linux/kernel/sched.c
<<
>>
Prefs
   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 */
  26
  27#include <linux/mm.h>
  28#include <linux/module.h>
  29#include <linux/nmi.h>
  30#include <linux/init.h>
  31#include <linux/uaccess.h>
  32#include <linux/highmem.h>
  33#include <linux/smp_lock.h>
  34#include <asm/mmu_context.h>
  35#include <linux/interrupt.h>
  36#include <linux/capability.h>
  37#include <linux/completion.h>
  38#include <linux/kernel_stat.h>
  39#include <linux/debug_locks.h>
  40#include <linux/security.h>
  41#include <linux/notifier.h>
  42#include <linux/profile.h>
  43#include <linux/freezer.h>
  44#include <linux/vmalloc.h>
  45#include <linux/blkdev.h>
  46#include <linux/delay.h>
  47#include <linux/pid_namespace.h>
  48#include <linux/smp.h>
  49#include <linux/threads.h>
  50#include <linux/timer.h>
  51#include <linux/rcupdate.h>
  52#include <linux/cpu.h>
  53#include <linux/cpuset.h>
  54#include <linux/percpu.h>
  55#include <linux/kthread.h>
  56#include <linux/seq_file.h>
  57#include <linux/sysctl.h>
  58#include <linux/syscalls.h>
  59#include <linux/times.h>
  60#include <linux/tsacct_kern.h>
  61#include <linux/kprobes.h>
  62#include <linux/delayacct.h>
  63#include <linux/reciprocal_div.h>
  64#include <linux/unistd.h>
  65#include <linux/pagemap.h>
  66
  67#include <asm/tlb.h>
  68#include <asm/irq_regs.h>
  69
  70/*
  71 * Scheduler clock - returns current time in nanosec units.
  72 * This is default implementation.
  73 * Architectures and sub-architectures can override this.
  74 */
  75unsigned long long __attribute__((weak)) sched_clock(void)
  76{
  77        return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
  78}
  79
  80/*
  81 * Convert user-nice values [ -20 ... 0 ... 19 ]
  82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
  83 * and back.
  84 */
  85#define NICE_TO_PRIO(nice)      (MAX_RT_PRIO + (nice) + 20)
  86#define PRIO_TO_NICE(prio)      ((prio) - MAX_RT_PRIO - 20)
  87#define TASK_NICE(p)            PRIO_TO_NICE((p)->static_prio)
  88
  89/*
  90 * 'User priority' is the nice value converted to something we
  91 * can work with better when scaling various scheduler parameters,
  92 * it's a [ 0 ... 39 ] range.
  93 */
  94#define USER_PRIO(p)            ((p)-MAX_RT_PRIO)
  95#define TASK_USER_PRIO(p)       USER_PRIO((p)->static_prio)
  96#define MAX_USER_PRIO           (USER_PRIO(MAX_PRIO))
  97
  98/*
  99 * Some helpers for converting nanosecond timing to jiffy resolution
 100 */
 101#define NS_TO_JIFFIES(TIME)     ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
 102#define JIFFIES_TO_NS(TIME)     ((TIME) * (NSEC_PER_SEC / HZ))
 103
 104#define NICE_0_LOAD             SCHED_LOAD_SCALE
 105#define NICE_0_SHIFT            SCHED_LOAD_SHIFT
 106
 107/*
 108 * These are the 'tuning knobs' of the scheduler:
 109 *
 110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
 111 * Timeslices get refilled after they expire.
 112 */
 113#define DEF_TIMESLICE           (100 * HZ / 1000)
 114
 115#ifdef CONFIG_SMP
 116/*
 117 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
 118 * Since cpu_power is a 'constant', we can use a reciprocal divide.
 119 */
 120static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
 121{
 122        return reciprocal_divide(load, sg->reciprocal_cpu_power);
 123}
 124
 125/*
 126 * Each time a sched group cpu_power is changed,
 127 * we must compute its reciprocal value
 128 */
 129static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
 130{
 131        sg->__cpu_power += val;
 132        sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
 133}
 134#endif
 135
 136static inline int rt_policy(int policy)
 137{
 138        if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
 139                return 1;
 140        return 0;
 141}
 142
 143static inline int task_has_rt_policy(struct task_struct *p)
 144{
 145        return rt_policy(p->policy);
 146}
 147
 148/*
 149 * This is the priority-queue data structure of the RT scheduling class:
 150 */
 151struct rt_prio_array {
 152        DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
 153        struct list_head queue[MAX_RT_PRIO];
 154};
 155
 156#ifdef CONFIG_FAIR_GROUP_SCHED
 157
 158#include <linux/cgroup.h>
 159
 160struct cfs_rq;
 161
 162/* task group related information */
 163struct task_group {
 164#ifdef CONFIG_FAIR_CGROUP_SCHED
 165        struct cgroup_subsys_state css;
 166#endif
 167        /* schedulable entities of this group on each cpu */
 168        struct sched_entity **se;
 169        /* runqueue "owned" by this group on each cpu */
 170        struct cfs_rq **cfs_rq;
 171        unsigned long shares;
 172        /* spinlock to serialize modification to shares */
 173        spinlock_t lock;
 174        struct rcu_head rcu;
 175};
 176
 177/* Default task group's sched entity on each cpu */
 178static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
 179/* Default task group's cfs_rq on each cpu */
 180static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
 181
 182static struct sched_entity *init_sched_entity_p[NR_CPUS];
 183static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
 184
 185/* Default task group.
 186 *      Every task in system belong to this group at bootup.
 187 */
 188struct task_group init_task_group = {
 189        .se     = init_sched_entity_p,
 190        .cfs_rq = init_cfs_rq_p,
 191};
 192
 193#ifdef CONFIG_FAIR_USER_SCHED
 194# define INIT_TASK_GRP_LOAD     2*NICE_0_LOAD
 195#else
 196# define INIT_TASK_GRP_LOAD     NICE_0_LOAD
 197#endif
 198
 199static int init_task_group_load = INIT_TASK_GRP_LOAD;
 200
 201/* return group to which a task belongs */
 202static inline struct task_group *task_group(struct task_struct *p)
 203{
 204        struct task_group *tg;
 205
 206#ifdef CONFIG_FAIR_USER_SCHED
 207        tg = p->user->tg;
 208#elif defined(CONFIG_FAIR_CGROUP_SCHED)
 209        tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
 210                                struct task_group, css);
 211#else
 212        tg = &init_task_group;
 213#endif
 214        return tg;
 215}
 216
 217/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
 218static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu)
 219{
 220        p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
 221        p->se.parent = task_group(p)->se[cpu];
 222}
 223
 224#else
 225
 226static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) { }
 227
 228#endif  /* CONFIG_FAIR_GROUP_SCHED */
 229
 230/* CFS-related fields in a runqueue */
 231struct cfs_rq {
 232        struct load_weight load;
 233        unsigned long nr_running;
 234
 235        u64 exec_clock;
 236        u64 min_vruntime;
 237
 238        struct rb_root tasks_timeline;
 239        struct rb_node *rb_leftmost;
 240        struct rb_node *rb_load_balance_curr;
 241        /* 'curr' points to currently running entity on this cfs_rq.
 242         * It is set to NULL otherwise (i.e when none are currently running).
 243         */
 244        struct sched_entity *curr;
 245
 246        unsigned long nr_spread_over;
 247
 248#ifdef CONFIG_FAIR_GROUP_SCHED
 249        struct rq *rq;  /* cpu runqueue to which this cfs_rq is attached */
 250
 251        /*
 252         * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
 253         * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
 254         * (like users, containers etc.)
 255         *
 256         * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
 257         * list is used during load balance.
 258         */
 259        struct list_head leaf_cfs_rq_list;
 260        struct task_group *tg;  /* group that "owns" this runqueue */
 261#endif
 262};
 263
 264/* Real-Time classes' related field in a runqueue: */
 265struct rt_rq {
 266        struct rt_prio_array active;
 267        int rt_load_balance_idx;
 268        struct list_head *rt_load_balance_head, *rt_load_balance_curr;
 269};
 270
 271/*
 272 * This is the main, per-CPU runqueue data structure.
 273 *
 274 * Locking rule: those places that want to lock multiple runqueues
 275 * (such as the load balancing or the thread migration code), lock
 276 * acquire operations must be ordered by ascending &runqueue.
 277 */
 278struct rq {
 279        /* runqueue lock: */
 280        spinlock_t lock;
 281
 282        /*
 283         * nr_running and cpu_load should be in the same cacheline because
 284         * remote CPUs use both these fields when doing load calculation.
 285         */
 286        unsigned long nr_running;
 287        #define CPU_LOAD_IDX_MAX 5
 288        unsigned long cpu_load[CPU_LOAD_IDX_MAX];
 289        unsigned char idle_at_tick;
 290#ifdef CONFIG_NO_HZ
 291        unsigned char in_nohz_recently;
 292#endif
 293        /* capture load from *all* tasks on this cpu: */
 294        struct load_weight load;
 295        unsigned long nr_load_updates;
 296        u64 nr_switches;
 297
 298        struct cfs_rq cfs;
 299#ifdef CONFIG_FAIR_GROUP_SCHED
 300        /* list of leaf cfs_rq on this cpu: */
 301        struct list_head leaf_cfs_rq_list;
 302#endif
 303        struct rt_rq rt;
 304
 305        /*
 306         * This is part of a global counter where only the total sum
 307         * over all CPUs matters. A task can increase this counter on
 308         * one CPU and if it got migrated afterwards it may decrease
 309         * it on another CPU. Always updated under the runqueue lock:
 310         */
 311        unsigned long nr_uninterruptible;
 312
 313        struct task_struct *curr, *idle;
 314        unsigned long next_balance;
 315        struct mm_struct *prev_mm;
 316
 317        u64 clock, prev_clock_raw;
 318        s64 clock_max_delta;
 319
 320        unsigned int clock_warps, clock_overflows;
 321        u64 idle_clock;
 322        unsigned int clock_deep_idle_events;
 323        u64 tick_timestamp;
 324
 325        atomic_t nr_iowait;
 326
 327#ifdef CONFIG_SMP
 328        struct sched_domain *sd;
 329
 330        /* For active balancing */
 331        int active_balance;
 332        int push_cpu;
 333        /* cpu of this runqueue: */
 334        int cpu;
 335
 336        struct task_struct *migration_thread;
 337        struct list_head migration_queue;
 338#endif
 339
 340#ifdef CONFIG_SCHEDSTATS
 341        /* latency stats */
 342        struct sched_info rq_sched_info;
 343
 344        /* sys_sched_yield() stats */
 345        unsigned int yld_exp_empty;
 346        unsigned int yld_act_empty;
 347        unsigned int yld_both_empty;
 348        unsigned int yld_count;
 349
 350        /* schedule() stats */
 351        unsigned int sched_switch;
 352        unsigned int sched_count;
 353        unsigned int sched_goidle;
 354
 355        /* try_to_wake_up() stats */
 356        unsigned int ttwu_count;
 357        unsigned int ttwu_local;
 358
 359        /* BKL stats */
 360        unsigned int bkl_count;
 361#endif
 362        struct lock_class_key rq_lock_key;
 363};
 364
 365static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
 366static DEFINE_MUTEX(sched_hotcpu_mutex);
 367
 368static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
 369{
 370        rq->curr->sched_class->check_preempt_curr(rq, p);
 371}
 372
 373static inline int cpu_of(struct rq *rq)
 374{
 375#ifdef CONFIG_SMP
 376        return rq->cpu;
 377#else
 378        return 0;
 379#endif
 380}
 381
 382/*
 383 * Update the per-runqueue clock, as finegrained as the platform can give
 384 * us, but without assuming monotonicity, etc.:
 385 */
 386static void __update_rq_clock(struct rq *rq)
 387{
 388        u64 prev_raw = rq->prev_clock_raw;
 389        u64 now = sched_clock();
 390        s64 delta = now - prev_raw;
 391        u64 clock = rq->clock;
 392
 393#ifdef CONFIG_SCHED_DEBUG
 394        WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
 395#endif
 396        /*
 397         * Protect against sched_clock() occasionally going backwards:
 398         */
 399        if (unlikely(delta < 0)) {
 400                clock++;
 401                rq->clock_warps++;
 402        } else {
 403                /*
 404                 * Catch too large forward jumps too:
 405                 */
 406                if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
 407                        if (clock < rq->tick_timestamp + TICK_NSEC)
 408                                clock = rq->tick_timestamp + TICK_NSEC;
 409                        else
 410                                clock++;
 411                        rq->clock_overflows++;
 412                } else {
 413                        if (unlikely(delta > rq->clock_max_delta))
 414                                rq->clock_max_delta = delta;
 415                        clock += delta;
 416                }
 417        }
 418
 419        rq->prev_clock_raw = now;
 420        rq->clock = clock;
 421}
 422
 423static void update_rq_clock(struct rq *rq)
 424{
 425        if (likely(smp_processor_id() == cpu_of(rq)))
 426                __update_rq_clock(rq);
 427}
 428
 429/*
 430 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
 431 * See detach_destroy_domains: synchronize_sched for details.
 432 *
 433 * The domain tree of any CPU may only be accessed from within
 434 * preempt-disabled sections.
 435 */
 436#define for_each_domain(cpu, __sd) \
 437        for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
 438
 439#define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
 440#define this_rq()               (&__get_cpu_var(runqueues))
 441#define task_rq(p)              cpu_rq(task_cpu(p))
 442#define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
 443
 444/*
 445 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
 446 */
 447#ifdef CONFIG_SCHED_DEBUG
 448# define const_debug __read_mostly
 449#else
 450# define const_debug static const
 451#endif
 452
 453/*
 454 * Debugging: various feature bits
 455 */
 456enum {
 457        SCHED_FEAT_NEW_FAIR_SLEEPERS    = 1,
 458        SCHED_FEAT_WAKEUP_PREEMPT       = 2,
 459        SCHED_FEAT_START_DEBIT          = 4,
 460        SCHED_FEAT_TREE_AVG             = 8,
 461        SCHED_FEAT_APPROX_AVG           = 16,
 462};
 463
 464const_debug unsigned int sysctl_sched_features =
 465                SCHED_FEAT_NEW_FAIR_SLEEPERS    * 1 |
 466                SCHED_FEAT_WAKEUP_PREEMPT       * 1 |
 467                SCHED_FEAT_START_DEBIT          * 1 |
 468                SCHED_FEAT_TREE_AVG             * 0 |
 469                SCHED_FEAT_APPROX_AVG           * 0;
 470
 471#define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
 472
 473/*
 474 * Number of tasks to iterate in a single balance run.
 475 * Limited because this is done with IRQs disabled.
 476 */
 477const_debug unsigned int sysctl_sched_nr_migrate = 32;
 478
 479/*
 480 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
 481 * clock constructed from sched_clock():
 482 */
 483unsigned long long cpu_clock(int cpu)
 484{
 485        unsigned long long now;
 486        unsigned long flags;
 487        struct rq *rq;
 488
 489        local_irq_save(flags);
 490        rq = cpu_rq(cpu);
 491        /*
 492         * Only call sched_clock() if the scheduler has already been
 493         * initialized (some code might call cpu_clock() very early):
 494         */
 495        if (rq->idle)
 496                update_rq_clock(rq);
 497        now = rq->clock;
 498        local_irq_restore(flags);
 499
 500        return now;
 501}
 502EXPORT_SYMBOL_GPL(cpu_clock);
 503
 504#ifndef prepare_arch_switch
 505# define prepare_arch_switch(next)      do { } while (0)
 506#endif
 507#ifndef finish_arch_switch
 508# define finish_arch_switch(prev)       do { } while (0)
 509#endif
 510
 511static inline int task_current(struct rq *rq, struct task_struct *p)
 512{
 513        return rq->curr == p;
 514}
 515
 516#ifndef __ARCH_WANT_UNLOCKED_CTXSW
 517static inline int task_running(struct rq *rq, struct task_struct *p)
 518{
 519        return task_current(rq, p);
 520}
 521
 522static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
 523{
 524}
 525
 526static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
 527{
 528#ifdef CONFIG_DEBUG_SPINLOCK
 529        /* this is a valid case when another task releases the spinlock */
 530        rq->lock.owner = current;
 531#endif
 532        /*
 533         * If we are tracking spinlock dependencies then we have to
 534         * fix up the runqueue lock - which gets 'carried over' from
 535         * prev into current:
 536         */
 537        spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
 538
 539        spin_unlock_irq(&rq->lock);
 540}
 541
 542#else /* __ARCH_WANT_UNLOCKED_CTXSW */
 543static inline int task_running(struct rq *rq, struct task_struct *p)
 544{
 545#ifdef CONFIG_SMP
 546        return p->oncpu;
 547#else
 548        return task_current(rq, p);
 549#endif
 550}
 551
 552static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
 553{
 554#ifdef CONFIG_SMP
 555        /*
 556         * We can optimise this out completely for !SMP, because the
 557         * SMP rebalancing from interrupt is the only thing that cares
 558         * here.
 559         */
 560        next->oncpu = 1;
 561#endif
 562#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
 563        spin_unlock_irq(&rq->lock);
 564#else
 565        spin_unlock(&rq->lock);
 566#endif
 567}
 568
 569static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
 570{
 571#ifdef CONFIG_SMP
 572        /*
 573         * After ->oncpu is cleared, the task can be moved to a different CPU.
 574         * We must ensure this doesn't happen until the switch is completely
 575         * finished.
 576         */
 577        smp_wmb();
 578        prev->oncpu = 0;
 579#endif
 580#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
 581        local_irq_enable();
 582#endif
 583}
 584#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
 585
 586/*
 587 * __task_rq_lock - lock the runqueue a given task resides on.
 588 * Must be called interrupts disabled.
 589 */
 590static inline struct rq *__task_rq_lock(struct task_struct *p)
 591        __acquires(rq->lock)
 592{
 593        for (;;) {
 594                struct rq *rq = task_rq(p);
 595                spin_lock(&rq->lock);
 596                if (likely(rq == task_rq(p)))
 597                        return rq;
 598                spin_unlock(&rq->lock);
 599        }
 600}
 601
 602/*
 603 * task_rq_lock - lock the runqueue a given task resides on and disable
 604 * interrupts. Note the ordering: we can safely lookup the task_rq without
 605 * explicitly disabling preemption.
 606 */
 607static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
 608        __acquires(rq->lock)
 609{
 610        struct rq *rq;
 611
 612        for (;;) {
 613                local_irq_save(*flags);
 614                rq = task_rq(p);
 615                spin_lock(&rq->lock);
 616                if (likely(rq == task_rq(p)))
 617                        return rq;
 618                spin_unlock_irqrestore(&rq->lock, *flags);
 619        }
 620}
 621
 622static void __task_rq_unlock(struct rq *rq)
 623        __releases(rq->lock)
 624{
 625        spin_unlock(&rq->lock);
 626}
 627
 628static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
 629        __releases(rq->lock)
 630{
 631        spin_unlock_irqrestore(&rq->lock, *flags);
 632}
 633
 634/*
 635 * this_rq_lock - lock this runqueue and disable interrupts.
 636 */
 637static struct rq *this_rq_lock(void)
 638        __acquires(rq->lock)
 639{
 640        struct rq *rq;
 641
 642        local_irq_disable();
 643        rq = this_rq();
 644        spin_lock(&rq->lock);
 645
 646        return rq;
 647}
 648
 649/*
 650 * We are going deep-idle (irqs are disabled):
 651 */
 652void sched_clock_idle_sleep_event(void)
 653{
 654        struct rq *rq = cpu_rq(smp_processor_id());
 655
 656        spin_lock(&rq->lock);
 657        __update_rq_clock(rq);
 658        spin_unlock(&rq->lock);
 659        rq->clock_deep_idle_events++;
 660}
 661EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 662
 663/*
 664 * We just idled delta nanoseconds (called with irqs disabled):
 665 */
 666void sched_clock_idle_wakeup_event(u64 delta_ns)
 667{
 668        struct rq *rq = cpu_rq(smp_processor_id());
 669        u64 now = sched_clock();
 670
 671        touch_softlockup_watchdog();
 672        rq->idle_clock += delta_ns;
 673        /*
 674         * Override the previous timestamp and ignore all
 675         * sched_clock() deltas that occured while we idled,
 676         * and use the PM-provided delta_ns to advance the
 677         * rq clock:
 678         */
 679        spin_lock(&rq->lock);
 680        rq->prev_clock_raw = now;
 681        rq->clock += delta_ns;
 682        spin_unlock(&rq->lock);
 683}
 684EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 685
 686/*
 687 * resched_task - mark a task 'to be rescheduled now'.
 688 *
 689 * On UP this means the setting of the need_resched flag, on SMP it
 690 * might also involve a cross-CPU call to trigger the scheduler on
 691 * the target CPU.
 692 */
 693#ifdef CONFIG_SMP
 694
 695#ifndef tsk_is_polling
 696#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
 697#endif
 698
 699static void resched_task(struct task_struct *p)
 700{
 701        int cpu;
 702
 703        assert_spin_locked(&task_rq(p)->lock);
 704
 705        if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
 706                return;
 707
 708        set_tsk_thread_flag(p, TIF_NEED_RESCHED);
 709
 710        cpu = task_cpu(p);
 711        if (cpu == smp_processor_id())
 712                return;
 713
 714        /* NEED_RESCHED must be visible before we test polling */
 715        smp_mb();
 716        if (!tsk_is_polling(p))
 717                smp_send_reschedule(cpu);
 718}
 719
 720static void resched_cpu(int cpu)
 721{
 722        struct rq *rq = cpu_rq(cpu);
 723        unsigned long flags;
 724
 725        if (!spin_trylock_irqsave(&rq->lock, flags))
 726                return;
 727        resched_task(cpu_curr(cpu));
 728        spin_unlock_irqrestore(&rq->lock, flags);
 729}
 730#else
 731static inline void resched_task(struct task_struct *p)
 732{
 733        assert_spin_locked(&task_rq(p)->lock);
 734        set_tsk_need_resched(p);
 735}
 736#endif
 737
 738#if BITS_PER_LONG == 32
 739# define WMULT_CONST    (~0UL)
 740#else
 741# define WMULT_CONST    (1UL << 32)
 742#endif
 743
 744#define WMULT_SHIFT     32
 745
 746/*
 747 * Shift right and round:
 748 */
 749#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
 750
 751static unsigned long
 752calc_delta_mine(unsigned long delta_exec, unsigned long weight,
 753                struct load_weight *lw)
 754{
 755        u64 tmp;
 756
 757        if (unlikely(!lw->inv_weight))
 758                lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
 759
 760        tmp = (u64)delta_exec * weight;
 761        /*
 762         * Check whether we'd overflow the 64-bit multiplication:
 763         */
 764        if (unlikely(tmp > WMULT_CONST))
 765                tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
 766                        WMULT_SHIFT/2);
 767        else
 768                tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
 769
 770        return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
 771}
 772
 773static inline unsigned long
 774calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
 775{
 776        return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
 777}
 778
 779static inline void update_load_add(struct load_weight *lw, unsigned long inc)
 780{
 781        lw->weight += inc;
 782}
 783
 784static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
 785{
 786        lw->weight -= dec;
 787}
 788
 789/*
 790 * To aid in avoiding the subversion of "niceness" due to uneven distribution
 791 * of tasks with abnormal "nice" values across CPUs the contribution that
 792 * each task makes to its run queue's load is weighted according to its
 793 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
 794 * scaled version of the new time slice allocation that they receive on time
 795 * slice expiry etc.
 796 */
 797
 798#define WEIGHT_IDLEPRIO         2
 799#define WMULT_IDLEPRIO          (1 << 31)
 800
 801/*
 802 * Nice levels are multiplicative, with a gentle 10% change for every
 803 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
 804 * nice 1, it will get ~10% less CPU time than another CPU-bound task
 805 * that remained on nice 0.
 806 *
 807 * The "10% effect" is relative and cumulative: from _any_ nice level,
 808 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
 809 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
 810 * If a task goes up by ~10% and another task goes down by ~10% then
 811 * the relative distance between them is ~25%.)
 812 */
 813static const int prio_to_weight[40] = {
 814 /* -20 */     88761,     71755,     56483,     46273,     36291,
 815 /* -15 */     29154,     23254,     18705,     14949,     11916,
 816 /* -10 */      9548,      7620,      6100,      4904,      3906,
 817 /*  -5 */      3121,      2501,      1991,      1586,      1277,
 818 /*   0 */      1024,       820,       655,       526,       423,
 819 /*   5 */       335,       272,       215,       172,       137,
 820 /*  10 */       110,        87,        70,        56,        45,
 821 /*  15 */        36,        29,        23,        18,        15,
 822};
 823
 824/*
 825 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
 826 *
 827 * In cases where the weight does not change often, we can use the
 828 * precalculated inverse to speed up arithmetics by turning divisions
 829 * into multiplications:
 830 */
 831static const u32 prio_to_wmult[40] = {
 832 /* -20 */     48388,     59856,     76040,     92818,    118348,
 833 /* -15 */    147320,    184698,    229616,    287308,    360437,
 834 /* -10 */    449829,    563644,    704093,    875809,   1099582,
 835 /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
 836 /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
 837 /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
 838 /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
 839 /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
 840};
 841
 842static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
 843
 844/*
 845 * runqueue iterator, to support SMP load-balancing between different
 846 * scheduling classes, without having to expose their internal data
 847 * structures to the load-balancing proper:
 848 */
 849struct rq_iterator {
 850        void *arg;
 851        struct task_struct *(*start)(void *);
 852        struct task_struct *(*next)(void *);
 853};
 854
 855#ifdef CONFIG_SMP
 856static unsigned long
 857balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
 858              unsigned long max_load_move, struct sched_domain *sd,
 859              enum cpu_idle_type idle, int *all_pinned,
 860              int *this_best_prio, struct rq_iterator *iterator);
 861
 862static int
 863iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
 864                   struct sched_domain *sd, enum cpu_idle_type idle,
 865                   struct rq_iterator *iterator);
 866#endif
 867
 868#ifdef CONFIG_CGROUP_CPUACCT
 869static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
 870#else
 871static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
 872#endif
 873
 874#include "sched_stats.h"
 875#include "sched_idletask.c"
 876#include "sched_fair.c"
 877#include "sched_rt.c"
 878#ifdef CONFIG_SCHED_DEBUG
 879# include "sched_debug.c"
 880#endif
 881
 882#define sched_class_highest (&rt_sched_class)
 883
 884/*
 885 * Update delta_exec, delta_fair fields for rq.
 886 *
 887 * delta_fair clock advances at a rate inversely proportional to
 888 * total load (rq->load.weight) on the runqueue, while
 889 * delta_exec advances at the same rate as wall-clock (provided
 890 * cpu is not idle).
 891 *
 892 * delta_exec / delta_fair is a measure of the (smoothened) load on this
 893 * runqueue over any given interval. This (smoothened) load is used
 894 * during load balance.
 895 *
 896 * This function is called /before/ updating rq->load
 897 * and when switching tasks.
 898 */
 899static inline void inc_load(struct rq *rq, const struct task_struct *p)
 900{
 901        update_load_add(&rq->load, p->se.load.weight);
 902}
 903
 904static inline void dec_load(struct rq *rq, const struct task_struct *p)
 905{
 906        update_load_sub(&rq->load, p->se.load.weight);
 907}
 908
 909static void inc_nr_running(struct task_struct *p, struct rq *rq)
 910{
 911        rq->nr_running++;
 912        inc_load(rq, p);
 913}
 914
 915static void dec_nr_running(struct task_struct *p, struct rq *rq)
 916{
 917        rq->nr_running--;
 918        dec_load(rq, p);
 919}
 920
 921static void set_load_weight(struct task_struct *p)
 922{
 923        if (task_has_rt_policy(p)) {
 924                p->se.load.weight = prio_to_weight[0] * 2;
 925                p->se.load.inv_weight = prio_to_wmult[0] >> 1;
 926                return;
 927        }
 928
 929        /*
 930         * SCHED_IDLE tasks get minimal weight:
 931         */
 932        if (p->policy == SCHED_IDLE) {
 933                p->se.load.weight = WEIGHT_IDLEPRIO;
 934                p->se.load.inv_weight = WMULT_IDLEPRIO;
 935                return;
 936        }
 937
 938        p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
 939        p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
 940}
 941
 942static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
 943{
 944        sched_info_queued(p);
 945        p->sched_class->enqueue_task(rq, p, wakeup);
 946        p->se.on_rq = 1;
 947}
 948
 949static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
 950{
 951        p->sched_class->dequeue_task(rq, p, sleep);
 952        p->se.on_rq = 0;
 953}
 954
 955/*
 956 * __normal_prio - return the priority that is based on the static prio
 957 */
 958static inline int __normal_prio(struct task_struct *p)
 959{
 960        return p->static_prio;
 961}
 962
 963/*
 964 * Calculate the expected normal priority: i.e. priority
 965 * without taking RT-inheritance into account. Might be
 966 * boosted by interactivity modifiers. Changes upon fork,
 967 * setprio syscalls, and whenever the interactivity
 968 * estimator recalculates.
 969 */
 970static inline int normal_prio(struct task_struct *p)
 971{
 972        int prio;
 973
 974        if (task_has_rt_policy(p))
 975                prio = MAX_RT_PRIO-1 - p->rt_priority;
 976        else
 977                prio = __normal_prio(p);
 978        return prio;
 979}
 980
 981/*
 982 * Calculate the current priority, i.e. the priority
 983 * taken into account by the scheduler. This value might
 984 * be boosted by RT tasks, or might be boosted by
 985 * interactivity modifiers. Will be RT if the task got
 986 * RT-boosted. If not then it returns p->normal_prio.
 987 */
 988static int effective_prio(struct task_struct *p)
 989{
 990        p->normal_prio = normal_prio(p);
 991        /*
 992         * If we are RT tasks or we were boosted to RT priority,
 993         * keep the priority unchanged. Otherwise, update priority
 994         * to the normal priority:
 995         */
 996        if (!rt_prio(p->prio))
 997                return p->normal_prio;
 998        return p->prio;
 999}
1000
1001/*
1002 * activate_task - move a task to the runqueue.
1003 */
1004static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1005{
1006        if (p->state == TASK_UNINTERRUPTIBLE)
1007                rq->nr_uninterruptible--;
1008
1009        enqueue_task(rq, p, wakeup);
1010        inc_nr_running(p, rq);
1011}
1012
1013/*
1014 * deactivate_task - remove a task from the runqueue.
1015 */
1016static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1017{
1018        if (p->state == TASK_UNINTERRUPTIBLE)
1019                rq->nr_uninterruptible++;
1020
1021        dequeue_task(rq, p, sleep);
1022        dec_nr_running(p, rq);
1023}
1024
1025/**
1026 * task_curr - is this task currently executing on a CPU?
1027 * @p: the task in question.
1028 */
1029inline int task_curr(const struct task_struct *p)
1030{
1031        return cpu_curr(task_cpu(p)) == p;
1032}
1033
1034/* Used instead of source_load when we know the type == 0 */
1035unsigned long weighted_cpuload(const int cpu)
1036{
1037        return cpu_rq(cpu)->load.weight;
1038}
1039
1040static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1041{
1042        set_task_cfs_rq(p, cpu);
1043#ifdef CONFIG_SMP
1044        /*
1045         * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1046         * successfuly executed on another CPU. We must ensure that updates of
1047         * per-task data have been completed by this moment.
1048         */
1049        smp_wmb();
1050        task_thread_info(p)->cpu = cpu;
1051#endif
1052}
1053
1054#ifdef CONFIG_SMP
1055
1056/*
1057 * Is this task likely cache-hot:
1058 */
1059static inline int
1060task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1061{
1062        s64 delta;
1063
1064        if (p->sched_class != &fair_sched_class)
1065                return 0;
1066
1067        if (sysctl_sched_migration_cost == -1)
1068                return 1;
1069        if (sysctl_sched_migration_cost == 0)
1070                return 0;
1071
1072        delta = now - p->se.exec_start;
1073
1074        return delta < (s64)sysctl_sched_migration_cost;
1075}
1076
1077
1078void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1079{
1080        int old_cpu = task_cpu(p);
1081        struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1082        struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1083                      *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1084        u64 clock_offset;
1085
1086        clock_offset = old_rq->clock - new_rq->clock;
1087
1088#ifdef CONFIG_SCHEDSTATS
1089        if (p->se.wait_start)
1090                p->se.wait_start -= clock_offset;
1091        if (p->se.sleep_start)
1092                p->se.sleep_start -= clock_offset;
1093        if (p->se.block_start)
1094                p->se.block_start -= clock_offset;
1095        if (old_cpu != new_cpu) {
1096                schedstat_inc(p, se.nr_migrations);
1097                if (task_hot(p, old_rq->clock, NULL))
1098                        schedstat_inc(p, se.nr_forced2_migrations);
1099        }
1100#endif
1101        p->se.vruntime -= old_cfsrq->min_vruntime -
1102                                         new_cfsrq->min_vruntime;
1103
1104        __set_task_cpu(p, new_cpu);
1105}
1106
1107struct migration_req {
1108        struct list_head list;
1109
1110        struct task_struct *task;
1111        int dest_cpu;
1112
1113        struct completion done;
1114};
1115
1116/*
1117 * The task's runqueue lock must be held.
1118 * Returns true if you have to wait for migration thread.
1119 */
1120static int
1121migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1122{
1123        struct rq *rq = task_rq(p);
1124
1125        /*
1126         * If the task is not on a runqueue (and not running), then
1127         * it is sufficient to simply update the task's cpu field.
1128         */
1129        if (!p->se.on_rq && !task_running(rq, p)) {
1130                set_task_cpu(p, dest_cpu);
1131                return 0;
1132        }
1133
1134        init_completion(&req->done);
1135        req->task = p;
1136        req->dest_cpu = dest_cpu;
1137        list_add(&req->list, &rq->migration_queue);
1138
1139        return 1;
1140}
1141
1142/*
1143 * wait_task_inactive - wait for a thread to unschedule.
1144 *
1145 * The caller must ensure that the task *will* unschedule sometime soon,
1146 * else this function might spin for a *long* time. This function can't
1147 * be called with interrupts off, or it may introduce deadlock with
1148 * smp_call_function() if an IPI is sent by the same process we are
1149 * waiting to become inactive.
1150 */
1151void wait_task_inactive(struct task_struct *p)
1152{
1153        unsigned long flags;
1154        int running, on_rq;
1155        struct rq *rq;
1156
1157        for (;;) {
1158                /*
1159                 * We do the initial early heuristics without holding
1160                 * any task-queue locks at all. We'll only try to get
1161                 * the runqueue lock when things look like they will
1162                 * work out!
1163                 */
1164                rq = task_rq(p);
1165
1166                /*
1167                 * If the task is actively running on another CPU
1168                 * still, just relax and busy-wait without holding
1169                 * any locks.
1170                 *
1171                 * NOTE! Since we don't hold any locks, it's not
1172                 * even sure that "rq" stays as the right runqueue!
1173                 * But we don't care, since "task_running()" will
1174                 * return false if the runqueue has changed and p
1175                 * is actually now running somewhere else!
1176                 */
1177                while (task_running(rq, p))
1178                        cpu_relax();
1179
1180                /*
1181                 * Ok, time to look more closely! We need the rq
1182                 * lock now, to be *sure*. If we're wrong, we'll
1183                 * just go back and repeat.
1184                 */
1185                rq = task_rq_lock(p, &flags);
1186                running = task_running(rq, p);
1187                on_rq = p->se.on_rq;
1188                task_rq_unlock(rq, &flags);
1189
1190                /*
1191                 * Was it really running after all now that we
1192                 * checked with the proper locks actually held?
1193                 *
1194                 * Oops. Go back and try again..
1195                 */
1196                if (unlikely(running)) {
1197                        cpu_relax();
1198                        continue;
1199                }
1200
1201                /*
1202                 * It's not enough that it's not actively running,
1203                 * it must be off the runqueue _entirely_, and not
1204                 * preempted!
1205                 *
1206                 * So if it wa still runnable (but just not actively
1207                 * running right now), it's preempted, and we should
1208                 * yield - it could be a while.
1209                 */
1210                if (unlikely(on_rq)) {
1211                        schedule_timeout_uninterruptible(1);
1212                        continue;
1213                }
1214
1215                /*
1216                 * Ahh, all good. It wasn't running, and it wasn't
1217                 * runnable, which means that it will never become
1218                 * running in the future either. We're all done!
1219                 */
1220                break;
1221        }
1222}
1223
1224/***
1225 * kick_process - kick a running thread to enter/exit the kernel
1226 * @p: the to-be-kicked thread
1227 *
1228 * Cause a process which is running on another CPU to enter
1229 * kernel-mode, without any delay. (to get signals handled.)
1230 *
1231 * NOTE: this function doesnt have to take the runqueue lock,
1232 * because all it wants to ensure is that the remote task enters
1233 * the kernel. If the IPI races and the task has been migrated
1234 * to another CPU then no harm is done and the purpose has been
1235 * achieved as well.
1236 */
1237void kick_process(struct task_struct *p)
1238{
1239        int cpu;
1240
1241        preempt_disable();
1242        cpu = task_cpu(p);
1243        if ((cpu != smp_processor_id()) && task_curr(p))
1244                smp_send_reschedule(cpu);
1245        preempt_enable();
1246}
1247
1248/*
1249 * Return a low guess at the load of a migration-source cpu weighted
1250 * according to the scheduling class and "nice" value.
1251 *
1252 * We want to under-estimate the load of migration sources, to
1253 * balance conservatively.
1254 */
1255static unsigned long source_load(int cpu, int type)
1256{
1257        struct rq *rq = cpu_rq(cpu);
1258        unsigned long total = weighted_cpuload(cpu);
1259
1260        if (type == 0)
1261                return total;
1262
1263        return min(rq->cpu_load[type-1], total);
1264}
1265
1266/*
1267 * Return a high guess at the load of a migration-target cpu weighted
1268 * according to the scheduling class and "nice" value.
1269 */
1270static unsigned long target_load(int cpu, int type)
1271{
1272        struct rq *rq = cpu_rq(cpu);
1273        unsigned long total = weighted_cpuload(cpu);
1274
1275        if (type == 0)
1276                return total;
1277
1278        return max(rq->cpu_load[type-1], total);
1279}
1280
1281/*
1282 * Return the average load per task on the cpu's run queue
1283 */
1284static inline unsigned long cpu_avg_load_per_task(int cpu)
1285{
1286        struct rq *rq = cpu_rq(cpu);
1287        unsigned long total = weighted_cpuload(cpu);
1288        unsigned long n = rq->nr_running;
1289
1290        return n ? total / n : SCHED_LOAD_SCALE;
1291}
1292
1293/*
1294 * find_idlest_group finds and returns the least busy CPU group within the
1295 * domain.
1296 */
1297static struct sched_group *
1298find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1299{
1300        struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1301        unsigned long min_load = ULONG_MAX, this_load = 0;
1302        int load_idx = sd->forkexec_idx;
1303        int imbalance = 100 + (sd->imbalance_pct-100)/2;
1304
1305        do {
1306                unsigned long load, avg_load;
1307                int local_group;
1308                int i;
1309
1310                /* Skip over this group if it has no CPUs allowed */
1311                if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1312                        continue;
1313
1314                local_group = cpu_isset(this_cpu, group->cpumask);
1315
1316                /* Tally up the load of all CPUs in the group */
1317                avg_load = 0;
1318
1319                for_each_cpu_mask(i, group->cpumask) {
1320                        /* Bias balancing toward cpus of our domain */
1321                        if (local_group)
1322                                load = source_load(i, load_idx);
1323                        else
1324                                load = target_load(i, load_idx);
1325
1326                        avg_load += load;
1327                }
1328
1329                /* Adjust by relative CPU power of the group */
1330                avg_load = sg_div_cpu_power(group,
1331                                avg_load * SCHED_LOAD_SCALE);
1332
1333                if (local_group) {
1334                        this_load = avg_load;
1335                        this = group;
1336                } else if (avg_load < min_load) {
1337                        min_load = avg_load;
1338                        idlest = group;
1339                }
1340        } while (group = group->next, group != sd->groups);
1341
1342        if (!idlest || 100*this_load < imbalance*min_load)
1343                return NULL;
1344        return idlest;
1345}
1346
1347/*
1348 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1349 */
1350static int
1351find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1352{
1353        cpumask_t tmp;
1354        unsigned long load, min_load = ULONG_MAX;
1355        int idlest = -1;
1356        int i;
1357
1358        /* Traverse only the allowed CPUs */
1359        cpus_and(tmp, group->cpumask, p->cpus_allowed);
1360
1361        for_each_cpu_mask(i, tmp) {
1362                load = weighted_cpuload(i);
1363
1364                if (load < min_load || (load == min_load && i == this_cpu)) {
1365                        min_load = load;
1366                        idlest = i;
1367                }
1368        }
1369
1370        return idlest;
1371}
1372
1373/*
1374 * sched_balance_self: balance the current task (running on cpu) in domains
1375 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1376 * SD_BALANCE_EXEC.
1377 *
1378 * Balance, ie. select the least loaded group.
1379 *
1380 * Returns the target CPU number, or the same CPU if no balancing is needed.
1381 *
1382 * preempt must be disabled.
1383 */
1384static int sched_balance_self(int cpu, int flag)
1385{
1386        struct task_struct *t = current;
1387        struct sched_domain *tmp, *sd = NULL;
1388
1389        for_each_domain(cpu, tmp) {
1390                /*
1391                 * If power savings logic is enabled for a domain, stop there.
1392                 */
1393                if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1394                        break;
1395                if (tmp->flags & flag)
1396                        sd = tmp;
1397        }
1398
1399        while (sd) {
1400                cpumask_t span;
1401                struct sched_group *group;
1402                int new_cpu, weight;
1403
1404                if (!(sd->flags & flag)) {
1405                        sd = sd->child;
1406                        continue;
1407                }
1408
1409                span = sd->span;
1410                group = find_idlest_group(sd, t, cpu);
1411                if (!group) {
1412                        sd = sd->child;
1413                        continue;
1414                }
1415
1416                new_cpu = find_idlest_cpu(group, t, cpu);
1417                if (new_cpu == -1 || new_cpu == cpu) {
1418                        /* Now try balancing at a lower domain level of cpu */
1419                        sd = sd->child;
1420                        continue;
1421                }
1422
1423                /* Now try balancing at a lower domain level of new_cpu */
1424                cpu = new_cpu;
1425                sd = NULL;
1426                weight = cpus_weight(span);
1427                for_each_domain(cpu, tmp) {
1428                        if (weight <= cpus_weight(tmp->span))
1429                                break;
1430                        if (tmp->flags & flag)
1431                                sd = tmp;
1432                }
1433                /* while loop will break here if sd == NULL */
1434        }
1435
1436        return cpu;
1437}
1438
1439#endif /* CONFIG_SMP */
1440
1441/*
1442 * wake_idle() will wake a task on an idle cpu if task->cpu is
1443 * not idle and an idle cpu is available.  The span of cpus to
1444 * search starts with cpus closest then further out as needed,
1445 * so we always favor a closer, idle cpu.
1446 *
1447 * Returns the CPU we should wake onto.
1448 */
1449#if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1450static int wake_idle(int cpu, struct task_struct *p)
1451{
1452        cpumask_t tmp;
1453        struct sched_domain *sd;
1454        int i;
1455
1456        /*
1457         * If it is idle, then it is the best cpu to run this task.
1458         *
1459         * This cpu is also the best, if it has more than one task already.
1460         * Siblings must be also busy(in most cases) as they didn't already
1461         * pickup the extra load from this cpu and hence we need not check
1462         * sibling runqueue info. This will avoid the checks and cache miss
1463         * penalities associated with that.
1464         */
1465        if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
1466                return cpu;
1467
1468        for_each_domain(cpu, sd) {
1469                if (sd->flags & SD_WAKE_IDLE) {
1470                        cpus_and(tmp, sd->span, p->cpus_allowed);
1471                        for_each_cpu_mask(i, tmp) {
1472                                if (idle_cpu(i)) {
1473                                        if (i != task_cpu(p)) {
1474                                                schedstat_inc(p,
1475                                                        se.nr_wakeups_idle);
1476                                        }
1477                                        return i;
1478                                }
1479                        }
1480                } else {
1481                        break;
1482                }
1483        }
1484        return cpu;
1485}
1486#else
1487static inline int wake_idle(int cpu, struct task_struct *p)
1488{
1489        return cpu;
1490}
1491#endif
1492
1493/***
1494 * try_to_wake_up - wake up a thread
1495 * @p: the to-be-woken-up thread
1496 * @state: the mask of task states that can be woken
1497 * @sync: do a synchronous wakeup?
1498 *
1499 * Put it on the run-queue if it's not already there. The "current"
1500 * thread is always on the run-queue (except when the actual
1501 * re-schedule is in progress), and as such you're allowed to do
1502 * the simpler "current->state = TASK_RUNNING" to mark yourself
1503 * runnable without the overhead of this.
1504 *
1505 * returns failure only if the task is already active.
1506 */
1507static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1508{
1509        int cpu, orig_cpu, this_cpu, success = 0;
1510        unsigned long flags;
1511        long old_state;
1512        struct rq *rq;
1513#ifdef CONFIG_SMP
1514        struct sched_domain *sd, *this_sd = NULL;
1515        unsigned long load, this_load;
1516        int new_cpu;
1517#endif
1518
1519        rq = task_rq_lock(p, &flags);
1520        old_state = p->state;
1521        if (!(old_state & state))
1522                goto out;
1523
1524        if (p->se.on_rq)
1525                goto out_running;
1526
1527        cpu = task_cpu(p);
1528        orig_cpu = cpu;
1529        this_cpu = smp_processor_id();
1530
1531#ifdef CONFIG_SMP
1532        if (unlikely(task_running(rq, p)))
1533                goto out_activate;
1534
1535        new_cpu = cpu;
1536
1537        schedstat_inc(rq, ttwu_count);
1538        if (cpu == this_cpu) {
1539                schedstat_inc(rq, ttwu_local);
1540                goto out_set_cpu;
1541        }
1542
1543        for_each_domain(this_cpu, sd) {
1544                if (cpu_isset(cpu, sd->span)) {
1545                        schedstat_inc(sd, ttwu_wake_remote);
1546                        this_sd = sd;
1547                        break;
1548                }
1549        }
1550
1551        if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1552                goto out_set_cpu;
1553
1554        /*
1555         * Check for affine wakeup and passive balancing possibilities.
1556         */
1557        if (this_sd) {
1558                int idx = this_sd->wake_idx;
1559                unsigned int imbalance;
1560
1561                imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1562
1563                load = source_load(cpu, idx);
1564                this_load = target_load(this_cpu, idx);
1565
1566                new_cpu = this_cpu; /* Wake to this CPU if we can */
1567
1568                if (this_sd->flags & SD_WAKE_AFFINE) {
1569                        unsigned long tl = this_load;
1570                        unsigned long tl_per_task;
1571
1572                        /*
1573                         * Attract cache-cold tasks on sync wakeups:
1574                         */
1575                        if (sync && !task_hot(p, rq->clock, this_sd))
1576                                goto out_set_cpu;
1577
1578                        schedstat_inc(p, se.nr_wakeups_affine_attempts);
1579                        tl_per_task = cpu_avg_load_per_task(this_cpu);
1580
1581                        /*
1582                         * If sync wakeup then subtract the (maximum possible)
1583                         * effect of the currently running task from the load
1584                         * of the current CPU:
1585                         */
1586                        if (sync)
1587                                tl -= current->se.load.weight;
1588
1589                        if ((tl <= load &&
1590                                tl + target_load(cpu, idx) <= tl_per_task) ||
1591                               100*(tl + p->se.load.weight) <= imbalance*load) {
1592                                /*
1593                                 * This domain has SD_WAKE_AFFINE and
1594                                 * p is cache cold in this domain, and
1595                                 * there is no bad imbalance.
1596                                 */
1597                                schedstat_inc(this_sd, ttwu_move_affine);
1598                                schedstat_inc(p, se.nr_wakeups_affine);
1599                                goto out_set_cpu;
1600                        }
1601                }
1602
1603                /*
1604                 * Start passive balancing when half the imbalance_pct
1605                 * limit is reached.
1606                 */
1607                if (this_sd->flags & SD_WAKE_BALANCE) {
1608                        if (imbalance*this_load <= 100*load) {
1609                                schedstat_inc(this_sd, ttwu_move_balance);
1610                                schedstat_inc(p, se.nr_wakeups_passive);
1611                                goto out_set_cpu;
1612                        }
1613                }
1614        }
1615
1616        new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1617out_set_cpu:
1618        new_cpu = wake_idle(new_cpu, p);
1619        if (new_cpu != cpu) {
1620                set_task_cpu(p, new_cpu);
1621                task_rq_unlock(rq, &flags);
1622                /* might preempt at this point */
1623                rq = task_rq_lock(p, &flags);
1624                old_state = p->state;
1625                if (!(old_state & state))
1626                        goto out;
1627                if (p->se.on_rq)
1628                        goto out_running;
1629
1630                this_cpu = smp_processor_id();
1631                cpu = task_cpu(p);
1632        }
1633
1634out_activate:
1635#endif /* CONFIG_SMP */
1636        schedstat_inc(p, se.nr_wakeups);
1637        if (sync)
1638                schedstat_inc(p, se.nr_wakeups_sync);
1639        if (orig_cpu != cpu)
1640                schedstat_inc(p, se.nr_wakeups_migrate);
1641        if (cpu == this_cpu)
1642                schedstat_inc(p, se.nr_wakeups_local);
1643        else
1644                schedstat_inc(p, se.nr_wakeups_remote);
1645        update_rq_clock(rq);
1646        activate_task(rq, p, 1);
1647        check_preempt_curr(rq, p);
1648        success = 1;
1649
1650out_running:
1651        p->state = TASK_RUNNING;
1652out:
1653        task_rq_unlock(rq, &flags);
1654
1655        return success;
1656}
1657
1658int fastcall wake_up_process(struct task_struct *p)
1659{
1660        return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1661                                 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1662}
1663EXPORT_SYMBOL(wake_up_process);
1664
1665int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1666{
1667        return try_to_wake_up(p, state, 0);
1668}
1669
1670/*
1671 * Perform scheduler related setup for a newly forked process p.
1672 * p is forked by current.
1673 *
1674 * __sched_fork() is basic setup used by init_idle() too:
1675 */
1676static void __sched_fork(struct task_struct *p)
1677{
1678        p->se.exec_start                = 0;
1679        p->se.sum_exec_runtime          = 0;
1680        p->se.prev_sum_exec_runtime     = 0;
1681
1682#ifdef CONFIG_SCHEDSTATS
1683        p->se.wait_start                = 0;
1684        p->se.sum_sleep_runtime         = 0;
1685        p->se.sleep_start               = 0;
1686        p->se.block_start               = 0;
1687        p->se.sleep_max                 = 0;
1688        p->se.block_max                 = 0;
1689        p->se.exec_max                  = 0;
1690        p->se.slice_max                 = 0;
1691        p->se.wait_max                  = 0;
1692#endif
1693
1694        INIT_LIST_HEAD(&p->run_list);
1695        p->se.on_rq = 0;
1696
1697#ifdef CONFIG_PREEMPT_NOTIFIERS
1698        INIT_HLIST_HEAD(&p->preempt_notifiers);
1699#endif
1700
1701        /*
1702         * We mark the process as running here, but have not actually
1703         * inserted it onto the runqueue yet. This guarantees that
1704         * nobody will actually run it, and a signal or other external
1705         * event cannot wake it up and insert it on the runqueue either.
1706         */
1707        p->state = TASK_RUNNING;
1708}
1709
1710/*
1711 * fork()/clone()-time setup:
1712 */
1713void sched_fork(struct task_struct *p, int clone_flags)
1714{
1715        int cpu = get_cpu();
1716
1717        __sched_fork(p);
1718
1719#ifdef CONFIG_SMP
1720        cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1721#endif
1722        set_task_cpu(p, cpu);
1723
1724        /*
1725         * Make sure we do not leak PI boosting priority to the child:
1726         */
1727        p->prio = current->normal_prio;
1728        if (!rt_prio(p->prio))
1729                p->sched_class = &fair_sched_class;
1730
1731#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1732        if (likely(sched_info_on()))
1733                memset(&p->sched_info, 0, sizeof(p->sched_info));
1734#endif
1735#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1736        p->oncpu = 0;
1737#endif
1738#ifdef CONFIG_PREEMPT
1739        /* Want to start with kernel preemption disabled. */
1740        task_thread_info(p)->preempt_count = 1;
1741#endif
1742        put_cpu();
1743}
1744
1745/*
1746 * wake_up_new_task - wake up a newly created task for the first time.
1747 *
1748 * This function will do some initial scheduler statistics housekeeping
1749 * that must be done for every newly created context, then puts the task
1750 * on the runqueue and wakes it.
1751 */
1752void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1753{
1754        unsigned long flags;
1755        struct rq *rq;
1756
1757        rq = task_rq_lock(p, &flags);
1758        BUG_ON(p->state != TASK_RUNNING);
1759        update_rq_clock(rq);
1760
1761        p->prio = effective_prio(p);
1762
1763        if (!p->sched_class->task_new || !current->se.on_rq) {
1764                activate_task(rq, p, 0);
1765        } else {
1766                /*
1767                 * Let the scheduling class do new task startup
1768                 * management (if any):
1769                 */
1770                p->sched_class->task_new(rq, p);
1771                inc_nr_running(p, rq);
1772        }
1773        check_preempt_curr(rq, p);
1774        task_rq_unlock(rq, &flags);
1775}
1776
1777#ifdef CONFIG_PREEMPT_NOTIFIERS
1778
1779/**
1780 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1781 * @notifier: notifier struct to register
1782 */
1783void preempt_notifier_register(struct preempt_notifier *notifier)
1784{
1785        hlist_add_head(&notifier->link, &current->preempt_notifiers);
1786}
1787EXPORT_SYMBOL_GPL(preempt_notifier_register);
1788
1789/**
1790 * preempt_notifier_unregister - no longer interested in preemption notifications
1791 * @notifier: notifier struct to unregister
1792 *
1793 * This is safe to call from within a preemption notifier.
1794 */
1795void preempt_notifier_unregister(struct preempt_notifier *notifier)
1796{
1797        hlist_del(&notifier->link);
1798}
1799EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1800
1801static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1802{
1803        struct preempt_notifier *notifier;
1804        struct hlist_node *node;
1805
1806        hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1807                notifier->ops->sched_in(notifier, raw_smp_processor_id());
1808}
1809
1810static void
1811fire_sched_out_preempt_notifiers(struct task_struct *curr,
1812                                 struct task_struct *next)
1813{
1814        struct preempt_notifier *notifier;
1815        struct hlist_node *node;
1816
1817        hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1818                notifier->ops->sched_out(notifier, next);
1819}
1820
1821#else
1822
1823static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1824{
1825}
1826
1827static void
1828fire_sched_out_preempt_notifiers(struct task_struct *curr,
1829                                 struct task_struct *next)
1830{
1831}
1832
1833#endif
1834
1835/**
1836 * prepare_task_switch - prepare to switch tasks
1837 * @rq: the runqueue preparing to switch
1838 * @prev: the current task that is being switched out
1839 * @next: the task we are going to switch to.
1840 *
1841 * This is called with the rq lock held and interrupts off. It must
1842 * be paired with a subsequent finish_task_switch after the context
1843 * switch.
1844 *
1845 * prepare_task_switch sets up locking and calls architecture specific
1846 * hooks.
1847 */
1848static inline void
1849prepare_task_switch(struct rq *rq, struct task_struct *prev,
1850                    struct task_struct *next)
1851{
1852        fire_sched_out_preempt_notifiers(prev, next);
1853        prepare_lock_switch(rq, next);
1854        prepare_arch_switch(next);
1855}
1856
1857/**
1858 * finish_task_switch - clean up after a task-switch
1859 * @rq: runqueue associated with task-switch
1860 * @prev: the thread we just switched away from.
1861 *
1862 * finish_task_switch must be called after the context switch, paired
1863 * with a prepare_task_switch call before the context switch.
1864 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1865 * and do any other architecture-specific cleanup actions.
1866 *
1867 * Note that we may have delayed dropping an mm in context_switch(). If
1868 * so, we finish that here outside of the runqueue lock. (Doing it
1869 * with the lock held can cause deadlocks; see schedule() for
1870 * details.)
1871 */
1872static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1873        __releases(rq->lock)
1874{
1875        struct mm_struct *mm = rq->prev_mm;
1876        long prev_state;
1877
1878        rq->prev_mm = NULL;
1879
1880        /*
1881         * A task struct has one reference for the use as "current".
1882         * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1883         * schedule one last time. The schedule call will never return, and
1884         * the scheduled task must drop that reference.
1885         * The test for TASK_DEAD must occur while the runqueue locks are
1886         * still held, otherwise prev could be scheduled on another cpu, die
1887         * there before we look at prev->state, and then the reference would
1888         * be dropped twice.
1889         *              Manfred Spraul <manfred@colorfullife.com>
1890         */
1891        prev_state = prev->state;
1892        finish_arch_switch(prev);
1893        finish_lock_switch(rq, prev);
1894        fire_sched_in_preempt_notifiers(current);
1895        if (mm)
1896                mmdrop(mm);
1897        if (unlikely(prev_state == TASK_DEAD)) {
1898                /*
1899                 * Remove function-return probe instances associated with this
1900                 * task and put them back on the free list.
1901                 */
1902                kprobe_flush_task(prev);
1903                put_task_struct(prev);
1904        }
1905}
1906
1907/**
1908 * schedule_tail - first thing a freshly forked thread must call.
1909 * @prev: the thread we just switched away from.
1910 */
1911asmlinkage void schedule_tail(struct task_struct *prev)
1912        __releases(rq->lock)
1913{
1914        struct rq *rq = this_rq();
1915
1916        finish_task_switch(rq, prev);
1917#ifdef __ARCH_WANT_UNLOCKED_CTXSW
1918        /* In this case, finish_task_switch does not reenable preemption */
1919        preempt_enable();
1920#endif
1921        if (current->set_child_tid)
1922                put_user(task_pid_vnr(current), current->set_child_tid);
1923}
1924
1925/*
1926 * context_switch - switch to the new MM and the new
1927 * thread's register state.
1928 */
1929static inline void
1930context_switch(struct rq *rq, struct task_struct *prev,
1931               struct task_struct *next)
1932{
1933        struct mm_struct *mm, *oldmm;
1934
1935        prepare_task_switch(rq, prev, next);
1936        mm = next->mm;
1937        oldmm = prev->active_mm;
1938        /*
1939         * For paravirt, this is coupled with an exit in switch_to to
1940         * combine the page table reload and the switch backend into
1941         * one hypercall.
1942         */
1943        arch_enter_lazy_cpu_mode();
1944
1945        if (unlikely(!mm)) {
1946                next->active_mm = oldmm;
1947                atomic_inc(&oldmm->mm_count);
1948                enter_lazy_tlb(oldmm, next);
1949        } else
1950                switch_mm(oldmm, mm, next);
1951
1952        if (unlikely(!prev->mm)) {
1953                prev->active_mm = NULL;
1954                rq->prev_mm = oldmm;
1955        }
1956        /*
1957         * Since the runqueue lock will be released by the next
1958         * task (which is an invalid locking op but in the case
1959         * of the scheduler it's an obvious special-case), so we
1960         * do an early lockdep release here:
1961         */
1962#ifndef __ARCH_WANT_UNLOCKED_CTXSW
1963        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1964#endif
1965
1966        /* Here we just switch the register state and the stack. */
1967        switch_to(prev, next, prev);
1968
1969        barrier();
1970        /*
1971         * this_rq must be evaluated again because prev may have moved
1972         * CPUs since it called schedule(), thus the 'rq' on its stack
1973         * frame will be invalid.
1974         */
1975        finish_task_switch(this_rq(), prev);
1976}
1977
1978/*
1979 * nr_running, nr_uninterruptible and nr_context_switches:
1980 *
1981 * externally visible scheduler statistics: current number of runnable
1982 * threads, current number of uninterruptible-sleeping threads, total
1983 * number of context switches performed since bootup.
1984 */
1985unsigned long nr_running(void)
1986{
1987        unsigned long i, sum = 0;
1988
1989        for_each_online_cpu(i)
1990                sum += cpu_rq(i)->nr_running;
1991
1992        return sum;
1993}
1994
1995unsigned long nr_uninterruptible(void)
1996{
1997        unsigned long i, sum = 0;
1998
1999        for_each_possible_cpu(i)
2000                sum += cpu_rq(i)->nr_uninterruptible;
2001
2002        /*
2003         * Since we read the counters lockless, it might be slightly
2004         * inaccurate. Do not allow it to go below zero though:
2005         */
2006        if (unlikely((long)sum < 0))
2007                sum = 0;
2008
2009        return sum;
2010}
2011
2012unsigned long long nr_context_switches(void)
2013{
2014        int i;
2015        unsigned long long sum = 0;
2016
2017        for_each_possible_cpu(i)
2018                sum += cpu_rq(i)->nr_switches;
2019
2020        return sum;
2021}
2022
2023unsigned long nr_iowait(void)
2024{
2025        unsigned long i, sum = 0;
2026
2027        for_each_possible_cpu(i)
2028                sum += atomic_read(&cpu_rq(i)->nr_iowait);
2029
2030        return sum;
2031}
2032
2033unsigned long nr_active(void)
2034{
2035        unsigned long i, running = 0, uninterruptible = 0;
2036
2037        for_each_online_cpu(i) {
2038                running += cpu_rq(i)->nr_running;
2039                uninterruptible += cpu_rq(i)->nr_uninterruptible;
2040        }
2041
2042        if (unlikely((long)uninterruptible < 0))
2043                uninterruptible = 0;
2044
2045        return running + uninterruptible;
2046}
2047
2048/*
2049 * Update rq->cpu_load[] statistics. This function is usually called every
2050 * scheduler tick (TICK_NSEC).
2051 */
2052static void update_cpu_load(struct rq *this_rq)
2053{
2054        unsigned long this_load = this_rq->load.weight;
2055        int i, scale;
2056
2057        this_rq->nr_load_updates++;
2058
2059        /* Update our load: */
2060        for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2061                unsigned long old_load, new_load;
2062
2063                /* scale is effectively 1 << i now, and >> i divides by scale */
2064
2065                old_load = this_rq->cpu_load[i];
2066                new_load = this_load;
2067                /*
2068                 * Round up the averaging division if load is increasing. This
2069                 * prevents us from getting stuck on 9 if the load is 10, for
2070                 * example.
2071                 */
2072                if (new_load > old_load)
2073                        new_load += scale-1;
2074                this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2075        }
2076}
2077
2078#ifdef CONFIG_SMP
2079
2080/*
2081 * double_rq_lock - safely lock two runqueues
2082 *
2083 * Note this does not disable interrupts like task_rq_lock,
2084 * you need to do so manually before calling.
2085 */
2086static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2087        __acquires(rq1->lock)
2088        __acquires(rq2->lock)
2089{
2090        BUG_ON(!irqs_disabled());
2091        if (rq1 == rq2) {
2092                spin_lock(&rq1->lock);
2093                __acquire(rq2->lock);   /* Fake it out ;) */
2094        } else {
2095                if (rq1 < rq2) {
2096                        spin_lock(&rq1->lock);
2097                        spin_lock(&rq2->lock);
2098                } else {
2099                        spin_lock(&rq2->lock);
2100                        spin_lock(&rq1->lock);
2101                }
2102        }
2103        update_rq_clock(rq1);
2104        update_rq_clock(rq2);
2105}
2106
2107/*
2108 * double_rq_unlock - safely unlock two runqueues
2109 *
2110 * Note this does not restore interrupts like task_rq_unlock,
2111 * you need to do so manually after calling.
2112 */
2113static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2114        __releases(rq1->lock)
2115        __releases(rq2->lock)
2116{
2117        spin_unlock(&rq1->lock);
2118        if (rq1 != rq2)
2119                spin_unlock(&rq2->lock);
2120        else
2121                __release(rq2->lock);
2122}
2123
2124/*
2125 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2126 */
2127static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
2128        __releases(this_rq->lock)
2129        __acquires(busiest->lock)
2130        __acquires(this_rq->lock)
2131{
2132        if (unlikely(!irqs_disabled())) {
2133                /* printk() doesn't work good under rq->lock */
2134                spin_unlock(&this_rq->lock);
2135                BUG_ON(1);
2136        }
2137        if (unlikely(!spin_trylock(&busiest->lock))) {
2138                if (busiest < this_rq) {
2139                        spin_unlock(&this_rq->lock);
2140                        spin_lock(&busiest->lock);
2141                        spin_lock(&this_rq->lock);
2142                } else
2143                        spin_lock(&busiest->lock);
2144        }
2145}
2146
2147/*
2148 * If dest_cpu is allowed for this process, migrate the task to it.
2149 * This is accomplished by forcing the cpu_allowed mask to only
2150 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2151 * the cpu_allowed mask is restored.
2152 */
2153static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2154{
2155        struct migration_req req;
2156        unsigned long flags;
2157        struct rq *rq;
2158
2159        rq = task_rq_lock(p, &flags);
2160        if (!cpu_isset(dest_cpu, p->cpus_allowed)
2161            || unlikely(cpu_is_offline(dest_cpu)))
2162                goto out;
2163
2164        /* force the process onto the specified CPU */
2165        if (migrate_task(p, dest_cpu, &req)) {
2166                /* Need to wait for migration thread (might exit: take ref). */
2167                struct task_struct *mt = rq->migration_thread;
2168
2169                get_task_struct(mt);
2170                task_rq_unlock(rq, &flags);
2171                wake_up_process(mt);
2172                put_task_struct(mt);
2173                wait_for_completion(&req.done);
2174
2175                return;
2176        }
2177out:
2178        task_rq_unlock(rq, &flags);
2179}
2180
2181/*
2182 * sched_exec - execve() is a valuable balancing opportunity, because at
2183 * this point the task has the smallest effective memory and cache footprint.
2184 */
2185void sched_exec(void)
2186{
2187        int new_cpu, this_cpu = get_cpu();
2188        new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2189        put_cpu();
2190        if (new_cpu != this_cpu)
2191                sched_migrate_task(current, new_cpu);
2192}
2193
2194/*
2195 * pull_task - move a task from a remote runqueue to the local runqueue.
2196 * Both runqueues must be locked.
2197 */
2198static void pull_task(struct rq *src_rq, struct task_struct *p,
2199                      struct rq *this_rq, int this_cpu)
2200{
2201        deactivate_task(src_rq, p, 0);
2202        set_task_cpu(p, this_cpu);
2203        activate_task(this_rq, p, 0);
2204        /*
2205         * Note that idle threads have a prio of MAX_PRIO, for this test
2206         * to be always true for them.
2207         */
2208        check_preempt_curr(this_rq, p);
2209}
2210
2211/*
2212 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2213 */
2214static
2215int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2216                     struct sched_domain *sd, enum cpu_idle_type idle,
2217                     int *all_pinned)
2218{
2219        /*
2220         * We do not migrate tasks that are:
2221         * 1) running (obviously), or
2222         * 2) cannot be migrated to this CPU due to cpus_allowed, or
2223         * 3) are cache-hot on their current CPU.
2224         */
2225        if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2226                schedstat_inc(p, se.nr_failed_migrations_affine);
2227                return 0;
2228        }
2229        *all_pinned = 0;
2230
2231        if (task_running(rq, p)) {
2232                schedstat_inc(p, se.nr_failed_migrations_running);
2233                return 0;
2234        }
2235
2236        /*
2237         * Aggressive migration if:
2238         * 1) task is cache cold, or
2239         * 2) too many balance attempts have failed.
2240         */
2241
2242        if (!task_hot(p, rq->clock, sd) ||
2243                        sd->nr_balance_failed > sd->cache_nice_tries) {
2244#ifdef CONFIG_SCHEDSTATS
2245                if (task_hot(p, rq->clock, sd)) {
2246                        schedstat_inc(sd, lb_hot_gained[idle]);
2247                        schedstat_inc(p, se.nr_forced_migrations);
2248                }
2249#endif
2250                return 1;
2251        }
2252
2253        if (task_hot(p, rq->clock, sd)) {
2254                schedstat_inc(p, se.nr_failed_migrations_hot);
2255                return 0;
2256        }
2257        return 1;
2258}
2259
2260static unsigned long
2261balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2262              unsigned long max_load_move, struct sched_domain *sd,
2263              enum cpu_idle_type idle, int *all_pinned,
2264              int *this_best_prio, struct rq_iterator *iterator)
2265{
2266        int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2267        struct task_struct *p;
2268        long rem_load_move = max_load_move;
2269
2270        if (max_load_move == 0)
2271                goto out;
2272
2273        pinned = 1;
2274
2275        /*
2276         * Start the load-balancing iterator:
2277         */
2278        p = iterator->start(iterator->arg);
2279next:
2280        if (!p || loops++ > sysctl_sched_nr_migrate)
2281                goto out;
2282        /*
2283         * To help distribute high priority tasks across CPUs we don't
2284         * skip a task if it will be the highest priority task (i.e. smallest
2285         * prio value) on its new queue regardless of its load weight
2286         */
2287        skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2288                                                         SCHED_LOAD_SCALE_FUZZ;
2289        if ((skip_for_load && p->prio >= *this_best_prio) ||
2290            !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2291                p = iterator->next(iterator->arg);
2292                goto next;
2293        }
2294
2295        pull_task(busiest, p, this_rq, this_cpu);
2296        pulled++;
2297        rem_load_move -= p->se.load.weight;
2298
2299        /*
2300         * We only want to steal up to the prescribed amount of weighted load.
2301         */
2302        if (rem_load_move > 0) {
2303                if (p->prio < *this_best_prio)
2304                        *this_best_prio = p->prio;
2305                p = iterator->next(iterator->arg);
2306                goto next;
2307        }
2308out:
2309        /*
2310         * Right now, this is one of only two places pull_task() is called,
2311         * so we can safely collect pull_task() stats here rather than
2312         * inside pull_task().
2313         */
2314        schedstat_add(sd, lb_gained[idle], pulled);
2315
2316        if (all_pinned)
2317                *all_pinned = pinned;
2318
2319        return max_load_move - rem_load_move;
2320}
2321
2322/*
2323 * move_tasks tries to move up to max_load_move weighted load from busiest to
2324 * this_rq, as part of a balancing operation within domain "sd".
2325 * Returns 1 if successful and 0 otherwise.
2326 *
2327 * Called with both runqueues locked.
2328 */
2329static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2330                      unsigned long max_load_move,
2331                      struct sched_domain *sd, enum cpu_idle_type idle,
2332                      int *all_pinned)
2333{
2334        const struct sched_class *class = sched_class_highest;
2335        unsigned long total_load_moved = 0;
2336        int this_best_prio = this_rq->curr->prio;
2337
2338        do {
2339                total_load_moved +=
2340                        class->load_balance(this_rq, this_cpu, busiest,
2341                                max_load_move - total_load_moved,
2342                                sd, idle, all_pinned, &this_best_prio);
2343                class = class->next;
2344        } while (class && max_load_move > total_load_moved);
2345
2346        return total_load_moved > 0;
2347}
2348
2349static int
2350iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2351                   struct sched_domain *sd, enum cpu_idle_type idle,
2352                   struct rq_iterator *iterator)
2353{
2354        struct task_struct *p = iterator->start(iterator->arg);
2355        int pinned = 0;
2356
2357        while (p) {
2358                if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2359                        pull_task(busiest, p, this_rq, this_cpu);
2360                        /*
2361                         * Right now, this is only the second place pull_task()
2362                         * is called, so we can safely collect pull_task()
2363                         * stats here rather than inside pull_task().
2364                         */
2365                        schedstat_inc(sd, lb_gained[idle]);
2366
2367                        return 1;
2368                }
2369                p = iterator->next(iterator->arg);
2370        }
2371
2372        return 0;
2373}
2374
2375/*
2376 * move_one_task tries to move exactly one task from busiest to this_rq, as
2377 * part of active balancing operations within "domain".
2378 * Returns 1 if successful and 0 otherwise.
2379 *
2380 * Called with both runqueues locked.
2381 */
2382static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2383                         struct sched_domain *sd, enum cpu_idle_type idle)
2384{
2385        const struct sched_class *class;
2386
2387        for (class = sched_class_highest; class; class = class->next)
2388                if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2389                        return 1;
2390
2391        return 0;
2392}
2393
2394/*
2395 * find_busiest_group finds and returns the busiest CPU group within the
2396 * domain. It calculates and returns the amount of weighted load which
2397 * should be moved to restore balance via the imbalance parameter.
2398 */
2399static struct sched_group *
2400find_busiest_group(struct sched_domain *sd, int this_cpu,
2401                   unsigned long *imbalance, enum cpu_idle_type idle,
2402                   int *sd_idle, cpumask_t *cpus, int *balance)
2403{
2404        struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2405        unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2406        unsigned long max_pull;
2407        unsigned long busiest_load_per_task, busiest_nr_running;
2408        unsigned long this_load_per_task, this_nr_running;
2409        int load_idx, group_imb = 0;
2410#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2411        int power_savings_balance = 1;
2412        unsigned long leader_nr_running = 0, min_load_per_task = 0;
2413        unsigned long min_nr_running = ULONG_MAX;
2414        struct sched_group *group_min = NULL, *group_leader = NULL;
2415#endif
2416
2417        max_load = this_load = total_load = total_pwr = 0;
2418        busiest_load_per_task = busiest_nr_running = 0;
2419        this_load_per_task = this_nr_running = 0;
2420        if (idle == CPU_NOT_IDLE)
2421                load_idx = sd->busy_idx;
2422        else if (idle == CPU_NEWLY_IDLE)
2423                load_idx = sd->newidle_idx;
2424        else
2425                load_idx = sd->idle_idx;
2426
2427        do {
2428                unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2429                int local_group;
2430                int i;
2431                int __group_imb = 0;
2432                unsigned int balance_cpu = -1, first_idle_cpu = 0;
2433                unsigned long sum_nr_running, sum_weighted_load;
2434
2435                local_group = cpu_isset(this_cpu, group->cpumask);
2436
2437                if (local_group)
2438                        balance_cpu = first_cpu(group->cpumask);
2439
2440                /* Tally up the load of all CPUs in the group */
2441                sum_weighted_load = sum_nr_running = avg_load = 0;
2442                max_cpu_load = 0;
2443                min_cpu_load = ~0UL;
2444
2445                for_each_cpu_mask(i, group->cpumask) {
2446                        struct rq *rq;
2447
2448                        if (!cpu_isset(i, *cpus))
2449                                continue;
2450
2451                        rq = cpu_rq(i);
2452
2453                        if (*sd_idle && rq->nr_running)
2454                                *sd_idle = 0;
2455
2456                        /* Bias balancing toward cpus of our domain */
2457                        if (local_group) {
2458                                if (idle_cpu(i) && !first_idle_cpu) {
2459                                        first_idle_cpu = 1;
2460                                        balance_cpu = i;
2461                                }
2462
2463                                load = target_load(i, load_idx);
2464                        } else {
2465                                load = source_load(i, load_idx);
2466                                if (load > max_cpu_load)
2467                                        max_cpu_load = load;
2468                                if (min_cpu_load > load)
2469                                        min_cpu_load = load;
2470                        }
2471
2472                        avg_load += load;
2473                        sum_nr_running += rq->nr_running;
2474                        sum_weighted_load += weighted_cpuload(i);
2475                }
2476
2477                /*
2478                 * First idle cpu or the first cpu(busiest) in this sched group
2479                 * is eligible for doing load balancing at this and above
2480                 * domains. In the newly idle case, we will allow all the cpu's
2481                 * to do the newly idle load balance.
2482                 */
2483                if (idle != CPU_NEWLY_IDLE && local_group &&
2484                    balance_cpu != this_cpu && balance) {
2485                        *balance = 0;
2486                        goto ret;
2487                }
2488
2489                total_load += avg_load;
2490                total_pwr += group->__cpu_power;
2491
2492                /* Adjust by relative CPU power of the group */
2493                avg_load = sg_div_cpu_power(group,
2494                                avg_load * SCHED_LOAD_SCALE);
2495
2496                if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2497                        __group_imb = 1;
2498
2499                group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2500
2501                if (local_group) {
2502                        this_load = avg_load;
2503                        this = group;
2504                        this_nr_running = sum_nr_running;
2505                        this_load_per_task = sum_weighted_load;
2506                } else if (avg_load > max_load &&
2507                           (sum_nr_running > group_capacity || __group_imb)) {
2508                        max_load = avg_load;
2509                        busiest = group;
2510                        busiest_nr_running = sum_nr_running;
2511                        busiest_load_per_task = sum_weighted_load;
2512                        group_imb = __group_imb;
2513                }
2514
2515#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2516                /*
2517                 * Busy processors will not participate in power savings
2518                 * balance.
2519                 */
2520                if (idle == CPU_NOT_IDLE ||
2521                                !(sd->flags & SD_POWERSAVINGS_BALANCE))
2522                        goto group_next;
2523
2524                /*
2525                 * If the local group is idle or completely loaded
2526                 * no need to do power savings balance at this domain
2527                 */
2528                if (local_group && (this_nr_running >= group_capacity ||
2529                                    !this_nr_running))
2530                        power_savings_balance = 0;
2531
2532                /*
2533                 * If a group is already running at full capacity or idle,
2534                 * don't include that group in power savings calculations
2535                 */
2536                if (!power_savings_balance || sum_nr_running >= group_capacity
2537                    || !sum_nr_running)
2538                        goto group_next;
2539
2540                /*
2541                 * Calculate the group which has the least non-idle load.
2542                 * This is the group from where we need to pick up the load
2543                 * for saving power
2544                 */
2545                if ((sum_nr_running < min_nr_running) ||
2546                    (sum_nr_running == min_nr_running &&
2547                     first_cpu(group->cpumask) <
2548                     first_cpu(group_min->cpumask))) {
2549                        group_min = group;
2550                        min_nr_running = sum_nr_running;
2551                        min_load_per_task = sum_weighted_load /
2552                                                sum_nr_running;
2553                }
2554
2555                /*
2556                 * Calculate the group which is almost near its
2557                 * capacity but still has some space to pick up some load
2558                 * from other group and save more power
2559                 */
2560                if (sum_nr_running <= group_capacity - 1) {
2561                        if (sum_nr_running > leader_nr_running ||
2562                            (sum_nr_running == leader_nr_running &&
2563                             first_cpu(group->cpumask) >
2564                              first_cpu(group_leader->cpumask))) {
2565                                group_leader = group;
2566                                leader_nr_running = sum_nr_running;
2567                        }
2568                }
2569group_next:
2570#endif
2571                group = group->next;
2572        } while (group != sd->groups);
2573
2574        if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2575                goto out_balanced;
2576
2577        avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2578
2579        if (this_load >= avg_load ||
2580                        100*max_load <= sd->imbalance_pct*this_load)
2581                goto out_balanced;
2582
2583        busiest_load_per_task /= busiest_nr_running;
2584        if (group_imb)
2585                busiest_load_per_task = min(busiest_load_per_task, avg_load);
2586
2587        /*
2588         * We're trying to get all the cpus to the average_load, so we don't
2589         * want to push ourselves above the average load, nor do we wish to
2590         * reduce the max loaded cpu below the average load, as either of these
2591         * actions would just result in more rebalancing later, and ping-pong
2592         * tasks around. Thus we look for the minimum possible imbalance.
2593         * Negative imbalances (*we* are more loaded than anyone else) will
2594         * be counted as no imbalance for these purposes -- we can't fix that
2595         * by pulling tasks to us. Be careful of negative numbers as they'll
2596         * appear as very large values with unsigned longs.
2597         */
2598        if (max_load <= busiest_load_per_task)
2599                goto out_balanced;
2600
2601        /*
2602         * In the presence of smp nice balancing, certain scenarios can have
2603         * max load less than avg load(as we skip the groups at or below
2604         * its cpu_power, while calculating max_load..)
2605         */
2606        if (max_load < avg_load) {
2607                *imbalance = 0;
2608                goto small_imbalance;
2609        }
2610
2611        /* Don't want to pull so many tasks that a group would go idle */
2612        max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2613
2614        /* How much load to actually move to equalise the imbalance */
2615        *imbalance = min(max_pull * busiest->__cpu_power,
2616                                (avg_load - this_load) * this->__cpu_power)
2617                        / SCHED_LOAD_SCALE;
2618
2619        /*
2620         * if *imbalance is less than the average load per runnable task
2621         * there is no gaurantee that any tasks will be moved so we'll have
2622         * a think about bumping its value to force at least one task to be
2623         * moved
2624         */
2625        if (*imbalance < busiest_load_per_task) {
2626                unsigned long tmp, pwr_now, pwr_move;
2627                unsigned int imbn;
2628
2629small_imbalance:
2630                pwr_move = pwr_now = 0;
2631                imbn = 2;
2632                if (this_nr_running) {
2633                        this_load_per_task /= this_nr_running;
2634                        if (busiest_load_per_task > this_load_per_task)
2635                                imbn = 1;
2636                } else
2637                        this_load_per_task = SCHED_LOAD_SCALE;
2638
2639                if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2640                                        busiest_load_per_task * imbn) {
2641                        *imbalance = busiest_load_per_task;
2642                        return busiest;
2643                }
2644
2645                /*
2646                 * OK, we don't have enough imbalance to justify moving tasks,
2647                 * however we may be able to increase total CPU power used by
2648                 * moving them.
2649                 */
2650
2651                pwr_now += busiest->__cpu_power *
2652                                min(busiest_load_per_task, max_load);
2653                pwr_now += this->__cpu_power *
2654                                min(this_load_per_task, this_load);
2655                pwr_now /= SCHED_LOAD_SCALE;
2656
2657                /* Amount of load we'd subtract */
2658                tmp = sg_div_cpu_power(busiest,
2659                                busiest_load_per_task * SCHED_LOAD_SCALE);
2660                if (max_load > tmp)
2661                        pwr_move += busiest->__cpu_power *
2662                                min(busiest_load_per_task, max_load - tmp);
2663
2664                /* Amount of load we'd add */
2665                if (max_load * busiest->__cpu_power <
2666                                busiest_load_per_task * SCHED_LOAD_SCALE)
2667                        tmp = sg_div_cpu_power(this,
2668                                        max_load * busiest->__cpu_power);
2669                else
2670                        tmp = sg_div_cpu_power(this,
2671                                busiest_load_per_task * SCHED_LOAD_SCALE);
2672                pwr_move += this->__cpu_power *
2673                                min(this_load_per_task, this_load + tmp);
2674                pwr_move /= SCHED_LOAD_SCALE;
2675
2676                /* Move if we gain throughput */
2677                if (pwr_move > pwr_now)
2678                        *imbalance = busiest_load_per_task;
2679        }
2680
2681        return busiest;
2682
2683out_balanced:
2684#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2685        if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2686                goto ret;
2687
2688        if (this == group_leader && group_leader != group_min) {
2689                *imbalance = min_load_per_task;
2690                return group_min;
2691        }
2692#endif
2693ret:
2694        *imbalance = 0;
2695        return NULL;
2696}
2697
2698/*
2699 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2700 */
2701static struct rq *
2702find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2703                   unsigned long imbalance, cpumask_t *cpus)
2704{
2705        struct rq *busiest = NULL, *rq;
2706        unsigned long max_load = 0;
2707        int i;
2708
2709        for_each_cpu_mask(i, group->cpumask) {
2710                unsigned long wl;
2711
2712                if (!cpu_isset(i, *cpus))
2713                        continue;
2714
2715                rq = cpu_rq(i);
2716                wl = weighted_cpuload(i);
2717
2718                if (rq->nr_running == 1 && wl > imbalance)
2719                        continue;
2720
2721                if (wl > max_load) {
2722                        max_load = wl;
2723                        busiest = rq;
2724                }
2725        }
2726
2727        return busiest;
2728}
2729
2730/*
2731 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2732 * so long as it is large enough.
2733 */
2734#define MAX_PINNED_INTERVAL     512
2735
2736/*
2737 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2738 * tasks if there is an imbalance.
2739 */
2740static int load_balance(int this_cpu, struct rq *this_rq,
2741                        struct sched_domain *sd, enum cpu_idle_type idle,
2742                        int *balance)
2743{
2744        int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2745        struct sched_group *group;
2746        unsigned long imbalance;
2747        struct rq *busiest;
2748        cpumask_t cpus = CPU_MASK_ALL;
2749        unsigned long flags;
2750
2751        /*
2752         * When power savings policy is enabled for the parent domain, idle
2753         * sibling can pick up load irrespective of busy siblings. In this case,
2754         * let the state of idle sibling percolate up as CPU_IDLE, instead of
2755         * portraying it as CPU_NOT_IDLE.
2756         */
2757        if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2758            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2759                sd_idle = 1;
2760
2761        schedstat_inc(sd, lb_count[idle]);
2762
2763redo:
2764        group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2765                                   &cpus, balance);
2766
2767        if (*balance == 0)
2768                goto out_balanced;
2769
2770        if (!group) {
2771                schedstat_inc(sd, lb_nobusyg[idle]);
2772                goto out_balanced;
2773        }
2774
2775        busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2776        if (!busiest) {
2777                schedstat_inc(sd, lb_nobusyq[idle]);
2778                goto out_balanced;
2779        }
2780
2781        BUG_ON(busiest == this_rq);
2782
2783        schedstat_add(sd, lb_imbalance[idle], imbalance);
2784
2785        ld_moved = 0;
2786        if (busiest->nr_running > 1) {
2787                /*
2788                 * Attempt to move tasks. If find_busiest_group has found
2789                 * an imbalance but busiest->nr_running <= 1, the group is
2790                 * still unbalanced. ld_moved simply stays zero, so it is
2791                 * correctly treated as an imbalance.
2792                 */
2793                local_irq_save(flags);
2794                double_rq_lock(this_rq, busiest);
2795                ld_moved = move_tasks(this_rq, this_cpu, busiest,
2796                                      imbalance, sd, idle, &all_pinned);
2797                double_rq_unlock(this_rq, busiest);
2798                local_irq_restore(flags);
2799
2800                /*
2801                 * some other cpu did the load balance for us.
2802                 */
2803                if (ld_moved && this_cpu != smp_processor_id())
2804                        resched_cpu(this_cpu);
2805
2806                /* All tasks on this runqueue were pinned by CPU affinity */
2807                if (unlikely(all_pinned)) {
2808                        cpu_clear(cpu_of(busiest), cpus);
2809                        if (!cpus_empty(cpus))
2810                                goto redo;
2811                        goto out_balanced;
2812                }
2813        }
2814
2815        if (!ld_moved) {
2816                schedstat_inc(sd, lb_failed[idle]);
2817                sd->nr_balance_failed++;
2818
2819                if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2820
2821                        spin_lock_irqsave(&busiest->lock, flags);
2822
2823                        /* don't kick the migration_thread, if the curr
2824                         * task on busiest cpu can't be moved to this_cpu
2825                         */
2826                        if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2827                                spin_unlock_irqrestore(&busiest->lock, flags);
2828                                all_pinned = 1;
2829                                goto out_one_pinned;
2830                        }
2831
2832                        if (!busiest->active_balance) {
2833                                busiest->active_balance = 1;
2834                                busiest->push_cpu = this_cpu;
2835                                active_balance = 1;
2836                        }
2837                        spin_unlock_irqrestore(&busiest->lock, flags);
2838                        if (active_balance)
2839                                wake_up_process(busiest->migration_thread);
2840
2841                        /*
2842                         * We've kicked active balancing, reset the failure
2843                         * counter.
2844                         */
2845                        sd->nr_balance_failed = sd->cache_nice_tries+1;
2846                }
2847        } else
2848                sd->nr_balance_failed = 0;
2849
2850        if (likely(!active_balance)) {
2851                /* We were unbalanced, so reset the balancing interval */
2852                sd->balance_interval = sd->min_interval;
2853        } else {
2854                /*
2855                 * If we've begun active balancing, start to back off. This
2856                 * case may not be covered by the all_pinned logic if there
2857                 * is only 1 task on the busy runqueue (because we don't call
2858                 * move_tasks).
2859                 */
2860                if (sd->balance_interval < sd->max_interval)
2861                        sd->balance_interval *= 2;
2862        }
2863
2864        if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2865            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2866                return -1;
2867        return ld_moved;
2868
2869out_balanced:
2870        schedstat_inc(sd, lb_balanced[idle]);
2871
2872        sd->nr_balance_failed = 0;
2873
2874out_one_pinned:
2875        /* tune up the balancing interval */
2876        if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2877                        (sd->balance_interval < sd->max_interval))
2878                sd->balance_interval *= 2;
2879
2880        if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2881            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2882                return -1;
2883        return 0;
2884}
2885
2886/*
2887 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2888 * tasks if there is an imbalance.
2889 *
2890 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
2891 * this_rq is locked.
2892 */
2893static int
2894load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2895{
2896        struct sched_group *group;
2897        struct rq *busiest = NULL;
2898        unsigned long imbalance;
2899        int ld_moved = 0;
2900        int sd_idle = 0;
2901        int all_pinned = 0;
2902        cpumask_t cpus = CPU_MASK_ALL;
2903
2904        /*
2905         * When power savings policy is enabled for the parent domain, idle
2906         * sibling can pick up load irrespective of busy siblings. In this case,
2907         * let the state of idle sibling percolate up as IDLE, instead of
2908         * portraying it as CPU_NOT_IDLE.
2909         */
2910        if (sd->flags & SD_SHARE_CPUPOWER &&
2911            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2912                sd_idle = 1;
2913
2914        schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
2915redo:
2916        group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
2917                                   &sd_idle, &cpus, NULL);
2918        if (!group) {
2919                schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
2920                goto out_balanced;
2921        }
2922
2923        busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
2924                                &cpus);
2925        if (!busiest) {
2926                schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
2927                goto out_balanced;
2928        }
2929
2930        BUG_ON(busiest == this_rq);
2931
2932        schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
2933
2934        ld_moved = 0;
2935        if (busiest->nr_running > 1) {
2936                /* Attempt to move tasks */
2937                double_lock_balance(this_rq, busiest);
2938                /* this_rq->clock is already updated */
2939                update_rq_clock(busiest);
2940                ld_moved = move_tasks(this_rq, this_cpu, busiest,
2941                                        imbalance, sd, CPU_NEWLY_IDLE,
2942                                        &all_pinned);
2943                spin_unlock(&busiest->lock);
2944
2945                if (unlikely(all_pinned)) {
2946                        cpu_clear(cpu_of(busiest), cpus);
2947                        if (!cpus_empty(cpus))
2948                                goto redo;
2949                }
2950        }
2951
2952        if (!ld_moved) {
2953                schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
2954                if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2955                    !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2956                        return -1;
2957        } else
2958                sd->nr_balance_failed = 0;
2959
2960        return ld_moved;
2961
2962out_balanced:
2963        schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
2964        if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2965            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2966                return -1;
2967        sd->nr_balance_failed = 0;
2968
2969        return 0;
2970}
2971
2972/*
2973 * idle_balance is called by schedule() if this_cpu is about to become
2974 * idle. Attempts to pull tasks from other CPUs.
2975 */
2976static void idle_balance(int this_cpu, struct rq *this_rq)
2977{
2978        struct sched_domain *sd;
2979        int pulled_task = -1;
2980        unsigned long next_balance = jiffies + HZ;
2981
2982        for_each_domain(this_cpu, sd) {
2983                unsigned long interval;
2984
2985                if (!(sd->flags & SD_LOAD_BALANCE))
2986                        continue;
2987
2988                if (sd->flags & SD_BALANCE_NEWIDLE)
2989                        /* If we've pulled tasks over stop searching: */
2990                        pulled_task = load_balance_newidle(this_cpu,
2991                                                                this_rq, sd);
2992
2993                interval = msecs_to_jiffies(sd->balance_interval);
2994                if (time_after(next_balance, sd->last_balance + interval))
2995                        next_balance = sd->last_balance + interval;
2996                if (pulled_task)
2997                        break;
2998        }
2999        if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3000                /*
3001                 * We are going idle. next_balance may be set based on
3002                 * a busy processor. So reset next_balance.
3003                 */
3004                this_rq->next_balance = next_balance;
3005        }
3006}
3007
3008/*
3009 * active_load_balance is run by migration threads. It pushes running tasks
3010 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3011 * running on each physical CPU where possible, and avoids physical /
3012 * logical imbalances.
3013 *
3014 * Called with busiest_rq locked.
3015 */
3016static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3017{
3018        int target_cpu = busiest_rq->push_cpu;
3019        struct sched_domain *sd;
3020        struct rq *target_rq;
3021
3022        /* Is there any task to move? */
3023        if (busiest_rq->nr_running <= 1)
3024                return;
3025
3026        target_rq = cpu_rq(target_cpu);
3027
3028        /*
3029         * This condition is "impossible", if it occurs
3030         * we need to fix it. Originally reported by
3031         * Bjorn Helgaas on a 128-cpu setup.
3032         */
3033        BUG_ON(busiest_rq == target_rq);
3034
3035        /* move a task from busiest_rq to target_rq */
3036        double_lock_balance(busiest_rq, target_rq);
3037        update_rq_clock(busiest_rq);
3038        update_rq_clock(target_rq);
3039
3040        /* Search for an sd spanning us and the target CPU. */
3041        for_each_domain(target_cpu, sd) {
3042                if ((sd->flags & SD_LOAD_BALANCE) &&
3043                    cpu_isset(busiest_cpu, sd->span))
3044                                break;
3045        }
3046
3047        if (likely(sd)) {
3048                schedstat_inc(sd, alb_count);
3049
3050                if (move_one_task(target_rq, target_cpu, busiest_rq,
3051                                  sd, CPU_IDLE))
3052                        schedstat_inc(sd, alb_pushed);
3053                else
3054                        schedstat_inc(sd, alb_failed);
3055        }
3056        spin_unlock(&target_rq->lock);
3057}
3058
3059#ifdef CONFIG_NO_HZ
3060static struct {
3061        atomic_t load_balancer;
3062        cpumask_t cpu_mask;
3063} nohz ____cacheline_aligned = {
3064        .load_balancer = ATOMIC_INIT(-1),
3065        .cpu_mask = CPU_MASK_NONE,
3066};
3067
3068/*
3069 * This routine will try to nominate the ilb (idle load balancing)
3070 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3071 * load balancing on behalf of all those cpus. If all the cpus in the system
3072 * go into this tickless mode, then there will be no ilb owner (as there is
3073 * no need for one) and all the cpus will sleep till the next wakeup event
3074 * arrives...
3075 *
3076 * For the ilb owner, tick is not stopped. And this tick will be used
3077 * for idle load balancing. ilb owner will still be part of
3078 * nohz.cpu_mask..
3079 *
3080 * While stopping the tick, this cpu will become the ilb owner if there
3081 * is no other owner. And will be the owner till that cpu becomes busy
3082 * or if all cpus in the system stop their ticks at which point
3083 * there is no need for ilb owner.
3084 *
3085 * When the ilb owner becomes busy, it nominates another owner, during the
3086 * next busy scheduler_tick()
3087 */
3088int select_nohz_load_balancer(int stop_tick)
3089{
3090        int cpu = smp_processor_id();
3091
3092        if (stop_tick) {
3093                cpu_set(cpu, nohz.cpu_mask);
3094                cpu_rq(cpu)->in_nohz_recently = 1;
3095
3096                /*
3097                 * If we are going offline and still the leader, give up!
3098                 */
3099                if (cpu_is_offline(cpu) &&
3100                    atomic_read(&nohz.load_balancer) == cpu) {
3101                        if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3102                                BUG();
3103                        return 0;
3104                }
3105
3106                /* time for ilb owner also to sleep */
3107                if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3108                        if (atomic_read(&nohz.load_balancer) == cpu)
3109                                atomic_set(&nohz.load_balancer, -1);
3110                        return 0;
3111                }
3112
3113                if (atomic_read(&nohz.load_balancer) == -1) {
3114                        /* make me the ilb owner */
3115                        if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3116                                return 1;
3117                } else if (atomic_read(&nohz.load_balancer) == cpu)
3118                        return 1;
3119        } else {
3120                if (!cpu_isset(cpu, nohz.cpu_mask))
3121                        return 0;
3122
3123                cpu_clear(cpu, nohz.cpu_mask);
3124
3125                if (atomic_read(&nohz.load_balancer) == cpu)
3126                        if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3127                                BUG();
3128        }
3129        return 0;
3130}
3131#endif
3132
3133static DEFINE_SPINLOCK(balancing);
3134
3135/*
3136 * It checks each scheduling domain to see if it is due to be balanced,
3137 * and initiates a balancing operation if so.
3138 *
3139 * Balancing parameters are set up in arch_init_sched_domains.
3140 */
3141static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3142{
3143        int balance = 1;
3144        struct rq *rq = cpu_rq(cpu);
3145        unsigned long interval;
3146        struct sched_domain *sd;
3147        /* Earliest time when we have to do rebalance again */
3148        unsigned long next_balance = jiffies + 60*HZ;
3149        int update_next_balance = 0;
3150
3151        for_each_domain(cpu, sd) {
3152                if (!(sd->flags & SD_LOAD_BALANCE))
3153                        continue;
3154
3155                interval = sd->balance_interval;
3156                if (idle != CPU_IDLE)
3157                        interval *= sd->busy_factor;
3158
3159                /* scale ms to jiffies */
3160                interval = msecs_to_jiffies(interval);
3161                if (unlikely(!interval))
3162                        interval = 1;
3163                if (interval > HZ*NR_CPUS/10)
3164                        interval = HZ*NR_CPUS/10;
3165
3166
3167                if (sd->flags & SD_SERIALIZE) {
3168                        if (!spin_trylock(&balancing))
3169                                goto out;
3170                }
3171
3172                if (time_after_eq(jiffies, sd->last_balance + interval)) {
3173                        if (load_balance(cpu, rq, sd, idle, &balance)) {
3174                                /*
3175                                 * We've pulled tasks over so either we're no
3176                                 * longer idle, or one of our SMT siblings is
3177                                 * not idle.
3178                                 */
3179                                idle = CPU_NOT_IDLE;
3180                        }
3181                        sd->last_balance = jiffies;
3182                }
3183                if (sd->flags & SD_SERIALIZE)
3184                        spin_unlock(&balancing);
3185out:
3186                if (time_after(next_balance, sd->last_balance + interval)) {
3187                        next_balance = sd->last_balance + interval;
3188                        update_next_balance = 1;
3189                }
3190
3191                /*
3192                 * Stop the load balance at this level. There is another
3193                 * CPU in our sched group which is doing load balancing more
3194                 * actively.
3195                 */
3196                if (!balance)
3197                        break;
3198        }
3199
3200        /*
3201         * next_balance will be updated only when there is a need.
3202         * When the cpu is attached to null domain for ex, it will not be
3203         * updated.
3204         */
3205        if (likely(update_next_balance))
3206                rq->next_balance = next_balance;
3207}
3208
3209/*
3210 * run_rebalance_domains is triggered when needed from the scheduler tick.
3211 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3212 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3213 */
3214static void run_rebalance_domains(struct softirq_action *h)
3215{
3216        int this_cpu = smp_processor_id();
3217        struct rq *this_rq = cpu_rq(this_cpu);
3218        enum cpu_idle_type idle = this_rq->idle_at_tick ?
3219                                                CPU_IDLE : CPU_NOT_IDLE;
3220
3221        rebalance_domains(this_cpu, idle);
3222
3223#ifdef CONFIG_NO_HZ
3224        /*
3225         * If this cpu is the owner for idle load balancing, then do the
3226         * balancing on behalf of the other idle cpus whose ticks are
3227         * stopped.
3228         */
3229        if (this_rq->idle_at_tick &&
3230            atomic_read(&nohz.load_balancer) == this_cpu) {
3231                cpumask_t cpus = nohz.cpu_mask;
3232                struct rq *rq;
3233                int balance_cpu;
3234
3235                cpu_clear(this_cpu, cpus);
3236                for_each_cpu_mask(balance_cpu, cpus) {
3237                        /*
3238                         * If this cpu gets work to do, stop the load balancing
3239                         * work being done for other cpus. Next load
3240                         * balancing owner will pick it up.
3241                         */
3242                        if (need_resched())
3243                                break;
3244
3245                        rebalance_domains(balance_cpu, CPU_IDLE);
3246
3247                        rq = cpu_rq(balance_cpu);
3248                        if (time_after(this_rq->next_balance, rq->next_balance))
3249                                this_rq->next_balance = rq->next_balance;
3250                }
3251        }
3252#endif
3253}
3254
3255/*
3256 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3257 *
3258 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3259 * idle load balancing owner or decide to stop the periodic load balancing,
3260 * if the whole system is idle.
3261 */
3262static inline void trigger_load_balance(struct rq *rq, int cpu)
3263{
3264#ifdef CONFIG_NO_HZ
3265        /*
3266         * If we were in the nohz mode recently and busy at the current
3267         * scheduler tick, then check if we need to nominate new idle
3268         * load balancer.
3269         */
3270        if (rq->in_nohz_recently && !rq->idle_at_tick) {
3271                rq->in_nohz_recently = 0;
3272
3273                if (atomic_read(&nohz.load_balancer) == cpu) {
3274                        cpu_clear(cpu, nohz.cpu_mask);
3275                        atomic_set(&nohz.load_balancer, -1);
3276                }
3277
3278                if (atomic_read(&nohz.load_balancer) == -1) {
3279                        /*
3280                         * simple selection for now: Nominate the
3281                         * first cpu in the nohz list to be the next
3282                         * ilb owner.
3283                         *
3284                         * TBD: Traverse the sched domains and nominate
3285                         * the nearest cpu in the nohz.cpu_mask.
3286                         */
3287                        int ilb = first_cpu(nohz.cpu_mask);
3288
3289                        if (ilb != NR_CPUS)
3290                                resched_cpu(ilb);
3291                }
3292        }
3293
3294        /*
3295         * If this cpu is idle and doing idle load balancing for all the
3296         * cpus with ticks stopped, is it time for that to stop?
3297         */
3298        if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3299            cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3300                resched_cpu(cpu);
3301                return;
3302        }
3303
3304        /*
3305         * If this cpu is idle and the idle load balancing is done by
3306         * someone else, then no need raise the SCHED_SOFTIRQ
3307         */
3308        if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3309            cpu_isset(cpu, nohz.cpu_mask))
3310                return;
3311#endif
3312        if (time_after_eq(jiffies, rq->next_balance))
3313                raise_softirq(SCHED_SOFTIRQ);
3314}
3315
3316#else   /* CONFIG_SMP */
3317
3318/*
3319 * on UP we do not need to balance between CPUs:
3320 */
3321static inline void idle_balance(int cpu, struct rq *rq)
3322{
3323}
3324
3325#endif
3326
3327DEFINE_PER_CPU(struct kernel_stat, kstat);
3328
3329EXPORT_PER_CPU_SYMBOL(kstat);
3330
3331/*
3332 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3333 * that have not yet been banked in case the task is currently running.
3334 */
3335unsigned long long task_sched_runtime(struct task_struct *p)
3336{
3337        unsigned long flags;
3338        u64 ns, delta_exec;
3339        struct rq *rq;
3340
3341        rq = task_rq_lock(p, &flags);
3342        ns = p->se.sum_exec_runtime;
3343        if (task_current(rq, p)) {
3344                update_rq_clock(rq);
3345                delta_exec = rq->clock - p->se.exec_start;
3346                if ((s64)delta_exec > 0)
3347                        ns += delta_exec;
3348        }
3349        task_rq_unlock(rq, &flags);
3350
3351        return ns;
3352}
3353
3354/*
3355 * Account user cpu time to a process.
3356 * @p: the process that the cpu time gets accounted to
3357 * @cputime: the cpu time spent in user space since the last update
3358 */
3359void account_user_time(struct task_struct *p, cputime_t cputime)
3360{
3361        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3362        cputime64_t tmp;
3363
3364        p->utime = cputime_add(p->utime, cputime);
3365
3366        /* Add user time to cpustat. */
3367        tmp = cputime_to_cputime64(cputime);
3368        if (TASK_NICE(p) > 0)
3369                cpustat->nice = cputime64_add(cpustat->nice, tmp);
3370        else
3371                cpustat->user = cputime64_add(cpustat->user, tmp);
3372}
3373
3374/*
3375 * Account guest cpu time to a process.
3376 * @p: the process that the cpu time gets accounted to
3377 * @cputime: the cpu time spent in virtual machine since the last update
3378 */
3379static void account_guest_time(struct task_struct *p, cputime_t cputime)
3380{
3381        cputime64_t tmp;
3382        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3383
3384        tmp = cputime_to_cputime64(cputime);
3385
3386        p->utime = cputime_add(p->utime, cputime);
3387        p->gtime = cputime_add(p->gtime, cputime);
3388
3389        cpustat->user = cputime64_add(cpustat->user, tmp);
3390        cpustat->guest = cputime64_add(cpustat->guest, tmp);
3391}
3392
3393/*
3394 * Account scaled user cpu time to a process.
3395 * @p: the process that the cpu time gets accounted to
3396 * @cputime: the cpu time spent in user space since the last update
3397 */
3398void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3399{
3400        p->utimescaled = cputime_add(p->utimescaled, cputime);
3401}
3402
3403/*
3404 * Account system cpu time to a process.
3405 * @p: the process that the cpu time gets accounted to
3406 * @hardirq_offset: the offset to subtract from hardirq_count()
3407 * @cputime: the cpu time spent in kernel space since the last update
3408 */
3409void account_system_time(struct task_struct *p, int hardirq_offset,
3410                         cputime_t cputime)
3411{
3412        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3413        struct rq *rq = this_rq();
3414        cputime64_t tmp;
3415
3416        if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3417                return account_guest_time(p, cputime);
3418
3419        p->stime = cputime_add(p->stime, cputime);
3420
3421        /* Add system time to cpustat. */
3422        tmp = cputime_to_cputime64(cputime);
3423        if (hardirq_count() - hardirq_offset)
3424                cpustat->irq = cputime64_add(cpustat->irq, tmp);
3425        else if (softirq_count())
3426                cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3427        else if (p != rq->idle)
3428                cpustat->system = cputime64_add(cpustat->system, tmp);
3429        else if (atomic_read(&rq->nr_iowait) > 0)
3430                cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3431        else
3432                cpustat->idle = cputime64_add(cpustat->idle, tmp);
3433        /* Account for system time used */
3434        acct_update_integrals(p);
3435}
3436
3437/*
3438 * Account scaled system cpu time to a process.
3439 * @p: the process that the cpu time gets accounted to
3440 * @hardirq_offset: the offset to subtract from hardirq_count()
3441 * @cputime: the cpu time spent in kernel space since the last update
3442 */
3443void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3444{
3445        p->stimescaled = cputime_add(p->stimescaled, cputime);
3446}
3447
3448/*
3449 * Account for involuntary wait time.
3450 * @p: the process from which the cpu time has been stolen
3451 * @steal: the cpu time spent in involuntary wait
3452 */
3453void account_steal_time(struct task_struct *p, cputime_t steal)
3454{
3455        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3456        cputime64_t tmp = cputime_to_cputime64(steal);
3457        struct rq *rq = this_rq();
3458
3459        if (p == rq->idle) {
3460                p->stime = cputime_add(p->stime, steal);
3461                if (atomic_read(&rq->nr_iowait) > 0)
3462                        cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3463                else
3464                        cpustat->idle = cputime64_add(cpustat->idle, tmp);
3465        } else
3466                cpustat->steal = cputime64_add(cpustat->steal, tmp);
3467}
3468
3469/*
3470 * This function gets called by the timer code, with HZ frequency.
3471 * We call it with interrupts disabled.
3472 *
3473 * It also gets called by the fork code, when changing the parent's
3474 * timeslices.
3475 */
3476void scheduler_tick(void)
3477{
3478        int cpu = smp_processor_id();
3479        struct rq *rq = cpu_rq(cpu);
3480        struct task_struct *curr = rq->curr;
3481        u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3482
3483        spin_lock(&rq->lock);
3484        __update_rq_clock(rq);
3485        /*
3486         * Let rq->clock advance by at least TICK_NSEC:
3487         */
3488        if (unlikely(rq->clock < next_tick))
3489                rq->clock = next_tick;
3490        rq->tick_timestamp = rq->clock;
3491        update_cpu_load(rq);
3492        if (curr != rq->idle) /* FIXME: needed? */
3493                curr->sched_class->task_tick(rq, curr);
3494        spin_unlock(&rq->lock);
3495
3496#ifdef CONFIG_SMP
3497        rq->idle_at_tick = idle_cpu(cpu);
3498        trigger_load_balance(rq, cpu);
3499#endif
3500}
3501
3502#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3503
3504void fastcall add_preempt_count(int val)
3505{
3506        /*
3507         * Underflow?
3508         */
3509        if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3510                return;
3511        preempt_count() += val;
3512        /*
3513         * Spinlock count overflowing soon?
3514         */
3515        DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3516                                PREEMPT_MASK - 10);
3517}
3518EXPORT_SYMBOL(add_preempt_count);
3519
3520void fastcall sub_preempt_count(int val)
3521{
3522        /*
3523         * Underflow?
3524         */
3525        if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3526                return;
3527        /*
3528         * Is the spinlock portion underflowing?
3529         */
3530        if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3531                        !(preempt_count() & PREEMPT_MASK)))
3532                return;
3533
3534        preempt_count() -= val;
3535}
3536EXPORT_SYMBOL(sub_preempt_count);
3537
3538#endif
3539
3540/*
3541 * Print scheduling while atomic bug:
3542 */
3543static noinline void __schedule_bug(struct task_struct *prev)
3544{
3545        struct pt_regs *regs = get_irq_regs();
3546
3547        printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3548                prev->comm, prev->pid, preempt_count());
3549
3550        debug_show_held_locks(prev);
3551        if (irqs_disabled())
3552                print_irqtrace_events(prev);
3553
3554        if (regs)
3555                show_regs(regs);
3556        else
3557                dump_stack();
3558}
3559
3560/*
3561 * Various schedule()-time debugging checks and statistics:
3562 */
3563static inline void schedule_debug(struct task_struct *prev)
3564{
3565        /*
3566         * Test if we are atomic. Since do_exit() needs to call into
3567         * schedule() atomically, we ignore that path for now.
3568         * Otherwise, whine if we are scheduling when we should not be.
3569         */
3570        if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3571                __schedule_bug(prev);
3572
3573        profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3574
3575        schedstat_inc(this_rq(), sched_count);
3576#ifdef CONFIG_SCHEDSTATS
3577        if (unlikely(prev->lock_depth >= 0)) {
3578                schedstat_inc(this_rq(), bkl_count);
3579                schedstat_inc(prev, sched_info.bkl_count);
3580        }
3581#endif
3582}
3583
3584/*
3585 * Pick up the highest-prio task:
3586 */
3587static inline struct task_struct *
3588pick_next_task(struct rq *rq, struct task_struct *prev)
3589{
3590        const struct sched_class *class;
3591        struct task_struct *p;
3592
3593        /*
3594         * Optimization: we know that if all tasks are in
3595         * the fair class we can call that function directly:
3596         */
3597        if (likely(rq->nr_running == rq->cfs.nr_running)) {
3598                p = fair_sched_class.pick_next_task(rq);
3599                if (likely(p))
3600                        return p;
3601        }
3602
3603        class = sched_class_highest;
3604        for ( ; ; ) {
3605                p = class->pick_next_task(rq);
3606                if (p)
3607                        return p;
3608                /*
3609                 * Will never be NULL as the idle class always
3610                 * returns a non-NULL p:
3611                 */
3612                class = class->next;
3613        }
3614}
3615
3616/*
3617 * schedule() is the main scheduler function.
3618 */
3619asmlinkage void __sched schedule(void)
3620{
3621        struct task_struct *prev, *next;
3622        long *switch_count;
3623        struct rq *rq;
3624        int cpu;
3625
3626need_resched:
3627        preempt_disable();
3628        cpu = smp_processor_id();
3629        rq = cpu_rq(cpu);
3630        rcu_qsctr_inc(cpu);
3631        prev = rq->curr;
3632        switch_count = &prev->nivcsw;
3633
3634        release_kernel_lock(prev);
3635need_resched_nonpreemptible:
3636
3637        schedule_debug(prev);
3638
3639        /*
3640         * Do the rq-clock update outside the rq lock:
3641         */
3642        local_irq_disable();
3643        __update_rq_clock(rq);
3644        spin_lock(&rq->lock);
3645        clear_tsk_need_resched(prev);
3646
3647        if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3648                if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3649                                unlikely(signal_pending(prev)))) {
3650                        prev->state = TASK_RUNNING;
3651                } else {
3652                        deactivate_task(rq, prev, 1);
3653                }
3654                switch_count = &prev->nvcsw;
3655        }
3656
3657        if (unlikely(!rq->nr_running))
3658                idle_balance(cpu, rq);
3659
3660        prev->sched_class->put_prev_task(rq, prev);
3661        next = pick_next_task(rq, prev);
3662
3663        sched_info_switch(prev, next);
3664
3665        if (likely(prev != next)) {
3666                rq->nr_switches++;
3667                rq->curr = next;
3668                ++*switch_count;
3669
3670                context_switch(rq, prev, next); /* unlocks the rq */
3671        } else
3672                spin_unlock_irq(&rq->lock);
3673
3674        if (unlikely(reacquire_kernel_lock(current) < 0)) {
3675                cpu = smp_processor_id();
3676                rq = cpu_rq(cpu);
3677                goto need_resched_nonpreemptible;
3678        }
3679        preempt_enable_no_resched();
3680        if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3681                goto need_resched;
3682}
3683EXPORT_SYMBOL(schedule);
3684
3685#ifdef CONFIG_PREEMPT
3686/*
3687 * this is the entry point to schedule() from in-kernel preemption
3688 * off of preempt_enable. Kernel preemptions off return from interrupt
3689 * occur there and call schedule directly.
3690 */
3691asmlinkage void __sched preempt_schedule(void)
3692{
3693        struct thread_info *ti = current_thread_info();
3694#ifdef CONFIG_PREEMPT_BKL
3695        struct task_struct *task = current;
3696        int saved_lock_depth;
3697#endif
3698        /*
3699         * If there is a non-zero preempt_count or interrupts are disabled,
3700         * we do not want to preempt the current task. Just return..
3701         */
3702        if (likely(ti->preempt_count || irqs_disabled()))
3703                return;
3704
3705        do {
3706                add_preempt_count(PREEMPT_ACTIVE);
3707
3708                /*
3709                 * We keep the big kernel semaphore locked, but we
3710                 * clear ->lock_depth so that schedule() doesnt
3711                 * auto-release the semaphore:
3712                 */
3713#ifdef CONFIG_PREEMPT_BKL
3714                saved_lock_depth = task->lock_depth;
3715                task->lock_depth = -1;
3716#endif
3717                schedule();
3718#ifdef CONFIG_PREEMPT_BKL
3719                task->lock_depth = saved_lock_depth;
3720#endif
3721                sub_preempt_count(PREEMPT_ACTIVE);
3722
3723                /*
3724                 * Check again in case we missed a preemption opportunity
3725                 * between schedule and now.
3726                 */
3727                barrier();
3728        } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3729}
3730EXPORT_SYMBOL(preempt_schedule);
3731
3732/*
3733 * this is the entry point to schedule() from kernel preemption
3734 * off of irq context.
3735 * Note, that this is called and return with irqs disabled. This will
3736 * protect us against recursive calling from irq.
3737 */
3738asmlinkage void __sched preempt_schedule_irq(void)
3739{
3740        struct thread_info *ti = current_thread_info();
3741#ifdef CONFIG_PREEMPT_BKL
3742        struct task_struct *task = current;
3743        int saved_lock_depth;
3744#endif
3745        /* Catch callers which need to be fixed */
3746        BUG_ON(ti->preempt_count || !irqs_disabled());
3747
3748        do {
3749                add_preempt_count(PREEMPT_ACTIVE);
3750
3751                /*
3752                 * We keep the big kernel semaphore locked, but we
3753                 * clear ->lock_depth so that schedule() doesnt
3754                 * auto-release the semaphore:
3755                 */
3756#ifdef CONFIG_PREEMPT_BKL
3757                saved_lock_depth = task->lock_depth;
3758                task->lock_depth = -1;
3759#endif
3760                local_irq_enable();
3761                schedule();
3762                local_irq_disable();
3763#ifdef CONFIG_PREEMPT_BKL
3764                task->lock_depth = saved_lock_depth;
3765#endif
3766                sub_preempt_count(PREEMPT_ACTIVE);
3767
3768                /*
3769                 * Check again in case we missed a preemption opportunity
3770                 * between schedule and now.
3771                 */
3772                barrier();
3773        } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
3774}
3775
3776#endif /* CONFIG_PREEMPT */
3777
3778int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3779                          void *key)
3780{
3781        return try_to_wake_up(curr->private, mode, sync);
3782}
3783EXPORT_SYMBOL(default_wake_function);
3784
3785/*
3786 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3787 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3788 * number) then we wake all the non-exclusive tasks and one exclusive task.
3789 *
3790 * There are circumstances in which we can try to wake a task which has already
3791 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3792 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3793 */
3794static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3795                             int nr_exclusive, int sync, void *key)
3796{
3797        wait_queue_t *curr, *next;
3798
3799        list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3800                unsigned flags = curr->flags;
3801
3802                if (curr->func(curr, mode, sync, key) &&
3803                                (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3804                        break;
3805        }
3806}
3807
3808/**
3809 * __wake_up - wake up threads blocked on a waitqueue.
3810 * @q: the waitqueue
3811 * @mode: which threads
3812 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3813 * @key: is directly passed to the wakeup function
3814 */
3815void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3816                        int nr_exclusive, void *key)
3817{
3818        unsigned long flags;
3819
3820        spin_lock_irqsave(&q->lock, flags);
3821        __wake_up_common(q, mode, nr_exclusive, 0, key);
3822        spin_unlock_irqrestore(&q->lock, flags);
3823}
3824EXPORT_SYMBOL(__wake_up);
3825
3826/*
3827 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3828 */
3829void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3830{
3831        __wake_up_common(q, mode, 1, 0, NULL);
3832}
3833
3834/**
3835 * __wake_up_sync - wake up threads blocked on a waitqueue.
3836 * @q: the waitqueue
3837 * @mode: which threads
3838 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3839 *
3840 * The sync wakeup differs that the waker knows that it will schedule
3841 * away soon, so while the target thread will be woken up, it will not
3842 * be migrated to another CPU - ie. the two threads are 'synchronized'
3843 * with each other. This can prevent needless bouncing between CPUs.
3844 *
3845 * On UP it can prevent extra preemption.
3846 */
3847void fastcall
3848__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3849{
3850        unsigned long flags;
3851        int sync = 1;
3852
3853        if (unlikely(!q))
3854                return;
3855
3856        if (unlikely(!nr_exclusive))
3857                sync = 0;
3858
3859        spin_lock_irqsave(&q->lock, flags);
3860        __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3861        spin_unlock_irqrestore(&q->lock, flags);
3862}
3863EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
3864
3865void complete(struct completion *x)
3866{
3867        unsigned long flags;
3868
3869        spin_lock_irqsave(&x->wait.lock, flags);
3870        x->done++;
3871        __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3872                         1, 0, NULL);
3873        spin_unlock_irqrestore(&x->wait.lock, flags);
3874}
3875EXPORT_SYMBOL(complete);
3876
3877void complete_all(struct completion *x)
3878{
3879        unsigned long flags;
3880
3881        spin_lock_irqsave(&x->wait.lock, flags);
3882        x->done += UINT_MAX/2;
3883        __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3884                         0, 0, NULL);
3885        spin_unlock_irqrestore(&x->wait.lock, flags);
3886}
3887EXPORT_SYMBOL(complete_all);
3888
3889static inline long __sched
3890do_wait_for_common(struct completion *x, long timeout, int state)
3891{
3892        if (!x->done) {
3893                DECLARE_WAITQUEUE(wait, current);
3894
3895                wait.flags |= WQ_FLAG_EXCLUSIVE;
3896                __add_wait_queue_tail(&x->wait, &wait);
3897                do {
3898                        if (state == TASK_INTERRUPTIBLE &&
3899                            signal_pending(current)) {
3900                                __remove_wait_queue(&x->wait, &wait);
3901                                return -ERESTARTSYS;
3902                        }
3903                        __set_current_state(state);
3904                        spin_unlock_irq(&x->wait.lock);
3905                        timeout = schedule_timeout(timeout);
3906                        spin_lock_irq(&x->wait.lock);
3907                        if (!timeout) {
3908                                __remove_wait_queue(&x->wait, &wait);
3909                                return timeout;
3910                        }
3911                } while (!x->done);
3912                __remove_wait_queue(&x->wait, &wait);
3913        }
3914        x->done--;
3915        return timeout;
3916}
3917
3918static long __sched
3919wait_for_common(struct completion *x, long timeout, int state)
3920{
3921        might_sleep();
3922
3923        spin_lock_irq(&x->wait.lock);
3924        timeout = do_wait_for_common(x, timeout, state);
3925        spin_unlock_irq(&x->wait.lock);
3926        return timeout;
3927}
3928
3929void __sched wait_for_completion(struct completion *x)
3930{
3931        wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3932}
3933EXPORT_SYMBOL(wait_for_completion);
3934
3935unsigned long __sched
3936wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3937{
3938        return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3939}
3940EXPORT_SYMBOL(wait_for_completion_timeout);
3941
3942int __sched wait_for_completion_interruptible(struct completion *x)
3943{
3944        long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3945        if (t == -ERESTARTSYS)
3946                return t;
3947        return 0;
3948}
3949EXPORT_SYMBOL(wait_for_completion_interruptible);
3950
3951unsigned long __sched
3952wait_for_completion_interruptible_timeout(struct completion *x,
3953                                          unsigned long timeout)
3954{
3955        return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3956}
3957EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3958
3959static long __sched
3960sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3961{
3962        unsigned long flags;
3963        wait_queue_t wait;
3964
3965        init_waitqueue_entry(&wait, current);
3966
3967        __set_current_state(state);
3968
3969        spin_lock_irqsave(&q->lock, flags);
3970        __add_wait_queue(q, &wait);
3971        spin_unlock(&q->lock);
3972        timeout = schedule_timeout(timeout);
3973        spin_lock_irq(&q->lock);
3974        __remove_wait_queue(q, &wait);
3975        spin_unlock_irqrestore(&q->lock, flags);
3976
3977        return timeout;
3978}
3979
3980void __sched interruptible_sleep_on(wait_queue_head_t *q)
3981{
3982        sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3983}
3984EXPORT_SYMBOL(interruptible_sleep_on);
3985
3986long __sched
3987interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3988{
3989        return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3990}
3991EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3992
3993void __sched sleep_on(wait_queue_head_t *q)
3994{
3995        sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3996}
3997EXPORT_SYMBOL(sleep_on);
3998
3999long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4000{
4001        return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4002}
4003EXPORT_SYMBOL(sleep_on_timeout);
4004
4005#ifdef CONFIG_RT_MUTEXES
4006
4007/*
4008 * rt_mutex_setprio - set the current priority of a task
4009 * @p: task
4010 * @prio: prio value (kernel-internal form)
4011 *
4012 * This function changes the 'effective' priority of a task. It does
4013 * not touch ->normal_prio like __setscheduler().
4014 *
4015 * Used by the rt_mutex code to implement priority inheritance logic.
4016 */
4017void rt_mutex_setprio(struct task_struct *p, int prio)
4018{
4019        unsigned long flags;
4020        int oldprio, on_rq, running;
4021        struct rq *rq;
4022
4023        BUG_ON(prio < 0 || prio > MAX_PRIO);
4024
4025        rq = task_rq_lock(p, &flags);
4026        update_rq_clock(rq);
4027
4028        oldprio = p->prio;
4029        on_rq = p->se.on_rq;
4030        running = task_current(rq, p);
4031        if (on_rq) {
4032                dequeue_task(rq, p, 0);
4033                if (running)
4034                        p->sched_class->put_prev_task(rq, p);
4035        }
4036
4037        if (rt_prio(prio))
4038                p->sched_class = &rt_sched_class;
4039        else
4040                p->sched_class = &fair_sched_class;
4041
4042        p->prio = prio;
4043
4044        if (on_rq) {
4045                if (running)
4046                        p->sched_class->set_curr_task(rq);
4047                enqueue_task(rq, p, 0);
4048                /*
4049                 * Reschedule if we are currently running on this runqueue and
4050                 * our priority decreased, or if we are not currently running on
4051                 * this runqueue and our priority is higher than the current's
4052                 */
4053                if (running) {
4054                        if (p->prio > oldprio)
4055                                resched_task(rq->curr);
4056                } else {
4057                        check_preempt_curr(rq, p);
4058                }
4059        }
4060        task_rq_unlock(rq, &flags);
4061}
4062
4063#endif
4064
4065void set_user_nice(struct task_struct *p, long nice)
4066{
4067        int old_prio, delta, on_rq;
4068        unsigned long flags;
4069        struct rq *rq;
4070
4071        if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4072                return;
4073        /*
4074         * We have to be careful, if called from sys_setpriority(),
4075         * the task might be in the middle of scheduling on another CPU.
4076         */
4077        rq = task_rq_lock(p, &flags);
4078        update_rq_clock(rq);
4079        /*
4080         * The RT priorities are set via sched_setscheduler(), but we still
4081         * allow the 'normal' nice value to be set - but as expected
4082         * it wont have any effect on scheduling until the task is
4083         * SCHED_FIFO/SCHED_RR:
4084         */
4085        if (task_has_rt_policy(p)) {
4086                p->static_prio = NICE_TO_PRIO(nice);
4087                goto out_unlock;
4088        }
4089        on_rq = p->se.on_rq;
4090        if (on_rq) {
4091                dequeue_task(rq, p, 0);
4092                dec_load(rq, p);
4093        }
4094
4095        p->static_prio = NICE_TO_PRIO(nice);
4096        set_load_weight(p);
4097        old_prio = p->prio;
4098        p->prio = effective_prio(p);
4099        delta = p->prio - old_prio;
4100
4101        if (on_rq) {
4102                enqueue_task(rq, p, 0);
4103                inc_load(rq, p);
4104                /*
4105                 * If the task increased its priority or is running and
4106                 * lowered its priority, then reschedule its CPU:
4107                 */
4108                if (delta < 0 || (delta > 0 && task_running(rq, p)))
4109                        resched_task(rq->curr);
4110        }
4111out_unlock:
4112        task_rq_unlock(rq, &flags);
4113}
4114EXPORT_SYMBOL(set_user_nice);
4115
4116/*
4117 * can_nice - check if a task can reduce its nice value
4118 * @p: task
4119 * @nice: nice value
4120 */
4121int can_nice(const struct task_struct *p, const int nice)
4122{
4123        /* convert nice value [19,-20] to rlimit style value [1,40] */
4124        int nice_rlim = 20 - nice;
4125
4126        return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4127                capable(CAP_SYS_NICE));
4128}
4129
4130#ifdef __ARCH_WANT_SYS_NICE
4131
4132/*
4133 * sys_nice - change the priority of the current process.
4134 * @increment: priority increment
4135 *
4136 * sys_setpriority is a more generic, but much slower function that
4137 * does similar things.
4138 */
4139asmlinkage long sys_nice(int increment)
4140{
4141        long nice, retval;
4142
4143        /*
4144         * Setpriority might change our priority at the same moment.
4145         * We don't have to worry. Conceptually one call occurs first
4146         * and we have a single winner.
4147         */
4148        if (increment < -40)
4149                increment = -40;
4150        if (increment > 40)
4151                increment = 40;
4152
4153        nice = PRIO_TO_NICE(current->static_prio) + increment;
4154        if (nice < -20)
4155                nice = -20;
4156        if (nice > 19)
4157                nice = 19;
4158
4159        if (increment < 0 && !can_nice(current, nice))
4160                return -EPERM;
4161
4162        retval = security_task_setnice(current, nice);
4163        if (retval)
4164                return retval;
4165
4166        set_user_nice(current, nice);
4167        return 0;
4168}
4169
4170#endif
4171
4172/**
4173 * task_prio - return the priority value of a given task.
4174 * @p: the task in question.
4175 *
4176 * This is the priority value as seen by users in /proc.
4177 * RT tasks are offset by -200. Normal tasks are centered
4178 * around 0, value goes from -16 to +15.
4179 */
4180int task_prio(const struct task_struct *p)
4181{
4182        return p->prio - MAX_RT_PRIO;
4183}
4184
4185/**
4186 * task_nice - return the nice value of a given task.
4187 * @p: the task in question.
4188 */
4189int task_nice(const struct task_struct *p)
4190{
4191        return TASK_NICE(p);
4192}
4193EXPORT_SYMBOL_GPL(task_nice);
4194
4195/**
4196 * idle_cpu - is a given cpu idle currently?
4197 * @cpu: the processor in question.
4198 */
4199int idle_cpu(int cpu)
4200{
4201        return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4202}
4203
4204/**
4205 * idle_task - return the idle task for a given cpu.
4206 * @cpu: the processor in question.
4207 */
4208struct task_struct *idle_task(int cpu)
4209{
4210        return cpu_rq(cpu)->idle;
4211}
4212
4213/**
4214 * find_process_by_pid - find a process with a matching PID value.
4215 * @pid: the pid in question.
4216 */
4217static struct task_struct *find_process_by_pid(pid_t pid)
4218{
4219        return pid ? find_task_by_vpid(pid) : current;
4220}
4221
4222/* Actually do priority change: must hold rq lock. */
4223static void
4224__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4225{
4226        BUG_ON(p->se.on_rq);
4227
4228        p->policy = policy;
4229        switch (p->policy) {
4230        case SCHED_NORMAL:
4231        case SCHED_BATCH:
4232        case SCHED_IDLE:
4233                p->sched_class = &fair_sched_class;
4234                break;
4235        case SCHED_FIFO:
4236        case SCHED_RR:
4237                p->sched_class = &rt_sched_class;
4238                break;
4239        }
4240
4241        p->rt_priority = prio;
4242        p->normal_prio = normal_prio(p);
4243        /* we are holding p->pi_lock already */
4244        p->prio = rt_mutex_getprio(p);
4245        set_load_weight(p);
4246}
4247
4248/**
4249 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4250 * @p: the task in question.
4251 * @policy: new policy.
4252 * @param: structure containing the new RT priority.
4253 *
4254 * NOTE that the task may be already dead.
4255 */
4256int sched_setscheduler(struct task_struct *p, int policy,
4257                       struct sched_param *param)
4258{
4259        int retval, oldprio, oldpolicy = -1, on_rq, running;
4260        unsigned long flags;
4261        struct rq *rq;
4262
4263        /* may grab non-irq protected spin_locks */
4264        BUG_ON(in_interrupt());
4265recheck:
4266        /* double check policy once rq lock held */
4267        if (policy < 0)
4268                policy = oldpolicy = p->policy;
4269        else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4270                        policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4271                        policy != SCHED_IDLE)
4272                return -EINVAL;
4273        /*
4274         * Valid priorities for SCHED_FIFO and SCHED_RR are
4275         * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4276         * SCHED_BATCH and SCHED_IDLE is 0.
4277         */
4278        if (param->sched_priority < 0 ||
4279            (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4280            (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4281                return -EINVAL;
4282        if (rt_policy(policy) != (param->sched_priority != 0))
4283                return -EINVAL;
4284
4285        /*
4286         * Allow unprivileged RT tasks to decrease priority:
4287         */
4288        if (!capable(CAP_SYS_NICE)) {
4289                if (rt_policy(policy)) {
4290                        unsigned long rlim_rtprio;
4291
4292                        if (!lock_task_sighand(p, &flags))
4293                                return -ESRCH;
4294                        rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4295                        unlock_task_sighand(p, &flags);
4296
4297                        /* can't set/change the rt policy */
4298                        if (policy != p->policy && !rlim_rtprio)
4299                                return -EPERM;
4300
4301                        /* can't increase priority */
4302                        if (param->sched_priority > p->rt_priority &&
4303                            param->sched_priority > rlim_rtprio)
4304                                return -EPERM;
4305                }
4306                /*
4307                 * Like positive nice levels, dont allow tasks to
4308                 * move out of SCHED_IDLE either:
4309                 */
4310                if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4311                        return -EPERM;
4312
4313                /* can't change other user's priorities */
4314                if ((current->euid != p->euid) &&
4315                    (current->euid != p->uid))
4316                        return -EPERM;
4317        }
4318
4319        retval = security_task_setscheduler(p, policy, param);
4320        if (retval)
4321                return retval;
4322        /*
4323         * make sure no PI-waiters arrive (or leave) while we are
4324         * changing the priority of the task:
4325         */
4326        spin_lock_irqsave(&p->pi_lock, flags);
4327        /*
4328         * To be able to change p->policy safely, the apropriate
4329         * runqueue lock must be held.
4330         */
4331        rq = __task_rq_lock(p);
4332        /* recheck policy now with rq lock held */
4333        if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4334                policy = oldpolicy = -1;
4335                __task_rq_unlock(rq);
4336                spin_unlock_irqrestore(&p->pi_lock, flags);
4337                goto recheck;
4338        }
4339        update_rq_clock(rq);
4340        on_rq = p->se.on_rq;
4341        running = task_current(rq, p);
4342        if (on_rq) {
4343                deactivate_task(rq, p, 0);
4344                if (running)
4345                        p->sched_class->put_prev_task(rq, p);
4346        }
4347
4348        oldprio = p->prio;
4349        __setscheduler(rq, p, policy, param->sched_priority);
4350
4351        if (on_rq) {
4352                if (running)
4353                        p->sched_class->set_curr_task(rq);
4354                activate_task(rq, p, 0);
4355                /*
4356                 * Reschedule if we are currently running on this runqueue and
4357                 * our priority decreased, or if we are not currently running on
4358                 * this runqueue and our priority is higher than the current's
4359                 */
4360                if (running) {
4361                        if (p->prio > oldprio)
4362                                resched_task(rq->curr);
4363                } else {
4364                        check_preempt_curr(rq, p);
4365                }
4366        }
4367        __task_rq_unlock(rq);
4368        spin_unlock_irqrestore(&p->pi_lock, flags);
4369
4370        rt_mutex_adjust_pi(p);
4371
4372        return 0;
4373}
4374EXPORT_SYMBOL_GPL(sched_setscheduler);
4375
4376static int
4377do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4378{
4379        struct sched_param lparam;
4380        struct task_struct *p;
4381        int retval;
4382
4383        if (!param || pid < 0)
4384                return -EINVAL;
4385        if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4386                return -EFAULT;
4387
4388        rcu_read_lock();
4389        retval = -ESRCH;
4390        p = find_process_by_pid(pid);
4391        if (p != NULL)
4392                retval = sched_setscheduler(p, policy, &lparam);
4393        rcu_read_unlock();
4394
4395        return retval;
4396}
4397
4398/**
4399 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4400 * @pid: the pid in question.
4401 * @policy: new policy.
4402 * @param: structure containing the new RT priority.
4403 */
4404asmlinkage long
4405sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4406{
4407        /* negative values for policy are not valid */
4408        if (policy < 0)
4409                return -EINVAL;
4410
4411        return do_sched_setscheduler(pid, policy, param);
4412}
4413
4414/**
4415 * sys_sched_setparam - set/change the RT priority of a thread
4416 * @pid: the pid in question.
4417 * @param: structure containing the new RT priority.
4418 */
4419asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4420{
4421        return do_sched_setscheduler(pid, -1, param);
4422}
4423
4424/**
4425 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4426 * @pid: the pid in question.
4427 */
4428asmlinkage long sys_sched_getscheduler(pid_t pid)
4429{
4430        struct task_struct *p;
4431        int retval;
4432
4433        if (pid < 0)
4434                return -EINVAL;
4435
4436        retval = -ESRCH;
4437        read_lock(&tasklist_lock);
4438        p = find_process_by_pid(pid);
4439        if (p) {
4440                retval = security_task_getscheduler(p);
4441                if (!retval)
4442                        retval = p->policy;
4443        }
4444        read_unlock(&tasklist_lock);
4445        return retval;
4446}
4447
4448/**
4449 * sys_sched_getscheduler - get the RT priority of a thread
4450 * @pid: the pid in question.
4451 * @param: structure containing the RT priority.
4452 */
4453asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4454{
4455        struct sched_param lp;
4456        struct task_struct *p;
4457        int retval;
4458
4459        if (!param || pid < 0)
4460                return -EINVAL;
4461
4462        read_lock(&tasklist_lock);
4463        p = find_process_by_pid(pid);
4464        retval = -ESRCH;
4465        if (!p)
4466                goto out_unlock;
4467
4468        retval = security_task_getscheduler(p);
4469        if (retval)
4470                goto out_unlock;
4471
4472        lp.sched_priority = p->rt_priority;
4473        read_unlock(&tasklist_lock);
4474
4475        /*
4476         * This one might sleep, we cannot do it with a spinlock held ...
4477         */
4478        retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4479
4480        return retval;
4481
4482out_unlock:
4483        read_unlock(&tasklist_lock);
4484        return retval;
4485}
4486
4487long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4488{
4489        cpumask_t cpus_allowed;
4490        struct task_struct *p;
4491        int retval;
4492
4493        mutex_lock(&sched_hotcpu_mutex);
4494        read_lock(&tasklist_lock);
4495
4496        p = find_process_by_pid(pid);
4497        if (!p) {
4498                read_unlock(&tasklist_lock);
4499                mutex_unlock(&sched_hotcpu_mutex);
4500                return -ESRCH;
4501        }
4502
4503        /*
4504         * It is not safe to call set_cpus_allowed with the
4505         * tasklist_lock held. We will bump the task_struct's
4506         * usage count and then drop tasklist_lock.
4507         */
4508        get_task_struct(p);
4509        read_unlock(&tasklist_lock);
4510
4511        retval = -EPERM;
4512        if ((current->euid != p->euid) && (current->euid != p->uid) &&
4513                        !capable(CAP_SYS_NICE))
4514                goto out_unlock;
4515
4516        retval = security_task_setscheduler(p, 0, NULL);
4517        if (retval)
4518                goto out_unlock;
4519
4520        cpus_allowed = cpuset_cpus_allowed(p);
4521        cpus_and(new_mask, new_mask, cpus_allowed);
4522 again:
4523        retval = set_cpus_allowed(p, new_mask);
4524
4525        if (!retval) {
4526                cpus_allowed = cpuset_cpus_allowed(p);
4527                if (!cpus_subset(new_mask, cpus_allowed)) {
4528                        /*
4529                         * We must have raced with a concurrent cpuset
4530                         * update. Just reset the cpus_allowed to the
4531                         * cpuset's cpus_allowed
4532                         */
4533                        new_mask = cpus_allowed;
4534                        goto again;
4535                }
4536        }
4537out_unlock:
4538        put_task_struct(p);
4539        mutex_unlock(&sched_hotcpu_mutex);
4540        return retval;
4541}
4542
4543static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4544                             cpumask_t *new_mask)
4545{
4546        if (len < sizeof(cpumask_t)) {
4547                memset(new_mask, 0, sizeof(cpumask_t));
4548        } else if (len > sizeof(cpumask_t)) {
4549                len = sizeof(cpumask_t);
4550        }
4551        return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4552}
4553
4554/**
4555 * sys_sched_setaffinity - set the cpu affinity of a process
4556 * @pid: pid of the process
4557 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4558 * @user_mask_ptr: user-space pointer to the new cpu mask
4559 */
4560asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4561                                      unsigned long __user *user_mask_ptr)
4562{
4563        cpumask_t new_mask;
4564        int retval;
4565
4566        retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4567        if (retval)
4568                return retval;
4569
4570        return sched_setaffinity(pid, new_mask);
4571}
4572
4573/*
4574 * Represents all cpu's present in the system
4575 * In systems capable of hotplug, this map could dynamically grow
4576 * as new cpu's are detected in the system via any platform specific
4577 * method, such as ACPI for e.g.
4578 */
4579
4580cpumask_t cpu_present_map __read_mostly;
4581EXPORT_SYMBOL(cpu_present_map);
4582
4583#ifndef CONFIG_SMP
4584cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4585EXPORT_SYMBOL(cpu_online_map);
4586
4587cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4588EXPORT_SYMBOL(cpu_possible_map);
4589#endif
4590
4591long sched_getaffinity(pid_t pid, cpumask_t *mask)
4592{
4593        struct task_struct *p;
4594        int retval;
4595
4596        mutex_lock(&sched_hotcpu_mutex);
4597        read_lock(&tasklist_lock);
4598
4599        retval = -ESRCH;
4600        p = find_process_by_pid(pid);
4601        if (!p)
4602                goto out_unlock;
4603
4604        retval = security_task_getscheduler(p);
4605        if (retval)
4606                goto out_unlock;
4607
4608        cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4609
4610out_unlock:
4611        read_unlock(&tasklist_lock);
4612        mutex_unlock(&sched_hotcpu_mutex);
4613
4614        return retval;
4615}
4616
4617/**
4618 * sys_sched_getaffinity - get the cpu affinity of a process
4619 * @pid: pid of the process
4620 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4621 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4622 */
4623asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4624                                      unsigned long __user *user_mask_ptr)
4625{
4626        int ret;
4627        cpumask_t mask;
4628
4629        if (len < sizeof(cpumask_t))
4630                return -EINVAL;
4631
4632        ret = sched_getaffinity(pid, &mask);
4633        if (ret < 0)
4634                return ret;
4635
4636        if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4637                return -EFAULT;
4638
4639        return sizeof(cpumask_t);
4640}
4641
4642/**
4643 * sys_sched_yield - yield the current processor to other threads.
4644 *
4645 * This function yields the current CPU to other tasks. If there are no
4646 * other threads running on this CPU then this function will return.
4647 */
4648asmlinkage long sys_sched_yield(void)
4649{
4650        struct rq *rq = this_rq_lock();
4651
4652        schedstat_inc(rq, yld_count);
4653        current->sched_class->yield_task(rq);
4654
4655        /*
4656         * Since we are going to call schedule() anyway, there's
4657         * no need to preempt or enable interrupts:
4658         */
4659        __release(rq->lock);
4660        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4661        _raw_spin_unlock(&rq->lock);
4662        preempt_enable_no_resched();
4663
4664        schedule();
4665
4666        return 0;
4667}
4668
4669static void __cond_resched(void)
4670{
4671#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4672        __might_sleep(__FILE__, __LINE__);
4673#endif
4674        /*
4675         * The BKS might be reacquired before we have dropped
4676         * PREEMPT_ACTIVE, which could trigger a second
4677         * cond_resched() call.
4678         */
4679        do {
4680                add_preempt_count(PREEMPT_ACTIVE);
4681                schedule();
4682                sub_preempt_count(PREEMPT_ACTIVE);
4683        } while (need_resched());
4684}
4685
4686int __sched cond_resched(void)
4687{
4688        if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4689                                        system_state == SYSTEM_RUNNING) {
4690                __cond_resched();
4691                return 1;
4692        }
4693        return 0;
4694}
4695EXPORT_SYMBOL(cond_resched);
4696
4697/*
4698 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4699 * call schedule, and on return reacquire the lock.
4700 *
4701 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4702 * operations here to prevent schedule() from being called twice (once via
4703 * spin_unlock(), once by hand).
4704 */
4705int cond_resched_lock(spinlock_t *lock)
4706{
4707        int ret = 0;
4708
4709        if (need_lockbreak(lock)) {
4710                spin_unlock(lock);
4711                cpu_relax();
4712                ret = 1;
4713                spin_lock(lock);
4714        }
4715        if (need_resched() && system_state == SYSTEM_RUNNING) {
4716                spin_release(&lock->dep_map, 1, _THIS_IP_);
4717                _raw_spin_unlock(lock);
4718                preempt_enable_no_resched();
4719                __cond_resched();
4720                ret = 1;
4721                spin_lock(lock);
4722        }
4723        return ret;
4724}
4725EXPORT_SYMBOL(cond_resched_lock);
4726
4727int __sched cond_resched_softirq(void)
4728{
4729        BUG_ON(!in_softirq());
4730
4731        if (need_resched() && system_state == SYSTEM_RUNNING) {
4732                local_bh_enable();
4733                __cond_resched();
4734                local_bh_disable();
4735                return 1;
4736        }
4737        return 0;
4738}
4739EXPORT_SYMBOL(cond_resched_softirq);
4740
4741/**
4742 * yield - yield the current processor to other threads.
4743 *
4744 * This is a shortcut for kernel-space yielding - it marks the
4745 * thread runnable and calls sys_sched_yield().
4746 */
4747void __sched yield(void)
4748{
4749        set_current_state(TASK_RUNNING);
4750        sys_sched_yield();
4751}
4752EXPORT_SYMBOL(yield);
4753
4754/*
4755 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4756 * that process accounting knows that this is a task in IO wait state.
4757 *
4758 * But don't do that if it is a deliberate, throttling IO wait (this task
4759 * has set its backing_dev_info: the queue against which it should throttle)
4760 */
4761void __sched io_schedule(void)
4762{
4763        struct rq *rq = &__raw_get_cpu_var(runqueues);
4764
4765        delayacct_blkio_start();
4766        atomic_inc(&rq->nr_iowait);
4767        schedule();
4768        atomic_dec(&rq->nr_iowait);
4769        delayacct_blkio_end();
4770}
4771EXPORT_SYMBOL(io_schedule);
4772
4773long __sched io_schedule_timeout(long timeout)
4774{
4775        struct rq *rq = &__raw_get_cpu_var(runqueues);
4776        long ret;
4777
4778        delayacct_blkio_start();
4779        atomic_inc(&rq->nr_iowait);
4780        ret = schedule_timeout(timeout);
4781        atomic_dec(&rq->nr_iowait);
4782        delayacct_blkio_end();
4783        return ret;
4784}
4785
4786/**
4787 * sys_sched_get_priority_max - return maximum RT priority.
4788 * @policy: scheduling class.
4789 *
4790 * this syscall returns the maximum rt_priority that can be used
4791 * by a given scheduling class.
4792 */
4793asmlinkage long sys_sched_get_priority_max(int policy)
4794{
4795        int ret = -EINVAL;
4796
4797        switch (policy) {
4798        case SCHED_FIFO:
4799        case SCHED_RR:
4800                ret = MAX_USER_RT_PRIO-1;
4801                break;
4802        case SCHED_NORMAL:
4803        case SCHED_BATCH:
4804        case SCHED_IDLE:
4805                ret = 0;
4806                break;
4807        }
4808        return ret;
4809}
4810
4811/**
4812 * sys_sched_get_priority_min - return minimum RT priority.
4813 * @policy: scheduling class.
4814 *
4815 * this syscall returns the minimum rt_priority that can be used
4816 * by a given scheduling class.
4817 */
4818asmlinkage long sys_sched_get_priority_min(int policy)
4819{
4820        int ret = -EINVAL;
4821
4822        switch (policy) {
4823        case SCHED_FIFO:
4824        case SCHED_RR:
4825                ret = 1;
4826                break;
4827        case SCHED_NORMAL:
4828        case SCHED_BATCH:
4829        case SCHED_IDLE:
4830                ret = 0;
4831        }
4832        return ret;
4833}
4834
4835/**
4836 * sys_sched_rr_get_interval - return the default timeslice of a process.
4837 * @pid: pid of the process.
4838 * @interval: userspace pointer to the timeslice value.
4839 *
4840 * this syscall writes the default timeslice value of a given process
4841 * into the user-space timespec buffer. A value of '0' means infinity.
4842 */
4843asmlinkage
4844long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4845{
4846        struct task_struct *p;
4847        unsigned int time_slice;
4848        int retval;
4849        struct timespec t;
4850
4851        if (pid < 0)
4852                return -EINVAL;
4853
4854        retval = -ESRCH;
4855        read_lock(&tasklist_lock);
4856        p = find_process_by_pid(pid);
4857        if (!p)
4858                goto out_unlock;
4859
4860        retval = security_task_getscheduler(p);
4861        if (retval)
4862                goto out_unlock;
4863
4864        /*
4865         * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4866         * tasks that are on an otherwise idle runqueue:
4867         */
4868        time_slice = 0;
4869        if (p->policy == SCHED_RR) {
4870                time_slice = DEF_TIMESLICE;
4871        } else {
4872                struct sched_entity *se = &p->se;
4873                unsigned long flags;
4874                struct rq *rq;
4875
4876                rq = task_rq_lock(p, &flags);
4877                if (rq->cfs.load.weight)
4878                        time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4879                task_rq_unlock(rq, &flags);
4880        }
4881        read_unlock(&tasklist_lock);
4882        jiffies_to_timespec(time_slice, &t);
4883        retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4884        return retval;
4885
4886out_unlock:
4887        read_unlock(&tasklist_lock);
4888        return retval;
4889}
4890
4891static const char stat_nam[] = "RSDTtZX";
4892
4893static void show_task(struct task_struct *p)
4894{
4895        unsigned long free = 0;
4896        unsigned state;
4897
4898        state = p->state ? __ffs(p->state) + 1 : 0;
4899        printk(KERN_INFO "%-13.13s %c", p->comm,
4900                state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4901#if BITS_PER_LONG == 32
4902        if (state == TASK_RUNNING)
4903                printk(KERN_CONT " running  ");
4904        else
4905                printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4906#else
4907        if (state == TASK_RUNNING)
4908                printk(KERN_CONT "  running task    ");
4909        else
4910                printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4911#endif
4912#ifdef CONFIG_DEBUG_STACK_USAGE
4913        {
4914                unsigned long *n = end_of_stack(p);
4915                while (!*n)
4916                        n++;
4917                free = (unsigned long)n - (unsigned long)end_of_stack(p);
4918        }
4919#endif
4920        printk(KERN_CONT "%5lu %5d %6d\n", free,
4921                task_pid_nr(p), task_pid_nr(p->real_parent));
4922
4923        if (state != TASK_RUNNING)
4924                show_stack(p, NULL);
4925}
4926
4927void show_state_filter(unsigned long state_filter)
4928{
4929        struct task_struct *g, *p;
4930
4931#if BITS_PER_LONG == 32
4932        printk(KERN_INFO
4933                "  task                PC stack   pid father\n");
4934#else
4935        printk(KERN_INFO
4936                "  task                        PC stack   pid father\n");
4937#endif
4938        read_lock(&tasklist_lock);
4939        do_each_thread(g, p) {
4940                /*
4941                 * reset the NMI-timeout, listing all files on a slow
4942                 * console might take alot of time:
4943                 */
4944                touch_nmi_watchdog();
4945                if (!state_filter || (p->state & state_filter))
4946                        show_task(p);
4947        } while_each_thread(g, p);
4948
4949        touch_all_softlockup_watchdogs();
4950
4951#ifdef CONFIG_SCHED_DEBUG
4952        sysrq_sched_debug_show();
4953#endif
4954        read_unlock(&tasklist_lock);
4955        /*
4956         * Only show locks if all tasks are dumped:
4957         */
4958        if (state_filter == -1)
4959                debug_show_all_locks();
4960}
4961
4962void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4963{
4964        idle->sched_class = &idle_sched_class;
4965}
4966
4967/**
4968 * init_idle - set up an idle thread for a given CPU
4969 * @idle: task in question
4970 * @cpu: cpu the idle task belongs to
4971 *
4972 * NOTE: this function does not set the idle thread's NEED_RESCHED
4973 * flag, to make booting more robust.
4974 */
4975void __cpuinit init_idle(struct task_struct *idle, int cpu)
4976{
4977        struct rq *rq = cpu_rq(cpu);
4978        unsigned long flags;
4979
4980        __sched_fork(idle);
4981        idle->se.exec_start = sched_clock();
4982
4983        idle->prio = idle->normal_prio = MAX_PRIO;
4984        idle->cpus_allowed = cpumask_of_cpu(cpu);
4985        __set_task_cpu(idle, cpu);
4986
4987        spin_lock_irqsave(&rq->lock, flags);
4988        rq->curr = rq->idle = idle;
4989#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4990        idle->oncpu = 1;
4991#endif
4992        spin_unlock_irqrestore(&rq->lock, flags);
4993
4994        /* Set the preempt count _outside_ the spinlocks! */
4995#if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4996        task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4997#else
4998        task_thread_info(idle)->preempt_count = 0;
4999#endif
5000        /*
5001         * The idle tasks have their own, simple scheduling class:
5002         */
5003        idle->sched_class = &idle_sched_class;
5004}
5005
5006/*
5007 * In a system that switches off the HZ timer nohz_cpu_mask
5008 * indicates which cpus entered this state. This is used
5009 * in the rcu update to wait only for active cpus. For system
5010 * which do not switch off the HZ timer nohz_cpu_mask should
5011 * always be CPU_MASK_NONE.
5012 */
5013cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5014
5015/*
5016 * Increase the granularity value when there are more CPUs,
5017 * because with more CPUs the 'effective latency' as visible
5018 * to users decreases. But the relationship is not linear,
5019 * so pick a second-best guess by going with the log2 of the
5020 * number of CPUs.
5021 *
5022 * This idea comes from the SD scheduler of Con Kolivas:
5023 */
5024static inline void sched_init_granularity(void)
5025{
5026        unsigned int factor = 1 + ilog2(num_online_cpus());
5027        const unsigned long limit = 200000000;
5028
5029        sysctl_sched_min_granularity *= factor;
5030        if (sysctl_sched_min_granularity > limit)
5031                sysctl_sched_min_granularity = limit;
5032
5033        sysctl_sched_latency *= factor;
5034        if (sysctl_sched_latency > limit)
5035                sysctl_sched_latency = limit;
5036
5037        sysctl_sched_wakeup_granularity *= factor;
5038        sysctl_sched_batch_wakeup_granularity *= factor;
5039}
5040
5041#ifdef CONFIG_SMP
5042/*
5043 * This is how migration works:
5044 *
5045 * 1) we queue a struct migration_req structure in the source CPU's
5046 *    runqueue and wake up that CPU's migration thread.
5047 * 2) we down() the locked semaphore => thread blocks.
5048 * 3) migration thread wakes up (implicitly it forces the migrated
5049 *    thread off the CPU)
5050 * 4) it gets the migration request and checks whether the migrated
5051 *    task is still in the wrong runqueue.
5052 * 5) if it's in the wrong runqueue then the migration thread removes
5053 *    it and puts it into the right queue.
5054 * 6) migration thread up()s the semaphore.
5055 * 7) we wake up and the migration is done.
5056 */
5057
5058/*
5059 * Change a given task's CPU affinity. Migrate the thread to a
5060 * proper CPU and schedule it away if the CPU it's executing on
5061 * is removed from the allowed bitmask.
5062 *
5063 * NOTE: the caller must have a valid reference to the task, the
5064 * task must not exit() & deallocate itself prematurely. The
5065 * call is not atomic; no spinlocks may be held.
5066 */
5067int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5068{
5069        struct migration_req req;
5070        unsigned long flags;
5071        struct rq *rq;
5072        int ret = 0;
5073
5074        rq = task_rq_lock(p, &flags);
5075        if (!cpus_intersects(new_mask, cpu_online_map)) {
5076                ret = -EINVAL;
5077                goto out;
5078        }
5079
5080        p->cpus_allowed = new_mask;
5081        /* Can the task run on the task's current CPU? If so, we're done */
5082        if (cpu_isset(task_cpu(p), new_mask))
5083                goto out;
5084
5085        if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5086                /* Need help from migration thread: drop lock and wait. */
5087                task_rq_unlock(rq, &flags);
5088                wake_up_process(rq->migration_thread);
5089                wait_for_completion(&req.done);
5090                tlb_migrate_finish(p->mm);
5091                return 0;
5092        }
5093out:
5094        task_rq_unlock(rq, &flags);
5095
5096        return ret;
5097}
5098EXPORT_SYMBOL_GPL(set_cpus_allowed);
5099
5100/*
5101 * Move (not current) task off this cpu, onto dest cpu. We're doing
5102 * this because either it can't run here any more (set_cpus_allowed()
5103 * away from this CPU, or CPU going down), or because we're
5104 * attempting to rebalance this task on exec (sched_exec).
5105 *
5106 * So we race with normal scheduler movements, but that's OK, as long
5107 * as the task is no longer on this CPU.
5108 *
5109 * Returns non-zero if task was successfully migrated.
5110 */
5111static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5112{
5113        struct rq *rq_dest, *rq_src;
5114        int ret = 0, on_rq;
5115
5116        if (unlikely(cpu_is_offline(dest_cpu)))
5117                return ret;
5118
5119        rq_src = cpu_rq(src_cpu);
5120        rq_dest = cpu_rq(dest_cpu);
5121
5122        double_rq_lock(rq_src, rq_dest);
5123        /* Already moved. */
5124        if (task_cpu(p) != src_cpu)
5125                goto out;
5126        /* Affinity changed (again). */
5127        if (!cpu_isset(dest_cpu, p->cpus_allowed))
5128                goto out;
5129
5130        on_rq = p->se.on_rq;
5131        if (on_rq)
5132                deactivate_task(rq_src, p, 0);
5133
5134        set_task_cpu(p, dest_cpu);
5135        if (on_rq) {
5136                activate_task(rq_dest, p, 0);
5137                check_preempt_curr(rq_dest, p);
5138        }
5139        ret = 1;
5140out:
5141        double_rq_unlock(rq_src, rq_dest);
5142        return ret;
5143}
5144
5145/*
5146 * migration_thread - this is a highprio system thread that performs
5147 * thread migration by bumping thread off CPU then 'pushing' onto
5148 * another runqueue.
5149 */
5150static int migration_thread(void *data)
5151{
5152        int cpu = (long)data;
5153        struct rq *rq;
5154
5155        rq = cpu_rq(cpu);
5156        BUG_ON(rq->migration_thread != current);
5157
5158        set_current_state(TASK_INTERRUPTIBLE);
5159        while (!kthread_should_stop()) {
5160                struct migration_req *req;
5161                struct list_head *head;
5162
5163                spin_lock_irq(&rq->lock);
5164
5165                if (cpu_is_offline(cpu)) {
5166                        spin_unlock_irq(&rq->lock);
5167                        goto wait_to_die;
5168                }
5169
5170                if (rq->active_balance) {
5171                        active_load_balance(rq, cpu);
5172                        rq->active_balance = 0;
5173                }
5174
5175                head = &rq->migration_queue;
5176
5177                if (list_empty(head)) {
5178                        spin_unlock_irq(&rq->lock);
5179                        schedule();
5180                        set_current_state(TASK_INTERRUPTIBLE);
5181                        continue;
5182                }
5183                req = list_entry(head->next, struct migration_req, list);
5184                list_del_init(head->next);
5185
5186                spin_unlock(&rq->lock);
5187                __migrate_task(req->task, cpu, req->dest_cpu);
5188                local_irq_enable();
5189
5190                complete(&req->done);
5191        }
5192        __set_current_state(TASK_RUNNING);
5193        return 0;
5194
5195wait_to_die:
5196        /* Wait for kthread_stop */
5197        set_current_state(TASK_INTERRUPTIBLE);
5198        while (!kthread_should_stop()) {
5199                schedule();
5200                set_current_state(TASK_INTERRUPTIBLE);
5201        }
5202        __set_current_state(TASK_RUNNING);
5203        return 0;
5204}
5205
5206#ifdef CONFIG_HOTPLUG_CPU
5207
5208static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5209{
5210        int ret;
5211
5212        local_irq_disable();
5213        ret = __migrate_task(p, src_cpu, dest_cpu);
5214        local_irq_enable();
5215        return ret;
5216}
5217
5218/*
5219 * Figure out where task on dead CPU should go, use force if necessary.
5220 * NOTE: interrupts should be disabled by the caller
5221 */
5222static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5223{
5224        unsigned long flags;
5225        cpumask_t mask;
5226        struct rq *rq;
5227        int dest_cpu;
5228
5229        do {
5230                /* On same node? */
5231                mask = node_to_cpumask(cpu_to_node(dead_cpu));
5232                cpus_and(mask, mask, p->cpus_allowed);
5233                dest_cpu = any_online_cpu(mask);
5234
5235                /* On any allowed CPU? */
5236                if (dest_cpu == NR_CPUS)
5237                        dest_cpu = any_online_cpu(p->cpus_allowed);
5238
5239                /* No more Mr. Nice Guy. */
5240                if (dest_cpu == NR_CPUS) {
5241                        cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5242                        /*
5243                         * Try to stay on the same cpuset, where the
5244                         * current cpuset may be a subset of all cpus.
5245                         * The cpuset_cpus_allowed_locked() variant of
5246                         * cpuset_cpus_allowed() will not block. It must be
5247                         * called within calls to cpuset_lock/cpuset_unlock.
5248                         */
5249                        rq = task_rq_lock(p, &flags);
5250                        p->cpus_allowed = cpus_allowed;
5251                        dest_cpu = any_online_cpu(p->cpus_allowed);
5252                        task_rq_unlock(rq, &flags);
5253
5254                        /*
5255                         * Don't tell them about moving exiting tasks or
5256                         * kernel threads (both mm NULL), since they never
5257                         * leave kernel.
5258                         */
5259                        if (p->mm && printk_ratelimit()) {
5260                                printk(KERN_INFO "process %d (%s) no "
5261                                       "longer affine to cpu%d\n",
5262                                        task_pid_nr(p), p->comm, dead_cpu);
5263                        }
5264                }
5265        } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5266}
5267
5268/*
5269 * While a dead CPU has no uninterruptible tasks queued at this point,
5270 * it might still have a nonzero ->nr_uninterruptible counter, because
5271 * for performance reasons the counter is not stricly tracking tasks to
5272 * their home CPUs. So we just add the counter to another CPU's counter,
5273 * to keep the global sum constant after CPU-down:
5274 */
5275static void migrate_nr_uninterruptible(struct rq *rq_src)
5276{
5277        struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5278        unsigned long flags;
5279
5280        local_irq_save(flags);
5281        double_rq_lock(rq_src, rq_dest);
5282        rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5283        rq_src->nr_uninterruptible = 0;
5284        double_rq_unlock(rq_src, rq_dest);
5285        local_irq_restore(flags);
5286}
5287
5288/* Run through task list and migrate tasks from the dead cpu. */
5289static void migrate_live_tasks(int src_cpu)
5290{
5291        struct task_struct *p, *t;
5292
5293        read_lock(&tasklist_lock);
5294
5295        do_each_thread(t, p) {
5296                if (p == current)
5297                        continue;
5298
5299                if (task_cpu(p) == src_cpu)
5300                        move_task_off_dead_cpu(src_cpu, p);
5301        } while_each_thread(t, p);
5302
5303        read_unlock(&tasklist_lock);
5304}
5305
5306/*
5307 * Schedules idle task to be the next runnable task on current CPU.
5308 * It does so by boosting its priority to highest possible.
5309 * Used by CPU offline code.
5310 */
5311void sched_idle_next(void)
5312{
5313        int this_cpu = smp_processor_id();
5314        struct rq *rq = cpu_rq(this_cpu);
5315        struct task_struct *p = rq->idle;
5316        unsigned long flags;
5317
5318        /* cpu has to be offline */
5319        BUG_ON(cpu_online(this_cpu));
5320
5321        /*
5322         * Strictly not necessary since rest of the CPUs are stopped by now
5323         * and interrupts disabled on the current cpu.
5324         */
5325        spin_lock_irqsave(&rq->lock, flags);
5326
5327        __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5328
5329        update_rq_clock(rq);
5330        activate_task(rq, p, 0);
5331
5332        spin_unlock_irqrestore(&rq->lock, flags);
5333}
5334
5335/*
5336 * Ensures that the idle task is using init_mm right before its cpu goes
5337 * offline.
5338 */
5339void idle_task_exit(void)
5340{
5341        struct mm_struct *mm = current->active_mm;
5342
5343        BUG_ON(cpu_online(smp_processor_id()));
5344
5345        if (mm != &init_mm)
5346                switch_mm(mm, &init_mm, current);
5347        mmdrop(mm);
5348}
5349
5350/* called under rq->lock with disabled interrupts */
5351static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5352{
5353        struct rq *rq = cpu_rq(dead_cpu);
5354
5355        /* Must be exiting, otherwise would be on tasklist. */
5356        BUG_ON(!p->exit_state);
5357
5358        /* Cannot have done final schedule yet: would have vanished. */
5359        BUG_ON(p->state == TASK_DEAD);
5360
5361        get_task_struct(p);
5362
5363        /*
5364         * Drop lock around migration; if someone else moves it,
5365         * that's OK. No task can be added to this CPU, so iteration is
5366         * fine.
5367         */
5368        spin_unlock_irq(&rq->lock);
5369        move_task_off_dead_cpu(dead_cpu, p);
5370        spin_lock_irq(&rq->lock);
5371
5372        put_task_struct(p);
5373}
5374
5375/* release_task() removes task from tasklist, so we won't find dead tasks. */
5376static void migrate_dead_tasks(unsigned int dead_cpu)
5377{
5378        struct rq *rq = cpu_rq(dead_cpu);
5379        struct task_struct *next;
5380
5381        for ( ; ; ) {
5382                if (!rq->nr_running)
5383                        break;
5384                update_rq_clock(rq);
5385                next = pick_next_task(rq, rq->curr);
5386                if (!next)
5387                        break;
5388                migrate_dead(dead_cpu, next);
5389
5390        }
5391}
5392#endif /* CONFIG_HOTPLUG_CPU */
5393
5394#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5395
5396static struct ctl_table sd_ctl_dir[] = {
5397        {
5398                .procname       = "sched_domain",
5399                .mode           = 0555,
5400        },
5401        {0, },
5402};
5403
5404static struct ctl_table sd_ctl_root[] = {
5405        {
5406                .ctl_name       = CTL_KERN,
5407                .procname       = "kernel",
5408                .mode           = 0555,
5409                .child          = sd_ctl_dir,
5410        },
5411        {0, },
5412};
5413
5414static struct ctl_table *sd_alloc_ctl_entry(int n)
5415{
5416        struct ctl_table *entry =
5417                kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5418
5419        return entry;
5420}
5421
5422static void sd_free_ctl_entry(struct ctl_table **tablep)
5423{
5424        struct ctl_table *entry;
5425
5426        /*
5427         * In the intermediate directories, both the child directory and
5428         * procname are dynamically allocated and could fail but the mode
5429         * will always be set. In the lowest directory the names are
5430         * static strings and all have proc handlers.
5431         */
5432        for (entry = *tablep; entry->mode; entry++) {
5433                if (entry->child)
5434                        sd_free_ctl_entry(&entry->child);
5435                if (entry->proc_handler == NULL)
5436                        kfree(entry->procname);
5437        }
5438
5439        kfree(*tablep);
5440        *tablep = NULL;
5441}
5442
5443static void
5444set_table_entry(struct ctl_table *entry,
5445                const char *procname, void *data, int maxlen,
5446                mode_t mode, proc_handler *proc_handler)
5447{
5448        entry->procname = procname;
5449        entry->data = data;
5450        entry->maxlen = maxlen;
5451        entry->mode = mode;
5452        entry->proc_handler = proc_handler;
5453}
5454
5455static struct ctl_table *
5456sd_alloc_ctl_domain_table(struct sched_domain *sd)
5457{
5458        struct ctl_table *table = sd_alloc_ctl_entry(12);
5459
5460        if (table == NULL)
5461                return NULL;
5462
5463        set_table_entry(&table[0], "min_interval", &sd->min_interval,
5464                sizeof(long), 0644, proc_doulongvec_minmax);
5465        set_table_entry(&table[1], "max_interval", &sd->max_interval,
5466                sizeof(long), 0644, proc_doulongvec_minmax);
5467        set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5468                sizeof(int), 0644, proc_dointvec_minmax);
5469        set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5470                sizeof(int), 0644, proc_dointvec_minmax);
5471        set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5472                sizeof(int), 0644, proc_dointvec_minmax);
5473        set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5474                sizeof(int), 0644, proc_dointvec_minmax);
5475        set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5476                sizeof(int), 0644, proc_dointvec_minmax);
5477        set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5478                sizeof(int), 0644, proc_dointvec_minmax);
5479        set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5480                sizeof(int), 0644, proc_dointvec_minmax);
5481        set_table_entry(&table[9], "cache_nice_tries",
5482                &sd->cache_nice_tries,
5483                sizeof(int), 0644, proc_dointvec_minmax);
5484        set_table_entry(&table[10], "flags", &sd->flags,
5485                sizeof(int), 0644, proc_dointvec_minmax);
5486        /* &table[11] is terminator */
5487
5488        return table;
5489}
5490
5491static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5492{
5493        struct ctl_table *entry, *table;
5494        struct sched_domain *sd;
5495        int domain_num = 0, i;
5496        char buf[32];
5497
5498        for_each_domain(cpu, sd)
5499                domain_num++;
5500        entry = table = sd_alloc_ctl_entry(domain_num + 1);
5501        if (table == NULL)
5502                return NULL;
5503
5504        i = 0;
5505        for_each_domain(cpu, sd) {
5506                snprintf(buf, 32, "domain%d", i);
5507                entry->procname = kstrdup(buf, GFP_KERNEL);
5508                entry->mode = 0555;
5509                entry->child = sd_alloc_ctl_domain_table(sd);
5510                entry++;
5511                i++;
5512        }
5513        return table;
5514}
5515
5516static struct ctl_table_header *sd_sysctl_header;
5517static void register_sched_domain_sysctl(void)
5518{
5519        int i, cpu_num = num_online_cpus();
5520        struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5521        char buf[32];
5522
5523        WARN_ON(sd_ctl_dir[0].child);
5524        sd_ctl_dir[0].child = entry;
5525
5526        if (entry == NULL)
5527                return;
5528
5529        for_each_online_cpu(i) {
5530                snprintf(buf, 32, "cpu%d", i);
5531                entry->procname = kstrdup(buf, GFP_KERNEL);
5532                entry->mode = 0555;
5533                entry->child = sd_alloc_ctl_cpu_table(i);
5534                entry++;
5535        }
5536
5537        WARN_ON(sd_sysctl_header);
5538        sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5539}
5540
5541/* may be called multiple times per register */
5542static void unregister_sched_domain_sysctl(void)
5543{
5544        if (sd_sysctl_header)
5545                unregister_sysctl_table(sd_sysctl_header);
5546        sd_sysctl_header = NULL;
5547        if (sd_ctl_dir[0].child)
5548                sd_free_ctl_entry(&sd_ctl_dir[0].child);
5549}
5550#else
5551static void register_sched_domain_sysctl(void)
5552{
5553}
5554static void unregister_sched_domain_sysctl(void)
5555{
5556}
5557#endif
5558
5559/*
5560 * migration_call - callback that gets triggered when a CPU is added.
5561 * Here we can start up the necessary migration thread for the new CPU.
5562 */
5563static int __cpuinit
5564migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5565{
5566        struct task_struct *p;
5567        int cpu = (long)hcpu;
5568        unsigned long flags;
5569        struct rq *rq;
5570
5571        switch (action) {
5572        case CPU_LOCK_ACQUIRE:
5573                mutex_lock(&sched_hotcpu_mutex);
5574                break;
5575
5576        case CPU_UP_PREPARE:
5577        case CPU_UP_PREPARE_FROZEN:
5578                p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5579                if (IS_ERR(p))
5580                        return NOTIFY_BAD;
5581                kthread_bind(p, cpu);
5582                /* Must be high prio: stop_machine expects to yield to it. */
5583                rq = task_rq_lock(p, &flags);
5584                __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5585                task_rq_unlock(rq, &flags);
5586                cpu_rq(cpu)->migration_thread = p;
5587                break;
5588
5589        case CPU_ONLINE:
5590        case CPU_ONLINE_FROZEN:
5591                /* Strictly unnecessary, as first user will wake it. */
5592                wake_up_process(cpu_rq(cpu)->migration_thread);
5593                break;
5594
5595#ifdef CONFIG_HOTPLUG_CPU
5596        case CPU_UP_CANCELED:
5597        case CPU_UP_CANCELED_FROZEN:
5598                if (!cpu_rq(cpu)->migration_thread)
5599                        break;
5600                /* Unbind it from offline cpu so it can run. Fall thru. */
5601                kthread_bind(cpu_rq(cpu)->migration_thread,
5602                             any_online_cpu(cpu_online_map));
5603                kthread_stop(cpu_rq(cpu)->migration_thread);
5604                cpu_rq(cpu)->migration_thread = NULL;
5605                break;
5606
5607        case CPU_DEAD:
5608        case CPU_DEAD_FROZEN:
5609                cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5610                migrate_live_tasks(cpu);
5611                rq = cpu_rq(cpu);
5612                kthread_stop(rq->migration_thread);
5613                rq->migration_thread = NULL;
5614                /* Idle task back to normal (off runqueue, low prio) */
5615                spin_lock_irq(&rq->lock);
5616                update_rq_clock(rq);
5617                deactivate_task(rq, rq->idle, 0);
5618                rq->idle->static_prio = MAX_PRIO;
5619                __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5620                rq->idle->sched_class = &idle_sched_class;
5621                migrate_dead_tasks(cpu);
5622                spin_unlock_irq(&rq->lock);
5623                cpuset_unlock();
5624                migrate_nr_uninterruptible(rq);
5625                BUG_ON(rq->nr_running != 0);
5626
5627