linux/kernel/sched.c
<<
>>
Prefs
   1/*
   2 *  kernel/sched.c
   3 *
   4 *  Kernel scheduler and related syscalls
   5 *
   6 *  Copyright (C) 1991-2002  Linus Torvalds
   7 *
   8 *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
   9 *              make semaphores SMP safe
  10 *  1998-11-19  Implemented schedule_timeout() and related stuff
  11 *              by Andrea Arcangeli
  12 *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
  13 *              hybrid priority-list and round-robin design with
  14 *              an array-switch method of distributing timeslices
  15 *              and per-CPU runqueues.  Cleanups and useful suggestions
  16 *              by Davide Libenzi, preemptible kernel bits by Robert Love.
  17 *  2003-09-03  Interactivity tuning by Con Kolivas.
  18 *  2004-04-02  Scheduler domains code by Nick Piggin
  19 *  2007-04-15  Work begun on replacing all interactivity tuning with a
  20 *              fair scheduling design by Con Kolivas.
  21 *  2007-05-05  Load balancing (smp-nice) and other improvements
  22 *              by Peter Williams
  23 *  2007-05-06  Interactivity improvements to CFS by Mike Galbraith
  24 *  2007-07-01  Group scheduling enhancements by Srivatsa Vaddagiri
  25 *  2007-11-29  RT balancing improvements by Steven Rostedt, Gregory Haskins,
  26 *              Thomas Gleixner, Mike Kravetz
  27 */
  28
  29#include <linux/mm.h>
  30#include <linux/module.h>
  31#include <linux/nmi.h>
  32#include <linux/init.h>
  33#include <linux/uaccess.h>
  34#include <linux/highmem.h>
  35#include <linux/smp_lock.h>
  36#include <asm/mmu_context.h>
  37#include <linux/interrupt.h>
  38#include <linux/capability.h>
  39#include <linux/completion.h>
  40#include <linux/kernel_stat.h>
  41#include <linux/debug_locks.h>
  42#include <linux/security.h>
  43#include <linux/notifier.h>
  44#include <linux/profile.h>
  45#include <linux/freezer.h>
  46#include <linux/vmalloc.h>
  47#include <linux/blkdev.h>
  48#include <linux/delay.h>
  49#include <linux/pid_namespace.h>
  50#include <linux/smp.h>
  51#include <linux/threads.h>
  52#include <linux/timer.h>
  53#include <linux/rcupdate.h>
  54#include <linux/cpu.h>
  55#include <linux/cpuset.h>
  56#include <linux/percpu.h>
  57#include <linux/kthread.h>
  58#include <linux/seq_file.h>
  59#include <linux/sysctl.h>
  60#include <linux/syscalls.h>
  61#include <linux/times.h>
  62#include <linux/tsacct_kern.h>
  63#include <linux/kprobes.h>
  64#include <linux/delayacct.h>
  65#include <linux/reciprocal_div.h>
  66#include <linux/unistd.h>
  67#include <linux/pagemap.h>
  68#include <linux/hrtimer.h>
  69#include <linux/tick.h>
  70#include <linux/bootmem.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/*
  81 * Convert user-nice values [ -20 ... 0 ... 19 ]
  82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
  83 * and back.
  84 */
  85#define NICE_TO_PRIO(nice)      (MAX_RT_PRIO + (nice) + 20)
  86#define PRIO_TO_NICE(prio)      ((prio) - MAX_RT_PRIO - 20)
  87#define TASK_NICE(p)            PRIO_TO_NICE((p)->static_prio)
  88
  89/*
  90 * 'User priority' is the nice value converted to something we
  91 * can work with better when scaling various scheduler parameters,
  92 * it's a [ 0 ... 39 ] range.
  93 */
  94#define USER_PRIO(p)            ((p)-MAX_RT_PRIO)
  95#define TASK_USER_PRIO(p)       USER_PRIO((p)->static_prio)
  96#define MAX_USER_PRIO           (USER_PRIO(MAX_PRIO))
  97
  98/*
  99 * Helpers for converting nanosecond timing to jiffy resolution
 100 */
 101#define NS_TO_JIFFIES(TIME)     ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
 102
 103#define NICE_0_LOAD             SCHED_LOAD_SCALE
 104#define NICE_0_SHIFT            SCHED_LOAD_SHIFT
 105
 106/*
 107 * These are the 'tuning knobs' of the scheduler:
 108 *
 109 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
 110 * Timeslices get refilled after they expire.
 111 */
 112#define DEF_TIMESLICE           (100 * HZ / 1000)
 113
 114/*
 115 * single value that denotes runtime == period, ie unlimited time.
 116 */
 117#define RUNTIME_INF     ((u64)~0ULL)
 118
 119#ifdef CONFIG_SMP
 120/*
 121 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
 122 * Since cpu_power is a 'constant', we can use a reciprocal divide.
 123 */
 124static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
 125{
 126        return reciprocal_divide(load, sg->reciprocal_cpu_power);
 127}
 128
 129/*
 130 * Each time a sched group cpu_power is changed,
 131 * we must compute its reciprocal value
 132 */
 133static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
 134{
 135        sg->__cpu_power += val;
 136        sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
 137}
 138#endif
 139
 140static inline int rt_policy(int policy)
 141{
 142        if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
 143                return 1;
 144        return 0;
 145}
 146
 147static inline int task_has_rt_policy(struct task_struct *p)
 148{
 149        return rt_policy(p->policy);
 150}
 151
 152/*
 153 * This is the priority-queue data structure of the RT scheduling class:
 154 */
 155struct rt_prio_array {
 156        DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
 157        struct list_head queue[MAX_RT_PRIO];
 158};
 159
 160struct rt_bandwidth {
 161        /* nests inside the rq lock: */
 162        spinlock_t              rt_runtime_lock;
 163        ktime_t                 rt_period;
 164        u64                     rt_runtime;
 165        struct hrtimer          rt_period_timer;
 166};
 167
 168static struct rt_bandwidth def_rt_bandwidth;
 169
 170static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
 171
 172static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
 173{
 174        struct rt_bandwidth *rt_b =
 175                container_of(timer, struct rt_bandwidth, rt_period_timer);
 176        ktime_t now;
 177        int overrun;
 178        int idle = 0;
 179
 180        for (;;) {
 181                now = hrtimer_cb_get_time(timer);
 182                overrun = hrtimer_forward(timer, now, rt_b->rt_period);
 183
 184                if (!overrun)
 185                        break;
 186
 187                idle = do_sched_rt_period_timer(rt_b, overrun);
 188        }
 189
 190        return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
 191}
 192
 193static
 194void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
 195{
 196        rt_b->rt_period = ns_to_ktime(period);
 197        rt_b->rt_runtime = runtime;
 198
 199        spin_lock_init(&rt_b->rt_runtime_lock);
 200
 201        hrtimer_init(&rt_b->rt_period_timer,
 202                        CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 203        rt_b->rt_period_timer.function = sched_rt_period_timer;
 204        rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
 205}
 206
 207static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
 208{
 209        ktime_t now;
 210
 211        if (rt_b->rt_runtime == RUNTIME_INF)
 212                return;
 213
 214        if (hrtimer_active(&rt_b->rt_period_timer))
 215                return;
 216
 217        spin_lock(&rt_b->rt_runtime_lock);
 218        for (;;) {
 219                if (hrtimer_active(&rt_b->rt_period_timer))
 220                        break;
 221
 222                now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
 223                hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
 224                hrtimer_start(&rt_b->rt_period_timer,
 225                              rt_b->rt_period_timer.expires,
 226                              HRTIMER_MODE_ABS);
 227        }
 228        spin_unlock(&rt_b->rt_runtime_lock);
 229}
 230
 231#ifdef CONFIG_RT_GROUP_SCHED
 232static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
 233{
 234        hrtimer_cancel(&rt_b->rt_period_timer);
 235}
 236#endif
 237
 238/*
 239 * sched_domains_mutex serializes calls to arch_init_sched_domains,
 240 * detach_destroy_domains and partition_sched_domains.
 241 */
 242static DEFINE_MUTEX(sched_domains_mutex);
 243
 244#ifdef CONFIG_GROUP_SCHED
 245
 246#include <linux/cgroup.h>
 247
 248struct cfs_rq;
 249
 250static LIST_HEAD(task_groups);
 251
 252/* task group related information */
 253struct task_group {
 254#ifdef CONFIG_CGROUP_SCHED
 255        struct cgroup_subsys_state css;
 256#endif
 257
 258#ifdef CONFIG_FAIR_GROUP_SCHED
 259        /* schedulable entities of this group on each cpu */
 260        struct sched_entity **se;
 261        /* runqueue "owned" by this group on each cpu */
 262        struct cfs_rq **cfs_rq;
 263        unsigned long shares;
 264#endif
 265
 266#ifdef CONFIG_RT_GROUP_SCHED
 267        struct sched_rt_entity **rt_se;
 268        struct rt_rq **rt_rq;
 269
 270        struct rt_bandwidth rt_bandwidth;
 271#endif
 272
 273        struct rcu_head rcu;
 274        struct list_head list;
 275
 276        struct task_group *parent;
 277        struct list_head siblings;
 278        struct list_head children;
 279};
 280
 281#ifdef CONFIG_USER_SCHED
 282
 283/*
 284 * Root task group.
 285 *      Every UID task group (including init_task_group aka UID-0) will
 286 *      be a child to this group.
 287 */
 288struct task_group root_task_group;
 289
 290#ifdef CONFIG_FAIR_GROUP_SCHED
 291/* Default task group's sched entity on each cpu */
 292static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
 293/* Default task group's cfs_rq on each cpu */
 294static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
 295#endif /* CONFIG_FAIR_GROUP_SCHED */
 296
 297#ifdef CONFIG_RT_GROUP_SCHED
 298static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
 299static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
 300#endif /* CONFIG_RT_GROUP_SCHED */
 301#else /* !CONFIG_FAIR_GROUP_SCHED */
 302#define root_task_group init_task_group
 303#endif /* CONFIG_FAIR_GROUP_SCHED */
 304
 305/* task_group_lock serializes add/remove of task groups and also changes to
 306 * a task group's cpu shares.
 307 */
 308static DEFINE_SPINLOCK(task_group_lock);
 309
 310#ifdef CONFIG_FAIR_GROUP_SCHED
 311#ifdef CONFIG_USER_SCHED
 312# define INIT_TASK_GROUP_LOAD   (2*NICE_0_LOAD)
 313#else /* !CONFIG_USER_SCHED */
 314# define INIT_TASK_GROUP_LOAD   NICE_0_LOAD
 315#endif /* CONFIG_USER_SCHED */
 316
 317/*
 318 * A weight of 0 or 1 can cause arithmetics problems.
 319 * A weight of a cfs_rq is the sum of weights of which entities
 320 * are queued on this cfs_rq, so a weight of a entity should not be
 321 * too large, so as the shares value of a task group.
 322 * (The default weight is 1024 - so there's no practical
 323 *  limitation from this.)
 324 */
 325#define MIN_SHARES      2
 326#define MAX_SHARES      (1UL << 18)
 327
 328static int init_task_group_load = INIT_TASK_GROUP_LOAD;
 329#endif
 330
 331/* Default task group.
 332 *      Every task in system belong to this group at bootup.
 333 */
 334struct task_group init_task_group;
 335
 336/* return group to which a task belongs */
 337static inline struct task_group *task_group(struct task_struct *p)
 338{
 339        struct task_group *tg;
 340
 341#ifdef CONFIG_USER_SCHED
 342        tg = p->user->tg;
 343#elif defined(CONFIG_CGROUP_SCHED)
 344        tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
 345                                struct task_group, css);
 346#else
 347        tg = &init_task_group;
 348#endif
 349        return tg;
 350}
 351
 352/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
 353static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
 354{
 355#ifdef CONFIG_FAIR_GROUP_SCHED
 356        p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
 357        p->se.parent = task_group(p)->se[cpu];
 358#endif
 359
 360#ifdef CONFIG_RT_GROUP_SCHED
 361        p->rt.rt_rq  = task_group(p)->rt_rq[cpu];
 362        p->rt.parent = task_group(p)->rt_se[cpu];
 363#endif
 364}
 365
 366#else
 367
 368static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
 369static inline struct task_group *task_group(struct task_struct *p)
 370{
 371        return NULL;
 372}
 373
 374#endif  /* CONFIG_GROUP_SCHED */
 375
 376/* CFS-related fields in a runqueue */
 377struct cfs_rq {
 378        struct load_weight load;
 379        unsigned long nr_running;
 380
 381        u64 exec_clock;
 382        u64 min_vruntime;
 383        u64 pair_start;
 384
 385        struct rb_root tasks_timeline;
 386        struct rb_node *rb_leftmost;
 387
 388        struct list_head tasks;
 389        struct list_head *balance_iterator;
 390
 391        /*
 392         * 'curr' points to currently running entity on this cfs_rq.
 393         * It is set to NULL otherwise (i.e when none are currently running).
 394         */
 395        struct sched_entity *curr, *next;
 396
 397        unsigned long nr_spread_over;
 398
 399#ifdef CONFIG_FAIR_GROUP_SCHED
 400        struct rq *rq;  /* cpu runqueue to which this cfs_rq is attached */
 401
 402        /*
 403         * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
 404         * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
 405         * (like users, containers etc.)
 406         *
 407         * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
 408         * list is used during load balance.
 409         */
 410        struct list_head leaf_cfs_rq_list;
 411        struct task_group *tg;  /* group that "owns" this runqueue */
 412
 413#ifdef CONFIG_SMP
 414        /*
 415         * the part of load.weight contributed by tasks
 416         */
 417        unsigned long task_weight;
 418
 419        /*
 420         *   h_load = weight * f(tg)
 421         *
 422         * Where f(tg) is the recursive weight fraction assigned to
 423         * this group.
 424         */
 425        unsigned long h_load;
 426
 427        /*
 428         * this cpu's part of tg->shares
 429         */
 430        unsigned long shares;
 431
 432        /*
 433         * load.weight at the time we set shares
 434         */
 435        unsigned long rq_weight;
 436#endif
 437#endif
 438};
 439
 440/* Real-Time classes' related field in a runqueue: */
 441struct rt_rq {
 442        struct rt_prio_array active;
 443        unsigned long rt_nr_running;
 444#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 445        int highest_prio; /* highest queued rt task prio */
 446#endif
 447#ifdef CONFIG_SMP
 448        unsigned long rt_nr_migratory;
 449        int overloaded;
 450#endif
 451        int rt_throttled;
 452        u64 rt_time;
 453        u64 rt_runtime;
 454        /* Nests inside the rq lock: */
 455        spinlock_t rt_runtime_lock;
 456
 457#ifdef CONFIG_RT_GROUP_SCHED
 458        unsigned long rt_nr_boosted;
 459
 460        struct rq *rq;
 461        struct list_head leaf_rt_rq_list;
 462        struct task_group *tg;
 463        struct sched_rt_entity *rt_se;
 464#endif
 465};
 466
 467#ifdef CONFIG_SMP
 468
 469/*
 470 * We add the notion of a root-domain which will be used to define per-domain
 471 * variables. Each exclusive cpuset essentially defines an island domain by
 472 * fully partitioning the member cpus from any other cpuset. Whenever a new
 473 * exclusive cpuset is created, we also create and attach a new root-domain
 474 * object.
 475 *
 476 */
 477struct root_domain {
 478        atomic_t refcount;
 479        cpumask_t span;
 480        cpumask_t online;
 481
 482        /*
 483         * The "RT overload" flag: it gets set if a CPU has more than
 484         * one runnable RT task.
 485         */
 486        cpumask_t rto_mask;
 487        atomic_t rto_count;
 488#ifdef CONFIG_SMP
 489        struct cpupri cpupri;
 490#endif
 491};
 492
 493/*
 494 * By default the system creates a single root-domain with all cpus as
 495 * members (mimicking the global state we have today).
 496 */
 497static struct root_domain def_root_domain;
 498
 499#endif
 500
 501/*
 502 * This is the main, per-CPU runqueue data structure.
 503 *
 504 * Locking rule: those places that want to lock multiple runqueues
 505 * (such as the load balancing or the thread migration code), lock
 506 * acquire operations must be ordered by ascending &runqueue.
 507 */
 508struct rq {
 509        /* runqueue lock: */
 510        spinlock_t lock;
 511
 512        /*
 513         * nr_running and cpu_load should be in the same cacheline because
 514         * remote CPUs use both these fields when doing load calculation.
 515         */
 516        unsigned long nr_running;
 517        #define CPU_LOAD_IDX_MAX 5
 518        unsigned long cpu_load[CPU_LOAD_IDX_MAX];
 519        unsigned char idle_at_tick;
 520#ifdef CONFIG_NO_HZ
 521        unsigned long last_tick_seen;
 522        unsigned char in_nohz_recently;
 523#endif
 524        /* capture load from *all* tasks on this cpu: */
 525        struct load_weight load;
 526        unsigned long nr_load_updates;
 527        u64 nr_switches;
 528
 529        struct cfs_rq cfs;
 530        struct rt_rq rt;
 531
 532#ifdef CONFIG_FAIR_GROUP_SCHED
 533        /* list of leaf cfs_rq on this cpu: */
 534        struct list_head leaf_cfs_rq_list;
 535#endif
 536#ifdef CONFIG_RT_GROUP_SCHED
 537        struct list_head leaf_rt_rq_list;
 538#endif
 539
 540        /*
 541         * This is part of a global counter where only the total sum
 542         * over all CPUs matters. A task can increase this counter on
 543         * one CPU and if it got migrated afterwards it may decrease
 544         * it on another CPU. Always updated under the runqueue lock:
 545         */
 546        unsigned long nr_uninterruptible;
 547
 548        struct task_struct *curr, *idle;
 549        unsigned long next_balance;
 550        struct mm_struct *prev_mm;
 551
 552        u64 clock;
 553
 554        atomic_t nr_iowait;
 555
 556#ifdef CONFIG_SMP
 557        struct root_domain *rd;
 558        struct sched_domain *sd;
 559
 560        /* For active balancing */
 561        int active_balance;
 562        int push_cpu;
 563        /* cpu of this runqueue: */
 564        int cpu;
 565        int online;
 566
 567        unsigned long avg_load_per_task;
 568
 569        struct task_struct *migration_thread;
 570        struct list_head migration_queue;
 571#endif
 572
 573#ifdef CONFIG_SCHED_HRTICK
 574#ifdef CONFIG_SMP
 575        int hrtick_csd_pending;
 576        struct call_single_data hrtick_csd;
 577#endif
 578        struct hrtimer hrtick_timer;
 579#endif
 580
 581#ifdef CONFIG_SCHEDSTATS
 582        /* latency stats */
 583        struct sched_info rq_sched_info;
 584
 585        /* sys_sched_yield() stats */
 586        unsigned int yld_exp_empty;
 587        unsigned int yld_act_empty;
 588        unsigned int yld_both_empty;
 589        unsigned int yld_count;
 590
 591        /* schedule() stats */
 592        unsigned int sched_switch;
 593        unsigned int sched_count;
 594        unsigned int sched_goidle;
 595
 596        /* try_to_wake_up() stats */
 597        unsigned int ttwu_count;
 598        unsigned int ttwu_local;
 599
 600        /* BKL stats */
 601        unsigned int bkl_count;
 602#endif
 603};
 604
 605static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
 606
 607static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
 608{
 609        rq->curr->sched_class->check_preempt_curr(rq, p);
 610}
 611
 612static inline int cpu_of(struct rq *rq)
 613{
 614#ifdef CONFIG_SMP
 615        return rq->cpu;
 616#else
 617        return 0;
 618#endif
 619}
 620
 621/*
 622 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
 623 * See detach_destroy_domains: synchronize_sched for details.
 624 *
 625 * The domain tree of any CPU may only be accessed from within
 626 * preempt-disabled sections.
 627 */
 628#define for_each_domain(cpu, __sd) \
 629        for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
 630
 631#define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
 632#define this_rq()               (&__get_cpu_var(runqueues))
 633#define task_rq(p)              cpu_rq(task_cpu(p))
 634#define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
 635
 636static inline void update_rq_clock(struct rq *rq)
 637{
 638        rq->clock = sched_clock_cpu(cpu_of(rq));
 639}
 640
 641/*
 642 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
 643 */
 644#ifdef CONFIG_SCHED_DEBUG
 645# define const_debug __read_mostly
 646#else
 647# define const_debug static const
 648#endif
 649
 650/**
 651 * runqueue_is_locked
 652 *
 653 * Returns true if the current cpu runqueue is locked.
 654 * This interface allows printk to be called with the runqueue lock
 655 * held and know whether or not it is OK to wake up the klogd.
 656 */
 657int runqueue_is_locked(void)
 658{
 659        int cpu = get_cpu();
 660        struct rq *rq = cpu_rq(cpu);
 661        int ret;
 662
 663        ret = spin_is_locked(&rq->lock);
 664        put_cpu();
 665        return ret;
 666}
 667
 668/*
 669 * Debugging: various feature bits
 670 */
 671
 672#define SCHED_FEAT(name, enabled)       \
 673        __SCHED_FEAT_##name ,
 674
 675enum {
 676#include "sched_features.h"
 677};
 678
 679#undef SCHED_FEAT
 680
 681#define SCHED_FEAT(name, enabled)       \
 682        (1UL << __SCHED_FEAT_##name) * enabled |
 683
 684const_debug unsigned int sysctl_sched_features =
 685#include "sched_features.h"
 686        0;
 687
 688#undef SCHED_FEAT
 689
 690#ifdef CONFIG_SCHED_DEBUG
 691#define SCHED_FEAT(name, enabled)       \
 692        #name ,
 693
 694static __read_mostly char *sched_feat_names[] = {
 695#include "sched_features.h"
 696        NULL
 697};
 698
 699#undef SCHED_FEAT
 700
 701static int sched_feat_open(struct inode *inode, struct file *filp)
 702{
 703        filp->private_data = inode->i_private;
 704        return 0;
 705}
 706
 707static ssize_t
 708sched_feat_read(struct file *filp, char __user *ubuf,
 709                size_t cnt, loff_t *ppos)
 710{
 711        char *buf;
 712        int r = 0;
 713        int len = 0;
 714        int i;
 715
 716        for (i = 0; sched_feat_names[i]; i++) {
 717                len += strlen(sched_feat_names[i]);
 718                len += 4;
 719        }
 720
 721        buf = kmalloc(len + 2, GFP_KERNEL);
 722        if (!buf)
 723                return -ENOMEM;
 724
 725        for (i = 0; sched_feat_names[i]; i++) {
 726                if (sysctl_sched_features & (1UL << i))
 727                        r += sprintf(buf + r, "%s ", sched_feat_names[i]);
 728                else
 729                        r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
 730        }
 731
 732        r += sprintf(buf + r, "\n");
 733        WARN_ON(r >= len + 2);
 734
 735        r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
 736
 737        kfree(buf);
 738
 739        return r;
 740}
 741
 742static ssize_t
 743sched_feat_write(struct file *filp, const char __user *ubuf,
 744                size_t cnt, loff_t *ppos)
 745{
 746        char buf[64];
 747        char *cmp = buf;
 748        int neg = 0;
 749        int i;
 750
 751        if (cnt > 63)
 752                cnt = 63;
 753
 754        if (copy_from_user(&buf, ubuf, cnt))
 755                return -EFAULT;
 756
 757        buf[cnt] = 0;
 758
 759        if (strncmp(buf, "NO_", 3) == 0) {
 760                neg = 1;
 761                cmp += 3;
 762        }
 763
 764        for (i = 0; sched_feat_names[i]; i++) {
 765                int len = strlen(sched_feat_names[i]);
 766
 767                if (strncmp(cmp, sched_feat_names[i], len) == 0) {
 768                        if (neg)
 769                                sysctl_sched_features &= ~(1UL << i);
 770                        else
 771                                sysctl_sched_features |= (1UL << i);
 772                        break;
 773                }
 774        }
 775
 776        if (!sched_feat_names[i])
 777                return -EINVAL;
 778
 779        filp->f_pos += cnt;
 780
 781        return cnt;
 782}
 783
 784static struct file_operations sched_feat_fops = {
 785        .open   = sched_feat_open,
 786        .read   = sched_feat_read,
 787        .write  = sched_feat_write,
 788};
 789
 790static __init int sched_init_debug(void)
 791{
 792        debugfs_create_file("sched_features", 0644, NULL, NULL,
 793                        &sched_feat_fops);
 794
 795        return 0;
 796}
 797late_initcall(sched_init_debug);
 798
 799#endif
 800
 801#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
 802
 803/*
 804 * Number of tasks to iterate in a single balance run.
 805 * Limited because this is done with IRQs disabled.
 806 */
 807const_debug unsigned int sysctl_sched_nr_migrate = 32;
 808
 809/*
 810 * ratelimit for updating the group shares.
 811 * default: 0.25ms
 812 */
 813unsigned int sysctl_sched_shares_ratelimit = 250000;
 814
 815/*
 816 * period over which we measure -rt task cpu usage in us.
 817 * default: 1s
 818 */
 819unsigned int sysctl_sched_rt_period = 1000000;
 820
 821static __read_mostly int scheduler_running;
 822
 823/*
 824 * part of the period that we allow rt tasks to run in us.
 825 * default: 0.95s
 826 */
 827int sysctl_sched_rt_runtime = 950000;
 828
 829static inline u64 global_rt_period(void)
 830{
 831        return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
 832}
 833
 834static inline u64 global_rt_runtime(void)
 835{
 836        if (sysctl_sched_rt_runtime < 0)
 837                return RUNTIME_INF;
 838
 839        return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
 840}
 841
 842#ifndef prepare_arch_switch
 843# define prepare_arch_switch(next)      do { } while (0)
 844#endif
 845#ifndef finish_arch_switch
 846# define finish_arch_switch(prev)       do { } while (0)
 847#endif
 848
 849static inline int task_current(struct rq *rq, struct task_struct *p)
 850{
 851        return rq->curr == p;
 852}
 853
 854#ifndef __ARCH_WANT_UNLOCKED_CTXSW
 855static inline int task_running(struct rq *rq, struct task_struct *p)
 856{
 857        return task_current(rq, p);
 858}
 859
 860static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
 861{
 862}
 863
 864static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
 865{
 866#ifdef CONFIG_DEBUG_SPINLOCK
 867        /* this is a valid case when another task releases the spinlock */
 868        rq->lock.owner = current;
 869#endif
 870        /*
 871         * If we are tracking spinlock dependencies then we have to
 872         * fix up the runqueue lock - which gets 'carried over' from
 873         * prev into current:
 874         */
 875        spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
 876
 877        spin_unlock_irq(&rq->lock);
 878}
 879
 880#else /* __ARCH_WANT_UNLOCKED_CTXSW */
 881static inline int task_running(struct rq *rq, struct task_struct *p)
 882{
 883#ifdef CONFIG_SMP
 884        return p->oncpu;
 885#else
 886        return task_current(rq, p);
 887#endif
 888}
 889
 890static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
 891{
 892#ifdef CONFIG_SMP
 893        /*
 894         * We can optimise this out completely for !SMP, because the
 895         * SMP rebalancing from interrupt is the only thing that cares
 896         * here.
 897         */
 898        next->oncpu = 1;
 899#endif
 900#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
 901        spin_unlock_irq(&rq->lock);
 902#else
 903        spin_unlock(&rq->lock);
 904#endif
 905}
 906
 907static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
 908{
 909#ifdef CONFIG_SMP
 910        /*
 911         * After ->oncpu is cleared, the task can be moved to a different CPU.
 912         * We must ensure this doesn't happen until the switch is completely
 913         * finished.
 914         */
 915        smp_wmb();
 916        prev->oncpu = 0;
 917#endif
 918#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
 919        local_irq_enable();
 920#endif
 921}
 922#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
 923
 924/*
 925 * __task_rq_lock - lock the runqueue a given task resides on.
 926 * Must be called interrupts disabled.
 927 */
 928static inline struct rq *__task_rq_lock(struct task_struct *p)
 929        __acquires(rq->lock)
 930{
 931        for (;;) {
 932                struct rq *rq = task_rq(p);
 933                spin_lock(&rq->lock);
 934                if (likely(rq == task_rq(p)))
 935                        return rq;
 936                spin_unlock(&rq->lock);
 937        }
 938}
 939
 940/*
 941 * task_rq_lock - lock the runqueue a given task resides on and disable
 942 * interrupts. Note the ordering: we can safely lookup the task_rq without
 943 * explicitly disabling preemption.
 944 */
 945static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
 946        __acquires(rq->lock)
 947{
 948        struct rq *rq;
 949
 950        for (;;) {
 951                local_irq_save(*flags);
 952                rq = task_rq(p);
 953                spin_lock(&rq->lock);
 954                if (likely(rq == task_rq(p)))
 955                        return rq;
 956                spin_unlock_irqrestore(&rq->lock, *flags);
 957        }
 958}
 959
 960static void __task_rq_unlock(struct rq *rq)
 961        __releases(rq->lock)
 962{
 963        spin_unlock(&rq->lock);
 964}
 965
 966static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
 967        __releases(rq->lock)
 968{
 969        spin_unlock_irqrestore(&rq->lock, *flags);
 970}
 971
 972/*
 973 * this_rq_lock - lock this runqueue and disable interrupts.
 974 */
 975static struct rq *this_rq_lock(void)
 976        __acquires(rq->lock)
 977{
 978        struct rq *rq;
 979
 980        local_irq_disable();
 981        rq = this_rq();
 982        spin_lock(&rq->lock);
 983
 984        return rq;
 985}
 986
 987#ifdef CONFIG_SCHED_HRTICK
 988/*
 989 * Use HR-timers to deliver accurate preemption points.
 990 *
 991 * Its all a bit involved since we cannot program an hrt while holding the
 992 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
 993 * reschedule event.
 994 *
 995 * When we get rescheduled we reprogram the hrtick_timer outside of the
 996 * rq->lock.
 997 */
 998
 999/*
1000 * Use hrtick when:
1001 *  - enabled by features
1002 *  - hrtimer is actually high res
1003 */
1004static inline int hrtick_enabled(struct rq *rq)
1005{
1006        if (!sched_feat(HRTICK))
1007                return 0;
1008        if (!cpu_active(cpu_of(rq)))
1009                return 0;
1010        return hrtimer_is_hres_active(&rq->hrtick_timer);
1011}
1012
1013static void hrtick_clear(struct rq *rq)
1014{
1015        if (hrtimer_active(&rq->hrtick_timer))
1016                hrtimer_cancel(&rq->hrtick_timer);
1017}
1018
1019/*
1020 * High-resolution timer tick.
1021 * Runs from hardirq context with interrupts disabled.
1022 */
1023static enum hrtimer_restart hrtick(struct hrtimer *timer)
1024{
1025        struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1026
1027        WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1028
1029        spin_lock(&rq->lock);
1030        update_rq_clock(rq);
1031        rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1032        spin_unlock(&rq->lock);
1033
1034        return HRTIMER_NORESTART;
1035}
1036
1037#ifdef CONFIG_SMP
1038/*
1039 * called from hardirq (IPI) context
1040 */
1041static void __hrtick_start(void *arg)
1042{
1043        struct rq *rq = arg;
1044
1045        spin_lock(&rq->lock);
1046        hrtimer_restart(&rq->hrtick_timer);
1047        rq->hrtick_csd_pending = 0;
1048        spin_unlock(&rq->lock);
1049}
1050
1051/*
1052 * Called to set the hrtick timer state.
1053 *
1054 * called with rq->lock held and irqs disabled
1055 */
1056static void hrtick_start(struct rq *rq, u64 delay)
1057{
1058        struct hrtimer *timer = &rq->hrtick_timer;
1059        ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1060
1061        timer->expires = time;
1062
1063        if (rq == this_rq()) {
1064                hrtimer_restart(timer);
1065        } else if (!rq->hrtick_csd_pending) {
1066                __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1067                rq->hrtick_csd_pending = 1;
1068        }
1069}
1070
1071static int
1072hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1073{
1074        int cpu = (int)(long)hcpu;
1075
1076        switch (action) {
1077        case CPU_UP_CANCELED:
1078        case CPU_UP_CANCELED_FROZEN:
1079        case CPU_DOWN_PREPARE:
1080        case CPU_DOWN_PREPARE_FROZEN:
1081        case CPU_DEAD:
1082        case CPU_DEAD_FROZEN:
1083                hrtick_clear(cpu_rq(cpu));
1084                return NOTIFY_OK;
1085        }
1086
1087        return NOTIFY_DONE;
1088}
1089
1090static __init void init_hrtick(void)
1091{
1092        hotcpu_notifier(hotplug_hrtick, 0);
1093}
1094#else
1095/*
1096 * Called to set the hrtick timer state.
1097 *
1098 * called with rq->lock held and irqs disabled
1099 */
1100static void hrtick_start(struct rq *rq, u64 delay)
1101{
1102        hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1103}
1104
1105static void init_hrtick(void)
1106{
1107}
1108#endif /* CONFIG_SMP */
1109
1110static void init_rq_hrtick(struct rq *rq)
1111{
1112#ifdef CONFIG_SMP
1113        rq->hrtick_csd_pending = 0;
1114
1115        rq->hrtick_csd.flags = 0;
1116        rq->hrtick_csd.func = __hrtick_start;
1117        rq->hrtick_csd.info = rq;
1118#endif
1119
1120        hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1121        rq->hrtick_timer.function = hrtick;
1122        rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1123}
1124#else
1125static inline void hrtick_clear(struct rq *rq)
1126{
1127}
1128
1129static inline void init_rq_hrtick(struct rq *rq)
1130{
1131}
1132
1133static inline void init_hrtick(void)
1134{
1135}
1136#endif
1137
1138/*
1139 * resched_task - mark a task 'to be rescheduled now'.
1140 *
1141 * On UP this means the setting of the need_resched flag, on SMP it
1142 * might also involve a cross-CPU call to trigger the scheduler on
1143 * the target CPU.
1144 */
1145#ifdef CONFIG_SMP
1146
1147#ifndef tsk_is_polling
1148#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1149#endif
1150
1151static void resched_task(struct task_struct *p)
1152{
1153        int cpu;
1154
1155        assert_spin_locked(&task_rq(p)->lock);
1156
1157        if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1158                return;
1159
1160        set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1161
1162        cpu = task_cpu(p);
1163        if (cpu == smp_processor_id())
1164                return;
1165
1166        /* NEED_RESCHED must be visible before we test polling */
1167        smp_mb();
1168        if (!tsk_is_polling(p))
1169                smp_send_reschedule(cpu);
1170}
1171
1172static void resched_cpu(int cpu)
1173{
1174        struct rq *rq = cpu_rq(cpu);
1175        unsigned long flags;
1176
1177        if (!spin_trylock_irqsave(&rq->lock, flags))
1178                return;
1179        resched_task(cpu_curr(cpu));
1180        spin_unlock_irqrestore(&rq->lock, flags);
1181}
1182
1183#ifdef CONFIG_NO_HZ
1184/*
1185 * When add_timer_on() enqueues a timer into the timer wheel of an
1186 * idle CPU then this timer might expire before the next timer event
1187 * which is scheduled to wake up that CPU. In case of a completely
1188 * idle system the next event might even be infinite time into the
1189 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1190 * leaves the inner idle loop so the newly added timer is taken into
1191 * account when the CPU goes back to idle and evaluates the timer
1192 * wheel for the next timer event.
1193 */
1194void wake_up_idle_cpu(int cpu)
1195{
1196        struct rq *rq = cpu_rq(cpu);
1197
1198        if (cpu == smp_processor_id())
1199                return;
1200
1201        /*
1202         * This is safe, as this function is called with the timer
1203         * wheel base lock of (cpu) held. When the CPU is on the way
1204         * to idle and has not yet set rq->curr to idle then it will
1205         * be serialized on the timer wheel base lock and take the new
1206         * timer into account automatically.
1207         */
1208        if (rq->curr != rq->idle)
1209                return;
1210
1211        /*
1212         * We can set TIF_RESCHED on the idle task of the other CPU
1213         * lockless. The worst case is that the other CPU runs the
1214         * idle task through an additional NOOP schedule()
1215         */
1216        set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1217
1218        /* NEED_RESCHED must be visible before we test polling */
1219        smp_mb();
1220        if (!tsk_is_polling(rq->idle))
1221                smp_send_reschedule(cpu);
1222}
1223#endif /* CONFIG_NO_HZ */
1224
1225#else /* !CONFIG_SMP */
1226static void resched_task(struct task_struct *p)
1227{
1228        assert_spin_locked(&task_rq(p)->lock);
1229        set_tsk_need_resched(p);
1230}
1231#endif /* CONFIG_SMP */
1232
1233#if BITS_PER_LONG == 32
1234# define WMULT_CONST    (~0UL)
1235#else
1236# define WMULT_CONST    (1UL << 32)
1237#endif
1238
1239#define WMULT_SHIFT     32
1240
1241/*
1242 * Shift right and round:
1243 */
1244#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1245
1246/*
1247 * delta *= weight / lw
1248 */
1249static unsigned long
1250calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1251                struct load_weight *lw)
1252{
1253        u64 tmp;
1254
1255        if (!lw->inv_weight) {
1256                if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1257                        lw->inv_weight = 1;
1258                else
1259                        lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1260                                / (lw->weight+1);
1261        }
1262
1263        tmp = (u64)delta_exec * weight;
1264        /*
1265         * Check whether we'd overflow the 64-bit multiplication:
1266         */
1267        if (unlikely(tmp > WMULT_CONST))
1268                tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1269                        WMULT_SHIFT/2);
1270        else
1271                tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1272
1273        return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1274}
1275
1276static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1277{
1278        lw->weight += inc;
1279        lw->inv_weight = 0;
1280}
1281
1282static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1283{
1284        lw->weight -= dec;
1285        lw->inv_weight = 0;
1286}
1287
1288/*
1289 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1290 * of tasks with abnormal "nice" values across CPUs the contribution that
1291 * each task makes to its run queue's load is weighted according to its
1292 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1293 * scaled version of the new time slice allocation that they receive on time
1294 * slice expiry etc.
1295 */
1296
1297#define WEIGHT_IDLEPRIO         2
1298#define WMULT_IDLEPRIO          (1 << 31)
1299
1300/*
1301 * Nice levels are multiplicative, with a gentle 10% change for every
1302 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1303 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1304 * that remained on nice 0.
1305 *
1306 * The "10% effect" is relative and cumulative: from _any_ nice level,
1307 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1308 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1309 * If a task goes up by ~10% and another task goes down by ~10% then
1310 * the relative distance between them is ~25%.)
1311 */
1312static const int prio_to_weight[40] = {
1313 /* -20 */     88761,     71755,     56483,     46273,     36291,
1314 /* -15 */     29154,     23254,     18705,     14949,     11916,
1315 /* -10 */      9548,      7620,      6100,      4904,      3906,
1316 /*  -5 */      3121,      2501,      1991,      1586,      1277,
1317 /*   0 */      1024,       820,       655,       526,       423,
1318 /*   5 */       335,       272,       215,       172,       137,
1319 /*  10 */       110,        87,        70,        56,        45,
1320 /*  15 */        36,        29,        23,        18,        15,
1321};
1322
1323/*
1324 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1325 *
1326 * In cases where the weight does not change often, we can use the
1327 * precalculated inverse to speed up arithmetics by turning divisions
1328 * into multiplications:
1329 */
1330static const u32 prio_to_wmult[40] = {
1331 /* -20 */     48388,     59856,     76040,     92818,    118348,
1332 /* -15 */    147320,    184698,    229616,    287308,    360437,
1333 /* -10 */    449829,    563644,    704093,    875809,   1099582,
1334 /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
1335 /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
1336 /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
1337 /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
1338 /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1339};
1340
1341static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1342
1343/*
1344 * runqueue iterator, to support SMP load-balancing between different
1345 * scheduling classes, without having to expose their internal data
1346 * structures to the load-balancing proper:
1347 */
1348struct rq_iterator {
1349        void *arg;
1350        struct task_struct *(*start)(void *);
1351        struct task_struct *(*next)(void *);
1352};
1353
1354#ifdef CONFIG_SMP
1355static unsigned long
1356balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1357              unsigned long max_load_move, struct sched_domain *sd,
1358              enum cpu_idle_type idle, int *all_pinned,
1359              int *this_best_prio, struct rq_iterator *iterator);
1360
1361static int
1362iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1363                   struct sched_domain *sd, enum cpu_idle_type idle,
1364                   struct rq_iterator *iterator);
1365#endif
1366
1367#ifdef CONFIG_CGROUP_CPUACCT
1368static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1369#else
1370static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1371#endif
1372
1373static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1374{
1375        update_load_add(&rq->load, load);
1376}
1377
1378static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1379{
1380        update_load_sub(&rq->load, load);
1381}
1382
1383#ifdef CONFIG_SMP
1384static unsigned long source_load(int cpu, int type);
1385static unsigned long target_load(int cpu, int type);
1386static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1387
1388static unsigned long cpu_avg_load_per_task(int cpu)
1389{
1390        struct rq *rq = cpu_rq(cpu);
1391
1392        if (rq->nr_running)
1393                rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1394
1395        return rq->avg_load_per_task;
1396}
1397
1398#ifdef CONFIG_FAIR_GROUP_SCHED
1399
1400typedef void (*tg_visitor)(struct task_group *, int, struct sched_domain *);
1401
1402/*
1403 * Iterate the full tree, calling @down when first entering a node and @up when
1404 * leaving it for the final time.
1405 */
1406static void
1407walk_tg_tree(tg_visitor down, tg_visitor up, int cpu, struct sched_domain *sd)
1408{
1409        struct task_group *parent, *child;
1410
1411        rcu_read_lock();
1412        parent = &root_task_group;
1413down:
1414        (*down)(parent, cpu, sd);
1415        list_for_each_entry_rcu(child, &parent->children, siblings) {
1416                parent = child;
1417                goto down;
1418
1419up:
1420                continue;
1421        }
1422        (*up)(parent, cpu, sd);
1423
1424        child = parent;
1425        parent = parent->parent;
1426        if (parent)
1427                goto up;
1428        rcu_read_unlock();
1429}
1430
1431static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1432
1433/*
1434 * Calculate and set the cpu's group shares.
1435 */
1436static void
1437__update_group_shares_cpu(struct task_group *tg, int cpu,
1438                          unsigned long sd_shares, unsigned long sd_rq_weight)
1439{
1440        int boost = 0;
1441        unsigned long shares;
1442        unsigned long rq_weight;
1443
1444        if (!tg->se[cpu])
1445                return;
1446
1447        rq_weight = tg->cfs_rq[cpu]->load.weight;
1448
1449        /*
1450         * If there are currently no tasks on the cpu pretend there is one of
1451         * average load so that when a new task gets to run here it will not
1452         * get delayed by group starvation.
1453         */
1454        if (!rq_weight) {
1455                boost = 1;
1456                rq_weight = NICE_0_LOAD;
1457        }
1458
1459        if (unlikely(rq_weight > sd_rq_weight))
1460                rq_weight = sd_rq_weight;
1461
1462        /*
1463         *           \Sum shares * rq_weight
1464         * shares =  -----------------------
1465         *               \Sum rq_weight
1466         *
1467         */
1468        shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1469
1470        /*
1471         * record the actual number of shares, not the boosted amount.
1472         */
1473        tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1474        tg->cfs_rq[cpu]->rq_weight = rq_weight;
1475
1476        if (shares < MIN_SHARES)
1477                shares = MIN_SHARES;
1478        else if (shares > MAX_SHARES)
1479                shares = MAX_SHARES;
1480
1481        __set_se_shares(tg->se[cpu], shares);
1482}
1483
1484/*
1485 * Re-compute the task group their per cpu shares over the given domain.
1486 * This needs to be done in a bottom-up fashion because the rq weight of a
1487 * parent group depends on the shares of its child groups.
1488 */
1489static void
1490tg_shares_up(struct task_group *tg, int cpu, struct sched_domain *sd)
1491{
1492        unsigned long rq_weight = 0;
1493        unsigned long shares = 0;
1494        int i;
1495
1496        for_each_cpu_mask(i, sd->span) {
1497                rq_weight += tg->cfs_rq[i]->load.weight;
1498                shares += tg->cfs_rq[i]->shares;
1499        }
1500
1501        if ((!shares && rq_weight) || shares > tg->shares)
1502                shares = tg->shares;
1503
1504        if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1505                shares = tg->shares;
1506
1507        if (!rq_weight)
1508                rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1509
1510        for_each_cpu_mask(i, sd->span) {
1511                struct rq *rq = cpu_rq(i);
1512                unsigned long flags;
1513
1514                spin_lock_irqsave(&rq->lock, flags);
1515                __update_group_shares_cpu(tg, i, shares, rq_weight);
1516                spin_unlock_irqrestore(&rq->lock, flags);
1517        }
1518}
1519
1520/*
1521 * Compute the cpu's hierarchical load factor for each task group.
1522 * This needs to be done in a top-down fashion because the load of a child
1523 * group is a fraction of its parents load.
1524 */
1525static void
1526tg_load_down(struct task_group *tg, int cpu, struct sched_domain *sd)
1527{
1528        unsigned long load;
1529
1530        if (!tg->parent) {
1531                load = cpu_rq(cpu)->load.weight;
1532        } else {
1533                load = tg->parent->cfs_rq[cpu]->h_load;
1534                load *= tg->cfs_rq[cpu]->shares;
1535                load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1536        }
1537
1538        tg->cfs_rq[cpu]->h_load = load;
1539}
1540
1541static void
1542tg_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
1543{
1544}
1545
1546static void update_shares(struct sched_domain *sd)
1547{
1548        u64 now = cpu_clock(raw_smp_processor_id());
1549        s64 elapsed = now - sd->last_update;
1550
1551        if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1552                sd->last_update = now;
1553                walk_tg_tree(tg_nop, tg_shares_up, 0, sd);
1554        }
1555}
1556
1557static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1558{
1559        spin_unlock(&rq->lock);
1560        update_shares(sd);
1561        spin_lock(&rq->lock);
1562}
1563
1564static void update_h_load(int cpu)
1565{
1566        walk_tg_tree(tg_load_down, tg_nop, cpu, NULL);
1567}
1568
1569#else
1570
1571static inline void update_shares(struct sched_domain *sd)
1572{
1573}
1574
1575static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1576{
1577}
1578
1579#endif
1580
1581#endif
1582
1583#ifdef CONFIG_FAIR_GROUP_SCHED
1584static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1585{
1586#ifdef CONFIG_SMP
1587        cfs_rq->shares = shares;
1588#endif
1589}
1590#endif
1591
1592#include "sched_stats.h"
1593#include "sched_idletask.c"
1594#include "sched_fair.c"
1595#include "sched_rt.c"
1596#ifdef CONFIG_SCHED_DEBUG
1597# include "sched_debug.c"
1598#endif
1599
1600#define sched_class_highest (&rt_sched_class)
1601#define for_each_class(class) \
1602   for (class = sched_class_highest; class; class = class->next)
1603
1604static void inc_nr_running(struct rq *rq)
1605{
1606        rq->nr_running++;
1607}
1608
1609static void dec_nr_running(struct rq *rq)
1610{
1611        rq->nr_running--;
1612}
1613
1614static void set_load_weight(struct task_struct *p)
1615{
1616        if (task_has_rt_policy(p)) {
1617                p->se.load.weight = prio_to_weight[0] * 2;
1618                p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1619                return;
1620        }
1621
1622        /*
1623         * SCHED_IDLE tasks get minimal weight:
1624         */
1625        if (p->policy == SCHED_IDLE) {
1626                p->se.load.weight = WEIGHT_IDLEPRIO;
1627                p->se.load.inv_weight = WMULT_IDLEPRIO;
1628                return;
1629        }
1630
1631        p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1632        p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1633}
1634
1635static void update_avg(u64 *avg, u64 sample)
1636{
1637        s64 diff = sample - *avg;
1638        *avg += diff >> 3;
1639}
1640
1641static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1642{
1643        sched_info_queued(p);
1644        p->sched_class->enqueue_task(rq, p, wakeup);
1645        p->se.on_rq = 1;
1646}
1647
1648static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1649{
1650        if (sleep && p->se.last_wakeup) {
1651                update_avg(&p->se.avg_overlap,
1652                           p->se.sum_exec_runtime - p->se.last_wakeup);
1653                p->se.last_wakeup = 0;
1654        }
1655
1656        sched_info_dequeued(p);
1657        p->sched_class->dequeue_task(rq, p, sleep);
1658        p->se.on_rq = 0;
1659}
1660
1661/*
1662 * __normal_prio - return the priority that is based on the static prio
1663 */
1664static inline int __normal_prio(struct task_struct *p)
1665{
1666        return p->static_prio;
1667}
1668
1669/*
1670 * Calculate the expected normal priority: i.e. priority
1671 * without taking RT-inheritance into account. Might be
1672 * boosted by interactivity modifiers. Changes upon fork,
1673 * setprio syscalls, and whenever the interactivity
1674 * estimator recalculates.
1675 */
1676static inline int normal_prio(struct task_struct *p)
1677{
1678        int prio;
1679
1680        if (task_has_rt_policy(p))
1681                prio = MAX_RT_PRIO-1 - p->rt_priority;
1682        else
1683                prio = __normal_prio(p);
1684        return prio;
1685}
1686
1687/*
1688 * Calculate the current priority, i.e. the priority
1689 * taken into account by the scheduler. This value might
1690 * be boosted by RT tasks, or might be boosted by
1691 * interactivity modifiers. Will be RT if the task got
1692 * RT-boosted. If not then it returns p->normal_prio.
1693 */
1694static int effective_prio(struct task_struct *p)
1695{
1696        p->normal_prio = normal_prio(p);
1697        /*
1698         * If we are RT tasks or we were boosted to RT priority,
1699         * keep the priority unchanged. Otherwise, update priority
1700         * to the normal priority:
1701         */
1702        if (!rt_prio(p->prio))
1703                return p->normal_prio;
1704        return p->prio;
1705}
1706
1707/*
1708 * activate_task - move a task to the runqueue.
1709 */
1710static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1711{
1712        if (task_contributes_to_load(p))
1713                rq->nr_uninterruptible--;
1714
1715        enqueue_task(rq, p, wakeup);
1716        inc_nr_running(rq);
1717}
1718
1719/*
1720 * deactivate_task - remove a task from the runqueue.
1721 */
1722static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1723{
1724        if (task_contributes_to_load(p))
1725                rq->nr_uninterruptible++;
1726
1727        dequeue_task(rq, p, sleep);
1728        dec_nr_running(rq);
1729}
1730
1731/**
1732 * task_curr - is this task currently executing on a CPU?
1733 * @p: the task in question.
1734 */
1735inline int task_curr(const struct task_struct *p)
1736{
1737        return cpu_curr(task_cpu(p)) == p;
1738}
1739
1740static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1741{
1742        set_task_rq(p, cpu);
1743#ifdef CONFIG_SMP
1744        /*
1745         * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1746         * successfuly executed on another CPU. We must ensure that updates of
1747         * per-task data have been completed by this moment.
1748         */
1749        smp_wmb();
1750        task_thread_info(p)->cpu = cpu;
1751#endif
1752}
1753
1754static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1755                                       const struct sched_class *prev_class,
1756                                       int oldprio, int running)
1757{
1758        if (prev_class != p->sched_class) {
1759                if (prev_class->switched_from)
1760                        prev_class->switched_from(rq, p, running);
1761                p->sched_class->switched_to(rq, p, running);
1762        } else
1763                p->sched_class->prio_changed(rq, p, oldprio, running);
1764}
1765
1766#ifdef CONFIG_SMP
1767
1768/* Used instead of source_load when we know the type == 0 */
1769static unsigned long weighted_cpuload(const int cpu)
1770{
1771        return cpu_rq(cpu)->load.weight;
1772}
1773
1774/*
1775 * Is this task likely cache-hot:
1776 */
1777static int
1778task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1779{
1780        s64 delta;
1781
1782        /*
1783         * Buddy candidates are cache hot:
1784         */
1785        if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1786                return 1;
1787
1788        if (p->sched_class != &fair_sched_class)
1789                return 0;
1790
1791        if (sysctl_sched_migration_cost == -1)
1792                return 1;
1793        if (sysctl_sched_migration_cost == 0)
1794                return 0;
1795
1796        delta = now - p->se.exec_start;
1797
1798        return delta < (s64)sysctl_sched_migration_cost;
1799}
1800
1801
1802void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1803{
1804        int old_cpu = task_cpu(p);
1805        struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1806        struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1807                      *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1808        u64 clock_offset;
1809
1810        clock_offset = old_rq->clock - new_rq->clock;
1811
1812#ifdef CONFIG_SCHEDSTATS
1813        if (p->se.wait_start)
1814                p->se.wait_start -= clock_offset;
1815        if (p->se.sleep_start)
1816                p->se.sleep_start -= clock_offset;
1817        if (p->se.block_start)
1818                p->se.block_start -= clock_offset;
1819        if (old_cpu != new_cpu) {
1820                schedstat_inc(p, se.nr_migrations);
1821                if (task_hot(p, old_rq->clock, NULL))
1822                        schedstat_inc(p, se.nr_forced2_migrations);
1823        }
1824#endif
1825        p->se.vruntime -= old_cfsrq->min_vruntime -
1826                                         new_cfsrq->min_vruntime;
1827
1828        __set_task_cpu(p, new_cpu);
1829}
1830
1831struct migration_req {
1832        struct list_head list;
1833
1834        struct task_struct *task;
1835        int dest_cpu;
1836
1837        struct completion done;
1838};
1839
1840/*
1841 * The task's runqueue lock must be held.
1842 * Returns true if you have to wait for migration thread.
1843 */
1844static int
1845migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1846{
1847        struct rq *rq = task_rq(p);
1848
1849        /*
1850         * If the task is not on a runqueue (and not running), then
1851         * it is sufficient to simply update the task's cpu field.
1852         */
1853        if (!p->se.on_rq && !task_running(rq, p)) {
1854                set_task_cpu(p, dest_cpu);
1855                return 0;
1856        }
1857
1858        init_completion(&req->done);
1859        req->task = p;
1860        req->dest_cpu = dest_cpu;
1861        list_add(&req->list, &rq->migration_queue);
1862
1863        return 1;
1864}
1865
1866/*
1867 * wait_task_inactive - wait for a thread to unschedule.
1868 *
1869 * If @match_state is nonzero, it's the @p->state value just checked and
1870 * not expected to change.  If it changes, i.e. @p might have woken up,
1871 * then return zero.  When we succeed in waiting for @p to be off its CPU,
1872 * we return a positive number (its total switch count).  If a second call
1873 * a short while later returns the same number, the caller can be sure that
1874 * @p has remained unscheduled the whole time.
1875 *
1876 * The caller must ensure that the task *will* unschedule sometime soon,
1877 * else this function might spin for a *long* time. This function can't
1878 * be called with interrupts off, or it may introduce deadlock with
1879 * smp_call_function() if an IPI is sent by the same process we are
1880 * waiting to become inactive.
1881 */
1882unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1883{
1884        unsigned long flags;
1885        int running, on_rq;
1886        unsigned long ncsw;
1887        struct rq *rq;
1888
1889        for (;;) {
1890                /*
1891                 * We do the initial early heuristics without holding
1892                 * any task-queue locks at all. We'll only try to get
1893                 * the runqueue lock when things look like they will
1894                 * work out!
1895                 */
1896                rq = task_rq(p);
1897
1898                /*
1899                 * If the task is actively running on another CPU
1900                 * still, just relax and busy-wait without holding
1901                 * any locks.
1902                 *
1903                 * NOTE! Since we don't hold any locks, it's not
1904                 * even sure that "rq" stays as the right runqueue!
1905                 * But we don't care, since "task_running()" will
1906                 * return false if the runqueue has changed and p
1907                 * is actually now running somewhere else!
1908                 */
1909                while (task_running(rq, p)) {
1910                        if (match_state && unlikely(p->state != match_state))
1911                                return 0;
1912                        cpu_relax();
1913                }
1914
1915                /*
1916                 * Ok, time to look more closely! We need the rq
1917                 * lock now, to be *sure*. If we're wrong, we'll
1918                 * just go back and repeat.
1919                 */
1920                rq = task_rq_lock(p, &flags);
1921                running = task_running(rq, p);
1922                on_rq = p->se.on_rq;
1923                ncsw = 0;
1924                if (!match_state || p->state == match_state) {
1925                        ncsw = p->nivcsw + p->nvcsw;
1926                        if (unlikely(!ncsw))
1927                                ncsw = 1;
1928                }
1929                task_rq_unlock(rq, &flags);
1930
1931                /*
1932                 * If it changed from the expected state, bail out now.
1933                 */
1934                if (unlikely(!ncsw))
1935                        break;
1936
1937                /*
1938                 * Was it really running after all now that we
1939                 * checked with the proper locks actually held?
1940                 *
1941                 * Oops. Go back and try again..
1942                 */
1943                if (unlikely(running)) {
1944                        cpu_relax();
1945                        continue;
1946                }
1947
1948                /*
1949                 * It's not enough that it's not actively running,
1950                 * it must be off the runqueue _entirely_, and not
1951                 * preempted!
1952                 *
1953                 * So if it wa still runnable (but just not actively
1954                 * running right now), it's preempted, and we should
1955                 * yield - it could be a while.
1956                 */
1957                if (unlikely(on_rq)) {
1958                        schedule_timeout_uninterruptible(1);
1959                        continue;
1960                }
1961
1962                /*
1963                 * Ahh, all good. It wasn't running, and it wasn't
1964                 * runnable, which means that it will never become
1965                 * running in the future either. We're all done!
1966                 */
1967                break;
1968        }
1969
1970        return ncsw;
1971}
1972
1973/***
1974 * kick_process - kick a running thread to enter/exit the kernel
1975 * @p: the to-be-kicked thread
1976 *
1977 * Cause a process which is running on another CPU to enter
1978 * kernel-mode, without any delay. (to get signals handled.)
1979 *
1980 * NOTE: this function doesnt have to take the runqueue lock,
1981 * because all it wants to ensure is that the remote task enters
1982 * the kernel. If the IPI races and the task has been migrated
1983 * to another CPU then no harm is done and the purpose has been
1984 * achieved as well.
1985 */
1986void kick_process(struct task_struct *p)
1987{
1988        int cpu;
1989
1990        preempt_disable();
1991        cpu = task_cpu(p);
1992        if ((cpu != smp_processor_id()) && task_curr(p))
1993                smp_send_reschedule(cpu);
1994        preempt_enable();
1995}
1996
1997/*
1998 * Return a low guess at the load of a migration-source cpu weighted
1999 * according to the scheduling class and "nice" value.
2000 *
2001 * We want to under-estimate the load of migration sources, to
2002 * balance conservatively.
2003 */
2004static unsigned long source_load(int cpu, int type)
2005{
2006        struct rq *rq = cpu_rq(cpu);
2007        unsigned long total = weighted_cpuload(cpu);
2008
2009        if (type == 0 || !sched_feat(LB_BIAS))
2010                return total;
2011
2012        return min(rq->cpu_load[type-1], total);
2013}
2014
2015/*
2016 * Return a high guess at the load of a migration-target cpu weighted
2017 * according to the scheduling class and "nice" value.
2018 */
2019static unsigned long target_load(int cpu, int type)
2020{
2021        struct rq *rq = cpu_rq(cpu);
2022        unsigned long total = weighted_cpuload(cpu);
2023
2024        if (type == 0 || !sched_feat(LB_BIAS))
2025                return total;
2026
2027        return max(rq->cpu_load[type-1], total);
2028}
2029
2030/*
2031 * find_idlest_group finds and returns the least busy CPU group within the
2032 * domain.
2033 */
2034static struct sched_group *
2035find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2036{
2037        struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2038        unsigned long min_load = ULONG_MAX, this_load = 0;
2039        int load_idx = sd->forkexec_idx;
2040        int imbalance = 100 + (sd->imbalance_pct-100)/2;
2041
2042        do {
2043                unsigned long load, avg_load;
2044                int local_group;
2045                int i;
2046
2047                /* Skip over this group if it has no CPUs allowed */
2048                if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2049                        continue;
2050
2051                local_group = cpu_isset(this_cpu, group->cpumask);
2052
2053                /* Tally up the load of all CPUs in the group */
2054                avg_load = 0;
2055
2056                for_each_cpu_mask_nr(i, group->cpumask) {
2057                        /* Bias balancing toward cpus of our domain */
2058                        if (local_group)
2059                                load = source_load(i, load_idx);
2060                        else
2061                                load = target_load(i, load_idx);
2062
2063                        avg_load += load;
2064                }
2065
2066                /* Adjust by relative CPU power of the group */
2067                avg_load = sg_div_cpu_power(group,
2068                                avg_load * SCHED_LOAD_SCALE);
2069
2070                if (local_group) {
2071                        this_load = avg_load;
2072                        this = group;
2073                } else if (avg_load < min_load) {
2074                        min_load = avg_load;
2075                        idlest = group;
2076                }
2077        } while (group = group->next, group != sd->groups);
2078
2079        if (!idlest || 100*this_load < imbalance*min_load)
2080                return NULL;
2081        return idlest;
2082}
2083
2084/*
2085 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2086 */
2087static int
2088find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2089                cpumask_t *tmp)
2090{
2091        unsigned long load, min_load = ULONG_MAX;
2092        int idlest = -1;
2093        int i;
2094
2095        /* Traverse only the allowed CPUs */
2096        cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2097
2098        for_each_cpu_mask_nr(i, *tmp) {
2099                load = weighted_cpuload(i);
2100
2101                if (load < min_load || (load == min_load && i == this_cpu)) {
2102                        min_load = load;
2103                        idlest = i;
2104                }
2105        }
2106
2107        return idlest;
2108}
2109
2110/*
2111 * sched_balance_self: balance the current task (running on cpu) in domains
2112 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2113 * SD_BALANCE_EXEC.
2114 *
2115 * Balance, ie. select the least loaded group.
2116 *
2117 * Returns the target CPU number, or the same CPU if no balancing is needed.
2118 *
2119 * preempt must be disabled.
2120 */
2121static int sched_balance_self(int cpu, int flag)
2122{
2123        struct task_struct *t = current;
2124        struct sched_domain *tmp, *sd = NULL;
2125
2126        for_each_domain(cpu, tmp) {
2127                /*
2128                 * If power savings logic is enabled for a domain, stop there.
2129                 */
2130                if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2131                        break;
2132                if (tmp->flags & flag)
2133                        sd = tmp;
2134        }
2135
2136        if (sd)
2137                update_shares(sd);
2138
2139        while (sd) {
2140                cpumask_t span, tmpmask;
2141                struct sched_group *group;
2142                int new_cpu, weight;
2143
2144                if (!(sd->flags & flag)) {
2145                        sd = sd->child;
2146                        continue;
2147                }
2148
2149                span = sd->span;
2150                group = find_idlest_group(sd, t, cpu);
2151                if (!group) {
2152                        sd = sd->child;
2153                        continue;
2154                }
2155
2156                new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2157                if (new_cpu == -1 || new_cpu == cpu) {
2158                        /* Now try balancing at a lower domain level of cpu */
2159                        sd = sd->child;
2160                        continue;
2161                }
2162
2163                /* Now try balancing at a lower domain level of new_cpu */
2164                cpu = new_cpu;
2165                sd = NULL;
2166                weight = cpus_weight(span);
2167                for_each_domain(cpu, tmp) {
2168                        if (weight <= cpus_weight(tmp->span))
2169                                break;
2170                        if (tmp->flags & flag)
2171                                sd = tmp;
2172                }
2173                /* while loop will break here if sd == NULL */
2174        }
2175
2176        return cpu;
2177}
2178
2179#endif /* CONFIG_SMP */
2180
2181/***
2182 * try_to_wake_up - wake up a thread
2183 * @p: the to-be-woken-up thread
2184 * @state: the mask of task states that can be woken
2185 * @sync: do a synchronous wakeup?
2186 *
2187 * Put it on the run-queue if it's not already there. The "current"
2188 * thread is always on the run-queue (except when the actual
2189 * re-schedule is in progress), and as such you're allowed to do
2190 * the simpler "current->state = TASK_RUNNING" to mark yourself
2191 * runnable without the overhead of this.
2192 *
2193 * returns failure only if the task is already active.
2194 */
2195static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2196{
2197        int cpu, orig_cpu, this_cpu, success = 0;
2198        unsigned long flags;
2199        long old_state;
2200        struct rq *rq;
2201
2202        if (!sched_feat(SYNC_WAKEUPS))
2203                sync = 0;
2204
2205#ifdef CONFIG_SMP
2206        if (sched_feat(LB_WAKEUP_UPDATE)) {
2207                struct sched_domain *sd;
2208
2209                this_cpu = raw_smp_processor_id();
2210                cpu = task_cpu(p);
2211
2212                for_each_domain(this_cpu, sd) {
2213                        if (cpu_isset(cpu, sd->span)) {
2214                                update_shares(sd);
2215                                break;
2216                        }
2217                }
2218        }
2219#endif
2220
2221        smp_wmb();
2222        rq = task_rq_lock(p, &flags);
2223        old_state = p->state;
2224        if (!(old_state & state))
2225                goto out;
2226
2227        if (p->se.on_rq)
2228                goto out_running;
2229
2230        cpu = task_cpu(p);
2231        orig_cpu = cpu;
2232        this_cpu = smp_processor_id();
2233
2234#ifdef CONFIG_SMP
2235        if (unlikely(task_running(rq, p)))
2236                goto out_activate;
2237
2238        cpu = p->sched_class->select_task_rq(p, sync);
2239        if (cpu != orig_cpu) {
2240                set_task_cpu(p, cpu);
2241                task_rq_unlock(rq, &flags);
2242                /* might preempt at this point */
2243                rq = task_rq_lock(p, &flags);
2244                old_state = p->state;
2245                if (!(old_state & state))
2246                        goto out;
2247                if (p->se.on_rq)
2248                        goto out_running;
2249
2250                this_cpu = smp_processor_id();
2251                cpu = task_cpu(p);
2252        }
2253
2254#ifdef CONFIG_SCHEDSTATS
2255        schedstat_inc(rq, ttwu_count);
2256        if (cpu == this_cpu)
2257                schedstat_inc(rq, ttwu_local);
2258        else {
2259                struct sched_domain *sd;
2260                for_each_domain(this_cpu, sd) {
2261                        if (cpu_isset(cpu, sd->span)) {
2262                                schedstat_inc(sd, ttwu_wake_remote);
2263                                break;
2264                        }
2265                }
2266        }
2267#endif /* CONFIG_SCHEDSTATS */
2268
2269out_activate:
2270#endif /* CONFIG_SMP */
2271        schedstat_inc(p, se.nr_wakeups);
2272        if (sync)
2273                schedstat_inc(p, se.nr_wakeups_sync);
2274        if (orig_cpu != cpu)
2275                schedstat_inc(p, se.nr_wakeups_migrate);
2276        if (cpu == this_cpu)
2277                schedstat_inc(p, se.nr_wakeups_local);
2278        else
2279                schedstat_inc(p, se.nr_wakeups_remote);
2280        update_rq_clock(rq);
2281        activate_task(rq, p, 1);
2282        success = 1;
2283
2284out_running:
2285        trace_mark(kernel_sched_wakeup,
2286                "pid %d state %ld ## rq %p task %p rq->curr %p",
2287                p->pid, p->state, rq, p, rq->curr);
2288        check_preempt_curr(rq, p);
2289
2290        p->state = TASK_RUNNING;
2291#ifdef CONFIG_SMP
2292        if (p->sched_class->task_wake_up)
2293                p->sched_class->task_wake_up(rq, p);
2294#endif
2295out:
2296        current->se.last_wakeup = current->se.sum_exec_runtime;
2297
2298        task_rq_unlock(rq, &flags);
2299
2300        return success;
2301}
2302
2303int wake_up_process(struct task_struct *p)
2304{
2305        return try_to_wake_up(p, TASK_ALL, 0);
2306}
2307EXPORT_SYMBOL(wake_up_process);
2308
2309int wake_up_state(struct task_struct *p, unsigned int state)
2310{
2311        return try_to_wake_up(p, state, 0);
2312}
2313
2314/*
2315 * Perform scheduler related setup for a newly forked process p.
2316 * p is forked by current.
2317 *
2318 * __sched_fork() is basic setup used by init_idle() too:
2319 */
2320static void __sched_fork(struct task_struct *p)
2321{
2322        p->se.exec_start                = 0;
2323        p->se.sum_exec_runtime          = 0;
2324        p->se.prev_sum_exec_runtime     = 0;
2325        p->se.last_wakeup               = 0;
2326        p->se.avg_overlap               = 0;
2327
2328#ifdef CONFIG_SCHEDSTATS
2329        p->se.wait_start                = 0;
2330        p->se.sum_sleep_runtime         = 0;
2331        p->se.sleep_start               = 0;
2332        p->se.block_start               = 0;
2333        p->se.sleep_max                 = 0;
2334        p->se.block_max                 = 0;
2335        p->se.exec_max                  = 0;
2336        p->se.slice_max                 = 0;
2337        p->se.wait_max                  = 0;
2338#endif
2339
2340        INIT_LIST_HEAD(&p->rt.run_list);
2341        p->se.on_rq = 0;
2342        INIT_LIST_HEAD(&p->se.group_node);
2343
2344#ifdef CONFIG_PREEMPT_NOTIFIERS
2345        INIT_HLIST_HEAD(&p->preempt_notifiers);
2346#endif
2347
2348        /*
2349         * We mark the process as running here, but have not actually
2350         * inserted it onto the runqueue yet. This guarantees that
2351         * nobody will actually run it, and a signal or other external
2352         * event cannot wake it up and insert it on the runqueue either.
2353         */
2354        p->state = TASK_RUNNING;
2355}
2356
2357/*
2358 * fork()/clone()-time setup:
2359 */
2360void sched_fork(struct task_struct *p, int clone_flags)
2361{
2362        int cpu = get_cpu();
2363
2364        __sched_fork(p);
2365
2366#ifdef CONFIG_SMP
2367        cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2368#endif
2369        set_task_cpu(p, cpu);
2370
2371        /*
2372         * Make sure we do not leak PI boosting priority to the child:
2373         */
2374        p->prio = current->normal_prio;
2375        if (!rt_prio(p->prio))
2376                p->sched_class = &fair_sched_class;
2377
2378#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2379        if (likely(sched_info_on()))
2380                memset(&p->sched_info, 0, sizeof(p->sched_info));
2381#endif
2382#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2383        p->oncpu = 0;
2384#endif
2385#ifdef CONFIG_PREEMPT
2386        /* Want to start with kernel preemption disabled. */
2387        task_thread_info(p)->preempt_count = 1;
2388#endif
2389        put_cpu();
2390}
2391
2392/*
2393 * wake_up_new_task - wake up a newly created task for the first time.
2394 *
2395 * This function will do some initial scheduler statistics housekeeping
2396 * that must be done for every newly created context, then puts the task
2397 * on the runqueue and wakes it.
2398 */
2399void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2400{
2401        unsigned long flags;
2402        struct rq *rq;
2403
2404        rq = task_rq_lock(p, &flags);
2405        BUG_ON(p->state != TASK_RUNNING);
2406        update_rq_clock(rq);
2407
2408        p->prio = effective_prio(p);
2409
2410        if (!p->sched_class->task_new || !current->se.on_rq) {
2411                activate_task(rq, p, 0);
2412        } else {
2413                /*
2414                 * Let the scheduling class do new task startup
2415                 * management (if any):
2416                 */
2417                p->sched_class->task_new(rq, p);
2418                inc_nr_running(rq);
2419        }
2420        trace_mark(kernel_sched_wakeup_new,
2421                "pid %d state %ld ## rq %p task %p rq->curr %p",
2422                p->pid, p->state, rq, p, rq->curr);
2423        check_preempt_curr(rq, p);
2424#ifdef CONFIG_SMP
2425        if (p->sched_class->task_wake_up)
2426                p->sched_class->task_wake_up(rq, p);
2427#endif
2428        task_rq_unlock(rq, &flags);
2429}
2430
2431#ifdef CONFIG_PREEMPT_NOTIFIERS
2432
2433/**
2434 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2435 * @notifier: notifier struct to register
2436 */
2437void preempt_notifier_register(struct preempt_notifier *notifier)
2438{
2439        hlist_add_head(&notifier->link, &current->preempt_notifiers);
2440}
2441EXPORT_SYMBOL_GPL(preempt_notifier_register);
2442
2443/**
2444 * preempt_notifier_unregister - no longer interested in preemption notifications
2445 * @notifier: notifier struct to unregister
2446 *
2447 * This is safe to call from within a preemption notifier.
2448 */
2449void preempt_notifier_unregister(struct preempt_notifier *notifier)
2450{
2451        hlist_del(&notifier->link);
2452}
2453EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2454
2455static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2456{
2457        struct preempt_notifier *notifier;
2458        struct hlist_node *node;
2459
2460        hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2461                notifier->ops->sched_in(notifier, raw_smp_processor_id());
2462}
2463
2464static void
2465fire_sched_out_preempt_notifiers(struct task_struct *curr,
2466                                 struct task_struct *next)
2467{
2468        struct preempt_notifier *notifier;
2469        struct hlist_node *node;
2470
2471        hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2472                notifier->ops->sched_out(notifier, next);
2473}
2474
2475#else /* !CONFIG_PREEMPT_NOTIFIERS */
2476
2477static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2478{
2479}
2480
2481static void
2482fire_sched_out_preempt_notifiers(struct task_struct *curr,
2483                                 struct task_struct *next)
2484{
2485}
2486
2487#endif /* CONFIG_PREEMPT_NOTIFIERS */
2488
2489/**
2490 * prepare_task_switch - prepare to switch tasks
2491 * @rq: the runqueue preparing to switch
2492 * @prev: the current task that is being switched out
2493 * @next: the task we are going to switch to.
2494 *
2495 * This is called with the rq lock held and interrupts off. It must
2496 * be paired with a subsequent finish_task_switch after the context
2497 * switch.
2498 *
2499 * prepare_task_switch sets up locking and calls architecture specific
2500 * hooks.
2501 */
2502static inline void
2503prepare_task_switch(struct rq *rq, struct task_struct *prev,
2504                    struct task_struct *next)
2505{
2506        fire_sched_out_preempt_notifiers(prev, next);
2507        prepare_lock_switch(rq, next);
2508        prepare_arch_switch(next);
2509}
2510
2511/**
2512 * finish_task_switch - clean up after a task-switch
2513 * @rq: runqueue associated with task-switch
2514 * @prev: the thread we just switched away from.
2515 *
2516 * finish_task_switch must be called after the context switch, paired
2517 * with a prepare_task_switch call before the context switch.
2518 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2519 * and do any other architecture-specific cleanup actions.
2520 *
2521 * Note that we may have delayed dropping an mm in context_switch(). If
2522 * so, we finish that here outside of the runqueue lock. (Doing it
2523 * with the lock held can cause deadlocks; see schedule() for
2524 * details.)
2525 */
2526static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2527        __releases(rq->lock)
2528{
2529        struct mm_struct *mm = rq->prev_mm;
2530        long prev_state;
2531
2532        rq->prev_mm = NULL;
2533
2534        /*
2535         * A task struct has one reference for the use as "current".
2536         * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2537         * schedule one last time. The schedule call will never return, and
2538         * the scheduled task must drop that reference.
2539         * The test for TASK_DEAD must occur while the runqueue locks are
2540         * still held, otherwise prev could be scheduled on another cpu, die
2541         * there before we look at prev->state, and then the reference would
2542         * be dropped twice.
2543         *              Manfred Spraul <manfred@colorfullife.com>
2544         */
2545        prev_state = prev->state;
2546        finish_arch_switch(prev);
2547        finish_lock_switch(rq, prev);
2548#ifdef CONFIG_SMP
2549        if (current->sched_class->post_schedule)
2550                current->sched_class->post_schedule(rq);
2551#endif
2552
2553        fire_sched_in_preempt_notifiers(current);
2554        if (mm)
2555                mmdrop(mm);
2556        if (unlikely(prev_state == TASK_DEAD)) {
2557                /*
2558                 * Remove function-return probe instances associated with this
2559                 * task and put them back on the free list.
2560                 */
2561                kprobe_flush_task(prev);
2562                put_task_struct(prev);
2563        }
2564}
2565
2566/**
2567 * schedule_tail - first thing a freshly forked thread must call.
2568 * @prev: the thread we just switched away from.
2569 */
2570asmlinkage void schedule_tail(struct task_struct *prev)
2571        __releases(rq->lock)
2572{
2573        struct rq *rq = this_rq();
2574
2575        finish_task_switch(rq, prev);
2576#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2577        /* In this case, finish_task_switch does not reenable preemption */
2578        preempt_enable();
2579#endif
2580        if (current->set_child_tid)
2581                put_user(task_pid_vnr(current), current->set_child_tid);
2582}
2583
2584/*
2585 * context_switch - switch to the new MM and the new
2586 * thread's register state.
2587 */
2588static inline void
2589context_switch(struct rq *rq, struct task_struct *prev,
2590               struct task_struct *next)
2591{
2592        struct mm_struct *mm, *oldmm;
2593
2594        prepare_task_switch(rq, prev, next);
2595        trace_mark(kernel_sched_schedule,
2596                "prev_pid %d next_pid %d prev_state %ld "
2597                "## rq %p prev %p next %p",
2598                prev->pid, next->pid, prev->state,
2599                rq, prev, next);
2600        mm = next->mm;
2601        oldmm = prev->active_mm;
2602        /*
2603         * For paravirt, this is coupled with an exit in switch_to to
2604         * combine the page table reload and the switch backend into
2605         * one hypercall.
2606         */
2607        arch_enter_lazy_cpu_mode();
2608
2609        if (unlikely(!mm)) {
2610                next->active_mm = oldmm;
2611                atomic_inc(&oldmm->mm_count);
2612                enter_lazy_tlb(oldmm, next);
2613        } else
2614                switch_mm(oldmm, mm, next);
2615
2616        if (unlikely(!prev->mm)) {
2617                prev->active_mm = NULL;
2618                rq->prev_mm = oldmm;
2619        }
2620        /*
2621         * Since the runqueue lock will be released by the next
2622         * task (which is an invalid locking op but in the case
2623         * of the scheduler it's an obvious special-case), so we
2624         * do an early lockdep release here:
2625         */
2626#ifndef __ARCH_WANT_UNLOCKED_CTXSW
2627        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2628#endif
2629
2630        /* Here we just switch the register state and the stack. */
2631        switch_to(prev, next, prev);
2632
2633        barrier();
2634        /*
2635         * this_rq must be evaluated again because prev may have moved
2636         * CPUs since it called schedule(), thus the 'rq' on its stack
2637         * frame will be invalid.
2638         */
2639        finish_task_switch(this_rq(), prev);
2640}
2641
2642/*
2643 * nr_running, nr_uninterruptible and nr_context_switches:
2644 *
2645 * externally visible scheduler statistics: current number of runnable
2646 * threads, current number of uninterruptible-sleeping threads, total
2647 * number of context switches performed since bootup.
2648 */
2649unsigned long nr_running(void)
2650{
2651        unsigned long i, sum = 0;
2652
2653        for_each_online_cpu(i)
2654                sum += cpu_rq(i)->nr_running;
2655
2656        return sum;
2657}
2658
2659unsigned long nr_uninterruptible(void)
2660{
2661        unsigned long i, sum = 0;
2662
2663        for_each_possible_cpu(i)
2664                sum += cpu_rq(i)->nr_uninterruptible;
2665
2666        /*
2667         * Since we read the counters lockless, it might be slightly
2668         * inaccurate. Do not allow it to go below zero though:
2669         */
2670        if (unlikely((long)sum < 0))
2671                sum = 0;
2672
2673        return sum;
2674}
2675
2676unsigned long long nr_context_switches(void)
2677{
2678        int i;
2679        unsigned long long sum = 0;
2680
2681        for_each_possible_cpu(i)
2682                sum += cpu_rq(i)->nr_switches;
2683
2684        return sum;
2685}
2686
2687unsigned long nr_iowait(void)
2688{
2689        unsigned long i, sum = 0;
2690
2691        for_each_possible_cpu(i)
2692                sum += atomic_read(&cpu_rq(i)->nr_iowait);
2693
2694        return sum;
2695}
2696
2697unsigned long nr_active(void)
2698{
2699        unsigned long i, running = 0, uninterruptible = 0;
2700
2701        for_each_online_cpu(i) {
2702                running += cpu_rq(i)->nr_running;
2703                uninterruptible += cpu_rq(i)->nr_uninterruptible;
2704        }
2705
2706        if (unlikely((long)uninterruptible < 0))
2707                uninterruptible = 0;
2708
2709        return running + uninterruptible;
2710}
2711
2712/*
2713 * Update rq->cpu_load[] statistics. This function is usually called every
2714 * scheduler tick (TICK_NSEC).
2715 */
2716static void update_cpu_load(struct rq *this_rq)
2717{
2718        unsigned long this_load = this_rq->load.weight;
2719        int i, scale;
2720
2721        this_rq->nr_load_updates++;
2722
2723        /* Update our load: */
2724        for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2725                unsigned long old_load, new_load;
2726
2727                /* scale is effectively 1 << i now, and >> i divides by scale */
2728
2729                old_load = this_rq->cpu_load[i];
2730                new_load = this_load;
2731                /*
2732                 * Round up the averaging division if load is increasing. This
2733                 * prevents us from getting stuck on 9 if the load is 10, for
2734                 * example.
2735                 */
2736                if (new_load > old_load)
2737                        new_load += scale-1;
2738                this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2739        }
2740}
2741
2742#ifdef CONFIG_SMP
2743
2744/*
2745 * double_rq_lock - safely lock two runqueues
2746 *
2747 * Note this does not disable interrupts like task_rq_lock,
2748 * you need to do so manually before calling.
2749 */
2750static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2751        __acquires(rq1->lock)
2752        __acquires(rq2->lock)
2753{
2754        BUG_ON(!irqs_disabled());
2755        if (rq1 == rq2) {
2756                spin_lock(&rq1->lock);
2757                __acquire(rq2->lock);   /* Fake it out ;) */
2758        } else {
2759                if (rq1 < rq2) {
2760                        spin_lock(&rq1->lock);
2761                        spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2762                } else {
2763                        spin_lock(&rq2->lock);
2764                        spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2765                }
2766        }
2767        update_rq_clock(rq1);
2768        update_rq_clock(rq2);
2769}
2770
2771/*
2772 * double_rq_unlock - safely unlock two runqueues
2773 *
2774 * Note this does not restore interrupts like task_rq_unlock,
2775 * you need to do so manually after calling.
2776 */
2777static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2778        __releases(rq1->lock)
2779        __releases(rq2->lock)
2780{
2781        spin_unlock(&rq1->lock);
2782        if (rq1 != rq2)
2783                spin_unlock(&rq2->lock);
2784        else
2785                __release(rq2->lock);
2786}
2787
2788/*
2789 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2790 */
2791static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2792        __releases(this_rq->lock)
2793        __acquires(busiest->lock)
2794        __acquires(this_rq->lock)
2795{
2796        int ret = 0;
2797
2798        if (unlikely(!irqs_disabled())) {
2799                /* printk() doesn't work good under rq->lock */
2800                spin_unlock(&this_rq->lock);
2801                BUG_ON(1);
2802        }
2803        if (unlikely(!spin_trylock(&busiest->lock))) {
2804                if (busiest < this_rq) {
2805                        spin_unlock(&this_rq->lock);
2806                        spin_lock(&busiest->lock);
2807                        spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2808                        ret = 1;
2809                } else
2810                        spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2811        }
2812        return ret;
2813}
2814
2815static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2816        __releases(busiest->lock)
2817{
2818        spin_unlock(&busiest->lock);
2819        lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2820}
2821
2822/*
2823 * If dest_cpu is allowed for this process, migrate the task to it.
2824 * This is accomplished by forcing the cpu_allowed mask to only
2825 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2826 * the cpu_allowed mask is restored.
2827 */
2828static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2829{
2830        struct migration_req req;
2831        unsigned long flags;
2832        struct rq *rq;
2833
2834        rq = task_rq_lock(p, &flags);
2835        if (!cpu_isset(dest_cpu, p->cpus_allowed)
2836            || unlikely(!cpu_active(dest_cpu)))
2837                goto out;
2838
2839        /* force the process onto the specified CPU */
2840        if (migrate_task(p, dest_cpu, &req)) {
2841                /* Need to wait for migration thread (might exit: take ref). */
2842                struct task_struct *mt = rq->migration_thread;
2843
2844                get_task_struct(mt);
2845                task_rq_unlock(rq, &flags);
2846                wake_up_process(mt);
2847                put_task_struct(mt);
2848                wait_for_completion(&req.done);
2849
2850                return;
2851        }
2852out:
2853        task_rq_unlock(rq, &flags);
2854}
2855
2856/*
2857 * sched_exec - execve() is a valuable balancing opportunity, because at
2858 * this point the task has the smallest effective memory and cache footprint.
2859 */
2860void sched_exec(void)
2861{
2862        int new_cpu, this_cpu = get_cpu();
2863        new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2864        put_cpu();
2865        if (new_cpu != this_cpu)
2866                sched_migrate_task(current, new_cpu);
2867}
2868
2869/*
2870 * pull_task - move a task from a remote runqueue to the local runqueue.
2871 * Both runqueues must be locked.
2872 */
2873static void pull_task(struct rq *src_rq, struct task_struct *p,
2874                      struct rq *this_rq, int this_cpu)
2875{
2876        deactivate_task(src_rq, p, 0);
2877        set_task_cpu(p, this_cpu);
2878        activate_task(this_rq, p, 0);
2879        /*
2880         * Note that idle threads have a prio of MAX_PRIO, for this test
2881         * to be always true for them.
2882         */
2883        check_preempt_curr(this_rq, p);
2884}
2885
2886/*
2887 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2888 */
2889static
2890int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2891                     struct sched_domain *sd, enum cpu_idle_type idle,
2892                     int *all_pinned)
2893{
2894        /*
2895         * We do not migrate tasks that are:
2896         * 1) running (obviously), or
2897         * 2) cannot be migrated to this CPU due to cpus_allowed, or
2898         * 3) are cache-hot on their current CPU.
2899         */
2900        if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2901                schedstat_inc(p, se.nr_failed_migrations_affine);
2902                return 0;
2903        }
2904        *all_pinned = 0;
2905
2906        if (task_running(rq, p)) {
2907                schedstat_inc(p, se.nr_failed_migrations_running);
2908                return 0;
2909        }
2910
2911        /*
2912         * Aggressive migration if:
2913         * 1) task is cache cold, or
2914         * 2) too many balance attempts have failed.
2915         */
2916
2917        if (!task_hot(p, rq->clock, sd) ||
2918                        sd->nr_balance_failed > sd->cache_nice_tries) {
2919#ifdef CONFIG_SCHEDSTATS
2920                if (task_hot(p, rq->clock, sd)) {
2921                        schedstat_inc(sd, lb_hot_gained[idle]);
2922                        schedstat_inc(p, se.nr_forced_migrations);
2923                }
2924#endif
2925                return 1;
2926        }
2927
2928        if (task_hot(p, rq->clock, sd)) {
2929                schedstat_inc(p, se.nr_failed_migrations_hot);
2930                return 0;
2931        }
2932        return 1;
2933}
2934
2935static unsigned long
2936balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2937              unsigned long max_load_move, struct sched_domain *sd,
2938              enum cpu_idle_type idle, int *all_pinned,
2939              int *this_best_prio, struct rq_iterator *iterator)
2940{
2941        int loops = 0, pulled = 0, pinned = 0;
2942        struct task_struct *p;
2943        long rem_load_move = max_load_move;
2944
2945        if (max_load_move == 0)
2946                goto out;
2947
2948        pinned = 1;
2949
2950        /*
2951         * Start the load-balancing iterator:
2952         */
2953        p = iterator->start(iterator->arg);
2954next:
2955        if (!p || loops++ > sysctl_sched_nr_migrate)
2956                goto out;
2957
2958        if ((p->se.load.weight >> 1) > rem_load_move ||
2959            !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2960                p = iterator->next(iterator->arg);
2961                goto next;
2962        }
2963
2964        pull_task(busiest, p, this_rq, this_cpu);
2965        pulled++;
2966        rem_load_move -= p->se.load.weight;
2967
2968        /*
2969         * We only want to steal up to the prescribed amount of weighted load.
2970         */
2971        if (rem_load_move > 0) {
2972                if (p->prio < *this_best_prio)
2973                        *this_best_prio = p->prio;
2974                p = iterator->next(iterator->arg);
2975                goto next;
2976        }
2977out:
2978        /*
2979         * Right now, this is one of only two places pull_task() is called,
2980         * so we can safely collect pull_task() stats here rather than
2981         * inside pull_task().
2982         */
2983        schedstat_add(sd, lb_gained[idle], pulled);
2984
2985        if (all_pinned)
2986                *all_pinned = pinned;
2987
2988        return max_load_move - rem_load_move;
2989}
2990
2991/*
2992 * move_tasks tries to move up to max_load_move weighted load from busiest to
2993 * this_rq, as part of a balancing operation within domain "sd".
2994 * Returns 1 if successful and 0 otherwise.
2995 *
2996 * Called with both runqueues locked.
2997 */
2998static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2999                      unsigned long max_load_move,
3000                      struct sched_domain *sd, enum cpu_idle_type idle,
3001                      int *all_pinned)
3002{
3003        const struct sched_class *class = sched_class_highest;
3004        unsigned long total_load_moved = 0;
3005        int this_best_prio = this_rq->curr->prio;
3006
3007        do {
3008                total_load_moved +=
3009                        class->load_balance(this_rq, this_cpu, busiest,
3010                                max_load_move - total_load_moved,
3011                                sd, idle, all_pinned, &this_best_prio);
3012                class = class->next;
3013
3014                if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3015                        break;
3016
3017        } while (class && max_load_move > total_load_moved);
3018
3019        return total_load_moved > 0;
3020}
3021
3022static int
3023iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3024                   struct sched_domain *sd, enum cpu_idle_type idle,
3025                   struct rq_iterator *iterator)
3026{
3027        struct task_struct *p = iterator->start(iterator->arg);
3028        int pinned = 0;
3029
3030        while (p) {
3031                if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3032                        pull_task(busiest, p, this_rq, this_cpu);
3033                        /*
3034                         * Right now, this is only the second place pull_task()
3035                         * is called, so we can safely collect pull_task()
3036                         * stats here rather than inside pull_task().
3037                         */
3038                        schedstat_inc(sd, lb_gained[idle]);
3039
3040                        return 1;
3041                }
3042                p = iterator->next(iterator->arg);
3043        }
3044
3045        return 0;
3046}
3047
3048/*
3049 * move_one_task tries to move exactly one task from busiest to this_rq, as
3050 * part of active balancing operations within "domain".
3051 * Returns 1 if successful and 0 otherwise.
3052 *
3053 * Called with both runqueues locked.
3054 */
3055static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3056                         struct sched_domain *sd, enum cpu_idle_type idle)
3057{
3058        const struct sched_class *class;
3059
3060        for (class = sched_class_highest; class; class = class->next)
3061                if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3062                        return 1;
3063
3064        return 0;
3065}
3066
3067/*
3068 * find_busiest_group finds and returns the busiest CPU group within the
3069 * domain. It calculates and returns the amount of weighted load which
3070 * should be moved to restore balance via the imbalance parameter.
3071 */
3072static struct sched_group *
3073find_busiest_group(struct sched_domain *sd, int this_cpu,
3074                   unsigned long *imbalance, enum cpu_idle_type idle,
3075                   int *sd_idle, const cpumask_t *cpus, int *balance)
3076{
3077        struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3078        unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3079        unsigned long max_pull;
3080        unsigned long busiest_load_per_task, busiest_nr_running;
3081        unsigned long this_load_per_task, this_nr_running;
3082        int load_idx, group_imb = 0;
3083#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3084        int power_savings_balance = 1;
3085        unsigned long leader_nr_running = 0, min_load_per_task = 0;
3086        unsigned long min_nr_running = ULONG_MAX;
3087        struct sched_group *group_min = NULL, *group_leader = NULL;
3088#endif
3089
3090        max_load = this_load = total_load = total_pwr = 0;
3091        busiest_load_per_task = busiest_nr_running = 0;
3092        this_load_per_task = this_nr_running = 0;
3093
3094        if (idle == CPU_NOT_IDLE)
3095                load_idx = sd->busy_idx;
3096        else if (idle == CPU_NEWLY_IDLE)
3097                load_idx = sd->newidle_idx;
3098        else
3099                load_idx = sd->idle_idx;
3100
3101        do {
3102                unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3103                int local_group;
3104                int i;
3105                int __group_imb = 0;
3106                unsigned int balance_cpu = -1, first_idle_cpu = 0;
3107                unsigned long sum_nr_running, sum_weighted_load;
3108                unsigned long sum_avg_load_per_task;
3109                unsigned long avg_load_per_task;
3110
3111                local_group = cpu_isset(this_cpu, group->cpumask);
3112
3113                if (local_group)
3114                        balance_cpu = first_cpu(group->cpumask);
3115
3116                /* Tally up the load of all CPUs in the group */
3117                sum_weighted_load = sum_nr_running = avg_load = 0;
3118                sum_avg_load_per_task = avg_load_per_task = 0;
3119
3120                max_cpu_load = 0;
3121                min_cpu_load = ~0UL;
3122
3123                for_each_cpu_mask_nr(i, group->cpumask) {
3124                        struct rq *rq;
3125
3126                        if (!cpu_isset(i, *cpus))
3127                                continue;
3128
3129                        rq = cpu_rq(i);
3130
3131                        if (*sd_idle && rq->nr_running)
3132                                *sd_idle = 0;
3133
3134                        /* Bias balancing toward cpus of our domain */
3135                        if (local_group) {
3136                                if (idle_cpu(i) && !first_idle_cpu) {
3137                                        first_idle_cpu = 1;
3138                                        balance_cpu = i;
3139                                }
3140
3141                                load = target_load(i, load_idx);
3142                        } else {
3143                                load = source_load(i, load_idx);
3144                                if (load > max_cpu_load)
3145                                        max_cpu_load = load;
3146                                if (min_cpu_load > load)
3147                                        min_cpu_load = load;
3148                        }
3149
3150                        avg_load += load;
3151                        sum_nr_running += rq->nr_running;
3152                        sum_weighted_load += weighted_cpuload(i);
3153
3154                        sum_avg_load_per_task += cpu_avg_load_per_task(i);
3155                }
3156
3157                /*
3158                 * First idle cpu or the first cpu(busiest) in this sched group
3159                 * is eligible for doing load balancing at this and above
3160                 * domains. In the newly idle case, we will allow all the cpu's
3161                 * to do the newly idle load balance.
3162                 */
3163                if (idle != CPU_NEWLY_IDLE && local_group &&
3164                    balance_cpu != this_cpu && balance) {
3165                        *balance = 0;
3166                        goto ret;
3167                }
3168
3169                total_load += avg_load;
3170                total_pwr += group->__cpu_power;
3171
3172                /* Adjust by relative CPU power of the group */
3173                avg_load = sg_div_cpu_power(group,
3174                                avg_load * SCHED_LOAD_SCALE);
3175
3176
3177                /*
3178                 * Consider the group unbalanced when the imbalance is larger
3179                 * than the average weight of two tasks.
3180                 *
3181                 * APZ: with cgroup the avg task weight can vary wildly and
3182                 *      might not be a suitable number - should we keep a
3183                 *      normalized nr_running number somewhere that negates
3184                 *      the hierarchy?
3185                 */
3186                avg_load_per_task = sg_div_cpu_power(group,
3187                                sum_avg_load_per_task * SCHED_LOAD_SCALE);
3188
3189                if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3190                        __group_imb = 1;
3191
3192                group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3193
3194                if (local_group) {
3195                        this_load = avg_load;
3196                        this = group;
3197                        this_nr_running = sum_nr_running;
3198                        this_load_per_task = sum_weighted_load;
3199                } else if (avg_load > max_load &&
3200                           (sum_nr_running > group_capacity || __group_imb)) {
3201                        max_load = avg_load;
3202                        busiest = group;
3203                        busiest_nr_running = sum_nr_running;
3204                        busiest_load_per_task = sum_weighted_load;
3205                        group_imb = __group_imb;
3206                }
3207
3208#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3209                /*
3210                 * Busy processors will not participate in power savings
3211                 * balance.
3212                 */
3213                if (idle == CPU_NOT_IDLE ||
3214                                !(sd->flags & SD_POWERSAVINGS_BALANCE))
3215                        goto group_next;
3216
3217                /*
3218                 * If the local group is idle or completely loaded
3219                 * no need to do power savings balance at this domain
3220                 */
3221                if (local_group && (this_nr_running >= group_capacity ||
3222                                    !this_nr_running))
3223                        power_savings_balance = 0;
3224
3225                /*
3226                 * If a group is already running at full capacity or idle,
3227                 * don't include that group in power savings calculations
3228                 */
3229                if (!power_savings_balance || sum_nr_running >= group_capacity
3230                    || !sum_nr_running)
3231                        goto group_next;
3232
3233                /*
3234                 * Calculate the group which has the least non-idle load.
3235                 * This is the group from where we need to pick up the load
3236                 * for saving power
3237                 */
3238                if ((sum_nr_running < min_nr_running) ||
3239                    (sum_nr_running == min_nr_running &&
3240                     first_cpu(group->cpumask) <
3241                     first_cpu(group_min->cpumask))) {
3242                        group_min = group;
3243                        min_nr_running = sum_nr_running;
3244                        min_load_per_task = sum_weighted_load /
3245                                                sum_nr_running;
3246                }
3247
3248                /*
3249                 * Calculate the group which is almost near its
3250                 * capacity but still has some space to pick up some load
3251                 * from other group and save more power
3252                 */
3253                if (sum_nr_running <= group_capacity - 1) {
3254                        if (sum_nr_running > leader_nr_running ||
3255                            (sum_nr_running == leader_nr_running &&
3256                             first_cpu(group->cpumask) >
3257                              first_cpu(group_leader->cpumask))) {
3258                                group_leader = group;
3259                                leader_nr_running = sum_nr_running;
3260                        }
3261                }
3262group_next:
3263#endif
3264                group = group->next;
3265        } while (group != sd->groups);
3266
3267        if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3268                goto out_balanced;
3269
3270        avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3271
3272        if (this_load >= avg_load ||
3273                        100*max_load <= sd->imbalance_pct*this_load)
3274                goto out_balanced;
3275
3276        busiest_load_per_task /= busiest_nr_running;
3277        if (group_imb)
3278                busiest_load_per_task = min(busiest_load_per_task, avg_load);
3279
3280        /*
3281         * We're trying to get all the cpus to the average_load, so we don't
3282         * want to push ourselves above the average load, nor do we wish to
3283         * reduce the max loaded cpu below the average load, as either of these
3284         * actions would just result in more rebalancing later, and ping-pong
3285         * tasks around. Thus we look for the minimum possible imbalance.
3286         * Negative imbalances (*we* are more loaded than anyone else) will
3287         * be counted as no imbalance for these purposes -- we can't fix that
3288         * by pulling tasks to us. Be careful of negative numbers as they'll
3289         * appear as very large values with unsigned longs.
3290         */
3291        if (max_load <= busiest_load_per_task)
3292                goto out_balanced;
3293
3294        /*
3295         * In the presence of smp nice balancing, certain scenarios can have
3296         * max load less than avg load(as we skip the groups at or below
3297         * its cpu_power, while calculating max_load..)
3298         */
3299        if (max_load < avg_load) {
3300                *imbalance = 0;
3301                goto small_imbalance;
3302        }
3303
3304        /* Don't want to pull so many tasks that a group would go idle */
3305        max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3306
3307        /* How much load to actually move to equalise the imbalance */
3308        *imbalance = min(max_pull * busiest->__cpu_power,
3309                                (avg_load - this_load) * this->__cpu_power)
3310                        / SCHED_LOAD_SCALE;
3311
3312        /*
3313         * if *imbalance is less than the average load per runnable task
3314         * there is no gaurantee that any tasks will be moved so we'll have
3315         * a think about bumping its value to force at least one task to be
3316         * moved
3317         */
3318        if (*imbalance < busiest_load_per_task) {
3319                unsigned long tmp, pwr_now, pwr_move;
3320                unsigned int imbn;
3321
3322small_imbalance:
3323                pwr_move = pwr_now = 0;
3324                imbn = 2;
3325                if (this_nr_running) {
3326                        this_load_per_task /= this_nr_running;
3327                        if (busiest_load_per_task > this_load_per_task)
3328                                imbn = 1;
3329                } else
3330                        this_load_per_task = cpu_avg_load_per_task(this_cpu);
3331
3332                if (max_load - this_load + 2*busiest_load_per_task >=
3333                                        busiest_load_per_task * imbn) {
3334                        *imbalance = busiest_load_per_task;
3335                        return busiest;
3336                }
3337
3338                /*
3339                 * OK, we don't have enough imbalance to justify moving tasks,
3340                 * however we may be able to increase total CPU power used by
3341                 * moving them.
3342                 */
3343
3344                pwr_now += busiest->__cpu_power *
3345                                min(busiest_load_per_task, max_load);
3346                pwr_now += this->__cpu_power *
3347                                min(this_load_per_task, this_load);
3348                pwr_now /= SCHED_LOAD_SCALE;
3349
3350                /* Amount of load we'd subtract */
3351                tmp = sg_div_cpu_power(busiest,
3352                                busiest_load_per_task * SCHED_LOAD_SCALE);
3353                if (max_load > tmp)
3354                        pwr_move += busiest->__cpu_power *
3355                                min(busiest_load_per_task, max_load - tmp);
3356
3357                /* Amount of load we'd add */
3358                if (max_load * busiest->__cpu_power <
3359                                busiest_load_per_task * SCHED_LOAD_SCALE)
3360                        tmp = sg_div_cpu_power(this,
3361                                        max_load * busiest->__cpu_power);
3362                else
3363                        tmp = sg_div_cpu_power(this,
3364                                busiest_load_per_task * SCHED_LOAD_SCALE);
3365                pwr_move += this->__cpu_power *
3366                                min(this_load_per_task, this_load + tmp);
3367                pwr_move /= SCHED_LOAD_SCALE;
3368
3369                /* Move if we gain throughput */
3370                if (pwr_move > pwr_now)
3371                        *imbalance = busiest_load_per_task;
3372        }
3373
3374        return busiest;
3375
3376out_balanced:
3377#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3378        if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3379                goto ret;
3380
3381        if (this == group_leader && group_leader != group_min) {
3382                *imbalance = min_load_per_task;
3383                return group_min;
3384        }
3385#endif
3386ret:
3387        *imbalance = 0;
3388        return NULL;
3389}
3390
3391/*
3392 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3393 */
3394static struct rq *
3395find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3396                   unsigned long imbalance, const cpumask_t *cpus)
3397{
3398        struct rq *busiest = NULL, *rq;
3399        unsigned long max_load = 0;
3400        int i;
3401
3402        for_each_cpu_mask_nr(i, group->cpumask) {
3403                unsigned long wl;
3404
3405                if (!cpu_isset(i, *cpus))
3406                        continue;
3407
3408                rq = cpu_rq(i);
3409                wl = weighted_cpuload(i);
3410
3411                if (rq->nr_running == 1 && wl > imbalance)
3412                        continue;
3413
3414                if (wl > max_load) {
3415                        max_load = wl;
3416                        busiest = rq;
3417                }
3418        }
3419
3420        return busiest;
3421}
3422
3423/*
3424 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3425 * so long as it is large enough.
3426 */
3427#define MAX_PINNED_INTERVAL     512
3428
3429/*
3430 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3431 * tasks if there is an imbalance.
3432 */
3433static int load_balance(int this_cpu, struct rq *this_rq,
3434                        struct sched_domain *sd, enum cpu_idle_type idle,
3435                        int *balance, cpumask_t *cpus)
3436{
3437        int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3438        struct sched_group *group;
3439        unsigned long imbalance;
3440        struct rq *busiest;
3441        unsigned long flags;
3442
3443        cpus_setall(*cpus);
3444
3445        /*
3446         * When power savings policy is enabled for the parent domain, idle
3447         * sibling can pick up load irrespective of busy siblings. In this case,
3448         * let the state of idle sibling percolate up as CPU_IDLE, instead of
3449         * portraying it as CPU_NOT_IDLE.
3450         */
3451        if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3452            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3453                sd_idle = 1;
3454
3455        schedstat_inc(sd, lb_count[idle]);
3456
3457redo:
3458        update_shares(sd);
3459        group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3460                                   cpus, balance);
3461
3462        if (*balance == 0)
3463                goto out_balanced;
3464
3465        if (!group) {
3466                schedstat_inc(sd, lb_nobusyg[idle]);
3467                goto out_balanced;
3468        }
3469
3470        busiest = find_busiest_queue(group, idle, imbalance, cpus);
3471        if (!busiest) {
3472                schedstat_inc(sd, lb_nobusyq[idle]);
3473                goto out_balanced;
3474        }
3475
3476        BUG_ON(busiest == this_rq);
3477
3478        schedstat_add(sd, lb_imbalance[idle], imbalance);
3479
3480        ld_moved = 0;
3481        if (busiest->nr_running > 1) {
3482                /*
3483                 * Attempt to move tasks. If find_busiest_group has found
3484                 * an imbalance but busiest->nr_running <= 1, the group is
3485                 * still unbalanced. ld_moved simply stays zero, so it is
3486                 * correctly treated as an imbalance.
3487                 */
3488                local_irq_save(flags);
3489                double_rq_lock(this_rq, busiest);
3490                ld_moved = move_tasks(this_rq, this_cpu, busiest,
3491                                      imbalance, sd, idle, &all_pinned);
3492                double_rq_unlock(this_rq, busiest);
3493                local_irq_restore(flags);
3494
3495                /*
3496                 * some other cpu did the load balance for us.
3497                 */
3498                if (ld_moved && this_cpu != smp_processor_id())
3499                        resched_cpu(this_cpu);
3500
3501                /* All tasks on this runqueue were pinned by CPU affinity */
3502                if (unlikely(all_pinned)) {
3503                        cpu_clear(cpu_of(busiest), *cpus);
3504                        if (!cpus_empty(*cpus))
3505                                goto redo;
3506                        goto out_balanced;
3507                }
3508        }
3509
3510        if (!ld_moved) {
3511                schedstat_inc(sd, lb_failed[idle]);
3512                sd->nr_balance_failed++;
3513
3514                if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3515
3516                        spin_lock_irqsave(&busiest->lock, flags);
3517
3518                        /* don't kick the migration_thread, if the curr
3519                         * task on busiest cpu can't be moved to this_cpu
3520                         */
3521                        if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3522                                spin_unlock_irqrestore(&busiest->lock, flags);
3523                                all_pinned = 1;
3524                                goto out_one_pinned;
3525                        }
3526
3527                        if (!busiest->active_balance) {
3528                                busiest->active_balance = 1;
3529                                busiest->push_cpu = this_cpu;
3530                                active_balance = 1;
3531                        }
3532                        spin_unlock_irqrestore(&busiest->lock, flags);
3533                        if (active_balance)
3534                                wake_up_process(busiest->migration_thread);
3535
3536                        /*
3537                         * We've kicked active balancing, reset the failure
3538                         * counter.
3539                         */
3540                        sd->nr_balance_failed = sd->cache_nice_tries+1;
3541                }
3542        } else
3543                sd->nr_balance_failed = 0;
3544
3545        if (likely(!active_balance)) {
3546                /* We were unbalanced, so reset the balancing interval */
3547                sd->balance_interval = sd->min_interval;
3548        } else {
3549                /*
3550                 * If we've begun active balancing, start to back off. This
3551                 * case may not be covered by the all_pinned logic if there
3552                 * is only 1 task on the busy runqueue (because we don't call
3553                 * move_tasks).
3554                 */
3555                if (sd->balance_interval < sd->max_interval)
3556                        sd->balance_interval *= 2;
3557        }
3558
3559        if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3560            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3561                ld_moved = -1;
3562
3563        goto out;
3564
3565out_balanced:
3566        schedstat_inc(sd, lb_balanced[idle]);
3567
3568        sd->nr_balance_failed = 0;
3569
3570out_one_pinned:
3571        /* tune up the balancing interval */
3572        if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3573                        (sd->balance_interval < sd->max_interval))
3574                sd->balance_interval *= 2;
3575
3576        if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3577            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3578                ld_moved = -1;
3579        else
3580                ld_moved = 0;
3581out:
3582        if (ld_moved)
3583                update_shares(sd);
3584        return ld_moved;
3585}
3586
3587/*
3588 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3589 * tasks if there is an imbalance.
3590 *
3591 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3592 * this_rq is locked.
3593 */
3594static int
3595load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3596                        cpumask_t *cpus)
3597{
3598        struct sched_group *group;
3599        struct rq *busiest = NULL;
3600        unsigned long imbalance;
3601        int ld_moved = 0;
3602        int sd_idle = 0;
3603        int all_pinned = 0;
3604
3605        cpus_setall(*cpus);
3606
3607        /*
3608         * When power savings policy is enabled for the parent domain, idle
3609         * sibling can pick up load irrespective of busy siblings. In this case,
3610         * let the state of idle sibling percolate up as IDLE, instead of
3611         * portraying it as CPU_NOT_IDLE.
3612         */
3613        if (sd->flags & SD_SHARE_CPUPOWER &&
3614            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3615                sd_idle = 1;
3616
3617        schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3618redo:
3619        update_shares_locked(this_rq, sd);
3620        group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3621                                   &sd_idle, cpus, NULL);
3622        if (!group) {
3623                schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3624                goto out_balanced;
3625        }
3626
3627        busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3628        if (!busiest) {
3629                schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3630                goto out_balanced;
3631        }
3632
3633        BUG_ON(busiest == this_rq);
3634
3635        schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3636
3637        ld_moved = 0;
3638        if (busiest->nr_running > 1) {
3639                /* Attempt to move tasks */
3640                double_lock_balance(this_rq, busiest);
3641                /* this_rq->clock is already updated */
3642                update_rq_clock(busiest);
3643                ld_moved = move_tasks(this_rq, this_cpu, busiest,
3644                                        imbalance, sd, CPU_NEWLY_IDLE,
3645                                        &all_pinned);
3646                double_unlock_balance(this_rq, busiest);
3647
3648                if (unlikely(all_pinned)) {
3649                        cpu_clear(cpu_of(busiest), *cpus);
3650                        if (!cpus_empty(*cpus))
3651                                goto redo;
3652                }
3653        }
3654
3655        if (!ld_moved) {
3656                schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3657                if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3658                    !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3659                        return -1;
3660        } else
3661                sd->nr_balance_failed = 0;
3662
3663        update_shares_locked(this_rq, sd);
3664        return ld_moved;
3665
3666out_balanced:
3667        schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3668        if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3669            !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3670                return -1;
3671        sd->nr_balance_failed = 0;
3672
3673        return 0;
3674}
3675
3676/*
3677 * idle_balance is called by schedule() if this_cpu is about to become
3678 * idle. Attempts to pull tasks from other CPUs.
3679 */
3680static void idle_balance(int this_cpu, struct rq *this_rq)
3681{
3682        struct sched_domain *sd;
3683        int pulled_task = -1;
3684        unsigned long next_balance = jiffies + HZ;
3685        cpumask_t tmpmask;
3686
3687        for_each_domain(this_cpu, sd) {
3688                unsigned long interval;
3689
3690                if (!(sd->flags & SD_LOAD_BALANCE))
3691                        continue;
3692
3693                if (sd->flags & SD_BALANCE_NEWIDLE)
3694                        /* If we've pulled tasks over stop searching: */
3695                        pulled_task = load_balance_newidle(this_cpu, this_rq,
3696                                                           sd, &tmpmask);
3697
3698                interval = msecs_to_jiffies(sd->balance_interval);
3699                if (time_after(next_balance, sd->last_balance + interval))
3700                        next_balance = sd->last_balance + interval;
3701                if (pulled_task)
3702                        break;
3703        }
3704        if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3705                /*
3706                 * We are going idle. next_balance may be set based on
3707                 * a busy processor. So reset next_balance.
3708                 */
3709                this_rq->next_balance = next_balance;
3710        }
3711}
3712
3713/*
3714 * active_load_balance is run by migration threads. It pushes running tasks
3715 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3716 * running on each physical CPU where possible, and avoids physical /
3717 * logical imbalances.
3718 *
3719 * Called with busiest_rq locked.
3720 */
3721static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3722{
3723        int target_cpu = busiest_rq->push_cpu;
3724        struct sched_domain *sd;
3725        struct rq *target_rq;
3726
3727        /* Is there any task to move? */
3728        if (busiest_rq->nr_running <= 1)
3729                return;
3730
3731        target_rq = cpu_rq(target_cpu);
3732
3733        /*
3734         * This condition is "impossible", if it occurs
3735         * we need to fix it. Originally reported by
3736         * Bjorn Helgaas on a 128-cpu setup.
3737         */
3738        BUG_ON(busiest_rq == target_rq);
3739
3740        /* move a task from busiest_rq to target_rq */
3741        double_lock_balance(busiest_rq, target_rq);
3742        update_rq_clock(busiest_rq);
3743        update_rq_clock(target_rq);
3744
3745        /* Search for an sd spanning us and the target CPU. */
3746        for_each_domain(target_cpu, sd) {
3747                if ((sd->flags & SD_LOAD_BALANCE) &&
3748                    cpu_isset(busiest_cpu, sd->span))
3749                                break;
3750        }
3751
3752        if (likely(sd)) {
3753                schedstat_inc(sd, alb_count);
3754
3755                if (move_one_task(target_rq, target_cpu, busiest_rq,
3756                                  sd, CPU_IDLE))
3757                        schedstat_inc(sd, alb_pushed);
3758                else
3759                        schedstat_inc(sd, alb_failed);
3760        }
3761        double_unlock_balance(busiest_rq, target_rq);
3762}
3763
3764#ifdef CONFIG_NO_HZ
3765static struct {
3766        atomic_t load_balancer;
3767        cpumask_t cpu_mask;
3768} nohz ____cacheline_aligned = {
3769        .load_balancer = ATOMIC_INIT(-1),
3770        .cpu_mask = CPU_MASK_NONE,
3771};
3772
3773/*
3774 * This routine will try to nominate the ilb (idle load balancing)
3775 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3776 * load balancing on behalf of all those cpus. If all the cpus in the system
3777 * go into this tickless mode, then there will be no ilb owner (as there is
3778 * no need for one) and all the cpus will sleep till the next wakeup event
3779 * arrives...
3780 *
3781 * For the ilb owner, tick is not stopped. And this tick will be used
3782 * for idle load balancing. ilb owner will still be part of
3783 * nohz.cpu_mask..
3784 *
3785 * While stopping the tick, this cpu will become the ilb owner if there
3786 * is no other owner. And will be the owner till that cpu becomes busy
3787 * or if all cpus in the system stop their ticks at which point
3788 * there is no need for ilb owner.
3789 *
3790 * When the ilb owner becomes busy, it nominates another owner, during the
3791 * next busy scheduler_tick()
3792 */
3793int select_nohz_load_balancer(int stop_tick)
3794{
3795        int cpu = smp_processor_id();
3796
3797        if (stop_tick) {
3798                cpu_set(cpu, nohz.cpu_mask);
3799                cpu_rq(cpu)->in_nohz_recently = 1;
3800
3801                /*
3802                 * If we are going offline and still the leader, give up!
3803                 */
3804                if (!cpu_active(cpu) &&
3805                    atomic_read(&nohz.load_balancer) == cpu) {
3806                        if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3807                                BUG();
3808                        return 0;
3809                }
3810
3811                /* time for ilb owner also to sleep */
3812                if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3813                        if (atomic_read(&nohz.load_balancer) == cpu)
3814                                atomic_set(&nohz.load_balancer, -1);
3815                        return 0;
3816                }
3817
3818                if (atomic_read(&nohz.load_balancer) == -1) {
3819                        /* make me the ilb owner */
3820                        if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3821                                return 1;
3822                } else if (atomic_read(&nohz.load_balancer) == cpu)
3823                        return 1;
3824        } else {
3825                if (!cpu_isset(cpu, nohz.cpu_mask))
3826                        return 0;
3827
3828                cpu_clear(cpu, nohz.cpu_mask);
3829
3830                if (atomic_read(&nohz.load_balancer) == cpu)
3831                        if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3832                                BUG();
3833        }
3834        return 0;
3835}
3836#endif
3837
3838static DEFINE_SPINLOCK(balancing);
3839
3840/*
3841 * It checks each scheduling domain to see if it is due to be balanced,
3842 * and initiates a balancing operation if so.
3843 *
3844 * Balancing parameters are set up in arch_init_sched_domains.
3845 */
3846static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3847{
3848        int balance = 1;
3849        struct rq *rq = cpu_rq(cpu);
3850        unsigned long interval;
3851        struct sched_domain *sd;
3852        /* Earliest time when we have to do rebalance again */
3853        unsigned long next_balance = jiffies + 60*HZ;
3854        int update_next_balance = 0;
3855        int need_serialize;
3856        cpumask_t tmp;
3857
3858        for_each_domain(cpu, sd) {
3859                if (!(sd->flags & SD_LOAD_BALANCE))
3860                        continue;
3861
3862                interval = sd->balance_interval;
3863                if (idle != CPU_IDLE)
3864                        interval *= sd->busy_factor;
3865
3866                /* scale ms to jiffies */
3867                interval = msecs_to_jiffies(interval);
3868                if (unlikely(!interval))
3869                        interval = 1;
3870                if (interval > HZ*NR_CPUS/10)
3871                        interval = HZ*NR_CPUS/10;
3872
3873                need_serialize = sd->flags & SD_SERIALIZE;
3874
3875                if (need_serialize) {
3876                        if (!spin_trylock(&balancing))
3877                                goto out;
3878                }
3879
3880                if (time_after_eq(jiffies, sd->last_balance + interval)) {
3881                        if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3882                                /*
3883                                 * We've pulled tasks over so either we're no
3884                                 * longer idle, or one of our SMT siblings is
3885                                 * not idle.
3886                                 */
3887                                idle = CPU_NOT_IDLE;
3888                        }
3889                        sd->last_balance = jiffies;
3890                }
3891                if (need_serialize)
3892                        spin_unlock(&balancing);
3893out:
3894                if (time_after(next_balance, sd->last_balance + interval)) {
3895                        next_balance = sd->last_balance + interval;
3896                        update_next_balance = 1;
3897                }
3898
3899                /*
3900                 * Stop the load balance at this level. There is another
3901                 * CPU in our sched group which is doing load balancing more
3902                 * actively.
3903                 */
3904                if (!balance)
3905                        break;
3906        }
3907
3908        /*
3909         * next_balance will be updated only when there is a need.
3910         * When the cpu is attached to null domain for ex, it will not be
3911         * updated.
3912         */
3913        if (likely(update_next_balance))
3914                rq->next_balance = next_balance;
3915}
3916
3917/*
3918 * run_rebalance_domains is triggered when needed from the scheduler tick.
3919 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3920 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3921 */
3922static void run_rebalance_domains(struct softirq_action *h)
3923{
3924        int this_cpu = smp_processor_id();
3925        struct rq *this_rq = cpu_rq(this_cpu);
3926        enum cpu_idle_type idle = this_rq->idle_at_tick ?
3927                                                CPU_IDLE : CPU_NOT_IDLE;
3928
3929        rebalance_domains(this_cpu, idle);
3930
3931#ifdef CONFIG_NO_HZ
3932        /*
3933         * If this cpu is the owner for idle load balancing, then do the
3934         * balancing on behalf of the other idle cpus whose ticks are
3935         * stopped.
3936         */
3937        if (this_rq->idle_at_tick &&
3938            atomic_read(&nohz.load_balancer) == this_cpu) {
3939                cpumask_t cpus = nohz.cpu_mask;
3940                struct rq *rq;
3941                int balance_cpu;
3942
3943                cpu_clear(this_cpu, cpus);
3944                for_each_cpu_mask_nr(balance_cpu, cpus) {
3945                        /*
3946                         * If this cpu gets work to do, stop the load balancing
3947                         * work being done for other cpus. Next load
3948                         * balancing owner will pick it up.
3949                         */
3950                        if (need_resched())
3951                                break;
3952
3953                        rebalance_domains(balance_cpu, CPU_IDLE);
3954
3955                        rq = cpu_rq(balance_cpu);
3956                        if (time_after(this_rq->next_balance, rq->next_balance))
3957                                this_rq->next_balance = rq->next_balance;
3958                }
3959        }
3960#endif
3961}
3962
3963/*
3964 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3965 *
3966 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3967 * idle load balancing owner or decide to stop the periodic load balancing,
3968 * if the whole system is idle.
3969 */
3970static inline void trigger_load_balance(struct rq *rq, int cpu)
3971{
3972#ifdef CONFIG_NO_HZ
3973        /*
3974         * If we were in the nohz mode recently and busy at the current
3975         * scheduler tick, then check if we need to nominate new idle
3976         * load balancer.
3977         */
3978        if (rq->in_nohz_recently && !rq->idle_at_tick) {
3979                rq->in_nohz_recently = 0;
3980
3981                if (atomic_read(&nohz.load_balancer) == cpu) {
3982                        cpu_clear(cpu, nohz.cpu_mask);
3983                        atomic_set(&nohz.load_balancer, -1);
3984                }
3985
3986                if (atomic_read(&nohz.load_balancer) == -1) {
3987                        /*
3988                         * simple selection for now: Nominate the
3989                         * first cpu in the nohz list to be the next
3990                         * ilb owner.
3991                         *
3992                         * TBD: Traverse the sched domains and nominate
3993                         * the nearest cpu in the nohz.cpu_mask.
3994                         */
3995                        int ilb = first_cpu(nohz.cpu_mask);
3996
3997                        if (ilb < nr_cpu_ids)
3998                                resched_cpu(ilb);
3999                }
4000        }
4001
4002        /*
4003         * If this cpu is idle and doing idle load balancing for all the
4004         * cpus with ticks stopped, is it time for that to stop?
4005         */
4006        if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4007            cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4008                resched_cpu(cpu);
4009                return;
4010        }
4011
4012        /*
4013         * If this cpu is idle and the idle load balancing is done by
4014         * someone else, then no need raise the SCHED_SOFTIRQ
4015         */
4016        if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4017            cpu_isset(cpu, nohz.cpu_mask))
4018                return;
4019#endif
4020        if (time_after_eq(jiffies, rq->next_balance))
4021                raise_softirq(SCHED_SOFTIRQ);
4022}
4023
4024#else   /* CONFIG_SMP */
4025
4026/*
4027 * on UP we do not need to balance between CPUs:
4028 */
4029static inline void idle_balance(int cpu, struct rq *rq)
4030{
4031}
4032
4033#endif
4034
4035DEFINE_PER_CPU(struct kernel_stat, kstat);
4036
4037EXPORT_PER_CPU_SYMBOL(kstat);
4038
4039/*
4040 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4041 * that have not yet been banked in case the task is currently running.
4042 */
4043unsigned long long task_sched_runtime(struct task_struct *p)
4044{
4045        unsigned long flags;
4046        u64 ns, delta_exec;
4047        struct rq *rq;
4048
4049        rq = task_rq_lock(p, &flags);
4050        ns = p->se.sum_exec_runtime;
4051        if (task_current(rq, p)) {
4052                update_rq_clock(rq);
4053                delta_exec = rq->clock - p->se.exec_start;
4054                if ((s64)delta_exec > 0)
4055                        ns += delta_exec;
4056        }
4057        task_rq_unlock(rq, &flags);
4058
4059        return ns;
4060}
4061
4062/*
4063 * Account user cpu time to a process.
4064 * @p: the process that the cpu time gets accounted to
4065 * @cputime: the cpu time spent in user space since the last update
4066 */
4067void account_user_time(struct task_struct *p, cputime_t cputime)
4068{
4069        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4070        cputime64_t tmp;
4071
4072        p->utime = cputime_add(p->utime, cputime);
4073
4074        /* Add user time to cpustat. */
4075        tmp = cputime_to_cputime64(cputime);
4076        if (TASK_NICE(p) > 0)
4077                cpustat->nice = cputime64_add(cpustat->nice, tmp);
4078        else
4079                cpustat->user = cputime64_add(cpustat->user, tmp);
4080        /* Account for user time used */
4081        acct_update_integrals(p);
4082}
4083
4084/*
4085 * Account guest cpu time to a process.
4086 * @p: the process that the cpu time gets accounted to
4087 * @cputime: the cpu time spent in virtual machine since the last update
4088 */
4089static void account_guest_time(struct task_struct *p, cputime_t cputime)
4090{
4091        cputime64_t tmp;
4092        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4093
4094        tmp = cputime_to_cputime64(cputime);
4095
4096        p->utime = cputime_add(p->utime, cputime);
4097        p->gtime = cputime_add(p->gtime, cputime);
4098
4099        cpustat->user = cputime64_add(cpustat->user, tmp);
4100        cpustat->guest = cputime64_add(cpustat->guest, tmp);
4101}
4102
4103/*
4104 * Account scaled user cpu time to a process.
4105 * @p: the process that the cpu time gets accounted to
4106 * @cputime: the cpu time spent in user space since the last update
4107 */
4108void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4109{
4110        p->utimescaled = cputime_add(p->utimescaled, cputime);
4111}
4112
4113/*
4114 * Account system cpu time to a process.
4115 * @p: the process that the cpu time gets accounted to
4116 * @hardirq_offset: the offset to subtract from hardirq_count()
4117 * @cputime: the cpu time spent in kernel space since the last update
4118 */
4119void account_system_time(struct task_struct *p, int hardirq_offset,
4120                         cputime_t cputime)
4121{
4122        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4123        struct rq *rq = this_rq();
4124        cputime64_t tmp;
4125
4126        if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4127                account_guest_time(p, cputime);
4128                return;
4129        }
4130
4131        p->stime = cputime_add(p->stime, cputime);
4132
4133        /* Add system time to cpustat. */
4134        tmp = cputime_to_cputime64(cputime);
4135        if (hardirq_count() - hardirq_offset)
4136                cpustat->irq = cputime64_add(cpustat->irq, tmp);
4137        else if (softirq_count())
4138                cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4139        else if (p != rq->idle)
4140                cpustat->system = cputime64_add(cpustat->system, tmp);
4141        else if (atomic_read(&rq->nr_iowait) > 0)
4142                cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4143        else
4144                cpustat->idle = cputime64_add(cpustat->idle, tmp);
4145        /* Account for system time used */
4146        acct_update_integrals(p);
4147}
4148
4149/*
4150 * Account scaled system cpu time to a process.
4151 * @p: the process that the cpu time gets accounted to
4152 * @hardirq_offset: the offset to subtract from hardirq_count()
4153 * @cputime: the cpu time spent in kernel space since the last update
4154 */
4155void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4156{
4157        p->stimescaled = cputime_add(p->stimescaled, cputime);
4158}
4159
4160/*
4161 * Account for involuntary wait time.
4162 * @p: the process from which the cpu time has been stolen
4163 * @steal: the cpu time spent in involuntary wait
4164 */
4165void account_steal_time(struct task_struct *p, cputime_t steal)
4166{
4167        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4168        cputime64_t tmp = cputime_to_cputime64(steal);
4169        struct rq *rq = this_rq();
4170
4171        if (p == rq->idle) {
4172                p->stime = cputime_add(p->stime, steal);
4173                if (atomic_read(&rq->nr_iowait) > 0)
4174                        cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4175                else
4176                        cpustat->idle = cputime64_add(cpustat->idle, tmp);
4177        } else
4178                cpustat->steal = cputime64_add(cpustat->steal, tmp);
4179}
4180
4181/*
4182 * Use precise platform statistics if available:
4183 */
4184#ifdef CONFIG_VIRT_CPU_ACCOUNTING
4185cputime_t task_utime(struct task_struct *p)
4186{
4187        return p->utime;
4188}
4189
4190cputime_t task_stime(struct task_struct *p)
4191{
4192        return p->stime;
4193}
4194#else
4195cputime_t task_utime(struct task_struct *p)
4196{
4197        clock_t utime = cputime_to_clock_t(p->utime),
4198                total = utime + cputime_to_clock_t(p->stime);
4199        u64 temp;
4200
4201        /*
4202         * Use CFS's precise accounting:
4203         */
4204        temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4205
4206        if (total) {
4207                temp *= utime;
4208                do_div(temp, total);
4209        }
4210        utime = (clock_t)temp;
4211
4212        p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4213        return p->prev_utime;
4214}
4215
4216cputime_t task_stime(struct task_struct *p)
4217{
4218        clock_t stime;
4219
4220        /*
4221         * Use CFS's precise accounting. (we subtract utime from
4222         * the total, to make sure the total observed by userspace
4223         * grows monotonically - apps rely on that):
4224         */
4225        stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4226                        cputime_to_clock_t(task_utime(p));
4227
4228        if (stime >= 0)
4229                p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4230
4231        return p->prev_stime;
4232}
4233#endif
4234
4235inline cputime_t task_gtime(struct task_struct *p)
4236{
4237        return p->gtime;
4238}
4239
4240/*
4241 * This function gets called by the timer code, with HZ frequency.
4242 * We call it with interrupts disabled.
4243 *
4244 * It also gets called by the fork code, when changing the parent's
4245 * timeslices.
4246 */
4247void scheduler_tick(void)
4248{
4249        int cpu = smp_processor_id();
4250        struct rq *rq = cpu_rq(cpu);
4251        struct task_struct *curr = rq->curr;
4252
4253        sched_clock_tick();
4254
4255        spin_lock(&rq->lock);
4256        update_rq_clock(rq);
4257        update_cpu_load(rq);
4258        curr->sched_class->task_tick(rq, curr, 0);
4259        spin_unlock(&rq->lock);
4260
4261#ifdef CONFIG_SMP
4262        rq->idle_at_tick = idle_cpu(cpu);
4263        trigger_load_balance(rq, cpu);
4264#endif
4265}
4266
4267#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4268                                defined(CONFIG_PREEMPT_TRACER))
4269
4270static inline unsigned long get_parent_ip(unsigned long addr)
4271{
4272        if (in_lock_functions(addr)) {
4273                addr = CALLER_ADDR2;
4274                if (in_lock_functions(addr))
4275                        addr = CALLER_ADDR3;
4276        }
4277        return addr;
4278}
4279
4280void __kprobes add_preempt_count(int val)
4281{
4282#ifdef CONFIG_DEBUG_PREEMPT
4283        /*
4284         * Underflow?
4285         */
4286        if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4287                return;
4288#endif
4289        preempt_count() += val;
4290#ifdef CONFIG_DEBUG_PREEMPT
4291        /*
4292         * Spinlock count overflowing soon?
4293         */
4294        DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4295                                PREEMPT_MASK - 10);
4296#endif
4297        if (preempt_count() == val)
4298                trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4299}
4300EXPORT_SYMBOL(add_preempt_count);
4301
4302void __kprobes sub_preempt_count(int val)
4303{
4304#ifdef CONFIG_DEBUG_PREEMPT
4305        /*
4306         * Underflow?
4307         */
4308        if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4309                return;
4310        /*
4311         * Is the spinlock portion underflowing?
4312         */
4313        if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4314                        !(preempt_count() & PREEMPT_MASK)))
4315                return;
4316#endif
4317
4318        if (preempt_count() == val)
4319                trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4320        preempt_count() -= val;
4321}
4322EXPORT_SYMBOL(sub_preempt_count);
4323
4324#endif
4325
4326/*
4327 * Print scheduling while atomic bug:
4328 */
4329static noinline void __schedule_bug(struct task_struct *prev)
4330{
4331        struct pt_regs *regs = get_irq_regs();
4332
4333        printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4334                prev->comm, prev->pid, preempt_count());
4335
4336        debug_show_held_locks(prev);
4337        print_modules();
4338        if (irqs_disabled())
4339                print_irqtrace_events(prev);
4340
4341        if (regs)
4342                show_regs(regs);
4343        else
4344                dump_stack();
4345}
4346
4347/*
4348 * Various schedule()-time debugging checks and statistics:
4349 */
4350static inline void schedule_debug(struct task_struct *prev)
4351{
4352        /*
4353         * Test if we are atomic. Since do_exit() needs to call into
4354         * schedule() atomically, we ignore that path for now.
4355         * Otherwise, whine if we are scheduling when we should not be.
4356         */
4357        if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4358                __schedule_bug(prev);
4359
4360        profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4361
4362        schedstat_inc(this_rq(), sched_count);
4363#ifdef CONFIG_SCHEDSTATS
4364        if (unlikely(prev->lock_depth >= 0)) {
4365                schedstat_inc(this_rq(), bkl_count);
4366                schedstat_inc(prev, sched_info.bkl_count);
4367        }
4368#endif
4369}
4370
4371/*
4372 * Pick up the highest-prio task:
4373 */
4374static inline struct task_struct *
4375pick_next_task(struct rq *rq, struct task_struct *prev)
4376{
4377        const struct sched_class *class;
4378        struct task_struct *p;
4379
4380        /*
4381         * Optimization: we know that if all tasks are in
4382         * the fair class we can call that function directly:
4383         */
4384        if (likely(rq->nr_running == rq->cfs.nr_running)) {
4385                p = fair_sched_class.pick_next_task(rq);
4386                if (likely(p))
4387                        return p;
4388        }
4389
4390        class = sched_class_highest;
4391        for ( ; ; ) {
4392                p = class->pick_next_task(rq);
4393                if (p)
4394                        return p;
4395                /*
4396                 * Will never be NULL as the idle class always
4397                 * returns a non-NULL p:
4398                 */
4399                class = class->next;
4400        }
4401}
4402
4403/*
4404 * schedule() is the main scheduler function.
4405 */
4406asmlinkage void __sched schedule(void)
4407{
4408        struct task_struct *prev, *next;
4409        unsigned long *switch_count;
4410        struct rq *rq;
4411        int cpu;
4412
4413need_resched:
4414        preempt_disable();
4415        cpu = smp_processor_id();
4416        rq = cpu_rq(cpu);
4417        rcu_qsctr_inc(cpu);
4418        prev = rq->curr;
4419        switch_count = &prev->nivcsw;
4420
4421        release_kernel_lock(prev);
4422need_resched_nonpreemptible:
4423
4424        schedule_debug(prev);
4425
4426        if (sched_feat(HRTICK))
4427                hrtick_clear(rq);
4428
4429        /*
4430         * Do the rq-clock update outside the rq lock:
4431         */
4432        local_irq_disable();
4433        update_rq_clock(rq);
4434        spin_lock(&rq->lock);
4435        clear_tsk_need_resched(prev);
4436
4437        if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4438                if (unlikely(signal_pending_state(prev->state, prev)))
4439                        prev->state = TASK_RUNNING;
4440                else
4441                        deactivate_task(rq, prev, 1);
4442                switch_count = &prev->nvcsw;
4443        }
4444
4445#ifdef CONFIG_SMP
4446        if (prev->sched_class->pre_schedule)
4447                prev->sched_class->pre_schedule(rq, prev);
4448#endif
4449
4450        if (unlikely(!rq->nr_running))
4451                idle_balance(cpu, rq);
4452
4453        prev->sched_class->put_prev_task(rq, prev);
4454        next = pick_next_task(rq, prev);
4455
4456        if (likely(prev != next)) {
4457                sched_info_switch(prev, next);
4458
4459                rq->nr_switches++;
4460                rq->curr = next;
4461                ++*switch_count;
4462
4463                context_switch(rq, prev, next); /* unlocks the rq */
4464                /*
4465                 * the context switch might have flipped the stack from under
4466                 * us, hence refresh the local variables.
4467                 */
4468                cpu = smp_processor_id();
4469                rq = cpu_rq(cpu);
4470        } else
4471                spin_unlock_irq(&rq->lock);
4472
4473        if (unlikely(reacquire_kernel_lock(current) < 0))
4474                goto need_resched_nonpreemptible;
4475
4476        preempt_enable_no_resched();
4477        if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4478                goto need_resched;
4479}
4480EXPORT_SYMBOL(schedule);
4481
4482#ifdef CONFIG_PREEMPT
4483/*
4484 * this is the entry point to schedule() from in-kernel preemption
4485 * off of preempt_enable. Kernel preemptions off return from interrupt
4486 * occur there and call schedule directly.
4487 */
4488asmlinkage void __sched preempt_schedule(void)
4489{
4490        struct thread_info *ti = current_thread_info();
4491
4492        /*
4493         * If there is a non-zero preempt_count or interrupts are disabled,
4494         * we do not want to preempt the current task. Just return..
4495         */
4496        if (likely(ti->preempt_count || irqs_disabled()))
4497                return;
4498
4499        do {
4500                add_preempt_count(PREEMPT_ACTIVE);
4501                schedule();
4502                sub_preempt_count(PREEMPT_ACTIVE);
4503
4504                /*
4505                 * Check again in case we missed a preemption opportunity
4506                 * between schedule and now.
4507                 */
4508                barrier();
4509        } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4510}
4511EXPORT_SYMBOL(preempt_schedule);
4512
4513/*
4514 * this is the entry point to schedule() from kernel preemption
4515 * off of irq context.
4516 * Note, that this is called and return with irqs disabled. This will
4517 * protect us against recursive calling from irq.
4518 */
4519asmlinkage void __sched preempt_schedule_irq(void)
4520{
4521        struct thread_info *ti = current_thread_info();
4522
4523        /* Catch callers which need to be fixed */
4524        BUG_ON(ti->preempt_count || !irqs_disabled());
4525
4526        do {
4527                add_preempt_count(PREEMPT_ACTIVE);
4528                local_irq_enable();
4529                schedule();
4530                local_irq_disable();
4531                sub_preempt_count(PREEMPT_ACTIVE);
4532
4533                /*
4534                 * Check again in case we missed a preemption opportunity
4535                 * between schedule and now.
4536                 */
4537                barrier();
4538        } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4539}
4540
4541#endif /* CONFIG_PREEMPT */
4542
4543int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4544                          void *key)
4545{
4546        return try_to_wake_up(curr->private, mode, sync);
4547}
4548EXPORT_SYMBOL(default_wake_function);
4549
4550/*
4551 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4552 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4553 * number) then we wake all the non-exclusive tasks and one exclusive task.
4554 *
4555 * There are circumstances in which we can try to wake a task which has already
4556 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4557 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4558 */
4559static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4560                             int nr_exclusive, int sync, void *key)
4561{
4562        wait_queue_t *curr, *next;
4563
4564        list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4565                unsigned flags = curr->flags;
4566
4567                if (curr->func(curr, mode, sync, key) &&
4568                                (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4569                        break;
4570        }
4571}
4572
4573/**
4574 * __wake_up - wake up threads blocked on a waitqueue.
4575 * @q: the waitqueue
4576 * @mode: which threads
4577 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4578 * @key: is directly passed to the wakeup function
4579 */
4580void __wake_up(wait_queue_head_t *q, unsigned int mode,
4581                        int nr_exclusive, void *key)
4582{
4583        unsigned long flags;
4584
4585        spin_lock_irqsave(&q->lock, flags);
4586        __wake_up_common(q, mode, nr_exclusive, 0, key);
4587        spin_unlock_irqrestore(&q->lock, flags);
4588}
4589EXPORT_SYMBOL(__wake_up);
4590
4591/*
4592 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4593 */
4594void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4595{
4596        __wake_up_common(q, mode, 1, 0, NULL);
4597}
4598
4599/**
4600 * __wake_up_sync - wake up threads blocked on a waitqueue.
4601 * @q: the waitqueue
4602 * @mode: which threads
4603 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4604 *
4605 * The sync wakeup differs that the waker knows that it will schedule
4606 * away soon, so while the target thread will be woken up, it will not
4607 * be migrated to another CPU - ie. the two threads are 'synchronized'
4608 * with each other. This can prevent needless bouncing between CPUs.
4609 *
4610 * On UP it can prevent extra preemption.
4611 */
4612void
4613__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4614{
4615        unsigned long flags;
4616        int sync = 1;
4617
4618        if (unlikely(!q))
4619                return;
4620
4621        if (unlikely(!nr_exclusive))
4622                sync = 0;
4623
4624        spin_lock_irqsave(&q->lock, flags);
4625        __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4626        spin_unlock_irqrestore(&q->lock, flags);
4627}
4628EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
4629
4630void complete(struct completion *x)
4631{
4632        unsigned long flags;
4633
4634        spin_lock_irqsave(&x->wait.lock, flags);
4635        x->done++;
4636        __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4637        spin_unlock_irqrestore(&x->wait.lock, flags);
4638}
4639EXPORT_SYMBOL(complete);
4640
4641void complete_all(struct completion *x)
4642{
4643        unsigned long flags;
4644
4645        spin_lock_irqsave(&x->wait.lock, flags);
4646        x->done += UINT_MAX/2;
4647        __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4648        spin_unlock_irqrestore(&x->wait.lock, flags);
4649}
4650EXPORT_SYMBOL(complete_all);
4651
4652static inline long __sched
4653do_wait_for_common(struct completion *x, long timeout, int state)
4654{
4655        if (!x->done) {
4656                DECLARE_WAITQUEUE(wait, current);
4657
4658                wait.flags |= WQ_FLAG_EXCLUSIVE;
4659                __add_wait_queue_tail(&x->wait, &wait);
4660                do {
4661                        if ((state == TASK_INTERRUPTIBLE &&
4662                             signal_pending(current)) ||
4663                            (state == TASK_KILLABLE &&
4664                             fatal_signal_pending(current))) {
4665                                timeout = -ERESTARTSYS;
4666                                break;
4667                        }
4668                        __set_current_state(state);
4669                        spin_unlock_irq(&x->wait.lock);
4670                        timeout = schedule_timeout(timeout);
4671                        spin_lock_irq(&x->wait.lock);
4672                } while (!x->done && timeout);
4673                __remove_wait_queue(&x->wait, &wait);
4674                if (!x->done)
4675                        return timeout;
4676        }
4677        x->done--;
4678        return timeout ?: 1;
4679}
4680
4681static long __sched
4682wait_for_common(struct completion *x, long timeout, int state)
4683{
4684        might_sleep();
4685
4686        spin_lock_irq(&x->wait.lock);
4687        timeout = do_wait_for_common(x, timeout, state);
4688        spin_unlock_irq(&x->wait.lock);
4689        return timeout;
4690}
4691
4692void __sched wait_for_completion(struct completion *x)
4693{
4694        wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4695}
4696EXPORT_SYMBOL(wait_for_completion);
4697
4698unsigned long __sched
4699wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4700{
4701        return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4702}
4703EXPORT_SYMBOL(wait_for_completion_timeout);
4704
4705int __sched wait_for_completion_interruptible(struct completion *x)
4706{
4707        long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4708        if (t == -ERESTARTSYS)
4709                return t;
4710        return 0;
4711}
4712EXPORT_SYMBOL(wait_for_completion_interruptible);
4713
4714unsigned long __sched
4715wait_for_completion_interruptible_timeout(struct completion *x,
4716                                          unsigned long timeout)
4717{
4718        return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4719}
4720EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4721
4722int __sched wait_for_completion_killable(struct completion *x)
4723{
4724        long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4725        if (t == -ERESTARTSYS)
4726                return t;
4727        return 0;
4728}
4729EXPORT_SYMBOL(wait_for_completion_killable);
4730
4731/**
4732 *      try_wait_for_completion - try to decrement a completion without blocking
4733 *      @x:     completion structure
4734 *
4735 *      Returns: 0 if a decrement cannot be done without blocking
4736 *               1 if a decrement succeeded.
4737 *
4738 *      If a completion is being used as a counting completion,
4739 *      attempt to decrement the counter without blocking. This
4740 *      enables us to avoid waiting if the resource the completion
4741 *      is protecting is not available.
4742 */
4743bool try_wait_for_completion(struct completion *x)
4744{
4745        int ret = 1;
4746
4747        spin_lock_irq(&x->wait.lock);
4748        if (!x->done)
4749                ret = 0;
4750        else
4751                x->done--;
4752        spin_unlock_irq(&x->wait.lock);
4753        return ret;
4754}
4755EXPORT_SYMBOL(try_wait_for_completion);
4756
4757/**
4758 *      completion_done - Test to see if a completion has any waiters
4759 *      @x:     completion structure
4760 *
4761 *      Returns: 0 if there are waiters (wait_for_completion() in progress)
4762 *               1 if there are no waiters.
4763 *
4764 */
4765bool completion_done(struct completion *x)
4766{
4767        int ret = 1;
4768
4769        spin_lock_irq(&x->wait.lock);
4770        if (!x->done)
4771                ret = 0;
4772        spin_unlock_irq(&x->wait.lock);
4773        return ret;
4774}
4775EXPORT_SYMBOL(completion_done);
4776
4777static long __sched
4778sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4779{
4780        unsigned long flags;
4781        wait_queue_t wait;
4782
4783        init_waitqueue_entry(&wait, current);
4784
4785        __set_current_state(state);
4786
4787        spin_lock_irqsave(&q->lock, flags);
4788        __add_wait_queue(q, &wait);
4789        spin_unlock(&q->lock);
4790        timeout = schedule_timeout(timeout);
4791        spin_lock_irq(&q->lock);
4792        __remove_wait_queue(q, &wait);
4793        spin_unlock_irqrestore(&q->lock, flags);
4794
4795        return timeout;
4796}
4797
4798void __sched interruptible_sleep_on(wait_queue_head_t *q)
4799{
4800        sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4801}
4802EXPORT_SYMBOL(interruptible_sleep_on);
4803
4804long __sched
4805interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4806{
4807        return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4808}
4809EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4810
4811void __sched sleep_on(wait_queue_head_t *q)
4812{
4813        sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4814}
4815EXPORT_SYMBOL(sleep_on);
4816
4817long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4818{
4819        return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4820}
4821EXPORT_SYMBOL(sleep_on_timeout);
4822
4823#ifdef CONFIG_RT_MUTEXES
4824
4825/*
4826 * rt_mutex_setprio - set the current priority of a task
4827 * @p: task
4828 * @prio: prio value (kernel-internal form)
4829 *
4830 * This function changes the 'effective' priority of a task. It does
4831 * not touch ->normal_prio like __setscheduler().
4832 *
4833 * Used by the rt_mutex code to implement priority inheritance logic.
4834 */
4835void rt_mutex_setprio(struct task_struct *p, int prio)
4836{
4837        unsigned long flags;
4838        int oldprio, on_rq, running;
4839        struct rq *rq;
4840        const struct sched_class *prev_class = p->sched_class;
4841
4842        BUG_ON(prio < 0 || prio > MAX_PRIO);
4843
4844        rq = task_rq_lock(p, &flags);
4845        update_rq_clock(rq);
4846
4847        oldprio = p->prio;
4848        on_rq = p->se.on_rq;
4849        running = task_current(rq, p);
4850        if (on_rq)
4851                dequeue_task(rq, p, 0);
4852        if (running)
4853                p->sched_class->put_prev_task(rq, p);
4854
4855        if (rt_prio(prio))
4856                p->sched_class = &rt_sched_class;
4857        else
4858                p->sched_class = &fair_sched_class;
4859
4860        p->prio = prio;
4861
4862        if (running)
4863                p->sched_class->set_curr_task(rq);
4864        if (on_rq) {
4865                enqueue_task(rq, p, 0);
4866
4867                check_class_changed(rq, p, prev_class, oldprio, running);
4868        }
4869        task_rq_unlock(rq, &flags);
4870}
4871
4872#endif
4873
4874void set_user_nice(struct task_struct *p, long nice)
4875{
4876        int old_prio, delta, on_rq;
4877        unsigned long flags;
4878        struct rq *rq;
4879
4880        if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4881                return;
4882        /*
4883         * We have to be careful, if called from sys_setpriority(),
4884         * the task might be in the middle of scheduling on another CPU.
4885         */
4886        rq = task_rq_lock(p, &flags);
4887        update_rq_clock(rq);
4888        /*
4889         * The RT priorities are set via sched_setscheduler(), but we still
4890         * allow the 'normal' nice value to be set - but as expected
4891         * it wont have any effect on scheduling until the task is
4892         * SCHED_FIFO/SCHED_RR:
4893         */
4894        if (task_has_rt_policy(p)) {
4895                p->static_prio = NICE_TO_PRIO(nice);
4896                goto out_unlock;
4897        }
4898        on_rq = p->se.on_rq;
4899        if (on_rq)
4900                dequeue_task(rq, p, 0);
4901
4902        p->static_prio = NICE_TO_PRIO(nice);
4903        set_load_weight(p);
4904        old_prio = p->prio;
4905        p->prio = effective_prio(p);
4906        delta = p->prio - old_prio;
4907
4908        if (on_rq) {
4909                enqueue_task(rq, p, 0);
4910                /*
4911                 * If the task increased its priority or is running and
4912                 * lowered its priority, then reschedule its CPU:
4913                 */
4914                if (delta < 0 || (delta > 0 && task_running(rq, p)))
4915                        resched_task(rq->curr);
4916        }
4917out_unlock:
4918        task_rq_unlock(rq, &flags);
4919}
4920EXPORT_SYMBOL(set_user_nice);
4921
4922/*
4923 * can_nice - check if a task can reduce its nice value
4924 * @p: task
4925 * @nice: nice value
4926 */
4927int can_nice(const struct task_struct *p, const int nice)
4928{
4929        /* convert nice value [19,-20] to rlimit style value [1,40] */
4930        int nice_rlim = 20 - nice;
4931
4932        return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4933                capable(CAP_SYS_NICE));
4934}
4935
4936#ifdef __ARCH_WANT_SYS_NICE
4937
4938/*
4939 * sys_nice - change the priority of the current process.
4940 * @increment: priority increment
4941 *
4942 * sys_setpriority is a more generic, but much slower function that
4943 * does similar things.
4944 */
4945asmlinkage long sys_nice(int increment)
4946{
4947        long nice, retval;
4948
4949        /*
4950         * Setpriority might change our priority at the same moment.
4951         * We don't have to worry. Conceptually one call occurs first
4952         * and we have a single winner.
4953         */
4954        if (increment < -40)
4955                increment = -40;
4956        if (increment > 40)
4957                increment = 40;
4958
4959        nice = PRIO_TO_NICE(current->static_prio) + increment;
4960        if (nice < -20)
4961                nice = -20;
4962        if (nice > 19)
4963                nice = 19;
4964
4965        if (increment < 0 && !can_nice(current, nice))
4966                return -EPERM;
4967
4968        retval = security_task_setnice(current, nice);
4969        if (retval)
4970                return retval;
4971
4972        set_user_nice(current, nice);
4973        return 0;
4974}
4975
4976#endif
4977
4978/**
4979 * task_prio - return the priority value of a given task.
4980 * @p: the task in question.
4981 *
4982 * This is the priority value as seen by users in /proc.
4983 * RT tasks are offset by -200. Normal tasks are centered
4984 * around 0, value goes from -16 to +15.
4985 */
4986int task_prio(const struct task_struct *p)
4987{
4988        return p->prio - MAX_RT_PRIO;
4989}
4990
4991/**
4992 * task_nice - return the nice value of a given task.
4993 * @p: the task in question.
4994 */
4995int task_nice(const struct task_struct *p)
4996{
4997        return TASK_NICE(p);
4998}
4999EXPORT_SYMBOL(task_nice);
5000
5001/**
5002 * idle_cpu - is a given cpu idle currently?
5003 * @cpu: the processor in question.
5004 */
5005int idle_cpu(int cpu)
5006{
5007        return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5008}
5009
5010/**
5011 * idle_task - return the idle task for a given cpu.
5012 * @cpu: the processor in question.
5013 */
5014struct task_struct *idle_task(int cpu)
5015{
5016        return cpu_rq(cpu)->idle;
5017}
5018
5019/**
5020 * find_process_by_pid - find a process with a matching PID value.
5021 * @pid: the pid in question.
5022 */
5023static struct task_struct *find_process_by_pid(pid_t pid)
5024{
5025        return pid ? find_task_by_vpid(pid) : current;
5026}
5027
5028/* Actually do priority change: must hold rq lock. */
5029static void
5030__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5031{
5032        BUG_ON(p->se.on_rq);
5033
5034        p->policy = policy;
5035        switch (p->policy) {
5036        case SCHED_NORMAL:
5037        case SCHED_BATCH:
5038        case SCHED_IDLE:
5039                p->sched_class = &fair_sched_class;
5040                break;
5041        case SCHED_FIFO:
5042        case SCHED_RR:
5043                p->sched_class = &rt_sched_class;
5044                break;
5045        }
5046
5047        p->rt_priority = prio;
5048        p->normal_prio = normal_prio(p);
5049        /* we are holding p->pi_lock already */
5050        p->prio = rt_mutex_getprio(p);
5051        set_load_weight(p);
5052}
5053
5054static int __sched_setscheduler(struct task_struct *p, int policy,
5055                                struct sched_param *param, bool user)
5056{
5057        int retval, oldprio, oldpolicy = -1, on_rq, running;
5058        unsigned long flags;
5059        const struct sched_class *prev_class = p->sched_class;
5060        struct rq *rq;
5061
5062        /* may grab non-irq protected spin_locks */
5063        BUG_ON(in_interrupt());
5064recheck:
5065        /* double check policy once rq lock held */
5066        if (policy < 0)
5067                policy = oldpolicy = p->policy;
5068        else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5069                        policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5070                        policy != SCHED_IDLE)
5071                return -EINVAL;
5072        /*
5073         * Valid priorities for SCHED_FIFO and SCHED_RR are
5074         * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5075         * SCHED_BATCH and SCHED_IDLE is 0.
5076         */
5077        if (param->sched_priority < 0 ||
5078            (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5079            (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5080                return -EINVAL;
5081        if (rt_policy(policy) != (param->sched_priority != 0))
5082                return -EINVAL;
5083
5084        /*
5085         * Allow unprivileged RT tasks to decrease priority:
5086         */
5087        if (user && !capable(CAP_SYS_NICE)) {
5088                if (rt_policy(policy)) {
5089                        unsigned long rlim_rtprio;
5090
5091                        if (!lock_task_sighand(p, &flags))
5092                                return -ESRCH;
5093                        rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5094                        unlock_task_sighand(p, &flags);
5095
5096                        /* can't set/change the rt policy */
5097                        if (policy != p->policy && !rlim_rtprio)
5098                                return -EPERM;
5099
5100                        /* can't increase priority */
5101                        if (param->sched_priority > p->rt_priority &&
5102                            param->sched_priority > rlim_rtprio)
5103                                return -EPERM;
5104                }
5105                /*
5106                 * Like positive nice levels, dont allow tasks to
5107                 * move out of SCHED_IDLE either:
5108                 */
5109                if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5110                        return -EPERM;
5111
5112                /* can't change other user's priorities */
5113                if ((current->euid != p->euid) &&
5114                    (current->euid != p->uid))
5115                        return -EPERM;
5116        }
5117
5118        if (user) {
5119#ifdef CONFIG_RT_GROUP_SCHED
5120                /*
5121                 * Do not allow realtime tasks into groups that have no runtime
5122                 * assigned.
5123                 */
5124                if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5125                        return -EPERM;
5126#endif
5127
5128                retval = security_task_setscheduler(p, policy, param);
5129                if (retval)
5130                        return retval;
5131        }
5132
5133        /*
5134         * make sure no PI-waiters arrive (or leave) while we are
5135         * changing the priority of the task:
5136         */
5137        spin_lock_irqsave(&p->pi_lock, flags);
5138        /*
5139         * To be able to change p->policy safely, the apropriate
5140         * runqueue lock must be held.
5141         */
5142        rq = __task_rq_lock(p);
5143        /* recheck policy now with rq lock held */
5144        if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5145                policy = oldpolicy = -1;
5146                __task_rq_unlock(rq);
5147                spin_unlock_irqrestore(&p->pi_lock, flags);
5148                goto recheck;
5149        }
5150        update_rq_clock(rq);
5151        on_rq = p->se.on_rq;
5152        running = task_current(rq, p);
5153        if (on_rq)
5154                deactivate_task(rq, p, 0);
5155        if (running)
5156                p->sched_class->put_prev_task(rq, p);
5157
5158        oldprio = p->prio;
5159        __setscheduler(rq, p, policy, param->sched_priority);
5160
5161        if (running)
5162                p->sched_class->set_curr_task(rq);
5163        if (on_rq) {
5164                activate_task(rq, p, 0);
5165
5166                check_class_changed(rq, p, prev_class, oldprio, running);
5167        }
5168        __task_rq_unlock(rq);
5169        spin_unlock_irqrestore(&p->pi_lock, flags);
5170
5171        rt_mutex_adjust_pi(p);
5172
5173        return 0;
5174}
5175
5176/**
5177 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5178 * @p: the task in question.
5179 * @policy: new policy.
5180 * @param: structure containing the new RT priority.
5181 *
5182 * NOTE that the task may be already dead.
5183 */
5184int sched_setscheduler(struct task_struct *p, int policy,
5185                       struct sched_param *param)
5186{
5187        return __sched_setscheduler(p, policy, param, true);
5188}
5189EXPORT_SYMBOL_GPL(sched_setscheduler);
5190
5191/**
5192 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5193 * @p: the task in question.
5194 * @policy: new policy.
5195 * @param: structure containing the new RT priority.
5196 *
5197 * Just like sched_setscheduler, only don't bother checking if the
5198 * current context has permission.  For example, this is needed in
5199 * stop_machine(): we create temporary high priority worker threads,
5200 * but our caller might not have that capability.
5201 */
5202int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5203                               struct sched_param *param)
5204{
5205        return __sched_setscheduler(p, policy, param, false);
5206}
5207
5208static int
5209do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5210{
5211        struct sched_param lparam;
5212        struct task_struct *p;
5213        int retval;
5214
5215        if (!param || pid < 0)
5216                return -EINVAL;
5217        if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5218                return -EFAULT;
5219
5220        rcu_read_lock();
5221        retval = -ESRCH;
5222        p = find_process_by_pid(pid);
5223        if (p != NULL)
5224                retval = sched_setscheduler(p, policy, &lparam);
5225        rcu_read_unlock();
5226
5227        return retval;
5228}
5229
5230/**
5231 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5232 * @pid: the pid in question.
5233 * @policy: new policy.
5234 * @param: structure containing the new RT priority.
5235 */
5236asmlinkage long
5237sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5238{
5239        /* negative values for policy are not valid */
5240        if (policy < 0)
5241                return -EINVAL;
5242
5243        return do_sched_setscheduler(pid, policy, param);
5244}
5245
5246/**
5247 * sys_sched_setparam - set/change the RT priority of a thread
5248 * @pid: the pid in question.
5249 * @param: structure containing the new RT priority.
5250 */
5251asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5252{
5253        return do_sched_setscheduler(pid, -1, param);
5254}
5255
5256/**
5257 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5258 * @pid: the pid in question.
5259 */
5260asmlinkage long sys_sched_getscheduler(pid_t pid)
5261{
5262        struct task_struct *p;
5263        int retval;
5264
5265        if (pid < 0)
5266                return -EINVAL;
5267
5268        retval = -ESRCH;
5269        read_lock(&tasklist_lock);
5270        p = find_process_by_pid(pid);
5271        if (p) {
5272                retval = security_task_getscheduler(p);
5273                if (!retval)
5274                        retval = p->policy;
5275        }
5276        read_unlock(&tasklist_lock);
5277        return retval;
5278}
5279
5280/**
5281 * sys_sched_getscheduler - get the RT priority of a thread
5282 * @pid: the pid in question.
5283 * @param: structure containing the RT priority.
5284 */
5285asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5286{
5287        struct sched_param lp;
5288        struct task_struct *p;
5289        int retval;
5290
5291        if (!param || pid < 0)
5292                return -EINVAL;
5293
5294        read_lock(&tasklist_lock);
5295        p = find_process_by_pid(pid);
5296        retval = -ESRCH;
5297        if (!p)
5298                goto out_unlock;
5299
5300        retval = security_task_getscheduler(p);
5301        if (retval)
5302                goto out_unlock;
5303
5304        lp.sched_priority = p->rt_priority;
5305        read_unlock(&tasklist_lock);
5306
5307        /*
5308         * This one might sleep, we cannot do it with a spinlock held ...
5309         */
5310        retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5311
5312        return retval;
5313
5314out_unlock:
5315        read_unlock(&tasklist_lock);
5316        return retval;
5317}
5318
5319long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5320{
5321        cpumask_t cpus_allowed;
5322        cpumask_t new_mask = *in_mask;
5323        struct task_struct *p;
5324        int retval;
5325
5326        get_online_cpus();
5327        read_lock(&tasklist_lock);
5328
5329        p = find_process_by_pid(pid);
5330        if (!p) {
5331                read_unlock(&tasklist_lock);
5332                put_online_cpus();
5333                return -ESRCH;
5334        }
5335
5336        /*
5337         * It is not safe to call set_cpus_allowed with the
5338         * tasklist_lock held. We will bump the task_struct's
5339         * usage count and then drop tasklist_lock.
5340         */
5341        get_task_struct(p);
5342        read_unlock(&tasklist_lock);
5343
5344        retval = -EPERM;
5345        if ((current->euid != p->euid) && (current->euid != p->uid) &&
5346                        !capable(CAP_SYS_NICE))
5347                goto out_unlock;
5348
5349        retval = security_task_setscheduler(p, 0, NULL);
5350        if (retval)
5351                goto out_unlock;
5352
5353        cpuset_cpus_allowed(p, &cpus_allowed);
5354        cpus_and(new_mask, new_mask, cpus_allowed);
5355 again:
5356        retval = set_cpus_allowed_ptr(p, &new_mask);
5357
5358        if (!retval) {
5359                cpuset_cpus_allowed(p, &cpus_allowed);
5360                if (!cpus_subset(new_mask, cpus_allowed)) {
5361                        /*
5362                         * We must have raced with a concurrent cpuset
5363                         * update. Just reset the cpus_allowed to the
5364                         * cpuset's cpus_allowed
5365                         */
5366                        new_mask = cpus_allowed;
5367                        goto again;
5368                }
5369        }
5370out_unlock:
5371        put_task_struct(p);
5372        put_online_cpus();
5373        return retval;
5374}
5375
5376static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5377                             cpumask_t *new_mask)
5378{
5379        if (len < sizeof(cpumask_t)) {
5380                memset(new_mask, 0, sizeof(cpumask_t));
5381        } else if (len > sizeof(cpumask_t)) {
5382                len = sizeof(cpumask_t);
5383        }
5384        return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5385}
5386
5387/**
5388 * sys_sched_setaffinity - set the cpu affinity of a process
5389 * @pid: pid of the process
5390 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5391 * @user_mask_ptr: user-space pointer to the new cpu mask
5392 */
5393asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5394                                      unsigned long __user *user_mask_ptr)
5395{
5396        cpumask_t new_mask;
5397        int retval;
5398
5399        retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5400        if (retval)
5401                return retval;
5402
5403        return sched_setaffinity(pid, &new_mask);
5404}
5405
5406long sched_getaffinity(pid_t pid, cpumask_t *mask)
5407{
5408        struct task_struct *p;
5409        int retval;
5410
5411        get_online_cpus();
5412        read_lock(&tasklist_lock);
5413
5414        retval = -ESRCH;
5415        p = find_process_by_pid(pid);
5416        if (!p)
5417                goto out_unlock;
5418
5419        retval = security_task_getscheduler(p);
5420        if (retval)
5421                goto out_unlock;
5422
5423        cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5424
5425out_unlock:
5426        read_unlock(&tasklist_lock);
5427        put_online_cpus();
5428
5429        return retval;
5430}
5431
5432/**
5433 * sys_sched_getaffinity - get the cpu affinity of a process
5434 * @pid: pid of the process
5435 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5436 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5437 */
5438asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5439                                      unsigned long __user *user_mask_ptr)
5440{
5441        int ret;
5442        cpumask_t mask;
5443
5444        if (len < sizeof(cpumask_t))
5445                return -EINVAL;
5446
5447        ret = sched_getaffinity(pid, &mask);
5448        if (ret < 0)
5449                return ret;
5450
5451        if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5452                return -EFAULT;
5453
5454        return sizeof(cpumask_t);
5455}
5456
5457/**
5458 * sys_sched_yield - yield the current processor to other threads.
5459 *
5460 * This function yields the current CPU to other tasks. If there are no
5461 * other threads running on this CPU then this function will return.
5462 */
5463asmlinkage long sys_sched_yield(void)
5464{
5465        struct rq *rq = this_rq_lock();
5466
5467        schedstat_inc(rq, yld_count);
5468        current->sched_class->yield_task(rq);
5469
5470        /*
5471         * Since we are going to call schedule() anyway, there's
5472         * no need to preempt or enable interrupts:
5473         */
5474        __release(rq->lock);
5475        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5476        _raw_spin_unlock(&rq->lock);
5477        preempt_enable_no_resched();
5478
5479        schedule();
5480
5481        return 0;
5482}
5483
5484static void __cond_resched(void)
5485{
5486#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5487        __might_sleep(__FILE__, __LINE__);
5488#endif
5489        /*
5490         * The BKS might be reacquired before we have dropped
5491         * PREEMPT_ACTIVE, which could trigger a second
5492         * cond_resched() call.
5493         */
5494        do {
5495                add_preempt_count(PREEMPT_ACTIVE);
5496                schedule();
5497                sub_preempt_count(PREEMPT_ACTIVE);
5498        } while (need_resched());
5499}
5500
5501int __sched _cond_resched(void)
5502{
5503        if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5504                                        system_state == SYSTEM_RUNNING) {
5505                __cond_resched();
5506                return 1;
5507        }
5508        return 0;
5509}
5510EXPORT_SYMBOL(_cond_resched);
5511
5512/*
5513 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5514 * call schedule, and on return reacquire the lock.
5515 *
5516 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5517 * operations here to prevent schedule() from being called twice (once via
5518 * spin_unlock(), once by hand).
5519 */
5520int cond_resched_lock(spinlock_t *lock)
5521{
5522        int resched = need_resched() && system_state == SYSTEM_RUNNING;
5523        int ret = 0;
5524
5525        if (spin_needbreak(lock) || resched) {
5526                spin_unlock(lock);
5527                if (resched && need_resched())
5528                        __cond_resched();
5529                else
5530                        cpu_relax();
5531                ret = 1;
5532                spin_lock(lock);
5533        }
5534        return ret;
5535}
5536EXPORT_SYMBOL(cond_resched_lock);
5537
5538int __sched cond_resched_softirq(void)
5539{
5540        BUG_ON(!in_softirq());
5541
5542        if (need_resched() && system_state == SYSTEM_RUNNING) {
5543                local_bh_enable();
5544                __cond_resched();
5545                local_bh_disable();
5546                return 1;
5547        }
5548        return 0;
5549}
5550EXPORT_SYMBOL(cond_resched_softirq);
5551
5552/**
5553 * yield - yield the current processor to other threads.
5554 *
5555 * This is a shortcut for kernel-space yielding - it marks the
5556 * thread runnable and calls sys_sched_yield().
5557 */
5558void __sched yield(void)
5559{
5560        set_current_state(TASK_RUNNING);
5561        sys_sched_yield();
5562}
5563EXPORT_SYMBOL(yield);
5564
5565/*
5566 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5567 * that process accounting knows that this is a task in IO wait state.
5568 *
5569 * But don't do that if it is a deliberate, throttling IO wait (this task
5570 * has set its backing_dev_info: the queue against which it should throttle)
5571 */
5572void __sched io_schedule(void)
5573{
5574        struct rq *rq = &__raw_get_cpu_var(runqueues);
5575
5576        delayacct_blkio_start();
5577        atomic_inc(&rq->nr_iowait);
5578        schedule();
5579        atomic_dec(&rq->nr_iowait);
5580        delayacct_blkio_end();
5581}
5582EXPORT_SYMBOL(io_schedule);
5583
5584long __sched io_schedule_timeout(long timeout)
5585{
5586        struct rq *rq = &__raw_get_cpu_var(runqueues);
5587        long ret;
5588
5589        delayacct_blkio_start();
5590        atomic_inc(&rq->nr_iowait);
5591        ret = schedule_timeout(timeout);
5592        atomic_dec(&rq->nr_iowait);
5593        delayacct_blkio_end();
5594        return ret;
5595}
5596
5597/**
5598 * sys_sched_get_priority_max - return maximum RT priority.
5599 * @policy: scheduling class.
5600 *
5601 * this syscall returns the maximum rt_priority that can be used
5602 * by a given scheduling class.
5603 */
5604asmlinkage long sys_sched_get_priority_max(int policy)
5605{
5606        int ret = -EINVAL;
5607
5608        switch (policy) {
5609        case SCHED_FIFO:
5610        case SCHED_RR:
5611                ret = MAX_USER_RT_PRIO-1;
5612                break;
5613        case SCHED_NORMAL:
5614        case SCHED_BATCH:
5615        case SCHED_IDLE:
5616                ret = 0;
5617                break;
5618        }
5619        return ret;
5620}
5621
5622/**
5623 * sys_sched_get_priority_min - return minimum RT priority.
5624 * @policy: scheduling class.
5625 *
5626 * this syscall returns the minimum rt_priority that can be used
5627 * by a given scheduling class.
5628 */
5629asmlinkage long sys_sched_get_priority_min(int policy)
5630{
5631        int ret = -EINVAL;
5632
5633        switch (policy) {
5634        case SCHED_FIFO:
5635        case SCHED_RR:
5636                ret = 1;
5637                break;
5638        case SCHED_NORMAL:
5639        case SCHED_BATCH:
5640        case SCHED_IDLE:
5641                ret = 0;
5642        }
5643        return ret;
5644}
5645
5646/**
5647 * sys_sched_rr_get_interval - return the default timeslice of a process.
5648 * @pid: pid of the process.
5649 * @interval: userspace pointer to the timeslice value.
5650 *
5651 * this syscall writes the default timeslice value of a given process
5652 * into the user-space timespec buffer. A value of '0' means infinity.
5653 */
5654asmlinkage
5655long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5656{
5657        struct task_struct *p;
5658        unsigned int time_slice;
5659        int retval;
5660        struct timespec t;
5661
5662        if (pid < 0)
5663                return -EINVAL;
5664
5665        retval = -ESRCH;
5666        read_lock(&tasklist_lock);
5667        p = find_process_by_pid(pid);
5668        if (!p)
5669                goto out_unlock;
5670
5671        retval = security_task_getscheduler(p);
5672        if (retval)
5673                goto out_unlock;
5674
5675        /*
5676         * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5677         * tasks that are on an otherwise idle runqueue:
5678         */
5679        time_slice = 0;
5680        if (p->policy == SCHED_RR) {
5681                time_slice = DEF_TIMESLICE;
5682        } else if (p->policy != SCHED_FIFO) {
5683                struct sched_entity *se = &p->se;
5684                unsigned long flags;
5685                struct rq *rq;
5686
5687                rq = task_rq_lock(p, &flags);
5688                if (rq->cfs.load.weight)
5689                        time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5690                task_rq_unlock(rq, &flags);
5691        }
5692        read_unlock(&tasklist_lock);
5693        jiffies_to_timespec(time_slice, &t);
5694        retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5695        return retval;