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

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

Version: ~ [ linux-5.16-rc3 ] ~ [ linux-5.15.5 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.82 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.162 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.218 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.256 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.291 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.293 ] ~ [ 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 <asm/mmu_context.h>
 36 #include <linux/interrupt.h>
 37 #include <linux/capability.h>
 38 #include <linux/completion.h>
 39 #include <linux/kernel_stat.h>
 40 #include <linux/debug_locks.h>
 41 #include <linux/perf_event.h>
 42 #include <linux/security.h>
 43 #include <linux/notifier.h>
 44 #include <linux/profile.h>
 45 #include <linux/freezer.h>
 46 #include <linux/vmalloc.h>
 47 #include <linux/blkdev.h>
 48 #include <linux/delay.h>
 49 #include <linux/pid_namespace.h>
 50 #include <linux/smp.h>
 51 #include <linux/threads.h>
 52 #include <linux/timer.h>
 53 #include <linux/rcupdate.h>
 54 #include <linux/cpu.h>
 55 #include <linux/cpuset.h>
 56 #include <linux/percpu.h>
 57 #include <linux/proc_fs.h>
 58 #include <linux/seq_file.h>
 59 #include <linux/sysctl.h>
 60 #include <linux/syscalls.h>
 61 #include <linux/times.h>
 62 #include <linux/tsacct_kern.h>
 63 #include <linux/kprobes.h>
 64 #include <linux/delayacct.h>
 65 #include <linux/unistd.h>
 66 #include <linux/pagemap.h>
 67 #include <linux/hrtimer.h>
 68 #include <linux/tick.h>
 69 #include <linux/debugfs.h>
 70 #include <linux/ctype.h>
 71 #include <linux/ftrace.h>
 72 #include <linux/slab.h>
 73 #include <linux/init_task.h>
 74 #include <linux/binfmts.h>
 75 #include <linux/context_tracking.h>
 76 #include <linux/compiler.h>
 77 
 78 #include <asm/switch_to.h>
 79 #include <asm/tlb.h>
 80 #include <asm/irq_regs.h>
 81 #include <asm/mutex.h>
 82 #ifdef CONFIG_PARAVIRT
 83 #include <asm/paravirt.h>
 84 #endif
 85 
 86 #include "sched.h"
 87 #include "../workqueue_internal.h"
 88 #include "../smpboot.h"
 89 
 90 #define CREATE_TRACE_POINTS
 91 #include <trace/events/sched.h>
 92 
 93 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
 94 {
 95         unsigned long delta;
 96         ktime_t soft, hard, now;
 97 
 98         for (;;) {
 99                 if (hrtimer_active(period_timer))
100                         break;
101 
102                 now = hrtimer_cb_get_time(period_timer);
103                 hrtimer_forward(period_timer, now, period);
104 
105                 soft = hrtimer_get_softexpires(period_timer);
106                 hard = hrtimer_get_expires(period_timer);
107                 delta = ktime_to_ns(ktime_sub(hard, soft));
108                 __hrtimer_start_range_ns(period_timer, soft, delta,
109                                          HRTIMER_MODE_ABS_PINNED, 0);
110         }
111 }
112 
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 
118 void update_rq_clock(struct rq *rq)
119 {
120         s64 delta;
121 
122         lockdep_assert_held(&rq->lock);
123 
124         if (rq->clock_skip_update & RQCF_ACT_SKIP)
125                 return;
126 
127         delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
128         if (delta < 0)
129                 return;
130         rq->clock += delta;
131         update_rq_clock_task(rq, delta);
132 }
133 
134 /*
135  * Debugging: various feature bits
136  */
137 
138 #define SCHED_FEAT(name, enabled)       \
139         (1UL << __SCHED_FEAT_##name) * enabled |
140 
141 const_debug unsigned int sysctl_sched_features =
142 #include "features.h"
143         0;
144 
145 #undef SCHED_FEAT
146 
147 #ifdef CONFIG_SCHED_DEBUG
148 #define SCHED_FEAT(name, enabled)       \
149         #name ,
150 
151 static const char * const sched_feat_names[] = {
152 #include "features.h"
153 };
154 
155 #undef SCHED_FEAT
156 
157 static int sched_feat_show(struct seq_file *m, void *v)
158 {
159         int i;
160 
161         for (i = 0; i < __SCHED_FEAT_NR; i++) {
162                 if (!(sysctl_sched_features & (1UL << i)))
163                         seq_puts(m, "NO_");
164                 seq_printf(m, "%s ", sched_feat_names[i]);
165         }
166         seq_puts(m, "\n");
167 
168         return 0;
169 }
170 
171 #ifdef HAVE_JUMP_LABEL
172 
173 #define jump_label_key__true  STATIC_KEY_INIT_TRUE
174 #define jump_label_key__false STATIC_KEY_INIT_FALSE
175 
176 #define SCHED_FEAT(name, enabled)       \
177         jump_label_key__##enabled ,
178 
179 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
180 #include "features.h"
181 };
182 
183 #undef SCHED_FEAT
184 
185 static void sched_feat_disable(int i)
186 {
187         if (static_key_enabled(&sched_feat_keys[i]))
188                 static_key_slow_dec(&sched_feat_keys[i]);
189 }
190 
191 static void sched_feat_enable(int i)
192 {
193         if (!static_key_enabled(&sched_feat_keys[i]))
194                 static_key_slow_inc(&sched_feat_keys[i]);
195 }
196 #else
197 static void sched_feat_disable(int i) { };
198 static void sched_feat_enable(int i) { };
199 #endif /* HAVE_JUMP_LABEL */
200 
201 static int sched_feat_set(char *cmp)
202 {
203         int i;
204         int neg = 0;
205 
206         if (strncmp(cmp, "NO_", 3) == 0) {
207                 neg = 1;
208                 cmp += 3;
209         }
210 
211         for (i = 0; i < __SCHED_FEAT_NR; i++) {
212                 if (strcmp(cmp, sched_feat_names[i]) == 0) {
213                         if (neg) {
214                                 sysctl_sched_features &= ~(1UL << i);
215                                 sched_feat_disable(i);
216                         } else {
217                                 sysctl_sched_features |= (1UL << i);
218                                 sched_feat_enable(i);
219                         }
220                         break;
221                 }
222         }
223 
224         return i;
225 }
226 
227 static ssize_t
228 sched_feat_write(struct file *filp, const char __user *ubuf,
229                 size_t cnt, loff_t *ppos)
230 {
231         char buf[64];
232         char *cmp;
233         int i;
234         struct inode *inode;
235 
236         if (cnt > 63)
237                 cnt = 63;
238 
239         if (copy_from_user(&buf, ubuf, cnt))
240                 return -EFAULT;
241 
242         buf[cnt] = 0;
243         cmp = strstrip(buf);
244 
245         /* Ensure the static_key remains in a consistent state */
246         inode = file_inode(filp);
247         mutex_lock(&inode->i_mutex);
248         i = sched_feat_set(cmp);
249         mutex_unlock(&inode->i_mutex);
250         if (i == __SCHED_FEAT_NR)
251                 return -EINVAL;
252 
253         *ppos += cnt;
254 
255         return cnt;
256 }
257 
258 static int sched_feat_open(struct inode *inode, struct file *filp)
259 {
260         return single_open(filp, sched_feat_show, NULL);
261 }
262 
263 static const struct file_operations sched_feat_fops = {
264         .open           = sched_feat_open,
265         .write          = sched_feat_write,
266         .read           = seq_read,
267         .llseek         = seq_lseek,
268         .release        = single_release,
269 };
270 
271 static __init int sched_init_debug(void)
272 {
273         debugfs_create_file("sched_features", 0644, NULL, NULL,
274                         &sched_feat_fops);
275 
276         return 0;
277 }
278 late_initcall(sched_init_debug);
279 #endif /* CONFIG_SCHED_DEBUG */
280 
281 /*
282  * Number of tasks to iterate in a single balance run.
283  * Limited because this is done with IRQs disabled.
284  */
285 const_debug unsigned int sysctl_sched_nr_migrate = 32;
286 
287 /*
288  * period over which we average the RT time consumption, measured
289  * in ms.
290  *
291  * default: 1s
292  */
293 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
294 
295 /*
296  * period over which we measure -rt task cpu usage in us.
297  * default: 1s
298  */
299 unsigned int sysctl_sched_rt_period = 1000000;
300 
301 __read_mostly int scheduler_running;
302 
303 /*
304  * part of the period that we allow rt tasks to run in us.
305  * default: 0.95s
306  */
307 int sysctl_sched_rt_runtime = 950000;
308 
309 /*
310  * this_rq_lock - lock this runqueue and disable interrupts.
311  */
312 static struct rq *this_rq_lock(void)
313         __acquires(rq->lock)
314 {
315         struct rq *rq;
316 
317         local_irq_disable();
318         rq = this_rq();
319         raw_spin_lock(&rq->lock);
320 
321         return rq;
322 }
323 
324 #ifdef CONFIG_SCHED_HRTICK
325 /*
326  * Use HR-timers to deliver accurate preemption points.
327  */
328 
329 static void hrtick_clear(struct rq *rq)
330 {
331         if (hrtimer_active(&rq->hrtick_timer))
332                 hrtimer_cancel(&rq->hrtick_timer);
333 }
334 
335 /*
336  * High-resolution timer tick.
337  * Runs from hardirq context with interrupts disabled.
338  */
339 static enum hrtimer_restart hrtick(struct hrtimer *timer)
340 {
341         struct rq *rq = container_of(timer, struct rq, hrtick_timer);
342 
343         WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
344 
345         raw_spin_lock(&rq->lock);
346         update_rq_clock(rq);
347         rq->curr->sched_class->task_tick(rq, rq->curr, 1);
348         raw_spin_unlock(&rq->lock);
349 
350         return HRTIMER_NORESTART;
351 }
352 
353 #ifdef CONFIG_SMP
354 
355 static int __hrtick_restart(struct rq *rq)
356 {
357         struct hrtimer *timer = &rq->hrtick_timer;
358         ktime_t time = hrtimer_get_softexpires(timer);
359 
360         return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
361 }
362 
363 /*
364  * called from hardirq (IPI) context
365  */
366 static void __hrtick_start(void *arg)
367 {
368         struct rq *rq = arg;
369 
370         raw_spin_lock(&rq->lock);
371         __hrtick_restart(rq);
372         rq->hrtick_csd_pending = 0;
373         raw_spin_unlock(&rq->lock);
374 }
375 
376 /*
377  * Called to set the hrtick timer state.
378  *
379  * called with rq->lock held and irqs disabled
380  */
381 void hrtick_start(struct rq *rq, u64 delay)
382 {
383         struct hrtimer *timer = &rq->hrtick_timer;
384         ktime_t time;
385         s64 delta;
386 
387         /*
388          * Don't schedule slices shorter than 10000ns, that just
389          * doesn't make sense and can cause timer DoS.
390          */
391         delta = max_t(s64, delay, 10000LL);
392         time = ktime_add_ns(timer->base->get_time(), delta);
393 
394         hrtimer_set_expires(timer, time);
395 
396         if (rq == this_rq()) {
397                 __hrtick_restart(rq);
398         } else if (!rq->hrtick_csd_pending) {
399                 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
400                 rq->hrtick_csd_pending = 1;
401         }
402 }
403 
404 static int
405 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
406 {
407         int cpu = (int)(long)hcpu;
408 
409         switch (action) {
410         case CPU_UP_CANCELED:
411         case CPU_UP_CANCELED_FROZEN:
412         case CPU_DOWN_PREPARE:
413         case CPU_DOWN_PREPARE_FROZEN:
414         case CPU_DEAD:
415         case CPU_DEAD_FROZEN:
416                 hrtick_clear(cpu_rq(cpu));
417                 return NOTIFY_OK;
418         }
419 
420         return NOTIFY_DONE;
421 }
422 
423 static __init void init_hrtick(void)
424 {
425         hotcpu_notifier(hotplug_hrtick, 0);
426 }
427 #else
428 /*
429  * Called to set the hrtick timer state.
430  *
431  * called with rq->lock held and irqs disabled
432  */
433 void hrtick_start(struct rq *rq, u64 delay)
434 {
435         /*
436          * Don't schedule slices shorter than 10000ns, that just
437          * doesn't make sense. Rely on vruntime for fairness.
438          */
439         delay = max_t(u64, delay, 10000LL);
440         __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
441                         HRTIMER_MODE_REL_PINNED, 0);
442 }
443 
444 static inline void init_hrtick(void)
445 {
446 }
447 #endif /* CONFIG_SMP */
448 
449 static void init_rq_hrtick(struct rq *rq)
450 {
451 #ifdef CONFIG_SMP
452         rq->hrtick_csd_pending = 0;
453 
454         rq->hrtick_csd.flags = 0;
455         rq->hrtick_csd.func = __hrtick_start;
456         rq->hrtick_csd.info = rq;
457 #endif
458 
459         hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
460         rq->hrtick_timer.function = hrtick;
461 }
462 #else   /* CONFIG_SCHED_HRTICK */
463 static inline void hrtick_clear(struct rq *rq)
464 {
465 }
466 
467 static inline void init_rq_hrtick(struct rq *rq)
468 {
469 }
470 
471 static inline void init_hrtick(void)
472 {
473 }
474 #endif  /* CONFIG_SCHED_HRTICK */
475 
476 /*
477  * cmpxchg based fetch_or, macro so it works for different integer types
478  */
479 #define fetch_or(ptr, val)                                              \
480 ({      typeof(*(ptr)) __old, __val = *(ptr);                           \
481         for (;;) {                                                      \
482                 __old = cmpxchg((ptr), __val, __val | (val));           \
483                 if (__old == __val)                                     \
484                         break;                                          \
485                 __val = __old;                                          \
486         }                                                               \
487         __old;                                                          \
488 })
489 
490 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
491 /*
492  * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
493  * this avoids any races wrt polling state changes and thereby avoids
494  * spurious IPIs.
495  */
496 static bool set_nr_and_not_polling(struct task_struct *p)
497 {
498         struct thread_info *ti = task_thread_info(p);
499         return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
500 }
501 
502 /*
503  * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
504  *
505  * If this returns true, then the idle task promises to call
506  * sched_ttwu_pending() and reschedule soon.
507  */
508 static bool set_nr_if_polling(struct task_struct *p)
509 {
510         struct thread_info *ti = task_thread_info(p);
511         typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
512 
513         for (;;) {
514                 if (!(val & _TIF_POLLING_NRFLAG))
515                         return false;
516                 if (val & _TIF_NEED_RESCHED)
517                         return true;
518                 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
519                 if (old == val)
520                         break;
521                 val = old;
522         }
523         return true;
524 }
525 
526 #else
527 static bool set_nr_and_not_polling(struct task_struct *p)
528 {
529         set_tsk_need_resched(p);
530         return true;
531 }
532 
533 #ifdef CONFIG_SMP
534 static bool set_nr_if_polling(struct task_struct *p)
535 {
536         return false;
537 }
538 #endif
539 #endif
540 
541 /*
542  * resched_curr - mark rq's current task 'to be rescheduled now'.
543  *
544  * On UP this means the setting of the need_resched flag, on SMP it
545  * might also involve a cross-CPU call to trigger the scheduler on
546  * the target CPU.
547  */
548 void resched_curr(struct rq *rq)
549 {
550         struct task_struct *curr = rq->curr;
551         int cpu;
552 
553         lockdep_assert_held(&rq->lock);
554 
555         if (test_tsk_need_resched(curr))
556                 return;
557 
558         cpu = cpu_of(rq);
559 
560         if (cpu == smp_processor_id()) {
561                 set_tsk_need_resched(curr);
562                 set_preempt_need_resched();
563                 return;
564         }
565 
566         if (set_nr_and_not_polling(curr))
567                 smp_send_reschedule(cpu);
568         else
569                 trace_sched_wake_idle_without_ipi(cpu);
570 }
571 
572 void resched_cpu(int cpu)
573 {
574         struct rq *rq = cpu_rq(cpu);
575         unsigned long flags;
576 
577         if (!raw_spin_trylock_irqsave(&rq->lock, flags))
578                 return;
579         resched_curr(rq);
580         raw_spin_unlock_irqrestore(&rq->lock, flags);
581 }
582 
583 #ifdef CONFIG_SMP
584 #ifdef CONFIG_NO_HZ_COMMON
585 /*
586  * In the semi idle case, use the nearest busy cpu for migrating timers
587  * from an idle cpu.  This is good for power-savings.
588  *
589  * We don't do similar optimization for completely idle system, as
590  * selecting an idle cpu will add more delays to the timers than intended
591  * (as that cpu's timer base may not be uptodate wrt jiffies etc).
592  */
593 int get_nohz_timer_target(int pinned)
594 {
595         int cpu = smp_processor_id();
596         int i;
597         struct sched_domain *sd;
598 
599         if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
600                 return cpu;
601 
602         rcu_read_lock();
603         for_each_domain(cpu, sd) {
604                 for_each_cpu(i, sched_domain_span(sd)) {
605                         if (!idle_cpu(i)) {
606                                 cpu = i;
607                                 goto unlock;
608                         }
609                 }
610         }
611 unlock:
612         rcu_read_unlock();
613         return cpu;
614 }
615 /*
616  * When add_timer_on() enqueues a timer into the timer wheel of an
617  * idle CPU then this timer might expire before the next timer event
618  * which is scheduled to wake up that CPU. In case of a completely
619  * idle system the next event might even be infinite time into the
620  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
621  * leaves the inner idle loop so the newly added timer is taken into
622  * account when the CPU goes back to idle and evaluates the timer
623  * wheel for the next timer event.
624  */
625 static void wake_up_idle_cpu(int cpu)
626 {
627         struct rq *rq = cpu_rq(cpu);
628 
629         if (cpu == smp_processor_id())
630                 return;
631 
632         if (set_nr_and_not_polling(rq->idle))
633                 smp_send_reschedule(cpu);
634         else
635                 trace_sched_wake_idle_without_ipi(cpu);
636 }
637 
638 static bool wake_up_full_nohz_cpu(int cpu)
639 {
640         /*
641          * We just need the target to call irq_exit() and re-evaluate
642          * the next tick. The nohz full kick at least implies that.
643          * If needed we can still optimize that later with an
644          * empty IRQ.
645          */
646         if (tick_nohz_full_cpu(cpu)) {
647                 if (cpu != smp_processor_id() ||
648                     tick_nohz_tick_stopped())
649                         tick_nohz_full_kick_cpu(cpu);
650                 return true;
651         }
652 
653         return false;
654 }
655 
656 void wake_up_nohz_cpu(int cpu)
657 {
658         if (!wake_up_full_nohz_cpu(cpu))
659                 wake_up_idle_cpu(cpu);
660 }
661 
662 static inline bool got_nohz_idle_kick(void)
663 {
664         int cpu = smp_processor_id();
665 
666         if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
667                 return false;
668 
669         if (idle_cpu(cpu) && !need_resched())
670                 return true;
671 
672         /*
673          * We can't run Idle Load Balance on this CPU for this time so we
674          * cancel it and clear NOHZ_BALANCE_KICK
675          */
676         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
677         return false;
678 }
679 
680 #else /* CONFIG_NO_HZ_COMMON */
681 
682 static inline bool got_nohz_idle_kick(void)
683 {
684         return false;
685 }
686 
687 #endif /* CONFIG_NO_HZ_COMMON */
688 
689 #ifdef CONFIG_NO_HZ_FULL
690 bool sched_can_stop_tick(void)
691 {
692         /*
693          * More than one running task need preemption.
694          * nr_running update is assumed to be visible
695          * after IPI is sent from wakers.
696          */
697         if (this_rq()->nr_running > 1)
698                 return false;
699 
700         return true;
701 }
702 #endif /* CONFIG_NO_HZ_FULL */
703 
704 void sched_avg_update(struct rq *rq)
705 {
706         s64 period = sched_avg_period();
707 
708         while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
709                 /*
710                  * Inline assembly required to prevent the compiler
711                  * optimising this loop into a divmod call.
712                  * See __iter_div_u64_rem() for another example of this.
713                  */
714                 asm("" : "+rm" (rq->age_stamp));
715                 rq->age_stamp += period;
716                 rq->rt_avg /= 2;
717         }
718 }
719 
720 #endif /* CONFIG_SMP */
721 
722 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
723                         (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
724 /*
725  * Iterate task_group tree rooted at *from, calling @down when first entering a
726  * node and @up when leaving it for the final time.
727  *
728  * Caller must hold rcu_lock or sufficient equivalent.
729  */
730 int walk_tg_tree_from(struct task_group *from,
731                              tg_visitor down, tg_visitor up, void *data)
732 {
733         struct task_group *parent, *child;
734         int ret;
735 
736         parent = from;
737 
738 down:
739         ret = (*down)(parent, data);
740         if (ret)
741                 goto out;
742         list_for_each_entry_rcu(child, &parent->children, siblings) {
743                 parent = child;
744                 goto down;
745 
746 up:
747                 continue;
748         }
749         ret = (*up)(parent, data);
750         if (ret || parent == from)
751                 goto out;
752 
753         child = parent;
754         parent = parent->parent;
755         if (parent)
756                 goto up;
757 out:
758         return ret;
759 }
760 
761 int tg_nop(struct task_group *tg, void *data)
762 {
763         return 0;
764 }
765 #endif
766 
767 static void set_load_weight(struct task_struct *p)
768 {
769         int prio = p->static_prio - MAX_RT_PRIO;
770         struct load_weight *load = &p->se.load;
771 
772         /*
773          * SCHED_IDLE tasks get minimal weight:
774          */
775         if (p->policy == SCHED_IDLE) {
776                 load->weight = scale_load(WEIGHT_IDLEPRIO);
777                 load->inv_weight = WMULT_IDLEPRIO;
778                 return;
779         }
780 
781         load->weight = scale_load(prio_to_weight[prio]);
782         load->inv_weight = prio_to_wmult[prio];
783 }
784 
785 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
786 {
787         update_rq_clock(rq);
788         sched_info_queued(rq, p);
789         p->sched_class->enqueue_task(rq, p, flags);
790 }
791 
792 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
793 {
794         update_rq_clock(rq);
795         sched_info_dequeued(rq, p);
796         p->sched_class->dequeue_task(rq, p, flags);
797 }
798 
799 void activate_task(struct rq *rq, struct task_struct *p, int flags)
800 {
801         if (task_contributes_to_load(p))
802                 rq->nr_uninterruptible--;
803 
804         enqueue_task(rq, p, flags);
805 }
806 
807 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
808 {
809         if (task_contributes_to_load(p))
810                 rq->nr_uninterruptible++;
811 
812         dequeue_task(rq, p, flags);
813 }
814 
815 static void update_rq_clock_task(struct rq *rq, s64 delta)
816 {
817 /*
818  * In theory, the compile should just see 0 here, and optimize out the call
819  * to sched_rt_avg_update. But I don't trust it...
820  */
821 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
822         s64 steal = 0, irq_delta = 0;
823 #endif
824 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
825         irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
826 
827         /*
828          * Since irq_time is only updated on {soft,}irq_exit, we might run into
829          * this case when a previous update_rq_clock() happened inside a
830          * {soft,}irq region.
831          *
832          * When this happens, we stop ->clock_task and only update the
833          * prev_irq_time stamp to account for the part that fit, so that a next
834          * update will consume the rest. This ensures ->clock_task is
835          * monotonic.
836          *
837          * It does however cause some slight miss-attribution of {soft,}irq
838          * time, a more accurate solution would be to update the irq_time using
839          * the current rq->clock timestamp, except that would require using
840          * atomic ops.
841          */
842         if (irq_delta > delta)
843                 irq_delta = delta;
844 
845         rq->prev_irq_time += irq_delta;
846         delta -= irq_delta;
847 #endif
848 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
849         if (static_key_false((&paravirt_steal_rq_enabled))) {
850                 steal = paravirt_steal_clock(cpu_of(rq));
851                 steal -= rq->prev_steal_time_rq;
852 
853                 if (unlikely(steal > delta))
854                         steal = delta;
855 
856                 rq->prev_steal_time_rq += steal;
857                 delta -= steal;
858         }
859 #endif
860 
861         rq->clock_task += delta;
862 
863 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
864         if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
865                 sched_rt_avg_update(rq, irq_delta + steal);
866 #endif
867 }
868 
869 void sched_set_stop_task(int cpu, struct task_struct *stop)
870 {
871         struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
872         struct task_struct *old_stop = cpu_rq(cpu)->stop;
873 
874         if (stop) {
875                 /*
876                  * Make it appear like a SCHED_FIFO task, its something
877                  * userspace knows about and won't get confused about.
878                  *
879                  * Also, it will make PI more or less work without too
880                  * much confusion -- but then, stop work should not
881                  * rely on PI working anyway.
882                  */
883                 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
884 
885                 stop->sched_class = &stop_sched_class;
886         }
887 
888         cpu_rq(cpu)->stop = stop;
889 
890         if (old_stop) {
891                 /*
892                  * Reset it back to a normal scheduling class so that
893                  * it can die in pieces.
894                  */
895                 old_stop->sched_class = &rt_sched_class;
896         }
897 }
898 
899 /*
900  * __normal_prio - return the priority that is based on the static prio
901  */
902 static inline int __normal_prio(struct task_struct *p)
903 {
904         return p->static_prio;
905 }
906 
907 /*
908  * Calculate the expected normal priority: i.e. priority
909  * without taking RT-inheritance into account. Might be
910  * boosted by interactivity modifiers. Changes upon fork,
911  * setprio syscalls, and whenever the interactivity
912  * estimator recalculates.
913  */
914 static inline int normal_prio(struct task_struct *p)
915 {
916         int prio;
917 
918         if (task_has_dl_policy(p))
919                 prio = MAX_DL_PRIO-1;
920         else if (task_has_rt_policy(p))
921                 prio = MAX_RT_PRIO-1 - p->rt_priority;
922         else
923                 prio = __normal_prio(p);
924         return prio;
925 }
926 
927 /*
928  * Calculate the current priority, i.e. the priority
929  * taken into account by the scheduler. This value might
930  * be boosted by RT tasks, or might be boosted by
931  * interactivity modifiers. Will be RT if the task got
932  * RT-boosted. If not then it returns p->normal_prio.
933  */
934 static int effective_prio(struct task_struct *p)
935 {
936         p->normal_prio = normal_prio(p);
937         /*
938          * If we are RT tasks or we were boosted to RT priority,
939          * keep the priority unchanged. Otherwise, update priority
940          * to the normal priority:
941          */
942         if (!rt_prio(p->prio))
943                 return p->normal_prio;
944         return p->prio;
945 }
946 
947 /**
948  * task_curr - is this task currently executing on a CPU?
949  * @p: the task in question.
950  *
951  * Return: 1 if the task is currently executing. 0 otherwise.
952  */
953 inline int task_curr(const struct task_struct *p)
954 {
955         return cpu_curr(task_cpu(p)) == p;
956 }
957 
958 /*
959  * Can drop rq->lock because from sched_class::switched_from() methods drop it.
960  */
961 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
962                                        const struct sched_class *prev_class,
963                                        int oldprio)
964 {
965         if (prev_class != p->sched_class) {
966                 if (prev_class->switched_from)
967                         prev_class->switched_from(rq, p);
968                 /* Possble rq->lock 'hole'.  */
969                 p->sched_class->switched_to(rq, p);
970         } else if (oldprio != p->prio || dl_task(p))
971                 p->sched_class->prio_changed(rq, p, oldprio);
972 }
973 
974 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
975 {
976         const struct sched_class *class;
977 
978         if (p->sched_class == rq->curr->sched_class) {
979                 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
980         } else {
981                 for_each_class(class) {
982                         if (class == rq->curr->sched_class)
983                                 break;
984                         if (class == p->sched_class) {
985                                 resched_curr(rq);
986                                 break;
987                         }
988                 }
989         }
990 
991         /*
992          * A queue event has occurred, and we're going to schedule.  In
993          * this case, we can save a useless back to back clock update.
994          */
995         if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
996                 rq_clock_skip_update(rq, true);
997 }
998 
999 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
1000 
1001 void register_task_migration_notifier(struct notifier_block *n)
1002 {
1003         atomic_notifier_chain_register(&task_migration_notifier, n);
1004 }
1005 
1006 #ifdef CONFIG_SMP
1007 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1008 {
1009 #ifdef CONFIG_SCHED_DEBUG
1010         /*
1011          * We should never call set_task_cpu() on a blocked task,
1012          * ttwu() will sort out the placement.
1013          */
1014         WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1015                         !p->on_rq);
1016 
1017 #ifdef CONFIG_LOCKDEP
1018         /*
1019          * The caller should hold either p->pi_lock or rq->lock, when changing
1020          * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1021          *
1022          * sched_move_task() holds both and thus holding either pins the cgroup,
1023          * see task_group().
1024          *
1025          * Furthermore, all task_rq users should acquire both locks, see
1026          * task_rq_lock().
1027          */
1028         WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1029                                       lockdep_is_held(&task_rq(p)->lock)));
1030 #endif
1031 #endif
1032 
1033         trace_sched_migrate_task(p, new_cpu);
1034 
1035         if (task_cpu(p) != new_cpu) {
1036                 struct task_migration_notifier tmn;
1037 
1038                 if (p->sched_class->migrate_task_rq)
1039                         p->sched_class->migrate_task_rq(p, new_cpu);
1040                 p->se.nr_migrations++;
1041                 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1042 
1043                 tmn.task = p;
1044                 tmn.from_cpu = task_cpu(p);
1045                 tmn.to_cpu = new_cpu;
1046 
1047                 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
1048         }
1049 
1050         __set_task_cpu(p, new_cpu);
1051 }
1052 
1053 static void __migrate_swap_task(struct task_struct *p, int cpu)
1054 {
1055         if (task_on_rq_queued(p)) {
1056                 struct rq *src_rq, *dst_rq;
1057 
1058                 src_rq = task_rq(p);
1059                 dst_rq = cpu_rq(cpu);
1060 
1061                 deactivate_task(src_rq, p, 0);
1062                 set_task_cpu(p, cpu);
1063                 activate_task(dst_rq, p, 0);
1064                 check_preempt_curr(dst_rq, p, 0);
1065         } else {
1066                 /*
1067                  * Task isn't running anymore; make it appear like we migrated
1068                  * it before it went to sleep. This means on wakeup we make the
1069                  * previous cpu our targer instead of where it really is.
1070                  */
1071                 p->wake_cpu = cpu;
1072         }
1073 }
1074 
1075 struct migration_swap_arg {
1076         struct task_struct *src_task, *dst_task;
1077         int src_cpu, dst_cpu;
1078 };
1079 
1080 static int migrate_swap_stop(void *data)
1081 {
1082         struct migration_swap_arg *arg = data;
1083         struct rq *src_rq, *dst_rq;
1084         int ret = -EAGAIN;
1085 
1086         src_rq = cpu_rq(arg->src_cpu);
1087         dst_rq = cpu_rq(arg->dst_cpu);
1088 
1089         double_raw_lock(&arg->src_task->pi_lock,
1090                         &arg->dst_task->pi_lock);
1091         double_rq_lock(src_rq, dst_rq);
1092         if (task_cpu(arg->dst_task) != arg->dst_cpu)
1093                 goto unlock;
1094 
1095         if (task_cpu(arg->src_task) != arg->src_cpu)
1096                 goto unlock;
1097 
1098         if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1099                 goto unlock;
1100 
1101         if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1102                 goto unlock;
1103 
1104         __migrate_swap_task(arg->src_task, arg->dst_cpu);
1105         __migrate_swap_task(arg->dst_task, arg->src_cpu);
1106 
1107         ret = 0;
1108 
1109 unlock:
1110         double_rq_unlock(src_rq, dst_rq);
1111         raw_spin_unlock(&arg->dst_task->pi_lock);
1112         raw_spin_unlock(&arg->src_task->pi_lock);
1113 
1114         return ret;
1115 }
1116 
1117 /*
1118  * Cross migrate two tasks
1119  */
1120 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1121 {
1122         struct migration_swap_arg arg;
1123         int ret = -EINVAL;
1124 
1125         arg = (struct migration_swap_arg){
1126                 .src_task = cur,
1127                 .src_cpu = task_cpu(cur),
1128                 .dst_task = p,
1129                 .dst_cpu = task_cpu(p),
1130         };
1131 
1132         if (arg.src_cpu == arg.dst_cpu)
1133                 goto out;
1134 
1135         /*
1136          * These three tests are all lockless; this is OK since all of them
1137          * will be re-checked with proper locks held further down the line.
1138          */
1139         if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1140                 goto out;
1141 
1142         if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1143                 goto out;
1144 
1145         if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1146                 goto out;
1147 
1148         trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1149         ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1150 
1151 out:
1152         return ret;
1153 }
1154 
1155 struct migration_arg {
1156         struct task_struct *task;
1157         int dest_cpu;
1158 };
1159 
1160 static int migration_cpu_stop(void *data);
1161 
1162 /*
1163  * wait_task_inactive - wait for a thread to unschedule.
1164  *
1165  * If @match_state is nonzero, it's the @p->state value just checked and
1166  * not expected to change.  If it changes, i.e. @p might have woken up,
1167  * then return zero.  When we succeed in waiting for @p to be off its CPU,
1168  * we return a positive number (its total switch count).  If a second call
1169  * a short while later returns the same number, the caller can be sure that
1170  * @p has remained unscheduled the whole time.
1171  *
1172  * The caller must ensure that the task *will* unschedule sometime soon,
1173  * else this function might spin for a *long* time. This function can't
1174  * be called with interrupts off, or it may introduce deadlock with
1175  * smp_call_function() if an IPI is sent by the same process we are
1176  * waiting to become inactive.
1177  */
1178 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1179 {
1180         unsigned long flags;
1181         int running, queued;
1182         unsigned long ncsw;
1183         struct rq *rq;
1184 
1185         for (;;) {
1186                 /*
1187                  * We do the initial early heuristics without holding
1188                  * any task-queue locks at all. We'll only try to get
1189                  * the runqueue lock when things look like they will
1190                  * work out!
1191                  */
1192                 rq = task_rq(p);
1193 
1194                 /*
1195                  * If the task is actively running on another CPU
1196                  * still, just relax and busy-wait without holding
1197                  * any locks.
1198                  *
1199                  * NOTE! Since we don't hold any locks, it's not
1200                  * even sure that "rq" stays as the right runqueue!
1201                  * But we don't care, since "task_running()" will
1202                  * return false if the runqueue has changed and p
1203                  * is actually now running somewhere else!
1204                  */
1205                 while (task_running(rq, p)) {
1206                         if (match_state && unlikely(p->state != match_state))
1207                                 return 0;
1208                         cpu_relax();
1209                 }
1210 
1211                 /*
1212                  * Ok, time to look more closely! We need the rq
1213                  * lock now, to be *sure*. If we're wrong, we'll
1214                  * just go back and repeat.
1215                  */
1216                 rq = task_rq_lock(p, &flags);
1217                 trace_sched_wait_task(p);
1218                 running = task_running(rq, p);
1219                 queued = task_on_rq_queued(p);
1220                 ncsw = 0;
1221                 if (!match_state || p->state == match_state)
1222                         ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1223                 task_rq_unlock(rq, p, &flags);
1224 
1225                 /*
1226                  * If it changed from the expected state, bail out now.
1227                  */
1228                 if (unlikely(!ncsw))
1229                         break;
1230 
1231                 /*
1232                  * Was it really running after all now that we
1233                  * checked with the proper locks actually held?
1234                  *
1235                  * Oops. Go back and try again..
1236                  */
1237                 if (unlikely(running)) {
1238                         cpu_relax();
1239                         continue;
1240                 }
1241 
1242                 /*
1243                  * It's not enough that it's not actively running,
1244                  * it must be off the runqueue _entirely_, and not
1245                  * preempted!
1246                  *
1247                  * So if it was still runnable (but just not actively
1248                  * running right now), it's preempted, and we should
1249                  * yield - it could be a while.
1250                  */
1251                 if (unlikely(queued)) {
1252                         ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1253 
1254                         set_current_state(TASK_UNINTERRUPTIBLE);
1255                         schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1256                         continue;
1257                 }
1258 
1259                 /*
1260                  * Ahh, all good. It wasn't running, and it wasn't
1261                  * runnable, which means that it will never become
1262                  * running in the future either. We're all done!
1263                  */
1264                 break;
1265         }
1266 
1267         return ncsw;
1268 }
1269 
1270 /***
1271  * kick_process - kick a running thread to enter/exit the kernel
1272  * @p: the to-be-kicked thread
1273  *
1274  * Cause a process which is running on another CPU to enter
1275  * kernel-mode, without any delay. (to get signals handled.)
1276  *
1277  * NOTE: this function doesn't have to take the runqueue lock,
1278  * because all it wants to ensure is that the remote task enters
1279  * the kernel. If the IPI races and the task has been migrated
1280  * to another CPU then no harm is done and the purpose has been
1281  * achieved as well.
1282  */
1283 void kick_process(struct task_struct *p)
1284 {
1285         int cpu;
1286 
1287         preempt_disable();
1288         cpu = task_cpu(p);
1289         if ((cpu != smp_processor_id()) && task_curr(p))
1290                 smp_send_reschedule(cpu);
1291         preempt_enable();
1292 }
1293 EXPORT_SYMBOL_GPL(kick_process);
1294 #endif /* CONFIG_SMP */
1295 
1296 #ifdef CONFIG_SMP
1297 /*
1298  * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1299  */
1300 static int select_fallback_rq(int cpu, struct task_struct *p)
1301 {
1302         int nid = cpu_to_node(cpu);
1303         const struct cpumask *nodemask = NULL;
1304         enum { cpuset, possible, fail } state = cpuset;
1305         int dest_cpu;
1306 
1307         /*
1308          * If the node that the cpu is on has been offlined, cpu_to_node()
1309          * will return -1. There is no cpu on the node, and we should
1310          * select the cpu on the other node.
1311          */
1312         if (nid != -1) {
1313                 nodemask = cpumask_of_node(nid);
1314 
1315                 /* Look for allowed, online CPU in same node. */
1316                 for_each_cpu(dest_cpu, nodemask) {
1317                         if (!cpu_online(dest_cpu))
1318                                 continue;
1319                         if (!cpu_active(dest_cpu))
1320                                 continue;
1321                         if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1322                                 return dest_cpu;
1323                 }
1324         }
1325 
1326         for (;;) {
1327                 /* Any allowed, online CPU? */
1328                 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1329                         if (!cpu_online(dest_cpu))
1330                                 continue;
1331                         if (!cpu_active(dest_cpu))
1332                                 continue;
1333                         goto out;
1334                 }
1335 
1336                 switch (state) {
1337                 case cpuset:
1338                         /* No more Mr. Nice Guy. */
1339                         cpuset_cpus_allowed_fallback(p);
1340                         state = possible;
1341                         break;
1342 
1343                 case possible:
1344                         do_set_cpus_allowed(p, cpu_possible_mask);
1345                         state = fail;
1346                         break;
1347 
1348                 case fail:
1349                         BUG();
1350                         break;
1351                 }
1352         }
1353 
1354 out:
1355         if (state != cpuset) {
1356                 /*
1357                  * Don't tell them about moving exiting tasks or
1358                  * kernel threads (both mm NULL), since they never
1359                  * leave kernel.
1360                  */
1361                 if (p->mm && printk_ratelimit()) {
1362                         printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1363                                         task_pid_nr(p), p->comm, cpu);
1364                 }
1365         }
1366 
1367         return dest_cpu;
1368 }
1369 
1370 /*
1371  * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1372  */
1373 static inline
1374 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1375 {
1376         if (p->nr_cpus_allowed > 1)
1377                 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1378 
1379         /*
1380          * In order not to call set_task_cpu() on a blocking task we need
1381          * to rely on ttwu() to place the task on a valid ->cpus_allowed
1382          * cpu.
1383          *
1384          * Since this is common to all placement strategies, this lives here.
1385          *
1386          * [ this allows ->select_task() to simply return task_cpu(p) and
1387          *   not worry about this generic constraint ]
1388          */
1389         if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1390                      !cpu_online(cpu)))
1391                 cpu = select_fallback_rq(task_cpu(p), p);
1392 
1393         return cpu;
1394 }
1395 
1396 static void update_avg(u64 *avg, u64 sample)
1397 {
1398         s64 diff = sample - *avg;
1399         *avg += diff >> 3;
1400 }
1401 #endif
1402 
1403 static void
1404 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1405 {
1406 #ifdef CONFIG_SCHEDSTATS
1407         struct rq *rq = this_rq();
1408 
1409 #ifdef CONFIG_SMP
1410         int this_cpu = smp_processor_id();
1411 
1412         if (cpu == this_cpu) {
1413                 schedstat_inc(rq, ttwu_local);
1414                 schedstat_inc(p, se.statistics.nr_wakeups_local);
1415         } else {
1416                 struct sched_domain *sd;
1417 
1418                 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1419                 rcu_read_lock();
1420                 for_each_domain(this_cpu, sd) {
1421                         if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1422                                 schedstat_inc(sd, ttwu_wake_remote);
1423                                 break;
1424                         }
1425                 }
1426                 rcu_read_unlock();
1427         }
1428 
1429         if (wake_flags & WF_MIGRATED)
1430                 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1431 
1432 #endif /* CONFIG_SMP */
1433 
1434         schedstat_inc(rq, ttwu_count);
1435         schedstat_inc(p, se.statistics.nr_wakeups);
1436 
1437         if (wake_flags & WF_SYNC)
1438                 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1439 
1440 #endif /* CONFIG_SCHEDSTATS */
1441 }
1442 
1443 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1444 {
1445         activate_task(rq, p, en_flags);
1446         p->on_rq = TASK_ON_RQ_QUEUED;
1447 
1448         /* if a worker is waking up, notify workqueue */
1449         if (p->flags & PF_WQ_WORKER)
1450                 wq_worker_waking_up(p, cpu_of(rq));
1451 }
1452 
1453 /*
1454  * Mark the task runnable and perform wakeup-preemption.
1455  */
1456 static void
1457 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1458 {
1459         check_preempt_curr(rq, p, wake_flags);
1460         trace_sched_wakeup(p, true);
1461 
1462         p->state = TASK_RUNNING;
1463 #ifdef CONFIG_SMP
1464         if (p->sched_class->task_woken)
1465                 p->sched_class->task_woken(rq, p);
1466 
1467         if (rq->idle_stamp) {
1468                 u64 delta = rq_clock(rq) - rq->idle_stamp;
1469                 u64 max = 2*rq->max_idle_balance_cost;
1470 
1471                 update_avg(&rq->avg_idle, delta);
1472 
1473                 if (rq->avg_idle > max)
1474                         rq->avg_idle = max;
1475 
1476                 rq->idle_stamp = 0;
1477         }
1478 #endif
1479 }
1480 
1481 static void
1482 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1483 {
1484 #ifdef CONFIG_SMP
1485         if (p->sched_contributes_to_load)
1486                 rq->nr_uninterruptible--;
1487 #endif
1488 
1489         ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1490         ttwu_do_wakeup(rq, p, wake_flags);
1491 }
1492 
1493 /*
1494  * Called in case the task @p isn't fully descheduled from its runqueue,
1495  * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1496  * since all we need to do is flip p->state to TASK_RUNNING, since
1497  * the task is still ->on_rq.
1498  */
1499 static int ttwu_remote(struct task_struct *p, int wake_flags)
1500 {
1501         struct rq *rq;
1502         int ret = 0;
1503 
1504         rq = __task_rq_lock(p);
1505         if (task_on_rq_queued(p)) {
1506                 /* check_preempt_curr() may use rq clock */
1507                 update_rq_clock(rq);
1508                 ttwu_do_wakeup(rq, p, wake_flags);
1509                 ret = 1;
1510         }
1511         __task_rq_unlock(rq);
1512 
1513         return ret;
1514 }
1515 
1516 #ifdef CONFIG_SMP
1517 void sched_ttwu_pending(void)
1518 {
1519         struct rq *rq = this_rq();
1520         struct llist_node *llist = llist_del_all(&rq->wake_list);
1521         struct task_struct *p;
1522         unsigned long flags;
1523 
1524         if (!llist)
1525                 return;
1526 
1527         raw_spin_lock_irqsave(&rq->lock, flags);
1528 
1529         while (llist) {
1530                 p = llist_entry(llist, struct task_struct, wake_entry);
1531                 llist = llist_next(llist);
1532                 ttwu_do_activate(rq, p, 0);
1533         }
1534 
1535         raw_spin_unlock_irqrestore(&rq->lock, flags);
1536 }
1537 
1538 void scheduler_ipi(void)
1539 {
1540         /*
1541          * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1542          * TIF_NEED_RESCHED remotely (for the first time) will also send
1543          * this IPI.
1544          */
1545         preempt_fold_need_resched();
1546 
1547         if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1548                 return;
1549 
1550         /*
1551          * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1552          * traditionally all their work was done from the interrupt return
1553          * path. Now that we actually do some work, we need to make sure
1554          * we do call them.
1555          *
1556          * Some archs already do call them, luckily irq_enter/exit nest
1557          * properly.
1558          *
1559          * Arguably we should visit all archs and update all handlers,
1560          * however a fair share of IPIs are still resched only so this would
1561          * somewhat pessimize the simple resched case.
1562          */
1563         irq_enter();
1564         sched_ttwu_pending();
1565 
1566         /*
1567          * Check if someone kicked us for doing the nohz idle load balance.
1568          */
1569         if (unlikely(got_nohz_idle_kick())) {
1570                 this_rq()->idle_balance = 1;
1571                 raise_softirq_irqoff(SCHED_SOFTIRQ);
1572         }
1573         irq_exit();
1574 }
1575 
1576 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1577 {
1578         struct rq *rq = cpu_rq(cpu);
1579 
1580         if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1581                 if (!set_nr_if_polling(rq->idle))
1582                         smp_send_reschedule(cpu);
1583                 else
1584                         trace_sched_wake_idle_without_ipi(cpu);
1585         }
1586 }
1587 
1588 void wake_up_if_idle(int cpu)
1589 {
1590         struct rq *rq = cpu_rq(cpu);
1591         unsigned long flags;
1592 
1593         rcu_read_lock();
1594 
1595         if (!is_idle_task(rcu_dereference(rq->curr)))
1596                 goto out;
1597 
1598         if (set_nr_if_polling(rq->idle)) {
1599                 trace_sched_wake_idle_without_ipi(cpu);
1600         } else {
1601                 raw_spin_lock_irqsave(&rq->lock, flags);
1602                 if (is_idle_task(rq->curr))
1603                         smp_send_reschedule(cpu);
1604                 /* Else cpu is not in idle, do nothing here */
1605                 raw_spin_unlock_irqrestore(&rq->lock, flags);
1606         }
1607 
1608 out:
1609         rcu_read_unlock();
1610 }
1611 
1612 bool cpus_share_cache(int this_cpu, int that_cpu)
1613 {
1614         return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1615 }
1616 #endif /* CONFIG_SMP */
1617 
1618 static void ttwu_queue(struct task_struct *p, int cpu)
1619 {
1620         struct rq *rq = cpu_rq(cpu);
1621 
1622 #if defined(CONFIG_SMP)
1623         if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1624                 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1625                 ttwu_queue_remote(p, cpu);
1626                 return;
1627         }
1628 #endif
1629 
1630         raw_spin_lock(&rq->lock);
1631         ttwu_do_activate(rq, p, 0);
1632         raw_spin_unlock(&rq->lock);
1633 }
1634 
1635 /**
1636  * try_to_wake_up - wake up a thread
1637  * @p: the thread to be awakened
1638  * @state: the mask of task states that can be woken
1639  * @wake_flags: wake modifier flags (WF_*)
1640  *
1641  * Put it on the run-queue if it's not already there. The "current"
1642  * thread is always on the run-queue (except when the actual
1643  * re-schedule is in progress), and as such you're allowed to do
1644  * the simpler "current->state = TASK_RUNNING" to mark yourself
1645  * runnable without the overhead of this.
1646  *
1647  * Return: %true if @p was woken up, %false if it was already running.
1648  * or @state didn't match @p's state.
1649  */
1650 static int
1651 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1652 {
1653         unsigned long flags;
1654         int cpu, success = 0;
1655 
1656         /*
1657          * If we are going to wake up a thread waiting for CONDITION we
1658          * need to ensure that CONDITION=1 done by the caller can not be
1659          * reordered with p->state check below. This pairs with mb() in
1660          * set_current_state() the waiting thread does.
1661          */
1662         smp_mb__before_spinlock();
1663         raw_spin_lock_irqsave(&p->pi_lock, flags);
1664         if (!(p->state & state))
1665                 goto out;
1666 
1667         success = 1; /* we're going to change ->state */
1668         cpu = task_cpu(p);
1669 
1670         if (p->on_rq && ttwu_remote(p, wake_flags))
1671                 goto stat;
1672 
1673 #ifdef CONFIG_SMP
1674         /*
1675          * If the owning (remote) cpu is still in the middle of schedule() with
1676          * this task as prev, wait until its done referencing the task.
1677          */
1678         while (p->on_cpu)
1679                 cpu_relax();
1680         /*
1681          * Pairs with the smp_wmb() in finish_lock_switch().
1682          */
1683         smp_rmb();
1684 
1685         p->sched_contributes_to_load = !!task_contributes_to_load(p);
1686         p->state = TASK_WAKING;
1687 
1688         if (p->sched_class->task_waking)
1689                 p->sched_class->task_waking(p);
1690 
1691         cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1692         if (task_cpu(p) != cpu) {
1693                 wake_flags |= WF_MIGRATED;
1694                 set_task_cpu(p, cpu);
1695         }
1696 #endif /* CONFIG_SMP */
1697 
1698         ttwu_queue(p, cpu);
1699 stat:
1700         ttwu_stat(p, cpu, wake_flags);
1701 out:
1702         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1703 
1704         return success;
1705 }
1706 
1707 /**
1708  * try_to_wake_up_local - try to wake up a local task with rq lock held
1709  * @p: the thread to be awakened
1710  *
1711  * Put @p on the run-queue if it's not already there. The caller must
1712  * ensure that this_rq() is locked, @p is bound to this_rq() and not
1713  * the current task.
1714  */
1715 static void try_to_wake_up_local(struct task_struct *p)
1716 {
1717         struct rq *rq = task_rq(p);
1718 
1719         if (WARN_ON_ONCE(rq != this_rq()) ||
1720             WARN_ON_ONCE(p == current))
1721                 return;
1722 
1723         lockdep_assert_held(&rq->lock);
1724 
1725         if (!raw_spin_trylock(&p->pi_lock)) {
1726                 raw_spin_unlock(&rq->lock);
1727                 raw_spin_lock(&p->pi_lock);
1728                 raw_spin_lock(&rq->lock);
1729         }
1730 
1731         if (!(p->state & TASK_NORMAL))
1732                 goto out;
1733 
1734         if (!task_on_rq_queued(p))
1735                 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1736 
1737         ttwu_do_wakeup(rq, p, 0);
1738         ttwu_stat(p, smp_processor_id(), 0);
1739 out:
1740         raw_spin_unlock(&p->pi_lock);
1741 }
1742 
1743 /**
1744  * wake_up_process - Wake up a specific process
1745  * @p: The process to be woken up.
1746  *
1747  * Attempt to wake up the nominated process and move it to the set of runnable
1748  * processes.
1749  *
1750  * Return: 1 if the process was woken up, 0 if it was already running.
1751  *
1752  * It may be assumed that this function implies a write memory barrier before
1753  * changing the task state if and only if any tasks are woken up.
1754  */
1755 int wake_up_process(struct task_struct *p)
1756 {
1757         WARN_ON(task_is_stopped_or_traced(p));
1758         return try_to_wake_up(p, TASK_NORMAL, 0);
1759 }
1760 EXPORT_SYMBOL(wake_up_process);
1761 
1762 int wake_up_state(struct task_struct *p, unsigned int state)
1763 {
1764         return try_to_wake_up(p, state, 0);
1765 }
1766 
1767 /*
1768  * This function clears the sched_dl_entity static params.
1769  */
1770 void __dl_clear_params(struct task_struct *p)
1771 {
1772         struct sched_dl_entity *dl_se = &p->dl;
1773 
1774         dl_se->dl_runtime = 0;
1775         dl_se->dl_deadline = 0;
1776         dl_se->dl_period = 0;
1777         dl_se->flags = 0;
1778         dl_se->dl_bw = 0;
1779 
1780         dl_se->dl_throttled = 0;
1781         dl_se->dl_new = 1;
1782         dl_se->dl_yielded = 0;
1783 }
1784 
1785 /*
1786  * Perform scheduler related setup for a newly forked process p.
1787  * p is forked by current.
1788  *
1789  * __sched_fork() is basic setup used by init_idle() too:
1790  */
1791 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1792 {
1793         p->on_rq                        = 0;
1794 
1795         p->se.on_rq                     = 0;
1796         p->se.exec_start                = 0;
1797         p->se.sum_exec_runtime          = 0;
1798         p->se.prev_sum_exec_runtime     = 0;
1799         p->se.nr_migrations             = 0;
1800         p->se.vruntime                  = 0;
1801 #ifdef CONFIG_SMP
1802         p->se.avg.decay_count           = 0;
1803 #endif
1804         INIT_LIST_HEAD(&p->se.group_node);
1805 
1806 #ifdef CONFIG_SCHEDSTATS
1807         memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1808 #endif
1809 
1810         RB_CLEAR_NODE(&p->dl.rb_node);
1811         init_dl_task_timer(&p->dl);
1812         __dl_clear_params(p);
1813 
1814         INIT_LIST_HEAD(&p->rt.run_list);
1815 
1816 #ifdef CONFIG_PREEMPT_NOTIFIERS
1817         INIT_HLIST_HEAD(&p->preempt_notifiers);
1818 #endif
1819 
1820 #ifdef CONFIG_NUMA_BALANCING
1821         if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1822                 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1823                 p->mm->numa_scan_seq = 0;
1824         }
1825 
1826         if (clone_flags & CLONE_VM)
1827                 p->numa_preferred_nid = current->numa_preferred_nid;
1828         else
1829                 p->numa_preferred_nid = -1;
1830 
1831         p->node_stamp = 0ULL;
1832         p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1833         p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1834         p->numa_work.next = &p->numa_work;
1835         p->numa_faults = NULL;
1836         p->last_task_numa_placement = 0;
1837         p->last_sum_exec_runtime = 0;
1838 
1839         p->numa_group = NULL;
1840 #endif /* CONFIG_NUMA_BALANCING */
1841 }
1842 
1843 #ifdef CONFIG_NUMA_BALANCING
1844 #ifdef CONFIG_SCHED_DEBUG
1845 void set_numabalancing_state(bool enabled)
1846 {
1847         if (enabled)
1848                 sched_feat_set("NUMA");
1849         else
1850                 sched_feat_set("NO_NUMA");
1851 }
1852 #else
1853 __read_mostly bool numabalancing_enabled;
1854 
1855 void set_numabalancing_state(bool enabled)
1856 {
1857         numabalancing_enabled = enabled;
1858 }
1859 #endif /* CONFIG_SCHED_DEBUG */
1860 
1861 #ifdef CONFIG_PROC_SYSCTL
1862 int sysctl_numa_balancing(struct ctl_table *table, int write,
1863                          void __user *buffer, size_t *lenp, loff_t *ppos)
1864 {
1865         struct ctl_table t;
1866         int err;
1867         int state = numabalancing_enabled;
1868 
1869         if (write && !capable(CAP_SYS_ADMIN))
1870                 return -EPERM;
1871 
1872         t = *table;
1873         t.data = &state;
1874         err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1875         if (err < 0)
1876                 return err;
1877         if (write)
1878                 set_numabalancing_state(state);
1879         return err;
1880 }
1881 #endif
1882 #endif
1883 
1884 /*
1885  * fork()/clone()-time setup:
1886  */
1887 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1888 {
1889         unsigned long flags;
1890         int cpu = get_cpu();
1891 
1892         __sched_fork(clone_flags, p);
1893         /*
1894          * We mark the process as running here. This guarantees that
1895          * nobody will actually run it, and a signal or other external
1896          * event cannot wake it up and insert it on the runqueue either.
1897          */
1898         p->state = TASK_RUNNING;
1899 
1900         /*
1901          * Make sure we do not leak PI boosting priority to the child.
1902          */
1903         p->prio = current->normal_prio;
1904 
1905         /*
1906          * Revert to default priority/policy on fork if requested.
1907          */
1908         if (unlikely(p->sched_reset_on_fork)) {
1909                 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1910                         p->policy = SCHED_NORMAL;
1911                         p->static_prio = NICE_TO_PRIO(0);
1912                         p->rt_priority = 0;
1913                 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1914                         p->static_prio = NICE_TO_PRIO(0);
1915 
1916                 p->prio = p->normal_prio = __normal_prio(p);
1917                 set_load_weight(p);
1918 
1919                 /*
1920                  * We don't need the reset flag anymore after the fork. It has
1921                  * fulfilled its duty:
1922                  */
1923                 p->sched_reset_on_fork = 0;
1924         }
1925 
1926         if (dl_prio(p->prio)) {
1927                 put_cpu();
1928                 return -EAGAIN;
1929         } else if (rt_prio(p->prio)) {
1930                 p->sched_class = &rt_sched_class;
1931         } else {
1932                 p->sched_class = &fair_sched_class;
1933         }
1934 
1935         if (p->sched_class->task_fork)
1936                 p->sched_class->task_fork(p);
1937 
1938         /*
1939          * The child is not yet in the pid-hash so no cgroup attach races,
1940          * and the cgroup is pinned to this child due to cgroup_fork()
1941          * is ran before sched_fork().
1942          *
1943          * Silence PROVE_RCU.
1944          */
1945         raw_spin_lock_irqsave(&p->pi_lock, flags);
1946         set_task_cpu(p, cpu);
1947         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1948 
1949 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1950         if (likely(sched_info_on()))
1951                 memset(&p->sched_info, 0, sizeof(p->sched_info));
1952 #endif
1953 #if defined(CONFIG_SMP)
1954         p->on_cpu = 0;
1955 #endif
1956         init_task_preempt_count(p);
1957 #ifdef CONFIG_SMP
1958         plist_node_init(&p->pushable_tasks, MAX_PRIO);
1959         RB_CLEAR_NODE(&p->pushable_dl_tasks);
1960 #endif
1961 
1962         put_cpu();
1963         return 0;
1964 }
1965 
1966 unsigned long to_ratio(u64 period, u64 runtime)
1967 {
1968         if (runtime == RUNTIME_INF)
1969                 return 1ULL << 20;
1970 
1971         /*
1972          * Doing this here saves a lot of checks in all
1973          * the calling paths, and returning zero seems
1974          * safe for them anyway.
1975          */
1976         if (period == 0)
1977                 return 0;
1978 
1979         return div64_u64(runtime << 20, period);
1980 }
1981 
1982 #ifdef CONFIG_SMP
1983 inline struct dl_bw *dl_bw_of(int i)
1984 {
1985         rcu_lockdep_assert(rcu_read_lock_sched_held(),
1986                            "sched RCU must be held");
1987         return &cpu_rq(i)->rd->dl_bw;
1988 }
1989 
1990 static inline int dl_bw_cpus(int i)
1991 {
1992         struct root_domain *rd = cpu_rq(i)->rd;
1993         int cpus = 0;
1994 
1995         rcu_lockdep_assert(rcu_read_lock_sched_held(),
1996                            "sched RCU must be held");
1997         for_each_cpu_and(i, rd->span, cpu_active_mask)
1998                 cpus++;
1999 
2000         return cpus;
2001 }
2002 #else
2003 inline struct dl_bw *dl_bw_of(int i)
2004 {
2005         return &cpu_rq(i)->dl.dl_bw;
2006 }
2007 
2008 static inline int dl_bw_cpus(int i)
2009 {
2010         return 1;
2011 }
2012 #endif
2013 
2014 /*
2015  * We must be sure that accepting a new task (or allowing changing the
2016  * parameters of an existing one) is consistent with the bandwidth
2017  * constraints. If yes, this function also accordingly updates the currently
2018  * allocated bandwidth to reflect the new situation.
2019  *
2020  * This function is called while holding p's rq->lock.
2021  *
2022  * XXX we should delay bw change until the task's 0-lag point, see
2023  * __setparam_dl().
2024  */
2025 static int dl_overflow(struct task_struct *p, int policy,
2026                        const struct sched_attr *attr)
2027 {
2028 
2029         struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2030         u64 period = attr->sched_period ?: attr->sched_deadline;
2031         u64 runtime = attr->sched_runtime;
2032         u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2033         int cpus, err = -1;
2034 
2035         if (new_bw == p->dl.dl_bw)
2036                 return 0;
2037 
2038         /*
2039          * Either if a task, enters, leave, or stays -deadline but changes
2040          * its parameters, we may need to update accordingly the total
2041          * allocated bandwidth of the container.
2042          */
2043         raw_spin_lock(&dl_b->lock);
2044         cpus = dl_bw_cpus(task_cpu(p));
2045         if (dl_policy(policy) && !task_has_dl_policy(p) &&
2046             !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2047                 __dl_add(dl_b, new_bw);
2048                 err = 0;
2049         } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2050                    !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2051                 __dl_clear(dl_b, p->dl.dl_bw);
2052                 __dl_add(dl_b, new_bw);
2053                 err = 0;
2054         } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2055                 __dl_clear(dl_b, p->dl.dl_bw);
2056                 err = 0;
2057         }
2058         raw_spin_unlock(&dl_b->lock);
2059 
2060         return err;
2061 }
2062 
2063 extern void init_dl_bw(struct dl_bw *dl_b);
2064 
2065 /*
2066  * wake_up_new_task - wake up a newly created task for the first time.
2067  *
2068  * This function will do some initial scheduler statistics housekeeping
2069  * that must be done for every newly created context, then puts the task
2070  * on the runqueue and wakes it.
2071  */
2072 void wake_up_new_task(struct task_struct *p)
2073 {
2074         unsigned long flags;
2075         struct rq *rq;
2076 
2077         raw_spin_lock_irqsave(&p->pi_lock, flags);
2078 #ifdef CONFIG_SMP
2079         /*
2080          * Fork balancing, do it here and not earlier because:
2081          *  - cpus_allowed can change in the fork path
2082          *  - any previously selected cpu might disappear through hotplug
2083          */
2084         set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2085 #endif
2086 
2087         /* Initialize new task's runnable average */
2088         init_task_runnable_average(p);
2089         rq = __task_rq_lock(p);
2090         activate_task(rq, p, 0);
2091         p->on_rq = TASK_ON_RQ_QUEUED;
2092         trace_sched_wakeup_new(p, true);
2093         check_preempt_curr(rq, p, WF_FORK);
2094 #ifdef CONFIG_SMP
2095         if (p->sched_class->task_woken)
2096                 p->sched_class->task_woken(rq, p);
2097 #endif
2098         task_rq_unlock(rq, p, &flags);
2099 }
2100 
2101 #ifdef CONFIG_PREEMPT_NOTIFIERS
2102 
2103 /**
2104  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2105  * @notifier: notifier struct to register
2106  */
2107 void preempt_notifier_register(struct preempt_notifier *notifier)
2108 {
2109         hlist_add_head(&notifier->link, &current->preempt_notifiers);
2110 }
2111 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2112 
2113 /**
2114  * preempt_notifier_unregister - no longer interested in preemption notifications
2115  * @notifier: notifier struct to unregister
2116  *
2117  * This is safe to call from within a preemption notifier.
2118  */
2119 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2120 {
2121         hlist_del(&notifier->link);
2122 }
2123 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2124 
2125 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2126 {
2127         struct preempt_notifier *notifier;
2128 
2129         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2130                 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2131 }
2132 
2133 static void
2134 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2135                                  struct task_struct *next)
2136 {
2137         struct preempt_notifier *notifier;
2138 
2139         hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2140                 notifier->ops->sched_out(notifier, next);
2141 }
2142 
2143 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2144 
2145 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2146 {
2147 }
2148 
2149 static void
2150 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2151                                  struct task_struct *next)
2152 {
2153 }
2154 
2155 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2156 
2157 /**
2158  * prepare_task_switch - prepare to switch tasks
2159  * @rq: the runqueue preparing to switch
2160  * @prev: the current task that is being switched out
2161  * @next: the task we are going to switch to.
2162  *
2163  * This is called with the rq lock held and interrupts off. It must
2164  * be paired with a subsequent finish_task_switch after the context
2165  * switch.
2166  *
2167  * prepare_task_switch sets up locking and calls architecture specific
2168  * hooks.
2169  */
2170 static inline void
2171 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2172                     struct task_struct *next)
2173 {
2174         trace_sched_switch(prev, next);
2175         sched_info_switch(rq, prev, next);
2176         perf_event_task_sched_out(prev, next);
2177         fire_sched_out_preempt_notifiers(prev, next);
2178         prepare_lock_switch(rq, next);
2179         prepare_arch_switch(next);
2180 }
2181 
2182 /**
2183  * finish_task_switch - clean up after a task-switch
2184  * @prev: the thread we just switched away from.
2185  *
2186  * finish_task_switch must be called after the context switch, paired
2187  * with a prepare_task_switch call before the context switch.
2188  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2189  * and do any other architecture-specific cleanup actions.
2190  *
2191  * Note that we may have delayed dropping an mm in context_switch(). If
2192  * so, we finish that here outside of the runqueue lock. (Doing it
2193  * with the lock held can cause deadlocks; see schedule() for
2194  * details.)
2195  *
2196  * The context switch have flipped the stack from under us and restored the
2197  * local variables which were saved when this task called schedule() in the
2198  * past. prev == current is still correct but we need to recalculate this_rq
2199  * because prev may have moved to another CPU.
2200  */
2201 static struct rq *finish_task_switch(struct task_struct *prev)
2202         __releases(rq->lock)
2203 {
2204         struct rq *rq = this_rq();
2205         struct mm_struct *mm = rq->prev_mm;
2206         long prev_state;
2207 
2208         rq->prev_mm = NULL;
2209 
2210         /*
2211          * A task struct has one reference for the use as "current".
2212          * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2213          * schedule one last time. The schedule call will never return, and
2214          * the scheduled task must drop that reference.
2215          * The test for TASK_DEAD must occur while the runqueue locks are
2216          * still held, otherwise prev could be scheduled on another cpu, die
2217          * there before we look at prev->state, and then the reference would
2218          * be dropped twice.
2219          *              Manfred Spraul <manfred@colorfullife.com>
2220          */
2221         prev_state = prev->state;
2222         vtime_task_switch(prev);
2223         finish_arch_switch(prev);
2224         perf_event_task_sched_in(prev, current);
2225         finish_lock_switch(rq, prev);
2226         finish_arch_post_lock_switch();
2227 
2228         fire_sched_in_preempt_notifiers(current);
2229         if (mm)
2230                 mmdrop(mm);
2231         if (unlikely(prev_state == TASK_DEAD)) {
2232                 if (prev->sched_class->task_dead)
2233                         prev->sched_class->task_dead(prev);
2234 
2235                 /*
2236                  * Remove function-return probe instances associated with this
2237                  * task and put them back on the free list.
2238                  */
2239                 kprobe_flush_task(prev);
2240                 put_task_struct(prev);
2241         }
2242 
2243         tick_nohz_task_switch(current);
2244         return rq;
2245 }
2246 
2247 #ifdef CONFIG_SMP
2248 
2249 /* rq->lock is NOT held, but preemption is disabled */
2250 static inline void post_schedule(struct rq *rq)
2251 {
2252         if (rq->post_schedule) {
2253                 unsigned long flags;
2254 
2255                 raw_spin_lock_irqsave(&rq->lock, flags);
2256                 if (rq->curr->sched_class->post_schedule)
2257                         rq->curr->sched_class->post_schedule(rq);
2258                 raw_spin_unlock_irqrestore(&rq->lock, flags);
2259 
2260                 rq->post_schedule = 0;
2261         }
2262 }
2263 
2264 #else
2265 
2266 static inline void post_schedule(struct rq *rq)
2267 {
2268 }
2269 
2270 #endif
2271 
2272 /**
2273  * schedule_tail - first thing a freshly forked thread must call.
2274  * @prev: the thread we just switched away from.
2275  */
2276 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2277         __releases(rq->lock)
2278 {
2279         struct rq *rq;
2280 
2281         /* finish_task_switch() drops rq->lock and enables preemtion */
2282         preempt_disable();
2283         rq = finish_task_switch(prev);
2284         post_schedule(rq);
2285         preempt_enable();
2286 
2287         if (current->set_child_tid)
2288                 put_user(task_pid_vnr(current), current->set_child_tid);
2289 }
2290 
2291 /*
2292  * context_switch - switch to the new MM and the new thread's register state.
2293  */
2294 static inline struct rq *
2295 context_switch(struct rq *rq, struct task_struct *prev,
2296                struct task_struct *next)
2297 {
2298         struct mm_struct *mm, *oldmm;
2299 
2300         prepare_task_switch(rq, prev, next);
2301 
2302         mm = next->mm;
2303         oldmm = prev->active_mm;
2304         /*
2305          * For paravirt, this is coupled with an exit in switch_to to
2306          * combine the page table reload and the switch backend into
2307          * one hypercall.
2308          */
2309         arch_start_context_switch(prev);
2310 
2311         if (!mm) {
2312                 next->active_mm = oldmm;
2313                 atomic_inc(&oldmm->mm_count);
2314                 enter_lazy_tlb(oldmm, next);
2315         } else
2316                 switch_mm(oldmm, mm, next);
2317 
2318         if (!prev->mm) {
2319                 prev->active_mm = NULL;
2320                 rq->prev_mm = oldmm;
2321         }
2322         /*
2323          * Since the runqueue lock will be released by the next
2324          * task (which is an invalid locking op but in the case
2325          * of the scheduler it's an obvious special-case), so we
2326          * do an early lockdep release here:
2327          */
2328         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2329 
2330         context_tracking_task_switch(prev, next);
2331         /* Here we just switch the register state and the stack. */
2332         switch_to(prev, next, prev);
2333         barrier();
2334 
2335         return finish_task_switch(prev);
2336 }
2337 
2338 /*
2339  * nr_running and nr_context_switches:
2340  *
2341  * externally visible scheduler statistics: current number of runnable
2342  * threads, total number of context switches performed since bootup.
2343  */
2344 unsigned long nr_running(void)
2345 {
2346         unsigned long i, sum = 0;
2347 
2348         for_each_online_cpu(i)
2349                 sum += cpu_rq(i)->nr_running;
2350 
2351         return sum;
2352 }
2353 
2354 /*
2355  * Check if only the current task is running on the cpu.
2356  */
2357 bool single_task_running(void)
2358 {
2359         if (cpu_rq(smp_processor_id())->nr_running == 1)
2360                 return true;
2361         else
2362                 return false;
2363 }
2364 EXPORT_SYMBOL(single_task_running);
2365 
2366 unsigned long long nr_context_switches(void)
2367 {
2368         int i;
2369         unsigned long long sum = 0;
2370 
2371         for_each_possible_cpu(i)
2372                 sum += cpu_rq(i)->nr_switches;
2373 
2374         return sum;
2375 }
2376 
2377 unsigned long nr_iowait(void)
2378 {
2379         unsigned long i, sum = 0;
2380 
2381         for_each_possible_cpu(i)
2382                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2383 
2384         return sum;
2385 }
2386 
2387 unsigned long nr_iowait_cpu(int cpu)
2388 {
2389         struct rq *this = cpu_rq(cpu);
2390         return atomic_read(&this->nr_iowait);
2391 }
2392 
2393 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2394 {
2395         struct rq *this = this_rq();
2396         *nr_waiters = atomic_read(&this->nr_iowait);
2397         *load = this->cpu_load[0];
2398 }
2399 
2400 #ifdef CONFIG_SMP
2401 
2402 /*
2403  * sched_exec - execve() is a valuable balancing opportunity, because at
2404  * this point the task has the smallest effective memory and cache footprint.
2405  */
2406 void sched_exec(void)
2407 {
2408         struct task_struct *p = current;
2409         unsigned long flags;
2410         int dest_cpu;
2411 
2412         raw_spin_lock_irqsave(&p->pi_lock, flags);
2413         dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2414         if (dest_cpu == smp_processor_id())
2415                 goto unlock;
2416 
2417         if (likely(cpu_active(dest_cpu))) {
2418                 struct migration_arg arg = { p, dest_cpu };
2419 
2420                 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2421                 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2422                 return;
2423         }
2424 unlock:
2425         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2426 }
2427 
2428 #endif
2429 
2430 DEFINE_PER_CPU(struct kernel_stat, kstat);
2431 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2432 
2433 EXPORT_PER_CPU_SYMBOL(kstat);
2434 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2435 
2436 /*
2437  * Return accounted runtime for the task.
2438  * In case the task is currently running, return the runtime plus current's
2439  * pending runtime that have not been accounted yet.
2440  */
2441 unsigned long long task_sched_runtime(struct task_struct *p)
2442 {
2443         unsigned long flags;
2444         struct rq *rq;
2445         u64 ns;
2446 
2447 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2448         /*
2449          * 64-bit doesn't need locks to atomically read a 64bit value.
2450          * So we have a optimization chance when the task's delta_exec is 0.
2451          * Reading ->on_cpu is racy, but this is ok.
2452          *
2453          * If we race with it leaving cpu, we'll take a lock. So we're correct.
2454          * If we race with it entering cpu, unaccounted time is 0. This is
2455          * indistinguishable from the read occurring a few cycles earlier.
2456          * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2457          * been accounted, so we're correct here as well.
2458          */
2459         if (!p->on_cpu || !task_on_rq_queued(p))
2460                 return p->se.sum_exec_runtime;
2461 #endif
2462 
2463         rq = task_rq_lock(p, &flags);
2464         /*
2465          * Must be ->curr _and_ ->on_rq.  If dequeued, we would
2466          * project cycles that may never be accounted to this
2467          * thread, breaking clock_gettime().
2468          */
2469         if (task_current(rq, p) && task_on_rq_queued(p)) {
2470                 update_rq_clock(rq);
2471                 p->sched_class->update_curr(rq);
2472         }
2473         ns = p->se.sum_exec_runtime;
2474         task_rq_unlock(rq, p, &flags);
2475 
2476         return ns;
2477 }
2478 
2479 /*
2480  * This function gets called by the timer code, with HZ frequency.
2481  * We call it with interrupts disabled.
2482  */
2483 void scheduler_tick(void)
2484 {
2485         int cpu = smp_processor_id();
2486         struct rq *rq = cpu_rq(cpu);
2487         struct task_struct *curr = rq->curr;
2488 
2489         sched_clock_tick();
2490 
2491         raw_spin_lock(&rq->lock);
2492         update_rq_clock(rq);
2493         curr->sched_class->task_tick(rq, curr, 0);
2494         update_cpu_load_active(rq);
2495         raw_spin_unlock(&rq->lock);
2496 
2497         perf_event_task_tick();
2498 
2499 #ifdef CONFIG_SMP
2500         rq->idle_balance = idle_cpu(cpu);
2501         trigger_load_balance(rq);
2502 #endif
2503         rq_last_tick_reset(rq);
2504 }
2505 
2506 #ifdef CONFIG_NO_HZ_FULL
2507 /**
2508  * scheduler_tick_max_deferment
2509  *
2510  * Keep at least one tick per second when a single
2511  * active task is running because the scheduler doesn't
2512  * yet completely support full dynticks environment.
2513  *
2514  * This makes sure that uptime, CFS vruntime, load
2515  * balancing, etc... continue to move forward, even
2516  * with a very low granularity.
2517  *
2518  * Return: Maximum deferment in nanoseconds.
2519  */
2520 u64 scheduler_tick_max_deferment(void)
2521 {
2522         struct rq *rq = this_rq();
2523         unsigned long next, now = ACCESS_ONCE(jiffies);
2524 
2525         next = rq->last_sched_tick + HZ;
2526 
2527         if (time_before_eq(next, now))
2528                 return 0;
2529 
2530         return jiffies_to_nsecs(next - now);
2531 }
2532 #endif
2533 
2534 notrace unsigned long get_parent_ip(unsigned long addr)
2535 {
2536         if (in_lock_functions(addr)) {
2537                 addr = CALLER_ADDR2;
2538                 if (in_lock_functions(addr))
2539                         addr = CALLER_ADDR3;
2540         }
2541         return addr;
2542 }
2543 
2544 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2545                                 defined(CONFIG_PREEMPT_TRACER))
2546 
2547 void preempt_count_add(int val)
2548 {
2549 #ifdef CONFIG_DEBUG_PREEMPT
2550         /*
2551          * Underflow?
2552          */
2553         if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2554                 return;
2555 #endif
2556         __preempt_count_add(val);
2557 #ifdef CONFIG_DEBUG_PREEMPT
2558         /*
2559          * Spinlock count overflowing soon?
2560          */
2561         DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2562                                 PREEMPT_MASK - 10);
2563 #endif
2564         if (preempt_count() == val) {
2565                 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2566 #ifdef CONFIG_DEBUG_PREEMPT
2567                 current->preempt_disable_ip = ip;
2568 #endif
2569                 trace_preempt_off(CALLER_ADDR0, ip);
2570         }
2571 }
2572 EXPORT_SYMBOL(preempt_count_add);
2573 NOKPROBE_SYMBOL(preempt_count_add);
2574 
2575 void preempt_count_sub(int val)
2576 {
2577 #ifdef CONFIG_DEBUG_PREEMPT
2578         /*
2579          * Underflow?
2580          */
2581         if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2582                 return;
2583         /*
2584          * Is the spinlock portion underflowing?
2585          */
2586         if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2587                         !(preempt_count() & PREEMPT_MASK)))
2588                 return;
2589 #endif
2590 
2591         if (preempt_count() == val)
2592                 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2593         __preempt_count_sub(val);
2594 }
2595 EXPORT_SYMBOL(preempt_count_sub);
2596 NOKPROBE_SYMBOL(preempt_count_sub);
2597 
2598 #endif
2599 
2600 /*
2601  * Print scheduling while atomic bug:
2602  */
2603 static noinline void __schedule_bug(struct task_struct *prev)
2604 {
2605         if (oops_in_progress)
2606                 return;
2607 
2608         printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2609                 prev->comm, prev->pid, preempt_count());
2610 
2611         debug_show_held_locks(prev);
2612         print_modules();
2613         if (irqs_disabled())
2614                 print_irqtrace_events(prev);
2615 #ifdef CONFIG_DEBUG_PREEMPT
2616         if (in_atomic_preempt_off()) {
2617                 pr_err("Preemption disabled at:");
2618                 print_ip_sym(current->preempt_disable_ip);
2619                 pr_cont("\n");
2620         }
2621 #endif
2622         dump_stack();
2623         add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2624 }
2625 
2626 /*
2627  * Various schedule()-time debugging checks and statistics:
2628  */
2629 static inline void schedule_debug(struct task_struct *prev)
2630 {
2631 #ifdef CONFIG_SCHED_STACK_END_CHECK
2632         BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2633 #endif
2634         /*
2635          * Test if we are atomic. Since do_exit() needs to call into
2636          * schedule() atomically, we ignore that path. Otherwise whine
2637          * if we are scheduling when we should not.
2638          */
2639         if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2640                 __schedule_bug(prev);
2641         rcu_sleep_check();
2642 
2643         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2644 
2645         schedstat_inc(this_rq(), sched_count);
2646 }
2647 
2648 /*
2649  * Pick up the highest-prio task:
2650  */
2651 static inline struct task_struct *
2652 pick_next_task(struct rq *rq, struct task_struct *prev)
2653 {
2654         const struct sched_class *class = &fair_sched_class;
2655         struct task_struct *p;
2656 
2657         /*
2658          * Optimization: we know that if all tasks are in
2659          * the fair class we can call that function directly:
2660          */
2661         if (likely(prev->sched_class == class &&
2662                    rq->nr_running == rq->cfs.h_nr_running)) {
2663                 p = fair_sched_class.pick_next_task(rq, prev);
2664                 if (unlikely(p == RETRY_TASK))
2665                         goto again;
2666 
2667                 /* assumes fair_sched_class->next == idle_sched_class */
2668                 if (unlikely(!p))
2669                         p = idle_sched_class.pick_next_task(rq, prev);
2670 
2671                 return p;
2672         }
2673 
2674 again:
2675         for_each_class(class) {
2676                 p = class->pick_next_task(rq, prev);
2677                 if (p) {
2678                         if (unlikely(p == RETRY_TASK))
2679                                 goto again;
2680                         return p;
2681                 }
2682         }
2683 
2684         BUG(); /* the idle class will always have a runnable task */
2685 }
2686 
2687 /*
2688  * __schedule() is the main scheduler function.
2689  *
2690  * The main means of driving the scheduler and thus entering this function are:
2691  *
2692  *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2693  *
2694  *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2695  *      paths. For example, see arch/x86/entry_64.S.
2696  *
2697  *      To drive preemption between tasks, the scheduler sets the flag in timer
2698  *      interrupt handler scheduler_tick().
2699  *
2700  *   3. Wakeups don't really cause entry into schedule(). They add a
2701  *      task to the run-queue and that's it.
2702  *
2703  *      Now, if the new task added to the run-queue preempts the current
2704  *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2705  *      called on the nearest possible occasion:
2706  *
2707  *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2708  *
2709  *         - in syscall or exception context, at the next outmost
2710  *           preempt_enable(). (this might be as soon as the wake_up()'s
2711  *           spin_unlock()!)
2712  *
2713  *         - in IRQ context, return from interrupt-handler to
2714  *           preemptible context
2715  *
2716  *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2717  *         then at the next:
2718  *
2719  *          - cond_resched() call
2720  *          - explicit schedule() call
2721  *          - return from syscall or exception to user-space
2722  *          - return from interrupt-handler to user-space
2723  *
2724  * WARNING: all callers must re-check need_resched() afterward and reschedule
2725  * accordingly in case an event triggered the need for rescheduling (such as
2726  * an interrupt waking up a task) while preemption was disabled in __schedule().
2727  */
2728 static void __sched __schedule(void)
2729 {
2730         struct task_struct *prev, *next;
2731         unsigned long *switch_count;
2732         struct rq *rq;
2733         int cpu;
2734 
2735         preempt_disable();
2736         cpu = smp_processor_id();
2737         rq = cpu_rq(cpu);
2738         rcu_note_context_switch();
2739         prev = rq->curr;
2740 
2741         schedule_debug(prev);
2742 
2743         if (sched_feat(HRTICK))
2744                 hrtick_clear(rq);
2745 
2746         /*
2747          * Make sure that signal_pending_state()->signal_pending() below
2748          * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2749          * done by the caller to avoid the race with signal_wake_up().
2750          */
2751         smp_mb__before_spinlock();
2752         raw_spin_lock_irq(&rq->lock);
2753 
2754         rq->clock_skip_update <<= 1; /* promote REQ to ACT */
2755 
2756         switch_count = &prev->nivcsw;
2757         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2758                 if (unlikely(signal_pending_state(prev->state, prev))) {
2759                         prev->state = TASK_RUNNING;
2760                 } else {
2761                         deactivate_task(rq, prev, DEQUEUE_SLEEP);
2762                         prev->on_rq = 0;
2763 
2764                         /*
2765                          * If a worker went to sleep, notify and ask workqueue
2766                          * whether it wants to wake up a task to maintain
2767                          * concurrency.
2768                          */
2769                         if (prev->flags & PF_WQ_WORKER) {
2770                                 struct task_struct *to_wakeup;
2771 
2772                                 to_wakeup = wq_worker_sleeping(prev, cpu);
2773                                 if (to_wakeup)
2774                                         try_to_wake_up_local(to_wakeup);
2775                         }
2776                 }
2777                 switch_count = &prev->nvcsw;
2778         }
2779 
2780         if (task_on_rq_queued(prev))
2781                 update_rq_clock(rq);
2782 
2783         next = pick_next_task(rq, prev);
2784         clear_tsk_need_resched(prev);
2785         clear_preempt_need_resched();
2786         rq->clock_skip_update = 0;
2787 
2788         if (likely(prev != next)) {
2789                 rq->nr_switches++;
2790                 rq->curr = next;
2791                 ++*switch_count;
2792 
2793                 rq = context_switch(rq, prev, next); /* unlocks the rq */
2794                 cpu = cpu_of(rq);
2795         } else
2796                 raw_spin_unlock_irq(&rq->lock);
2797 
2798         post_schedule(rq);
2799 
2800         sched_preempt_enable_no_resched();
2801 }
2802 
2803 static inline void sched_submit_work(struct task_struct *tsk)
2804 {
2805         if (!tsk->state || tsk_is_pi_blocked(tsk))
2806                 return;
2807         /*
2808          * If we are going to sleep and we have plugged IO queued,
2809          * make sure to submit it to avoid deadlocks.
2810          */
2811         if (blk_needs_flush_plug(tsk))
2812                 blk_schedule_flush_plug(tsk);
2813 }
2814 
2815 asmlinkage __visible void __sched schedule(void)
2816 {
2817         struct task_struct *tsk = current;
2818 
2819         sched_submit_work(tsk);
2820         do {
2821                 __schedule();
2822         } while (need_resched());
2823 }
2824 EXPORT_SYMBOL(schedule);
2825 
2826 #ifdef CONFIG_CONTEXT_TRACKING
2827 asmlinkage __visible void __sched schedule_user(void)
2828 {
2829         /*
2830          * If we come here after a random call to set_need_resched(),
2831          * or we have been woken up remotely but the IPI has not yet arrived,
2832          * we haven't yet exited the RCU idle mode. Do it here manually until
2833          * we find a better solution.
2834          *
2835          * NB: There are buggy callers of this function.  Ideally we
2836          * should warn if prev_state != IN_USER, but that will trigger
2837          * too frequently to make sense yet.
2838          */
2839         enum ctx_state prev_state = exception_enter();
2840         schedule();
2841         exception_exit(prev_state);
2842 }
2843 #endif
2844 
2845 /**
2846  * schedule_preempt_disabled - called with preemption disabled
2847  *
2848  * Returns with preemption disabled. Note: preempt_count must be 1
2849  */
2850 void __sched schedule_preempt_disabled(void)
2851 {
2852         sched_preempt_enable_no_resched();
2853         schedule();
2854         preempt_disable();
2855 }
2856 
2857 static void __sched notrace preempt_schedule_common(void)
2858 {
2859         do {
2860                 __preempt_count_add(PREEMPT_ACTIVE);
2861                 __schedule();
2862                 __preempt_count_sub(PREEMPT_ACTIVE);
2863 
2864                 /*
2865                  * Check again in case we missed a preemption opportunity
2866                  * between schedule and now.
2867                  */
2868                 barrier();
2869         } while (need_resched());
2870 }
2871 
2872 #ifdef CONFIG_PREEMPT
2873 /*
2874  * this is the entry point to schedule() from in-kernel preemption
2875  * off of preempt_enable. Kernel preemptions off return from interrupt
2876  * occur there and call schedule directly.
2877  */
2878 asmlinkage __visible void __sched notrace preempt_schedule(void)
2879 {
2880         /*
2881          * If there is a non-zero preempt_count or interrupts are disabled,
2882          * we do not want to preempt the current task. Just return..
2883          */
2884         if (likely(!preemptible()))
2885                 return;
2886 
2887         preempt_schedule_common();
2888 }
2889 NOKPROBE_SYMBOL(preempt_schedule);
2890 EXPORT_SYMBOL(preempt_schedule);
2891 
2892 #ifdef CONFIG_CONTEXT_TRACKING
2893 /**
2894  * preempt_schedule_context - preempt_schedule called by tracing
2895  *
2896  * The tracing infrastructure uses preempt_enable_notrace to prevent
2897  * recursion and tracing preempt enabling caused by the tracing
2898  * infrastructure itself. But as tracing can happen in areas coming
2899  * from userspace or just about to enter userspace, a preempt enable
2900  * can occur before user_exit() is called. This will cause the scheduler
2901  * to be called when the system is still in usermode.
2902  *
2903  * To prevent this, the preempt_enable_notrace will use this function
2904  * instead of preempt_schedule() to exit user context if needed before
2905  * calling the scheduler.
2906  */
2907 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2908 {
2909         enum ctx_state prev_ctx;
2910 
2911         if (likely(!preemptible()))
2912                 return;
2913 
2914         do {
2915                 __preempt_count_add(PREEMPT_ACTIVE);
2916                 /*
2917                  * Needs preempt disabled in case user_exit() is traced
2918                  * and the tracer calls preempt_enable_notrace() causing
2919                  * an infinite recursion.
2920                  */
2921                 prev_ctx = exception_enter();
2922                 __schedule();
2923                 exception_exit(prev_ctx);
2924 
2925                 __preempt_count_sub(PREEMPT_ACTIVE);
2926                 barrier();
2927         } while (need_resched());
2928 }
2929 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2930 #endif /* CONFIG_CONTEXT_TRACKING */
2931 
2932 #endif /* CONFIG_PREEMPT */
2933 
2934 /*
2935  * this is the entry point to schedule() from kernel preemption
2936  * off of irq context.
2937  * Note, that this is called and return with irqs disabled. This will
2938  * protect us against recursive calling from irq.
2939  */
2940 asmlinkage __visible void __sched preempt_schedule_irq(void)
2941 {
2942         enum ctx_state prev_state;
2943 
2944         /* Catch callers which need to be fixed */
2945         BUG_ON(preempt_count() || !irqs_disabled());
2946 
2947         prev_state = exception_enter();
2948 
2949         do {
2950                 __preempt_count_add(PREEMPT_ACTIVE);
2951                 local_irq_enable();
2952                 __schedule();
2953                 local_irq_disable();
2954                 __preempt_count_sub(PREEMPT_ACTIVE);
2955 
2956                 /*
2957                  * Check again in case we missed a preemption opportunity
2958                  * between schedule and now.
2959                  */
2960                 barrier();
2961         } while (need_resched());
2962 
2963         exception_exit(prev_state);
2964 }
2965 
2966 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2967                           void *key)
2968 {
2969         return try_to_wake_up(curr->private, mode, wake_flags);
2970 }
2971 EXPORT_SYMBOL(default_wake_function);
2972 
2973 #ifdef CONFIG_RT_MUTEXES
2974 
2975 /*
2976  * rt_mutex_setprio - set the current priority of a task
2977  * @p: task
2978  * @prio: prio value (kernel-internal form)
2979  *
2980  * This function changes the 'effective' priority of a task. It does
2981  * not touch ->normal_prio like __setscheduler().
2982  *
2983  * Used by the rt_mutex code to implement priority inheritance
2984  * logic. Call site only calls if the priority of the task changed.
2985  */
2986 void rt_mutex_setprio(struct task_struct *p, int prio)
2987 {
2988         int oldprio, queued, running, enqueue_flag = 0;
2989         struct rq *rq;
2990         const struct sched_class *prev_class;
2991 
2992         BUG_ON(prio > MAX_PRIO);
2993 
2994         rq = __task_rq_lock(p);
2995 
2996         /*
2997          * Idle task boosting is a nono in general. There is one
2998          * exception, when PREEMPT_RT and NOHZ is active:
2999          *
3000          * The idle task calls get_next_timer_interrupt() and holds
3001          * the timer wheel base->lock on the CPU and another CPU wants
3002          * to access the timer (probably to cancel it). We can safely
3003          * ignore the boosting request, as the idle CPU runs this code
3004          * with interrupts disabled and will complete the lock
3005          * protected section without being interrupted. So there is no
3006          * real need to boost.
3007          */
3008         if (unlikely(p == rq->idle)) {
3009                 WARN_ON(p != rq->curr);
3010                 WARN_ON(p->pi_blocked_on);
3011                 goto out_unlock;
3012         }
3013 
3014         trace_sched_pi_setprio(p, prio);
3015         oldprio = p->prio;
3016         prev_class = p->sched_class;
3017         queued = task_on_rq_queued(p);
3018         running = task_current(rq, p);
3019         if (queued)
3020                 dequeue_task(rq, p, 0);
3021         if (running)
3022                 put_prev_task(rq, p);
3023 
3024         /*
3025          * Boosting condition are:
3026          * 1. -rt task is running and holds mutex A
3027          *      --> -dl task blocks on mutex A
3028          *
3029          * 2. -dl task is running and holds mutex A
3030          *      --> -dl task blocks on mutex A and could preempt the
3031          *          running task
3032          */
3033         if (dl_prio(prio)) {
3034                 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3035                 if (!dl_prio(p->normal_prio) ||
3036                     (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3037                         p->dl.dl_boosted = 1;
3038                         p->dl.dl_throttled = 0;
3039                         enqueue_flag = ENQUEUE_REPLENISH;
3040                 } else
3041                         p->dl.dl_boosted = 0;
3042                 p->sched_class = &dl_sched_class;
3043         } else if (rt_prio(prio)) {
3044                 if (dl_prio(oldprio))
3045                         p->dl.dl_boosted = 0;
3046                 if (oldprio < prio)
3047                         enqueue_flag = ENQUEUE_HEAD;
3048                 p->sched_class = &rt_sched_class;
3049         } else {
3050                 if (dl_prio(oldprio))
3051                         p->dl.dl_boosted = 0;
3052                 if (rt_prio(oldprio))
3053                         p->rt.timeout = 0;
3054                 p->sched_class = &fair_sched_class;
3055         }
3056 
3057         p->prio = prio;
3058 
3059         if (running)
3060                 p->sched_class->set_curr_task(rq);
3061         if (queued)
3062                 enqueue_task(rq, p, enqueue_flag);
3063 
3064         check_class_changed(rq, p, prev_class, oldprio);
3065 out_unlock:
3066         __task_rq_unlock(rq);
3067 }
3068 #endif
3069 
3070 void set_user_nice(struct task_struct *p, long nice)
3071 {
3072         int old_prio, delta, queued;
3073         unsigned long flags;
3074         struct rq *rq;
3075 
3076         if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3077                 return;
3078         /*
3079          * We have to be careful, if called from sys_setpriority(),
3080          * the task might be in the middle of scheduling on another CPU.
3081          */
3082         rq = task_rq_lock(p, &flags);
3083         /*
3084          * The RT priorities are set via sched_setscheduler(), but we still
3085          * allow the 'normal' nice value to be set - but as expected
3086          * it wont have any effect on scheduling until the task is
3087          * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3088          */
3089         if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3090                 p->static_prio = NICE_TO_PRIO(nice);
3091                 goto out_unlock;
3092         }
3093         queued = task_on_rq_queued(p);
3094         if (queued)
3095                 dequeue_task(rq, p, 0);
3096 
3097         p->static_prio = NICE_TO_PRIO(nice);
3098         set_load_weight(p);
3099         old_prio = p->prio;
3100         p->prio = effective_prio(p);
3101         delta = p->prio - old_prio;
3102 
3103         if (queued) {
3104                 enqueue_task(rq, p, 0);
3105                 /*
3106                  * If the task increased its priority or is running and
3107                  * lowered its priority, then reschedule its CPU:
3108                  */
3109                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3110                         resched_curr(rq);
3111         }
3112 out_unlock:
3113         task_rq_unlock(rq, p, &flags);
3114 }
3115 EXPORT_SYMBOL(set_user_nice);
3116 
3117 /*
3118  * can_nice - check if a task can reduce its nice value
3119  * @p: task
3120  * @nice: nice value
3121  */
3122 int can_nice(const struct task_struct *p, const int nice)
3123 {
3124         /* convert nice value [19,-20] to rlimit style value [1,40] */
3125         int nice_rlim = nice_to_rlimit(nice);
3126 
3127         return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3128                 capable(CAP_SYS_NICE));
3129 }
3130 
3131 #ifdef __ARCH_WANT_SYS_NICE
3132 
3133 /*
3134  * sys_nice - change the priority of the current process.
3135  * @increment: priority increment
3136  *
3137  * sys_setpriority is a more generic, but much slower function that
3138  * does similar things.
3139  */
3140 SYSCALL_DEFINE1(nice, int, increment)
3141 {
3142         long nice, retval;
3143         if (!ccs_capable(CCS_SYS_NICE))
3144                 return -EPERM;
3145 
3146         /*
3147          * Setpriority might change our priority at the same moment.
3148          * We don't have to worry. Conceptually one call occurs first
3149          * and we have a single winner.
3150          */
3151         increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3152         nice = task_nice(current) + increment;
3153 
3154         nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3155         if (increment < 0 && !can_nice(current, nice))
3156                 return -EPERM;
3157 
3158         retval = security_task_setnice(current, nice);
3159         if (retval)
3160                 return retval;
3161 
3162         set_user_nice(current, nice);
3163         return 0;
3164 }
3165 
3166 #endif
3167 
3168 /**
3169  * task_prio - return the priority value of a given task.
3170  * @p: the task in question.
3171  *
3172  * Return: The priority value as seen by users in /proc.
3173  * RT tasks are offset by -200. Normal tasks are centered
3174  * around 0, value goes from -16 to +15.
3175  */
3176 int task_prio(const struct task_struct *p)
3177 {
3178         return p->prio - MAX_RT_PRIO;
3179 }
3180 
3181 /**
3182  * idle_cpu - is a given cpu idle currently?
3183  * @cpu: the processor in question.
3184  *
3185  * Return: 1 if the CPU is currently idle. 0 otherwise.
3186  */
3187 int idle_cpu(int cpu)
3188 {
3189         struct rq *rq = cpu_rq(cpu);
3190 
3191         if (rq->curr != rq->idle)
3192                 return 0;
3193 
3194         if (rq->nr_running)
3195                 return 0;
3196 
3197 #ifdef CONFIG_SMP
3198         if (!llist_empty(&rq->wake_list))
3199                 return 0;
3200 #endif
3201 
3202         return 1;
3203 }
3204 
3205 /**
3206  * idle_task - return the idle task for a given cpu.
3207  * @cpu: the processor in question.
3208  *
3209  * Return: The idle task for the cpu @cpu.
3210  */
3211 struct task_struct *idle_task(int cpu)
3212 {
3213         return cpu_rq(cpu)->idle;
3214 }
3215 
3216 /**
3217  * find_process_by_pid - find a process with a matching PID value.
3218  * @pid: the pid in question.
3219  *
3220  * The task of @pid, if found. %NULL otherwise.
3221  */
3222 static struct task_struct *find_process_by_pid(pid_t pid)
3223 {
3224         return pid ? find_task_by_vpid(pid) : current;
3225 }
3226 
3227 /*
3228  * This function initializes the sched_dl_entity of a newly becoming
3229  * SCHED_DEADLINE task.
3230  *
3231  * Only the static values are considered here, the actual runtime and the
3232  * absolute deadline will be properly calculated when the task is enqueued
3233  * for the first time with its new policy.
3234  */
3235 static void
3236 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3237 {
3238         struct sched_dl_entity *dl_se = &p->dl;
3239 
3240         dl_se->dl_runtime = attr->sched_runtime;
3241         dl_se->dl_deadline = attr->sched_deadline;
3242         dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3243         dl_se->flags = attr->sched_flags;
3244         dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3245 
3246         /*
3247          * Changing the parameters of a task is 'tricky' and we're not doing
3248          * the correct thing -- also see task_dead_dl() and switched_from_dl().
3249          *
3250          * What we SHOULD do is delay the bandwidth release until the 0-lag
3251          * point. This would include retaining the task_struct until that time
3252          * and change dl_overflow() to not immediately decrement the current
3253          * amount.
3254          *
3255          * Instead we retain the current runtime/deadline and let the new
3256          * parameters take effect after the current reservation period lapses.
3257          * This is safe (albeit pessimistic) because the 0-lag point is always
3258          * before the current scheduling deadline.
3259          *
3260          * We can still have temporary overloads because we do not delay the
3261          * change in bandwidth until that time; so admission control is
3262          * not on the safe side. It does however guarantee tasks will never
3263          * consume more than promised.
3264          */
3265 }
3266 
3267 /*
3268  * sched_setparam() passes in -1 for its policy, to let the functions
3269  * it calls know not to change it.
3270  */
3271 #define SETPARAM_POLICY -1
3272 
3273 static void __setscheduler_params(struct task_struct *p,
3274                 const struct sched_attr *attr)
3275 {
3276         int policy = attr->sched_policy;
3277 
3278         if (policy == SETPARAM_POLICY)
3279                 policy = p->policy;
3280 
3281         p->policy = policy;
3282 
3283         if (dl_policy(policy))
3284                 __setparam_dl(p, attr);
3285         else if (fair_policy(policy))
3286                 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3287 
3288         /*
3289          * __sched_setscheduler() ensures attr->sched_priority == 0 when
3290          * !rt_policy. Always setting this ensures that things like
3291          * getparam()/getattr() don't report silly values for !rt tasks.
3292          */
3293         p->rt_priority = attr->sched_priority;
3294         p->normal_prio = normal_prio(p);
3295         set_load_weight(p);
3296 }
3297 
3298 /* Actually do priority change: must hold pi & rq lock. */
3299 static void __setscheduler(struct rq *rq, struct task_struct *p,
3300                            const struct sched_attr *attr, bool keep_boost)
3301 {
3302         __setscheduler_params(p, attr);
3303 
3304         /*
3305          * Keep a potential priority boosting if called from
3306          * sched_setscheduler().
3307          */
3308         if (keep_boost)
3309                 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3310         else
3311                 p->prio = normal_prio(p);
3312 
3313         if (dl_prio(p->prio))
3314                 p->sched_class = &dl_sched_class;
3315         else if (rt_prio(p->prio))
3316                 p->sched_class = &rt_sched_class;
3317         else
3318                 p->sched_class = &fair_sched_class;
3319 }
3320 
3321 static void
3322 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3323 {
3324         struct sched_dl_entity *dl_se = &p->dl;
3325 
3326         attr->sched_priority = p->rt_priority;
3327         attr->sched_runtime = dl_se->dl_runtime;
3328         attr->sched_deadline = dl_se->dl_deadline;
3329         attr->sched_period = dl_se->dl_period;
3330         attr->sched_flags = dl_se->flags;
3331 }
3332 
3333 /*
3334  * This function validates the new parameters of a -deadline task.
3335  * We ask for the deadline not being zero, and greater or equal
3336  * than the runtime, as well as the period of being zero or
3337  * greater than deadline. Furthermore, we have to be sure that
3338  * user parameters are above the internal resolution of 1us (we
3339  * check sched_runtime only since it is always the smaller one) and
3340  * below 2^63 ns (we have to check both sched_deadline and
3341  * sched_period, as the latter can be zero).
3342  */
3343 static bool
3344 __checkparam_dl(const struct sched_attr *attr)
3345 {
3346         /* deadline != 0 */
3347         if (attr->sched_deadline == 0)
3348                 return false;
3349 
3350         /*
3351          * Since we truncate DL_SCALE bits, make sure we're at least
3352          * that big.
3353          */
3354         if (attr->sched_runtime < (1ULL << DL_SCALE))
3355                 return false;
3356 
3357         /*
3358          * Since we use the MSB for wrap-around and sign issues, make
3359          * sure it's not set (mind that period can be equal to zero).
3360          */
3361         if (attr->sched_deadline & (1ULL << 63) ||
3362             attr->sched_period & (1ULL << 63))
3363                 return false;
3364 
3365         /* runtime <= deadline <= period (if period != 0) */
3366         if ((attr->sched_period != 0 &&
3367              attr->sched_period < attr->sched_deadline) ||
3368             attr->sched_deadline < attr->sched_runtime)
3369                 return false;
3370 
3371         return true;
3372 }
3373 
3374 /*
3375  * check the target process has a UID that matches the current process's
3376  */
3377 static bool check_same_owner(struct task_struct *p)
3378 {
3379         const struct cred *cred = current_cred(), *pcred;
3380         bool match;
3381 
3382         rcu_read_lock();
3383         pcred = __task_cred(p);
3384         match = (uid_eq(cred->euid, pcred->euid) ||
3385                  uid_eq(cred->euid, pcred->uid));
3386         rcu_read_unlock();
3387         return match;
3388 }
3389 
3390 static bool dl_param_changed(struct task_struct *p,
3391                 const struct sched_attr *attr)
3392 {
3393         struct sched_dl_entity *dl_se = &p->dl;
3394 
3395         if (dl_se->dl_runtime != attr->sched_runtime ||
3396                 dl_se->dl_deadline != attr->sched_deadline ||
3397                 dl_se->dl_period != attr->sched_period ||
3398                 dl_se->flags != attr->sched_flags)
3399                 return true;
3400 
3401         return false;
3402 }
3403 
3404 static int __sched_setscheduler(struct task_struct *p,
3405                                 const struct sched_attr *attr,
3406                                 bool user)
3407 {
3408         int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3409                       MAX_RT_PRIO - 1 - attr->sched_priority;
3410         int retval, oldprio, oldpolicy = -1, queued, running;
3411         int new_effective_prio, policy = attr->sched_policy;
3412         unsigned long flags;
3413         const struct sched_class *prev_class;
3414         struct rq *rq;
3415         int reset_on_fork;
3416 
3417         /* may grab non-irq protected spin_locks */
3418         BUG_ON(in_interrupt());
3419 recheck:
3420         /* double check policy once rq lock held */
3421         if (policy < 0) {
3422                 reset_on_fork = p->sched_reset_on_fork;
3423                 policy = oldpolicy = p->policy;
3424         } else {
3425                 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3426 
3427                 if (policy != SCHED_DEADLINE &&
3428                                 policy != SCHED_FIFO && policy != SCHED_RR &&
3429                                 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3430                                 policy != SCHED_IDLE)
3431                         return -EINVAL;
3432         }
3433 
3434         if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3435                 return -EINVAL;
3436 
3437         /*
3438          * Valid priorities for SCHED_FIFO and SCHED_RR are
3439          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3440          * SCHED_BATCH and SCHED_IDLE is 0.
3441          */
3442         if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3443             (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3444                 return -EINVAL;
3445         if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3446             (rt_policy(policy) != (attr->sched_priority != 0)))
3447                 return -EINVAL;
3448 
3449         /*
3450          * Allow unprivileged RT tasks to decrease priority:
3451          */
3452         if (user && !capable(CAP_SYS_NICE)) {
3453                 if (fair_policy(policy)) {
3454                         if (attr->sched_nice < task_nice(p) &&
3455                             !can_nice(p, attr->sched_nice))
3456                                 return -EPERM;
3457                 }
3458 
3459                 if (rt_policy(policy)) {
3460                         unsigned long rlim_rtprio =
3461                                         task_rlimit(p, RLIMIT_RTPRIO);
3462 
3463                         /* can't set/change the rt policy */
3464                         if (policy != p->policy && !rlim_rtprio)
3465                                 return -EPERM;
3466 
3467                         /* can't increase priority */
3468                         if (attr->sched_priority > p->rt_priority &&
3469                             attr->sched_priority > rlim_rtprio)
3470                                 return -EPERM;
3471                 }
3472 
3473                  /*
3474                   * Can't set/change SCHED_DEADLINE policy at all for now
3475                   * (safest behavior); in the future we would like to allow
3476                   * unprivileged DL tasks to increase their relative deadline
3477                   * or reduce their runtime (both ways reducing utilization)
3478                   */
3479                 if (dl_policy(policy))
3480                         return -EPERM;
3481 
3482                 /*
3483                  * Treat SCHED_IDLE as nice 20. Only allow a switch to
3484                  * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3485                  */
3486                 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3487                         if (!can_nice(p, task_nice(p)))
3488                                 return -EPERM;
3489                 }
3490 
3491                 /* can't change other user's priorities */
3492                 if (!check_same_owner(p))
3493                         return -EPERM;
3494 
3495                 /* Normal users shall not reset the sched_reset_on_fork flag */
3496                 if (p->sched_reset_on_fork && !reset_on_fork)
3497                         return -EPERM;
3498         }
3499 
3500         if (user) {
3501                 retval = security_task_setscheduler(p);
3502                 if (retval)
3503                         return retval;
3504         }
3505 
3506         /*
3507          * make sure no PI-waiters arrive (or leave) while we are
3508          * changing the priority of the task:
3509          *
3510          * To be able to change p->policy safely, the appropriate
3511          * runqueue lock must be held.
3512          */
3513         rq = task_rq_lock(p, &flags);
3514 
3515         /*
3516          * Changing the policy of the stop threads its a very bad idea
3517          */
3518         if (p == rq->stop) {
3519                 task_rq_unlock(rq, p, &flags);
3520                 return -EINVAL;
3521         }
3522 
3523         /*
3524          * If not changing anything there's no need to proceed further,
3525          * but store a possible modification of reset_on_fork.
3526          */
3527         if (unlikely(policy == p->policy)) {
3528                 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3529                         goto change;
3530                 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3531                         goto change;
3532                 if (dl_policy(policy) && dl_param_changed(p, attr))
3533                         goto change;
3534 
3535                 p->sched_reset_on_fork = reset_on_fork;
3536                 task_rq_unlock(rq, p, &flags);
3537                 return 0;
3538         }
3539 change:
3540 
3541         if (user) {
3542 #ifdef CONFIG_RT_GROUP_SCHED
3543                 /*
3544                  * Do not allow realtime tasks into groups that have no runtime
3545                  * assigned.
3546                  */
3547                 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3548                                 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3549                                 !task_group_is_autogroup(task_group(p))) {
3550                         task_rq_unlock(rq, p, &flags);
3551                         return -EPERM;
3552                 }
3553 #endif
3554 #ifdef CONFIG_SMP
3555                 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3556                         cpumask_t *span = rq->rd->span;
3557 
3558                         /*
3559                          * Don't allow tasks with an affinity mask smaller than
3560                          * the entire root_domain to become SCHED_DEADLINE. We
3561                          * will also fail if there's no bandwidth available.
3562                          */
3563                         if (!cpumask_subset(span, &p->cpus_allowed) ||
3564                             rq->rd->dl_bw.bw == 0) {
3565                                 task_rq_unlock(rq, p, &flags);
3566                                 return -EPERM;
3567                         }
3568                 }
3569 #endif
3570         }
3571 
3572         /* recheck policy now with rq lock held */
3573         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3574                 policy = oldpolicy = -1;
3575                 task_rq_unlock(rq, p, &flags);
3576                 goto recheck;
3577         }
3578 
3579         /*
3580          * If setscheduling to SCHED_DEADLINE (or changing the parameters
3581          * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3582          * is available.
3583          */
3584         if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3585                 task_rq_unlock(rq, p, &flags);
3586                 return -EBUSY;
3587         }
3588 
3589         p->sched_reset_on_fork = reset_on_fork;
3590         oldprio = p->prio;
3591 
3592         /*
3593          * Take priority boosted tasks into account. If the new
3594          * effective priority is unchanged, we just store the new
3595          * normal parameters and do not touch the scheduler class and
3596          * the runqueue. This will be done when the task deboost
3597          * itself.
3598          */
3599         new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3600         if (new_effective_prio == oldprio) {
3601                 __setscheduler_params(p, attr);
3602                 task_rq_unlock(rq, p, &flags);
3603                 return 0;
3604         }
3605 
3606         queued = task_on_rq_queued(p);
3607         running = task_current(rq, p);
3608         if (queued)
3609                 dequeue_task(rq, p, 0);
3610         if (running)
3611                 put_prev_task(rq, p);
3612 
3613         prev_class = p->sched_class;
3614         __setscheduler(rq, p, attr, true);
3615 
3616         if (running)
3617                 p->sched_class->set_curr_task(rq);
3618         if (queued) {
3619                 /*
3620                  * We enqueue to tail when the priority of a task is
3621                  * increased (user space view).
3622                  */
3623                 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3624         }
3625 
3626         check_class_changed(rq, p, prev_class, oldprio);
3627         task_rq_unlock(rq, p, &flags);
3628 
3629         rt_mutex_adjust_pi(p);
3630 
3631         return 0;
3632 }
3633 
3634 static int _sched_setscheduler(struct task_struct *p, int policy,
3635                                const struct sched_param *param, bool check)
3636 {
3637         struct sched_attr attr = {
3638                 .sched_policy   = policy,
3639                 .sched_priority = param->sched_priority,
3640                 .sched_nice     = PRIO_TO_NICE(p->static_prio),
3641         };
3642 
3643         /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3644         if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3645                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3646                 policy &= ~SCHED_RESET_ON_FORK;
3647                 attr.sched_policy = policy;
3648         }
3649 
3650         return __sched_setscheduler(p, &attr, check);
3651 }
3652 /**
3653  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3654  * @p: the task in question.
3655  * @policy: new policy.
3656  * @param: structure containing the new RT priority.
3657  *
3658  * Return: 0 on success. An error code otherwise.
3659  *
3660  * NOTE that the task may be already dead.
3661  */
3662 int sched_setscheduler(struct task_struct *p, int policy,
3663                        const struct sched_param *param)
3664 {
3665         return _sched_setscheduler(p, policy, param, true);
3666 }
3667 EXPORT_SYMBOL_GPL(sched_setscheduler);
3668 
3669 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3670 {
3671         return __sched_setscheduler(p, attr, true);
3672 }
3673 EXPORT_SYMBOL_GPL(sched_setattr);
3674 
3675 /**
3676  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3677  * @p: the task in question.
3678  * @policy: new policy.
3679  * @param: structure containing the new RT priority.
3680  *
3681  * Just like sched_setscheduler, only don't bother checking if the
3682  * current context has permission.  For example, this is needed in
3683  * stop_machine(): we create temporary high priority worker threads,
3684  * but our caller might not have that capability.
3685  *
3686  * Return: 0 on success. An error code otherwise.
3687  */
3688 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3689                                const struct sched_param *param)
3690 {
3691         return _sched_setscheduler(p, policy, param, false);
3692 }
3693 
3694 static int
3695 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3696 {
3697         struct sched_param lparam;
3698         struct task_struct *p;
3699         int retval;
3700 
3701         if (!param || pid < 0)
3702                 return -EINVAL;
3703         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3704                 return -EFAULT;
3705 
3706         rcu_read_lock();
3707         retval = -ESRCH;
3708         p = find_process_by_pid(pid);
3709         if (p != NULL)
3710                 retval = sched_setscheduler(p, policy, &lparam);
3711         rcu_read_unlock();
3712 
3713         return retval;
3714 }
3715 
3716 /*
3717  * Mimics kernel/events/core.c perf_copy_attr().
3718  */
3719 static int sched_copy_attr(struct sched_attr __user *uattr,
3720                            struct sched_attr *attr)
3721 {
3722         u32 size;
3723         int ret;
3724 
3725         if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3726                 return -EFAULT;
3727 
3728         /*
3729          * zero the full structure, so that a short copy will be nice.
3730          */
3731         memset(attr, 0, sizeof(*attr));
3732 
3733         ret = get_user(size, &uattr->size);
3734         if (ret)
3735                 return ret;
3736 
3737         if (size > PAGE_SIZE)   /* silly large */
3738                 goto err_size;
3739 
3740         if (!size)              /* abi compat */
3741                 size = SCHED_ATTR_SIZE_VER0;
3742 
3743         if (size < SCHED_ATTR_SIZE_VER0)
3744                 goto err_size;
3745 
3746         /*
3747          * If we're handed a bigger struct than we know of,
3748          * ensure all the unknown bits are 0 - i.e. new
3749          * user-space does not rely on any kernel feature
3750          * extensions we dont know about yet.
3751          */
3752         if (size > sizeof(*attr)) {
3753                 unsigned char __user *addr;
3754                 unsigned char __user *end;
3755                 unsigned char val;
3756 
3757                 addr = (void __user *)uattr + sizeof(*attr);
3758                 end  = (void __user *)uattr + size;
3759 
3760                 for (; addr < end; addr++) {
3761                         ret = get_user(val, addr);
3762                         if (ret)
3763                                 return ret;
3764                         if (val)
3765                                 goto err_size;
3766                 }
3767                 size = sizeof(*attr);
3768         }
3769 
3770         ret = copy_from_user(attr, uattr, size);
3771         if (ret)
3772                 return -EFAULT;
3773 
3774         /*
3775          * XXX: do we want to be lenient like existing syscalls; or do we want
3776          * to be strict and return an error on out-of-bounds values?
3777          */
3778         attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3779 
3780         return 0;
3781 
3782 err_size:
3783         put_user(sizeof(*attr), &uattr->size);
3784         return -E2BIG;
3785 }
3786 
3787 /**
3788  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3789  * @pid: the pid in question.
3790  * @policy: new policy.
3791  * @param: structure containing the new RT priority.
3792  *
3793  * Return: 0 on success. An error code otherwise.
3794  */
3795 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3796                 struct sched_param __user *, param)
3797 {
3798         /* negative values for policy are not valid */
3799         if (policy < 0)
3800                 return -EINVAL;
3801 
3802         return do_sched_setscheduler(pid, policy, param);
3803 }
3804 
3805 /**
3806  * sys_sched_setparam - set/change the RT priority of a thread
3807  * @pid: the pid in question.
3808  * @param: structure containing the new RT priority.
3809  *
3810  * Return: 0 on success. An error code otherwise.
3811  */
3812 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3813 {
3814         return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3815 }
3816 
3817 /**
3818  * sys_sched_setattr - same as above, but with extended sched_attr
3819  * @pid: the pid in question.
3820  * @uattr: structure containing the extended parameters.
3821  * @flags: for future extension.
3822  */
3823 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3824                                unsigned int, flags)
3825 {
3826         struct sched_attr attr;
3827         struct task_struct *p;
3828         int retval;
3829 
3830         if (!uattr || pid < 0 || flags)
3831                 return -EINVAL;
3832 
3833         retval = sched_copy_attr(uattr, &attr);
3834         if (retval)
3835                 return retval;
3836 
3837         if ((int)attr.sched_policy < 0)
3838                 return -EINVAL;
3839 
3840         rcu_read_lock();
3841         retval = -ESRCH;
3842         p = find_process_by_pid(pid);
3843         if (p != NULL)
3844                 retval = sched_setattr(p, &attr);
3845         rcu_read_unlock();
3846 
3847         return retval;
3848 }
3849 
3850 /**
3851  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3852  * @pid: the pid in question.
3853  *
3854  * Return: On success, the policy of the thread. Otherwise, a negative error
3855  * code.
3856  */
3857 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3858 {
3859         struct task_struct *p;
3860         int retval;
3861 
3862         if (pid < 0)
3863                 return -EINVAL;
3864 
3865         retval = -ESRCH;
3866         rcu_read_lock();
3867         p = find_process_by_pid(pid);
3868         if (p) {
3869                 retval = security_task_getscheduler(p);
3870                 if (!retval)
3871                         retval = p->policy
3872                                 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3873         }
3874         rcu_read_unlock();
3875         return retval;
3876 }
3877 
3878 /**
3879  * sys_sched_getparam - get the RT priority of a thread
3880  * @pid: the pid in question.
3881  * @param: structure containing the RT priority.
3882  *
3883  * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3884  * code.
3885  */
3886 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3887 {
3888         struct sched_param lp = { .sched_priority = 0 };
3889         struct task_struct *p;
3890         int retval;
3891 
3892         if (!param || pid < 0)
3893                 return -EINVAL;
3894 
3895         rcu_read_lock();
3896         p = find_process_by_pid(pid);
3897         retval = -ESRCH;
3898         if (!p)
3899                 goto out_unlock;
3900 
3901         retval = security_task_getscheduler(p);
3902         if (retval)
3903                 goto out_unlock;
3904 
3905         if (task_has_rt_policy(p))
3906                 lp.sched_priority = p->rt_priority;
3907         rcu_read_unlock();
3908 
3909         /*
3910          * This one might sleep, we cannot do it with a spinlock held ...
3911          */
3912         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3913 
3914         return retval;
3915 
3916 out_unlock:
3917         rcu_read_unlock();
3918         return retval;
3919 }
3920 
3921 static int sched_read_attr(struct sched_attr __user *uattr,
3922                            struct sched_attr *attr,
3923                            unsigned int usize)
3924 {
3925         int ret;
3926 
3927         if (!access_ok(VERIFY_WRITE, uattr, usize))
3928                 return -EFAULT;
3929 
3930         /*
3931          * If we're handed a smaller struct than we know of,
3932          * ensure all the unknown bits are 0 - i.e. old
3933          * user-space does not get uncomplete information.
3934          */
3935         if (usize < sizeof(*attr)) {
3936                 unsigned char *addr;
3937                 unsigned char *end;
3938 
3939                 addr = (void *)attr + usize;
3940                 end  = (void *)attr + sizeof(*attr);
3941 
3942                 for (; addr < end; addr++) {
3943                         if (*addr)
3944                                 return -EFBIG;
3945                 }
3946 
3947                 attr->size = usize;
3948         }
3949 
3950         ret = copy_to_user(uattr, attr, attr->size);
3951         if (ret)
3952                 return -EFAULT;
3953 
3954         return 0;
3955 }
3956 
3957 /**
3958  * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3959  * @pid: the pid in question.
3960  * @uattr: structure containing the extended parameters.
3961  * @size: sizeof(attr) for fwd/bwd comp.
3962  * @flags: for future extension.
3963  */
3964 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3965                 unsigned int, size, unsigned int, flags)
3966 {
3967         struct sched_attr attr = {
3968                 .size = sizeof(struct sched_attr),
3969         };
3970         struct task_struct *p;
3971         int retval;
3972 
3973         if (!uattr || pid < 0 || size > PAGE_SIZE ||
3974             size < SCHED_ATTR_SIZE_VER0 || flags)
3975                 return -EINVAL;
3976 
3977         rcu_read_lock();
3978         p = find_process_by_pid(pid);
3979         retval = -ESRCH;
3980         if (!p)
3981                 goto out_unlock;
3982 
3983         retval = security_task_getscheduler(p);
3984         if (retval)
3985                 goto out_unlock;
3986 
3987         attr.sched_policy = p->policy;
3988         if (p->sched_reset_on_fork)
3989                 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3990         if (task_has_dl_policy(p))
3991                 __getparam_dl(p, &attr);
3992         else if (task_has_rt_policy(p))
3993                 attr.sched_priority = p->rt_priority;
3994         else
3995                 attr.sched_nice = task_nice(p);
3996 
3997         rcu_read_unlock();
3998 
3999         retval = sched_read_attr(uattr, &attr, size);
4000         return retval;
4001 
4002 out_unlock:
4003         rcu_read_unlock();
4004         return retval;
4005 }
4006 
4007 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4008 {
4009         cpumask_var_t cpus_allowed, new_mask;
4010         struct task_struct *p;
4011         int retval;
4012 
4013         rcu_read_lock();
4014 
4015         p = find_process_by_pid(pid);
4016         if (!p) {
4017                 rcu_read_unlock();
4018                 return -ESRCH;
4019         }
4020 
4021         /* Prevent p going away */
4022         get_task_struct(p);
4023         rcu_read_unlock();
4024 
4025         if (p->flags & PF_NO_SETAFFINITY) {
4026                 retval = -EINVAL;
4027                 goto out_put_task;
4028         }
4029         if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4030                 retval = -ENOMEM;
4031                 goto out_put_task;
4032         }
4033         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4034                 retval = -ENOMEM;
4035                 goto out_free_cpus_allowed;
4036         }
4037         retval = -EPERM;
4038         if (!check_same_owner(p)) {
4039                 rcu_read_lock();
4040                 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4041                         rcu_read_unlock();
4042                         goto out_free_new_mask;
4043                 }
4044                 rcu_read_unlock();
4045         }
4046 
4047         retval = security_task_setscheduler(p);
4048         if (retval)
4049                 goto out_free_new_mask;
4050 
4051 
4052         cpuset_cpus_allowed(p, cpus_allowed);
4053         cpumask_and(new_mask, in_mask, cpus_allowed);
4054 
4055         /*
4056          * Since bandwidth control happens on root_domain basis,
4057          * if admission test is enabled, we only admit -deadline
4058          * tasks allowed to run on all the CPUs in the task's
4059          * root_domain.
4060          */
4061 #ifdef CONFIG_SMP
4062         if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4063                 rcu_read_lock();
4064                 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4065                         retval = -EBUSY;
4066                         rcu_read_unlock();
4067                         goto out_free_new_mask;
4068                 }
4069                 rcu_read_unlock();
4070         }
4071 #endif
4072 again:
4073         retval = set_cpus_allowed_ptr(p, new_mask);
4074 
4075         if (!retval) {
4076                 cpuset_cpus_allowed(p, cpus_allowed);
4077                 if (!cpumask_subset(new_mask, cpus_allowed)) {
4078                         /*
4079                          * We must have raced with a concurrent cpuset
4080                          * update. Just reset the cpus_allowed to the
4081                          * cpuset's cpus_allowed
4082                          */
4083                         cpumask_copy(new_mask, cpus_allowed);
4084                         goto again;
4085                 }
4086         }
4087 out_free_new_mask:
4088         free_cpumask_var(new_mask);
4089 out_free_cpus_allowed:
4090         free_cpumask_var(cpus_allowed);
4091 out_put_task:
4092         put_task_struct(p);
4093         return retval;
4094 }
4095 
4096 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4097                              struct cpumask *new_mask)
4098 {
4099         if (len < cpumask_size())
4100                 cpumask_clear(new_mask);
4101         else if (len > cpumask_size())
4102                 len = cpumask_size();
4103 
4104         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4105 }
4106 
4107 /**
4108  * sys_sched_setaffinity - set the cpu affinity of a process
4109  * @pid: pid of the process
4110  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4111  * @user_mask_ptr: user-space pointer to the new cpu mask
4112  *
4113  * Return: 0 on success. An error code otherwise.
4114  */
4115 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4116                 unsigned long __user *, user_mask_ptr)
4117 {
4118         cpumask_var_t new_mask;
4119         int retval;
4120 
4121         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4122                 return -ENOMEM;
4123 
4124         retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4125         if (retval == 0)
4126                 retval = sched_setaffinity(pid, new_mask);
4127         free_cpumask_var(new_mask);
4128         return retval;
4129 }
4130 
4131 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4132 {
4133         struct task_struct *p;
4134         unsigned long flags;
4135         int retval;
4136 
4137         rcu_read_lock();
4138 
4139         retval = -ESRCH;
4140         p = find_process_by_pid(pid);
4141         if (!p)
4142                 goto out_unlock;
4143 
4144         retval = security_task_getscheduler(p);
4145         if (retval)
4146                 goto out_unlock;
4147 
4148         raw_spin_lock_irqsave(&p->pi_lock, flags);
4149         cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4150         raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4151 
4152 out_unlock:
4153         rcu_read_unlock();
4154 
4155         return retval;
4156 }
4157 
4158 /**
4159  * sys_sched_getaffinity - get the cpu affinity of a process
4160  * @pid: pid of the process
4161  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4162  * @user_mask_ptr: user-space pointer to hold the current cpu mask
4163  *
4164  * Return: 0 on success. An error code otherwise.
4165  */
4166 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4167                 unsigned long __user *, user_mask_ptr)
4168 {
4169         int ret;
4170         cpumask_var_t mask;
4171 
4172         if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4173                 return -EINVAL;
4174         if (len & (sizeof(unsigned long)-1))
4175                 return -EINVAL;
4176 
4177         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4178                 return -ENOMEM;
4179 
4180         ret = sched_getaffinity(pid, mask);
4181         if (ret == 0) {
4182                 size_t retlen = min_t(size_t, len, cpumask_size());
4183 
4184                 if (copy_to_user(user_mask_ptr, mask, retlen))
4185                         ret = -EFAULT;
4186                 else
4187                         ret = retlen;
4188         }
4189         free_cpumask_var(mask);
4190 
4191         return ret;
4192 }
4193 
4194 /**
4195  * sys_sched_yield - yield the current processor to other threads.
4196  *
4197  * This function yields the current CPU to other tasks. If there are no
4198  * other threads running on this CPU then this function will return.
4199  *
4200  * Return: 0.
4201  */
4202 SYSCALL_DEFINE0(sched_yield)
4203 {
4204         struct rq *rq = this_rq_lock();
4205 
4206         schedstat_inc(rq, yld_count);
4207         current->sched_class->yield_task(rq);
4208 
4209         /*
4210          * Since we are going to call schedule() anyway, there's
4211          * no need to preempt or enable interrupts:
4212          */
4213         __release(rq->lock);
4214         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4215         do_raw_spin_unlock(&rq->lock);
4216         sched_preempt_enable_no_resched();
4217 
4218         schedule();
4219 
4220         return 0;
4221 }
4222 
4223 int __sched _cond_resched(void)
4224 {
4225         if (should_resched()) {
4226                 preempt_schedule_common();
4227                 return 1;
4228         }
4229         return 0;
4230 }
4231 EXPORT_SYMBOL(_cond_resched);
4232 
4233 /*
4234  * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4235  * call schedule, and on return reacquire the lock.
4236  *
4237  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4238  * operations here to prevent schedule() from being called twice (once via
4239  * spin_unlock(), once by hand).
4240  */
4241 int __cond_resched_lock(spinlock_t *lock)
4242 {
4243         int resched = should_resched();
4244         int ret = 0;
4245 
4246         lockdep_assert_held(lock);
4247 
4248         if (spin_needbreak(lock) || resched) {
4249                 spin_unlock(lock);
4250                 if (resched)
4251                         preempt_schedule_common();
4252                 else
4253                         cpu_relax();
4254                 ret = 1;
4255                 spin_lock(lock);
4256         }
4257         return ret;
4258 }
4259 EXPORT_SYMBOL(__cond_resched_lock);
4260 
4261 int __sched __cond_resched_softirq(void)
4262 {
4263         BUG_ON(!in_softirq());
4264 
4265         if (should_resched()) {
4266                 local_bh_enable();
4267                 preempt_schedule_common();
4268                 local_bh_disable();
4269                 return 1;
4270         }
4271         return 0;
4272 }
4273 EXPORT_SYMBOL(__cond_resched_softirq);
4274 
4275 /**
4276  * yield - yield the current processor to other threads.
4277  *
4278  * Do not ever use this function, there's a 99% chance you're doing it wrong.
4279  *
4280  * The scheduler is at all times free to pick the calling task as the most
4281  * eligible task to run, if removing the yield() call from your code breaks
4282  * it, its already broken.
4283  *
4284  * Typical broken usage is:
4285  *
4286  * while (!event)
4287  *      yield();
4288  *
4289  * where one assumes that yield() will let 'the other' process run that will
4290  * make event true. If the current task is a SCHED_FIFO task that will never
4291  * happen. Never use yield() as a progress guarantee!!
4292  *
4293  * If you want to use yield() to wait for something, use wait_event().
4294  * If you want to use yield() to be 'nice' for others, use cond_resched().
4295  * If you still want to use yield(), do not!
4296  */
4297 void __sched yield(void)
4298 {
4299         set_current_state(TASK_RUNNING);
4300         sys_sched_yield();
4301 }
4302 EXPORT_SYMBOL(yield);
4303 
4304 /**
4305  * yield_to - yield the current processor to another thread in
4306  * your thread group, or accelerate that thread toward the
4307  * processor it's on.
4308  * @p: target task
4309  * @preempt: whether task preemption is allowed or not
4310  *
4311  * It's the caller's job to ensure that the target task struct
4312  * can't go away on us before we can do any checks.
4313  *
4314  * Return:
4315  *      true (>0) if we indeed boosted the target task.
4316  *      false (0) if we failed to boost the target.
4317  *      -ESRCH if there's no task to yield to.
4318  */
4319 int __sched yield_to(struct task_struct *p, bool preempt)
4320 {
4321         struct task_struct *curr = current;
4322         struct rq *rq, *p_rq;
4323         unsigned long flags;
4324         int yielded = 0;
4325 
4326         local_irq_save(flags);
4327         rq = this_rq();
4328 
4329 again:
4330         p_rq = task_rq(p);
4331         /*
4332          * If we're the only runnable task on the rq and target rq also
4333          * has only one task, there's absolutely no point in yielding.
4334          */
4335         if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4336                 yielded = -ESRCH;
4337                 goto out_irq;
4338         }
4339 
4340         double_rq_lock(rq, p_rq);
4341         if (task_rq(p) != p_rq) {
4342                 double_rq_unlock(rq, p_rq);
4343                 goto again;
4344         }
4345 
4346         if (!curr->sched_class->yield_to_task)
4347                 goto out_unlock;
4348 
4349         if (curr->sched_class != p->sched_class)
4350                 goto out_unlock;
4351 
4352         if (task_running(p_rq, p) || p->state)
4353                 goto out_unlock;
4354 
4355         yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4356         if (yielded) {
4357                 schedstat_inc(rq, yld_count);
4358                 /*
4359                  * Make p's CPU reschedule; pick_next_entity takes care of
4360                  * fairness.
4361                  */
4362                 if (preempt && rq != p_rq)
4363                         resched_curr(p_rq);
4364         }
4365 
4366 out_unlock:
4367         double_rq_unlock(rq, p_rq);
4368 out_irq:
4369         local_irq_restore(flags);
4370 
4371         if (yielded > 0)
4372                 schedule();
4373 
4374         return yielded;
4375 }
4376 EXPORT_SYMBOL_GPL(yield_to);
4377 
4378 /*
4379  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4380  * that process accounting knows that this is a task in IO wait state.
4381  */
4382 long __sched io_schedule_timeout(long timeout)
4383 {
4384         int old_iowait = current->in_iowait;
4385         struct rq *rq;
4386         long ret;
4387 
4388         current->in_iowait = 1;
4389         blk_schedule_flush_plug(current);
4390 
4391         delayacct_blkio_start();
4392         rq = raw_rq();
4393         atomic_inc(&rq->nr_iowait);
4394         ret = schedule_timeout(timeout);
4395         current->in_iowait = old_iowait;
4396         atomic_dec(&rq->nr_iowait);
4397         delayacct_blkio_end();
4398 
4399         return ret;
4400 }
4401 EXPORT_SYMBOL(io_schedule_timeout);
4402 
4403 /**
4404  * sys_sched_get_priority_max - return maximum RT priority.
4405  * @policy: scheduling class.
4406  *
4407  * Return: On success, this syscall returns the maximum
4408  * rt_priority that can be used by a given scheduling class.
4409  * On failure, a negative error code is returned.
4410  */
4411 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4412 {
4413         int ret = -EINVAL;
4414 
4415         switch (policy) {
4416         case SCHED_FIFO:
4417         case SCHED_RR:
4418                 ret = MAX_USER_RT_PRIO-1;
4419                 break;
4420         case SCHED_DEADLINE:
4421         case SCHED_NORMAL:
4422         case SCHED_BATCH:
4423         case SCHED_IDLE:
4424                 ret = 0;
4425                 break;
4426         }
4427         return ret;
4428 }
4429 
4430 /**
4431  * sys_sched_get_priority_min - return minimum RT priority.
4432  * @policy: scheduling class.
4433  *
4434  * Return: On success, this syscall returns the minimum
4435  * rt_priority that can be used by a given scheduling class.
4436  * On failure, a negative error code is returned.
4437  */
4438 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4439 {
4440         int ret = -EINVAL;
4441 
4442         switch (policy) {
4443         case SCHED_FIFO:
4444         case SCHED_RR:
4445                 ret = 1;
4446                 break;
4447         case SCHED_DEADLINE:
4448         case SCHED_NORMAL:
4449         case SCHED_BATCH:
4450         case SCHED_IDLE:
4451                 ret = 0;
4452         }
4453         return ret;
4454 }
4455 
4456 /**
4457  * sys_sched_rr_get_interval - return the default timeslice of a process.
4458  * @pid: pid of the process.
4459  * @interval: userspace pointer to the timeslice value.
4460  *
4461  * this syscall writes the default timeslice value of a given process
4462  * into the user-space timespec buffer. A value of '' means infinity.
4463  *
4464  * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4465  * an error code.
4466  */
4467 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4468                 struct timespec __user *, interval)
4469 {
4470         struct task_struct *p;
4471         unsigned int time_slice;
4472         unsigned long flags;
4473         struct rq *rq;
4474         int retval;
4475         struct timespec t;
4476 
4477         if (pid < 0)
4478                 return -EINVAL;
4479 
4480         retval = -ESRCH;
4481         rcu_read_lock();
4482         p = find_process_by_pid(pid);
4483         if (!p)
4484                 goto out_unlock;
4485 
4486         retval = security_task_getscheduler(p);
4487         if (retval)
4488                 goto out_unlock;
4489 
4490         rq = task_rq_lock(p, &flags);
4491         time_slice = 0;
4492         if (p->sched_class->get_rr_interval)
4493                 time_slice = p->sched_class->get_rr_interval(rq, p);
4494         task_rq_unlock(rq, p, &flags);
4495 
4496         rcu_read_unlock();
4497         jiffies_to_timespec(time_slice, &t);
4498         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4499         return retval;
4500 
4501 out_unlock:
4502         rcu_read_unlock();
4503         return retval;
4504 }
4505 
4506 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4507 
4508 void sched_show_task(struct task_struct *p)
4509 {
4510         unsigned long free = 0;
4511         int ppid;
4512         unsigned long state = p->state;
4513 
4514         if (state)
4515                 state = __ffs(state) + 1;
4516         printk(KERN_INFO "%-15.15s %c", p->comm,
4517                 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4518 #if BITS_PER_LONG == 32
4519         if (state == TASK_RUNNING)
4520                 printk(KERN_CONT " running  ");
4521         else
4522                 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4523 #else
4524         if (state == TASK_RUNNING)
4525                 printk(KERN_CONT "  running task    ");
4526         else
4527                 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4528 #endif
4529 #ifdef CONFIG_DEBUG_STACK_USAGE
4530         free = stack_not_used(p);
4531 #endif
4532         ppid = 0;
4533         rcu_read_lock();
4534         if (pid_alive(p))
4535                 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4536         rcu_read_unlock();
4537         printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4538                 task_pid_nr(p), ppid,
4539                 (unsigned long)task_thread_info(p)->flags);
4540 
4541         print_worker_info(KERN_INFO, p);
4542         show_stack(p, NULL);
4543 }
4544 
4545 void show_state_filter(unsigned long state_filter)
4546 {
4547         struct task_struct *g, *p;
4548 
4549 #if BITS_PER_LONG == 32
4550         printk(KERN_INFO
4551                 "  task                PC stack   pid father\n");
4552 #else
4553         printk(KERN_INFO
4554                 "  task                        PC stack   pid father\n");
4555 #endif
4556         rcu_read_lock();
4557         for_each_process_thread(g, p) {
4558                 /*
4559                  * reset the NMI-timeout, listing all files on a slow
4560                  * console might take a lot of time:
4561                  */
4562                 touch_nmi_watchdog();
4563                 if (!state_filter || (p->state & state_filter))
4564                         sched_show_task(p);
4565         }
4566 
4567         touch_all_softlockup_watchdogs();
4568 
4569 #ifdef CONFIG_SCHED_DEBUG
4570         sysrq_sched_debug_show();
4571 #endif
4572         rcu_read_unlock();
4573         /*
4574          * Only show locks if all tasks are dumped:
4575          */
4576         if (!state_filter)
4577                 debug_show_all_locks();
4578 }
4579 
4580 void init_idle_bootup_task(struct task_struct *idle)
4581 {
4582         idle->sched_class = &idle_sched_class;
4583 }
4584 
4585 /**
4586  * init_idle - set up an idle thread for a given CPU
4587  * @idle: task in question
4588  * @cpu: cpu the idle task belongs to
4589  *
4590  * NOTE: this function does not set the idle thread's NEED_RESCHED
4591  * flag, to make booting more robust.
4592  */
4593 void init_idle(struct task_struct *idle, int cpu)
4594 {
4595         struct rq *rq = cpu_rq(cpu);
4596         unsigned long flags;
4597 
4598         raw_spin_lock_irqsave(&rq->lock, flags);
4599 
4600         __sched_fork(0, idle);
4601         idle->state = TASK_RUNNING;
4602         idle->se.exec_start = sched_clock();
4603 
4604         do_set_cpus_allowed(idle, cpumask_of(cpu));
4605         /*
4606          * We're having a chicken and egg problem, even though we are
4607          * holding rq->lock, the cpu isn't yet set to this cpu so the
4608          * lockdep check in task_group() will fail.
4609          *
4610          * Similar case to sched_fork(). / Alternatively we could
4611          * use task_rq_lock() here and obtain the other rq->lock.
4612          *
4613          * Silence PROVE_RCU
4614          */
4615         rcu_read_lock();
4616         __set_task_cpu(idle, cpu);
4617         rcu_read_unlock();
4618 
4619         rq->curr = rq->idle = idle;
4620         idle->on_rq = TASK_ON_RQ_QUEUED;
4621 #if defined(CONFIG_SMP)
4622         idle->on_cpu = 1;
4623 #endif
4624         raw_spin_unlock_irqrestore(&rq->lock, flags);
4625 
4626         /* Set the preempt count _outside_ the spinlocks! */
4627         init_idle_preempt_count(idle, cpu);
4628 
4629         /*
4630          * The idle tasks have their own, simple scheduling class:
4631          */
4632         idle->sched_class = &idle_sched_class;
4633         ftrace_graph_init_idle_task(idle, cpu);
4634         vtime_init_idle(idle, cpu);
4635 #if defined(CONFIG_SMP)
4636         sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4637 #endif
4638 }
4639 
4640 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4641                               const struct cpumask *trial)
4642 {
4643         int ret = 1, trial_cpus;
4644         struct dl_bw *cur_dl_b;
4645         unsigned long flags;
4646 
4647         if (!cpumask_weight(cur))
4648                 return ret;
4649 
4650         rcu_read_lock_sched();
4651         cur_dl_b = dl_bw_of(cpumask_any(cur));
4652         trial_cpus = cpumask_weight(trial);
4653 
4654         raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4655         if (cur_dl_b->bw != -1 &&
4656             cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4657                 ret = 0;
4658         raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4659         rcu_read_unlock_sched();
4660 
4661         return ret;
4662 }
4663 
4664 int task_can_attach(struct task_struct *p,
4665                     const struct cpumask *cs_cpus_allowed)
4666 {
4667         int ret = 0;
4668 
4669         /*
4670          * Kthreads which disallow setaffinity shouldn't be moved
4671          * to a new cpuset; we don't want to change their cpu
4672          * affinity and isolating such threads by their set of
4673          * allowed nodes is unnecessary.  Thus, cpusets are not
4674          * applicable for such threads.  This prevents checking for
4675          * success of set_cpus_allowed_ptr() on all attached tasks
4676          * before cpus_allowed may be changed.
4677          */
4678         if (p->flags & PF_NO_SETAFFINITY) {
4679                 ret = -EINVAL;
4680                 goto out;
4681         }
4682 
4683 #ifdef CONFIG_SMP
4684         if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4685                                               cs_cpus_allowed)) {
4686                 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4687                                                         cs_cpus_allowed);
4688                 struct dl_bw *dl_b;
4689                 bool overflow;
4690                 int cpus;
4691                 unsigned long flags;
4692 
4693                 rcu_read_lock_sched();
4694                 dl_b = dl_bw_of(dest_cpu);
4695                 raw_spin_lock_irqsave(&dl_b->lock, flags);
4696                 cpus = dl_bw_cpus(dest_cpu);
4697                 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4698                 if (overflow)
4699                         ret = -EBUSY;
4700                 else {
4701                         /*
4702                          * We reserve space for this task in the destination
4703                          * root_domain, as we can't fail after this point.
4704                          * We will free resources in the source root_domain
4705                          * later on (see set_cpus_allowed_dl()).
4706                          */
4707                         __dl_add(dl_b, p->dl.dl_bw);
4708                 }
4709                 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4710                 rcu_read_unlock_sched();
4711 
4712         }
4713 #endif
4714 out:
4715         return ret;
4716 }
4717 
4718 #ifdef CONFIG_SMP
4719 /*
4720  * move_queued_task - move a queued task to new rq.
4721  *
4722  * Returns (locked) new rq. Old rq's lock is released.
4723  */
4724 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4725 {
4726         struct rq *rq = task_rq(p);
4727 
4728         lockdep_assert_held(&rq->lock);
4729 
4730         dequeue_task(rq, p, 0);
4731         p->on_rq = TASK_ON_RQ_MIGRATING;
4732         set_task_cpu(p, new_cpu);
4733         raw_spin_unlock(&rq->lock);
4734 
4735         rq = cpu_rq(new_cpu);
4736 
4737         raw_spin_lock(&rq->lock);
4738         BUG_ON(task_cpu(p) != new_cpu);
4739         p->on_rq = TASK_ON_RQ_QUEUED;
4740         enqueue_task(rq, p, 0);
4741         check_preempt_curr(rq, p, 0);
4742 
4743         return rq;
4744 }
4745 
4746 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4747 {
4748         if (p->sched_class->set_cpus_allowed)
4749                 p->sched_class->set_cpus_allowed(p, new_mask);
4750 
4751         cpumask_copy(&p->cpus_allowed, new_mask);
4752         p->nr_cpus_allowed = cpumask_weight(new_mask);
4753 }
4754 
4755 /*
4756  * This is how migration works:
4757  *
4758  * 1) we invoke migration_cpu_stop() on the target CPU using
4759  *    stop_one_cpu().
4760  * 2) stopper starts to run (implicitly forcing the migrated thread
4761  *    off the CPU)
4762  * 3) it checks whether the migrated task is still in the wrong runqueue.
4763  * 4) if it's in the wrong runqueue then the migration thread removes
4764  *    it and puts it into the right queue.
4765  * 5) stopper completes and stop_one_cpu() returns and the migration
4766  *    is done.
4767  */
4768 
4769 /*
4770  * Change a given task's CPU affinity. Migrate the thread to a
4771  * proper CPU and schedule it away if the CPU it's executing on
4772  * is removed from the allowed bitmask.
4773  *
4774  * NOTE: the caller must have a valid reference to the task, the
4775  * task must not exit() & deallocate itself prematurely. The
4776  * call is not atomic; no spinlocks may be held.
4777  */
4778 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4779 {
4780         unsigned long flags;
4781         struct rq *rq;
4782         unsigned int dest_cpu;
4783         int ret = 0;
4784 
4785         rq = task_rq_lock(p, &flags);
4786 
4787         if (cpumask_equal(&p->cpus_allowed, new_mask))
4788                 goto out;
4789 
4790         if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4791                 ret = -EINVAL;
4792                 goto out;
4793         }
4794 
4795         do_set_cpus_allowed(p, new_mask);
4796 
4797         /* Can the task run on the task's current CPU? If so, we're done */
4798         if (cpumask_test_cpu(task_cpu(p), new_mask))
4799                 goto out;
4800 
4801         dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4802         if (task_running(rq, p) || p->state == TASK_WAKING) {
4803                 struct migration_arg arg = { p, dest_cpu };
4804                 /* Need help from migration thread: drop lock and wait. */
4805                 task_rq_unlock(rq, p, &flags);
4806                 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4807                 tlb_migrate_finish(p->mm);
4808                 return 0;
4809         } else if (task_on_rq_queued(p))
4810                 rq = move_queued_task(p, dest_cpu);
4811 out:
4812         task_rq_unlock(rq, p, &flags);
4813 
4814         return ret;
4815 }
4816 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4817 
4818 /*
4819  * Move (not current) task off this cpu, onto dest cpu. We're doing
4820  * this because either it can't run here any more (set_cpus_allowed()
4821  * away from this CPU, or CPU going down), or because we're
4822  * attempting to rebalance this task on exec (sched_exec).
4823  *
4824  * So we race with normal scheduler movements, but that's OK, as long
4825  * as the task is no longer on this CPU.
4826  *
4827  * Returns non-zero if task was successfully migrated.
4828  */
4829 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4830 {
4831         struct rq *rq;
4832         int ret = 0;
4833 
4834         if (unlikely(!cpu_active(dest_cpu)))
4835                 return ret;
4836 
4837         rq = cpu_rq(src_cpu);
4838 
4839         raw_spin_lock(&p->pi_lock);
4840         raw_spin_lock(&rq->lock);
4841         /* Already moved. */
4842         if (task_cpu(p) != src_cpu)
4843                 goto done;
4844 
4845         /* Affinity changed (again). */
4846         if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4847                 goto fail;
4848 
4849         /*
4850          * If we're not on a rq, the next wake-up will ensure we're
4851          * placed properly.
4852          */
4853         if (task_on_rq_queued(p))
4854                 rq = move_queued_task(p, dest_cpu);
4855 done:
4856         ret = 1;
4857 fail:
4858         raw_spin_unlock(&rq->lock);
4859         raw_spin_unlock(&p->pi_lock);
4860         return ret;
4861 }
4862 
4863 #ifdef CONFIG_NUMA_BALANCING
4864 /* Migrate current task p to target_cpu */
4865 int migrate_task_to(struct task_struct *p, int target_cpu)
4866 {
4867         struct migration_arg arg = { p, target_cpu };
4868         int curr_cpu = task_cpu(p);
4869 
4870         if (curr_cpu == target_cpu)
4871                 return 0;
4872 
4873         if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4874                 return -EINVAL;
4875 
4876         /* TODO: This is not properly updating schedstats */
4877 
4878         trace_sched_move_numa(p, curr_cpu, target_cpu);
4879         return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4880 }
4881 
4882 /*
4883  * Requeue a task on a given node and accurately track the number of NUMA
4884  * tasks on the runqueues
4885  */
4886 void sched_setnuma(struct task_struct *p, int nid)
4887 {
4888         struct rq *rq;
4889         unsigned long flags;
4890         bool queued, running;
4891 
4892         rq = task_rq_lock(p, &flags);
4893         queued = task_on_rq_queued(p);
4894         running = task_current(rq, p);
4895 
4896         if (queued)
4897                 dequeue_task(rq, p, 0);
4898         if (running)
4899                 put_prev_task(rq, p);
4900 
4901         p->numa_preferred_nid = nid;
4902 
4903         if (running)
4904                 p->sched_class->set_curr_task(rq);
4905         if (queued)
4906                 enqueue_task(rq, p, 0);
4907         task_rq_unlock(rq, p, &flags);
4908 }
4909 #endif
4910 
4911 /*
4912  * migration_cpu_stop - this will be executed by a highprio stopper thread
4913  * and performs thread migration by bumping thread off CPU then
4914  * 'pushing' onto another runqueue.
4915  */
4916 static int migration_cpu_stop(void *data)
4917 {
4918         struct migration_arg *arg = data;
4919 
4920         /*
4921          * The original target cpu might have gone down and we might
4922          * be on another cpu but it doesn't matter.
4923          */
4924         local_irq_disable();
4925         /*
4926          * We need to explicitly wake pending tasks before running
4927          * __migrate_task() such that we will not miss enforcing cpus_allowed
4928          * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4929          */
4930         sched_ttwu_pending();
4931         __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4932         local_irq_enable();
4933         return 0;
4934 }
4935 
4936 #ifdef CONFIG_HOTPLUG_CPU
4937 
4938 /*
4939  * Ensures that the idle task is using init_mm right before its cpu goes
4940  * offline.
4941  */
4942 void idle_task_exit(void)
4943 {
4944         struct mm_struct *mm = current->active_mm;
4945 
4946         BUG_ON(cpu_online(smp_processor_id()));
4947 
4948         if (mm != &init_mm) {
4949                 switch_mm(mm, &init_mm, current);
4950                 finish_arch_post_lock_switch();
4951         }
4952         mmdrop(mm);
4953 }
4954 
4955 /*
4956  * Since this CPU is going 'away' for a while, fold any nr_active delta
4957  * we might have. Assumes we're called after migrate_tasks() so that the
4958  * nr_active count is stable.
4959  *
4960  * Also see the comment "Global load-average calculations".
4961  */
4962 static void calc_load_migrate(struct rq *rq)
4963 {
4964         long delta = calc_load_fold_active(rq);
4965         if (delta)
4966                 atomic_long_add(delta, &calc_load_tasks);
4967 }
4968 
4969 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4970 {
4971 }
4972 
4973 static const struct sched_class fake_sched_class = {
4974         .put_prev_task = put_prev_task_fake,
4975 };
4976 
4977 static struct task_struct fake_task = {
4978         /*
4979          * Avoid pull_{rt,dl}_task()
4980          */
4981         .prio = MAX_PRIO + 1,
4982         .sched_class = &fake_sched_class,
4983 };
4984 
4985 /*
4986  * Migrate all tasks from the rq, sleeping tasks will be migrated by
4987  * try_to_wake_up()->select_task_rq().
4988  *
4989  * Called with rq->lock held even though we'er in stop_machine() and
4990  * there's no concurrency possible, we hold the required locks anyway
4991  * because of lock validation efforts.
4992  */
4993 static void migrate_tasks(unsigned int dead_cpu)
4994 {
4995         struct rq *rq = cpu_rq(dead_cpu);
4996         struct task_struct *next, *stop = rq->stop;
4997         int dest_cpu;
4998 
4999         /*
5000          * Fudge the rq selection such that the below task selection loop
5001          * doesn't get stuck on the currently eligible stop task.
5002          *
5003          * We're currently inside stop_machine() and the rq is either stuck
5004          * in the stop_machine_cpu_stop() loop, or we're executing this code,
5005          * either way we should never end up calling schedule() until we're
5006          * done here.
5007          */
5008         rq->stop = NULL;
5009 
5010         /*
5011          * put_prev_task() and pick_next_task() sched
5012          * class method both need to have an up-to-date
5013          * value of rq->clock[_task]
5014          */
5015         update_rq_clock(rq);
5016 
5017         for ( ; ; ) {
5018                 /*
5019                  * There's this thread running, bail when that's the only
5020                  * remaining thread.
5021                  */
5022                 if (rq->nr_running == 1)
5023                         break;
5024 
5025                 next = pick_next_task(rq, &fake_task);
5026                 BUG_ON(!next);
5027                 next->sched_class->put_prev_task(rq, next);
5028 
5029                 /* Find suitable destination for @next, with force if needed. */
5030                 dest_cpu = select_fallback_rq(dead_cpu, next);
5031                 raw_spin_unlock(&rq->lock);
5032 
5033                 __migrate_task(next, dead_cpu, dest_cpu);
5034 
5035                 raw_spin_lock(&rq->lock);
5036         }
5037 
5038         rq->stop = stop;
5039 }
5040 
5041 #endif /* CONFIG_HOTPLUG_CPU */
5042 
5043 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5044 
5045 static struct ctl_table sd_ctl_dir[] = {
5046         {
5047                 .procname       = "sched_domain",
5048                 .mode           = 0555,
5049         },
5050         {}
5051 };
5052 
5053 static struct ctl_table sd_ctl_root[] = {
5054         {
5055                 .procname       = "kernel",
5056                 .mode           = 0555,
5057                 .child          = sd_ctl_dir,
5058         },
5059         {}
5060 };
5061 
5062 static struct ctl_table *sd_alloc_ctl_entry(int n)
5063 {
5064         struct ctl_table *entry =
5065                 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5066 
5067         return entry;
5068 }
5069 
5070 static void sd_free_ctl_entry(struct ctl_table **tablep)
5071 {
5072         struct ctl_table *entry;
5073 
5074         /*
5075          * In the intermediate directories, both the child directory and
5076          * procname are dynamically allocated and could fail but the mode
5077          * will always be set. In the lowest directory the names are
5078          * static strings and all have proc handlers.
5079          */
5080         for (entry = *tablep; entry->mode; entry++) {
5081                 if (entry->child)
5082                         sd_free_ctl_entry(&entry->child);
5083                 if (entry->proc_handler == NULL)
5084                         kfree(entry->procname);
5085         }
5086 
5087         kfree(*tablep);
5088         *tablep = NULL;
5089 }
5090 
5091 static int min_load_idx = 0;
5092 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5093 
5094 static void
5095 set_table_entry(struct ctl_table *entry,
5096                 const char *procname, void *data, int maxlen,
5097                 umode_t mode, proc_handler *proc_handler,
5098                 bool load_idx)
5099 {
5100         entry->procname = procname;
5101         entry->data = data;
5102         entry->maxlen = maxlen;
5103         entry->mode = mode;
5104         entry->proc_handler = proc_handler;
5105 
5106         if (load_idx) {
5107                 entry->extra1 = &min_load_idx;
5108                 entry->extra2 = &max_load_idx;
5109         }
5110 }
5111 
5112 static struct ctl_table *
5113 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5114 {
5115         struct ctl_table *table = sd_alloc_ctl_entry(14);
5116 
5117         if (table == NULL)
5118                 return NULL;
5119 
5120         set_table_entry(&table[0], "min_interval", &sd->min_interval,
5121                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5122         set_table_entry(&table[1], "max_interval", &sd->max_interval,
5123                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5124         set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5125                 sizeof(int), 0644, proc_dointvec_minmax, true);
5126         set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5127                 sizeof(int), 0644, proc_dointvec_minmax, true);
5128         set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5129                 sizeof(int), 0644, proc_dointvec_minmax, true);
5130         set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5131                 sizeof(int), 0644, proc_dointvec_minmax, true);
5132         set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5133                 sizeof(int), 0644, proc_dointvec_minmax, true);
5134         set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5135                 sizeof(int), 0644, proc_dointvec_minmax, false);
5136         set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5137                 sizeof(int), 0644, proc_dointvec_minmax, false);
5138         set_table_entry(&table[9], "cache_nice_tries",
5139                 &sd->cache_nice_tries,
5140                 sizeof(int), 0644, proc_dointvec_minmax, false);
5141         set_table_entry(&table[10], "flags", &sd->flags,
5142                 sizeof(int), 0644, proc_dointvec_minmax, false);
5143         set_table_entry(&table[11], "max_newidle_lb_cost",
5144                 &sd->max_newidle_lb_cost,
5145                 sizeof(long), 0644, proc_doulongvec_minmax, false);
5146         set_table_entry(&table[12], "name", sd->name,
5147                 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5148         /* &table[13] is terminator */
5149 
5150         return table;
5151 }
5152 
5153 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5154 {
5155         struct ctl_table *entry, *table;
5156         struct sched_domain *sd;
5157         int domain_num = 0, i;
5158         char buf[32];
5159 
5160         for_each_domain(cpu, sd)
5161                 domain_num++;
5162         entry = table = sd_alloc_ctl_entry(domain_num + 1);
5163         if (table == NULL)
5164                 return NULL;
5165 
5166         i = 0;
5167         for_each_domain(cpu, sd) {
5168                 snprintf(buf, 32, "domain%d", i);
5169                 entry->procname = kstrdup(buf, GFP_KERNEL);
5170                 entry->mode = 0555;
5171                 entry->child = sd_alloc_ctl_domain_table(sd);
5172                 entry++;
5173                 i++;
5174         }
5175         return table;
5176 }
5177 
5178 static struct ctl_table_header *sd_sysctl_header;
5179 static void register_sched_domain_sysctl(void)
5180 {
5181         int i, cpu_num = num_possible_cpus();
5182         struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5183         char buf[32];
5184 
5185         WARN_ON(sd_ctl_dir[0].child);
5186         sd_ctl_dir[0].child = entry;
5187 
5188         if (entry == NULL)
5189                 return;
5190 
5191         for_each_possible_cpu(i) {
5192                 snprintf(buf, 32, "cpu%d", i);
5193                 entry->procname = kstrdup(buf, GFP_KERNEL);
5194                 entry->mode = 0555;
5195                 entry->child = sd_alloc_ctl_cpu_table(i);
5196                 entry++;
5197         }
5198 
5199         WARN_ON(sd_sysctl_header);
5200         sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5201 }
5202 
5203 /* may be called multiple times per register */
5204 static void unregister_sched_domain_sysctl(void)
5205 {
5206         if (sd_sysctl_header)
5207                 unregister_sysctl_table(sd_sysctl_header);
5208         sd_sysctl_header = NULL;
5209         if (sd_ctl_dir[0].child)
5210                 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5211 }
5212 #else
5213 static void register_sched_domain_sysctl(void)
5214 {
5215 }
5216 static void unregister_sched_domain_sysctl(void)
5217 {
5218 }
5219 #endif
5220 
5221 static void set_rq_online(struct rq *rq)
5222 {
5223         if (!rq->online) {
5224                 const struct sched_class *class;
5225 
5226                 cpumask_set_cpu(rq->cpu, rq->rd->online);
5227                 rq->online = 1;
5228 
5229                 for_each_class(class) {
5230                         if (class->rq_online)
5231                                 class->rq_online(rq);
5232                 }
5233         }
5234 }
5235 
5236 static void set_rq_offline(struct rq *rq)
5237 {
5238         if (rq->online) {
5239                 const struct sched_class *class;
5240 
5241                 for_each_class(class) {
5242                         if (class->rq_offline)
5243                                 class->rq_offline(rq);
5244                 }
5245 
5246                 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5247                 rq->online = 0;
5248         }
5249 }
5250 
5251 /*
5252  * migration_call - callback that gets triggered when a CPU is added.
5253  * Here we can start up the necessary migration thread for the new CPU.
5254  */
5255 static int
5256 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5257 {
5258         int cpu = (long)hcpu;
5259         unsigned long flags;
5260         struct rq *rq = cpu_rq(cpu);
5261 
5262         switch (action & ~CPU_TASKS_FROZEN) {
5263 
5264         case CPU_UP_PREPARE:
5265                 rq->calc_load_update = calc_load_update;
5266                 break;
5267 
5268         case CPU_ONLINE:
5269                 /* Update our root-domain */
5270                 raw_spin_lock_irqsave(&rq->lock, flags);
5271                 if (rq->rd) {
5272                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5273 
5274                         set_rq_online(rq);
5275                 }
5276                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5277                 break;
5278 
5279 #ifdef CONFIG_HOTPLUG_CPU
5280         case CPU_DYING:
5281                 sched_ttwu_pending();
5282                 /* Update our root-domain */
5283                 raw_spin_lock_irqsave(&rq->lock, flags);
5284                 if (rq->rd) {
5285                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5286                         set_rq_offline(rq);
5287                 }
5288                 migrate_tasks(cpu);
5289                 BUG_ON(rq->nr_running != 1); /* the migration thread */
5290                 raw_spin_unlock_irqrestore(&rq->lock, flags);
5291                 break;
5292 
5293         case CPU_DEAD:
5294                 calc_load_migrate(rq);
5295                 break;
5296 #endif
5297         }
5298 
5299         update_max_interval();
5300 
5301         return NOTIFY_OK;
5302 }
5303 
5304 /*
5305  * Register at high priority so that task migration (migrate_all_tasks)
5306  * happens before everything else.  This has to be lower priority than
5307  * the notifier in the perf_event subsystem, though.
5308  */
5309 static struct notifier_block migration_notifier = {
5310         .notifier_call = migration_call,
5311         .priority = CPU_PRI_MIGRATION,
5312 };
5313 
5314 static void __cpuinit set_cpu_rq_start_time(void)
5315 {
5316         int cpu = smp_processor_id();
5317         struct rq *rq = cpu_rq(cpu);
5318         rq->age_stamp = sched_clock_cpu(cpu);
5319 }
5320 
5321 static int sched_cpu_active(struct notifier_block *nfb,
5322                                       unsigned long action, void *hcpu)
5323 {
5324         switch (action & ~CPU_TASKS_FROZEN) {
5325         case CPU_STARTING:
5326                 set_cpu_rq_start_time();
5327                 return NOTIFY_OK;
5328         case CPU_DOWN_FAILED:
5329                 set_cpu_active((long)hcpu, true);
5330                 return NOTIFY_OK;
5331         default:
5332                 return NOTIFY_DONE;
5333         }
5334 }
5335 
5336 static int sched_cpu_inactive(struct notifier_block *nfb,
5337                                         unsigned long action, void *hcpu)
5338 {
5339         unsigned long flags;
5340         long cpu = (long)hcpu;
5341         struct dl_bw *dl_b;
5342 
5343         switch (action & ~CPU_TASKS_FROZEN) {
5344         case CPU_DOWN_PREPARE:
5345                 set_cpu_active(cpu, false);
5346 
5347                 /* explicitly allow suspend */
5348                 if (!(action & CPU_TASKS_FROZEN)) {
5349                         bool overflow;
5350                         int cpus;
5351 
5352                         rcu_read_lock_sched();
5353                         dl_b = dl_bw_of(cpu);
5354 
5355                         raw_spin_lock_irqsave(&dl_b->lock, flags);
5356                         cpus = dl_bw_cpus(cpu);
5357                         overflow = __dl_overflow(dl_b, cpus, 0, 0);
5358                         raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5359 
5360                         rcu_read_unlock_sched();
5361 
5362                         if (overflow)
5363                                 return notifier_from_errno(-EBUSY);
5364                 }
5365                 return NOTIFY_OK;
5366         }
5367 
5368         return NOTIFY_DONE;
5369 }
5370 
5371 static int __init migration_init(void)
5372 {
5373         void *cpu = (void *)(long)smp_processor_id();
5374         int err;
5375 
5376         /* Initialize migration for the boot CPU */
5377         err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5378         BUG_ON(err == NOTIFY_BAD);
5379         migration_call(&migration_notifier, CPU_ONLINE, cpu);
5380         register_cpu_notifier(&migration_notifier);
5381 
5382         /* Register cpu active notifiers */
5383         cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5384         cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5385 
5386         return 0;
5387 }
5388 early_initcall(migration_init);
5389 #endif
5390 
5391 #ifdef CONFIG_SMP
5392 
5393 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5394 
5395 #ifdef CONFIG_SCHED_DEBUG
5396 
5397 static __read_mostly int sched_debug_enabled;
5398 
5399 static int __init sched_debug_setup(char *str)
5400 {
5401         sched_debug_enabled = 1;
5402 
5403         return 0;
5404 }
5405 early_param("sched_debug", sched_debug_setup);
5406 
5407 static inline bool sched_debug(void)
5408 {
5409         return sched_debug_enabled;
5410 }
5411 
5412 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5413                                   struct cpumask *groupmask)
5414 {
5415         struct sched_group *group = sd->groups;
5416 
5417         cpumask_clear(groupmask);
5418 
5419         printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5420 
5421         if (!(sd->flags & SD_LOAD_BALANCE)) {
5422                 printk("does not load-balance\n");
5423                 if (sd->parent)
5424                         printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5425                                         " has parent");
5426                 return -1;
5427         }
5428 
5429         printk(KERN_CONT "span %*pbl level %s\n",
5430                cpumask_pr_args(sched_domain_span(sd)), sd->name);
5431 
5432         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5433                 printk(KERN_ERR "ERROR: domain->span does not contain "
5434                                 "CPU%d\n", cpu);
5435         }
5436         if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5437                 printk(KERN_ERR "ERROR: domain->groups does not contain"
5438                                 " CPU%d\n", cpu);
5439         }
5440 
5441         printk(KERN_DEBUG "%*s groups:", level + 1, "");
5442         do {
5443                 if (!group) {
5444                         printk("\n");
5445                         printk(KERN_ERR "ERROR: group is NULL\n");
5446                         break;
5447                 }
5448 
5449                 /*
5450                  * Even though we initialize ->capacity to something semi-sane,
5451                  * we leave capacity_orig unset. This allows us to detect if
5452                  * domain iteration is still funny without causing /0 traps.
5453                  */
5454                 if (!group->sgc->capacity_orig) {
5455                         printk(KERN_CONT "\n");
5456                         printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5457                         break;
5458                 }
5459 
5460                 if (!cpumask_weight(sched_group_cpus(group))) {
5461                         printk(KERN_CONT "\n");
5462                         printk(KERN_ERR "ERROR: empty group\n");
5463                         break;
5464                 }
5465 
5466                 if (!(sd->flags & SD_OVERLAP) &&
5467                     cpumask_intersects(groupmask, sched_group_cpus(group))) {
5468                         printk(KERN_CONT "\n");
5469                         printk(KERN_ERR "ERROR: repeated CPUs\n");
5470                         break;
5471                 }
5472 
5473                 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5474 
5475                 printk(KERN_CONT " %*pbl",
5476                        cpumask_pr_args(sched_group_cpus(group)));
5477                 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5478                         printk(KERN_CONT " (cpu_capacity = %d)",
5479                                 group->sgc->capacity);
5480                 }
5481 
5482                 group = group->next;
5483         } while (group != sd->groups);
5484         printk(KERN_CONT "\n");
5485 
5486         if (!cpumask_equal(sched_domain_span(sd), groupmask))
5487                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5488 
5489         if (sd->parent &&
5490             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5491                 printk(KERN_ERR "ERROR: parent span is not a superset "
5492                         "of domain->span\n");
5493         return 0;
5494 }
5495 
5496 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5497 {
5498         int level = 0;
5499 
5500         if (!sched_debug_enabled)
5501                 return;
5502 
5503         if (!sd) {
5504                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5505                 return;
5506         }
5507 
5508         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5509 
5510         for (;;) {
5511                 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5512                         break;
5513                 level++;
5514                 sd = sd->parent;
5515                 if (!sd)
5516                         break;
5517         }
5518 }
5519 #else /* !CONFIG_SCHED_DEBUG */
5520 # define sched_domain_debug(sd, cpu) do { } while (0)
5521 static inline bool sched_debug(void)
5522 {
5523         return false;
5524 }
5525 #endif /* CONFIG_SCHED_DEBUG */
5526 
5527 static int sd_degenerate(struct sched_domain *sd)
5528 {
5529         if (cpumask_weight(sched_domain_span(sd)) == 1)
5530                 return 1;
5531 
5532         /* Following flags need at least 2 groups */
5533         if (sd->flags & (SD_LOAD_BALANCE |
5534                          SD_BALANCE_NEWIDLE |
5535                          SD_BALANCE_FORK |
5536                          SD_BALANCE_EXEC |
5537                          SD_SHARE_CPUCAPACITY |
5538                          SD_SHARE_PKG_RESOURCES |
5539                          SD_SHARE_POWERDOMAIN)) {
5540                 if (sd->groups != sd->groups->next)
5541                         return 0;
5542         }
5543 
5544         /* Following flags don't use groups */
5545         if (sd->flags & (SD_WAKE_AFFINE))
5546                 return 0;
5547 
5548         return 1;
5549 }
5550 
5551 static int
5552 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5553 {
5554         unsigned long cflags = sd->flags, pflags = parent->flags;
5555 
5556         if (sd_degenerate(parent))
5557                 return 1;
5558 
5559         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5560                 return 0;
5561 
5562         /* Flags needing groups don't count if only 1 group in parent */
5563         if (parent->groups == parent->groups->next) {
5564                 pflags &= ~(SD_LOAD_BALANCE |
5565                                 SD_BALANCE_NEWIDLE |
5566                                 SD_BALANCE_FORK |
5567                                 SD_BALANCE_EXEC |
5568                                 SD_SHARE_CPUCAPACITY |
5569                                 SD_SHARE_PKG_RESOURCES |
5570                                 SD_PREFER_SIBLING |
5571                                 SD_SHARE_POWERDOMAIN);
5572                 if (nr_node_ids == 1)
5573                         pflags &= ~SD_SERIALIZE;
5574         }
5575         if (~cflags & pflags)
5576                 return 0;
5577 
5578         return 1;
5579 }
5580 
5581 static void free_rootdomain(struct rcu_head *rcu)
5582 {
5583         struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5584 
5585         cpupri_cleanup(&rd->cpupri);
5586         cpudl_cleanup(&rd->cpudl);
5587         free_cpumask_var(rd->dlo_mask);
5588         free_cpumask_var(rd->rto_mask);
5589         free_cpumask_var(rd->online);
5590         free_cpumask_var(rd->span);
5591         kfree(rd);
5592 }
5593 
5594 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5595 {
5596         struct root_domain *old_rd = NULL;
5597         unsigned long flags;
5598 
5599         raw_spin_lock_irqsave(&rq->lock, flags);
5600 
5601         if (rq->rd) {
5602                 old_rd = rq->rd;
5603 
5604                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5605                         set_rq_offline(rq);
5606 
5607                 cpumask_clear_cpu(rq->cpu, old_rd->span);
5608 
5609                 /*
5610                  * If we dont want to free the old_rd yet then
5611                  * set old_rd to NULL to skip the freeing later
5612                  * in this function:
5613                  */
5614                 if (!atomic_dec_and_test(&old_rd->refcount))
5615                         old_rd = NULL;
5616         }
5617 
5618         atomic_inc(&rd->refcount);
5619         rq->rd = rd;
5620 
5621         cpumask_set_cpu(rq->cpu, rd->span);
5622         if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5623                 set_rq_online(rq);
5624 
5625         raw_spin_unlock_irqrestore(&rq->lock, flags);
5626 
5627         if (old_rd)
5628                 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5629 }
5630 
5631 static int init_rootdomain(struct root_domain *rd)
5632 {
5633         memset(rd, 0, sizeof(*rd));
5634 
5635         if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5636                 goto out;
5637         if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5638                 goto free_span;
5639         if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5640                 goto free_online;
5641         if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5642                 goto free_dlo_mask;
5643 
5644         init_dl_bw(&rd->dl_bw);
5645         if (cpudl_init(&rd->cpudl) != 0)
5646                 goto free_dlo_mask;
5647 
5648         if (cpupri_init(&rd->cpupri) != 0)
5649                 goto free_rto_mask;
5650         return 0;
5651 
5652 free_rto_mask:
5653         free_cpumask_var(rd->rto_mask);
5654 free_dlo_mask:
5655         free_cpumask_var(rd->dlo_mask);
5656 free_online:
5657         free_cpumask_var(rd->online);
5658 free_span:
5659         free_cpumask_var(rd->span);
5660 out:
5661         return -ENOMEM;
5662 }
5663 
5664 /*
5665  * By default the system creates a single root-domain with all cpus as
5666  * members (mimicking the global state we have today).
5667  */
5668 struct root_domain def_root_domain;
5669 
5670 static void init_defrootdomain(void)
5671 {
5672         init_rootdomain(&def_root_domain);
5673 
5674         atomic_set(&def_root_domain.refcount, 1);
5675 }
5676 
5677 static struct root_domain *alloc_rootdomain(void)
5678 {
5679         struct root_domain *rd;
5680 
5681         rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5682         if (!rd)
5683                 return NULL;
5684 
5685         if (init_rootdomain(rd) != 0) {
5686                 kfree(rd);
5687                 return NULL;
5688         }
5689 
5690         return rd;
5691 }
5692 
5693 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5694 {
5695         struct sched_group *tmp, *first;
5696 
5697         if (!sg)
5698                 return;
5699 
5700         first = sg;
5701         do {
5702                 tmp = sg->next;
5703 
5704                 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5705                         kfree(sg->sgc);
5706 
5707                 kfree(sg);
5708                 sg = tmp;
5709         } while (sg != first);
5710 }
5711 
5712 static void free_sched_domain(struct rcu_head *rcu)
5713 {
5714         struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5715 
5716         /*
5717          * If its an overlapping domain it has private groups, iterate and
5718          * nuke them all.
5719          */
5720         if (sd->flags & SD_OVERLAP) {
5721                 free_sched_groups(sd->groups, 1);
5722         } else if (atomic_dec_and_test(&sd->groups->ref)) {
5723                 kfree(sd->groups->sgc);
5724                 kfree(sd->groups);
5725         }
5726         kfree(sd);
5727 }
5728 
5729 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5730 {
5731         call_rcu(&sd->rcu, free_sched_domain);
5732 }
5733 
5734 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5735 {
5736         for (; sd; sd = sd->parent)
5737                 destroy_sched_domain(sd, cpu);
5738 }
5739 
5740 /*
5741  * Keep a special pointer to the highest sched_domain that has
5742  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5743  * allows us to avoid some pointer chasing select_idle_sibling().
5744  *
5745  * Also keep a unique ID per domain (we use the first cpu number in
5746  * the cpumask of the domain), this allows us to quickly tell if
5747  * two cpus are in the same cache domain, see cpus_share_cache().
5748  */
5749 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5750 DEFINE_PER_CPU(int, sd_llc_size);
5751 DEFINE_PER_CPU(int, sd_llc_id);
5752 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5753 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5754 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5755 
5756 static void update_top_cache_domain(int cpu)
5757 {
5758         struct sched_domain *sd;
5759         struct sched_domain *busy_sd = NULL;
5760         int id = cpu;
5761         int size = 1;
5762 
5763         sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5764         if (sd) {
5765                 id = cpumask_first(sched_domain_span(sd));
5766                 size = cpumask_weight(sched_domain_span(sd));
5767                 busy_sd = sd->parent; /* sd_busy */
5768         }
5769         rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5770 
5771         rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5772         per_cpu(sd_llc_size, cpu) = size;
5773         per_cpu(sd_llc_id, cpu) = id;
5774 
5775         sd = lowest_flag_domain(cpu, SD_NUMA);
5776         rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5777 
5778         sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5779         rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5780 }
5781 
5782 /*
5783  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5784  * hold the hotplug lock.
5785  */
5786 static void
5787 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5788 {
5789         struct rq *rq = cpu_rq(cpu);
5790         struct sched_domain *tmp;
5791 
5792         /* Remove the sched domains which do not contribute to scheduling. */
5793         for (tmp = sd; tmp; ) {
5794                 struct sched_domain *parent = tmp->parent;
5795                 if (!parent)
5796                         break;
5797 
5798                 if (sd_parent_degenerate(tmp, parent)) {
5799                         tmp->parent = parent->parent;
5800                         if (parent->parent)
5801                                 parent->parent->child = tmp;
5802                         /*
5803                          * Transfer SD_PREFER_SIBLING down in case of a
5804                          * degenerate parent; the spans match for this
5805                          * so the property transfers.
5806                          */
5807                         if (parent->flags & SD_PREFER_SIBLING)
5808                                 tmp->flags |= SD_PREFER_SIBLING;
5809                         destroy_sched_domain(parent, cpu);
5810                 } else
5811                         tmp = tmp->parent;
5812         }
5813 
5814         if (sd && sd_degenerate(sd)) {
5815                 tmp = sd;
5816                 sd = sd->parent;
5817                 destroy_sched_domain(tmp, cpu);
5818                 if (sd)
5819                         sd->child = NULL;
5820         }
5821 
5822         sched_domain_debug(sd, cpu);
5823 
5824         rq_attach_root(rq, rd);
5825         tmp = rq->sd;
5826         rcu_assign_pointer(rq->sd, sd);
5827         destroy_sched_domains(tmp, cpu);
5828 
5829         update_top_cache_domain(cpu);
5830 }
5831 
5832 /* cpus with isolated domains */
5833 static cpumask_var_t cpu_isolated_map;
5834 
5835 /* Setup the mask of cpus configured for isolated domains */
5836 static int __init isolated_cpu_setup(char *str)
5837 {
5838         alloc_bootmem_cpumask_var(&cpu_isolated_map);
5839         cpulist_parse(str, cpu_isolated_map);
5840         return 1;
5841 }
5842 
5843 __setup("isolcpus=", isolated_cpu_setup);
5844 
5845 struct s_data {
5846         struct sched_domain ** __percpu sd;
5847         struct root_domain      *rd;
5848 };
5849 
5850 enum s_alloc {
5851         sa_rootdomain,
5852         sa_sd,
5853         sa_sd_storage,
5854         sa_none,
5855 };
5856 
5857 /*
5858  * Build an iteration mask that can exclude certain CPUs from the upwards
5859  * domain traversal.
5860  *
5861  * Asymmetric node setups can result in situations where the domain tree is of
5862  * unequal depth, make sure to skip domains that already cover the entire
5863  * range.
5864  *
5865  * In that case build_sched_domains() will have terminated the iteration early
5866  * and our sibling sd spans will be empty. Domains should always include the
5867  * cpu they're built on, so check that.
5868  *
5869  */
5870 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5871 {
5872         const struct cpumask *span = sched_domain_span(sd);
5873         struct sd_data *sdd = sd->private;
5874         struct sched_domain *sibling;
5875         int i;
5876 
5877         for_each_cpu(i, span) {
5878                 sibling = *per_cpu_ptr(sdd->sd, i);
5879                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5880                         continue;
5881 
5882                 cpumask_set_cpu(i, sched_group_mask(sg));
5883         }
5884 }
5885 
5886 /*
5887  * Return the canonical balance cpu for this group, this is the first cpu
5888  * of this group that's also in the iteration mask.
5889  */
5890 int group_balance_cpu(struct sched_group *sg)
5891 {
5892         return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5893 }
5894 
5895 static int
5896 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5897 {
5898         struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5899         const struct cpumask *span = sched_domain_span(sd);
5900         struct cpumask *covered = sched_domains_tmpmask;
5901         struct sd_data *sdd = sd->private;
5902         struct sched_domain *sibling;
5903         int i;
5904 
5905         cpumask_clear(covered);
5906 
5907         for_each_cpu(i, span) {
5908                 struct cpumask *sg_span;
5909 
5910                 if (cpumask_test_cpu(i, covered))
5911                         continue;
5912 
5913                 sibling = *per_cpu_ptr(sdd->sd, i);
5914 
5915                 /* See the comment near build_group_mask(). */
5916                 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5917                         continue;
5918 
5919                 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5920                                 GFP_KERNEL, cpu_to_node(cpu));
5921 
5922                 if (!sg)
5923                         goto fail;
5924 
5925                 sg_span = sched_group_cpus(sg);
5926                 if (sibling->child)
5927                         cpumask_copy(sg_span, sched_domain_span(sibling->child));
5928                 else
5929                         cpumask_set_cpu(i, sg_span);
5930 
5931                 cpumask_or(covered, covered, sg_span);
5932 
5933                 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5934                 if (atomic_inc_return(&sg->sgc->ref) == 1)
5935                         build_group_mask(sd, sg);
5936 
5937                 /*
5938                  * Initialize sgc->capacity such that even if we mess up the
5939                  * domains and no possible iteration will get us here, we won't
5940                  * die on a /0 trap.
5941                  */
5942                 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5943                 sg->sgc->capacity_orig = sg->sgc->capacity;
5944 
5945                 /*
5946                  * Make sure the first group of this domain contains the
5947                  * canonical balance cpu. Otherwise the sched_domain iteration
5948                  * breaks. See update_sg_lb_stats().
5949                  */
5950                 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5951                     group_balance_cpu(sg) == cpu)
5952                         groups = sg;
5953 
5954                 if (!first)
5955                         first = sg;
5956                 if (last)
5957                         last->next = sg;
5958                 last = sg;
5959                 last->next = first;
5960         }
5961         sd->groups = groups;
5962 
5963         return 0;
5964 
5965 fail:
5966         free_sched_groups(first, 0);
5967 
5968         return -ENOMEM;
5969 }
5970 
5971 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5972 {
5973         struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5974         struct sched_domain *child = sd->child;
5975 
5976         if (child)
5977                 cpu = cpumask_first(sched_domain_span(child));
5978 
5979         if (sg) {
5980                 *sg = *per_cpu_ptr(sdd->sg, cpu);
5981                 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5982                 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5983         }
5984 
5985         return cpu;
5986 }
5987 
5988 /*
5989  * build_sched_groups will build a circular linked list of the groups
5990  * covered by the given span, and will set each group's ->cpumask correctly,
5991  * and ->cpu_capacity to 0.
5992  *
5993  * Assumes the sched_domain tree is fully constructed
5994  */
5995 static int
5996 build_sched_groups(struct sched_domain *sd, int cpu)
5997 {
5998         struct sched_group *first = NULL, *last = NULL;
5999         struct sd_data *sdd = sd->private;
6000         const struct cpumask *span = sched_domain_span(sd);
6001         struct cpumask *covered;
6002         int i;
6003 
6004         get_group(cpu, sdd, &sd->groups);
6005         atomic_inc(&sd->groups->ref);
6006 
6007         if (cpu != cpumask_first(span))
6008                 return 0;
6009 
6010         lockdep_assert_held(&sched_domains_mutex);
6011         covered = sched_domains_tmpmask;
6012 
6013         cpumask_clear(covered);
6014 
6015         for_each_cpu(i, span) {
6016                 struct sched_group *sg;
6017                 int group, j;
6018 
6019                 if (cpumask_test_cpu(i, covered))
6020                         continue;
6021 
6022                 group = get_group(i, sdd, &sg);
6023                 cpumask_setall(sched_group_mask(sg));
6024 
6025                 for_each_cpu(j, span) {
6026                         if (get_group(j, sdd, NULL) != group)
6027                                 continue;
6028 
6029                         cpumask_set_cpu(j, covered);
6030                         cpumask_set_cpu(j, sched_group_cpus(sg));
6031                 }
6032 
6033                 if (!first)
6034                         first = sg;
6035                 if (last)
6036                         last->next = sg;
6037                 last = sg;
6038         }
6039         last->next = first;
6040 
6041         return 0;
6042 }
6043 
6044 /*
6045  * Initialize sched groups cpu_capacity.
6046  *
6047  * cpu_capacity indicates the capacity of sched group, which is used while
6048  * distributing the load between different sched groups in a sched domain.
6049  * Typically cpu_capacity for all the groups in a sched domain will be same
6050  * unless there are asymmetries in the topology. If there are asymmetries,
6051  * group having more cpu_capacity will pickup more load compared to the
6052  * group having less cpu_capacity.
6053  */
6054 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6055 {
6056         struct sched_group *sg = sd->groups;
6057 
6058         WARN_ON(!sg);
6059 
6060         do {
6061                 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6062                 sg = sg->next;
6063         } while (sg != sd->groups);
6064 
6065         if (cpu != group_balance_cpu(sg))
6066                 return;
6067 
6068         update_group_capacity(sd, cpu);
6069         atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6070 }
6071 
6072 /*
6073  * Initializers for schedule domains
6074  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6075  */
6076 
6077 static int default_relax_domain_level = -1;
6078 int sched_domain_level_max;
6079 
6080 static int __init setup_relax_domain_level(char *str)
6081 {
6082         if (kstrtoint(str, 0, &default_relax_domain_level))
6083                 pr_warn("Unable to set relax_domain_level\n");
6084 
6085         return 1;
6086 }
6087 __setup("relax_domain_level=", setup_relax_domain_level);
6088 
6089 static void set_domain_attribute(struct sched_domain *sd,
6090                                  struct sched_domain_attr *attr)
6091 {
6092         int request;
6093 
6094         if (!attr || attr->relax_domain_level < 0) {
6095                 if (default_relax_domain_level < 0)
6096                         return;
6097                 else
6098                         request = default_relax_domain_level;
6099         } else
6100                 request = attr->relax_domain_level;
6101         if (request < sd->level) {
6102                 /* turn off idle balance on this domain */
6103                 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6104         } else {
6105                 /* turn on idle balance on this domain */
6106                 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6107         }
6108 }
6109 
6110 static void __sdt_free(const struct cpumask *cpu_map);
6111 static int __sdt_alloc(const struct cpumask *cpu_map);
6112 
6113 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6114                                  const struct cpumask *cpu_map)
6115 {
6116         switch (what) {
6117         case sa_rootdomain:
6118                 if (!atomic_read(&d->rd->refcount))
6119                         free_rootdomain(&d->rd->rcu); /* fall through */
6120         case sa_sd:
6121                 free_percpu(d->sd); /* fall through */
6122         case sa_sd_storage:
6123                 __sdt_free(cpu_map); /* fall through */
6124         case sa_none:
6125                 break;
6126         }
6127 }
6128 
6129 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6130                                                    const struct cpumask *cpu_map)
6131 {
6132         memset(d, 0, sizeof(*d));
6133 
6134         if (__sdt_alloc(cpu_map))
6135                 return sa_sd_storage;
6136         d->sd = alloc_percpu(struct sched_domain *);
6137         if (!d->sd)
6138                 return sa_sd_storage;
6139         d->rd = alloc_rootdomain();
6140         if (!d->rd)
6141                 return sa_sd;
6142         return sa_rootdomain;
6143 }
6144 
6145 /*
6146  * NULL the sd_data elements we've used to build the sched_domain and
6147  * sched_group structure so that the subsequent __free_domain_allocs()
6148  * will not free the data we're using.
6149  */
6150 static void claim_allocations(int cpu, struct sched_domain *sd)
6151 {
6152         struct sd_data *sdd = sd->private;
6153 
6154         WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6155         *per_cpu_ptr(sdd->sd, cpu) = NULL;
6156 
6157         if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6158                 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6159 
6160         if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6161                 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6162 }
6163 
6164 #ifdef CONFIG_NUMA
6165 static int sched_domains_numa_levels;
6166 enum numa_topology_type sched_numa_topology_type;
6167 static int *sched_domains_numa_distance;
6168 int sched_max_numa_distance;
6169 static struct cpumask ***sched_domains_numa_masks;
6170 static int sched_domains_curr_level;
6171 #endif
6172 
6173 /*
6174  * SD_flags allowed in topology descriptions.
6175  *
6176  * SD_SHARE_CPUCAPACITY      - describes SMT topologies
6177  * SD_SHARE_PKG_RESOURCES - describes shared caches
6178  * SD_NUMA                - describes NUMA topologies
6179  * SD_SHARE_POWERDOMAIN   - describes shared power domain
6180  *
6181  * Odd one out:
6182  * SD_ASYM_PACKING        - describes SMT quirks
6183  */
6184 #define TOPOLOGY_SD_FLAGS               \
6185         (SD_SHARE_CPUCAPACITY |         \
6186          SD_SHARE_PKG_RESOURCES |       \
6187          SD_NUMA |                      \
6188          SD_ASYM_PACKING |              \
6189          SD_SHARE_POWERDOMAIN)
6190 
6191 static struct sched_domain *
6192 sd_init(struct sched_domain_topology_level *tl, int cpu)
6193 {
6194         struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6195         int sd_weight, sd_flags = 0;
6196 
6197 #ifdef CONFIG_NUMA
6198         /*
6199          * Ugly hack to pass state to sd_numa_mask()...
6200          */
6201         sched_domains_curr_level = tl->numa_level;
6202 #endif
6203 
6204         sd_weight = cpumask_weight(tl->mask(cpu));
6205 
6206         if (tl->sd_flags)
6207                 sd_flags = (*tl->sd_flags)();
6208         if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6209                         "wrong sd_flags in topology description\n"))
6210                 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6211 
6212         *sd = (struct sched_domain){
6213                 .min_interval           = sd_weight,
6214                 .max_interval           = 2*sd_weight,
6215                 .busy_factor            = 32,
6216                 .imbalance_pct          = 125,
6217 
6218                 .cache_nice_tries       = 0,
6219                 .busy_idx               = 0,
6220                 .idle_idx               = 0,
6221                 .newidle_idx            = 0,
6222                 .wake_idx               = 0,
6223                 .forkexec_idx           = 0,
6224 
6225                 .flags                  = 1*SD_LOAD_BALANCE
6226                                         | 1*SD_BALANCE_NEWIDLE
6227                                         | 1*SD_BALANCE_EXEC
6228                                         | 1*SD_BALANCE_FORK
6229                                         | 0*SD_BALANCE_WAKE
6230                                         | 1*SD_WAKE_AFFINE
6231                                         | 0*SD_SHARE_CPUCAPACITY
6232                                         | 0*SD_SHARE_PKG_RESOURCES
6233                                         | 0*SD_SERIALIZE
6234                                         | 0*SD_PREFER_SIBLING
6235                                         | 0*SD_NUMA
6236                                         | sd_flags
6237                                         ,
6238 
6239                 .last_balance           = jiffies,
6240                 .balance_interval       = sd_weight,
6241                 .smt_gain               = 0,
6242                 .max_newidle_lb_cost    = 0,
6243                 .next_decay_max_lb_cost = jiffies,
6244 #ifdef CONFIG_SCHED_DEBUG
6245                 .name                   = tl->name,
6246 #endif
6247         };
6248 
6249         /*
6250          * Convert topological properties into behaviour.
6251          */
6252 
6253         if (sd->flags & SD_SHARE_CPUCAPACITY) {
6254                 sd->imbalance_pct = 110;
6255                 sd->smt_gain = 1178; /* ~15% */
6256 
6257         } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6258                 sd->imbalance_pct = 117;
6259                 sd->cache_nice_tries = 1;
6260                 sd->busy_idx = 2;
6261 
6262 #ifdef CONFIG_NUMA
6263         } else if (sd->flags & SD_NUMA) {
6264                 sd->cache_nice_tries = 2;
6265                 sd->busy_idx = 3;
6266                 sd->idle_idx = 2;
6267 
6268                 sd->flags |= SD_SERIALIZE;
6269                 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6270                         sd->flags &= ~(SD_BALANCE_EXEC |
6271                                        SD_BALANCE_FORK |
6272                                        SD_WAKE_AFFINE);
6273                 }
6274 
6275 #endif
6276         } else {
6277                 sd->flags |= SD_PREFER_SIBLING;
6278                 sd->cache_nice_tries = 1;
6279                 sd->busy_idx = 2;
6280                 sd->idle_idx = 1;
6281         }
6282 
6283         sd->private = &tl->data;
6284 
6285         return sd;
6286 }
6287 
6288 /*
6289  * Topology list, bottom-up.
6290  */
6291 static struct sched_domain_topology_level default_topology[] = {
6292 #ifdef CONFIG_SCHED_SMT
6293         { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6294 #endif
6295 #ifdef CONFIG_SCHED_MC
6296         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6297 #endif
6298         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6299         { NULL, },
6300 };
6301 
6302 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6303 
6304 #define for_each_sd_topology(tl)                        \
6305         for (tl = sched_domain_topology; tl->mask; tl++)
6306 
6307 void set_sched_topology(struct sched_domain_topology_level *tl)
6308 {
6309         sched_domain_topology = tl;
6310 }
6311 
6312 #ifdef CONFIG_NUMA
6313 
6314 static const struct cpumask *sd_numa_mask(int cpu)
6315 {
6316         return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6317 }
6318 
6319 static void sched_numa_warn(const char *str)
6320 {
6321         static int done = false;
6322         int i,j;
6323 
6324         if (done)
6325                 return;
6326 
6327         done = true;
6328 
6329         printk(KERN_WARNING "ERROR: %s\n\n", str);
6330 
6331         for (i = 0; i < nr_node_ids; i++) {
6332                 printk(KERN_WARNING "  ");
6333                 for (j = 0; j < nr_node_ids; j++)
6334                         printk(KERN_CONT "%02d ", node_distance(i,j));
6335                 printk(KERN_CONT "\n");
6336         }
6337         printk(KERN_WARNING "\n");
6338 }
6339 
6340 bool find_numa_distance(int distance)
6341 {
6342         int i;
6343 
6344         if (distance == node_distance(0, 0))
6345                 return true;
6346 
6347         for (i = 0; i < sched_domains_numa_levels; i++) {
6348                 if (sched_domains_numa_distance[i] == distance)
6349                         return true;
6350         }
6351 
6352         return false;
6353 }
6354 
6355 /*
6356  * A system can have three types of NUMA topology:
6357  * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6358  * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6359  * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6360  *
6361  * The difference between a glueless mesh topology and a backplane
6362  * topology lies in whether communication between not directly
6363  * connected nodes goes through intermediary nodes (where programs
6364  * could run), or through backplane controllers. This affects
6365  * placement of programs.
6366  *
6367  * The type of topology can be discerned with the following tests:
6368  * - If the maximum distance between any nodes is 1 hop, the system
6369  *   is directly connected.
6370  * - If for two nodes A and B, located N > 1 hops away from each other,
6371  *   there is an intermediary node C, which is < N hops away from both
6372  *   nodes A and B, the system is a glueless mesh.
6373  */
6374 static void init_numa_topology_type(void)
6375 {
6376         int a, b, c, n;
6377 
6378         n = sched_max_numa_distance;
6379 
6380         if (n <= 1)
6381                 sched_numa_topology_type = NUMA_DIRECT;
6382 
6383         for_each_online_node(a) {
6384                 for_each_online_node(b) {
6385                         /* Find two nodes furthest removed from each other. */
6386                         if (node_distance(a, b) < n)
6387                                 continue;
6388 
6389                         /* Is there an intermediary node between a and b? */
6390                         for_each_online_node(c) {
6391                                 if (node_distance(a, c) < n &&
6392                                     node_distance(b, c) < n) {
6393                                         sched_numa_topology_type =
6394                                                         NUMA_GLUELESS_MESH;
6395                                         return;
6396                                 }
6397                         }
6398 
6399                         sched_numa_topology_type = NUMA_BACKPLANE;
6400                         return;
6401                 }
6402         }
6403 }
6404 
6405 static void sched_init_numa(void)
6406 {
6407         int next_distance, curr_distance = node_distance(0, 0);