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 *  2007-11-29  RT balancing improvements by Steven Rostedt, Gregory Haskins,
  26 *              Thomas Gleixner, Mike Kravetz
  27 */
  28
  29#include <linux/mm.h>
  30#include <linux/module.h>
  31#include <linux/nmi.h>
  32#include <linux/init.h>
  33#include <linux/uaccess.h>
  34#include <linux/highmem.h>
  35#include <linux/smp_lock.h>
  36#include <asm/mmu_context.h>
  37#include <linux/interrupt.h>
  38#include <linux/capability.h>
  39#include <linux/completion.h>
  40#include <linux/kernel_stat.h>
  41#include <linux/debug_locks.h>
  42#include <linux/security.h>
  43#include <linux/notifier.h>
  44#include <linux/profile.h>
  45#include <linux/freezer.h>
  46#include <linux/vmalloc.h>
  47#include <linux/blkdev.h>
  48#include <linux/delay.h>
  49#include <linux/pid_namespace.h>
  50#include <linux/smp.h>
  51#include <linux/threads.h>
  52#include <linux/timer.h>
  53#include <linux/rcupdate.h>
  54#include <linux/cpu.h>
  55#include <linux/cpuset.h>
  56#include <linux/percpu.h>
  57#include <linux/kthread.h>
  58#include <linux/seq_file.h>
  59#include <linux/sysctl.h>
  60#include <linux/syscalls.h>
  61#include <linux/times.h>
  62#include <linux/tsacct_kern.h>
  63#include <linux/kprobes.h>
  64#include <linux/delayacct.h>
  65#include <linux/reciprocal_div.h>
  66#include <linux/unistd.h>
  67#include <linux/pagemap.h>
  68#include <linux/hrtimer.h>
  69
  70#include <asm/tlb.h>
  71#include <asm/irq_regs.h>
  72
  73/*
  74 * Scheduler clock - returns current time in nanosec units.
  75 * This is default implementation.
  76 * Architectures and sub-architectures can override this.
  77 */
  78unsigned long long __attribute__((weak)) sched_clock(void)
  79{
  80        return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
  81}
  82
  83/*
  84 * Convert user-nice values [ -20 ... 0 ... 19 ]
  85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
  86 * and back.
  87 */
  88#define NICE_TO_PRIO(nice)      (MAX_RT_PRIO + (nice) + 20)
  89#define PRIO_TO_NICE(prio)      ((prio) - MAX_RT_PRIO - 20)
  90#define TASK_NICE(p)            PRIO_TO_NICE((p)->static_prio)
  91
  92/*
  93 * 'User priority' is the nice value converted to something we
  94 * can work with better when scaling various scheduler parameters,
  95 * it's a [ 0 ... 39 ] range.
  96 */
  97#define USER_PRIO(p)            ((p)-MAX_RT_PRIO)
  98#define TASK_USER_PRIO(p)       USER_PRIO((p)->static_prio)
  99#define MAX_USER_PRIO           (USER_PRIO(MAX_PRIO))
 100
 101/*
 102 * Helpers for converting nanosecond timing to jiffy resolution
 103 */
 104#define NS_TO_JIFFIES(TIME)     ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
 105
 106#define NICE_0_LOAD             SCHED_LOAD_SCALE
 107#define NICE_0_SHIFT            SCHED_LOAD_SHIFT
 108
 109/*
 110 * These are the 'tuning knobs' of the scheduler:
 111 *
 112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
 113 * Timeslices get refilled after they expire.
 114 */
 115#define DEF_TIMESLICE           (100 * HZ / 1000)
 116
 117#ifdef CONFIG_SMP
 118/*
 119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
 120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
 121 */
 122static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
 123{
 124        return reciprocal_divide(load, sg->reciprocal_cpu_power);
 125}
 126
 127/*
 128 * Each time a sched group cpu_power is changed,
 129 * we must compute its reciprocal value
 130 */
 131static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
 132{
 133        sg->__cpu_power += val;
 134        sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
 135}
 136#endif
 137
 138static inline int rt_policy(int policy)
 139{
 140        if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
 141                return 1;
 142        return 0;
 143}
 144
 145static inline int task_has_rt_policy(struct task_struct *p)
 146{
 147        return rt_policy(p->policy);
 148}
 149
 150/*
 151 * This is the priority-queue data structure of the RT scheduling class:
 152 */
 153struct rt_prio_array {
 154        DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
 155        struct list_head queue[MAX_RT_PRIO];
 156};
 157
 158#ifdef CONFIG_GROUP_SCHED
 159
 160#include <linux/cgroup.h>
 161
 162struct cfs_rq;
 163
 164static LIST_HEAD(task_groups);
 165
 166/* task group related information */
 167struct task_group {
 168#ifdef CONFIG_CGROUP_SCHED
 169        struct cgroup_subsys_state css;
 170#endif
 171
 172#ifdef CONFIG_FAIR_GROUP_SCHED
 173        /* schedulable entities of this group on each cpu */
 174        struct sched_entity **se;
 175        /* runqueue "owned" by this group on each cpu */
 176        struct cfs_rq **cfs_rq;
 177        unsigned long shares;
 178#endif
 179
 180#ifdef CONFIG_RT_GROUP_SCHED
 181        struct sched_rt_entity **rt_se;
 182        struct rt_rq **rt_rq;
 183
 184        u64 rt_runtime;
 185#endif
 186
 187        struct rcu_head rcu;
 188        struct list_head list;
 189};
 190
 191#ifdef CONFIG_FAIR_GROUP_SCHED
 192/* Default task group's sched entity on each cpu */
 193static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
 194/* Default task group's cfs_rq on each cpu */
 195static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
 196
 197static struct sched_entity *init_sched_entity_p[NR_CPUS];
 198static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
 199#endif
 200
 201#ifdef CONFIG_RT_GROUP_SCHED
 202static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
 203static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
 204
 205static struct sched_rt_entity *init_sched_rt_entity_p[NR_CPUS];
 206static struct rt_rq *init_rt_rq_p[NR_CPUS];
 207#endif
 208
 209/* task_group_lock serializes add/remove of task groups and also changes to
 210 * a task group's cpu shares.
 211 */
 212static DEFINE_SPINLOCK(task_group_lock);
 213
 214/* doms_cur_mutex serializes access to doms_cur[] array */
 215static DEFINE_MUTEX(doms_cur_mutex);
 216
 217#ifdef CONFIG_FAIR_GROUP_SCHED
 218#ifdef CONFIG_USER_SCHED
 219# define INIT_TASK_GROUP_LOAD   (2*NICE_0_LOAD)
 220#else
 221# define INIT_TASK_GROUP_LOAD   NICE_0_LOAD
 222#endif
 223
 224static int init_task_group_load = INIT_TASK_GROUP_LOAD;
 225#endif
 226
 227/* Default task group.
 228 *      Every task in system belong to this group at bootup.
 229 */
 230struct task_group init_task_group = {
 231#ifdef CONFIG_FAIR_GROUP_SCHED
 232        .se     = init_sched_entity_p,
 233        .cfs_rq = init_cfs_rq_p,
 234#endif
 235
 236#ifdef CONFIG_RT_GROUP_SCHED
 237        .rt_se  = init_sched_rt_entity_p,
 238        .rt_rq  = init_rt_rq_p,
 239#endif
 240};
 241
 242/* return group to which a task belongs */
 243static inline struct task_group *task_group(struct task_struct *p)
 244{
 245        struct task_group *tg;
 246
 247#ifdef CONFIG_USER_SCHED
 248        tg = p->user->tg;
 249#elif defined(CONFIG_CGROUP_SCHED)
 250        tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
 251                                struct task_group, css);
 252#else
 253        tg = &init_task_group;
 254#endif
 255        return tg;
 256}
 257
 258/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
 259static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
 260{
 261#ifdef CONFIG_FAIR_GROUP_SCHED
 262        p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
 263        p->se.parent = task_group(p)->se[cpu];
 264#endif
 265
 266#ifdef CONFIG_RT_GROUP_SCHED
 267        p->rt.rt_rq  = task_group(p)->rt_rq[cpu];
 268        p->rt.parent = task_group(p)->rt_se[cpu];
 269#endif
 270}
 271
 272static inline void lock_doms_cur(void)
 273{
 274        mutex_lock(&doms_cur_mutex);
 275}
 276
 277static inline void unlock_doms_cur(void)
 278{
 279        mutex_unlock(&doms_cur_mutex);
 280}
 281
 282#else
 283
 284static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
 285static inline void lock_doms_cur(void) { }
 286static inline void unlock_doms_cur(void) { }
 287
 288#endif  /* CONFIG_GROUP_SCHED */
 289
 290/* CFS-related fields in a runqueue */
 291struct cfs_rq {
 292        struct load_weight load;
 293        unsigned long nr_running;
 294
 295        u64 exec_clock;
 296        u64 min_vruntime;
 297
 298        struct rb_root tasks_timeline;
 299        struct rb_node *rb_leftmost;
 300        struct rb_node *rb_load_balance_curr;
 301        /* 'curr' points to currently running entity on this cfs_rq.
 302         * It is set to NULL otherwise (i.e when none are currently running).
 303         */
 304        struct sched_entity *curr, *next;
 305
 306        unsigned long nr_spread_over;
 307
 308#ifdef CONFIG_FAIR_GROUP_SCHED
 309        struct rq *rq;  /* cpu runqueue to which this cfs_rq is attached */
 310
 311        /*
 312         * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
 313         * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
 314         * (like users, containers etc.)
 315         *
 316         * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
 317         * list is used during load balance.
 318         */
 319        struct list_head leaf_cfs_rq_list;
 320        struct task_group *tg;  /* group that "owns" this runqueue */
 321#endif
 322};
 323
 324/* Real-Time classes' related field in a runqueue: */
 325struct rt_rq {
 326        struct rt_prio_array active;
 327        unsigned long rt_nr_running;
 328#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 329        int highest_prio; /* highest queued rt task prio */
 330#endif
 331#ifdef CONFIG_SMP
 332        unsigned long rt_nr_migratory;
 333        int overloaded;
 334#endif
 335        int rt_throttled;
 336        u64 rt_time;
 337
 338#ifdef CONFIG_RT_GROUP_SCHED
 339        unsigned long rt_nr_boosted;
 340
 341        struct rq *rq;
 342        struct list_head leaf_rt_rq_list;
 343        struct task_group *tg;
 344        struct sched_rt_entity *rt_se;
 345#endif
 346};
 347
 348#ifdef CONFIG_SMP
 349
 350/*
 351 * We add the notion of a root-domain which will be used to define per-domain
 352 * variables. Each exclusive cpuset essentially defines an island domain by
 353 * fully partitioning the member cpus from any other cpuset. Whenever a new
 354 * exclusive cpuset is created, we also create and attach a new root-domain
 355 * object.
 356 *
 357 */
 358struct root_domain {
 359        atomic_t refcount;
 360        cpumask_t span;
 361        cpumask_t online;
 362
 363        /*
 364         * The "RT overload" flag: it gets set if a CPU has more than
 365         * one runnable RT task.
 366         */
 367        cpumask_t rto_mask;
 368        atomic_t rto_count;
 369};
 370
 371/*
 372 * By default the system creates a single root-domain with all cpus as
 373 * members (mimicking the global state we have today).
 374 */
 375static struct root_domain def_root_domain;
 376
 377#endif
 378
 379/*
 380 * This is the main, per-CPU runqueue data structure.
 381 *
 382 * Locking rule: those places that want to lock multiple runqueues
 383 * (such as the load balancing or the thread migration code), lock
 384 * acquire operations must be ordered by ascending &runqueue.
 385 */
 386struct rq {
 387        /* runqueue lock: */
 388        spinlock_t lock;
 389
 390        /*
 391         * nr_running and cpu_load should be in the same cacheline because
 392         * remote CPUs use both these fields when doing load calculation.
 393         */
 394        unsigned long nr_running;
 395        #define CPU_LOAD_IDX_MAX 5
 396        unsigned long cpu_load[CPU_LOAD_IDX_MAX];
 397        unsigned char idle_at_tick;
 398#ifdef CONFIG_NO_HZ
 399        unsigned char in_nohz_recently;
 400#endif
 401        /* capture load from *all* tasks on this cpu: */
 402        struct load_weight load;
 403        unsigned long nr_load_updates;
 404        u64 nr_switches;
 405
 406        struct cfs_rq cfs;
 407        struct rt_rq rt;
 408        u64 rt_period_expire;
 409        int rt_throttled;
 410
 411#ifdef CONFIG_FAIR_GROUP_SCHED
 412        /* list of leaf cfs_rq on this cpu: */
 413        struct list_head leaf_cfs_rq_list;
 414#endif
 415#ifdef CONFIG_RT_GROUP_SCHED
 416        struct list_head leaf_rt_rq_list;
 417#endif
 418
 419        /*
 420         * This is part of a global counter where only the total sum
 421         * over all CPUs matters. A task can increase this counter on
 422         * one CPU and if it got migrated afterwards it may decrease
 423         * it on another CPU. Always updated under the runqueue lock:
 424         */
 425        unsigned long nr_uninterruptible;
 426
 427        struct task_struct *curr, *idle;
 428        unsigned long next_balance;
 429        struct mm_struct *prev_mm;
 430
 431        u64 clock, prev_clock_raw;
 432        s64 clock_max_delta;
 433
 434        unsigned int clock_warps, clock_overflows, clock_underflows;
 435        u64 idle_clock;
 436        unsigned int clock_deep_idle_events;
 437        u64 tick_timestamp;
 438
 439        atomic_t nr_iowait;
 440
 441#ifdef CONFIG_SMP
 442        struct root_domain *rd;
 443        struct sched_domain *sd;
 444
 445        /* For active balancing */
 446        int active_balance;
 447        int push_cpu;
 448        /* cpu of this runqueue: */
 449        int cpu;
 450
 451        struct task_struct *migration_thread;
 452        struct list_head migration_queue;
 453#endif
 454
 455#ifdef CONFIG_SCHED_HRTICK
 456        unsigned long hrtick_flags;
 457        ktime_t hrtick_expire;
 458        struct hrtimer hrtick_timer;
 459#endif
 460
 461#ifdef CONFIG_SCHEDSTATS
 462        /* latency stats */
 463        struct sched_info rq_sched_info;
 464
 465        /* sys_sched_yield() stats */
 466        unsigned int yld_exp_empty;
 467        unsigned int yld_act_empty;
 468        unsigned int yld_both_empty;
 469        unsigned int yld_count;
 470
 471        /* schedule() stats */
 472        unsigned int sched_switch;
 473        unsigned int sched_count;
 474        unsigned int sched_goidle;
 475
 476        /* try_to_wake_up() stats */
 477        unsigned int ttwu_count;
 478        unsigned int ttwu_local;
 479
 480        /* BKL stats */
 481        unsigned int bkl_count;
 482#endif
 483        struct lock_class_key rq_lock_key;
 484};
 485
 486static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
 487
 488static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
 489{
 490        rq->curr->sched_class->check_preempt_curr(rq, p);
 491}
 492
 493static inline int cpu_of(struct rq *rq)
 494{
 495#ifdef CONFIG_SMP
 496        return rq->cpu;
 497#else
 498        return 0;
 499#endif
 500}
 501
 502/*
 503 * Update the per-runqueue clock, as finegrained as the platform can give
 504 * us, but without assuming monotonicity, etc.:
 505 */
 506static void __update_rq_clock(struct rq *rq)
 507{
 508        u64 prev_raw = rq->prev_clock_raw;
 509        u64 now = sched_clock();
 510        s64 delta = now - prev_raw;
 511        u64 clock = rq->clock;
 512
 513#ifdef CONFIG_SCHED_DEBUG
 514        WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
 515#endif
 516        /*
 517         * Protect against sched_clock() occasionally going backwards:
 518         */
 519        if (unlikely(delta < 0)) {
 520                clock++;
 521                rq->clock_warps++;
 522        } else {
 523                /*
 524                 * Catch too large forward jumps too:
 525                 */
 526                if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
 527                        if (clock < rq->tick_timestamp + TICK_NSEC)
 528                                clock = rq->tick_timestamp + TICK_NSEC;
 529                        else
 530                                clock++;
 531                        rq->clock_overflows++;
 532                } else {
 533                        if (unlikely(delta > rq->clock_max_delta))
 534                                rq->clock_max_delta = delta;
 535                        clock += delta;
 536                }
 537        }
 538
 539        rq->prev_clock_raw = now;
 540        rq->clock = clock;
 541}
 542
 543static void update_rq_clock(struct rq *rq)
 544{
 545        if (likely(smp_processor_id() == cpu_of(rq)))
 546                __update_rq_clock(rq);
 547}
 548
 549/*
 550 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
 551 * See detach_destroy_domains: synchronize_sched for details.
 552 *
 553 * The domain tree of any CPU may only be accessed from within
 554 * preempt-disabled sections.
 555 */
 556#define for_each_domain(cpu, __sd) \
 557        for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
 558
 559#define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
 560#define this_rq()               (&__get_cpu_var(runqueues))
 561#define task_rq(p)              cpu_rq(task_cpu(p))
 562#define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
 563
 564unsigned long rt_needs_cpu(int cpu)
 565{
 566        struct rq *rq = cpu_rq(cpu);
 567        u64 delta;
 568
 569        if (!rq->rt_throttled)
 570                return 0;
 571
 572        if (rq->clock > rq->rt_period_expire)
 573                return 1;
 574
 575        delta = rq->rt_period_expire - rq->clock;
 576        do_div(delta, NSEC_PER_SEC / HZ);
 577
 578        return (unsigned long)delta;
 579}
 580
 581/*
 582 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
 583 */
 584#ifdef CONFIG_SCHED_DEBUG
 585# define const_debug __read_mostly
 586#else
 587# define const_debug static const
 588#endif
 589
 590/*
 591 * Debugging: various feature bits
 592 */
 593enum {
 594        SCHED_FEAT_NEW_FAIR_SLEEPERS    = 1,
 595        SCHED_FEAT_WAKEUP_PREEMPT       = 2,
 596        SCHED_FEAT_START_DEBIT          = 4,
 597        SCHED_FEAT_HRTICK               = 8,
 598        SCHED_FEAT_DOUBLE_TICK          = 16,
 599};
 600
 601const_debug unsigned int sysctl_sched_features =
 602                SCHED_FEAT_NEW_FAIR_SLEEPERS    * 1 |
 603                SCHED_FEAT_WAKEUP_PREEMPT       * 1 |
 604                SCHED_FEAT_START_DEBIT          * 1 |
 605                SCHED_FEAT_HRTICK               * 1 |
 606                SCHED_FEAT_DOUBLE_TICK          * 0;
 607
 608#define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
 609
 610/*
 611 * Number of tasks to iterate in a single balance run.
 612 * Limited because this is done with IRQs disabled.
 613 */
 614const_debug unsigned int sysctl_sched_nr_migrate = 32;
 615
 616/*
 617 * period over which we measure -rt task cpu usage in us.
 618 * default: 1s
 619 */
 620unsigned int sysctl_sched_rt_period = 1000000;
 621
 622static __read_mostly int scheduler_running;
 623
 624/*
 625 * part of the period that we allow rt tasks to run in us.
 626 * default: 0.95s
 627 */
 628int sysctl_sched_rt_runtime = 950000;
 629
 630/*
 631 * single value that denotes runtime == period, ie unlimited time.
 632 */
 633#define RUNTIME_INF     ((u64)~0ULL)
 634
 635/*
 636 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
 637 * clock constructed from sched_clock():
 638 */
 639unsigned long long cpu_clock(int cpu)
 640{
 641        unsigned long long now;
 642        unsigned long flags;
 643        struct rq *rq;
 644
 645        /*
 646         * Only call sched_clock() if the scheduler has already been
 647         * initialized (some code might call cpu_clock() very early):
 648         */
 649        if (unlikely(!scheduler_running))
 650                return 0;
 651
 652        local_irq_save(flags);
 653        rq = cpu_rq(cpu);
 654        update_rq_clock(rq);
 655        now = rq->clock;
 656        local_irq_restore(flags);
 657
 658        return now;
 659}
 660EXPORT_SYMBOL_GPL(cpu_clock);
 661
 662#ifndef prepare_arch_switch
 663# define prepare_arch_switch(next)      do { } while (0)
 664#endif
 665#ifndef finish_arch_switch
 666# define finish_arch_switch(prev)       do { } while (0)
 667#endif
 668
 669static inline int task_current(struct rq *rq, struct task_struct *p)
 670{
 671        return rq->curr == p;
 672}
 673
 674#ifndef __ARCH_WANT_UNLOCKED_CTXSW
 675static inline int task_running(struct rq *rq, struct task_struct *p)
 676{
 677        return task_current(rq, p);
 678}
 679
 680static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
 681{
 682}
 683
 684static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
 685{
 686#ifdef CONFIG_DEBUG_SPINLOCK
 687        /* this is a valid case when another task releases the spinlock */
 688        rq->lock.owner = current;
 689#endif
 690        /*
 691         * If we are tracking spinlock dependencies then we have to
 692         * fix up the runqueue lock - which gets 'carried over' from
 693         * prev into current:
 694         */
 695        spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
 696
 697        spin_unlock_irq(&rq->lock);
 698}
 699
 700#else /* __ARCH_WANT_UNLOCKED_CTXSW */
 701static inline int task_running(struct rq *rq, struct task_struct *p)
 702{
 703#ifdef CONFIG_SMP
 704        return p->oncpu;
 705#else
 706        return task_current(rq, p);
 707#endif
 708}
 709
 710static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
 711{
 712#ifdef CONFIG_SMP
 713        /*
 714         * We can optimise this out completely for !SMP, because the
 715         * SMP rebalancing from interrupt is the only thing that cares
 716         * here.
 717         */
 718        next->oncpu = 1;
 719#endif
 720#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
 721        spin_unlock_irq(&rq->lock);
 722#else
 723        spin_unlock(&rq->lock);
 724#endif
 725}
 726
 727static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
 728{
 729#ifdef CONFIG_SMP
 730        /*
 731         * After ->oncpu is cleared, the task can be moved to a different CPU.
 732         * We must ensure this doesn't happen until the switch is completely
 733         * finished.
 734         */
 735        smp_wmb();
 736        prev->oncpu = 0;
 737#endif
 738#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
 739        local_irq_enable();
 740#endif
 741}
 742#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
 743
 744/*
 745 * __task_rq_lock - lock the runqueue a given task resides on.
 746 * Must be called interrupts disabled.
 747 */
 748static inline struct rq *__task_rq_lock(struct task_struct *p)
 749        __acquires(rq->lock)
 750{
 751        for (;;) {
 752                struct rq *rq = task_rq(p);
 753                spin_lock(&rq->lock);
 754                if (likely(rq == task_rq(p)))
 755                        return rq;
 756                spin_unlock(&rq->lock);
 757        }
 758}
 759
 760/*
 761 * task_rq_lock - lock the runqueue a given task resides on and disable
 762 * interrupts. Note the ordering: we can safely lookup the task_rq without
 763 * explicitly disabling preemption.
 764 */
 765static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
 766        __acquires(rq->lock)
 767{
 768        struct rq *rq;
 769
 770        for (;;) {
 771                local_irq_save(*flags);
 772                rq = task_rq(p);
 773                spin_lock(&rq->lock);
 774                if (likely(rq == task_rq(p)))
 775                        return rq;
 776                spin_unlock_irqrestore(&rq->lock, *flags);
 777        }
 778}
 779
 780static void __task_rq_unlock(struct rq *rq)
 781        __releases(rq->lock)
 782{
 783        spin_unlock(&rq->lock);
 784}
 785
 786static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
 787        __releases(rq->lock)
 788{
 789        spin_unlock_irqrestore(&rq->lock, *flags);
 790}
 791
 792/*
 793 * this_rq_lock - lock this runqueue and disable interrupts.
 794 */
 795static struct rq *this_rq_lock(void)
 796        __acquires(rq->lock)
 797{
 798        struct rq *rq;
 799
 800        local_irq_disable();
 801        rq = this_rq();
 802        spin_lock(&rq->lock);
 803
 804        return rq;
 805}
 806
 807/*
 808 * We are going deep-idle (irqs are disabled):
 809 */
 810void sched_clock_idle_sleep_event(void)
 811{
 812        struct rq *rq = cpu_rq(smp_processor_id());
 813
 814        spin_lock(&rq->lock);
 815        __update_rq_clock(rq);
 816        spin_unlock(&rq->lock);
 817        rq->clock_deep_idle_events++;
 818}
 819EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
 820
 821/*
 822 * We just idled delta nanoseconds (called with irqs disabled):
 823 */
 824void sched_clock_idle_wakeup_event(u64 delta_ns)
 825{
 826        struct rq *rq = cpu_rq(smp_processor_id());
 827        u64 now = sched_clock();
 828
 829        rq->idle_clock += delta_ns;
 830        /*
 831         * Override the previous timestamp and ignore all
 832         * sched_clock() deltas that occured while we idled,
 833         * and use the PM-provided delta_ns to advance the
 834         * rq clock:
 835         */
 836        spin_lock(&rq->lock);
 837        rq->prev_clock_raw = now;
 838        rq->clock += delta_ns;
 839        spin_unlock(&rq->lock);
 840        touch_softlockup_watchdog();
 841}
 842EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
 843
 844static void __resched_task(struct task_struct *p, int tif_bit);
 845
 846static inline void resched_task(struct task_struct *p)
 847{
 848        __resched_task(p, TIF_NEED_RESCHED);
 849}
 850
 851#ifdef CONFIG_SCHED_HRTICK
 852/*
 853 * Use HR-timers to deliver accurate preemption points.
 854 *
 855 * Its all a bit involved since we cannot program an hrt while holding the
 856 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
 857 * reschedule event.
 858 *
 859 * When we get rescheduled we reprogram the hrtick_timer outside of the
 860 * rq->lock.
 861 */
 862static inline void resched_hrt(struct task_struct *p)
 863{
 864        __resched_task(p, TIF_HRTICK_RESCHED);
 865}
 866
 867static inline void resched_rq(struct rq *rq)
 868{
 869        unsigned long flags;
 870
 871        spin_lock_irqsave(&rq->lock, flags);
 872        resched_task(rq->curr);
 873        spin_unlock_irqrestore(&rq->lock, flags);
 874}
 875
 876enum {
 877        HRTICK_SET,             /* re-programm hrtick_timer */
 878        HRTICK_RESET,           /* not a new slice */
 879        HRTICK_BLOCK,           /* stop hrtick operations */
 880};
 881
 882/*
 883 * Use hrtick when:
 884 *  - enabled by features
 885 *  - hrtimer is actually high res
 886 */
 887static inline int hrtick_enabled(struct rq *rq)
 888{
 889        if (!sched_feat(HRTICK))
 890                return 0;
 891        if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
 892                return 0;
 893        return hrtimer_is_hres_active(&rq->hrtick_timer);
 894}
 895
 896/*
 897 * Called to set the hrtick timer state.
 898 *
 899 * called with rq->lock held and irqs disabled
 900 */
 901static void hrtick_start(struct rq *rq, u64 delay, int reset)
 902{
 903        assert_spin_locked(&rq->lock);
 904
 905        /*
 906         * preempt at: now + delay
 907         */
 908        rq->hrtick_expire =
 909                ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
 910        /*
 911         * indicate we need to program the timer
 912         */
 913        __set_bit(HRTICK_SET, &rq->hrtick_flags);
 914        if (reset)
 915                __set_bit(HRTICK_RESET, &rq->hrtick_flags);
 916
 917        /*
 918         * New slices are called from the schedule path and don't need a
 919         * forced reschedule.
 920         */
 921        if (reset)
 922                resched_hrt(rq->curr);
 923}
 924
 925static void hrtick_clear(struct rq *rq)
 926{
 927        if (hrtimer_active(&rq->hrtick_timer))
 928                hrtimer_cancel(&rq->hrtick_timer);
 929}
 930
 931/*
 932 * Update the timer from the possible pending state.
 933 */
 934static void hrtick_set(struct rq *rq)
 935{
 936        ktime_t time;
 937        int set, reset;
 938        unsigned long flags;
 939
 940        WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
 941
 942        spin_lock_irqsave(&rq->lock, flags);
 943        set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
 944        reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
 945        time = rq->hrtick_expire;
 946        clear_thread_flag(TIF_HRTICK_RESCHED);
 947        spin_unlock_irqrestore(&rq->lock, flags);
 948
 949        if (set) {
 950                hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
 951                if (reset && !hrtimer_active(&rq->hrtick_timer))
 952                        resched_rq(rq);
 953        } else
 954                hrtick_clear(rq);
 955}
 956
 957/*
 958 * High-resolution timer tick.
 959 * Runs from hardirq context with interrupts disabled.
 960 */
 961static enum hrtimer_restart hrtick(struct hrtimer *timer)
 962{
 963        struct rq *rq = container_of(timer, struct rq, hrtick_timer);
 964
 965        WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
 966
 967        spin_lock(&rq->lock);
 968        __update_rq_clock(rq);
 969        rq->curr->sched_class->task_tick(rq, rq->curr, 1);
 970        spin_unlock(&rq->lock);
 971
 972        return HRTIMER_NORESTART;
 973}
 974
 975static void hotplug_hrtick_disable(int cpu)
 976{
 977        struct rq *rq = cpu_rq(cpu);
 978        unsigned long flags;
 979
 980        spin_lock_irqsave(&rq->lock, flags);
 981        rq->hrtick_flags = 0;
 982        __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
 983        spin_unlock_irqrestore(&rq->lock, flags);
 984
 985        hrtick_clear(rq);
 986}
 987
 988static void hotplug_hrtick_enable(int cpu)
 989{
 990        struct rq *rq = cpu_rq(cpu);
 991        unsigned long flags;
 992
 993        spin_lock_irqsave(&rq->lock, flags);
 994        __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
 995        spin_unlock_irqrestore(&rq->lock, flags);
 996}
 997
 998static int
 999hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1000{
1001        int cpu = (int)(long)hcpu;
1002
1003        switch (action) {
1004        case CPU_UP_CANCELED:
1005        case CPU_UP_CANCELED_FROZEN:
1006        case CPU_DOWN_PREPARE:
1007        case CPU_DOWN_PREPARE_FROZEN:
1008        case CPU_DEAD:
1009        case CPU_DEAD_FROZEN:
1010                hotplug_hrtick_disable(cpu);
1011                return NOTIFY_OK;
1012
1013        case CPU_UP_PREPARE:
1014        case CPU_UP_PREPARE_FROZEN:
1015        case CPU_DOWN_FAILED:
1016        case CPU_DOWN_FAILED_FROZEN:
1017        case CPU_ONLINE:
1018        case CPU_ONLINE_FROZEN:
1019                hotplug_hrtick_enable(cpu);
1020                return NOTIFY_OK;
1021        }
1022
1023        return NOTIFY_DONE;
1024}
1025
1026static void init_hrtick(void)
1027{
1028        hotcpu_notifier(hotplug_hrtick, 0);
1029}
1030
1031static void init_rq_hrtick(struct rq *rq)
1032{
1033        rq->hrtick_flags = 0;
1034        hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1035        rq->hrtick_timer.function = hrtick;
1036        rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1037}
1038
1039void hrtick_resched(void)
1040{
1041        struct rq *rq;
1042        unsigned long flags;
1043
1044        if (!test_thread_flag(TIF_HRTICK_RESCHED))
1045                return;
1046
1047        local_irq_save(flags);
1048        rq = cpu_rq(smp_processor_id());
1049        hrtick_set(rq);
1050        local_irq_restore(flags);
1051}
1052#else
1053static inline void hrtick_clear(struct rq *rq)
1054{
1055}
1056
1057static inline void hrtick_set(struct rq *rq)
1058{
1059}
1060
1061static inline void init_rq_hrtick(struct rq *rq)
1062{
1063}
1064
1065void hrtick_resched(void)
1066{
1067}
1068
1069static inline void init_hrtick(void)
1070{
1071}
1072#endif
1073
1074/*
1075 * resched_task - mark a task 'to be rescheduled now'.
1076 *
1077 * On UP this means the setting of the need_resched flag, on SMP it
1078 * might also involve a cross-CPU call to trigger the scheduler on
1079 * the target CPU.
1080 */
1081#ifdef CONFIG_SMP
1082
1083#ifndef tsk_is_polling
1084#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1085#endif
1086
1087static void __resched_task(struct task_struct *p, int tif_bit)
1088{
1089        int cpu;
1090
1091        assert_spin_locked(&task_rq(p)->lock);
1092
1093        if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1094                return;
1095
1096        set_tsk_thread_flag(p, tif_bit);
1097
1098        cpu = task_cpu(p);
1099        if (cpu == smp_processor_id())
1100                return;
1101
1102        /* NEED_RESCHED must be visible before we test polling */
1103        smp_mb();
1104        if (!tsk_is_polling(p))
1105                smp_send_reschedule(cpu);
1106}
1107
1108static void resched_cpu(int cpu)
1109{
1110        struct rq *rq = cpu_rq(cpu);
1111        unsigned long flags;
1112
1113        if (!spin_trylock_irqsave(&rq->lock, flags))
1114                return;
1115        resched_task(cpu_curr(cpu));
1116        spin_unlock_irqrestore(&rq->lock, flags);
1117}
1118
1119#ifdef CONFIG_NO_HZ
1120/*
1121 * When add_timer_on() enqueues a timer into the timer wheel of an
1122 * idle CPU then this timer might expire before the next timer event
1123 * which is scheduled to wake up that CPU. In case of a completely
1124 * idle system the next event might even be infinite time into the
1125 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1126 * leaves the inner idle loop so the newly added timer is taken into
1127 * account when the CPU goes back to idle and evaluates the timer
1128 * wheel for the next timer event.
1129 */
1130void wake_up_idle_cpu(int cpu)
1131{
1132        struct rq *rq = cpu_rq(cpu);
1133
1134        if (cpu == smp_processor_id())
1135                return;
1136
1137        /*
1138         * This is safe, as this function is called with the timer
1139         * wheel base lock of (cpu) held. When the CPU is on the way
1140         * to idle and has not yet set rq->curr to idle then it will
1141         * be serialized on the timer wheel base lock and take the new
1142         * timer into account automatically.
1143         */
1144        if (rq->curr != rq->idle)
1145                return;
1146
1147        /*
1148         * We can set TIF_RESCHED on the idle task of the other CPU
1149         * lockless. The worst case is that the other CPU runs the
1150         * idle task through an additional NOOP schedule()
1151         */
1152        set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1153
1154        /* NEED_RESCHED must be visible before we test polling */
1155        smp_mb();
1156        if (!tsk_is_polling(rq->idle))
1157                smp_send_reschedule(cpu);
1158}
1159#endif
1160
1161#else
1162static void __resched_task(struct task_struct *p, int tif_bit)
1163{
1164        assert_spin_locked(&task_rq(p)->lock);
1165        set_tsk_thread_flag(p, tif_bit);
1166}
1167#endif
1168
1169#if BITS_PER_LONG == 32
1170# define WMULT_CONST    (~0UL)
1171#else
1172# define WMULT_CONST    (1UL << 32)
1173#endif
1174
1175#define WMULT_SHIFT     32
1176
1177/*
1178 * Shift right and round:
1179 */
1180#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1181
1182static unsigned long
1183calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1184                struct load_weight *lw)
1185{
1186        u64 tmp;
1187
1188        if (unlikely(!lw->inv_weight))
1189                lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
1190
1191        tmp = (u64)delta_exec * weight;
1192        /*
1193         * Check whether we'd overflow the 64-bit multiplication:
1194         */
1195        if (unlikely(tmp > WMULT_CONST))
1196                tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1197                        WMULT_SHIFT/2);
1198        else
1199                tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1200
1201        return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1202}
1203
1204static inline unsigned long
1205calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1206{
1207        return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1208}
1209
1210static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1211{
1212        lw->weight += inc;
1213        lw->inv_weight = 0;
1214}
1215
1216static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1217{
1218        lw->weight -= dec;
1219        lw->inv_weight = 0;
1220}
1221
1222/*
1223 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1224 * of tasks with abnormal "nice" values across CPUs the contribution that
1225 * each task makes to its run queue's load is weighted according to its
1226 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1227 * scaled version of the new time slice allocation that they receive on time
1228 * slice expiry etc.
1229 */
1230
1231#define WEIGHT_IDLEPRIO         2
1232#define WMULT_IDLEPRIO          (1 << 31)
1233
1234/*
1235 * Nice levels are multiplicative, with a gentle 10% change for every
1236 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1237 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1238 * that remained on nice 0.
1239 *
1240 * The "10% effect" is relative and cumulative: from _any_ nice level,
1241 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1242 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1243 * If a task goes up by ~10% and another task goes down by ~10% then
1244 * the relative distance between them is ~25%.)
1245 */
1246static const int prio_to_weight[40] = {
1247 /* -20 */     88761,     71755,     56483,     46273,     36291,
1248 /* -15 */     29154,     23254,     18705,     14949,     11916,
1249 /* -10 */      9548,      7620,      6100,      4904,      3906,
1250 /*  -5 */      3121,      2501,      1991,      1586,      1277,
1251 /*   0 */      1024,       820,       655,       526,       423,
1252 /*   5 */       335,       272,       215,       172,       137,
1253 /*  10 */       110,        87,        70,        56,        45,
1254 /*  15 */        36,        29,        23,        18,        15,
1255};
1256
1257/*
1258 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1259 *
1260 * In cases where the weight does not change often, we can use the
1261 * precalculated inverse to speed up arithmetics by turning divisions
1262 * into multiplications:
1263 */
1264static const u32 prio_to_wmult[40] = {
1265 /* -20 */     48388,     59856,     76040,     92818,    118348,
1266 /* -15 */    147320,    184698,    229616,    287308,    360437,
1267 /* -10 */    449829,    563644,    704093,    875809,   1099582,
1268 /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
1269 /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
1270 /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
1271 /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
1272 /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1273};
1274
1275static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1276
1277/*
1278 * runqueue iterator, to support SMP load-balancing between different
1279 * scheduling classes, without having to expose their internal data
1280 * structures to the load-balancing proper:
1281 */
1282struct rq_iterator {
1283        void *arg;
1284        struct task_struct *(*start)(void *);
1285        struct task_struct *(*next)(void *);
1286};
1287
1288#ifdef CONFIG_SMP
1289static unsigned long
1290balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1291              unsigned long max_load_move, struct sched_domain *sd,
1292              enum cpu_idle_type idle, int *all_pinned,
1293              int *this_best_prio, struct rq_iterator *iterator);
1294
1295static int
1296iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1297                   struct sched_domain *sd, enum cpu_idle_type idle,
1298                   struct rq_iterator *iterator);
1299#endif
1300
1301#ifdef CONFIG_CGROUP_CPUACCT
1302static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1303#else
1304static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1305#endif
1306
1307#ifdef CONFIG_SMP
1308static unsigned long source_load(int cpu, int type);
1309static unsigned long target_load(int cpu, int type);
1310static unsigned long cpu_avg_load_per_task(int cpu);
1311static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1312#endif /* CONFIG_SMP */
1313
1314#include "sched_stats.h"
1315#include "sched_idletask.c"
1316#include "sched_fair.c"
1317#include "sched_rt.c"
1318#ifdef CONFIG_SCHED_DEBUG
1319# include "sched_debug.c"
1320#endif
1321
1322#define sched_class_highest (&rt_sched_class)
1323
1324static inline void inc_load(struct rq *rq, const struct task_struct *p)
1325{
1326        update_load_add(&rq->load, p->se.load.weight);
1327}
1328
1329static inline void dec_load(struct rq *rq, const struct task_struct *p)
1330{
1331        update_load_sub(&rq->load, p->se.load.weight);
1332}
1333
1334static void inc_nr_running(struct task_struct *p, struct rq *rq)
1335{
1336        rq->nr_running++;
1337        inc_load(rq, p);
1338}
1339
1340static void dec_nr_running(struct task_struct *p, struct rq *rq)
1341{
1342        rq->nr_running--;
1343        dec_load(rq, p);
1344}
1345
1346static void set_load_weight(struct task_struct *p)
1347{
1348        if (task_has_rt_policy(p)) {
1349                p->se.load.weight = prio_to_weight[0] * 2;
1350                p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1351                return;
1352        }
1353
1354        /*
1355         * SCHED_IDLE tasks get minimal weight:
1356         */
1357        if (p->policy == SCHED_IDLE) {
1358                p->se.load.weight = WEIGHT_IDLEPRIO;
1359                p->se.load.inv_weight = WMULT_IDLEPRIO;
1360                return;
1361        }
1362
1363        p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1364        p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1365}
1366
1367static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1368{
1369        sched_info_queued(p);
1370        p->sched_class->enqueue_task(rq, p, wakeup);
1371        p->se.on_rq = 1;
1372}
1373
1374static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1375{
1376        p->sched_class->dequeue_task(rq, p, sleep);
1377        p->se.on_rq = 0;
1378}
1379
1380/*
1381 * __normal_prio - return the priority that is based on the static prio
1382 */
1383static inline int __normal_prio(struct task_struct *p)
1384{
1385        return p->static_prio;
1386}
1387
1388/*
1389 * Calculate the expected normal priority: i.e. priority
1390 * without taking RT-inheritance into account. Might be
1391 * boosted by interactivity modifiers. Changes upon fork,
1392 * setprio syscalls, and whenever the interactivity
1393 * estimator recalculates.
1394 */
1395static inline int normal_prio(struct task_struct *p)
1396{
1397        int prio;
1398
1399        if (task_has_rt_policy(p))
1400                prio = MAX_RT_PRIO-1 - p->rt_priority;
1401        else
1402                prio = __normal_prio(p);
1403        return prio;
1404}
1405
1406/*
1407 * Calculate the current priority, i.e. the priority
1408 * taken into account by the scheduler. This value might
1409 * be boosted by RT tasks, or might be boosted by
1410 * interactivity modifiers. Will be RT if the task got
1411 * RT-boosted. If not then it returns p->normal_prio.
1412 */
1413static int effective_prio(struct task_struct *p)
1414{
1415        p->normal_prio = normal_prio(p);
1416        /*
1417         * If we are RT tasks or we were boosted to RT priority,
1418         * keep the priority unchanged. Otherwise, update priority
1419         * to the normal priority:
1420         */
1421        if (!rt_prio(p->prio))
1422                return p->normal_prio;
1423        return p->prio;
1424}
1425
1426/*
1427 * activate_task - move a task to the runqueue.
1428 */
1429static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1430{
1431        if (task_contributes_to_load(p))
1432                rq->nr_uninterruptible--;
1433
1434        enqueue_task(rq, p, wakeup);
1435        inc_nr_running(p, rq);
1436}
1437
1438/*
1439 * deactivate_task - remove a task from the runqueue.
1440 */
1441static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1442{
1443        if (task_contributes_to_load(p))
1444                rq->nr_uninterruptible++;
1445
1446        dequeue_task(rq, p, sleep);
1447        dec_nr_running(p, rq);
1448}
1449
1450/**
1451 * task_curr - is this task currently executing on a CPU?
1452 * @p: the task in question.
1453 */
1454inline int task_curr(const struct task_struct *p)
1455{
1456        return cpu_curr(task_cpu(p)) == p;
1457}
1458
1459/* Used instead of source_load when we know the type == 0 */
1460unsigned long weighted_cpuload(const int cpu)
1461{
1462        return cpu_rq(cpu)->load.weight;
1463}
1464
1465static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1466{
1467        set_task_rq(p, cpu);
1468#ifdef CONFIG_SMP
1469        /*
1470         * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1471         * successfuly executed on another CPU. We must ensure that updates of
1472         * per-task data have been completed by this moment.
1473         */
1474        smp_wmb();
1475        task_thread_info(p)->cpu = cpu;
1476#endif
1477}
1478
1479static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1480                                       const struct sched_class *prev_class,
1481                                       int oldprio, int running)
1482{
1483        if (prev_class != p->sched_class) {
1484                if (prev_class->switched_from)
1485                        prev_class->switched_from(rq, p, running);
1486                p->sched_class->switched_to(rq, p, running);
1487        } else
1488                p->sched_class->prio_changed(rq, p, oldprio, running);
1489}
1490
1491#ifdef CONFIG_SMP
1492
1493/*
1494 * Is this task likely cache-hot:
1495 */
1496static int
1497task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1498{
1499        s64 delta;
1500
1501        /*
1502         * Buddy candidates are cache hot:
1503         */
1504        if (&p->se == cfs_rq_of(&p->se)->next)
1505                return 1;
1506
1507        if (p->sched_class != &fair_sched_class)
1508                return 0;
1509
1510        if (sysctl_sched_migration_cost == -1)
1511                return 1;
1512        if (sysctl_sched_migration_cost == 0)
1513                return 0;
1514
1515        delta = now - p->se.exec_start;
1516
1517        return delta < (s64)sysctl_sched_migration_cost;
1518}
1519
1520
1521void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1522{
1523        int old_cpu = task_cpu(p);
1524        struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1525        struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1526                      *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1527        u64 clock_offset;
1528
1529        clock_offset = old_rq->clock - new_rq->clock;
1530
1531#ifdef CONFIG_SCHEDSTATS
1532        if (p->se.wait_start)
1533                p->se.wait_start -= clock_offset;
1534        if (p->se.sleep_start)
1535                p->se.sleep_start -= clock_offset;
1536        if (p->se.block_start)
1537                p->se.block_start -= clock_offset;
1538        if (old_cpu != new_cpu) {
1539                schedstat_inc(p, se.nr_migrations);
1540                if (task_hot(p, old_rq->clock, NULL))
1541                        schedstat_inc(p, se.nr_forced2_migrations);
1542        }
1543#endif
1544        p->se.vruntime -= old_cfsrq->min_vruntime -
1545                                         new_cfsrq->min_vruntime;
1546
1547        __set_task_cpu(p, new_cpu);
1548}
1549
1550struct migration_req {
1551        struct list_head list;
1552
1553        struct task_struct *task;
1554        int dest_cpu;
1555
1556        struct completion done;
1557};
1558
1559/*
1560 * The task's runqueue lock must be held.
1561 * Returns true if you have to wait for migration thread.
1562 */
1563static int
1564migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1565{
1566        struct rq *rq = task_rq(p);
1567
1568        /*
1569         * If the task is not on a runqueue (and not running), then
1570         * it is sufficient to simply update the task's cpu field.
1571         */
1572        if (!p->se.on_rq && !task_running(rq, p)) {
1573                set_task_cpu(p, dest_cpu);
1574                return 0;
1575        }
1576
1577        init_completion(&req->done);
1578        req->task = p;
1579        req->dest_cpu = dest_cpu;
1580        list_add(&req->list, &rq->migration_queue);
1581
1582        return 1;
1583}
1584
1585/*
1586 * wait_task_inactive - wait for a thread to unschedule.
1587 *
1588 * The caller must ensure that the task *will* unschedule sometime soon,
1589 * else this function might spin for a *long* time. This function can't
1590 * be called with interrupts off, or it may introduce deadlock with
1591 * smp_call_function() if an IPI is sent by the same process we are
1592 * waiting to become inactive.
1593 */
1594void wait_task_inactive(struct task_struct *p)
1595{
1596        unsigned long flags;
1597        int running, on_rq;
1598        struct rq *rq;
1599
1600        for (;;) {
1601                /*
1602                 * We do the initial early heuristics without holding
1603                 * any task-queue locks at all. We'll only try to get
1604                 * the runqueue lock when things look like they will
1605                 * work out!
1606                 */
1607                rq = task_rq(p);
1608
1609                /*
1610                 * If the task is actively running on another CPU
1611                 * still, just relax and busy-wait without holding
1612                 * any locks.
1613                 *
1614                 * NOTE! Since we don't hold any locks, it's not
1615                 * even sure that "rq" stays as the right runqueue!
1616                 * But we don't care, since "task_running()" will
1617                 * return false if the runqueue has changed and p
1618                 * is actually now running somewhere else!
1619                 */
1620                while (task_running(rq, p))
1621                        cpu_relax();
1622
1623                /*
1624                 * Ok, time to look more closely! We need the rq
1625                 * lock now, to be *sure*. If we're wrong, we'll
1626                 * just go back and repeat.
1627                 */
1628                rq = task_rq_lock(p, &flags);
1629                running = task_running(rq, p);
1630                on_rq = p->se.on_rq;
1631                task_rq_unlock(rq, &flags);
1632
1633                /*
1634                 * Was it really running after all now that we
1635                 * checked with the proper locks actually held?
1636                 *
1637                 * Oops. Go back and try again..
1638                 */
1639                if (unlikely(running)) {
1640                        cpu_relax();
1641                        continue;
1642                }
1643
1644                /*
1645                 * It's not enough that it's not actively running,
1646                 * it must be off the runqueue _entirely_, and not
1647                 * preempted!
1648                 *
1649                 * So if it wa still runnable (but just not actively
1650                 * running right now), it's preempted, and we should
1651                 * yield - it could be a while.
1652                 */
1653                if (unlikely(on_rq)) {
1654                        schedule_timeout_uninterruptible(1);
1655                        continue;
1656                }
1657
1658                /*
1659                 * Ahh, all good. It wasn't running, and it wasn't
1660                 * runnable, which means that it will never become
1661                 * running in the future either. We're all done!
1662                 */
1663                break;
1664        }
1665}
1666
1667/***
1668 * kick_process - kick a running thread to enter/exit the kernel
1669 * @p: the to-be-kicked thread
1670 *
1671 * Cause a process which is running on another CPU to enter
1672 * kernel-mode, without any delay. (to get signals handled.)
1673 *
1674 * NOTE: this function doesnt have to take the runqueue lock,
1675 * because all it wants to ensure is that the remote task enters
1676 * the kernel. If the IPI races and the task has been migrated
1677 * to another CPU then no harm is done and the purpose has been
1678 * achieved as well.
1679 */
1680void kick_process(struct task_struct *p)
1681{
1682        int cpu;
1683
1684        preempt_disable();
1685        cpu = task_cpu(p);
1686        if ((cpu != smp_processor_id()) && task_curr(p))
1687                smp_send_reschedule(cpu);
1688        preempt_enable();
1689}
1690
1691/*
1692 * Return a low guess at the load of a migration-source cpu weighted
1693 * according to the scheduling class and "nice" value.
1694 *
1695 * We want to under-estimate the load of migration sources, to
1696 * balance conservatively.
1697 */
1698static unsigned long source_load(int cpu, int type)
1699{
1700        struct rq *rq = cpu_rq(cpu);
1701        unsigned long total = weighted_cpuload(cpu);
1702
1703        if (type == 0)
1704                return total;
1705
1706        return min(rq->cpu_load[type-1], total);
1707}
1708
1709/*
1710 * Return a high guess at the load of a migration-target cpu weighted
1711 * according to the scheduling class and "nice" value.
1712 */
1713static unsigned long target_load(int cpu, int type)
1714{
1715        struct rq *rq = cpu_rq(cpu);
1716        unsigned long total = weighted_cpuload(cpu);
1717
1718        if (type == 0)
1719                return total;
1720
1721        return max(rq->cpu_load[type-1], total);
1722}
1723
1724/*
1725 * Return the average load per task on the cpu's run queue
1726 */
1727static unsigned long cpu_avg_load_per_task(int cpu)
1728{
1729        struct rq *rq = cpu_rq(cpu);
1730        unsigned long total = weighted_cpuload(cpu);
1731        unsigned long n = rq->nr_running;
1732
1733        return n ? total / n : SCHED_LOAD_SCALE;
1734}
1735
1736/*
1737 * find_idlest_group finds and returns the least busy CPU group within the
1738 * domain.
1739 */
1740static struct sched_group *
1741find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1742{
1743        struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1744        unsigned long min_load = ULONG_MAX, this_load = 0;
1745        int load_idx = sd->forkexec_idx;
1746        int imbalance = 100 + (sd->imbalance_pct-100)/2;
1747
1748        do {
1749                unsigned long load, avg_load;
1750                int local_group;
1751                int i;
1752
1753                /* Skip over this group if it has no CPUs allowed */
1754                if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1755                        continue;
1756
1757                local_group = cpu_isset(this_cpu, group->cpumask);
1758
1759                /* Tally up the load of all CPUs in the group */
1760                avg_load = 0;
1761
1762                for_each_cpu_mask(i, group->cpumask) {
1763                        /* Bias balancing toward cpus of our domain */
1764                        if (local_group)
1765                                load = source_load(i, load_idx);
1766                        else
1767                                load = target_load(i, load_idx);
1768
1769                        avg_load += load;
1770                }
1771
1772                /* Adjust by relative CPU power of the group */
1773                avg_load = sg_div_cpu_power(group,
1774                                avg_load * SCHED_LOAD_SCALE);
1775
1776                if (local_group) {
1777                        this_load = avg_load;
1778                        this = group;
1779                } else if (avg_load < min_load) {
1780                        min_load = avg_load;
1781                        idlest = group;
1782                }
1783        } while (group = group->next, group != sd->groups);
1784
1785        if (!idlest || 100*this_load < imbalance*min_load)
1786                return NULL;
1787        return idlest;
1788}
1789
1790/*
1791 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1792 */
1793static int
1794find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1795{
1796        cpumask_t tmp;
1797        unsigned long load, min_load = ULONG_MAX;
1798        int idlest = -1;
1799        int i;
1800
1801        /* Traverse only the allowed CPUs */
1802        cpus_and(tmp, group->cpumask, p->cpus_allowed);
1803
1804        for_each_cpu_mask(i, tmp) {
1805                load = weighted_cpuload(i);
1806
1807                if (load < min_load || (load == min_load && i == this_cpu)) {
1808                        min_load = load;
1809                        idlest = i;
1810                }
1811        }
1812
1813        return idlest;
1814}
1815
1816/*
1817 * sched_balance_self: balance the current task (running on cpu) in domains
1818 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1819 * SD_BALANCE_EXEC.
1820 *
1821 * Balance, ie. select the least loaded group.
1822 *
1823 * Returns the target CPU number, or the same CPU if no balancing is needed.
1824 *
1825 * preempt must be disabled.
1826 */
1827static int sched_balance_self(int cpu, int flag)
1828{
1829        struct task_struct *t = current;
1830        struct sched_domain *tmp, *sd = NULL;
1831
1832        for_each_domain(cpu, tmp) {
1833                /*
1834                 * If power savings logic is enabled for a domain, stop there.
1835                 */
1836                if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1837                        break;
1838                if (tmp->flags & flag)
1839                        sd = tmp;
1840        }
1841
1842        while (sd) {
1843                cpumask_t span;
1844                struct sched_group *group;
1845                int new_cpu, weight;
1846
1847                if (!(sd->flags & flag)) {
1848                        sd = sd->child;
1849                        continue;
1850                }
1851
1852                span = sd->span;
1853                group = find_idlest_group(sd, t, cpu);
1854                if (!group) {
1855                        sd = sd->child;
1856                        continue;
1857                }
1858
1859                new_cpu = find_idlest_cpu(group, t, cpu);
1860                if (new_cpu == -1 || new_cpu == cpu) {
1861                        /* Now try balancing at a lower domain level of cpu */
1862                        sd = sd->child;
1863                        continue;
1864                }
1865
1866                /* Now try balancing at a lower domain level of new_cpu */
1867                cpu = new_cpu;
1868                sd = NULL;
1869                weight = cpus_weight(span);
1870                for_each_domain(cpu, tmp) {
1871                        if (weight <= cpus_weight(tmp->span))
1872                                break;
1873                        if (tmp->flags & flag)
1874                                sd = tmp;
1875                }
1876                /* while loop will break here if sd == NULL */
1877        }
1878
1879        return cpu;
1880}
1881
1882#endif /* CONFIG_SMP */
1883
1884/***
1885 * try_to_wake_up - wake up a thread
1886 * @p: the to-be-woken-up thread
1887 * @state: the mask of task states that can be woken
1888 * @sync: do a synchronous wakeup?
1889 *
1890 * Put it on the run-queue if it's not already there. The "current"
1891 * thread is always on the run-queue (except when the actual
1892 * re-schedule is in progress), and as such you're allowed to do
1893 * the simpler "current->state = TASK_RUNNING" to mark yourself
1894 * runnable without the overhead of this.
1895 *
1896 * returns failure only if the task is already active.
1897 */
1898static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1899{
1900        int cpu, orig_cpu, this_cpu, success = 0;
1901        unsigned long flags;
1902        long old_state;
1903        struct rq *rq;
1904
1905        smp_wmb();
1906        rq = task_rq_lock(p, &flags);
1907        old_state = p->state;
1908        if (!(old_state & state))
1909                goto out;
1910
1911        if (p->se.on_rq)
1912                goto out_running;
1913
1914        cpu = task_cpu(p);
1915        orig_cpu = cpu;
1916        this_cpu = smp_processor_id();
1917
1918#ifdef CONFIG_SMP
1919        if (unlikely(task_running(rq, p)))
1920                goto out_activate;
1921
1922        cpu = p->sched_class->select_task_rq(p, sync);
1923        if (cpu != orig_cpu) {
1924                set_task_cpu(p, cpu);
1925                task_rq_unlock(rq, &flags);
1926                /* might preempt at this point */
1927                rq = task_rq_lock(p, &flags);
1928                old_state = p->state;
1929                if (!(old_state & state))
1930                        goto out;
1931                if (p->se.on_rq)
1932                        goto out_running;
1933
1934                this_cpu = smp_processor_id();
1935                cpu = task_cpu(p);
1936        }
1937
1938#ifdef CONFIG_SCHEDSTATS
1939        schedstat_inc(rq, ttwu_count);
1940        if (cpu == this_cpu)
1941                schedstat_inc(rq, ttwu_local);
1942        else {
1943                struct sched_domain *sd;
1944                for_each_domain(this_cpu, sd) {
1945                        if (cpu_isset(cpu, sd->span)) {
1946                                schedstat_inc(sd, ttwu_wake_remote);
1947                                break;
1948                        }
1949                }
1950        }
1951#endif
1952
1953out_activate:
1954#endif /* CONFIG_SMP */
1955        schedstat_inc(p, se.nr_wakeups);
1956        if (sync)
1957                schedstat_inc(p, se.nr_wakeups_sync);
1958        if (orig_cpu != cpu)
1959                schedstat_inc(p, se.nr_wakeups_migrate);
1960        if (cpu == this_cpu)
1961                schedstat_inc(p, se.nr_wakeups_local);
1962        else
1963                schedstat_inc(p, se.nr_wakeups_remote);
1964        update_rq_clock(rq);
1965        activate_task(rq, p, 1);
1966        success = 1;
1967
1968out_running:
1969        check_preempt_curr(rq, p);
1970
1971        p->state = TASK_RUNNING;
1972#ifdef CONFIG_SMP
1973        if (p->sched_class->task_wake_up)
1974                p->sched_class->task_wake_up(rq, p);
1975#endif
1976out:
1977        task_rq_unlock(rq, &flags);
1978
1979        return success;
1980}
1981
1982int wake_up_process(struct task_struct *p)
1983{
1984        return try_to_wake_up(p, TASK_ALL, 0);
1985}
1986EXPORT_SYMBOL(wake_up_process);
1987
1988int wake_up_state(struct task_struct *p, unsigned int state)
1989{
1990        return try_to_wake_up(p, state, 0);
1991}
1992
1993/*
1994 * Perform scheduler related setup for a newly forked process p.
1995 * p is forked by current.
1996 *
1997 * __sched_fork() is basic setup used by init_idle() too:
1998 */
1999static void __sched_fork(struct task_struct *p)
2000{
2001        p->se.exec_start                = 0;
2002        p->se.sum_exec_runtime          = 0;
2003        p->se.prev_sum_exec_runtime     = 0;
2004        p->se.last_wakeup               = 0;
2005        p->se.avg_overlap               = 0;
2006
2007#ifdef CONFIG_SCHEDSTATS
2008        p->se.wait_start                = 0;
2009        p->se.sum_sleep_runtime         = 0;
2010        p->se.sleep_start               = 0;
2011        p->se.block_start               = 0;
2012        p->se.sleep_max                 = 0;
2013        p->se.block_max                 = 0;
2014        p->se.exec_max                  = 0;
2015        p->se.slice_max                 = 0;
2016        p->se.wait_max                  = 0;
2017#endif
2018
2019        INIT_LIST_HEAD(&p->rt.run_list);
2020        p->se.on_rq = 0;
2021
2022#ifdef CONFIG_PREEMPT_NOTIFIERS
2023        INIT_HLIST_HEAD(&p->preempt_notifiers);
2024#endif
2025
2026        /*
2027         * We mark the process as running here, but have not actually
2028         * inserted it onto the runqueue yet. This guarantees that
2029         * nobody will actually run it, and a signal or other external
2030         * event cannot wake it up and insert it on the runqueue either.
2031         */
2032        p->state = TASK_RUNNING;
2033}
2034
2035/*
2036 * fork()/clone()-time setup:
2037 */
2038void sched_fork(struct task_struct *p, int clone_flags)
2039{
2040        int cpu = get_cpu();
2041
2042        __sched_fork(p);
2043
2044#ifdef CONFIG_SMP
2045        cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2046#endif
2047        set_task_cpu(p, cpu);
2048
2049        /*
2050         * Make sure we do not leak PI boosting priority to the child:
2051         */
2052        p->prio = current->normal_prio;
2053        if (!rt_prio(p->prio))
2054                p->sched_class = &fair_sched_class;
2055
2056#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2057        if (likely(sched_info_on()))
2058                memset(&p->sched_info, 0, sizeof(p->sched_info));
2059#endif
2060#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2061        p->oncpu = 0;
2062#endif
2063#ifdef CONFIG_PREEMPT
2064        /* Want to start with kernel preemption disabled. */
2065        task_thread_info(p)->preempt_count = 1;
2066#endif
2067        put_cpu();
2068}
2069
2070/*
2071 * wake_up_new_task - wake up a newly created task for the first time.
2072 *
2073 * This function will do some initial scheduler statistics housekeeping
2074 * that must be done for every newly created context, then puts the task
2075 * on the runqueue and wakes it.
2076 */
2077void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2078{
2079        unsigned long flags;
2080        struct rq *rq;
2081
2082        rq = task_rq_lock(p, &flags);
2083        BUG_ON(p->state != TASK_RUNNING);
2084        update_rq_clock(rq);
2085
2086        p->prio = effective_prio(p);
2087
2088        if (!p->sched_class->task_new || !current->se.on_rq) {
2089                activate_task(rq, p, 0);
2090        } else {
2091                /*
2092                 * Let the scheduling class do new task startup
2093                 * management (if any):
2094                 */
2095                p->sched_class->task_new(rq, p);
2096                inc_nr_running(p, rq);
2097        }
2098        check_preempt_curr(rq, p);
2099#ifdef CONFIG_SMP
2100        if (p->sched_class->task_wake_up)
2101                p->sched_class->task_wake_up(rq, p);
2102#endif
2103        task_rq_unlock(rq, &flags);
2104}
2105
2106#ifdef CONFIG_PREEMPT_NOTIFIERS
2107
2108/**
2109 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2110 * @notifier: notifier struct to register
2111 */
2112void preempt_notifier_register(struct preempt_notifier *notifier)
2113{
2114        hlist_add_head(&notifier->link, &current->preempt_notifiers);
2115}
2116EXPORT_SYMBOL_GPL(preempt_notifier_register);
2117
2118/**
2119 * preempt_notifier_unregister - no longer interested in preemption notifications
2120 * @notifier: notifier struct to unregister
2121 *
2122 * This is safe to call from within a preemption notifier.
2123 */
2124void preempt_notifier_unregister(struct preempt_notifier *notifier)
2125{
2126        hlist_del(&notifier->link);
2127}
2128EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2129
2130static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2131{
2132        struct preempt_notifier *notifier;
2133        struct hlist_node *node;
2134
2135        hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2136                notifier->ops->sched_in(notifier, raw_smp_processor_id());
2137}
2138
2139static void
2140fire_sched_out_preempt_notifiers(struct task_struct *curr,
2141                                 struct task_struct *next)
2142{
2143        struct preempt_notifier *notifier;
2144        struct hlist_node *node;
2145
2146        hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2147                notifier->ops->sched_out(notifier, next);
2148}
2149
2150#else
2151
2152static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2153{
2154}
2155
2156static void
2157fire_sched_out_preempt_notifiers(struct task_struct *curr,
2158                                 struct task_struct *next)
2159{
2160}
2161
2162#endif
2163
2164/**
2165 * prepare_task_switch - prepare to switch tasks
2166 * @rq: the runqueue preparing to switch
2167 * @prev: the current task that is being switched out
2168 * @next: the task we are going to switch to.
2169 *
2170 * This is called with the rq lock held and interrupts off. It must
2171 * be paired with a subsequent finish_task_switch after the context
2172 * switch.
2173 *
2174 * prepare_task_switch sets up locking and calls architecture specific
2175 * hooks.
2176 */
2177static inline void
2178prepare_task_switch(struct rq *rq, struct task_struct *prev,
2179                    struct task_struct *next)
2180{
2181        fire_sched_out_preempt_notifiers(prev, next);
2182        prepare_lock_switch(rq, next);
2183        prepare_arch_switch(next);
2184}
2185
2186/**
2187 * finish_task_switch - clean up after a task-switch
2188 * @rq: runqueue associated with task-switch
2189 * @prev: the thread we just switched away from.
2190 *
2191 * finish_task_switch must be called after the context switch, paired
2192 * with a prepare_task_switch call before the context switch.
2193 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2194 * and do any other architecture-specific cleanup actions.
2195 *
2196 * Note that we may have delayed dropping an mm in context_switch(). If
2197 * so, we finish that here outside of the runqueue lock. (Doing it
2198 * with the lock held can cause deadlocks; see schedule() for
2199 * details.)
2200 */
2201static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2202        __releases(rq->lock)
2203{
2204        struct mm_struct *mm = rq->prev_mm;
2205        long prev_state;
2206
2207        rq->prev_mm = NULL;
2208
2209        /*
2210         * A task struct has one reference for the use as "current".
2211         * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2212         * schedule one last time. The schedule call will never return, and
2213         * the scheduled task must drop that reference.
2214         * The test for TASK_DEAD must occur while the runqueue locks are
2215         * still held, otherwise prev could be scheduled on another cpu, die
2216         * there before we look at prev->state, and then the reference would
2217         * be dropped twice.
2218         *              Manfred Spraul <manfred@colorfullife.com>
2219         */
2220        prev_state = prev->state;
2221        finish_arch_switch(prev);
2222        finish_lock_switch(rq, prev);
2223#ifdef CONFIG_SMP
2224        if (current->sched_class->post_schedule)
2225                current->sched_class->post_schedule(rq);
2226#endif
2227
2228        fire_sched_in_preempt_notifiers(current);
2229        if (mm)
2230                mmdrop(mm);
2231        if (unlikely(prev_state == TASK_DEAD)) {
2232                /*
2233                 * Remove function-return probe instances associated with this
2234                 * task and put them back on the free list.
2235                 */
2236                kprobe_flush_task(prev);
2237                put_task_struct(prev);
2238        }
2239}
2240
2241/**
2242 * schedule_tail - first thing a freshly forked thread must call.
2243 * @prev: the thread we just switched away from.
2244 */
2245asmlinkage void schedule_tail(struct task_struct *prev)
2246        __releases(rq->lock)
2247{
2248        struct rq *rq = this_rq();
2249
2250        finish_task_switch(rq, prev);
2251#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2252        /* In this case, finish_task_switch does not reenable preemption */
2253        preempt_enable();
2254#endif
2255        if (current->set_child_tid)
2256                put_user(task_pid_vnr(current), current->set_child_tid);
2257}
2258
2259/*
2260 * context_switch - switch to the new MM and the new
2261 * thread's register state.
2262 */
2263static inline void
2264context_switch(struct rq *rq, struct task_struct *prev,
2265               struct task_struct *next)
2266{
2267        struct mm_struct *mm, *oldmm;
2268
2269        prepare_task_switch(rq, prev, next);
2270        mm = next->mm;
2271        oldmm = prev->active_mm;
2272        /*
2273         * For paravirt, this is coupled with an exit in switch_to to
2274         * combine the page table reload and the switch backend into
2275         * one hypercall.
2276         */
2277        arch_enter_lazy_cpu_mode();
2278
2279        if (unlikely(!mm)) {
2280                next->active_mm = oldmm;
2281                atomic_inc(&oldmm->mm_count);
2282                enter_lazy_tlb(oldmm, next);
2283        } else
2284                switch_mm(oldmm, mm, next);
2285
2286        if (unlikely(!prev->mm)) {
2287                prev->active_mm = NULL;
2288                rq->prev_mm = oldmm;
2289        }
2290        /*
2291         * Since the runqueue lock will be released by the next
2292         * task (which is an invalid locking op but in the case
2293         * of the scheduler it's an obvious special-case), so we
2294         * do an early lockdep release here:
2295         */
2296#ifndef __ARCH_WANT_UNLOCKED_CTXSW
2297        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2298#endif
2299
2300        /* Here we just switch the register state and the stack. */
2301        switch_to(prev, next, prev);
2302
2303        barrier();
2304        /*
2305         * this_rq must be evaluated again because prev may have moved
2306         * CPUs since it called schedule(), thus the 'rq' on its stack
2307         * frame will be invalid.
2308         */
2309        finish_task_switch(this_rq(), prev);
2310}
2311
2312/*
2313 * nr_running, nr_uninterruptible and nr_context_switches:
2314 *
2315 * externally visible scheduler statistics: current number of runnable
2316 * threads, current number of uninterruptible-sleeping threads, total
2317 * number of context switches performed since bootup.
2318 */
2319unsigned long nr_running(void)
2320{
2321        unsigned long i, sum = 0;
2322
2323        for_each_online_cpu(i)
2324                sum += cpu_rq(i)->nr_running;
2325
2326        return sum;
2327}
2328
2329unsigned long nr_uninterruptible(void)
2330{
2331        unsigned long i, sum = 0;
2332
2333        for_each_possible_cpu(i)
2334                sum += cpu_rq(i)->nr_uninterruptible;
2335
2336        /*
2337         * Since we read the counters lockless, it might be slightly
2338         * inaccurate. Do not allow it to go below zero though:
2339         */
2340        if (unlikely((long)sum < 0))
2341                sum = 0;
2342
2343        return sum;
2344}
2345
2346unsigned long long nr_context_switches(void)
2347{
2348        int i;
2349        unsigned long long sum = 0;
2350
2351        for_each_possible_cpu(i)
2352                sum += cpu_rq(i)->nr_switches;
2353
2354        return sum;
2355}
2356
2357unsigned long nr_iowait(void)
2358{
2359        unsigned long i, sum = 0;
2360
2361        for_each_possible_cpu(i)
2362                sum += atomic_read(&cpu_rq(i)->nr_iowait);
2363
2364        return sum;
2365}
2366
2367unsigned long nr_active(void)
2368{
2369        unsigned long i, running = 0, uninterruptible = 0;
2370
2371        for_each_online_cpu(i) {
2372                running += cpu_rq(i)->nr_running;
2373                uninterruptible += cpu_rq(i)->nr_uninterruptible;
2374        }
2375
2376        if (unlikely((long)uninterruptible < 0))
2377                uninterruptible = 0;
2378
2379        return running + uninterruptible;
2380}
2381
2382/*
2383 * Update rq->cpu_load[] statistics. This function is usually called every
2384 * scheduler tick (TICK_NSEC).
2385 */
2386static void update_cpu_load(struct rq *this_rq)
2387{
2388        unsigned long this_load = this_rq->load.weight;
2389        int i, scale;
2390
2391        this_rq->nr_load_updates++;
2392
2393        /* Update our load: */
2394        for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2395                unsigned long old_load, new_load;
2396
2397                /* scale is effectively 1 << i now, and >> i divides by scale */
2398
2399                old_load = this_rq->cpu_load[i];
2400                new_load = this_load;
2401                /*
2402                 * Round up the averaging division if load is increasing. This
2403                 * prevents us from getting stuck on 9 if the load is 10, for
2404                 * example.
2405                 */
2406                if (new_load > old_load)
2407                        new_load += scale-1;
2408                this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2409        }
2410}
2411
2412#ifdef CONFIG_SMP
2413
2414/*
2415 * double_rq_lock - safely lock two runqueues
2416 *
2417 * Note this does not disable interrupts like task_rq_lock,
2418 * you need to do so manually before calling.
2419 */
2420static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2421        __acquires(rq1->lock)
2422        __acquires(rq2->lock)
2423{
2424        BUG_ON(!irqs_disabled());
2425        if (rq1 == rq2) {
2426                spin_lock(&rq1->lock);
2427                __acquire(rq2->lock);   /* Fake it out ;) */
2428        } else {
2429                if (rq1 < rq2) {
2430                        spin_lock(&rq1->lock);
2431                        spin_lock(&rq2->lock);
2432                } else {
2433                        spin_lock(&rq2->lock);
2434                        spin_lock(&rq1->lock);
2435                }
2436        }
2437        update_rq_clock(rq1);
2438        update_rq_clock(rq2);
2439}
2440
2441/*
2442 * double_rq_unlock - safely unlock two runqueues
2443 *
2444 * Note this does not restore interrupts like task_rq_unlock,
2445 * you need to do so manually after calling.
2446 */
2447static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2448        __releases(rq1->lock)
2449        __releases(rq2->lock)
2450{
2451        spin_unlock(&rq1->lock);
2452        if (rq1 != rq2)
2453                spin_unlock(&rq2->lock);
2454        else
2455                __release(rq2->lock);
2456}
2457
2458/*
2459 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2460 */
2461static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2462        __releases(this_rq->lock)
2463        __acquires(busiest->lock)
2464        __acquires(this_rq->lock)
2465{
2466        int ret = 0;
2467
2468        if (unlikely(!irqs_disabled())) {
2469                /* printk() doesn't work good under rq->lock */
2470                spin_unlock(&this_rq->lock);
2471                BUG_ON(1);
2472        }
2473        if (unlikely(!spin_trylock(&busiest->lock))) {
2474                if (busiest < this_rq) {
2475                        spin_unlock(&this_rq->lock);
2476                        spin_lock(&busiest->lock);
2477                        spin_lock(&this_rq->lock);
2478                        ret = 1;
2479                } else
2480                        spin_lock(&busiest->lock);
2481        }
2482        return ret;
2483}
2484
2485/*
2486 * If dest_cpu is allowed for this process, migrate the task to it.
2487 * This is accomplished by forcing the cpu_allowed mask to only
2488 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2489 * the cpu_allowed mask is restored.
2490 */
2491static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2492{
2493        struct migration_req req;
2494        unsigned long flags;
2495        struct rq *rq;
2496
2497        rq = task_rq_lock(p, &flags);
2498        if (!cpu_isset(dest_cpu, p->cpus_allowed)
2499            || unlikely(cpu_is_offline(dest_cpu)))
2500                goto out;
2501
2502        /* force the process onto the specified CPU */
2503        if (migrate_task(p, dest_cpu, &req)) {
2504                /* Need to wait for migration thread (might exit: take ref). */
2505                struct task_struct *mt = rq->migration_thread;
2506
2507                get_task_struct(mt);
2508                task_rq_unlock(rq, &flags);
2509                wake_up_process(mt);
2510                put_task_struct(mt);
2511                wait_for_completion(&req.done);
2512
2513                return;
2514        }
2515out:
2516        task_rq_unlock(rq, &flags);
2517}
2518
2519/*
2520 * sched_exec - execve() is a valuable balancing opportunity, because at
2521 * this point the task has the smallest effective memory and cache footprint.
2522 */
2523void sched_exec(void)
2524{
2525        int new_cpu, this_cpu = get_cpu();
2526        new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2527        put_cpu();
2528        if (new_cpu != this_cpu)
2529                sched_migrate_task(current, new_cpu);
2530}
2531
2532/*
2533 * pull_task - move a task from a remote runqueue to the local runqueue.
2534 * Both runqueues must be locked.
2535 */
2536static void pull_task(struct rq *src_rq, struct task_struct *p,
2537                      struct rq *this_rq, int this_cpu)
2538{
2539        deactivate_task(src_rq, p, 0);
2540        set_task_cpu(p, this_cpu);
2541        activate_task(this_rq, p, 0);
2542        /*
2543         * Note that idle threads have a prio of MAX_PRIO, for this test
2544         * to be always true for them.
2545         */
2546        check_preempt_curr(this_rq, p);
2547}
2548
2549/*
2550 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2551 */
2552static
2553int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2554                     struct sched_domain *sd, enum cpu_idle_type idle,
2555                     int *all_pinned)
2556{
2557        /*
2558         * We do not migrate tasks that are:
2559         * 1) running (obviously), or
2560         * 2) cannot be migrated to this CPU due to cpus_allowed, or
2561         * 3) are cache-hot on their current CPU.
2562         */
2563        if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2564                schedstat_inc(p, se.nr_failed_migrations_affine);
2565                return 0;
2566        }
2567        *all_pinned = 0;
2568
2569        if (task_running(rq, p)) {
2570                schedstat_inc(p, se.nr_failed_migrations_running);
2571                return 0;
2572        }
2573
2574        /*
2575         * Aggressive migration if:
2576         * 1) task is cache cold, or
2577         * 2) too many balance attempts have failed.
2578         */
2579
2580        if (!task_hot(p, rq->clock, sd) ||
2581                        sd->nr_balance_failed > sd->cache_nice_tries) {
2582#ifdef CONFIG_SCHEDSTATS
2583                if (task_hot(p, rq->clock, sd)) {
2584                        schedstat_inc(sd, lb_hot_gained[idle]);
2585                        schedstat_inc(p, se.nr_forced_migrations);
2586                }
2587#endif
2588                return 1;
2589        }
2590
2591        if (task_hot(p, rq->clock, sd)) {
2592                schedstat_inc(p, se.nr_failed_migrations_hot);
2593                return 0;
2594        }
2595        return 1;
2596}
2597
2598static unsigned long
2599balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2600              unsigned long max_load_move, struct sched_domain *sd,
2601              enum cpu_idle_type idle, int *all_pinned,
2602              int *this_best_prio, struct rq_iterator *iterator)
2603{
2604        int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2605        struct task_struct *p;
2606        long rem_load_move = max_load_move;
2607
2608        if (max_load_move == 0)
2609                goto out;
2610
2611        pinned = 1;
2612
2613        /*
2614         * Start the load-balancing iterator:
2615         */
2616        p = iterator->start(iterator->arg);
2617next:
2618        if (!p || loops++ > sysctl_sched_nr_migrate)
2619                goto out;
2620        /*
2621         * To help distribute high priority tasks across CPUs we don't
2622         * skip a task if it will be the highest priority task (i.e. smallest
2623         * prio value) on its new queue regardless of its load weight
2624         */
2625        skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2626                                                         SCHED_LOAD_SCALE_FUZZ;
2627        if ((skip_for_load && p->prio >= *this_best_prio) ||
2628            !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2629                p = iterator->next(iterator->arg);
2630                goto next;
2631        }
2632
2633        pull_task(busiest, p, this_rq, this_cpu);
2634        pulled++;
2635        rem_load_move -= p->se.load.weight;
2636
2637        /*
2638         * We only want to steal up to the prescribed amount of weighted load.
2639         */
2640        if (rem_load_move > 0) {
2641                if (p->prio < *this_best_prio)
2642                        *this_best_prio = p->prio;
2643                p = iterator->next(iterator->arg);
2644                goto next;
2645        }
2646out:
2647        /*
2648         * Right now, this is one of only two places pull_task() is called,
2649         * so we can safely collect pull_task() stats here rather than
2650         * inside pull_task().
2651         */
2652        schedstat_add(sd, lb_gained[idle], pulled);
2653
2654        if (all_pinned)
2655                *all_pinned = pinned;
2656
2657        return max_load_move - rem_load_move;
2658}
2659
2660/*
2661 * move_tasks tries to move up to max_load_move weighted load from busiest to
2662 * this_rq, as part of a balancing operation within domain "sd".
2663 * Returns 1 if successful and 0 otherwise.
2664 *
2665 * Called with both runqueues locked.
2666 */
2667static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2668                      unsigned long max_load_move,
2669                      struct sched_domain *sd, enum cpu_idle_type idle,
2670                      int *all_pinned)
2671{
2672        const struct sched_class *class = sched_class_highest;
2673        unsigned long total_load_moved = 0;
2674        int this_best_prio = this_rq->curr->prio;
2675
2676        do {
2677                total_load_moved +=
2678                        class->load_balance(this_rq, this_cpu, busiest,
2679                                max_load_move - total_load_moved,
2680                                sd, idle, all_pinned, &this_best_prio);
2681                class = class->next;
2682        } while (class && max_load_move > total_load_moved);
2683
2684        return total_load_moved > 0;
2685}
2686
2687static int
2688iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2689                   struct sched_domain *sd, enum cpu_idle_type idle,
2690                   struct rq_iterator *iterator)
2691{
2692        struct task_struct *p = iterator->start(iterator->arg);
2693        int pinned = 0;
2694
2695        while (p) {
2696                if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2697                        pull_task(busiest, p, this_rq, this_cpu);
2698                        /*
2699                         * Right now, this is only the second place pull_task()
2700                         * is called, so we can safely collect pull_task()
2701                         * stats here rather than inside pull_task().
2702                         */
2703                        schedstat_inc(sd, lb_gained[idle]);
2704
2705                        return 1;
2706                }
2707                p = iterator->next(iterator->arg);
2708        }
2709
2710        return 0;
2711}
2712
2713/*
2714 * move_one_task tries to move exactly one task from busiest to this_rq, as
2715 * part of active balancing operations within "domain".
2716 * Returns 1 if successful and 0 otherwise.
2717 *
2718 * Called with both runqueues locked.
2719 */
2720static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2721                         struct sched_domain *sd, enum cpu_idle_type idle)
2722{
2723        const struct sched_class *class;
2724
2725        for (class = sched_class_highest; class; class = class->next)
2726                if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2727                        return 1;
2728
2729        return 0;
2730}
2731
2732/*
2733 * find_busiest_group finds and returns the busiest CPU group within the
2734 * domain. It calculates and returns the amount of weighted load which
2735 * should be moved to restore balance via the imbalance parameter.
2736 */
2737static struct sched_group *
2738find_busiest_group(struct sched_domain *sd, int this_cpu,
2739                   unsigned long *imbalance, enum cpu_idle_type idle,
2740                   int *sd_idle, cpumask_t *cpus, int *balance)
2741{
2742        struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2743        unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2744        unsigned long max_pull;
2745        unsigned long busiest_load_per_task, busiest_nr_running;
2746        unsigned long this_load_per_task, this_nr_running;
2747        int load_idx, group_imb = 0;
2748#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2749        int power_savings_balance = 1;
2750        unsigned long leader_nr_running = 0, min_load_per_task = 0;
2751        unsigned long min_nr_running = ULONG_MAX;
2752        struct sched_group *group_min = NULL, *group_leader = NULL;
2753#endif
2754
2755        max_load = this_load = total_load = total_pwr = 0;
2756        busiest_load_per_task = busiest_nr_running = 0;
2757        this_load_per_task = this_nr_running = 0;
2758        if (idle == CPU_NOT_IDLE)
2759                load_idx = sd->busy_idx;
2760        else if (idle == CPU_NEWLY_IDLE)
2761                load_idx = sd->newidle_idx;
2762        else
2763                load_idx = sd->idle_idx;
2764
2765        do {
2766                unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2767                int local_group;
2768                int i;
2769                int __group_imb = 0;
2770                unsigned int balance_cpu = -1, first_idle_cpu = 0;
2771                unsigned long sum_nr_running, sum_weighted_load;
2772
2773                local_group = cpu_isset(this_cpu, group->cpumask);
2774
2775                if (local_group)
2776                        balance_cpu = first_cpu(group->cpumask);
2777
2778                /* Tally up the load of all CPUs in the group */
2779                sum_weighted_load = sum_nr_running = avg_load = 0;
2780                max_cpu_load = 0;
2781                min_cpu_load = ~0UL;
2782
2783                for_each_cpu_mask(i, group->cpumask) {
2784                        struct rq *rq;
2785
2786                        if (!cpu_isset(i, *cpus))
2787                                continue;
2788
2789                        rq = cpu_rq(i);
2790
2791                        if (*sd_idle && rq->nr_running)
2792                                *sd_idle = 0;
2793
2794                        /* Bias balancing toward cpus of our domain */
2795                        if (local_group) {
2796                                if (idle_cpu(i) && !first_idle_cpu) {
2797                                        first_idle_cpu = 1;
2798                                        balance_cpu = i;
2799                                }
2800
2801                                load = target_load(i, load_idx);
2802                        } else {
2803                                load = source_load(i, load_idx);
2804                                if (load > max_cpu_load)
2805                                        max_cpu_load = load;
2806                                if (min_cpu_load > load)
2807                                        min_cpu_load = load;
2808                        }
2809
2810                        avg_load += load;
2811                        sum_nr_running += rq->nr_running;
2812                        sum_weighted_load += weighted_cpuload(i);
2813                }
2814
2815                /*
2816                 * First idle cpu or the first cpu(busiest) in this sched group
2817                 * is eligible for doing load balancing at this and above
2818                 * domains. In the newly idle case, we will allow all the cpu's
2819                 * to do the newly idle load balance.
2820                 */
2821                if (idle != CPU_NEWLY_IDLE && local_group &&
2822                    balance_cpu != this_cpu && balance) {
2823                        *balance = 0;
2824                        goto ret;
2825                }
2826
2827                total_load += avg_load;
2828                total_pwr += group->__cpu_power;
2829
2830                /* Adjust by relative CPU power of the group */
2831                avg_load = sg_div_cpu_power(group,
2832                                avg_load * SCHED_LOAD_SCALE);
2833
2834                if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2835                        __group_imb = 1;
2836
2837                group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
2838
2839                if (local_group) {
2840                        this_load = avg_load;
2841                        this = group;
2842                        this_nr_running = sum_nr_running;
2843                        this_load_per_task = sum_weighted_load;
2844                } else if (avg_load > max_load &&
2845                           (sum_nr_running > group_capacity || __group_imb)) {
2846                        max_load = avg_load;
2847                        busiest = group;
2848                        busiest_nr_running = sum_nr_running;
2849                        busiest_load_per_task = sum_weighted_load;
2850                        group_imb = __group_imb;
2851                }
2852
2853#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2854                /*
2855                 * Busy processors will not participate in power savings
2856                 * balance.
2857                 */
2858                if (idle == CPU_NOT_IDLE ||
2859                                !(sd->flags & SD_POWERSAVINGS_BALANCE))
2860                        goto group_next;
2861
2862                /*
2863                 * If the local group is idle or completely loaded
2864                 * no need to do power savings balance at this domain
2865                 */
2866                if (local_group && (this_nr_running >= group_capacity ||
2867                                    !this_nr_running))
2868                        power_savings_balance = 0;
2869
2870                /*
2871                 * If a group is already running at full capacity or idle,
2872                 * don't include that group in power savings calculations
2873                 */
2874                if (!power_savings_balance || sum_nr_running >= group_capacity
2875                    || !sum_nr_running)
2876                        goto group_next;
2877
2878                /*
2879                 * Calculate the group which has the least non-idle load.
2880                 * This is the group from where we need to pick up the load
2881                 * for saving power
2882                 */
2883                if ((sum_nr_running < min_nr_running) ||
2884                    (sum_nr_running == min_nr_running &&
2885                     first_cpu(group->cpumask) <
2886                     first_cpu(group_min->cpumask))) {
2887                        group_min = group;
2888                        min_nr_running = sum_nr_running;
2889                        min_load_per_task = sum_weighted_load /
2890                                                sum_nr_running;
2891                }
2892
2893                /*
2894                 * Calculate the group which is almost near its
2895                 * capacity but still has some space to pick up some load
2896                 * from other group and save more power
2897                 */
2898                if (sum_nr_running <= group_capacity - 1) {
2899                        if (sum_nr_running > leader_nr_running ||
2900                            (sum_nr_running == leader_nr_running &&
2901                             first_cpu(group->cpumask) >
2902                              first_cpu(group_leader->cpumask))) {
2903                                group_leader = group;
2904                                leader_nr_running = sum_nr_running;
2905                        }
2906                }
2907group_next:
2908#endif
2909                group = group->next;
2910        } while (group != sd->groups);
2911
2912        if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2913                goto out_balanced;
2914
2915        avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2916
2917        if (this_load >= avg_load ||
2918                        100*max_load <= sd->imbalance_pct*this_load)
2919                goto out_balanced;
2920
2921        busiest_load_per_task /= busiest_nr_running;
2922        if (group_imb)
2923                busiest_load_per_task = min(busiest_load_per_task, avg_load);
2924
2925        /*
2926         * We're trying to get all the cpus to the average_load, so we don't
2927         * want to push ourselves above the average load, nor do we wish to
2928         * reduce the max loaded cpu below the average load, as either of these
2929         * actions would just result in more rebalancing later, and ping-pong
2930         * tasks around. Thus we look for the minimum possible imbalance.
2931         * Negative imbalances (*we* are more loaded than anyone else) will
2932         * be counted as no imbalance for these purposes -- we can't fix that
2933         * by pulling tasks to us. Be careful of negative numbers as they'll
2934         * appear as very large values with unsigned longs.
2935         */
2936        if (max_load <= busiest_load_per_task)
2937                goto out_balanced;
2938
2939        /*
2940         * In the presence of smp nice balancing, certain scenarios can have
2941         * max load less than avg load(as we skip the groups at or below
2942         * its cpu_power, while calculating max_load..)
2943         */
2944        if (max_load < avg_load) {
2945                *imbalance = 0;
2946                goto small_imbalance;
2947        }
2948
2949        /* Don't want to pull so many tasks that a group would go idle */
2950        max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2951
2952        /* How much load to actually move to equalise the imbalance */
2953        *imbalance = min(max_pull * busiest->__cpu_power,
2954                                (avg_load - this_load) * this->__cpu_power)
2955                        / SCHED_LOAD_SCALE;
2956
2957        /*
2958         * if *imbalance is less than the average load per runnable task
2959         * there is no gaurantee that any tasks will be moved so we'll have
2960         * a think about bumping its value to force at least one task to be
2961         * moved
2962         */
2963        if (*imbalance < busiest_load_per_task) {
2964                unsigned long tmp, pwr_now, pwr_move;
2965                unsigned int imbn;
2966
2967small_imbalance:
2968                pwr_move = pwr_now = 0;
2969                imbn = 2;
2970                if (this_nr_running) {
2971                        this_load_per_task /= this_nr_running;
2972                        if (busiest_load_per_task > this_load_per_task)
2973                                imbn = 1;
2974                } else
2975                        this_load_per_task = SCHED_LOAD_SCALE;
2976
2977                if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2978                                        busiest_load_per_task * imbn) {
2979                        *imbalance = busiest_load_per_task;
2980                        return busiest;
2981                }
2982
2983                /*
2984                 * OK, we don't have enough imbalance to justify moving tasks,
2985                 * however we may be able to increase total CPU power used by
2986                 * moving them.
2987                 */
2988
2989                pwr_now += busiest->__cpu_power *
2990                                min(busiest_load_per_task, max_load);
2991                pwr_now += this->__cpu_power *
2992                                min(this_load_per_task, this_load);
2993                pwr_now /= SCHED_LOAD_SCALE;
2994
2995                /* Amount of load we'd subtract */
2996                tmp = sg_div_cpu_power(busiest,
2997                                busiest_load_per_task * SCHED_LOAD_SCALE);
2998                if (max_load > tmp)
2999                        pwr_move += busiest->__cpu_power *
3000                                min(busiest_load_per_task, max_load - tmp);
3001
3002                /* Amount of load we'd add */
3003                if (max_load * busiest->__cpu_power <
3004                                busiest_load_per_task * SCHED_LOAD_SCALE)
3005                        tmp = sg_div_cpu_power(this,
3006                                        max_load * busiest->__cpu_power);
3007                else
3008                        tmp = sg_div_cpu_power(this,
3009                                busiest_load_per_task * SCHED_LOAD_SCALE);
3010                pwr_move += this->__cpu_power *
3011                                min(this_load_per_task, this_load + tmp);
3012                pwr_move /= SCHED_LOAD_SCALE;
3013
3014                /* Move if we gain throughput */
3015                if (pwr_move > pwr_now)
3016                        *imbalance = busiest_load_per_task;
3017        }
3018
3019        return busiest;
3020
3021out_balanced:
3022#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3023        if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3024                goto ret;
3025
3026        if (this == group_leader && group_leader != group_min) {
3027                *imbalance = min_load_per_task;
3028                return group_min;
3029        }
3030#endif
3031ret:
3032        *imbalance = 0;
3033        return NULL;
3034}
3035
3036/*
3037 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3038 */
3039static struct rq *
3040find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3041                   unsigned long imbalance, cpumask_t *cpus)
3042{
3043        struct rq *busiest = NULL, *rq;
3044        unsigned long max_load = 0;
3045        int i;
3046
3047        for_each_cpu_mask(i, group->cpumask) {
3048                unsigned long wl;
3049
3050                if (!cpu_isset(i, *cpus))
3051                        continue;
3052
3053                rq = cpu_rq(i);
3054                wl = weighted_cpuload(i);
3055
3056                if (rq->nr_running == 1 && wl > imbalance)
3057                        continue;
3058
3059                if (wl > max_load) {
3060                        max_load = wl;
3061                        busiest = rq;
3062                }
3063        }
3064
3065        return busiest;
3066}
3067
3068/*
3069 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3070 * so long as it is large enough.
3071 */
3072#define MAX_PINNED_INTERVAL     512
3073
3074/*
3075 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3076 * tasks if there is an imbalance.
3077 */
3078static int load_balance(int this_cpu, struct rq *this_rq,
3079                        struct sched_domain *sd, enum cpu_idle_type idle,
3080                        int *balance)
3081{
3082        int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3083        struct sched_group *group;
3084        unsigned long imbalance;
3085        struct rq *busiest;
3086        cpumask_t cpus = CPU_MASK_ALL;
3087        unsigned long flags;
3088
3089        /*
3090         * When power savings policy is enabled for the parent domain, idle
3091         * sibling can pick up load irrespective of busy siblings. In this case,
3092         * let the state of idle sibling percolate up as CPU_IDLE, instead of
3093         * portraying it as CPU_NOT_IDLE.
3094         */
3095        if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3096            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3097                sd_idle = 1;
3098
3099        schedstat_inc(sd, lb_count[idle]);
3100
3101redo:
3102        group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3103                                   &cpus, balance);
3104
3105        if (*balance == 0)
3106                goto out_balanced;
3107
3108        if (!group) {
3109                schedstat_inc(sd, lb_nobusyg[idle]);
3110                goto out_balanced;
3111        }
3112
3113        busiest = find_busiest_queue(group, idle, imbalance, &cpus);
3114        if (!busiest) {
3115                schedstat_inc(sd, lb_nobusyq[idle]);
3116                goto out_balanced;
3117        }
3118
3119        BUG_ON(busiest == this_rq);
3120
3121        schedstat_add(sd, lb_imbalance[idle], imbalance);
3122
3123        ld_moved = 0;
3124        if (busiest->nr_running > 1) {
3125                /*
3126                 * Attempt to move tasks. If find_busiest_group has found
3127                 * an imbalance but busiest->nr_running <= 1, the group is
3128                 * still unbalanced. ld_moved simply stays zero, so it is
3129                 * correctly treated as an imbalance.
3130                 */
3131                local_irq_save(flags);
3132                double_rq_lock(this_rq, busiest);
3133                ld_moved = move_tasks(this_rq, this_cpu, busiest,
3134                                      imbalance, sd, idle, &all_pinned);
3135                double_rq_unlock(this_rq, busiest);
3136                local_irq_restore(flags);
3137
3138                /*
3139                 * some other cpu did the load balance for us.
3140                 */
3141                if (ld_moved && this_cpu != smp_processor_id())
3142                        resched_cpu(this_cpu);
3143
3144                /* All tasks on this runqueue were pinned by CPU affinity */
3145                if (unlikely(all_pinned)) {
3146                        cpu_clear(cpu_of(busiest), cpus);
3147                        if (!cpus_empty(cpus))
3148                                goto redo;
3149                        goto out_balanced;
3150                }
3151        }
3152
3153        if (!ld_moved) {
3154                schedstat_inc(sd, lb_failed[idle]);
3155                sd->nr_balance_failed++;
3156
3157                if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3158
3159                        spin_lock_irqsave(&busiest->lock, flags);
3160
3161                        /* don't kick the migration_thread, if the curr
3162                         * task on busiest cpu can't be moved to this_cpu
3163                         */
3164                        if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3165                                spin_unlock_irqrestore(&busiest->lock, flags);
3166                                all_pinned = 1;
3167                                goto out_one_pinned;
3168                        }
3169
3170                        if (!busiest->active_balance) {
3171                                busiest->active_balance = 1;
3172                                busiest->push_cpu = this_cpu;
3173                                active_balance = 1;
3174                        }
3175                        spin_unlock_irqrestore(&busiest->lock, flags);
3176                        if (active_balance)
3177                                wake_up_process(busiest->migration_thread);
3178
3179                        /*
3180                         * We've kicked active balancing, reset the failure
3181                         * counter.
3182                         */
3183                        sd->nr_balance_failed = sd->cache_nice_tries+1;
3184                }
3185        } else
3186                sd->nr_balance_failed = 0;
3187
3188        if (likely(!active_balance)) {
3189                /* We were unbalanced, so reset the balancing interval */
3190                sd->balance_interval = sd->min_interval;
3191        } else {
3192                /*
3193                 * If we've begun active balancing, start to back off. This
3194                 * case may not be covered by the all_pinned logic if there
3195                 * is only 1 task on the busy runqueue (because we don't call
3196                 * move_tasks).
3197                 */
3198                if (sd->balance_interval < sd->max_interval)
3199                        sd->balance_interval *= 2;
3200        }
3201
3202        if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3203            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3204                return -1;
3205        return ld_moved;
3206
3207out_balanced:
3208        schedstat_inc(sd, lb_balanced[idle]);
3209
3210        sd->nr_balance_failed = 0;
3211
3212out_one_pinned:
3213        /* tune up the balancing interval */
3214        if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3215                        (sd->balance_interval < sd->max_interval))
3216                sd->balance_interval *= 2;
3217
3218        if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3219            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3220                return -1;
3221        return 0;
3222}
3223
3224/*
3225 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3226 * tasks if there is an imbalance.
3227 *
3228 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3229 * this_rq is locked.
3230 */
3231static int
3232load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
3233{
3234        struct sched_group *group;
3235        struct rq *busiest = NULL;
3236        unsigned long imbalance;
3237        int ld_moved = 0;
3238        int sd_idle = 0;
3239        int all_pinned = 0;
3240        cpumask_t cpus = CPU_MASK_ALL;
3241
3242        /*
3243         * When power savings policy is enabled for the parent domain, idle
3244         * sibling can pick up load irrespective of busy siblings. In this case,
3245         * let the state of idle sibling percolate up as IDLE, instead of
3246         * portraying it as CPU_NOT_IDLE.
3247         */
3248        if (sd->flags & SD_SHARE_CPUPOWER &&
3249            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3250                sd_idle = 1;
3251
3252        schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3253redo:
3254        group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3255                                   &sd_idle, &cpus, NULL);
3256        if (!group) {
3257                schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3258                goto out_balanced;
3259        }
3260
3261        busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
3262                                &cpus);
3263        if (!busiest) {
3264                schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3265                goto out_balanced;
3266        }
3267
3268        BUG_ON(busiest == this_rq);
3269
3270        schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3271
3272        ld_moved = 0;
3273        if (busiest->nr_running > 1) {
3274                /* Attempt to move tasks */
3275                double_lock_balance(this_rq, busiest);
3276                /* this_rq->clock is already updated */
3277                update_rq_clock(busiest);
3278                ld_moved = move_tasks(this_rq, this_cpu, busiest,
3279                                        imbalance, sd, CPU_NEWLY_IDLE,
3280                                        &all_pinned);
3281                spin_unlock(&busiest->lock);
3282
3283                if (unlikely(all_pinned)) {
3284                        cpu_clear(cpu_of(busiest), cpus);
3285                        if (!cpus_empty(cpus))
3286                                goto redo;
3287                }
3288        }
3289
3290        if (!ld_moved) {
3291                schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3292                if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3293                    !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3294                        return -1;
3295        } else
3296                sd->nr_balance_failed = 0;
3297
3298        return ld_moved;
3299
3300out_balanced:
3301        schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3302        if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3303            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3304                return -1;
3305        sd->nr_balance_failed = 0;
3306
3307        return 0;
3308}
3309
3310/*
3311 * idle_balance is called by schedule() if this_cpu is about to become
3312 * idle. Attempts to pull tasks from other CPUs.
3313 */
3314static void idle_balance(int this_cpu, struct rq *this_rq)
3315{
3316        struct sched_domain *sd;
3317        int pulled_task = -1;
3318        unsigned long next_balance = jiffies + HZ;
3319
3320        for_each_domain(this_cpu, sd) {
3321                unsigned long interval;
3322
3323                if (!(sd->flags & SD_LOAD_BALANCE))
3324                        continue;
3325
3326                if (sd->flags & SD_BALANCE_NEWIDLE)
3327                        /* If we've pulled tasks over stop searching: */
3328                        pulled_task = load_balance_newidle(this_cpu,
3329                                                                this_rq, sd);
3330
3331                interval = msecs_to_jiffies(sd->balance_interval);
3332                if (time_after(next_balance, sd->last_balance + interval))
3333                        next_balance = sd->last_balance + interval;
3334                if (pulled_task)
3335                        break;
3336        }
3337        if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3338                /*
3339                 * We are going idle. next_balance may be set based on
3340                 * a busy processor. So reset next_balance.
3341                 */
3342                this_rq->next_balance = next_balance;
3343        }
3344}
3345
3346/*
3347 * active_load_balance is run by migration threads. It pushes running tasks
3348 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3349 * running on each physical CPU where possible, and avoids physical /
3350 * logical imbalances.
3351 *
3352 * Called with busiest_rq locked.
3353 */
3354static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3355{
3356        int target_cpu = busiest_rq->push_cpu;
3357        struct sched_domain *sd;
3358        struct rq *target_rq;
3359
3360        /* Is there any task to move? */
3361        if (busiest_rq->nr_running <= 1)
3362                return;
3363
3364        target_rq = cpu_rq(target_cpu);
3365
3366        /*
3367         * This condition is "impossible", if it occurs
3368         * we need to fix it. Originally reported by
3369         * Bjorn Helgaas on a 128-cpu setup.
3370         */
3371        BUG_ON(busiest_rq == target_rq);
3372
3373        /* move a task from busiest_rq to target_rq */
3374        double_lock_balance(busiest_rq, target_rq);
3375        update_rq_clock(busiest_rq);
3376        update_rq_clock(target_rq);
3377
3378        /* Search for an sd spanning us and the target CPU. */
3379        for_each_domain(target_cpu, sd) {
3380                if ((sd->flags & SD_LOAD_BALANCE) &&
3381                    cpu_isset(busiest_cpu, sd->span))
3382                                break;
3383        }
3384
3385        if (likely(sd)) {
3386                schedstat_inc(sd, alb_count);
3387
3388                if (move_one_task(target_rq, target_cpu, busiest_rq,
3389                                  sd, CPU_IDLE))
3390                        schedstat_inc(sd, alb_pushed);
3391                else
3392                        schedstat_inc(sd, alb_failed);
3393        }
3394        spin_unlock(&target_rq->lock);
3395}
3396
3397#ifdef CONFIG_NO_HZ
3398static struct {
3399        atomic_t load_balancer;
3400        cpumask_t cpu_mask;
3401} nohz ____cacheline_aligned = {
3402        .load_balancer = ATOMIC_INIT(-1),
3403        .cpu_mask = CPU_MASK_NONE,
3404};
3405
3406/*
3407 * This routine will try to nominate the ilb (idle load balancing)
3408 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3409 * load balancing on behalf of all those cpus. If all the cpus in the system
3410 * go into this tickless mode, then there will be no ilb owner (as there is
3411 * no need for one) and all the cpus will sleep till the next wakeup event
3412 * arrives...
3413 *
3414 * For the ilb owner, tick is not stopped. And this tick will be used
3415 * for idle load balancing. ilb owner will still be part of
3416 * nohz.cpu_mask..
3417 *
3418 * While stopping the tick, this cpu will become the ilb owner if there
3419 * is no other owner. And will be the owner till that cpu becomes busy
3420 * or if all cpus in the system stop their ticks at which point
3421 * there is no need for ilb owner.
3422 *
3423 * When the ilb owner becomes busy, it nominates another owner, during the
3424 * next busy scheduler_tick()
3425 */
3426int select_nohz_load_balancer(int stop_tick)
3427{
3428        int cpu = smp_processor_id();
3429
3430        if (stop_tick) {
3431                cpu_set(cpu, nohz.cpu_mask);
3432                cpu_rq(cpu)->in_nohz_recently = 1;
3433
3434                /*
3435                 * If we are going offline and still the leader, give up!
3436                 */
3437                if (cpu_is_offline(cpu) &&
3438                    atomic_read(&nohz.load_balancer) == cpu) {
3439                        if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3440                                BUG();
3441                        return 0;
3442                }
3443
3444                /* time for ilb owner also to sleep */
3445                if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3446                        if (atomic_read(&nohz.load_balancer) == cpu)
3447                                atomic_set(&nohz.load_balancer, -1);
3448                        return 0;
3449                }
3450
3451                if (atomic_read(&nohz.load_balancer) == -1) {
3452                        /* make me the ilb owner */
3453                        if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3454                                return 1;
3455                } else if (atomic_read(&nohz.load_balancer) == cpu)
3456                        return 1;
3457        } else {
3458                if (!cpu_isset(cpu, nohz.cpu_mask))
3459                        return 0;
3460
3461                cpu_clear(cpu, nohz.cpu_mask);
3462
3463                if (atomic_read(&nohz.load_balancer) == cpu)
3464                        if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3465                                BUG();
3466        }
3467        return 0;
3468}
3469#endif
3470
3471static DEFINE_SPINLOCK(balancing);
3472
3473/*
3474 * It checks each scheduling domain to see if it is due to be balanced,
3475 * and initiates a balancing operation if so.
3476 *
3477 * Balancing parameters are set up in arch_init_sched_domains.
3478 */
3479static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3480{
3481        int balance = 1;
3482        struct rq *rq = cpu_rq(cpu);
3483        unsigned long interval;
3484        struct sched_domain *sd;
3485        /* Earliest time when we have to do rebalance again */
3486        unsigned long next_balance = jiffies + 60*HZ;
3487        int update_next_balance = 0;
3488
3489        for_each_domain(cpu, sd) {
3490                if (!(sd->flags & SD_LOAD_BALANCE))
3491                        continue;
3492
3493                interval = sd->balance_interval;
3494                if (idle != CPU_IDLE)
3495                        interval *= sd->busy_factor;
3496
3497                /* scale ms to jiffies */
3498                interval = msecs_to_jiffies(interval);
3499                if (unlikely(!interval))
3500                        interval = 1;
3501                if (interval > HZ*NR_CPUS/10)
3502                        interval = HZ*NR_CPUS/10;
3503
3504
3505                if (sd->flags & SD_SERIALIZE) {
3506                        if (!spin_trylock(&balancing))
3507                                goto out;
3508                }
3509
3510                if (time_after_eq(jiffies, sd->last_balance + interval)) {
3511                        if (load_balance(cpu, rq, sd, idle, &balance)) {
3512                                /*
3513                                 * We've pulled tasks over so either we're no
3514                                 * longer idle, or one of our SMT siblings is
3515                                 * not idle.
3516                                 */
3517                                idle = CPU_NOT_IDLE;
3518                        }
3519                        sd->last_balance = jiffies;
3520                }
3521                if (sd->flags & SD_SERIALIZE)
3522                        spin_unlock(&balancing);
3523out:
3524                if (time_after(next_balance, sd->last_balance + interval)) {
3525                        next_balance = sd->last_balance + interval;
3526                        update_next_balance = 1;
3527                }
3528
3529                /*
3530                 * Stop the load balance at this level. There is another
3531                 * CPU in our sched group which is doing load balancing more
3532                 * actively.
3533                 */
3534                if (!balance)
3535                        break;
3536        }
3537
3538        /*
3539         * next_balance will be updated only when there is a need.
3540         * When the cpu is attached to null domain for ex, it will not be
3541         * updated.
3542         */
3543        if (likely(update_next_balance))
3544                rq->next_balance = next_balance;
3545}
3546
3547/*
3548 * run_rebalance_domains is triggered when needed from the scheduler tick.
3549 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3550 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3551 */
3552static void run_rebalance_domains(struct softirq_action *h)
3553{
3554        int this_cpu = smp_processor_id();
3555        struct rq *this_rq = cpu_rq(this_cpu);
3556        enum cpu_idle_type idle = this_rq->idle_at_tick ?
3557                                                CPU_IDLE : CPU_NOT_IDLE;
3558
3559        rebalance_domains(this_cpu, idle);
3560
3561#ifdef CONFIG_NO_HZ
3562        /*
3563         * If this cpu is the owner for idle load balancing, then do the
3564         * balancing on behalf of the other idle cpus whose ticks are
3565         * stopped.
3566         */
3567        if (this_rq->idle_at_tick &&
3568            atomic_read(&nohz.load_balancer) == this_cpu) {
3569                cpumask_t cpus = nohz.cpu_mask;
3570                struct rq *rq;
3571                int balance_cpu;
3572
3573                cpu_clear(this_cpu, cpus);
3574                for_each_cpu_mask(balance_cpu, cpus) {
3575                        /*
3576                         * If this cpu gets work to do, stop the load balancing
3577                         * work being done for other cpus. Next load
3578                         * balancing owner will pick it up.
3579                         */
3580                        if (need_resched())
3581                                break;
3582
3583                        rebalance_domains(balance_cpu, CPU_IDLE);
3584
3585                        rq = cpu_rq(balance_cpu);
3586                        if (time_after(this_rq->next_balance, rq->next_balance))
3587                                this_rq->next_balance = rq->next_balance;
3588                }
3589        }
3590#endif
3591}
3592
3593/*
3594 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3595 *
3596 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3597 * idle load balancing owner or decide to stop the periodic load balancing,
3598 * if the whole system is idle.
3599 */
3600static inline void trigger_load_balance(struct rq *rq, int cpu)
3601{
3602#ifdef CONFIG_NO_HZ
3603        /*
3604         * If we were in the nohz mode recently and busy at the current
3605         * scheduler tick, then check if we need to nominate new idle
3606         * load balancer.
3607         */
3608        if (rq->in_nohz_recently && !rq->idle_at_tick) {
3609                rq->in_nohz_recently = 0;
3610
3611                if (atomic_read(&nohz.load_balancer) == cpu) {
3612                        cpu_clear(cpu, nohz.cpu_mask);
3613                        atomic_set(&nohz.load_balancer, -1);
3614                }
3615
3616                if (atomic_read(&nohz.load_balancer) == -1) {
3617                        /*
3618                         * simple selection for now: Nominate the
3619                         * first cpu in the nohz list to be the next
3620                         * ilb owner.
3621                         *
3622                         * TBD: Traverse the sched domains and nominate
3623                         * the nearest cpu in the nohz.cpu_mask.
3624                         */
3625                        int ilb = first_cpu(nohz.cpu_mask);
3626
3627                        if (ilb != NR_CPUS)
3628                                resched_cpu(ilb);
3629                }
3630        }
3631
3632        /*
3633         * If this cpu is idle and doing idle load balancing for all the
3634         * cpus with ticks stopped, is it time for that to stop?
3635         */
3636        if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3637            cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3638                resched_cpu(cpu);
3639                return;
3640        }
3641
3642        /*
3643         * If this cpu is idle and the idle load balancing is done by
3644         * someone else, then no need raise the SCHED_SOFTIRQ
3645         */
3646        if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3647            cpu_isset(cpu, nohz.cpu_mask))
3648                return;
3649#endif
3650        if (time_after_eq(jiffies, rq->next_balance))
3651                raise_softirq(SCHED_SOFTIRQ);
3652}
3653
3654#else   /* CONFIG_SMP */
3655
3656/*
3657 * on UP we do not need to balance between CPUs:
3658 */
3659static inline void idle_balance(int cpu, struct rq *rq)
3660{
3661}
3662
3663#endif
3664
3665DEFINE_PER_CPU(struct kernel_stat, kstat);
3666
3667EXPORT_PER_CPU_SYMBOL(kstat);
3668
3669/*
3670 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3671 * that have not yet been banked in case the task is currently running.
3672 */
3673unsigned long long task_sched_runtime(struct task_struct *p)
3674{
3675        unsigned long flags;
3676        u64 ns, delta_exec;
3677        struct rq *rq;
3678
3679        rq = task_rq_lock(p, &flags);
3680        ns = p->se.sum_exec_runtime;
3681        if (task_current(rq, p)) {
3682                update_rq_clock(rq);
3683                delta_exec = rq->clock - p->se.exec_start;
3684                if ((s64)delta_exec > 0)
3685                        ns += delta_exec;
3686        }
3687        task_rq_unlock(rq, &flags);
3688
3689        return ns;
3690}
3691
3692/*
3693 * Account user cpu time to a process.
3694 * @p: the process that the cpu time gets accounted to
3695 * @cputime: the cpu time spent in user space since the last update
3696 */
3697void account_user_time(struct task_struct *p, cputime_t cputime)
3698{
3699        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3700        cputime64_t tmp;
3701
3702        p->utime = cputime_add(p->utime, cputime);
3703
3704        /* Add user time to cpustat. */
3705        tmp = cputime_to_cputime64(cputime);
3706        if (TASK_NICE(p) > 0)
3707                cpustat->nice = cputime64_add(cpustat->nice, tmp);
3708        else
3709                cpustat->user = cputime64_add(cpustat->user, tmp);
3710}
3711
3712/*
3713 * Account guest cpu time to a process.
3714 * @p: the process that the cpu time gets accounted to
3715 * @cputime: the cpu time spent in virtual machine since the last update
3716 */
3717static void account_guest_time(struct task_struct *p, cputime_t cputime)
3718{
3719        cputime64_t tmp;
3720        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3721
3722        tmp = cputime_to_cputime64(cputime);
3723
3724        p->utime = cputime_add(p->utime, cputime);
3725        p->gtime = cputime_add(p->gtime, cputime);
3726
3727        cpustat->user = cputime64_add(cpustat->user, tmp);
3728        cpustat->guest = cputime64_add(cpustat->guest, tmp);
3729}
3730
3731/*
3732 * Account scaled user cpu time to a process.
3733 * @p: the process that the cpu time gets accounted to
3734 * @cputime: the cpu time spent in user space since the last update
3735 */
3736void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3737{
3738        p->utimescaled = cputime_add(p->utimescaled, cputime);
3739}
3740
3741/*
3742 * Account system cpu time to a process.
3743 * @p: the process that the cpu time gets accounted to
3744 * @hardirq_offset: the offset to subtract from hardirq_count()
3745 * @cputime: the cpu time spent in kernel space since the last update
3746 */
3747void account_system_time(struct task_struct *p, int hardirq_offset,
3748                         cputime_t cputime)
3749{
3750        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3751        struct rq *rq = this_rq();
3752        cputime64_t tmp;
3753
3754        if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3755                return account_guest_time(p, cputime);
3756
3757        p->stime = cputime_add(p->stime, cputime);
3758
3759        /* Add system time to cpustat. */
3760        tmp = cputime_to_cputime64(cputime);
3761        if (hardirq_count() - hardirq_offset)
3762                cpustat->irq = cputime64_add(cpustat->irq, tmp);
3763        else if (softirq_count())
3764                cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3765        else if (p != rq->idle)
3766                cpustat->system = cputime64_add(cpustat->system, tmp);
3767        else if (atomic_read(&rq->nr_iowait) > 0)
3768                cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3769        else
3770                cpustat->idle = cputime64_add(cpustat->idle, tmp);
3771        /* Account for system time used */
3772        acct_update_integrals(p);
3773}
3774
3775/*
3776 * Account scaled system cpu time to a process.
3777 * @p: the process that the cpu time gets accounted to
3778 * @hardirq_offset: the offset to subtract from hardirq_count()
3779 * @cputime: the cpu time spent in kernel space since the last update
3780 */
3781void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3782{
3783        p->stimescaled = cputime_add(p->stimescaled, cputime);
3784}
3785
3786/*
3787 * Account for involuntary wait time.
3788 * @p: the process from which the cpu time has been stolen
3789 * @steal: the cpu time spent in involuntary wait
3790 */
3791void account_steal_time(struct task_struct *p, cputime_t steal)
3792{
3793        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3794        cputime64_t tmp = cputime_to_cputime64(steal);
3795        struct rq *rq = this_rq();
3796
3797        if (p == rq->idle) {
3798                p->stime = cputime_add(p->stime, steal);
3799                if (atomic_read(&rq->nr_iowait) > 0)
3800                        cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3801                else
3802                        cpustat->idle = cputime64_add(cpustat->idle, tmp);
3803        } else
3804                cpustat->steal = cputime64_add(cpustat->steal, tmp);
3805}
3806
3807/*
3808 * This function gets called by the timer code, with HZ frequency.
3809 * We call it with interrupts disabled.
3810 *
3811 * It also gets called by the fork code, when changing the parent's
3812 * timeslices.
3813 */
3814void scheduler_tick(void)
3815{
3816        int cpu = smp_processor_id();
3817        struct rq *rq = cpu_rq(cpu);
3818        struct task_struct *curr = rq->curr;
3819        u64 next_tick = rq->tick_timestamp + TICK_NSEC;
3820
3821        spin_lock(&rq->lock);
3822        __update_rq_clock(rq);
3823        /*
3824         * Let rq->clock advance by at least TICK_NSEC:
3825         */
3826        if (unlikely(rq->clock < next_tick)) {
3827                rq->clock = next_tick;
3828                rq->clock_underflows++;
3829        }
3830        rq->tick_timestamp = rq->clock;
3831        update_cpu_load(rq);
3832        curr->sched_class->task_tick(rq, curr, 0);
3833        update_sched_rt_period(rq);
3834        spin_unlock(&rq->lock);
3835
3836#ifdef CONFIG_SMP
3837        rq->idle_at_tick = idle_cpu(cpu);
3838        trigger_load_balance(rq, cpu);
3839#endif
3840}
3841
3842#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3843
3844void __kprobes add_preempt_count(int val)
3845{
3846        /*
3847         * Underflow?
3848         */
3849        if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3850                return;
3851        preempt_count() += val;
3852        /*
3853         * Spinlock count overflowing soon?
3854         */
3855        DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3856                                PREEMPT_MASK - 10);
3857}
3858EXPORT_SYMBOL(add_preempt_count);
3859
3860void __kprobes sub_preempt_count(int val)
3861{
3862        /*
3863         * Underflow?
3864         */
3865        if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3866                return;
3867        /*
3868         * Is the spinlock portion underflowing?
3869         */
3870        if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3871                        !(preempt_count() & PREEMPT_MASK)))
3872                return;
3873
3874        preempt_count() -= val;
3875}
3876EXPORT_SYMBOL(sub_preempt_count);
3877
3878#endif
3879
3880/*
3881 * Print scheduling while atomic bug:
3882 */
3883static noinline void __schedule_bug(struct task_struct *prev)
3884{
3885        struct pt_regs *regs = get_irq_regs();
3886
3887        printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3888                prev->comm, prev->pid, preempt_count());
3889
3890        debug_show_held_locks(prev);
3891        if (irqs_disabled())
3892                print_irqtrace_events(prev);
3893
3894        if (regs)
3895                show_regs(regs);
3896        else
3897                dump_stack();
3898}
3899
3900/*
3901 * Various schedule()-time debugging checks and statistics:
3902 */
3903static inline void schedule_debug(struct task_struct *prev)
3904{
3905        /*
3906         * Test if we are atomic. Since do_exit() needs to call into
3907         * schedule() atomically, we ignore that path for now.
3908         * Otherwise, whine if we are scheduling when we should not be.
3909         */
3910        if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3911                __schedule_bug(prev);
3912
3913        profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3914
3915        schedstat_inc(this_rq(), sched_count);
3916#ifdef CONFIG_SCHEDSTATS
3917        if (unlikely(prev->lock_depth >= 0)) {
3918                schedstat_inc(this_rq(), bkl_count);
3919                schedstat_inc(prev, sched_info.bkl_count);
3920        }
3921#endif
3922}
3923
3924/*
3925 * Pick up the highest-prio task:
3926 */
3927static inline struct task_struct *
3928pick_next_task(struct rq *rq, struct task_struct *prev)
3929{
3930        const struct sched_class *class;
3931        struct task_struct *p;
3932
3933        /*
3934         * Optimization: we know that if all tasks are in
3935         * the fair class we can call that function directly:
3936         */
3937        if (likely(rq->nr_running == rq->cfs.nr_running)) {
3938                p = fair_sched_class.pick_next_task(rq);
3939                if (likely(p))
3940                        return p;
3941        }
3942
3943        class = sched_class_highest;
3944        for ( ; ; ) {
3945                p = class->pick_next_task(rq);
3946                if (p)
3947                        return p;
3948                /*
3949                 * Will never be NULL as the idle class always
3950                 * returns a non-NULL p:
3951                 */
3952                class = class->next;
3953        }
3954}
3955
3956/*
3957 * schedule() is the main scheduler function.
3958 */
3959asmlinkage void __sched schedule(void)
3960{
3961        struct task_struct *prev, *next;
3962        unsigned long *switch_count;
3963        struct rq *rq;
3964        int cpu;
3965
3966need_resched:
3967        preempt_disable();
3968        cpu = smp_processor_id();
3969        rq = cpu_rq(cpu);
3970        rcu_qsctr_inc(cpu);
3971        prev = rq->curr;
3972        switch_count = &prev->nivcsw;
3973
3974        release_kernel_lock(prev);
3975need_resched_nonpreemptible:
3976
3977        schedule_debug(prev);
3978
3979        hrtick_clear(rq);
3980
3981        /*
3982         * Do the rq-clock update outside the rq lock:
3983         */
3984        local_irq_disable();
3985        __update_rq_clock(rq);
3986        spin_lock(&rq->lock);
3987        clear_tsk_need_resched(prev);
3988
3989        if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3990                if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3991                                signal_pending(prev))) {
3992                        prev->state = TASK_RUNNING;
3993                } else {
3994                        deactivate_task(rq, prev, 1);
3995                }
3996                switch_count = &prev->nvcsw;
3997        }
3998
3999#ifdef CONFIG_SMP
4000        if (prev->sched_class->pre_schedule)
4001                prev->sched_class->pre_schedule(rq, prev);
4002#endif
4003
4004        if (unlikely(!rq->nr_running))
4005                idle_balance(cpu, rq);
4006
4007        prev->sched_class->put_prev_task(rq, prev);
4008        next = pick_next_task(rq, prev);
4009
4010        sched_info_switch(prev, next);
4011
4012        if (likely(prev != next)) {
4013                rq->nr_switches++;
4014                rq->curr = next;
4015                ++*switch_count;
4016
4017                context_switch(rq, prev, next); /* unlocks the rq */
4018                /*
4019                 * the context switch might have flipped the stack from under
4020                 * us, hence refresh the local variables.
4021                 */
4022                cpu = smp_processor_id();
4023                rq = cpu_rq(cpu);
4024        } else
4025                spin_unlock_irq(&rq->lock);
4026
4027        hrtick_set(rq);
4028
4029        if (unlikely(reacquire_kernel_lock(current) < 0))
4030                goto need_resched_nonpreemptible;
4031
4032        preempt_enable_no_resched();
4033        if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4034                goto need_resched;
4035}
4036EXPORT_SYMBOL(schedule);
4037
4038#ifdef CONFIG_PREEMPT
4039/*
4040 * this is the entry point to schedule() from in-kernel preemption
4041 * off of preempt_enable. Kernel preemptions off return from interrupt
4042 * occur there and call schedule directly.
4043 */
4044asmlinkage void __sched preempt_schedule(void)
4045{
4046        struct thread_info *ti = current_thread_info();
4047        struct task_struct *task = current;
4048        int saved_lock_depth;
4049
4050        /*
4051         * If there is a non-zero preempt_count or interrupts are disabled,
4052         * we do not want to preempt the current task. Just return..
4053         */
4054        if (likely(ti->preempt_count || irqs_disabled()))
4055                return;
4056
4057        do {
4058                add_preempt_count(PREEMPT_ACTIVE);
4059
4060                /*
4061                 * We keep the big kernel semaphore locked, but we
4062                 * clear ->lock_depth so that schedule() doesnt
4063                 * auto-release the semaphore:
4064                 */
4065                saved_lock_depth = task->lock_depth;
4066                task->lock_depth = -1;
4067                schedule();
4068                task->lock_depth = saved_lock_depth;
4069                sub_preempt_count(PREEMPT_ACTIVE);
4070
4071                /*
4072                 * Check again in case we missed a preemption opportunity
4073                 * between schedule and now.
4074                 */
4075                barrier();
4076        } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4077}
4078EXPORT_SYMBOL(preempt_schedule);
4079
4080/*
4081 * this is the entry point to schedule() from kernel preemption
4082 * off of irq context.
4083 * Note, that this is called and return with irqs disabled. This will
4084 * protect us against recursive calling from irq.
4085 */
4086asmlinkage void __sched preempt_schedule_irq(void)
4087{
4088        struct thread_info *ti = current_thread_info();
4089        struct task_struct *task = current;
4090        int saved_lock_depth;
4091
4092        /* Catch callers which need to be fixed */
4093        BUG_ON(ti->preempt_count || !irqs_disabled());
4094
4095        do {
4096                add_preempt_count(PREEMPT_ACTIVE);
4097
4098                /*
4099                 * We keep the big kernel semaphore locked, but we
4100                 * clear ->lock_depth so that schedule() doesnt
4101                 * auto-release the semaphore:
4102                 */
4103                saved_lock_depth = task->lock_depth;
4104                task->lock_depth = -1;
4105                local_irq_enable();
4106                schedule();
4107                local_irq_disable();
4108                task->lock_depth = saved_lock_depth;
4109                sub_preempt_count(PREEMPT_ACTIVE);
4110
4111                /*
4112                 * Check again in case we missed a preemption opportunity
4113                 * between schedule and now.
4114                 */
4115                barrier();
4116        } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4117}
4118
4119#endif /* CONFIG_PREEMPT */
4120
4121int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4122                          void *key)
4123{
4124        return try_to_wake_up(curr->private, mode, sync);
4125}
4126EXPORT_SYMBOL(default_wake_function);
4127
4128/*
4129 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4130 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4131 * number) then we wake all the non-exclusive tasks and one exclusive task.
4132 *
4133 * There are circumstances in which we can try to wake a task which has already
4134 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4135 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4136 */
4137static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4138                             int nr_exclusive, int sync, void *key)
4139{
4140        wait_queue_t *curr, *next;
4141
4142        list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4143                unsigned flags = curr->flags;
4144
4145                if (curr->func(curr, mode, sync, key) &&
4146                                (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4147                        break;
4148        }
4149}
4150
4151/**
4152 * __wake_up - wake up threads blocked on a waitqueue.
4153 * @q: the waitqueue
4154 * @mode: which threads
4155 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4156 * @key: is directly passed to the wakeup function
4157 */
4158void __wake_up(wait_queue_head_t *q, unsigned int mode,
4159                        int nr_exclusive, void *key)
4160{
4161        unsigned long flags;
4162
4163        spin_lock_irqsave(&q->lock, flags);
4164        __wake_up_common(q, mode, nr_exclusive, 0, key);
4165        spin_unlock_irqrestore(&q->lock, flags);
4166}
4167EXPORT_SYMBOL(__wake_up);
4168
4169/*
4170 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4171 */
4172void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4173{
4174        __wake_up_common(q, mode, 1, 0, NULL);
4175}
4176
4177/**
4178 * __wake_up_sync - wake up threads blocked on a waitqueue.
4179 * @q: the waitqueue
4180 * @mode: which threads
4181 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4182 *
4183 * The sync wakeup differs that the waker knows that it will schedule
4184 * away soon, so while the target thread will be woken up, it will not
4185 * be migrated to another CPU - ie. the two threads are 'synchronized'
4186 * with each other. This can prevent needless bouncing between CPUs.
4187 *
4188 * On UP it can prevent extra preemption.
4189 */
4190void
4191__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4192{
4193        unsigned long flags;
4194        int sync = 1;
4195
4196        if (unlikely(!q))
4197                return;
4198
4199        if (unlikely(!nr_exclusive))
4200                sync = 0;
4201
4202        spin_lock_irqsave(&q->lock, flags);
4203        __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4204        spin_unlock_irqrestore(&q->lock, flags);
4205}
4206EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
4207
4208void complete(struct completion *x)
4209{
4210        unsigned long flags;
4211
4212        spin_lock_irqsave(&x->wait.lock, flags);
4213        x->done++;
4214        __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4215        spin_unlock_irqrestore(&x->wait.lock, flags);
4216}
4217EXPORT_SYMBOL(complete);
4218
4219void complete_all(struct completion *x)
4220{
4221        unsigned long flags;
4222
4223        spin_lock_irqsave(&x->wait.lock, flags);
4224        x->done += UINT_MAX/2;
4225        __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4226        spin_unlock_irqrestore(&x->wait.lock, flags);
4227}
4228EXPORT_SYMBOL(complete_all);
4229
4230static inline long __sched
4231do_wait_for_common(struct completion *x, long timeout, int state)
4232{
4233        if (!x->done) {
4234                DECLARE_WAITQUEUE(wait, current);
4235
4236                wait.flags |= WQ_FLAG_EXCLUSIVE;
4237                __add_wait_queue_tail(&x->wait, &wait);
4238                do {
4239                        if ((state == TASK_INTERRUPTIBLE &&
4240                             signal_pending(current)) ||
4241                            (state == TASK_KILLABLE &&
4242                             fatal_signal_pending(current))) {
4243                                __remove_wait_queue(&x->wait, &wait);
4244                                return -ERESTARTSYS;
4245                        }
4246                        __set_current_state(state);
4247                        spin_unlock_irq(&x->wait.lock);
4248                        timeout = schedule_timeout(timeout);
4249                        spin_lock_irq(&x->wait.lock);
4250                        if (!timeout) {
4251                                __remove_wait_queue(&x->wait, &wait);
4252                                return timeout;
4253                        }
4254                } while (!x->done);
4255                __remove_wait_queue(&x->wait, &wait);
4256        }
4257        x->done--;
4258        return timeout;
4259}
4260
4261static long __sched
4262wait_for_common(struct completion *x, long timeout, int state)
4263{
4264        might_sleep();
4265
4266        spin_lock_irq(&x->wait.lock);
4267        timeout = do_wait_for_common(x, timeout, state);
4268        spin_unlock_irq(&x->wait.lock);
4269        return timeout;
4270}
4271
4272void __sched wait_for_completion(struct completion *x)
4273{
4274        wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4275}
4276EXPORT_SYMBOL(wait_for_completion);
4277
4278unsigned long __sched
4279wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4280{
4281        return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4282}
4283EXPORT_SYMBOL(wait_for_completion_timeout);
4284
4285int __sched wait_for_completion_interruptible(struct completion *x)
4286{
4287        long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4288        if (t == -ERESTARTSYS)
4289                return t;
4290        return 0;
4291}
4292EXPORT_SYMBOL(wait_for_completion_interruptible);
4293
4294unsigned long __sched
4295wait_for_completion_interruptible_timeout(struct completion *x,
4296                                          unsigned long timeout)
4297{
4298        return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4299}
4300EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4301
4302int __sched wait_for_completion_killable(struct completion *x)
4303{
4304        long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4305        if (t == -ERESTARTSYS)
4306                return t;
4307        return 0;
4308}
4309EXPORT_SYMBOL(wait_for_completion_killable);
4310
4311static long __sched
4312sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4313{
4314        unsigned long flags;
4315        wait_queue_t wait;
4316
4317        init_waitqueue_entry(&wait, current);
4318
4319        __set_current_state(state);
4320
4321        spin_lock_irqsave(&q->lock, flags);
4322        __add_wait_queue(q, &wait);
4323        spin_unlock(&q->lock);
4324        timeout = schedule_timeout(timeout);
4325        spin_lock_irq(&q->lock);
4326        __remove_wait_queue(q, &wait);
4327        spin_unlock_irqrestore(&q->lock, flags);
4328
4329        return timeout;
4330}
4331
4332void __sched interruptible_sleep_on(wait_queue_head_t *q)
4333{
4334        sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4335}
4336EXPORT_SYMBOL(interruptible_sleep_on);
4337
4338long __sched
4339interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4340{
4341        return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4342}
4343EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4344
4345void __sched sleep_on(wait_queue_head_t *q)
4346{
4347        sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4348}
4349EXPORT_SYMBOL(sleep_on);
4350
4351long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4352{
4353        return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4354}
4355EXPORT_SYMBOL(sleep_on_timeout);
4356
4357#ifdef CONFIG_RT_MUTEXES
4358
4359/*
4360 * rt_mutex_setprio - set the current priority of a task
4361 * @p: task
4362 * @prio: prio value (kernel-internal form)
4363 *
4364 * This function changes the 'effective' priority of a task. It does
4365 * not touch ->normal_prio like __setscheduler().
4366 *
4367 * Used by the rt_mutex code to implement priority inheritance logic.
4368 */
4369void rt_mutex_setprio(struct task_struct *p, int prio)
4370{
4371        unsigned long flags;
4372        int oldprio, on_rq, running;
4373        struct rq *rq;
4374        const struct sched_class *prev_class = p->sched_class;
4375
4376        BUG_ON(prio < 0 || prio > MAX_PRIO);
4377
4378        rq = task_rq_lock(p, &flags);
4379        update_rq_clock(rq);
4380
4381        oldprio = p->prio;
4382        on_rq = p->se.on_rq;
4383        running = task_current(rq, p);
4384        if (on_rq)
4385                dequeue_task(rq, p, 0);
4386        if (running)
4387                p->sched_class->put_prev_task(rq, p);
4388
4389        if (rt_prio(prio))
4390                p->sched_class = &rt_sched_class;
4391        else
4392                p->sched_class = &fair_sched_class;
4393
4394        p->prio = prio;
4395
4396        if (running)
4397                p->sched_class->set_curr_task(rq);
4398        if (on_rq) {
4399                enqueue_task(rq, p, 0);
4400
4401                check_class_changed(rq, p, prev_class, oldprio, running);
4402        }
4403        task_rq_unlock(rq, &flags);
4404}
4405
4406#endif
4407
4408void set_user_nice(struct task_struct *p, long nice)
4409{
4410        int old_prio, delta, on_rq;
4411        unsigned long flags;
4412        struct rq *rq;
4413
4414        if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4415                return;
4416        /*
4417         * We have to be careful, if called from sys_setpriority(),
4418         * the task might be in the middle of scheduling on another CPU.
4419         */
4420        rq = task_rq_lock(p, &flags);
4421        update_rq_clock(rq);
4422        /*
4423         * The RT priorities are set via sched_setscheduler(), but we still
4424         * allow the 'normal' nice value to be set - but as expected
4425         * it wont have any effect on scheduling until the task is
4426         * SCHED_FIFO/SCHED_RR:
4427         */
4428        if (task_has_rt_policy(p)) {
4429                p->static_prio = NICE_TO_PRIO(nice);
4430                goto out_unlock;
4431        }
4432        on_rq = p->se.on_rq;
4433        if (on_rq) {
4434                dequeue_task(rq, p, 0);
4435                dec_load(rq, p);
4436        }
4437
4438        p->static_prio = NICE_TO_PRIO(nice);
4439        set_load_weight(p);
4440        old_prio = p->prio;
4441        p->prio = effective_prio(p);
4442        delta = p->prio - old_prio;
4443
4444        if (on_rq) {
4445                enqueue_task(rq, p, 0);
4446                inc_load(rq, p);
4447                /*
4448                 * If the task increased its priority or is running and
4449                 * lowered its priority, then reschedule its CPU:
4450                 */
4451                if (delta < 0 || (delta > 0 && task_running(rq, p)))
4452                        resched_task(rq->curr);
4453        }
4454out_unlock:
4455        task_rq_unlock(rq, &flags);
4456}
4457EXPORT_SYMBOL(set_user_nice);
4458
4459/*
4460 * can_nice - check if a task can reduce its nice value
4461 * @p: task
4462 * @nice: nice value
4463 */
4464int can_nice(const struct task_struct *p, const int nice)
4465{
4466        /* convert nice value [19,-20] to rlimit style value [1,40] */
4467        int nice_rlim = 20 - nice;
4468
4469        return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4470                capable(CAP_SYS_NICE));
4471}
4472
4473#ifdef __ARCH_WANT_SYS_NICE
4474
4475/*
4476 * sys_nice - change the priority of the current process.
4477 * @increment: priority increment
4478 *
4479 * sys_setpriority is a more generic, but much slower function that
4480 * does similar things.
4481 */
4482asmlinkage long sys_nice(int increment)
4483{
4484        long nice, retval;
4485
4486        /*
4487         * Setpriority might change our priority at the same moment.
4488         * We don't have to worry. Conceptually one call occurs first
4489         * and we have a single winner.
4490         */
4491        if (increment < -40)
4492                increment = -40;
4493        if (increment > 40)
4494                increment = 40;
4495
4496        nice = PRIO_TO_NICE(current->static_prio) + increment;
4497        if (nice < -20)
4498                nice = -20;
4499        if (nice > 19)
4500                nice = 19;
4501
4502        if (increment < 0 && !can_nice(current, nice))
4503                return -EPERM;
4504
4505        retval = security_task_setnice(current, nice);
4506        if (retval)
4507                return retval;
4508
4509        set_user_nice(current, nice);
4510        return 0;
4511}
4512
4513#endif
4514
4515/**
4516 * task_prio - return the priority value of a given task.
4517 * @p: the task in question.
4518 *
4519 * This is the priority value as seen by users in /proc.
4520 * RT tasks are offset by -200. Normal tasks are centered
4521 * around 0, value goes from -16 to +15.
4522 */
4523int task_prio(const struct task_struct *p)
4524{
4525        return p->prio - MAX_RT_PRIO;
4526}
4527
4528/**
4529 * task_nice - return the nice value of a given task.
4530 * @p: the task in question.
4531 */
4532int task_nice(const struct task_struct *p)
4533{
4534        return TASK_NICE(p);
4535}
4536EXPORT_SYMBOL(task_nice);
4537
4538/**
4539 * idle_cpu - is a given cpu idle currently?
4540 * @cpu: the processor in question.
4541 */
4542int idle_cpu(int cpu)
4543{
4544        return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4545}
4546
4547/**
4548 * idle_task - return the idle task for a given cpu.
4549 * @cpu: the processor in question.
4550 */
4551struct task_struct *idle_task(int cpu)
4552{
4553        return cpu_rq(cpu)->idle;
4554}
4555
4556/**
4557 * find_process_by_pid - find a process with a matching PID value.
4558 * @pid: the pid in question.
4559 */
4560static struct task_struct *find_process_by_pid(pid_t pid)
4561{
4562        return pid ? find_task_by_vpid(pid) : current;
4563}
4564
4565/* Actually do priority change: must hold rq lock. */
4566static void
4567__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4568{
4569        BUG_ON(p->se.on_rq);
4570
4571        p->policy = policy;
4572        switch (p->policy) {
4573        case SCHED_NORMAL:
4574        case SCHED_BATCH:
4575        case SCHED_IDLE:
4576                p->sched_class = &fair_sched_class;
4577                break;
4578        case SCHED_FIFO:
4579        case SCHED_RR:
4580                p->sched_class = &rt_sched_class;
4581                break;
4582        }
4583
4584        p->rt_priority = prio;
4585        p->normal_prio = normal_prio(p);
4586        /* we are holding p->pi_lock already */
4587        p->prio = rt_mutex_getprio(p);
4588        set_load_weight(p);
4589}
4590
4591/**
4592 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4593 * @p: the task in question.
4594 * @policy: new policy.
4595 * @param: structure containing the new RT priority.
4596 *
4597 * NOTE that the task may be already dead.
4598 */
4599int sched_setscheduler(struct task_struct *p, int policy,
4600                       struct sched_param *param)
4601{
4602        int retval, oldprio, oldpolicy = -1, on_rq, running;
4603        unsigned long flags;
4604        const struct sched_class *prev_class = p->sched_class;
4605        struct rq *rq;
4606
4607        /* may grab non-irq protected spin_locks */
4608        BUG_ON(in_interrupt());
4609recheck:
4610        /* double check policy once rq lock held */
4611        if (policy < 0)
4612                policy = oldpolicy = p->policy;
4613        else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4614                        policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4615                        policy != SCHED_IDLE)
4616                return -EINVAL;
4617        /*
4618         * Valid priorities for SCHED_FIFO and SCHED_RR are
4619         * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4620         * SCHED_BATCH and SCHED_IDLE is 0.
4621         */
4622        if (param->sched_priority < 0 ||
4623            (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4624            (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4625                return -EINVAL;
4626        if (rt_policy(policy) != (param->sched_priority != 0))
4627                return -EINVAL;
4628
4629        /*
4630         * Allow unprivileged RT tasks to decrease priority:
4631         */
4632        if (!capable(CAP_SYS_NICE)) {
4633                if (rt_policy(policy)) {
4634                        unsigned long rlim_rtprio;
4635
4636                        if (!lock_task_sighand(p, &flags))
4637                                return -ESRCH;
4638                        rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4639                        unlock_task_sighand(p, &flags);
4640
4641                        /* can't set/change the rt policy */
4642                        if (policy != p->policy && !rlim_rtprio)
4643                                return -EPERM;
4644
4645                        /* can't increase priority */
4646                        if (param->sched_priority > p->rt_priority &&
4647                            param->sched_priority > rlim_rtprio)
4648                                return -EPERM;
4649                }
4650                /*
4651                 * Like positive nice levels, dont allow tasks to
4652                 * move out of SCHED_IDLE either:
4653                 */
4654                if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4655                        return -EPERM;
4656
4657                /* can't change other user's priorities */
4658                if ((current->euid != p->euid) &&
4659                    (current->euid != p->uid))
4660                        return -EPERM;
4661        }
4662
4663#ifdef CONFIG_RT_GROUP_SCHED
4664        /*
4665         * Do not allow realtime tasks into groups that have no runtime
4666         * assigned.
4667         */
4668        if (rt_policy(policy) && task_group(p)->rt_runtime == 0)
4669                return -EPERM;
4670#endif
4671
4672        retval = security_task_setscheduler(p, policy, param);
4673        if (retval)
4674                return retval;
4675        /*
4676         * make sure no PI-waiters arrive (or leave) while we are
4677         * changing the priority of the task:
4678         */
4679        spin_lock_irqsave(&p->pi_lock, flags);
4680        /*
4681         * To be able to change p->policy safely, the apropriate
4682         * runqueue lock must be held.
4683         */
4684        rq = __task_rq_lock(p);
4685        /* recheck policy now with rq lock held */
4686        if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4687                policy = oldpolicy = -1;
4688                __task_rq_unlock(rq);
4689                spin_unlock_irqrestore(&p->pi_lock, flags);
4690                goto recheck;
4691        }
4692        update_rq_clock(rq);
4693        on_rq = p->se.on_rq;
4694        running = task_current(rq, p);
4695        if (on_rq)
4696                deactivate_task(rq, p, 0);
4697        if (running)
4698                p->sched_class->put_prev_task(rq, p);
4699
4700        oldprio = p->prio;
4701        __setscheduler(rq, p, policy, param->sched_priority);
4702
4703        if (running)
4704                p->sched_class->set_curr_task(rq);
4705        if (on_rq) {
4706                activate_task(rq, p, 0);
4707
4708                check_class_changed(rq, p, prev_class, oldprio, running);
4709        }
4710        __task_rq_unlock(rq);
4711        spin_unlock_irqrestore(&p->pi_lock, flags);
4712
4713        rt_mutex_adjust_pi(p);
4714
4715        return 0;
4716}
4717EXPORT_SYMBOL_GPL(sched_setscheduler);
4718
4719static int
4720do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4721{
4722        struct sched_param lparam;
4723        struct task_struct *p;
4724        int retval;
4725
4726        if (!param || pid < 0)
4727                return -EINVAL;
4728        if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4729                return -EFAULT;
4730
4731        rcu_read_lock();
4732        retval = -ESRCH;
4733        p = find_process_by_pid(pid);
4734        if (p != NULL)
4735                retval = sched_setscheduler(p, policy, &lparam);
4736        rcu_read_unlock();
4737
4738        return retval;
4739}
4740
4741/**
4742 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4743 * @pid: the pid in question.
4744 * @policy: new policy.
4745 * @param: structure containing the new RT priority.
4746 */
4747asmlinkage long
4748sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4749{
4750        /* negative values for policy are not valid */
4751        if (policy < 0)
4752                return -EINVAL;
4753
4754        return do_sched_setscheduler(pid, policy, param);
4755}
4756
4757/**
4758 * sys_sched_setparam - set/change the RT priority of a thread
4759 * @pid: the pid in question.
4760 * @param: structure containing the new RT priority.
4761 */
4762asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4763{
4764        return do_sched_setscheduler(pid, -1, param);
4765}
4766
4767/**
4768 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4769 * @pid: the pid in question.
4770 */
4771asmlinkage long sys_sched_getscheduler(pid_t pid)
4772{
4773        struct task_struct *p;
4774        int retval;
4775
4776        if (pid < 0)
4777                return -EINVAL;
4778
4779        retval = -ESRCH;
4780        read_lock(&tasklist_lock);
4781        p = find_process_by_pid(pid);
4782        if (p) {
4783                retval = security_task_getscheduler(p);
4784                if (!retval)
4785                        retval = p->policy;
4786        }
4787        read_unlock(&tasklist_lock);
4788        return retval;
4789}
4790
4791/**
4792 * sys_sched_getscheduler - get the RT priority of a thread
4793 * @pid: the pid in question.
4794 * @param: structure containing the RT priority.
4795 */
4796asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4797{
4798        struct sched_param lp;
4799        struct task_struct *p;
4800        int retval;
4801
4802        if (!param || pid < 0)
4803                return -EINVAL;
4804
4805        read_lock(&tasklist_lock);
4806        p = find_process_by_pid(pid);
4807        retval = -ESRCH;
4808        if (!p)
4809                goto out_unlock;
4810
4811        retval = security_task_getscheduler(p);
4812        if (retval)
4813                goto out_unlock;
4814
4815        lp.sched_priority = p->rt_priority;
4816        read_unlock(&tasklist_lock);
4817
4818        /*
4819         * This one might sleep, we cannot do it with a spinlock held ...
4820         */
4821        retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4822
4823        return retval;
4824
4825out_unlock:
4826        read_unlock(&tasklist_lock);
4827        return retval;
4828}
4829
4830long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4831{
4832        cpumask_t cpus_allowed;
4833        struct task_struct *p;
4834        int retval;
4835
4836        get_online_cpus();
4837        read_lock(&tasklist_lock);
4838
4839        p = find_process_by_pid(pid);
4840        if (!p) {
4841                read_unlock(&tasklist_lock);
4842                put_online_cpus();
4843                return -ESRCH;
4844        }
4845
4846        /*
4847         * It is not safe to call set_cpus_allowed with the
4848         * tasklist_lock held. We will bump the task_struct's
4849         * usage count and then drop tasklist_lock.
4850         */
4851        get_task_struct(p);
4852        read_unlock(&tasklist_lock);
4853
4854        retval = -EPERM;
4855        if ((current->euid != p->euid) && (current->euid != p->uid) &&
4856                        !capable(CAP_SYS_NICE))
4857                goto out_unlock;
4858
4859        retval = security_task_setscheduler(p, 0, NULL);
4860        if (retval)
4861                goto out_unlock;
4862
4863        cpus_allowed = cpuset_cpus_allowed(p);
4864        cpus_and(new_mask, new_mask, cpus_allowed);
4865 again:
4866        retval = set_cpus_allowed(p, new_mask);
4867
4868        if (!retval) {
4869                cpus_allowed = cpuset_cpus_allowed(p);
4870                if (!cpus_subset(new_mask, cpus_allowed)) {
4871                        /*
4872                         * We must have raced with a concurrent cpuset
4873                         * update. Just reset the cpus_allowed to the
4874                         * cpuset's cpus_allowed
4875                         */
4876                        new_mask = cpus_allowed;
4877                        goto again;
4878                }
4879        }
4880out_unlock:
4881        put_task_struct(p);
4882        put_online_cpus();
4883        return retval;
4884}
4885
4886static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4887                             cpumask_t *new_mask)
4888{
4889        if (len < sizeof(cpumask_t)) {
4890                memset(new_mask, 0, sizeof(cpumask_t));
4891        } else if (len > sizeof(cpumask_t)) {
4892                len = sizeof(cpumask_t);
4893        }
4894        return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4895}
4896
4897/**
4898 * sys_sched_setaffinity - set the cpu affinity of a process
4899 * @pid: pid of the process
4900 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4901 * @user_mask_ptr: user-space pointer to the new cpu mask
4902 */
4903asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4904                                      unsigned long __user *user_mask_ptr)
4905{
4906        cpumask_t new_mask;
4907        int retval;
4908
4909        retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4910        if (retval)
4911                return retval;
4912
4913        return sched_setaffinity(pid, new_mask);
4914}
4915
4916/*
4917 * Represents all cpu's present in the system
4918 * In systems capable of hotplug, this map could dynamically grow
4919 * as new cpu's are detected in the system via any platform specific
4920 * method, such as ACPI for e.g.
4921 */
4922
4923cpumask_t cpu_present_map __read_mostly;
4924EXPORT_SYMBOL(cpu_present_map);
4925
4926#ifndef CONFIG_SMP
4927cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4928EXPORT_SYMBOL(cpu_online_map);
4929
4930cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4931EXPORT_SYMBOL(cpu_possible_map);
4932#endif
4933
4934long sched_getaffinity(pid_t pid, cpumask_t *mask)
4935{
4936        struct task_struct *p;
4937        int retval;
4938
4939        get_online_cpus();
4940        read_lock(&tasklist_lock);
4941
4942        retval = -ESRCH;
4943        p = find_process_by_pid(pid);
4944        if (!p)
4945                goto out_unlock;
4946
4947        retval = security_task_getscheduler(p);
4948        if (retval)
4949                goto out_unlock;
4950
4951        cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4952
4953out_unlock:
4954        read_unlock(&tasklist_lock);
4955        put_online_cpus();
4956
4957        return retval;
4958}
4959
4960/**
4961 * sys_sched_getaffinity - get the cpu affinity of a process
4962 * @pid: pid of the process
4963 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4964 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4965 */
4966asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4967                                      unsigned long __user *user_mask_ptr)
4968{
4969        int ret;
4970        cpumask_t mask;
4971
4972        if (len < sizeof(cpumask_t))
4973                return -EINVAL;
4974
4975        ret = sched_getaffinity(pid, &mask);
4976        if (ret < 0)
4977                return ret;
4978
4979        if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4980                return -EFAULT;
4981
4982        return sizeof(cpumask_t);
4983}
4984
4985/**
4986 * sys_sched_yield - yield the current processor to other threads.
4987 *
4988 * This function yields the current CPU to other tasks. If there are no
4989 * other threads running on this CPU then this function will return.
4990 */
4991asmlinkage long sys_sched_yield(void)
4992{
4993        struct rq *rq = this_rq_lock();
4994
4995        schedstat_inc(rq, yld_count);
4996        current->sched_class->yield_task(rq);
4997
4998        /*
4999         * Since we are going to call schedule() anyway, there's
5000         * no need to preempt or enable interrupts:
5001         */
5002        __release(rq->lock);
5003        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5004        _raw_spin_unlock(&rq->lock);
5005        preempt_enable_no_resched();
5006
5007        schedule();
5008
5009        return 0;
5010}
5011
5012static void __cond_resched(void)
5013{
5014#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5015        __might_sleep(__FILE__, __LINE__);
5016#endif
5017        /*
5018         * The BKS might be reacquired before we have dropped
5019         * PREEMPT_ACTIVE, which could trigger a second
5020         * cond_resched() call.
5021         */
5022        do {
5023                add_preempt_count(PREEMPT_ACTIVE);
5024                schedule();
5025                sub_preempt_count(PREEMPT_ACTIVE);
5026        } while (need_resched());
5027}
5028
5029#if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5030int __sched _cond_resched(void)
5031{
5032        if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5033                                        system_state == SYSTEM_RUNNING) {
5034                __cond_resched();
5035                return 1;
5036        }
5037        return 0;
5038}
5039EXPORT_SYMBOL(_cond_resched);
5040#endif
5041
5042/*
5043 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5044 * call schedule, and on return reacquire the lock.
5045 *
5046 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5047 * operations here to prevent schedule() from being called twice (once via
5048 * spin_unlock(), once by hand).
5049 */
5050int cond_resched_lock(spinlock_t *lock)
5051{
5052        int resched = need_resched() && system_state == SYSTEM_RUNNING;
5053        int ret = 0;
5054
5055        if (spin_needbreak(lock) || resched) {
5056                spin_unlock(lock);
5057                if (resched && need_resched())
5058                        __cond_resched();
5059                else
5060                        cpu_relax();
5061                ret = 1;
5062                spin_lock(lock);
5063        }
5064        return ret;
5065}
5066EXPORT_SYMBOL(cond_resched_lock);
5067
5068int __sched cond_resched_softirq(void)
5069{
5070        BUG_ON(!in_softirq());
5071
5072        if (need_resched() && system_state == SYSTEM_RUNNING) {
5073                local_bh_enable();
5074                __cond_resched();
5075                local_bh_disable();
5076                return 1;
5077        }
5078        return 0;
5079}
5080EXPORT_SYMBOL(cond_resched_softirq);
5081
5082/**
5083 * yield - yield the current processor to other threads.
5084 *
5085 * This is a shortcut for kernel-space yielding - it marks the
5086 * thread runnable and calls sys_sched_yield().
5087 */
5088void __sched yield(void)
5089{
5090        set_current_state(TASK_RUNNING);
5091        sys_sched_yield();
5092}
5093EXPORT_SYMBOL(yield);
5094
5095/*
5096 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5097 * that process accounting knows that this is a task in IO wait state.
5098 *
5099 * But don't do that if it is a deliberate, throttling IO wait (this task
5100 * has set its backing_dev_info: the queue against which it should throttle)
5101 */
5102void __sched io_schedule(void)
5103{
5104        struct rq *rq = &__raw_get_cpu_var(runqueues);
5105
5106        delayacct_blkio_start();
5107        atomic_inc(&rq->nr_iowait);
5108        schedule();
5109        atomic_dec(&rq->nr_iowait);
5110        delayacct_blkio_end();
5111}
5112EXPORT_SYMBOL(io_schedule);
5113
5114long __sched io_schedule_timeout(long timeout)
5115{
5116        struct rq *rq = &__raw_get_cpu_var(runqueues);
5117        long ret;
5118
5119        delayacct_blkio_start();
5120        atomic_inc(&rq->nr_iowait);
5121        ret = schedule_timeout(timeout);
5122        atomic_dec(&rq->nr_iowait);
5123        delayacct_blkio_end();
5124        return ret;
5125}
5126
5127/**
5128 * sys_sched_get_priority_max - return maximum RT priority.
5129 * @policy: scheduling class.
5130 *
5131 * this syscall returns the maximum rt_priority that can be used
5132 * by a given scheduling class.
5133 */
5134asmlinkage long sys_sched_get_priority_max(int policy)
5135{
5136        int ret = -EINVAL;
5137
5138        switch (policy) {
5139        case SCHED_FIFO:
5140        case SCHED_RR:
5141                ret = MAX_USER_RT_PRIO-1;
5142                break;
5143        case SCHED_NORMAL:
5144        case SCHED_BATCH:
5145        case SCHED_IDLE:
5146                ret = 0;
5147                break;
5148        }
5149        return ret;
5150}
5151
5152/**
5153 * sys_sched_get_priority_min - return minimum RT priority.
5154 * @policy: scheduling class.
5155 *
5156 * this syscall returns the minimum rt_priority that can be used
5157 * by a given scheduling class.
5158 */
5159asmlinkage long sys_sched_get_priority_min(int policy)
5160{
5161        int ret = -EINVAL;
5162
5163        switch (policy) {
5164        case SCHED_FIFO:
5165        case SCHED_RR:
5166                ret = 1;
5167                break;
5168        case SCHED_NORMAL:
5169        case SCHED_BATCH:
5170        case SCHED_IDLE:
5171                ret = 0;
5172        }
5173        return ret;
5174}
5175
5176/**
5177 * sys_sched_rr_get_interval - return the default timeslice of a process.
5178 * @pid: pid of the process.
5179 * @interval: userspace pointer to the timeslice value.
5180 *
5181 * this syscall writes the default timeslice value of a given process
5182 * into the user-space timespec buffer. A value of '0' means infinity.
5183 */
5184asmlinkage
5185long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5186{
5187        struct task_struct *p;
5188        unsigned int time_slice;
5189        int retval;
5190        struct timespec t;
5191
5192        if (pid < 0)
5193                return -EINVAL;
5194
5195        retval = -ESRCH;
5196        read_lock(&tasklist_lock);
5197        p = find_process_by_pid(pid);
5198        if (!p)
5199                goto out_unlock;
5200
5201        retval = security_task_getscheduler(p);
5202        if (retval)
5203                goto out_unlock;
5204
5205        /*
5206         * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5207         * tasks that are on an otherwise idle runqueue:
5208         */
5209        time_slice = 0;
5210        if (p->policy == SCHED_RR) {
5211                time_slice = DEF_TIMESLICE;
5212        } else if (p->policy != SCHED_FIFO) {
5213                struct sched_entity *se = &p->se;
5214                unsigned long flags;
5215                struct rq *rq;
5216
5217                rq = task_rq_lock(p, &flags);
5218                if (rq->cfs.load.weight)
5219                        time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5220                task_rq_unlock(rq, &flags);
5221        }
5222        read_unlock(&tasklist_lock);
5223        jiffies_to_timespec(time_slice, &t);
5224        retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5225        return retval;
5226
5227out_unlock:
5228        read_unlock(&tasklist_lock);
5229        return retval;
5230}
5231
5232static const char stat_nam[] = "RSDTtZX";
5233
5234void sched_show_task(struct task_struct *p)
5235{
5236        unsigned long free = 0;
5237        unsigned state;
5238
5239        state = p->state ? __ffs(p->state) + 1 : 0;
5240        printk(KERN_INFO "%-13.13s %c", p->comm,
5241                state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5242#if BITS_PER_LONG == 32
5243        if (state == TASK_RUNNING)
5244                printk(KERN_CONT " running  ");
5245        else
5246                printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5247#else
5248        if (state == TASK_RUNNING)
5249                printk(KERN_CONT "  running task    ");
5250        else
5251                printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5252#endif
5253#ifdef CONFIG_DEBUG_STACK_USAGE
5254        {
5255                unsigned long *n = end_of_stack(p);
5256                while (!*n)
5257                        n++;
5258                free = (unsigned long)n - (unsigned long)end_of_stack(p);
5259        }
5260#endif
5261        printk(KERN_CONT "%5lu %5d %6d\n", free,
5262                task_pid_nr(p), task_pid_nr(p->real_parent));
5263
5264        show_stack(p, NULL);
5265}
5266
5267void show_state_filter(unsigned long state_filter)
5268{
5269        struct task_struct *g, *p;
5270
5271#if BITS_PER_LONG == 32
5272        printk(KERN_INFO
5273                "  task                PC stack   pid father\n");
5274#else
5275        printk(KERN_INFO
5276                "  task                        PC stack   pid father\n");
5277#endif
5278        read_lock(&tasklist_lock);
5279        do_each_thread(g, p) {
5280                /*
5281                 * reset the NMI-timeout, listing all files on a slow
5282                 * console might take alot of time:
5283                 */
5284                touch_nmi_watchdog();
5285                if (!state_filter || (p->state & state_filter))
5286                        sched_show_task(p);
5287        } while_each_thread(g, p);
5288
5289        touch_all_softlockup_watchdogs();
5290
5291#ifdef CONFIG_SCHED_DEBUG
5292        sysrq_sched_debug_show();
5293#endif
5294        read_unlock(&tasklist_lock);
5295        /*
5296         * Only show locks if all tasks are dumped:
5297         */
5298        if (state_filter == -1)
5299                debug_show_all_locks();
5300}
5301
5302void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5303{
5304        idle->sched_class = &idle_sched_class;
5305}
5306
5307/**
5308 * init_idle - set up an idle thread for a given CPU
5309 * @idle: task in question
5310 * @cpu: cpu the idle task belongs to
5311 *
5312 * NOTE: this function does not set the idle thread's NEED_RESCHED
5313 * flag, to make booting more robust.
5314 */
5315void __cpuinit init_idle(struct task_struct *idle, int cpu)
5316{
5317        struct rq *rq = cpu_rq(cpu);
5318        unsigned long flags;
5319
5320        __sched_fork(idle);
5321        idle->se.exec_start = sched_clock();
5322
5323        idle->prio = idle->normal_prio = MAX_PRIO;
5324        idle->cpus_allowed = cpumask_of_cpu(cpu);
5325        __set_task_cpu(idle, cpu);
5326
5327        spin_lock_irqsave(&rq->lock, flags);
5328        rq->curr = rq->idle = idle;
5329#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5330        idle->oncpu = 1;
5331#endif
5332        spin_unlock_irqrestore(&rq->lock, flags);
5333
5334        /* Set the preempt count _outside_ the spinlocks! */
5335        task_thread_info(idle)->preempt_count = 0;
5336
5337        /*
5338         * The idle tasks have their own, simple scheduling class:
5339         */
5340        idle->sched_class = &idle_sched_class;
5341}
5342
5343/*
5344 * In a system that switches off the HZ timer nohz_cpu_mask
5345 * indicates which cpus entered this state. This is used
5346 * in the rcu update to wait only for active cpus. For system
5347 * which do not switch off the HZ timer nohz_cpu_mask should
5348 * always be CPU_MASK_NONE.
5349 */
5350cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5351
5352/*
5353 * Increase the granularity value when there are more CPUs,
5354 * because with more CPUs the 'effective latency' as visible
5355 * to users decreases. But the relationship is not linear,
5356 * so pick a second-best guess by going with the log2 of the
5357 * number of CPUs.
5358 *
5359 * This idea comes from the SD scheduler of Con Kolivas:
5360 */
5361static inline void sched_init_granularity(void)
5362{
5363        unsigned int factor = 1 + ilog2(num_online_cpus());
5364        const unsigned long limit = 200000000;
5365
5366        sysctl_sched_min_granularity *= factor;
5367        if (sysctl_sched_min_granularity > limit)
5368                sysctl_sched_min_granularity = limit;
5369
5370        sysctl_sched_latency *= factor;
5371        if (sysctl_sched_latency > limit)
5372                sysctl_sched_latency = limit;
5373
5374        sysctl_sched_wakeup_granularity *= factor;
5375        sysctl_sched_batch_wakeup_granularity *= factor;
5376}
5377
5378#ifdef CONFIG_SMP
5379/*
5380 * This is how migration works:
5381 *
5382 * 1) we queue a struct migration_req structure in the source CPU's
5383 *    runqueue and wake up that CPU's migration thread.
5384 * 2) we down() the locked semaphore => thread blocks.
5385 * 3) migration thread wakes up (implicitly it forces the migrated
5386 *    thread off the CPU)
5387 * 4) it gets the migration request and checks whether the migrated
5388 *    task is still in the wrong runqueue.
5389 * 5) if it's in the wrong runqueue then the migration thread removes
5390 *    it and puts it into the right queue.
5391 * 6) migration thread up()s the semaphore.
5392 * 7) we wake up and the migration is done.
5393 */
5394
5395/*
5396 * Change a given task's CPU affinity. Migrate the thread to a
5397 * proper CPU and schedule it away if the CPU it's executing on
5398 * is removed from the allowed bitmask.
5399 *
5400 * NOTE: the caller must have a valid reference to the task, the
5401 * task must not exit() & deallocate itself prematurely. The
5402 * call is not atomic; no spinlocks may be held.
5403 */
5404int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
5405{
5406        struct migration_req req;
5407        unsigned long flags;
5408        struct rq *rq;
5409        int ret = 0;
5410
5411        rq = task_rq_lock(p, &flags);
5412        if (!cpus_intersects(new_mask, cpu_online_map)) {
5413                ret = -EINVAL;
5414                goto out;
5415        }
5416
5417        if (p->sched_class->set_cpus_allowed)
5418                p->sched_class->set_cpus_allowed(p, &new_mask);
5419        else {
5420                p->cpus_allowed = new_mask;
5421                p->rt.nr_cpus_allowed = cpus_weight(new_mask);
5422        }
5423
5424        /* Can the task run on the task's current CPU? If so, we're done */
5425        if (cpu_isset(task_cpu(p), new_mask))
5426                goto out;
5427
5428        if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5429                /* Need help from migration thread: drop lock and wait. */
5430                task_rq_unlock(rq, &flags);
5431                wake_up_process(rq->migration_thread);
5432                wait_for_completion(&req.done);
5433                tlb_migrate_finish(p->mm);
5434                return 0;
5435        }
5436out:
5437        task_rq_unlock(rq, &flags);
5438
5439        return ret;
5440}
5441EXPORT_SYMBOL_GPL(set_cpus_allowed);
5442
5443/*
5444 * Move (not current) task off this cpu, onto dest cpu. We're doing
5445 * this because either it can't run here any more (set_cpus_allowed()
5446 * away from this CPU, or CPU going down), or because we're
5447 * attempting to rebalance this task on exec (sched_exec).
5448 *
5449 * So we race with normal scheduler movements, but that's OK, as long
5450 * as the task is no longer on this CPU.
5451 *
5452 * Returns non-zero if task was successfully migrated.
5453 */
5454static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5455{
5456        struct rq *rq_dest, *rq_src;
5457        int ret = 0, on_rq;
5458
5459        if (unlikely(cpu_is_offline(dest_cpu)))
5460                return ret;
5461
5462        rq_src = cpu_rq(src_cpu);
5463        rq_dest = cpu_rq(dest_cpu);
5464
5465        double_rq_lock(rq_src, rq_dest);
5466        /* Already moved. */
5467        if (task_cpu(p) != src_cpu)
5468                goto out;
5469        /* Affinity changed (again). */
5470        if (!cpu_isset(dest_cpu, p->cpus_allowed))
5471                goto out;
5472
5473        on_rq = p->se.on_rq;
5474        if (on_rq)
5475                deactivate_task(rq_src, p, 0);
5476
5477        set_task_cpu(p, dest_cpu);
5478        if (on_rq) {
5479                activate_task(rq_dest, p, 0);
5480                check_preempt_curr(rq_dest, p);
5481        }
5482        ret = 1;
5483out:
5484        double_rq_unlock(rq_src, rq_dest);
5485        return ret;
5486}
5487
5488/*
5489 * migration_thread - this is a highprio system thread that performs
5490 * thread migration by bumping thread off CPU then 'pushing' onto
5491 * another runqueue.
5492 */
5493static int migration_thread(void *data)
5494{
5495        int cpu = (long)data;
5496        struct rq *rq;
5497
5498        rq = cpu_rq(cpu);
5499        BUG_ON(rq->migration_thread != current);
5500
5501        set_current_state(TASK_INTERRUPTIBLE);
5502        while (!kthread_should_stop()) {
5503                struct migration_req *req;
5504                struct list_head *head;
5505
5506                spin_lock_irq(&rq->lock);
5507
5508                if (cpu_is_offline(cpu)) {
5509                        spin_unlock_irq(&rq->lock);
5510                        goto wait_to_die;
5511                }
5512
5513                if (rq->active_balance) {
5514                        active_load_balance(rq, cpu);
5515                        rq->active_balance = 0;
5516                }
5517
5518                head = &rq->migration_queue;
5519
5520                if (list_empty(head)) {
5521                        spin_unlock_irq(&rq->lock);
5522                        schedule();
5523                        set_current_state(TASK_INTERRUPTIBLE);
5524                        continue;
5525                }
5526                req = list_entry(head->next, struct migration_req, list);
5527                list_del_init(head->next);
5528
5529                spin_unlock(&rq->lock);
5530                __migrate_task(req->task, cpu, req->dest_cpu);
5531                local_irq_enable();
5532
5533                complete(&req->done);
5534        }
5535        __set_current_state(TASK_RUNNING);
5536        return 0;
5537
5538wait_to_die:
5539        /* Wait for kthread_stop */
5540        set_current_state(TASK_INTERRUPTIBLE);
5541        while (!kthread_should_stop()) {
5542                schedule();
5543                set_current_state(TASK_INTERRUPTIBLE);
5544        }
5545        __set_current_state(TASK_RUNNING);
5546        return 0;
5547}
5548
5549#ifdef CONFIG_HOTPLUG_CPU
5550
5551static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5552{
5553        int ret;
5554
5555        local_irq_disable();
5556        ret = __migrate_task(p, src_cpu, dest_cpu);
5557        local_irq_enable();
5558        return ret;
5559}
5560
5561/*
5562 * Figure out where task on dead CPU should go, use force if necessary.
5563 * NOTE: interrupts should be disabled by the caller
5564 */
5565static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5566{
5567        unsigned long flags;
5568        cpumask_t mask;
5569        struct rq *rq;
5570        int dest_cpu;
5571
5572        do {
5573                /* On same node? */
5574                mask = node_to_cpumask(cpu_to_node(dead_cpu));
5575                cpus_and(mask, mask, p->cpus_allowed);
5576                dest_cpu = any_online_cpu(mask);
5577
5578                /* On any allowed CPU? */
5579                if (dest_cpu == NR_CPUS)
5580                        dest_cpu = any_online_cpu(p->cpus_allowed);
5581
5582                /* No more Mr. Nice Guy. */
5583                if (dest_cpu == NR_CPUS) {
5584                        cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5585                        /*
5586                         * Try to stay on the same cpuset, where the
5587                         * current cpuset may be a subset of all cpus.
5588                         * The cpuset_cpus_allowed_locked() variant of
5589                         * cpuset_cpus_allowed() will not block. It must be
5590                         * called within calls to cpuset_lock/cpuset_unlock.
5591                         */
5592                        rq = task_rq_lock(p, &flags);
5593                        p->cpus_allowed = cpus_allowed;
5594                        dest_cpu = any_online_cpu(p->cpus_allowed);
5595                        task_rq_unlock(rq, &flags);
5596
5597                        /*
5598                         * Don't tell them about moving exiting tasks or
5599                         * kernel threads (both mm NULL), since they never
5600                         * leave kernel.
5601                         */
5602                        if (p->mm && printk_ratelimit()) {
5603                                printk(KERN_INFO "process %d (%s) no "
5604                                       "longer affine to cpu%d\n",
5605                                        task_pid_nr(p), p->comm, dead_cpu);
5606                        }
5607                }
5608        } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5609}
5610
5611/*
5612 * While a dead CPU has no uninterruptible tasks queued at this point,
5613 * it might still have a nonzero ->nr_uninterruptible counter, because
5614 * for performance reasons the counter is not stricly tracking tasks to
5615 * their home CPUs. So we just add the counter to another CPU's counter,
5616 * to keep the global sum constant after CPU-down:
5617 */
5618static void migrate_nr_uninterruptible(struct rq *rq_src)
5619{
5620        struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5621        unsigned long flags;
5622
5623        local_irq_save(flags);
5624        double_rq_lock(rq_src, rq_dest);
5625        rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5626        rq_src->nr_uninterruptible = 0;
5627        double_rq_unlock(rq_src, rq_dest);
5628        local_irq_restore(flags);
5629}
5630
5631/* Run through task list and migrate tasks from the dead cpu. */
5632static void migrate_live_tasks(int src_cpu)
5633{
5634        struct task_struct *p, *t;
5635
5636        read_lock(&tasklist_lock);
5637
5638        do_each_thread(t, p) {
5639                if (p == current)
5640                        continue;
5641
5642                if (task_cpu(p) == src_cpu)
5643                        move_task_off_dead_cpu(src_cpu, p);
5644        } while_each_thread(t, p);
5645
5646        read_unlock(&tasklist_lock);
5647}
5648
5649/*
5650 * Schedules idle task to be the next runnable task on current CPU.
5651 * It does so by boosting its priority to highest possible.
5652 * Used by CPU offline code.
5653 */
5654void sched_idle_next(void)
5655{
5656        int this_cpu = smp_processor_id();
5657        struct rq *rq = cpu_rq(this_cpu);
5658        struct task_struct *p = rq->idle;
5659        unsigned long flags;
5660
5661        /* cpu has to be offline */
5662        BUG_ON(cpu_online(this_cpu));
5663
5664        /*
5665         * Strictly not necessary since rest of the CPUs are stopped by now
5666         * and interrupts disabled on the current cpu.
5667         */
5668        spin_lock_irqsave(&rq->lock, flags);
5669
5670        __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5671
5672        update_rq_clock(rq);
5673        activate_task(rq, p, 0);
5674
5675        spin_unlock_irqrestore(&rq->lock, flags);
5676}
5677
5678/*
5679 * Ensures that the idle task is using init_mm right before its cpu goes
5680 * offline.
5681 */
5682void idle_task_exit(void)
5683{
5684        struct mm_struct *mm = current->active_mm;
5685
5686        BUG_ON(cpu_online(smp_processor_id()));
5687
5688        if (mm != &init_mm)
5689                switch_mm(mm, &init_mm, current);
5690        mmdrop(mm);
5691}
5692
5693/* called under rq->lock with disabled interrupts */
5694static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5695{
5696        struct rq *rq = cpu_rq(dead_cpu);
5697
5698        /* Must be exiting, otherwise would be on tasklist. */
5699        BUG_ON(!p->exit_state);
5700
5701        /* Cannot have done final schedule yet: would have vanished. */
5702        BUG_ON(p->state == TASK_DEAD);
5703
5704        get_task_struct(p);
5705
5706        /*
5707         * Drop lock around migration; if someone else moves it,
5708         * that's OK. No task can be added to this CPU, so iteration is
5709         * fine.
5710         */
5711        spin_unlock_irq(&rq->lock);
5712        move_task_off_dead_cpu(dead_cpu, p);
5713        spin_lock_irq(&rq->lock);
5714
5715        put_task_struct(p);
5716}
5717
5718/* release_task() removes task from tasklist, so we won't find dead tasks. */
5719static void migrate_dead_tasks(unsigned int dead_cpu)
5720{
5721        struct rq *rq = cpu_rq(dead_cpu);
5722        struct task_struct *next;
5723
5724        for ( ; ; ) {
5725                if (!rq->nr_running)
5726                        break;
5727                update_rq_clock(rq);
5728                next = pick_next_task(rq, rq->curr);
5729                if (!next)
5730                        break;
5731                migrate_dead(dead_cpu, next);
5732
5733        }
5734}
5735#endif /* CONFIG_HOTPLUG_CPU */
5736
5737#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5738
5739static struct ctl_table sd_ctl_dir[] = {
5740        {
5741                .procname       = "sched_domain",
5742                .mode           = 0555,
5743        },
5744        {0, },
5745};
5746
5747static struct ctl_table sd_ctl_root[] = {
5748        {
5749                .ctl_name       = CTL_KERN,
5750                .procname       = "kernel",
5751                .mode           = 0555,
5752                .child          = sd_ctl_dir,
5753        },
5754        {0, },
5755};
5756
5757static struct ctl_table *sd_alloc_ctl_entry(int n)
5758{
5759        struct ctl_table *entry =
5760                kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5761
5762        return entry;
5763}
5764
5765static void sd_free_ctl_entry(struct ctl_table **tablep)
5766{
5767        struct ctl_table *entry;
5768
5769        /*
5770         * In the intermediate directories, both the child directory and
5771         * procname are dynamically allocated and could fail but the mode
5772         * will always be set. In the lowest directory the names are
5773         * static strings and all have proc handlers.
5774         */
5775        for (entry = *tablep; entry->mode; entry++) {
5776                if (entry->child)
5777                        sd_free_ctl_entry(&entry->child);
5778                if (entry->proc_handler == NULL)
5779                        kfree(entry->procname);
5780        }
5781
5782        kfree(*tablep);
5783        *tablep = NULL;
5784}
5785
5786static void
5787set_table_entry(struct ctl_table *entry,
5788                const char *procname, void *data, int maxlen,
5789                mode_t mode, proc_handler *proc_handler)
5790{
5791        entry->procname = procname;
5792        entry->data = data;
5793        entry->maxlen = maxlen;
5794        entry->mode = mode;
5795        entry->proc_handler = proc_handler;
5796}
5797
5798static struct ctl_table *
5799sd_alloc_ctl_domain_table(struct sched_domain *sd)
5800{
5801        struct ctl_table *table = sd_alloc_ctl_entry(12);
5802
5803        if (table == NULL)
5804                return NULL;
5805
5806        set_table_entry(&table[0], "min_interval", &sd->min_interval,
5807                sizeof(long), 0644, proc_doulongvec_minmax);
5808        set_table_entry(&table[1], "max_interval", &sd->max_interval,
5809                sizeof(long), 0644, proc_doulongvec_minmax);
5810        set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5811                sizeof(int), 0644, proc_dointvec_minmax);
5812        set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5813                sizeof(int), 0644, proc_dointvec_minmax);
5814        set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5815                sizeof(int), 0644, proc_dointvec_minmax);
5816        set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5817                sizeof(int), 0644, proc_dointvec_minmax);
5818        set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5819                sizeof(int), 0644, proc_dointvec_minmax);
5820        set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5821                sizeof(int), 0644, proc_dointvec_minmax);
5822        set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5823                sizeof(int), 0644, proc_dointvec_minmax);
5824        set_table_entry(&table[9], "cache_nice_tries",
5825                &sd->cache_nice_tries,
5826                sizeof(int), 0644, proc_dointvec_minmax);
5827        set_table_entry(&table[10], "flags", &sd->flags,
5828                sizeof(int), 0644, proc_dointvec_minmax);
5829        /* &table[11] is terminator */
5830
5831        return table;
5832}
5833
5834static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5835{
5836        struct ctl_table *entry, *table;
5837        struct sched_domain *sd;
5838        int domain_num = 0, i;
5839        char buf[32];
5840
5841        for_each_domain(cpu, sd)
5842                domain_num++;
5843        entry = table = sd_alloc_ctl_entry(domain_num + 1);
5844        if (table == NULL)
5845                return NULL;
5846
5847        i = 0;
5848        for_each_domain(cpu, sd) {
5849                snprintf(buf, 32, "domain%d", i);
5850                entry->procname = kstrdup(buf, GFP_KERNEL);
5851                entry->mode = 0555;
5852                entry->child = sd_alloc_ctl_domain_table(sd);
5853                entry++;
5854                i++;
5855        }
5856        return table;
5857}
5858
5859static struct ctl_table_header *sd_sysctl_header;
5860static void register_sched_domain_sysctl(void)
5861{
5862        int i, cpu_num = num_online_cpus();
5863        struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5864        char buf[32];
5865
5866        WARN_ON(sd_ctl_dir[0].child);
5867        sd_ctl_dir[0].child = entry;
5868
5869        if (entry == NULL)
5870                return;
5871
5872        for_each_online_cpu(i) {
5873                snprintf(buf, 32, "cpu%d", i);
5874                entry->procname = kstrdup(buf, GFP_KERNEL);
5875                entry->mode = 0555;
5876                entry->child = sd_alloc_ctl_cpu_table(i);
5877                entry++;
5878        }
5879
5880        WARN_ON(sd_sysctl_header);
5881        sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5882}
5883
5884/* may be called multiple times per register */
5885static void unregister_sched_domain_sysctl(void)
5886{
5887        if (sd_sysctl_header)
5888                unregister_sysctl_table(sd_sysctl_header);
5889        sd_sysctl_header = NULL;
5890        if (sd_ctl_dir[0].child)
5891                sd_free_ctl_entry(&sd_ctl_dir[0].child);
5892}
5893#else
5894static void register_sched_domain_sysctl(void)
5895{
5896}
5897static void unregister_sched_domain_sysctl(void)
5898{
5899}
5900#endif
5901
5902/*
5903 * migration_call - callback that gets triggered when a CPU is added.
5904 * Here we can start up the necessary migration thread for the new CPU.
5905 */
5906static int __cpuinit
5907migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5908{
5909        struct task_struct *p;
5910        int cpu = (long)hcpu;
5911        unsigned long flags;
5912        struct rq *rq;
5913
5914        switch (action) {
5915
5916        case CPU_UP_PREPARE:
5917        case CPU_UP_PREPARE_FROZEN:
5918                p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5919                if (IS_ERR(p))
5920                        return NOTIFY_BAD;
5921                kthread_bind(p, cpu);
5922                /* Must be high prio: stop_machine expects to yield to it. */
5923                rq = task_rq_lock(p, &flags);
5924                __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5925                task_rq_unlock(rq, &flags);
5926                cpu_rq(cpu)->migration_thread = p;
5927                break;
5928
5929        case CPU_ONLINE:
5930        case CPU_ONLINE_FROZEN:
5931                /* Strictly unnecessary, as first user will wake it. */
5932                wake_up_process(cpu_rq(cpu)->migration_thread);
5933
5934                /* Update our root-domain */
5935                rq = cpu_rq(cpu);
5936                spin_lock_irqsave(&rq->lock, flags);
5937                if (rq->rd) {
5938                        BUG_ON(!cpu_isset(cpu, rq->rd->span));
5939                        cpu_set(cpu, rq->rd->online);
5940                }
5941                spin_unlock_irqrestore(&rq->lock, flags);
5942                break;
5943
5944#ifdef CONFIG_HOTPLUG_CPU
5945        case CPU_UP_CANCELED:
5946        case CPU_UP_CANCELED_FROZEN:
5947                if (!cpu_rq(cpu)->migration_thread)
5948                        break;
5949                /* Unbind it from offline cpu so it can run. Fall thru. */
5950                kthread_bind(cpu_rq(cpu)->migration_thread,
5951                             any_online_cpu(cpu_online_map));
5952                kthread_stop(cpu_rq(cpu)->migration_thread);
5953                cpu_rq(cpu)->migration_thread = NULL;
5954                break;
5955
5956        case CPU_DEAD:
5957        case CPU_DEAD_FROZEN:
5958                cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5959                migrate_live_tasks(cpu);
5960                rq = cpu_rq(cpu);
5961                kthread_stop(rq->migration_thread);
5962                rq->migration_thread = NULL;
5963                /* Idle task back to normal (off runqueue, low prio) */
5964                spin_lock_irq(&rq->lock);
5965                update_rq_clock(rq);
5966                deactivate_task(rq, rq->idle, 0);
5967                rq->idle->static_prio = MAX_PRIO;
5968                __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5969                rq->idle->sched_class = &idle_sched_class;
5970                migrate_dead_tasks(cpu);
5971                spin_unlock_irq(&rq->lock);
5972                cpuset_unlock();
5973                migrate_nr_uninterruptible(rq);
5974                BUG_ON(rq->nr_running != 0);
5975
5976                /*
5977                 * No need to migrate the tasks: it was best-effort if
5978                 * they didn't take sched_hotcpu_mutex. Just wake up
5979                 * the requestors.
5980                 */
5981                spin_lock_irq(&rq->lock);
5982                while (!list_empty(&rq->migration_queue)) {
5983                        struct migration_req *req;
5984
5985                        req = list_entry(rq->migration_queue.next,
5986                                         struct migration_req, list);
5987                        list_del_init(&req->list);
5988                        complete(&req->done);
5989                }
5990                spin_unlock_irq(&rq->lock);
5991                break;
5992
5993        case CPU_DYING:
5994        case CPU_DYING_FROZEN:
5995                /* Update our root-domain */
5996                rq = cpu_rq(cpu);
5997                spin_lock_irqsave(&rq->lock, flags);
5998                if (rq->rd) {
5999                        BUG_ON(!cpu_isset(cpu, rq->rd->span));
6000                        cpu_clear(cpu, rq->rd->online);
6001                }
6002                spin_unlock_irqrestore(&rq->lock, flags);
6003                break;
6004#endif
6005        }
6006        return NOTIFY_OK;
6007}
6008
6009/* Register at highest priority so that task migration (migrate_all_tasks)
6010 * happens before everything else.
6011 */
6012static struct notifier_block __cpuinitdata migration_notifier = {
6013        .notifier_call = migration_call,
6014        .priority = 10
6015};
6016
6017void __init migration_init(void)
6018{
6019        void *cpu = (void *)(long)smp_processor_id();
6020        int err;
6021
6022        /* Start one for the boot CPU: */
6023        err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6024        BUG_ON(err == NOTIFY_BAD);
6025        migration_call(&migration_notifier, CPU_ONLINE, cpu);
6026        register_cpu_notifier(&migration_notifier);
6027}
6028#endif
6029
6030#ifdef CONFIG_SMP
6031
6032/* Number of possible processor ids */
6033int nr_cpu_ids __read_mostly = NR_CPUS;
6034EXPORT_SYMBOL(nr_cpu_ids);
6035
6036#ifdef CONFIG_SCHED_DEBUG
6037
6038static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
6039{
6040        struct sched_group *group = sd->groups;
6041        cpumask_t groupmask;
6042        char str[NR_CPUS];
6043
6044        cpumask_scnprintf(str, NR_CPUS, sd->span);
6045        cpus_clear(groupmask);
6046
6047        printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6048
6049        if (!(sd->flags & SD_LOAD_BALANCE)) {
6050                printk("does not load-balance\n");
6051                if (sd->parent)
6052                        printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6053                                        " has parent");
6054                return -1;
6055        }
6056
6057        printk(KERN_CONT "span %s\n", str);
6058
6059        if (!cpu_isset(cpu, sd->span)) {
6060                printk(KERN_ERR "ERROR: domain->span does not contain "
6061                                "CPU%d\n", cpu);
6062        }
6063        if (!cpu_isset(cpu, group->cpumask)) {
6064                printk(KERN_ERR "ERROR: domain->groups does not contain"
6065                                " CPU%d\n", cpu);
6066        }
6067
6068        printk(KERN_DEBUG "%*s groups:", level + 1, "");
6069        do {
6070                if (!group) {
6071                        printk("\n");
6072                        printk(KERN_ERR "ERROR: group is NULL\n");
6073                        break;
6074                }
6075
6076                if (!group->__cpu_power) {
6077                        printk(KERN_CONT "\n");
6078                        printk(KERN_ERR "ERROR: domain->cpu_power not "
6079                                        "set\n");
6080                        break;
6081                }
6082
6083                if (!cpus_weight(group->cpumask)) {
6084                        printk(KERN_CONT "\n");
6085                        printk(KERN_ERR "ERROR: empty group\n");
6086                        break;
6087                }
6088
6089                if (cpus_intersects(groupmask, group->cpumask)) {
6090                        printk(KERN_CONT "\n");
6091                        printk(KERN_ERR "ERROR: repeated CPUs\n");
6092                        break;
6093                }
6094
6095                cpus_or(groupmask, groupmask, group->cpumask);
6096
6097                cpumask_scnprintf(str, NR_CPUS, group->cpumask);
6098                printk(KERN_CONT " %s", str);
6099
6100                group = group->next;
6101        } while (group != sd->groups);
6102        printk(KERN_CONT "\n");
6103
6104        if (!cpus_equal(sd->span, groupmask))
6105                printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6106
6107        if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
6108                printk(KERN_ERR "ERROR: parent span is not a superset "
6109                        "of domain->span\n");
6110        return 0;
6111}
6112
6113static void sched_domain_debug(struct sched_domain *sd, int cpu)
6114{
6115        int level = 0;
6116
6117        if (!sd) {
6118                printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6119                return;
6120        }
6121
6122        printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6123
6124        for (;;) {
6125                if (sched_domain_debug_one(sd, cpu, level))
6126                        break;
6127                level++;
6128                sd = sd->parent;
6129                if (!sd)
6130                        break;
6131        }
6132}
6133#else
6134# define sched_domain_debug(sd, cpu) do { } while (0)
6135#endif
6136
6137static int sd_degenerate(struct sched_domain *sd)
6138{
6139        if (cpus_weight(sd->span) == 1)
6140                return 1;
6141
6142        /* Following flags need at least 2 groups */
6143        if (sd->flags & (SD_LOAD_BALANCE |
6144                         SD_BALANCE_NEWIDLE |
6145                         SD_BALANCE_FORK |
6146                         SD_BALANCE_EXEC |
6147                         SD_SHARE_CPUPOWER |
6148                         SD_SHARE_PKG_RESOURCES)) {
6149                if (sd->groups != sd->groups->next)
6150                        return 0;
6151        }
6152
6153        /* Following flags don't use groups */
6154        if (sd->flags & (SD_WAKE_IDLE |
6155                         SD_WAKE_AFFINE |
6156                         SD_WAKE_BALANCE))
6157                return 0;
6158
6159        return 1;
6160}
6161
6162static int
6163sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6164{
6165        unsigned long cflags = sd->flags, pflags = parent->flags;
6166
6167        if (sd_degenerate(parent))
6168                return 1;
6169
6170        if (!cpus_equal(sd->span, parent->span))
6171                return 0;
6172
6173        /* Does parent contain flags not in child? */
6174        /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6175        if (cflags & SD_WAKE_AFFINE)
6176                pflags &= ~SD_WAKE_BALANCE;
6177        /* Flags needing groups don't count if only 1 group in parent */
6178        if (parent->groups == parent->groups->next) {
6179                pflags &= ~(SD_LOAD_BALANCE |
6180                                SD_BALANCE_NEWIDLE |
6181                                SD_BALANCE_FORK |
6182                                SD_BALANCE_EXEC |
6183                                SD_SHARE_CPUPOWER |
6184                                SD_SHARE_PKG_RESOURCES);
6185        }
6186        if (~cflags & pflags)
6187                return 0;
6188
6189        return 1;
6190}
6191
6192static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6193{
6194        unsigned long flags;
6195        const struct sched_class *class;
6196
6197        spin_lock_irqsave(&rq->lock, flags);
6198
6199        if (rq->rd) {
6200                struct root_domain *old_rd = rq->rd;
6201
6202                for (class = sched_class_highest; class; class = class->next) {
6203                        if (class->leave_domain)
6204                                class->leave_domain(rq);
6205                }
6206
6207                cpu_clear(rq->cpu, old_rd->span);
6208                cpu_clear(rq->cpu, old_rd->online);
6209
6210                if (atomic_dec_and_test(&old_rd->refcount))
6211                        kfree(old_rd);
6212        }
6213
6214        atomic_inc(&rd->refcount);
6215        rq->rd = rd;
6216
6217        cpu_set(rq->cpu, rd->span);
6218        if (cpu_isset(rq->cpu, cpu_online_map))
6219                cpu_set(rq->cpu, rd->online);
6220
6221        for (class = sched_class_highest; class; class = class->next) {
6222                if (class->join_domain)
6223                        class->join_domain(rq);
6224        }
6225
6226        spin_unlock_irqrestore(&rq->lock, flags);
6227}
6228
6229static void init_rootdomain(struct root_domain *rd)
6230{
6231        memset(rd, 0, sizeof(*rd));
6232
6233        cpus_clear(rd->span);
6234        cpus_clear(rd->online);
6235}
6236
6237static void init_defrootdomain(void)
6238{
6239        init_rootdomain(&def_root_domain);
6240        atomic_set(&def_root_domain.refcount, 1);
6241}
6242
6243static struct root_domain *alloc_rootdomain(void)
6244{
6245        struct root_domain *rd;
6246
6247        rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6248        if (!rd)
6249                return NULL;
6250
6251        init_rootdomain(rd);
6252
6253        return rd;
6254}
6255
6256/*
6257 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6258 * hold the hotplug lock.
6259 */
6260static void
6261cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6262{
6263        struct rq *rq = cpu_rq(cpu);
6264        struct sched_domain *tmp;
6265
6266        /* Remove the sched domains which do not contribute to scheduling. */
6267        for (tmp = sd; tmp; tmp = tmp->parent) {
6268                struct sched_domain *parent = tmp->parent;
6269                if (!parent)
6270                        break;
6271                if (sd_parent_degenerate(tmp, parent)) {
6272                        tmp->parent = parent->parent;
6273                        if (parent->parent)
6274                                parent->parent->child = tmp;
6275                }
6276        }
6277
6278        if (sd && sd_degenerate(sd)) {
6279                sd = sd->parent;
6280                if (sd)
6281                        sd->child = NULL;
6282        }
6283
6284        sched_domain_debug(sd, cpu);
6285
6286        rq_attach_root(rq, rd);
6287        rcu_assign_pointer(rq->sd, sd);
6288}
6289
6290/* cpus with isolated domains */
6291static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6292
6293/* Setup the mask of cpus configured for isolated domains */
6294static int __init isolated_cpu_setup(char *str)
6295{
6296        int ints[NR_CPUS], i;
6297
6298        str = get_options(str, ARRAY_SIZE(ints), ints);
6299        cpus_clear(cpu_isolated_map);
6300        for (i = 1; i <= ints[0]; i++)
6301                if (ints[i] < NR_CPUS)
6302                        cpu_set(ints[i], cpu_isolated_map);
6303        return 1;
6304}
6305
6306__setup("isolcpus=", isolated_cpu_setup);
6307
6308/*
6309 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6310 * to a function which identifies what group(along with sched group) a CPU
6311 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6312 * (due to the fact that we keep track of groups covered with a cpumask_t).
6313 *
6314 * init_sched_build_groups will build a circular linked list of the groups
6315 * covered by the given span, and will set each group's ->cpumask correctly,
6316 * and ->cpu_power to 0.
6317 */
6318static void
6319init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
6320                        int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6321                                        struct sched_group **sg))
6322{
6323        struct sched_group *first = NULL, *last = NULL;
6324        cpumask_t covered = CPU_MASK_NONE;
6325        int i;
6326
6327        for_each_cpu_mask(i, span) {
6328                struct sched_group *sg;
6329                int group = group_fn(i, cpu_map, &sg);
6330                int j;
6331
6332                if (cpu_isset(i, covered))
6333                        continue;
6334
6335                sg->cpumask = CPU_MASK_NONE;
6336                sg->__cpu_power = 0;
6337
6338                for_each_cpu_mask(j, span) {
6339                        if (group_fn(j, cpu_map, NULL) != group)
6340                                continue;
6341
6342                        cpu_set(j, covered);
6343                        cpu_set(j, sg->cpumask);
6344                }
6345                if (!first)
6346                        first = sg;
6347                if (last)
6348                        last->next = sg;
6349                last = sg;
6350        }
6351        last->next = first;
6352}
6353
6354#define SD_NODES_PER_DOMAIN 16
6355
6356#ifdef CONFIG_NUMA
6357
6358/**
6359 * find_next_best_node - find the next node to include in a sched_domain
6360 * @node: node whose sched_domain we're building
6361 * @used_nodes: nodes already in the sched_domain
6362 *
6363 * Find the next node to include in a given scheduling domain. Simply
6364 * finds the closest node not already in the @used_nodes map.
6365 *
6366 * Should use nodemask_t.
6367 */
6368static int find_next_best_node(int node, unsigned long *used_nodes)
6369{
6370        int i, n, val, min_val, best_node = 0;
6371
6372        min_val = INT_MAX;
6373
6374        for (i = 0; i < MAX_NUMNODES; i++) {
6375                /* Start at @node */
6376                n = (node + i) % MAX_NUMNODES;
6377
6378                if (!nr_cpus_node(n))
6379                        continue;
6380
6381                /* Skip already used nodes */
6382                if (test_bit(n, used_nodes))
6383                        continue;
6384
6385                /* Simple min distance search */
6386                val = node_distance(node, n);
6387
6388                if (val < min_val) {
6389                        min_val = val;
6390                        best_node = n;
6391                }
6392        }
6393
6394        set_bit(best_node, used_nodes);
6395        return best_node;
6396}
6397
6398/**
6399 * sched_domain_node_span - get a cpumask for a node's sched_domain
6400 * @node: node whose cpumask we're constructing
6401 * @size: number of nodes to include in this span
6402 *
6403 * Given a node, construct a good cpumask for its sched_domain to span. It
6404 * should be one that prevents unnecessary balancing, but also spreads tasks
6405 * out optimally.
6406 */
6407static cpumask_t sched_domain_node_span(int node)
6408{
6409        DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
6410        cpumask_t span, nodemask;
6411        int i;
6412
6413        cpus_clear(span);
6414        bitmap_zero(used_nodes, MAX_NUMNODES);
6415
6416        nodemask = node_to_cpumask(node);
6417        cpus_or(span, span, nodemask);
6418        set_bit(node, used_nodes);
6419
6420        for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6421                int next_node = find_next_best_node(node, used_nodes);
6422
6423                nodemask = node_to_cpumask(next_node);
6424                cpus_or(span, span, nodemask);
6425        }
6426
6427        return span;
6428}
6429#endif
6430
6431int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6432
6433/*
6434 * SMT sched-domains:
6435 */
6436#ifdef CONFIG_SCHED_SMT
6437static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6438static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6439
6440static int
6441cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6442{
6443        if (sg)
6444                *sg = &per_cpu(sched_group_cpus, cpu);
6445        return cpu;
6446}
6447#endif
6448
6449/*
6450 * multi-core sched-domains:
6451 */
6452#ifdef CONFIG_SCHED_MC
6453static DEFINE_PER_CPU(struct sched_domain, core_domains);
6454static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6455#endif
6456
6457#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6458static int
6459cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6460{
6461        int group;
6462        cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6463        cpus_and(mask, mask, *cpu_map);
6464        group = first_cpu(mask);
6465        if (sg)
6466                *sg = &per_cpu(sched_group_core, group);
6467        return group;
6468}
6469#elif defined(CONFIG_SCHED_MC)
6470static int
6471cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6472{
6473        if (sg)
6474                *sg = &per_cpu(sched_group_core, cpu);
6475        return cpu;
6476}
6477#endif
6478
6479static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6480static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6481
6482static int
6483cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
6484{
6485        int group;
6486#ifdef CONFIG_SCHED_MC
6487        cpumask_t mask = cpu_coregroup_map(cpu);
6488        cpus_and(mask, mask, *cpu_map);
6489        group = first_cpu(mask);
6490#elif defined(CONFIG_SCHED_SMT)
6491        cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
6492        cpus_and(mask, mask, *cpu_map);
6493        group = first_cpu(mask);
6494#else
6495        group = cpu;
6496#endif
6497        if (sg)
6498                *sg = &per_cpu(sched_group_phys, group);
6499        return group;
6500}
6501
6502#ifdef CONFIG_NUMA
6503/*
6504 * The init_sched_build_groups can't handle what we want to do with node
6505 * groups, so roll our own. Now each node has its own list of groups which
6506 * gets dynamically allocated.
6507 */
6508static DEFINE_PER_CPU(struct sched_domain, node_domains);
6509static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
6510
6511static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6512static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6513
6514static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6515                                 struct sched_group **sg)
6516{
6517        cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6518        int group;
6519
6520        cpus_and(nodemask, nodemask, *cpu_map);
6521        group = first_cpu(nodemask);
6522
6523        if (sg)
6524                *sg = &per_cpu(sched_group_allnodes, group);
6525        return group;
6526}
6527
6528static void init_numa_sched_groups_power(struct sched_group *group_head)
6529{
6530        struct sched_group *sg = group_head;
6531        int j;
6532
6533        if (!sg)
6534                return;
6535        do {
6536                for_each_cpu_mask(j, sg->cpumask) {
6537                        struct sched_domain *sd;
6538
6539                        sd = &per_cpu(phys_domains, j);
6540                        if (j != first_cpu(sd->groups->cpumask)) {
6541                                /*
6542                                 * Only add "power" once for each
6543                                 * physical package.
6544                                 */
6545                                continue;
6546                        }
6547
6548                        sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6549                }
6550                sg = sg->next;
6551        } while (sg != group_head);
6552}
6553#endif
6554
6555#ifdef CONFIG_NUMA
6556/* Free memory allocated for various sched_group structures */
6557static void free_sched_groups(const cpumask_t *cpu_map)
6558{
6559        int cpu, i;
6560
6561        for_each_cpu_mask(cpu, *cpu_map) {
6562                struct sched_group **sched_group_nodes
6563                        = sched_group_nodes_bycpu[cpu];
6564
6565                if (!sched_group_nodes)
6566                        continue;
6567
6568                for (i = 0; i < MAX_NUMNODES; i++) {
6569                        cpumask_t nodemask = node_to_cpumask(i);
6570                        struct sched_group *oldsg, *sg = sched_group_nodes[i];
6571
6572                        cpus_and(nodemask, nodemask, *cpu_map);
6573                        if (cpus_empty(nodemask))
6574                                continue;
6575
6576                        if (sg == NULL)
6577                                continue;
6578                        sg = sg->next;
6579next_sg:
6580                        oldsg = sg;
6581                        sg = sg->next;
6582                        kfree(oldsg);
6583                        if (oldsg != sched_group_nodes[i])
6584                                goto next_sg;
6585                }
6586                kfree(sched_group_nodes);
6587                sched_group_nodes_bycpu[cpu] = NULL;
6588        }
6589}
6590#else
6591static void free_sched_groups(const cpumask_t *cpu_map)
6592{
6593}
6594#endif
6595
6596/*
6597 * Initialize sched groups cpu_power.
6598 *
6599 * cpu_power indicates the capacity of sched group, which is used while
6600 * distributing the load between different sched groups in a sched domain.
6601 * Typically cpu_power for all the groups in a sched domain will be same unless
6602 * there are asymmetries in the topology. If there are asymmetries, group
6603 * having more cpu_power will pickup more load compared to the group having
6604 * less cpu_power.
6605 *
6606 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6607 * the maximum number of tasks a group can handle in the presence of other idle
6608 * or lightly loaded groups in the same sched domain.
6609 */
6610static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6611{
6612        struct sched_domain *child;
6613        struct sched_group *group;
6614
6615        WARN_ON(!sd || !sd->groups);
6616
6617        if (cpu != first_cpu(sd->groups->cpumask))
6618                return;
6619
6620        child = sd->child;
6621
6622        sd->groups->__cpu_power = 0;
6623
6624        /*
6625         * For perf policy, if the groups in child domain share resources
6626         * (for example cores sharing some portions of the cache hierarchy
6627         * or SMT), then set this domain groups cpu_power such that each group
6628         * can handle only one task, when there are other idle groups in the
6629         * same sched domain.
6630         */
6631        if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6632                       (child->flags &
6633                        (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6634                sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6635                return;
6636        }
6637
6638        /*
6639         * add cpu_power of each child group to this groups cpu_power
6640         */
6641        group = child->groups;
6642        do {
6643                sg_inc_cpu_power(sd->groups, group->__cpu_power);
6644                group = group->next;
6645        } while (group != child->groups);
6646}
6647
6648/*
6649 * Build sched domains for a given set of cpus and attach the sched domains
6650 * to the individual cpus
6651 */
6652static int build_sched_domains(const cpumask_t *cpu_map)
6653{
6654        int i;
6655        struct root_domain *rd;
6656#ifdef CONFIG_NUMA
6657        struct sched_group **sched_group_nodes = NULL;
6658        int sd_allnodes = 0;
6659
6660        /*
6661         * Allocate the per-node list of sched groups
6662         */
6663        sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6664                                    GFP_KERNEL);
6665        if (!sched_group_nodes) {
6666                printk(KERN_WARNING "Can not alloc sched group node list\n");
6667                return -ENOMEM;
6668        }
6669        sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6670#endif
6671
6672        rd = alloc_rootdomain();
6673        if (!rd) {
6674                printk(KERN_WARNING "Cannot alloc root domain\n");
6675                return -ENOMEM;
6676        }
6677
6678        /*
6679         * Set up domains for cpus specified by the cpu_map.
6680         */
6681        for_each_cpu_mask(i, *cpu_map) {
6682                struct sched_domain *sd = NULL, *p;
6683                cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6684
6685                cpus_and(nodemask, nodemask, *cpu_map);
6686
6687#ifdef CONFIG_NUMA
6688                if (cpus_weight(*cpu_map) >
6689                                SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
6690                        sd = &per_cpu(allnodes_domains, i);
6691                        *sd = SD_ALLNODES_INIT;
6692                        sd->span = *cpu_map;
6693                        cpu_to_allnodes_group(i, cpu_map, &sd->groups);
6694                        p = sd;
6695                        sd_allnodes = 1;
6696                } else
6697                        p = NULL;
6698
6699                sd = &per_cpu(node_domains, i);
6700                *sd = SD_NODE_INIT;
6701                sd->span = sched_domain_node_span(cpu_to_node(i));
6702                sd->parent = p;
6703                if (p)
6704                        p->child = sd;
6705                cpus_and(sd->span, sd->span, *cpu_map);
6706#endif
6707
6708                p = sd;
6709                sd = &per_cpu(phys_domains, i);
6710                *sd = SD_CPU_INIT;
6711                sd->span = nodemask;
6712                sd->parent = p;
6713                if (p)
6714                        p->child = sd;
6715                cpu_to_phys_group(i, cpu_map, &sd->groups);
6716
6717#ifdef CONFIG_SCHED_MC
6718                p = sd;
6719                sd = &per_cpu(core_domains, i);
6720                *sd = SD_MC_INIT;
6721                sd->span = cpu_coregroup_map(i);
6722                cpus_and(sd->span, sd->span, *cpu_map);
6723                sd->parent = p;
6724                p->child = sd;
6725                cpu_to_core_group(i, cpu_map, &sd->groups);
6726#endif
6727
6728#ifdef CONFIG_SCHED_SMT
6729                p = sd;
6730                sd = &per_cpu(cpu_domains, i);
6731                *sd = SD_SIBLING_INIT;
6732                sd->span = per_cpu(cpu_sibling_map, i);
6733                cpus_and(sd->span, sd->span, *cpu_map);
6734                sd->parent = p;
6735                p->child = sd;
6736                cpu_to_cpu_group(i, cpu_map, &sd->groups);
6737#endif
6738        }
6739
6740#ifdef CONFIG_SCHED_SMT
6741        /* Set up CPU (sibling) groups */
6742        for_each_cpu_mask(i, *cpu_map) {
6743                cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
6744                cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
6745                if (i != first_cpu(this_sibling_map))
6746                        continue;
6747
6748                init_sched_build_groups(this_sibling_map, cpu_map,
6749                                        &cpu_to_cpu_group);
6750        }
6751#endif
6752
6753#ifdef CONFIG_SCHED_MC
6754        /* Set up multi-core groups */
6755        for_each_cpu_mask(i, *cpu_map) {
6756                cpumask_t this_core_map = cpu_coregroup_map(i);
6757                cpus_and(this_core_map, this_core_map, *cpu_map);
6758                if (i != first_cpu(this_core_map))
6759                        continue;
6760                init_sched_build_groups(this_core_map, cpu_map,
6761                                        &cpu_to_core_group);
6762        }
6763#endif
6764
6765        /* Set up physical groups */
6766        for (i = 0; i < MAX_NUMNODES; i++) {
6767                cpumask_t nodemask = node_to_cpumask(i);
6768
6769                cpus_and(nodemask, nodemask, *cpu_map);
6770                if (cpus_empty(nodemask))
6771                        continue;
6772
6773                init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
6774        }
6775
6776#ifdef CONFIG_NUMA
6777        /* Set up node groups */
6778        if (sd_allnodes)
6779                init_sched_build_groups(*cpu_map, cpu_map,
6780                                        &cpu_to_allnodes_group);
6781
6782        for (i = 0; i < MAX_NUMNODES; i++) {
6783                /* Set up node groups */
6784                struct sched_group *sg, *prev;
6785                cpumask_t nodemask = node_to_cpumask(i);
6786                cpumask_t domainspan;
6787                cpumask_t covered = CPU_MASK_NONE;
6788                int j;
6789
6790                cpus_and(nodemask, nodemask, *cpu_map);
6791                if (cpus_empty(nodemask)) {
6792                        sched_group_nodes[i] = NULL;
6793                        continue;
6794                }
6795
6796                domainspan = sched_domain_node_span(i);
6797                cpus_and(domainspan, domainspan, *cpu_map);
6798
6799                sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
6800                if (!sg) {
6801                        printk(KERN_WARNING "Can not alloc domain group for "
6802                                "node %d\n", i);
6803                        goto error;
6804                }
6805                sched_group_nodes[i] = sg;
6806                for_each_cpu_mask(j, nodemask) {
6807                        struct sched_domain *sd;
6808
6809                        sd = &per_cpu(node_domains, j);
6810                        sd->groups = sg;
6811                }
6812                sg->__cpu_power = 0;
6813                sg->cpumask = nodemask;
6814                sg->next = sg;
6815                cpus_or(covered, covered, nodemask);
6816                prev = sg;
6817
6818                for (j = 0; j < MAX_NUMNODES; j++) {
6819                        cpumask_t tmp, notcovered;
6820                        int n = (i + j) % MAX_NUMNODES;
6821
6822                        cpus_complement(notcovered, covered);
6823                        cpus_and(tmp, notcovered, *cpu_map);
6824                        cpus_and(tmp, tmp, domainspan);
6825                        if (cpus_empty(tmp))
6826                                break;
6827
6828                        nodemask = node_to_cpumask(n);
6829                        cpus_and(tmp, tmp, nodemask);
6830                        if (cpus_empty(tmp))
6831                                continue;
6832
6833                        sg = kmalloc_node(sizeof(struct sched_group),
6834                                          GFP_KERNEL, i);
6835                        if (!sg) {
6836                                printk(KERN_WARNING
6837                                "Can not alloc domain group for node %d\n", j);
6838                                goto error;
6839                        }
6840                        sg->__cpu_power = 0;
6841                        sg->cpumask = tmp;
6842                        sg->next = prev->next;
6843                        cpus_or(covered, covered, tmp);
6844                        prev->next = sg;
6845                        prev = sg;
6846                }
6847        }
6848#endif
6849
6850        /* Calculate CPU power for physical packages and nodes */
6851#ifdef CONFIG_SCHED_SMT
6852        for_each_cpu_mask(i, *cpu_map) {
6853                struct sched_domain *sd = &per_cpu(cpu_domains, i);
6854
6855                init_sched_groups_power(i, sd);
6856        }
6857#endif
6858#ifdef CONFIG_SCHED_MC
6859        for_each_cpu_mask(i, *cpu_map) {
6860                struct sched_domain *sd = &per_cpu(core_domains, i);
6861
6862                init_sched_groups_power(i, sd);
6863        }
6864#endif
6865
6866        for_each_cpu_mask(i, *cpu_map) {
6867                struct sched_domain *sd = &per_cpu(phys_domains, i);
6868
6869                init_sched_groups_power(i, sd);
6870        }
6871
6872#ifdef CONFIG_NUMA
6873        for (i = 0; i < MAX_NUMNODES; i++)
6874                init_numa_sched_groups_power(sched_group_nodes[i]);
6875
6876        if (sd_allnodes) {
6877                struct sched_group *sg;
6878
6879                cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
6880                init_numa_sched_groups_power(sg);
6881        }
6882#endif
6883
6884        /* Attach the domains */
6885        for_each_cpu_mask(i, *cpu_map) {
6886                struct sched_domain *sd;
6887#ifdef CONFIG_SCHED_SMT
6888                sd = &per_cpu(cpu_domains, i);
6889#elif defined(CONFIG_SCHED_MC)
6890                sd = &per_cpu(core_domains, i);
6891#else
6892                sd = &per_cpu(phys_domains, i);
6893#endif
6894                cpu_attach_domain(sd, rd, i);
6895        }
6896
6897        return 0;
6898
6899#ifdef CONFIG_NUMA
6900error:
6901        free_sched_groups(cpu_map);
6902        return -ENOMEM;
6903#endif
6904}
6905
6906static cpumask_t *doms_cur;     /* current sched domains */
6907static int ndoms_cur;           /* number of sched domains in 'doms_cur' */
6908
6909/*
6910 * Special case: If a kmalloc of a doms_cur partition (array of
6911 * cpumask_t) fails, then fallback to a single sched domain,
6912 * as determined by the single cpumask_t fallback_doms.
6913 */
6914static cpumask_t fallback_doms;
6915
6916void __attribute__((weak)) arch_update_cpu_topology(void)
6917{
6918}
6919
6920/*
6921 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6922 * For now this just excludes isolated cpus, but could be used to
6923 * exclude other special cases in the future.
6924 */
6925static int arch_init_sched_domains(const cpumask_t *cpu_map)
6926{
6927        int err;
6928
6929        arch_update_cpu_topology();
6930        ndoms_cur = 1;
6931        doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6932        if (!doms_cur)
6933                doms_cur = &fallback_doms;
6934        cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
6935        err = build_sched_domains(doms_cur);
6936        register_sched_domain_sysctl();
6937
6938        return err;
6939}
6940
6941static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
6942{
6943        free_sched_groups(cpu_map);
6944}
6945
6946/*
6947 * Detach sched domains from a group of cpus specified in cpu_map
6948 * These cpus will now be attached to the NULL domain
6949 */
6950static void detach_destroy_domains(const cpumask_t *cpu_map)
6951{
6952        int i;
6953
6954        unregister_sched_domain_sysctl();
6955
6956        for_each_cpu_mask(i, *cpu_map)
6957                cpu_attach_domain(NULL, &def_root_domain, i);
6958        synchronize_sched();
6959        arch_destroy_sched_domains(cpu_map);
6960}
6961
6962/*
6963 * Partition sched domains as specified by the 'ndoms_new'
6964 * cpumasks in the array doms_new[] of cpumasks. This compares
6965 * doms_new[] to the current sched domain partitioning, doms_cur[].
6966 * It destroys each deleted domain and builds each new domain.
6967 *
6968 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
6969 * The masks don't intersect (don't overlap.) We should setup one
6970 * sched domain for each mask. CPUs not in any of the cpumasks will
6971 * not be load balanced. If the same cpumask appears both in the
6972 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6973 * it as it is.
6974 *
6975 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6976 * ownership of it and will kfree it when done with it. If the caller
6977 * failed the kmalloc call, then it can pass in doms_new == NULL,
6978 * and partition_sched_domains() will fallback to the single partition
6979 * 'fallback_doms'.
6980 *
6981 * Call with hotplug lock held
6982 */
6983void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6984{
6985        int i, j;
6986
6987        lock_doms_cur();
6988
6989        /* always unregister in case we don't destroy any domains */
6990        unregister_sched_domain_sysctl();
6991
6992        if (doms_new == NULL) {
6993                ndoms_new = 1;
6994                doms_new = &fallback_doms;
6995                cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6996        }
6997
6998        /* Destroy deleted domains */
6999        for (i = 0; i < ndoms_cur; i++) {
7000                for (j = 0; j < ndoms_new; j++) {
7001                        if (cpus_equal(doms_cur[i], doms_new[j]))
7002                                goto match1;
7003                }
7004                /* no match - a current sched domain not in new doms_new[] */
7005                detach_destroy_domains(doms_cur + i);
7006match1:
7007                ;
7008        }
7009
7010        /* Build new domains */
7011        for (i = 0; i < ndoms_new; i++) {
7012                for (j = 0; j < ndoms_cur; j++) {
7013                        if (cpus_equal(doms_new[i], doms_cur[j]))
7014                                goto match2;
7015                }
7016                /* no match - add a new doms_new */
7017                build_sched_domains(doms_new + i);
7018match2:
7019                ;
7020        }
7021
7022        /* Remember the new sched domains */
7023        if (doms_cur != &fallback_doms)
7024                kfree(doms_cur);
7025        doms_cur = doms_new;
7026        ndoms_cur = ndoms_new;
7027
7028        register_sched_domain_sysctl();
7029
7030        unlock_doms_cur();
7031}
7032
7033#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7034int arch_reinit_sched_domains(void)
7035{
7036        int err;
7037
7038        get_online_cpus();
7039        detach_destroy_domains(&cpu_online_map);
7040        err = arch_init_sched_domains(&cpu_online_map);
7041        put_online_cpus();
7042
7043        return err;
7044}
7045
7046static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7047{
7048        int ret;
7049
7050        if (buf[0] != '0' && buf[0] != '1')
7051                return -EINVAL;
7052
7053        if (smt)
7054                sched_smt_power_savings = (buf[0] == '1');
7055        else
7056                sched_mc_power_savings = (buf[0] == '1');
7057
7058        ret = arch_reinit_sched_domains();
7059
7060        return ret ? ret : count;
7061}
7062
7063#ifdef CONFIG_SCHED_MC
7064static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7065{
7066        return sprintf(page, "%u\n", sched_mc_power_savings);
7067}
7068static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7069                                            const char *buf, size_t count)
7070{
7071        return sched_power_savings_store(buf, count, 0);
7072}
7073static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7074                   sched_mc_power_savings_store);
7075#endif
7076
7077#ifdef CONFIG_SCHED_SMT
7078static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7079{
7080        return sprintf(page, "%u\n", sched_smt_power_savings);
7081}
7082static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7083                                             const char *buf, size_t count)
7084{
7085        return sched_power_savings_store(buf, count, 1);
7086}
7087static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7088                   sched_smt_power_savings_store);
7089#endif
7090
7091int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7092{
7093        int err = 0;
7094
7095#ifdef CONFIG_SCHED_SMT
7096        if (smt_capable())
7097                err = sysfs_create_file(&cls->kset.kobj,
7098                                        &attr_sched_smt_power_savings.attr);
7099#endif
7100#ifdef CONFIG_SCHED_MC
7101        if (!err && mc_capable())
7102                err = sysfs_create_file(&cls->kset.kobj,
7103                                        &attr_sched_mc_power_savings.attr);
7104#endif
7105        return err;
7106}
7107#endif
7108
7109/*
7110 * Force a reinitialization of the sched domains hierarchy. The domains
7111 * and groups cannot be updated in place without racing with the balancing
7112 * code, so we temporarily attach all running cpus to the NULL domain
7113 * which will prevent rebalancing while the sched domains are recalculated.
7114 */
7115static int update_sched_domains(struct notifier_block *nfb,
7116                                unsigned long action, void *hcpu)
7117{
7118        switch (action) {
7119        case CPU_UP_PREPARE:
7120        case CPU_UP_PREPARE_FROZEN:
7121        case CPU_DOWN_PREPARE:
7122        case CPU_DOWN_PREPARE_FROZEN:
7123                detach_destroy_domains(&cpu_online_map);
7124                return NOTIFY_OK;
7125
7126        case CPU_UP_CANCELED:
7127        case CPU_UP_CANCELED_FROZEN:
7128        case CPU_DOWN_FAILED:
7129        case CPU_DOWN_FAILED_FROZEN:
7130        case CPU_ONLINE:
7131        case CPU_ONLINE_FROZEN:
7132        case CPU_DEAD:
7133        case CPU_DEAD_FROZEN:
7134                /*
7135                 * Fall through and re-initialise the domains.
7136                 */
7137                break;
7138        default:
7139                return NOTIFY_DONE;
7140        }
7141
7142        /* The hotplug lock is already held by cpu_up/cpu_down */
7143        arch_init_sched_domains(&cpu_online_map);
7144
7145        return NOTIFY_OK;
7146}
7147
7148void __init sched_init_smp(void)
7149{
7150        cpumask_t non_isolated_cpus;
7151
7152        get_online_cpus();
7153        arch_init_sched_domains(&cpu_online_map);
7154        cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7155        if (cpus_empty(non_isolated_cpus))
7156                cpu_set(smp_processor_id(), non_isolated_cpus);
7157        put_online_cpus();
7158        /* XXX: Theoretical race here - CPU may be hotplugged now */
7159        hotcpu_notifier(update_sched_domains, 0);
7160        init_hrtick();
7161
7162        /* Move init over to a non-isolated CPU */
7163        if (set_cpus_allowed(current, non_isolated_cpus) < 0)
7164                BUG();
7165        sched_init_granularity();
7166}
7167#else
7168void __init sched_init_smp(void)
7169{
7170        sched_init_granularity();
7171}
7172#endif /* CONFIG_SMP */
7173
7174int in_sched_functions(unsigned long addr)
7175{
7176        return in_lock_functions(addr) ||
7177                (addr >= (unsigned long)__sched_text_start
7178                && addr < (unsigned long)__sched_text_end);
7179}
7180
7181static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7182{
7183        cfs_rq->tasks_timeline = RB_ROOT;
7184#ifdef CONFIG_FAIR_GROUP_SCHED
7185        cfs_rq->rq = rq;
7186#endif
7187        cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7188}
7189
7190static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7191{
7192        struct rt_prio_array *array;
7193        int i;
7194
7195        array = &rt_rq->active;
7196        for (i = 0; i < MAX_RT_PRIO; i++) {
7197                INIT_LIST_HEAD(array->queue + i);
7198                __clear_bit(i, array->bitmap);
7199        }
7200        /* delimiter for bitsearch: */
7201        __set_bit(MAX_RT_PRIO, array->bitmap);
7202
7203#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7204        rt_rq->highest_prio = MAX_RT_PRIO;
7205#endif
7206#ifdef CONFIG_SMP
7207        rt_rq->rt_nr_migratory = 0;
7208        rt_rq->overloaded = 0;
7209#endif
7210
7211        rt_rq->rt_time = 0;
7212        rt_rq->rt_throttled = 0;
7213
7214#ifdef CONFIG_RT_GROUP_SCHED
7215        rt_rq->rt_nr_boosted = 0;
7216        rt_rq->rq = rq;
7217#endif
7218}
7219
7220#ifdef CONFIG_FAIR_GROUP_SCHED
7221static void init_tg_cfs_entry(struct rq *rq, struct task_group *tg,
7222                struct cfs_rq *cfs_rq, struct sched_entity *se,
7223                int cpu, int add)
7224{
7225        tg->cfs_rq[cpu] = cfs_rq;
7226        init_cfs_rq(cfs_rq, rq);
7227        cfs_rq->tg = tg;
7228        if (add)
7229                list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7230
7231        tg->se[cpu] = se;
7232        se->cfs_rq = &rq->cfs;
7233        se->my_q = cfs_rq;
7234        se->load.weight = tg->shares;
7235        se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
7236        se->parent = NULL;
7237}
7238#endif
7239
7240#ifdef CONFIG_RT_GROUP_SCHED
7241static void init_tg_rt_entry(struct rq *rq, struct task_group *tg,
7242                struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
7243                int cpu, int add)
7244{
7245        tg->rt_rq[cpu] = rt_rq;
7246        init_rt_rq(rt_rq, rq);
7247        rt_rq->tg = tg;
7248        rt_rq->rt_se = rt_se;
7249        if (add)
7250                list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7251
7252        tg->rt_se[cpu] = rt_se;
7253        rt_se->rt_rq = &rq->rt;
7254        rt_se->my_q = rt_rq;
7255        rt_se->parent = NULL;
7256        INIT_LIST_HEAD(&rt_se->run_list);
7257}
7258#endif
7259
7260void __init sched_init(void)
7261{
7262        int highest_cpu = 0;
7263        int i, j;
7264
7265#ifdef CONFIG_SMP
7266        init_defrootdomain();
7267#endif
7268
7269#ifdef CONFIG_GROUP_SCHED
7270        list_add(&init_task_group.list, &task_groups);
7271#endif
7272
7273        for_each_possible_cpu(i) {
7274                struct rq *rq;
7275
7276                rq = cpu_rq(i);
7277                spin_lock_init(&rq->lock);
7278                lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7279                rq->nr_running = 0;
7280                rq->clock = 1;
7281                init_cfs_rq(&rq->cfs, rq);
7282                init_rt_rq(&rq->rt, rq);
7283#ifdef CONFIG_FAIR_GROUP_SCHED
7284                init_task_group.shares = init_task_group_load;
7285                INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7286                init_tg_cfs_entry(rq, &init_task_group,
7287                                &per_cpu(init_cfs_rq, i),
7288                                &per_cpu(init_sched_entity, i), i, 1);
7289
7290#endif
7291#ifdef CONFIG_RT_GROUP_SCHED
7292                init_task_group.rt_runtime =
7293                        sysctl_sched_rt_runtime * NSEC_PER_USEC;
7294                INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7295                init_tg_rt_entry(rq, &init_task_group,
7296                                &per_cpu(init_rt_rq, i),
7297                                &per_cpu(init_sched_rt_entity, i), i, 1);
7298#endif
7299                rq->rt_period_expire = 0;
7300                rq->rt_throttled = 0;
7301
7302                for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7303                        rq->cpu_load[j] = 0;
7304#ifdef CONFIG_SMP
7305                rq->sd = NULL;
7306                rq->rd = NULL;
7307                rq->active_balance = 0;
7308                rq->next_balance = jiffies;
7309                rq->push_cpu = 0;
7310                rq->cpu = i;
7311                rq->migration_thread = NULL;
7312                INIT_LIST_HEAD(&rq->migration_queue);
7313                rq_attach_root(rq, &def_root_domain);
7314#endif
7315                init_rq_hrtick(rq);
7316                atomic_set(&rq->nr_iowait, 0);
7317                highest_cpu = i;
7318        }
7319
7320        set_load_weight(&init_task);
7321
7322#ifdef CONFIG_PREEMPT_NOTIFIERS
7323        INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7324#endif
7325
7326#ifdef CONFIG_SMP
7327        nr_cpu_ids = highest_cpu + 1;
7328        open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7329#endif
7330
7331#ifdef CONFIG_RT_MUTEXES
7332        plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7333#endif
7334
7335        /*
7336         * The boot idle thread does lazy MMU switching as well:
7337         */
7338        atomic_inc(&init_mm.mm_count);
7339        enter_lazy_tlb(&init_mm, current);
7340
7341        /*
7342         * Make us the idle thread. Technically, schedule() should not be
7343         * called from this thread, however somewhere below it might be,
7344         * but because we are the idle thread, we just pick up running again
7345         * when this runqueue becomes "idle".
7346         */
7347        init_idle(current, smp_processor_id());
7348        /*
7349         * During early bootup we pretend to be a normal task:
7350         */
7351        current->sched_class = &fair_sched_class;
7352
7353        scheduler_running = 1;
7354}
7355
7356#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7357void __might_sleep(char *file, int line)
7358{
7359#ifdef in_atomic
7360        static unsigned long prev_jiffy;        /* ratelimiting */
7361
7362        if ((in_atomic() || irqs_disabled()) &&
7363            system_state == SYSTEM_RUNNING && !oops_in_progress) {
7364                if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7365                        return;
7366                prev_jiffy = jiffies;
7367                printk(KERN_ERR "BUG: sleeping function called from invalid"
7368                                " context at %s:%d\n", file, line);
7369                printk("in_atomic():%d, irqs_disabled():%d\n",
7370                        in_atomic(), irqs_disabled());
7371                debug_show_held_locks(current);
7372                if (irqs_disabled())
7373                        print_irqtrace_events(current);
7374                dump_stack();
7375        }
7376#endif
7377}
7378EXPORT_SYMBOL(__might_sleep);
7379#endif
7380
7381#ifdef CONFIG_MAGIC_SYSRQ
7382static void normalize_task(struct rq *rq, struct task_struct *p)
7383{
7384        int on_rq;
7385        update_rq_clock(rq);
7386        on_rq = p->se.on_rq;
7387        if (on_rq)
7388                deactivate_task(rq, p, 0);
7389        __setscheduler(rq, p, SCHED_NORMAL, 0);
7390        if (on_rq) {
7391                activate_task(rq, p, 0);
7392                resched_task(rq->curr);
7393        }
7394}
7395
7396void normalize_rt_tasks(void)
7397{
7398        struct task_struct *g, *p;
7399        unsigned long flags;
7400        struct rq *rq;
7401
7402        read_lock_irqsave(&tasklist_lock, flags);
7403        do_each_thread(g, p) {
7404                /*
7405                 * Only normalize user tasks:
7406                 */
7407                if (!p->mm)
7408                        continue;
7409
7410                p->se.exec_start                = 0;
7411#ifdef CONFIG_SCHEDSTATS
7412                p->se.wait_start                = 0;
7413                p->se.sleep_start               = 0;
7414                p->se.block_start               = 0;
7415#endif
7416                task_rq(p)->clock               = 0;
7417
7418                if (!rt_task(p)) {
7419                        /*
7420                         * Renice negative nice level userspace
7421                         * tasks back to 0:
7422                         */
7423                        if (TASK_NICE(p) < 0 && p->mm)
7424                                set_user_nice(p, 0);
7425                        continue;
7426                }
7427
7428                spin_lock(&p->pi_lock);
7429                rq = __task_rq_lock(p);
7430
7431                normalize_task(rq, p);
7432
7433                __task_rq_unlock(rq);
7434                spin_unlock(&p->pi_lock);
7435        } while_each_thread(g, p);
7436
7437        read_unlock_irqrestore(&tasklist_lock, flags);
7438}
7439
7440#endif /* CONFIG_MAGIC_SYSRQ */
7441
7442#ifdef CONFIG_IA64
7443/*
7444 * These functions are only useful for the IA64 MCA handling.
7445 *
7446 * They can only be called when the whole system has been
7447 * stopped - every CPU needs to be quiescent, and no scheduling
7448 * activity can take place. Using them for anything else would
7449 * be a serious bug, and as a result, they aren't even visible
7450 * under any other configuration.
7451 */
7452
7453/**
7454 * curr_task - return the current task for a given cpu.
7455 * @cpu: the processor in question.
7456 *
7457 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7458 */
7459struct task_struct *curr_task(int cpu)
7460{
7461        return cpu_curr(cpu);
7462}
7463
7464/**
7465 * set_curr_task - set the current task for a given cpu.
7466 * @cpu: the processor in question.
7467 * @p: the task pointer to set.
7468 *
7469 * Description: This function must only be used when non-maskable interrupts
7470 * are serviced on a separate stack. It allows the architecture to switch the
7471 * notion of the current task on a cpu in a non-blocking manner. This function
7472 * must be called with all CPU's synchronized, and interrupts disabled, the
7473 * and caller must save the original value of the current task (see
7474 * curr_task() above) and restore that value before reenabling interrupts and
7475 * re-starting the system.
7476 *
7477 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7478 */
7479void set_curr_task(int cpu, struct task_struct *p)
7480{
7481        cpu_curr(cpu) = p;
7482}
7483
7484#endif
7485
7486#ifdef CONFIG_GROUP_SCHED
7487
7488#ifdef CONFIG_FAIR_GROUP_SCHED
7489static void free_fair_sched_group(struct task_group *tg)
7490{
7491        int i;
7492
7493        for_each_possible_cpu(i) {
7494                if (tg->cfs_rq)
7495                        kfree(tg->cfs_rq[i]);
7496                if (tg->se)
7497                        kfree(tg->se[i]);
7498        }
7499
7500        kfree(tg->cfs_rq);
7501        kfree(tg->se);
7502}
7503
7504static int alloc_fair_sched_group(struct task_group *tg)
7505{
7506        struct cfs_rq *cfs_rq;
7507        struct sched_entity *se;
7508        struct rq *rq;
7509        int i;
7510
7511        tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
7512        if (!tg->cfs_rq)
7513                goto err;
7514        tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
7515        if (!tg->se)
7516                goto err;
7517
7518        tg->shares = NICE_0_LOAD;
7519
7520        for_each_possible_cpu(i) {
7521                rq = cpu_rq(i);
7522
7523                cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
7524                                GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7525                if (!cfs_rq)
7526                        goto err;
7527
7528                se = kmalloc_node(sizeof(struct sched_entity),
7529                                GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7530                if (!se)
7531                        goto err;
7532
7533                init_tg_cfs_entry(rq, tg, cfs_rq, se, i, 0);
7534        }
7535
7536        return 1;
7537
7538 err:
7539        return 0;
7540}
7541
7542static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7543{
7544        list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7545                        &cpu_rq(cpu)->leaf_cfs_rq_list);
7546}
7547
7548static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7549{
7550        list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7551}
7552#else
7553static inline void free_fair_sched_group(struct task_group *tg)
7554{
7555}
7556
7557static inline int alloc_fair_sched_group(struct task_group *tg)
7558{
7559        return 1;
7560}
7561
7562static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7563{
7564}
7565
7566static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7567{
7568}
7569#endif
7570
7571#ifdef CONFIG_RT_GROUP_SCHED
7572static void free_rt_sched_group(struct task_group *tg)
7573{
7574        int i;
7575
7576        for_each_possible_cpu(i) {
7577                if (tg->rt_rq)
7578                        kfree(tg->rt_rq[i]);
7579                if (tg->rt_se)
7580                        kfree(tg->rt_se[i]);
7581        }
7582
7583        kfree(tg->rt_rq);
7584        kfree(tg->rt_se);
7585}
7586
7587static int alloc_rt_sched_group(struct task_group *tg)
7588{
7589        struct rt_rq *rt_rq;
7590        struct sched_rt_entity *rt_se;
7591        struct rq *rq;
7592        int i;
7593
7594        tg->rt_rq = kzalloc(sizeof(rt_rq) * NR_CPUS, GFP_KERNEL);
7595        if (!tg->rt_rq)
7596                goto err;
7597        tg->rt_se = kzalloc(sizeof(rt_se) * NR_CPUS, GFP_KERNEL);
7598        if (!tg->rt_se)
7599                goto err;
7600
7601        tg->rt_runtime = 0;
7602
7603        for_each_possible_cpu(i) {
7604                rq = cpu_rq(i);
7605
7606                rt_rq = kmalloc_node(sizeof(struct rt_rq),
7607                                GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7608                if (!rt_rq)
7609                        goto err;
7610
7611                rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
7612                                GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
7613                if (!rt_se)
7614                        goto err;
7615
7616                init_tg_rt_entry(rq, tg, rt_rq, rt_se, i, 0);
7617        }
7618
7619        return 1;
7620
7621 err:
7622        return 0;
7623}
7624
7625static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7626{
7627        list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7628                        &cpu_rq(cpu)->leaf_rt_rq_list);
7629}
7630
7631static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7632{
7633        list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7634}
7635#else
7636static inline void free_rt_sched_group(struct task_group *tg)
7637{
7638}
7639
7640static inline int alloc_rt_sched_group(struct task_group *tg)
7641{
7642        return 1;
7643}
7644
7645static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7646{
7647}
7648
7649static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7650{
7651}
7652#endif
7653
7654static void free_sched_group(struct task_group *tg)
7655{
7656        free_fair_sched_group(tg);
7657        free_rt_sched_group(tg);
7658        kfree(tg);
7659}
7660
7661/* allocate runqueue etc for a new task group */
7662struct task_group *sched_create_group(void)
7663{
7664        struct task_group *tg;
7665        unsigned long flags;
7666        int i;
7667
7668        tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7669        if (!tg)
7670                return ERR_PTR(-ENOMEM);
7671
7672        if (!alloc_fair_sched_group(tg))
7673                goto err;
7674
7675        if (!alloc_rt_sched_group(tg))
7676                goto err;
7677
7678        spin_lock_irqsave(&task_group_lock, flags);
7679        for_each_possible_cpu(i) {
7680                register_fair_sched_group(tg, i);
7681                register_rt_sched_group(tg, i);
7682        }
7683        list_add_rcu(&tg->list, &task_groups);
7684        spin_unlock_irqrestore(&task_group_lock, flags);
7685
7686        return tg;
7687
7688err:
7689        free_sched_group(tg);
7690        return ERR_PTR(-ENOMEM);
7691}
7692
7693/* rcu callback to free various structures associated with a task group */
7694static void free_sched_group_rcu(struct rcu_head *rhp)
7695{
7696        /* now it should be safe to free those cfs_rqs */
7697        free_sched_group(container_of(rhp, struct task_group, rcu));
7698}
7699
7700/* Destroy runqueue etc associated with a task group */
7701void sched_destroy_group(struct task_group *tg)
7702{
7703        unsigned long flags;
7704        int i;
7705
7706        spin_lock_irqsave(&task_group_lock, flags);
7707        for_each_possible_cpu(i) {
7708                unregister_fair_sched_group(tg, i);
7709                unregister_rt_sched_group(tg, i);
7710        }
7711        list_del_rcu(&tg->list);
7712        spin_unlock_irqrestore(&task_group_lock, flags);
7713
7714        /* wait for possible concurrent references to cfs_rqs complete */
7715        call_rcu(&tg->rcu, free_sched_group_rcu);
7716}
7717
7718/* change task's runqueue when it moves between groups.
7719 *      The caller of this function should have put the task in its new group
7720 *      by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7721 *      reflect its new group.
7722 */
7723void sched_move_task(struct task_struct *tsk)
7724{
7725        int on_rq, running;
7726        unsigned long flags;
7727        struct rq *rq;
7728
7729        rq = task_rq_lock(tsk, &flags);
7730
7731        update_rq_clock(rq);
7732
7733        running = task_current(rq, tsk);
7734        on_rq = tsk->se.on_rq;
7735
7736        if (on_rq)
7737                dequeue_task(rq, tsk, 0);
7738        if (unlikely(running))
7739                tsk->sched_class->put_prev_task(rq, tsk);
7740
7741        set_task_rq(tsk, task_cpu(tsk));
7742
7743#ifdef CONFIG_FAIR_GROUP_SCHED
7744        if (tsk->sched_class->moved_group)
7745                tsk->sched_class->moved_group(tsk);
7746#endif
7747
7748        if (unlikely(running))
7749                tsk->sched_class->set_curr_task(rq);
7750        if (on_rq)
7751                enqueue_task(rq, tsk, 0);
7752
7753        task_rq_unlock(rq, &flags);
7754}
7755
7756#ifdef CONFIG_FAIR_GROUP_SCHED
7757static void set_se_shares(struct sched_entity *se, unsigned long shares)
7758{
7759        struct cfs_rq *cfs_rq = se->cfs_rq;
7760        struct rq *rq = cfs_rq->rq;
7761        int on_rq;
7762
7763        spin_lock_irq(&rq->lock);
7764
7765        on_rq = se->on_rq;
7766        if (on_rq)
7767                dequeue_entity(cfs_rq, se, 0);
7768
7769        se->load.weight = shares;
7770        se->load.inv_weight = div64_64((1ULL<<32), shares);
7771
7772        if (on_rq)
7773                enqueue_entity(cfs_rq, se, 0);
7774
7775        spin_unlock_irq(&rq->lock);
7776}
7777
7778static DEFINE_MUTEX(shares_mutex);
7779
7780int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7781{
7782        int i;
7783        unsigned long flags;
7784
7785        /*
7786         * A weight of 0 or 1 can cause arithmetics problems.
7787         * (The default weight is 1024 - so there's no practical
7788         *  limitation from this.)
7789         */
7790        if (shares < 2)
7791                shares = 2;
7792
7793        mutex_lock(&shares_mutex);
7794        if (tg->shares == shares)
7795                goto done;
7796
7797        spin_lock_irqsave(&task_group_lock, flags);
7798        for_each_possible_cpu(i)
7799                unregister_fair_sched_group(tg, i);
7800        spin_unlock_irqrestore(&task_group_lock, flags);
7801
7802        /* wait for any ongoing reference to this group to finish */
7803        synchronize_sched();
7804
7805        /*
7806         * Now we are free to modify the group's share on each cpu
7807         * w/o tripping rebalance_share or load_balance_fair.
7808         */
7809        tg->shares = shares;
7810        for_each_possible_cpu(i)
7811                set_se_shares(tg->se[i], shares);
7812
7813        /*
7814         * Enable load balance activity on this group, by inserting it back on
7815         * each cpu's rq->leaf_cfs_rq_list.
7816         */
7817        spin_lock_irqsave(&task_group_lock, flags);
7818        for_each_possible_cpu(i)
7819                register_fair_sched_group(tg, i);
7820        spin_unlock_irqrestore(&task_group_lock, flags);
7821done:
7822        mutex_unlock(&shares_mutex);
7823        return 0;
7824}
7825
7826unsigned long sched_group_shares(struct task_group *tg)
7827{
7828        return tg->shares;
7829}
7830#endif
7831
7832#ifdef CONFIG_RT_GROUP_SCHED
7833/*
7834 * Ensure that the real time constraints are schedulable.
7835 */
7836static DEFINE_MUTEX(rt_constraints_mutex);
7837
7838static unsigned long to_ratio(u64 period, u64 runtime)
7839{
7840        if (runtime == RUNTIME_INF)
7841                return 1ULL << 16;
7842
7843        return div64_64(runtime << 16, period);
7844}
7845
7846static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7847{
7848        struct task_group *tgi;
7849        unsigned long total = 0;
7850        unsigned long global_ratio =
7851                to_ratio(sysctl_sched_rt_period,
7852                         sysctl_sched_rt_runtime < 0 ?
7853                                RUNTIME_INF : sysctl_sched_rt_runtime);
7854
7855        rcu_read_lock();
7856        list_for_each_entry_rcu(tgi, &task_groups, list) {
7857                if (tgi == tg)
7858                        continue;
7859
7860                total += to_ratio(period, tgi->rt_runtime);
7861        }
7862        rcu_read_unlock();
7863
7864        return total + to_ratio(period, runtime) < global_ratio;
7865}
7866
7867/* Must be called with tasklist_lock held */
7868static inline int tg_has_rt_tasks(struct task_group *tg)
7869{
7870        struct task_struct *g, *p;
7871        do_each_thread(g, p) {
7872                if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
7873                        return 1;
7874        } while_each_thread(g, p);
7875        return 0;
7876}
7877
7878int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7879{
7880        u64 rt_runtime, rt_period;
7881        int err = 0;
7882
7883        rt_period = (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
7884        rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7885        if (rt_runtime_us == -1)
7886                rt_runtime = RUNTIME_INF;
7887
7888        mutex_lock(&rt_constraints_mutex);
7889        read_lock(&tasklist_lock);
7890        if (rt_runtime_us == 0 && tg_has_rt_tasks(tg)) {
7891                err = -EBUSY;
7892                goto unlock;
7893        }
7894        if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
7895                err = -EINVAL;
7896                goto unlock;
7897        }
7898        tg->rt_runtime = rt_runtime;
7899 unlock:
7900        read_unlock(&tasklist_lock);
7901        mutex_unlock(&rt_constraints_mutex);
7902
7903        return err;
7904}
7905
7906long sched_group_rt_runtime(struct task_group *tg)
7907{
7908        u64 rt_runtime_us;
7909
7910        if (tg->rt_runtime == RUNTIME_INF)
7911                return -1;
7912
7913        rt_runtime_us = tg->rt_runtime;
7914        do_div(rt_runtime_us, NSEC_PER_USEC);
7915        return rt_runtime_us;
7916}
7917#endif
7918#endif  /* CONFIG_GROUP_SCHED */
7919
7920#ifdef CONFIG_CGROUP_SCHED
7921
7922/* return corresponding task_group object of a cgroup */
7923static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7924{
7925        return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7926                            struct task_group, css);
7927}
7928
7929static struct cgroup_subsys_state *
7930cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7931{
7932        struct task_group *tg;
7933
7934        if (!cgrp->parent) {
7935                /* This is early initialization for the top cgroup */
7936                init_task_group.css.cgroup = cgrp;
7937                return &init_task_group.css;
7938        }
7939
7940        /* we support only 1-level deep hierarchical scheduler atm */
7941        if (cgrp->parent->parent)
7942                return ERR_PTR(-EINVAL);
7943
7944        tg = sched_create_group();
7945        if (IS_ERR(tg))
7946                return ERR_PTR(-ENOMEM);
7947
7948        /* Bind the cgroup to task_group object we just created */
7949        tg->css.cgroup = cgrp;
7950
7951        return &tg->css;
7952}
7953
7954static void
7955cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7956{
7957        struct task_group *tg = cgroup_tg(cgrp);
7958
7959        sched_destroy_group(tg);
7960}
7961
7962static int
7963cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7964                      struct task_struct *tsk)
7965{
7966#ifdef CONFIG_RT_GROUP_SCHED
7967        /* Don't accept realtime tasks when there is no way for them to run */
7968        if (rt_task(tsk) && cgroup_tg(cgrp)->rt_runtime == 0)
7969                return -EINVAL;
7970#else
7971        /* We don't support RT-tasks being in separate groups */
7972        if (tsk->sched_class != &fair_sched_class)
7973                return -EINVAL;
7974#endif
7975
7976        return 0;
7977}
7978
7979static void
7980cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7981                        struct cgroup *old_cont, struct task_struct *tsk)
7982{
7983        sched_move_task(tsk);
7984}
7985
7986#ifdef CONFIG_FAIR_GROUP_SCHED
7987static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7988                                u64 shareval)
7989{
7990        return sched_group_set_shares(cgroup_tg(cgrp), shareval);
7991}
7992
7993static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
7994{
7995        struct task_group *tg = cgroup_tg(cgrp);
7996
7997        return (u64) tg->shares;
7998}
7999#endif
8000
8001#ifdef CONFIG_RT_GROUP_SCHED
8002static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8003                                struct file *file,
8004                                const char __user *userbuf,
8005                                size_t nbytes, loff_t *unused_ppos)
8006{
8007        char buffer[64];
8008        int retval = 0;
8009        s64 val;
8010        char *end;
8011
8012        if (!nbytes)
8013                return -EINVAL;
8014        if (nbytes >= sizeof(buffer))
8015                return -E2BIG;
8016        if (copy_from_user(buffer, userbuf, nbytes))
8017                return -EFAULT;
8018
8019        buffer[nbytes] = 0;     /* nul-terminate */
8020
8021        /* strip newline if necessary */
8022        if (nbytes && (buffer[nbytes-1] == '\n'))
8023                buffer[nbytes-1] = 0;
8024        val = simple_strtoll(buffer, &end, 0);
8025        if (*end)
8026                return -EINVAL;
8027
8028        /* Pass to subsystem */
8029        retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8030        if (!retval)
8031                retval = nbytes;
8032        return retval;
8033}
8034
8035static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
8036                                   struct file *file,
8037                                   char __user *buf, size_t nbytes,
8038                                   loff_t *ppos)
8039{
8040        char tmp[64];
8041        long val = sched_group_rt_runtime(cgroup_tg(cgrp));
8042        int len = sprintf(tmp, "%ld\n", val);
8043
8044        return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
8045}
8046#endif
8047
8048static struct cftype cpu_files[] = {
8049#ifdef CONFIG_FAIR_GROUP_SCHED
8050        {
8051                .name = "shares",
8052                .read_uint = cpu_shares_read_uint,
8053                .write_uint = cpu_shares_write_uint,
8054        },
8055#endif
8056#ifdef CONFIG_RT_GROUP_SCHED
8057        {
8058                .name = "rt_runtime_us",
8059                .read = cpu_rt_runtime_read,
8060                .write = cpu_rt_runtime_write,
8061        },
8062#endif
8063};
8064
8065static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8066{
8067        return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8068}
8069
8070struct cgroup_subsys cpu_cgroup_subsys = {
8071        .name           = "cpu",
8072        .create         = cpu_cgroup_create,
8073        .destroy        = cpu_cgroup_destroy,
8074        .can_attach     = cpu_cgroup_can_attach,
8075        .attach         = cpu_cgroup_attach,
8076        .populate       = cpu_cgroup_populate,
8077        .subsys_id      = cpu_cgroup_subsys_id,
8078        .early_init     = 1,
8079};
8080
8081#endif  /* CONFIG_CGROUP_SCHED */
8082
8083#ifdef CONFIG_CGROUP_CPUACCT
8084
8085/*
8086 * CPU accounting code for task groups.
8087 *
8088 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8089 * (balbir@in.ibm.com).
8090 */
8091
8092/* track cpu usage of a group of tasks */
8093struct cpuacct {
8094        struct cgroup_subsys_state css;
8095        /* cpuusage holds pointer to a u64-type object on every cpu */
8096        u64 *cpuusage;
8097};
8098
8099struct cgroup_subsys cpuacct_subsys;
8100
8101/* return cpu accounting group corresponding to this container */
8102static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
8103{
8104        return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
8105                            struct cpuacct, css);
8106}
8107
8108/* return cpu accounting group to which this task belongs */
8109static inline struct cpuacct *task_ca(struct task_struct *tsk)
8110{
8111        return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8112                            struct cpuacct, css);
8113}
8114
8115/* create a new cpu accounting group */
8116static struct cgroup_subsys_state *cpuacct_create(
8117        struct cgroup_subsys *ss, struct cgroup *cont)
8118{
8119        struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8120
8121        if (!ca)
8122                return ERR_PTR(-ENOMEM);
8123
8124        ca->cpuusage = alloc_percpu(u64);
8125        if (!ca->cpuusage) {
8126                kfree(ca);
8127                return ERR_PTR(-ENOMEM);
8128        }
8129
8130        return &ca->css;
8131}
8132
8133/* destroy an existing cpu accounting group */
8134static void
8135cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
8136{
8137        struct cpuacct *ca = cgroup_ca(cont);
8138
8139        free_percpu(ca->cpuusage);
8140        kfree(ca);
8141}
8142
8143/* return total cpu usage (in nanoseconds) of a group */
8144static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
8145{
8146        struct cpuacct *ca = cgroup_ca(cont);
8147        u64 totalcpuusage = 0;
8148        int i;
8149
8150        for_each_possible_cpu(i) {
8151                u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8152
8153                /*
8154                 * Take rq->lock to make 64-bit addition safe on 32-bit
8155                 * platforms.
8156                 */
8157                spin_lock_irq(&cpu_rq(i)->lock);
8158                totalcpuusage += *cpuusage;
8159                spin_unlock_irq(&cpu_rq(i)->lock);
8160        }
8161
8162        return totalcpuusage;
8163}
8164
8165static struct cftype files[] = {
8166        {
8167                .name = "usage",
8168                .read_uint = cpuusage_read,
8169        },
8170};
8171
8172static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8173{
8174        return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
8175}
8176
8177/*
8178 * charge this task's execution time to its accounting group.
8179 *
8180 * called with rq->lock held.
8181 */
8182static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8183{
8184        struct cpuacct *ca;
8185
8186        if (!cpuacct_subsys.active)
8187                return;
8188
8189        ca = task_ca(tsk);
8190        if (ca) {
8191                u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8192
8193                *cpuusage += cputime;
8194        }
8195}
8196
8197struct cgroup_subsys cpuacct_subsys = {
8198        .name = "cpuacct",
8199        .create = cpuacct_create,
8200        .destroy = cpuacct_destroy,
8201        .populate = cpuacct_populate,
8202        .subsys_id = cpuacct_subsys_id,
8203};
8204#endif  /* CONFIG_CGROUP_CPUACCT */
8205