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

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

Version: ~ [ linux-5.17-rc1 ] ~ [ linux-5.16.2 ] ~ [ linux-5.15.16 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.93 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.173 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.225 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.262 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.297 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.299 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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