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/perf_event.h>
  43#include <linux/security.h>
  44#include <linux/notifier.h>
  45#include <linux/profile.h>
  46#include <linux/freezer.h>
  47#include <linux/vmalloc.h>
  48#include <linux/blkdev.h>
  49#include <linux/delay.h>
  50#include <linux/pid_namespace.h>
  51#include <linux/smp.h>
  52#include <linux/threads.h>
  53#include <linux/timer.h>
  54#include <linux/rcupdate.h>
  55#include <linux/cpu.h>
  56#include <linux/cpuset.h>
  57#include <linux/percpu.h>
  58#include <linux/kthread.h>
  59#include <linux/proc_fs.h>
  60#include <linux/seq_file.h>
  61#include <linux/sysctl.h>
  62#include <linux/syscalls.h>
  63#include <linux/times.h>
  64#include <linux/tsacct_kern.h>
  65#include <linux/kprobes.h>
  66#include <linux/delayacct.h>
  67#include <linux/unistd.h>
  68#include <linux/pagemap.h>
  69#include <linux/hrtimer.h>
  70#include <linux/tick.h>
  71#include <linux/debugfs.h>
  72#include <linux/ctype.h>
  73#include <linux/ftrace.h>
  74
  75#include <asm/tlb.h>
  76#include <asm/irq_regs.h>
  77
  78#include "sched_cpupri.h"
  79
  80#define CREATE_TRACE_POINTS
  81#include <trace/events/sched.h>
  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/*
 118 * single value that denotes runtime == period, ie unlimited time.
 119 */
 120#define RUNTIME_INF     ((u64)~0ULL)
 121
 122static inline int rt_policy(int policy)
 123{
 124        if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
 125                return 1;
 126        return 0;
 127}
 128
 129static inline int task_has_rt_policy(struct task_struct *p)
 130{
 131        return rt_policy(p->policy);
 132}
 133
 134/*
 135 * This is the priority-queue data structure of the RT scheduling class:
 136 */
 137struct rt_prio_array {
 138        DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
 139        struct list_head queue[MAX_RT_PRIO];
 140};
 141
 142struct rt_bandwidth {
 143        /* nests inside the rq lock: */
 144        raw_spinlock_t          rt_runtime_lock;
 145        ktime_t                 rt_period;
 146        u64                     rt_runtime;
 147        struct hrtimer          rt_period_timer;
 148};
 149
 150static struct rt_bandwidth def_rt_bandwidth;
 151
 152static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
 153
 154static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
 155{
 156        struct rt_bandwidth *rt_b =
 157                container_of(timer, struct rt_bandwidth, rt_period_timer);
 158        ktime_t now;
 159        int overrun;
 160        int idle = 0;
 161
 162        for (;;) {
 163                now = hrtimer_cb_get_time(timer);
 164                overrun = hrtimer_forward(timer, now, rt_b->rt_period);
 165
 166                if (!overrun)
 167                        break;
 168
 169                idle = do_sched_rt_period_timer(rt_b, overrun);
 170        }
 171
 172        return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
 173}
 174
 175static
 176void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
 177{
 178        rt_b->rt_period = ns_to_ktime(period);
 179        rt_b->rt_runtime = runtime;
 180
 181        raw_spin_lock_init(&rt_b->rt_runtime_lock);
 182
 183        hrtimer_init(&rt_b->rt_period_timer,
 184                        CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 185        rt_b->rt_period_timer.function = sched_rt_period_timer;
 186}
 187
 188static inline int rt_bandwidth_enabled(void)
 189{
 190        return sysctl_sched_rt_runtime >= 0;
 191}
 192
 193static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
 194{
 195        ktime_t now;
 196
 197        if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
 198                return;
 199
 200        if (hrtimer_active(&rt_b->rt_period_timer))
 201                return;
 202
 203        raw_spin_lock(&rt_b->rt_runtime_lock);
 204        for (;;) {
 205                unsigned long delta;
 206                ktime_t soft, hard;
 207
 208                if (hrtimer_active(&rt_b->rt_period_timer))
 209                        break;
 210
 211                now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
 212                hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
 213
 214                soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
 215                hard = hrtimer_get_expires(&rt_b->rt_period_timer);
 216                delta = ktime_to_ns(ktime_sub(hard, soft));
 217                __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
 218                                HRTIMER_MODE_ABS_PINNED, 0);
 219        }
 220        raw_spin_unlock(&rt_b->rt_runtime_lock);
 221}
 222
 223#ifdef CONFIG_RT_GROUP_SCHED
 224static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
 225{
 226        hrtimer_cancel(&rt_b->rt_period_timer);
 227}
 228#endif
 229
 230/*
 231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
 232 * detach_destroy_domains and partition_sched_domains.
 233 */
 234static DEFINE_MUTEX(sched_domains_mutex);
 235
 236#ifdef CONFIG_GROUP_SCHED
 237
 238#include <linux/cgroup.h>
 239
 240struct cfs_rq;
 241
 242static LIST_HEAD(task_groups);
 243
 244/* task group related information */
 245struct task_group {
 246#ifdef CONFIG_CGROUP_SCHED
 247        struct cgroup_subsys_state css;
 248#endif
 249
 250#ifdef CONFIG_USER_SCHED
 251        uid_t uid;
 252#endif
 253
 254#ifdef CONFIG_FAIR_GROUP_SCHED
 255        /* schedulable entities of this group on each cpu */
 256        struct sched_entity **se;
 257        /* runqueue "owned" by this group on each cpu */
 258        struct cfs_rq **cfs_rq;
 259        unsigned long shares;
 260#endif
 261
 262#ifdef CONFIG_RT_GROUP_SCHED
 263        struct sched_rt_entity **rt_se;
 264        struct rt_rq **rt_rq;
 265
 266        struct rt_bandwidth rt_bandwidth;
 267#endif
 268
 269        struct rcu_head rcu;
 270        struct list_head list;
 271
 272        struct task_group *parent;
 273        struct list_head siblings;
 274        struct list_head children;
 275};
 276
 277#ifdef CONFIG_USER_SCHED
 278
 279/* Helper function to pass uid information to create_sched_user() */
 280void set_tg_uid(struct user_struct *user)
 281{
 282        user->tg->uid = user->uid;
 283}
 284
 285/*
 286 * Root task group.
 287 *      Every UID task group (including init_task_group aka UID-0) will
 288 *      be a child to this group.
 289 */
 290struct task_group root_task_group;
 291
 292#ifdef CONFIG_FAIR_GROUP_SCHED
 293/* Default task group's sched entity on each cpu */
 294static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
 295/* Default task group's cfs_rq on each cpu */
 296static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
 297#endif /* CONFIG_FAIR_GROUP_SCHED */
 298
 299#ifdef CONFIG_RT_GROUP_SCHED
 300static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
 301static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq_var);
 302#endif /* CONFIG_RT_GROUP_SCHED */
 303#else /* !CONFIG_USER_SCHED */
 304#define root_task_group init_task_group
 305#endif /* CONFIG_USER_SCHED */
 306
 307/* task_group_lock serializes add/remove of task groups and also changes to
 308 * a task group's cpu shares.
 309 */
 310static DEFINE_SPINLOCK(task_group_lock);
 311
 312#ifdef CONFIG_FAIR_GROUP_SCHED
 313
 314#ifdef CONFIG_SMP
 315static int root_task_group_empty(void)
 316{
 317        return list_empty(&root_task_group.children);
 318}
 319#endif
 320
 321#ifdef CONFIG_USER_SCHED
 322# define INIT_TASK_GROUP_LOAD   (2*NICE_0_LOAD)
 323#else /* !CONFIG_USER_SCHED */
 324# define INIT_TASK_GROUP_LOAD   NICE_0_LOAD
 325#endif /* CONFIG_USER_SCHED */
 326
 327/*
 328 * A weight of 0 or 1 can cause arithmetics problems.
 329 * A weight of a cfs_rq is the sum of weights of which entities
 330 * are queued on this cfs_rq, so a weight of a entity should not be
 331 * too large, so as the shares value of a task group.
 332 * (The default weight is 1024 - so there's no practical
 333 *  limitation from this.)
 334 */
 335#define MIN_SHARES      2
 336#define MAX_SHARES      (1UL << 18)
 337
 338static int init_task_group_load = INIT_TASK_GROUP_LOAD;
 339#endif
 340
 341/* Default task group.
 342 *      Every task in system belong to this group at bootup.
 343 */
 344struct task_group init_task_group;
 345
 346/* return group to which a task belongs */
 347static inline struct task_group *task_group(struct task_struct *p)
 348{
 349        struct task_group *tg;
 350
 351#ifdef CONFIG_USER_SCHED
 352        rcu_read_lock();
 353        tg = __task_cred(p)->user->tg;
 354        rcu_read_unlock();
 355#elif defined(CONFIG_CGROUP_SCHED)
 356        tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
 357                                struct task_group, css);
 358#else
 359        tg = &init_task_group;
 360#endif
 361        return tg;
 362}
 363
 364/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
 365static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
 366{
 367#ifdef CONFIG_FAIR_GROUP_SCHED
 368        p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
 369        p->se.parent = task_group(p)->se[cpu];
 370#endif
 371
 372#ifdef CONFIG_RT_GROUP_SCHED
 373        p->rt.rt_rq  = task_group(p)->rt_rq[cpu];
 374        p->rt.parent = task_group(p)->rt_se[cpu];
 375#endif
 376}
 377
 378#else
 379
 380static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
 381static inline struct task_group *task_group(struct task_struct *p)
 382{
 383        return NULL;
 384}
 385
 386#endif  /* CONFIG_GROUP_SCHED */
 387
 388/* CFS-related fields in a runqueue */
 389struct cfs_rq {
 390        struct load_weight load;
 391        unsigned long nr_running;
 392
 393        u64 exec_clock;
 394        u64 min_vruntime;
 395
 396        struct rb_root tasks_timeline;
 397        struct rb_node *rb_leftmost;
 398
 399        struct list_head tasks;
 400        struct list_head *balance_iterator;
 401
 402        /*
 403         * 'curr' points to currently running entity on this cfs_rq.
 404         * It is set to NULL otherwise (i.e when none are currently running).
 405         */
 406        struct sched_entity *curr, *next, *last;
 407
 408        unsigned int nr_spread_over;
 409
 410#ifdef CONFIG_FAIR_GROUP_SCHED
 411        struct rq *rq;  /* cpu runqueue to which this cfs_rq is attached */
 412
 413        /*
 414         * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
 415         * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
 416         * (like users, containers etc.)
 417         *
 418         * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
 419         * list is used during load balance.
 420         */
 421        struct list_head leaf_cfs_rq_list;
 422        struct task_group *tg;  /* group that "owns" this runqueue */
 423
 424#ifdef CONFIG_SMP
 425        /*
 426         * the part of load.weight contributed by tasks
 427         */
 428        unsigned long task_weight;
 429
 430        /*
 431         *   h_load = weight * f(tg)
 432         *
 433         * Where f(tg) is the recursive weight fraction assigned to
 434         * this group.
 435         */
 436        unsigned long h_load;
 437
 438        /*
 439         * this cpu's part of tg->shares
 440         */
 441        unsigned long shares;
 442
 443        /*
 444         * load.weight at the time we set shares
 445         */
 446        unsigned long rq_weight;
 447#endif
 448#endif
 449};
 450
 451/* Real-Time classes' related field in a runqueue: */
 452struct rt_rq {
 453        struct rt_prio_array active;
 454        unsigned long rt_nr_running;
 455#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 456        struct {
 457                int curr; /* highest queued rt task prio */
 458#ifdef CONFIG_SMP
 459                int next; /* next highest */
 460#endif
 461        } highest_prio;
 462#endif
 463#ifdef CONFIG_SMP
 464        unsigned long rt_nr_migratory;
 465        unsigned long rt_nr_total;
 466        int overloaded;
 467        struct plist_head pushable_tasks;
 468#endif
 469        int rt_throttled;
 470        u64 rt_time;
 471        u64 rt_runtime;
 472        /* Nests inside the rq lock: */
 473        raw_spinlock_t rt_runtime_lock;
 474
 475#ifdef CONFIG_RT_GROUP_SCHED
 476        unsigned long rt_nr_boosted;
 477
 478        struct rq *rq;
 479        struct list_head leaf_rt_rq_list;
 480        struct task_group *tg;
 481        struct sched_rt_entity *rt_se;
 482#endif
 483};
 484
 485#ifdef CONFIG_SMP
 486
 487/*
 488 * We add the notion of a root-domain which will be used to define per-domain
 489 * variables. Each exclusive cpuset essentially defines an island domain by
 490 * fully partitioning the member cpus from any other cpuset. Whenever a new
 491 * exclusive cpuset is created, we also create and attach a new root-domain
 492 * object.
 493 *
 494 */
 495struct root_domain {
 496        atomic_t refcount;
 497        cpumask_var_t span;
 498        cpumask_var_t online;
 499
 500        /*
 501         * The "RT overload" flag: it gets set if a CPU has more than
 502         * one runnable RT task.
 503         */
 504        cpumask_var_t rto_mask;
 505        atomic_t rto_count;
 506#ifdef CONFIG_SMP
 507        struct cpupri cpupri;
 508#endif
 509};
 510
 511/*
 512 * By default the system creates a single root-domain with all cpus as
 513 * members (mimicking the global state we have today).
 514 */
 515static struct root_domain def_root_domain;
 516
 517#endif
 518
 519/*
 520 * This is the main, per-CPU runqueue data structure.
 521 *
 522 * Locking rule: those places that want to lock multiple runqueues
 523 * (such as the load balancing or the thread migration code), lock
 524 * acquire operations must be ordered by ascending &runqueue.
 525 */
 526struct rq {
 527        /* runqueue lock: */
 528        raw_spinlock_t lock;
 529
 530        /*
 531         * nr_running and cpu_load should be in the same cacheline because
 532         * remote CPUs use both these fields when doing load calculation.
 533         */
 534        unsigned long nr_running;
 535        #define CPU_LOAD_IDX_MAX 5
 536        unsigned long cpu_load[CPU_LOAD_IDX_MAX];
 537#ifdef CONFIG_NO_HZ
 538        unsigned char in_nohz_recently;
 539#endif
 540        /* capture load from *all* tasks on this cpu: */
 541        struct load_weight load;
 542        unsigned long nr_load_updates;
 543        u64 nr_switches;
 544
 545        struct cfs_rq cfs;
 546        struct rt_rq rt;
 547
 548#ifdef CONFIG_FAIR_GROUP_SCHED
 549        /* list of leaf cfs_rq on this cpu: */
 550        struct list_head leaf_cfs_rq_list;
 551#endif
 552#ifdef CONFIG_RT_GROUP_SCHED
 553        struct list_head leaf_rt_rq_list;
 554#endif
 555
 556        /*
 557         * This is part of a global counter where only the total sum
 558         * over all CPUs matters. A task can increase this counter on
 559         * one CPU and if it got migrated afterwards it may decrease
 560         * it on another CPU. Always updated under the runqueue lock:
 561         */
 562        unsigned long nr_uninterruptible;
 563
 564        struct task_struct *curr, *idle;
 565        unsigned long next_balance;
 566        struct mm_struct *prev_mm;
 567
 568        u64 clock;
 569
 570        atomic_t nr_iowait;
 571
 572#ifdef CONFIG_SMP
 573        struct root_domain *rd;
 574        struct sched_domain *sd;
 575
 576        unsigned char idle_at_tick;
 577        /* For active balancing */
 578        int post_schedule;
 579        int active_balance;
 580        int push_cpu;
 581        /* cpu of this runqueue: */
 582        int cpu;
 583        int online;
 584
 585        unsigned long avg_load_per_task;
 586
 587        struct task_struct *migration_thread;
 588        struct list_head migration_queue;
 589
 590        u64 rt_avg;
 591        u64 age_stamp;
 592        u64 idle_stamp;
 593        u64 avg_idle;
 594#endif
 595
 596        /* calc_load related fields */
 597        unsigned long calc_load_update;
 598        long calc_load_active;
 599
 600#ifdef CONFIG_SCHED_HRTICK
 601#ifdef CONFIG_SMP
 602        int hrtick_csd_pending;
 603        struct call_single_data hrtick_csd;
 604#endif
 605        struct hrtimer hrtick_timer;
 606#endif
 607
 608#ifdef CONFIG_SCHEDSTATS
 609        /* latency stats */
 610        struct sched_info rq_sched_info;
 611        unsigned long long rq_cpu_time;
 612        /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
 613
 614        /* sys_sched_yield() stats */
 615        unsigned int yld_count;
 616
 617        /* schedule() stats */
 618        unsigned int sched_switch;
 619        unsigned int sched_count;
 620        unsigned int sched_goidle;
 621
 622        /* try_to_wake_up() stats */
 623        unsigned int ttwu_count;
 624        unsigned int ttwu_local;
 625
 626        /* BKL stats */
 627        unsigned int bkl_count;
 628#endif
 629};
 630
 631static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
 632
 633static inline
 634void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
 635{
 636        rq->curr->sched_class->check_preempt_curr(rq, p, flags);
 637}
 638
 639static inline int cpu_of(struct rq *rq)
 640{
 641#ifdef CONFIG_SMP
 642        return rq->cpu;
 643#else
 644        return 0;
 645#endif
 646}
 647
 648/*
 649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
 650 * See detach_destroy_domains: synchronize_sched for details.
 651 *
 652 * The domain tree of any CPU may only be accessed from within
 653 * preempt-disabled sections.
 654 */
 655#define for_each_domain(cpu, __sd) \
 656        for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
 657
 658#define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
 659#define this_rq()               (&__get_cpu_var(runqueues))
 660#define task_rq(p)              cpu_rq(task_cpu(p))
 661#define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
 662#define raw_rq()                (&__raw_get_cpu_var(runqueues))
 663
 664inline void update_rq_clock(struct rq *rq)
 665{
 666        rq->clock = sched_clock_cpu(cpu_of(rq));
 667}
 668
 669/*
 670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
 671 */
 672#ifdef CONFIG_SCHED_DEBUG
 673# define const_debug __read_mostly
 674#else
 675# define const_debug static const
 676#endif
 677
 678/**
 679 * runqueue_is_locked
 680 * @cpu: the processor in question.
 681 *
 682 * Returns true if the current cpu runqueue is locked.
 683 * This interface allows printk to be called with the runqueue lock
 684 * held and know whether or not it is OK to wake up the klogd.
 685 */
 686int runqueue_is_locked(int cpu)
 687{
 688        return raw_spin_is_locked(&cpu_rq(cpu)->lock);
 689}
 690
 691/*
 692 * Debugging: various feature bits
 693 */
 694
 695#define SCHED_FEAT(name, enabled)       \
 696        __SCHED_FEAT_##name ,
 697
 698enum {
 699#include "sched_features.h"
 700};
 701
 702#undef SCHED_FEAT
 703
 704#define SCHED_FEAT(name, enabled)       \
 705        (1UL << __SCHED_FEAT_##name) * enabled |
 706
 707const_debug unsigned int sysctl_sched_features =
 708#include "sched_features.h"
 709        0;
 710
 711#undef SCHED_FEAT
 712
 713#ifdef CONFIG_SCHED_DEBUG
 714#define SCHED_FEAT(name, enabled)       \
 715        #name ,
 716
 717static __read_mostly char *sched_feat_names[] = {
 718#include "sched_features.h"
 719        NULL
 720};
 721
 722#undef SCHED_FEAT
 723
 724static int sched_feat_show(struct seq_file *m, void *v)
 725{
 726        int i;
 727
 728        for (i = 0; sched_feat_names[i]; i++) {
 729                if (!(sysctl_sched_features & (1UL << i)))
 730                        seq_puts(m, "NO_");
 731                seq_printf(m, "%s ", sched_feat_names[i]);
 732        }
 733        seq_puts(m, "\n");
 734
 735        return 0;
 736}
 737
 738static ssize_t
 739sched_feat_write(struct file *filp, const char __user *ubuf,
 740                size_t cnt, loff_t *ppos)
 741{
 742        char buf[64];
 743        char *cmp = buf;
 744        int neg = 0;
 745        int i;
 746
 747        if (cnt > 63)
 748                cnt = 63;
 749
 750        if (copy_from_user(&buf, ubuf, cnt))
 751                return -EFAULT;
 752
 753        buf[cnt] = 0;
 754
 755        if (strncmp(buf, "NO_", 3) == 0) {
 756                neg = 1;
 757                cmp += 3;
 758        }
 759
 760        for (i = 0; sched_feat_names[i]; i++) {
 761                int len = strlen(sched_feat_names[i]);
 762
 763                if (strncmp(cmp, sched_feat_names[i], len) == 0) {
 764                        if (neg)
 765                                sysctl_sched_features &= ~(1UL << i);
 766                        else
 767                                sysctl_sched_features |= (1UL << i);
 768                        break;
 769                }
 770        }
 771
 772        if (!sched_feat_names[i])
 773                return -EINVAL;
 774
 775        *ppos += cnt;
 776
 777        return cnt;
 778}
 779
 780static int sched_feat_open(struct inode *inode, struct file *filp)
 781{
 782        return single_open(filp, sched_feat_show, NULL);
 783}
 784
 785static const struct file_operations sched_feat_fops = {
 786        .open           = sched_feat_open,
 787        .write          = sched_feat_write,
 788        .read           = seq_read,
 789        .llseek         = seq_lseek,
 790        .release        = single_release,
 791};
 792
 793static __init int sched_init_debug(void)
 794{
 795        debugfs_create_file("sched_features", 0644, NULL, NULL,
 796                        &sched_feat_fops);
 797
 798        return 0;
 799}
 800late_initcall(sched_init_debug);
 801
 802#endif
 803
 804#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
 805
 806/*
 807 * Number of tasks to iterate in a single balance run.
 808 * Limited because this is done with IRQs disabled.
 809 */
 810const_debug unsigned int sysctl_sched_nr_migrate = 32;
 811
 812/*
 813 * ratelimit for updating the group shares.
 814 * default: 0.25ms
 815 */
 816unsigned int sysctl_sched_shares_ratelimit = 250000;
 817unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
 818
 819/*
 820 * Inject some fuzzyness into changing the per-cpu group shares
 821 * this avoids remote rq-locks at the expense of fairness.
 822 * default: 4
 823 */
 824unsigned int sysctl_sched_shares_thresh = 4;
 825
 826/*
 827 * period over which we average the RT time consumption, measured
 828 * in ms.
 829 *
 830 * default: 1s
 831 */
 832const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
 833
 834/*
 835 * period over which we measure -rt task cpu usage in us.
 836 * default: 1s
 837 */
 838unsigned int sysctl_sched_rt_period = 1000000;
 839
 840static __read_mostly int scheduler_running;
 841
 842/*
 843 * part of the period that we allow rt tasks to run in us.
 844 * default: 0.95s
 845 */
 846int sysctl_sched_rt_runtime = 950000;
 847
 848static inline u64 global_rt_period(void)
 849{
 850        return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
 851}
 852
 853static inline u64 global_rt_runtime(void)
 854{
 855        if (sysctl_sched_rt_runtime < 0)
 856                return RUNTIME_INF;
 857
 858        return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
 859}
 860
 861#ifndef prepare_arch_switch
 862# define prepare_arch_switch(next)      do { } while (0)
 863#endif
 864#ifndef finish_arch_switch
 865# define finish_arch_switch(prev)       do { } while (0)
 866#endif
 867
 868static inline int task_current(struct rq *rq, struct task_struct *p)
 869{
 870        return rq->curr == p;
 871}
 872
 873#ifndef __ARCH_WANT_UNLOCKED_CTXSW
 874static inline int task_running(struct rq *rq, struct task_struct *p)
 875{
 876        return task_current(rq, p);
 877}
 878
 879static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
 880{
 881}
 882
 883static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
 884{
 885#ifdef CONFIG_DEBUG_SPINLOCK
 886        /* this is a valid case when another task releases the spinlock */
 887        rq->lock.owner = current;
 888#endif
 889        /*
 890         * If we are tracking spinlock dependencies then we have to
 891         * fix up the runqueue lock - which gets 'carried over' from
 892         * prev into current:
 893         */
 894        spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
 895
 896        raw_spin_unlock_irq(&rq->lock);
 897}
 898
 899#else /* __ARCH_WANT_UNLOCKED_CTXSW */
 900static inline int task_running(struct rq *rq, struct task_struct *p)
 901{
 902#ifdef CONFIG_SMP
 903        return p->oncpu;
 904#else
 905        return task_current(rq, p);
 906#endif
 907}
 908
 909static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
 910{
 911#ifdef CONFIG_SMP
 912        /*
 913         * We can optimise this out completely for !SMP, because the
 914         * SMP rebalancing from interrupt is the only thing that cares
 915         * here.
 916         */
 917        next->oncpu = 1;
 918#endif
 919#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
 920        raw_spin_unlock_irq(&rq->lock);
 921#else
 922        raw_spin_unlock(&rq->lock);
 923#endif
 924}
 925
 926static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
 927{
 928#ifdef CONFIG_SMP
 929        /*
 930         * After ->oncpu is cleared, the task can be moved to a different CPU.
 931         * We must ensure this doesn't happen until the switch is completely
 932         * finished.
 933         */
 934        smp_wmb();
 935        prev->oncpu = 0;
 936#endif
 937#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
 938        local_irq_enable();
 939#endif
 940}
 941#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
 942
 943/*
 944 * __task_rq_lock - lock the runqueue a given task resides on.
 945 * Must be called interrupts disabled.
 946 */
 947static inline struct rq *__task_rq_lock(struct task_struct *p)
 948        __acquires(rq->lock)
 949{
 950        for (;;) {
 951                struct rq *rq = task_rq(p);
 952                raw_spin_lock(&rq->lock);
 953                if (likely(rq == task_rq(p)))
 954                        return rq;
 955                raw_spin_unlock(&rq->lock);
 956        }
 957}
 958
 959/*
 960 * task_rq_lock - lock the runqueue a given task resides on and disable
 961 * interrupts. Note the ordering: we can safely lookup the task_rq without
 962 * explicitly disabling preemption.
 963 */
 964static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
 965        __acquires(rq->lock)
 966{
 967        struct rq *rq;
 968
 969        for (;;) {
 970                local_irq_save(*flags);
 971                rq = task_rq(p);
 972                raw_spin_lock(&rq->lock);
 973                if (likely(rq == task_rq(p)))
 974                        return rq;
 975                raw_spin_unlock_irqrestore(&rq->lock, *flags);
 976        }
 977}
 978
 979void task_rq_unlock_wait(struct task_struct *p)
 980{
 981        struct rq *rq = task_rq(p);
 982
 983        smp_mb(); /* spin-unlock-wait is not a full memory barrier */
 984        raw_spin_unlock_wait(&rq->lock);
 985}
 986
 987static void __task_rq_unlock(struct rq *rq)
 988        __releases(rq->lock)
 989{
 990        raw_spin_unlock(&rq->lock);
 991}
 992
 993static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
 994        __releases(rq->lock)
 995{
 996        raw_spin_unlock_irqrestore(&rq->lock, *flags);
 997}
 998
 999/*
1000 * this_rq_lock - lock this runqueue and disable interrupts.
1001 */
1002static struct rq *this_rq_lock(void)
1003        __acquires(rq->lock)
1004{
1005        struct rq *rq;
1006
1007        local_irq_disable();
1008        rq = this_rq();
1009        raw_spin_lock(&rq->lock);
1010
1011        return rq;
1012}
1013
1014#ifdef CONFIG_SCHED_HRTICK
1015/*
1016 * Use HR-timers to deliver accurate preemption points.
1017 *
1018 * Its all a bit involved since we cannot program an hrt while holding the
1019 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1020 * reschedule event.
1021 *
1022 * When we get rescheduled we reprogram the hrtick_timer outside of the
1023 * rq->lock.
1024 */
1025
1026/*
1027 * Use hrtick when:
1028 *  - enabled by features
1029 *  - hrtimer is actually high res
1030 */
1031static inline int hrtick_enabled(struct rq *rq)
1032{
1033        if (!sched_feat(HRTICK))
1034                return 0;
1035        if (!cpu_active(cpu_of(rq)))
1036                return 0;
1037        return hrtimer_is_hres_active(&rq->hrtick_timer);
1038}
1039
1040static void hrtick_clear(struct rq *rq)
1041{
1042        if (hrtimer_active(&rq->hrtick_timer))
1043                hrtimer_cancel(&rq->hrtick_timer);
1044}
1045
1046/*
1047 * High-resolution timer tick.
1048 * Runs from hardirq context with interrupts disabled.
1049 */
1050static enum hrtimer_restart hrtick(struct hrtimer *timer)
1051{
1052        struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1053
1054        WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1055
1056        raw_spin_lock(&rq->lock);
1057        update_rq_clock(rq);
1058        rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1059        raw_spin_unlock(&rq->lock);
1060
1061        return HRTIMER_NORESTART;
1062}
1063
1064#ifdef CONFIG_SMP
1065/*
1066 * called from hardirq (IPI) context
1067 */
1068static void __hrtick_start(void *arg)
1069{
1070        struct rq *rq = arg;
1071
1072        raw_spin_lock(&rq->lock);
1073        hrtimer_restart(&rq->hrtick_timer);
1074        rq->hrtick_csd_pending = 0;
1075        raw_spin_unlock(&rq->lock);
1076}
1077
1078/*
1079 * Called to set the hrtick timer state.
1080 *
1081 * called with rq->lock held and irqs disabled
1082 */
1083static void hrtick_start(struct rq *rq, u64 delay)
1084{
1085        struct hrtimer *timer = &rq->hrtick_timer;
1086        ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1087
1088        hrtimer_set_expires(timer, time);
1089
1090        if (rq == this_rq()) {
1091                hrtimer_restart(timer);
1092        } else if (!rq->hrtick_csd_pending) {
1093                __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1094                rq->hrtick_csd_pending = 1;
1095        }
1096}
1097
1098static int
1099hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1100{
1101        int cpu = (int)(long)hcpu;
1102
1103        switch (action) {
1104        case CPU_UP_CANCELED:
1105        case CPU_UP_CANCELED_FROZEN:
1106        case CPU_DOWN_PREPARE:
1107        case CPU_DOWN_PREPARE_FROZEN:
1108        case CPU_DEAD:
1109        case CPU_DEAD_FROZEN:
1110                hrtick_clear(cpu_rq(cpu));
1111                return NOTIFY_OK;
1112        }
1113
1114        return NOTIFY_DONE;
1115}
1116
1117static __init void init_hrtick(void)
1118{
1119        hotcpu_notifier(hotplug_hrtick, 0);
1120}
1121#else
1122/*
1123 * Called to set the hrtick timer state.
1124 *
1125 * called with rq->lock held and irqs disabled
1126 */
1127static void hrtick_start(struct rq *rq, u64 delay)
1128{
1129        __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1130                        HRTIMER_MODE_REL_PINNED, 0);
1131}
1132
1133static inline void init_hrtick(void)
1134{
1135}
1136#endif /* CONFIG_SMP */
1137
1138static void init_rq_hrtick(struct rq *rq)
1139{
1140#ifdef CONFIG_SMP
1141        rq->hrtick_csd_pending = 0;
1142
1143        rq->hrtick_csd.flags = 0;
1144        rq->hrtick_csd.func = __hrtick_start;
1145        rq->hrtick_csd.info = rq;
1146#endif
1147
1148        hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1149        rq->hrtick_timer.function = hrtick;
1150}
1151#else   /* CONFIG_SCHED_HRTICK */
1152static inline void hrtick_clear(struct rq *rq)
1153{
1154}
1155
1156static inline void init_rq_hrtick(struct rq *rq)
1157{
1158}
1159
1160static inline void init_hrtick(void)
1161{
1162}
1163#endif  /* CONFIG_SCHED_HRTICK */
1164
1165/*
1166 * resched_task - mark a task 'to be rescheduled now'.
1167 *
1168 * On UP this means the setting of the need_resched flag, on SMP it
1169 * might also involve a cross-CPU call to trigger the scheduler on
1170 * the target CPU.
1171 */
1172#ifdef CONFIG_SMP
1173
1174#ifndef tsk_is_polling
1175#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1176#endif
1177
1178static void resched_task(struct task_struct *p)
1179{
1180        int cpu;
1181
1182        assert_raw_spin_locked(&task_rq(p)->lock);
1183
1184        if (test_tsk_need_resched(p))
1185                return;
1186
1187        set_tsk_need_resched(p);
1188
1189        cpu = task_cpu(p);
1190        if (cpu == smp_processor_id())
1191                return;
1192
1193        /* NEED_RESCHED must be visible before we test polling */
1194        smp_mb();
1195        if (!tsk_is_polling(p))
1196                smp_send_reschedule(cpu);
1197}
1198
1199static void resched_cpu(int cpu)
1200{
1201        struct rq *rq = cpu_rq(cpu);
1202        unsigned long flags;
1203
1204        if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1205                return;
1206        resched_task(cpu_curr(cpu));
1207        raw_spin_unlock_irqrestore(&rq->lock, flags);
1208}
1209
1210#ifdef CONFIG_NO_HZ
1211/*
1212 * When add_timer_on() enqueues a timer into the timer wheel of an
1213 * idle CPU then this timer might expire before the next timer event
1214 * which is scheduled to wake up that CPU. In case of a completely
1215 * idle system the next event might even be infinite time into the
1216 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1217 * leaves the inner idle loop so the newly added timer is taken into
1218 * account when the CPU goes back to idle and evaluates the timer
1219 * wheel for the next timer event.
1220 */
1221void wake_up_idle_cpu(int cpu)
1222{
1223        struct rq *rq = cpu_rq(cpu);
1224
1225        if (cpu == smp_processor_id())
1226                return;
1227
1228        /*
1229         * This is safe, as this function is called with the timer
1230         * wheel base lock of (cpu) held. When the CPU is on the way
1231         * to idle and has not yet set rq->curr to idle then it will
1232         * be serialized on the timer wheel base lock and take the new
1233         * timer into account automatically.
1234         */
1235        if (rq->curr != rq->idle)
1236                return;
1237
1238        /*
1239         * We can set TIF_RESCHED on the idle task of the other CPU
1240         * lockless. The worst case is that the other CPU runs the
1241         * idle task through an additional NOOP schedule()
1242         */
1243        set_tsk_need_resched(rq->idle);
1244
1245        /* NEED_RESCHED must be visible before we test polling */
1246        smp_mb();
1247        if (!tsk_is_polling(rq->idle))
1248                smp_send_reschedule(cpu);
1249}
1250#endif /* CONFIG_NO_HZ */
1251
1252static u64 sched_avg_period(void)
1253{
1254        return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1255}
1256
1257static void sched_avg_update(struct rq *rq)
1258{
1259        s64 period = sched_avg_period();
1260
1261        while ((s64)(rq->clock - rq->age_stamp) > period) {
1262                rq->age_stamp += period;
1263                rq->rt_avg /= 2;
1264        }
1265}
1266
1267static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1268{
1269        rq->rt_avg += rt_delta;
1270        sched_avg_update(rq);
1271}
1272
1273#else /* !CONFIG_SMP */
1274static void resched_task(struct task_struct *p)
1275{
1276        assert_raw_spin_locked(&task_rq(p)->lock);
1277        set_tsk_need_resched(p);
1278}
1279
1280static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1281{
1282}
1283#endif /* CONFIG_SMP */
1284
1285#if BITS_PER_LONG == 32
1286# define WMULT_CONST    (~0UL)
1287#else
1288# define WMULT_CONST    (1UL << 32)
1289#endif
1290
1291#define WMULT_SHIFT     32
1292
1293/*
1294 * Shift right and round:
1295 */
1296#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1297
1298/*
1299 * delta *= weight / lw
1300 */
1301static unsigned long
1302calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1303                struct load_weight *lw)
1304{
1305        u64 tmp;
1306
1307        if (!lw->inv_weight) {
1308                if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1309                        lw->inv_weight = 1;
1310                else
1311                        lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1312                                / (lw->weight+1);
1313        }
1314
1315        tmp = (u64)delta_exec * weight;
1316        /*
1317         * Check whether we'd overflow the 64-bit multiplication:
1318         */
1319        if (unlikely(tmp > WMULT_CONST))
1320                tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1321                        WMULT_SHIFT/2);
1322        else
1323                tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1324
1325        return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1326}
1327
1328static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1329{
1330        lw->weight += inc;
1331        lw->inv_weight = 0;
1332}
1333
1334static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1335{
1336        lw->weight -= dec;
1337        lw->inv_weight = 0;
1338}
1339
1340/*
1341 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1342 * of tasks with abnormal "nice" values across CPUs the contribution that
1343 * each task makes to its run queue's load is weighted according to its
1344 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1345 * scaled version of the new time slice allocation that they receive on time
1346 * slice expiry etc.
1347 */
1348
1349#define WEIGHT_IDLEPRIO                3
1350#define WMULT_IDLEPRIO         1431655765
1351
1352/*
1353 * Nice levels are multiplicative, with a gentle 10% change for every
1354 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1355 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1356 * that remained on nice 0.
1357 *
1358 * The "10% effect" is relative and cumulative: from _any_ nice level,
1359 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1360 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1361 * If a task goes up by ~10% and another task goes down by ~10% then
1362 * the relative distance between them is ~25%.)
1363 */
1364static const int prio_to_weight[40] = {
1365 /* -20 */     88761,     71755,     56483,     46273,     36291,
1366 /* -15 */     29154,     23254,     18705,     14949,     11916,
1367 /* -10 */      9548,      7620,      6100,      4904,      3906,
1368 /*  -5 */      3121,      2501,      1991,      1586,      1277,
1369 /*   0 */      1024,       820,       655,       526,       423,
1370 /*   5 */       335,       272,       215,       172,       137,
1371 /*  10 */       110,        87,        70,        56,        45,
1372 /*  15 */        36,        29,        23,        18,        15,
1373};
1374
1375/*
1376 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1377 *
1378 * In cases where the weight does not change often, we can use the
1379 * precalculated inverse to speed up arithmetics by turning divisions
1380 * into multiplications:
1381 */
1382static const u32 prio_to_wmult[40] = {
1383 /* -20 */     48388,     59856,     76040,     92818,    118348,
1384 /* -15 */    147320,    184698,    229616,    287308,    360437,
1385 /* -10 */    449829,    563644,    704093,    875809,   1099582,
1386 /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
1387 /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
1388 /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
1389 /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
1390 /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1391};
1392
1393static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1394
1395/*
1396 * runqueue iterator, to support SMP load-balancing between different
1397 * scheduling classes, without having to expose their internal data
1398 * structures to the load-balancing proper:
1399 */
1400struct rq_iterator {
1401        void *arg;
1402        struct task_struct *(*start)(void *);
1403        struct task_struct *(*next)(void *);
1404};
1405
1406#ifdef CONFIG_SMP
1407static unsigned long
1408balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1409              unsigned long max_load_move, struct sched_domain *sd,
1410              enum cpu_idle_type idle, int *all_pinned,
1411              int *this_best_prio, struct rq_iterator *iterator);
1412
1413static int
1414iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1415                   struct sched_domain *sd, enum cpu_idle_type idle,
1416                   struct rq_iterator *iterator);
1417#endif
1418
1419/* Time spent by the tasks of the cpu accounting group executing in ... */
1420enum cpuacct_stat_index {
1421        CPUACCT_STAT_USER,      /* ... user mode */
1422        CPUACCT_STAT_SYSTEM,    /* ... kernel mode */
1423
1424        CPUACCT_STAT_NSTATS,
1425};
1426
1427#ifdef CONFIG_CGROUP_CPUACCT
1428static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1429static void cpuacct_update_stats(struct task_struct *tsk,
1430                enum cpuacct_stat_index idx, cputime_t val);
1431#else
1432static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1433static inline void cpuacct_update_stats(struct task_struct *tsk,
1434                enum cpuacct_stat_index idx, cputime_t val) {}
1435#endif
1436
1437static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1438{
1439        update_load_add(&rq->load, load);
1440}
1441
1442static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1443{
1444        update_load_sub(&rq->load, load);
1445}
1446
1447#if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1448typedef int (*tg_visitor)(struct task_group *, void *);
1449
1450/*
1451 * Iterate the full tree, calling @down when first entering a node and @up when
1452 * leaving it for the final time.
1453 */
1454static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1455{
1456        struct task_group *parent, *child;
1457        int ret;
1458
1459        rcu_read_lock();
1460        parent = &root_task_group;
1461down:
1462        ret = (*down)(parent, data);
1463        if (ret)
1464                goto out_unlock;
1465        list_for_each_entry_rcu(child, &parent->children, siblings) {
1466                parent = child;
1467                goto down;
1468
1469up:
1470                continue;
1471        }
1472        ret = (*up)(parent, data);
1473        if (ret)
1474                goto out_unlock;
1475
1476        child = parent;
1477        parent = parent->parent;
1478        if (parent)
1479                goto up;
1480out_unlock:
1481        rcu_read_unlock();
1482
1483        return ret;
1484}
1485
1486static int tg_nop(struct task_group *tg, void *data)
1487{
1488        return 0;
1489}
1490#endif
1491
1492#ifdef CONFIG_SMP
1493/* Used instead of source_load when we know the type == 0 */
1494static unsigned long weighted_cpuload(const int cpu)
1495{
1496        return cpu_rq(cpu)->load.weight;
1497}
1498
1499/*
1500 * Return a low guess at the load of a migration-source cpu weighted
1501 * according to the scheduling class and "nice" value.
1502 *
1503 * We want to under-estimate the load of migration sources, to
1504 * balance conservatively.
1505 */
1506static unsigned long source_load(int cpu, int type)
1507{
1508        struct rq *rq = cpu_rq(cpu);
1509        unsigned long total = weighted_cpuload(cpu);
1510
1511        if (type == 0 || !sched_feat(LB_BIAS))
1512                return total;
1513
1514        return min(rq->cpu_load[type-1], total);
1515}
1516
1517/*
1518 * Return a high guess at the load of a migration-target cpu weighted
1519 * according to the scheduling class and "nice" value.
1520 */
1521static unsigned long target_load(int cpu, int type)
1522{
1523        struct rq *rq = cpu_rq(cpu);
1524        unsigned long total = weighted_cpuload(cpu);
1525
1526        if (type == 0 || !sched_feat(LB_BIAS))
1527                return total;
1528
1529        return max(rq->cpu_load[type-1], total);
1530}
1531
1532static struct sched_group *group_of(int cpu)
1533{
1534        struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1535
1536        if (!sd)
1537                return NULL;
1538
1539        return sd->groups;
1540}
1541
1542static unsigned long power_of(int cpu)
1543{
1544        struct sched_group *group = group_of(cpu);
1545
1546        if (!group)
1547                return SCHED_LOAD_SCALE;
1548
1549        return group->cpu_power;
1550}
1551
1552static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1553
1554static unsigned long cpu_avg_load_per_task(int cpu)
1555{
1556        struct rq *rq = cpu_rq(cpu);
1557        unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1558
1559        if (nr_running)
1560                rq->avg_load_per_task = rq->load.weight / nr_running;
1561        else
1562                rq->avg_load_per_task = 0;
1563
1564        return rq->avg_load_per_task;
1565}
1566
1567#ifdef CONFIG_FAIR_GROUP_SCHED
1568
1569static __read_mostly unsigned long *update_shares_data;
1570
1571static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1572
1573/*
1574 * Calculate and set the cpu's group shares.
1575 */
1576static void update_group_shares_cpu(struct task_group *tg, int cpu,
1577                                    unsigned long sd_shares,
1578                                    unsigned long sd_rq_weight,
1579                                    unsigned long *usd_rq_weight)
1580{
1581        unsigned long shares, rq_weight;
1582        int boost = 0;
1583
1584        rq_weight = usd_rq_weight[cpu];
1585        if (!rq_weight) {
1586                boost = 1;
1587                rq_weight = NICE_0_LOAD;
1588        }
1589
1590        /*
1591         *             \Sum_j shares_j * rq_weight_i
1592         * shares_i =  -----------------------------
1593         *                  \Sum_j rq_weight_j
1594         */
1595        shares = (sd_shares * rq_weight) / sd_rq_weight;
1596        shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1597
1598        if (abs(shares - tg->se[cpu]->load.weight) >
1599                        sysctl_sched_shares_thresh) {
1600                struct rq *rq = cpu_rq(cpu);
1601                unsigned long flags;
1602
1603                raw_spin_lock_irqsave(&rq->lock, flags);
1604                tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1605                tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1606                __set_se_shares(tg->se[cpu], shares);
1607                raw_spin_unlock_irqrestore(&rq->lock, flags);
1608        }
1609}
1610
1611/*
1612 * Re-compute the task group their per cpu shares over the given domain.
1613 * This needs to be done in a bottom-up fashion because the rq weight of a
1614 * parent group depends on the shares of its child groups.
1615 */
1616static int tg_shares_up(struct task_group *tg, void *data)
1617{
1618        unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1619        unsigned long *usd_rq_weight;
1620        struct sched_domain *sd = data;
1621        unsigned long flags;
1622        int i;
1623
1624        if (!tg->se[0])
1625                return 0;
1626
1627        local_irq_save(flags);
1628        usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1629
1630        for_each_cpu(i, sched_domain_span(sd)) {
1631                weight = tg->cfs_rq[i]->load.weight;
1632                usd_rq_weight[i] = weight;
1633
1634                rq_weight += weight;
1635                /*
1636                 * If there are currently no tasks on the cpu pretend there
1637                 * is one of average load so that when a new task gets to
1638                 * run here it will not get delayed by group starvation.
1639                 */
1640                if (!weight)
1641                        weight = NICE_0_LOAD;
1642
1643                sum_weight += weight;
1644                shares += tg->cfs_rq[i]->shares;
1645        }
1646
1647        if (!rq_weight)
1648                rq_weight = sum_weight;
1649
1650        if ((!shares && rq_weight) || shares > tg->shares)
1651                shares = tg->shares;
1652
1653        if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1654                shares = tg->shares;
1655
1656        for_each_cpu(i, sched_domain_span(sd))
1657                update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1658
1659        local_irq_restore(flags);
1660
1661        return 0;
1662}
1663
1664/*
1665 * Compute the cpu's hierarchical load factor for each task group.
1666 * This needs to be done in a top-down fashion because the load of a child
1667 * group is a fraction of its parents load.
1668 */
1669static int tg_load_down(struct task_group *tg, void *data)
1670{
1671        unsigned long load;
1672        long cpu = (long)data;
1673
1674        if (!tg->parent) {
1675                load = cpu_rq(cpu)->load.weight;
1676        } else {
1677                load = tg->parent->cfs_rq[cpu]->h_load;
1678                load *= tg->cfs_rq[cpu]->shares;
1679                load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1680        }
1681
1682        tg->cfs_rq[cpu]->h_load = load;
1683
1684        return 0;
1685}
1686
1687static void update_shares(struct sched_domain *sd)
1688{
1689        s64 elapsed;
1690        u64 now;
1691
1692        if (root_task_group_empty())
1693                return;
1694
1695        now = cpu_clock(raw_smp_processor_id());
1696        elapsed = now - sd->last_update;
1697
1698        if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1699                sd->last_update = now;
1700                walk_tg_tree(tg_nop, tg_shares_up, sd);
1701        }
1702}
1703
1704static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1705{
1706        if (root_task_group_empty())
1707                return;
1708
1709        raw_spin_unlock(&rq->lock);
1710        update_shares(sd);
1711        raw_spin_lock(&rq->lock);
1712}
1713
1714static void update_h_load(long cpu)
1715{
1716        if (root_task_group_empty())
1717                return;
1718
1719        walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1720}
1721
1722#else
1723
1724static inline void update_shares(struct sched_domain *sd)
1725{
1726}
1727
1728static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1729{
1730}
1731
1732#endif
1733
1734#ifdef CONFIG_PREEMPT
1735
1736static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1737
1738/*
1739 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1740 * way at the expense of forcing extra atomic operations in all
1741 * invocations.  This assures that the double_lock is acquired using the
1742 * same underlying policy as the spinlock_t on this architecture, which
1743 * reduces latency compared to the unfair variant below.  However, it
1744 * also adds more overhead and therefore may reduce throughput.
1745 */
1746static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1747        __releases(this_rq->lock)
1748        __acquires(busiest->lock)
1749        __acquires(this_rq->lock)
1750{
1751        raw_spin_unlock(&this_rq->lock);
1752        double_rq_lock(this_rq, busiest);
1753
1754        return 1;
1755}
1756
1757#else
1758/*
1759 * Unfair double_lock_balance: Optimizes throughput at the expense of
1760 * latency by eliminating extra atomic operations when the locks are
1761 * already in proper order on entry.  This favors lower cpu-ids and will
1762 * grant the double lock to lower cpus over higher ids under contention,
1763 * regardless of entry order into the function.
1764 */
1765static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1766        __releases(this_rq->lock)
1767        __acquires(busiest->lock)
1768        __acquires(this_rq->lock)
1769{
1770        int ret = 0;
1771
1772        if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1773                if (busiest < this_rq) {
1774                        raw_spin_unlock(&this_rq->lock);
1775                        raw_spin_lock(&busiest->lock);
1776                        raw_spin_lock_nested(&this_rq->lock,
1777                                              SINGLE_DEPTH_NESTING);
1778                        ret = 1;
1779                } else
1780                        raw_spin_lock_nested(&busiest->lock,
1781                                              SINGLE_DEPTH_NESTING);
1782        }
1783        return ret;
1784}
1785
1786#endif /* CONFIG_PREEMPT */
1787
1788/*
1789 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1790 */
1791static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1792{
1793        if (unlikely(!irqs_disabled())) {
1794                /* printk() doesn't work good under rq->lock */
1795                raw_spin_unlock(&this_rq->lock);
1796                BUG_ON(1);
1797        }
1798
1799        return _double_lock_balance(this_rq, busiest);
1800}
1801
1802static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1803        __releases(busiest->lock)
1804{
1805        raw_spin_unlock(&busiest->lock);
1806        lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1807}
1808#endif
1809
1810#ifdef CONFIG_FAIR_GROUP_SCHED
1811static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1812{
1813#ifdef CONFIG_SMP
1814        cfs_rq->shares = shares;
1815#endif
1816}
1817#endif
1818
1819static void calc_load_account_active(struct rq *this_rq);
1820static void update_sysctl(void);
1821static int get_update_sysctl_factor(void);
1822
1823static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1824{
1825        set_task_rq(p, cpu);
1826#ifdef CONFIG_SMP
1827        /*
1828         * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1829         * successfuly executed on another CPU. We must ensure that updates of
1830         * per-task data have been completed by this moment.
1831         */
1832        smp_wmb();
1833        task_thread_info(p)->cpu = cpu;
1834#endif
1835}
1836
1837#include "sched_stats.h"
1838#include "sched_idletask.c"
1839#include "sched_fair.c"
1840#include "sched_rt.c"
1841#ifdef CONFIG_SCHED_DEBUG
1842# include "sched_debug.c"
1843#endif
1844
1845#define sched_class_highest (&rt_sched_class)
1846#define for_each_class(class) \
1847   for (class = sched_class_highest; class; class = class->next)
1848
1849static void inc_nr_running(struct rq *rq)
1850{
1851        rq->nr_running++;
1852}
1853
1854static void dec_nr_running(struct rq *rq)
1855{
1856        rq->nr_running--;
1857}
1858
1859static void set_load_weight(struct task_struct *p)
1860{
1861        if (task_has_rt_policy(p)) {
1862                p->se.load.weight = prio_to_weight[0] * 2;
1863                p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1864                return;
1865        }
1866
1867        /*
1868         * SCHED_IDLE tasks get minimal weight:
1869         */
1870        if (p->policy == SCHED_IDLE) {
1871                p->se.load.weight = WEIGHT_IDLEPRIO;
1872                p->se.load.inv_weight = WMULT_IDLEPRIO;
1873                return;
1874        }
1875
1876        p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1877        p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1878}
1879
1880static void update_avg(u64 *avg, u64 sample)
1881{
1882        s64 diff = sample - *avg;
1883        *avg += diff >> 3;
1884}
1885
1886static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1887{
1888        if (wakeup)
1889                p->se.start_runtime = p->se.sum_exec_runtime;
1890
1891        sched_info_queued(p);
1892        p->sched_class->enqueue_task(rq, p, wakeup);
1893        p->se.on_rq = 1;
1894}
1895
1896static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1897{
1898        if (sleep) {
1899                if (p->se.last_wakeup) {
1900                        update_avg(&p->se.avg_overlap,
1901                                p->se.sum_exec_runtime - p->se.last_wakeup);
1902                        p->se.last_wakeup = 0;
1903                } else {
1904                        update_avg(&p->se.avg_wakeup,
1905                                sysctl_sched_wakeup_granularity);
1906                }
1907        }
1908
1909        sched_info_dequeued(p);
1910        p->sched_class->dequeue_task(rq, p, sleep);
1911        p->se.on_rq = 0;
1912}
1913
1914/*
1915 * __normal_prio - return the priority that is based on the static prio
1916 */
1917static inline int __normal_prio(struct task_struct *p)
1918{
1919        return p->static_prio;
1920}
1921
1922/*
1923 * Calculate the expected normal priority: i.e. priority
1924 * without taking RT-inheritance into account. Might be
1925 * boosted by interactivity modifiers. Changes upon fork,
1926 * setprio syscalls, and whenever the interactivity
1927 * estimator recalculates.
1928 */
1929static inline int normal_prio(struct task_struct *p)
1930{
1931        int prio;
1932
1933        if (task_has_rt_policy(p))
1934                prio = MAX_RT_PRIO-1 - p->rt_priority;
1935        else
1936                prio = __normal_prio(p);
1937        return prio;
1938}
1939
1940/*
1941 * Calculate the current priority, i.e. the priority
1942 * taken into account by the scheduler. This value might
1943 * be boosted by RT tasks, or might be boosted by
1944 * interactivity modifiers. Will be RT if the task got
1945 * RT-boosted. If not then it returns p->normal_prio.
1946 */
1947static int effective_prio(struct task_struct *p)
1948{
1949        p->normal_prio = normal_prio(p);
1950        /*
1951         * If we are RT tasks or we were boosted to RT priority,
1952         * keep the priority unchanged. Otherwise, update priority
1953         * to the normal priority:
1954         */
1955        if (!rt_prio(p->prio))
1956                return p->normal_prio;
1957        return p->prio;
1958}
1959
1960/*
1961 * activate_task - move a task to the runqueue.
1962 */
1963static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1964{
1965        if (task_contributes_to_load(p))
1966                rq->nr_uninterruptible--;
1967
1968        enqueue_task(rq, p, wakeup);
1969        inc_nr_running(rq);
1970}
1971
1972/*
1973 * deactivate_task - remove a task from the runqueue.
1974 */
1975static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1976{
1977        if (task_contributes_to_load(p))
1978                rq->nr_uninterruptible++;
1979
1980        dequeue_task(rq, p, sleep);
1981        dec_nr_running(rq);
1982}
1983
1984/**
1985 * task_curr - is this task currently executing on a CPU?
1986 * @p: the task in question.
1987 */
1988inline int task_curr(const struct task_struct *p)
1989{
1990        return cpu_curr(task_cpu(p)) == p;
1991}
1992
1993static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1994                                       const struct sched_class *prev_class,
1995                                       int oldprio, int running)
1996{
1997        if (prev_class != p->sched_class) {
1998                if (prev_class->switched_from)
1999                        prev_class->switched_from(rq, p, running);
2000                p->sched_class->switched_to(rq, p, running);
2001        } else
2002                p->sched_class->prio_changed(rq, p, oldprio, running);
2003}
2004
2005#ifdef CONFIG_SMP
2006/*
2007 * Is this task likely cache-hot:
2008 */
2009static int
2010task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2011{
2012        s64 delta;
2013
2014        if (p->sched_class != &fair_sched_class)
2015                return 0;
2016
2017        /*
2018         * Buddy candidates are cache hot:
2019         */
2020        if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2021                        (&p->se == cfs_rq_of(&p->se)->next ||
2022                         &p->se == cfs_rq_of(&p->se)->last))
2023                return 1;
2024
2025        if (sysctl_sched_migration_cost == -1)
2026                return 1;
2027        if (sysctl_sched_migration_cost == 0)
2028                return 0;
2029
2030        delta = now - p->se.exec_start;
2031
2032        return delta < (s64)sysctl_sched_migration_cost;
2033}
2034
2035void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2036{
2037#ifdef CONFIG_SCHED_DEBUG
2038        /*
2039         * We should never call set_task_cpu() on a blocked task,
2040         * ttwu() will sort out the placement.
2041         */
2042        WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2043                        !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2044#endif
2045
2046        trace_sched_migrate_task(p, new_cpu);
2047
2048        if (task_cpu(p) != new_cpu) {
2049                p->se.nr_migrations++;
2050                perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2051        }
2052
2053        __set_task_cpu(p, new_cpu);
2054}
2055
2056struct migration_req {
2057        struct list_head list;
2058
2059        struct task_struct *task;
2060        int dest_cpu;
2061
2062        struct completion done;
2063};
2064
2065/*
2066 * The task's runqueue lock must be held.
2067 * Returns true if you have to wait for migration thread.
2068 */
2069static int
2070migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2071{
2072        struct rq *rq = task_rq(p);
2073
2074        /*
2075         * If the task is not on a runqueue (and not running), then
2076         * the next wake-up will properly place the task.
2077         */
2078        if (!p->se.on_rq && !task_running(rq, p))
2079                return 0;
2080
2081        init_completion(&req->done);
2082        req->task = p;
2083        req->dest_cpu = dest_cpu;
2084        list_add(&req->list, &rq->migration_queue);
2085
2086        return 1;
2087}
2088
2089/*
2090 * wait_task_context_switch -   wait for a thread to complete at least one
2091 *                              context switch.
2092 *
2093 * @p must not be current.
2094 */
2095void wait_task_context_switch(struct task_struct *p)
2096{
2097        unsigned long nvcsw, nivcsw, flags;
2098        int running;
2099        struct rq *rq;
2100
2101        nvcsw   = p->nvcsw;
2102        nivcsw  = p->nivcsw;
2103        for (;;) {
2104                /*
2105                 * The runqueue is assigned before the actual context
2106                 * switch. We need to take the runqueue lock.
2107                 *
2108                 * We could check initially without the lock but it is
2109                 * very likely that we need to take the lock in every
2110                 * iteration.
2111                 */
2112                rq = task_rq_lock(p, &flags);
2113                running = task_running(rq, p);
2114                task_rq_unlock(rq, &flags);
2115
2116                if (likely(!running))
2117                        break;
2118                /*
2119                 * The switch count is incremented before the actual
2120                 * context switch. We thus wait for two switches to be
2121                 * sure at least one completed.
2122                 */
2123                if ((p->nvcsw - nvcsw) > 1)
2124                        break;
2125                if ((p->nivcsw - nivcsw) > 1)
2126                        break;
2127
2128                cpu_relax();
2129        }
2130}
2131
2132/*
2133 * wait_task_inactive - wait for a thread to unschedule.
2134 *
2135 * If @match_state is nonzero, it's the @p->state value just checked and
2136 * not expected to change.  If it changes, i.e. @p might have woken up,
2137 * then return zero.  When we succeed in waiting for @p to be off its CPU,
2138 * we return a positive number (its total switch count).  If a second call
2139 * a short while later returns the same number, the caller can be sure that
2140 * @p has remained unscheduled the whole time.
2141 *
2142 * The caller must ensure that the task *will* unschedule sometime soon,
2143 * else this function might spin for a *long* time. This function can't
2144 * be called with interrupts off, or it may introduce deadlock with
2145 * smp_call_function() if an IPI is sent by the same process we are
2146 * waiting to become inactive.
2147 */
2148unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2149{
2150        unsigned long flags;
2151        int running, on_rq;
2152        unsigned long ncsw;
2153        struct rq *rq;
2154
2155        for (;;) {
2156                /*
2157                 * We do the initial early heuristics without holding
2158                 * any task-queue locks at all. We'll only try to get
2159                 * the runqueue lock when things look like they will
2160                 * work out!
2161                 */
2162                rq = task_rq(p);
2163
2164                /*
2165                 * If the task is actively running on another CPU
2166                 * still, just relax and busy-wait without holding
2167                 * any locks.
2168                 *
2169                 * NOTE! Since we don't hold any locks, it's not
2170                 * even sure that "rq" stays as the right runqueue!
2171                 * But we don't care, since "task_running()" will
2172                 * return false if the runqueue has changed and p
2173                 * is actually now running somewhere else!
2174                 */
2175                while (task_running(rq, p)) {
2176                        if (match_state && unlikely(p->state != match_state))
2177                                return 0;
2178                        cpu_relax();
2179                }
2180
2181                /*
2182                 * Ok, time to look more closely! We need the rq
2183                 * lock now, to be *sure*. If we're wrong, we'll
2184                 * just go back and repeat.
2185                 */
2186                rq = task_rq_lock(p, &flags);
2187                trace_sched_wait_task(rq, p);
2188                running = task_running(rq, p);
2189                on_rq = p->se.on_rq;
2190                ncsw = 0;
2191                if (!match_state || p->state == match_state)
2192                        ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2193                task_rq_unlock(rq, &flags);
2194
2195                /*
2196                 * If it changed from the expected state, bail out now.
2197                 */
2198                if (unlikely(!ncsw))
2199                        break;
2200
2201                /*
2202                 * Was it really running after all now that we
2203                 * checked with the proper locks actually held?
2204                 *
2205                 * Oops. Go back and try again..
2206                 */
2207                if (unlikely(running)) {
2208                        cpu_relax();
2209                        continue;
2210                }
2211
2212                /*
2213                 * It's not enough that it's not actively running,
2214                 * it must be off the runqueue _entirely_, and not
2215                 * preempted!
2216                 *
2217                 * So if it was still runnable (but just not actively
2218                 * running right now), it's preempted, and we should
2219                 * yield - it could be a while.
2220                 */
2221                if (unlikely(on_rq)) {
2222                        schedule_timeout_uninterruptible(1);
2223                        continue;
2224                }
2225
2226                /*
2227                 * Ahh, all good. It wasn't running, and it wasn't
2228                 * runnable, which means that it will never become
2229                 * running in the future either. We're all done!
2230                 */
2231                break;
2232        }
2233
2234        return ncsw;
2235}
2236
2237/***
2238 * kick_process - kick a running thread to enter/exit the kernel
2239 * @p: the to-be-kicked thread
2240 *
2241 * Cause a process which is running on another CPU to enter
2242 * kernel-mode, without any delay. (to get signals handled.)
2243 *
2244 * NOTE: this function doesnt have to take the runqueue lock,
2245 * because all it wants to ensure is that the remote task enters
2246 * the kernel. If the IPI races and the task has been migrated
2247 * to another CPU then no harm is done and the purpose has been
2248 * achieved as well.
2249 */
2250void kick_process(struct task_struct *p)
2251{
2252        int cpu;
2253
2254        preempt_disable();
2255        cpu = task_cpu(p);
2256        if ((cpu != smp_processor_id()) && task_curr(p))
2257                smp_send_reschedule(cpu);
2258        preempt_enable();
2259}
2260EXPORT_SYMBOL_GPL(kick_process);
2261#endif /* CONFIG_SMP */
2262
2263/**
2264 * task_oncpu_function_call - call a function on the cpu on which a task runs
2265 * @p:          the task to evaluate
2266 * @func:       the function to be called
2267 * @info:       the function call argument
2268 *
2269 * Calls the function @func when the task is currently running. This might
2270 * be on the current CPU, which just calls the function directly
2271 */
2272void task_oncpu_function_call(struct task_struct *p,
2273                              void (*func) (void *info), void *info)
2274{
2275        int cpu;
2276
2277        preempt_disable();
2278        cpu = task_cpu(p);
2279        if (task_curr(p))
2280                smp_call_function_single(cpu, func, info, 1);
2281        preempt_enable();
2282}
2283
2284#ifdef CONFIG_SMP
2285static int select_fallback_rq(int cpu, struct task_struct *p)
2286{
2287        int dest_cpu;
2288        const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2289
2290        /* Look for allowed, online CPU in same node. */
2291        for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2292                if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2293                        return dest_cpu;
2294
2295        /* Any allowed, online CPU? */
2296        dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2297        if (dest_cpu < nr_cpu_ids)
2298                return dest_cpu;
2299
2300        /* No more Mr. Nice Guy. */
2301        if (dest_cpu >= nr_cpu_ids) {
2302                rcu_read_lock();
2303                cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2304                rcu_read_unlock();
2305                dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2306
2307                /*
2308                 * Don't tell them about moving exiting tasks or
2309                 * kernel threads (both mm NULL), since they never
2310                 * leave kernel.
2311                 */
2312                if (p->mm && printk_ratelimit()) {
2313                        printk(KERN_INFO "process %d (%s) no "
2314                               "longer affine to cpu%d\n",
2315                               task_pid_nr(p), p->comm, cpu);
2316                }
2317        }
2318
2319        return dest_cpu;
2320}
2321
2322/*
2323 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2324 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2325 * by:
2326 *
2327 *  exec:           is unstable, retry loop
2328 *  fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2329 */
2330static inline
2331int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2332{
2333        int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2334
2335        /*
2336         * In order not to call set_task_cpu() on a blocking task we need
2337         * to rely on ttwu() to place the task on a valid ->cpus_allowed
2338         * cpu.
2339         *
2340         * Since this is common to all placement strategies, this lives here.
2341         *
2342         * [ this allows ->select_task() to simply return task_cpu(p) and
2343         *   not worry about this generic constraint ]
2344         */
2345        if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2346                     !cpu_online(cpu)))
2347                cpu = select_fallback_rq(task_cpu(p), p);
2348
2349        return cpu;
2350}
2351#endif
2352
2353/***
2354 * try_to_wake_up - wake up a thread
2355 * @p: the to-be-woken-up thread
2356 * @state: the mask of task states that can be woken
2357 * @sync: do a synchronous wakeup?
2358 *
2359 * Put it on the run-queue if it's not already there. The "current"
2360 * thread is always on the run-queue (except when the actual
2361 * re-schedule is in progress), and as such you're allowed to do
2362 * the simpler "current->state = TASK_RUNNING" to mark yourself
2363 * runnable without the overhead of this.
2364 *
2365 * returns failure only if the task is already active.
2366 */
2367static int try_to_wake_up(struct task_struct *p, unsigned int state,
2368                          int wake_flags)
2369{
2370        int cpu, orig_cpu, this_cpu, success = 0;
2371        unsigned long flags;
2372        struct rq *rq, *orig_rq;
2373
2374        if (!sched_feat(SYNC_WAKEUPS))
2375                wake_flags &= ~WF_SYNC;
2376
2377        this_cpu = get_cpu();
2378
2379        smp_wmb();
2380        rq = orig_rq = task_rq_lock(p, &flags);
2381        update_rq_clock(rq);
2382        if (!(p->state & state))
2383                goto out;
2384
2385        if (p->se.on_rq)
2386                goto out_running;
2387
2388        cpu = task_cpu(p);
2389        orig_cpu = cpu;
2390
2391#ifdef CONFIG_SMP
2392        if (unlikely(task_running(rq, p)))
2393                goto out_activate;
2394
2395        /*
2396         * In order to handle concurrent wakeups and release the rq->lock
2397         * we put the task in TASK_WAKING state.
2398         *
2399         * First fix up the nr_uninterruptible count:
2400         */
2401        if (task_contributes_to_load(p))
2402                rq->nr_uninterruptible--;
2403        p->state = TASK_WAKING;
2404
2405        if (p->sched_class->task_waking)
2406                p->sched_class->task_waking(rq, p);
2407
2408        __task_rq_unlock(rq);
2409
2410        cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2411        if (cpu != orig_cpu)
2412                set_task_cpu(p, cpu);
2413
2414        rq = __task_rq_lock(p);
2415        update_rq_clock(rq);
2416
2417        WARN_ON(p->state != TASK_WAKING);
2418        cpu = task_cpu(p);
2419
2420#ifdef CONFIG_SCHEDSTATS
2421        schedstat_inc(rq, ttwu_count);
2422        if (cpu == this_cpu)
2423                schedstat_inc(rq, ttwu_local);
2424        else {
2425                struct sched_domain *sd;
2426                for_each_domain(this_cpu, sd) {
2427                        if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2428                                schedstat_inc(sd, ttwu_wake_remote);
2429                                break;
2430                        }
2431                }
2432        }
2433#endif /* CONFIG_SCHEDSTATS */
2434
2435out_activate:
2436#endif /* CONFIG_SMP */
2437        schedstat_inc(p, se.nr_wakeups);
2438        if (wake_flags & WF_SYNC)
2439                schedstat_inc(p, se.nr_wakeups_sync);
2440        if (orig_cpu != cpu)
2441                schedstat_inc(p, se.nr_wakeups_migrate);
2442        if (cpu == this_cpu)
2443                schedstat_inc(p, se.nr_wakeups_local);
2444        else
2445                schedstat_inc(p, se.nr_wakeups_remote);
2446        activate_task(rq, p, 1);
2447        success = 1;
2448
2449        /*
2450         * Only attribute actual wakeups done by this task.
2451         */
2452        if (!in_interrupt()) {
2453                struct sched_entity *se = &current->se;
2454                u64 sample = se->sum_exec_runtime;
2455
2456                if (se->last_wakeup)
2457                        sample -= se->last_wakeup;
2458                else
2459                        sample -= se->start_runtime;
2460                update_avg(&se->avg_wakeup, sample);
2461
2462                se->last_wakeup = se->sum_exec_runtime;
2463        }
2464
2465out_running:
2466        trace_sched_wakeup(rq, p, success);
2467        check_preempt_curr(rq, p, wake_flags);
2468
2469        p->state = TASK_RUNNING;
2470#ifdef CONFIG_SMP
2471        if (p->sched_class->task_woken)
2472                p->sched_class->task_woken(rq, p);
2473
2474        if (unlikely(rq->idle_stamp)) {
2475                u64 delta = rq->clock - rq->idle_stamp;
2476                u64 max = 2*sysctl_sched_migration_cost;
2477
2478                if (delta > max)
2479                        rq->avg_idle = max;
2480                else
2481                        update_avg(&rq->avg_idle, delta);
2482                rq->idle_stamp = 0;
2483        }
2484#endif
2485out:
2486        task_rq_unlock(rq, &flags);
2487        put_cpu();
2488
2489        return success;
2490}
2491
2492/**
2493 * wake_up_process - Wake up a specific process
2494 * @p: The process to be woken up.
2495 *
2496 * Attempt to wake up the nominated process and move it to the set of runnable
2497 * processes.  Returns 1 if the process was woken up, 0 if it was already
2498 * running.
2499 *
2500 * It may be assumed that this function implies a write memory barrier before
2501 * changing the task state if and only if any tasks are woken up.
2502 */
2503int wake_up_process(struct task_struct *p)
2504{
2505        return try_to_wake_up(p, TASK_ALL, 0);
2506}
2507EXPORT_SYMBOL(wake_up_process);
2508
2509int wake_up_state(struct task_struct *p, unsigned int state)
2510{
2511        return try_to_wake_up(p, state, 0);
2512}
2513
2514/*
2515 * Perform scheduler related setup for a newly forked process p.
2516 * p is forked by current.
2517 *
2518 * __sched_fork() is basic setup used by init_idle() too:
2519 */
2520static void __sched_fork(struct task_struct *p)
2521{
2522        p->se.exec_start                = 0;
2523        p->se.sum_exec_runtime          = 0;
2524        p->se.prev_sum_exec_runtime     = 0;
2525        p->se.nr_migrations             = 0;
2526        p->se.last_wakeup               = 0;
2527        p->se.avg_overlap               = 0;
2528        p->se.start_runtime             = 0;
2529        p->se.avg_wakeup                = sysctl_sched_wakeup_granularity;
2530
2531#ifdef CONFIG_SCHEDSTATS
2532        p->se.wait_start                        = 0;
2533        p->se.wait_max                          = 0;
2534        p->se.wait_count                        = 0;
2535        p->se.wait_sum                          = 0;
2536
2537        p->se.sleep_start                       = 0;
2538        p->se.sleep_max                         = 0;
2539        p->se.sum_sleep_runtime                 = 0;
2540
2541        p->se.block_start                       = 0;
2542        p->se.block_max                         = 0;
2543        p->se.exec_max                          = 0;
2544        p->se.slice_max                         = 0;
2545
2546        p->se.nr_migrations_cold                = 0;
2547        p->se.nr_failed_migrations_affine       = 0;
2548        p->se.nr_failed_migrations_running      = 0;
2549        p->se.nr_failed_migrations_hot          = 0;
2550        p->se.nr_forced_migrations              = 0;
2551
2552        p->se.nr_wakeups                        = 0;
2553        p->se.nr_wakeups_sync                   = 0;
2554        p->se.nr_wakeups_migrate                = 0;
2555        p->se.nr_wakeups_local                  = 0;
2556        p->se.nr_wakeups_remote                 = 0;
2557        p->se.nr_wakeups_affine                 = 0;
2558        p->se.nr_wakeups_affine_attempts        = 0;
2559        p->se.nr_wakeups_passive                = 0;
2560        p->se.nr_wakeups_idle                   = 0;
2561
2562#endif
2563
2564        INIT_LIST_HEAD(&p->rt.run_list);
2565        p->se.on_rq = 0;
2566        INIT_LIST_HEAD(&p->se.group_node);
2567
2568#ifdef CONFIG_PREEMPT_NOTIFIERS
2569        INIT_HLIST_HEAD(&p->preempt_notifiers);
2570#endif
2571}
2572
2573/*
2574 * fork()/clone()-time setup:
2575 */
2576void sched_fork(struct task_struct *p, int clone_flags)
2577{
2578        int cpu = get_cpu();
2579
2580        __sched_fork(p);
2581        /*
2582         * We mark the process as waking here. This guarantees that
2583         * nobody will actually run it, and a signal or other external
2584         * event cannot wake it up and insert it on the runqueue either.
2585         */
2586        p->state = TASK_WAKING;
2587
2588        /*
2589         * Revert to default priority/policy on fork if requested.
2590         */
2591        if (unlikely(p->sched_reset_on_fork)) {
2592                if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2593                        p->policy = SCHED_NORMAL;
2594                        p->normal_prio = p->static_prio;
2595                }
2596
2597                if (PRIO_TO_NICE(p->static_prio) < 0) {
2598                        p->static_prio = NICE_TO_PRIO(0);
2599                        p->normal_prio = p->static_prio;
2600                        set_load_weight(p);
2601                }
2602
2603                /*
2604                 * We don't need the reset flag anymore after the fork. It has
2605                 * fulfilled its duty:
2606                 */
2607                p->sched_reset_on_fork = 0;
2608        }
2609
2610        /*
2611         * Make sure we do not leak PI boosting priority to the child.
2612         */
2613        p->prio = current->normal_prio;
2614
2615        if (!rt_prio(p->prio))
2616                p->sched_class = &fair_sched_class;
2617
2618        if (p->sched_class->task_fork)
2619                p->sched_class->task_fork(p);
2620
2621        set_task_cpu(p, cpu);
2622
2623#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2624        if (likely(sched_info_on()))
2625                memset(&p->sched_info, 0, sizeof(p->sched_info));
2626#endif
2627#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2628        p->oncpu = 0;
2629#endif
2630#ifdef CONFIG_PREEMPT
2631        /* Want to start with kernel preemption disabled. */
2632        task_thread_info(p)->preempt_count = 1;
2633#endif
2634        plist_node_init(&p->pushable_tasks, MAX_PRIO);
2635
2636        put_cpu();
2637}
2638
2639/*
2640 * wake_up_new_task - wake up a newly created task for the first time.
2641 *
2642 * This function will do some initial scheduler statistics housekeeping
2643 * that must be done for every newly created context, then puts the task
2644 * on the runqueue and wakes it.
2645 */
2646void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2647{
2648        unsigned long flags;
2649        struct rq *rq;
2650        int cpu = get_cpu();
2651
2652#ifdef CONFIG_SMP
2653        /*
2654         * Fork balancing, do it here and not earlier because:
2655         *  - cpus_allowed can change in the fork path
2656         *  - any previously selected cpu might disappear through hotplug
2657         *
2658         * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2659         * ->cpus_allowed is stable, we have preemption disabled, meaning
2660         * cpu_online_mask is stable.
2661         */
2662        cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2663        set_task_cpu(p, cpu);
2664#endif
2665
2666        rq = task_rq_lock(p, &flags);
2667        BUG_ON(p->state != TASK_WAKING);
2668        p->state = TASK_RUNNING;
2669        update_rq_clock(rq);
2670        activate_task(rq, p, 0);
2671        trace_sched_wakeup_new(rq, p, 1);
2672        check_preempt_curr(rq, p, WF_FORK);
2673#ifdef CONFIG_SMP
2674        if (p->sched_class->task_woken)
2675                p->sched_class->task_woken(rq, p);
2676#endif
2677        task_rq_unlock(rq, &flags);
2678        put_cpu();
2679}
2680
2681#ifdef CONFIG_PREEMPT_NOTIFIERS
2682
2683/**
2684 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2685 * @notifier: notifier struct to register
2686 */
2687void preempt_notifier_register(struct preempt_notifier *notifier)
2688{
2689        hlist_add_head(&notifier->link, &current->preempt_notifiers);
2690}
2691EXPORT_SYMBOL_GPL(preempt_notifier_register);
2692
2693/**
2694 * preempt_notifier_unregister - no longer interested in preemption notifications
2695 * @notifier: notifier struct to unregister
2696 *
2697 * This is safe to call from within a preemption notifier.
2698 */
2699void preempt_notifier_unregister(struct preempt_notifier *notifier)
2700{
2701        hlist_del(&notifier->link);
2702}
2703EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2704
2705static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2706{
2707        struct preempt_notifier *notifier;
2708        struct hlist_node *node;
2709
2710        hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2711                notifier->ops->sched_in(notifier, raw_smp_processor_id());
2712}
2713
2714static void
2715fire_sched_out_preempt_notifiers(struct task_struct *curr,
2716                                 struct task_struct *next)
2717{
2718        struct preempt_notifier *notifier;
2719        struct hlist_node *node;
2720
2721        hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2722                notifier->ops->sched_out(notifier, next);
2723}
2724
2725#else /* !CONFIG_PREEMPT_NOTIFIERS */
2726
2727static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2728{
2729}
2730
2731static void
2732fire_sched_out_preempt_notifiers(struct task_struct *curr,
2733                                 struct task_struct *next)
2734{
2735}
2736
2737#endif /* CONFIG_PREEMPT_NOTIFIERS */
2738
2739/**
2740 * prepare_task_switch - prepare to switch tasks
2741 * @rq: the runqueue preparing to switch
2742 * @prev: the current task that is being switched out
2743 * @next: the task we are going to switch to.
2744 *
2745 * This is called with the rq lock held and interrupts off. It must
2746 * be paired with a subsequent finish_task_switch after the context
2747 * switch.
2748 *
2749 * prepare_task_switch sets up locking and calls architecture specific
2750 * hooks.
2751 */
2752static inline void
2753prepare_task_switch(struct rq *rq, struct task_struct *prev,
2754                    struct task_struct *next)
2755{
2756        fire_sched_out_preempt_notifiers(prev, next);
2757        prepare_lock_switch(rq, next);
2758        prepare_arch_switch(next);
2759}
2760
2761/**
2762 * finish_task_switch - clean up after a task-switch
2763 * @rq: runqueue associated with task-switch
2764 * @prev: the thread we just switched away from.
2765 *
2766 * finish_task_switch must be called after the context switch, paired
2767 * with a prepare_task_switch call before the context switch.
2768 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2769 * and do any other architecture-specific cleanup actions.
2770 *
2771 * Note that we may have delayed dropping an mm in context_switch(). If
2772 * so, we finish that here outside of the runqueue lock. (Doing it
2773 * with the lock held can cause deadlocks; see schedule() for
2774 * details.)
2775 */
2776static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2777        __releases(rq->lock)
2778{
2779        struct mm_struct *mm = rq->prev_mm;
2780        long prev_state;
2781
2782        rq->prev_mm = NULL;
2783
2784        /*
2785         * A task struct has one reference for the use as "current".
2786         * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2787         * schedule one last time. The schedule call will never return, and
2788         * the scheduled task must drop that reference.
2789         * The test for TASK_DEAD must occur while the runqueue locks are
2790         * still held, otherwise prev could be scheduled on another cpu, die
2791         * there before we look at prev->state, and then the reference would
2792         * be dropped twice.
2793         *              Manfred Spraul <manfred@colorfullife.com>
2794         */
2795        prev_state = prev->state;
2796        finish_arch_switch(prev);
2797        perf_event_task_sched_in(current, cpu_of(rq));
2798        finish_lock_switch(rq, prev);
2799
2800        fire_sched_in_preempt_notifiers(current);
2801        if (mm)
2802                mmdrop(mm);
2803        if (unlikely(prev_state == TASK_DEAD)) {
2804                /*
2805                 * Remove function-return probe instances associated with this
2806                 * task and put them back on the free list.
2807                 */
2808                kprobe_flush_task(prev);
2809                put_task_struct(prev);
2810        }
2811}
2812
2813#ifdef CONFIG_SMP
2814
2815/* assumes rq->lock is held */
2816static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2817{
2818        if (prev->sched_class->pre_schedule)
2819                prev->sched_class->pre_schedule(rq, prev);
2820}
2821
2822/* rq->lock is NOT held, but preemption is disabled */
2823static inline void post_schedule(struct rq *rq)
2824{
2825        if (rq->post_schedule) {
2826                unsigned long flags;
2827
2828                raw_spin_lock_irqsave(&rq->lock, flags);
2829                if (rq->curr->sched_class->post_schedule)
2830                        rq->curr->sched_class->post_schedule(rq);
2831                raw_spin_unlock_irqrestore(&rq->lock, flags);
2832
2833                rq->post_schedule = 0;
2834        }
2835}
2836
2837#else
2838
2839static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2840{
2841}
2842
2843static inline void post_schedule(struct rq *rq)
2844{
2845}
2846
2847#endif
2848
2849/**
2850 * schedule_tail - first thing a freshly forked thread must call.
2851 * @prev: the thread we just switched away from.
2852 */
2853asmlinkage void schedule_tail(struct task_struct *prev)
2854        __releases(rq->lock)
2855{
2856        struct rq *rq = this_rq();
2857
2858        finish_task_switch(rq, prev);
2859
2860        /*
2861         * FIXME: do we need to worry about rq being invalidated by the
2862         * task_switch?
2863         */
2864        post_schedule(rq);
2865
2866#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2867        /* In this case, finish_task_switch does not reenable preemption */
2868        preempt_enable();
2869#endif
2870        if (current->set_child_tid)
2871                put_user(task_pid_vnr(current), current->set_child_tid);
2872}
2873
2874/*
2875 * context_switch - switch to the new MM and the new
2876 * thread's register state.
2877 */
2878static inline void
2879context_switch(struct rq *rq, struct task_struct *prev,
2880               struct task_struct *next)
2881{
2882        struct mm_struct *mm, *oldmm;
2883
2884        prepare_task_switch(rq, prev, next);
2885        trace_sched_switch(rq, prev, next);
2886        mm = next->mm;
2887        oldmm = prev->active_mm;
2888        /*
2889         * For paravirt, this is coupled with an exit in switch_to to
2890         * combine the page table reload and the switch backend into
2891         * one hypercall.
2892         */
2893        arch_start_context_switch(prev);
2894
2895        if (likely(!mm)) {
2896                next->active_mm = oldmm;
2897                atomic_inc(&oldmm->mm_count);
2898                enter_lazy_tlb(oldmm, next);
2899        } else
2900                switch_mm(oldmm, mm, next);
2901
2902        if (likely(!prev->mm)) {
2903                prev->active_mm = NULL;
2904                rq->prev_mm = oldmm;
2905        }
2906        /*
2907         * Since the runqueue lock will be released by the next
2908         * task (which is an invalid locking op but in the case
2909         * of the scheduler it's an obvious special-case), so we
2910         * do an early lockdep release here:
2911         */
2912#ifndef __ARCH_WANT_UNLOCKED_CTXSW
2913        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2914#endif
2915
2916        /* Here we just switch the register state and the stack. */
2917        switch_to(prev, next, prev);
2918
2919        barrier();
2920        /*
2921         * this_rq must be evaluated again because prev may have moved
2922         * CPUs since it called schedule(), thus the 'rq' on its stack
2923         * frame will be invalid.
2924         */
2925        finish_task_switch(this_rq(), prev);
2926}
2927
2928/*
2929 * nr_running, nr_uninterruptible and nr_context_switches:
2930 *
2931 * externally visible scheduler statistics: current number of runnable
2932 * threads, current number of uninterruptible-sleeping threads, total
2933 * number of context switches performed since bootup.
2934 */
2935unsigned long nr_running(void)
2936{
2937        unsigned long i, sum = 0;
2938
2939        for_each_online_cpu(i)
2940                sum += cpu_rq(i)->nr_running;
2941
2942        return sum;
2943}
2944
2945unsigned long nr_uninterruptible(void)
2946{
2947        unsigned long i, sum = 0;
2948
2949        for_each_possible_cpu(i)
2950                sum += cpu_rq(i)->nr_uninterruptible;
2951
2952        /*
2953         * Since we read the counters lockless, it might be slightly
2954         * inaccurate. Do not allow it to go below zero though:
2955         */
2956        if (unlikely((long)sum < 0))
2957                sum = 0;
2958
2959        return sum;
2960}
2961
2962unsigned long long nr_context_switches(void)
2963{
2964        int i;
2965        unsigned long long sum = 0;
2966
2967        for_each_possible_cpu(i)
2968                sum += cpu_rq(i)->nr_switches;
2969
2970        return sum;
2971}
2972
2973unsigned long nr_iowait(void)
2974{
2975        unsigned long i, sum = 0;
2976
2977        for_each_possible_cpu(i)
2978                sum += atomic_read(&cpu_rq(i)->nr_iowait);
2979
2980        return sum;
2981}
2982
2983unsigned long nr_iowait_cpu(void)
2984{
2985        struct rq *this = this_rq();
2986        return atomic_read(&this->nr_iowait);
2987}
2988
2989unsigned long this_cpu_load(void)
2990{
2991        struct rq *this = this_rq();
2992        return this->cpu_load[0];
2993}
2994
2995
2996/* Variables and functions for calc_load */
2997static atomic_long_t calc_load_tasks;
2998static unsigned long calc_load_update;
2999unsigned long avenrun[3];
3000EXPORT_SYMBOL(avenrun);
3001
3002/**
3003 * get_avenrun - get the load average array
3004 * @loads:      pointer to dest load array
3005 * @offset:     offset to add
3006 * @shift:      shift count to shift the result left
3007 *
3008 * These values are estimates at best, so no need for locking.
3009 */
3010void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3011{
3012        loads[0] = (avenrun[0] + offset) << shift;
3013        loads[1] = (avenrun[1] + offset) << shift;
3014        loads[2] = (avenrun[2] + offset) << shift;
3015}
3016
3017static unsigned long
3018calc_load(unsigned long load, unsigned long exp, unsigned long active)
3019{
3020        load *= exp;
3021        load += active * (FIXED_1 - exp);
3022        return load >> FSHIFT;
3023}
3024
3025/*
3026 * calc_load - update the avenrun load estimates 10 ticks after the
3027 * CPUs have updated calc_load_tasks.
3028 */
3029void calc_global_load(void)
3030{
3031        unsigned long upd = calc_load_update + 10;
3032        long active;
3033
3034        if (time_before(jiffies, upd))
3035                return;
3036
3037        active = atomic_long_read(&calc_load_tasks);
3038        active = active > 0 ? active * FIXED_1 : 0;
3039
3040        avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3041        avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3042        avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3043
3044        calc_load_update += LOAD_FREQ;
3045}
3046
3047/*
3048 * Either called from update_cpu_load() or from a cpu going idle
3049 */
3050static void calc_load_account_active(struct rq *this_rq)
3051{
3052        long nr_active, delta;
3053
3054        nr_active = this_rq->nr_running;
3055        nr_active += (long) this_rq->nr_uninterruptible;
3056
3057        if (nr_active != this_rq->calc_load_active) {
3058                delta = nr_active - this_rq->calc_load_active;
3059                this_rq->calc_load_active = nr_active;
3060                atomic_long_add(delta, &calc_load_tasks);
3061        }
3062}
3063
3064/*
3065 * Update rq->cpu_load[] statistics. This function is usually called every
3066 * scheduler tick (TICK_NSEC).
3067 */
3068static void update_cpu_load(struct rq *this_rq)
3069{
3070        unsigned long this_load = this_rq->load.weight;
3071        int i, scale;
3072
3073        this_rq->nr_load_updates++;
3074
3075        /* Update our load: */
3076        for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3077                unsigned long old_load, new_load;
3078
3079                /* scale is effectively 1 << i now, and >> i divides by scale */
3080
3081                old_load = this_rq->cpu_load[i];
3082                new_load = this_load;
3083                /*
3084                 * Round up the averaging division if load is increasing. This
3085                 * prevents us from getting stuck on 9 if the load is 10, for
3086                 * example.
3087                 */
3088                if (new_load > old_load)
3089                        new_load += scale-1;
3090                this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3091        }
3092
3093        if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3094                this_rq->calc_load_update += LOAD_FREQ;
3095                calc_load_account_active(this_rq);
3096        }
3097}
3098
3099#ifdef CONFIG_SMP
3100
3101/*
3102 * double_rq_lock - safely lock two runqueues
3103 *
3104 * Note this does not disable interrupts like task_rq_lock,
3105 * you need to do so manually before calling.
3106 */
3107static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3108        __acquires(rq1->lock)
3109        __acquires(rq2->lock)
3110{
3111        BUG_ON(!irqs_disabled());
3112        if (rq1 == rq2) {
3113                raw_spin_lock(&rq1->lock);
3114                __acquire(rq2->lock);   /* Fake it out ;) */
3115        } else {
3116                if (rq1 < rq2) {
3117                        raw_spin_lock(&rq1->lock);
3118                        raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3119                } else {
3120                        raw_spin_lock(&rq2->lock);
3121                        raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3122                }
3123        }
3124        update_rq_clock(rq1);
3125        update_rq_clock(rq2);
3126}
3127
3128/*
3129 * double_rq_unlock - safely unlock two runqueues
3130 *
3131 * Note this does not restore interrupts like task_rq_unlock,
3132 * you need to do so manually after calling.
3133 */
3134static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3135        __releases(rq1->lock)
3136        __releases(rq2->lock)
3137{
3138        raw_spin_unlock(&rq1->lock);
3139        if (rq1 != rq2)
3140                raw_spin_unlock(&rq2->lock);
3141        else
3142                __release(rq2->lock);
3143}
3144
3145/*
3146 * sched_exec - execve() is a valuable balancing opportunity, because at
3147 * this point the task has the smallest effective memory and cache footprint.
3148 */
3149void sched_exec(void)
3150{
3151        struct task_struct *p = current;
3152        struct migration_req req;
3153        int dest_cpu, this_cpu;
3154        unsigned long flags;
3155        struct rq *rq;
3156
3157again:
3158        this_cpu = get_cpu();
3159        dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3160        if (dest_cpu == this_cpu) {
3161                put_cpu();
3162                return;
3163        }
3164
3165        rq = task_rq_lock(p, &flags);
3166        put_cpu();
3167
3168        /*
3169         * select_task_rq() can race against ->cpus_allowed
3170         */
3171        if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3172            || unlikely(!cpu_active(dest_cpu))) {
3173                task_rq_unlock(rq, &flags);
3174                goto again;
3175        }
3176
3177        /* force the process onto the specified CPU */
3178        if (migrate_task(p, dest_cpu, &req)) {
3179                /* Need to wait for migration thread (might exit: take ref). */
3180                struct task_struct *mt = rq->migration_thread;
3181
3182                get_task_struct(mt);
3183                task_rq_unlock(rq, &flags);
3184                wake_up_process(mt);
3185                put_task_struct(mt);
3186                wait_for_completion(&req.done);
3187
3188                return;
3189        }
3190        task_rq_unlock(rq, &flags);
3191}
3192
3193/*
3194 * pull_task - move a task from a remote runqueue to the local runqueue.
3195 * Both runqueues must be locked.
3196 */
3197static void pull_task(struct rq *src_rq, struct task_struct *p,
3198                      struct rq *this_rq, int this_cpu)
3199{
3200        deactivate_task(src_rq, p, 0);
3201        set_task_cpu(p, this_cpu);
3202        activate_task(this_rq, p, 0);
3203        check_preempt_curr(this_rq, p, 0);
3204}
3205
3206/*
3207 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3208 */
3209static
3210int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3211                     struct sched_domain *sd, enum cpu_idle_type idle,
3212                     int *all_pinned)
3213{
3214        int tsk_cache_hot = 0;
3215        /*
3216         * We do not migrate tasks that are:
3217         * 1) running (obviously), or
3218         * 2) cannot be migrated to this CPU due to cpus_allowed, or
3219         * 3) are cache-hot on their current CPU.
3220         */
3221        if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3222                schedstat_inc(p, se.nr_failed_migrations_affine);
3223                return 0;
3224        }
3225        *all_pinned = 0;
3226
3227        if (task_running(rq, p)) {
3228                schedstat_inc(p, se.nr_failed_migrations_running);
3229                return 0;
3230        }
3231
3232        /*
3233         * Aggressive migration if:
3234         * 1) task is cache cold, or
3235         * 2) too many balance attempts have failed.
3236         */
3237
3238        tsk_cache_hot = task_hot(p, rq->clock, sd);
3239        if (!tsk_cache_hot ||
3240                sd->nr_balance_failed > sd->cache_nice_tries) {
3241#ifdef CONFIG_SCHEDSTATS
3242                if (tsk_cache_hot) {
3243                        schedstat_inc(sd, lb_hot_gained[idle]);
3244                        schedstat_inc(p, se.nr_forced_migrations);
3245                }
3246#endif
3247                return 1;
3248        }
3249
3250        if (tsk_cache_hot) {
3251                schedstat_inc(p, se.nr_failed_migrations_hot);
3252                return 0;
3253        }
3254        return 1;
3255}
3256
3257static unsigned long
3258balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3259              unsigned long max_load_move, struct sched_domain *sd,
3260              enum cpu_idle_type idle, int *all_pinned,
3261              int *this_best_prio, struct rq_iterator *iterator)
3262{
3263        int loops = 0, pulled = 0, pinned = 0;
3264        struct task_struct *p;
3265        long rem_load_move = max_load_move;
3266
3267        if (max_load_move == 0)
3268                goto out;
3269
3270        pinned = 1;
3271
3272        /*
3273         * Start the load-balancing iterator:
3274         */
3275        p = iterator->start(iterator->arg);
3276next:
3277        if (!p || loops++ > sysctl_sched_nr_migrate)
3278                goto out;
3279
3280        if ((p->se.load.weight >> 1) > rem_load_move ||
3281            !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3282                p = iterator->next(iterator->arg);
3283                goto next;
3284        }
3285
3286        pull_task(busiest, p, this_rq, this_cpu);
3287        pulled++;
3288        rem_load_move -= p->se.load.weight;
3289
3290#ifdef CONFIG_PREEMPT
3291        /*
3292         * NEWIDLE balancing is a source of latency, so preemptible kernels
3293         * will stop after the first task is pulled to minimize the critical
3294         * section.
3295         */
3296        if (idle == CPU_NEWLY_IDLE)
3297                goto out;
3298#endif
3299
3300        /*
3301         * We only want to steal up to the prescribed amount of weighted load.
3302         */
3303        if (rem_load_move > 0) {
3304                if (p->prio < *this_best_prio)
3305                        *this_best_prio = p->prio;
3306                p = iterator->next(iterator->arg);
3307                goto next;
3308        }
3309out:
3310        /*
3311         * Right now, this is one of only two places pull_task() is called,
3312         * so we can safely collect pull_task() stats here rather than
3313         * inside pull_task().
3314         */
3315        schedstat_add(sd, lb_gained[idle], pulled);
3316
3317        if (all_pinned)
3318                *all_pinned = pinned;
3319
3320        return max_load_move - rem_load_move;
3321}
3322
3323/*
3324 * move_tasks tries to move up to max_load_move weighted load from busiest to
3325 * this_rq, as part of a balancing operation within domain "sd".
3326 * Returns 1 if successful and 0 otherwise.
3327 *
3328 * Called with both runqueues locked.
3329 */
3330static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3331                      unsigned long max_load_move,
3332                      struct sched_domain *sd, enum cpu_idle_type idle,
3333                      int *all_pinned)
3334{
3335        const struct sched_class *class = sched_class_highest;
3336        unsigned long total_load_moved = 0;
3337        int this_best_prio = this_rq->curr->prio;
3338
3339        do {
3340                total_load_moved +=
3341                        class->load_balance(this_rq, this_cpu, busiest,
3342                                max_load_move - total_load_moved,
3343                                sd, idle, all_pinned, &this_best_prio);
3344                class = class->next;
3345
3346#ifdef CONFIG_PREEMPT
3347                /*
3348                 * NEWIDLE balancing is a source of latency, so preemptible
3349                 * kernels will stop after the first task is pulled to minimize
3350                 * the critical section.
3351                 */
3352                if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3353                        break;
3354#endif
3355        } while (class && max_load_move > total_load_moved);
3356
3357        return total_load_moved > 0;
3358}
3359
3360static int
3361iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3362                   struct sched_domain *sd, enum cpu_idle_type idle,
3363                   struct rq_iterator *iterator)
3364{
3365        struct task_struct *p = iterator->start(iterator->arg);
3366        int pinned = 0;
3367
3368        while (p) {
3369                if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3370                        pull_task(busiest, p, this_rq, this_cpu);
3371                        /*
3372                         * Right now, this is only the second place pull_task()
3373                         * is called, so we can safely collect pull_task()
3374                         * stats here rather than inside pull_task().
3375                         */
3376                        schedstat_inc(sd, lb_gained[idle]);
3377
3378                        return 1;
3379                }
3380                p = iterator->next(iterator->arg);
3381        }
3382
3383        return 0;
3384}
3385
3386/*
3387 * move_one_task tries to move exactly one task from busiest to this_rq, as
3388 * part of active balancing operations within "domain".
3389 * Returns 1 if successful and 0 otherwise.
3390 *
3391 * Called with both runqueues locked.
3392 */
3393static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3394                         struct sched_domain *sd, enum cpu_idle_type idle)
3395{
3396        const struct sched_class *class;
3397
3398        for_each_class(class) {
3399                if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3400                        return 1;
3401        }
3402
3403        return 0;
3404}
3405/********** Helpers for find_busiest_group ************************/
3406/*
3407 * sd_lb_stats - Structure to store the statistics of a sched_domain
3408 *              during load balancing.
3409 */
3410struct sd_lb_stats {
3411        struct sched_group *busiest; /* Busiest group in this sd */
3412        struct sched_group *this;  /* Local group in this sd */
3413        unsigned long total_load;  /* Total load of all groups in sd */
3414        unsigned long total_pwr;   /*   Total power of all groups in sd */
3415        unsigned long avg_load;    /* Average load across all groups in sd */
3416
3417        /** Statistics of this group */
3418        unsigned long this_load;
3419        unsigned long this_load_per_task;
3420        unsigned long this_nr_running;
3421
3422        /* Statistics of the busiest group */
3423        unsigned long max_load;
3424        unsigned long busiest_load_per_task;
3425        unsigned long busiest_nr_running;
3426
3427        int group_imb; /* Is there imbalance in this sd */
3428#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3429        int power_savings_balance; /* Is powersave balance needed for this sd */
3430        struct sched_group *group_min; /* Least loaded group in sd */
3431        struct sched_group *group_leader; /* Group which relieves group_min */
3432        unsigned long min_load_per_task; /* load_per_task in group_min */
3433        unsigned long leader_nr_running; /* Nr running of group_leader */
3434        unsigned long min_nr_running; /* Nr running of group_min */
3435#endif
3436};
3437
3438/*
3439 * sg_lb_stats - stats of a sched_group required for load_balancing
3440 */
3441struct sg_lb_stats {
3442        unsigned long avg_load; /*Avg load across the CPUs of the group */
3443        unsigned long group_load; /* Total load over the CPUs of the group */
3444        unsigned long sum_nr_running; /* Nr tasks running in the group */
3445        unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3446        unsigned long group_capacity;
3447        int group_imb; /* Is there an imbalance in the group ? */
3448};
3449
3450/**
3451 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3452 * @group: The group whose first cpu is to be returned.
3453 */
3454static inline unsigned int group_first_cpu(struct sched_group *group)
3455{
3456        return cpumask_first(sched_group_cpus(group));
3457}
3458
3459/**
3460 * get_sd_load_idx - Obtain the load index for a given sched domain.
3461 * @sd: The sched_domain whose load_idx is to be obtained.
3462 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3463 */
3464static inline int get_sd_load_idx(struct sched_domain *sd,
3465                                        enum cpu_idle_type idle)
3466{
3467        int load_idx;
3468
3469        switch (idle) {
3470        case CPU_NOT_IDLE:
3471                load_idx = sd->busy_idx;
3472                break;
3473
3474        case CPU_NEWLY_IDLE:
3475                load_idx = sd->newidle_idx;
3476                break;
3477        default:
3478                load_idx = sd->idle_idx;
3479                break;
3480        }
3481
3482        return load_idx;
3483}
3484
3485
3486#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3487/**
3488 * init_sd_power_savings_stats - Initialize power savings statistics for
3489 * the given sched_domain, during load balancing.
3490 *
3491 * @sd: Sched domain whose power-savings statistics are to be initialized.
3492 * @sds: Variable containing the statistics for sd.
3493 * @idle: Idle status of the CPU at which we're performing load-balancing.
3494 */
3495static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3496        struct sd_lb_stats *sds, enum cpu_idle_type idle)
3497{
3498        /*
3499         * Busy processors will not participate in power savings
3500         * balance.
3501         */
3502        if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3503                sds->power_savings_balance = 0;
3504        else {
3505                sds->power_savings_balance = 1;
3506                sds->min_nr_running = ULONG_MAX;
3507                sds->leader_nr_running = 0;
3508        }
3509}
3510
3511/**
3512 * update_sd_power_savings_stats - Update the power saving stats for a
3513 * sched_domain while performing load balancing.
3514 *
3515 * @group: sched_group belonging to the sched_domain under consideration.
3516 * @sds: Variable containing the statistics of the sched_domain
3517 * @local_group: Does group contain the CPU for which we're performing
3518 *              load balancing ?
3519 * @sgs: Variable containing the statistics of the group.
3520 */
3521static inline void update_sd_power_savings_stats(struct sched_group *group,
3522        struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3523{
3524
3525        if (!sds->power_savings_balance)
3526                return;
3527
3528        /*
3529         * If the local group is idle or completely loaded
3530         * no need to do power savings balance at this domain
3531         */
3532        if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3533                                !sds->this_nr_running))
3534                sds->power_savings_balance = 0;
3535
3536        /*
3537         * If a group is already running at full capacity or idle,
3538         * don't include that group in power savings calculations
3539         */
3540        if (!sds->power_savings_balance ||
3541                sgs->sum_nr_running >= sgs->group_capacity ||
3542                !sgs->sum_nr_running)
3543                return;
3544
3545        /*
3546         * Calculate the group which has the least non-idle load.
3547         * This is the group from where we need to pick up the load
3548         * for saving power
3549         */
3550        if ((sgs->sum_nr_running < sds->min_nr_running) ||
3551            (sgs->sum_nr_running == sds->min_nr_running &&
3552             group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3553                sds->group_min = group;
3554                sds->min_nr_running = sgs->sum_nr_running;
3555                sds->min_load_per_task = sgs->sum_weighted_load /
3556                                                sgs->sum_nr_running;
3557        }
3558
3559        /*
3560         * Calculate the group which is almost near its
3561         * capacity but still has some space to pick up some load
3562         * from other group and save more power
3563         */
3564        if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3565                return;
3566
3567        if (sgs->sum_nr_running > sds->leader_nr_running ||
3568            (sgs->sum_nr_running == sds->leader_nr_running &&
3569             group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3570                sds->group_leader = group;
3571                sds->leader_nr_running = sgs->sum_nr_running;
3572        }
3573}
3574
3575/**
3576 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3577 * @sds: Variable containing the statistics of the sched_domain
3578 *      under consideration.
3579 * @this_cpu: Cpu at which we're currently performing load-balancing.
3580 * @imbalance: Variable to store the imbalance.
3581 *
3582 * Description:
3583 * Check if we have potential to perform some power-savings balance.
3584 * If yes, set the busiest group to be the least loaded group in the
3585 * sched_domain, so that it's CPUs can be put to idle.
3586 *
3587 * Returns 1 if there is potential to perform power-savings balance.
3588 * Else returns 0.
3589 */
3590static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3591                                        int this_cpu, unsigned long *imbalance)
3592{
3593        if (!sds->power_savings_balance)
3594                return 0;
3595
3596        if (sds->this != sds->group_leader ||
3597                        sds->group_leader == sds->group_min)
3598                return 0;
3599
3600        *imbalance = sds->min_load_per_task;
3601        sds->busiest = sds->group_min;
3602
3603        return 1;
3604
3605}
3606#else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3607static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3608        struct sd_lb_stats *sds, enum cpu_idle_type idle)
3609{
3610        return;
3611}
3612
3613static inline void update_sd_power_savings_stats(struct sched_group *group,
3614        struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3615{
3616        return;
3617}
3618
3619static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3620                                        int this_cpu, unsigned long *imbalance)
3621{
3622        return 0;
3623}
3624#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3625
3626
3627unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3628{
3629        return SCHED_LOAD_SCALE;
3630}
3631
3632unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3633{
3634        return default_scale_freq_power(sd, cpu);
3635}
3636
3637unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3638{
3639        unsigned long weight = cpumask_weight(sched_domain_span(sd));
3640        unsigned long smt_gain = sd->smt_gain;
3641
3642        smt_gain /= weight;
3643
3644        return smt_gain;
3645}
3646
3647unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3648{
3649        return default_scale_smt_power(sd, cpu);
3650}
3651
3652unsigned long scale_rt_power(int cpu)
3653{
3654        struct rq *rq = cpu_rq(cpu);
3655        u64 total, available;
3656
3657        sched_avg_update(rq);
3658
3659        total = sched_avg_period() + (rq->clock - rq->age_stamp);
3660        available = total - rq->rt_avg;
3661
3662        if (unlikely((s64)total < SCHED_LOAD_SCALE))
3663                total = SCHED_LOAD_SCALE;
3664
3665        total >>= SCHED_LOAD_SHIFT;
3666
3667        return div_u64(available, total);
3668}
3669
3670static void update_cpu_power(struct sched_domain *sd, int cpu)
3671{
3672        unsigned long weight = cpumask_weight(sched_domain_span(sd));
3673        unsigned long power = SCHED_LOAD_SCALE;
3674        struct sched_group *sdg = sd->groups;
3675
3676        if (sched_feat(ARCH_POWER))
3677                power *= arch_scale_freq_power(sd, cpu);
3678        else
3679                power *= default_scale_freq_power(sd, cpu);
3680
3681        power >>= SCHED_LOAD_SHIFT;
3682
3683        if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3684                if (sched_feat(ARCH_POWER))
3685                        power *= arch_scale_smt_power(sd, cpu);
3686                else
3687                        power *= default_scale_smt_power(sd, cpu);
3688
3689                power >>= SCHED_LOAD_SHIFT;
3690        }
3691
3692        power *= scale_rt_power(cpu);
3693        power >>= SCHED_LOAD_SHIFT;
3694
3695        if (!power)
3696                power = 1;
3697
3698        sdg->cpu_power = power;
3699}
3700
3701static void update_group_power(struct sched_domain *sd, int cpu)
3702{
3703        struct sched_domain *child = sd->child;
3704        struct sched_group *group, *sdg = sd->groups;
3705        unsigned long power;
3706
3707        if (!child) {
3708                update_cpu_power(sd, cpu);
3709                return;
3710        }
3711
3712        power = 0;
3713
3714        group = child->groups;
3715        do {
3716                power += group->cpu_power;
3717                group = group->next;
3718        } while (group != child->groups);
3719
3720        sdg->cpu_power = power;
3721}
3722
3723/**
3724 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3725 * @sd: The sched_domain whose statistics are to be updated.
3726 * @group: sched_group whose statistics are to be updated.
3727 * @this_cpu: Cpu for which load balance is currently performed.
3728 * @idle: Idle status of this_cpu
3729 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3730 * @sd_idle: Idle status of the sched_domain containing group.
3731 * @local_group: Does group contain this_cpu.
3732 * @cpus: Set of cpus considered for load balancing.
3733 * @balance: Should we balance.
3734 * @sgs: variable to hold the statistics for this group.
3735 */
3736static inline void update_sg_lb_stats(struct sched_domain *sd,
3737                        struct sched_group *group, int this_cpu,
3738                        enum cpu_idle_type idle, int load_idx, int *sd_idle,
3739                        int local_group, const struct cpumask *cpus,
3740                        int *balance, struct sg_lb_stats *sgs)
3741{
3742        unsigned long load, max_cpu_load, min_cpu_load;
3743        int i;
3744        unsigned int balance_cpu = -1, first_idle_cpu = 0;
3745        unsigned long sum_avg_load_per_task;
3746        unsigned long avg_load_per_task;
3747
3748        if (local_group) {
3749                balance_cpu = group_first_cpu(group);
3750                if (balance_cpu == this_cpu)
3751                        update_group_power(sd, this_cpu);
3752        }
3753
3754        /* Tally up the load of all CPUs in the group */
3755        sum_avg_load_per_task = avg_load_per_task = 0;
3756        max_cpu_load = 0;
3757        min_cpu_load = ~0UL;
3758
3759        for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3760                struct rq *rq = cpu_rq(i);
3761
3762                if (*sd_idle && rq->nr_running)
3763                        *sd_idle = 0;
3764
3765                /* Bias balancing toward cpus of our domain */
3766                if (local_group) {
3767                        if (idle_cpu(i) && !first_idle_cpu) {
3768                                first_idle_cpu = 1;
3769                                balance_cpu = i;
3770                        }
3771
3772                        load = target_load(i, load_idx);
3773                } else {
3774                        load = source_load(i, load_idx);
3775                        if (load > max_cpu_load)
3776                                max_cpu_load = load;
3777                        if (min_cpu_load > load)
3778                                min_cpu_load = load;
3779                }
3780
3781                sgs->group_load += load;
3782                sgs->sum_nr_running += rq->nr_running;
3783                sgs->sum_weighted_load += weighted_cpuload(i);
3784
3785                sum_avg_load_per_task += cpu_avg_load_per_task(i);
3786        }
3787
3788        /*
3789         * First idle cpu or the first cpu(busiest) in this sched group
3790         * is eligible for doing load balancing at this and above
3791         * domains. In the newly idle case, we will allow all the cpu's
3792         * to do the newly idle load balance.
3793         */
3794        if (idle != CPU_NEWLY_IDLE && local_group &&
3795            balance_cpu != this_cpu && balance) {
3796                *balance = 0;
3797                return;
3798        }
3799
3800        /* Adjust by relative CPU power of the group */
3801        sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3802
3803
3804        /*
3805         * Consider the group unbalanced when the imbalance is larger
3806         * than the average weight of two tasks.
3807         *
3808         * APZ: with cgroup the avg task weight can vary wildly and
3809         *      might not be a suitable number - should we keep a
3810         *      normalized nr_running number somewhere that negates
3811         *      the hierarchy?
3812         */
3813        avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3814                group->cpu_power;
3815
3816        if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3817                sgs->group_imb = 1;
3818
3819        sgs->group_capacity =
3820                DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3821}
3822
3823/**
3824 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3825 * @sd: sched_domain whose statistics are to be updated.
3826 * @this_cpu: Cpu for which load balance is currently performed.
3827 * @idle: Idle status of this_cpu
3828 * @sd_idle: Idle status of the sched_domain containing group.
3829 * @cpus: Set of cpus considered for load balancing.
3830 * @balance: Should we balance.
3831 * @sds: variable to hold the statistics for this sched_domain.
3832 */
3833static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3834                        enum cpu_idle_type idle, int *sd_idle,
3835                        const struct cpumask *cpus, int *balance,
3836                        struct sd_lb_stats *sds)
3837{
3838        struct sched_domain *child = sd->child;
3839        struct sched_group *group = sd->groups;
3840        struct sg_lb_stats sgs;
3841        int load_idx, prefer_sibling = 0;
3842
3843        if (child && child->flags & SD_PREFER_SIBLING)
3844                prefer_sibling = 1;
3845
3846        init_sd_power_savings_stats(sd, sds, idle);
3847        load_idx = get_sd_load_idx(sd, idle);
3848
3849        do {
3850                int local_group;
3851
3852                local_group = cpumask_test_cpu(this_cpu,
3853                                               sched_group_cpus(group));
3854                memset(&sgs, 0, sizeof(sgs));
3855                update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3856                                local_group, cpus, balance, &sgs);
3857
3858                if (local_group && balance && !(*balance))
3859                        return;
3860
3861                sds->total_load += sgs.group_load;
3862                sds->total_pwr += group->cpu_power;
3863
3864                /*
3865                 * In case the child domain prefers tasks go to siblings
3866                 * first, lower the group capacity to one so that we'll try
3867                 * and move all the excess tasks away.
3868                 */
3869                if (prefer_sibling)
3870                        sgs.group_capacity = min(sgs.group_capacity, 1UL);
3871
3872                if (local_group) {
3873                        sds->this_load = sgs.avg_load;
3874                        sds->this = group;
3875                        sds->this_nr_running = sgs.sum_nr_running;
3876                        sds->this_load_per_task = sgs.sum_weighted_load;
3877                } else if (sgs.avg_load > sds->max_load &&
3878                           (sgs.sum_nr_running > sgs.group_capacity ||
3879                                sgs.group_imb)) {
3880                        sds->max_load = sgs.avg_load;
3881                        sds->busiest = group;
3882                        sds->busiest_nr_running = sgs.sum_nr_running;
3883                        sds->busiest_load_per_task = sgs.sum_weighted_load;
3884                        sds->group_imb = sgs.group_imb;
3885                }
3886
3887                update_sd_power_savings_stats(group, sds, local_group, &sgs);
3888                group = group->next;
3889        } while (group != sd->groups);
3890}
3891
3892/**
3893 * fix_small_imbalance - Calculate the minor imbalance that exists
3894 *                      amongst the groups of a sched_domain, during
3895 *                      load balancing.
3896 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3897 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3898 * @imbalance: Variable to store the imbalance.
3899 */
3900static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3901                                int this_cpu, unsigned long *imbalance)
3902{
3903        unsigned long tmp, pwr_now = 0, pwr_move = 0;
3904        unsigned int imbn = 2;
3905
3906        if (sds->this_nr_running) {
3907                sds->this_load_per_task /= sds->this_nr_running;
3908                if (sds->busiest_load_per_task >
3909                                sds->this_load_per_task)
3910                        imbn = 1;
3911        } else
3912                sds->this_load_per_task =
3913                        cpu_avg_load_per_task(this_cpu);
3914
3915        if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3916                        sds->busiest_load_per_task * imbn) {
3917                *imbalance = sds->busiest_load_per_task;
3918                return;
3919        }
3920
3921        /*
3922         * OK, we don't have enough imbalance to justify moving tasks,
3923         * however we may be able to increase total CPU power used by
3924         * moving them.
3925         */
3926
3927        pwr_now += sds->busiest->cpu_power *
3928                        min(sds->busiest_load_per_task, sds->max_load);
3929        pwr_now += sds->this->cpu_power *
3930                        min(sds->this_load_per_task, sds->this_load);
3931        pwr_now /= SCHED_LOAD_SCALE;
3932
3933        /* Amount of load we'd subtract */
3934        tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3935                sds->busiest->cpu_power;
3936        if (sds->max_load > tmp)
3937                pwr_move += sds->busiest->cpu_power *
3938                        min(sds->busiest_load_per_task, sds->max_load - tmp);
3939
3940        /* Amount of load we'd add */
3941        if (sds->max_load * sds->busiest->cpu_power <
3942                sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3943                tmp = (sds->max_load * sds->busiest->cpu_power) /
3944                        sds->this->cpu_power;
3945        else
3946                tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3947                        sds->this->cpu_power;
3948        pwr_move += sds->this->cpu_power *
3949                        min(sds->this_load_per_task, sds->this_load + tmp);
3950        pwr_move /= SCHED_LOAD_SCALE;
3951
3952        /* Move if we gain throughput */
3953        if (pwr_move > pwr_now)
3954                *imbalance = sds->busiest_load_per_task;
3955}
3956
3957/**
3958 * calculate_imbalance - Calculate the amount of imbalance present within the
3959 *                       groups of a given sched_domain during load balance.
3960 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3961 * @this_cpu: Cpu for which currently load balance is being performed.
3962 * @imbalance: The variable to store the imbalance.
3963 */
3964static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3965                unsigned long *imbalance)
3966{
3967        unsigned long max_pull;
3968        /*
3969         * In the presence of smp nice balancing, certain scenarios can have
3970         * max load less than avg load(as we skip the groups at or below
3971         * its cpu_power, while calculating max_load..)
3972         */
3973        if (sds->max_load < sds->avg_load) {
3974                *imbalance = 0;
3975                return fix_small_imbalance(sds, this_cpu, imbalance);
3976        }
3977
3978        /* Don't want to pull so many tasks that a group would go idle */
3979        max_pull = min(sds->max_load - sds->avg_load,
3980                        sds->max_load - sds->busiest_load_per_task);
3981
3982        /* How much load to actually move to equalise the imbalance */
3983        *imbalance = min(max_pull * sds->busiest->cpu_power,
3984                (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3985                        / SCHED_LOAD_SCALE;
3986
3987        /*
3988         * if *imbalance is less than the average load per runnable task
3989         * there is no gaurantee that any tasks will be moved so we'll have
3990         * a think about bumping its value to force at least one task to be
3991         * moved
3992         */
3993        if (*imbalance < sds->busiest_load_per_task)
3994                return fix_small_imbalance(sds, this_cpu, imbalance);
3995
3996}
3997/******* find_busiest_group() helpers end here *********************/
3998
3999/**
4000 * find_busiest_group - Returns the busiest group within the sched_domain
4001 * if there is an imbalance. If there isn't an imbalance, and
4002 * the user has opted for power-savings, it returns a group whose
4003 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4004 * such a group exists.
4005 *
4006 * Also calculates the amount of weighted load which should be moved
4007 * to restore balance.
4008 *
4009 * @sd: The sched_domain whose busiest group is to be returned.
4010 * @this_cpu: The cpu for which load balancing is currently being performed.
4011 * @imbalance: Variable which stores amount of weighted load which should
4012 *              be moved to restore balance/put a group to idle.
4013 * @idle: The idle status of this_cpu.
4014 * @sd_idle: The idleness of sd
4015 * @cpus: The set of CPUs under consideration for load-balancing.
4016 * @balance: Pointer to a variable indicating if this_cpu
4017 *      is the appropriate cpu to perform load balancing at this_level.
4018 *
4019 * Returns:     - the busiest group if imbalance exists.
4020 *              - If no imbalance and user has opted for power-savings balance,
4021 *                 return the least loaded group whose CPUs can be
4022 *                 put to idle by rebalancing its tasks onto our group.
4023 */
4024static struct sched_group *
4025find_busiest_group(struct sched_domain *sd, int this_cpu,
4026                   unsigned long *imbalance, enum cpu_idle_type idle,
4027                   int *sd_idle, const struct cpumask *cpus, int *balance)
4028{
4029        struct sd_lb_stats sds;
4030
4031        memset(&sds, 0, sizeof(sds));
4032
4033        /*
4034         * Compute the various statistics relavent for load balancing at
4035         * this level.
4036         */
4037        update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4038                                        balance, &sds);
4039
4040        /* Cases where imbalance does not exist from POV of this_cpu */
4041        /* 1) this_cpu is not the appropriate cpu to perform load balancing
4042         *    at this level.
4043         * 2) There is no busy sibling group to pull from.
4044         * 3) This group is the busiest group.
4045         * 4) This group is more busy than the avg busieness at this
4046         *    sched_domain.
4047         * 5) The imbalance is within the specified limit.
4048         * 6) Any rebalance would lead to ping-pong
4049         */
4050        if (balance && !(*balance))
4051                goto ret;
4052
4053        if (!sds.busiest || sds.busiest_nr_running == 0)
4054                goto out_balanced;
4055
4056        if (sds.this_load >= sds.max_load)
4057                goto out_balanced;
4058
4059        sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4060
4061        if (sds.this_load >= sds.avg_load)
4062                goto out_balanced;
4063
4064        if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4065                goto out_balanced;
4066
4067        sds.busiest_load_per_task /= sds.busiest_nr_running;
4068        if (sds.group_imb)
4069                sds.busiest_load_per_task =
4070                        min(sds.busiest_load_per_task, sds.avg_load);
4071
4072        /*
4073         * We're trying to get all the cpus to the average_load, so we don't
4074         * want to push ourselves above the average load, nor do we wish to
4075         * reduce the max loaded cpu below the average load, as either of these
4076         * actions would just result in more rebalancing later, and ping-pong
4077         * tasks around. Thus we look for the minimum possible imbalance.
4078         * Negative imbalances (*we* are more loaded than anyone else) will
4079         * be counted as no imbalance for these purposes -- we can't fix that
4080         * by pulling tasks to us. Be careful of negative numbers as they'll
4081         * appear as very large values with unsigned longs.
4082         */
4083        if (sds.max_load <= sds.busiest_load_per_task)
4084                goto out_balanced;
4085
4086        /* Looks like there is an imbalance. Compute it */
4087        calculate_imbalance(&sds, this_cpu, imbalance);
4088        return sds.busiest;
4089
4090out_balanced:
4091        /*
4092         * There is no obvious imbalance. But check if we can do some balancing
4093         * to save power.
4094         */
4095        if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4096                return sds.busiest;
4097ret:
4098        *imbalance = 0;
4099        return NULL;
4100}
4101
4102/*
4103 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4104 */
4105static struct rq *
4106find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4107                   unsigned long imbalance, const struct cpumask *cpus)
4108{
4109        struct rq *busiest = NULL, *rq;
4110        unsigned long max_load = 0;
4111        int i;
4112
4113        for_each_cpu(i, sched_group_cpus(group)) {
4114                unsigned long power = power_of(i);
4115                unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4116                unsigned long wl;
4117
4118                if (!cpumask_test_cpu(i, cpus))
4119                        continue;
4120
4121                rq = cpu_rq(i);
4122                wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4123                wl /= power;
4124
4125                if (capacity && rq->nr_running == 1 && wl > imbalance)
4126                        continue;
4127
4128                if (wl > max_load) {
4129                        max_load = wl;
4130                        busiest = rq;
4131                }
4132        }
4133
4134        return busiest;
4135}
4136
4137/*
4138 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4139 * so long as it is large enough.
4140 */
4141#define MAX_PINNED_INTERVAL     512
4142
4143/* Working cpumask for load_balance and load_balance_newidle. */
4144static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4145
4146/*
4147 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4148 * tasks if there is an imbalance.
4149 */
4150static int load_balance(int this_cpu, struct rq *this_rq,
4151                        struct sched_domain *sd, enum cpu_idle_type idle,
4152                        int *balance)
4153{
4154        int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4155        struct sched_group *group;
4156        unsigned long imbalance;
4157        struct rq *busiest;
4158        unsigned long flags;
4159        struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4160
4161        cpumask_copy(cpus, cpu_active_mask);
4162
4163        /*
4164         * When power savings policy is enabled for the parent domain, idle
4165         * sibling can pick up load irrespective of busy siblings. In this case,
4166         * let the state of idle sibling percolate up as CPU_IDLE, instead of
4167         * portraying it as CPU_NOT_IDLE.
4168         */
4169        if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4170            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4171                sd_idle = 1;
4172
4173        schedstat_inc(sd, lb_count[idle]);
4174
4175redo:
4176        update_shares(sd);
4177        group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4178                                   cpus, balance);
4179
4180        if (*balance == 0)
4181                goto out_balanced;
4182
4183        if (!group) {
4184                schedstat_inc(sd, lb_nobusyg[idle]);
4185                goto out_balanced;
4186        }
4187
4188        busiest = find_busiest_queue(group, idle, imbalance, cpus);
4189        if (!busiest) {
4190                schedstat_inc(sd, lb_nobusyq[idle]);
4191                goto out_balanced;
4192        }
4193
4194        BUG_ON(busiest == this_rq);
4195
4196        schedstat_add(sd, lb_imbalance[idle], imbalance);
4197
4198        ld_moved = 0;
4199        if (busiest->nr_running > 1) {
4200                /*
4201                 * Attempt to move tasks. If find_busiest_group has found
4202                 * an imbalance but busiest->nr_running <= 1, the group is
4203                 * still unbalanced. ld_moved simply stays zero, so it is
4204                 * correctly treated as an imbalance.
4205                 */
4206                local_irq_save(flags);
4207                double_rq_lock(this_rq, busiest);
4208                ld_moved = move_tasks(this_rq, this_cpu, busiest,
4209                                      imbalance, sd, idle, &all_pinned);
4210                double_rq_unlock(this_rq, busiest);
4211                local_irq_restore(flags);
4212
4213                /*
4214                 * some other cpu did the load balance for us.
4215                 */
4216                if (ld_moved && this_cpu != smp_processor_id())
4217                        resched_cpu(this_cpu);
4218
4219                /* All tasks on this runqueue were pinned by CPU affinity */
4220                if (unlikely(all_pinned)) {
4221                        cpumask_clear_cpu(cpu_of(busiest), cpus);
4222                        if (!cpumask_empty(cpus))
4223                                goto redo;
4224                        goto out_balanced;
4225                }
4226        }
4227
4228        if (!ld_moved) {
4229                schedstat_inc(sd, lb_failed[idle]);
4230                sd->nr_balance_failed++;
4231
4232                if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4233
4234                        raw_spin_lock_irqsave(&busiest->lock, flags);
4235
4236                        /* don't kick the migration_thread, if the curr
4237                         * task on busiest cpu can't be moved to this_cpu
4238                         */
4239                        if (!cpumask_test_cpu(this_cpu,
4240                                              &busiest->curr->cpus_allowed)) {
4241                                raw_spin_unlock_irqrestore(&busiest->lock,
4242                                                            flags);
4243                                all_pinned = 1;
4244                                goto out_one_pinned;
4245                        }
4246
4247                        if (!busiest->active_balance) {
4248                                busiest->active_balance = 1;
4249                                busiest->push_cpu = this_cpu;
4250                                active_balance = 1;
4251                        }
4252                        raw_spin_unlock_irqrestore(&busiest->lock, flags);
4253                        if (active_balance)
4254                                wake_up_process(busiest->migration_thread);
4255
4256                        /*
4257                         * We've kicked active balancing, reset the failure
4258                         * counter.
4259                         */
4260                        sd->nr_balance_failed = sd->cache_nice_tries+1;
4261                }
4262        } else
4263                sd->nr_balance_failed = 0;
4264
4265        if (likely(!active_balance)) {
4266                /* We were unbalanced, so reset the balancing interval */
4267                sd->balance_interval = sd->min_interval;
4268        } else {
4269                /*
4270                 * If we've begun active balancing, start to back off. This
4271                 * case may not be covered by the all_pinned logic if there
4272                 * is only 1 task on the busy runqueue (because we don't call
4273                 * move_tasks).
4274                 */
4275                if (sd->balance_interval < sd->max_interval)
4276                        sd->balance_interval *= 2;
4277        }
4278
4279        if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4280            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4281                ld_moved = -1;
4282
4283        goto out;
4284
4285out_balanced:
4286        schedstat_inc(sd, lb_balanced[idle]);
4287
4288        sd->nr_balance_failed = 0;
4289
4290out_one_pinned:
4291        /* tune up the balancing interval */
4292        if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4293                        (sd->balance_interval < sd->max_interval))
4294                sd->balance_interval *= 2;
4295
4296        if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4297            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4298                ld_moved = -1;
4299        else
4300                ld_moved = 0;
4301out:
4302        if (ld_moved)
4303                update_shares(sd);
4304        return ld_moved;
4305}
4306
4307/*
4308 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4309 * tasks if there is an imbalance.
4310 *
4311 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4312 * this_rq is locked.
4313 */
4314static int
4315load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4316{
4317        struct sched_group *group;
4318        struct rq *busiest = NULL;
4319        unsigned long imbalance;
4320        int ld_moved = 0;
4321        int sd_idle = 0;
4322        int all_pinned = 0;
4323        struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4324
4325        cpumask_copy(cpus, cpu_active_mask);
4326
4327        /*
4328         * When power savings policy is enabled for the parent domain, idle
4329         * sibling can pick up load irrespective of busy siblings. In this case,
4330         * let the state of idle sibling percolate up as IDLE, instead of
4331         * portraying it as CPU_NOT_IDLE.
4332         */
4333        if (sd->flags & SD_SHARE_CPUPOWER &&
4334            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4335                sd_idle = 1;
4336
4337        schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4338redo:
4339        update_shares_locked(this_rq, sd);
4340        group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4341                                   &sd_idle, cpus, NULL);
4342        if (!group) {
4343                schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4344                goto out_balanced;
4345        }
4346
4347        busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4348        if (!busiest) {
4349                schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4350                goto out_balanced;
4351        }
4352
4353        BUG_ON(busiest == this_rq);
4354
4355        schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4356
4357        ld_moved = 0;
4358        if (busiest->nr_running > 1) {
4359                /* Attempt to move tasks */
4360                double_lock_balance(this_rq, busiest);
4361                /* this_rq->clock is already updated */
4362                update_rq_clock(busiest);
4363                ld_moved = move_tasks(this_rq, this_cpu, busiest,
4364                                        imbalance, sd, CPU_NEWLY_IDLE,
4365                                        &all_pinned);
4366                double_unlock_balance(this_rq, busiest);
4367
4368                if (unlikely(all_pinned)) {
4369                        cpumask_clear_cpu(cpu_of(busiest), cpus);
4370                        if (!cpumask_empty(cpus))
4371                                goto redo;
4372                }
4373        }
4374
4375        if (!ld_moved) {
4376                int active_balance = 0;
4377
4378                schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4379                if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4380                    !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4381                        return -1;
4382
4383                if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4384                        return -1;
4385
4386                if (sd->nr_balance_failed++ < 2)
4387                        return -1;
4388
4389                /*
4390                 * The only task running in a non-idle cpu can be moved to this
4391                 * cpu in an attempt to completely freeup the other CPU
4392                 * package. The same method used to move task in load_balance()
4393                 * have been extended for load_balance_newidle() to speedup
4394                 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4395                 *
4396                 * The package power saving logic comes from
4397                 * find_busiest_group().  If there are no imbalance, then
4398                 * f_b_g() will return NULL.  However when sched_mc={1,2} then
4399                 * f_b_g() will select a group from which a running task may be
4400                 * pulled to this cpu in order to make the other package idle.
4401                 * If there is no opportunity to make a package idle and if
4402                 * there are no imbalance, then f_b_g() will return NULL and no
4403                 * action will be taken in load_balance_newidle().
4404                 *
4405                 * Under normal task pull operation due to imbalance, there
4406                 * will be more than one task in the source run queue and
4407                 * move_tasks() will succeed.  ld_moved will be true and this
4408                 * active balance code will not be triggered.
4409                 */
4410
4411                /* Lock busiest in correct order while this_rq is held */
4412                double_lock_balance(this_rq, busiest);
4413
4414                /*
4415                 * don't kick the migration_thread, if the curr
4416                 * task on busiest cpu can't be moved to this_cpu
4417                 */
4418                if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4419                        double_unlock_balance(this_rq, busiest);
4420                        all_pinned = 1;
4421                        return ld_moved;
4422                }
4423
4424                if (!busiest->active_balance) {
4425                        busiest->active_balance = 1;
4426                        busiest->push_cpu = this_cpu;
4427                        active_balance = 1;
4428                }
4429
4430                double_unlock_balance(this_rq, busiest);
4431                /*
4432                 * Should not call ttwu while holding a rq->lock
4433                 */
4434                raw_spin_unlock(&this_rq->lock);
4435                if (active_balance)
4436                        wake_up_process(busiest->migration_thread);
4437                raw_spin_lock(&this_rq->lock);
4438
4439        } else
4440                sd->nr_balance_failed = 0;
4441
4442        update_shares_locked(this_rq, sd);
4443        return ld_moved;
4444
4445out_balanced:
4446        schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4447        if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4448            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4449                return -1;
4450        sd->nr_balance_failed = 0;
4451
4452        return 0;
4453}
4454
4455/*
4456 * idle_balance is called by schedule() if this_cpu is about to become
4457 * idle. Attempts to pull tasks from other CPUs.
4458 */
4459static void idle_balance(int this_cpu, struct rq *this_rq)
4460{
4461        struct sched_domain *sd;
4462        int pulled_task = 0;
4463        unsigned long next_balance = jiffies + HZ;
4464
4465        this_rq->idle_stamp = this_rq->clock;
4466
4467        if (this_rq->avg_idle < sysctl_sched_migration_cost)
4468                return;
4469
4470        for_each_domain(this_cpu, sd) {
4471                unsigned long interval;
4472
4473                if (!(sd->flags & SD_LOAD_BALANCE))
4474                        continue;
4475
4476                if (sd->flags & SD_BALANCE_NEWIDLE)
4477                        /* If we've pulled tasks over stop searching: */
4478                        pulled_task = load_balance_newidle(this_cpu, this_rq,
4479                                                           sd);
4480
4481                interval = msecs_to_jiffies(sd->balance_interval);
4482                if (time_after(next_balance, sd->last_balance + interval))
4483                        next_balance = sd->last_balance + interval;
4484                if (pulled_task) {
4485                        this_rq->idle_stamp = 0;
4486                        break;
4487                }
4488        }
4489        if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4490                /*
4491                 * We are going idle. next_balance may be set based on
4492                 * a busy processor. So reset next_balance.
4493                 */
4494                this_rq->next_balance = next_balance;
4495        }
4496}
4497
4498/*
4499 * active_load_balance is run by migration threads. It pushes running tasks
4500 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4501 * running on each physical CPU where possible, and avoids physical /
4502 * logical imbalances.
4503 *
4504 * Called with busiest_rq locked.
4505 */
4506static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4507{
4508        int target_cpu = busiest_rq->push_cpu;
4509        struct sched_domain *sd;
4510        struct rq *target_rq;
4511
4512        /* Is there any task to move? */
4513        if (busiest_rq->nr_running <= 1)
4514                return;
4515
4516        target_rq = cpu_rq(target_cpu);
4517
4518        /*
4519         * This condition is "impossible", if it occurs
4520         * we need to fix it. Originally reported by
4521         * Bjorn Helgaas on a 128-cpu setup.
4522         */
4523        BUG_ON(busiest_rq == target_rq);
4524
4525        /* move a task from busiest_rq to target_rq */
4526        double_lock_balance(busiest_rq, target_rq);
4527        update_rq_clock(busiest_rq);
4528        update_rq_clock(target_rq);
4529
4530        /* Search for an sd spanning us and the target CPU. */
4531        for_each_domain(target_cpu, sd) {
4532                if ((sd->flags & SD_LOAD_BALANCE) &&
4533                    cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4534                                break;
4535        }
4536
4537        if (likely(sd)) {
4538                schedstat_inc(sd, alb_count);
4539
4540                if (move_one_task(target_rq, target_cpu, busiest_rq,
4541                                  sd, CPU_IDLE))
4542                        schedstat_inc(sd, alb_pushed);
4543                else
4544                        schedstat_inc(sd, alb_failed);
4545        }
4546        double_unlock_balance(busiest_rq, target_rq);
4547}
4548
4549#ifdef CONFIG_NO_HZ
4550static struct {
4551        atomic_t load_balancer;
4552        cpumask_var_t cpu_mask;
4553        cpumask_var_t ilb_grp_nohz_mask;
4554} nohz ____cacheline_aligned = {
4555        .load_balancer = ATOMIC_INIT(-1),
4556};
4557
4558int get_nohz_load_balancer(void)
4559{
4560        return atomic_read(&nohz.load_balancer);
4561}
4562
4563#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4564/**
4565 * lowest_flag_domain - Return lowest sched_domain containing flag.
4566 * @cpu:        The cpu whose lowest level of sched domain is to
4567 *              be returned.
4568 * @flag:       The flag to check for the lowest sched_domain
4569 *              for the given cpu.
4570 *
4571 * Returns the lowest sched_domain of a cpu which contains the given flag.
4572 */
4573static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4574{
4575        struct sched_domain *sd;
4576
4577        for_each_domain(cpu, sd)
4578                if (sd && (sd->flags & flag))
4579                        break;
4580
4581        return sd;
4582}
4583
4584/**
4585 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4586 * @cpu:        The cpu whose domains we're iterating over.
4587 * @sd:         variable holding the value of the power_savings_sd
4588 *              for cpu.
4589 * @flag:       The flag to filter the sched_domains to be iterated.
4590 *
4591 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4592 * set, starting from the lowest sched_domain to the highest.
4593 */
4594#define for_each_flag_domain(cpu, sd, flag) \
4595        for (sd = lowest_flag_domain(cpu, flag); \
4596                (sd && (sd->flags & flag)); sd = sd->parent)
4597
4598/**
4599 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4600 * @ilb_group:  group to be checked for semi-idleness
4601 *
4602 * Returns:     1 if the group is semi-idle. 0 otherwise.
4603 *
4604 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4605 * and atleast one non-idle CPU. This helper function checks if the given
4606 * sched_group is semi-idle or not.
4607 */
4608static inline int is_semi_idle_group(struct sched_group *ilb_group)
4609{
4610        cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4611                                        sched_group_cpus(ilb_group));
4612
4613        /*
4614         * A sched_group is semi-idle when it has atleast one busy cpu
4615         * and atleast one idle cpu.
4616         */
4617        if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4618                return 0;
4619
4620        if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4621                return 0;
4622
4623        return 1;
4624}
4625/**
4626 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4627 * @cpu:        The cpu which is nominating a new idle_load_balancer.
4628 *
4629 * Returns:     Returns the id of the idle load balancer if it exists,
4630 *              Else, returns >= nr_cpu_ids.
4631 *
4632 * This algorithm picks the idle load balancer such that it belongs to a
4633 * semi-idle powersavings sched_domain. The idea is to try and avoid
4634 * completely idle packages/cores just for the purpose of idle load balancing
4635 * when there are other idle cpu's which are better suited for that job.
4636 */
4637static int find_new_ilb(int cpu)
4638{
4639        struct sched_domain *sd;
4640        struct sched_group *ilb_group;
4641
4642        /*
4643         * Have idle load balancer selection from semi-idle packages only
4644         * when power-aware load balancing is enabled
4645         */
4646        if (!(sched_smt_power_savings || sched_mc_power_savings))
4647                goto out_done;
4648
4649        /*
4650         * Optimize for the case when we have no idle CPUs or only one
4651         * idle CPU. Don't walk the sched_domain hierarchy in such cases
4652         */
4653        if (cpumask_weight(nohz.cpu_mask) < 2)
4654                goto out_done;
4655
4656        for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4657                ilb_group = sd->groups;
4658
4659                do {
4660                        if (is_semi_idle_group(ilb_group))
4661                                return cpumask_first(nohz.ilb_grp_nohz_mask);
4662
4663                        ilb_group = ilb_group->next;
4664
4665                } while (ilb_group != sd->groups);
4666        }
4667
4668out_done:
4669        return cpumask_first(nohz.cpu_mask);
4670}
4671#else /*  (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4672static inline int find_new_ilb(int call_cpu)
4673{
4674        return cpumask_first(nohz.cpu_mask);
4675}
4676#endif
4677
4678/*
4679 * This routine will try to nominate the ilb (idle load balancing)
4680 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4681 * load balancing on behalf of all those cpus. If all the cpus in the system
4682 * go into this tickless mode, then there will be no ilb owner (as there is
4683 * no need for one) and all the cpus will sleep till the next wakeup event
4684 * arrives...
4685 *
4686 * For the ilb owner, tick is not stopped. And this tick will be used
4687 * for idle load balancing. ilb owner will still be part of
4688 * nohz.cpu_mask..
4689 *
4690 * While stopping the tick, this cpu will become the ilb owner if there
4691 * is no other owner. And will be the owner till that cpu becomes busy
4692 * or if all cpus in the system stop their ticks at which point
4693 * there is no need for ilb owner.
4694 *
4695 * When the ilb owner becomes busy, it nominates another owner, during the
4696 * next busy scheduler_tick()
4697 */
4698int select_nohz_load_balancer(int stop_tick)
4699{
4700        int cpu = smp_processor_id();
4701
4702        if (stop_tick) {
4703                cpu_rq(cpu)->in_nohz_recently = 1;
4704
4705                if (!cpu_active(cpu)) {
4706                        if (atomic_read(&nohz.load_balancer) != cpu)
4707                                return 0;
4708
4709                        /*
4710                         * If we are going offline and still the leader,
4711                         * give up!
4712                         */
4713                        if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4714                                BUG();
4715
4716                        return 0;
4717                }
4718
4719                cpumask_set_cpu(cpu, nohz.cpu_mask);
4720
4721                /* time for ilb owner also to sleep */
4722                if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4723                        if (atomic_read(&nohz.load_balancer) == cpu)
4724                                atomic_set(&nohz.load_balancer, -1);
4725                        return 0;
4726                }
4727
4728                if (atomic_read(&nohz.load_balancer) == -1) {
4729                        /* make me the ilb owner */
4730                        if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4731                                return 1;
4732                } else if (atomic_read(&nohz.load_balancer) == cpu) {
4733                        int new_ilb;
4734
4735                        if (!(sched_smt_power_savings ||
4736                                                sched_mc_power_savings))
4737                                return 1;
4738                        /*
4739                         * Check to see if there is a more power-efficient
4740                         * ilb.
4741                         */
4742                        new_ilb = find_new_ilb(cpu);
4743                        if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4744                                atomic_set(&nohz.load_balancer, -1);
4745                                resched_cpu(new_ilb);
4746                                return 0;
4747                        }
4748                        return 1;
4749                }
4750        } else {
4751                if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4752                        return 0;
4753
4754                cpumask_clear_cpu(cpu, nohz.cpu_mask);
4755
4756                if (atomic_read(&nohz.load_balancer) == cpu)
4757                        if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4758                                BUG();
4759        }
4760        return 0;
4761}
4762#endif
4763
4764static DEFINE_SPINLOCK(balancing);
4765
4766/*
4767 * It checks each scheduling domain to see if it is due to be balanced,
4768 * and initiates a balancing operation if so.
4769 *
4770 * Balancing parameters are set up in arch_init_sched_domains.
4771 */
4772static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4773{
4774        int balance = 1;
4775        struct rq *rq = cpu_rq(cpu);
4776        unsigned long interval;
4777        struct sched_domain *sd;
4778        /* Earliest time when we have to do rebalance again */
4779        unsigned long next_balance = jiffies + 60*HZ;
4780        int update_next_balance = 0;
4781        int need_serialize;
4782
4783        for_each_domain(cpu, sd) {
4784                if (!(sd->flags & SD_LOAD_BALANCE))
4785                        continue;
4786
4787                interval = sd->balance_interval;
4788                if (idle != CPU_IDLE)
4789                        interval *= sd->busy_factor;
4790
4791                /* scale ms to jiffies */
4792                interval = msecs_to_jiffies(interval);
4793                if (unlikely(!interval))
4794                        interval = 1;
4795                if (interval > HZ*NR_CPUS/10)
4796                        interval = HZ*NR_CPUS/10;
4797
4798                need_serialize = sd->flags & SD_SERIALIZE;
4799
4800                if (need_serialize) {
4801                        if (!spin_trylock(&balancing))
4802                                goto out;
4803                }
4804
4805                if (time_after_eq(jiffies, sd->last_balance + interval)) {
4806                        if (load_balance(cpu, rq, sd, idle, &balance)) {
4807                                /*
4808                                 * We've pulled tasks over so either we're no
4809                                 * longer idle, or one of our SMT siblings is
4810                                 * not idle.
4811                                 */
4812                                idle = CPU_NOT_IDLE;
4813                        }
4814                        sd->last_balance = jiffies;
4815                }
4816                if (need_serialize)
4817                        spin_unlock(&balancing);
4818out:
4819                if (time_after(next_balance, sd->last_balance + interval)) {
4820                        next_balance = sd->last_balance + interval;
4821                        update_next_balance = 1;
4822                }
4823
4824                /*
4825                 * Stop the load balance at this level. There is another
4826                 * CPU in our sched group which is doing load balancing more
4827                 * actively.
4828                 */
4829                if (!balance)
4830                        break;
4831        }
4832
4833        /*
4834         * next_balance will be updated only when there is a need.
4835         * When the cpu is attached to null domain for ex, it will not be
4836         * updated.
4837         */
4838        if (likely(update_next_balance))
4839                rq->next_balance = next_balance;
4840}
4841
4842/*
4843 * run_rebalance_domains is triggered when needed from the scheduler tick.
4844 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4845 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4846 */
4847static void run_rebalance_domains(struct softirq_action *h)
4848{
4849        int this_cpu = smp_processor_id();
4850        struct rq *this_rq = cpu_rq(this_cpu);
4851        enum cpu_idle_type idle = this_rq->idle_at_tick ?
4852                                                CPU_IDLE : CPU_NOT_IDLE;
4853
4854        rebalance_domains(this_cpu, idle);
4855
4856#ifdef CONFIG_NO_HZ
4857        /*
4858         * If this cpu is the owner for idle load balancing, then do the
4859         * balancing on behalf of the other idle cpus whose ticks are
4860         * stopped.
4861         */
4862        if (this_rq->idle_at_tick &&
4863            atomic_read(&nohz.load_balancer) == this_cpu) {
4864                struct rq *rq;
4865                int balance_cpu;
4866
4867                for_each_cpu(balance_cpu, nohz.cpu_mask) {
4868                        if (balance_cpu == this_cpu)
4869                                continue;
4870
4871                        /*
4872                         * If this cpu gets work to do, stop the load balancing
4873                         * work being done for other cpus. Next load
4874                         * balancing owner will pick it up.
4875                         */
4876                        if (need_resched())
4877                                break;
4878
4879                        rebalance_domains(balance_cpu, CPU_IDLE);
4880
4881                        rq = cpu_rq(balance_cpu);
4882                        if (time_after(this_rq->next_balance, rq->next_balance))
4883                                this_rq->next_balance = rq->next_balance;
4884                }
4885        }
4886#endif
4887}
4888
4889static inline int on_null_domain(int cpu)
4890{
4891        return !rcu_dereference(cpu_rq(cpu)->sd);
4892}
4893
4894/*
4895 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4896 *
4897 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4898 * idle load balancing owner or decide to stop the periodic load balancing,
4899 * if the whole system is idle.
4900 */
4901static inline void trigger_load_balance(struct rq *rq, int cpu)
4902{
4903#ifdef CONFIG_NO_HZ
4904        /*
4905         * If we were in the nohz mode recently and busy at the current
4906         * scheduler tick, then check if we need to nominate new idle
4907         * load balancer.
4908         */
4909        if (rq->in_nohz_recently && !rq->idle_at_tick) {
4910                rq->in_nohz_recently = 0;
4911
4912                if (atomic_read(&nohz.load_balancer) == cpu) {
4913                        cpumask_clear_cpu(cpu, nohz.cpu_mask);
4914                        atomic_set(&nohz.load_balancer, -1);
4915                }
4916
4917                if (atomic_read(&nohz.load_balancer) == -1) {
4918                        int ilb = find_new_ilb(cpu);
4919
4920                        if (ilb < nr_cpu_ids)
4921                                resched_cpu(ilb);
4922                }
4923        }
4924
4925        /*
4926         * If this cpu is idle and doing idle load balancing for all the
4927         * cpus with ticks stopped, is it time for that to stop?
4928         */
4929        if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4930            cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4931                resched_cpu(cpu);
4932                return;
4933        }
4934
4935        /*
4936         * If this cpu is idle and the idle load balancing is done by
4937         * someone else, then no need raise the SCHED_SOFTIRQ
4938         */
4939        if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4940            cpumask_test_cpu(cpu, nohz.cpu_mask))
4941                return;
4942#endif
4943        /* Don't need to rebalance while attached to NULL domain */
4944        if (time_after_eq(jiffies, rq->next_balance) &&
4945            likely(!on_null_domain(cpu)))
4946                raise_softirq(SCHED_SOFTIRQ);
4947}
4948
4949#else   /* CONFIG_SMP */
4950
4951/*
4952 * on UP we do not need to balance between CPUs:
4953 */
4954static inline void idle_balance(int cpu, struct rq *rq)
4955{
4956}
4957
4958#endif
4959
4960DEFINE_PER_CPU(struct kernel_stat, kstat);
4961
4962EXPORT_PER_CPU_SYMBOL(kstat);
4963
4964/*
4965 * Return any ns on the sched_clock that have not yet been accounted in
4966 * @p in case that task is currently running.
4967 *
4968 * Called with task_rq_lock() held on @rq.
4969 */
4970static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4971{
4972        u64 ns = 0;
4973
4974        if (task_current(rq, p)) {
4975                update_rq_clock(rq);
4976                ns = rq->clock - p->se.exec_start;
4977                if ((s64)ns < 0)
4978                        ns = 0;
4979        }
4980
4981        return ns;
4982}
4983
4984unsigned long long task_delta_exec(struct task_struct *p)
4985{
4986        unsigned long flags;
4987        struct rq *rq;
4988        u64 ns = 0;
4989
4990        rq = task_rq_lock(p, &flags);
4991        ns = do_task_delta_exec(p, rq);
4992        task_rq_unlock(rq, &flags);
4993
4994        return ns;
4995}
4996
4997/*
4998 * Return accounted runtime for the task.
4999 * In case the task is currently running, return the runtime plus current's
5000 * pending runtime that have not been accounted yet.
5001 */
5002unsigned long long task_sched_runtime(struct task_struct *p)
5003{
5004        unsigned long flags;
5005        struct rq *rq;
5006        u64 ns = 0;
5007
5008        rq = task_rq_lock(p, &flags);
5009        ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5010        task_rq_unlock(rq, &flags);
5011
5012        return ns;
5013}
5014
5015/*
5016 * Return sum_exec_runtime for the thread group.
5017 * In case the task is currently running, return the sum plus current's
5018 * pending runtime that have not been accounted yet.
5019 *
5020 * Note that the thread group might have other running tasks as well,
5021 * so the return value not includes other pending runtime that other
5022 * running tasks might have.
5023 */
5024unsigned long long thread_group_sched_runtime(struct task_struct *p)
5025{
5026        struct task_cputime totals;
5027        unsigned long flags;
5028        struct rq *rq;
5029        u64 ns;
5030
5031        rq = task_rq_lock(p, &flags);
5032        thread_group_cputime(p, &totals);
5033        ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5034        task_rq_unlock(rq, &flags);
5035
5036        return ns;
5037}
5038
5039/*
5040 * Account user cpu time to a process.
5041 * @p: the process that the cpu time gets accounted to
5042 * @cputime: the cpu time spent in user space since the last update
5043 * @cputime_scaled: cputime scaled by cpu frequency
5044 */
5045void account_user_time(struct task_struct *p, cputime_t cputime,
5046                       cputime_t cputime_scaled)
5047{
5048        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5049        cputime64_t tmp;
5050
5051        /* Add user time to process. */
5052        p->utime = cputime_add(p->utime, cputime);
5053        p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5054        account_group_user_time(p, cputime);
5055
5056        /* Add user time to cpustat. */
5057        tmp = cputime_to_cputime64(cputime);
5058        if (TASK_NICE(p) > 0)
5059                cpustat->nice = cputime64_add(cpustat->nice, tmp);
5060        else
5061                cpustat->user = cputime64_add(cpustat->user, tmp);
5062
5063        cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5064        /* Account for user time used */
5065        acct_update_integrals(p);
5066}
5067
5068/*
5069 * Account guest cpu time to a process.
5070 * @p: the process that the cpu time gets accounted to
5071 * @cputime: the cpu time spent in virtual machine since the last update
5072 * @cputime_scaled: cputime scaled by cpu frequency
5073 */
5074static void account_guest_time(struct task_struct *p, cputime_t cputime,
5075                               cputime_t cputime_scaled)
5076{
5077        cputime64_t tmp;
5078        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5079
5080        tmp = cputime_to_cputime64(cputime);
5081
5082        /* Add guest time to process. */
5083        p->utime = cputime_add(p->utime, cputime);
5084        p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5085        account_group_user_time(p, cputime);
5086        p->gtime = cputime_add(p->gtime, cputime);
5087
5088        /* Add guest time to cpustat. */
5089        if (TASK_NICE(p) > 0) {
5090                cpustat->nice = cputime64_add(cpustat->nice, tmp);
5091                cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5092        } else {
5093                cpustat->user = cputime64_add(cpustat->user, tmp);
5094                cpustat->guest = cputime64_add(cpustat->guest, tmp);
5095        }
5096}
5097
5098/*
5099 * Account system cpu time to a process.
5100 * @p: the process that the cpu time gets accounted to
5101 * @hardirq_offset: the offset to subtract from hardirq_count()
5102 * @cputime: the cpu time spent in kernel space since the last update
5103 * @cputime_scaled: cputime scaled by cpu frequency
5104 */
5105void account_system_time(struct task_struct *p, int hardirq_offset,
5106                         cputime_t cputime, cputime_t cputime_scaled)
5107{
5108        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5109        cputime64_t tmp;
5110
5111        if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5112                account_guest_time(p, cputime, cputime_scaled);
5113                return;
5114        }
5115
5116        /* Add system time to process. */
5117        p->stime = cputime_add(p->stime, cputime);
5118        p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5119        account_group_system_time(p, cputime);
5120
5121        /* Add system time to cpustat. */
5122        tmp = cputime_to_cputime64(cputime);
5123        if (hardirq_count() - hardirq_offset)
5124                cpustat->irq = cputime64_add(cpustat->irq, tmp);
5125        else if (softirq_count())
5126                cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5127        else
5128                cpustat->system = cputime64_add(cpustat->system, tmp);
5129
5130        cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5131
5132        /* Account for system time used */
5133        acct_update_integrals(p);
5134}
5135
5136/*
5137 * Account for involuntary wait time.
5138 * @steal: the cpu time spent in involuntary wait
5139 */
5140void account_steal_time(cputime_t cputime)
5141{
5142        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5143        cputime64_t cputime64 = cputime_to_cputime64(cputime);
5144
5145        cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5146}
5147
5148/*
5149 * Account for idle time.
5150 * @cputime: the cpu time spent in idle wait
5151 */
5152void account_idle_time(cputime_t cputime)
5153{
5154        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5155        cputime64_t cputime64 = cputime_to_cputime64(cputime);
5156        struct rq *rq = this_rq();
5157
5158        if (atomic_read(&rq->nr_iowait) > 0)
5159                cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5160        else
5161                cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5162}
5163
5164#ifndef CONFIG_VIRT_CPU_ACCOUNTING
5165
5166/*
5167 * Account a single tick of cpu time.
5168 * @p: the process that the cpu time gets accounted to
5169 * @user_tick: indicates if the tick is a user or a system tick
5170 */
5171void account_process_tick(struct task_struct *p, int user_tick)
5172{
5173        cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5174        struct rq *rq = this_rq();
5175
5176        if (user_tick)
5177                account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5178        else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5179                account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5180                                    one_jiffy_scaled);
5181        else
5182                account_idle_time(cputime_one_jiffy);
5183}
5184
5185/*
5186 * Account multiple ticks of steal time.
5187 * @p: the process from which the cpu time has been stolen
5188 * @ticks: number of stolen ticks
5189 */
5190void account_steal_ticks(unsigned long ticks)
5191{
5192        account_steal_time(jiffies_to_cputime(ticks));
5193}
5194
5195/*
5196 * Account multiple ticks of idle time.
5197 * @ticks: number of stolen ticks
5198 */
5199void account_idle_ticks(unsigned long ticks)
5200{
5201        account_idle_time(jiffies_to_cputime(ticks));
5202}
5203
5204#endif
5205
5206/*
5207 * Use precise platform statistics if available:
5208 */
5209#ifdef CONFIG_VIRT_CPU_ACCOUNTING
5210void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5211{
5212        *ut = p->utime;
5213        *st = p->stime;
5214}
5215
5216void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5217{
5218        struct task_cputime cputime;
5219
5220        thread_group_cputime(p, &cputime);
5221
5222        *ut = cputime.utime;
5223        *st = cputime.stime;
5224}
5225#else
5226
5227#ifndef nsecs_to_cputime
5228# define nsecs_to_cputime(__nsecs)      nsecs_to_jiffies(__nsecs)
5229#endif
5230
5231void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5232{
5233        cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5234
5235        /*
5236         * Use CFS's precise accounting:
5237         */
5238        rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5239
5240        if (total) {
5241                u64 temp;
5242
5243                temp = (u64)(rtime * utime);
5244                do_div(temp, total);
5245                utime = (cputime_t)temp;
5246        } else
5247                utime = rtime;
5248
5249        /*
5250         * Compare with previous values, to keep monotonicity:
5251         */
5252        p->prev_utime = max(p->prev_utime, utime);
5253        p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5254
5255        *ut = p->prev_utime;
5256        *st = p->prev_stime;
5257}
5258
5259/*
5260 * Must be called with siglock held.
5261 */
5262void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5263{
5264        struct signal_struct *sig = p->signal;
5265        struct task_cputime cputime;
5266        cputime_t rtime, utime, total;
5267
5268        thread_group_cputime(p, &cputime);
5269
5270        total = cputime_add(cputime.utime, cputime.stime);
5271        rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5272
5273        if (total) {
5274                u64 temp;
5275
5276                temp = (u64)(rtime * cputime.utime);
5277                do_div(temp, total);
5278                utime = (cputime_t)temp;
5279        } else
5280                utime = rtime;
5281
5282        sig->prev_utime = max(sig->prev_utime, utime);
5283        sig->prev_stime = max(sig->prev_stime,
5284                              cputime_sub(rtime, sig->prev_utime));
5285
5286        *ut = sig->prev_utime;
5287        *st = sig->prev_stime;
5288}
5289#endif
5290
5291/*
5292 * This function gets called by the timer code, with HZ frequency.
5293 * We call it with interrupts disabled.
5294 *
5295 * It also gets called by the fork code, when changing the parent's
5296 * timeslices.
5297 */
5298void scheduler_tick(void)
5299{
5300        int cpu = smp_processor_id();
5301        struct rq *rq = cpu_rq(cpu);
5302        struct task_struct *curr = rq->curr;
5303
5304        sched_clock_tick();
5305
5306        raw_spin_lock(&rq->lock);
5307        update_rq_clock(rq);
5308        update_cpu_load(rq);
5309        curr->sched_class->task_tick(rq, curr, 0);
5310        raw_spin_unlock(&rq->lock);
5311
5312        perf_event_task_tick(curr, cpu);
5313
5314#ifdef CONFIG_SMP
5315        rq->idle_at_tick = idle_cpu(cpu);
5316        trigger_load_balance(rq, cpu);
5317#endif
5318}
5319
5320notrace unsigned long get_parent_ip(unsigned long addr)
5321{
5322        if (in_lock_functions(addr)) {
5323                addr = CALLER_ADDR2;
5324                if (in_lock_functions(addr))
5325                        addr = CALLER_ADDR3;
5326        }
5327        return addr;
5328}
5329
5330#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5331                                defined(CONFIG_PREEMPT_TRACER))
5332
5333void __kprobes add_preempt_count(int val)
5334{
5335#ifdef CONFIG_DEBUG_PREEMPT
5336        /*
5337         * Underflow?
5338         */
5339        if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5340                return;
5341#endif
5342        preempt_count() += val;
5343#ifdef CONFIG_DEBUG_PREEMPT
5344        /*
5345         * Spinlock count overflowing soon?
5346         */
5347        DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5348                                PREEMPT_MASK - 10);
5349#endif
5350        if (preempt_count() == val)
5351                trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5352}
5353EXPORT_SYMBOL(add_preempt_count);
5354
5355void __kprobes sub_preempt_count(int val)
5356{
5357#ifdef CONFIG_DEBUG_PREEMPT
5358        /*
5359         * Underflow?
5360         */
5361        if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5362                return;
5363        /*
5364         * Is the spinlock portion underflowing?
5365         */
5366        if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5367                        !(preempt_count() & PREEMPT_MASK)))
5368                return;
5369#endif
5370
5371        if (preempt_count() == val)
5372                trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5373        preempt_count() -= val;
5374}
5375EXPORT_SYMBOL(sub_preempt_count);
5376
5377#endif
5378
5379/*
5380 * Print scheduling while atomic bug:
5381 */
5382static noinline void __schedule_bug(struct task_struct *prev)
5383{
5384        struct pt_regs *regs = get_irq_regs();
5385
5386        printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5387                prev->comm, prev->pid, preempt_count());
5388
5389        debug_show_held_locks(prev);
5390        print_modules();
5391        if (irqs_disabled())
5392                print_irqtrace_events(prev);
5393
5394        if (regs)
5395                show_regs(regs);
5396        else
5397                dump_stack();
5398}
5399
5400/*
5401 * Various schedule()-time debugging checks and statistics:
5402 */
5403static inline void schedule_debug(struct task_struct *prev)
5404{
5405        /*
5406         * Test if we are atomic. Since do_exit() needs to call into
5407         * schedule() atomically, we ignore that path for now.
5408         * Otherwise, whine if we are scheduling when we should not be.
5409         */
5410        if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5411                __schedule_bug(prev);
5412
5413        profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5414
5415        schedstat_inc(this_rq(), sched_count);
5416#ifdef CONFIG_SCHEDSTATS
5417        if (unlikely(prev->lock_depth >= 0)) {
5418                schedstat_inc(this_rq(), bkl_count);
5419                schedstat_inc(prev, sched_info.bkl_count);
5420        }
5421#endif
5422}
5423
5424static void put_prev_task(struct rq *rq, struct task_struct *prev)
5425{
5426        if (prev->state == TASK_RUNNING) {
5427                u64 runtime = prev->se.sum_exec_runtime;
5428
5429                runtime -= prev->se.prev_sum_exec_runtime;
5430                runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5431
5432                /*
5433                 * In order to avoid avg_overlap growing stale when we are
5434                 * indeed overlapping and hence not getting put to sleep, grow
5435                 * the avg_overlap on preemption.
5436                 *
5437                 * We use the average preemption runtime because that
5438                 * correlates to the amount of cache footprint a task can
5439                 * build up.
5440                 */
5441                update_avg(&prev->se.avg_overlap, runtime);
5442        }
5443        prev->sched_class->put_prev_task(rq, prev);
5444}
5445
5446/*
5447 * Pick up the highest-prio task:
5448 */
5449static inline struct task_struct *
5450pick_next_task(struct rq *rq)
5451{
5452        const struct sched_class *class;
5453        struct task_struct *p;
5454
5455        /*
5456         * Optimization: we know that if all tasks are in
5457         * the fair class we can call that function directly:
5458         */
5459        if (likely(rq->nr_running == rq->cfs.nr_running)) {
5460                p = fair_sched_class.pick_next_task(rq);
5461                if (likely(p))
5462                        return p;
5463        }
5464
5465        class = sched_class_highest;
5466        for ( ; ; ) {
5467                p = class->pick_next_task(rq);
5468                if (p)
5469                        return p;
5470                /*
5471                 * Will never be NULL as the idle class always
5472                 * returns a non-NULL p:
5473                 */
5474                class = class->next;
5475        }
5476}
5477
5478/*
5479 * schedule() is the main scheduler function.
5480 */
5481asmlinkage void __sched schedule(void)
5482{
5483        struct task_struct *prev, *next;
5484        unsigned long *switch_count;
5485        struct rq *rq;
5486        int cpu;
5487
5488need_resched:
5489        preempt_disable();
5490        cpu = smp_processor_id();
5491        rq = cpu_rq(cpu);
5492        rcu_sched_qs(cpu);
5493        prev = rq->curr;
5494        switch_count = &prev->nivcsw;
5495
5496        release_kernel_lock(prev);
5497need_resched_nonpreemptible:
5498
5499        schedule_debug(prev);
5500
5501        if (sched_feat(HRTICK))
5502                hrtick_clear(rq);
5503
5504        raw_spin_lock_irq(&rq->lock);
5505        update_rq_clock(rq);
5506        clear_tsk_need_resched(prev);
5507
5508        if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5509                if (unlikely(signal_pending_state(prev->state, prev)))
5510                        prev->state = TASK_RUNNING;
5511                else
5512                        deactivate_task(rq, prev, 1);
5513                switch_count = &prev->nvcsw;
5514        }
5515
5516        pre_schedule(rq, prev);
5517
5518        if (unlikely(!rq->nr_running))
5519                idle_balance(cpu, rq);
5520
5521        put_prev_task(rq, prev);
5522        next = pick_next_task(rq);
5523
5524        if (likely(prev != next)) {
5525                sched_info_switch(prev, next);
5526                perf_event_task_sched_out(prev, next, cpu);
5527
5528                rq->nr_switches++;
5529                rq->curr = next;
5530                ++*switch_count;
5531
5532                context_switch(rq, prev, next); /* unlocks the rq */
5533                /*
5534                 * the context switch might have flipped the stack from under
5535                 * us, hence refresh the local variables.
5536                 */
5537                cpu = smp_processor_id();
5538                rq = cpu_rq(cpu);
5539        } else
5540                raw_spin_unlock_irq(&rq->lock);
5541
5542        post_schedule(rq);
5543
5544        if (unlikely(reacquire_kernel_lock(current) < 0)) {
5545                prev = rq->curr;
5546                switch_count = &prev->nivcsw;
5547                goto need_resched_nonpreemptible;
5548        }
5549
5550        preempt_enable_no_resched();
5551        if (need_resched())
5552                goto need_resched;
5553}
5554EXPORT_SYMBOL(schedule);
5555
5556#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5557/*
5558 * Look out! "owner" is an entirely speculative pointer
5559 * access and not reliable.
5560 */
5561int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5562{
5563        unsigned int cpu;
5564        struct rq *rq;
5565
5566        if (!sched_feat(OWNER_SPIN))
5567                return 0;
5568
5569#ifdef CONFIG_DEBUG_PAGEALLOC
5570        /*
5571         * Need to access the cpu field knowing that
5572         * DEBUG_PAGEALLOC could have unmapped it if
5573         * the mutex owner just released it and exited.
5574         */
5575        if (probe_kernel_address(&owner->cpu, cpu))
5576                goto out;
5577#else
5578        cpu = owner->cpu;
5579#endif
5580
5581        /*
5582         * Even if the access succeeded (likely case),
5583         * the cpu field may no longer be valid.
5584         */
5585        if (cpu >= nr_cpumask_bits)
5586                goto out;
5587
5588        /*
5589         * We need to validate that we can do a
5590         * get_cpu() and that we have the percpu area.
5591         */
5592        if (!cpu_online(cpu))
5593                goto out;
5594
5595        rq = cpu_rq(cpu);
5596
5597        for (;;) {
5598                /*
5599                 * Owner changed, break to re-assess state.
5600                 */
5601                if (lock->owner != owner)
5602                        break;
5603
5604                /*
5605                 * Is that owner really running on that cpu?
5606                 */
5607                if (task_thread_info(rq->curr) != owner || need_resched())
5608                        return 0;
5609
5610                cpu_relax();
5611        }
5612out:
5613        return 1;
5614}
5615#endif
5616
5617#ifdef CONFIG_PREEMPT
5618/*
5619 * this is the entry point to schedule() from in-kernel preemption
5620 * off of preempt_enable. Kernel preemptions off return from interrupt
5621 * occur there and call schedule directly.
5622 */
5623asmlinkage void __sched preempt_schedule(void)
5624{
5625        struct thread_info *ti = current_thread_info();
5626
5627        /*
5628         * If there is a non-zero preempt_count or interrupts are disabled,
5629         * we do not want to preempt the current task. Just return..
5630         */
5631        if (likely(ti->preempt_count || irqs_disabled()))
5632                return;
5633
5634        do {
5635                add_preempt_count(PREEMPT_ACTIVE);
5636                schedule();
5637                sub_preempt_count(PREEMPT_ACTIVE);
5638
5639                /*
5640                 * Check again in case we missed a preemption opportunity
5641                 * between schedule and now.
5642                 */
5643                barrier();
5644        } while (need_resched());
5645}
5646EXPORT_SYMBOL(preempt_schedule);
5647
5648/*
5649 * this is the entry point to schedule() from kernel preemption
5650 * off of irq context.
5651 * Note, that this is called and return with irqs disabled. This will
5652 * protect us against recursive calling from irq.
5653 */
5654asmlinkage void __sched preempt_schedule_irq(void)
5655{
5656        struct thread_info *ti = current_thread_info();
5657
5658        /* Catch callers which need to be fixed */
5659        BUG_ON(ti->preempt_count || !irqs_disabled());
5660
5661        do {
5662                add_preempt_count(PREEMPT_ACTIVE);
5663                local_irq_enable();
5664                schedule();
5665                local_irq_disable();
5666                sub_preempt_count(PREEMPT_ACTIVE);
5667
5668                /*
5669                 * Check again in case we missed a preemption opportunity
5670                 * between schedule and now.
5671                 */
5672                barrier();
5673        } while (need_resched());
5674}
5675
5676#endif /* CONFIG_PREEMPT */
5677
5678int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5679                          void *key)
5680{
5681        return try_to_wake_up(curr->private, mode, wake_flags);
5682}
5683EXPORT_SYMBOL(default_wake_function);
5684
5685/*
5686 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5687 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5688 * number) then we wake all the non-exclusive tasks and one exclusive task.
5689 *
5690 * There are circumstances in which we can try to wake a task which has already
5691 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5692 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5693 */
5694static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5695                        int nr_exclusive, int wake_flags, void *key)
5696{
5697        wait_queue_t *curr, *next;
5698
5699        list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5700                unsigned flags = curr->flags;
5701
5702                if (curr->func(curr, mode, wake_flags, key) &&
5703                                (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5704                        break;
5705        }
5706}
5707
5708/**
5709 * __wake_up - wake up threads blocked on a waitqueue.
5710 * @q: the waitqueue
5711 * @mode: which threads
5712 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5713 * @key: is directly passed to the wakeup function
5714 *
5715 * It may be assumed that this function implies a write memory barrier before
5716 * changing the task state if and only if any tasks are woken up.
5717 */
5718void __wake_up(wait_queue_head_t *q, unsigned int mode,
5719                        int nr_exclusive, void *key)
5720{
5721        unsigned long flags;
5722
5723        spin_lock_irqsave(&q->lock, flags);
5724        __wake_up_common(q, mode, nr_exclusive, 0, key);
5725        spin_unlock_irqrestore(&q->lock, flags);
5726}
5727EXPORT_SYMBOL(__wake_up);
5728
5729/*
5730 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5731 */
5732void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5733{
5734        __wake_up_common(q, mode, 1, 0, NULL);
5735}
5736
5737void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5738{
5739        __wake_up_common(q, mode, 1, 0, key);
5740}
5741
5742/**
5743 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5744 * @q: the waitqueue
5745 * @mode: which threads
5746 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5747 * @key: opaque value to be passed to wakeup targets
5748 *
5749 * The sync wakeup differs that the waker knows that it will schedule
5750 * away soon, so while the target thread will be woken up, it will not
5751 * be migrated to another CPU - ie. the two threads are 'synchronized'
5752 * with each other. This can prevent needless bouncing between CPUs.
5753 *
5754 * On UP it can prevent extra preemption.
5755 *
5756 * It may be assumed that this function implies a write memory barrier before
5757 * changing the task state if and only if any tasks are woken up.
5758 */
5759void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5760                        int nr_exclusive, void *key)
5761{
5762        unsigned long flags;
5763        int wake_flags = WF_SYNC;
5764
5765        if (unlikely(!q))
5766                return;
5767
5768        if (unlikely(!nr_exclusive))
5769                wake_flags = 0;
5770
5771        spin_lock_irqsave(&q->lock, flags);
5772        __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5773        spin_unlock_irqrestore(&q->lock, flags);
5774}
5775EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5776
5777/*
5778 * __wake_up_sync - see __wake_up_sync_key()
5779 */
5780void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5781{
5782        __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5783}
5784EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
5785
5786/**
5787 * complete: - signals a single thread waiting on this completion
5788 * @x:  holds the state of this particular completion
5789 *
5790 * This will wake up a single thread waiting on this completion. Threads will be
5791 * awakened in the same order in which they were queued.
5792 *
5793 * See also complete_all(), wait_for_completion() and related routines.
5794 *
5795 * It may be assumed that this function implies a write memory barrier before
5796 * changing the task state if and only if any tasks are woken up.
5797 */
5798void complete(struct completion *x)
5799{
5800        unsigned long flags;
5801
5802        spin_lock_irqsave(&x->wait.lock, flags);
5803        x->done++;
5804        __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5805        spin_unlock_irqrestore(&x->wait.lock, flags);
5806}
5807EXPORT_SYMBOL(complete);
5808
5809/**
5810 * complete_all: - signals all threads waiting on this completion
5811 * @x:  holds the state of this particular completion
5812 *
5813 * This will wake up all threads waiting on this particular completion event.
5814 *
5815 * It may be assumed that this function implies a write memory barrier before
5816 * changing the task state if and only if any tasks are woken up.
5817 */
5818void complete_all(struct completion *x)
5819{
5820        unsigned long flags;
5821
5822        spin_lock_irqsave(&x->wait.lock, flags);
5823        x->done += UINT_MAX/2;
5824        __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5825        spin_unlock_irqrestore(&x->wait.lock, flags);
5826}
5827EXPORT_SYMBOL(complete_all);
5828
5829static inline long __sched
5830do_wait_for_common(struct completion *x, long timeout, int state)
5831{
5832        if (!x->done) {
5833                DECLARE_WAITQUEUE(wait, current);
5834
5835                wait.flags |= WQ_FLAG_EXCLUSIVE;
5836                __add_wait_queue_tail(&x->wait, &wait);
5837                do {
5838                        if (signal_pending_state(state, current)) {
5839                                timeout = -ERESTARTSYS;
5840                                break;
5841                        }
5842                        __set_current_state(state);
5843                        spin_unlock_irq(&x->wait.lock);
5844                        timeout = schedule_timeout(timeout);
5845                        spin_lock_irq(&x->wait.lock);
5846                } while (!x->done && timeout);
5847                __remove_wait_queue(&x->wait, &wait);
5848                if (!x->done)
5849                        return timeout;
5850        }
5851        x->done--;
5852        return timeout ?: 1;
5853}
5854
5855static long __sched
5856wait_for_common(struct completion *x, long timeout, int state)
5857{
5858        might_sleep();
5859
5860        spin_lock_irq(&x->wait.lock);
5861        timeout = do_wait_for_common(x, timeout, state);
5862        spin_unlock_irq(&x->wait.lock);
5863        return timeout;
5864}
5865
5866/**
5867 * wait_for_completion: - waits for completion of a task
5868 * @x:  holds the state of this particular completion
5869 *
5870 * This waits to be signaled for completion of a specific task. It is NOT
5871 * interruptible and there is no timeout.
5872 *
5873 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5874 * and interrupt capability. Also see complete().
5875 */
5876void __sched wait_for_completion(struct completion *x)
5877{
5878        wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5879}
5880EXPORT_SYMBOL(wait_for_completion);
5881
5882/**
5883 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5884 * @x:  holds the state of this particular completion
5885 * @timeout:  timeout value in jiffies
5886 *
5887 * This waits for either a completion of a specific task to be signaled or for a
5888 * specified timeout to expire. The timeout is in jiffies. It is not
5889 * interruptible.
5890 */
5891unsigned long __sched
5892wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5893{
5894        return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5895}
5896EXPORT_SYMBOL(wait_for_completion_timeout);
5897
5898/**
5899 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5900 * @x:  holds the state of this particular completion
5901 *
5902 * This waits for completion of a specific task to be signaled. It is
5903 * interruptible.
5904 */
5905int __sched wait_for_completion_interruptible(struct completion *x)
5906{
5907        long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5908        if (t == -ERESTARTSYS)
5909                return t;
5910        return 0;
5911}
5912EXPORT_SYMBOL(wait_for_completion_interruptible);
5913
5914/**
5915 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5916 * @x:  holds the state of this particular completion
5917 * @timeout:  timeout value in jiffies
5918 *
5919 * This waits for either a completion of a specific task to be signaled or for a
5920 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5921 */
5922unsigned long __sched
5923wait_for_completion_interruptible_timeout(struct completion *x,
5924                                          unsigned long timeout)
5925{
5926        return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5927}
5928EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5929
5930/**
5931 * wait_for_completion_killable: - waits for completion of a task (killable)
5932 * @x:  holds the state of this particular completion
5933 *
5934 * This waits to be signaled for completion of a specific task. It can be
5935 * interrupted by a kill signal.
5936 */
5937int __sched wait_for_completion_killable(struct completion *x)
5938{
5939        long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5940        if (t == -ERESTARTSYS)
5941                return t;
5942        return 0;
5943}
5944EXPORT_SYMBOL(wait_for_completion_killable);
5945
5946/**
5947 *      try_wait_for_completion - try to decrement a completion without blocking
5948 *      @x:     completion structure
5949 *
5950 *      Returns: 0 if a decrement cannot be done without blocking
5951 *               1 if a decrement succeeded.
5952 *
5953 *      If a completion is being used as a counting completion,
5954 *      attempt to decrement the counter without blocking. This
5955 *      enables us to avoid waiting if the resource the completion
5956 *      is protecting is not available.
5957 */
5958bool try_wait_for_completion(struct completion *x)
5959{
5960        unsigned long flags;
5961        int ret = 1;
5962
5963        spin_lock_irqsave(&x->wait.lock, flags);
5964        if (!x->done)
5965                ret = 0;
5966        else
5967                x->done--;
5968        spin_unlock_irqrestore(&x->wait.lock, flags);
5969        return ret;
5970}
5971EXPORT_SYMBOL(try_wait_for_completion);
5972
5973/**
5974 *      completion_done - Test to see if a completion has any waiters
5975 *      @x:     completion structure
5976 *
5977 *      Returns: 0 if there are waiters (wait_for_completion() in progress)
5978 *               1 if there are no waiters.
5979 *
5980 */
5981bool completion_done(struct completion *x)
5982{
5983        unsigned long flags;
5984        int ret = 1;
5985
5986        spin_lock_irqsave(&x->wait.lock, flags);
5987        if (!x->done)
5988                ret = 0;
5989        spin_unlock_irqrestore(&x->wait.