Linux kernel & device driver programming

Cross-Referenced Linux and Device Driver Code

[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ]
Version: [ 2.6.11.8 ] [ 2.6.25 ] [ 2.6.25.8 ] [ 2.6.31.13 ] Architecture: [ i386 ]
  1 /*
  2  *  kernel/sched.c
  3  *
  4  *  Kernel scheduler and related syscalls
  5  *
  6  *  Copyright (C) 1991-2002  Linus Torvalds
  7  *
  8  *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
  9  *              make semaphores SMP safe
 10  *  1998-11-19  Implemented schedule_timeout() and related stuff
 11  *              by Andrea Arcangeli
 12  *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
 13  *              hybrid priority-list and round-robin design with
 14  *              an array-switch method of distributing timeslices
 15  *              and per-CPU runqueues.  Cleanups and useful suggestions
 16  *              by Davide Libenzi, preemptible kernel bits by Robert Love.
 17  *  2003-09-03  Interactivity tuning by Con Kolivas.
 18  *  2004-04-02  Scheduler domains code by Nick Piggin
 19  *  2007-04-15  Work begun on replacing all interactivity tuning with a
 20  *              fair scheduling design by Con Kolivas.
 21  *  2007-05-05  Load balancing (smp-nice) and other improvements
 22  *              by Peter Williams
 23  *  2007-05-06  Interactivity improvements to CFS by Mike Galbraith
 24  *  2007-07-01  Group scheduling enhancements by Srivatsa Vaddagiri
 25  *  2007-11-29  RT balancing improvements by Steven Rostedt, Gregory Haskins,
 26  *              Thomas Gleixner, Mike Kravetz
 27  */
 28 
 29 #include <linux/mm.h>
 30 #include <linux/module.h>
 31 #include <linux/nmi.h>
 32 #include <linux/init.h>
 33 #include <linux/uaccess.h>
 34 #include <linux/highmem.h>
 35 #include <linux/smp_lock.h>
 36 #include <asm/mmu_context.h>
 37 #include <linux/interrupt.h>
 38 #include <linux/capability.h>
 39 #include <linux/completion.h>
 40 #include <linux/kernel_stat.h>
 41 #include <linux/debug_locks.h>
 42 #include <linux/perf_counter.h>
 43 #include <linux/security.h>
 44 #include <linux/notifier.h>
 45 #include <linux/profile.h>
 46 #include <linux/freezer.h>
 47 #include <linux/vmalloc.h>
 48 #include <linux/blkdev.h>
 49 #include <linux/delay.h>
 50 #include <linux/pid_namespace.h>
 51 #include <linux/smp.h>
 52 #include <linux/threads.h>
 53 #include <linux/timer.h>
 54 #include <linux/rcupdate.h>
 55 #include <linux/cpu.h>
 56 #include <linux/cpuset.h>
 57 #include <linux/percpu.h>
 58 #include <linux/kthread.h>
 59 #include <linux/proc_fs.h>
 60 #include <linux/seq_file.h>
 61 #include <linux/sysctl.h>
 62 #include <linux/syscalls.h>
 63 #include <linux/times.h>
 64 #include <linux/tsacct_kern.h>
 65 #include <linux/kprobes.h>
 66 #include <linux/delayacct.h>
 67 #include <linux/reciprocal_div.h>
 68 #include <linux/unistd.h>
 69 #include <linux/pagemap.h>
 70 #include <linux/hrtimer.h>
 71 #include <linux/tick.h>
 72 #include <linux/debugfs.h>
 73 #include <linux/ctype.h>
 74 #include <linux/ftrace.h>
 75 
 76 #include <asm/tlb.h>
 77 #include <asm/irq_regs.h>
 78 
 79 #include "sched_cpupri.h"
 80 
 81 #define CREATE_TRACE_POINTS
 82 #include <trace/events/sched.h>
 83 
 84 /*
 85  * Convert user-nice values [ -20 ... 0 ... 19 ]
 86  * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
 87  * and back.
 88  */
 89 #define NICE_TO_PRIO(nice)      (MAX_RT_PRIO + (nice) + 20)
 90 #define PRIO_TO_NICE(prio)      ((prio) - MAX_RT_PRIO - 20)
 91 #define TASK_NICE(p)            PRIO_TO_NICE((p)->static_prio)
 92 
 93 /*
 94  * 'User priority' is the nice value converted to something we
 95  * can work with better when scaling various scheduler parameters,
 96  * it's a [ 0 ... 39 ] range.
 97  */
 98 #define USER_PRIO(p)            ((p)-MAX_RT_PRIO)
 99 #define TASK_USER_PRIO(p)       USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO           (USER_PRIO(MAX_PRIO))
101 
102 /*
103  * Helpers for converting nanosecond timing to jiffy resolution
104  */
105 #define NS_TO_JIFFIES(TIME)     ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 
107 #define NICE_0_LOAD             SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT            SCHED_LOAD_SHIFT
109 
110 /*
111  * These are the 'tuning knobs' of the scheduler:
112  *
113  * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114  * Timeslices get refilled after they expire.
115  */
116 #define DEF_TIMESLICE           (100 * HZ / 1000)
117 
118 /*
119  * single value that denotes runtime == period, ie unlimited time.
120  */
121 #define RUNTIME_INF     ((u64)~0ULL)
122 
123 #ifdef CONFIG_SMP
124 
125 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
126 
127 /*
128  * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129  * Since cpu_power is a 'constant', we can use a reciprocal divide.
130  */
131 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
132 {
133         return reciprocal_divide(load, sg->reciprocal_cpu_power);
134 }
135 
136 /*
137  * Each time a sched group cpu_power is changed,
138  * we must compute its reciprocal value
139  */
140 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
141 {
142         sg->__cpu_power += val;
143         sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
144 }
145 #endif
146 
147 static inline int rt_policy(int policy)
148 {
149         if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
150                 return 1;
151         return 0;
152 }
153 
154 static inline int task_has_rt_policy(struct task_struct *p)
155 {
156         return rt_policy(p->policy);
157 }
158 
159 /*
160  * This is the priority-queue data structure of the RT scheduling class:
161  */
162 struct rt_prio_array {
163         DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
164         struct list_head queue[MAX_RT_PRIO];
165 };
166 
167 struct rt_bandwidth {
168         /* nests inside the rq lock: */
169         spinlock_t              rt_runtime_lock;
170         ktime_t                 rt_period;
171         u64                     rt_runtime;
172         struct hrtimer          rt_period_timer;
173 };
174 
175 static struct rt_bandwidth def_rt_bandwidth;
176 
177 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
178 
179 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
180 {
181         struct rt_bandwidth *rt_b =
182                 container_of(timer, struct rt_bandwidth, rt_period_timer);
183         ktime_t now;
184         int overrun;
185         int idle = 0;
186 
187         for (;;) {
188                 now = hrtimer_cb_get_time(timer);
189                 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
190 
191                 if (!overrun)
192                         break;
193 
194                 idle = do_sched_rt_period_timer(rt_b, overrun);
195         }
196 
197         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
198 }
199 
200 static
201 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
202 {
203         rt_b->rt_period = ns_to_ktime(period);
204         rt_b->rt_runtime = runtime;
205 
206         spin_lock_init(&rt_b->rt_runtime_lock);
207 
208         hrtimer_init(&rt_b->rt_period_timer,
209                         CLOCK_MONOTONIC, HRTIMER_MODE_REL);
210         rt_b->rt_period_timer.function = sched_rt_period_timer;
211 }
212 
213 static inline int rt_bandwidth_enabled(void)
214 {
215         return sysctl_sched_rt_runtime >= 0;
216 }
217 
218 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
219 {
220         ktime_t now;
221 
222         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
223                 return;
224 
225         if (hrtimer_active(&rt_b->rt_period_timer))
226                 return;
227 
228         spin_lock(&rt_b->rt_runtime_lock);
229         for (;;) {
230                 unsigned long delta;
231                 ktime_t soft, hard;
232 
233                 if (hrtimer_active(&rt_b->rt_period_timer))
234                         break;
235 
236                 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
237                 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
238 
239                 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
240                 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
241                 delta = ktime_to_ns(ktime_sub(hard, soft));
242                 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
243                                 HRTIMER_MODE_ABS_PINNED, 0);
244         }
245         spin_unlock(&rt_b->rt_runtime_lock);
246 }
247 
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
250 {
251         hrtimer_cancel(&rt_b->rt_period_timer);
252 }
253 #endif
254 
255 /*
256  * sched_domains_mutex serializes calls to arch_init_sched_domains,
257  * detach_destroy_domains and partition_sched_domains.
258  */
259 static DEFINE_MUTEX(sched_domains_mutex);
260 
261 #ifdef CONFIG_GROUP_SCHED
262 
263 #include <linux/cgroup.h>
264 
265 struct cfs_rq;
266 
267 static LIST_HEAD(task_groups);
268 
269 /* task group related information */
270 struct task_group {
271 #ifdef CONFIG_CGROUP_SCHED
272         struct cgroup_subsys_state css;
273 #endif
274 
275 #ifdef CONFIG_USER_SCHED
276         uid_t uid;
277 #endif
278 
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280         /* schedulable entities of this group on each cpu */
281         struct sched_entity **se;
282         /* runqueue "owned" by this group on each cpu */
283         struct cfs_rq **cfs_rq;
284         unsigned long shares;
285 #endif
286 
287 #ifdef CONFIG_RT_GROUP_SCHED
288         struct sched_rt_entity **rt_se;
289         struct rt_rq **rt_rq;
290 
291         struct rt_bandwidth rt_bandwidth;
292 #endif
293 
294         struct rcu_head rcu;
295         struct list_head list;
296 
297         struct task_group *parent;
298         struct list_head siblings;
299         struct list_head children;
300 };
301 
302 #ifdef CONFIG_USER_SCHED
303 
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct *user)
306 {
307         user->tg->uid = user->uid;
308 }
309 
310 /*
311  * Root task group.
312  *      Every UID task group (including init_task_group aka UID-0) will
313  *      be a child to this group.
314  */
315 struct task_group root_task_group;
316 
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
323 
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
326 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
331 
332 /* task_group_lock serializes add/remove of task groups and also changes to
333  * a task group's cpu shares.
334  */
335 static DEFINE_SPINLOCK(task_group_lock);
336 
337 #ifdef CONFIG_SMP
338 static int root_task_group_empty(void)
339 {
340         return list_empty(&root_task_group.children);
341 }
342 #endif
343 
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD   (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD   NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
350 
351 /*
352  * A weight of 0 or 1 can cause arithmetics problems.
353  * A weight of a cfs_rq is the sum of weights of which entities
354  * are queued on this cfs_rq, so a weight of a entity should not be
355  * too large, so as the shares value of a task group.
356  * (The default weight is 1024 - so there's no practical
357  *  limitation from this.)
358  */
359 #define MIN_SHARES      2
360 #define MAX_SHARES      (1UL << 18)
361 
362 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
363 #endif
364 
365 /* Default task group.
366  *      Every task in system belong to this group at bootup.
367  */
368 struct task_group init_task_group;
369 
370 /* return group to which a task belongs */
371 static inline struct task_group *task_group(struct task_struct *p)
372 {
373         struct task_group *tg;
374 
375 #ifdef CONFIG_USER_SCHED
376         rcu_read_lock();
377         tg = __task_cred(p)->user->tg;
378         rcu_read_unlock();
379 #elif defined(CONFIG_CGROUP_SCHED)
380         tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
381                                 struct task_group, css);
382 #else
383         tg = &init_task_group;
384 #endif
385         return tg;
386 }
387 
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
390 {
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392         p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
393         p->se.parent = task_group(p)->se[cpu];
394 #endif
395 
396 #ifdef CONFIG_RT_GROUP_SCHED
397         p->rt.rt_rq  = task_group(p)->rt_rq[cpu];
398         p->rt.parent = task_group(p)->rt_se[cpu];
399 #endif
400 }
401 
402 #else
403 
404 #ifdef CONFIG_SMP
405 static int root_task_group_empty(void)
406 {
407         return 1;
408 }
409 #endif
410 
411 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
412 static inline struct task_group *task_group(struct task_struct *p)
413 {
414         return NULL;
415 }
416 
417 #endif  /* CONFIG_GROUP_SCHED */
418 
419 /* CFS-related fields in a runqueue */
420 struct cfs_rq {
421         struct load_weight load;
422         unsigned long nr_running;
423 
424         u64 exec_clock;
425         u64 min_vruntime;
426 
427         struct rb_root tasks_timeline;
428         struct rb_node *rb_leftmost;
429 
430         struct list_head tasks;
431         struct list_head *balance_iterator;
432 
433         /*
434          * 'curr' points to currently running entity on this cfs_rq.
435          * It is set to NULL otherwise (i.e when none are currently running).
436          */
437         struct sched_entity *curr, *next, *last;
438 
439         unsigned int nr_spread_over;
440 
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442         struct rq *rq;  /* cpu runqueue to which this cfs_rq is attached */
443 
444         /*
445          * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446          * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447          * (like users, containers etc.)
448          *
449          * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450          * list is used during load balance.
451          */
452         struct list_head leaf_cfs_rq_list;
453         struct task_group *tg;  /* group that "owns" this runqueue */
454 
455 #ifdef CONFIG_SMP
456         /*
457          * the part of load.weight contributed by tasks
458          */
459         unsigned long task_weight;
460 
461         /*
462          *   h_load = weight * f(tg)
463          *
464          * Where f(tg) is the recursive weight fraction assigned to
465          * this group.
466          */
467         unsigned long h_load;
468 
469         /*
470          * this cpu's part of tg->shares
471          */
472         unsigned long shares;
473 
474         /*
475          * load.weight at the time we set shares
476          */
477         unsigned long rq_weight;
478 #endif
479 #endif
480 };
481 
482 /* Real-Time classes' related field in a runqueue: */
483 struct rt_rq {
484         struct rt_prio_array active;
485         unsigned long rt_nr_running;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
487         struct {
488                 int curr; /* highest queued rt task prio */
489 #ifdef CONFIG_SMP
490                 int next; /* next highest */
491 #endif
492         } highest_prio;
493 #endif
494 #ifdef CONFIG_SMP
495         unsigned long rt_nr_migratory;
496         unsigned long rt_nr_total;
497         int overloaded;
498         struct plist_head pushable_tasks;
499 #endif
500         int rt_throttled;
501         u64 rt_time;
502         u64 rt_runtime;
503         /* Nests inside the rq lock: */
504         spinlock_t rt_runtime_lock;
505 
506 #ifdef CONFIG_RT_GROUP_SCHED
507         unsigned long rt_nr_boosted;
508 
509         struct rq *rq;
510         struct list_head leaf_rt_rq_list;
511         struct task_group *tg;
512         struct sched_rt_entity *rt_se;
513 #endif
514 };
515 
516 #ifdef CONFIG_SMP
517 
518 /*
519  * We add the notion of a root-domain which will be used to define per-domain
520  * variables. Each exclusive cpuset essentially defines an island domain by
521  * fully partitioning the member cpus from any other cpuset. Whenever a new
522  * exclusive cpuset is created, we also create and attach a new root-domain
523  * object.
524  *
525  */
526 struct root_domain {
527         atomic_t refcount;
528         cpumask_var_t span;
529         cpumask_var_t online;
530 
531         /*
532          * The "RT overload" flag: it gets set if a CPU has more than
533          * one runnable RT task.
534          */
535         cpumask_var_t rto_mask;
536         atomic_t rto_count;
537 #ifdef CONFIG_SMP
538         struct cpupri cpupri;
539 #endif
540 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
541         /*
542          * Preferred wake up cpu nominated by sched_mc balance that will be
543          * used when most cpus are idle in the system indicating overall very
544          * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
545          */
546         unsigned int sched_mc_preferred_wakeup_cpu;
547 #endif
548 };
549 
550 /*
551  * By default the system creates a single root-domain with all cpus as
552  * members (mimicking the global state we have today).
553  */
554 static struct root_domain def_root_domain;
555 
556 #endif
557 
558 /*
559  * This is the main, per-CPU runqueue data structure.
560  *
561  * Locking rule: those places that want to lock multiple runqueues
562  * (such as the load balancing or the thread migration code), lock
563  * acquire operations must be ordered by ascending &runqueue.
564  */
565 struct rq {
566         /* runqueue lock: */
567         spinlock_t lock;
568 
569         /*
570          * nr_running and cpu_load should be in the same cacheline because
571          * remote CPUs use both these fields when doing load calculation.
572          */
573         unsigned long nr_running;
574         #define CPU_LOAD_IDX_MAX 5
575         unsigned long cpu_load[CPU_LOAD_IDX_MAX];
576 #ifdef CONFIG_NO_HZ
577         unsigned long last_tick_seen;
578         unsigned char in_nohz_recently;
579 #endif
580         /* capture load from *all* tasks on this cpu: */
581         struct load_weight load;
582         unsigned long nr_load_updates;
583         u64 nr_switches;
584         u64 nr_migrations_in;
585 
586         struct cfs_rq cfs;
587         struct rt_rq rt;
588 
589 #ifdef CONFIG_FAIR_GROUP_SCHED
590         /* list of leaf cfs_rq on this cpu: */
591         struct list_head leaf_cfs_rq_list;
592 #endif
593 #ifdef CONFIG_RT_GROUP_SCHED
594         struct list_head leaf_rt_rq_list;
595 #endif
596 
597         /*
598          * This is part of a global counter where only the total sum
599          * over all CPUs matters. A task can increase this counter on
600          * one CPU and if it got migrated afterwards it may decrease
601          * it on another CPU. Always updated under the runqueue lock:
602          */
603         unsigned long nr_uninterruptible;
604 
605         struct task_struct *curr, *idle;
606         unsigned long next_balance;
607         struct mm_struct *prev_mm;
608 
609         u64 clock;
610 
611         atomic_t nr_iowait;
612 
613 #ifdef CONFIG_SMP
614         struct root_domain *rd;
615         struct sched_domain *sd;
616 
617         unsigned char idle_at_tick;
618         /* For active balancing */
619         int active_balance;
620         int push_cpu;
621         /* cpu of this runqueue: */
622         int cpu;
623         int online;
624 
625         unsigned long avg_load_per_task;
626 
627         struct task_struct *migration_thread;
628         struct list_head migration_queue;
629 #endif
630 
631         /* calc_load related fields */
632         unsigned long calc_load_update;
633         long calc_load_active;
634 
635 #ifdef CONFIG_SCHED_HRTICK
636 #ifdef CONFIG_SMP
637         int hrtick_csd_pending;
638         struct call_single_data hrtick_csd;
639 #endif
640         struct hrtimer hrtick_timer;
641 #endif
642 
643 #ifdef CONFIG_SCHEDSTATS
644         /* latency stats */
645         struct sched_info rq_sched_info;
646         unsigned long long rq_cpu_time;
647         /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
648 
649         /* sys_sched_yield() stats */
650         unsigned int yld_count;
651 
652         /* schedule() stats */
653         unsigned int sched_switch;
654         unsigned int sched_count;
655         unsigned int sched_goidle;
656 
657         /* try_to_wake_up() stats */
658         unsigned int ttwu_count;
659         unsigned int ttwu_local;
660 
661         /* BKL stats */
662         unsigned int bkl_count;
663 #endif
664 };
665 
666 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
667 
668 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
669 {
670         rq->curr->sched_class->check_preempt_curr(rq, p, sync);
671 }
672 
673 static inline int cpu_of(struct rq *rq)
674 {
675 #ifdef CONFIG_SMP
676         return rq->cpu;
677 #else
678         return 0;
679 #endif
680 }
681 
682 /*
683  * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
684  * See detach_destroy_domains: synchronize_sched for details.
685  *
686  * The domain tree of any CPU may only be accessed from within
687  * preempt-disabled sections.
688  */
689 #define for_each_domain(cpu, __sd) \
690         for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
691 
692 #define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
693 #define this_rq()               (&__get_cpu_var(runqueues))
694 #define task_rq(p)              cpu_rq(task_cpu(p))
695 #define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
696 
697 inline void update_rq_clock(struct rq *rq)
698 {
699         rq->clock = sched_clock_cpu(cpu_of(rq));
700 }
701 
702 /*
703  * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
704  */
705 #ifdef CONFIG_SCHED_DEBUG
706 # define const_debug __read_mostly
707 #else
708 # define const_debug static const
709 #endif
710 
711 /**
712  * runqueue_is_locked
713  *
714  * Returns true if the current cpu runqueue is locked.
715  * This interface allows printk to be called with the runqueue lock
716  * held and know whether or not it is OK to wake up the klogd.
717  */
718 int runqueue_is_locked(void)
719 {
720         int cpu = get_cpu();
721         struct rq *rq = cpu_rq(cpu);
722         int ret;
723 
724         ret = spin_is_locked(&rq->lock);
725         put_cpu();
726         return ret;
727 }
728 
729 /*
730  * Debugging: various feature bits
731  */
732 
733 #define SCHED_FEAT(name, enabled)       \
734         __SCHED_FEAT_##name ,
735 
736 enum {
737 #include "sched_features.h"
738 };
739 
740 #undef SCHED_FEAT
741 
742 #define SCHED_FEAT(name, enabled)       \
743         (1UL << __SCHED_FEAT_##name) * enabled |
744 
745 const_debug unsigned int sysctl_sched_features =
746 #include "sched_features.h"
747         0;
748 
749 #undef SCHED_FEAT
750 
751 #ifdef CONFIG_SCHED_DEBUG
752 #define SCHED_FEAT(name, enabled)       \
753         #name ,
754 
755 static __read_mostly char *sched_feat_names[] = {
756 #include "sched_features.h"
757         NULL
758 };
759 
760 #undef SCHED_FEAT
761 
762 static int sched_feat_show(struct seq_file *m, void *v)
763 {
764         int i;
765 
766         for (i = 0; sched_feat_names[i]; i++) {
767                 if (!(sysctl_sched_features & (1UL << i)))
768                         seq_puts(m, "NO_");
769                 seq_printf(m, "%s ", sched_feat_names[i]);
770         }
771         seq_puts(m, "\n");
772 
773         return 0;
774 }
775 
776 static ssize_t
777 sched_feat_write(struct file *filp, const char __user *ubuf,
778                 size_t cnt, loff_t *ppos)
779 {
780         char buf[64];
781         char *cmp = buf;
782         int neg = 0;
783         int i;
784 
785         if (cnt > 63)
786                 cnt = 63;
787 
788         if (copy_from_user(&buf, ubuf, cnt))
789                 return -EFAULT;
790 
791         buf[cnt] = 0;
792 
793         if (strncmp(buf, "NO_", 3) == 0) {
794                 neg = 1;
795                 cmp += 3;
796         }
797 
798         for (i = 0; sched_feat_names[i]; i++) {
799                 int len = strlen(sched_feat_names[i]);
800 
801                 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
802                         if (neg)
803                                 sysctl_sched_features &= ~(1UL << i);
804                         else
805                                 sysctl_sched_features |= (1UL << i);
806                         break;
807                 }
808         }
809 
810         if (!sched_feat_names[i])
811                 return -EINVAL;
812 
813         filp->f_pos += cnt;
814 
815         return cnt;
816 }
817 
818 static int sched_feat_open(struct inode *inode, struct file *filp)
819 {
820         return single_open(filp, sched_feat_show, NULL);
821 }
822 
823 static struct file_operations sched_feat_fops = {
824         .open           = sched_feat_open,
825         .write          = sched_feat_write,
826         .read           = seq_read,
827         .llseek         = seq_lseek,
828         .release        = single_release,
829 };
830 
831 static __init int sched_init_debug(void)
832 {
833         debugfs_create_file("sched_features", 0644, NULL, NULL,
834                         &sched_feat_fops);
835 
836         return 0;
837 }
838 late_initcall(sched_init_debug);
839 
840 #endif
841 
842 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
843 
844 /*
845  * Number of tasks to iterate in a single balance run.
846  * Limited because this is done with IRQs disabled.
847  */
848 const_debug unsigned int sysctl_sched_nr_migrate = 32;
849 
850 /*
851  * ratelimit for updating the group shares.
852  * default: 0.25ms
853  */
854 unsigned int sysctl_sched_shares_ratelimit = 250000;
855 
856 /*
857  * Inject some fuzzyness into changing the per-cpu group shares
858  * this avoids remote rq-locks at the expense of fairness.
859  * default: 4
860  */
861 unsigned int sysctl_sched_shares_thresh = 4;
862 
863 /*
864  * period over which we measure -rt task cpu usage in us.
865  * default: 1s
866  */
867 unsigned int sysctl_sched_rt_period = 1000000;
868 
869 static __read_mostly int scheduler_running;
870 
871 /*
872  * part of the period that we allow rt tasks to run in us.
873  * default: 0.95s
874  */
875 int sysctl_sched_rt_runtime = 950000;
876 
877 static inline u64 global_rt_period(void)
878 {
879         return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
880 }
881 
882 static inline u64 global_rt_runtime(void)
883 {
884         if (sysctl_sched_rt_runtime < 0)
885                 return RUNTIME_INF;
886 
887         return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
888 }
889 
890 #ifndef prepare_arch_switch
891 # define prepare_arch_switch(next)      do { } while (0)
892 #endif
893 #ifndef finish_arch_switch
894 # define finish_arch_switch(prev)       do { } while (0)
895 #endif
896 
897 static inline int task_current(struct rq *rq, struct task_struct *p)
898 {
899         return rq->curr == p;
900 }
901 
902 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
903 static inline int task_running(struct rq *rq, struct task_struct *p)
904 {
905         return task_current(rq, p);
906 }
907 
908 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
909 {
910 }
911 
912 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 {
914 #ifdef CONFIG_DEBUG_SPINLOCK
915         /* this is a valid case when another task releases the spinlock */
916         rq->lock.owner = current;
917 #endif
918         /*
919          * If we are tracking spinlock dependencies then we have to
920          * fix up the runqueue lock - which gets 'carried over' from
921          * prev into current:
922          */
923         spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
924 
925         spin_unlock_irq(&rq->lock);
926 }
927 
928 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
929 static inline int task_running(struct rq *rq, struct task_struct *p)
930 {
931 #ifdef CONFIG_SMP
932         return p->oncpu;
933 #else
934         return task_current(rq, p);
935 #endif
936 }
937 
938 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
939 {
940 #ifdef CONFIG_SMP
941         /*
942          * We can optimise this out completely for !SMP, because the
943          * SMP rebalancing from interrupt is the only thing that cares
944          * here.
945          */
946         next->oncpu = 1;
947 #endif
948 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
949         spin_unlock_irq(&rq->lock);
950 #else
951         spin_unlock(&rq->lock);
952 #endif
953 }
954 
955 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
956 {
957 #ifdef CONFIG_SMP
958         /*
959          * After ->oncpu is cleared, the task can be moved to a different CPU.
960          * We must ensure this doesn't happen until the switch is completely
961          * finished.
962          */
963         smp_wmb();
964         prev->oncpu = 0;
965 #endif
966 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
967         local_irq_enable();
968 #endif
969 }
970 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
971 
972 /*
973  * __task_rq_lock - lock the runqueue a given task resides on.
974  * Must be called interrupts disabled.
975  */
976 static inline struct rq *__task_rq_lock(struct task_struct *p)
977         __acquires(rq->lock)
978 {
979         for (;;) {
980                 struct rq *rq = task_rq(p);
981                 spin_lock(&rq->lock);
982                 if (likely(rq == task_rq(p)))
983                         return rq;
984                 spin_unlock(&rq->lock);
985         }
986 }
987 
988 /*
989  * task_rq_lock - lock the runqueue a given task resides on and disable
990  * interrupts. Note the ordering: we can safely lookup the task_rq without
991  * explicitly disabling preemption.
992  */
993 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
994         __acquires(rq->lock)
995 {
996         struct rq *rq;
997 
998         for (;;) {
999                 local_irq_save(*flags);
1000                 rq = task_rq(p);
1001                 spin_lock(&rq->lock);
1002                 if (likely(rq == task_rq(p)))
1003                         return rq;
1004                 spin_unlock_irqrestore(&rq->lock, *flags);
1005         }
1006 }
1007 
1008 void task_rq_unlock_wait(struct task_struct *p)
1009 {
1010         struct rq *rq = task_rq(p);
1011 
1012         smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1013         spin_unlock_wait(&rq->lock);
1014 }
1015 
1016 static void __task_rq_unlock(struct rq *rq)
1017         __releases(rq->lock)
1018 {
1019         spin_unlock(&rq->lock);
1020 }
1021 
1022 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1023         __releases(rq->lock)
1024 {
1025         spin_unlock_irqrestore(&rq->lock, *flags);
1026 }
1027 
1028 /*
1029  * this_rq_lock - lock this runqueue and disable interrupts.
1030  */
1031 static struct rq *this_rq_lock(void)
1032         __acquires(rq->lock)
1033 {
1034         struct rq *rq;
1035 
1036         local_irq_disable();
1037         rq = this_rq();
1038         spin_lock(&rq->lock);
1039 
1040         return rq;
1041 }
1042 
1043 #ifdef CONFIG_SCHED_HRTICK
1044 /*
1045  * Use HR-timers to deliver accurate preemption points.
1046  *
1047  * Its all a bit involved since we cannot program an hrt while holding the
1048  * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1049  * reschedule event.
1050  *
1051  * When we get rescheduled we reprogram the hrtick_timer outside of the
1052  * rq->lock.
1053  */
1054 
1055 /*
1056  * Use hrtick when:
1057  *  - enabled by features
1058  *  - hrtimer is actually high res
1059  */
1060 static inline int hrtick_enabled(struct rq *rq)
1061 {
1062         if (!sched_feat(HRTICK))
1063                 return 0;
1064         if (!cpu_active(cpu_of(rq)))
1065                 return 0;
1066         return hrtimer_is_hres_active(&rq->hrtick_timer);
1067 }
1068 
1069 static void hrtick_clear(struct rq *rq)
1070 {
1071         if (hrtimer_active(&rq->hrtick_timer))
1072                 hrtimer_cancel(&rq->hrtick_timer);
1073 }
1074 
1075 /*
1076  * High-resolution timer tick.
1077  * Runs from hardirq context with interrupts disabled.
1078  */
1079 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1080 {
1081         struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1082 
1083         WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1084 
1085         spin_lock(&rq->lock);
1086         update_rq_clock(rq);
1087         rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1088         spin_unlock(&rq->lock);
1089 
1090         return HRTIMER_NORESTART;
1091 }
1092 
1093 #ifdef CONFIG_SMP
1094 /*
1095  * called from hardirq (IPI) context
1096  */
1097 static void __hrtick_start(void *arg)
1098 {
1099         struct rq *rq = arg;
1100 
1101         spin_lock(&rq->lock);
1102         hrtimer_restart(&rq->hrtick_timer);
1103         rq->hrtick_csd_pending = 0;
1104         spin_unlock(&rq->lock);
1105 }
1106 
1107 /*
1108  * Called to set the hrtick timer state.
1109  *
1110  * called with rq->lock held and irqs disabled
1111  */
1112 static void hrtick_start(struct rq *rq, u64 delay)
1113 {
1114         struct hrtimer *timer = &rq->hrtick_timer;
1115         ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1116 
1117         hrtimer_set_expires(timer, time);
1118 
1119         if (rq == this_rq()) {
1120                 hrtimer_restart(timer);
1121         } else if (!rq->hrtick_csd_pending) {
1122                 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1123                 rq->hrtick_csd_pending = 1;
1124         }
1125 }
1126 
1127 static int
1128 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1129 {
1130         int cpu = (int)(long)hcpu;
1131 
1132         switch (action) {
1133         case CPU_UP_CANCELED:
1134         case CPU_UP_CANCELED_FROZEN:
1135         case CPU_DOWN_PREPARE:
1136         case CPU_DOWN_PREPARE_FROZEN:
1137         case CPU_DEAD:
1138         case CPU_DEAD_FROZEN:
1139                 hrtick_clear(cpu_rq(cpu));
1140                 return NOTIFY_OK;
1141         }
1142 
1143         return NOTIFY_DONE;
1144 }
1145 
1146 static __init void init_hrtick(void)
1147 {
1148         hotcpu_notifier(hotplug_hrtick, 0);
1149 }
1150 #else
1151 /*
1152  * Called to set the hrtick timer state.
1153  *
1154  * called with rq->lock held and irqs disabled
1155  */
1156 static void hrtick_start(struct rq *rq, u64 delay)
1157 {
1158         __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1159                         HRTIMER_MODE_REL_PINNED, 0);
1160 }
1161 
1162 static inline void init_hrtick(void)
1163 {
1164 }
1165 #endif /* CONFIG_SMP */
1166 
1167 static void init_rq_hrtick(struct rq *rq)
1168 {
1169 #ifdef CONFIG_SMP
1170         rq->hrtick_csd_pending = 0;
1171 
1172         rq->hrtick_csd.flags = 0;
1173         rq->hrtick_csd.func = __hrtick_start;
1174         rq->hrtick_csd.info = rq;
1175 #endif
1176 
1177         hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1178         rq->hrtick_timer.function = hrtick;
1179 }
1180 #else   /* CONFIG_SCHED_HRTICK */
1181 static inline void hrtick_clear(struct rq *rq)
1182 {
1183 }
1184 
1185 static inline void init_rq_hrtick(struct rq *rq)
1186 {
1187 }
1188 
1189 static inline void init_hrtick(void)
1190 {
1191 }
1192 #endif  /* CONFIG_SCHED_HRTICK */
1193 
1194 /*
1195  * resched_task - mark a task 'to be rescheduled now'.
1196  *
1197  * On UP this means the setting of the need_resched flag, on SMP it
1198  * might also involve a cross-CPU call to trigger the scheduler on
1199  * the target CPU.
1200  */
1201 #ifdef CONFIG_SMP
1202 
1203 #ifndef tsk_is_polling
1204 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1205 #endif
1206 
1207 static void resched_task(struct task_struct *p)
1208 {
1209         int cpu;
1210 
1211         assert_spin_locked(&task_rq(p)->lock);
1212 
1213         if (test_tsk_need_resched(p))
1214                 return;
1215 
1216         set_tsk_need_resched(p);
1217 
1218         cpu = task_cpu(p);
1219         if (cpu == smp_processor_id())
1220                 return;
1221 
1222         /* NEED_RESCHED must be visible before we test polling */
1223         smp_mb();
1224         if (!tsk_is_polling(p))
1225                 smp_send_reschedule(cpu);
1226 }
1227 
1228 static void resched_cpu(int cpu)
1229 {
1230         struct rq *rq = cpu_rq(cpu);
1231         unsigned long flags;
1232 
1233         if (!spin_trylock_irqsave(&rq->lock, flags))
1234                 return;
1235         resched_task(cpu_curr(cpu));
1236         spin_unlock_irqrestore(&rq->lock, flags);
1237 }
1238 
1239 #ifdef CONFIG_NO_HZ
1240 /*
1241  * When add_timer_on() enqueues a timer into the timer wheel of an
1242  * idle CPU then this timer might expire before the next timer event
1243  * which is scheduled to wake up that CPU. In case of a completely
1244  * idle system the next event might even be infinite time into the
1245  * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1246  * leaves the inner idle loop so the newly added timer is taken into
1247  * account when the CPU goes back to idle and evaluates the timer
1248  * wheel for the next timer event.
1249  */
1250 void wake_up_idle_cpu(int cpu)
1251 {
1252         struct rq *rq = cpu_rq(cpu);
1253 
1254         if (cpu == smp_processor_id())
1255                 return;
1256 
1257         /*
1258          * This is safe, as this function is called with the timer
1259          * wheel base lock of (cpu) held. When the CPU is on the way
1260          * to idle and has not yet set rq->curr to idle then it will
1261          * be serialized on the timer wheel base lock and take the new
1262          * timer into account automatically.
1263          */
1264         if (rq->curr != rq->idle)
1265                 return;
1266 
1267         /*
1268          * We can set TIF_RESCHED on the idle task of the other CPU
1269          * lockless. The worst case is that the other CPU runs the
1270          * idle task through an additional NOOP schedule()
1271          */
1272         set_tsk_need_resched(rq->idle);
1273 
1274         /* NEED_RESCHED must be visible before we test polling */
1275         smp_mb();
1276         if (!tsk_is_polling(rq->idle))
1277                 smp_send_reschedule(cpu);
1278 }
1279 #endif /* CONFIG_NO_HZ */
1280 
1281 #else /* !CONFIG_SMP */
1282 static void resched_task(struct task_struct *p)
1283 {
1284         assert_spin_locked(&task_rq(p)->lock);
1285         set_tsk_need_resched(p);
1286 }
1287 #endif /* CONFIG_SMP */
1288 
1289 #if BITS_PER_LONG == 32
1290 # define WMULT_CONST    (~0UL)
1291 #else
1292 # define WMULT_CONST    (1UL << 32)
1293 #endif
1294 
1295 #define WMULT_SHIFT     32
1296 
1297 /*
1298  * Shift right and round:
1299  */
1300 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1301 
1302 /*
1303  * delta *= weight / lw
1304  */
1305 static unsigned long
1306 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1307                 struct load_weight *lw)
1308 {
1309         u64 tmp;
1310 
1311         if (!lw->inv_weight) {
1312                 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1313                         lw->inv_weight = 1;
1314                 else
1315                         lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1316                                 / (lw->weight+1);
1317         }
1318 
1319         tmp = (u64)delta_exec * weight;
1320         /*
1321          * Check whether we'd overflow the 64-bit multiplication:
1322          */
1323         if (unlikely(tmp > WMULT_CONST))
1324                 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1325                         WMULT_SHIFT/2);
1326         else
1327                 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1328 
1329         return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1330 }
1331 
1332 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1333 {
1334         lw->weight += inc;
1335         lw->inv_weight = 0;
1336 }
1337 
1338 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1339 {
1340         lw->weight -= dec;
1341         lw->inv_weight = 0;
1342 }
1343 
1344 /*
1345  * To aid in avoiding the subversion of "niceness" due to uneven distribution
1346  * of tasks with abnormal "nice" values across CPUs the contribution that
1347  * each task makes to its run queue's load is weighted according to its
1348  * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1349  * scaled version of the new time slice allocation that they receive on time
1350  * slice expiry etc.
1351  */
1352 
1353 #define WEIGHT_IDLEPRIO                3
1354 #define WMULT_IDLEPRIO         1431655765
1355 
1356 /*
1357  * Nice levels are multiplicative, with a gentle 10% change for every
1358  * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1359  * nice 1, it will get ~10% less CPU time than another CPU-bound task
1360  * that remained on nice 0.
1361  *
1362  * The "10% effect" is relative and cumulative: from _any_ nice level,
1363  * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1364  * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1365  * If a task goes up by ~10% and another task goes down by ~10% then
1366  * the relative distance between them is ~25%.)
1367  */
1368 static const int prio_to_weight[40] = {
1369  /* -20 */     88761,     71755,     56483,     46273,     36291,
1370  /* -15 */     29154,     23254,     18705,     14949,     11916,
1371  /* -10 */      9548,      7620,      6100,      4904,      3906,
1372  /*  -5 */      3121,      2501,      1991,      1586,      1277,
1373  /*   0 */      1024,       820,       655,       526,       423,
1374  /*   5 */       335,       272,       215,       172,       137,
1375  /*  10 */       110,        87,        70,        56,        45,
1376  /*  15 */        36,        29,        23,        18,        15,
1377 };
1378 
1379 /*
1380  * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1381  *
1382  * In cases where the weight does not change often, we can use the
1383  * precalculated inverse to speed up arithmetics by turning divisions
1384  * into multiplications:
1385  */
1386 static const u32 prio_to_wmult[40] = {
1387  /* -20 */     48388,     59856,     76040,     92818,    118348,
1388  /* -15 */    147320,    184698,    229616,    287308,    360437,
1389  /* -10 */    449829,    563644,    704093,    875809,   1099582,
1390  /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
1391  /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
1392  /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
1393  /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
1394  /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1395 };
1396 
1397 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1398 
1399 /*
1400  * runqueue iterator, to support SMP load-balancing between different
1401  * scheduling classes, without having to expose their internal data
1402  * structures to the load-balancing proper:
1403  */
1404 struct rq_iterator {
1405         void *arg;
1406         struct task_struct *(*start)(void *);
1407         struct task_struct *(*next)(void *);
1408 };
1409 
1410 #ifdef CONFIG_SMP
1411 static unsigned long
1412 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1413               unsigned long max_load_move, struct sched_domain *sd,
1414               enum cpu_idle_type idle, int *all_pinned,
1415               int *this_best_prio, struct rq_iterator *iterator);
1416 
1417 static int
1418 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1419                    struct sched_domain *sd, enum cpu_idle_type idle,
1420                    struct rq_iterator *iterator);
1421 #endif
1422 
1423 /* Time spent by the tasks of the cpu accounting group executing in ... */
1424 enum cpuacct_stat_index {
1425         CPUACCT_STAT_USER,      /* ... user mode */
1426         CPUACCT_STAT_SYSTEM,    /* ... kernel mode */
1427 
1428         CPUACCT_STAT_NSTATS,
1429 };
1430 
1431 #ifdef CONFIG_CGROUP_CPUACCT
1432 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1433 static void cpuacct_update_stats(struct task_struct *tsk,
1434                 enum cpuacct_stat_index idx, cputime_t val);
1435 #else
1436 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1437 static inline void cpuacct_update_stats(struct task_struct *tsk,
1438                 enum cpuacct_stat_index idx, cputime_t val) {}
1439 #endif
1440 
1441 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1442 {
1443         update_load_add(&rq->load, load);
1444 }
1445 
1446 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1447 {
1448         update_load_sub(&rq->load, load);
1449 }
1450 
1451 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1452 typedef int (*tg_visitor)(struct task_group *, void *);
1453 
1454 /*
1455  * Iterate the full tree, calling @down when first entering a node and @up when
1456  * leaving it for the final time.
1457  */
1458 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1459 {
1460         struct task_group *parent, *child;
1461         int ret;
1462 
1463         rcu_read_lock();
1464         parent = &root_task_group;
1465 down:
1466         ret = (*down)(parent, data);
1467         if (ret)
1468                 goto out_unlock;
1469         list_for_each_entry_rcu(child, &parent->children, siblings) {
1470                 parent = child;
1471                 goto down;
1472 
1473 up:
1474                 continue;
1475         }
1476         ret = (*up)(parent, data);
1477         if (ret)
1478                 goto out_unlock;
1479 
1480         child = parent;
1481         parent = parent->parent;
1482         if (parent)
1483                 goto up;
1484 out_unlock:
1485         rcu_read_unlock();
1486 
1487         return ret;
1488 }
1489 
1490 static int tg_nop(struct task_group *tg, void *data)
1491 {
1492         return 0;
1493 }
1494 #endif
1495 
1496 #ifdef CONFIG_SMP
1497 static unsigned long source_load(int cpu, int type);
1498 static unsigned long target_load(int cpu, int type);
1499 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1500 
1501 static unsigned long cpu_avg_load_per_task(int cpu)
1502 {
1503         struct rq *rq = cpu_rq(cpu);
1504         unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1505 
1506         if (nr_running)
1507                 rq->avg_load_per_task = rq->load.weight / nr_running;
1508         else
1509                 rq->avg_load_per_task = 0;
1510 
1511         return rq->avg_load_per_task;
1512 }
1513 
1514 #ifdef CONFIG_FAIR_GROUP_SCHED
1515 
1516 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1517 
1518 /*
1519  * Calculate and set the cpu's group shares.
1520  */
1521 static void
1522 update_group_shares_cpu(struct task_group *tg, int cpu,
1523                         unsigned long sd_shares, unsigned long sd_rq_weight)
1524 {
1525         unsigned long shares;
1526         unsigned long rq_weight;
1527 
1528         if (!tg->se[cpu])
1529                 return;
1530 
1531         rq_weight = tg->cfs_rq[cpu]->rq_weight;
1532 
1533         /*
1534          *           \Sum shares * rq_weight
1535          * shares =  -----------------------
1536          *               \Sum rq_weight
1537          *
1538          */
1539         shares = (sd_shares * rq_weight) / sd_rq_weight;
1540         shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1541 
1542         if (abs(shares - tg->se[cpu]->load.weight) >
1543                         sysctl_sched_shares_thresh) {
1544                 struct rq *rq = cpu_rq(cpu);
1545                 unsigned long flags;
1546 
1547                 spin_lock_irqsave(&rq->lock, flags);
1548                 tg->cfs_rq[cpu]->shares = shares;
1549 
1550                 __set_se_shares(tg->se[cpu], shares);
1551                 spin_unlock_irqrestore(&rq->lock, flags);
1552         }
1553 }
1554 
1555 /*
1556  * Re-compute the task group their per cpu shares over the given domain.
1557  * This needs to be done in a bottom-up fashion because the rq weight of a
1558  * parent group depends on the shares of its child groups.
1559  */
1560 static int tg_shares_up(struct task_group *tg, void *data)
1561 {
1562         unsigned long weight, rq_weight = 0;
1563         unsigned long shares = 0;
1564         struct sched_domain *sd = data;
1565         int i;
1566 
1567         for_each_cpu(i, sched_domain_span(sd)) {
1568                 /*
1569                  * If there are currently no tasks on the cpu pretend there
1570                  * is one of average load so that when a new task gets to
1571                  * run here it will not get delayed by group starvation.
1572                  */
1573                 weight = tg->cfs_rq[i]->load.weight;
1574                 if (!weight)
1575                         weight = NICE_0_LOAD;
1576 
1577                 tg->cfs_rq[i]->rq_weight = weight;
1578                 rq_weight += weight;
1579                 shares += tg->cfs_rq[i]->shares;
1580         }
1581 
1582         if ((!shares && rq_weight) || shares > tg->shares)
1583                 shares = tg->shares;
1584 
1585         if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1586                 shares = tg->shares;
1587 
1588         for_each_cpu(i, sched_domain_span(sd))
1589                 update_group_shares_cpu(tg, i, shares, rq_weight);
1590 
1591         return 0;
1592 }
1593 
1594 /*
1595  * Compute the cpu's hierarchical load factor for each task group.
1596  * This needs to be done in a top-down fashion because the load of a child
1597  * group is a fraction of its parents load.
1598  */
1599 static int tg_load_down(struct task_group *tg, void *data)
1600 {
1601         unsigned long load;
1602         long cpu = (long)data;
1603 
1604         if (!tg->parent) {
1605                 load = cpu_rq(cpu)->load.weight;
1606         } else {
1607                 load = tg->parent->cfs_rq[cpu]->h_load;
1608                 load *= tg->cfs_rq[cpu]->shares;
1609                 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1610         }
1611 
1612         tg->cfs_rq[cpu]->h_load = load;
1613 
1614         return 0;
1615 }
1616 
1617 static void update_shares(struct sched_domain *sd)
1618 {
1619         u64 now = cpu_clock(raw_smp_processor_id());
1620         s64 elapsed = now - sd->last_update;
1621 
1622         if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1623                 sd->last_update = now;
1624                 walk_tg_tree(tg_nop, tg_shares_up, sd);
1625         }
1626 }
1627 
1628 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1629 {
1630         spin_unlock(&rq->lock);
1631         update_shares(sd);
1632         spin_lock(&rq->lock);
1633 }
1634 
1635 static void update_h_load(long cpu)
1636 {
1637         walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1638 }
1639 
1640 #else
1641 
1642 static inline void update_shares(struct sched_domain *sd)
1643 {
1644 }
1645 
1646 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1647 {
1648 }
1649 
1650 #endif
1651 
1652 #ifdef CONFIG_PREEMPT
1653 
1654 /*
1655  * fair double_lock_balance: Safely acquires both rq->locks in a fair
1656  * way at the expense of forcing extra atomic operations in all
1657  * invocations.  This assures that the double_lock is acquired using the
1658  * same underlying policy as the spinlock_t on this architecture, which
1659  * reduces latency compared to the unfair variant below.  However, it
1660  * also adds more overhead and therefore may reduce throughput.
1661  */
1662 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1663         __releases(this_rq->lock)
1664         __acquires(busiest->lock)
1665         __acquires(this_rq->lock)
1666 {
1667         spin_unlock(&this_rq->lock);
1668         double_rq_lock(this_rq, busiest);
1669 
1670         return 1;
1671 }
1672 
1673 #else
1674 /*
1675  * Unfair double_lock_balance: Optimizes throughput at the expense of
1676  * latency by eliminating extra atomic operations when the locks are
1677  * already in proper order on entry.  This favors lower cpu-ids and will
1678  * grant the double lock to lower cpus over higher ids under contention,
1679  * regardless of entry order into the function.
1680  */
1681 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1682         __releases(this_rq->lock)
1683         __acquires(busiest->lock)
1684         __acquires(this_rq->lock)
1685 {
1686         int ret = 0;
1687 
1688         if (unlikely(!spin_trylock(&busiest->lock))) {
1689                 if (busiest < this_rq) {
1690                         spin_unlock(&this_rq->lock);
1691                         spin_lock(&busiest->lock);
1692                         spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1693                         ret = 1;
1694                 } else
1695                         spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1696         }
1697         return ret;
1698 }
1699 
1700 #endif /* CONFIG_PREEMPT */
1701 
1702 /*
1703  * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1704  */
1705 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1706 {
1707         if (unlikely(!irqs_disabled())) {
1708                 /* printk() doesn't work good under rq->lock */
1709                 spin_unlock(&this_rq->lock);
1710                 BUG_ON(1);
1711         }
1712 
1713         return _double_lock_balance(this_rq, busiest);
1714 }
1715 
1716 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1717         __releases(busiest->lock)
1718 {
1719         spin_unlock(&busiest->lock);
1720         lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1721 }
1722 #endif
1723 
1724 #ifdef CONFIG_FAIR_GROUP_SCHED
1725 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1726 {
1727 #ifdef CONFIG_SMP
1728         cfs_rq->shares = shares;
1729 #endif
1730 }
1731 #endif
1732 
1733 static void calc_load_account_active(struct rq *this_rq);
1734 
1735 #include "sched_stats.h"
1736 #include "sched_idletask.c"
1737 #include "sched_fair.c"
1738 #include "sched_rt.c"
1739 #ifdef CONFIG_SCHED_DEBUG
1740 # include "sched_debug.c"
1741 #endif
1742 
1743 #define sched_class_highest (&rt_sched_class)
1744 #define for_each_class(class) \
1745    for (class = sched_class_highest; class; class = class->next)
1746 
1747 static void inc_nr_running(struct rq *rq)
1748 {
1749         rq->nr_running++;
1750 }
1751 
1752 static void dec_nr_running(struct rq *rq)
1753 {
1754         rq->nr_running--;
1755 }
1756 
1757 static void set_load_weight(struct task_struct *p)
1758 {
1759         if (task_has_rt_policy(p)) {
1760                 p->se.load.weight = prio_to_weight[0] * 2;
1761                 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1762                 return;
1763         }
1764 
1765         /*
1766          * SCHED_IDLE tasks get minimal weight:
1767          */
1768         if (p->policy == SCHED_IDLE) {
1769                 p->se.load.weight = WEIGHT_IDLEPRIO;
1770                 p->se.load.inv_weight = WMULT_IDLEPRIO;
1771                 return;
1772         }
1773 
1774         p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1775         p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1776 }
1777 
1778 static void update_avg(u64 *avg, u64 sample)
1779 {
1780         s64 diff = sample - *avg;
1781         *avg += diff >> 3;
1782 }
1783 
1784 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1785 {
1786         if (wakeup)
1787                 p->se.start_runtime = p->se.sum_exec_runtime;
1788 
1789         sched_info_queued(p);
1790         p->sched_class->enqueue_task(rq, p, wakeup);
1791         p->se.on_rq = 1;
1792 }
1793 
1794 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1795 {
1796         if (sleep) {
1797                 if (p->se.last_wakeup) {
1798                         update_avg(&p->se.avg_overlap,
1799                                 p->se.sum_exec_runtime - p->se.last_wakeup);
1800                         p->se.last_wakeup = 0;
1801                 } else {
1802                         update_avg(&p->se.avg_wakeup,
1803                                 sysctl_sched_wakeup_granularity);
1804                 }
1805         }
1806 
1807         sched_info_dequeued(p);
1808         p->sched_class->dequeue_task(rq, p, sleep);
1809         p->se.on_rq = 0;
1810 }
1811 
1812 /*
1813  * __normal_prio - return the priority that is based on the static prio
1814  */
1815 static inline int __normal_prio(struct task_struct *p)
1816 {
1817         return p->static_prio;
1818 }
1819 
1820 /*
1821  * Calculate the expected normal priority: i.e. priority
1822  * without taking RT-inheritance into account. Might be
1823  * boosted by interactivity modifiers. Changes upon fork,
1824  * setprio syscalls, and whenever the interactivity
1825  * estimator recalculates.
1826  */
1827 static inline int normal_prio(struct task_struct *p)
1828 {
1829         int prio;
1830 
1831         if (task_has_rt_policy(p))
1832                 prio = MAX_RT_PRIO-1 - p->rt_priority;
1833         else
1834                 prio = __normal_prio(p);
1835         return prio;
1836 }
1837 
1838 /*
1839  * Calculate the current priority, i.e. the priority
1840  * taken into account by the scheduler. This value might
1841  * be boosted by RT tasks, or might be boosted by
1842  * interactivity modifiers. Will be RT if the task got
1843  * RT-boosted. If not then it returns p->normal_prio.
1844  */
1845 static int effective_prio(struct task_struct *p)
1846 {
1847         p->normal_prio = normal_prio(p);
1848         /*
1849          * If we are RT tasks or we were boosted to RT priority,
1850          * keep the priority unchanged. Otherwise, update priority
1851          * to the normal priority:
1852          */
1853         if (!rt_prio(p->prio))
1854                 return p->normal_prio;
1855         return p->prio;
1856 }
1857 
1858 /*
1859  * activate_task - move a task to the runqueue.
1860  */
1861 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1862 {
1863         if (task_contributes_to_load(p))
1864                 rq->nr_uninterruptible--;
1865 
1866         enqueue_task(rq, p, wakeup);
1867         inc_nr_running(rq);
1868 }
1869 
1870 /*
1871  * deactivate_task - remove a task from the runqueue.
1872  */
1873 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1874 {
1875         if (task_contributes_to_load(p))
1876                 rq->nr_uninterruptible++;
1877 
1878         dequeue_task(rq, p, sleep);
1879         dec_nr_running(rq);
1880 }
1881 
1882 /**
1883  * task_curr - is this task currently executing on a CPU?
1884  * @p: the task in question.
1885  */
1886 inline int task_curr(const struct task_struct *p)
1887 {
1888         return cpu_curr(task_cpu(p)) == p;
1889 }
1890 
1891 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1892 {
1893         set_task_rq(p, cpu);
1894 #ifdef CONFIG_SMP
1895         /*
1896          * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1897          * successfuly executed on another CPU. We must ensure that updates of
1898          * per-task data have been completed by this moment.
1899          */
1900         smp_wmb();
1901         task_thread_info(p)->cpu = cpu;
1902 #endif
1903 }
1904 
1905 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1906                                        const struct sched_class *prev_class,
1907                                        int oldprio, int running)
1908 {
1909         if (prev_class != p->sched_class) {
1910                 if (prev_class->switched_from)
1911                         prev_class->switched_from(rq, p, running);
1912                 p->sched_class->switched_to(rq, p, running);
1913         } else
1914                 p->sched_class->prio_changed(rq, p, oldprio, running);
1915 }
1916 
1917 #ifdef CONFIG_SMP
1918 
1919 /* Used instead of source_load when we know the type == 0 */
1920 static unsigned long weighted_cpuload(const int cpu)
1921 {
1922         return cpu_rq(cpu)->load.weight;
1923 }
1924 
1925 /*
1926  * Is this task likely cache-hot:
1927  */
1928 static int
1929 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1930 {
1931         s64 delta;
1932 
1933         /*
1934          * Buddy candidates are cache hot:
1935          */
1936         if (sched_feat(CACHE_HOT_BUDDY) &&
1937                         (&p->se == cfs_rq_of(&p->se)->next ||
1938                          &p->se == cfs_rq_of(&p->se)->last))
1939                 return 1;
1940 
1941         if (p->sched_class != &fair_sched_class)
1942                 return 0;
1943 
1944         if (sysctl_sched_migration_cost == -1)
1945                 return 1;
1946         if (sysctl_sched_migration_cost == 0)
1947                 return 0;
1948 
1949         delta = now - p->se.exec_start;
1950 
1951         return delta < (s64)sysctl_sched_migration_cost;
1952 }
1953 
1954 
1955 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1956 {
1957         int old_cpu = task_cpu(p);
1958         struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1959         struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1960                       *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1961         u64 clock_offset;
1962 
1963         clock_offset = old_rq->clock - new_rq->clock;
1964 
1965         trace_sched_migrate_task(p, new_cpu);
1966 
1967 #ifdef CONFIG_SCHEDSTATS
1968         if (p->se.wait_start)
1969                 p->se.wait_start -= clock_offset;
1970         if (p->se.sleep_start)
1971                 p->se.sleep_start -= clock_offset;
1972         if (p->se.block_start)
1973                 p->se.block_start -= clock_offset;
1974 #endif
1975         if (old_cpu != new_cpu) {
1976                 p->se.nr_migrations++;
1977                 new_rq->nr_migrations_in++;
1978 #ifdef CONFIG_SCHEDSTATS
1979                 if (task_hot(p, old_rq->clock, NULL))
1980                         schedstat_inc(p, se.nr_forced2_migrations);
1981 #endif
1982                 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
1983                                      1, 1, NULL, 0);
1984         }
1985         p->se.vruntime -= old_cfsrq->min_vruntime -
1986                                          new_cfsrq->min_vruntime;
1987 
1988         __set_task_cpu(p, new_cpu);
1989 }
1990 
1991 struct migration_req {
1992         struct list_head list;
1993 
1994         struct task_struct *task;
1995         int dest_cpu;
1996 
1997         struct completion done;
1998 };
1999 
2000 /*
2001  * The task's runqueue lock must be held.
2002  * Returns true if you have to wait for migration thread.
2003  */
2004 static int
2005 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2006 {
2007         struct rq *rq = task_rq(p);
2008 
2009         /*
2010          * If the task is not on a runqueue (and not running), then
2011          * it is sufficient to simply update the task's cpu field.
2012          */
2013         if (!p->se.on_rq && !task_running(rq, p)) {
2014                 set_task_cpu(p, dest_cpu);
2015                 return 0;
2016         }
2017 
2018         init_completion(&req->done);
2019         req->task = p;
2020         req->dest_cpu = dest_cpu;
2021         list_add(&req->list, &rq->migration_queue);
2022 
2023         return 1;
2024 }
2025 
2026 /*
2027  * wait_task_context_switch -   wait for a thread to complete at least one
2028  *                              context switch.
2029  *
2030  * @p must not be current.
2031  */
2032 void wait_task_context_switch(struct task_struct *p)
2033 {
2034         unsigned long nvcsw, nivcsw, flags;
2035         int running;
2036         struct rq *rq;
2037 
2038         nvcsw   = p->nvcsw;
2039         nivcsw  = p->nivcsw;
2040         for (;;) {
2041                 /*
2042                  * The runqueue is assigned before the actual context
2043                  * switch. We need to take the runqueue lock.
2044                  *
2045                  * We could check initially without the lock but it is
2046                  * very likely that we need to take the lock in every
2047                  * iteration.
2048                  */
2049                 rq = task_rq_lock(p, &flags);
2050                 running = task_running(rq, p);
2051                 task_rq_unlock(rq, &flags);
2052 
2053                 if (likely(!running))
2054                         break;
2055                 /*
2056                  * The switch count is incremented before the actual
2057                  * context switch. We thus wait for two switches to be
2058                  * sure at least one completed.
2059                  */
2060                 if ((p->nvcsw - nvcsw) > 1)
2061                         break;
2062                 if ((p->nivcsw - nivcsw) > 1)
2063                         break;
2064 
2065                 cpu_relax();
2066         }
2067 }
2068 
2069 /*
2070  * wait_task_inactive - wait for a thread to unschedule.
2071  *
2072  * If @match_state is nonzero, it's the @p->state value just checked and
2073  * not expected to change.  If it changes, i.e. @p might have woken up,
2074  * then return zero.  When we succeed in waiting for @p to be off its CPU,
2075  * we return a positive number (its total switch count).  If a second call
2076  * a short while later returns the same number, the caller can be sure that
2077  * @p has remained unscheduled the whole time.
2078  *
2079  * The caller must ensure that the task *will* unschedule sometime soon,
2080  * else this function might spin for a *long* time. This function can't
2081  * be called with interrupts off, or it may introduce deadlock with
2082  * smp_call_function() if an IPI is sent by the same process we are
2083  * waiting to become inactive.
2084  */
2085 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2086 {
2087         unsigned long flags;
2088         int running, on_rq;
2089         unsigned long ncsw;
2090         struct rq *rq;
2091 
2092         for (;;) {
2093                 /*
2094                  * We do the initial early heuristics without holding
2095                  * any task-queue locks at all. We'll only try to get
2096                  * the runqueue lock when things look like they will
2097                  * work out!
2098                  */
2099                 rq = task_rq(p);
2100 
2101                 /*
2102                  * If the task is actively running on another CPU
2103                  * still, just relax and busy-wait without holding
2104                  * any locks.
2105                  *
2106                  * NOTE! Since we don't hold any locks, it's not
2107                  * even sure that "rq" stays as the right runqueue!
2108                  * But we don't care, since "task_running()" will
2109                  * return false if the runqueue has changed and p
2110                  * is actually now running somewhere else!
2111                  */
2112                 while (task_running(rq, p)) {
2113                         if (match_state && unlikely(p->state != match_state))
2114                                 return 0;
2115                         cpu_relax();
2116                 }
2117 
2118                 /*
2119                  * Ok, time to look more closely! We need the rq
2120                  * lock now, to be *sure*. If we're wrong, we'll
2121                  * just go back and repeat.
2122                  */
2123                 rq = task_rq_lock(p, &flags);
2124                 trace_sched_wait_task(rq, p);
2125                 running = task_running(rq, p);
2126                 on_rq = p->se.on_rq;
2127                 ncsw = 0;
2128                 if (!match_state || p->state == match_state)
2129                         ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2130                 task_rq_unlock(rq, &flags);
2131 
2132                 /*
2133                  * If it changed from the expected state, bail out now.
2134                  */
2135                 if (unlikely(!ncsw))
2136                         break;
2137 
2138                 /*
2139                  * Was it really running after all now that we
2140                  * checked with the proper locks actually held?
2141                  *
2142                  * Oops. Go back and try again..
2143                  */
2144                 if (unlikely(running)) {
2145                         cpu_relax();
2146                         continue;
2147                 }
2148 
2149                 /*
2150                  * It's not enough that it's not actively running,
2151                  * it must be off the runqueue _entirely_, and not
2152                  * preempted!
2153                  *
2154                  * So if it was still runnable (but just not actively
2155                  * running right now), it's preempted, and we should
2156                  * yield - it could be a while.
2157                  */
2158                 if (unlikely(on_rq)) {
2159                         schedule_timeout_uninterruptible(1);
2160                         continue;
2161                 }
2162 
2163                 /*
2164                  * Ahh, all good. It wasn't running, and it wasn't
2165                  * runnable, which means that it will never become
2166                  * running in the future either. We're all done!
2167                  */
2168                 break;
2169         }
2170 
2171         return ncsw;
2172 }
2173 
2174 /***
2175  * kick_process - kick a running thread to enter/exit the kernel
2176  * @p: the to-be-kicked thread
2177  *
2178  * Cause a process which is running on another CPU to enter
2179  * kernel-mode, without any delay. (to get signals handled.)
2180  *
2181  * NOTE: this function doesnt have to take the runqueue lock,
2182  * because all it wants to ensure is that the remote task enters
2183  * the kernel. If the IPI races and the task has been migrated
2184  * to another CPU then no harm is done and the purpose has been
2185  * achieved as well.
2186  */
2187 void kick_process(struct task_struct *p)
2188 {
2189         int cpu;
2190 
2191         preempt_disable();
2192         cpu = task_cpu(p);
2193         if ((cpu != smp_processor_id()) && task_curr(p))
2194                 smp_send_reschedule(cpu);
2195         preempt_enable();
2196 }
2197 EXPORT_SYMBOL_GPL(kick_process);
2198 
2199 /*
2200  * Return a low guess at the load of a migration-source cpu weighted
2201  * according to the scheduling class and "nice" value.
2202  *
2203  * We want to under-estimate the load of migration sources, to
2204  * balance conservatively.
2205  */
2206 static unsigned long source_load(int cpu, int type)
2207 {
2208         struct rq *rq = cpu_rq(cpu);
2209         unsigned long total = weighted_cpuload(cpu);
2210 
2211         if (type == 0 || !sched_feat(LB_BIAS))
2212                 return total;
2213 
2214         return min(rq->cpu_load[type-1], total);
2215 }
2216 
2217 /*
2218  * Return a high guess at the load of a migration-target cpu weighted
2219  * according to the scheduling class and "nice" value.
2220  */
2221 static unsigned long target_load(int cpu, int type)
2222 {
2223         struct rq *rq = cpu_rq(cpu);
2224         unsigned long total = weighted_cpuload(cpu);
2225 
2226         if (type == 0 || !sched_feat(LB_BIAS))
2227                 return total;
2228 
2229         return max(rq->cpu_load[type-1], total);
2230 }
2231 
2232 /*
2233  * find_idlest_group finds and returns the least busy CPU group within the
2234  * domain.
2235  */
2236 static struct sched_group *
2237 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2238 {
2239         struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2240         unsigned long min_load = ULONG_MAX, this_load = 0;
2241         int load_idx = sd->forkexec_idx;
2242         int imbalance = 100 + (sd->imbalance_pct-100)/2;
2243 
2244         do {
2245                 unsigned long load, avg_load;
2246                 int local_group;
2247                 int i;
2248 
2249                 /* Skip over this group if it has no CPUs allowed */
2250                 if (!cpumask_intersects(sched_group_cpus(group),
2251                                         &p->cpus_allowed))
2252                         continue;
2253 
2254                 local_group = cpumask_test_cpu(this_cpu,
2255                                                sched_group_cpus(group));
2256 
2257                 /* Tally up the load of all CPUs in the group */
2258                 avg_load = 0;
2259 
2260                 for_each_cpu(i, sched_group_cpus(group)) {
2261                         /* Bias balancing toward cpus of our domain */
2262                         if (local_group)
2263                                 load = source_load(i, load_idx);
2264                         else
2265                                 load = target_load(i, load_idx);
2266 
2267                         avg_load += load;
2268                 }
2269 
2270                 /* Adjust by relative CPU power of the group */
2271                 avg_load = sg_div_cpu_power(group,
2272                                 avg_load * SCHED_LOAD_SCALE);
2273 
2274                 if (local_group) {
2275                         this_load = avg_load;
2276                         this = group;
2277                 } else if (avg_load < min_load) {
2278                         min_load = avg_load;
2279                         idlest = group;
2280                 }
2281         } while (group = group->next, group != sd->groups);
2282 
2283         if (!idlest || 100*this_load < imbalance*min_load)
2284                 return NULL;
2285         return idlest;
2286 }
2287 
2288 /*
2289  * find_idlest_cpu - find the idlest cpu among the cpus in group.
2290  */
2291 static int
2292 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2293 {
2294         unsigned long load, min_load = ULONG_MAX;
2295         int idlest = -1;
2296         int i;
2297 
2298         /* Traverse only the allowed CPUs */
2299         for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2300                 load = weighted_cpuload(i);
2301 
2302                 if (load < min_load || (load == min_load && i == this_cpu)) {
2303                         min_load = load;
2304                         idlest = i;
2305                 }
2306         }
2307 
2308         return idlest;
2309 }
2310 
2311 /*
2312  * sched_balance_self: balance the current task (running on cpu) in domains
2313  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2314  * SD_BALANCE_EXEC.
2315  *
2316  * Balance, ie. select the least loaded group.
2317  *
2318  * Returns the target CPU number, or the same CPU if no balancing is needed.
2319  *
2320  * preempt must be disabled.
2321  */
2322 static int sched_balance_self(int cpu, int flag)
2323 {
2324         struct task_struct *t = current;
2325         struct sched_domain *tmp, *sd = NULL;
2326 
2327         for_each_domain(cpu, tmp) {
2328                 /*
2329                  * If power savings logic is enabled for a domain, stop there.
2330                  */
2331                 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2332                         break;
2333                 if (tmp->flags & flag)
2334                         sd = tmp;
2335         }
2336 
2337         if (sd)
2338                 update_shares(sd);
2339 
2340         while (sd) {
2341                 struct sched_group *group;
2342                 int new_cpu, weight;
2343 
2344                 if (!(sd->flags & flag)) {
2345                         sd = sd->child;
2346                         continue;
2347                 }
2348 
2349                 group = find_idlest_group(sd, t, cpu);
2350                 if (!group) {
2351                         sd = sd->child;
2352                         continue;
2353                 }
2354 
2355                 new_cpu = find_idlest_cpu(group, t, cpu);
2356                 if (new_cpu == -1 || new_cpu == cpu) {
2357                         /* Now try balancing at a lower domain level of cpu */
2358                         sd = sd->child;
2359                         continue;
2360                 }
2361 
2362                 /* Now try balancing at a lower domain level of new_cpu */
2363                 cpu = new_cpu;
2364                 weight = cpumask_weight(sched_domain_span(sd));
2365                 sd = NULL;
2366                 for_each_domain(cpu, tmp) {
2367                         if (weight <= cpumask_weight(sched_domain_span(tmp)))
2368                                 break;
2369                         if (tmp->flags & flag)
2370                                 sd = tmp;
2371                 }
2372                 /* while loop will break here if sd == NULL */
2373         }
2374 
2375         return cpu;
2376 }
2377 
2378 #endif /* CONFIG_SMP */
2379 
2380 /**
2381  * task_oncpu_function_call - call a function on the cpu on which a task runs
2382  * @p:          the task to evaluate
2383  * @func:       the function to be called
2384  * @info:       the function call argument
2385  *
2386  * Calls the function @func when the task is currently running. This might
2387  * be on the current CPU, which just calls the function directly
2388  */
2389 void task_oncpu_function_call(struct task_struct *p,
2390                               void (*func) (void *info), void *info)
2391 {
2392         int cpu;
2393 
2394         preempt_disable();
2395         cpu = task_cpu(p);
2396         if (task_curr(p))
2397                 smp_call_function_single(cpu, func, info, 1);
2398         preempt_enable();
2399 }
2400 
2401 /***
2402  * try_to_wake_up - wake up a thread
2403  * @p: the to-be-woken-up thread
2404  * @state: the mask of task states that can be woken
2405  * @sync: do a synchronous wakeup?
2406  *
2407  * Put it on the run-queue if it's not already there. The "current"
2408  * thread is always on the run-queue (except when the actual
2409  * re-schedule is in progress), and as such you're allowed to do
2410  * the simpler "current->state = TASK_RUNNING" to mark yourself
2411  * runnable without the overhead of this.
2412  *
2413  * returns failure only if the task is already active.
2414  */
2415 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2416 {
2417         int cpu, orig_cpu, this_cpu, success = 0;
2418         unsigned long flags;
2419         long old_state;
2420         struct rq *rq;
2421 
2422         if (!sched_feat(SYNC_WAKEUPS))
2423                 sync = 0;
2424 
2425 #ifdef CONFIG_SMP
2426         if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2427                 struct sched_domain *sd;
2428 
2429                 this_cpu = raw_smp_processor_id();
2430                 cpu = task_cpu(p);
2431 
2432                 for_each_domain(this_cpu, sd) {
2433                         if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2434                                 update_shares(sd);
2435                                 break;
2436                         }
2437                 }
2438         }
2439 #endif
2440 
2441         smp_wmb();
2442         rq = task_rq_lock(p, &flags);
2443         update_rq_clock(rq);
2444         old_state = p->state;
2445         if (!(old_state & state))
2446                 goto out;
2447 
2448         if (p->se.on_rq)
2449                 goto out_running;
2450 
2451         cpu = task_cpu(p);
2452         orig_cpu = cpu;
2453         this_cpu = smp_processor_id();
2454 
2455 #ifdef CONFIG_SMP
2456         if (unlikely(task_running(rq, p)))
2457                 goto out_activate;
2458 
2459         cpu = p->sched_class->select_task_rq(p, sync);
2460         if (cpu != orig_cpu) {
2461                 set_task_cpu(p, cpu);
2462                 task_rq_unlock(rq, &flags);
2463                 /* might preempt at this point */
2464                 rq = task_rq_lock(p, &flags);
2465                 old_state = p->state;
2466                 if (!(old_state & state))
2467                         goto out;
2468                 if (p->se.on_rq)
2469                         goto out_running;
2470 
2471                 this_cpu = smp_processor_id();
2472                 cpu = task_cpu(p);
2473         }
2474 
2475 #ifdef CONFIG_SCHEDSTATS
2476         schedstat_inc(rq, ttwu_count);
2477         if (cpu == this_cpu)
2478                 schedstat_inc(rq, ttwu_local);
2479         else {
2480                 struct sched_domain *sd;
2481                 for_each_domain(this_cpu, sd) {
2482                         if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2483                                 schedstat_inc(sd, ttwu_wake_remote);
2484                                 break;
2485                         }
2486                 }
2487         }
2488 #endif /* CONFIG_SCHEDSTATS */
2489 
2490 out_activate:
2491 #endif /* CONFIG_SMP */
2492         schedstat_inc(p, se.nr_wakeups);
2493         if (sync)
2494                 schedstat_inc(p, se.nr_wakeups_sync);
2495         if (orig_cpu != cpu)
2496                 schedstat_inc(p, se.nr_wakeups_migrate);
2497         if (cpu == this_cpu)
2498                 schedstat_inc(p, se.nr_wakeups_local);
2499         else
2500                 schedstat_inc(p, se.nr_wakeups_remote);
2501         activate_task(rq, p, 1);
2502         success = 1;
2503 
2504         /*
2505          * Only attribute actual wakeups done by this task.
2506          */
2507         if (!in_interrupt()) {
2508                 struct sched_entity *se = &current->se;
2509                 u64 sample = se->sum_exec_runtime;
2510 
2511                 if (se->last_wakeup)
2512                         sample -= se->last_wakeup;
2513                 else
2514                         sample -= se->start_runtime;
2515                 update_avg(&se->avg_wakeup, sample);
2516 
2517                 se->last_wakeup = se->sum_exec_runtime;
2518         }
2519 
2520 out_running:
2521         trace_sched_wakeup(rq, p, success);
2522         check_preempt_curr(rq, p, sync);
2523 
2524         p->state = TASK_RUNNING;
2525 #ifdef CONFIG_SMP
2526         if (p->sched_class->task_wake_up)
2527                 p->sched_class->task_wake_up(rq, p);
2528 #endif
2529 out:
2530         task_rq_unlock(rq, &flags);
2531 
2532         return success;
2533 }
2534 
2535 /**
2536  * wake_up_process - Wake up a specific process
2537  * @p: The process to be woken up.
2538  *
2539  * Attempt to wake up the nominated process and move it to the set of runnable
2540  * processes.  Returns 1 if the process was woken up, 0 if it was already
2541  * running.
2542  *
2543  * It may be assumed that this function implies a write memory barrier before
2544  * changing the task state if and only if any tasks are woken up.
2545  */
2546 int wake_up_process(struct task_struct *p)
2547 {
2548         return try_to_wake_up(p, TASK_ALL, 0);
2549 }
2550 EXPORT_SYMBOL(wake_up_process);
2551 
2552 int wake_up_state(struct task_struct *p, unsigned int state)
2553 {
2554         return try_to_wake_up(p, state, 0);
2555 }
2556 
2557 /*
2558  * Perform scheduler related setup for a newly forked process p.
2559  * p is forked by current.
2560  *
2561  * __sched_fork() is basic setup used by init_idle() too:
2562  */
2563 static void __sched_fork(struct task_struct *p)
2564 {
2565         p->se.exec_start                = 0;
2566         p->se.sum_exec_runtime          = 0;
2567         p->se.prev_sum_exec_runtime     = 0;
2568         p->se.nr_migrations             = 0;
2569         p->se.last_wakeup               = 0;
2570         p->se.avg_overlap               = 0;
2571         p->se.start_runtime             = 0;
2572         p->se.avg_wakeup                = sysctl_sched_wakeup_granularity;
2573 
2574 #ifdef CONFIG_SCHEDSTATS
2575         p->se.wait_start                        = 0;
2576         p->se.wait_max                          = 0;
2577         p->se.wait_count                        = 0;
2578         p->se.wait_sum                          = 0;
2579 
2580         p->se.sleep_start                       = 0;
2581         p->se.sleep_max                         = 0;
2582         p->se.sum_sleep_runtime                 = 0;
2583 
2584         p->se.block_start                       = 0;
2585         p->se.block_max                         = 0;
2586         p->se.exec_max                          = 0;
2587         p->se.slice_max                         = 0;
2588 
2589         p->se.nr_migrations_cold                = 0;
2590         p->se.nr_failed_migrations_affine       = 0;
2591         p->se.nr_failed_migrations_running      = 0;
2592         p->se.nr_failed_migrations_hot          = 0;
2593         p->se.nr_forced_migrations              = 0;
2594         p->se.nr_forced2_migrations             = 0;
2595 
2596         p->se.nr_wakeups                        = 0;
2597         p->se.nr_wakeups_sync                   = 0;
2598         p->se.nr_wakeups_migrate                = 0;
2599         p->se.nr_wakeups_local                  = 0;
2600         p->se.nr_wakeups_remote                 = 0;
2601         p->se.nr_wakeups_affine                 = 0;
2602         p->se.nr_wakeups_affine_attempts        = 0;
2603         p->se.nr_wakeups_passive                = 0;
2604         p->se.nr_wakeups_idle                   = 0;
2605 
2606 #endif
2607 
2608         INIT_LIST_HEAD(&p->rt.run_list);
2609         p->se.on_rq = 0;
2610         INIT_LIST_HEAD(&p->se.group_node);
2611 
2612 #ifdef CONFIG_PREEMPT_NOTIFIERS
2613         INIT_HLIST_HEAD(&p->preempt_notifiers);
2614 #endif
2615 
2616         /*
2617          * We mark the process as running here, but have not actually
2618          * inserted it onto the runqueue yet. This guarantees that
2619          * nobody will actually run it, and a signal or other external
2620          * event cannot wake it up and insert it on the runqueue either.
2621          */
2622         p->state = TASK_RUNNING;
2623 }
2624 
2625 /*
2626  * fork()/clone()-time setup:
2627  */
2628 void sched_fork(struct task_struct *p, int clone_flags)
2629 {
2630         int cpu = get_cpu();
2631 
2632         __sched_fork(p);
2633 
2634 #ifdef CONFIG_SMP
2635         cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2636 #endif
2637         set_task_cpu(p, cpu);
2638 
2639         /*
2640          * Make sure we do not leak PI boosting priority to the child:
2641          */
2642         p->prio = current->normal_prio;
2643         if (!rt_prio(p->prio))
2644                 p->sched_class = &fair_sched_class;
2645 
2646 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2647         if (likely(sched_info_on()))
2648                 memset(&p->sched_info, 0, sizeof(p->sched_info));
2649 #endif
2650 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2651         p->oncpu = 0;
2652 #endif
2653 #ifdef CONFIG_PREEMPT
2654         /* Want to start with kernel preemption disabled. */
2655         task_thread_info(p)->preempt_count = 1;
2656 #endif
2657         plist_node_init(&p->pushable_tasks, MAX_PRIO);
2658 
2659         put_cpu();
2660 }
2661 
2662 /*
2663  * wake_up_new_task - wake up a newly created task for the first time.
2664  *
2665  * This function will do some initial scheduler statistics housekeeping
2666  * that must be done for every newly created context, then puts the task
2667  * on the runqueue and wakes it.
2668  */
2669 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2670 {
2671         unsigned long flags;
2672         struct rq *rq;
2673 
2674         rq = task_rq_lock(p, &flags);
2675         BUG_ON(p->state != TASK_RUNNING);
2676         update_rq_clock(rq);
2677 
2678         p->prio = effective_prio(p);
2679 
2680         if (!p->sched_class->task_new || !current->se.on_rq) {
2681                 activate_task(rq, p, 0);
2682         } else {
2683                 /*
2684                  * Let the scheduling class do new task startup
2685                  * management (if any):
2686                  */
2687                 p->sched_class->task_new(rq, p);
2688                 inc_nr_running(rq);
2689         }
2690         trace_sched_wakeup_new(rq, p, 1);
2691         check_preempt_curr(rq, p, 0);
2692 #ifdef CONFIG_SMP
2693         if (p->sched_class->task_wake_up)
2694                 p->sched_class->task_wake_up(rq, p);
2695 #endif
2696         task_rq_unlock(rq, &flags);
2697 }
2698 
2699 #ifdef CONFIG_PREEMPT_NOTIFIERS
2700 
2701 /**
2702  * preempt_notifier_register - tell me when current is being preempted & rescheduled
2703  * @notifier: notifier struct to register
2704  */
2705 void preempt_notifier_register(struct preempt_notifier *notifier)
2706 {
2707         hlist_add_head(&notifier->link, &current->preempt_notifiers);
2708 }
2709 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2710 
2711 /**
2712  * preempt_notifier_unregister - no longer interested in preemption notifications
2713  * @notifier: notifier struct to unregister
2714  *
2715  * This is safe to call from within a preemption notifier.
2716  */
2717 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2718 {
2719         hlist_del(&notifier->link);
2720 }
2721 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2722 
2723 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2724 {
2725         struct preempt_notifier *notifier;
2726         struct hlist_node *node;
2727 
2728         hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2729                 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2730 }
2731 
2732 static void
2733 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2734                                  struct task_struct *next)
2735 {
2736         struct preempt_notifier *notifier;
2737         struct hlist_node *node;
2738 
2739         hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2740                 notifier->ops->sched_out(notifier, next);
2741 }
2742 
2743 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2744 
2745 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2746 {
2747 }
2748 
2749 static void
2750 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2751                                  struct task_struct *next)
2752 {
2753 }
2754 
2755 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2756 
2757 /**
2758  * prepare_task_switch - prepare to switch tasks
2759  * @rq: the runqueue preparing to switch
2760  * @prev: the current task that is being switched out
2761  * @next: the task we are going to switch to.
2762  *
2763  * This is called with the rq lock held and interrupts off. It must
2764  * be paired with a subsequent finish_task_switch after the context
2765  * switch.
2766  *
2767  * prepare_task_switch sets up locking and calls architecture specific
2768  * hooks.
2769  */
2770 static inline void
2771 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2772                     struct task_struct *next)
2773 {
2774         fire_sched_out_preempt_notifiers(prev, next);
2775         prepare_lock_switch(rq, next);
2776         prepare_arch_switch(next);
2777 }
2778 
2779 /**
2780  * finish_task_switch - clean up after a task-switch
2781  * @rq: runqueue associated with task-switch
2782  * @prev: the thread we just switched away from.
2783  *
2784  * finish_task_switch must be called after the context switch, paired
2785  * with a prepare_task_switch call before the context switch.
2786  * finish_task_switch will reconcile locking set up by prepare_task_switch,
2787  * and do any other architecture-specific cleanup actions.
2788  *
2789  * Note that we may have delayed dropping an mm in context_switch(). If
2790  * so, we finish that here outside of the runqueue lock. (Doing it
2791  * with the lock held can cause deadlocks; see schedule() for
2792  * details.)
2793  */
2794 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2795         __releases(rq->lock)
2796 {
2797         struct mm_struct *mm = rq->prev_mm;
2798         long prev_state;
2799 #ifdef CONFIG_SMP
2800         int post_schedule = 0;
2801 
2802         if (current->sched_class->needs_post_schedule)
2803                 post_schedule = current->sched_class->needs_post_schedule(rq);
2804 #endif
2805 
2806         rq->prev_mm = NULL;
2807 
2808         /*
2809          * A task struct has one reference for the use as "current".
2810          * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2811          * schedule one last time. The schedule call will never return, and
2812          * the scheduled task must drop that reference.
2813          * The test for TASK_DEAD must occur while the runqueue locks are
2814          * still held, otherwise prev could be scheduled on another cpu, die
2815          * there before we look at prev->state, and then the reference would
2816          * be dropped twice.
2817          *              Manfred Spraul <manfred@colorfullife.com>
2818          */
2819         prev_state = prev->state;
2820         finish_arch_switch(prev);
2821         perf_counter_task_sched_in(current, cpu_of(rq));
2822         finish_lock_switch(rq, prev);
2823 #ifdef CONFIG_SMP
2824         if (post_schedule)
2825                 current->sched_class->post_schedule(rq);
2826 #endif
2827 
2828         fire_sched_in_preempt_notifiers(current);
2829         if (mm)
2830                 mmdrop(mm);
2831         if (unlikely(prev_state == TASK_DEAD)) {
2832                 /*
2833                  * Remove function-return probe instances associated with this
2834                  * task and put them back on the free list.
2835                  */
2836                 kprobe_flush_task(prev);
2837                 put_task_struct(prev);
2838         }
2839 }
2840 
2841 /**
2842  * schedule_tail - first thing a freshly forked thread must call.
2843  * @prev: the thread we just switched away from.
2844  */
2845 asmlinkage void schedule_tail(struct task_struct *prev)
2846         __releases(rq->lock)
2847 {
2848         struct rq *rq = this_rq();
2849 
2850         finish_task_switch(rq, prev);
2851 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2852         /* In this case, finish_task_switch does not reenable preemption */
2853         preempt_enable();
2854 #endif
2855         if (current->set_child_tid)
2856                 put_user(task_pid_vnr(current), current->set_child_tid);
2857 }
2858 
2859 /*
2860  * context_switch - switch to the new MM and the new
2861  * thread's register state.
2862  */
2863 static inline void
2864 context_switch(struct rq *rq, struct task_struct *prev,
2865                struct task_struct *next)
2866 {
2867         struct mm_struct *mm, *oldmm;
2868 
2869         prepare_task_switch(rq, prev, next);
2870         trace_sched_switch(rq, prev, next);
2871         mm = next->mm;
2872         oldmm = prev->active_mm;
2873         /*
2874          * For paravirt, this is coupled with an exit in switch_to to
2875          * combine the page table reload and the switch backend into
2876          * one hypercall.
2877          */
2878         arch_start_context_switch(prev);
2879 
2880         if (unlikely(!mm)) {
2881                 next->active_mm = oldmm;
2882                 atomic_inc(&oldmm->mm_count);
2883                 enter_lazy_tlb(oldmm, next);
2884         } else
2885                 switch_mm(oldmm, mm, next);
2886 
2887         if (unlikely(!prev->mm)) {
2888                 prev->active_mm = NULL;
2889                 rq->prev_mm = oldmm;
2890         }
2891         /*
2892          * Since the runqueue lock will be released by the next
2893          * task (which is an invalid locking op but in the case
2894          * of the scheduler it's an obvious special-case), so we
2895          * do an early lockdep release here:
2896          */
2897 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2898         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2899 #endif
2900 
2901         /* Here we just switch the register state and the stack. */
2902         switch_to(prev, next, prev);
2903 
2904         barrier();
2905         /*
2906          * this_rq must be evaluated again because prev may have moved
2907          * CPUs since it called schedule(), thus the 'rq' on its stack
2908          * frame will be invalid.
2909          */
2910         finish_task_switch(this_rq(), prev);
2911 }
2912 
2913 /*
2914  * nr_running, nr_uninterruptible and nr_context_switches:
2915  *
2916  * externally visible scheduler statistics: current number of runnable
2917  * threads, current number of uninterruptible-sleeping threads, total
2918  * number of context switches performed since bootup.
2919  */
2920 unsigned long nr_running(void)
2921 {
2922         unsigned long i, sum = 0;
2923 
2924         for_each_online_cpu(i)
2925                 sum += cpu_rq(i)->nr_running;
2926 
2927         return sum;
2928 }
2929 
2930 unsigned long nr_uninterruptible(void)
2931 {
2932         unsigned long i, sum = 0;
2933 
2934         for_each_possible_cpu(i)
2935                 sum += cpu_rq(i)->nr_uninterruptible;
2936 
2937         /*
2938          * Since we read the counters lockless, it might be slightly
2939          * inaccurate. Do not allow it to go below zero though:
2940          */
2941         if (unlikely((long)sum < 0))
2942                 sum = 0;
2943 
2944         return sum;
2945 }
2946 
2947 unsigned long long nr_context_switches(void)
2948 {
2949         int i;
2950         unsigned long long sum = 0;
2951 
2952         for_each_possible_cpu(i)
2953                 sum += cpu_rq(i)->nr_switches;
2954 
2955         return sum;
2956 }
2957 
2958 unsigned long nr_iowait(void)
2959 {
2960         unsigned long i, sum = 0;
2961 
2962         for_each_possible_cpu(i)
2963                 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2964 
2965         return sum;
2966 }
2967 
2968 /* Variables and functions for calc_load */
2969 static atomic_long_t calc_load_tasks;
2970 static unsigned long calc_load_update;
2971 unsigned long avenrun[3];
2972 EXPORT_SYMBOL(avenrun);
2973 
2974 /**
2975  * get_avenrun - get the load average array
2976  * @loads:      pointer to dest load array
2977  * @offset:     offset to add
2978  * @shift:      shift count to shift the result left
2979  *
2980  * These values are estimates at best, so no need for locking.
2981  */
2982 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2983 {
2984         loads[0] = (avenrun[0] + offset) << shift;
2985         loads[1] = (avenrun[1] + offset) << shift;
2986         loads[2] = (avenrun[2] + offset) << shift;
2987 }
2988 
2989 static unsigned long
2990 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2991 {
2992         load *= exp;
2993         load += active * (FIXED_1 - exp);
2994         return load >> FSHIFT;
2995 }
2996 
2997 /*
2998  * calc_load - update the avenrun load estimates 10 ticks after the
2999  * CPUs have updated calc_load_tasks.
3000  */
3001 void calc_global_load(void)
3002 {
3003         unsigned long upd = calc_load_update + 10;
3004         long active;
3005 
3006         if (time_before(jiffies, upd))
3007                 return;
3008 
3009         active = atomic_long_read(&calc_load_tasks);
3010         active = active > 0 ? active * FIXED_1 : 0;
3011 
3012         avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3013         avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3014         avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3015 
3016         calc_load_update += LOAD_FREQ;
3017 }
3018 
3019 /*
3020  * Either called from update_cpu_load() or from a cpu going idle
3021  */
3022 static void calc_load_account_active(struct rq *this_rq)
3023 {
3024         long nr_active, delta;
3025 
3026         nr_active = this_rq->nr_running;
3027         nr_active += (long) this_rq->nr_uninterruptible;
3028 
3029         if (nr_active != this_rq->calc_load_active) {
3030                 delta = nr_active - this_rq->calc_load_active;
3031                 this_rq->calc_load_active = nr_active;
3032                 atomic_long_add(delta, &calc_load_tasks);
3033         }
3034 }
3035 
3036 /*
3037  * Externally visible per-cpu scheduler statistics:
3038  * cpu_nr_migrations(cpu) - number of migrations into that cpu
3039  */
3040 u64 cpu_nr_migrations(int cpu)
3041 {
3042         return cpu_rq(cpu)->nr_migrations_in;
3043 }
3044 
3045 /*
3046  * Update rq->cpu_load[] statistics. This function is usually called every
3047  * scheduler tick (TICK_NSEC).
3048  */
3049 static void update_cpu_load(struct rq *this_rq)
3050 {
3051         unsigned long this_load = this_rq->load.weight;
3052         int i, scale;
3053 
3054         this_rq->nr_load_updates++;
3055 
3056         /* Update our load: */
3057         for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3058                 unsigned long old_load, new_load;
3059 
3060                 /* scale is effectively 1 << i now, and >> i divides by scale */
3061 
3062                 old_load = this_rq->cpu_load[i];
3063                 new_load = this_load;
3064                 /*
3065                  * Round up the averaging division if load is increasing. This
3066                  * prevents us from getting stuck on 9 if the load is 10, for
3067                  * example.
3068                  */
3069                 if (new_load > old_load)
3070                         new_load += scale-1;
3071                 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3072         }
3073 
3074         if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3075                 this_rq->calc_load_update += LOAD_FREQ;
3076                 calc_load_account_active(this_rq);
3077         }
3078 }
3079 
3080 #ifdef CONFIG_SMP
3081 
3082 /*
3083  * double_rq_lock - safely lock two runqueues
3084  *
3085  * Note this does not disable interrupts like task_rq_lock,
3086  * you need to do so manually before calling.
3087  */
3088 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3089         __acquires(rq1->lock)
3090         __acquires(rq2->lock)
3091 {
3092         BUG_ON(!irqs_disabled());
3093         if (rq1 == rq2) {
3094                 spin_lock(&rq1->lock);
3095                 __acquire(rq2->lock);   /* Fake it out ;) */
3096         } else {
3097                 if (rq1 < rq2) {
3098                         spin_lock(&rq1->lock);
3099                         spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3100                 } else {
3101                         spin_lock(&rq2->lock);
3102                         spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3103                 }
3104         }
3105         update_rq_clock(rq1);
3106         update_rq_clock(rq2);
3107 }
3108 
3109 /*
3110  * double_rq_unlock - safely unlock two runqueues
3111  *
3112  * Note this does not restore interrupts like task_rq_unlock,
3113  * you need to do so manually after calling.
3114  */
3115 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3116         __releases(rq1->lock)
3117         __releases(rq2->lock)
3118 {
3119         spin_unlock(&rq1->lock);
3120         if (rq1 != rq2)
3121                 spin_unlock(&rq2->lock);
3122         else
3123                 __release(rq2->lock);
3124 }
3125 
3126 /*
3127  * If dest_cpu is allowed for this process, migrate the task to it.
3128  * This is accomplished by forcing the cpu_allowed mask to only
3129  * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3130  * the cpu_allowed mask is restored.
3131  */
3132 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3133 {
3134         struct migration_req req;
3135         unsigned long flags;
3136         struct rq *rq;
3137 
3138         rq = task_rq_lock(p, &flags);
3139         if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3140             || unlikely(!cpu_active(dest_cpu)))
3141                 goto out;
3142 
3143         /* force the process onto the specified CPU */
3144         if (migrate_task(p, dest_cpu, &req)) {
3145                 /* Need to wait for migration thread (might exit: take ref). */
3146                 struct task_struct *mt = rq->migration_thread;
3147 
3148                 get_task_struct(mt);
3149                 task_rq_unlock(rq, &flags);
3150                 wake_up_process(mt);
3151                 put_task_struct(mt);
3152                 wait_for_completion(&req.done);
3153 
3154                 return;
3155         }
3156 out:
3157         task_rq_unlock(rq, &flags);
3158 }
3159 
3160 /*
3161  * sched_exec - execve() is a valuable balancing opportunity, because at
3162  * this point the task has the smallest effective memory and cache footprint.
3163  */
3164 void sched_exec(void)
3165 {
3166         int new_cpu, this_cpu = get_cpu();
3167         new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3168         put_cpu();
3169         if (new_cpu != this_cpu)
3170                 sched_migrate_task(current, new_cpu);
3171 }
3172 
3173 /*
3174  * pull_task - move a task from a remote runqueue to the local runqueue.
3175  * Both runqueues must be locked.
3176  */
3177 static void pull_task(struct rq *src_rq, struct task_struct *p,
3178                       struct rq *this_rq, int this_cpu)
3179 {
3180         deactivate_task(src_rq, p, 0);
3181         set_task_cpu(p, this_cpu);
3182         activate_task(this_rq, p, 0);
3183         /*
3184          * Note that idle threads have a prio of MAX_PRIO, for this test
3185          * to be always true for them.
3186          */
3187         check_preempt_curr(this_rq, p, 0);
3188 }
3189 
3190 /*
3191  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3192  */
3193 static
3194 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3195                      struct sched_domain *sd, enum cpu_idle_type idle,
3196                      int *all_pinned)
3197 {
3198         int tsk_cache_hot = 0;
3199         /*
3200          * We do not migrate tasks that are:
3201          * 1) running (obviously), or
3202          * 2) cannot be migrated to this CPU due to cpus_allowed, or
3203          * 3) are cache-hot on their current CPU.
3204          */
3205         if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3206                 schedstat_inc(p, se.nr_failed_migrations_affine);
3207                 return 0;
3208         }
3209         *all_pinned = 0;
3210 
3211         if (task_running(rq, p)) {
3212                 schedstat_inc(p, se.nr_failed_migrations_running);
3213                 return 0;
3214         }
3215 
3216         /*
3217          * Aggressive migration if:
3218          * 1) task is cache cold, or
3219          * 2) too many balance attempts have failed.
3220          */
3221 
3222         tsk_cache_hot = task_hot(p, rq->clock, sd);
3223         if (!tsk_cache_hot ||
3224                 sd->nr_balance_failed > sd->cache_nice_tries) {
3225 #ifdef CONFIG_SCHEDSTATS
3226                 if (tsk_cache_hot) {
3227                         schedstat_inc(sd, lb_hot_gained[idle]);
3228                         schedstat_inc(p, se.nr_forced_migrations);
3229                 }
3230 #endif
3231                 return 1;
3232         }
3233 
3234         if (tsk_cache_hot) {
3235                 schedstat_inc(p, se.nr_failed_migrations_hot);
3236                 return 0;
3237         }
3238         return 1;
3239 }
3240 
3241 static unsigned long
3242 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3243               unsigned long max_load_move, struct sched_domain *sd,
3244               enum cpu_idle_type idle, int *all_pinned,
3245               int *this_best_prio, struct rq_iterator *iterator)
3246 {
3247         int loops = 0, pulled = 0, pinned = 0;
3248         struct task_struct *p;
3249         long rem_load_move = max_load_move;
3250 
3251         if (max_load_move == 0)
3252                 goto out;
3253 
3254         pinned = 1;
3255 
3256         /*
3257          * Start the load-balancing iterator:
3258          */
3259         p = iterator->start(iterator->arg);
3260 next:
3261         if (!p || loops++ > sysctl_sched_nr_migrate)
3262                 goto out;
3263 
3264         if ((p->se.load.weight >> 1) > rem_load_move ||
3265             !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3266                 p = iterator->next(iterator->arg);
3267                 goto next;
3268         }
3269 
3270         pull_task(busiest, p, this_rq, this_cpu);
3271         pulled++;
3272         rem_load_move -= p->se.load.weight;
3273 
3274 #ifdef CONFIG_PREEMPT
3275         /*
3276          * NEWIDLE balancing is a source of latency, so preemptible kernels
3277          * will stop after the first task is pulled to minimize the critical
3278          * section.
3279          */
3280         if (idle == CPU_NEWLY_IDLE)
3281                 goto out;
3282 #endif
3283 
3284         /*
3285          * We only want to steal up to the prescribed amount of weighted load.
3286          */
3287         if (rem_load_move > 0) {
3288                 if (p->prio < *this_best_prio)
3289                         *this_best_prio = p->prio;
3290                 p = iterator->next(iterator->arg);
3291                 goto next;
3292         }
3293 out:
3294         /*
3295          * Right now, this is one of only two places pull_task() is called,
3296          * so we can safely collect pull_task() stats here rather than
3297          * inside pull_task().
3298          */
3299         schedstat_add(sd, lb_gained[idle], pulled);
3300 
3301         if (all_pinned)
3302                 *all_pinned = pinned;
3303 
3304         return max_load_move - rem_load_move;
3305 }
3306 
3307 /*
3308  * move_tasks tries to move up to max_load_move weighted load from busiest to
3309  * this_rq, as part of a balancing operation within domain "sd".
3310  * Returns 1 if successful and 0 otherwise.
3311  *
3312  * Called with both runqueues locked.
3313  */
3314 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3315                       unsigned long max_load_move,
3316                       struct sched_domain *sd, enum cpu_idle_type idle,
3317                       int *all_pinned)
3318 {
3319         const struct sched_class *class = sched_class_highest;
3320         unsigned long total_load_moved = 0;
3321         int this_best_prio = this_rq->curr->prio;
3322 
3323         do {
3324                 total_load_moved +=
3325                         class->load_balance(this_rq, this_cpu, busiest,
3326                                 max_load_move - total_load_moved,
3327                                 sd, idle, all_pinned, &this_best_prio);
3328                 class = class->next;
3329 
3330 #ifdef CONFIG_PREEMPT
3331                 /*
3332                  * NEWIDLE balancing is a source of latency, so preemptible
3333                  * kernels will stop after the first task is pulled to minimize
3334                  * the critical section.
3335                  */
3336                 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3337                         break;
3338 #endif
3339         } while (class && max_load_move > total_load_moved);
3340 
3341         return total_load_moved > 0;
3342 }
3343 
3344 static int
3345 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3346                    struct sched_domain *sd, enum cpu_idle_type idle,
3347                    struct rq_iterator *iterator)
3348 {
3349         struct task_struct *p = iterator->start(iterator->arg);
3350         int pinned = 0;
3351 
3352         while (p) {
3353                 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3354                         pull_task(busiest, p, this_rq, this_cpu);
3355                         /*
3356                          * Right now, this is only the second place pull_task()
3357                          * is called, so we can safely collect pull_task()
3358                          * stats here rather than inside pull_task().
3359                          */
3360                         schedstat_inc(sd, lb_gained[idle]);
3361 
3362                         return 1;
3363                 }
3364                 p = iterator->next(iterator->arg);
3365         }
3366 
3367         return 0;
3368 }
3369 
3370 /*
3371  * move_one_task tries to move exactly one task from busiest to this_rq, as
3372  * part of active balancing operations within "domain".
3373  * Returns 1 if successful and 0 otherwise.
3374  *
3375  * Called with both runqueues locked.
3376  */
3377 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3378                          struct sched_domain *sd, enum cpu_idle_type idle)
3379 {
3380         const struct sched_class *class;
3381 
3382         for (class = sched_class_highest; class; class = class->next)
3383                 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3384                         return 1;
3385 
3386         return 0;
3387 }
3388 /********** Helpers for find_busiest_group ************************/
3389 /*
3390  * sd_lb_stats - Structure to store the statistics of a sched_domain
3391  *              during load balancing.
3392  */
3393 struct sd_lb_stats {
3394         struct sched_group *busiest; /* Busiest group in this sd */
3395         struct sched_group *this;  /* Local group in this sd */
3396         unsigned long total_load;  /* Total load of all groups in sd */
3397         unsigned long total_pwr;   /*   Total power of all groups in sd */
3398         unsigned long avg_load;    /* Average load across all groups in sd */
3399 
3400         /** Statistics of this group */
3401         unsigned long this_load;
3402         unsigned long this_load_per_task;
3403         unsigned long this_nr_running;
3404 
3405         /* Statistics of the busiest group */
3406         unsigned long max_load;
3407         unsigned long busiest_load_per_task;
3408         unsigned long busiest_nr_running;
3409 
3410         int group_imb; /* Is there imbalance in this sd */
3411 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3412         int power_savings_balance; /* Is powersave balance needed for this sd */
3413         struct sched_group *group_min; /* Least loaded group in sd */
3414         struct sched_group *group_leader; /* Group which relieves group_min */
3415         unsigned long min_load_per_task; /* load_per_task in group_min */
3416         unsigned long leader_nr_running; /* Nr running of group_leader */
3417         unsigned long min_nr_running; /* Nr running of group_min */
3418 #endif
3419 };
3420 
3421 /*
3422  * sg_lb_stats - stats of a sched_group required for load_balancing
3423  */
3424 struct sg_lb_stats {
3425         unsigned long avg_load; /*Avg load across the CPUs of the group */
3426         unsigned long group_load; /* Total load over the CPUs of the group */
3427         unsigned long sum_nr_running; /* Nr tasks running in the group */
3428         unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3429         unsigned long group_capacity;
3430         int group_imb; /* Is there an imbalance in the group ? */
3431 };
3432 
3433 /**
3434  * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3435  * @group: The group whose first cpu is to be returned.
3436  */
3437 static inline unsigned int group_first_cpu(struct sched_group *group)
3438 {
3439         return cpumask_first(sched_group_cpus(group));
3440 }
3441 
3442 /**
3443  * get_sd_load_idx - Obtain the load index for a given sched domain.
3444  * @sd: The sched_domain whose load_idx is to be obtained.
3445  * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3446  */
3447 static inline int get_sd_load_idx(struct sched_domain *sd,
3448                                         enum cpu_idle_type idle)
3449 {
3450         int load_idx;
3451 
3452         switch (idle) {
3453         case CPU_NOT_IDLE:
3454                 load_idx = sd->busy_idx;
3455                 break;
3456 
3457         case CPU_NEWLY_IDLE:
3458                 load_idx = sd->newidle_idx;
3459                 break;
3460         default:
3461                 load_idx = sd->idle_idx;
3462                 break;
3463         }
3464 
3465         return load_idx;
3466 }
3467 
3468 
3469 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3470 /**
3471  * init_sd_power_savings_stats - Initialize power savings statistics for
3472  * the given sched_domain, during load balancing.
3473  *
3474  * @sd: Sched domain whose power-savings statistics are to be initialized.
3475  * @sds: Variable containing the statistics for sd.
3476  * @idle: Idle status of the CPU at which we're performing load-balancing.
3477  */
3478 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3479         struct sd_lb_stats *sds, enum cpu_idle_type idle)
3480 {
3481         /*
3482          * Busy processors will not participate in power savings
3483          * balance.
3484          */
3485         if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3486                 sds->power_savings_balance = 0;
3487         else {
3488                 sds->power_savings_balance = 1;
3489                 sds->min_nr_running = ULONG_MAX;
3490                 sds->leader_nr_running = 0;
3491         }
3492 }
3493 
3494 /**
3495  * update_sd_power_savings_stats - Update the power saving stats for a
3496  * sched_domain while performing load balancing.
3497  *
3498  * @group: sched_group belonging to the sched_domain under consideration.
3499  * @sds: Variable containing the statistics of the sched_domain
3500  * @local_group: Does group contain the CPU for which we're performing
3501  *              load balancing ?
3502  * @sgs: Variable containing the statistics of the group.
3503  */
3504 static inline void update_sd_power_savings_stats(struct sched_group *group,
3505         struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3506 {
3507 
3508         if (!sds->power_savings_balance)
3509                 return;
3510 
3511         /*
3512          * If the local group is idle or completely loaded
3513          * no need to do power savings balance at this domain
3514          */
3515         if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3516                                 !sds->this_nr_running))
3517                 sds->power_savings_balance = 0;
3518 
3519         /*
3520          * If a group is already running at full capacity or idle,
3521          * don't include that group in power savings calculations
3522          */
3523         if (!sds->power_savings_balance ||
3524                 sgs->sum_nr_running >= sgs->group_capacity ||
3525                 !sgs->sum_nr_running)
3526                 return;
3527 
3528         /*
3529          * Calculate the group which has the least non-idle load.
3530          * This is the group from where we need to pick up the load
3531          * for saving power
3532          */
3533         if ((sgs->sum_nr_running < sds->min_nr_running) ||
3534             (sgs->sum_nr_running == sds->min_nr_running &&
3535              group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3536                 sds->group_min = group;
3537                 sds->min_nr_running = sgs->sum_nr_running;
3538                 sds->min_load_per_task = sgs->sum_weighted_load /
3539                                                 sgs->sum_nr_running;
3540         }
3541 
3542         /*
3543          * Calculate the group which is almost near its
3544          * capacity but still has some space to pick up some load
3545          * from other group and save more power
3546          */
3547         if (sgs->sum_nr_running > sgs->group_capacity - 1)
3548                 return;
3549 
3550         if (sgs->sum_nr_running > sds->leader_nr_running ||
3551             (sgs->sum_nr_running == sds->leader_nr_running &&
3552              group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3553                 sds->group_leader = group;
3554                 sds->leader_nr_running = sgs->sum_nr_running;
3555         }
3556 }
3557 
3558 /**
3559  * check_power_save_busiest_group - see if there is potential for some power-savings balance
3560  * @sds: Variable containing the statistics of the sched_domain
3561  *      under consideration.
3562  * @this_cpu: Cpu at which we're currently performing load-balancing.
3563  * @imbalance: Variable to store the imbalance.
3564  *
3565  * Description:
3566  * Check if we have potential to perform some power-savings balance.
3567  * If yes, set the busiest group to be the least loaded group in the
3568  * sched_domain, so that it's CPUs can be put to idle.
3569  *
3570  * Returns 1 if there is potential to perform power-savings balance.
3571  * Else returns 0.
3572  */
3573 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3574                                         int this_cpu, unsigned long *imbalance)
3575 {
3576         if (!sds->power_savings_balance)
3577                 return 0;
3578 
3579         if (sds->this != sds->group_leader ||
3580                         sds->group_leader == sds->group_min)
3581                 return 0;
3582 
3583         *imbalance = sds->min_load_per_task;
3584         sds->busiest = sds->group_min;
3585 
3586         if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3587                 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3588                         group_first_cpu(sds->group_leader);
3589         }
3590 
3591         return 1;
3592 
3593 }
3594 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3595 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3596         struct sd_lb_stats *sds, enum cpu_idle_type idle)
3597 {
3598         return;
3599 }
3600 
3601 static inline void update_sd_power_savings_stats(struct sched_group *group,
3602         struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3603 {
3604         return;
3605 }
3606 
3607 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3608                                         int this_cpu, unsigned long *imbalance)
3609 {
3610         return 0;
3611 }
3612 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3613 
3614 
3615 /**
3616  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3617  * @group: sched_group whose statistics are to be updated.
3618  * @this_cpu: Cpu for which load balance is currently performed.
3619  * @idle: Idle status of this_cpu
3620  * @load_idx: Load index of sched_domain of this_cpu for load calc.
3621  * @sd_idle: Idle status of the sched_domain containing group.
3622  * @local_group: Does group contain this_cpu.
3623  * @cpus: Set of cpus considered for load balancing.
3624  * @balance: Should we balance.
3625  * @sgs: variable to hold the statistics for this group.
3626  */
3627 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3628                         enum cpu_idle_type idle, int load_idx, int *sd_idle,
3629                         int local_group, const struct cpumask *cpus,
3630                         int *balance, struct sg_lb_stats *sgs)
3631 {
3632         unsigned long load, max_cpu_load, min_cpu_load;
3633         int i;
3634         unsigned int balance_cpu = -1, first_idle_cpu = 0;
3635         unsigned long sum_avg_load_per_task;
3636         unsigned long avg_load_per_task;
3637 
3638         if (local_group)
3639                 balance_cpu = group_first_cpu(group);
3640 
3641         /* Tally up the load of all CPUs in the group */
3642         sum_avg_load_per_task = avg_load_per_task = 0;
3643         max_cpu_load = 0;
3644         min_cpu_load = ~0UL;
3645 
3646         for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3647                 struct rq *rq = cpu_rq(i);
3648 
3649                 if (*sd_idle && rq->nr_running)
3650                         *sd_idle = 0;
3651 
3652                 /* Bias balancing toward cpus of our domain */
3653                 if (local_group) {
3654                         if (idle_cpu(i) && !first_idle_cpu) {
3655                                 first_idle_cpu = 1;
3656                                 balance_cpu = i;
3657                         }
3658 
3659                         load = target_load(i, load_idx);
3660                 } else {
3661                         load = source_load(i, load_idx);
3662                         if (load > max_cpu_load)
3663                                 max_cpu_load = load;
3664                         if (min_cpu_load > load)
3665                                 min_cpu_load = load;
3666                 }
3667 
3668                 sgs->group_load += load;
3669                 sgs->sum_nr_running += rq->nr_running;
3670                 sgs->sum_weighted_load += weighted_cpuload(i);
3671 
3672                 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3673         }
3674 
3675         /*
3676          * First idle cpu or the first cpu(busiest) in this sched group
3677          * is eligible for doing load balancing at this and above
3678          * domains. In the newly idle case, we will allow all the cpu's
3679          * to do the newly idle load balance.
3680          */
3681         if (idle != CPU_NEWLY_IDLE && local_group &&
3682             balance_cpu != this_cpu && balance) {
3683                 *balance = 0;
3684                 return;
3685         }
3686 
3687         /* Adjust by relative CPU power of the group */
3688         sgs->avg_load = sg_div_cpu_power(group,
3689                         sgs->group_load * SCHED_LOAD_SCALE);
3690 
3691 
3692         /*
3693          * Consider the group unbalanced when the imbalance is larger
3694          * than the average weight of two tasks.
3695          *
3696          * APZ: with cgroup the avg task weight can vary wildly and
3697          *      might not be a suitable number - should we keep a
3698          *      normalized nr_running number somewhere that negates
3699          *      the hierarchy?
3700          */
3701         avg_load_per_task = sg_div_cpu_power(group,
3702                         sum_avg_load_per_task * SCHED_LOAD_SCALE);
3703 
3704         if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3705                 sgs->group_imb = 1;
3706 
3707         sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3708 
3709 }
3710 
3711 /**
3712  * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3713  * @sd: sched_domain whose statistics are to be updated.
3714  * @this_cpu: Cpu for which load balance is currently performed.
3715  * @idle: Idle status of this_cpu
3716  * @sd_idle: Idle status of the sched_domain containing group.
3717  * @cpus: Set of cpus considered for load balancing.
3718  * @balance: Should we balance.
3719  * @sds: variable to hold the statistics for this sched_domain.
3720  */
3721 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3722                         enum cpu_idle_type idle, int *sd_idle,
3723                         const struct cpumask *cpus, int *balance,
3724                         struct sd_lb_stats *sds)
3725 {
3726         struct sched_group *group = sd->groups;
3727         struct sg_lb_stats sgs;
3728         int load_idx;
3729 
3730         init_sd_power_savings_stats(sd, sds, idle);
3731         load_idx = get_sd_load_idx(sd, idle);
3732 
3733         do {
3734                 int local_group;
3735 
3736                 local_group = cpumask_test_cpu(this_cpu,
3737                                                sched_group_cpus(group));
3738                 memset(&sgs, 0, sizeof(sgs));
3739                 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3740                                 local_group, cpus, balance, &sgs);
3741 
3742                 if (local_group && balance && !(*balance))
3743                         return;
3744 
3745                 sds->total_load += sgs.group_load;
3746                 sds->total_pwr += group->__cpu_power;
3747 
3748                 if (local_group) {
3749                         sds->this_load = sgs.avg_load;
3750                         sds->this = group;
3751                         sds->this_nr_running = sgs.sum_nr_running;
3752                         sds->this_load_per_task = sgs.sum_weighted_load;
3753                 } else if (sgs.avg_load > sds->max_load &&
3754                            (sgs.sum_nr_running > sgs.group_capacity ||
3755                                 sgs.group_imb)) {
3756                         sds->max_load = sgs.avg_load;
3757                         sds->busiest = group;
3758                         sds->busiest_nr_running = sgs.sum_nr_running;
3759                         sds->busiest_load_per_task = sgs.sum_weighted_load;
3760                         sds->group_imb = sgs.group_imb;
3761                 }
3762 
3763                 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3764                 group = group->next;
3765         } while (group != sd->groups);
3766 
3767 }
3768 
3769 /**
3770  * fix_small_imbalance - Calculate the minor imbalance that exists
3771  *                      amongst the groups of a sched_domain, during
3772  *                      load balancing.
3773  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3774  * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3775  * @imbalance: Variable to store the imbalance.
3776  */
3777 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3778                                 int this_cpu, unsigned long *imbalance)
3779 {
3780         unsigned long tmp, pwr_now = 0, pwr_move = 0;
3781         unsigned int imbn = 2;
3782 
3783         if (sds->this_nr_running) {
3784                 sds->this_load_per_task /= sds->this_nr_running;
3785                 if (sds->busiest_load_per_task >
3786                                 sds->this_load_per_task)
3787                         imbn = 1;
3788         } else
3789                 sds->this_load_per_task =
3790                         cpu_avg_load_per_task(this_cpu);
3791 
3792         if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3793                         sds->busiest_load_per_task * imbn) {
3794                 *imbalance = sds->busiest_load_per_task;
3795                 return;
3796         }
3797 
3798         /*
3799          * OK, we don't have enough imbalance to justify moving tasks,
3800          * however we may be able to increase total CPU power used by
3801          * moving them.
3802          */
3803 
3804         pwr_now += sds->busiest->__cpu_power *
3805                         min(sds->busiest_load_per_task, sds->max_load);
3806         pwr_now += sds->this->__cpu_power *
3807                         min(sds->this_load_per_task, sds->this_load);
3808         pwr_now /= SCHED_LOAD_SCALE;
3809 
3810         /* Amount of load we'd subtract */
3811         tmp = sg_div_cpu_power(sds->busiest,
3812                         sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3813         if (sds->max_load > tmp)
3814                 pwr_move += sds->busiest->__cpu_power *
3815                         min(sds->busiest_load_per_task, sds->max_load - tmp);
3816 
3817         /* Amount of load we'd add */
3818         if (sds->max_load * sds->busiest->__cpu_power <
3819                 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3820                 tmp = sg_div_cpu_power(sds->this,
3821                         sds->max_load * sds->busiest->__cpu_power);
3822         else
3823                 tmp = sg_div_cpu_power(sds->this,
3824                         sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3825         pwr_move += sds->this->__cpu_power *
3826                         min(sds->this_load_per_task, sds->this_load + tmp);
3827         pwr_move /= SCHED_LOAD_SCALE;
3828 
3829         /* Move if we gain throughput */
3830         if (pwr_move > pwr_now)
3831                 *imbalance = sds->busiest_load_per_task;
3832 }
3833 
3834 /**
3835  * calculate_imbalance - Calculate the amount of imbalance present within the
3836  *                       groups of a given sched_domain during load balance.
3837  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3838  * @this_cpu: Cpu for which currently load balance is being performed.
3839  * @imbalance: The variable to store the imbalance.
3840  */
3841 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3842                 unsigned long *imbalance)
3843 {
3844         unsigned long max_pull;
3845         /*
3846          * In the presence of smp nice balancing, certain scenarios can have
3847          * max load less than avg load(as we skip the groups at or below
3848          * its cpu_power, while calculating max_load..)
3849          */
3850         if (sds->max_load < sds->avg_load) {
3851                 *imbalance = 0;
3852                 return fix_small_imbalance(sds, this_cpu, imbalance);
3853         }
3854 
3855         /* Don't want to pull so many tasks that a group would go idle */
3856         max_pull = min(sds->max_load - sds->avg_load,
3857                         sds->max_load - sds->busiest_load_per_task);
3858 
3859         /* How much load to actually move to equalise the imbalance */
3860         *imbalance = min(max_pull * sds->busiest->__cpu_power,
3861                 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3862                         / SCHED_LOAD_SCALE;
3863 
3864         /*
3865          * if *imbalance is less than the average load per runnable task
3866          * there is no gaurantee that any tasks will be moved so we'll have
3867          * a think about bumping its value to force at least one task to be
3868          * moved
3869          */
3870         if (*imbalance < sds->busiest_load_per_task)
3871                 return fix_small_imbalance(sds, this_cpu, imbalance);
3872 
3873 }
3874 /******* find_busiest_group() helpers end here *********************/
3875 
3876 /**
3877  * find_busiest_group - Returns the busiest group within the sched_domain
3878  * if there is an imbalance. If there isn't an imbalance, and
3879  * the user has opted for power-savings, it returns a group whose
3880  * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3881  * such a group exists.
3882  *
3883  * Also calculates the amount of weighted load which should be moved
3884  * to restore balance.
3885  *
3886  * @sd: The sched_domain whose busiest group is to be returned.
3887  * @this_cpu: The cpu for which load balancing is currently being performed.
3888  * @imbalance: Variable which stores amount of weighted load which should
3889  *              be moved to restore balance/put a group to idle.
3890  * @idle: The idle status of this_cpu.
3891  * @sd_idle: The idleness of sd
3892  * @cpus: The set of CPUs under consideration for load-balancing.
3893  * @balance: Pointer to a variable indicating if this_cpu
3894  *      is the appropriate cpu to perform load balancing at this_level.
3895  *
3896  * Returns:     - the busiest group if imbalance exists.
3897  *              - If no imbalance and user has opted for power-savings balance,
3898  *                 return the least loaded group whose CPUs can be
3899  *                 put to idle by rebalancing its tasks onto our group.
3900  */
3901 static struct sched_group *
3902 find_busiest_group(struct sched_domain *sd, int this_cpu,
3903                    unsigned long *imbalance, enum cpu_idle_type idle,
3904                    int *sd_idle, const struct cpumask *cpus, int *balance)
3905 {
3906         struct sd_lb_stats sds;
3907 
3908         memset(&sds, 0, sizeof(sds));
3909 
3910         /*
3911          * Compute the various statistics relavent for load balancing at
3912          * this level.
3913          */
3914         update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3915                                         balance, &sds);
3916 
3917         /* Cases where imbalance does not exist from POV of this_cpu */
3918         /* 1) this_cpu is not the appropriate cpu to perform load balancing
3919          *    at this level.
3920          * 2) There is no busy sibling group to pull from.
3921          * 3) This group is the busiest group.
3922          * 4) This group is more busy than the avg busieness at this
3923          *    sched_domain.
3924          * 5) The imbalance is within the specified limit.
3925          * 6) Any rebalance would lead to ping-pong
3926          */
3927         if (balance && !(*balance))
3928                 goto ret;
3929 
3930         if (!sds.busiest || sds.busiest_nr_running == 0)
3931                 goto out_balanced;
3932 
3933         if (sds.this_load >= sds.max_load)
3934                 goto out_balanced;
3935 
3936         sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3937 
3938         if (sds.this_load >= sds.avg_load)
3939                 goto out_balanced;
3940 
3941         if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3942                 goto out_balanced;
3943 
3944         sds.busiest_load_per_task /= sds.busiest_nr_running;
3945         if (sds.group_imb)
3946                 sds.busiest_load_per_task =
3947                         min(sds.busiest_load_per_task, sds.avg_load);
3948 
3949         /*
3950          * We're trying to get all the cpus to the average_load, so we don't
3951          * want to push ourselves above the average load, nor do we wish to
3952          * reduce the max loaded cpu below the average load, as either of these
3953          * actions would just result in more rebalancing later, and ping-pong
3954          * tasks around. Thus we look for the minimum possible imbalance.
3955          * Negative imbalances (*we* are more loaded than anyone else) will
3956          * be counted as no imbalance for these purposes -- we can't fix that
3957          * by pulling tasks to us. Be careful of negative numbers as they'll
3958          * appear as very large values with unsigned longs.
3959          */
3960         if (sds.max_load <= sds.busiest_load_per_task)
3961                 goto out_balanced;
3962 
3963         /* Looks like there is an imbalance. Compute it */
3964         calculate_imbalance(&sds, this_cpu, imbalance);
3965         return sds.busiest;
3966 
3967 out_balanced:
3968         /*
3969          * There is no obvious imbalance. But check if we can do some balancing
3970          * to save power.
3971          */
3972         if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3973                 return sds.busiest;
3974 ret:
3975         *imbalance = 0;
3976         return NULL;
3977 }
3978 
3979 /*
3980  * find_busiest_queue - find the busiest runqueue among the cpus in group.
3981  */
3982 static struct rq *
3983 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3984                    unsigned long imbalance, const struct cpumask *cpus)
3985 {
3986         struct rq *busiest = NULL, *rq;
3987         unsigned long max_load = 0;
3988         int i;
3989 
3990         for_each_cpu(i, sched_group_cpus(group)) {
3991                 unsigned long wl;
3992 
3993                 if (!cpumask_test_cpu(i, cpus))
3994                         continue;
3995 
3996                 rq = cpu_rq(i);
3997                 wl = weighted_cpuload(i);
3998 
3999                 if (rq->nr_running == 1 && wl > imbalance)
4000                         continue;
4001 
4002                 if (wl > max_load) {
4003                         max_load = wl;
4004                         busiest = rq;
4005                 }
4006         }
4007 
4008         return busiest;
4009 }
4010 
4011 /*
4012  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4013  * so long as it is large enough.
4014  */
4015 #define MAX_PINNED_INTERVAL     512
4016 
4017 /* Working cpumask for load_balance and load_balance_newidle. */
4018 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4019 
4020 /*
4021  * Check this_cpu to ensure it is balanced within domain. Attempt to move
4022  * tasks if there is an imbalance.
4023  */
4024 static int load_balance(int this_cpu, struct rq *this_rq,
4025                         struct sched_domain *sd, enum cpu_idle_type idle,
4026                         int *balance)
4027 {
4028         int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4029         struct sched_group *group;
4030         unsigned long imbalance;
4031         struct rq *busiest;
4032         unsigned long flags;
4033         struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4034 
4035         cpumask_setall(cpus);
4036 
4037         /*
4038          * When power savings policy is enabled for the parent domain, idle
4039          * sibling can pick up load irrespective of busy siblings. In this case,
4040          * let the state of idle sibling percolate up as CPU_IDLE, instead of
4041          * portraying it as CPU_NOT_IDLE.
4042          */
4043         if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4044             !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4045                 sd_idle = 1;
4046 
4047         schedstat_inc(sd, lb_count[idle]);
4048 
4049 redo:
4050         update_shares(sd);
4051         group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4052                                    cpus, balance);
4053 
4054         if (*balance == 0)
4055                 goto out_balanced;
4056 
4057         if (!group) {
4058                 schedstat_inc(sd, lb_nobusyg[idle]);
4059                 goto out_balanced;
4060         }
4061 
4062         busiest = find_busiest_queue(group, idle, imbalance, cpus);
4063         if (!busiest) {
4064                 schedstat_inc(sd, lb_nobusyq[idle]);
4065                 goto out_balanced;
4066         }
4067 
4068         BUG_ON(busiest == this_rq);
4069 
4070         schedstat_add(sd, lb_imbalance[idle], imbalance);
4071 
4072         ld_moved = 0;
4073         if (busiest->nr_running > 1) {
4074                 /*
4075                  * Attempt to move tasks. If find_busiest_group has found
4076                  * an imbalance but busiest->nr_running <= 1, the group is
4077                  * still unbalanced. ld_moved simply stays zero, so it is
4078                  * correctly treated as an imbalance.
4079                  */
4080                 local_irq_save(flags);
4081                 double_rq_lock(this_rq, busiest);
4082                 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4083                                       imbalance, sd, idle, &all_pinned);
4084                 double_rq_unlock(this_rq, busiest);
4085                 local_irq_restore(flags);
4086 
4087                 /*
4088                  * some other cpu did the load balance for us.
4089                  */
4090                 if (ld_moved && this_cpu != smp_processor_id())
4091                         resched_cpu(this_cpu);
4092 
4093                 /* All tasks on this runqueue were pinned by CPU affinity */
4094                 if (unlikely(all_pinned)) {
4095                         cpumask_clear_cpu(cpu_of(busiest), cpus);
4096                         if (!cpumask_empty(cpus))
4097                                 goto redo;
4098                         goto out_balanced;
4099                 }
4100         }
4101 
4102         if (!ld_moved) {
4103                 schedstat_inc(sd, lb_failed[idle]);
4104                 sd->nr_balance_failed++;
4105 
4106                 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4107 
4108                         spin_lock_irqsave(&busiest->lock, flags);
4109 
4110                         /* don't kick the migration_thread, if the curr
4111                          * task on busiest cpu can't be moved to this_cpu
4112                          */
4113                         if (!cpumask_test_cpu(this_cpu,
4114                                               &busiest->curr->cpus_allowed)) {
4115                                 spin_unlock_irqrestore(&busiest->lock, flags);
4116                                 all_pinned = 1;
4117                                 goto out_one_pinned;
4118                         }
4119 
4120                         if (!busiest->active_balance) {
4121                                 busiest->active_balance = 1;
4122                                 busiest->push_cpu = this_cpu;
4123                                 active_balance = 1;
4124                         }
4125                         spin_unlock_irqrestore(&busiest->lock, flags);
4126                         if (active_balance)
4127                                 wake_up_process(busiest->migration_thread);
4128 
4129                         /*
4130                          * We've kicked active balancing, reset the failure
4131                          * counter.
4132                          */
4133                         sd->nr_balance_failed = sd->cache_nice_tries+1;
4134                 }
4135         } else
4136                 sd->nr_balance_failed = 0;
4137 
4138         if (likely(!active_balance)) {
4139                 /* We were unbalanced, so reset the balancing interval */
4140                 sd->balance_interval = sd->min_interval;
4141         } else {
4142                 /*
4143                  * If we've begun active balancing, start to back off. This
4144                  * case may not be covered by the all_pinned logic if there
4145                  * is only 1 task on the busy runqueue (because we don't call
4146                  * move_tasks).
4147                  */
4148                 if (sd->balance_interval < sd->max_interval)
4149                         sd->balance_interval *= 2;
4150         }
4151 
4152         if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4153             !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4154                 ld_moved = -1;
4155 
4156         goto out;
4157 
4158 out_balanced:
4159         schedstat_inc(sd, lb_balanced[idle]);
4160 
4161         sd->nr_balance_failed = 0;
4162 
4163 out_one_pinned:
4164         /* tune up the balancing interval */
4165         if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4166                         (sd->balance_interval < sd->max_interval))
4167                 sd->balance_interval *= 2;
4168 
4169         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4170             !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4171                 ld_moved = -1;
4172         else
4173                 ld_moved = 0;
4174 out:
4175         if (ld_moved)
4176                 update_shares(sd);
4177         return ld_moved;
4178 }
4179 
4180 /*
4181  * Check this_cpu to ensure it is balanced within domain. Attempt to move
4182  * tasks if there is an imbalance.
4183  *
4184  * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4185  * this_rq is locked.
4186  */
4187 static int
4188 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4189 {
4190         struct sched_group *group;
4191         struct rq *busiest = NULL;
4192         unsigned long imbalance;
4193         int ld_moved = 0;
4194         int sd_idle = 0;
4195         int all_pinned = 0;
4196         struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4197 
4198         cpumask_setall(cpus);
4199 
4200         /*
4201          * When power savings policy is enabled for the parent domain, idle
4202          * sibling can pick up load irrespective of busy siblings. In this case,
4203          * let the state of idle sibling percolate up as IDLE, instead of
4204          * portraying it as CPU_NOT_IDLE.
4205          */
4206         if (sd->flags & SD_SHARE_CPUPOWER &&
4207             !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4208                 sd_idle = 1;
4209 
4210         schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4211 redo:
4212         update_shares_locked(this_rq, sd);
4213         group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4214                                    &sd_idle, cpus, NULL);
4215         if (!group) {
4216                 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4217                 goto out_balanced;
4218         }
4219 
4220         busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4221         if (!busiest) {
4222                 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4223                 goto out_balanced;
4224         }
4225 
4226         BUG_ON(busiest == this_rq);
4227 
4228         schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4229 
4230         ld_moved = 0;
4231         if (busiest->nr_running > 1) {
4232                 /* Attempt to move tasks */
4233                 double_lock_balance(this_rq, busiest);
4234                 /* this_rq->clock is already updated */
4235                 update_rq_clock(busiest);
4236                 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4237                                         imbalance, sd, CPU_NEWLY_IDLE,
4238                                         &all_pinned);
4239                 double_unlock_balance(this_rq, busiest);
4240 
4241                 if (unlikely(all_pinned)) {
4242                         cpumask_clear_cpu(cpu_of(busiest), cpus);
4243                         if (!cpumask_empty(cpus))
4244                                 goto redo;
4245                 }
4246         }
4247 
4248         if (!ld_moved) {
4249                 int active_balance = 0;
4250 
4251                 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4252                 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4253                     !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4254                         return -1;
4255 
4256                 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4257                         return -1;
4258 
4259                 if (sd->nr_balance_failed++ < 2)
4260                         return -1;
4261 
4262                 /*
4263                  * The only task running in a non-idle cpu can be moved to this
4264                  * cpu in an attempt to completely freeup the other CPU
4265                  * package. The same method used to move task in load_balance()
4266                  * have been extended for load_balance_newidle() to speedup
4267                  * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4268                  *
4269                  * The package power saving logic comes from
4270                  * find_busiest_group().  If there are no imbalance, then
4271                  * f_b_g() will return NULL.  However when sched_mc={1,2} then
4272                  * f_b_g() will select a group from which a running task may be
4273                  * pulled to this cpu in order to make the other package idle.
4274                  * If there is no opportunity to make a package idle and if
4275                  * there are no imbalance, then f_b_g() will return NULL and no
4276                  * action will be taken in load_balance_newidle().
4277                  *
4278                  * Under normal task pull operation due to imbalance, there
4279                  * will be more than one task in the source run queue and
4280                  * move_tasks() will succeed.  ld_moved will be true and this
4281                  * active balance code will not be triggered.
4282                  */
4283 
4284                 /* Lock busiest in correct order while this_rq is held */
4285                 double_lock_balance(this_rq, busiest);
4286 
4287                 /*
4288                  * don't kick the migration_thread, if the curr
4289                  * task on busiest cpu can't be moved to this_cpu
4290                  */
4291                 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4292                         double_unlock_balance(this_rq, busiest);
4293                         all_pinned = 1;
4294                         return ld_moved;
4295                 }
4296 
4297                 if (!busiest->active_balance) {
4298                         busiest->active_balance = 1;
4299                         busiest->push_cpu = this_cpu;
4300                         active_balance = 1;
4301                 }
4302 
4303                 double_unlock_balance(this_rq, busiest);
4304                 /*
4305                  * Should not call ttwu while holding a rq->lock
4306                  */
4307                 spin_unlock(&this_rq->lock);
4308                 if (active_balance)
4309                         wake_up_process(busiest->migration_thread);
4310                 spin_lock(&this_rq->lock);
4311 
4312         } else
4313                 sd->nr_balance_failed = 0;
4314 
4315         update_shares_locked(this_rq, sd);
4316         return ld_moved;
4317 
4318 out_balanced:
4319         schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4320         if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4321             !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4322                 return -1;
4323         sd->nr_balance_failed = 0;
4324 
4325         return 0;
4326 }
4327 
4328 /*
4329  * idle_balance is called by schedule() if this_cpu is about to become
4330  * idle. Attempts to pull tasks from other CPUs.
4331  */
4332 static void idle_balance(int this_cpu, struct rq *this_rq)
4333 {
4334         struct sched_domain *sd;
4335         int pulled_task = 0;
4336         unsigned long next_balance = jiffies + HZ;
4337 
4338         for_each_domain(this_cpu, sd) {
4339                 unsigned long interval;
4340 
4341                 if (!(sd->flags & SD_LOAD_BALANCE))
4342                         continue;
4343 
4344                 if (sd->flags & SD_BALANCE_NEWIDLE)
4345                         /* If we've pulled tasks over stop searching: */
4346                         pulled_task = load_balance_newidle(this_cpu, this_rq,
4347                                                            sd);
4348 
4349                 interval = msecs_to_jiffies(sd->balance_interval);
4350                 if (time_after(next_balance, sd->last_balance + interval))
4351                         next_balance = sd->last_balance + interval;
4352                 if (pulled_task)
4353                         break;
4354         }
4355         if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4356                 /*
4357                  * We are going idle. next_balance may be set based on
4358                  * a busy processor. So reset next_balance.
4359                  */
4360                 this_rq->next_balance = next_balance;
4361         }
4362 }
4363 
4364 /*
4365  * active_load_balance is run by migration threads. It pushes running tasks
4366  * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4367  * running on each physical CPU where possible, and avoids physical /
4368  * logical imbalances.
4369  *
4370  * Called with busiest_rq locked.
4371  */
4372 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4373 {
4374         int target_cpu = busiest_rq->push_cpu;
4375         struct sched_domain *sd;
4376         struct rq *target_rq;
4377 
4378         /* Is there any task to move? */
4379         if (busiest_rq->nr_running <= 1)
4380                 return;
4381 
4382         target_rq = cpu_rq(target_cpu);
4383 
4384         /*
4385          * This condition is "impossible", if it occurs
4386          * we need to fix it. Originally reported by
4387          * Bjorn Helgaas on a 128-cpu setup.
4388          */
4389         BUG_ON(busiest_rq == target_rq);
4390 
4391         /* move a task from busiest_rq to target_rq */
4392         double_lock_balance(busiest_rq, target_rq);
4393         update_rq_clock(busiest_rq);
4394         update_rq_clock(target_rq);
4395 
4396         /* Search for an sd spanning us and the target CPU. */
4397         for_each_domain(target_cpu, sd) {
4398                 if ((sd->flags & SD_LOAD_BALANCE) &&
4399                     cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4400                                 break;
4401         }
4402 
4403         if (likely(sd)) {
4404                 schedstat_inc(sd, alb_count);
4405 
4406                 if (move_one_task(target_rq, target_cpu, busiest_rq,
4407                                   sd, CPU_IDLE))
4408                         schedstat_inc(sd, alb_pushed);
4409                 else
4410                         schedstat_inc(sd, alb_failed);
4411         }
4412         double_unlock_balance(busiest_rq, target_rq);
4413 }
4414 
4415 #ifdef CONFIG_NO_HZ
4416 static struct {
4417         atomic_t load_balancer;
4418         cpumask_var_t cpu_mask;
4419         cpumask_var_t ilb_grp_nohz_mask;
4420 } nohz ____cacheline_aligned = {
4421         .load_balancer = ATOMIC_INIT(-1),
4422 };
4423 
4424 int get_nohz_load_balancer(void)
4425 {
4426         return atomic_read(&nohz.load_balancer);
4427 }
4428 
4429 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4430 /**
4431  * lowest_flag_domain - Return lowest sched_domain containing flag.
4432  * @cpu:        The cpu whose lowest level of sched domain is to
4433  *              be returned.
4434  * @flag:       The flag to check for the lowest sched_domain
4435  *              for the given cpu.
4436  *
4437  * Returns the lowest sched_domain of a cpu which contains the given flag.
4438  */
4439 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4440 {
4441         struct sched_domain *sd;
4442 
4443         for_each_domain(cpu, sd)
4444                 if (sd && (sd->flags & flag))
4445                         break;
4446 
4447         return sd;
4448 }
4449 
4450 /**
4451  * for_each_flag_domain - Iterates over sched_domains containing the flag.
4452  * @cpu:        The cpu whose domains we're iterating over.
4453  * @sd:         variable holding the value of the power_savings_sd
4454  *              for cpu.
4455  * @flag:       The flag to filter the sched_domains to be iterated.
4456  *
4457  * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4458  * set, starting from the lowest sched_domain to the highest.
4459  */
4460 #define for_each_flag_domain(cpu, sd, flag) \
4461         for (sd = lowest_flag_domain(cpu, flag); \
4462                 (sd && (sd->flags & flag)); sd = sd->parent)
4463 
4464 /**
4465  * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4466  * @ilb_group:  group to be checked for semi-idleness
4467  *
4468  * Returns:     1 if the group is semi-idle. 0 otherwise.
4469  *
4470  * We define a sched_group to be semi idle if it has atleast one idle-CPU
4471  * and atleast one non-idle CPU. This helper function checks if the given
4472  * sched_group is semi-idle or not.
4473  */
4474 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4475 {
4476         cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4477                                         sched_group_cpus(ilb_group));
4478 
4479         /*
4480          * A sched_group is semi-idle when it has atleast one busy cpu
4481          * and atleast one idle cpu.
4482          */
4483         if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4484                 return 0;
4485 
4486         if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4487                 return 0;
4488 
4489         return 1;
4490 }
4491 /**
4492  * find_new_ilb - Finds the optimum idle load balancer for nomination.
4493  * @cpu:        The cpu which is nominating a new idle_load_balancer.
4494  *
4495  * Returns:     Returns the id of the idle load balancer if it exists,
4496  *              Else, returns >= nr_cpu_ids.
4497  *
4498  * This algorithm picks the idle load balancer such that it belongs to a
4499  * semi-idle powersavings sched_domain. The idea is to try and avoid
4500  * completely idle packages/cores just for the purpose of idle load balancing
4501  * when there are other idle cpu's which are better suited for that job.
4502  */
4503 static int find_new_ilb(int cpu)
4504 {
4505         struct sched_domain *sd;
4506         struct sched_group *ilb_group;
4507 
4508         /*
4509          * Have idle load balancer selection from semi-idle packages only
4510          * when power-aware load balancing is enabled
4511          */
4512         if (!(sched_smt_power_savings || sched_mc_power_savings))
4513                 goto out_done;
4514 
4515         /*
4516          * Optimize for the case when we have no idle CPUs or only one
4517          * idle CPU. Don't walk the sched_domain hierarchy in such cases
4518          */
4519         if (cpumask_weight(nohz.cpu_mask) < 2)
4520                 goto out_done;
4521 
4522         for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4523                 ilb_group = sd->groups;
4524 
4525                 do {
4526                         if (is_semi_idle_group(ilb_group))
4527                                 return cpumask_first(nohz.ilb_grp_nohz_mask);
4528 
4529                         ilb_group = ilb_group->next;
4530 
4531                 } while (ilb_group != sd->groups);
4532         }
4533 
4534 out_done:
4535         return cpumask_first(nohz.cpu_mask);
4536 }
4537 #else /*  (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4538 static inline int find_new_ilb(int call_cpu)
4539 {
4540         return cpumask_first(nohz.cpu_mask);
4541 }
4542 #endif
4543 
4544 /*
4545  * This routine will try to nominate the ilb (idle load balancing)
4546  * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4547  * load balancing on behalf of all those cpus. If all the cpus in the system
4548  * go into this tickless mode, then there will be no ilb owner (as there is
4549  * no need for one) and all the cpus will sleep till the next wakeup event
4550  * arrives...
4551  *
4552  * For the ilb owner, tick is not stopped. And this tick will be used
4553  * for idle load balancing. ilb owner will still be part of
4554  * nohz.cpu_mask..
4555  *
4556  * While stopping the tick, this cpu will become the ilb owner if there
4557  * is no other owner. And will be the owner till that cpu becomes busy
4558  * or if all cpus in the system stop their ticks at which point
4559  * there is no need for ilb owner.
4560  *
4561  * When the ilb owner becomes busy, it nominates another owner, during the
4562  * next busy scheduler_tick()
4563  */
4564 int select_nohz_load_balancer(int stop_tick)
4565 {
4566         int cpu = smp_processor_id();
4567 
4568         if (stop_tick) {
4569                 cpu_rq(cpu)->in_nohz_recently = 1;
4570 
4571                 if (!cpu_active(cpu)) {
4572                         if (atomic_read(&nohz.load_balancer) != cpu)
4573                                 return 0;
4574 
4575                         /*
4576                          * If we are going offline and still the leader,
4577                          * give up!
4578                          */
4579                         if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4580                                 BUG();
4581 
4582                         return 0;
4583                 }
4584 
4585                 cpumask_set_cpu(cpu, nohz.cpu_mask);
4586 
4587                 /* time for ilb owner also to sleep */
4588                 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4589                         if (atomic_read(&nohz.load_balancer) == cpu)
4590                                 atomic_set(&nohz.load_balancer, -1);
4591                         return 0;
4592                 }
4593 
4594                 if (atomic_read(&nohz.load_balancer) == -1) {
4595                         /* make me the ilb owner */
4596                         if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4597                                 return 1;
4598                 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4599                         int new_ilb;
4600 
4601                         if (!(sched_smt_power_savings ||
4602                                                 sched_mc_power_savings))
4603                                 return 1;
4604                         /*
4605                          * Check to see if there is a more power-efficient
4606                          * ilb.
4607                          */
4608                         new_ilb = find_new_ilb(cpu);
4609                         if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4610                                 atomic_set(&nohz.load_balancer, -1);
4611                                 resched_cpu(new_ilb);
4612                                 return 0;
4613                         }
4614                         return 1;
4615                 }
4616         } else {
4617                 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4618                         return 0;
4619 
4620                 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4621 
4622                 if (atomic_read(&nohz.load_balancer) == cpu)
4623                         if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4624                                 BUG();
4625         }
4626         return 0;
4627 }
4628 #endif
4629 
4630 static DEFINE_SPINLOCK(balancing);
4631 
4632 /*
4633  * It checks each scheduling domain to see if it is due to be balanced,
4634  * and initiates a balancing operation if so.
4635  *
4636  * Balancing parameters are set up in arch_init_sched_domains.
4637  */
4638 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4639 {
4640         int balance = 1;
4641         struct rq *rq = cpu_rq(cpu);
4642         unsigned long interval;
4643         struct sched_domain *sd;
4644         /* Earliest time when we have to do rebalance again */
4645         unsigned long next_balance = jiffies + 60*HZ;
4646         int update_next_balance = 0;
4647         int need_serialize;
4648 
4649         for_each_domain(cpu, sd) {
4650                 if (!(sd->flags & SD_LOAD_BALANCE))
4651                         continue;
4652 
4653                 interval = sd->balance_interval;
4654                 if (idle != CPU_IDLE)
4655                         interval *= sd->busy_factor;
4656 
4657                 /* scale ms to jiffies */
4658                 interval = msecs_to_jiffies(interval);
4659                 if (unlikely(!interval))
4660                         interval = 1;
4661                 if (interval > HZ*NR_CPUS/10)
4662                         interval = HZ*NR_CPUS/10;
4663 
4664                 need_serialize = sd->flags & SD_SERIALIZE;
4665 
4666                 if (need_serialize) {
4667                         if (!spin_trylock(&balancing))
4668                                 goto out;
4669                 }
4670 
4671                 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4672                         if (load_balance(cpu, rq, sd, idle, &balance)) {
4673                                 /*
4674                                  * We've pulled tasks over so either we're no
4675                                  * longer idle, or one of our SMT siblings is
4676                                  * not idle.
4677                                  */
4678                                 idle = CPU_NOT_IDLE;
4679                         }
4680                         sd->last_balance = jiffies;
4681                 }
4682                 if (need_serialize)
4683                         spin_unlock(&balancing);
4684 out:
4685                 if (time_after(next_balance, sd->last_balance + interval)) {
4686                         next_balance = sd->last_balance + interval;
4687                         update_next_balance = 1;
4688                 }
4689 
4690                 /*
4691                  * Stop the load balance at this level. There is another
4692                  * CPU in our sched group which is doing load balancing more
4693                  * actively.
4694                  */
4695                 if (!balance)
4696                         break;
4697         }
4698 
4699         /*
4700          * next_balance will be updated only when there is a need.
4701          * When the cpu is attached to null domain for ex, it will not be
4702          * updated.
4703          */
4704         if (likely(update_next_balance))
4705                 rq->next_balance = next_balance;
4706 }
4707 
4708 /*
4709  * run_rebalance_domains is triggered when needed from the scheduler tick.
4710  * In CONFIG_NO_HZ case, the idle load balance owner will do the
4711  * rebalancing for all the cpus for whom scheduler ticks are stopped.
4712  */
4713 static void run_rebalance_domains(struct softirq_action *h)
4714 {
4715         int this_cpu = smp_processor_id();
4716         struct rq *this_rq = cpu_rq(this_cpu);
4717         enum cpu_idle_type idle = this_rq->idle_at_tick ?
4718                                                 CPU_IDLE : CPU_NOT_IDLE;
4719 
4720         rebalance_domains(this_cpu, idle);
4721 
4722 #ifdef CONFIG_NO_HZ
4723         /*
4724          * If this cpu is the owner for idle load balancing, then do the
4725          * balancing on behalf of the other idle cpus whose ticks are
4726          * stopped.
4727          */
4728         if (this_rq->idle_at_tick &&
4729             atomic_read(&nohz.load_balancer) == this_cpu) {
4730                 struct rq *rq;
4731                 int balance_cpu;
4732 
4733                 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4734                         if (balance_cpu == this_cpu)
4735                                 continue;
4736 
4737                         /*
4738                          * If this cpu gets work to do, stop the load balancing
4739                          * work being done for other cpus. Next load
4740                          * balancing owner will pick it up.
4741                          */
4742                         if (need_resched())
4743                                 break;
4744 
4745                         rebalance_domains(balance_cpu, CPU_IDLE);
4746 
4747                         rq = cpu_rq(balance_cpu);
4748                         if (time_after(this_rq->next_balance, rq->next_balance))
4749                                 this_rq->next_balance = rq->next_balance;
4750                 }
4751         }
4752 #endif
4753 }
4754 
4755 static inline int on_null_domain(int cpu)
4756 {
4757         return !rcu_dereference(cpu_rq(cpu)->sd);
4758 }
4759 
4760 /*
4761  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4762  *
4763  * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4764  * idle load balancing owner or decide to stop the periodic load balancing,
4765  * if the whole system is idle.
4766  */
4767 static inline void trigger_load_balance(struct rq *rq, int cpu)
4768 {
4769 #ifdef CONFIG_NO_HZ
4770         /*
4771          * If we were in the nohz mode recently and busy at the current
4772          * scheduler tick, then check if we need to nominate new idle
4773          * load balancer.
4774          */
4775         if (rq->in_nohz_recently && !rq->idle_at_tick) {
4776                 rq->in_nohz_recently = 0;
4777 
4778                 if (atomic_read(&nohz.load_balancer) == cpu) {
4779                         cpumask_clear_cpu(cpu, nohz.cpu_mask);
4780                         atomic_set(&nohz.load_balancer, -1);
4781                 }
4782 
4783                 if (atomic_read(&nohz.load_balancer) == -1) {
4784                         int ilb = find_new_ilb(cpu);
4785 
4786                         if (ilb < nr_cpu_ids)
4787                                 resched_cpu(ilb);
4788                 }
4789         }
4790 
4791         /*
4792          * If this cpu is idle and doing idle load balancing for all the
4793          * cpus with ticks stopped, is it time for that to stop?
4794          */
4795         if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4796             cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4797                 resched_cpu(cpu);
4798                 return;
4799         }
4800 
4801         /*
4802          * If this cpu is idle and the idle load balancing is done by
4803          * someone else, then no need raise the SCHED_SOFTIRQ
4804          */
4805         if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4806             cpumask_test_cpu(cpu, nohz.cpu_mask))
4807                 return;
4808 #endif
4809         /* Don't need to rebalance while attached to NULL domain */
4810         if (time_after_eq(jiffies, rq->next_balance) &&
4811             likely(!on_null_domain(cpu)))
4812                 raise_softirq(SCHED_SOFTIRQ);
4813 }
4814 
4815 #else   /* CONFIG_SMP */
4816 
4817 /*
4818  * on UP we do not need to balance between CPUs:
4819  */
4820 static inline void idle_balance(int cpu, struct rq *rq)
4821 {
4822 }
4823 
4824 #endif
4825 
4826 DEFINE_PER_CPU(struct kernel_stat, kstat);
4827 
4828 EXPORT_PER_CPU_SYMBOL(kstat);
4829 
4830 /*
4831  * Return any ns on the sched_clock that have not yet been accounted in
4832  * @p in case that task is currently running.
4833  *
4834  * Called with task_rq_lock() held on @rq.
4835  */
4836 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4837 {
4838         u64 ns = 0;
4839 
4840         if (task_current(rq, p)) {
4841                 update_rq_clock(rq);
4842                 ns = rq->clock - p->se.exec_start;
4843                 if ((s64)ns < 0)
4844                         ns = 0;
4845         }
4846 
4847         return ns;
4848 }
4849 
4850 unsigned long long task_delta_exec(struct task_struct *p)
4851 {
4852         unsigned long flags;
4853         struct rq *rq;
4854         u64 ns = 0;
4855 
4856         rq = task_rq_lock(p, &flags);
4857         ns = do_task_delta_exec(p, rq);
4858         task_rq_unlock(rq, &flags);
4859 
4860         return ns;
4861 }
4862 
4863 /*
4864  * Return accounted runtime for the task.
4865  * In case the task is currently running, return the runtime plus current's
4866  * pending runtime that have not been accounted yet.
4867  */
4868 unsigned long long task_sched_runtime(struct task_struct *p)
4869 {
4870         unsigned long flags;
4871         struct rq *rq;
4872         u64 ns = 0;
4873 
4874         rq = task_rq_lock(p, &flags);
4875         ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4876         task_rq_unlock(rq, &flags);
4877 
4878         return ns;
4879 }
4880 
4881 /*
4882  * Return sum_exec_runtime for the thread group.
4883  * In case the task is currently running, return the sum plus current's
4884  * pending runtime that have not been accounted yet.
4885  *
4886  * Note that the thread group might have other running tasks as well,
4887  * so the return value not includes other pending runtime that other
4888  * running tasks might have.
4889  */
4890 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4891 {
4892         struct task_cputime totals;
4893         unsigned long flags;
4894         struct rq *rq;
4895         u64 ns;
4896 
4897         rq = task_rq_lock(p, &flags);
4898         thread_group_cputime(p, &totals);
4899         ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4900         task_rq_unlock(rq, &flags);
4901 
4902         return ns;
4903 }
4904 
4905 /*
4906  * Account user cpu time to a process.
4907  * @p: the process that the cpu time gets accounted to
4908  * @cputime: the cpu time spent in user space since the last update
4909  * @cputime_scaled: cputime scaled by cpu frequency
4910  */
4911 void account_user_time(struct task_struct *p, cputime_t cputime,
4912                        cputime_t cputime_scaled)
4913 {
4914         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4915         cputime64_t tmp;
4916 
4917         /* Add user time to process. */
4918         p->utime = cputime_add(p->utime, cputime);
4919         p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4920         account_group_user_time(p, cputime);
4921 
4922         /* Add user time to cpustat. */
4923         tmp = cputime_to_cputime64(cputime);
4924         if (TASK_NICE(p) > 0)
4925                 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4926         else
4927                 cpustat->user = cputime64_add(cpustat->user, tmp);
4928 
4929         cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4930         /* Account for user time used */
4931         acct_update_integrals(p);
4932 }
4933 
4934 /*
4935  * Account guest cpu time to a process.
4936  * @p: the process that the cpu time gets accounted to
4937  * @cputime: the cpu time spent in virtual machine since the last update
4938  * @cputime_scaled: cputime scaled by cpu frequency
4939  */
4940 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4941                                cputime_t cputime_scaled)
4942 {
4943         cputime64_t tmp;
4944         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4945 
4946         tmp = cputime_to_cputime64(cputime);
4947 
4948         /* Add guest time to process. */
4949         p->utime = cputime_add(p->utime, cputime);
4950         p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4951         account_group_user_time(p, cputime);
4952         p->gtime = cputime_add(p->gtime, cputime);
4953 
4954         /* Add guest time to cpustat. */
4955         cpustat->user = cputime64_add(cpustat->user, tmp);
4956         cpustat->guest = cputime64_add(cpustat->guest, tmp);
4957 }
4958 
4959 /*
4960  * Account system cpu time to a process.
4961  * @p: the process that the cpu time gets accounted to
4962  * @hardirq_offset: the offset to subtract from hardirq_count()
4963  * @cputime: the cpu time spent in kernel space since the last update
4964  * @cputime_scaled: cputime scaled by cpu frequency
4965  */
4966 void account_system_time(struct task_struct *p, int hardirq_offset,
4967                          cputime_t cputime, cputime_t cputime_scaled)
4968 {
4969         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4970         cputime64_t tmp;
4971 
4972         if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4973                 account_guest_time(p, cputime, cputime_scaled);
4974                 return;
4975         }
4976 
4977         /* Add system time to process. */
4978         p->stime = cputime_add(p->stime, cputime);
4979         p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4980         account_group_system_time(p, cputime);
4981 
4982         /* Add system time to cpustat. */
4983         tmp = cputime_to_cputime64(cputime);
4984         if (hardirq_count() - hardirq_offset)
4985                 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4986         else if (softirq_count())
4987                 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4988         else
4989                 cpustat->system = cputime64_add(cpustat->system, tmp);
4990 
4991         cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4992 
4993         /* Account for system time used */
4994         acct_update_integrals(p);
4995 }
4996 
4997 /*
4998  * Account for involuntary wait time.
4999  * @steal: the cpu time spent in involuntary wait
5000  */
5001 void account_steal_time(cputime_t cputime)
5002 {
5003         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5004         cputime64_t cputime64 = cputime_to_cputime64(cputime);
5005 
5006         cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5007 }
5008 
5009 /*
5010  * Account for idle time.
5011  * @cputime: the cpu time spent in idle wait
5012  */
5013 void account_idle_time(cputime_t cputime)
5014 {
5015         struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5016         cputime64_t cputime64 = cputime_to_cputime64(cputime);
5017         struct rq *rq = this_rq();
5018 
5019         if (atomic_read(&rq->nr_iowait) > 0)
5020                 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5021         else
5022                 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5023 }
5024 
5025 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5026 
5027 /*
5028  * Account a single tick of cpu time.
5029  * @p: the process that the cpu time gets accounted to
5030  * @user_tick: indicates if the tick is a user or a system tick
5031  */
5032 void account_process_tick(struct task_struct *p, int user_tick)
5033 {
5034         cputime_t one_jiffy = jiffies_to_cputime(1);
5035         cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5036         struct rq *rq = this_rq();
5037 
5038         if (user_tick)
5039                 account_user_time(p, one_jiffy, one_jiffy_scaled);
5040         else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5041                 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5042                                     one_jiffy_scaled);
5043         else
5044                 account_idle_time(one_jiffy);
5045 }
5046 
5047 /*
5048  * Account multiple ticks of steal time.
5049  * @p: the process from which the cpu time has been stolen
5050  * @ticks: number of stolen ticks
5051  */
5052 void account_steal_ticks(unsigned long ticks)
5053 {
5054         account_steal_time(jiffies_to_cputime(ticks));
5055 }
5056 
5057 /*
5058  * Account multiple ticks of idle time.
5059  * @ticks: number of stolen ticks
5060  */
5061 void account_idle_ticks(unsigned long ticks)
5062 {
5063         account_idle_time(jiffies_to_cputime(ticks));
5064 }
5065 
5066 #endif
5067 
5068 /*
5069  * Use precise platform statistics if available:
5070  */
5071 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5072 cputime_t task_utime(struct task_struct *p)
5073 {
5074         return p->utime;
5075 }
5076 
5077 cputime_t task_stime(struct task_struct *p)
5078 {
5079         return p->stime;
5080 }
5081 #else
5082 cputime_t task_utime(struct task_struct *p)
5083 {
5084         clock_t utime = cputime_to_clock_t(p->utime),
5085                 total = utime + cputime_to_clock_t(p->stime);
5086         u64 temp;
5087 
5088         /*
5089          * Use CFS's precise accounting:
5090          */
5091         temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5092 
5093         if (total) {
5094                 temp *= utime;
5095                 do_div(temp, total);
5096         }
5097         utime = (clock_t)temp;
5098 
5099         p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5100         return p->prev_utime;
5101 }
5102 
5103 cputime_t task_stime(struct task_struct *p)
5104 {
5105         clock_t stime;
5106 
5107         /*
5108          * Use CFS's precise accounting. (we subtract utime from
5109          * the total, to make sure the total observed by userspace
5110          * grows monotonically - apps rely on that):
5111          */
5112         stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5113                         cputime_to_clock_t(task_utime(p));
5114 
5115         if (stime >= 0)
5116                 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5117 
5118         return p->prev_stime;
5119 }
5120 #endif
5121 
5122 inline cputime_t task_gtime(struct task_struct *p)
5123 {
5124         return p->gtime;
5125 }
5126 
5127 /*
5128  * This function gets called by the timer code, with HZ frequency.
5129  * We call it with interrupts disabled.
5130  *
5131  * It also gets called by the fork code, when changing the parent's
5132  * timeslices.
5133  */
5134 void scheduler_tick(void)
5135 {
5136         int cpu = smp_processor_id();
5137         struct rq *rq = cpu_rq(cpu);
5138         struct task_struct *curr = rq->curr;
5139 
5140         sched_clock_tick();
5141 
5142         spin_lock(&rq->lock);
5143         update_rq_clock(rq);
5144         update_cpu_load(rq);
5145         curr->sched_class->task_tick(rq, curr, 0);
5146         spin_unlock(&rq->lock);
5147 
5148         perf_counter_task_tick(curr, cpu);
5149 
5150 #ifdef CONFIG_SMP
5151         rq->idle_at_tick = idle_cpu(cpu);
5152         trigger_load_balance(rq, cpu);
5153 #endif
5154 }
5155 
5156 notrace unsigned long get_parent_ip(unsigned long addr)
5157 {
5158         if (in_lock_functions(addr)) {
5159                 addr = CALLER_ADDR2;
5160                 if (in_lock_functions(addr))
5161                         addr = CALLER_ADDR3;
5162         }
5163         return addr;
5164 }
5165 
5166 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5167                                 defined(CONFIG_PREEMPT_TRACER))
5168 
5169 void __kprobes add_preempt_count(int val)
5170 {
5171 #ifdef CONFIG_DEBUG_PREEMPT
5172         /*
5173          * Underflow?
5174          */
5175         if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5176                 return;
5177 #endif
5178         preempt_count() += val;
5179 #ifdef CONFIG_DEBUG_PREEMPT
5180         /*
5181          * Spinlock count overflowing soon?
5182          */
5183         DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5184                                 PREEMPT_MASK - 10);
5185 #endif
5186         if (preempt_count() == val)
5187                 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5188 }
5189 EXPORT_SYMBOL(add_preempt_count);
5190 
5191 void __kprobes sub_preempt_count(int val)
5192 {
5193 #ifdef CONFIG_DEBUG_PREEMPT
5194         /*
5195          * Underflow?
5196          */
5197         if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5198                 return;
5199         /*
5200          * Is the spinlock portion underflowing?
5201          */
5202         if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5203                         !(preempt_count() & PREEMPT_MASK)))
5204                 return;
5205 #endif
5206 
5207         if (preempt_count() == val)
5208                 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5209         preempt_count() -= val;
5210 }
5211 EXPORT_SYMBOL(sub_preempt_count);
5212 
5213 #endif
5214 
5215 /*
5216  * Print scheduling while atomic bug:
5217  */
5218 static noinline void __schedule_bug(struct task_struct *prev)
5219 {
5220         struct pt_regs *regs = get_irq_regs();
5221 
5222         printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5223                 prev->comm, prev->pid, preempt_count());
5224 
5225         debug_show_held_locks(prev);
5226         print_modules();
5227         if (irqs_disabled())
5228                 print_irqtrace_events(prev);
5229 
5230         if (regs)
5231                 show_regs(regs);
5232         else
5233                 dump_stack();
5234 }
5235 
5236 /*
5237  * Various schedule()-time debugging checks and statistics:
5238  */
5239 static inline void schedule_debug(struct task_struct *prev)
5240 {
5241         /*
5242          * Test if we are atomic. Since do_exit() needs to call into
5243          * schedule() atomically, we ignore that path for now.
5244          * Otherwise, whine if we are scheduling when we should not be.
5245          */
5246         if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5247                 __schedule_bug(prev);
5248 
5249         profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5250 
5251         schedstat_inc(this_rq(), sched_count);
5252 #ifdef CONFIG_SCHEDSTATS
5253         if (unlikely(prev->lock_depth >= 0)) {
5254                 schedstat_inc(this_rq(), bkl_count);
5255                 schedstat_inc(prev, sched_info.bkl_count);
5256         }
5257 #endif
5258 }
5259 
5260 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5261 {
5262         if (prev->state == TASK_RUNNING) {
5263                 u64 runtime = prev->se.sum_exec_runtime;
5264 
5265                 runtime -= prev->se.prev_sum_exec_runtime;
5266                 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5267 
5268                 /*
5269                  * In order to avoid avg_overlap growing stale when we are
5270                  * indeed overlapping and hence not getting put to sleep, grow
5271                  * the avg_overlap on preemption.
5272                  *
5273                  * We use the average preemption runtime because that
5274                  * correlates to the amount of cache footprint a task can
5275                  * build up.
5276                  */
5277                 update_avg(&prev->se.avg_overlap, runtime);
5278         }
5279         prev->sched_class->put_prev_task(rq, prev);
5280 }
5281 
5282 /*
5283  * Pick up the highest-prio task:
5284  */
5285 static inline struct task_struct *
5286 pick_next_task(struct rq *rq)
5287 {
5288         const struct sched_class *class;
5289         struct task_struct *p;
5290 
5291         /*
5292          * Optimization: we know that if all tasks are in
5293          * the fair class we can call that function directly:
5294          */
5295         if (likely(rq->nr_running == rq->cfs.nr_running)) {
5296                 p = fair_sched_class.pick_next_task(rq);
5297                 if (likely(p))
5298                         return p;
5299         }
5300 
5301         class = sched_class_highest;
5302         for ( ; ; ) {
5303                 p = class->pick_next_task(rq);
5304                 if (p)
5305                         return p;
5306                 /*
5307                  * Will never be NULL as the idle class always
5308                  * returns a non-NULL p:
5309                  */
5310                 class = class->next;
5311         }
5312 }
5313 
5314 /*
5315  * schedule() is the main scheduler function.
5316  */
5317 asmlinkage void __sched schedule(void)
5318 {
5319         struct task_struct *prev, *next;
5320         unsigned long *switch_count;
5321         struct rq *rq;
5322         int cpu;
5323 
5324 need_resched:
5325         preempt_disable();
5326         cpu = smp_processor_id();
5327         rq = cpu_rq(cpu);
5328         rcu_qsctr_inc(cpu);
5329         prev = rq->curr;
5330         switch_count = &prev->nivcsw;
5331 
5332         release_kernel_lock(prev);
5333 need_resched_nonpreemptible:
5334 
5335         schedule_debug(prev);
5336 
5337         if (sched_feat(HRTICK))
5338                 hrtick_clear(rq);
5339 
5340         spin_lock_irq(&rq->lock);
5341         update_rq_clock(rq);
5342         clear_tsk_need_resched(prev);
5343 
5344         if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5345                 if (unlikely(signal_pending_state(prev->state, prev)))
5346                         prev->state = TASK_RUNNING;
5347                 else
5348                         deactivate_task(rq, prev, 1);
5349                 switch_count = &prev->nvcsw;
5350         }
5351 
5352 #ifdef CONFIG_SMP
5353         if (prev->sched_class->pre_schedule)
5354                 prev->sched_class->pre_schedule(rq, prev);
5355 #endif
5356 
5357         if (unlikely(!rq->nr_running))
5358                 idle_balance(cpu, rq);
5359 
5360         put_prev_task(rq, prev);
5361         next = pick_next_task(rq);
5362 
5363         if (likely(prev != next)) {
5364                 sched_info_switch(prev, next);
5365                 perf_counter_task_sched_out(prev, next, cpu);
5366 
5367                 rq->nr_switches++;
5368                 rq->curr = next;
5369                 ++*switch_count;
5370 
5371                 context_switch(rq, prev, next); /* unlocks the rq */
5372                 /*
5373                  * the context switch might have flipped the stack from under
5374                  * us, hence refresh the local variables.
5375                  */
5376                 cpu = smp_processor_id();
5377                 rq = cpu_rq(cpu);
5378         } else
5379                 spin_unlock_irq(&rq->lock);
5380 
5381         if (unlikely(reacquire_kernel_lock(current) < 0))
5382                 goto need_resched_nonpreemptible;
5383 
5384         preempt_enable_no_resched();
5385         if (need_resched())
5386                 goto need_resched;
5387 }
5388 EXPORT_SYMBOL(schedule);
5389 
5390 #ifdef CONFIG_SMP
5391 /*
5392  * Look out! "owner" is an entirely speculative pointer
5393  * access and not reliable.
5394  */
5395 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5396 {
5397         unsigned int cpu;
5398         struct rq *rq;
5399 
5400         if (!sched_feat(OWNER_SPIN))
5401                 return 0;
5402 
5403 #ifdef CONFIG_DEBUG_PAGEALLOC
5404         /*
5405          * Need to access the cpu field knowing that
5406          * DEBUG_PAGEALLOC could have unmapped it if
5407          * the mutex owner just released it and exited.
5408          */
5409         if (probe_kernel_address(&owner->cpu, cpu))
5410                 goto out;
5411 #else
5412         cpu = owner->cpu;
5413 #endif
5414 
5415         /*
5416          * Even if the access succeeded (likely case),
5417          * the cpu field may no longer be valid.
5418          */
5419         if (cpu >= nr_cpumask_bits)
5420                 goto out;
5421 
5422         /*
5423          * We need to validate that we can do a
5424          * get_cpu() and that we have the percpu area.
5425          */
5426         if (!cpu_online(cpu))
5427                 goto out;
5428 
5429         rq = cpu_rq(cpu);
5430 
5431         for (;;) {
5432                 /*
5433                  * Owner changed, break to re-assess state.
5434                  */
5435                 if (lock->owner != owner)
5436                         break;
5437 
5438                 /*
5439                  * Is that owner really running on that cpu?
5440                  */
5441                 if (task_thread_info(rq->curr) != owner || need_resched())
5442                         return 0;
5443 
5444                 cpu_relax();
5445         }
5446 out:
5447         return 1;
5448 }
5449 #endif
5450 
5451 #ifdef CONFIG_PREEMPT
5452 /*
5453  * this is the entry point to schedule() from in-kernel preemption
5454  * off of preempt_enable. Kernel preemptions off return from interrupt
5455  * occur there and call schedule directly.
5456  */
5457 asmlinkage void __sched preempt_schedule(void)
5458 {
5459         struct thread_info *ti = current_thread_info();
5460 
5461         /*
5462          * If there is a non-zero preempt_count or interrupts are disabled,
5463          * we do not want to preempt the current task. Just return..
5464          */
5465         if (likely(ti->preempt_count || irqs_disabled()))
5466                 return;
5467 
5468         do {
5469                 add_preempt_count(PREEMPT_ACTIVE);
5470                 schedule();
5471                 sub_preempt_count(PREEMPT_ACTIVE);
5472 
5473                 /*
5474                  * Check again in case we missed a preemption opportunity
5475                  * between schedule and now.
5476                  */
5477                 barrier();
5478         } while (need_resched());
5479 }
5480 EXPORT_SYMBOL(preempt_schedule);
5481 
5482 /*
5483  * this is the entry point to schedule() from kernel preemption
5484  * off of irq context.
5485  * Note, that this is called and return with irqs disabled. This will
5486  * protect us against recursive calling from irq.
5487  */
5488 asmlinkage void __sched preempt_schedule_irq(void)
5489 {
5490         struct thread_info *ti = current_thread_info();
5491 
5492         /* Catch callers which need to be fixed */
5493         BUG_ON(ti->preempt_count || !irqs_disabled());
5494 
5495         do {
5496                 add_preempt_count(PREEMPT_ACTIVE);
5497                 local_irq_enable();
5498                 schedule();
5499                 local_irq_disable();
5500                 sub_preempt_count(PREEMPT_ACTIVE);
5501 
5502                 /*
5503                  * Check again in case we missed a preemption opportunity
5504                  * between schedule and now.
5505                  */
5506                 barrier();
5507         } while (need_resched());
5508 }
5509 
5510 #endif /* CONFIG_PREEMPT */
5511 
5512 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5513                           void *key)
5514 {
5515         return try_to_wake_up(curr->private, mode, sync);
5516 }
5517 EXPORT_SYMBOL(default_wake_function);
5518 
5519 /*
5520  * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5521  * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5522  * number) then we wake all the non-exclusive tasks and one exclusive task.
5523  *
5524  * There are circumstances in which we can try to wake a task which has already
5525  * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5526  * zero in this (rare) case, and we handle it by continuing to scan the queue.
5527  */
5528 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5529                         int nr_exclusive, int sync, void *key)
5530 {
5531         wait_queue_t *curr, *next;
5532 
5533         list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5534                 unsigned flags = curr->flags;
5535 
5536                 if (curr->func(curr, mode, sync, key) &&
5537                                 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5538                         break;
5539         }
5540 }
5541 
5542 /**
5543  * __wake_up - wake up threads blocked on a waitqueue.
5544  * @q: the waitqueue
5545  * @mode: which threads
5546  * @nr_exclusive: how many wake-one or wake-many threads to wake up
5547  * @key: is directly passed to the wakeup function
5548  *
5549  * It may be assumed that this function implies a write memory barrier before
5550  * changing the task state if and only if any tasks are woken up.
5551  */
5552 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5553                         int nr_exclusive, void *key)
5554 {
5555         unsigned long flags;
5556 
5557         spin_lock_irqsave(&q->lock, flags);
5558         __wake_up_common(q, mode, nr_exclusive, 0, key);
5559         spin_unlock_irqrestore(&q->lock, flags);
5560 }
5561 EXPORT_SYMBOL(__wake_up);
5562 
5563 /*
5564  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5565  */
5566 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5567 {
5568         __wake_up_common(q, mode, 1, 0, NULL);
5569 }
5570 
5571 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5572 {
5573         __wake_up_common(q, mode, 1, 0, key);
5574 }
5575 
5576 /**
5577  * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5578  * @q: the waitqueue
5579  * @mode: which threads
5580  * @nr_exclusive: how many wake-one or wake-many threads to wake up
5581  * @key: opaque value to be passed to wakeup targets
5582  *
5583  * The sync wakeup differs that the waker knows that it will schedule
5584  * away soon, so while the target thread will be woken up, it will not
5585  * be migrated to another CPU - ie. the two threads are 'synchronized'
5586  * with each other. This can prevent needless bouncing between CPUs.
5587  *
5588  * On UP it can prevent extra preemption.
5589  *
5590  * It may be assumed that this function implies a write memory barrier before
5591  * changing the task state if and only if any tasks are woken up.
5592  */
5593 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5594                         int nr_exclusive, void *key)
5595 {
5596         unsigned long flags;
5597         int sync = 1;
5598 
5599         if (unlikely(!q))
5600                 return;
5601 
5602         if (unlikely(!nr_exclusive))
5603                 sync = 0;
5604 
5605         spin_lock_irqsave(&q->lock, flags);
5606         __wake_up_common(q, mode, nr_exclusive, sync, key);
5607         spin_unlock_irqrestore(&q->lock, flags);
5608 }
5609 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5610 
5611 /*
5612  * __wake_up_sync - see __wake_up_sync_key()
5613  */
5614 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5615 {
5616         __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5617 }
5618 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
5619 
5620 /**
5621  * complete: - signals a single thread waiting on this completion
5622  * @x:  holds the state of this particular completion
5623  *
5624  * This will wake up a single thread waiting on this completion. Threads will be
5625  * awakened in the same order in which they were queued.
5626  *
5627  * See also complete_all(), wait_for_completion() and related routines.
5628  *
5629  * It may be assumed that this function implies a write memory barrier before
5630  * changing the task state if and only if any tasks are woken up.
5631  */
5632 void complete(struct completion *x)
5633 {
5634         unsigned long flags;
5635 
5636         spin_lock_irqsave(&x->wait.lock, flags);
5637         x->done++;
5638         __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5639         spin_unlock_irqrestore(&x->wait.lock, flags);
5640 }
5641 EXPORT_SYMBOL(complete);
5642 
5643 /**
5644  * complete_all: - signals all threads waiting on this completion
5645  * @x:  holds the state of this particular completion
5646  *
5647  * This will wake up all threads waiting on this particular completion event.
5648  *
5649  * It may be assumed that this function implies a write memory barrier before
5650  * changing the task state if and only if any tasks are woken up.
5651  */
5652 void complete_all(struct completion *x)
5653 {
5654         unsigned long flags;
5655 
5656         spin_lock_irqsave(&x->wait.lock, flags);
5657         x->done += UINT_MAX/2;
5658         __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5659         spin_unlock_irqrestore(&x->wait.lock, flags);
5660 }
5661 EXPORT_SYMBOL(complete_all);
5662 
5663 static inline long __sched
5664 do_wait_for_common(struct completion *x, long timeout, int state)
5665 {
5666         if (!x->done) {
5667                 DECLARE_WAITQUEUE(wait, current);
5668 
5669                 wait.flags |= WQ_FLAG_EXCLUSIVE;
5670                 __add_wait_queue_tail(&x->wait, &wait);
5671                 do {
5672                         if (signal_pending_state(state, current)) {
5673                                 timeout = -ERESTARTSYS;
5674                                 break;
5675                         }
5676                         __set_current_state(state);
5677                         spin_unlock_irq(&x->wait.lock);
5678                         timeout = schedule_timeout(timeout);
5679                         spin_lock_irq(&x->wait.lock);
5680                 } while (!x->done && timeout);
5681                 __remove_wait_queue(&x->wait, &wait);
5682                 if (!x->done)
5683                         return timeout;
5684         }
5685         x->done--;
5686         return timeout ?: 1;
5687 }
5688 
5689 static long __sched
5690 wait_for_common(struct completion *x, long timeout, int state)
5691 {
5692         might_sleep();
5693 
5694         spin_lock_irq(&x->wait.lock);
5695         timeout = do_wait_for_common(x, timeout, state);
5696         spin_unlock_irq(&x->wait.lock);
5697         return timeout;
5698 }
5699 
5700 /**
5701  * wait_for_completion: - waits for completion of a task
5702  * @x:  holds the state of this particular completion
5703  *
5704  * This waits to be signaled for completion of a specific task. It is NOT
5705  * interruptible and there is no timeout.
5706  *
5707  * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5708  * and interrupt capability. Also see complete().
5709  */
5710 void __sched wait_for_completion(struct completion *x)
5711 {
5712         wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5713 }
5714 EXPORT_SYMBOL(wait_for_completion);
5715 
5716 /**
5717  * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5718  * @x:  holds the state of this particular completion
5719  * @timeout:  timeout value in jiffies
5720  *
5721  * This waits for either a completion of a specific task to be signaled or for a
5722  * specified timeout to expire. The timeout is in jiffies. It is not
5723  * interruptible.
5724  */
5725 unsigned long __sched
5726 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5727 {
5728         return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5729 }
5730 EXPORT_SYMBOL(wait_for_completion_timeout);
5731 
5732 /**
5733  * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5734  * @x:  holds the state of this particular completion
5735  *
5736  * This waits for completion of a specific task to be signaled. It is
5737  * interruptible.
5738  */
5739 int __sched wait_for_completion_interruptible(struct completion *x)
5740 {
5741         long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5742         if (t == -ERESTARTSYS)
5743                 return t;
5744         return 0;
5745 }
5746 EXPORT_SYMBOL(wait_for_completion_interruptible);
5747 
5748 /**
5749  * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5750  * @x:  holds the state of this particular completion
5751  * @timeout:  timeout value in jiffies
5752  *
5753  * This waits for either a completion of a specific task to be signaled or for a
5754  * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5755  */
5756 unsigned long __sched
5757 wait_for_completion_interruptible_timeout(struct completion *x,
5758                                           unsigned long timeout)
5759 {
5760         return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5761 }
5762 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5763 
5764 /**
5765  * wait_for_completion_killable: - waits for completion of a task (killable)
5766  * @x:  holds the state of this particular completion
5767  *
5768  * This waits to be signaled for completion of a specific task. It can be
5769  * interrupted by a kill signal.
5770  */
5771 int __sched wait_for_completion_killable(struct completion *x)
5772 {
5773         long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5774         if (t == -ERESTARTSYS)
5775                 return t;
5776         return 0;
5777 }
5778 EXPORT_SYMBOL(wait_for_completion_killable);
5779 
5780 /**
5781  *      try_wait_for_completion - try to decrement a completion without blocking
5782  *      @x:     completion structure
5783  *
5784  *      Returns: 0 if a decrement cannot be done without blocking
5785  *               1 if a decrement succeeded.
5786  *
5787  *      If a completion is being used as a counting completion,
5788  *      attempt to decrement the counter without blocking. This
5789  *      enables us to avoid waiting if the resource the completion
5790  *      is protecting is not available.
5791  */
5792 bool try_wait_for_completion(struct completion *x)
5793 {
5794         int ret = 1;
5795 
5796         spin_lock_irq(&x->wait.lock);
5797         if (!x->done)
5798                 ret = 0;
5799         else
5800                 x->done--;
5801         spin_unlock_irq(&x->wait.lock);
5802         return ret;
5803 }
5804 EXPORT_SYMBOL(try_wait_for_completion);
5805 
5806 /**
5807  *      completion_done - Test to see if a completion has any waiters
5808  *      @x:     completion structure
5809  *
5810  *      Returns: 0 if there are waiters (wait_for_completion() in progress)
5811  *               1 if there are no waiters.
5812  *
5813  */
5814 bool completion_done(struct completion *x)
5815 {
5816         int ret = 1;
5817 
5818         spin_lock_irq(&x->wait.lock);
5819         if (!x->done)
5820                 ret = 0;
5821         spin_unlock_irq(&x->wait.lock);
5822         return ret;
5823 }
5824 EXPORT_SYMBOL(completion_done);
5825 
5826 static long __sched
5827 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5828 {
5829         unsigned long flags;
5830         wait_queue_t wait;
5831 
5832         init_waitqueue_entry(&wait, current);
5833 
5834         __set_current_state(state);
5835 
5836         spin_lock_irqsave(&q->lock, flags);
5837         __add_wait_queue(q, &wait);
5838         spin_unlock(&q->lock);
5839         timeout = schedule_timeout(timeout);
5840         spin_lock_irq(&q->lock);
5841         __remove_wait_queue(q, &wait);
5842         spin_unlock_irqrestore(&q->lock, flags);
5843 
5844         return timeout;
5845 }
5846 
5847 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5848 {
5849         sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5850 }
5851 EXPORT_SYMBOL(interruptible_sleep_on);
5852 
5853 long __sched
5854 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5855 {
5856         return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5857 }
5858 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5859 
5860 void __sched sleep_on(wait_queue_head_t *q)
5861 {
5862         sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5863 }
5864 EXPORT_SYMBOL(sleep_on);
5865 
5866 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5867 {
5868         return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5869 }
5870 EXPORT_SYMBOL(sleep_on_timeout);
5871 
5872 #ifdef CONFIG_RT_MUTEXES
5873 
5874 /*
5875  * rt_mutex_setprio - set the current priority of a task
5876  * @p: task
5877  * @prio: prio value (kernel-internal form)
5878  *
5879  * This function changes the 'effective' priority of a task. It does
5880  * not touch ->normal_prio like __setscheduler().
5881  *
5882  * Used by the rt_mutex code to implement priority inheritance logic.
5883  */
5884 void rt_mutex_setprio(struct task_struct *p, int prio)
5885 {
5886         unsigned long flags;
5887         int oldprio, on_rq, running;
5888         struct rq *rq;
5889         const struct sched_class *prev_class = p->sched_class;
5890 
5891         BUG_ON(prio < 0 || prio > MAX_PRIO);
5892 
5893         rq = task_rq_lock(p, &flags);
5894         update_rq_clock(rq);
5895 
5896         oldprio = p->prio;
5897         on_rq = p->se.on_rq;
5898         running = task_current(rq, p);
5899         if (on_rq)
5900                 dequeue_task(rq, p, 0);
5901         if (running)
5902                 p->sched_class->put_prev_task(rq, p);
5903 
5904         if (rt_prio(prio))
5905                 p->sched_class = &rt_sched_class;
5906         else
5907                 p->sched_class = &fair_sched_class;
5908 
5909         p->prio = prio;
5910 
5911         if (running)
5912                 p->sched_class->set_curr_task(rq);
5913         if (on_rq) {
5914                 enqueue_task(rq, p, 0);
5915 
5916                 check_class_changed(rq, p, prev_class, oldprio, running);
5917         }
5918         task_rq_unlock(rq, &flags);
5919 }
5920 
5921 #endif
5922 
5923 void set_user_nice(struct task_struct *p, long nice)
5924 {
5925         int old_prio, delta, on_rq;
5926         unsigned long flags;
5927         struct rq *rq;
5928 
5929         if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5930                 return;
5931         /*
5932          * We have to be careful, if called from sys_setpriority(),
5933          * the task might be in the middle of scheduling on another CPU.
5934          */
5935         rq = task_rq_lock(p, &flags);
5936         update_rq_clock(rq);
5937         /*
5938          * The RT priorities are set via sched_setscheduler(), but we still
5939          * allow the 'normal' nice value to be set - but as expected
5940          * it wont have any effect on scheduling until the task is
5941          * SCHED_FIFO/SCHED_RR:
5942          */
5943         if (task_has_rt_policy(p)) {
5944                 p->static_prio = NICE_TO_PRIO(nice);
5945                 goto out_unlock;
5946         }
5947         on_rq = p->se.on_rq;
5948         if (on_rq)
5949                 dequeue_task(rq, p, 0);
5950 
5951         p->static_prio = NICE_TO_PRIO(nice);
5952         set_load_weight(p);
5953         old_prio = p->prio;
5954         p->prio = effective_prio(p);
5955         delta = p->prio - old_prio;
5956 
5957         if (on_rq) {
5958                 enqueue_task(rq, p, 0);
5959                 /*
5960                  * If the task increased its priority or is running and
5961                  * lowered its priority, then reschedule its CPU:
5962                  */
5963                 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5964                         resched_task(rq->curr);
5965         }
5966 out_unlock:
5967         task_rq_unlock(rq, &flags);
5968 }
5969 EXPORT_SYMBOL(set_user_nice);
5970 
5971 /*
5972  * can_nice - check if a task can reduce its nice value
5973  * @p: task
5974  * @nice: nice value
5975  */
5976 int can_nice(const struct task_struct *p, const int nice)
5977 {
5978         /* convert nice value [19,-20] to rlimit style value [1,40] */
5979         int nice_rlim = 20 - nice;
5980 
5981         return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5982                 capable(CAP_SYS_NICE));
5983 }
5984 
5985 #ifdef __ARCH_WANT_SYS_NICE
5986 
5987 /*
5988  * sys_nice - change the priority of the current process.
5989  * @increment: priority increment
5990  *
5991  * sys_setpriority is a more generic, but much slower function that
5992  * does similar things.
5993  */
5994 SYSCALL_DEFINE1(nice, int, increment)
5995 {
5996         long nice, retval;
5997 
5998         /*
5999          * Setpriority might change our priority at the same moment.
6000          * We don't have to worry. Conceptually one call occurs first
6001          * and we have a single winner.
6002          */
6003         if (increment < -40)
6004                 increment = -40;
6005         if (increment > 40)
6006                 increment = 40;
6007 
6008         nice = TASK_NICE(current) + increment;
6009         if (nice < -20)
6010                 nice = -20;
6011         if (nice > 19)
6012                 nice = 19;
6013 
6014         if (increment < 0 && !can_nice(current, nice))
6015                 return -EPERM;
6016 
6017         retval = security_task_setnice(current, nice);
6018         if (retval)
6019                 return retval;
6020 
6021         set_user_nice(current, nice);
6022         return 0;
6023 }
6024 
6025 #endif
6026 
6027 /**
6028  * task_prio - return the priority value of a given task.
6029  * @p: the task in question.
6030  *
6031  * This is the priority value as seen by users in /proc.
6032  * RT tasks are offset by -200. Normal tasks are centered
6033  * around 0, value goes from -16 to +15.
6034  */
6035 int task_prio(const struct task_struct *p)
6036 {
6037         return p->prio - MAX_RT_PRIO;
6038 }
6039 
6040 /**
6041  * task_nice - return the nice value of a given task.
6042  * @p: the task in question.
6043  */
6044 int task_nice(const struct task_struct *p)
6045 {
6046         return TASK_NICE(p);
6047 }
6048 EXPORT_SYMBOL(task_nice);
6049 
6050 /**
6051  * idle_cpu - is a given cpu idle currently?
6052  * @cpu: the processor in question.
6053  */
6054 int idle_cpu(int cpu)
6055 {
6056         return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6057 }
6058 
6059 /**
6060  * idle_task - return the idle task for a given cpu.
6061  * @cpu: the processor in question.
6062  */
6063 struct task_struct *idle_task(int cpu)
6064 {
6065         return cpu_rq(cpu)->idle;
6066 }
6067 
6068 /**
6069  * find_process_by_pid - find a process with a matching PID value.
6070  * @pid: the pid in question.
6071  */
6072 static struct task_struct *find_process_by_pid(pid_t pid)
6073 {
6074         return pid ? find_task_by_vpid(pid) : current;
6075 }
6076 
6077 /* Actually do priority change: must hold rq lock. */
6078 static void
6079 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6080 {
6081         BUG_ON(p->se.on_rq);
6082 
6083         p->policy = policy;
6084         switch (p->policy) {
6085         case SCHED_NORMAL:
6086         case SCHED_BATCH:
6087         case SCHED_IDLE:
6088                 p->sched_class = &fair_sched_class;
6089                 break;
6090         case SCHED_FIFO:
6091         case SCHED_RR:
6092                 p->sched_class = &rt_sched_class;
6093                 break;
6094         }
6095 
6096         p->rt_priority = prio;
6097         p->normal_prio = normal_prio(p);
6098         /* we are holding p->pi_lock already */
6099         p->prio = rt_mutex_getprio(p);
6100         set_load_weight(p);
6101 }
6102 
6103 /*
6104  * check the target process has a UID that matches the current process's
6105  */
6106 static bool check_same_owner(struct task_struct *p)
6107 {
6108         const struct cred *cred = current_cred(), *pcred;
6109         bool match;
6110 
6111         rcu_read_lock();
6112         pcred = __task_cred(p);
6113         match = (cred->euid == pcred->euid ||
6114                  cred->euid == pcred->uid);
6115         rcu_read_unlock();
6116         return match;
6117 }
6118 
6119 static int __sched_setscheduler(struct task_struct *p, int policy,
6120                                 struct sched_param *param, bool user)
6121 {
6122         int retval, oldprio, oldpolicy = -1, on_rq, running;
6123         unsigned long flags;
6124         const struct sched_class *prev_class = p->sched_class;
6125         struct rq *rq;
6126 
6127         /* may grab non-irq protected spin_locks */
6128         BUG_ON(in_interrupt());
6129 recheck:
6130         /* double check policy once rq lock held */
6131         if (policy < 0)
6132                 policy = oldpolicy = p->policy;
6133         else if (policy != SCHED_FIFO && policy != SCHED_RR &&
6134                         policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6135                         policy != SCHED_IDLE)
6136                 return -EINVAL;
6137         /*
6138          * Valid priorities for SCHED_FIFO and SCHED_RR are
6139          * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6140          * SCHED_BATCH and SCHED_IDLE is 0.
6141          */
6142         if (param->sched_priority < 0 ||
6143             (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6144             (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6145                 return -EINVAL;
6146         if (rt_policy(policy) != (param->sched_priority != 0))
6147                 return -EINVAL;
6148 
6149         /*
6150          * Allow unprivileged RT tasks to decrease priority:
6151          */
6152         if (user && !capable(CAP_SYS_NICE)) {
6153                 if (rt_policy(policy)) {
6154                         unsigned long rlim_rtprio;
6155 
6156                         if (!lock_task_sighand(p, &flags))
6157                                 return -ESRCH;
6158                         rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6159                         unlock_task_sighand(p, &flags);
6160 
6161                         /* can't set/change the rt policy */
6162                         if (policy != p->policy && !rlim_rtprio)
6163                                 return -EPERM;
6164 
6165                         /* can't increase priority */
6166                         if (param->sched_priority > p->rt_priority &&
6167                             param->sched_priority > rlim_rtprio)
6168                                 return -EPERM;
6169                 }
6170                 /*
6171                  * Like positive nice levels, dont allow tasks to
6172                  * move out of SCHED_IDLE either:
6173                  */
6174                 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6175                         return -EPERM;
6176 
6177                 /* can't change other user's priorities */
6178                 if (!check_same_owner(p))
6179                         return -EPERM;
6180         }
6181 
6182         if (user) {
6183 #ifdef CONFIG_RT_GROUP_SCHED
6184                 /*
6185                  * Do not allow realtime tasks into groups that have no runtime
6186                  * assigned.
6187                  */
6188                 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6189                                 task_group(p)->rt_bandwidth.rt_runtime == 0)
6190                         return -EPERM;
6191 #endif
6192 
6193                 retval = security_task_setscheduler(p, policy, param);
6194                 if (retval)
6195                         return retval;
6196         }
6197 
6198         /*
6199          * make sure no PI-waiters arrive (or leave) while we are
6200          * changing the priority of the task:
6201          */
6202         spin_lock_irqsave(&p->pi_lock, flags);
6203         /*
6204          * To be able to change p->policy safely, the apropriate
6205          * runqueue lock must be held.
6206          */
6207         rq = __task_rq_lock(p);
6208         /* recheck policy now with rq lock held */
6209         if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6210                 policy = oldpolicy = -1;
6211                 __task_rq_unlock(rq);
6212                 spin_unlock_irqrestore(&p->pi_lock, flags);
6213                 goto recheck;
6214         }
6215         update_rq_clock(rq);
6216         on_rq = p->se.on_rq;
6217         running = task_current(rq, p);
6218         if (on_rq)
6219                 deactivate_task(rq, p, 0);
6220         if (running)
6221                 p->sched_class->put_prev_task(rq, p);
6222 
6223         oldprio = p->prio;
6224         __setscheduler(rq, p, policy, param->sched_priority);
6225 
6226         if (running)
6227                 p->sched_class->set_curr_task(rq);
6228         if (on_rq) {
6229                 activate_task(rq, p, 0);
6230 
6231                 check_class_changed(rq, p, prev_class, oldprio, running);
6232         }
6233         __task_rq_unlock(rq);
6234         spin_unlock_irqrestore(&p->pi_lock, flags);
6235 
6236         rt_mutex_adjust_pi(p);
6237 
6238         return 0;
6239 }
6240 
6241 /**
6242  * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6243  * @p: the task in question.
6244  * @policy: new policy.
6245  * @param: structure containing the new RT priority.
6246  *
6247  * NOTE that the task may be already dead.
6248  */
6249 int sched_setscheduler(struct task_struct *p, int policy,
6250                        struct sched_param *param)
6251 {
6252         return __sched_setscheduler(p, policy, param, true);
6253 }
6254 EXPORT_SYMBOL_GPL(sched_setscheduler);
6255 
6256 /**
6257  * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6258  * @p: the task in question.
6259  * @policy: new policy.
6260  * @param: structure containing the new RT priority.
6261  *
6262  * Just like sched_setscheduler, only don't bother checking if the
6263  * current context has permission.  For example, this is needed in
6264  * stop_machine(): we create temporary high priority worker threads,
6265  * but our caller might not have that capability.
6266  */
6267 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6268                                struct sched_param *param)
6269 {
6270         return __sched_setscheduler(p, policy, param, false);
6271 }
6272 
6273 static int
6274 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6275 {
6276         struct sched_param lparam;
6277         struct task_struct *p;
6278         int retval;
6279 
6280         if (!param || pid < 0)
6281                 return -EINVAL;
6282         if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6283                 return -EFAULT;
6284 
6285         rcu_read_lock();
6286         retval = -ESRCH;
6287         p = find_process_by_pid(pid);
6288         if (p != NULL)
6289                 retval = sched_setscheduler(p, policy, &lparam);
6290         rcu_read_unlock();
6291 
6292         return retval;
6293 }
6294 
6295 /**
6296  * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6297  * @pid: the pid in question.
6298  * @policy: new policy.
6299  * @param: structure containing the new RT priority.
6300  */
6301 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6302                 struct sched_param __user *, param)
6303 {
6304         /* negative values for policy are not valid */
6305         if (policy < 0)
6306                 return -EINVAL;
6307 
6308         return do_sched_setscheduler(pid, policy, param);
6309 }
6310 
6311 /**
6312  * sys_sched_setparam - set/change the RT priority of a thread
6313  * @pid: the pid in question.
6314  * @param: structure containing the new RT priority.
6315  */
6316 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6317 {
6318         return do_sched_setscheduler(pid, -1, param);
6319 }
6320 
6321 /**
6322  * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6323  * @pid: the pid in question.
6324  */
6325 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6326 {
6327         struct task_struct *p;
6328         int retval;
6329 
6330         if (pid < 0)
6331                 return -EINVAL;
6332 
6333         retval = -ESRCH;
6334         read_lock(&tasklist_lock);
6335         p = find_process_by_pid(pid);
6336         if (p) {
6337                 retval = security_task_getscheduler(p);
6338                 if (!retval)
6339                         retval = p->policy;
6340         }
6341         read_unlock(&tasklist_lock);
6342         return retval;
6343 }
6344 
6345 /**
6346  * sys_sched_getscheduler - get the RT priority of a thread
6347  * @pid: the pid in question.
6348  * @param: structure containing the RT priority.
6349  */
6350 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6351 {
6352         struct sched_param lp;
6353         struct task_struct *p;
6354         int retval;
6355 
6356         if (!param || pid < 0)
6357                 return -EINVAL;
6358 
6359         read_lock(&tasklist_lock);
6360         p = find_process_by_pid(pid);
6361         retval = -ESRCH;
6362         if (!p)
6363                 goto out_unlock;
6364 
6365         retval = security_task_getscheduler(p);
6366         if (retval)
6367                 goto out_unlock;
6368 
6369         lp.sched_priority = p->rt_priority;
6370         read_unlock(&tasklist_lock);
6371 
6372         /*
6373          * This one might sleep, we cannot do it with a spinlock held ...
6374          */
6375         retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6376 
6377         return retval;
6378 
6379 out_unlock:
6380         read_unlock(&tasklist_lock);
6381         return retval;
6382 }
6383 
6384 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6385 {
6386         cpumask_var_t cpus_allowed, new_mask;
6387         struct task_struct *p;
6388         int retval;
6389 
6390         get_online_cpus();
6391         read_lock(&tasklist_lock);
6392 
6393         p = find_process_by_pid(pid);
6394         if (!p) {
6395                 read_unlock(&tasklist_lock);
6396                 put_online_cpus();
6397                 return -ESRCH;
6398         }
6399 
6400         /*
6401          * It is not safe to call set_cpus_allowed with the
6402          * tasklist_lock held. We will bump the task_struct's
6403          * usage count and then drop tasklist_lock.
6404          */
6405         get_task_struct(p);
6406         read_unlock(&tasklist_lock);
6407 
6408         if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6409                 retval = -ENOMEM;
6410                 goto out_put_task;
6411         }
6412         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6413                 retval = -ENOMEM;
6414                 goto out_free_cpus_allowed;
6415         }
6416         retval = -EPERM;
6417         if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6418                 goto out_unlock;
6419 
6420         retval = security_task_setscheduler(p, 0, NULL);
6421         if (retval)
6422                 goto out_unlock;
6423 
6424         cpuset_cpus_allowed(p, cpus_allowed);
6425         cpumask_and(new_mask, in_mask, cpus_allowed);
6426  again:
6427         retval = set_cpus_allowed_ptr(p, new_mask);
6428 
6429         if (!retval) {
6430                 cpuset_cpus_allowed(p, cpus_allowed);
6431                 if (!cpumask_subset(new_mask, cpus_allowed)) {
6432                         /*
6433                          * We must have raced with a concurrent cpuset
6434                          * update. Just reset the cpus_allowed to the
6435                          * cpuset's cpus_allowed
6436                          */
6437                         cpumask_copy(new_mask, cpus_allowed);
6438                         goto again;
6439                 }
6440         }
6441 out_unlock:
6442         free_cpumask_var(new_mask);
6443 out_free_cpus_allowed:
6444         free_cpumask_var(cpus_allowed);
6445 out_put_task:
6446         put_task_struct(p);
6447         put_online_cpus();
6448         return retval;
6449 }
6450 
6451 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6452                              struct cpumask *new_mask)
6453 {
6454         if (len < cpumask_size())
6455                 cpumask_clear(new_mask);
6456         else if (len > cpumask_size())
6457                 len = cpumask_size();
6458 
6459         return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6460 }
6461 
6462 /**
6463  * sys_sched_setaffinity - set the cpu affinity of a process
6464  * @pid: pid of the process
6465  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6466  * @user_mask_ptr: user-space pointer to the new cpu mask
6467  */
6468 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6469                 unsigned long __user *, user_mask_ptr)
6470 {
6471         cpumask_var_t new_mask;
6472         int retval;
6473 
6474         if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6475                 return -ENOMEM;
6476 
6477         retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6478         if (retval == 0)
6479                 retval = sched_setaffinity(pid, new_mask);
6480         free_cpumask_var(new_mask);
6481         return retval;
6482 }
6483 
6484 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6485 {
6486         struct task_struct *p;
6487         int retval;
6488 
6489         get_online_cpus();
6490         read_lock(&tasklist_lock);
6491 
6492         retval = -ESRCH;
6493         p = find_process_by_pid(pid);
6494         if (!p)
6495                 goto out_unlock;
6496 
6497         retval = security_task_getscheduler(p);
6498         if (retval)
6499                 goto out_unlock;
6500 
6501         cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6502 
6503 out_unlock:
6504         read_unlock(&tasklist_lock);
6505         put_online_cpus();
6506 
6507         return retval;
6508 }
6509 
6510 /**
6511  * sys_sched_getaffinity - get the cpu affinity of a process
6512  * @pid: pid of the process
6513  * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6514  * @user_mask_ptr: user-space pointer to hold the current cpu mask
6515  */
6516 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6517                 unsigned long __user *, user_mask_ptr)
6518 {
6519         int ret;
6520         cpumask_var_t mask;
6521 
6522         if (len < cpumask_size())
6523                 return -EINVAL;
6524 
6525         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6526                 return -ENOMEM;
6527 
6528         ret = sched_getaffinity(pid, mask);
6529         if (ret == 0) {
6530                 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6531                         ret = -EFAULT;
6532                 else
6533                         ret = cpumask_size();
6534         }
6535         free_cpumask_var(mask);
6536 
6537         return ret;
6538 }
6539 
6540 /**
6541  * sys_sched_yield - yield the current processor to other threads.
6542  *
6543  * This function yields the current CPU to other tasks. If there are no
6544  * other threads running on this CPU then this function will return.
6545  */
6546 SYSCALL_DEFINE0(sched_yield)
6547 {
6548         struct rq *rq = this_rq_lock();
6549 
6550         schedstat_inc(rq, yld_count);
6551         current->sched_class->yield_task(rq);
6552 
6553         /*
6554          * Since we are going to call schedule() anyway, there's
6555          * no need to preempt or enable interrupts:
6556          */
6557         __release(rq->lock);
6558         spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6559         _raw_spin_unlock(&rq->lock);
6560         preempt_enable_no_resched();
6561 
6562         schedule();
6563 
6564         return 0;
6565 }
6566 
6567 static inline int should_resched(void)
6568 {
6569         return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6570 }
6571 
6572 static void __cond_resched(void)
6573 {
6574 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6575         __might_sleep(__FILE__, __LINE__);
6576 #endif
6577         /*
6578          * The BKS might be reacquired before we have dropped
6579          * PREEMPT_ACTIVE, which could trigger a second
6580          * cond_resched() call.
6581          */
6582         do {
6583                 add_preempt_count(PREEMPT_ACTIVE);
6584                 schedule();
6585                 sub_preempt_count(PREEMPT_ACTIVE);
6586         } while (need_resched());
6587 }
6588 
6589 int __sched _cond_resched(void)
6590 {
6591         if (should_resched()) {
6592                 __cond_resched();
6593                 return 1;
6594         }
6595         return 0;
6596 }
6597 EXPORT_SYMBOL(_cond_resched);
6598 
6599 /*
6600  * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6601  * call schedule, and on return reacquire the lock.
6602  *
6603  * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6604  * operations here to prevent schedule() from being called twice (once via
6605  * spin_unlock(), once by hand).
6606  */
6607 int cond_resched_lock(spinlock_t *lock)
6608 {
6609         int resched = should_resched();
6610         int ret = 0;
6611 
6612         if (spin_needbreak(lock) || resched) {
6613                 spin_unlock(lock);
6614                 if (resched)
6615                         __cond_resched();
6616                 else
6617                         cpu_relax();
6618                 ret = 1;
6619                 spin_lock(lock);
6620         }
6621         return ret;
6622 }
6623 EXPORT_SYMBOL(cond_resched_lock);
6624 
6625 int __sched cond_resched_softirq(void)
6626 {
6627         BUG_ON(!in_softirq());
6628 
6629         if (should_resched()) {
6630                 local_bh_enable();
6631                 __cond_resched();
6632                 local_bh_disable();
6633                 return 1;
6634         }
6635         return 0;
6636 }
6637 EXPORT_SYMBOL(cond_resched_softirq);
6638 
6639 /**
6640  * yield - yield the current processor to other threads.
6641  *
6642  * This is a shortcut for kernel-space yielding - it marks the
6643  * thread runnable and calls sys_sched_yield().
6644  */
6645 void __sched yield(void)
6646 {
6647         set_current_state(TASK_RUNNING);
6648         sys_sched_yield();
6649 }
6650 EXPORT_SYMBOL(yield);
6651 
6652 /*
6653  * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6654  * that process accounting knows that this is a task in IO wait state.
6655  *
6656  * But don't do that if it is a deliberate, throttling IO wait (this task
6657  * has set its backing_dev_info: the queue against which it should throttle)
6658  */
6659 void __sched io_schedule(void)
6660 {
6661         struct rq *rq = &__raw_get_cpu_var(runqueues);
6662 
6663         delayacct_blkio_start();
6664         atomic_inc(&rq->nr_iowait);
6665         schedule();
6666         atomic_dec(&rq->nr_iowait);
6667         delayacct_blkio_end();
6668 }
6669 EXPORT_SYMBOL(io_schedule);
6670 
6671 long __sched io_schedule_timeout(long timeout)
6672 {
6673         struct rq *rq = &__raw_get_cpu_var(runqueues);
6674         long ret;
6675 
6676         delayacct_blkio_start();
6677         atomic_inc(&rq->nr_iowait);
6678         ret = schedule_timeout(timeout);
6679         atomic_dec(&rq->nr_iowait);
6680         delayacct_blkio_end();
6681         return ret;
6682 }
6683 
6684 /**
6685  * sys_sched_get_priority_max - return maximum RT priority.
6686  * @policy: scheduling class.
6687  *
6688  * this syscall returns the maximum rt_priority that can be used
6689  * by a given scheduling class.
6690  */
6691 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6692 {
6693         int ret = -EINVAL;
6694 
6695         switch (policy) {
6696         case SCHED_FIFO:
6697         case SCHED_RR:
6698                 ret = MAX_USER_RT_PRIO-1;
6699                 break;
6700         case SCHED_NORMAL:
6701         case SCHED_BATCH:
6702         case SCHED_IDLE:
6703                 ret = 0;
6704                 break;
6705         }
6706         return ret;
6707 }
6708 
6709 /**
6710  * sys_sched_get_priority_min - return minimum RT priority.
6711  * @policy: scheduling class.
6712  *
6713  * this syscall returns the minimum rt_priority that can be used
6714  * by a given scheduling class.
6715  */
6716 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6717 {
6718         int ret = -EINVAL;
6719 
6720         switch (policy) {
6721         case SCHED_FIFO:
6722         case SCHED_RR:
6723                 ret = 1;
6724                 break;
6725         case SCHED_NORMAL:
6726         case SCHED_BATCH:
6727         case SCHED_IDLE:
6728                 ret = 0;
6729         }
6730         return ret;
6731 }
6732 
6733 /**
6734  * sys_sched_rr_get_interval - return the default timeslice of a process.
6735  * @pid: pid of the process.
6736  * @interval: userspace pointer to the timeslice value.
6737  *
6738  * this syscall writes the default timeslice value of a given process
6739  * into the user-space timespec buffer. A value of '' means infinity.
6740  */
6741 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6742                 struct timespec __user *, interval)
6743 {
6744         struct task_struct *p;
6745         unsigned int time_slice;
6746         int retval;
6747         struct timespec t;
6748 
6749         if (pid < 0)
6750                 return -EINVAL;
6751 
6752         retval = -ESRCH;
6753         read_lock(&tasklist_lock);
6754         p = find_process_by_pid(pid);
6755         if (!p)
6756                 goto out_unlock;
6757 
6758         retval = security_task_getscheduler(p);
6759         if (retval)
6760                 goto out_unlock;
6761 
6762         /*
6763          * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6764          * tasks that are on an otherwise idle runqueue:
6765          */
6766         time_slice = 0;
6767         if (p->policy == SCHED_RR) {
6768                 time_slice = DEF_TIMESLICE;
6769         } else if (p->policy != SCHED_FIFO) {
6770                 struct sched_entity *se = &p->se;
6771                 unsigned long flags;
6772                 struct rq *rq;
6773 
6774                 rq = task_rq_lock(p, &flags);
6775                 if (rq->cfs.load.weight)
6776                         time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6777                 task_rq_unlock(rq, &flags);
6778         }
6779         read_unlock(&tasklist_lock);
6780         jiffies_to_timespec(time_slice, &t);
6781         retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6782         return retval;
6783 
6784 out_unlock:
6785         read_unlock(&tasklist_lock);
6786         return retval;
6787 }
6788 
6789 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6790 
6791 void sched_show_task(struct task_struct *p)
6792 {
6793         unsigned long free = 0;
6794         unsigned state;
6795 
6796         state = p->state ? __ffs(p->state) + 1 : 0;
6797         printk(KERN_INFO "%-13.13s %c", p->comm,
6798                 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6799 #if BITS_PER_LONG == 32
6800         if (state == TASK_RUNNING)
6801                 printk(KERN_CONT " running  ");
6802         else
6803                 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6804 #else
6805         if (state == TASK_RUNNING)
6806                 printk(KERN_CONT "  running task    ");
6807         else
6808                 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6809 #endif
6810 #ifdef CONFIG_DEBUG_STACK_USAGE
6811         free = stack_not_used(p);
6812 #endif
6813         printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6814                 task_pid_nr(p), task_pid_nr(p->real_parent),
6815                 (unsigned long)task_thread_info(p)->flags);
6816 
6817         show_stack(p, NULL);
6818 }
6819 
6820 void show_state_filter(unsigned long state_filter)
6821 {
6822         struct task_struct *g, *p;
6823 
6824 #if BITS_PER_LONG == 32
6825         printk(KERN_INFO
6826                 "  task                PC stack   pid father\n");
6827 #else
6828         printk(KERN_INFO
6829                 "  task                        PC stack   pid father\n");
6830 #endif
6831         read_lock(&tasklist_lock);
6832         do_each_thread(g, p) {
6833                 /*
6834                  * reset the NMI-timeout, listing all files on a slow
6835                  * console might take alot of time:
6836                  */
6837                 touch_nmi_watchdog();
6838                 if (!state_filter || (p->state & state_filter))
6839                         sched_show_task(p);
6840         } while_each_thread(g, p);
6841 
6842         touch_all_softlockup_watchdogs();
6843 
6844 #ifdef CONFIG_SCHED_DEBUG
6845         sysrq_sched_debug_show();
6846 #endif
6847         read_unlock(&tasklist_lock);
6848         /*
6849          * Only show locks if all tasks are dumped:
6850          */
6851         if (state_filter == -1)
6852                 debug_show_all_locks();
6853 }
6854 
6855 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6856 {
6857         idle->sched_class = &idle_sched_class;
6858 }
6859 
6860 /**
6861  * init_idle - set up an idle thread for a given CPU
6862  * @idle: task in question
6863  * @cpu: cpu the idle task belongs to
6864  *
6865  * NOTE: this function does not set the idle thread's NEED_RESCHED
6866  * flag, to make booting more robust.
6867  */
6868 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6869 {
6870         struct rq *rq = cpu_rq(cpu);
6871         unsigned long flags;
6872 
6873         spin_lock_irqsave(&rq->lock, flags);
6874 
6875         __sched_fork(idle);
6876         idle->se.exec_start = sched_clock();
6877 
6878         idle->prio = idle->normal_prio = MAX_PRIO;
6879         cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6880         __set_task_cpu(idle, cpu);
6881 
6882         rq->curr = rq->idle = idle;
6883 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6884         idle->oncpu = 1;
6885 #endif
6886         spin_unlock_irqrestore(&rq->lock, flags);
6887 
6888         /* Set the preempt count _outside_ the spinlocks! */
6889 #if defined(CONFIG_PREEMPT)
6890         task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6891 #else
6892         task_thread_info(idle)->preempt_count = 0;
6893 #endif
6894         /*
6895          * The idle tasks have their own, simple scheduling class:
6896          */
6897         idle->sched_class = &idle_sched_class;
6898         ftrace_graph_init_task(idle);
6899 }
6900 
6901 /*
6902  * In a system that switches off the HZ timer nohz_cpu_mask
6903  * indicates which cpus entered this state. This is used
6904  * in the rcu update to wait only for active cpus. For system
6905  * which do not switch off the HZ timer nohz_cpu_mask should
6906  * always be CPU_BITS_NONE.
6907  */
6908 cpumask_var_t nohz_cpu_mask;
6909 
6910 /*
6911  * Increase the granularity value when there are more CPUs,
6912  * because with more CPUs the 'effective latency' as visible
6913  * to users decreases. But the relationship is not linear,
6914  * so pick a second-best guess by going with the log2 of the
6915  * number of CPUs.
6916  *
6917  * This idea comes from the SD scheduler of Con Kolivas:
6918  */
6919 static inline void sched_init_granularity(void)
6920 {
6921         unsigned int factor = 1 + ilog2(num_online_cpus());
6922         const unsigned long limit = 200000000;
6923 
6924         sysctl_sched_min_granularity *= factor;
6925         if (sysctl_sched_min_granularity > limit)
6926                 sysctl_sched_min_granularity = limit;
6927 
6928         sysctl_sched_latency *= factor;
6929         if (sysctl_sched_latency > limit)
6930                 sysctl_sched_latency = limit;
6931 
6932         sysctl_sched_wakeup_granularity *= factor;
6933 
6934         sysctl_sched_shares_ratelimit *= factor;
6935 }
6936 
6937 #ifdef CONFIG_SMP
6938 /*
6939  * This is how migration works:
6940  *
6941  * 1) we queue a struct migration_req structure in the source CPU's
6942  *    runqueue and wake up that CPU's migration thread.
6943  * 2) we down() the locked semaphore => thread blocks.
6944  * 3) migration thread wakes up (implicitly it forces the migrated
6945  *    thread off the CPU)
6946  * 4) it gets the migration request and checks whether the migrated
6947  *    task is still in the wrong runqueue.
6948  * 5) if it's in the wrong runqueue then the migration thread removes
6949  *    it and puts it into the right queue.
6950  * 6) migration thread up()s the semaphore.
6951  * 7) we wake up and the migration is done.
6952  */
6953 
6954 /*
6955  * Change a given task's CPU affinity. Migrate the thread to a
6956  * proper CPU and schedule it away if the CPU it's executing on
6957  * is removed from the allowed bitmask.
6958  *
6959  * NOTE: the caller must have a valid reference to the task, the
6960  * task must not exit() & deallocate itself prematurely. The
6961  * call is not atomic; no spinlocks may be held.
6962  */
6963 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6964 {
6965         struct migration_req req;
6966         unsigned long flags;
6967         struct rq *rq;
6968         int ret = 0;
6969 
6970         rq = task_rq_lock(p, &flags);
6971         if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6972                 ret = -EINVAL;
6973                 goto out;
6974         }
6975 
6976         if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6977                      !cpumask_equal(&p->cpus_allowed, new_mask))) {
6978                 ret = -EINVAL;
6979                 goto out;
6980         }
6981 
6982         if (p->sched_class->set_cpus_allowed)
6983                 p->sched_class->set_cpus_allowed(p, new_mask);
6984         else {
6985                 cpumask_copy(&p->cpus_allowed, new_mask);
6986                 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6987         }
6988 
6989         /* Can the task run on the task's current CPU? If so, we're done */
6990         if (cpumask_test_cpu(task_cpu(p), new_mask))
6991                 goto out;
6992 
6993         if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6994                 /* Need help from migration thread: drop lock and wait. */
6995                 task_rq_unlock(rq, &flags);
6996                 wake_up_process(rq->migration_thread);
6997                 wait_for_completion(&req.done);
6998                 tlb_migrate_finish(p->mm);
6999                 return 0;
7000         }
7001 out:
7002         task_rq_unlock(rq, &flags);
7003 
7004         return ret;
7005 }
7006 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7007 
7008 /*
7009  * Move (not current) task off this cpu, onto dest cpu. We're doing
7010  * this because either it can't run here any more (set_cpus_allowed()
7011  * away from this CPU, or CPU going down), or because we're
7012  * attempting to rebalance this task on exec (sched_exec).
7013  *
7014  * So we race with normal scheduler movements, but that's OK, as long
7015  * as the task is no longer on this CPU.
7016  *
7017  * Returns non-zero if task was successfully migrated.
7018  */
7019 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7020 {
7021         struct rq *rq_dest, *rq_src;
7022         int ret = 0, on_rq;
7023 
7024         if (unlikely(!cpu_active(dest_cpu)))
7025                 return ret;
7026 
7027         rq_src = cpu_rq(src_cpu);
7028         rq_dest = cpu_rq(dest_cpu);
7029 
7030         double_rq_lock(rq_src, rq_dest);
7031         /* Already moved. */
7032         if (task_cpu(p) != src_cpu)
7033                 goto done;
7034         /* Affinity changed (again). */
7035         if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7036                 goto fail;
7037 
7038         on_rq = p->se.on_rq;
7039         if (on_rq)
7040                 deactivate_task(rq_src, p, 0);
7041 
7042         set_task_cpu(p, dest_cpu);
7043         if (on_rq) {
7044                 activate_task(rq_dest, p, 0);
7045                 check_preempt_curr(rq_dest, p, 0);
7046         }
7047 done:
7048         ret = 1;
7049 fail:
7050         double_rq_unlock(rq_src, rq_dest);
7051         return ret;
7052 }
7053 
7054 /*
7055  * migration_thread - this is a highprio system thread that performs
7056  * thread migration by bumping thread off CPU then 'pushing' onto
7057  * another runqueue.
7058  */
7059 static int migration_thread(void *data)
7060 {
7061         int cpu = (long)data;
7062         struct rq *rq;
7063 
7064         rq = cpu_rq(cpu);
7065         BUG_ON(rq->migration_thread != current);
7066 
7067         set_current_state(TASK_INTERRUPTIBLE);
7068         while (!kthread_should_stop()) {
7069                 struct migration_req *req;
7070                 struct list_head *head;
7071 
7072                 spin_lock_irq(&rq->lock);
7073 
7074                 if (cpu_is_offline(cpu)) {
7075                         spin_unlock_irq(&rq->lock);
7076                         break;
7077                 }
7078 
7079                 if (rq->active_balance) {
7080                         active_load_balance(rq, cpu);
7081                         rq->active_balance = 0;
7082                 }
7083 
7084                 head = &rq->migration_queue;
7085 
7086                 if (list_empty(head)) {
7087                         spin_unlock_irq(&rq->lock);
7088                         schedule();
7089                         set_current_state(TASK_INTERRUPTIBLE);
7090                         continue;
7091                 }
7092                 req = list_entry(head->next, struct migration_req, list);
7093                 list_del_init(head->next);
7094 
7095                 spin_unlock(&rq->lock);
7096                 __migrate_task(req->task, cpu, req->dest_cpu);
7097                 local_irq_enable();
7098 
7099                 complete(&req->done);
7100         }
7101         __set_current_state(TASK_RUNNING);
7102 
7103         return 0;
7104 }
7105 
7106 #ifdef CONFIG_HOTPLUG_CPU
7107 
7108 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7109 {
7110         int ret;
7111 
7112         local_irq_disable();
7113         ret = __migrate_task(p, src_cpu, dest_cpu);
7114         local_irq_enable();
7115         return ret;
7116 }
7117 
7118 /*
7119  * Figure out where task on dead CPU should go, use force if necessary.
7120  */
7121 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7122 {
7123         int dest_cpu;
7124         const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7125 
7126 again:
7127         /* Look for allowed, online CPU in same node. */
7128         for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7129                 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7130                         goto move;
7131 
7132         /* Any allowed, online CPU? */
7133         dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7134         if (dest_cpu < nr_cpu_ids)
7135                 goto move;
7136 
7137         /* No more Mr. Nice Guy. */
7138         if (dest_cpu >= nr_cpu_ids) {
7139                 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7140                 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7141 
7142                 /*
7143                  * Don't tell them about moving exiting tasks or
7144                  * kernel threads (both mm NULL), since they never
7145                  * leave kernel.
7146                  */
7147                 if (p->mm && printk_ratelimit()) {
7148                         printk(KERN_INFO "process %d (%s) no "
7149                                "longer affine to cpu%d\n",
7150                                task_pid_nr(p), p->comm, dead_cpu);
7151                 }
7152         }
7153 
7154 move:
7155         /* It can have affinity changed while we were choosing. */
7156         if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7157                 goto again;
7158 }
7159 
7160 /*
7161  * While a dead CPU has no uninterruptible tasks queued at this point,
7162  * it might still have a nonzero ->nr_uninterruptible counter, because
7163  * for performance reasons the counter is not stricly tracking tasks to
7164  * their home CPUs. So we just add the counter to another CPU's counter,
7165  * to keep the global sum constant after CPU-down:
7166  */
7167 static void migrate_nr_uninterruptible(struct rq *rq_src)
7168 {
7169         struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7170         unsigned long flags;
7171 
7172         local_irq_save(flags);
7173         double_rq_lock(rq_src, rq_dest);
7174         rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7175         rq_src->nr_uninterruptible = 0;
7176         double_rq_unlock(rq_src, rq_dest);
7177         local_irq_restore(flags);
7178 }
7179 
7180 /* Run through task list and migrate tasks from the dead cpu. */
7181 static void migrate_live_tasks(int src_cpu)
7182 {
7183         struct task_struct *p, *t;
7184 
7185         read_lock(&tasklist_lock);
7186 
7187         do_each_thread(t, p) {
7188                 if (p == current)
7189                         continue;
7190 
7191                 if (task_cpu(p) == src_cpu)
7192                         move_task_off_dead_cpu(src_cpu, p);
7193         } while_each_thread(t, p);
7194 
7195         read_unlock(&tasklist_lock);
7196 }
7197 
7198 /*
7199  * Schedules idle task to be the next runnable task on current CPU.
7200  * It does so by boosting its priority to highest possible.
7201  * Used by CPU offline code.
7202  */
7203 void sched_idle_next(void)
7204 {
7205         int this_cpu = smp_processor_id();
7206         struct rq *rq = cpu_rq(this_cpu);
7207         struct task_struct *p = rq->idle;
7208         unsigned long flags;
7209 
7210         /* cpu has to be offline */
7211         BUG_ON(cpu_online(this_cpu));
7212 
7213         /*
7214          * Strictly not necessary since rest of the CPUs are stopped by now
7215          * and interrupts disabled on the current cpu.
7216          */
7217         spin_lock_irqsave(&rq->lock, flags);
7218 
7219         __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7220 
7221         update_rq_clock(rq);
7222         activate_task(rq, p, 0);
7223 
7224         spin_unlock_irqrestore(&rq->lock, flags);
7225 }
7226 
7227 /*
7228  * Ensures that the idle task is using init_mm right before its cpu goes
7229  * offline.
7230  */
7231 void idle_task_exit(void)
7232 {
7233         struct mm_struct *mm = current->active_mm;
7234 
7235         BUG_ON(cpu_online(smp_processor_id()));
7236 
7237         if (mm != &init_mm)
7238                 switch_mm(mm, &init_mm, current);
7239         mmdrop(mm);
7240 }
7241 
7242 /* called under rq->lock with disabled interrupts */
7243 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7244 {
7245         struct rq *rq = cpu_rq(dead_cpu);
7246 
7247         /* Must be exiting, otherwise would be on tasklist. */
7248         BUG_ON(!p->exit_state);
7249 
7250         /* Cannot have done final schedule yet: would have vanished. */
7251         BUG_ON(p->state == TASK_DEAD);
7252 
7253         get_task_struct(p);
7254 
7255         /*
7256          * Drop lock around migration; if someone else moves it,
7257          * that's OK. No task can be added to this CPU, so iteration is
7258          * fine.
7259          */
7260         spin_unlock_irq(&rq->lock);
7261         move_task_off_dead_cpu(dead_cpu, p);
7262         spin_lock_irq(&rq->lock);
7263 
7264         put_task_struct(p);
7265 }
7266 
7267 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7268 static void migrate_dead_tasks(unsigned int dead_cpu)
7269 {
7270         struct rq *rq = cpu_rq(dead_cpu);
7271         struct task_struct *next;
7272 
7273         for ( ; ; ) {
7274                 if (!rq->nr_running)
7275                         break;
7276                 update_rq_clock(rq);
7277                 next = pick_next_task(rq);
7278                 if (!next)
7279                         break;
7280                 next->sched_class->put_prev_task(rq, next);
7281                 migrate_dead(dead_cpu, next);
7282 
7283         }
7284 }
7285 
7286 /*
7287  * remove the tasks which were accounted by rq from calc_load_tasks.
7288  */
7289 static void calc_global_load_remove(struct rq *rq)
7290 {
7291         atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7292         rq->calc_load_active = 0;
7293 }
7294 #endif /* CONFIG_HOTPLUG_CPU */
7295 
7296 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7297 
7298 static struct ctl_table sd_ctl_dir[] = {
7299         {
7300                 .procname       = "sched_domain",
7301                 .mode           = 0555,
7302         },
7303         {0, },
7304 };
7305 
7306 static struct ctl_table sd_ctl_root[] = {
7307         {
7308                 .ctl_name       = CTL_KERN,
7309                 .procname       = "kernel",
7310                 .mode           = 0555,
7311                 .child          = sd_ctl_dir,
7312         },
7313         {0, },
7314 };
7315 
7316 static struct ctl_table *sd_alloc_ctl_entry(int n)
7317 {
7318         struct ctl_table *entry =
7319                 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7320 
7321         return entry;
7322 }
7323 
7324 static void sd_free_ctl_entry(struct ctl_table **tablep)
7325 {
7326         struct ctl_table *entry;
7327 
7328         /*
7329          * In the intermediate directories, both the child directory and
7330          * procname are dynamically allocated and could fail but the mode
7331          * will always be set. In the lowest directory the names are
7332          * static strings and all have proc handlers.
7333          */
7334         for (entry = *tablep; entry->mode; entry++) {
7335                 if (entry->child)
7336                         sd_free_ctl_entry(&entry->child);
7337                 if (entry->proc_handler == NULL)
7338                         kfree(entry->procname);
7339         }
7340 
7341         kfree(*tablep);
7342         *tablep = NULL;
7343 }
7344 
7345 static void
7346 set_table_entry(struct ctl_table *entry,
7347                 const char *procname, void *data, int maxlen,
7348                 mode_t mode, proc_handler *proc_handler)
7349 {
7350         entry->procname = procname;
7351         entry->data = data;
7352         entry->maxlen = maxlen;
7353         entry->mode = mode;
7354         entry->proc_handler = proc_handler;
7355 }
7356 
7357 static struct ctl_table *
7358 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7359 {
7360         struct ctl_table *table = sd_alloc_ctl_entry(13);
7361 
7362         if (table == NULL)
7363                 return NULL;
7364 
7365         set_table_entry(&table[0], "min_interval", &sd->min_interval,
7366                 sizeof(long), 0644, proc_doulongvec_minmax);
7367         set_table_entry(&table[1], "max_interval", &sd->max_interval,
7368                 sizeof(long), 0644, proc_doulongvec_minmax);
7369         set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7370                 sizeof(int), 0644, proc_dointvec_minmax);
7371         set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7372                 sizeof(int), 0644, proc_dointvec_minmax);
7373         set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7374                 sizeof(int), 0644, proc_dointvec_minmax);
7375         set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7376                 sizeof(int), 0644, proc_dointvec_minmax);
7377         set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7378                 sizeof(int), 0644, proc_dointvec_minmax);
7379         set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7380                 sizeof(int), 0644, proc_dointvec_minmax);
7381         set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7382                 sizeof(int), 0644, proc_dointvec_minmax);
7383         set_table_entry(&table[9], "cache_nice_tries",
7384                 &sd->cache_nice_tries,
7385                 sizeof(int), 0644, proc_dointvec_minmax);
7386         set_table_entry(&table[10], "flags", &sd->flags,
7387                 sizeof(int), 0644, proc_dointvec_minmax);
7388         set_table_entry(&table[11], "name", sd->name,
7389                 CORENAME_MAX_SIZE, 0444, proc_dostring);
7390         /* &table[12] is terminator */
7391 
7392         return table;
7393 }
7394 
7395 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7396 {
7397         struct ctl_table *entry, *table;
7398         struct sched_domain *sd;
7399         int domain_num = 0, i;
7400         char buf[32];
7401 
7402         for_each_domain(cpu, sd)
7403                 domain_num++;
7404         entry = table = sd_alloc_ctl_entry(domain_num + 1);
7405         if (table == NULL)
7406                 return NULL;
7407 
7408         i = 0;
7409         for_each_domain(cpu, sd) {
7410                 snprintf(buf, 32, "domain%d", i);
7411                 entry->procname = kstrdup(buf, GFP_KERNEL);
7412                 entry->mode = 0555;
7413                 entry->child = sd_alloc_ctl_domain_table(sd);
7414                 entry++;
7415                 i++;
7416         }
7417         return table;
7418 }
7419 
7420 static struct ctl_table_header *sd_sysctl_header;
7421 static void register_sched_domain_sysctl(void)
7422 {
7423         int i, cpu_num = num_online_cpus();
7424         struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7425         char buf[32];
7426 
7427         WARN_ON(sd_ctl_dir[0].child);
7428         sd_ctl_dir[0].child = entry;
7429 
7430         if (entry == NULL)
7431                 return;
7432 
7433         for_each_online_cpu(i) {
7434                 snprintf(buf, 32, "cpu%d", i);
7435                 entry->procname = kstrdup(buf, GFP_KERNEL);
7436                 entry->mode = 0555;
7437                 entry->child = sd_alloc_ctl_cpu_table(i);
7438                 entry++;
7439         }
7440 
7441         WARN_ON(sd_sysctl_header);
7442         sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7443 }
7444 
7445 /* may be called multiple times per register */
7446 static void unregister_sched_domain_sysctl(void)
7447 {
7448         if (sd_sysctl_header)
7449                 unregister_sysctl_table(sd_sysctl_header);
7450         sd_sysctl_header = NULL;
7451         if (sd_ctl_dir[0].child)
7452                 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7453 }
7454 #else
7455 static void register_sched_domain_sysctl(void)
7456 {
7457 }
7458 static void unregister_sched_domain_sysctl(void)
7459 {
7460 }
7461 #endif
7462 
7463 static void set_rq_online(struct rq *rq)
7464 {
7465         if (!rq->online) {
7466                 const struct sched_class *class;
7467 
7468                 cpumask_set_cpu(rq->cpu, rq->rd->online);
7469                 rq->online = 1;
7470 
7471                 for_each_class(class) {
7472                         if (class->rq_online)
7473                                 class->rq_online(rq);
7474                 }
7475         }
7476 }
7477 
7478 static void set_rq_offline(struct rq *rq)
7479 {
7480         if (rq->online) {
7481                 const struct sched_class *class;
7482 
7483                 for_each_class(class) {
7484                         if (class->rq_offline)
7485                                 class->rq_offline(rq);
7486                 }
7487 
7488                 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7489                 rq->online = 0;
7490         }
7491 }
7492 
7493 /*
7494  * migration_call - callback that gets triggered when a CPU is added.
7495  * Here we can start up the necessary migration thread for the new CPU.
7496  */
7497 static int __cpuinit
7498 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7499 {
7500         struct task_struct *p;
7501         int cpu = (long)hcpu;
7502         unsigned long flags;
7503         struct rq *rq;
7504 
7505         switch (action) {
7506 
7507         case CPU_UP_PREPARE:
7508         case CPU_UP_PREPARE_FROZEN:
7509                 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7510                 if (IS_ERR(p))
7511                         return NOTIFY_BAD;
7512                 kthread_bind(p, cpu);
7513                 /* Must be high prio: stop_machine expects to yield to it. */
7514                 rq = task_rq_lock(p, &flags);
7515                 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7516                 task_rq_unlock(rq, &flags);
7517                 get_task_struct(p);
7518                 cpu_rq(cpu)->migration_thread = p;
7519                 rq->calc_load_update = calc_load_update;
7520                 break;
7521 
7522         case CPU_ONLINE:
7523         case CPU_ONLINE_FROZEN:
7524                 /* Strictly unnecessary, as first user will wake it. */
7525                 wake_up_process(cpu_rq(cpu)->migration_thread);
7526 
7527                 /* Update our root-domain */
7528                 rq = cpu_rq(cpu);
7529                 spin_lock_irqsave(&rq->lock, flags);
7530                 if (rq->rd) {
7531                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7532 
7533                         set_rq_online(rq);
7534                 }
7535                 spin_unlock_irqrestore(&rq->lock, flags);
7536                 break;
7537 
7538 #ifdef CONFIG_HOTPLUG_CPU
7539         case CPU_UP_CANCELED:
7540         case CPU_UP_CANCELED_FROZEN:
7541                 if (!cpu_rq(cpu)->migration_thread)
7542                         break;
7543                 /* Unbind it from offline cpu so it can run. Fall thru. */
7544                 kthread_bind(cpu_rq(cpu)->migration_thread,
7545                              cpumask_any(cpu_online_mask));
7546                 kthread_stop(cpu_rq(cpu)->migration_thread);
7547                 put_task_struct(cpu_rq(cpu)->migration_thread);
7548                 cpu_rq(cpu)->migration_thread = NULL;
7549                 break;
7550 
7551         case CPU_DEAD:
7552         case CPU_DEAD_FROZEN:
7553                 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7554                 migrate_live_tasks(cpu);
7555                 rq = cpu_rq(cpu);
7556                 kthread_stop(rq->migration_thread);
7557                 put_task_struct(rq->migration_thread);
7558                 rq->migration_thread = NULL;
7559                 /* Idle task back to normal (off runqueue, low prio) */
7560                 spin_lock_irq(&rq->lock);
7561                 update_rq_clock(rq);
7562                 deactivate_task(rq, rq->idle, 0);
7563                 rq->idle->static_prio = MAX_PRIO;
7564                 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7565                 rq->idle->sched_class = &idle_sched_class;
7566                 migrate_dead_tasks(cpu);
7567                 spin_unlock_irq(&rq->lock);
7568                 cpuset_unlock();
7569                 migrate_nr_uninterruptible(rq);
7570                 BUG_ON(rq->nr_running != 0);
7571                 calc_global_load_remove(rq);
7572                 /*
7573                  * No need to migrate the tasks: it was best-effort if
7574                  * they didn't take sched_hotcpu_mutex. Just wake up
7575                  * the requestors.
7576                  */
7577                 spin_lock_irq(&rq->lock);
7578                 while (!list_empty(&rq->migration_queue)) {
7579                         struct migration_req *req;
7580 
7581                         req = list_entry(rq->migration_queue.next,
7582                                          struct migration_req, list);
7583                         list_del_init(&req->list);
7584                         spin_unlock_irq(&rq->lock);
7585                         complete(&req->done);
7586                         spin_lock_irq(&rq->lock);
7587                 }
7588                 spin_unlock_irq(&rq->lock);
7589                 break;
7590 
7591         case CPU_DYING:
7592         case CPU_DYING_FROZEN:
7593                 /* Update our root-domain */
7594                 rq = cpu_rq(cpu);
7595                 spin_lock_irqsave(&rq->lock, flags);
7596                 if (rq->rd) {
7597                         BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7598                         set_rq_offline(rq);
7599                 }
7600                 spin_unlock_irqrestore(&rq->lock, flags);
7601                 break;
7602 #endif
7603         }
7604         return NOTIFY_OK;
7605 }
7606 
7607 /*
7608  * Register at high priority so that task migration (migrate_all_tasks)
7609  * happens before everything else.  This has to be lower priority than
7610  * the notifier in the perf_counter subsystem, though.
7611  */
7612 static struct notifier_block __cpuinitdata migration_notifier = {
7613         .notifier_call = migration_call,
7614         .priority = 10
7615 };
7616 
7617 static int __init migration_init(void)
7618 {
7619         void *cpu = (void *)(long)smp_processor_id();
7620         int err;
7621 
7622         /* Start one for the boot CPU: */
7623         err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7624         BUG_ON(err == NOTIFY_BAD);
7625         migration_call(&migration_notifier, CPU_ONLINE, cpu);
7626         register_cpu_notifier(&migration_notifier);
7627 
7628         return err;
7629 }
7630 early_initcall(migration_init);
7631 #endif
7632 
7633 #ifdef CONFIG_SMP
7634 
7635 #ifdef CONFIG_SCHED_DEBUG
7636 
7637 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7638                                   struct cpumask *groupmask)
7639 {
7640         struct sched_group *group = sd->groups;
7641         char str[256];
7642 
7643         cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7644         cpumask_clear(groupmask);
7645 
7646         printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7647 
7648         if (!(sd->flags & SD_LOAD_BALANCE)) {
7649                 printk("does not load-balance\n");
7650                 if (sd->parent)
7651                         printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7652                                         " has parent");
7653                 return -1;
7654         }
7655 
7656         printk(KERN_CONT "span %s level %s\n", str, sd->name);
7657 
7658         if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7659                 printk(KERN_ERR "ERROR: domain->span does not contain "
7660                                 "CPU%d\n", cpu);
7661         }
7662         if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7663                 printk(KERN_ERR "ERROR: domain->groups does not contain"
7664                                 " CPU%d\n", cpu);
7665         }
7666 
7667         printk(KERN_DEBUG "%*s groups:", level + 1, "");
7668         do {
7669                 if (!group) {
7670                         printk("\n");
7671                         printk(KERN_ERR "ERROR: group is NULL\n");
7672                         break;
7673                 }
7674 
7675                 if (!group->__cpu_power) {
7676                         printk(KERN_CONT "\n");
7677                         printk(KERN_ERR "ERROR: domain->cpu_power not "
7678                                         "set\n");
7679                         break;
7680                 }
7681 
7682                 if (!cpumask_weight(sched_group_cpus(group))) {
7683                         printk(KERN_CONT "\n");
7684                         printk(KERN_ERR "ERROR: empty group\n");
7685                         break;
7686                 }
7687 
7688                 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7689                         printk(KERN_CONT "\n");
7690                         printk(KERN_ERR "ERROR: repeated CPUs\n");
7691                         break;
7692                 }
7693 
7694                 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7695 
7696                 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7697 
7698                 printk(KERN_CONT " %s", str);
7699                 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7700                         printk(KERN_CONT " (__cpu_power = %d)",
7701                                 group->__cpu_power);
7702                 }
7703 
7704                 group = group->next;
7705         } while (group != sd->groups);
7706         printk(KERN_CONT "\n");
7707 
7708         if (!cpumask_equal(sched_domain_span(sd), groupmask))
7709                 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7710 
7711         if (sd->parent &&
7712             !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7713                 printk(KERN_ERR "ERROR: parent span is not a superset "
7714                         "of domain->span\n");
7715         return 0;
7716 }
7717 
7718 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7719 {
7720         cpumask_var_t groupmask;
7721         int level = 0;
7722 
7723         if (!sd) {
7724                 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7725                 return;
7726         }
7727 
7728         printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7729 
7730         if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7731                 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7732                 return;
7733         }
7734 
7735         for (;;) {
7736                 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7737                         break;
7738                 level++;
7739                 sd = sd->parent;
7740                 if (!sd)
7741                         break;
7742         }
7743         free_cpumask_var(groupmask);
7744 }
7745 #else /* !CONFIG_SCHED_DEBUG */
7746 # define sched_domain_debug(sd, cpu) do { } while (0)
7747 #endif /* CONFIG_SCHED_DEBUG */
7748 
7749 static int sd_degenerate(struct sched_domain *sd)
7750 {
7751         if (cpumask_weight(sched_domain_span(sd)) == 1)
7752                 return 1;
7753 
7754         /* Following flags need at least 2 groups */
7755         if (sd->flags & (SD_LOAD_BALANCE |
7756                          SD_BALANCE_NEWIDLE |
7757                          SD_BALANCE_FORK |
7758                          SD_BALANCE_EXEC |
7759                          SD_SHARE_CPUPOWER |
7760                          SD_SHARE_PKG_RESOURCES)) {
7761                 if (sd->groups != sd->groups->next)
7762                         return 0;
7763         }
7764 
7765         /* Following flags don't use groups */
7766         if (sd->flags & (SD_WAKE_IDLE |
7767                          SD_WAKE_AFFINE |
7768                          SD_WAKE_BALANCE))
7769                 return 0;
7770 
7771         return 1;
7772 }
7773 
7774 static int
7775 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7776 {
7777         unsigned long cflags = sd->flags, pflags = parent->flags;
7778 
7779         if (sd_degenerate(parent))
7780                 return 1;
7781 
7782         if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7783                 return 0;
7784 
7785         /* Does parent contain flags not in child? */
7786         /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7787         if (cflags & SD_WAKE_AFFINE)
7788                 pflags &= ~SD_WAKE_BALANCE;
7789         /* Flags needing groups don't count if only 1 group in parent */
7790         if (parent->groups == parent->groups->next) {
7791                 pflags &= ~(SD_LOAD_BALANCE |
7792                                 SD_BALANCE_NEWIDLE |
7793                                 SD_BALANCE_FORK |
7794                                 SD_BALANCE_EXEC |
7795                                 SD_SHARE_CPUPOWER |
7796                                 SD_SHARE_PKG_RESOURCES);
7797                 if (nr_node_ids == 1)
7798                         pflags &= ~SD_SERIALIZE;
7799         }
7800         if (~cflags & pflags)
7801                 return 0;
7802 
7803         return 1;
7804 }
7805 
7806 static void free_rootdomain(struct root_domain *rd)
7807 {
7808         cpupri_cleanup(&rd->cpupri);
7809 
7810         free_cpumask_var(rd->rto_mask);
7811         free_cpumask_var(rd->online);
7812         free_cpumask_var(rd->span);
7813         kfree(rd);
7814 }
7815 
7816 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7817 {
7818         struct root_domain *old_rd = NULL;
7819         unsigned long flags;
7820 
7821         spin_lock_irqsave(&rq->lock, flags);
7822 
7823         if (rq->rd) {
7824                 old_rd = rq->rd;
7825 
7826                 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7827                         set_rq_offline(rq);
7828 
7829                 cpumask_clear_cpu(rq->cpu, old_rd->span);
7830 
7831                 /*
7832                  * If we dont want to free the old_rt yet then
7833                  * set old_rd to NULL to skip the freeing later
7834                  * in this function:
7835                  */
7836                 if (!atomic_dec_and_test(&old_rd->refcount))
7837                         old_rd = NULL;
7838         }
7839 
7840         atomic_inc(&rd->refcount);
7841         rq->rd = rd;
7842 
7843         cpumask_set_cpu(rq->cpu, rd->span);
7844         if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7845                 set_rq_online(rq);
7846 
7847         spin_unlock_irqrestore(&rq->lock, flags);
7848 
7849         if (old_rd)
7850                 free_rootdomain(old_rd);
7851 }
7852 
7853 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7854 {
7855         gfp_t gfp = GFP_KERNEL;
7856 
7857         memset(rd, 0, sizeof(*rd));
7858 
7859         if (bootmem)
7860                 gfp = GFP_NOWAIT;
7861 
7862         if (!alloc_cpumask_var(&rd->span, gfp))
7863                 goto out;
7864         if (!alloc_cpumask_var(&rd->online, gfp))
7865                 goto free_span;
7866         if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7867                 goto free_online;
7868 
7869         if (cpupri_init(&rd->cpupri, bootmem) != 0)
7870                 goto free_rto_mask;
7871         return 0;
7872 
7873 free_rto_mask:
7874         free_cpumask_var(rd->rto_mask);
7875 free_online:
7876         free_cpumask_var(rd->online);
7877 free_span:
7878         free_cpumask_var(rd->span);
7879 out:
7880         return -ENOMEM;
7881 }
7882 
7883 static void init_defrootdomain(void)
7884 {
7885         init_rootdomain(&def_root_domain, true);
7886 
7887         atomic_set(&def_root_domain.refcount, 1);
7888 }
7889 
7890 static struct root_domain *alloc_rootdomain(void)
7891 {
7892         struct root_domain *rd;
7893 
7894         rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7895         if (!rd)
7896                 return NULL;
7897 
7898         if (init_rootdomain(rd, false) != 0) {
7899                 kfree(rd);
7900                 return NULL;
7901         }
7902 
7903         return rd;
7904 }
7905 
7906 /*
7907  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7908  * hold the hotplug lock.
7909  */
7910 static void
7911 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7912 {
7913         struct rq *rq = cpu_rq(cpu);
7914         struct sched_domain *tmp;
7915 
7916         /* Remove the sched domains which do not contribute to scheduling. */
7917         for (tmp = sd; tmp; ) {
7918                 struct sched_domain *parent = tmp->parent;
7919                 if (!parent)
7920                         break;
7921 
7922                 if (sd_parent_degenerate(tmp, parent)) {
7923                         tmp->parent = parent->parent;
7924                         if (parent->parent)
7925                                 parent->parent->child = tmp;
7926                 } else
7927                         tmp = tmp->parent;
7928         }
7929 
7930         if (sd && sd_degenerate(sd)) {
7931                 sd = sd->parent;
7932                 if (sd)
7933                         sd->child = NULL;
7934         }
7935 
7936         sched_domain_debug(sd, cpu);
7937 
7938         rq_attach_root(rq, rd);
7939         rcu_assign_pointer(rq->sd, sd);
7940 }
7941 
7942 /* cpus with isolated domains */
7943 static cpumask_var_t cpu_isolated_map;
7944 
7945 /* Setup the mask of cpus configured for isolated domains */
7946 static int __init isolated_cpu_setup(char *str)
7947 {
7948         alloc_bootmem_cpumask_var(&cpu_isolated_map);
7949         cpulist_parse(str, cpu_isolated_map);
7950         return 1;
7951 }
7952 
7953 __setup("isolcpus=", isolated_cpu_setup);
7954 
7955 /*
7956  * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7957  * to a function which identifies what group(along with sched group) a CPU
7958  * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7959  * (due to the fact that we keep track of groups covered with a struct cpumask).
7960  *
7961  * init_sched_build_groups will build a circular linked list of the groups
7962  * covered by the given span, and will set each group's ->cpumask correctly,
7963  * and ->cpu_power to 0.
7964  */
7965 static void
7966 init_sched_build_groups(const struct cpumask *span,
7967                         const struct cpumask *cpu_map,
7968                         int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7969                                         struct sched_group **sg,
7970                                         struct cpumask *tmpmask),
7971                         struct cpumask *covered, struct cpumask *tmpmask)
7972 {
7973         struct sched_group *first = NULL, *last = NULL;
7974         int i;
7975 
7976         cpumask_clear(covered);
7977 
7978         for_each_cpu(i, span) {
7979                 struct sched_group *sg;
7980                 int group = group_fn(i, cpu_map, &sg, tmpmask);
7981                 int j;
7982 
7983                 if (cpumask_test_cpu(i, covered))
7984                         continue;
7985 
7986                 cpumask_clear(sched_group_cpus(sg));
7987                 sg->__cpu_power = 0;
7988 
7989                 for_each_cpu(j, span) {
7990                         if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7991                                 continue;
7992 
7993                         cpumask_set_cpu(j, covered);
7994                         cpumask_set_cpu(j, sched_group_cpus(sg));
7995                 }
7996                 if (!first)
7997                         first = sg;
7998                 if (last)
7999                         last->next = sg;
8000                 last = sg;
8001         }
8002         last->next = first;
8003 }
8004 
8005 #define SD_NODES_PER_DOMAIN 16
8006 
8007 #ifdef CONFIG_NUMA
8008 
8009 /**
8010  * find_next_best_node - find the next node to include in a sched_domain
8011  * @node: node whose sched_domain we're building
8012  * @used_nodes: nodes already in the sched_domain
8013  *
8014  * Find the next node to include in a given scheduling domain. Simply
8015  * finds the closest node not already in the @used_nodes map.
8016  *
8017  * Should use nodemask_t.
8018  */
8019 static int find_next_best_node(int node, nodemask_t *used_nodes)
8020 {
8021         int i, n, val, min_val, best_node = 0;
8022 
8023         min_val = INT_MAX;
8024 
8025         for (i = 0; i < nr_node_ids; i++) {
8026                 /* Start at @node */
8027                 n = (node + i) % nr_node_ids;
8028 
8029                 if (!nr_cpus_node(n))
8030                         continue;
8031 
8032                 /* Skip already used nodes */
8033                 if (node_isset(n, *used_nodes))
8034                         continue;
8035 
8036                 /* Simple min distance search */
8037                 val = node_distance(node, n);
8038 
8039                 if (val < min_val) {
8040                         min_val = val;
8041                         best_node = n;
8042                 }
8043         }
8044 
8045         node_set(best_node, *used_nodes);
8046         return best_node;
8047 }
8048 
8049 /**
8050  * sched_domain_node_span - get a cpumask for a node's sched_domain
8051  * @node: node whose cpumask we're constructing
8052  * @span: resulting cpumask
8053  *
8054  * Given a node, construct a good cpumask for its sched_domain to span. It
8055  * should be one that prevents unnecessary balancing, but also spreads tasks
8056  * out optimally.
8057  */
8058 static void sched_domain_node_span(int node, struct cpumask *span)
8059 {
8060         nodemask_t used_nodes;
8061         int i;
8062 
8063         cpumask_clear(span);
8064         nodes_clear(used_nodes);
8065 
8066         cpumask_or(span, span, cpumask_of_node(node));
8067         node_set(node, used_nodes);
8068 
8069         for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8070                 int next_node = find_next_best_node(node, &used_nodes);
8071 
8072                 cpumask_or(span, span, cpumask_of_node(next_node));
8073         }
8074 }
8075 #endif /* CONFIG_NUMA */
8076 
8077 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8078 
8079 /*
8080  * The cpus mask in sched_group and sched_domain hangs off the end.
8081  *
8082  * ( See the the comments in include/linux/sched.h:struct sched_group
8083  *   and struct sched_domain. )
8084  */
8085 struct static_sched_group {
8086         struct sched_group sg;
8087         DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8088 };
8089 
8090 struct static_sched_domain {
8091         struct sched_domain sd;
8092         DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8093 };
8094 
8095 /*
8096  * SMT sched-domains:
8097  */
8098 #ifdef CONFIG_SCHED_SMT
8099 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8100 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8101 
8102 static int
8103 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8104                  struct sched_group **sg, struct cpumask *unused)
8105 {
8106         if (sg)
8107                 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8108         return cpu;
8109 }
8110 #endif /* CONFIG_SCHED_SMT */
8111 
8112 /*
8113  * multi-core sched-domains:
8114  */
8115 #ifdef CONFIG_SCHED_MC
8116 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8117 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8118 #endif /* CONFIG_SCHED_MC */
8119 
8120 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8121 static int
8122 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8123                   struct sched_group **sg, struct cpumask *mask)
8124 {
8125         int group;
8126 
8127         cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8128         group = cpumask_first(mask);
8129         if (sg)
8130                 *sg = &per_cpu(sched_group_core, group).sg;
8131         return group;
8132 }
8133 #elif defined(CONFIG_SCHED_MC)
8134 static int
8135 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8136                   struct sched_group **sg, struct cpumask *unused)
8137 {
8138         if (sg)
8139                 *sg = &per_cpu(sched_group_core, cpu).sg;
8140         return cpu;
8141 }
8142 #endif
8143 
8144 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8145 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8146 
8147 static int
8148 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8149                   struct sched_group **sg, struct cpumask *mask)
8150 {
8151         int group;
8152 #ifdef CONFIG_SCHED_MC
8153         cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8154         group = cpumask_first(mask);
8155 #elif defined(CONFIG_SCHED_SMT)
8156         cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8157         group = cpumask_first(mask);
8158 #else
8159         group = cpu;
8160 #endif
8161         if (sg)
8162                 *sg = &per_cpu(sched_group_phys, group).sg;
8163         return group;
8164 }
8165 
8166 #ifdef CONFIG_NUMA
8167 /*
8168  * The init_sched_build_groups can't handle what we want to do with node
8169  * groups, so roll our own. Now each node has its own list of groups which
8170  * gets dynamically allocated.
8171  */
8172 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8173 static struct sched_group ***sched_group_nodes_bycpu;
8174 
8175 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8176 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8177 
8178 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8179                                  struct sched_group **sg,
8180                                  struct cpumask *nodemask)
8181 {
8182         int group;
8183 
8184         cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8185         group = cpumask_first(nodemask);
8186 
8187         if (sg)
8188                 *sg = &per_cpu(sched_group_allnodes, group).sg;
8189         return group;
8190 }
8191 
8192 static void init_numa_sched_groups_power(struct sched_group *group_head)
8193 {
8194         struct sched_group *sg = group_head;
8195         int j;
8196 
8197         if (!sg)
8198                 return;
8199         do {
8200                 for_each_cpu(j, sched_group_cpus(sg)) {
8201                         struct sched_domain *sd;
8202 
8203                         sd = &per_cpu(phys_domains, j).sd;
8204                         if (j != group_first_cpu(sd->groups)) {
8205                                 /*
8206                                  * Only add "power" once for each
8207                                  * physical package.
8208                                  */
8209                                 continue;
8210                         }
8211 
8212                         sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8213                 }
8214                 sg = sg->next;
8215         } while (sg != group_head);
8216 }
8217 #endif /* CONFIG_NUMA */
8218 
8219 #ifdef CONFIG_NUMA
8220 /* Free memory allocated for various sched_group structures */
8221 static void free_sched_groups(const struct cpumask *cpu_map,
8222                               struct cpumask *nodemask)
8223 {
8224         int cpu, i;
8225 
8226         for_each_cpu(cpu, cpu_map) {
8227                 struct sched_group **sched_group_nodes
8228                         = sched_group_nodes_bycpu[cpu];
8229 
8230                 if (!sched_group_nodes)
8231                         continue;
8232 
8233                 for (i = 0; i < nr_node_ids; i++) {
8234                         struct sched_group *oldsg, *sg = sched_group_nodes[i];
8235 
8236                         cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8237                         if (cpumask_empty(nodemask))
8238                                 continue;
8239 
8240                         if (sg == NULL)
8241                                 continue;
8242                         sg = sg->next;
8243 next_sg:
8244                         oldsg = sg;
8245                         sg = sg->next;
8246                         kfree(oldsg);
8247                         if (oldsg != sched_group_nodes[i])
8248                                 goto next_sg;
8249                 }
8250                 kfree(sched_group_nodes);
8251                 sched_group_nodes_bycpu[cpu] = NULL;
8252         }
8253 }
8254 #else /* !CONFIG_NUMA */
8255 static void free_sched_groups(const struct cpumask *cpu_map,
8256                               struct cpumask *nodemask)
8257 {
8258 }
8259 #endif /* CONFIG_NUMA */
8260 
8261 /*
8262  * Initialize sched groups cpu_power.
8263  *
8264  * cpu_power indicates the capacity of sched group, which is used while
8265  * distributing the load between different sched groups in a sched domain.
8266  * Typically cpu_power for all the groups in a sched domain will be same unless
8267  * there are asymmetries in the topology. If there are asymmetries, group
8268  * having more cpu_power will pickup more load compared to the group having
8269  * less cpu_power.
8270  *
8271  * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8272  * the maximum number of tasks a group can handle in the presence of other idle
8273  * or lightly loaded groups in the same sched domain.
8274  */
8275 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8276 {
8277         struct sched_domain *child;
8278         struct sched_group *group;
8279 
8280         WARN_ON(!sd || !sd->groups);
8281 
8282         if (cpu != group_first_cpu(sd->groups))
8283                 return;
8284 
8285         child = sd->child;
8286 
8287         sd->groups->__cpu_power = 0;
8288 
8289         /*
8290          * For perf policy, if the groups in child domain share resources
8291          * (for example cores sharing some portions of the cache hierarchy
8292          * or SMT), then set this domain groups cpu_power such that each group
8293          * can handle only one task, when there are other idle groups in the
8294          * same sched domain.
8295          */
8296         if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8297                        (child->flags &
8298                         (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8299                 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8300                 return;
8301         }
8302 
8303         /*
8304          * add cpu_power of each child group to this groups cpu_power
8305          */
8306         group = child->groups;
8307         do {
8308                 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8309                 group = group->next;
8310         } while (group != child->groups);
8311 }
8312 
8313 /*
8314  * Initializers for schedule domains
8315  * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8316  */
8317 
8318 #ifdef CONFIG_SCHED_DEBUG
8319 # define SD_INIT_NAME(sd, type)         sd->name = #type
8320 #else
8321 # define SD_INIT_NAME(sd, type)         do { } while (0)
8322 #endif
8323 
8324 #define SD_INIT(sd, type)       sd_init_##type(sd)
8325 
8326 #define SD_INIT_FUNC(type)      \
8327 static noinline void sd_init_##type(struct sched_domain *sd)    \
8328 {                                                               \
8329         memset(sd, 0, sizeof(*sd));                             \
8330         *sd = SD_##type##_INIT;                                 \
8331         sd->level = SD_LV_##type;                               \
8332         SD_INIT_NAME(sd, type);                                 \
8333 }
8334 
8335 SD_INIT_FUNC(CPU)
8336 #ifdef CONFIG_NUMA
8337  SD_INIT_FUNC(ALLNODES)
8338  SD_INIT_FUNC(NODE)
8339 #endif
8340 #ifdef CONFIG_SCHED_SMT
8341  SD_INIT_FUNC(SIBLING)
8342 #endif
8343 #ifdef CONFIG_SCHED_MC
8344  SD_INIT_FUNC(MC)
8345 #endif
8346 
8347 static int default_relax_domain_level = -1;
8348 
8349 static int __init setup_relax_domain_level(char *str)
8350 {
8351         unsigned long val;
8352 
8353         val = simple_strtoul(str, NULL, 0);
8354         if (val < SD_LV_MAX)
8355                 default_relax_domain_level = val;
8356 
8357         return 1;
8358 }
8359 __setup("relax_domain_level=", setup_relax_domain_level);
8360 
8361 static void set_domain_attribute(struct sched_domain *sd,
8362                                  struct sched_domain_attr *attr)
8363 {
8364         int request;
8365