
3205 {
3206         /* deadline != 0 */
3207         if (attr->sched_deadline == 0)
3208                 return false;
3209 
3210         /*
3211          * Since we truncate DL_SCALE bits, make sure we're at least
3212          * that big.
3213          */
3214         if (attr->sched_runtime < (1ULL << DL_SCALE))
3215                 return false;
3216 
3217         /*
3218          * Since we use the MSB for wrap-around and sign issues, make
3219          * sure it's not set (mind that period can be equal to zero).
3220          */
3221         if (attr->sched_deadline & (1ULL << 63) ||
3222             attr->sched_period & (1ULL << 63))
3223                 return false;
3224 
3225         /* runtime <= deadline <= period (if period != 0) */
3226         if ((attr->sched_period != 0 &&
3227              attr->sched_period < attr->sched_deadline) ||
3228             attr->sched_deadline < attr->sched_runtime)
3229                 return false;
3230 
3231         return true;
3232 }
3233 
3234 /*
3235  * check the target process has a UID that matches the current process's
3236  */
3237 static bool check_same_owner(struct task_struct *p)
3238 {
3239         const struct cred *cred = current_cred(), *pcred;
3240         bool match;
3241 
3242         rcu_read_lock();
3243         pcred = __task_cred(p);
3244         match = (uid_eq(cred->euid, pcred->euid) ||
3245                  uid_eq(cred->euid, pcred->uid));
3246         rcu_read_unlock();
3247         return match;
3248 }
3249 
3250 static int __sched_setscheduler(struct task_struct *p,
3251                                 const struct sched_attr *attr,
3252                                 bool user)
3253 {
3254         int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3255                       MAX_RT_PRIO - 1 - attr->sched_priority;
3256         int retval, oldprio, oldpolicy = -1, on_rq, running;
3257         int policy = attr->sched_policy;
3258         unsigned long flags;
3259         const struct sched_class *prev_class;
3260         struct rq *rq;
3261         int reset_on_fork;
3262 
3263         /* may grab non-irq protected spin_locks */
3264         BUG_ON(in_interrupt());
3265 recheck:
3266         /* double check policy once rq lock held */
3267         if (policy < 0) {
3268                 reset_on_fork = p->sched_reset_on_fork;
3269                 policy = oldpolicy = p->policy;
3270         } else {
3271                 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3272 
3273                 if (policy != SCHED_DEADLINE &&
3274                                 policy != SCHED_FIFO && policy != SCHED_RR &&
3275                                 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3276                                 policy != SCHED_IDLE)
3277                         return -EINVAL;
3278         }
3279 
3280         if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3281                 return -EINVAL;
3282 
3283         /*
3284          * Valid priorities for SCHED_FIFO and SCHED_RR are
3285          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3286          * SCHED_BATCH and SCHED_IDLE is 0.
3287          */
3288         if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3289             (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3290                 return -EINVAL;
3291         if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3292             (rt_policy(policy) != (attr->sched_priority != 0)))
3293                 return -EINVAL;
3294 
3295         /*
3296          * Allow unprivileged RT tasks to decrease priority:
3297          */
3298         if (user && !capable(CAP_SYS_NICE)) {
3299                 if (fair_policy(policy)) {
3300                         if (attr->sched_nice < task_nice(p) &&
3301                             !can_nice(p, attr->sched_nice))
3302                                 return -EPERM;
3303                 }
3304 
3305                 if (rt_policy(policy)) {
3306                         unsigned long rlim_rtprio =
3307                                         task_rlimit(p, RLIMIT_RTPRIO);
3308 
3309                         /* can't set/change the rt policy */
3310                         if (policy != p->policy && !rlim_rtprio)
3311                                 return -EPERM;
3312 
3313                         /* can't increase priority */
3314                         if (attr->sched_priority > p->rt_priority &&
3315                             attr->sched_priority > rlim_rtprio)
3316                                 return -EPERM;
3317                 }
3318 
3319                  /*
3320                   * Can't set/change SCHED_DEADLINE policy at all for now
3321                   * (safest behavior); in the future we would like to allow
3322                   * unprivileged DL tasks to increase their relative deadline
3323                   * or reduce their runtime (both ways reducing utilization)
3324                   */
3325                 if (dl_policy(policy))
3326                         return -EPERM;
3327 
3328                 /*
3329                  * Treat SCHED_IDLE as nice 20. Only allow a switch to
3330                  * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3331                  */
3332                 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3333                         if (!can_nice(p, task_nice(p)))
3334                                 return -EPERM;
3335                 }
3336 
3337                 /* can't change other user's priorities */
3338                 if (!check_same_owner(p))
3339                         return -EPERM;
3340 
3341                 /* Normal users shall not reset the sched_reset_on_fork flag */
3342                 if (p->sched_reset_on_fork && !reset_on_fork)
3343                         return -EPERM;
3344         }
3345 
3346         if (user) {
3347                 retval = security_task_setscheduler(p);
3348                 if (retval)
3349                         return retval;
3350         }
3351 
3352         /*
3353          * make sure no PI-waiters arrive (or leave) while we are
3354          * changing the priority of the task:
3355          *
3356          * To be able to change p->policy safely, the appropriate
3357          * runqueue lock must be held.
3358          */
3359         rq = task_rq_lock(p, &flags);
3360 
3361         /*
3362          * Changing the policy of the stop threads its a very bad idea
3363          */
3364         if (p == rq->stop) {
3365                 task_rq_unlock(rq, p, &flags);
3366                 return -EINVAL;
3367         }
3368 
3369         /*
3370          * If not changing anything there's no need to proceed further,
3371          * but store a possible modification of reset_on_fork.
3372          */
3373         if (unlikely(policy == p->policy)) {
3374                 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3375                         goto change;
3376                 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3377                         goto change;
3378                 if (dl_policy(policy))
3379                         goto change;
3380 
3381                 p->sched_reset_on_fork = reset_on_fork;
3382                 task_rq_unlock(rq, p, &flags);
3383                 return 0;
3384         }
3385 change:
3386 
3387         if (user) {
3388 #ifdef CONFIG_RT_GROUP_SCHED
3389                 /*
3390                  * Do not allow realtime tasks into groups that have no runtime
3391                  * assigned.
3392                  */
3393                 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3394                                 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3395                                 !task_group_is_autogroup(task_group(p))) {
3396                         task_rq_unlock(rq, p, &flags);
3397                         return -EPERM;
3398                 }
3399 #endif
3400 #ifdef CONFIG_SMP
3401                 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3402                         cpumask_t *span = rq->rd->span;
3403 
3404                         /*
3405                          * Don't allow tasks with an affinity mask smaller than
3406                          * the entire root_domain to become SCHED_DEADLINE. We
3407                          * will also fail if there's no bandwidth available.
3408                          */
3409                         if (!cpumask_subset(span, &p->cpus_allowed) ||
3410                             rq->rd->dl_bw.bw == 0) {
3411                                 task_rq_unlock(rq, p, &flags);
3412                                 return -EPERM;
3413                         }
3414                 }
3415 #endif
3416         }
3417 
3418         /* recheck policy now with rq lock held */
3419         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3420                 policy = oldpolicy = -1;
3421                 task_rq_unlock(rq, p, &flags);
3422                 goto recheck;
3423         }
3424 
3425         /*
3426          * If setscheduling to SCHED_DEADLINE (or changing the parameters
3427          * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3428          * is available.
3429          */
3430         if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3431                 task_rq_unlock(rq, p, &flags);
3432                 return -EBUSY;
3433         }
3434 
3435         p->sched_reset_on_fork = reset_on_fork;
3436         oldprio = p->prio;
3437 
3438         /*
3439          * Special case for priority boosted tasks.
3440          *
3441          * If the new priority is lower or equal (user space view)
3442          * than the current (boosted) priority, we just store the new
3443          * normal parameters and do not touch the scheduler class and
3444          * the runqueue. This will be done when the task deboost
3445          * itself.
3446          */
3447         if (rt_mutex_check_prio(p, newprio)) {
3448                 __setscheduler_params(p, attr);
3449                 task_rq_unlock(rq, p, &flags);
3450                 return 0;
3451         }
3452 
3453         on_rq = p->on_rq;
3454         running = task_current(rq, p);
3455         if (on_rq)
3456                 dequeue_task(rq, p, 0);
3457         if (running)
3458                 p->sched_class->put_prev_task(rq, p);
3459 
3460         prev_class = p->sched_class;
3461         __setscheduler(rq, p, attr);
3462 
3463         if (running)
3464                 p->sched_class->set_curr_task(rq);
3465         if (on_rq) {
3466                 /*
3467                  * We enqueue to tail when the priority of a task is
3468                  * increased (user space view).
3469                  */
3470                 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3471         }
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, MIN_NICE, MAX_NICE);
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  * @flags: for future extension.
3673  */
3674 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3675                                unsigned int, flags)
3676 {
3677         struct sched_attr attr;
3678         struct task_struct *p;
3679         int retval;
3680 
3681         if (!uattr || pid < 0 || flags)
3682                 return -EINVAL;
3683 
3684         retval = sched_copy_attr(uattr, &attr);
3685         if (retval)
3686                 return retval;
3687 
3688         if ((int)attr.sched_policy < 0)
3689                 return -EINVAL;
3690 
3691         rcu_read_lock();
3692         retval = -ESRCH;
3693         p = find_process_by_pid(pid);
3694         if (p != NULL)
3695                 retval = sched_setattr(p, &attr);
3696         rcu_read_unlock();
3697 
3698         return retval;
3699 }
3700 
3701 /**
3702  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3703  * @pid: the pid in question.
3704  *
3705  * Return: On success, the policy of the thread. Otherwise, a negative error
3706  * code.
3707  */
3708 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3709 {
3710         struct task_struct *p;
3711         int retval;
3712 
3713         if (pid < 0)
3714                 return -EINVAL;
3715 
3716         retval = -ESRCH;
3717         rcu_read_lock();
3718         p = find_process_by_pid(pid);
3719         if (p) {
3720                 retval = security_task_getscheduler(p);
3721                 if (!retval)
3722                         retval = p->policy
3723                                 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3724         }
3725         rcu_read_unlock();
3726         return retval;
3727 }
3728 
3729 /**
3730  * sys_sched_getparam - get the RT priority of a thread
3731  * @pid: the pid in question.
3732  * @param: structure containing the RT priority.
3733  *
3734  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3735  * code.
3736  */
3737 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3738 {
3739         struct sched_param lp = { .sched_priority = 0 };
3740         struct task_struct *p;
3741         int retval;
3742 
3743         if (!param || pid < 0)
3744                 return -EINVAL;
3745 
3746         rcu_read_lock();
3747         p = find_process_by_pid(pid);
3748         retval = -ESRCH;
3749         if (!p)
3750                 goto out_unlock;
3751 
3752         retval = security_task_getscheduler(p);
3753         if (retval)
3754                 goto out_unlock;
3755 
3756         if (task_has_rt_policy(p))
3757                 lp.sched_priority = p->rt_priority;
3758         rcu_read_unlock();
3759 
3760         /*
3761          * This one might sleep, we cannot do it with a spinlock held ...
3762          */
3763         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3764 
3765         return retval;
3766 
3767 out_unlock:
3768         rcu_read_unlock();
3769         return retval;
3770 }
3771 
3772 static int sched_read_attr(struct sched_attr __user *uattr,
3773                            struct sched_attr *attr,
3774                            unsigned int usize)
3775 {
3776         int ret;
3777 
3778         if (!access_ok(VERIFY_WRITE, uattr, usize))
3779                 return -EFAULT;
3780 
3781         /*
3782          * If we're handed a smaller struct than we know of,
3783          * ensure all the unknown bits are 0 - i.e. old
3784          * user-space does not get uncomplete information.
3785          */
3786         if (usize < sizeof(*attr)) {
3787                 unsigned char *addr;
3788                 unsigned char *end;
3789 
3790                 addr = (void *)attr + usize;
3791                 end  = (void *)attr + sizeof(*attr);
3792 
3793                 for (; addr < end; addr++) {
3794                         if (*addr)
3795                                 goto err_size;
3796                 }
3797 
3798                 attr->size = usize;
3799         }
3800 
3801         ret = copy_to_user(uattr, attr, attr->size);
3802         if (ret)
3803                 return -EFAULT;
3804 
3805 out:
3806         return ret;
3807 
3808 err_size:
3809         ret = -E2BIG;
3810         goto out;
3811 }
3812 
3813 /**
3814  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3815  * @pid: the pid in question.
3816  * @uattr: structure containing the extended parameters.
3817  * @size: sizeof(attr) for fwd/bwd comp.
3818  * @flags: for future extension.
3819  */
3820 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3821                 unsigned int, size, unsigned int, flags)
3822 {
3823         struct sched_attr attr = {
3824                 .size = sizeof(struct sched_attr),
3825         };
3826         struct task_struct *p;
3827         int retval;
3828 
3829         if (!uattr || pid < 0 || size > PAGE_SIZE ||
3830             size < SCHED_ATTR_SIZE_VER0 || flags)
3831                 return -EINVAL;
3832 
3833         rcu_read_lock();
3834         p = find_process_by_pid(pid);
3835         retval = -ESRCH;
3836         if (!p)
3837                 goto out_unlock;
3838 
3839         retval = security_task_getscheduler(p);
3840         if (retval)
3841                 goto out_unlock;
3842 
3843         attr.sched_policy = p->policy;
3844         if (p->sched_reset_on_fork)
3845                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3846         if (task_has_dl_policy(p))
3847                 __getparam_dl(p, &attr);
3848         else if (task_has_rt_policy(p))
3849                 attr.sched_priority = p->rt_priority;
3850         else
3851                 attr.sched_nice = task_nice(p);
3852 
3853         rcu_read_unlock();
3854 
3855         retval = sched_read_attr(uattr, &attr, size);
3856         return retval;
3857 
3858 out_unlock:
3859         rcu_read_unlock();
3860         return retval;
3861 }
3862 
3863 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3864 {
3865         cpumask_var_t cpus_allowed, new_mask;
3866         struct task_struct *p;
3867         int retval;
3868 
3869         rcu_read_lock();
3870 
3871         p = find_process_by_pid(pid);
3872         if (!p) {
3873                 rcu_read_unlock();
3874                 return -ESRCH;
3875         }
3876 
3877         /* Prevent p going away */
3878         get_task_struct(p);
3879         rcu_read_unlock();
3880 
3881         if (p->flags & PF_NO_SETAFFINITY) {
3882                 retval = -EINVAL;
3883                 goto out_put_task;
3884         }
3885         if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3886                 retval = -ENOMEM;
3887                 goto out_put_task;
3888         }
3889         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3890                 retval = -ENOMEM;
3891                 goto out_free_cpus_allowed;
3892         }
3893         retval = -EPERM;
3894         if (!check_same_owner(p)) {
3895                 rcu_read_lock();
3896                 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3897                         rcu_read_unlock();
3898                         goto out_unlock;
3899                 }
3900                 rcu_read_unlock();
3901         }
3902 
3903         retval = security_task_setscheduler(p);
3904         if (retval)
3905                 goto out_unlock;
3906 
3907 
3908         cpuset_cpus_allowed(p, cpus_allowed);
3909         cpumask_and(new_mask, in_mask, cpus_allowed);
3910 
3911         /*
3912          * Since bandwidth control happens on root_domain basis,
3913          * if admission test is enabled, we only admit -deadline
3914          * tasks allowed to run on all the CPUs in the task's
3915          * root_domain.
3916          */
3917 #ifdef CONFIG_SMP
3918         if (task_has_dl_policy(p)) {
3919                 const struct cpumask *span = task_rq(p)->rd->span;
3920 
3921                 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3922                         retval = -EBUSY;
3923                         goto out_unlock;
3924                 }
3925         }
3926 #endif
3927 again:
3928         retval = set_cpus_allowed_ptr(p, new_mask);
3929 
3930         if (!retval) {
3931                 cpuset_cpus_allowed(p, cpus_allowed);
3932                 if (!cpumask_subset(new_mask, cpus_allowed)) {
3933                         /*
3934                          * We must have raced with a concurrent cpuset
3935                          * update. Just reset the cpus_allowed to the
3936                          * cpuset's cpus_allowed
3937                          */
3938                         cpumask_copy(new_mask, cpus_allowed);
3939                         goto again;
3940                 }
3941         }
3942 out_unlock:
3943         free_cpumask_var(new_mask);
3944 out_free_cpus_allowed:
3945         free_cpumask_var(cpus_allowed);
3946 out_put_task:
3947         put_task_struct(p);
3948         return retval;
3949 }
3950 
3951 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3952                              struct cpumask *new_mask)
3953 {
3954         if (len < cpumask_size())
3955                 cpumask_clear(new_mask);
3956         else if (len > cpumask_size())
3957                 len = cpumask_size();
3958 
3959         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3960 }
3961 
3962 /**
3963  * sys_sched_setaffinity - set the cpu affinity of a process
3964  * @pid: pid of the process
3965  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3966  * @user_mask_ptr: user-space pointer to the new cpu mask
3967  *
3968  * Return: 0 on success. An error code otherwise.
3969  */
3970 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3971                 unsigned long __user *, user_mask_ptr)
3972 {
3973         cpumask_var_t new_mask;
3974         int retval;
3975 
3976         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3977                 return -ENOMEM;
3978 
3979         retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3980         if (retval == 0)
3981                 retval = sched_setaffinity(pid, new_mask);
3982         free_cpumask_var(new_mask);
3983         return retval;
3984 }
3985 
3986 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3987 {
3988         struct task_struct *p;
3989         unsigned long flags;
3990         int retval;
3991 
3992         rcu_read_lock();
3993 
3994         retval = -ESRCH;
3995         p = find_process_by_pid(pid);
3996         if (!p)
3997                 goto out_unlock;
3998 
3999         retval = security_task_getscheduler(p);
4000         if (retval)
4001                 goto out_unlock;
4002 
4003         raw_spin_lock_irqsave(&p->pi_lock, flags);
4004         cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4005         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4006 
4007 out_unlock:
4008         rcu_read_unlock();
4009 
4010         return retval;
4011 }
4012 
4013 /**
4014  * sys_sched_getaffinity - get the cpu affinity of a process
4015  * @pid: pid of the process
4016  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4017  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4018  *
4019  * Return: 0 on success. An error code otherwise.
4020  */
4021 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4022                 unsigned long __user *, user_mask_ptr)
4023 {
4024         int ret;
4025         cpumask_var_t mask;
4026 
4027         if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4028                 return -EINVAL;
4029         if (len & (sizeof(unsigned long)-1))
4030                 return -EINVAL;
4031 
4032         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4033                 return -ENOMEM;
4034 
4035         ret = sched_getaffinity(pid, mask);
4036         if (ret == 0) {
4037                 size_t retlen = min_t(size_t, len, cpumask_size());
4038 
4039                 if (copy_to_user(user_mask_ptr, mask, retlen))
4040                         ret = -EFAULT;
4041                 else
4042                         ret = retlen;
4043         }
4044         free_cpumask_var(mask);
4045 
4046         return ret;
4047 }
4048 
4049 /**
4050  * sys_sched_yield - yield the current processor to other threads.
4051  *
4052  * This function yields the current CPU to other tasks. If there are no
4053  * other threads running on this CPU then this function will return.
4054  *
4055  * Return: 0.
4056  */
4057 SYSCALL_DEFINE0(sched_yield)
4058 {
4059         struct rq *rq = this_rq_lock();
4060 
4061         schedstat_inc(rq, yld_count);
4062         current->sched_class->yield_task(rq);
4063 
4064         /*
4065          * Since we are going to call schedule() anyway, there's
4066          * no need to preempt or enable interrupts:
4067          */
4068         __release(rq->lock);
4069         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4070         do_raw_spin_unlock(&rq->lock);
4071         sched_preempt_enable_no_resched();
4072 
4073         schedule();
4074 
4075         return 0;
4076 }
4077 
4078 static void __cond_resched(void)
4079 {
4080         __preempt_count_add(PREEMPT_ACTIVE);
4081         __schedule();
4082         __preempt_count_sub(PREEMPT_ACTIVE);
4083 }
4084 
4085 int __sched _cond_resched(void)
4086 {
4087         if (should_resched()) {
4088                 __cond_resched();
4089                 return 1;
4090         }
4091         return 0;
4092 }
4093 EXPORT_SYMBOL(_cond_resched);
4094 
4095 /*
4096  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4097  * call schedule, and on return reacquire the lock.
4098  *
4099  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4100  * operations here to prevent schedule() from being called twice (once via
4101  * spin_unlock(), once by hand).
4102  */
4103 int __cond_resched_lock(spinlock_t *lock)
4104 {
4105         int resched = should_resched();
4106         int ret = 0;
4107 
4108         lockdep_assert_held(lock);
4109 
4110         if (spin_needbreak(lock) || resched) {
4111                 spin_unlock(lock);
4112                 if (resched)
4113                         __cond_resched();
4114                 else
4115                         cpu_relax();
4116                 ret = 1;
4117                 spin_lock(lock);
4118         }
4119         return ret;
4120 }
4121 EXPORT_SYMBOL(__cond_resched_lock);
4122 
4123 int __sched __cond_resched_softirq(void)
4124 {
4125         BUG_ON(!in_softirq());
4126 
4127         if (should_resched()) {
4128                 local_bh_enable();
4129                 __cond_resched();
4130                 local_bh_disable();
4131                 return 1;
4132         }
4133         return 0;
4134 }
4135 EXPORT_SYMBOL(__cond_resched_softirq);
4136 
4137 /**
4138  * yield - yield the current processor to other threads.
4139  *
4140  * Do not ever use this function, there's a 99% chance you're doing it wrong.
4141  *
4142  * The scheduler is at all times free to pick the calling task as the most
4143  * eligible task to run, if removing the yield() call from your code breaks
4144  * it, its already broken.
4145  *
4146  * Typical broken usage is:
4147  *
4148  * while (!event)
4149  *      yield();
4150  *
4151  * where one assumes that yield() will let 'the other' process run that will
4152  * make event true. If the current task is a SCHED_FIFO task that will never
4153  * happen. Never use yield() as a progress guarantee!!
4154  *
4155  * If you want to use yield() to wait for something, use wait_event().
4156  * If you want to use yield() to be 'nice' for others, use cond_resched().
4157  * If you still want to use yield(), do not!
4158  */
4159 void __sched yield(void)
4160 {
4161         set_current_state(TASK_RUNNING);
4162         sys_sched_yield();
4163 }
4164 EXPORT_SYMBOL(yield);
4165 
4166 /**
4167  * yield_to - yield the current processor to another thread in
4168  * your thread group, or accelerate that thread toward the
4169  * processor it's on.
4170  * @p: target task
4171  * @preempt: whether task preemption is allowed or not
4172  *
4173  * It's the caller's job to ensure that the target task struct
4174  * can't go away on us before we can do any checks.
4175  *
4176  * Return:
4177  *      true (>0) if we indeed boosted the target task.
4178  *      false (0) if we failed to boost the target.
4179  *      -ESRCH if there's no task to yield to.
4180  */
4181 bool __sched yield_to(struct task_struct *p, bool preempt)
4182 {
4183         struct task_struct *curr = current;
4184         struct rq *rq, *p_rq;
4185         unsigned long flags;
4186         int yielded = 0;
4187 
4188         local_irq_save(flags);
4189         rq = this_rq();
4190 
4191 again:
4192         p_rq = task_rq(p);
4193         /*
4194          * If we're the only runnable task on the rq and target rq also
4195          * has only one task, there's absolutely no point in yielding.
4196          */
4197         if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4198                 yielded = -ESRCH;
4199                 goto out_irq;
4200         }
4201 
4202         double_rq_lock(rq, p_rq);
4203         if (task_rq(p) != p_rq) {
4204                 double_rq_unlock(rq, p_rq);
4205                 goto again;
4206         }
4207 
4208         if (!curr->sched_class->yield_to_task)
4209                 goto out_unlock;
4210 
4211         if (curr->sched_class != p->sched_class)
4212                 goto out_unlock;
4213 
4214         if (task_running(p_rq, p) || p->state)
4215                 goto out_unlock;
4216 
4217         yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4218         if (yielded) {
4219                 schedstat_inc(rq, yld_count);
4220                 /*
4221                  * Make p's CPU reschedule; pick_next_entity takes care of
4222                  * fairness.
4223                  */
4224                 if (preempt && rq != p_rq)
4225                         resched_task(p_rq->curr);
4226         }
4227 
4228 out_unlock:
4229         double_rq_unlock(rq, p_rq);
4230 out_irq:
4231         local_irq_restore(flags);
4232 
4233         if (yielded > 0)
4234                 schedule();
4235 
4236         return yielded;
4237 }
4238 EXPORT_SYMBOL_GPL(yield_to);
4239 
4240 /*
4241  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4242  * that process accounting knows that this is a task in IO wait state.
4243  */
4244 void __sched io_schedule(void)
4245 {
4246         struct rq *rq = raw_rq();
4247 
4248         delayacct_blkio_start();
4249         atomic_inc(&rq->nr_iowait);
4250         blk_flush_plug(current);
4251         current->in_iowait = 1;
4252         schedule();
4253         current->in_iowait = 0;
4254         atomic_dec(&rq->nr_iowait);
4255         delayacct_blkio_end();
4256 }
4257 EXPORT_SYMBOL(io_schedule);
4258 
4259 long __sched io_schedule_timeout(long timeout)
4260 {
4261         struct rq *rq = raw_rq();
4262         long ret;
4263 
4264         delayacct_blkio_start();
4265         atomic_inc(&rq->nr_iowait);
4266         blk_flush_plug(current);
4267         current->in_iowait = 1;
4268         ret = schedule_timeout(timeout);
4269         current->in_iowait = 0;
4270         atomic_dec(&rq->nr_iowait);
4271         delayacct_blkio_end();
4272         return ret;
4273 }
4274 
4275 /**
4276  * sys_sched_get_priority_max - return maximum RT priority.
4277  * @policy: scheduling class.
4278  *
4279  * Return: On success, this syscall returns the maximum
4280  * rt_priority that can be used by a given scheduling class.
4281  * On failure, a negative error code is returned.
4282  */
4283 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4284 {
4285         int ret = -EINVAL;
4286 
4287         switch (policy) {
4288         case SCHED_FIFO:
4289         case SCHED_RR:
4290                 ret = MAX_USER_RT_PRIO-1;
4291                 break;
4292         case SCHED_DEADLINE:
4293         case SCHED_NORMAL:
4294         case SCHED_BATCH:
4295         case SCHED_IDLE:
4296                 ret = 0;
4297                 break;
4298         }
4299         return ret;
4300 }
4301 
4302 /**
4303  * sys_sched_get_priority_min - return minimum RT priority.
4304  * @policy: scheduling class.
4305  *
4306  * Return: On success, this syscall returns the minimum
4307  * rt_priority that can be used by a given scheduling class.
4308  * On failure, a negative error code is returned.
4309  */
4310 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4311 {
4312         int ret = -EINVAL;
4313 
4314         switch (policy) {
4315         case SCHED_FIFO:
4316         case SCHED_RR:
4317                 ret = 1;
4318                 break;
4319         case SCHED_DEADLINE:
4320         case SCHED_NORMAL:
4321         case SCHED_BATCH:
4322         case SCHED_IDLE:
4323                 ret = 0;
4324         }
4325         return ret;
4326 }
4327 
4328 /**
4329  * sys_sched_rr_get_interval - return the default timeslice of a process.
4330  * @pid: pid of the process.
4331  * @interval: userspace pointer to the timeslice value.
4332  *
4333  * this syscall writes the default timeslice value of a given process
4334  * into the user-space timespec buffer. A value of '' means infinity.
4335  *
4336  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4337  * an error code.
4338  */
4339 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4340                 struct timespec __user *, interval)
4341 {
4342         struct task_struct *p;
4343         unsigned int time_slice;
4344         unsigned long flags;
4345         struct rq *rq;
4346         int retval;
4347         struct timespec t;
4348 
4349         if (pid < 0)
4350                 return -EINVAL;
4351 
4352         retval = -ESRCH;
4353         rcu_read_lock();
4354         p = find_process_by_pid(pid);
4355         if (!p)
4356                 goto out_unlock;
4357 
4358         retval = security_task_getscheduler(p);
4359         if (retval)
4360                 goto out_unlock;
4361 
4362         rq = task_rq_lock(p, &flags);
4363         time_slice = 0;
4364         if (p->sched_class->get_rr_interval)
4365                 time_slice = p->sched_class->get_rr_interval(rq, p);
4366         task_rq_unlock(rq, p, &flags);
4367 
4368         rcu_read_unlock();
4369         jiffies_to_timespec(time_slice, &t);
4370         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4371         return retval;
4372 
4373 out_unlock:
4374         rcu_read_unlock();
4375         return retval;
4376 }
4377 
4378 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4379 
4380 void sched_show_task(struct task_struct *p)
4381 {
4382         unsigned long free = 0;
4383         int ppid;
4384         unsigned state;
4385 
4386         state = p->state ? __ffs(p->state) + 1 : 0;
4387         printk(KERN_INFO "%-15.15s %c", p->comm,
4388                 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4389 #if BITS_PER_LONG == 32
4390         if (state == TASK_RUNNING)
4391                 printk(KERN_CONT " running  ");
4392         else
4393                 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4394 #else
4395         if (state == TASK_RUNNING)
4396                 printk(KERN_CONT "  running task    ");
4397         else
4398                 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4399 #endif
4400 #ifdef CONFIG_DEBUG_STACK_USAGE
4401         free = stack_not_used(p);
4402 #endif
4403         rcu_read_lock();
4404         ppid = task_pid_nr(rcu_dereference(p->real_parent));
4405         rcu_read_unlock();
4406         printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4407                 task_pid_nr(p), ppid,
4408                 (unsigned long)task_thread_info(p)->flags);
4409 
4410         print_worker_info(KERN_INFO, p);
4411         show_stack(p, NULL);
4412 }
4413 
4414 void show_state_filter(unsigned long state_filter)
4415 {
4416         struct task_struct *g, *p;
4417 
4418 #if BITS_PER_LONG == 32
4419         printk(KERN_INFO
4420                 "  task                PC stack   pid father\n");
4421 #else
4422         printk(KERN_INFO
4423                 "  task                        PC stack   pid father\n");
4424 #endif
4425         rcu_read_lock();
4426         do_each_thread(g, p) {
4427                 /*
4428                  * reset the NMI-timeout, listing all files on a slow
4429                  * console might take a lot of time:
4430                  */
4431                 touch_nmi_watchdog();
4432                 if (!state_filter || (p->state & state_filter))
4433                         sched_show_task(p);
4434         } while_each_thread(g, p);
4435 
4436         touch_all_softlockup_watchdogs();
4437 
4438 #ifdef CONFIG_SCHED_DEBUG
4439         sysrq_sched_debug_show();
4440 #endif
4441         rcu_read_unlock();
4442         /*
4443          * Only show locks if all tasks are dumped:
4444          */
4445         if (!state_filter)
4446                 debug_show_all_locks();
4447 }
4448 
4449 void init_idle_bootup_task(struct task_struct *idle)
4450 {
4451         idle->sched_class = &idle_sched_class;
4452 }
4453 
4454 /**
4455  * init_idle - set up an idle thread for a given CPU
4456  * @idle: task in question
4457  * @cpu: cpu the idle task belongs to
4458  *
4459  * NOTE: this function does not set the idle thread's NEED_RESCHED
4460  * flag, to make booting more robust.
4461  */
4462 void init_idle(struct task_struct *idle, int cpu)
4463 {
4464         struct rq *rq = cpu_rq(cpu);
4465         unsigned long flags;
4466 
4467         raw_spin_lock_irqsave(&rq->lock, flags);
4468 
4469         __sched_fork(0, idle);
4470         idle->state = TASK_RUNNING;
4471         idle->se.exec_start = sched_clock();
4472 
4473         do_set_cpus_allowed(idle, cpumask_of(cpu));
4474         /*
4475          * We're having a chicken and egg problem, even though we are
4476          * holding rq->lock, the cpu isn't yet set to this cpu so the
4477          * lockdep check in task_group() will fail.
4478          *
4479          * Similar case to sched_fork(). / Alternatively we could
4480          * use task_rq_lock() here and obtain the other rq->lock.
4481          *
4482          * Silence PROVE_RCU
4483          */
4484         rcu_read_lock();
4485         __set_task_cpu(idle, cpu);
4486         rcu_read_unlock();
4487 
4488         rq->curr = rq->idle = idle;
4489         idle->on_rq = 1;
4490 #if defined(CONFIG_SMP)
4491         idle->on_cpu = 1;
4492 #endif
4493         raw_spin_unlock_irqrestore(&rq->lock, flags);
4494 
4495         /* Set the preempt count _outside_ the spinlocks! */
4496         init_idle_preempt_count(idle, cpu);
4497 
4498         /*
4499          * The idle tasks have their own, simple scheduling class:
4500          */
4501         idle->sched_class = &idle_sched_class;
4502         ftrace_graph_init_idle_task(idle, cpu);
4503         vtime_init_idle(idle, cpu);
4504 #if defined(CONFIG_SMP)
4505         sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4506 #endif
4507 }
4508 
4509 #ifdef CONFIG_SMP
4510 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4511 {
4512         if (p->sched_class && p->sched_class->set_cpus_allowed)
4513                 p->sched_class->set_cpus_allowed(p, new_mask);
4514 
4515         cpumask_copy(&p->cpus_allowed, new_mask);
4516         p->nr_cpus_allowed = cpumask_weight(new_mask);
4517 }
4518 
4519 /*
4520  * This is how migration works:
4521  *
4522  * 1) we invoke migration_cpu_stop() on the target CPU using
4523  *    stop_one_cpu().
4524  * 2) stopper starts to run (implicitly forcing the migrated thread
4525  *    off the CPU)
4526  * 3) it checks whether the migrated task is still in the wrong runqueue.
4527  * 4) if it's in the wrong runqueue then the migration thread removes
4528  *    it and puts it into the right queue.
4529  * 5) stopper completes and stop_one_cpu() returns and the migration
4530  *    is done.
4531  */
4532 
4533 /*
4534  * Change a given task's CPU affinity. Migrate the thread to a
4535  * proper CPU and schedule it away if the CPU it's executing on
4536  * is removed from the allowed bitmask.
4537  *
4538  * NOTE: the caller must have a valid reference to the task, the
4539  * task must not exit() & deallocate itself prematurely. The
4540  * call is not atomic; no spinlocks may be held.
4541  */
4542 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4543 {
4544         unsigned long flags;
4545         struct rq *rq;
4546         unsigned int dest_cpu;
4547         int ret = 0;
4548 
4549         rq = task_rq_lock(p, &flags);
4550 
4551         if (cpumask_equal(&p->cpus_allowed, new_mask))
4552                 goto out;
4553 
4554         if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4555                 ret = -EINVAL;
4556                 goto out;
4557         }
4558 
4559         do_set_cpus_allowed(p, new_mask);
4560 
4561         /* Can the task run on the task's current CPU? If so, we're done */
4562         if (cpumask_test_cpu(task_cpu(p), new_mask))
4563                 goto out;
4564 
4565         dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4566         if (p->on_rq) {
4567                 struct migration_arg arg = { p, dest_cpu };
4568                 /* Need help from migration thread: drop lock and wait. */
4569                 task_rq_unlock(rq, p, &flags);
4570                 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4571                 tlb_migrate_finish(p->mm);
4572                 return 0;
4573         }
4574 out:
4575         task_rq_unlock(rq, p, &flags);
4576 
4577         return ret;
4578 }
4579 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4580 
4581 /*
4582  * Move (not current) task off this cpu, onto dest cpu. We're doing
4583  * this because either it can't run here any more (set_cpus_allowed()
4584  * away from this CPU, or CPU going down), or because we're
4585  * attempting to rebalance this task on exec (sched_exec).
4586  *
4587  * So we race with normal scheduler movements, but that's OK, as long
4588  * as the task is no longer on this CPU.
4589  *
4590  * Returns non-zero if task was successfully migrated.
4591  */
4592 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4593 {
4594         struct rq *rq_dest, *rq_src;
4595         int ret = 0;
4596 
4597         if (unlikely(!cpu_active(dest_cpu)))
4598                 return ret;
4599 
4600         rq_src = cpu_rq(src_cpu);
4601         rq_dest = cpu_rq(dest_cpu);
4602 
4603         raw_spin_lock(&p->pi_lock);
4604         double_rq_lock(rq_src, rq_dest);
4605         /* Already moved. */
4606         if (task_cpu(p) != src_cpu)
4607                 goto done;
4608         /* Affinity changed (again). */
4609         if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4610                 goto fail;
4611 
4612         /*
4613          * If we're not on a rq, the next wake-up will ensure we're
4614          * placed properly.
4615          */
4616         if (p->on_rq) {
4617                 dequeue_task(rq_src, p, 0);
4618                 set_task_cpu(p, dest_cpu);
4619                 enqueue_task(rq_dest, p, 0);
4620                 check_preempt_curr(rq_dest, p, 0);
4621         }
4622 done:
4623         ret = 1;
4624 fail:
4625         double_rq_unlock(rq_src, rq_dest);
4626         raw_spin_unlock(&p->pi_lock);
4627         return ret;
4628 }
4629 
4630 #ifdef CONFIG_NUMA_BALANCING
4631 /* Migrate current task p to target_cpu */
4632 int migrate_task_to(struct task_struct *p, int target_cpu)
4633 {
4634         struct migration_arg arg = { p, target_cpu };
4635         int curr_cpu = task_cpu(p);
4636 
4637         if (curr_cpu == target_cpu)
4638                 return 0;
4639 
4640         if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4641                 return -EINVAL;
4642 
4643         /* TODO: This is not properly updating schedstats */
4644 
4645         trace_sched_move_numa(p, curr_cpu, target_cpu);
4646         return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4647 }
4648 
4649 /*
4650  * Requeue a task on a given node and accurately track the number of NUMA
4651  * tasks on the runqueues
4652  */
4653 void sched_setnuma(struct task_struct *p, int nid)
4654 {
4655         struct rq *rq;
4656         unsigned long flags;
4657         bool on_rq, running;
4658 
4659         rq = task_rq_lock(p, &flags);
4660         on_rq = p->on_rq;
4661         running = task_current(rq, p);
4662 
4663         if (on_rq)
4664                 dequeue_task(rq, p, 0);
4665         if (running)
4666                 p->sched_class->put_prev_task(rq, p);
4667 
4668         p->numa_preferred_nid = nid;
4669 
4670         if (running)
4671                 p->sched_class->set_curr_task(rq);
4672         if (on_rq)
4673                 enqueue_task(rq, p, 0);
4674         task_rq_unlock(rq, p, &flags);
4675 }
4676 #endif
4677 
4678 /*
4679  * migration_cpu_stop - this will be executed by a highprio stopper thread
4680  * and performs thread migration by bumping thread off CPU then
4681  * 'pushing' onto another runqueue.
4682  */
4683 static int migration_cpu_stop(void *data)
4684 {
4685         struct migration_arg *arg = data;
4686 
4687         /*
4688          * The original target cpu might have gone down and we might
4689          * be on another cpu but it doesn't matter.
4690          */
4691         local_irq_disable();
4692         __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4693         local_irq_enable();
4694         return 0;
4695 }
4696 
4697 #ifdef CONFIG_HOTPLUG_CPU
4698 
4699 /*
4700  * Ensures that the idle task is using init_mm right before its cpu goes
4701  * offline.
4702  */
4703 void idle_task_exit(void)
4704 {
4705         struct mm_struct *mm = current->active_mm;
4706 
4707         BUG_ON(cpu_online(smp_processor_id()));
4708 
4709         if (mm != &init_mm) {
4710                 switch_mm(mm, &init_mm, current);
4711                 finish_arch_post_lock_switch();
4712         }
4713         mmdrop(mm);
4714 }
4715 
4716 /*
4717  * Since this CPU is going 'away' for a while, fold any nr_active delta
4718  * we might have. Assumes we're called after migrate_tasks() so that the
4719  * nr_active count is stable.
4720  *
4721  * Also see the comment "Global load-average calculations".
4722  */
4723 static void calc_load_migrate(struct rq *rq)
4724 {
4725         long delta = calc_load_fold_active(rq);
4726         if (delta)
4727                 atomic_long_add(delta, &calc_load_tasks);
4728 }
4729 
4730 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4731 {
4732 }
4733 
4734 static const struct sched_class fake_sched_class = {
4735         .put_prev_task = put_prev_task_fake,
4736 };
4737 
4738 static struct task_struct fake_task = {
4739         /*
4740          * Avoid pull_{rt,dl}_task()
4741          */
4742         .prio = MAX_PRIO + 1,
4743         .sched_class = &fake_sched_class,
4744 };
4745 
4746 /*
4747  * Migrate all tasks from the rq, sleeping tasks will be migrated by
4748  * try_to_wake_up()->select_task_rq().
4749  *
4750  * Called with rq->lock held even though we'er in stop_machine() and
4751  * there's no concurrency possible, we hold the required locks anyway
4752  * because of lock validation efforts.
4753  */
4754 static void migrate_tasks(unsigned int dead_cpu)
4755 {
4756         struct rq *rq = cpu_rq(dead_cpu);
4757         struct task_struct *next, *stop = rq->stop;
4758         int dest_cpu;
4759 
4760         /*
4761          * Fudge the rq selection such that the below task selection loop
4762          * doesn't get stuck on the currently eligible stop task.
4763          *
4764          * We're currently inside stop_machine() and the rq is either stuck
4765          * in the stop_machine_cpu_stop() loop, or we're executing this code,
4766          * either way we should never end up calling schedule() until we're
4767          * done here.
4768          */
4769         rq->stop = NULL;
4770 
4771         /*
4772          * put_prev_task() and pick_next_task() sched
4773          * class method both need to have an up-to-date
4774          * value of rq->clock[_task]
4775          */
4776         update_rq_clock(rq);
4777 
4778         for ( ; ; ) {
4779                 /*
4780                  * There's this thread running, bail when that's the only
4781                  * remaining thread.
4782                  */
4783                 if (rq->nr_running == 1)
4784                         break;
4785 
4786                 next = pick_next_task(rq, &fake_task);
4787                 BUG_ON(!next);
4788                 next->sched_class->put_prev_task(rq, next);
4789 
4790                 /* Find suitable destination for @next, with force if needed. */
4791                 dest_cpu = select_fallback_rq(dead_cpu, next);
4792                 raw_spin_unlock(&rq->lock);
4793 
4794                 __migrate_task(next, dead_cpu, dest_cpu);
4795 
4796                 raw_spin_lock(&rq->lock);
4797         }
4798 
4799         rq->stop = stop;
4800 }
4801 
4802 #endif /* CONFIG_HOTPLUG_CPU */
4803 
4804 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4805 
4806 static struct ctl_table sd_ctl_dir[] = {
4807         {
4808                 .procname       = "sched_domain",
4809                 .mode           = 0555,
4810         },
4811         {}
4812 };
4813 
4814 static struct ctl_table sd_ctl_root[] = {
4815         {
4816                 .procname       = "kernel",
4817                 .mode           = 0555,
4818                 .child          = sd_ctl_dir,
4819         },
4820         {}
4821 };
4822 
4823 static struct ctl_table *sd_alloc_ctl_entry(int n)
4824 {
4825         struct ctl_table *entry =
4826                 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4827 
4828         return entry;
4829 }
4830 
4831 static void sd_free_ctl_entry(struct ctl_table **tablep)
4832 {
4833         struct ctl_table *entry;
4834 
4835         /*
4836          * In the intermediate directories, both the child directory and
4837          * procname are dynamically allocated and could fail but the mode
4838          * will always be set. In the lowest directory the names are
4839          * static strings and all have proc handlers.
4840          */
4841         for (entry = *tablep; entry->mode; entry++) {
4842                 if (entry->child)
4843                         sd_free_ctl_entry(&entry->child);
4844                 if (entry->proc_handler == NULL)
4845                         kfree(entry->procname);
4846         }
4847 
4848         kfree(*tablep);
4849         *tablep = NULL;
4850 }
4851 
4852 static int min_load_idx = 0;
4853 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4854 
4855 static void
4856 set_table_entry(struct ctl_table *entry,
4857                 const char *procname, void *data, int maxlen,
4858                 umode_t mode, proc_handler *proc_handler,
4859                 bool load_idx)
4860 {
4861         entry->procname = procname;
4862         entry->data = data;
4863         entry->maxlen = maxlen;
4864         entry->mode = mode;
4865         entry->proc_handler = proc_handler;
4866 
4867         if (load_idx) {
4868                 entry->extra1 = &min_load_idx;
4869                 entry->extra2 = &max_load_idx;
4870         }
4871 }
4872 
4873 static struct ctl_table *
4874 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4875 {
4876         struct ctl_table *table = sd_alloc_ctl_entry(14);
4877 
4878         if (table == NULL)
4879                 return NULL;
4880 
4881         set_table_entry(&table[0], "min_interval", &sd->min_interval,
4882                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4883         set_table_entry(&table[1], "max_interval", &sd->max_interval,
4884                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4885         set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4886                 sizeof(int), 0644, proc_dointvec_minmax, true);
4887         set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4888                 sizeof(int), 0644, proc_dointvec_minmax, true);
4889         set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4890                 sizeof(int), 0644, proc_dointvec_minmax, true);
4891         set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4892                 sizeof(int), 0644, proc_dointvec_minmax, true);
4893         set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4894                 sizeof(int), 0644, proc_dointvec_minmax, true);
4895         set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4896                 sizeof(int), 0644, proc_dointvec_minmax, false);
4897         set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4898                 sizeof(int), 0644, proc_dointvec_minmax, false);
4899         set_table_entry(&table[9], "cache_nice_tries",
4900                 &sd->cache_nice_tries,
4901                 sizeof(int), 0644, proc_dointvec_minmax, false);
4902         set_table_entry(&table[10], "flags", &sd->flags,
4903                 sizeof(int), 0644, proc_dointvec_minmax, false);
4904         set_table_entry(&table[11], "max_newidle_lb_cost",
4905                 &sd->max_newidle_lb_cost,
4906                 sizeof(long), 0644, proc_doulongvec_minmax, false);
4907         set_table_entry(&table[12], "name", sd->name,
4908                 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4909         /* &table[13] is terminator */
4910 
4911         return table;
4912 }
4913 
4914 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4915 {
4916         struct ctl_table *entry, *table;
4917         struct sched_domain *sd;
4918         int domain_num = 0, i;
4919         char buf[32];
4920 
4921         for_each_domain(cpu, sd)
4922                 domain_num++;
4923         entry = table = sd_alloc_ctl_entry(domain_num + 1);
4924         if (table == NULL)
4925                 return NULL;
4926 
4927         i = 0;
4928         for_each_domain(cpu, sd) {
4929                 snprintf(buf, 32, "domain%d", i);
4930                 entry->procname = kstrdup(buf, GFP_KERNEL);
4931                 entry->mode = 0555;
4932                 entry->child = sd_alloc_ctl_domain_table(sd);
4933                 entry++;
4934                 i++;
4935         }
4936         return table;
4937 }
4938 
4939 static struct ctl_table_header *sd_sysctl_header;
4940 static void register_sched_domain_sysctl(void)
4941 {
4942         int i, cpu_num = num_possible_cpus();
4943         struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4944         char buf[32];
4945 
4946         WARN_ON(sd_ctl_dir[0].child);
4947         sd_ctl_dir[0].child = entry;
4948 
4949         if (entry == NULL)
4950                 return;
4951 
4952         for_each_possible_cpu(i) {
4953                 snprintf(buf, 32, "cpu%d", i);
4954                 entry->procname = kstrdup(buf, GFP_KERNEL);
4955                 entry->mode = 0555;
4956                 entry->child = sd_alloc_ctl_cpu_table(i);
4957                 entry++;
4958         }
4959 
4960         WARN_ON(sd_sysctl_header);
4961         sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4962 }
4963 
4964 /* may be called multiple times per register */
4965 static void unregister_sched_domain_sysctl(void)
4966 {
4967         if (sd_sysctl_header)
4968                 unregister_sysctl_table(sd_sysctl_header);
4969         sd_sysctl_header = NULL;
4970         if (sd_ctl_dir[0].child)
4971                 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4972 }
4973 #else
4974 static void register_sched_domain_sysctl(void)
4975 {
4976 }
4977 static void unregister_sched_domain_sysctl(void)
4978 {
4979 }
4980 #endif
4981 
4982 static void set_rq_online(struct rq *rq)
4983 {
4984         if (!rq->online) {
4985                 const struct sched_class *class;
4986 
4987                 cpumask_set_cpu(rq->cpu, rq->rd->online);
4988                 rq->online = 1;
4989 
4990                 for_each_class(class) {
4991                         if (class->rq_online)
4992                                 class->rq_online(rq);
4993                 }
4994         }
4995 }
4996 
4997 static void set_rq_offline(struct rq *rq)
4998 {
4999         if (rq->online) {
5000                 const struct sched_class *class;
5001 
5002                 for_each_class(class) {
5003                         if (class->rq_offline)
5004                                 class->rq_offline(rq);
5005                 }
5006 
5007                 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5008                 rq->online = 0;
5009         }
5010 }
5011 
5012 /*
5013  * migration_call - callback that gets triggered when a CPU is added.
5014  * Here we can start up the necessary migration thread for the new CPU.
5015  */
5016 static int
5017 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5018 {
5019         int cpu = (long)hcpu;
5020         unsigned long flags;
5021         struct rq *rq = cpu_rq(cpu);
5022 
5023         switch (action & ~CPU_TASKS_FROZEN) {
5024 
5025         case CPU_UP_PREPARE:
5026                 rq->calc_load_update = calc_load_update;
5027                 break;
5028 
5029         case CPU_ONLINE:
5030                 /* Update our root-domain */
5031                 raw_spin_lock_irqsave(&rq->lock, flags);
5032                 if (rq->rd) {
5033                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5034 
5035                         set_rq_online(rq);
5036                 }
5037                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5038                 break;
5039 
5040 #ifdef CONFIG_HOTPLUG_CPU
5041         case CPU_DYING:
5042                 sched_ttwu_pending();
5043                 /* Update our root-domain */
5044                 raw_spin_lock_irqsave(&rq->lock, flags);
5045                 if (rq->rd) {
5046                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5047                         set_rq_offline(rq);
5048                 }
5049                 migrate_tasks(cpu);
5050                 BUG_ON(rq->nr_running != 1); /* the migration thread */
5051                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5052                 break;
5053 
5054         case CPU_DEAD:
5055                 calc_load_migrate(rq);
5056                 break;
5057 #endif
5058         }
5059 
5060         update_max_interval();
5061 
5062         return NOTIFY_OK;
5063 }
5064 
5065 /*
5066  * Register at high priority so that task migration (migrate_all_tasks)
5067  * happens before everything else.  This has to be lower priority than
5068  * the notifier in the perf_event subsystem, though.
5069  */
5070 static struct notifier_block migration_notifier = {
5071         .notifier_call = migration_call,
5072         .priority = CPU_PRI_MIGRATION,
5073 };
5074 
5075 static int sched_cpu_active(struct notifier_block *nfb,
5076                                       unsigned long action, void *hcpu)
5077 {
5078         switch (action & ~CPU_TASKS_FROZEN) {
5079         case CPU_DOWN_FAILED:
5080                 set_cpu_active((long)hcpu, true);
5081                 return NOTIFY_OK;
5082         default:
5083                 return NOTIFY_DONE;
5084         }
5085 }
5086 
5087 static int sched_cpu_inactive(struct notifier_block *nfb,
5088                                         unsigned long action, void *hcpu)
5089 {
5090         unsigned long flags;
5091         long cpu = (long)hcpu;
5092 
5093         switch (action & ~CPU_TASKS_FROZEN) {
5094         case CPU_DOWN_PREPARE:
5095                 set_cpu_active(cpu, false);
5096 
5097                 /* explicitly allow suspend */
5098                 if (!(action & CPU_TASKS_FROZEN)) {
5099                         struct dl_bw *dl_b = dl_bw_of(cpu);
5100                         bool overflow;
5101                         int cpus;
5102 
5103                         raw_spin_lock_irqsave(&dl_b->lock, flags);
5104                         cpus = dl_bw_cpus(cpu);
5105                         overflow = __dl_overflow(dl_b, cpus, 0, 0);
5106                         raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5107 
5108                         if (overflow)
5109                                 return notifier_from_errno(-EBUSY);
5110                 }
5111                 return NOTIFY_OK;
5112         }
5113 
5114         return NOTIFY_DONE;
5115 }
5116 
5117 static int __init migration_init(void)
5118 {
5119         void *cpu = (void *)(long)smp_processor_id();
5120         int err;
5121 
5122         /* Initialize migration for the boot CPU */
5123         err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5124         BUG_ON(err == NOTIFY_BAD);
5125         migration_call(&migration_notifier, CPU_ONLINE, cpu);
5126         register_cpu_notifier(&migration_notifier);
5127 
5128         /* Register cpu active notifiers */
5129         cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5130         cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5131 
5132         return 0;
5133 }
5134 early_initcall(migration_init);
5135 #endif
5136 
5137 #ifdef CONFIG_SMP
5138 
5139 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5140 
5141 #ifdef CONFIG_SCHED_DEBUG
5142 
5143 static __read_mostly int sched_debug_enabled;
5144 
5145 static int __init sched_debug_setup(char *str)
5146 {
5147         sched_debug_enabled = 1;
5148 
5149         return 0;
5150 }
5151 early_param("sched_debug", sched_debug_setup);
5152 
5153 static inline bool sched_debug(void)
5154 {
5155         return sched_debug_enabled;
5156 }
5157 
5158 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5159                                   struct cpumask *groupmask)
5160 {
5161         struct sched_group *group = sd->groups;
5162         char str[256];
5163 
5164         cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5165         cpumask_clear(groupmask);
5166 
5167         printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5168 
5169         if (!(sd->flags & SD_LOAD_BALANCE)) {
5170                 printk("does not load-balance\n");
5171                 if (sd->parent)
5172                         printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5173                                         " has parent");
5174                 return -1;
5175         }
5176 
5177         printk(KERN_CONT "span %s level %s\n", str, sd->name);
5178 
5179         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5180                 printk(KERN_ERR "ERROR: domain->span does not contain "
5181                                 "CPU%d\n", cpu);
5182         }
5183         if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5184                 printk(KERN_ERR "ERROR: domain->groups does not contain"
5185                                 " CPU%d\n", cpu);
5186         }
5187 
5188         printk(KERN_DEBUG "%*s groups:", level + 1, "");
5189         do {
5190                 if (!group) {
5191                         printk("\n");
5192                         printk(KERN_ERR "ERROR: group is NULL\n");
5193                         break;
5194                 }
5195 
5196                 /*
5197                  * Even though we initialize ->power to something semi-sane,
5198                  * we leave power_orig unset. This allows us to detect if
5199                  * domain iteration is still funny without causing /0 traps.
5200                  */
5201                 if (!group->sgp->power_orig) {
5202                         printk(KERN_CONT "\n");
5203                         printk(KERN_ERR "ERROR: domain->cpu_power not "
5204                                         "set\n");
5205                         break;
5206                 }
5207 
5208                 if (!cpumask_weight(sched_group_cpus(group))) {
5209                         printk(KERN_CONT "\n");
5210                         printk(KERN_ERR "ERROR: empty group\n");
5211                         break;
5212                 }
5213 
5214                 if (!(sd->flags & SD_OVERLAP) &&
5215                     cpumask_intersects(groupmask, sched_group_cpus(group))) {
5216                         printk(KERN_CONT "\n");
5217                         printk(KERN_ERR "ERROR: repeated CPUs\n");
5218                         break;
5219                 }
5220 
5221                 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5222 
5223                 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5224 
5225                 printk(KERN_CONT " %s", str);
5226                 if (group->sgp->power != SCHED_POWER_SCALE) {
5227                         printk(KERN_CONT " (cpu_power = %d)",
5228                                 group->sgp->power);
5229                 }
5230 
5231                 group = group->next;
5232         } while (group != sd->groups);
5233         printk(KERN_CONT "\n");
5234 
5235         if (!cpumask_equal(sched_domain_span(sd), groupmask))
5236                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5237 
5238         if (sd->parent &&
5239             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5240                 printk(KERN_ERR "ERROR: parent span is not a superset "
5241                         "of domain->span\n");
5242         return 0;
5243 }
5244 
5245 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5246 {
5247         int level = 0;
5248 
5249         if (!sched_debug_enabled)
5250                 return;
5251 
5252         if (!sd) {
5253                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5254                 return;
5255         }
5256 
5257         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5258 
5259         for (;;) {
5260                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5261                         break;
5262                 level++;
5263                 sd = sd->parent;
5264                 if (!sd)
5265                         break;
5266         }
5267 }
5268 #else /* !CONFIG_SCHED_DEBUG */
5269 # define sched_domain_debug(sd, cpu) do { } while (0)
5270 static inline bool sched_debug(void)
5271 {
5272         return false;
5273 }
5274 #endif /* CONFIG_SCHED_DEBUG */
5275 
5276 static int sd_degenerate(struct sched_domain *sd)
5277 {
5278         if (cpumask_weight(sched_domain_span(sd)) == 1)
5279                 return 1;
5280 
5281         /* Following flags need at least 2 groups */
5282         if (sd->flags & (SD_LOAD_BALANCE |
5283                          SD_BALANCE_NEWIDLE |
5284                          SD_BALANCE_FORK |
5285                          SD_BALANCE_EXEC |
5286                          SD_SHARE_CPUPOWER |
5287                          SD_SHARE_PKG_RESOURCES)) {
5288                 if (sd->groups != sd->groups->next)
5289                         return 0;
5290         }
5291 
5292         /* Following flags don't use groups */
5293         if (sd->flags & (SD_WAKE_AFFINE))
5294                 return 0;
5295 
5296         return 1;
5297 }
5298 
5299 static int
5300 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5301 {
5302         unsigned long cflags = sd->flags, pflags = parent->flags;
5303 
5304         if (sd_degenerate(parent))
5305                 return 1;
5306 
5307         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5308                 return 0;
5309 
5310         /* Flags needing groups don't count if only 1 group in parent */
5311         if (parent->groups == parent->groups->next) {
5312                 pflags &= ~(SD_LOAD_BALANCE |
5313                                 SD_BALANCE_NEWIDLE |
5314                                 SD_BALANCE_FORK |
5315                                 SD_BALANCE_EXEC |
5316                                 SD_SHARE_CPUPOWER |
5317                                 SD_SHARE_PKG_RESOURCES |
5318                                 SD_PREFER_SIBLING);
5319                 if (nr_node_ids == 1)
5320                         pflags &= ~SD_SERIALIZE;
5321         }
5322         if (~cflags & pflags)
5323                 return 0;
5324 
5325         return 1;
5326 }
5327 
5328 static void free_rootdomain(struct rcu_head *rcu)
5329 {
5330         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5331 
5332         cpupri_cleanup(&rd->cpupri);
5333         cpudl_cleanup(&rd->cpudl);
5334         free_cpumask_var(rd->dlo_mask);
5335         free_cpumask_var(rd->rto_mask);
5336         free_cpumask_var(rd->online);
5337         free_cpumask_var(rd->span);
5338         kfree(rd);
5339 }
5340 
5341 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5342 {
5343         struct root_domain *old_rd = NULL;
5344         unsigned long flags;
5345 
5346         raw_spin_lock_irqsave(&rq->lock, flags);
5347 
5348         if (rq->rd) {
5349                 old_rd = rq->rd;
5350 
5351                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5352                         set_rq_offline(rq);
5353 
5354                 cpumask_clear_cpu(rq->cpu, old_rd->span);
5355 
5356                 /*
5357                  * If we dont want to free the old_rd yet then
5358                  * set old_rd to NULL to skip the freeing later
5359                  * in this function:
5360                  */
5361                 if (!atomic_dec_and_test(&old_rd->refcount))
5362                         old_rd = NULL;
5363         }
5364 
5365         atomic_inc(&rd->refcount);
5366         rq->rd = rd;
5367 
5368         cpumask_set_cpu(rq->cpu, rd->span);
5369         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5370                 set_rq_online(rq);
5371 
5372         raw_spin_unlock_irqrestore(&rq->lock, flags);
5373 
5374         if (old_rd)
5375                 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5376 }
5377 
5378 static int init_rootdomain(struct root_domain *rd)
5379 {
5380         memset(rd, 0, sizeof(*rd));
5381 
5382         if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5383                 goto out;
5384         if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5385                 goto free_span;
5386         if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5387                 goto free_online;
5388         if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5389                 goto free_dlo_mask;
5390 
5391         init_dl_bw(&rd->dl_bw);
5392         if (cpudl_init(&rd->cpudl) != 0)
5393                 goto free_dlo_mask;
5394 
5395         if (cpupri_init(&rd->cpupri) != 0)
5396                 goto free_rto_mask;
5397         return 0;
5398 
5399 free_rto_mask:
5400         free_cpumask_var(rd->rto_mask);
5401 free_dlo_mask:
5402         free_cpumask_var(rd->dlo_mask);
5403 free_online:
5404         free_cpumask_var(rd->online);
5405 free_span:
5406         free_cpumask_var(rd->span);
5407 out:
5408         return -ENOMEM;
5409 }
5410 
5411 /*
5412  * By default the system creates a single root-domain with all cpus as
5413  * members (mimicking the global state we have today).
5414  */
5415 struct root_domain def_root_domain;
5416 
5417 static void init_defrootdomain(void)
5418 {
5419         init_rootdomain(&def_root_domain);
5420 
5421         atomic_set(&def_root_domain.refcount, 1);
5422 }
5423 
5424 static struct root_domain *alloc_rootdomain(void)
5425 {
5426         struct root_domain *rd;
5427 
5428         rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5429         if (!rd)
5430                 return NULL;
5431 
5432         if (init_rootdomain(rd) != 0) {
5433                 kfree(rd);
5434                 return NULL;
5435         }
5436 
5437         return rd;
5438 }
5439 
5440 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5441 {
5442         struct sched_group *tmp, *first;
5443 
5444         if (!sg)
5445                 return;
5446 
5447         first = sg;
5448         do {
5449                 tmp = sg->next;
5450 
5451                 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5452                         kfree(sg->sgp);
5453 
5454                 kfree(sg);
5455                 sg = tmp;
5456         } while (sg != first);
5457 }
5458 
5459 static void free_sched_domain(struct rcu_head *rcu)
5460 {
5461         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5462 
5463         /*
5464          * If its an overlapping domain it has private groups, iterate and
5465          * nuke them all.
5466          */
5467         if (sd->flags & SD_OVERLAP) {
5468                 free_sched_groups(sd->groups, 1);
5469         } else if (atomic_dec_and_test(&sd->groups->ref)) {
5470                 kfree(sd->groups->sgp);
5471                 kfree(sd->groups);
5472         }
5473         kfree(sd);
5474 }
5475 
5476 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5477 {
5478         call_rcu(&sd->rcu, free_sched_domain);
5479 }
5480 
5481 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5482 {
5483         for (; sd; sd = sd->parent)
5484                 destroy_sched_domain(sd, cpu);
5485 }
5486 
5487 /*
5488  * Keep a special pointer to the highest sched_domain that has
5489  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5490  * allows us to avoid some pointer chasing select_idle_sibling().
5491  *
5492  * Also keep a unique ID per domain (we use the first cpu number in
5493  * the cpumask of the domain), this allows us to quickly tell if
5494  * two cpus are in the same cache domain, see cpus_share_cache().
5495  */
5496 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5497 DEFINE_PER_CPU(int, sd_llc_size);
5498 DEFINE_PER_CPU(int, sd_llc_id);
5499 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5500 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5501 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5502 
5503 static void update_top_cache_domain(int cpu)
5504 {
5505         struct sched_domain *sd;
5506         struct sched_domain *busy_sd = NULL;
5507         int id = cpu;
5508         int size = 1;
5509 
5510         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5511         if (sd) {
5512                 id = cpumask_first(sched_domain_span(sd));
5513                 size = cpumask_weight(sched_domain_span(sd));
5514                 busy_sd = sd->parent; /* sd_busy */
5515         }
5516         rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5517 
5518         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5519         per_cpu(sd_llc_size, cpu) = size;
5520         per_cpu(sd_llc_id, cpu) = id;
5521 
5522         sd = lowest_flag_domain(cpu, SD_NUMA);
5523         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5524 
5525         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5526         rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5527 }
5528 
5529 /*
5530  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5531  * hold the hotplug lock.
5532  */
5533 static void
5534 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5535 {
5536         struct rq *rq = cpu_rq(cpu);
5537         struct sched_domain *tmp;
5538 
5539         /* Remove the sched domains which do not contribute to scheduling. */
5540         for (tmp = sd; tmp; ) {
5541                 struct sched_domain *parent = tmp->parent;
5542                 if (!parent)
5543                         break;
5544 
5545                 if (sd_parent_degenerate(tmp, parent)) {
5546                         tmp->parent = parent->parent;
5547                         if (parent->parent)
5548                                 parent->parent->child = tmp;
5549                         /*
5550                          * Transfer SD_PREFER_SIBLING down in case of a
5551                          * degenerate parent; the spans match for this
5552                          * so the property transfers.
5553                          */
5554                         if (parent->flags & SD_PREFER_SIBLING)
5555                                 tmp->flags |= SD_PREFER_SIBLING;
5556                         destroy_sched_domain(parent, cpu);
5557                 } else
5558                         tmp = tmp->parent;
5559         }
5560 
5561         if (sd && sd_degenerate(sd)) {
5562                 tmp = sd;
5563                 sd = sd->parent;
5564                 destroy_sched_domain(tmp, cpu);
5565                 if (sd)
5566                         sd->child = NULL;
5567         }
5568 
5569         sched_domain_debug(sd, cpu);
5570 
5571         rq_attach_root(rq, rd);
5572         tmp = rq->sd;
5573         rcu_assign_pointer(rq->sd, sd);
5574         destroy_sched_domains(tmp, cpu);
5575 
5576         update_top_cache_domain(cpu);
5577 }
5578 
5579 /* cpus with isolated domains */
5580 static cpumask_var_t cpu_isolated_map;
5581 
5582 /* Setup the mask of cpus configured for isolated domains */
5583 static int __init isolated_cpu_setup(char *str)
5584 {
5585         alloc_bootmem_cpumask_var(&cpu_isolated_map);
5586         cpulist_parse(str, cpu_isolated_map);
5587         return 1;
5588 }
5589 
5590 __setup("isolcpus=", isolated_cpu_setup);
5591 
5592 static const struct cpumask *cpu_cpu_mask(int cpu)
5593 {
5594         return cpumask_of_node(cpu_to_node(cpu));
5595 }
5596 
5597 struct sd_data {
5598         struct sched_domain **__percpu sd;
5599         struct sched_group **__percpu sg;
5600         struct sched_group_power **__percpu sgp;
5601 };
5602 
5603 struct s_data {
5604         struct sched_domain ** __percpu sd;
5605         struct root_domain      *rd;
5606 };
5607 
5608 enum s_alloc {
5609         sa_rootdomain,
5610         sa_sd,
5611         sa_sd_storage,
5612         sa_none,
5613 };
5614 
5615 struct sched_domain_topology_level;
5616 
5617 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5618 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5619 
5620 #define SDTL_OVERLAP    0x01
5621 
5622 struct sched_domain_topology_level {
5623         sched_domain_init_f init;
5624         sched_domain_mask_f mask;
5625         int                 flags;
5626         int                 numa_level;
5627         struct sd_data      data;
5628 };
5629 
5630 /*
5631  * Build an iteration mask that can exclude certain CPUs from the upwards
5632  * domain traversal.
5633  *
5634  * Asymmetric node setups can result in situations where the domain tree is of
5635  * unequal depth, make sure to skip domains that already cover the entire
5636  * range.
5637  *
5638  * In that case build_sched_domains() will have terminated the iteration early
5639  * and our sibling sd spans will be empty. Domains should always include the
5640  * cpu they're built on, so check that.
5641  *
5642  */
5643 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5644 {
5645         const struct cpumask *span = sched_domain_span(sd);
5646         struct sd_data *sdd = sd->private;
5647         struct sched_domain *sibling;
5648         int i;
5649 
5650         for_each_cpu(i, span) {
5651                 sibling = *per_cpu_ptr(sdd->sd, i);
5652                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5653                         continue;
5654 
5655                 cpumask_set_cpu(i, sched_group_mask(sg));
5656         }
5657 }
5658 
5659 /*
5660  * Return the canonical balance cpu for this group, this is the first cpu
5661  * of this group that's also in the iteration mask.
5662  */
5663 int group_balance_cpu(struct sched_group *sg)
5664 {
5665         return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5666 }
5667 
5668 static int
5669 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5670 {
5671         struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5672         const struct cpumask *span = sched_domain_span(sd);
5673         struct cpumask *covered = sched_domains_tmpmask;
5674         struct sd_data *sdd = sd->private;
5675         struct sched_domain *child;
5676         int i;
5677 
5678         cpumask_clear(covered);
5679 
5680         for_each_cpu(i, span) {
5681                 struct cpumask *sg_span;
5682 
5683                 if (cpumask_test_cpu(i, covered))
5684                         continue;
5685 
5686                 child = *per_cpu_ptr(sdd->sd, i);
5687 
5688                 /* See the comment near build_group_mask(). */
5689                 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5690                         continue;
5691 
5692                 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5693                                 GFP_KERNEL, cpu_to_node(cpu));
5694 
5695                 if (!sg)
5696                         goto fail;
5697 
5698                 sg_span = sched_group_cpus(sg);
5699                 if (child->child) {
5700                         child = child->child;
5701                         cpumask_copy(sg_span, sched_domain_span(child));
5702                 } else
5703                         cpumask_set_cpu(i, sg_span);
5704 
5705                 cpumask_or(covered, covered, sg_span);
5706 
5707                 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5708                 if (atomic_inc_return(&sg->sgp->ref) == 1)
5709                         build_group_mask(sd, sg);
5710 
5711                 /*
5712                  * Initialize sgp->power such that even if we mess up the
5713                  * domains and no possible iteration will get us here, we won't
5714                  * die on a /0 trap.
5715                  */
5716                 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5717                 sg->sgp->power_orig = sg->sgp->power;
5718 
5719                 /*
5720                  * Make sure the first group of this domain contains the
5721                  * canonical balance cpu. Otherwise the sched_domain iteration
5722                  * breaks. See update_sg_lb_stats().
5723                  */
5724                 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5725                     group_balance_cpu(sg) == cpu)
5726                         groups = sg;
5727 
5728                 if (!first)
5729                         first = sg;
5730                 if (last)
5731                         last->next = sg;
5732                 last = sg;
5733                 last->next = first;
5734         }
5735         sd->groups = groups;
5736 
5737         return 0;
5738 
5739 fail:
5740         free_sched_groups(first, 0);
5741 
5742         return -ENOMEM;
5743 }
5744 
5745 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5746 {
5747         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5748         struct sched_domain *child = sd->child;
5749 
5750         if (child)
5751                 cpu = cpumask_first(sched_domain_span(child));
5752 
5753         if (sg) {
5754                 *sg = *per_cpu_ptr(sdd->sg, cpu);
5755                 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5756                 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5757         }
5758 
5759         return cpu;
5760 }
5761 
5762 /*
5763  * build_sched_groups will build a circular linked list of the groups
5764  * covered by the given span, and will set each group's ->cpumask correctly,
5765  * and ->cpu_power to 0.
5766  *
5767  * Assumes the sched_domain tree is fully constructed
5768  */
5769 static int
5770 build_sched_groups(struct sched_domain *sd, int cpu)
5771 {
5772         struct sched_group *first = NULL, *last = NULL;
5773         struct sd_data *sdd = sd->private;
5774         const struct cpumask *span = sched_domain_span(sd);
5775         struct cpumask *covered;
5776         int i;
5777 
5778         get_group(cpu, sdd, &sd->groups);
5779         atomic_inc(&sd->groups->ref);
5780 
5781         if (cpu != cpumask_first(span))
5782                 return 0;
5783 
5784         lockdep_assert_held(&sched_domains_mutex);
5785         covered = sched_domains_tmpmask;
5786 
5787         cpumask_clear(covered);
5788 
5789         for_each_cpu(i, span) {
5790                 struct sched_group *sg;
5791                 int group, j;
5792 
5793                 if (cpumask_test_cpu(i, covered))
5794                         continue;
5795 
5796                 group = get_group(i, sdd, &sg);
5797                 cpumask_clear(sched_group_cpus(sg));
5798                 sg->sgp->power = 0;
5799                 cpumask_setall(sched_group_mask(sg));
5800 
5801                 for_each_cpu(j, span) {
5802                         if (get_group(j, sdd, NULL) != group)
5803                                 continue;
5804 
5805                         cpumask_set_cpu(j, covered);
5806                         cpumask_set_cpu(j, sched_group_cpus(sg));
5807                 }
5808 
5809                 if (!first)
5810                         first = sg;
5811                 if (last)
5812                         last->next = sg;
5813                 last = sg;
5814         }
5815         last->next = first;
5816 
5817         return 0;
5818 }
5819 
5820 /*
5821  * Initialize sched groups cpu_power.
5822  *
5823  * cpu_power indicates the capacity of sched group, which is used while
5824  * distributing the load between different sched groups in a sched domain.
5825  * Typically cpu_power for all the groups in a sched domain will be same unless
5826  * there are asymmetries in the topology. If there are asymmetries, group
5827  * having more cpu_power will pickup more load compared to the group having
5828  * less cpu_power.
5829  */
5830 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5831 {
5832         struct sched_group *sg = sd->groups;
5833 
5834         WARN_ON(!sg);
5835 
5836         do {
5837                 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5838                 sg = sg->next;
5839         } while (sg != sd->groups);
5840 
5841         if (cpu != group_balance_cpu(sg))
5842                 return;
5843 
5844         update_group_power(sd, cpu);
5845         atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5846 }
5847 
5848 int __weak arch_sd_sibling_asym_packing(void)
5849 {
5850        return 0*SD_ASYM_PACKING;
5851 }
5852 
5853 /*
5854  * Initializers for schedule domains
5855  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5856  */
5857 
5858 #ifdef CONFIG_SCHED_DEBUG
5859 # define SD_INIT_NAME(sd, type)         sd->name = #type
5860 #else
5861 # define SD_INIT_NAME(sd, type)         do { } while (0)
5862 #endif
5863 
5864 #define SD_INIT_FUNC(type)                                              \
5865 static noinline struct sched_domain *                                   \
5866 sd_init_##type(struct sched_domain_topology_level *tl, int cpu)         \
5867 {                                                                       \
5868         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);       \
5869         *sd = SD_##type##_INIT;                                         \
5870         SD_INIT_NAME(sd, type);                                         \
5871         sd->private = &tl->data;                                        \
5872         return sd;                                                      \
5873 }
5874 
5875 SD_INIT_FUNC(CPU)
5876 #ifdef CONFIG_SCHED_SMT
5877  SD_INIT_FUNC(SIBLING)
5878 #endif
5879 #ifdef CONFIG_SCHED_MC
5880  SD_INIT_FUNC(MC)
5881 #endif
5882 #ifdef CONFIG_SCHED_BOOK
5883  SD_INIT_FUNC(BOOK)
5884 #endif
5885 
5886 static int default_relax_domain_level = -1;
5887 int sched_domain_level_max;
5888 
5889 static int __init setup_relax_domain_level(char *str)
5890 {
5891         if (kstrtoint(str, 0, &default_relax_domain_level))
5892                 pr_warn("Unable to set relax_domain_level\n");
5893 
5894         return 1;
5895 }
5896 __setup("relax_domain_level=", setup_relax_domain_level);
5897 
5898 static void set_domain_attribute(struct sched_domain *sd,
5899                                  struct sched_domain_attr *attr)
5900 {
5901         int request;
5902 
5903         if (!attr || attr->relax_domain_level < 0) {
5904                 if (default_relax_domain_level < 0)
5905                         return;
5906                 else
5907                         request = default_relax_domain_level;
5908         } else
5909                 request = attr->relax_domain_level;
5910         if (request < sd->level) {
5911                 /* turn off idle balance on this domain */
5912                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5913         } else {
5914                 /* turn on idle balance on this domain */
5915                 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5916         }
5917 }
5918 
5919 static void __sdt_free(const struct cpumask *cpu_map);
5920 static int __sdt_alloc(const struct cpumask *cpu_map);
5921 
5922 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5923                                  const struct cpumask *cpu_map)
5924 {
5925         switch (what) {
5926         case sa_rootdomain:
5927                 if (!atomic_read(&d->rd->refcount))
5928                         free_rootdomain(&d->rd->rcu); /* fall through */
5929         case sa_sd:
5930                 free_percpu(d->sd); /* fall through */
5931         case sa_sd_storage:
5932                 __sdt_free(cpu_map); /* fall through */
5933         case sa_none:
5934                 break;
5935         }
5936 }
5937 
5938 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5939                                                    const struct cpumask *cpu_map)
5940 {
5941         memset(d, 0, sizeof(*d));
5942 
5943         if (__sdt_alloc(cpu_map))
5944                 return sa_sd_storage;
5945         d->sd = alloc_percpu(struct sched_domain *);
5946         if (!d->sd)
5947                 return sa_sd_storage;
5948         d->rd = alloc_rootdomain();
5949         if (!d->rd)
5950                 return sa_sd;
5951         return sa_rootdomain;
5952 }
5953 
5954 /*
5955  * NULL the sd_data elements we've used to build the sched_domain and
5956  * sched_group structure so that the subsequent __free_domain_allocs()
5957  * will not free the data we're using.
5958  */
5959 static void claim_allocations(int cpu, struct sched_domain *sd)
5960 {
5961         struct sd_data *sdd = sd->private;
5962 
5963         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5964         *per_cpu_ptr(sdd->sd, cpu) = NULL;
5965 
5966         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5967                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5968 
5969         if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5970                 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5971 }
5972 
5973 #ifdef CONFIG_SCHED_SMT
5974 static const struct cpumask *cpu_smt_mask(int cpu)
5975 {
5976         return topology_thread_cpumask(cpu);
5977 }
5978 #endif
5979 
5980 /*
5981  * Topology list, bottom-up.
5982  */
5983 static struct sched_domain_topology_level default_topology[] = {
5984 #ifdef CONFIG_SCHED_SMT
5985         { sd_init_SIBLING, cpu_smt_mask, },
5986 #endif
5987 #ifdef CONFIG_SCHED_MC
5988         { sd_init_MC, cpu_coregroup_mask, },
5989 #endif
5990 #ifdef CONFIG_SCHED_BOOK
5991         { sd_init_BOOK, cpu_book_mask, },
5992 #endif
5993         { sd_init_CPU, cpu_cpu_mask, },
5994         { NULL, },
5995 };
5996 
5997 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5998 
5999 #define for_each_sd_topology(tl)                        \
6000         for (tl = sched_domain_topology; tl->init; tl++)
6001 
6002 #ifdef CONFIG_NUMA
6003 
6004 static int sched_domains_numa_levels;
6005 static int *sched_domains_numa_distance;
6006 static struct cpumask ***sched_domains_numa_masks;
6007 static int sched_domains_curr_level;
6008 
6009 static inline int sd_local_flags(int level)
6010 {
6011         if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6012                 return 0;
6013 
6014         return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6015 }
6016 
6017 static struct sched_domain *
6018 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6019 {
6020         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6021         int level = tl->numa_level;
6022         int sd_weight = cpumask_weight(
6023                         sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6024 
6025         *sd = (struct sched_domain){
6026                 .min_interval           = sd_weight,
6027                 .max_interval           = 2*sd_weight,
6028                 .busy_factor            = 32,
6029                 .imbalance_pct          = 125,
6030                 .cache_nice_tries       = 2,
6031                 .busy_idx               = 3,
6032                 .idle_idx               = 2,
6033                 .newidle_idx            = 0,
6034                 .wake_idx               = 0,
6035                 .forkexec_idx           = 0,
6036 
6037                 .flags                  = 1*SD_LOAD_BALANCE
6038                                         | 1*SD_BALANCE_NEWIDLE
6039                                         | 0*SD_BALANCE_EXEC
6040                                         | 0*SD_BALANCE_FORK
6041                                         | 0*SD_BALANCE_WAKE
6042                                         | 0*SD_WAKE_AFFINE
6043                                         | 0*SD_SHARE_CPUPOWER
6044                                         | 0*SD_SHARE_PKG_RESOURCES
6045                                         | 1*SD_SERIALIZE
6046                                         | 0*SD_PREFER_SIBLING
6047                                         | 1*SD_NUMA
6048                                         | sd_local_flags(level)
6049                                         ,
6050                 .last_balance           = jiffies,
6051                 .balance_interval       = sd_weight,
6052                 .max_newidle_lb_cost    = 0,
6053                 .next_decay_max_lb_cost = jiffies,
6054         };
6055         SD_INIT_NAME(sd, NUMA);
6056         sd->private = &tl->data;
6057 
6058         /*
6059          * Ugly hack to pass state to sd_numa_mask()...
6060          */
6061         sched_domains_curr_level = tl->numa_level;
6062 
6063         return sd;
6064 }
6065 
6066 static const struct cpumask *sd_numa_mask(int cpu)
6067 {
6068         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6069 }
6070 
6071 static void sched_numa_warn(const char *str)
6072 {
6073         static int done = false;
6074         int i,j;
6075 
6076         if (done)
6077                 return;
6078 
6079         done = true;
6080 
6081         printk(KERN_WARNING "ERROR: %s\n\n", str);
6082 
6083         for (i = 0; i < nr_node_ids; i++) {
6084                 printk(KERN_WARNING "  ");
6085                 for (j = 0; j < nr_node_ids; j++)
6086                         printk(KERN_CONT "%02d ", node_distance(i,j));
6087                 printk(KERN_CONT "\n");
6088         }
6089         printk(KERN_WARNING "\n");
6090 }
6091 
6092 static bool find_numa_distance(int distance)
6093 {
6094         int i;
6095 
6096         if (distance == node_distance(0, 0))
6097                 return true;
6098 
6099         for (i = 0; i < sched_domains_numa_levels; i++) {
6100                 if (sched_domains_numa_distance[i] == distance)
6101                         return true;
6102         }
6103 
6104         return false;
6105 }
6106 
6107 static void sched_init_numa(void)
6108 {
6109         int next_distance, curr_distance = node_distance(0, 0);
6110         struct sched_domain_topology_level *tl;
6111         int level = 0;
6112         int i, j, k;
6113 
6114         sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6115         if (!sched_domains_numa_distance)
6116                 return;
6117 
6118         /*
6119          * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6120          * unique distances in the node_distance() table.
6121          *
6122          * Assumes node_distance(0,j) includes all distances in
6123          * node_distance(i,j) in order to avoid cubic time.
6124          */
6125         next_distance = curr_distance;
6126         for (i = 0; i < nr_node_ids; i++) {
6127                 for (j = 0; j < nr_node_ids; j++) {
6128                         for (k = 0; k < nr_node_ids; k++) {
6129                                 int distance = node_distance(i, k);
6130 
6131                                 if (distance > curr_distance &&
6132                                     (distance < next_distance ||
6133                                      next_distance == curr_distance))
6134                                         next_distance = distance;
6135 
6136                                 /*
6137                                  * While not a strong assumption it would be nice to know
6138                                  * about cases where if node A is connected to B, B is not
6139                                  * equally connected to A.
6140                                  */
6141                                 if (sched_debug() && node_distance(k, i) != distance)
6142                                         sched_numa_warn("Node-distance not symmetric");
6143 
6144                                 if (sched_debug() && i && !find_numa_distance(distance))
6145                                         sched_numa_warn("Node-0 not representative");
6146                         }
6147                         if (next_distance != curr_distance) {
6148                                 sched_domains_numa_distance[level++] = next_distance;
6149                                 sched_domains_numa_levels = level;
6150                                 curr_distance = next_distance;
6151                         } else break;
6152                 }
6153 
6154                 /*
6155                  * In case of sched_debug() we verify the above assumption.
6156                  */
6157                 if (!sched_debug())
6158                         break;
6159         }
6160         /*
6161          * 'level' contains the number of unique distances, excluding the
6162          * identity distance node_distance(i,i).
6163          *
6164          * The sched_domains_numa_distance[] array includes the actual distance
6165          * numbers.
6166          */
6167 
6168         /*
6169          * Here, we should temporarily reset sched_domains_numa_levels to 0.
6170          * If it fails to allocate memory for array sched_domains_numa_masks[][],
6171          * the array will contain less then 'level' members. This could be
6172          * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6173          * in other functions.
6174          *
6175          * We reset it to 'level' at the end of this function.
6176          */
6177         sched_domains_numa_levels = 0;
6178 
6179         sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6180         if (!sched_domains_numa_masks)
6181                 return;
6182 
6183         /*
6184          * Now for each level, construct a mask per node which contains all
6185          * cpus of nodes that are that many hops away from us.
6186          */
6187         for (i = 0; i < level; i++) {
6188                 sched_domains_numa_masks[i] =
6189                         kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6190                 if (!sched_domains_numa_masks[i])
6191                         return;
6192 
6193                 for (j = 0; j < nr_node_ids; j++) {
6194                         struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6195                         if (!mask)
6196                                 return;
6197 
6198                         sched_domains_numa_masks[i][j] = mask;
6199 
6200                         for (k = 0; k < nr_node_ids; k++) {
6201                                 if (node_distance(j, k) > sched_domains_numa_distance[i])
6202                                         continue;
6203 
6204                                 cpumask_or(mask, mask, cpumask_of_node(k));
6205                         }
6206                 }
6207         }
6208 
6209         tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6210                         sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6211         if (!tl)
6212                 return;
6213 
6214         /*
6215          * Copy the default topology bits..
6216          */
6217         for (i = 0; default_topology[i].init; i++)
6218                 tl[i] = default_topology[i];
6219 
6220         /*
6221          * .. and append 'j' levels of NUMA goodness.
6222          */
6223         for (j = 0; j < level; i++, j++) {
6224                 tl[i] = (struct sched_domain_topology_level){
6225                         .init = sd_numa_init,
6226                         .mask = sd_numa_mask,
6227                         .flags = SDTL_OVERLAP,
6228                         .numa_level = j,
6229                 };
6230         }
6231 
6232         sched_domain_topology = tl;
6233 
6234         sched_domains_numa_levels = level;
6235 }
6236 
6237 static void sched_domains_numa_masks_set(int cpu)
6238 {
6239         int i, j;
6240         int node = cpu_to_node(cpu);
6241 
6242         for (i = 0; i < sched_domains_numa_levels; i++) {
6243                 for (j = 0; j < nr_node_ids; j++) {
6244                         if (node_distance(j, node) <= sched_domains_numa_distance[i])
6245                                 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6246                 }
6247         }
6248 }
6249 
6250 static void sched_domains_numa_masks_clear(int cpu)
6251 {
6252         int i, j;
6253         for (i = 0; i < sched_domains_numa_levels; i++) {
6254                 for (j = 0; j < nr_node_ids; j++)
6255                         cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6256         }
6257 }
6258 
6259 /*
6260  * Update sched_domains_numa_masks[level][node] array when new cpus
6261  * are onlined.
6262  */
6263 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6264                                            unsigned long action,
6265                                            void *hcpu)
6266 {
6267         int cpu = (long)hcpu;
6268 
6269         switch (action & ~CPU_TASKS_FROZEN) {
6270         case CPU_ONLINE:
6271                 sched_domains_numa_masks_set(cpu);
6272                 break;
6273 
6274         case CPU_DEAD:
6275                 sched_domains_numa_masks_clear(cpu);
6276                 break;
6277 
6278         default:
6279                 return NOTIFY_DONE;
6280         }
6281 
6282         return NOTIFY_OK;
6283 }
6284 #else
6285 static inline void sched_init_numa(void)
6286 {
6287 }
6288 
6289 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6290                                            unsigned long action,
6291                                            void *hcpu)
6292 {
6293         return 0;
6294 }
6295 #endif /* CONFIG_NUMA */
6296 
6297 static int __sdt_alloc(const struct cpumask *cpu_map)
6298 {
6299         struct sched_domain_topology_level *tl;
6300         int j;
6301 
6302         for_each_sd_topology(tl) {
6303                 struct sd_data *sdd = &tl->data;
6304 
6305                 sdd->sd = alloc_percpu(struct sched_domain *);
6306                 if (!sdd->sd)
6307                         return -ENOMEM;
6308 
6309                 sdd->sg = alloc_percpu(struct sched_group *);
6310                 if (!sdd->sg)
6311                         return -ENOMEM;
6312 
6313                 sdd->sgp = alloc_percpu(struct sched_group_power *);
6314                 if (!sdd->sgp)
6315                         return -ENOMEM;
6316 
6317                 for_each_cpu(j, cpu_map) {
6318                         struct sched_domain *sd;
6319                         struct sched_group *sg;
6320                         struct sched_group_power *sgp;
6321 
6322                         sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6323                                         GFP_KERNEL, cpu_to_node(j));
6324                         if (!sd)
6325                                 return -ENOMEM;
6326 
6327                         *per_cpu_ptr(sdd->sd, j) = sd;
6328 
6329                         sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6330                                         GFP_KERNEL, cpu_to_node(j));
6331                         if (!sg)
6332                                 return -ENOMEM;
6333 
6334                         sg->next = sg;
6335 
6336                         *per_cpu_ptr(sdd->sg, j) = sg;
6337 
6338                         sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6339                                         GFP_KERNEL, cpu_to_node(j));
6340                         if (!sgp)
6341                                 return -ENOMEM;
6342 
6343                         *per_cpu_ptr(sdd->sgp, j) = sgp;
6344                 }
6345         }
6346 
6347         return 0;
6348 }
6349 
6350 static void __sdt_free(const struct cpumask *cpu_map)
6351 {
6352         struct sched_domain_topology_level *tl;
6353         int j;
6354 
6355         for_each_sd_topology(tl) {
6356                 struct sd_data *sdd = &tl->data;
6357 
6358                 for_each_cpu(j, cpu_map) {
6359                         struct sched_domain *sd;
6360 
6361                         if (sdd->sd) {
6362                                 sd = *per_cpu_ptr(sdd->sd, j);
6363                                 if (sd && (sd->flags & SD_OVERLAP))
6364                                         free_sched_groups(sd->groups, 0);
6365                                 kfree(*per_cpu_ptr(sdd->sd, j));
6366                         }
6367 
6368                         if (sdd->sg)
6369                                 kfree(*per_cpu_ptr(sdd->sg, j));
6370                         if (sdd->sgp)
6371                                 kfree(*per_cpu_ptr(sdd->sgp, j));
6372                 }
6373                 free_percpu(sdd->sd);
6374                 sdd->sd = NULL;
6375                 free_percpu(sdd->sg);
6376                 sdd->sg = NULL;
6377                 free_percpu(sdd->sgp);
6378                 sdd->sgp = NULL;
6379         }
6380 }
6381 
6382 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6383                 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6384                 struct sched_domain *child, int cpu)
6385 {
6386         struct sched_domain *sd = tl->init(tl, cpu);
6387         if (!sd)
6388                 return child;
6389 
6390         cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6391         if (child) {
6392                 sd->level = child->level + 1;
6393                 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6394                 child->parent = sd;
6395                 sd->child = child;
6396         }
6397         set_domain_attribute(sd, attr);
6398 
6399         return sd;
6400 }
6401 
6402 /*
6403  * Build sched domains for a given set of cpus and attach the sched domains
6404  * to the individual cpus
6405  */
6406 static int build_sched_domains(const struct cpumask *cpu_map,
6407                                struct sched_domain_attr *attr)
6408 {
6409         enum s_alloc alloc_state;
6410         struct sched_domain *sd;
6411         struct s_data d;
6412         int i, ret = -ENOMEM;
6413 
6414         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6415         if (alloc_state != sa_rootdomain)
6416                 goto error;
6417 
6418         /* Set up domains for cpus specified by the cpu_map. */
6419         for_each_cpu(i, cpu_map) {
6420                 struct sched_domain_topology_level *tl;
6421 
6422                 sd = NULL;
6423                 for_each_sd_topology(tl) {
6424                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6425                         if (tl == sched_domain_topology)
6426                                 *per_cpu_ptr(d.sd, i) = sd;
6427                         if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6428                                 sd->flags |= SD_OVERLAP;
6429                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6430                                 break;
6431                 }
6432         }
6433 
6434         /* Build the groups for the domains */
6435         for_each_cpu(i, cpu_map) {
6436                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6437                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
6438                         if (sd->flags & SD_OVERLAP) {
6439                                 if (build_overlap_sched_groups(sd, i))
6440                                         goto error;
6441                         } else {
6442                                 if (build_sched_groups(sd, i))
6443                                         goto error;
6444                         }
6445                 }
6446         }
6447 
6448         /* Calculate CPU power for physical packages and nodes */
6449         for (i = nr_cpumask_bits-1; i >= 0; i--) {
6450                 if (!cpumask_test_cpu(i, cpu_map))
6451                         continue;
6452 
6453                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6454                         claim_allocations(i, sd);
6455                         init_sched_groups_power(i, sd);
6456                 }
6457         }
6458 
6459         /* Attach the domains */
6460         rcu_read_lock();
6461         for_each_cpu(i, cpu_map) {
6462                 sd = *per_cpu_ptr(d.sd, i);
6463                 cpu_attach_domain(sd, d.rd, i);
6464         }
6465         rcu_read_unlock();
6466 
6467         ret = 0;
6468 error:
6469         __free_domain_allocs(&d, alloc_state, cpu_map);
6470         return ret;
6471 }
6472 
6473 static cpumask_var_t *doms_cur; /* current sched domains */
6474 static int ndoms_cur;           /* number of sched domains in 'doms_cur' */
6475 static struct sched_domain_attr *dattr_cur;
6476                                 /* attribues of custom domains in 'doms_cur' */
6477 
6478 /*
6479  * Special case: If a kmalloc of a doms_cur partition (array of
6480  * cpumask) fails, then fallback to a single sched domain,
6481  * as determined by the single cpumask fallback_doms.
6482  */
6483 static cpumask_var_t fallback_doms;
6484 
6485 /*
6486  * arch_update_cpu_topology lets virtualized architectures update the
6487  * cpu core maps. It is supposed to return 1 if the topology changed
6488  * or 0 if it stayed the same.
6489  */
6490 int __weak arch_update_cpu_topology(void)
6491 {
6492         return 0;
6493 }
6494 
6495 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6496 {
6497         int i;
6498         cpumask_var_t *doms;
6499 
6500         doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6501         if (!doms)
6502                 return NULL;
6503         for (i = 0; i < ndoms; i++) {
6504                 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6505                         free_sched_domains(doms, i);
6506                         return NULL;
6507                 }
6508         }
6509         return doms;
6510 }
6511 
6512 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6513 {
6514         unsigned int i;
6515         for (i = 0; i < ndoms; i++)
6516                 free_cpumask_var(doms[i]);
6517         kfree(doms);
6518 }
6519 
6520 /*
6521  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6522  * For now this just excludes isolated cpus, but could be used to
6523  * exclude other special cases in the future.
6524  */
6525 static int init_sched_domains(const struct cpumask *cpu_map)
6526 {
6527         int err;
6528 
6529         arch_update_cpu_topology();
6530         ndoms_cur = 1;
6531         doms_cur = alloc_sched_domains(ndoms_cur);
6532         if (!doms_cur)
6533                 doms_cur = &fallback_doms;
6534         cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6535         err = build_sched_domains(doms_cur[0], NULL);
6536         register_sched_domain_sysctl();
6537 
6538         return err;
6539 }
6540 
6541 /*
6542  * Detach sched domains from a group of cpus specified in cpu_map
6543  * These cpus will now be attached to the NULL domain
6544  */
6545 static void detach_destroy_domains(const struct cpumask *cpu_map)
6546 {
6547         int i;
6548 
6549         rcu_read_lock();
6550         for_each_cpu(i, cpu_map)
6551                 cpu_attach_domain(NULL, &def_root_domain, i);
6552         rcu_read_unlock();
6553 }
6554 
6555 /* handle null as "default" */
6556 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6557                         struct sched_domain_attr *new, int idx_new)
6558 {
6559         struct sched_domain_attr tmp;
6560 
6561         /* fast path */
6562         if (!new && !cur)
6563                 return 1;
6564 
6565         tmp = SD_ATTR_INIT;
6566         return !memcmp(cur ? (cur + idx_cur) : &tmp,
6567                         new ? (new + idx_new) : &tmp,
6568                         sizeof(struct sched_domain_attr));
6569 }
6570 
6571 /*
6572  * Partition sched domains as specified by the 'ndoms_new'
6573  * cpumasks in the array doms_new[] of cpumasks. This compares
6574  * doms_new[] to the current sched domain partitioning, doms_cur[].
6575  * It destroys each deleted domain and builds each new domain.
6576  *
6577  * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6578  * The masks don't intersect (don't overlap.) We should setup one
6579  * sched domain for each mask. CPUs not in any of the cpumasks will
6580  * not be load balanced. If the same cpumask appears both in the
6581  * current 'doms_cur' domains and in the new 'doms_new', we can leave
6582  * it as it is.
6583  *
6584  * The passed in 'doms_new' should be allocated using
6585  * alloc_sched_domains.  This routine takes ownership of it and will
6586  * free_sched_domains it when done with it. If the caller failed the
6587  * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6588  * and partition_sched_domains() will fallback to the single partition
6589  * 'fallback_doms', it also forces the domains to be rebuilt.
6590  *
6591  * If doms_new == NULL it will be replaced with cpu_online_mask.
6592  * ndoms_new == 0 is a special case for destroying existing domains,
6593  * and it will not create the default domain.
6594  *
6595  * Call with hotplug lock held
6596  */
6597 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6598                              struct sched_domain_attr *dattr_new)
6599 {
6600         int i, j, n;
6601         int new_topology;
6602 
6603         mutex_lock(&sched_domains_mutex);
6604 
6605         /* always unregister in case we don't destroy any domains */
6606         unregister_sched_domain_sysctl();
6607 
6608         /* Let architecture update cpu core mappings. */
6609         new_topology = arch_update_cpu_topology();
6610 
6611         n = doms_new ? ndoms_new : 0;
6612 
6613         /* Destroy deleted domains */
6614         for (i = 0; i < ndoms_cur; i++) {
6615                 for (j = 0; j < n && !new_topology; j++) {
6616                         if (cpumask_equal(doms_cur[i], doms_new[j])
6617                             && dattrs_equal(dattr_cur, i, dattr_new, j))
6618                                 goto match1;
6619                 }
6620                 /* no match - a current sched domain not in new doms_new[] */
6621                 detach_destroy_domains(doms_cur[i]);
6622 match1:
6623                 ;
6624         }
6625 
6626         n = ndoms_cur;
6627         if (doms_new == NULL) {
6628                 n = 0;
6629                 doms_new = &fallback_doms;
6630                 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6631                 WARN_ON_ONCE(dattr_new);
6632         }
6633 
6634         /* Build new domains */
6635         for (i = 0; i < ndoms_new; i++) {
6636                 for (j = 0; j < n && !new_topology; j++) {
6637                         if (cpumask_equal(doms_new[i], doms_cur[j])
6638                             && dattrs_equal(dattr_new, i, dattr_cur, j))
6639                                 goto match2;
6640                 }
6641                 /* no match - add a new doms_new */
6642                 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6643 match2:
6644                 ;
6645         }
6646 
6647         /* Remember the new sched domains */
6648         if (doms_cur != &fallback_doms)
6649                 free_sched_domains(doms_cur, ndoms_cur);
6650         kfree(dattr_cur);       /* kfree(NULL) is safe */
6651         doms_cur = doms_new;
6652         dattr_cur = dattr_new;
6653         ndoms_cur = ndoms_new;
6654 
6655         register_sched_domain_sysctl();
6656 
6657         mutex_unlock(&sched_domains_mutex);
6658 }
6659 
6660 static int num_cpus_frozen;     /* used to mark begin/end of suspend/resume */
6661 
6662 /*
6663  * Update cpusets according to cpu_active mask.  If cpusets are
6664  * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6665  * around partition_sched_domains().
6666  *
6667  * If we come here as part of a suspend/resume, don't touch cpusets because we
6668  * want to restore it back to its original state upon resume anyway.
6669  */
6670 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6671                              void *hcpu)
6672 {
6673         switch (action) {
6674         case CPU_ONLINE_FROZEN:
6675         case CPU_DOWN_FAILED_FROZEN:
6676 
6677                 /*
6678                  * num_cpus_frozen tracks how many CPUs are involved in suspend
6679                  * resume sequence. As long as this is not the last online
6680                  * operation in the resume sequence, just build a single sched
6681                  * domain, ignoring cpusets.
6682                  */
6683                 num_cpus_frozen--;
6684                 if (likely(num_cpus_frozen)) {
6685                         partition_sched_domains(1, NULL, NULL);
6686                         break;
6687                 }
6688 
6689                 /*
6690                  * This is the last CPU online operation. So fall through and
6691                  * restore the original sched domains by considering the
6692                  * cpuset configurations.
6693                  */
6694 
6695         case CPU_ONLINE:
6696         case CPU_DOWN_FAILED:
6697                 cpuset_update_active_cpus(true);
6698                 break;
6699         default:
6700                 return NOTIFY_DONE;
6701         }
6702         return NOTIFY_OK;
6703 }
6704 
6705 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6706                                void *hcpu)
6707 {
6708         switch (action) {
6709         case CPU_DOWN_PREPARE:
6710                 cpuset_update_active_cpus(false);
6711                 break;
6712         case CPU_DOWN_PREPARE_FROZEN:
6713                 num_cpus_frozen++;
6714                 partition_sched_domains(1, NULL, NULL);
6715                 break;
6716         default:
6717                 return NOTIFY_DONE;
6718         }
6719         return NOTIFY_OK;
6720 }
6721 
6722 void __init sched_init_smp(void)
6723 {
6724         cpumask_var_t non_isolated_cpus;
6725 
6726         alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6727         alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6728 
6729         sched_init_numa();
6730 
6731         /*
6732          * There's no userspace yet to cause hotplug operations; hence all the
6733          * cpu masks are stable and all blatant races in the below code cannot
6734          * happen.
6735          */
6736         mutex_lock(&sched_domains_mutex);
6737         init_sched_domains(cpu_active_mask);
6738         cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6739         if (cpumask_empty(non_isolated_cpus))
6740                 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6741         mutex_unlock(&sched_domains_mutex);
6742 
6743         hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6744         hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6745         hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6746 
6747         init_hrtick();
6748 
6749         /* Move init over to a non-isolated CPU */
6750         if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6751                 BUG();
6752         sched_init_granularity();
6753         free_cpumask_var(non_isolated_cpus);
6754 
6755         init_sched_rt_class();
6756         init_sched_dl_class();
6757 }
6758 #else
6759 void __init sched_init_smp(void)
6760 {
6761         sched_init_granularity();
6762 }
6763 #endif /* CONFIG_SMP */
6764 
6765 const_debug unsigned int sysctl_timer_migration = 1;
6766 
6767 int in_sched_functions(unsigned long addr)
6768 {
6769         return in_lock_functions(addr) ||
6770                 (addr >= (unsigned long)__sched_text_start
6771                 && addr < (unsigned long)__sched_text_end);
6772 }
6773 
6774 #ifdef CONFIG_CGROUP_SCHED
6775 /*
6776  * Default task group.
6777  * Every task in system belongs to this group at bootup.
6778  */
6779 struct task_group root_task_group;
6780 LIST_HEAD(task_groups);
6781 #endif
6782 
6783 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6784 
6785 void __init sched_init(void)
6786 {
6787         int i, j;
6788         unsigned long alloc_size = 0, ptr;
6789 
6790 #ifdef CONFIG_FAIR_GROUP_SCHED
6791         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6792 #endif
6793 #ifdef CONFIG_RT_GROUP_SCHED
6794         alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6795 #endif
6796 #ifdef CONFIG_CPUMASK_OFFSTACK
6797         alloc_size += num_possible_cpus() * cpumask_size();
6798 #endif
6799         if (alloc_size) {
6800                 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6801 
6802 #ifdef CONFIG_FAIR_GROUP_SCHED
6803                 root_task_group.se = (struct sched_entity **)ptr;
6804                 ptr += nr_cpu_ids * sizeof(void **);
6805 
6806                 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6807                 ptr += nr_cpu_ids * sizeof(void **);
6808 
6809 #endif /* CONFIG_FAIR_GROUP_SCHED */
6810 #ifdef CONFIG_RT_GROUP_SCHED
6811                 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6812                 ptr += nr_cpu_ids * sizeof(void **);
6813 
6814                 root_task_group.rt_rq = (struct rt_rq **)ptr;
6815                 ptr += nr_cpu_ids * sizeof(void **);
6816 
6817 #endif /* CONFIG_RT_GROUP_SCHED */
6818 #ifdef CONFIG_CPUMASK_OFFSTACK
6819                 for_each_possible_cpu(i) {
6820                         per_cpu(load_balance_mask, i) = (void *)ptr;
6821                         ptr += cpumask_size();
6822                 }
6823 #endif /* CONFIG_CPUMASK_OFFSTACK */
6824         }
6825 
6826         init_rt_bandwidth(&def_rt_bandwidth,
6827                         global_rt_period(), global_rt_runtime());
6828         init_dl_bandwidth(&def_dl_bandwidth,
6829                         global_rt_period(), global_rt_runtime());
6830 
6831 #ifdef CONFIG_SMP
6832         init_defrootdomain();
6833 #endif
6834 
6835 #ifdef CONFIG_RT_GROUP_SCHED
6836         init_rt_bandwidth(&root_task_group.rt_bandwidth,
6837                         global_rt_period(), global_rt_runtime());
6838 #endif /* CONFIG_RT_GROUP_SCHED */
6839 
6840 #ifdef CONFIG_CGROUP_SCHED
6841         list_add(&root_task_group.list, &task_groups);
6842         INIT_LIST_HEAD(&root_task_group.children);
6843         INIT_LIST_HEAD(&root_task_group.siblings);
6844         autogroup_init(&init_task);
6845 
6846 #endif /* CONFIG_CGROUP_SCHED */
6847 
6848         for_each_possible_cpu(i) {
6849                 struct rq *rq;
6850 
6851                 rq = cpu_rq(i);
6852                 raw_spin_lock_init(&rq->lock);
6853                 rq->nr_running = 0;
6854                 rq->calc_load_active = 0;
6855                 rq->calc_load_update = jiffies + LOAD_FREQ;
6856                 init_cfs_rq(&rq->cfs);
6857                 init_rt_rq(&rq->rt, rq);
6858                 init_dl_rq(&rq->dl, rq);
6859 #ifdef CONFIG_FAIR_GROUP_SCHED
6860                 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6861                 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6862                 /*
6863                  * How much cpu bandwidth does root_task_group get?
6864                  *
6865                  * In case of task-groups formed thr' the cgroup filesystem, it
6866                  * gets 100% of the cpu resources in the system. This overall
6867                  * system cpu resource is divided among the tasks of
6868                  * root_task_group and its child task-groups in a fair manner,
6869                  * based on each entity's (task or task-group's) weight
6870                  * (se->load.weight).
6871                  *
6872                  * In other words, if root_task_group has 10 tasks of weight
6873                  * 1024) and two child groups A0 and A1 (of weight 1024 each),
6874                  * then A0's share of the cpu resource is:
6875                  *
6876                  *      A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6877                  *
6878                  * We achieve this by letting root_task_group's tasks sit
6879                  * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6880                  */
6881                 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6882                 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6883 #endif /* CONFIG_FAIR_GROUP_SCHED */
6884 
6885                 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6886 #ifdef CONFIG_RT_GROUP_SCHED
6887                 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6888 #endif
6889 
6890                 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6891                         rq->cpu_load[j] = 0;
6892 
6893                 rq->last_load_update_tick = jiffies;
6894 
6895 #ifdef CONFIG_SMP
6896                 rq->sd = NULL;
6897                 rq->rd = NULL;
6898                 rq->cpu_power = SCHED_POWER_SCALE;
6899                 rq->post_schedule = 0;
6900                 rq->active_balance = 0;
6901                 rq->next_balance = jiffies;
6902                 rq->push_cpu = 0;
6903                 rq->cpu = i;
6904                 rq->online = 0;
6905                 rq->idle_stamp = 0;
6906                 rq->avg_idle = 2*sysctl_sched_migration_cost;
6907                 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6908 
6909                 INIT_LIST_HEAD(&rq->cfs_tasks);
6910 
6911                 rq_attach_root(rq, &def_root_domain);
6912 #ifdef CONFIG_NO_HZ_COMMON
6913                 rq->nohz_flags = 0;
6914 #endif
6915 #ifdef CONFIG_NO_HZ_FULL
6916                 rq->last_sched_tick = 0;
6917 #endif
6918 #endif
6919                 init_rq_hrtick(rq);
6920                 atomic_set(&rq->nr_iowait, 0);
6921         }
6922 
6923         set_load_weight(&init_task);
6924 
6925 #ifdef CONFIG_PREEMPT_NOTIFIERS
6926         INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6927 #endif
6928 
6929         /*
6930          * The boot idle thread does lazy MMU switching as well:
6931          */
6932         atomic_inc(&init_mm.mm_count);
6933         enter_lazy_tlb(&init_mm, current);
6934 
6935         /*
6936          * Make us the idle thread. Technically, schedule() should not be
6937          * called from this thread, however somewhere below it might be,
6938          * but because we are the idle thread, we just pick up running again
6939          * when this runqueue becomes "idle".
6940          */
6941         init_idle(current, smp_processor_id());
6942 
6943         calc_load_update = jiffies + LOAD_FREQ;
6944 
6945         /*
6946          * During early bootup we pretend to be a normal task:
6947          */
6948         current->sched_class = &fair_sched_class;
6949 
6950 #ifdef CONFIG_SMP
6951         zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6952         /* May be allocated at isolcpus cmdline parse time */
6953         if (cpu_isolated_map == NULL)
6954                 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6955         idle_thread_set_boot_cpu();
6956 #endif
6957         init_sched_fair_class();
6958 
6959         scheduler_running = 1;
6960 }
6961 
6962 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6963 static inline int preempt_count_equals(int preempt_offset)
6964 {
6965         int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6966 
6967         return (nested == preempt_offset);
6968 }
6969 
6970 void __might_sleep(const char *file, int line, int preempt_offset)
6971 {
6972         static unsigned long prev_jiffy;        /* ratelimiting */
6973 
6974         rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6975         if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6976              !is_idle_task(current)) ||
6977             system_state != SYSTEM_RUNNING || oops_in_progress)
6978                 return;
6979         if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6980                 return;
6981         prev_jiffy = jiffies;
6982 
6983         printk(KERN_ERR
6984                 "BUG: sleeping function called from invalid context at %s:%d\n",
6985                         file, line);
6986         printk(KERN_ERR
6987                 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6988                         in_atomic(), irqs_disabled(),
6989                         current->pid, current->comm);
6990 
6991         debug_show_held_locks(current);
6992         if (irqs_disabled())
6993                 print_irqtrace_events(current);
6994 #ifdef CONFIG_DEBUG_PREEMPT
6995         if (!preempt_count_equals(preempt_offset)) {
6996                 pr_err("Preemption disabled at:");
6997                 print_ip_sym(current->preempt_disable_ip);
6998                 pr_cont("\n");
6999         }
7000 #endif
7001         dump_stack();
7002 }
7003 EXPORT_SYMBOL(__might_sleep);
7004 #endif
7005 
7006 #ifdef CONFIG_MAGIC_SYSRQ
7007 static void normalize_task(struct rq *rq, struct task_struct *p)
7008 {
7009         const struct sched_class *prev_class = p->sched_class;
7010         struct sched_attr attr = {
7011                 .sched_policy = SCHED_NORMAL,
7012         };
7013         int old_prio = p->prio;
7014         int on_rq;
7015 
7016         on_rq = p->on_rq;
7017         if (on_rq)
7018                 dequeue_task(rq, p, 0);
7019         __setscheduler(rq, p, &attr);
7020         if (on_rq) {
7021                 enqueue_task(rq, p, 0);
7022                 resched_task(rq->curr);
7023         }
7024 
7025         check_class_changed(rq, p, prev_class, old_prio);
7026 }
7027 
7028 void normalize_rt_tasks(void)
7029 {
7030         struct task_struct *g, *p;
7031         unsigned long flags;
7032         struct rq *rq;
7033 
7034         read_lock_irqsave(&tasklist_lock, flags);
7035         do_each_thread(g, p) {
7036                 /*
7037                  * Only normalize user tasks:
7038                  */
7039                 if (!p->mm)
7040                         continue;
7041 
7042                 p->se.exec_start                = 0;
7043 #ifdef CONFIG_SCHEDSTATS
7044                 p->se.statistics.wait_start     = 0;
7045                 p->se.statistics.sleep_start    = 0;
7046                 p->se.statistics.block_start    = 0;
7047 #endif
7048 
7049                 if (!dl_task(p) && !rt_task(p)) {
7050                         /*
7051                          * Renice negative nice level userspace
7052                          * tasks back to 0:
7053                          */
7054                         if (task_nice(p) < 0 && p->mm)
7055                                 set_user_nice(p, 0);
7056                         continue;
7057                 }
7058 
7059                 raw_spin_lock(&p->pi_lock);
7060                 rq = __task_rq_lock(p);
7061 
7062                 normalize_task(rq, p);
7063 
7064                 __task_rq_unlock(rq);
7065                 raw_spin_unlock(&p->pi_lock);
7066         } while_each_thread(g, p);
7067 
7068         read_unlock_irqrestore(&tasklist_lock, flags);
7069 }
7070 
7071 #endif /* CONFIG_MAGIC_SYSRQ */
7072 
7073 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7074 /*
7075  * These functions are only useful for the IA64 MCA handling, or kdb.
7076  *
7077  * They can only be called when the whole system has been
7078  * stopped - every CPU needs to be quiescent, and no scheduling
7079  * activity can take place. Using them for anything else would
7080  * be a serious bug, and as a result, they aren't even visible
7081  * under any other configuration.
7082  */
7083 
7084 /**
7085  * curr_task - return the current task for a given cpu.
7086  * @cpu: the processor in question.
7087  *
7088  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7089  *
7090  * Return: The current task for @cpu.
7091  */
7092 struct task_struct *curr_task(int cpu)
7093 {
7094         return cpu_curr(cpu);
7095 }
7096 
7097 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7098 
7099 #ifdef CONFIG_IA64
7100 /**
7101  * set_curr_task - set the current task for a given cpu.
7102  * @cpu: the processor in question.
7103  * @p: the task pointer to set.
7104  *
7105  * Description: This function must only be used when non-maskable interrupts
7106  * are serviced on a separate stack. It allows the architecture to switch the
7107  * notion of the current task on a cpu in a non-blocking manner. This function
7108  * must be called with all CPU's synchronized, and interrupts disabled, the
7109  * and caller must save the original value of the current task (see
7110  * curr_task() above) and restore that value before reenabling interrupts and
7111  * re-starting the system.
7112  *
7113  * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7114  */
7115 void set_curr_task(int cpu, struct task_struct *p)
7116 {
7117         cpu_curr(cpu) = p;
7118 }
7119 
7120 #endif
7121 
7122 #ifdef CONFIG_CGROUP_SCHED
7123 /* task_group_lock serializes the addition/removal of task groups */
7124 static DEFINE_SPINLOCK(task_group_lock);
7125 
7126 static void free_sched_group(struct task_group *tg)
7127 {
7128         free_fair_sched_group(tg);
7129         free_rt_sched_group(tg);
7130         autogroup_free(tg);
7131         kfree(tg);
7132 }
7133 
7134 /* allocate runqueue etc for a new task group */
7135 struct task_group *sched_create_group(struct task_group *parent)
7136 {
7137         struct task_group *tg;
7138 
7139         tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7140         if (!tg)
7141                 return ERR_PTR(-ENOMEM);
7142 
7143         if (!alloc_fair_sched_group(tg, parent))
7144                 goto err;
7145 
7146         if (!alloc_rt_sched_group(tg, parent))
7147                 goto err;
7148 
7149         return tg;
7150 
7151 err:
7152         free_sched_group(tg);
7153         return ERR_PTR(-ENOMEM);
7154 }
7155 
7156 void sched_online_group(struct task_group *tg, struct task_group *parent)
7157 {
7158         unsigned long flags;
7159 
7160         spin_lock_irqsave(&task_group_lock, flags);
7161         list_add_rcu(&tg->list, &task_groups);
7162 
7163         WARN_ON(!parent); /* root should already exist */
7164 
7165         tg->parent = parent;
7166         INIT_LIST_HEAD(&tg->children);
7167         list_add_rcu(&tg->siblings, &parent->children);
7168         spin_unlock_irqrestore(&task_group_lock, flags);
7169 }
7170 
7171 /* rcu callback to free various structures associated with a task group */
7172 static void free_sched_group_rcu(struct rcu_head *rhp)
7173 {
7174         /* now it should be safe to free those cfs_rqs */
7175         free_sched_group(container_of(rhp, struct task_group, rcu));
7176 }
7177 
7178 /* Destroy runqueue etc associated with a task group */
7179 void sched_destroy_group(struct task_group *tg)
7180 {
7181         /* wait for possible concurrent references to cfs_rqs complete */
7182         call_rcu(&tg->rcu, free_sched_group_rcu);
7183 }
7184 
7185 void sched_offline_group(struct task_group *tg)
7186 {
7187         unsigned long flags;
7188         int i;
7189 
7190         /* end participation in shares distribution */
7191         for_each_possible_cpu(i)
7192                 unregister_fair_sched_group(tg, i);
7193 
7194         spin_lock_irqsave(&task_group_lock, flags);
7195         list_del_rcu(&tg->list);
7196         list_del_rcu(&tg->siblings);
7197         spin_unlock_irqrestore(&task_group_lock, flags);
7198 }
7199 
7200 /* change task's runqueue when it moves between groups.
7201  *      The caller of this function should have put the task in its new group
7202  *      by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7203  *      reflect its new group.
7204  */
7205 void sched_move_task(struct task_struct *tsk)
7206 {
7207         struct task_group *tg;
7208         int on_rq, running;
7209         unsigned long flags;
7210         struct rq *rq;
7211 
7212         rq = task_rq_lock(tsk, &flags);
7213 
7214         running = task_current(rq, tsk);
7215         on_rq = tsk->on_rq;
7216 
7217         if (on_rq)
7218                 dequeue_task(rq, tsk, 0);
7219         if (unlikely(running))
7220                 tsk->sched_class->put_prev_task(rq, tsk);
7221 
7222         tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7223                                 lockdep_is_held(&tsk->sighand->siglock)),
7224                           struct task_group, css);
7225         tg = autogroup_task_group(tsk, tg);
7226         tsk->sched_task_group = tg;
7227 
7228 #ifdef CONFIG_FAIR_GROUP_SCHED
7229         if (tsk->sched_class->task_move_group)
7230                 tsk->sched_class->task_move_group(tsk, on_rq);
7231         else
7232 #endif
7233                 set_task_rq(tsk, task_cpu(tsk));
7234 
7235         if (unlikely(running))
7236                 tsk->sched_class->set_curr_task(rq);
7237         if (on_rq)
7238                 enqueue_task(rq, tsk, 0);
7239 
7240         task_rq_unlock(rq, tsk, &flags);
7241 }
7242 #endif /* CONFIG_CGROUP_SCHED */
7243 
7244 #ifdef CONFIG_RT_GROUP_SCHED
7245 /*
7246  * Ensure that the real time constraints are schedulable.
7247  */
7248 static DEFINE_MUTEX(rt_constraints_mutex);
7249 
7250 /* Must be called with tasklist_lock held */
7251 static inline int tg_has_rt_tasks(struct task_group *tg)
7252 {
7253         struct task_struct *g, *p;
7254 
7255         do_each_thread(g, p) {
7256                 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7257                         return 1;
7258         } while_each_thread(g, p);
7259 
7260         return 0;
7261 }
7262 
7263 struct rt_schedulable_data {
7264         struct task_group *tg;
7265         u64 rt_period;
7266         u64 rt_runtime;
7267 };
7268 
7269 static int tg_rt_schedulable(struct task_group *tg, void *data)
7270 {
7271         struct rt_schedulable_data *d = data;
7272         struct task_group *child;
7273         unsigned long total, sum = 0;
7274         u64 period, runtime;
7275 
7276         period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7277         runtime = tg->rt_bandwidth.rt_runtime;
7278 
7279         if (tg == d->tg) {
7280                 period = d->rt_period;
7281                 runtime = d->rt_runtime;
7282         }
7283 
7284         /*
7285          * Cannot have more runtime than the period.
7286          */
7287         if (runtime > period && runtime != RUNTIME_INF)
7288                 return -EINVAL;
7289 
7290         /*
7291          * Ensure we don't starve existing RT tasks.
7292          */
7293         if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7294                 return -EBUSY;
7295 
7296         total = to_ratio(period, runtime);
7297 
7298         /*
7299          * Nobody can have more than the global setting allows.
7300          */
7301         if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7302                 return -EINVAL;
7303 
7304         /*
7305          * The sum of our children's runtime should not exceed our own.
7306          */
7307         list_for_each_entry_rcu(child, &tg->children, siblings) {
7308                 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7309                 runtime = child->rt_bandwidth.rt_runtime;
7310 
7311                 if (child == d->tg) {
7312                         period = d->rt_period;
7313                         runtime = d->rt_runtime;
7314                 }
7315 
7316                 sum += to_ratio(period, runtime);
7317         }
7318 
7319         if (sum > total)
7320                 return -EINVAL;
7321 
7322         return 0;
7323 }
7324 
7325 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7326 {
7327         int ret;
7328 
7329         struct rt_schedulable_data data = {
7330                 .tg = tg,
7331                 .rt_period = period,
7332                 .rt_runtime = runtime,
7333         };
7334 
7335         rcu_read_lock();
7336         ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7337         rcu_read_unlock();
7338 
7339         return ret;
7340 }
7341 
7342 static int tg_set_rt_bandwidth(struct task_group *tg,
7343                 u64 rt_period, u64 rt_runtime)
7344 {
7345         int i, err = 0;
7346 
7347         mutex_lock(&rt_constraints_mutex);
7348         read_lock(&tasklist_lock);
7349         err = __rt_schedulable(tg, rt_period, rt_runtime);
7350         if (err)
7351                 goto unlock;
7352 
7353         raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7354         tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7355         tg->rt_bandwidth.rt_runtime = rt_runtime;
7356 
7357         for_each_possible_cpu(i) {
7358                 struct rt_rq *rt_rq = tg->rt_rq[i];
7359 
7360                 raw_spin_lock(&rt_rq->rt_runtime_lock);
7361                 rt_rq->rt_runtime = rt_runtime;
7362                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7363         }
7364         raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7365 unlock:
7366         read_unlock(&tasklist_lock);
7367         mutex_unlock(&rt_constraints_mutex);
7368 
7369         return err;
7370 }
7371 
7372 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7373 {
7374         u64 rt_runtime, rt_period;
7375 
7376         rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7377         rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7378         if (rt_runtime_us < 0)
7379                 rt_runtime = RUNTIME_INF;
7380 
7381         return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7382 }
7383 
7384 static long sched_group_rt_runtime(struct task_group *tg)
7385 {
7386         u64 rt_runtime_us;
7387 
7388         if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7389                 return -1;
7390 
7391         rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7392         do_div(rt_runtime_us, NSEC_PER_USEC);
7393         return rt_runtime_us;
7394 }
7395 
7396 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7397 {
7398         u64 rt_runtime, rt_period;
7399 
7400         rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7401         rt_runtime = tg->rt_bandwidth.rt_runtime;
7402 
7403         if (rt_period == 0)
7404                 return -EINVAL;
7405 
7406         return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7407 }
7408 
7409 static long sched_group_rt_period(struct task_group *tg)
7410 {
7411         u64 rt_period_us;
7412 
7413         rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7414         do_div(rt_period_us, NSEC_PER_USEC);