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 */
  20
  21#include <linux/mm.h>
  22#include <linux/module.h>
  23#include <linux/nmi.h>
  24#include <linux/init.h>
  25#include <asm/uaccess.h>
  26#include <linux/highmem.h>
  27#include <linux/smp_lock.h>
  28#include <asm/mmu_context.h>
  29#include <linux/interrupt.h>
  30#include <linux/capability.h>
  31#include <linux/completion.h>
  32#include <linux/kernel_stat.h>
  33#include <linux/debug_locks.h>
  34#include <linux/security.h>
  35#include <linux/notifier.h>
  36#include <linux/profile.h>
  37#include <linux/suspend.h>
  38#include <linux/vmalloc.h>
  39#include <linux/blkdev.h>
  40#include <linux/delay.h>
  41#include <linux/smp.h>
  42#include <linux/threads.h>
  43#include <linux/timer.h>
  44#include <linux/rcupdate.h>
  45#include <linux/cpu.h>
  46#include <linux/cpuset.h>
  47#include <linux/percpu.h>
  48#include <linux/kthread.h>
  49#include <linux/seq_file.h>
  50#include <linux/syscalls.h>
  51#include <linux/times.h>
  52#include <linux/acct.h>
  53#include <linux/kprobes.h>
  54#include <linux/delayacct.h>
  55#include <asm/tlb.h>
  56
  57#include <asm/unistd.h>
  58
  59/*
  60 * Convert user-nice values [ -20 ... 0 ... 19 ]
  61 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
  62 * and back.
  63 */
  64#define NICE_TO_PRIO(nice)      (MAX_RT_PRIO + (nice) + 20)
  65#define PRIO_TO_NICE(prio)      ((prio) - MAX_RT_PRIO - 20)
  66#define TASK_NICE(p)            PRIO_TO_NICE((p)->static_prio)
  67
  68/*
  69 * 'User priority' is the nice value converted to something we
  70 * can work with better when scaling various scheduler parameters,
  71 * it's a [ 0 ... 39 ] range.
  72 */
  73#define USER_PRIO(p)            ((p)-MAX_RT_PRIO)
  74#define TASK_USER_PRIO(p)       USER_PRIO((p)->static_prio)
  75#define MAX_USER_PRIO           (USER_PRIO(MAX_PRIO))
  76
  77/*
  78 * Some helpers for converting nanosecond timing to jiffy resolution
  79 */
  80#define NS_TO_JIFFIES(TIME)     ((TIME) / (1000000000 / HZ))
  81#define JIFFIES_TO_NS(TIME)     ((TIME) * (1000000000 / HZ))
  82
  83/*
  84 * These are the 'tuning knobs' of the scheduler:
  85 *
  86 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
  87 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
  88 * Timeslices get refilled after they expire.
  89 */
  90#define MIN_TIMESLICE           max(5 * HZ / 1000, 1)
  91#define DEF_TIMESLICE           (100 * HZ / 1000)
  92#define ON_RUNQUEUE_WEIGHT       30
  93#define CHILD_PENALTY            95
  94#define PARENT_PENALTY          100
  95#define EXIT_WEIGHT               3
  96#define PRIO_BONUS_RATIO         25
  97#define MAX_BONUS               (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
  98#define INTERACTIVE_DELTA         2
  99#define MAX_SLEEP_AVG           (DEF_TIMESLICE * MAX_BONUS)
 100#define STARVATION_LIMIT        (MAX_SLEEP_AVG)
 101#define NS_MAX_SLEEP_AVG        (JIFFIES_TO_NS(MAX_SLEEP_AVG))
 102
 103/*
 104 * If a task is 'interactive' then we reinsert it in the active
 105 * array after it has expired its current timeslice. (it will not
 106 * continue to run immediately, it will still roundrobin with
 107 * other interactive tasks.)
 108 *
 109 * This part scales the interactivity limit depending on niceness.
 110 *
 111 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
 112 * Here are a few examples of different nice levels:
 113 *
 114 *  TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
 115 *  TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
 116 *  TASK_INTERACTIVE(  0): [1,1,1,1,0,0,0,0,0,0,0]
 117 *  TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
 118 *  TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
 119 *
 120 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
 121 *  priority range a task can explore, a value of '1' means the
 122 *  task is rated interactive.)
 123 *
 124 * Ie. nice +19 tasks can never get 'interactive' enough to be
 125 * reinserted into the active array. And only heavily CPU-hog nice -20
 126 * tasks will be expired. Default nice 0 tasks are somewhere between,
 127 * it takes some effort for them to get interactive, but it's not
 128 * too hard.
 129 */
 130
 131#define CURRENT_BONUS(p) \
 132        (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
 133                MAX_SLEEP_AVG)
 134
 135#define GRANULARITY     (10 * HZ / 1000 ? : 1)
 136
 137#ifdef CONFIG_SMP
 138#define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
 139                (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
 140                        num_online_cpus())
 141#else
 142#define TIMESLICE_GRANULARITY(p)        (GRANULARITY * \
 143                (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
 144#endif
 145
 146#define SCALE(v1,v1_max,v2_max) \
 147        (v1) * (v2_max) / (v1_max)
 148
 149#define DELTA(p) \
 150        (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
 151                INTERACTIVE_DELTA)
 152
 153#define TASK_INTERACTIVE(p) \
 154        ((p)->prio <= (p)->static_prio - DELTA(p))
 155
 156#define INTERACTIVE_SLEEP(p) \
 157        (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
 158                (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
 159
 160#define TASK_PREEMPTS_CURR(p, rq) \
 161        ((p)->prio < (rq)->curr->prio)
 162
 163/*
 164 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
 165 * to time slice values: [800ms ... 100ms ... 5ms]
 166 *
 167 * The higher a thread's priority, the bigger timeslices
 168 * it gets during one round of execution. But even the lowest
 169 * priority thread gets MIN_TIMESLICE worth of execution time.
 170 */
 171
 172#define SCALE_PRIO(x, prio) \
 173        max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
 174
 175static unsigned int static_prio_timeslice(int static_prio)
 176{
 177        if (static_prio < NICE_TO_PRIO(0))
 178                return SCALE_PRIO(DEF_TIMESLICE * 4, static_prio);
 179        else
 180                return SCALE_PRIO(DEF_TIMESLICE, static_prio);
 181}
 182
 183static inline unsigned int task_timeslice(struct task_struct *p)
 184{
 185        return static_prio_timeslice(p->static_prio);
 186}
 187
 188/*
 189 * These are the runqueue data structures:
 190 */
 191
 192struct prio_array {
 193        unsigned int nr_active;
 194        DECLARE_BITMAP(bitmap, MAX_PRIO+1); /* include 1 bit for delimiter */
 195        struct list_head queue[MAX_PRIO];
 196};
 197
 198/*
 199 * This is the main, per-CPU runqueue data structure.
 200 *
 201 * Locking rule: those places that want to lock multiple runqueues
 202 * (such as the load balancing or the thread migration code), lock
 203 * acquire operations must be ordered by ascending &runqueue.
 204 */
 205struct rq {
 206        spinlock_t lock;
 207
 208        /*
 209         * nr_running and cpu_load should be in the same cacheline because
 210         * remote CPUs use both these fields when doing load calculation.
 211         */
 212        unsigned long nr_running;
 213        unsigned long raw_weighted_load;
 214#ifdef CONFIG_SMP
 215        unsigned long cpu_load[3];
 216#endif
 217        unsigned long long nr_switches;
 218
 219        /*
 220         * This is part of a global counter where only the total sum
 221         * over all CPUs matters. A task can increase this counter on
 222         * one CPU and if it got migrated afterwards it may decrease
 223         * it on another CPU. Always updated under the runqueue lock:
 224         */
 225        unsigned long nr_uninterruptible;
 226
 227        unsigned long expired_timestamp;
 228        unsigned long long timestamp_last_tick;
 229        struct task_struct *curr, *idle;
 230        struct mm_struct *prev_mm;
 231        struct prio_array *active, *expired, arrays[2];
 232        int best_expired_prio;
 233        atomic_t nr_iowait;
 234
 235#ifdef CONFIG_SMP
 236        struct sched_domain *sd;
 237
 238        /* For active balancing */
 239        int active_balance;
 240        int push_cpu;
 241        int cpu;                /* cpu of this runqueue */
 242
 243        struct task_struct *migration_thread;
 244        struct list_head migration_queue;
 245#endif
 246
 247#ifdef CONFIG_SCHEDSTATS
 248        /* latency stats */
 249        struct sched_info rq_sched_info;
 250
 251        /* sys_sched_yield() stats */
 252        unsigned long yld_exp_empty;
 253        unsigned long yld_act_empty;
 254        unsigned long yld_both_empty;
 255        unsigned long yld_cnt;
 256
 257        /* schedule() stats */
 258        unsigned long sched_switch;
 259        unsigned long sched_cnt;
 260        unsigned long sched_goidle;
 261
 262        /* try_to_wake_up() stats */
 263        unsigned long ttwu_cnt;
 264        unsigned long ttwu_local;
 265#endif
 266        struct lock_class_key rq_lock_key;
 267};
 268
 269static DEFINE_PER_CPU(struct rq, runqueues);
 270
 271static inline int cpu_of(struct rq *rq)
 272{
 273#ifdef CONFIG_SMP
 274        return rq->cpu;
 275#else
 276        return 0;
 277#endif
 278}
 279
 280/*
 281 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
 282 * See detach_destroy_domains: synchronize_sched for details.
 283 *
 284 * The domain tree of any CPU may only be accessed from within
 285 * preempt-disabled sections.
 286 */
 287#define for_each_domain(cpu, __sd) \
 288        for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
 289
 290#define cpu_rq(cpu)             (&per_cpu(runqueues, (cpu)))
 291#define this_rq()               (&__get_cpu_var(runqueues))
 292#define task_rq(p)              cpu_rq(task_cpu(p))
 293#define cpu_curr(cpu)           (cpu_rq(cpu)->curr)
 294
 295#ifndef prepare_arch_switch
 296# define prepare_arch_switch(next)      do { } while (0)
 297#endif
 298#ifndef finish_arch_switch
 299# define finish_arch_switch(prev)       do { } while (0)
 300#endif
 301
 302#ifndef __ARCH_WANT_UNLOCKED_CTXSW
 303static inline int task_running(struct rq *rq, struct task_struct *p)
 304{
 305        return rq->curr == p;
 306}
 307
 308static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
 309{
 310}
 311
 312static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
 313{
 314#ifdef CONFIG_DEBUG_SPINLOCK
 315        /* this is a valid case when another task releases the spinlock */
 316        rq->lock.owner = current;
 317#endif
 318        /*
 319         * If we are tracking spinlock dependencies then we have to
 320         * fix up the runqueue lock - which gets 'carried over' from
 321         * prev into current:
 322         */
 323        spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
 324
 325        spin_unlock_irq(&rq->lock);
 326}
 327
 328#else /* __ARCH_WANT_UNLOCKED_CTXSW */
 329static inline int task_running(struct rq *rq, struct task_struct *p)
 330{
 331#ifdef CONFIG_SMP
 332        return p->oncpu;
 333#else
 334        return rq->curr == p;
 335#endif
 336}
 337
 338static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
 339{
 340#ifdef CONFIG_SMP
 341        /*
 342         * We can optimise this out completely for !SMP, because the
 343         * SMP rebalancing from interrupt is the only thing that cares
 344         * here.
 345         */
 346        next->oncpu = 1;
 347#endif
 348#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
 349        spin_unlock_irq(&rq->lock);
 350#else
 351        spin_unlock(&rq->lock);
 352#endif
 353}
 354
 355static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
 356{
 357#ifdef CONFIG_SMP
 358        /*
 359         * After ->oncpu is cleared, the task can be moved to a different CPU.
 360         * We must ensure this doesn't happen until the switch is completely
 361         * finished.
 362         */
 363        smp_wmb();
 364        prev->oncpu = 0;
 365#endif
 366#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
 367        local_irq_enable();
 368#endif
 369}
 370#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
 371
 372/*
 373 * __task_rq_lock - lock the runqueue a given task resides on.
 374 * Must be called interrupts disabled.
 375 */
 376static inline struct rq *__task_rq_lock(struct task_struct *p)
 377        __acquires(rq->lock)
 378{
 379        struct rq *rq;
 380
 381repeat_lock_task:
 382        rq = task_rq(p);
 383        spin_lock(&rq->lock);
 384        if (unlikely(rq != task_rq(p))) {
 385                spin_unlock(&rq->lock);
 386                goto repeat_lock_task;
 387        }
 388        return rq;
 389}
 390
 391/*
 392 * task_rq_lock - lock the runqueue a given task resides on and disable
 393 * interrupts.  Note the ordering: we can safely lookup the task_rq without
 394 * explicitly disabling preemption.
 395 */
 396static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
 397        __acquires(rq->lock)
 398{
 399        struct rq *rq;
 400
 401repeat_lock_task:
 402        local_irq_save(*flags);
 403        rq = task_rq(p);
 404        spin_lock(&rq->lock);
 405        if (unlikely(rq != task_rq(p))) {
 406                spin_unlock_irqrestore(&rq->lock, *flags);
 407                goto repeat_lock_task;
 408        }
 409        return rq;
 410}
 411
 412static inline void __task_rq_unlock(struct rq *rq)
 413        __releases(rq->lock)
 414{
 415        spin_unlock(&rq->lock);
 416}
 417
 418static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
 419        __releases(rq->lock)
 420{
 421        spin_unlock_irqrestore(&rq->lock, *flags);
 422}
 423
 424#ifdef CONFIG_SCHEDSTATS
 425/*
 426 * bump this up when changing the output format or the meaning of an existing
 427 * format, so that tools can adapt (or abort)
 428 */
 429#define SCHEDSTAT_VERSION 12
 430
 431static int show_schedstat(struct seq_file *seq, void *v)
 432{
 433        int cpu;
 434
 435        seq_printf(seq, "version %d\n", SCHEDSTAT_VERSION);
 436        seq_printf(seq, "timestamp %lu\n", jiffies);
 437        for_each_online_cpu(cpu) {
 438                struct rq *rq = cpu_rq(cpu);
 439#ifdef CONFIG_SMP
 440                struct sched_domain *sd;
 441                int dcnt = 0;
 442#endif
 443
 444                /* runqueue-specific stats */
 445                seq_printf(seq,
 446                    "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
 447                    cpu, rq->yld_both_empty,
 448                    rq->yld_act_empty, rq->yld_exp_empty, rq->yld_cnt,
 449                    rq->sched_switch, rq->sched_cnt, rq->sched_goidle,
 450                    rq->ttwu_cnt, rq->ttwu_local,
 451                    rq->rq_sched_info.cpu_time,
 452                    rq->rq_sched_info.run_delay, rq->rq_sched_info.pcnt);
 453
 454                seq_printf(seq, "\n");
 455
 456#ifdef CONFIG_SMP
 457                /* domain-specific stats */
 458                preempt_disable();
 459                for_each_domain(cpu, sd) {
 460                        enum idle_type itype;
 461                        char mask_str[NR_CPUS];
 462
 463                        cpumask_scnprintf(mask_str, NR_CPUS, sd->span);
 464                        seq_printf(seq, "domain%d %s", dcnt++, mask_str);
 465                        for (itype = SCHED_IDLE; itype < MAX_IDLE_TYPES;
 466                                        itype++) {
 467                                seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu",
 468                                    sd->lb_cnt[itype],
 469                                    sd->lb_balanced[itype],
 470                                    sd->lb_failed[itype],
 471                                    sd->lb_imbalance[itype],
 472                                    sd->lb_gained[itype],
 473                                    sd->lb_hot_gained[itype],
 474                                    sd->lb_nobusyq[itype],
 475                                    sd->lb_nobusyg[itype]);
 476                        }
 477                        seq_printf(seq, " %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu\n",
 478                            sd->alb_cnt, sd->alb_failed, sd->alb_pushed,
 479                            sd->sbe_cnt, sd->sbe_balanced, sd->sbe_pushed,
 480                            sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
 481                            sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
 482                }
 483                preempt_enable();
 484#endif
 485        }
 486        return 0;
 487}
 488
 489static int schedstat_open(struct inode *inode, struct file *file)
 490{
 491        unsigned int size = PAGE_SIZE * (1 + num_online_cpus() / 32);
 492        char *buf = kmalloc(size, GFP_KERNEL);
 493        struct seq_file *m;
 494        int res;
 495
 496        if (!buf)
 497                return -ENOMEM;
 498        res = single_open(file, show_schedstat, NULL);
 499        if (!res) {
 500                m = file->private_data;
 501                m->buf = buf;
 502                m->size = size;
 503        } else
 504                kfree(buf);
 505        return res;
 506}
 507
 508struct file_operations proc_schedstat_operations = {
 509        .open    = schedstat_open,
 510        .read    = seq_read,
 511        .llseek  = seq_lseek,
 512        .release = single_release,
 513};
 514
 515/*
 516 * Expects runqueue lock to be held for atomicity of update
 517 */
 518static inline void
 519rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
 520{
 521        if (rq) {
 522                rq->rq_sched_info.run_delay += delta_jiffies;
 523                rq->rq_sched_info.pcnt++;
 524        }
 525}
 526
 527/*
 528 * Expects runqueue lock to be held for atomicity of update
 529 */
 530static inline void
 531rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
 532{
 533        if (rq)
 534                rq->rq_sched_info.cpu_time += delta_jiffies;
 535}
 536# define schedstat_inc(rq, field)       do { (rq)->field++; } while (0)
 537# define schedstat_add(rq, field, amt)  do { (rq)->field += (amt); } while (0)
 538#else /* !CONFIG_SCHEDSTATS */
 539static inline void
 540rq_sched_info_arrive(struct rq *rq, unsigned long delta_jiffies)
 541{}
 542static inline void
 543rq_sched_info_depart(struct rq *rq, unsigned long delta_jiffies)
 544{}
 545# define schedstat_inc(rq, field)       do { } while (0)
 546# define schedstat_add(rq, field, amt)  do { } while (0)
 547#endif
 548
 549/*
 550 * rq_lock - lock a given runqueue and disable interrupts.
 551 */
 552static inline struct rq *this_rq_lock(void)
 553        __acquires(rq->lock)
 554{
 555        struct rq *rq;
 556
 557        local_irq_disable();
 558        rq = this_rq();
 559        spin_lock(&rq->lock);
 560
 561        return rq;
 562}
 563
 564#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
 565/*
 566 * Called when a process is dequeued from the active array and given
 567 * the cpu.  We should note that with the exception of interactive
 568 * tasks, the expired queue will become the active queue after the active
 569 * queue is empty, without explicitly dequeuing and requeuing tasks in the
 570 * expired queue.  (Interactive tasks may be requeued directly to the
 571 * active queue, thus delaying tasks in the expired queue from running;
 572 * see scheduler_tick()).
 573 *
 574 * This function is only called from sched_info_arrive(), rather than
 575 * dequeue_task(). Even though a task may be queued and dequeued multiple
 576 * times as it is shuffled about, we're really interested in knowing how
 577 * long it was from the *first* time it was queued to the time that it
 578 * finally hit a cpu.
 579 */
 580static inline void sched_info_dequeued(struct task_struct *t)
 581{
 582        t->sched_info.last_queued = 0;
 583}
 584
 585/*
 586 * Called when a task finally hits the cpu.  We can now calculate how
 587 * long it was waiting to run.  We also note when it began so that we
 588 * can keep stats on how long its timeslice is.
 589 */
 590static void sched_info_arrive(struct task_struct *t)
 591{
 592        unsigned long now = jiffies, delta_jiffies = 0;
 593
 594        if (t->sched_info.last_queued)
 595                delta_jiffies = now - t->sched_info.last_queued;
 596        sched_info_dequeued(t);
 597        t->sched_info.run_delay += delta_jiffies;
 598        t->sched_info.last_arrival = now;
 599        t->sched_info.pcnt++;
 600
 601        rq_sched_info_arrive(task_rq(t), delta_jiffies);
 602}
 603
 604/*
 605 * Called when a process is queued into either the active or expired
 606 * array.  The time is noted and later used to determine how long we
 607 * had to wait for us to reach the cpu.  Since the expired queue will
 608 * become the active queue after active queue is empty, without dequeuing
 609 * and requeuing any tasks, we are interested in queuing to either. It
 610 * is unusual but not impossible for tasks to be dequeued and immediately
 611 * requeued in the same or another array: this can happen in sched_yield(),
 612 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
 613 * to runqueue.
 614 *
 615 * This function is only called from enqueue_task(), but also only updates
 616 * the timestamp if it is already not set.  It's assumed that
 617 * sched_info_dequeued() will clear that stamp when appropriate.
 618 */
 619static inline void sched_info_queued(struct task_struct *t)
 620{
 621        if (unlikely(sched_info_on()))
 622                if (!t->sched_info.last_queued)
 623                        t->sched_info.last_queued = jiffies;
 624}
 625
 626/*
 627 * Called when a process ceases being the active-running process, either
 628 * voluntarily or involuntarily.  Now we can calculate how long we ran.
 629 */
 630static inline void sched_info_depart(struct task_struct *t)
 631{
 632        unsigned long delta_jiffies = jiffies - t->sched_info.last_arrival;
 633
 634        t->sched_info.cpu_time += delta_jiffies;
 635        rq_sched_info_depart(task_rq(t), delta_jiffies);
 636}
 637
 638/*
 639 * Called when tasks are switched involuntarily due, typically, to expiring
 640 * their time slice.  (This may also be called when switching to or from
 641 * the idle task.)  We are only called when prev != next.
 642 */
 643static inline void
 644__sched_info_switch(struct task_struct *prev, struct task_struct *next)
 645{
 646        struct rq *rq = task_rq(prev);
 647
 648        /*
 649         * prev now departs the cpu.  It's not interesting to record
 650         * stats about how efficient we were at scheduling the idle
 651         * process, however.
 652         */
 653        if (prev != rq->idle)
 654                sched_info_depart(prev);
 655
 656        if (next != rq->idle)
 657                sched_info_arrive(next);
 658}
 659static inline void
 660sched_info_switch(struct task_struct *prev, struct task_struct *next)
 661{
 662        if (unlikely(sched_info_on()))
 663                __sched_info_switch(prev, next);
 664}
 665#else
 666#define sched_info_queued(t)            do { } while (0)
 667#define sched_info_switch(t, next)      do { } while (0)
 668#endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
 669
 670/*
 671 * Adding/removing a task to/from a priority array:
 672 */
 673static void dequeue_task(struct task_struct *p, struct prio_array *array)
 674{
 675        array->nr_active--;
 676        list_del(&p->run_list);
 677        if (list_empty(array->queue + p->prio))
 678                __clear_bit(p->prio, array->bitmap);
 679}
 680
 681static void enqueue_task(struct task_struct *p, struct prio_array *array)
 682{
 683        sched_info_queued(p);
 684        list_add_tail(&p->run_list, array->queue + p->prio);
 685        __set_bit(p->prio, array->bitmap);
 686        array->nr_active++;
 687        p->array = array;
 688}
 689
 690/*
 691 * Put task to the end of the run list without the overhead of dequeue
 692 * followed by enqueue.
 693 */
 694static void requeue_task(struct task_struct *p, struct prio_array *array)
 695{
 696        list_move_tail(&p->run_list, array->queue + p->prio);
 697}
 698
 699static inline void
 700enqueue_task_head(struct task_struct *p, struct prio_array *array)
 701{
 702        list_add(&p->run_list, array->queue + p->prio);
 703        __set_bit(p->prio, array->bitmap);
 704        array->nr_active++;
 705        p->array = array;
 706}
 707
 708/*
 709 * __normal_prio - return the priority that is based on the static
 710 * priority but is modified by bonuses/penalties.
 711 *
 712 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
 713 * into the -5 ... 0 ... +5 bonus/penalty range.
 714 *
 715 * We use 25% of the full 0...39 priority range so that:
 716 *
 717 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
 718 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
 719 *
 720 * Both properties are important to certain workloads.
 721 */
 722
 723static inline int __normal_prio(struct task_struct *p)
 724{
 725        int bonus, prio;
 726
 727        bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
 728
 729        prio = p->static_prio - bonus;
 730        if (prio < MAX_RT_PRIO)
 731                prio = MAX_RT_PRIO;
 732        if (prio > MAX_PRIO-1)
 733                prio = MAX_PRIO-1;
 734        return prio;
 735}
 736
 737/*
 738 * To aid in avoiding the subversion of "niceness" due to uneven distribution
 739 * of tasks with abnormal "nice" values across CPUs the contribution that
 740 * each task makes to its run queue's load is weighted according to its
 741 * scheduling class and "nice" value.  For SCHED_NORMAL tasks this is just a
 742 * scaled version of the new time slice allocation that they receive on time
 743 * slice expiry etc.
 744 */
 745
 746/*
 747 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
 748 * If static_prio_timeslice() is ever changed to break this assumption then
 749 * this code will need modification
 750 */
 751#define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
 752#define LOAD_WEIGHT(lp) \
 753        (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
 754#define PRIO_TO_LOAD_WEIGHT(prio) \
 755        LOAD_WEIGHT(static_prio_timeslice(prio))
 756#define RTPRIO_TO_LOAD_WEIGHT(rp) \
 757        (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
 758
 759static void set_load_weight(struct task_struct *p)
 760{
 761        if (has_rt_policy(p)) {
 762#ifdef CONFIG_SMP
 763                if (p == task_rq(p)->migration_thread)
 764                        /*
 765                         * The migration thread does the actual balancing.
 766                         * Giving its load any weight will skew balancing
 767                         * adversely.
 768                         */
 769                        p->load_weight = 0;
 770                else
 771#endif
 772                        p->load_weight = RTPRIO_TO_LOAD_WEIGHT(p->rt_priority);
 773        } else
 774                p->load_weight = PRIO_TO_LOAD_WEIGHT(p->static_prio);
 775}
 776
 777static inline void
 778inc_raw_weighted_load(struct rq *rq, const struct task_struct *p)
 779{
 780        rq->raw_weighted_load += p->load_weight;
 781}
 782
 783static inline void
 784dec_raw_weighted_load(struct rq *rq, const struct task_struct *p)
 785{
 786        rq->raw_weighted_load -= p->load_weight;
 787}
 788
 789static inline void inc_nr_running(struct task_struct *p, struct rq *rq)
 790{
 791        rq->nr_running++;
 792        inc_raw_weighted_load(rq, p);
 793}
 794
 795static inline void dec_nr_running(struct task_struct *p, struct rq *rq)
 796{
 797        rq->nr_running--;
 798        dec_raw_weighted_load(rq, p);
 799}
 800
 801/*
 802 * Calculate the expected normal priority: i.e. priority
 803 * without taking RT-inheritance into account. Might be
 804 * boosted by interactivity modifiers. Changes upon fork,
 805 * setprio syscalls, and whenever the interactivity
 806 * estimator recalculates.
 807 */
 808static inline int normal_prio(struct task_struct *p)
 809{
 810        int prio;
 811
 812        if (has_rt_policy(p))
 813                prio = MAX_RT_PRIO-1 - p->rt_priority;
 814        else
 815                prio = __normal_prio(p);
 816        return prio;
 817}
 818
 819/*
 820 * Calculate the current priority, i.e. the priority
 821 * taken into account by the scheduler. This value might
 822 * be boosted by RT tasks, or might be boosted by
 823 * interactivity modifiers. Will be RT if the task got
 824 * RT-boosted. If not then it returns p->normal_prio.
 825 */
 826static int effective_prio(struct task_struct *p)
 827{
 828        p->normal_prio = normal_prio(p);
 829        /*
 830         * If we are RT tasks or we were boosted to RT priority,
 831         * keep the priority unchanged. Otherwise, update priority
 832         * to the normal priority:
 833         */
 834        if (!rt_prio(p->prio))
 835                return p->normal_prio;
 836        return p->prio;
 837}
 838
 839/*
 840 * __activate_task - move a task to the runqueue.
 841 */
 842static void __activate_task(struct task_struct *p, struct rq *rq)
 843{
 844        struct prio_array *target = rq->active;
 845
 846        if (batch_task(p))
 847                target = rq->expired;
 848        enqueue_task(p, target);
 849        inc_nr_running(p, rq);
 850}
 851
 852/*
 853 * __activate_idle_task - move idle task to the _front_ of runqueue.
 854 */
 855static inline void __activate_idle_task(struct task_struct *p, struct rq *rq)
 856{
 857        enqueue_task_head(p, rq->active);
 858        inc_nr_running(p, rq);
 859}
 860
 861/*
 862 * Recalculate p->normal_prio and p->prio after having slept,
 863 * updating the sleep-average too:
 864 */
 865static int recalc_task_prio(struct task_struct *p, unsigned long long now)
 866{
 867        /* Caller must always ensure 'now >= p->timestamp' */
 868        unsigned long sleep_time = now - p->timestamp;
 869
 870        if (batch_task(p))
 871                sleep_time = 0;
 872
 873        if (likely(sleep_time > 0)) {
 874                /*
 875                 * This ceiling is set to the lowest priority that would allow
 876                 * a task to be reinserted into the active array on timeslice
 877                 * completion.
 878                 */
 879                unsigned long ceiling = INTERACTIVE_SLEEP(p);
 880
 881                if (p->mm && sleep_time > ceiling && p->sleep_avg < ceiling) {
 882                        /*
 883                         * Prevents user tasks from achieving best priority
 884                         * with one single large enough sleep.
 885                         */
 886                        p->sleep_avg = ceiling;
 887                        /*
 888                         * Using INTERACTIVE_SLEEP() as a ceiling places a
 889                         * nice(0) task 1ms sleep away from promotion, and
 890                         * gives it 700ms to round-robin with no chance of
 891                         * being demoted.  This is more than generous, so
 892                         * mark this sleep as non-interactive to prevent the
 893                         * on-runqueue bonus logic from intervening should
 894                         * this task not receive cpu immediately.
 895                         */
 896                        p->sleep_type = SLEEP_NONINTERACTIVE;
 897                } else {
 898                        /*
 899                         * Tasks waking from uninterruptible sleep are
 900                         * limited in their sleep_avg rise as they
 901                         * are likely to be waiting on I/O
 902                         */
 903                        if (p->sleep_type == SLEEP_NONINTERACTIVE && p->mm) {
 904                                if (p->sleep_avg >= ceiling)
 905                                        sleep_time = 0;
 906                                else if (p->sleep_avg + sleep_time >=
 907                                         ceiling) {
 908                                                p->sleep_avg = ceiling;
 909                                                sleep_time = 0;
 910                                }
 911                        }
 912
 913                        /*
 914                         * This code gives a bonus to interactive tasks.
 915                         *
 916                         * The boost works by updating the 'average sleep time'
 917                         * value here, based on ->timestamp. The more time a
 918                         * task spends sleeping, the higher the average gets -
 919                         * and the higher the priority boost gets as well.
 920                         */
 921                        p->sleep_avg += sleep_time;
 922
 923                }
 924                if (p->sleep_avg > NS_MAX_SLEEP_AVG)
 925                        p->sleep_avg = NS_MAX_SLEEP_AVG;
 926        }
 927
 928        return effective_prio(p);
 929}
 930
 931/*
 932 * activate_task - move a task to the runqueue and do priority recalculation
 933 *
 934 * Update all the scheduling statistics stuff. (sleep average
 935 * calculation, priority modifiers, etc.)
 936 */
 937static void activate_task(struct task_struct *p, struct rq *rq, int local)
 938{
 939        unsigned long long now;
 940
 941        now = sched_clock();
 942#ifdef CONFIG_SMP
 943        if (!local) {
 944                /* Compensate for drifting sched_clock */
 945                struct rq *this_rq = this_rq();
 946                now = (now - this_rq->timestamp_last_tick)
 947                        + rq->timestamp_last_tick;
 948        }
 949#endif
 950
 951        if (!rt_task(p))
 952                p->prio = recalc_task_prio(p, now);
 953
 954        /*
 955         * This checks to make sure it's not an uninterruptible task
 956         * that is now waking up.
 957         */
 958        if (p->sleep_type == SLEEP_NORMAL) {
 959                /*
 960                 * Tasks which were woken up by interrupts (ie. hw events)
 961                 * are most likely of interactive nature. So we give them
 962                 * the credit of extending their sleep time to the period
 963                 * of time they spend on the runqueue, waiting for execution
 964                 * on a CPU, first time around:
 965                 */
 966                if (in_interrupt())
 967                        p->sleep_type = SLEEP_INTERRUPTED;
 968                else {
 969                        /*
 970                         * Normal first-time wakeups get a credit too for
 971                         * on-runqueue time, but it will be weighted down:
 972                         */
 973                        p->sleep_type = SLEEP_INTERACTIVE;
 974                }
 975        }
 976        p->timestamp = now;
 977
 978        __activate_task(p, rq);
 979}
 980
 981/*
 982 * deactivate_task - remove a task from the runqueue.
 983 */
 984static void deactivate_task(struct task_struct *p, struct rq *rq)
 985{
 986        dec_nr_running(p, rq);
 987        dequeue_task(p, p->array);
 988        p->array = NULL;
 989}
 990
 991/*
 992 * resched_task - mark a task 'to be rescheduled now'.
 993 *
 994 * On UP this means the setting of the need_resched flag, on SMP it
 995 * might also involve a cross-CPU call to trigger the scheduler on
 996 * the target CPU.
 997 */
 998#ifdef CONFIG_SMP
 999
1000#ifndef tsk_is_polling
1001#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1002#endif
1003
1004static void resched_task(struct task_struct *p)
1005{
1006        int cpu;
1007
1008        assert_spin_locked(&task_rq(p)->lock);
1009
1010        if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1011                return;
1012
1013        set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1014
1015        cpu = task_cpu(p);
1016        if (cpu == smp_processor_id())
1017                return;
1018
1019        /* NEED_RESCHED must be visible before we test polling */
1020        smp_mb();
1021        if (!tsk_is_polling(p))
1022                smp_send_reschedule(cpu);
1023}
1024#else
1025static inline void resched_task(struct task_struct *p)
1026{
1027        assert_spin_locked(&task_rq(p)->lock);
1028        set_tsk_need_resched(p);
1029}
1030#endif
1031
1032/**
1033 * task_curr - is this task currently executing on a CPU?
1034 * @p: the task in question.
1035 */
1036inline int task_curr(const struct task_struct *p)
1037{
1038        return cpu_curr(task_cpu(p)) == p;
1039}
1040
1041/* Used instead of source_load when we know the type == 0 */
1042unsigned long weighted_cpuload(const int cpu)
1043{
1044        return cpu_rq(cpu)->raw_weighted_load;
1045}
1046
1047#ifdef CONFIG_SMP
1048struct migration_req {
1049        struct list_head list;
1050
1051        struct task_struct *task;
1052        int dest_cpu;
1053
1054        struct completion done;
1055};
1056
1057/*
1058 * The task's runqueue lock must be held.
1059 * Returns true if you have to wait for migration thread.
1060 */
1061static int
1062migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1063{
1064        struct rq *rq = task_rq(p);
1065
1066        /*
1067         * If the task is not on a runqueue (and not running), then
1068         * it is sufficient to simply update the task's cpu field.
1069         */
1070        if (!p->array && !task_running(rq, p)) {
1071                set_task_cpu(p, dest_cpu);
1072                return 0;
1073        }
1074
1075        init_completion(&req->done);
1076        req->task = p;
1077        req->dest_cpu = dest_cpu;
1078        list_add(&req->list, &rq->migration_queue);
1079
1080        return 1;
1081}
1082
1083/*
1084 * wait_task_inactive - wait for a thread to unschedule.
1085 *
1086 * The caller must ensure that the task *will* unschedule sometime soon,
1087 * else this function might spin for a *long* time. This function can't
1088 * be called with interrupts off, or it may introduce deadlock with
1089 * smp_call_function() if an IPI is sent by the same process we are
1090 * waiting to become inactive.
1091 */
1092void wait_task_inactive(struct task_struct *p)
1093{
1094        unsigned long flags;
1095        struct rq *rq;
1096        int preempted;
1097
1098repeat:
1099        rq = task_rq_lock(p, &flags);
1100        /* Must be off runqueue entirely, not preempted. */
1101        if (unlikely(p->array || task_running(rq, p))) {
1102                /* If it's preempted, we yield.  It could be a while. */
1103                preempted = !task_running(rq, p);
1104                task_rq_unlock(rq, &flags);
1105                cpu_relax();
1106                if (preempted)
1107                        yield();
1108                goto repeat;
1109        }
1110        task_rq_unlock(rq, &flags);
1111}
1112
1113/***
1114 * kick_process - kick a running thread to enter/exit the kernel
1115 * @p: the to-be-kicked thread
1116 *
1117 * Cause a process which is running on another CPU to enter
1118 * kernel-mode, without any delay. (to get signals handled.)
1119 *
1120 * NOTE: this function doesnt have to take the runqueue lock,
1121 * because all it wants to ensure is that the remote task enters
1122 * the kernel. If the IPI races and the task has been migrated
1123 * to another CPU then no harm is done and the purpose has been
1124 * achieved as well.
1125 */
1126void kick_process(struct task_struct *p)
1127{
1128        int cpu;
1129
1130        preempt_disable();
1131        cpu = task_cpu(p);
1132        if ((cpu != smp_processor_id()) && task_curr(p))
1133                smp_send_reschedule(cpu);
1134        preempt_enable();
1135}
1136
1137/*
1138 * Return a low guess at the load of a migration-source cpu weighted
1139 * according to the scheduling class and "nice" value.
1140 *
1141 * We want to under-estimate the load of migration sources, to
1142 * balance conservatively.
1143 */
1144static inline unsigned long source_load(int cpu, int type)
1145{
1146        struct rq *rq = cpu_rq(cpu);
1147
1148        if (type == 0)
1149                return rq->raw_weighted_load;
1150
1151        return min(rq->cpu_load[type-1], rq->raw_weighted_load);
1152}
1153
1154/*
1155 * Return a high guess at the load of a migration-target cpu weighted
1156 * according to the scheduling class and "nice" value.
1157 */
1158static inline unsigned long target_load(int cpu, int type)
1159{
1160        struct rq *rq = cpu_rq(cpu);
1161
1162        if (type == 0)
1163                return rq->raw_weighted_load;
1164
1165        return max(rq->cpu_load[type-1], rq->raw_weighted_load);
1166}
1167
1168/*
1169 * Return the average load per task on the cpu's run queue
1170 */
1171static inline unsigned long cpu_avg_load_per_task(int cpu)
1172{
1173        struct rq *rq = cpu_rq(cpu);
1174        unsigned long n = rq->nr_running;
1175
1176        return n ? rq->raw_weighted_load / n : SCHED_LOAD_SCALE;
1177}
1178
1179/*
1180 * find_idlest_group finds and returns the least busy CPU group within the
1181 * domain.
1182 */
1183static struct sched_group *
1184find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1185{
1186        struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1187        unsigned long min_load = ULONG_MAX, this_load = 0;
1188        int load_idx = sd->forkexec_idx;
1189        int imbalance = 100 + (sd->imbalance_pct-100)/2;
1190
1191        do {
1192                unsigned long load, avg_load;
1193                int local_group;
1194                int i;
1195
1196                /* Skip over this group if it has no CPUs allowed */
1197                if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1198                        goto nextgroup;
1199
1200                local_group = cpu_isset(this_cpu, group->cpumask);
1201
1202                /* Tally up the load of all CPUs in the group */
1203                avg_load = 0;
1204
1205                for_each_cpu_mask(i, group->cpumask) {
1206                        /* Bias balancing toward cpus of our domain */
1207                        if (local_group)
1208                                load = source_load(i, load_idx);
1209                        else
1210                                load = target_load(i, load_idx);
1211
1212                        avg_load += load;
1213                }
1214
1215                /* Adjust by relative CPU power of the group */
1216                avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1217
1218                if (local_group) {
1219                        this_load = avg_load;
1220                        this = group;
1221                } else if (avg_load < min_load) {
1222                        min_load = avg_load;
1223                        idlest = group;
1224                }
1225nextgroup:
1226                group = group->next;
1227        } while (group != sd->groups);
1228
1229        if (!idlest || 100*this_load < imbalance*min_load)
1230                return NULL;
1231        return idlest;
1232}
1233
1234/*
1235 * find_idlest_queue - find the idlest runqueue among the cpus in group.
1236 */
1237static int
1238find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1239{
1240        cpumask_t tmp;
1241        unsigned long load, min_load = ULONG_MAX;
1242        int idlest = -1;
1243        int i;
1244
1245        /* Traverse only the allowed CPUs */
1246        cpus_and(tmp, group->cpumask, p->cpus_allowed);
1247
1248        for_each_cpu_mask(i, tmp) {
1249                load = weighted_cpuload(i);
1250
1251                if (load < min_load || (load == min_load && i == this_cpu)) {
1252                        min_load = load;
1253                        idlest = i;
1254                }
1255        }
1256
1257        return idlest;
1258}
1259
1260/*
1261 * sched_balance_self: balance the current task (running on cpu) in domains
1262 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1263 * SD_BALANCE_EXEC.
1264 *
1265 * Balance, ie. select the least loaded group.
1266 *
1267 * Returns the target CPU number, or the same CPU if no balancing is needed.
1268 *
1269 * preempt must be disabled.
1270 */
1271static int sched_balance_self(int cpu, int flag)
1272{
1273        struct task_struct *t = current;
1274        struct sched_domain *tmp, *sd = NULL;
1275
1276        for_each_domain(cpu, tmp) {
1277                /*
1278                 * If power savings logic is enabled for a domain, stop there.
1279                 */
1280                if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1281                        break;
1282                if (tmp->flags & flag)
1283                        sd = tmp;
1284        }
1285
1286        while (sd) {
1287                cpumask_t span;
1288                struct sched_group *group;
1289                int new_cpu;
1290                int weight;
1291
1292                span = sd->span;
1293                group = find_idlest_group(sd, t, cpu);
1294                if (!group)
1295                        goto nextlevel;
1296
1297                new_cpu = find_idlest_cpu(group, t, cpu);
1298                if (new_cpu == -1 || new_cpu == cpu)
1299                        goto nextlevel;
1300
1301                /* Now try balancing at a lower domain level */
1302                cpu = new_cpu;
1303nextlevel:
1304                sd = NULL;
1305                weight = cpus_weight(span);
1306                for_each_domain(cpu, tmp) {
1307                        if (weight <= cpus_weight(tmp->span))
1308                                break;
1309                        if (tmp->flags & flag)
1310                                sd = tmp;
1311                }
1312                /* while loop will break here if sd == NULL */
1313        }
1314
1315        return cpu;
1316}
1317
1318#endif /* CONFIG_SMP */
1319
1320/*
1321 * wake_idle() will wake a task on an idle cpu if task->cpu is
1322 * not idle and an idle cpu is available.  The span of cpus to
1323 * search starts with cpus closest then further out as needed,
1324 * so we always favor a closer, idle cpu.
1325 *
1326 * Returns the CPU we should wake onto.
1327 */
1328#if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1329static int wake_idle(int cpu, struct task_struct *p)
1330{
1331        cpumask_t tmp;
1332        struct sched_domain *sd;
1333        int i;
1334
1335        if (idle_cpu(cpu))
1336                return cpu;
1337
1338        for_each_domain(cpu, sd) {
1339                if (sd->flags & SD_WAKE_IDLE) {
1340                        cpus_and(tmp, sd->span, p->cpus_allowed);
1341                        for_each_cpu_mask(i, tmp) {
1342                                if (idle_cpu(i))
1343                                        return i;
1344                        }
1345                }
1346                else
1347                        break;
1348        }
1349        return cpu;
1350}
1351#else
1352static inline int wake_idle(int cpu, struct task_struct *p)
1353{
1354        return cpu;
1355}
1356#endif
1357
1358/***
1359 * try_to_wake_up - wake up a thread
1360 * @p: the to-be-woken-up thread
1361 * @state: the mask of task states that can be woken
1362 * @sync: do a synchronous wakeup?
1363 *
1364 * Put it on the run-queue if it's not already there. The "current"
1365 * thread is always on the run-queue (except when the actual
1366 * re-schedule is in progress), and as such you're allowed to do
1367 * the simpler "current->state = TASK_RUNNING" to mark yourself
1368 * runnable without the overhead of this.
1369 *
1370 * returns failure only if the task is already active.
1371 */
1372static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1373{
1374        int cpu, this_cpu, success = 0;
1375        unsigned long flags;
1376        long old_state;
1377        struct rq *rq;
1378#ifdef CONFIG_SMP
1379        struct sched_domain *sd, *this_sd = NULL;
1380        unsigned long load, this_load;
1381        int new_cpu;
1382#endif
1383
1384        rq = task_rq_lock(p, &flags);
1385        old_state = p->state;
1386        if (!(old_state & state))
1387                goto out;
1388
1389        if (p->array)
1390                goto out_running;
1391
1392        cpu = task_cpu(p);
1393        this_cpu = smp_processor_id();
1394
1395#ifdef CONFIG_SMP
1396        if (unlikely(task_running(rq, p)))
1397                goto out_activate;
1398
1399        new_cpu = cpu;
1400
1401        schedstat_inc(rq, ttwu_cnt);
1402        if (cpu == this_cpu) {
1403                schedstat_inc(rq, ttwu_local);
1404                goto out_set_cpu;
1405        }
1406
1407        for_each_domain(this_cpu, sd) {
1408                if (cpu_isset(cpu, sd->span)) {
1409                        schedstat_inc(sd, ttwu_wake_remote);
1410                        this_sd = sd;
1411                        break;
1412                }
1413        }
1414
1415        if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1416                goto out_set_cpu;
1417
1418        /*
1419         * Check for affine wakeup and passive balancing possibilities.
1420         */
1421        if (this_sd) {
1422                int idx = this_sd->wake_idx;
1423                unsigned int imbalance;
1424
1425                imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1426
1427                load = source_load(cpu, idx);
1428                this_load = target_load(this_cpu, idx);
1429
1430                new_cpu = this_cpu; /* Wake to this CPU if we can */
1431
1432                if (this_sd->flags & SD_WAKE_AFFINE) {
1433                        unsigned long tl = this_load;
1434                        unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu);
1435
1436                        /*
1437                         * If sync wakeup then subtract the (maximum possible)
1438                         * effect of the currently running task from the load
1439                         * of the current CPU:
1440                         */
1441                        if (sync)
1442                                tl -= current->load_weight;
1443
1444                        if ((tl <= load &&
1445                                tl + target_load(cpu, idx) <= tl_per_task) ||
1446                                100*(tl + p->load_weight) <= imbalance*load) {
1447                                /*
1448                                 * This domain has SD_WAKE_AFFINE and
1449                                 * p is cache cold in this domain, and
1450                                 * there is no bad imbalance.
1451                                 */
1452                                schedstat_inc(this_sd, ttwu_move_affine);
1453                                goto out_set_cpu;
1454                        }
1455                }
1456
1457                /*
1458                 * Start passive balancing when half the imbalance_pct
1459                 * limit is reached.
1460                 */
1461                if (this_sd->flags & SD_WAKE_BALANCE) {
1462                        if (imbalance*this_load <= 100*load) {
1463                                schedstat_inc(this_sd, ttwu_move_balance);
1464                                goto out_set_cpu;
1465                        }
1466                }
1467        }
1468
1469        new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1470out_set_cpu:
1471        new_cpu = wake_idle(new_cpu, p);
1472        if (new_cpu != cpu) {
1473                set_task_cpu(p, new_cpu);
1474                task_rq_unlock(rq, &flags);
1475                /* might preempt at this point */
1476                rq = task_rq_lock(p, &flags);
1477                old_state = p->state;
1478                if (!(old_state & state))
1479                        goto out;
1480                if (p->array)
1481                        goto out_running;
1482
1483                this_cpu = smp_processor_id();
1484                cpu = task_cpu(p);
1485        }
1486
1487out_activate:
1488#endif /* CONFIG_SMP */
1489        if (old_state == TASK_UNINTERRUPTIBLE) {
1490                rq->nr_uninterruptible--;
1491                /*
1492                 * Tasks on involuntary sleep don't earn
1493                 * sleep_avg beyond just interactive state.
1494                 */
1495                p->sleep_type = SLEEP_NONINTERACTIVE;
1496        } else
1497
1498        /*
1499         * Tasks that have marked their sleep as noninteractive get
1500         * woken up with their sleep average not weighted in an
1501         * interactive way.
1502         */
1503                if (old_state & TASK_NONINTERACTIVE)
1504                        p->sleep_type = SLEEP_NONINTERACTIVE;
1505
1506
1507        activate_task(p, rq, cpu == this_cpu);
1508        /*
1509         * Sync wakeups (i.e. those types of wakeups where the waker
1510         * has indicated that it will leave the CPU in short order)
1511         * don't trigger a preemption, if the woken up task will run on
1512         * this cpu. (in this case the 'I will reschedule' promise of
1513         * the waker guarantees that the freshly woken up task is going
1514         * to be considered on this CPU.)
1515         */
1516        if (!sync || cpu != this_cpu) {
1517                if (TASK_PREEMPTS_CURR(p, rq))
1518                        resched_task(rq->curr);
1519        }
1520        success = 1;
1521
1522out_running:
1523        p->state = TASK_RUNNING;
1524out:
1525        task_rq_unlock(rq, &flags);
1526
1527        return success;
1528}
1529
1530int fastcall wake_up_process(struct task_struct *p)
1531{
1532        return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1533                                 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1534}
1535EXPORT_SYMBOL(wake_up_process);
1536
1537int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1538{
1539        return try_to_wake_up(p, state, 0);
1540}
1541
1542/*
1543 * Perform scheduler related setup for a newly forked process p.
1544 * p is forked by current.
1545 */
1546void fastcall sched_fork(struct task_struct *p, int clone_flags)
1547{
1548        int cpu = get_cpu();
1549
1550#ifdef CONFIG_SMP
1551        cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1552#endif
1553        set_task_cpu(p, cpu);
1554
1555        /*
1556         * We mark the process as running here, but have not actually
1557         * inserted it onto the runqueue yet. This guarantees that
1558         * nobody will actually run it, and a signal or other external
1559         * event cannot wake it up and insert it on the runqueue either.
1560         */
1561        p->state = TASK_RUNNING;
1562
1563        /*
1564         * Make sure we do not leak PI boosting priority to the child:
1565         */
1566        p->prio = current->normal_prio;
1567
1568        INIT_LIST_HEAD(&p->run_list);
1569        p->array = NULL;
1570#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1571        if (unlikely(sched_info_on()))
1572                memset(&p->sched_info, 0, sizeof(p->sched_info));
1573#endif
1574#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1575        p->oncpu = 0;
1576#endif
1577#ifdef CONFIG_PREEMPT
1578        /* Want to start with kernel preemption disabled. */
1579        task_thread_info(p)->preempt_count = 1;
1580#endif
1581        /*
1582         * Share the timeslice between parent and child, thus the
1583         * total amount of pending timeslices in the system doesn't change,
1584         * resulting in more scheduling fairness.
1585         */
1586        local_irq_disable();
1587        p->time_slice = (current->time_slice + 1) >> 1;
1588        /*
1589         * The remainder of the first timeslice might be recovered by
1590         * the parent if the child exits early enough.
1591         */
1592        p->first_time_slice = 1;
1593        current->time_slice >>= 1;
1594        p->timestamp = sched_clock();
1595        if (unlikely(!current->time_slice)) {
1596                /*
1597                 * This case is rare, it happens when the parent has only
1598                 * a single jiffy left from its timeslice. Taking the
1599                 * runqueue lock is not a problem.
1600                 */
1601                current->time_slice = 1;
1602                scheduler_tick();
1603        }
1604        local_irq_enable();
1605        put_cpu();
1606}
1607
1608/*
1609 * wake_up_new_task - wake up a newly created task for the first time.
1610 *
1611 * This function will do some initial scheduler statistics housekeeping
1612 * that must be done for every newly created context, then puts the task
1613 * on the runqueue and wakes it.
1614 */
1615void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1616{
1617        struct rq *rq, *this_rq;
1618        unsigned long flags;
1619        int this_cpu, cpu;
1620
1621        rq = task_rq_lock(p, &flags);
1622        BUG_ON(p->state != TASK_RUNNING);
1623        this_cpu = smp_processor_id();
1624        cpu = task_cpu(p);
1625
1626        /*
1627         * We decrease the sleep average of forking parents
1628         * and children as well, to keep max-interactive tasks
1629         * from forking tasks that are max-interactive. The parent
1630         * (current) is done further down, under its lock.
1631         */
1632        p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1633                CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1634
1635        p->prio = effective_prio(p);
1636
1637        if (likely(cpu == this_cpu)) {
1638                if (!(clone_flags & CLONE_VM)) {
1639                        /*
1640                         * The VM isn't cloned, so we're in a good position to
1641                         * do child-runs-first in anticipation of an exec. This
1642                         * usually avoids a lot of COW overhead.
1643                         */
1644                        if (unlikely(!current->array))
1645                                __activate_task(p, rq);
1646                        else {
1647                                p->prio = current->prio;
1648                                p->normal_prio = current->normal_prio;
1649                                list_add_tail(&p->run_list, &current->run_list);
1650                                p->array = current->array;
1651                                p->array->nr_active++;
1652                                inc_nr_running(p, rq);
1653                        }
1654                        set_need_resched();
1655                } else
1656                        /* Run child last */
1657                        __activate_task(p, rq);
1658                /*
1659                 * We skip the following code due to cpu == this_cpu
1660                 *
1661                 *   task_rq_unlock(rq, &flags);
1662                 *   this_rq = task_rq_lock(current, &flags);
1663                 */
1664                this_rq = rq;
1665        } else {
1666                this_rq = cpu_rq(this_cpu);
1667
1668                /*
1669                 * Not the local CPU - must adjust timestamp. This should
1670                 * get optimised away in the !CONFIG_SMP case.
1671                 */
1672                p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1673                                        + rq->timestamp_last_tick;
1674                __activate_task(p, rq);
1675                if (TASK_PREEMPTS_CURR(p, rq))
1676                        resched_task(rq->curr);
1677
1678                /*
1679                 * Parent and child are on different CPUs, now get the
1680                 * parent runqueue to update the parent's ->sleep_avg:
1681                 */
1682                task_rq_unlock(rq, &flags);
1683                this_rq = task_rq_lock(current, &flags);
1684        }
1685        current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1686                PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1687        task_rq_unlock(this_rq, &flags);
1688}
1689
1690/*
1691 * Potentially available exiting-child timeslices are
1692 * retrieved here - this way the parent does not get
1693 * penalized for creating too many threads.
1694 *
1695 * (this cannot be used to 'generate' timeslices
1696 * artificially, because any timeslice recovered here
1697 * was given away by the parent in the first place.)
1698 */
1699void fastcall sched_exit(struct task_struct *p)
1700{
1701        unsigned long flags;
1702        struct rq *rq;
1703
1704        /*
1705         * If the child was a (relative-) CPU hog then decrease
1706         * the sleep_avg of the parent as well.
1707         */
1708        rq = task_rq_lock(p->parent, &flags);
1709        if (p->first_time_slice && task_cpu(p) == task_cpu(p->parent)) {
1710                p->parent->time_slice += p->time_slice;
1711                if (unlikely(p->parent->time_slice > task_timeslice(p)))
1712                        p->parent->time_slice = task_timeslice(p);
1713        }
1714        if (p->sleep_avg < p->parent->sleep_avg)
1715                p->parent->sleep_avg = p->parent->sleep_avg /
1716                (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1717                (EXIT_WEIGHT + 1);
1718        task_rq_unlock(rq, &flags);
1719}
1720
1721/**
1722 * prepare_task_switch - prepare to switch tasks
1723 * @rq: the runqueue preparing to switch
1724 * @next: the task we are going to switch to.
1725 *
1726 * This is called with the rq lock held and interrupts off. It must
1727 * be paired with a subsequent finish_task_switch after the context
1728 * switch.
1729 *
1730 * prepare_task_switch sets up locking and calls architecture specific
1731 * hooks.
1732 */
1733static inline void prepare_task_switch(struct rq *rq, struct task_struct *next)
1734{
1735        prepare_lock_switch(rq, next);
1736        prepare_arch_switch(next);
1737}
1738
1739/**
1740 * finish_task_switch - clean up after a task-switch
1741 * @rq: runqueue associated with task-switch
1742 * @prev: the thread we just switched away from.
1743 *
1744 * finish_task_switch must be called after the context switch, paired
1745 * with a prepare_task_switch call before the context switch.
1746 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1747 * and do any other architecture-specific cleanup actions.
1748 *
1749 * Note that we may have delayed dropping an mm in context_switch(). If
1750 * so, we finish that here outside of the runqueue lock.  (Doing it
1751 * with the lock held can cause deadlocks; see schedule() for
1752 * details.)
1753 */
1754static inline void finish_task_switch(struct rq *rq, struct task_struct *prev)
1755        __releases(rq->lock)
1756{
1757        struct mm_struct *mm = rq->prev_mm;
1758        unsigned long prev_task_flags;
1759
1760        rq->prev_mm = NULL;
1761
1762        /*
1763         * A task struct has one reference for the use as "current".
1764         * If a task dies, then it sets EXIT_ZOMBIE in tsk->exit_state and
1765         * calls schedule one last time. The schedule call will never return,
1766         * and the scheduled task must drop that reference.
1767         * The test for EXIT_ZOMBIE must occur while the runqueue locks are
1768         * still held, otherwise prev could be scheduled on another cpu, die
1769         * there before we look at prev->state, and then the reference would
1770         * be dropped twice.
1771         *              Manfred Spraul <manfred@colorfullife.com>
1772         */
1773        prev_task_flags = prev->flags;
1774        finish_arch_switch(prev);
1775        finish_lock_switch(rq, prev);
1776        if (mm)
1777                mmdrop(mm);
1778        if (unlikely(prev_task_flags & PF_DEAD)) {
1779                /*
1780                 * Remove function-return probe instances associated with this
1781                 * task and put them back on the free list.
1782                 */
1783                kprobe_flush_task(prev);
1784                put_task_struct(prev);
1785        }
1786}
1787
1788/**
1789 * schedule_tail - first thing a freshly forked thread must call.
1790 * @prev: the thread we just switched away from.
1791 */
1792asmlinkage void schedule_tail(struct task_struct *prev)
1793        __releases(rq->lock)
1794{
1795        struct rq *rq = this_rq();
1796
1797        finish_task_switch(rq, prev);
1798#ifdef __ARCH_WANT_UNLOCKED_CTXSW
1799        /* In this case, finish_task_switch does not reenable preemption */
1800        preempt_enable();
1801#endif
1802        if (current->set_child_tid)
1803                put_user(current->pid, current->set_child_tid);
1804}
1805
1806/*
1807 * context_switch - switch to the new MM and the new
1808 * thread's register state.
1809 */
1810static inline struct task_struct *
1811context_switch(struct rq *rq, struct task_struct *prev,
1812               struct task_struct *next)
1813{
1814        struct mm_struct *mm = next->mm;
1815        struct mm_struct *oldmm = prev->active_mm;
1816
1817        if (unlikely(!mm)) {
1818                next->active_mm = oldmm;
1819                atomic_inc(&oldmm->mm_count);
1820                enter_lazy_tlb(oldmm, next);
1821        } else
1822                switch_mm(oldmm, mm, next);
1823
1824        if (unlikely(!prev->mm)) {
1825                prev->active_mm = NULL;
1826                WARN_ON(rq->prev_mm);
1827                rq->prev_mm = oldmm;
1828        }
1829        /*
1830         * Since the runqueue lock will be released by the next
1831         * task (which is an invalid locking op but in the case
1832         * of the scheduler it's an obvious special-case), so we
1833         * do an early lockdep release here:
1834         */
1835#ifndef __ARCH_WANT_UNLOCKED_CTXSW
1836        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1837#endif
1838
1839        /* Here we just switch the register state and the stack. */
1840        switch_to(prev, next, prev);
1841
1842        return prev;
1843}
1844
1845/*
1846 * nr_running, nr_uninterruptible and nr_context_switches:
1847 *
1848 * externally visible scheduler statistics: current number of runnable
1849 * threads, current number of uninterruptible-sleeping threads, total
1850 * number of context switches performed since bootup.
1851 */
1852unsigned long nr_running(void)
1853{
1854        unsigned long i, sum = 0;
1855
1856        for_each_online_cpu(i)
1857                sum += cpu_rq(i)->nr_running;
1858
1859        return sum;
1860}
1861
1862unsigned long nr_uninterruptible(void)
1863{
1864        unsigned long i, sum = 0;
1865
1866        for_each_possible_cpu(i)
1867                sum += cpu_rq(i)->nr_uninterruptible;
1868
1869        /*
1870         * Since we read the counters lockless, it might be slightly
1871         * inaccurate. Do not allow it to go below zero though:
1872         */
1873        if (unlikely((long)sum < 0))
1874                sum = 0;
1875
1876        return sum;
1877}
1878
1879unsigned long long nr_context_switches(void)
1880{
1881        int i;
1882        unsigned long long sum = 0;
1883
1884        for_each_possible_cpu(i)
1885                sum += cpu_rq(i)->nr_switches;
1886
1887        return sum;
1888}
1889
1890unsigned long nr_iowait(void)
1891{
1892        unsigned long i, sum = 0;
1893
1894        for_each_possible_cpu(i)
1895                sum += atomic_read(&cpu_rq(i)->nr_iowait);
1896
1897        return sum;
1898}
1899
1900unsigned long nr_active(void)
1901{
1902        unsigned long i, running = 0, uninterruptible = 0;
1903
1904        for_each_online_cpu(i) {
1905                running += cpu_rq(i)->nr_running;
1906                uninterruptible += cpu_rq(i)->nr_uninterruptible;
1907        }
1908
1909        if (unlikely((long)uninterruptible < 0))
1910                uninterruptible = 0;
1911
1912        return running + uninterruptible;
1913}
1914
1915#ifdef CONFIG_SMP
1916
1917/*
1918 * Is this task likely cache-hot:
1919 */
1920static inline int
1921task_hot(struct task_struct *p, unsigned long long now, struct sched_domain *sd)
1922{
1923        return (long long)(now - p->last_ran) < (long long)sd->cache_hot_time;
1924}
1925
1926/*
1927 * double_rq_lock - safely lock two runqueues
1928 *
1929 * Note this does not disable interrupts like task_rq_lock,
1930 * you need to do so manually before calling.
1931 */
1932static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1933        __acquires(rq1->lock)
1934        __acquires(rq2->lock)
1935{
1936        if (rq1 == rq2) {
1937                spin_lock(&rq1->lock);
1938                __acquire(rq2->lock);   /* Fake it out ;) */
1939        } else {
1940                if (rq1 < rq2) {
1941                        spin_lock(&rq1->lock);
1942                        spin_lock(&rq2->lock);
1943                } else {
1944                        spin_lock(&rq2->lock);
1945                        spin_lock(&rq1->lock);
1946                }
1947        }
1948}
1949
1950/*
1951 * double_rq_unlock - safely unlock two runqueues
1952 *
1953 * Note this does not restore interrupts like task_rq_unlock,
1954 * you need to do so manually after calling.
1955 */
1956static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1957        __releases(rq1->lock)
1958        __releases(rq2->lock)
1959{
1960        spin_unlock(&rq1->lock);
1961        if (rq1 != rq2)
1962                spin_unlock(&rq2->lock);
1963        else
1964                __release(rq2->lock);
1965}
1966
1967/*
1968 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1969 */
1970static void double_lock_balance(struct rq *this_rq, struct rq *busiest)
1971        __releases(this_rq->lock)
1972        __acquires(busiest->lock)
1973        __acquires(this_rq->lock)
1974{
1975        if (unlikely(!spin_trylock(&busiest->lock))) {
1976                if (busiest < this_rq) {
1977                        spin_unlock(&this_rq->lock);
1978                        spin_lock(&busiest->lock);
1979                        spin_lock(&this_rq->lock);
1980                } else
1981                        spin_lock(&busiest->lock);
1982        }
1983}
1984
1985/*
1986 * If dest_cpu is allowed for this process, migrate the task to it.
1987 * This is accomplished by forcing the cpu_allowed mask to only
1988 * allow dest_cpu, which will force the cpu onto dest_cpu.  Then
1989 * the cpu_allowed mask is restored.
1990 */
1991static void sched_migrate_task(struct task_struct *p, int dest_cpu)
1992{
1993        struct migration_req req;
1994        unsigned long flags;
1995        struct rq *rq;
1996
1997        rq = task_rq_lock(p, &flags);
1998        if (!cpu_isset(dest_cpu, p->cpus_allowed)
1999            || unlikely(cpu_is_offline(dest_cpu)))
2000                goto out;
2001
2002        /* force the process onto the specified CPU */
2003        if (migrate_task(p, dest_cpu, &req)) {
2004                /* Need to wait for migration thread (might exit: take ref). */
2005                struct task_struct *mt = rq->migration_thread;
2006
2007                get_task_struct(mt);
2008                task_rq_unlock(rq, &flags);
2009                wake_up_process(mt);
2010                put_task_struct(mt);
2011                wait_for_completion(&req.done);
2012
2013                return;
2014        }
2015out:
2016        task_rq_unlock(rq, &flags);
2017}
2018
2019/*
2020 * sched_exec - execve() is a valuable balancing opportunity, because at
2021 * this point the task has the smallest effective memory and cache footprint.
2022 */
2023void sched_exec(void)
2024{
2025        int new_cpu, this_cpu = get_cpu();
2026        new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2027        put_cpu();
2028        if (new_cpu != this_cpu)
2029                sched_migrate_task(current, new_cpu);
2030}
2031
2032/*
2033 * pull_task - move a task from a remote runqueue to the local runqueue.
2034 * Both runqueues must be locked.
2035 */
2036static void pull_task(struct rq *src_rq, struct prio_array *src_array,
2037                      struct task_struct *p, struct rq *this_rq,
2038                      struct prio_array *this_array, int this_cpu)
2039{
2040        dequeue_task(p, src_array);
2041        dec_nr_running(p, src_rq);
2042        set_task_cpu(p, this_cpu);
2043        inc_nr_running(p, this_rq);
2044        enqueue_task(p, this_array);
2045        p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
2046                                + this_rq->timestamp_last_tick;
2047        /*
2048         * Note that idle threads have a prio of MAX_PRIO, for this test
2049         * to be always true for them.
2050         */
2051        if (TASK_PREEMPTS_CURR(p, this_rq))
2052                resched_task(this_rq->curr);
2053}
2054
2055/*
2056 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2057 */
2058static
2059int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2060                     struct sched_domain *sd, enum idle_type idle,
2061                     int *all_pinned)
2062{
2063        /*
2064         * We do not migrate tasks that are:
2065         * 1) running (obviously), or
2066         * 2) cannot be migrated to this CPU due to cpus_allowed, or
2067         * 3) are cache-hot on their current CPU.
2068         */
2069        if (!cpu_isset(this_cpu, p->cpus_allowed))
2070                return 0;
2071        *all_pinned = 0;
2072
2073        if (task_running(rq, p))
2074                return 0;
2075
2076        /*
2077         * Aggressive migration if:
2078         * 1) task is cache cold, or
2079         * 2) too many balance attempts have failed.
2080         */
2081
2082        if (sd->nr_balance_failed > sd->cache_nice_tries)
2083                return 1;
2084
2085        if (task_hot(p, rq->timestamp_last_tick, sd))
2086                return 0;
2087        return 1;
2088}
2089
2090#define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2091
2092/*
2093 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2094 * load from busiest to this_rq, as part of a balancing operation within
2095 * "domain". Returns the number of tasks moved.
2096 *
2097 * Called with both runqueues locked.
2098 */
2099static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2100                      unsigned long max_nr_move, unsigned long max_load_move,
2101                      struct sched_domain *sd, enum idle_type idle,
2102                      int *all_pinned)
2103{
2104        int idx, pulled = 0, pinned = 0, this_best_prio, best_prio,
2105            best_prio_seen, skip_for_load;
2106        struct prio_array *array, *dst_array;
2107        struct list_head *head, *curr;
2108        struct task_struct *tmp;
2109        long rem_load_move;
2110
2111        if (max_nr_move == 0 || max_load_move == 0)
2112                goto out;
2113
2114        rem_load_move = max_load_move;
2115        pinned = 1;
2116        this_best_prio = rq_best_prio(this_rq);
2117        best_prio = rq_best_prio(busiest);
2118        /*
2119         * Enable handling of the case where there is more than one task
2120         * with the best priority.   If the current running task is one
2121         * of those with prio==best_prio we know it won't be moved
2122         * and therefore it's safe to override the skip (based on load) of
2123         * any task we find with that prio.
2124         */
2125        best_prio_seen = best_prio == busiest->curr->prio;
2126
2127        /*
2128         * We first consider expired tasks. Those will likely not be
2129         * executed in the near future, and they are most likely to
2130         * be cache-cold, thus switching CPUs has the least effect
2131         * on them.
2132         */
2133        if (busiest->expired->nr_active) {
2134                array = busiest->expired;
2135                dst_array = this_rq->expired;
2136        } else {
2137                array = busiest->active;
2138                dst_array = this_rq->active;
2139        }
2140
2141new_array:
2142        /* Start searching at priority 0: */
2143        idx = 0;
2144skip_bitmap:
2145        if (!idx)
2146                idx = sched_find_first_bit(array->bitmap);
2147        else
2148                idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
2149        if (idx >= MAX_PRIO) {
2150                if (array == busiest->expired && busiest->active->nr_active) {
2151                        array = busiest->active;
2152                        dst_array = this_rq->active;
2153                        goto new_array;
2154                }
2155                goto out;
2156        }
2157
2158        head = array->queue + idx;
2159        curr = head->prev;
2160skip_queue:
2161        tmp = list_entry(curr, struct task_struct, run_list);
2162
2163        curr = curr->prev;
2164
2165        /*
2166         * To help distribute high priority tasks accross CPUs we don't
2167         * skip a task if it will be the highest priority task (i.e. smallest
2168         * prio value) on its new queue regardless of its load weight
2169         */
2170        skip_for_load = tmp->load_weight > rem_load_move;
2171        if (skip_for_load && idx < this_best_prio)
2172                skip_for_load = !best_prio_seen && idx == best_prio;
2173        if (skip_for_load ||
2174            !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
2175
2176                best_prio_seen |= idx == best_prio;
2177                if (curr != head)
2178                        goto skip_queue;
2179                idx++;
2180                goto skip_bitmap;
2181        }
2182
2183#ifdef CONFIG_SCHEDSTATS
2184        if (task_hot(tmp, busiest->timestamp_last_tick, sd))
2185                schedstat_inc(sd, lb_hot_gained[idle]);
2186#endif
2187
2188        pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
2189        pulled++;
2190        rem_load_move -= tmp->load_weight;
2191
2192        /*
2193         * We only want to steal up to the prescribed number of tasks
2194         * and the prescribed amount of weighted load.
2195         */
2196        if (pulled < max_nr_move && rem_load_move > 0) {
2197                if (idx < this_best_prio)
2198                        this_best_prio = idx;
2199                if (curr != head)
2200                        goto skip_queue;
2201                idx++;
2202                goto skip_bitmap;
2203        }
2204out:
2205        /*
2206         * Right now, this is the only place pull_task() is called,
2207         * so we can safely collect pull_task() stats here rather than
2208         * inside pull_task().
2209         */
2210        schedstat_add(sd, lb_gained[idle], pulled);
2211
2212        if (all_pinned)
2213                *all_pinned = pinned;
2214        return pulled;
2215}
2216
2217/*
2218 * find_busiest_group finds and returns the busiest CPU group within the
2219 * domain. It calculates and returns the amount of weighted load which
2220 * should be moved to restore balance via the imbalance parameter.
2221 */
2222static struct sched_group *
2223find_busiest_group(struct sched_domain *sd, int this_cpu,
2224                   unsigned long *imbalance, enum idle_type idle, int *sd_idle,
2225                   cpumask_t *cpus)
2226{
2227        struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2228        unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2229        unsigned long max_pull;
2230        unsigned long busiest_load_per_task, busiest_nr_running;
2231        unsigned long this_load_per_task, this_nr_running;
2232        int load_idx;
2233#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2234        int power_savings_balance = 1;
2235        unsigned long leader_nr_running = 0, min_load_per_task = 0;
2236        unsigned long min_nr_running = ULONG_MAX;
2237        struct sched_group *group_min = NULL, *group_leader = NULL;
2238#endif
2239
2240        max_load = this_load = total_load = total_pwr = 0;
2241        busiest_load_per_task = busiest_nr_running = 0;
2242        this_load_per_task = this_nr_running = 0;
2243        if (idle == NOT_IDLE)
2244                load_idx = sd->busy_idx;
2245        else if (idle == NEWLY_IDLE)
2246                load_idx = sd->newidle_idx;
2247        else
2248                load_idx = sd->idle_idx;
2249
2250        do {
2251                unsigned long load, group_capacity;
2252                int local_group;
2253                int i;
2254                unsigned long sum_nr_running, sum_weighted_load;
2255
2256                local_group = cpu_isset(this_cpu, group->cpumask);
2257
2258                /* Tally up the load of all CPUs in the group */
2259                sum_weighted_load = sum_nr_running = avg_load = 0;
2260
2261                for_each_cpu_mask(i, group->cpumask) {
2262                        struct rq *rq;
2263
2264                        if (!cpu_isset(i, *cpus))
2265                                continue;
2266
2267                        rq = cpu_rq(i);
2268
2269                        if (*sd_idle && !idle_cpu(i))
2270                                *sd_idle = 0;
2271
2272                        /* Bias balancing toward cpus of our domain */
2273                        if (local_group)
2274                                load = target_load(i, load_idx);
2275                        else
2276                                load = source_load(i, load_idx);
2277
2278                        avg_load += load;
2279                        sum_nr_running += rq->nr_running;
2280                        sum_weighted_load += rq->raw_weighted_load;
2281                }
2282
2283                total_load += avg_load;
2284                total_pwr += group->cpu_power;
2285
2286                /* Adjust by relative CPU power of the group */
2287                avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2288
2289                group_capacity = group->cpu_power / SCHED_LOAD_SCALE;
2290
2291                if (local_group) {
2292                        this_load = avg_load;
2293                        this = group;
2294                        this_nr_running = sum_nr_running;
2295                        this_load_per_task = sum_weighted_load;
2296                } else if (avg_load > max_load &&
2297                           sum_nr_running > group_capacity) {
2298                        max_load = avg_load;
2299                        busiest = group;
2300                        busiest_nr_running = sum_nr_running;
2301                        busiest_load_per_task = sum_weighted_load;
2302                }
2303
2304#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2305                /*
2306                 * Busy processors will not participate in power savings
2307                 * balance.
2308                 */
2309                if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2310                        goto group_next;
2311
2312                /*
2313                 * If the local group is idle or completely loaded
2314                 * no need to do power savings balance at this domain
2315                 */
2316                if (local_group && (this_nr_running >= group_capacity ||
2317                                    !this_nr_running))
2318                        power_savings_balance = 0;
2319
2320                /*
2321                 * If a group is already running at full capacity or idle,
2322                 * don't include that group in power savings calculations
2323                 */
2324                if (!power_savings_balance || sum_nr_running >= group_capacity
2325                    || !sum_nr_running)
2326                        goto group_next;
2327
2328                /*
2329                 * Calculate the group which has the least non-idle load.
2330                 * This is the group from where we need to pick up the load
2331                 * for saving power
2332                 */
2333                if ((sum_nr_running < min_nr_running) ||
2334                    (sum_nr_running == min_nr_running &&
2335                     first_cpu(group->cpumask) <
2336                     first_cpu(group_min->cpumask))) {
2337                        group_min = group;
2338                        min_nr_running = sum_nr_running;
2339                        min_load_per_task = sum_weighted_load /
2340                                                sum_nr_running;
2341                }
2342
2343                /*
2344                 * Calculate the group which is almost near its
2345                 * capacity but still has some space to pick up some load
2346                 * from other group and save more power
2347                 */
2348                if (sum_nr_running <= group_capacity - 1) {
2349                        if (sum_nr_running > leader_nr_running ||
2350                            (sum_nr_running == leader_nr_running &&
2351                             first_cpu(group->cpumask) >
2352                              first_cpu(group_leader->cpumask))) {
2353                                group_leader = group;
2354                                leader_nr_running = sum_nr_running;
2355                        }
2356                }
2357group_next:
2358#endif
2359                group = group->next;
2360        } while (group != sd->groups);
2361
2362        if (!busiest || this_load >= max_load || busiest_nr_running == 0)
2363                goto out_balanced;
2364
2365        avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2366
2367        if (this_load >= avg_load ||
2368                        100*max_load <= sd->imbalance_pct*this_load)
2369                goto out_balanced;
2370
2371        busiest_load_per_task /= busiest_nr_running;
2372        /*
2373         * We're trying to get all the cpus to the average_load, so we don't
2374         * want to push ourselves above the average load, nor do we wish to
2375         * reduce the max loaded cpu below the average load, as either of these
2376         * actions would just result in more rebalancing later, and ping-pong
2377         * tasks around. Thus we look for the minimum possible imbalance.
2378         * Negative imbalances (*we* are more loaded than anyone else) will
2379         * be counted as no imbalance for these purposes -- we can't fix that
2380         * by pulling tasks to us.  Be careful of negative numbers as they'll
2381         * appear as very large values with unsigned longs.
2382         */
2383        if (max_load <= busiest_load_per_task)
2384                goto out_balanced;
2385
2386        /*
2387         * In the presence of smp nice balancing, certain scenarios can have
2388         * max load less than avg load(as we skip the groups at or below
2389         * its cpu_power, while calculating max_load..)
2390         */
2391        if (max_load < avg_load) {
2392                *imbalance = 0;
2393                goto small_imbalance;
2394        }
2395
2396        /* Don't want to pull so many tasks that a group would go idle */
2397        max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
2398
2399        /* How much load to actually move to equalise the imbalance */
2400        *imbalance = min(max_pull * busiest->cpu_power,
2401                                (avg_load - this_load) * this->cpu_power)
2402                        / SCHED_LOAD_SCALE;
2403
2404        /*
2405         * if *imbalance is less than the average load per runnable task
2406         * there is no gaurantee that any tasks will be moved so we'll have
2407         * a think about bumping its value to force at least one task to be
2408         * moved
2409         */
2410        if (*imbalance < busiest_load_per_task) {
2411                unsigned long tmp, pwr_now, pwr_move;
2412                unsigned int imbn;
2413
2414small_imbalance:
2415                pwr_move = pwr_now = 0;
2416                imbn = 2;
2417                if (this_nr_running) {
2418                        this_load_per_task /= this_nr_running;
2419                        if (busiest_load_per_task > this_load_per_task)
2420                                imbn = 1;
2421                } else
2422                        this_load_per_task = SCHED_LOAD_SCALE;
2423
2424                if (max_load - this_load >= busiest_load_per_task * imbn) {
2425                        *imbalance = busiest_load_per_task;
2426                        return busiest;
2427                }
2428
2429                /*
2430                 * OK, we don't have enough imbalance to justify moving tasks,
2431                 * however we may be able to increase total CPU power used by
2432                 * moving them.
2433                 */
2434
2435                pwr_now += busiest->cpu_power *
2436                        min(busiest_load_per_task, max_load);
2437                pwr_now += this->cpu_power *
2438                        min(this_load_per_task, this_load);
2439                pwr_now /= SCHED_LOAD_SCALE;
2440
2441                /* Amount of load we'd subtract */
2442                tmp = busiest_load_per_task*SCHED_LOAD_SCALE/busiest->cpu_power;
2443                if (max_load > tmp)
2444                        pwr_move += busiest->cpu_power *
2445                                min(busiest_load_per_task, max_load - tmp);
2446
2447                /* Amount of load we'd add */
2448                if (max_load*busiest->cpu_power <
2449                                busiest_load_per_task*SCHED_LOAD_SCALE)
2450                        tmp = max_load*busiest->cpu_power/this->cpu_power;
2451                else
2452                        tmp = busiest_load_per_task*SCHED_LOAD_SCALE/this->cpu_power;
2453                pwr_move += this->cpu_power*min(this_load_per_task, this_load + tmp);
2454                pwr_move /= SCHED_LOAD_SCALE;
2455
2456                /* Move if we gain throughput */
2457                if (pwr_move <= pwr_now)
2458                        goto out_balanced;
2459
2460                *imbalance = busiest_load_per_task;
2461        }
2462
2463        return busiest;
2464
2465out_balanced:
2466#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2467        if (idle == NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2468                goto ret;
2469
2470        if (this == group_leader && group_leader != group_min) {
2471                *imbalance = min_load_per_task;
2472                return group_min;
2473        }
2474ret:
2475#endif
2476        *imbalance = 0;
2477        return NULL;
2478}
2479
2480/*
2481 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2482 */
2483static struct rq *
2484find_busiest_queue(struct sched_group *group, enum idle_type idle,
2485                   unsigned long imbalance, cpumask_t *cpus)
2486{
2487        struct rq *busiest = NULL, *rq;
2488        unsigned long max_load = 0;
2489        int i;
2490
2491        for_each_cpu_mask(i, group->cpumask) {
2492
2493                if (!cpu_isset(i, *cpus))
2494                        continue;
2495
2496                rq = cpu_rq(i);
2497
2498                if (rq->nr_running == 1 && rq->raw_weighted_load > imbalance)
2499                        continue;
2500
2501                if (rq->raw_weighted_load > max_load) {
2502                        max_load = rq->raw_weighted_load;
2503                        busiest = rq;
2504                }
2505        }
2506
2507        return busiest;
2508}
2509
2510/*
2511 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2512 * so long as it is large enough.
2513 */
2514#define MAX_PINNED_INTERVAL     512
2515
2516static inline unsigned long minus_1_or_zero(unsigned long n)
2517{
2518        return n > 0 ? n - 1 : 0;
2519}
2520
2521/*
2522 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2523 * tasks if there is an imbalance.
2524 *
2525 * Called with this_rq unlocked.
2526 */
2527static int load_balance(int this_cpu, struct rq *this_rq,
2528                        struct sched_domain *sd, enum idle_type idle)
2529{
2530        int nr_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2531        struct sched_group *group;
2532        unsigned long imbalance;
2533        struct rq *busiest;
2534        cpumask_t cpus = CPU_MASK_ALL;
2535
2536        if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2537            !sched_smt_power_savings)
2538                sd_idle = 1;
2539
2540        schedstat_inc(sd, lb_cnt[idle]);
2541
2542redo:
2543        group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2544                                                        &cpus);
2545        if (!group) {
2546                schedstat_inc(sd, lb_nobusyg[idle]);
2547                goto out_balanced;
2548        }
2549
2550        busiest = find_busiest_queue(group, idle, imbalance, &cpus);
2551        if (!busiest) {
2552                schedstat_inc(sd, lb_nobusyq[idle]);
2553                goto out_balanced;
2554        }
2555
2556        BUG_ON(busiest == this_rq);
2557
2558        schedstat_add(sd, lb_imbalance[idle], imbalance);
2559
2560        nr_moved = 0;
2561        if (busiest->nr_running > 1) {
2562                /*
2563                 * Attempt to move tasks. If find_busiest_group has found
2564                 * an imbalance but busiest->nr_running <= 1, the group is
2565                 * still unbalanced. nr_moved simply stays zero, so it is
2566                 * correctly treated as an imbalance.
2567                 */
2568                double_rq_lock(this_rq, busiest);
2569                nr_moved = move_tasks(this_rq, this_cpu, busiest,
2570                                      minus_1_or_zero(busiest->nr_running),
2571                                      imbalance, sd, idle, &all_pinned);
2572                double_rq_unlock(this_rq, busiest);
2573
2574                /* All tasks on this runqueue were pinned by CPU affinity */
2575                if (unlikely(all_pinned)) {
2576                        cpu_clear(cpu_of(busiest), cpus);
2577                        if (!cpus_empty(cpus))
2578                                goto redo;
2579                        goto out_balanced;
2580                }
2581        }
2582
2583        if (!nr_moved) {
2584                schedstat_inc(sd, lb_failed[idle]);
2585                sd->nr_balance_failed++;
2586
2587                if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2588
2589                        spin_lock(&busiest->lock);
2590
2591                        /* don't kick the migration_thread, if the curr
2592                         * task on busiest cpu can't be moved to this_cpu
2593                         */
2594                        if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
2595                                spin_unlock(&busiest->lock);
2596                                all_pinned = 1;
2597                                goto out_one_pinned;
2598                        }
2599
2600                        if (!busiest->active_balance) {
2601                                busiest->active_balance = 1;
2602                                busiest->push_cpu = this_cpu;
2603                                active_balance = 1;
2604                        }
2605                        spin_unlock(&busiest->lock);
2606                        if (active_balance)
2607                                wake_up_process(busiest->migration_thread);
2608
2609                        /*
2610                         * We've kicked active balancing, reset the failure
2611                         * counter.
2612                         */
2613                        sd->nr_balance_failed = sd->cache_nice_tries+1;
2614                }
2615        } else
2616                sd->nr_balance_failed = 0;
2617
2618        if (likely(!active_balance)) {
2619                /* We were unbalanced, so reset the balancing interval */
2620                sd->balance_interval = sd->min_interval;
2621        } else {
2622                /*
2623                 * If we've begun active balancing, start to back off. This
2624                 * case may not be covered by the all_pinned logic if there
2625                 * is only 1 task on the busy runqueue (because we don't call
2626                 * move_tasks).
2627                 */
2628                if (sd->balance_interval < sd->max_interval)
2629                        sd->balance_interval *= 2;
2630        }
2631
2632        if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2633            !sched_smt_power_savings)
2634                return -1;
2635        return nr_moved;
2636
2637out_balanced:
2638        schedstat_inc(sd, lb_balanced[idle]);
2639
2640        sd->nr_balance_failed = 0;
2641
2642out_one_pinned:
2643        /* tune up the balancing interval */
2644        if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2645                        (sd->balance_interval < sd->max_interval))
2646                sd->balance_interval *= 2;
2647
2648        if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2649                        !sched_smt_power_savings)
2650                return -1;
2651        return 0;
2652}
2653
2654/*
2655 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2656 * tasks if there is an imbalance.
2657 *
2658 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2659 * this_rq is locked.
2660 */
2661static int
2662load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
2663{
2664        struct sched_group *group;
2665        struct rq *busiest = NULL;
2666        unsigned long imbalance;
2667        int nr_moved = 0;
2668        int sd_idle = 0;
2669        cpumask_t cpus = CPU_MASK_ALL;
2670
2671        if (sd->flags & SD_SHARE_CPUPOWER && !sched_smt_power_savings)
2672                sd_idle = 1;
2673
2674        schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
2675redo:
2676        group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE,
2677                                &sd_idle, &cpus);
2678        if (!group) {
2679                schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
2680                goto out_balanced;
2681        }
2682
2683        busiest = find_busiest_queue(group, NEWLY_IDLE, imbalance,
2684                                &cpus);
2685        if (!busiest) {
2686                schedstat_inc(sd, lb_nobusyq[NEWLY_IDLE]);
2687                goto out_balanced;
2688        }
2689
2690        BUG_ON(busiest == this_rq);
2691
2692        schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
2693
2694        nr_moved = 0;
2695        if (busiest->nr_running > 1) {
2696                /* Attempt to move tasks */
2697                double_lock_balance(this_rq, busiest);
2698                nr_moved = move_tasks(this_rq, this_cpu, busiest,
2699                                        minus_1_or_zero(busiest->nr_running),
2700                                        imbalance, sd, NEWLY_IDLE, NULL);
2701                spin_unlock(&busiest->lock);
2702
2703                if (!nr_moved) {
2704                        cpu_clear(cpu_of(busiest), cpus);
2705                        if (!cpus_empty(cpus))
2706                                goto redo;
2707                }
2708        }
2709
2710        if (!nr_moved) {
2711                schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
2712                if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
2713                        return -1;
2714        } else
2715                sd->nr_balance_failed = 0;
2716
2717        return nr_moved;
2718
2719out_balanced:
2720        schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
2721        if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2722                                        !sched_smt_power_savings)
2723                return -1;
2724        sd->nr_balance_failed = 0;
2725
2726        return 0;
2727}
2728
2729/*
2730 * idle_balance is called by schedule() if this_cpu is about to become
2731 * idle. Attempts to pull tasks from other CPUs.
2732 */
2733static void idle_balance(int this_cpu, struct rq *this_rq)
2734{
2735        struct sched_domain *sd;
2736
2737        for_each_domain(this_cpu, sd) {
2738                if (sd->flags & SD_BALANCE_NEWIDLE) {
2739                        /* If we've pulled tasks over stop searching: */
2740                        if (load_balance_newidle(this_cpu, this_rq, sd))
2741                                break;
2742                }
2743        }
2744}
2745
2746/*
2747 * active_load_balance is run by migration threads. It pushes running tasks
2748 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2749 * running on each physical CPU where possible, and avoids physical /
2750 * logical imbalances.
2751 *
2752 * Called with busiest_rq locked.
2753 */
2754static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
2755{
2756        int target_cpu = busiest_rq->push_cpu;
2757        struct sched_domain *sd;
2758        struct rq *target_rq;
2759
2760        /* Is there any task to move? */
2761        if (busiest_rq->nr_running <= 1)
2762                return;
2763
2764        target_rq = cpu_rq(target_cpu);
2765
2766        /*
2767         * This condition is "impossible", if it occurs
2768         * we need to fix it.  Originally reported by
2769         * Bjorn Helgaas on a 128-cpu setup.
2770         */
2771        BUG_ON(busiest_rq == target_rq);
2772
2773        /* move a task from busiest_rq to target_rq */
2774        double_lock_balance(busiest_rq, target_rq);
2775
2776        /* Search for an sd spanning us and the target CPU. */
2777        for_each_domain(target_cpu, sd) {
2778                if ((sd->flags & SD_LOAD_BALANCE) &&
2779                    cpu_isset(busiest_cpu, sd->span))
2780                                break;
2781        }
2782
2783        if (likely(sd)) {
2784                schedstat_inc(sd, alb_cnt);
2785
2786                if (move_tasks(target_rq, target_cpu, busiest_rq, 1,
2787                               RTPRIO_TO_LOAD_WEIGHT(100), sd, SCHED_IDLE,
2788                               NULL))
2789                        schedstat_inc(sd, alb_pushed);
2790                else
2791                        schedstat_inc(sd, alb_failed);
2792        }
2793        spin_unlock(&target_rq->lock);
2794}
2795
2796/*
2797 * rebalance_tick will get called every timer tick, on every CPU.
2798 *
2799 * It checks each scheduling domain to see if it is due to be balanced,
2800 * and initiates a balancing operation if so.
2801 *
2802 * Balancing parameters are set up in arch_init_sched_domains.
2803 */
2804
2805/* Don't have all balancing operations going off at once: */
2806static inline unsigned long cpu_offset(int cpu)
2807{
2808        return jiffies + cpu * HZ / NR_CPUS;
2809}
2810
2811static void
2812rebalance_tick(int this_cpu, struct rq *this_rq, enum idle_type idle)
2813{
2814        unsigned long this_load, interval, j = cpu_offset(this_cpu);
2815        struct sched_domain *sd;
2816        int i, scale;
2817
2818        this_load = this_rq->raw_weighted_load;
2819
2820        /* Update our load: */
2821        for (i = 0, scale = 1; i < 3; i++, scale <<= 1) {
2822                unsigned long old_load, new_load;
2823
2824                old_load = this_rq->cpu_load[i];
2825                new_load = this_load;
2826                /*
2827                 * Round up the averaging division if load is increasing. This
2828                 * prevents us from getting stuck on 9 if the load is 10, for
2829                 * example.
2830                 */
2831                if (new_load > old_load)
2832                        new_load += scale-1;
2833                this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
2834        }
2835
2836        for_each_domain(this_cpu, sd) {
2837                if (!(sd->flags & SD_LOAD_BALANCE))
2838                        continue;
2839
2840                interval = sd->balance_interval;
2841                if (idle != SCHED_IDLE)
2842                        interval *= sd->busy_factor;
2843
2844                /* scale ms to jiffies */
2845                interval = msecs_to_jiffies(interval);
2846                if (unlikely(!interval))
2847                        interval = 1;
2848
2849                if (j - sd->last_balance >= interval) {
2850                        if (load_balance(this_cpu, this_rq, sd, idle)) {
2851                                /*
2852                                 * We've pulled tasks over so either we're no
2853                                 * longer idle, or one of our SMT siblings is
2854                                 * not idle.
2855                                 */
2856                                idle = NOT_IDLE;
2857                        }
2858                        sd->last_balance += interval;
2859                }
2860        }
2861}
2862#else
2863/*
2864 * on UP we do not need to balance between CPUs:
2865 */
2866static inline void rebalance_tick(int cpu, struct rq *rq, enum idle_type idle)
2867{
2868}
2869static inline void idle_balance(int cpu, struct rq *rq)
2870{
2871}
2872#endif
2873
2874static inline int wake_priority_sleeper(struct rq *rq)
2875{
2876        int ret = 0;
2877
2878#ifdef CONFIG_SCHED_SMT
2879        spin_lock(&rq->lock);
2880        /*
2881         * If an SMT sibling task has been put to sleep for priority
2882         * reasons reschedule the idle task to see if it can now run.
2883         */
2884        if (rq->nr_running) {
2885                resched_task(rq->idle);
2886                ret = 1;
2887        }
2888        spin_unlock(&rq->lock);
2889#endif
2890        return ret;
2891}
2892
2893DEFINE_PER_CPU(struct kernel_stat, kstat);
2894
2895EXPORT_PER_CPU_SYMBOL(kstat);
2896
2897/*
2898 * This is called on clock ticks and on context switches.
2899 * Bank in p->sched_time the ns elapsed since the last tick or switch.
2900 */
2901static inline void
2902update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
2903{
2904        p->sched_time += now - max(p->timestamp, rq->timestamp_last_tick);
2905}
2906
2907/*
2908 * Return current->sched_time plus any more ns on the sched_clock
2909 * that have not yet been banked.
2910 */
2911unsigned long long current_sched_time(const struct task_struct *p)
2912{
2913        unsigned long long ns;
2914        unsigned long flags;
2915
2916        local_irq_save(flags);
2917        ns = max(p->timestamp, task_rq(p)->timestamp_last_tick);
2918        ns = p->sched_time + sched_clock() - ns;
2919        local_irq_restore(flags);
2920
2921        return ns;
2922}
2923
2924/*
2925 * We place interactive tasks back into the active array, if possible.
2926 *
2927 * To guarantee that this does not starve expired tasks we ignore the
2928 * interactivity of a task if the first expired task had to wait more
2929 * than a 'reasonable' amount of time. This deadline timeout is
2930 * load-dependent, as the frequency of array switched decreases with
2931 * increasing number of running tasks. We also ignore the interactivity
2932 * if a better static_prio task has expired:
2933 */
2934static inline int expired_starving(struct rq *rq)
2935{
2936        if (rq->curr->static_prio > rq->best_expired_prio)
2937                return 1;
2938        if (!STARVATION_LIMIT || !rq->expired_timestamp)
2939                return 0;
2940        if (jiffies - rq->expired_timestamp > STARVATION_LIMIT * rq->nr_running)
2941                return 1;
2942        return 0;
2943}
2944
2945/*
2946 * Account user cpu time to a process.
2947 * @p: the process that the cpu time gets accounted to
2948 * @hardirq_offset: the offset to subtract from hardirq_count()
2949 * @cputime: the cpu time spent in user space since the last update
2950 */
2951void account_user_time(struct task_struct *p, cputime_t cputime)
2952{
2953        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2954        cputime64_t tmp;
2955
2956        p->utime = cputime_add(p->utime, cputime);
2957
2958        /* Add user time to cpustat. */
2959        tmp = cputime_to_cputime64(cputime);
2960        if (TASK_NICE(p) > 0)
2961                cpustat->nice = cputime64_add(cpustat->nice, tmp);
2962        else
2963                cpustat->user = cputime64_add(cpustat->user, tmp);
2964}
2965
2966/*
2967 * Account system cpu time to a process.
2968 * @p: the process that the cpu time gets accounted to
2969 * @hardirq_offset: the offset to subtract from hardirq_count()
2970 * @cputime: the cpu time spent in kernel space since the last update
2971 */
2972void account_system_time(struct task_struct *p, int hardirq_offset,
2973                         cputime_t cputime)
2974{
2975        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2976        struct rq *rq = this_rq();
2977        cputime64_t tmp;
2978
2979        p->stime = cputime_add(p->stime, cputime);
2980
2981        /* Add system time to cpustat. */
2982        tmp = cputime_to_cputime64(cputime);
2983        if (hardirq_count() - hardirq_offset)
2984                cpustat->irq = cputime64_add(cpustat->irq, tmp);
2985        else if (softirq_count())
2986                cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2987        else if (p != rq->idle)
2988                cpustat->system = cputime64_add(cpustat->system, tmp);
2989        else if (atomic_read(&rq->nr_iowait) > 0)
2990                cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
2991        else
2992                cpustat->idle = cputime64_add(cpustat->idle, tmp);
2993        /* Account for system time used */
2994        acct_update_integrals(p);
2995}
2996
2997/*
2998 * Account for involuntary wait time.
2999 * @p: the process from which the cpu time has been stolen
3000 * @steal: the cpu time spent in involuntary wait
3001 */
3002void account_steal_time(struct task_struct *p, cputime_t steal)
3003{
3004        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3005        cputime64_t tmp = cputime_to_cputime64(steal);
3006        struct rq *rq = this_rq();
3007
3008        if (p == rq->idle) {
3009                p->stime = cputime_add(p->stime, steal);
3010                if (atomic_read(&rq->nr_iowait) > 0)
3011                        cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3012                else
3013                        cpustat->idle = cputime64_add(cpustat->idle, tmp);
3014        } else
3015                cpustat->steal = cputime64_add(cpustat->steal, tmp);
3016}
3017
3018/*
3019 * This function gets called by the timer code, with HZ frequency.
3020 * We call it with interrupts disabled.
3021 *
3022 * It also gets called by the fork code, when changing the parent's
3023 * timeslices.
3024 */
3025void scheduler_tick(void)
3026{
3027        unsigned long long now = sched_clock();
3028        struct task_struct *p = current;
3029        int cpu = smp_processor_id();
3030        struct rq *rq = cpu_rq(cpu);
3031
3032        update_cpu_clock(p, rq, now);
3033
3034        rq->timestamp_last_tick = now;
3035
3036        if (p == rq->idle) {
3037                if (wake_priority_sleeper(rq))
3038                        goto out;
3039                rebalance_tick(cpu, rq, SCHED_IDLE);
3040                return;
3041        }
3042
3043        /* Task might have expired already, but not scheduled off yet */
3044        if (p->array != rq->active) {
3045                set_tsk_need_resched(p);
3046                goto out;
3047        }
3048        spin_lock(&rq->lock);
3049        /*
3050         * The task was running during this tick - update the
3051         * time slice counter. Note: we do not update a thread's
3052         * priority until it either goes to sleep or uses up its
3053         * timeslice. This makes it possible for interactive tasks
3054         * to use up their timeslices at their highest priority levels.
3055         */
3056        if (rt_task(p)) {
3057                /*
3058                 * RR tasks need a special form of timeslice management.
3059                 * FIFO tasks have no timeslices.
3060                 */
3061                if ((p->policy == SCHED_RR) && !--p->time_slice) {
3062                        p->time_slice = task_timeslice(p);
3063                        p->first_time_slice = 0;
3064                        set_tsk_need_resched(p);
3065
3066                        /* put it at the end of the queue: */
3067                        requeue_task(p, rq->active);
3068                }
3069                goto out_unlock;
3070        }
3071        if (!--p->time_slice) {
3072                dequeue_task(p, rq->active);
3073                set_tsk_need_resched(p);
3074                p->prio = effective_prio(p);
3075                p->time_slice = task_timeslice(p);
3076                p->first_time_slice = 0;
3077
3078                if (!rq->expired_timestamp)
3079                        rq->expired_timestamp = jiffies;
3080                if (!TASK_INTERACTIVE(p) || expired_starving(rq)) {
3081                        enqueue_task(p, rq->expired);
3082                        if (p->static_prio < rq->best_expired_prio)
3083                                rq->best_expired_prio = p->static_prio;
3084                } else
3085                        enqueue_task(p, rq->active);
3086        } else {
3087                /*
3088                 * Prevent a too long timeslice allowing a task to monopolize
3089                 * the CPU. We do this by splitting up the timeslice into
3090                 * smaller pieces.
3091                 *
3092                 * Note: this does not mean the task's timeslices expire or
3093                 * get lost in any way, they just might be preempted by
3094                 * another task of equal priority. (one with higher
3095                 * priority would have preempted this task already.) We
3096                 * requeue this task to the end of the list on this priority
3097                 * level, which is in essence a round-robin of tasks with
3098                 * equal priority.
3099                 *
3100                 * This only applies to tasks in the interactive
3101                 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3102                 */
3103                if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
3104                        p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
3105                        (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
3106                        (p->array == rq->active)) {
3107
3108                        requeue_task(p, rq->active);
3109                        set_tsk_need_resched(p);
3110                }
3111        }
3112out_unlock:
3113        spin_unlock(&rq->lock);
3114out:
3115        rebalance_tick(cpu, rq, NOT_IDLE);
3116}
3117
3118#ifdef CONFIG_SCHED_SMT
3119static inline void wakeup_busy_runqueue(struct rq *rq)
3120{
3121        /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3122        if (rq->curr == rq->idle && rq->nr_running)
3123                resched_task(rq->idle);
3124}
3125
3126/*
3127 * Called with interrupt disabled and this_rq's runqueue locked.
3128 */
3129static void wake_sleeping_dependent(int this_cpu)
3130{
3131        struct sched_domain *tmp, *sd = NULL;
3132        int i;
3133
3134        for_each_domain(this_cpu, tmp) {
3135                if (tmp->flags & SD_SHARE_CPUPOWER) {
3136                        sd = tmp;
3137                        break;
3138                }
3139        }
3140
3141        if (!sd)
3142                return;
3143
3144        for_each_cpu_mask(i, sd->span) {
3145                struct rq *smt_rq = cpu_rq(i);
3146
3147                if (i == this_cpu)
3148                        continue;
3149                if (unlikely(!spin_trylock(&smt_rq->lock)))
3150                        continue;
3151
3152                wakeup_busy_runqueue(smt_rq);
3153                spin_unlock(&smt_rq->lock);
3154        }
3155}
3156
3157/*
3158 * number of 'lost' timeslices this task wont be able to fully
3159 * utilize, if another task runs on a sibling. This models the
3160 * slowdown effect of other tasks running on siblings:
3161 */
3162static inline unsigned long
3163smt_slice(struct task_struct *p, struct sched_domain *sd)
3164{
3165        return p->time_slice * (100 - sd->per_cpu_gain) / 100;
3166}
3167
3168/*
3169 * To minimise lock contention and not have to drop this_rq's runlock we only
3170 * trylock the sibling runqueues and bypass those runqueues if we fail to
3171 * acquire their lock. As we only trylock the normal locking order does not
3172 * need to be obeyed.
3173 */
3174static int
3175dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3176{
3177        struct sched_domain *tmp, *sd = NULL;
3178        int ret = 0, i;
3179
3180        /* kernel/rt threads do not participate in dependent sleeping */
3181        if (!p->mm || rt_task(p))
3182                return 0;
3183
3184        for_each_domain(this_cpu, tmp) {
3185                if (tmp->flags & SD_SHARE_CPUPOWER) {
3186                        sd = tmp;
3187                        break;
3188                }
3189        }
3190
3191        if (!sd)
3192                return 0;
3193
3194        for_each_cpu_mask(i, sd->span) {
3195                struct task_struct *smt_curr;
3196                struct rq *smt_rq;
3197
3198                if (i == this_cpu)
3199                        continue;
3200
3201                smt_rq = cpu_rq(i);
3202                if (unlikely(!spin_trylock(&smt_rq->lock)))
3203                        continue;
3204
3205                smt_curr = smt_rq->curr;
3206
3207                if (!smt_curr->mm)
3208                        goto unlock;
3209
3210                /*
3211                 * If a user task with lower static priority than the
3212                 * running task on the SMT sibling is trying to schedule,
3213                 * delay it till there is proportionately less timeslice
3214                 * left of the sibling task to prevent a lower priority
3215                 * task from using an unfair proportion of the
3216                 * physical cpu's resources. -ck
3217                 */
3218                if (rt_task(smt_curr)) {
3219                        /*
3220                         * With real time tasks we run non-rt tasks only
3221                         * per_cpu_gain% of the time.
3222                         */
3223                        if ((jiffies % DEF_TIMESLICE) >
3224                                (sd->per_cpu_gain * DEF_TIMESLICE / 100))
3225                                        ret = 1;
3226                } else {
3227                        if (smt_curr->static_prio < p->static_prio &&
3228                                !TASK_PREEMPTS_CURR(p, smt_rq) &&
3229                                smt_slice(smt_curr, sd) > task_timeslice(p))
3230                                        ret = 1;
3231                }
3232unlock:
3233                spin_unlock(&smt_rq->lock);
3234        }
3235        return ret;
3236}
3237#else
3238static inline void wake_sleeping_dependent(int this_cpu)
3239{
3240}
3241static inline int
3242dependent_sleeper(int this_cpu, struct rq *this_rq, struct task_struct *p)
3243{
3244        return 0;
3245}
3246#endif
3247
3248#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3249
3250void fastcall add_preempt_count(int val)
3251{
3252        /*
3253         * Underflow?
3254         */
3255        if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3256                return;
3257        preempt_count() += val;
3258        /*
3259         * Spinlock count overflowing soon?
3260         */
3261        DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= PREEMPT_MASK-10);
3262}
3263EXPORT_SYMBOL(add_preempt_count);
3264
3265void fastcall sub_preempt_count(int val)
3266{
3267        /*
3268         * Underflow?
3269         */
3270        if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3271                return;
3272        /*
3273         * Is the spinlock portion underflowing?
3274         */
3275        if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3276                        !(preempt_count() & PREEMPT_MASK)))
3277                return;
3278
3279        preempt_count() -= val;
3280}
3281EXPORT_SYMBOL(sub_preempt_count);
3282
3283#endif
3284
3285static inline int interactive_sleep(enum sleep_type sleep_type)
3286{
3287        return (sleep_type == SLEEP_INTERACTIVE ||
3288                sleep_type == SLEEP_INTERRUPTED);
3289}
3290
3291/*
3292 * schedule() is the main scheduler function.
3293 */
3294asmlinkage void __sched schedule(void)
3295{
3296        struct task_struct *prev, *next;
3297        struct prio_array *array;
3298        struct list_head *queue;
3299        unsigned long long now;
3300        unsigned long run_time;
3301        int cpu, idx, new_prio;
3302        long *switch_count;
3303        struct rq *rq;
3304
3305        /*
3306         * Test if we are atomic.  Since do_exit() needs to call into
3307         * schedule() atomically, we ignore that path for now.
3308         * Otherwise, whine if we are scheduling when we should not be.
3309         */
3310        if (unlikely(in_atomic() && !current->exit_state)) {
3311                printk(KERN_ERR "BUG: scheduling while atomic: "
3312                        "%s/0x%08x/%d\n",
3313                        current->comm, preempt_count(), current->pid);
3314                dump_stack();
3315        }
3316        profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3317
3318need_resched:
3319        preempt_disable();
3320        prev = current;
3321        release_kernel_lock(prev);
3322need_resched_nonpreemptible:
3323        rq = this_rq();
3324
3325        /*
3326         * The idle thread is not allowed to schedule!
3327         * Remove this check after it has been exercised a bit.
3328         */
3329        if (unlikely(prev == rq->idle) && prev->state != TASK_RUNNING) {
3330                printk(KERN_ERR "bad: scheduling from the idle thread!\n");
3331                dump_stack();
3332        }
3333
3334        schedstat_inc(rq, sched_cnt);
3335        now = sched_clock();
3336        if (likely((long long)(now - prev->timestamp) < NS_MAX_SLEEP_AVG)) {
3337                run_time = now - prev->timestamp;
3338                if (unlikely((long long)(now - prev->timestamp) < 0))
3339                        run_time = 0;
3340        } else
3341                run_time = NS_MAX_SLEEP_AVG;
3342
3343        /*
3344         * Tasks charged proportionately less run_time at high sleep_avg to
3345         * delay them losing their interactive status
3346         */
3347        run_time /= (CURRENT_BONUS(prev) ? : 1);
3348
3349        spin_lock_irq(&rq->lock);
3350
3351        if (unlikely(prev->flags & PF_DEAD))
3352                prev->state = EXIT_DEAD;
3353
3354        switch_count = &prev->nivcsw;
3355        if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3356                switch_count = &prev->nvcsw;
3357                if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
3358                                unlikely(signal_pending(prev))))
3359                        prev->state = TASK_RUNNING;
3360                else {
3361                        if (prev->state == TASK_UNINTERRUPTIBLE)
3362                                rq->nr_uninterruptible++;
3363                        deactivate_task(prev, rq);
3364                }
3365        }
3366
3367        cpu = smp_processor_id();
3368        if (unlikely(!rq->nr_running)) {
3369                idle_balance(cpu, rq);
3370                if (!rq->nr_running) {
3371                        next = rq->idle;
3372                        rq->expired_timestamp = 0;
3373                        wake_sleeping_dependent(cpu);
3374                        goto switch_tasks;
3375                }
3376        }
3377
3378        array = rq->active;
3379        if (unlikely(!array->nr_active)) {
3380                /*
3381                 * Switch the active and expired arrays.
3382                 */
3383                schedstat_inc(rq, sched_switch);
3384                rq->active = rq->expired;
3385                rq->expired = array;
3386                array = rq->active;
3387                rq->expired_timestamp = 0;
3388                rq->best_expired_prio = MAX_PRIO;
3389        }
3390
3391        idx = sched_find_first_bit(array->bitmap);
3392        queue = array->queue + idx;
3393        next = list_entry(queue->next, struct task_struct, run_list);
3394
3395        if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
3396                unsigned long long delta = now - next->timestamp;
3397                if (unlikely((long long)(now - next->timestamp) < 0))
3398                        delta = 0;
3399
3400                if (next->sleep_type == SLEEP_INTERACTIVE)
3401                        delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
3402
3403                array = next->array;
3404                new_prio = recalc_task_prio(next, next->timestamp + delta);
3405
3406                if (unlikely(next->prio != new_prio)) {
3407                        dequeue_task(next, array);
3408                        next->prio = new_prio;
3409                        enqueue_task(next, array);
3410                }
3411        }
3412        next->sleep_type = SLEEP_NORMAL;
3413        if (dependent_sleeper(cpu, rq, next))
3414                next = rq->idle;
3415switch_tasks:
3416        if (next == rq->idle)
3417                schedstat_inc(rq, sched_goidle);
3418        prefetch(next);
3419        prefetch_stack(next);
3420        clear_tsk_need_resched(prev);
3421        rcu_qsctr_inc(task_cpu(prev));
3422
3423        update_cpu_clock(prev, rq, now);
3424
3425        prev->sleep_avg -= run_time;
3426        if ((long)prev->sleep_avg <= 0)
3427                prev->sleep_avg = 0;
3428        prev->timestamp = prev->last_ran = now;
3429
3430        sched_info_switch(prev, next);
3431        if (likely(prev != next)) {
3432                next->timestamp = now;
3433                rq->nr_switches++;
3434                rq->curr = next;
3435                ++*switch_count;
3436
3437                prepare_task_switch(rq, next);
3438                prev = context_switch(rq, prev, next);
3439                barrier();
3440                /*
3441                 * this_rq must be evaluated again because prev may have moved
3442                 * CPUs since it called schedule(), thus the 'rq' on its stack
3443                 * frame will be invalid.
3444                 */
3445                finish_task_switch(this_rq(), prev);
3446        } else
3447                spin_unlock_irq(&rq->lock);
3448
3449        prev = current;
3450        if (unlikely(reacquire_kernel_lock(prev) < 0))
3451                goto need_resched_nonpreemptible;
3452        preempt_enable_no_resched();
3453        if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3454                goto need_resched;
3455}
3456EXPORT_SYMBOL(schedule);
3457
3458#ifdef CONFIG_PREEMPT
3459/*
3460 * this is the entry point to schedule() from in-kernel preemption
3461 * off of preempt_enable.  Kernel preemptions off return from interrupt
3462 * occur there and call schedule directly.
3463 */
3464asmlinkage void __sched preempt_schedule(void)
3465{
3466        struct thread_info *ti = current_thread_info();
3467#ifdef CONFIG_PREEMPT_BKL
3468        struct task_struct *task = current;
3469        int saved_lock_depth;
3470#endif
3471        /*
3472         * If there is a non-zero preempt_count or interrupts are disabled,
3473         * we do not want to preempt the current task.  Just return..
3474         */
3475        if (unlikely(ti->preempt_count || irqs_disabled()))
3476                return;
3477
3478need_resched:
3479        add_preempt_count(PREEMPT_ACTIVE);
3480        /*
3481         * We keep the big kernel semaphore locked, but we
3482         * clear ->lock_depth so that schedule() doesnt
3483         * auto-release the semaphore:
3484         */
3485#ifdef CONFIG_PREEMPT_BKL
3486        saved_lock_depth = task->lock_depth;
3487        task->lock_depth = -1;
3488#endif
3489        schedule();
3490#ifdef CONFIG_PREEMPT_BKL
3491        task->lock_depth = saved_lock_depth;
3492#endif
3493        sub_preempt_count(PREEMPT_ACTIVE);
3494
3495        /* we could miss a preemption opportunity between schedule and now */
3496        barrier();
3497        if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3498                goto need_resched;
3499}
3500EXPORT_SYMBOL(preempt_schedule);
3501
3502/*
3503 * this is the entry point to schedule() from kernel preemption
3504 * off of irq context.
3505 * Note, that this is called and return with irqs disabled. This will
3506 * protect us against recursive calling from irq.
3507 */
3508asmlinkage void __sched preempt_schedule_irq(void)
3509{
3510        struct thread_info *ti = current_thread_info();
3511#ifdef CONFIG_PREEMPT_BKL
3512        struct task_struct *task = current;
3513        int saved_lock_depth;
3514#endif
3515        /* Catch callers which need to be fixed */
3516        BUG_ON(ti->preempt_count || !irqs_disabled());
3517
3518need_resched:
3519        add_preempt_count(PREEMPT_ACTIVE);
3520        /*
3521         * We keep the big kernel semaphore locked, but we
3522         * clear ->lock_depth so that schedule() doesnt
3523         * auto-release the semaphore:
3524         */
3525#ifdef CONFIG_PREEMPT_BKL
3526        saved_lock_depth = task->lock_depth;
3527        task->lock_depth = -1;
3528#endif
3529        local_irq_enable();
3530        schedule();
3531        local_irq_disable();
3532#ifdef CONFIG_PREEMPT_BKL
3533        task->lock_depth = saved_lock_depth;
3534#endif
3535        sub_preempt_count(PREEMPT_ACTIVE);
3536
3537        /* we could miss a preemption opportunity between schedule and now */
3538        barrier();
3539        if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3540                goto need_resched;
3541}
3542
3543#endif /* CONFIG_PREEMPT */
3544
3545int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3546                          void *key)
3547{
3548        return try_to_wake_up(curr->private, mode, sync);
3549}
3550EXPORT_SYMBOL(default_wake_function);
3551
3552/*
3553 * The core wakeup function.  Non-exclusive wakeups (nr_exclusive == 0) just
3554 * wake everything up.  If it's an exclusive wakeup (nr_exclusive == small +ve
3555 * number) then we wake all the non-exclusive tasks and one exclusive task.
3556 *
3557 * There are circumstances in which we can try to wake a task which has already
3558 * started to run but is not in state TASK_RUNNING.  try_to_wake_up() returns
3559 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3560 */
3561static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3562                             int nr_exclusive, int sync, void *key)
3563{
3564        struct list_head *tmp, *next;
3565
3566        list_for_each_safe(tmp, next, &q->task_list) {
3567                wait_queue_t *curr = list_entry(tmp, wait_queue_t, task_list);
3568                unsigned flags = curr->flags;
3569
3570                if (curr->func(curr, mode, sync, key) &&
3571                                (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3572                        break;
3573        }
3574}
3575
3576/**
3577 * __wake_up - wake up threads blocked on a waitqueue.
3578 * @q: the waitqueue
3579 * @mode: which threads
3580 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3581 * @key: is directly passed to the wakeup function
3582 */
3583void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
3584                        int nr_exclusive, void *key)
3585{
3586        unsigned long flags;
3587
3588        spin_lock_irqsave(&q->lock, flags);
3589        __wake_up_common(q, mode, nr_exclusive, 0, key);
3590        spin_unlock_irqrestore(&q->lock, flags);
3591}
3592EXPORT_SYMBOL(__wake_up);
3593
3594/*
3595 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3596 */
3597void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3598{
3599        __wake_up_common(q, mode, 1, 0, NULL);
3600}
3601
3602/**
3603 * __wake_up_sync - wake up threads blocked on a waitqueue.
3604 * @q: the waitqueue
3605 * @mode: which threads
3606 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3607 *
3608 * The sync wakeup differs that the waker knows that it will schedule
3609 * away soon, so while the target thread will be woken up, it will not
3610 * be migrated to another CPU - ie. the two threads are 'synchronized'
3611 * with each other. This can prevent needless bouncing between CPUs.
3612 *
3613 * On UP it can prevent extra preemption.
3614 */
3615void fastcall
3616__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3617{
3618        unsigned long flags;
3619        int sync = 1;
3620
3621        if (unlikely(!q))
3622                return;
3623
3624        if (unlikely(!nr_exclusive))
3625                sync = 0;
3626
3627        spin_lock_irqsave(&q->lock, flags);
3628        __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3629        spin_unlock_irqrestore(&q->lock, flags);
3630}
3631EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
3632
3633void fastcall complete(struct completion *x)
3634{
3635        unsigned long flags;
3636
3637        spin_lock_irqsave(&x->wait.lock, flags);
3638        x->done++;
3639        __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3640                         1, 0, NULL);
3641        spin_unlock_irqrestore(&x->wait.lock, flags);
3642}
3643EXPORT_SYMBOL(complete);
3644
3645void fastcall complete_all(struct completion *x)
3646{
3647        unsigned long flags;
3648
3649        spin_lock_irqsave(&x->wait.lock, flags);
3650        x->done += UINT_MAX/2;
3651        __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3652                         0, 0, NULL);
3653        spin_unlock_irqrestore(&x->wait.lock, flags);
3654}
3655EXPORT_SYMBOL(complete_all);
3656
3657void fastcall __sched wait_for_completion(struct completion *x)
3658{
3659        might_sleep();
3660
3661        spin_lock_irq(&x->wait.lock);
3662        if (!x->done) {
3663                DECLARE_WAITQUEUE(wait, current);
3664
3665                wait.flags |= WQ_FLAG_EXCLUSIVE;
3666                __add_wait_queue_tail(&x->wait, &wait);
3667                do {
3668                        __set_current_state(TASK_UNINTERRUPTIBLE);
3669                        spin_unlock_irq(&x->wait.lock);
3670                        schedule();
3671                        spin_lock_irq(&x->wait.lock);
3672                } while (!x->done);
3673                __remove_wait_queue(&x->wait, &wait);
3674        }
3675        x->done--;
3676        spin_unlock_irq(&x->wait.lock);
3677}
3678EXPORT_SYMBOL(wait_for_completion);
3679
3680unsigned long fastcall __sched
3681wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3682{
3683        might_sleep();
3684
3685        spin_lock_irq(&x->wait.lock);
3686        if (!x->done) {
3687                DECLARE_WAITQUEUE(wait, current);
3688
3689                wait.flags |= WQ_FLAG_EXCLUSIVE;
3690                __add_wait_queue_tail(&x->wait, &wait);
3691                do {
3692                        __set_current_state(TASK_UNINTERRUPTIBLE);
3693                        spin_unlock_irq(&x->wait.lock);
3694                        timeout = schedule_timeout(timeout);
3695                        spin_lock_irq(&x->wait.lock);
3696                        if (!timeout) {
3697                                __remove_wait_queue(&x->wait, &wait);
3698                                goto out;
3699                        }
3700                } while (!x->done);
3701                __remove_wait_queue(&x->wait, &wait);
3702        }
3703        x->done--;
3704out:
3705        spin_unlock_irq(&x->wait.lock);
3706        return timeout;
3707}
3708EXPORT_SYMBOL(wait_for_completion_timeout);
3709
3710int fastcall __sched wait_for_completion_interruptible(struct completion *x)
3711{
3712        int ret = 0;
3713
3714        might_sleep();
3715
3716        spin_lock_irq(&x->wait.lock);
3717        if (!x->done) {
3718                DECLARE_WAITQUEUE(wait, current);
3719
3720                wait.flags |= WQ_FLAG_EXCLUSIVE;
3721                __add_wait_queue_tail(&x->wait, &wait);
3722                do {
3723                        if (signal_pending(current)) {
3724                                ret = -ERESTARTSYS;
3725                                __remove_wait_queue(&x->wait, &wait);
3726                                goto out;
3727                        }
3728                        __set_current_state(TASK_INTERRUPTIBLE);
3729                        spin_unlock_irq(&x->wait.lock);
3730                        schedule();
3731                        spin_lock_irq(&x->wait.lock);
3732                } while (!x->done);
3733                __remove_wait_queue(&x->wait, &wait);
3734        }
3735        x->done--;
3736out:
3737        spin_unlock_irq(&x->wait.lock);
3738
3739        return ret;
3740}
3741EXPORT_SYMBOL(wait_for_completion_interruptible);
3742
3743unsigned long fastcall __sched
3744wait_for_completion_interruptible_timeout(struct completion *x,
3745                                          unsigned long timeout)
3746{
3747        might_sleep();
3748
3749        spin_lock_irq(&x->wait.lock);
3750        if (!x->done) {
3751                DECLARE_WAITQUEUE(wait, current);
3752
3753                wait.flags |= WQ_FLAG_EXCLUSIVE;
3754                __add_wait_queue_tail(&x->wait, &wait);
3755                do {
3756                        if (signal_pending(current)) {
3757                                timeout = -ERESTARTSYS;
3758                                __remove_wait_queue(&x->wait, &wait);
3759                                goto out;
3760                        }
3761                        __set_current_state(TASK_INTERRUPTIBLE);
3762                        spin_unlock_irq(&x->wait.lock);
3763                        timeout = schedule_timeout(timeout);
3764                        spin_lock_irq(&x->wait.lock);
3765                        if (!timeout) {
3766                                __remove_wait_queue(&x->wait, &wait);
3767                                goto out;
3768                        }
3769                } while (!x->done);
3770                __remove_wait_queue(&x->wait, &wait);
3771        }
3772        x->done--;
3773out:
3774        spin_unlock_irq(&x->wait.lock);
3775        return timeout;
3776}
3777EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3778
3779
3780#define SLEEP_ON_VAR                                    \
3781        unsigned long flags;                            \
3782        wait_queue_t wait;                              \
3783        init_waitqueue_entry(&wait, current);
3784
3785#define SLEEP_ON_HEAD                                   \
3786        spin_lock_irqsave(&q->lock,flags);              \
3787        __add_wait_queue(q, &wait);                     \
3788        spin_unlock(&q->lock);
3789
3790#define SLEEP_ON_TAIL                                   \
3791        spin_lock_irq(&q->lock);                        \
3792        __remove_wait_queue(q, &wait);                  \
3793        spin_unlock_irqrestore(&q->lock, flags);
3794
3795void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
3796{
3797        SLEEP_ON_VAR
3798
3799        current->state = TASK_INTERRUPTIBLE;
3800
3801        SLEEP_ON_HEAD
3802        schedule();
3803        SLEEP_ON_TAIL
3804}
3805EXPORT_SYMBOL(interruptible_sleep_on);
3806
3807long fastcall __sched
3808interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3809{
3810        SLEEP_ON_VAR
3811
3812        current->state = TASK_INTERRUPTIBLE;
3813
3814        SLEEP_ON_HEAD
3815        timeout = schedule_timeout(timeout);
3816        SLEEP_ON_TAIL
3817
3818        return timeout;
3819}
3820EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3821
3822void fastcall __sched sleep_on(wait_queue_head_t *q)
3823{
3824        SLEEP_ON_VAR
3825
3826        current->state = TASK_UNINTERRUPTIBLE;
3827
3828        SLEEP_ON_HEAD
3829        schedule();
3830        SLEEP_ON_TAIL
3831}
3832EXPORT_SYMBOL(sleep_on);
3833
3834long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3835{
3836        SLEEP_ON_VAR
3837
3838        current->state = TASK_UNINTERRUPTIBLE;
3839
3840        SLEEP_ON_HEAD
3841        timeout = schedule_timeout(timeout);
3842        SLEEP_ON_TAIL
3843
3844        return timeout;
3845}
3846
3847EXPORT_SYMBOL(sleep_on_timeout);
3848
3849#ifdef CONFIG_RT_MUTEXES
3850
3851/*
3852 * rt_mutex_setprio - set the current priority of a task
3853 * @p: task
3854 * @prio: prio value (kernel-internal form)
3855 *
3856 * This function changes the 'effective' priority of a task. It does
3857 * not touch ->normal_prio like __setscheduler().
3858 *
3859 * Used by the rt_mutex code to implement priority inheritance logic.
3860 */
3861void rt_mutex_setprio(struct task_struct *p, int prio)
3862{
3863        struct prio_array *array;
3864        unsigned long flags;
3865        struct rq *rq;
3866        int oldprio;
3867
3868        BUG_ON(prio < 0 || prio > MAX_PRIO);
3869
3870        rq = task_rq_lock(p, &flags);
3871
3872        oldprio = p->prio;
3873        array = p->array;
3874        if (array)
3875                dequeue_task(p, array);
3876        p->prio = prio;
3877
3878        if (array) {
3879                /*
3880                 * If changing to an RT priority then queue it
3881                 * in the active array!
3882                 */
3883                if (rt_task(p))
3884                        array = rq->active;
3885                enqueue_task(p, array);
3886                /*
3887                 * Reschedule if we are currently running on this runqueue and
3888                 * our priority decreased, or if we are not currently running on
3889                 * this runqueue and our priority is higher than the current's
3890                 */
3891                if (task_running(rq, p)) {
3892                        if (p->prio > oldprio)
3893                                resched_task(rq->curr);
3894                } else if (TASK_PREEMPTS_CURR(p, rq))
3895                        resched_task(rq->curr);
3896        }
3897        task_rq_unlock(rq, &flags);
3898}
3899
3900#endif
3901
3902void set_user_nice(struct task_struct *p, long nice)
3903{
3904        struct prio_array *array;
3905        int old_prio, delta;
3906        unsigned long flags;
3907        struct rq *rq;
3908
3909        if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3910                return;
3911        /*
3912         * We have to be careful, if called from sys_setpriority(),
3913         * the task might be in the middle of scheduling on another CPU.
3914         */
3915        rq = task_rq_lock(p, &flags);
3916        /*
3917         * The RT priorities are set via sched_setscheduler(), but we still
3918         * allow the 'normal' nice value to be set - but as expected
3919         * it wont have any effect on scheduling until the task is
3920         * not SCHED_NORMAL/SCHED_BATCH:
3921         */
3922        if (has_rt_policy(p)) {
3923                p->static_prio = NICE_TO_PRIO(nice);
3924                goto out_unlock;
3925        }
3926        array = p->array;
3927        if (array) {
3928                dequeue_task(p, array);
3929                dec_raw_weighted_load(rq, p);
3930        }
3931
3932        p->static_prio = NICE_TO_PRIO(nice);
3933        set_load_weight(p);
3934        old_prio = p->prio;
3935        p->prio = effective_prio(p);
3936        delta = p->prio - old_prio;
3937
3938        if (array) {
3939                enqueue_task(p, array);
3940                inc_raw_weighted_load(rq, p);
3941                /*
3942                 * If the task increased its priority or is running and
3943                 * lowered its priority, then reschedule its CPU:
3944                 */
3945                if (delta < 0 || (delta > 0 && task_running(rq, p)))
3946                        resched_task(rq->curr);
3947        }
3948out_unlock:
3949        task_rq_unlock(rq, &flags);
3950}
3951EXPORT_SYMBOL(set_user_nice);
3952
3953/*
3954 * can_nice - check if a task can reduce its nice value
3955 * @p: task
3956 * @nice: nice value
3957 */
3958int can_nice(const struct task_struct *p, const int nice)
3959{
3960        /* convert nice value [19,-20] to rlimit style value [1,40] */
3961        int nice_rlim = 20 - nice;
3962
3963        return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
3964                capable(CAP_SYS_NICE));
3965}
3966
3967#ifdef __ARCH_WANT_SYS_NICE
3968
3969/*
3970 * sys_nice - change the priority of the current process.
3971 * @increment: priority increment
3972 *
3973 * sys_setpriority is a more generic, but much slower function that
3974 * does similar things.
3975 */
3976asmlinkage long sys_nice(int increment)
3977{
3978        long nice, retval;
3979
3980        /*
3981         * Setpriority might change our priority at the same moment.
3982         * We don't have to worry. Conceptually one call occurs first
3983         * and we have a single winner.
3984         */
3985        if (increment < -40)
3986                increment = -40;
3987        if (increment > 40)
3988                increment = 40;
3989
3990        nice = PRIO_TO_NICE(current->static_prio) + increment;
3991        if (nice < -20)
3992                nice = -20;
3993        if (nice > 19)
3994                nice = 19;
3995
3996        if (increment < 0 && !can_nice(current, nice))
3997                return -EPERM;
3998
3999        retval = security_task_setnice(current, nice);
4000        if (retval)
4001                return retval;
4002
4003        set_user_nice(current, nice);
4004        return 0;
4005}
4006
4007#endif
4008
4009/**
4010 * task_prio - return the priority value of a given task.
4011 * @p: the task in question.
4012 *
4013 * This is the priority value as seen by users in /proc.
4014 * RT tasks are offset by -200. Normal tasks are centered
4015 * around 0, value goes from -16 to +15.
4016 */
4017int task_prio(const struct task_struct *p)
4018{
4019        return p->prio - MAX_RT_PRIO;
4020}
4021
4022/**
4023 * task_nice - return the nice value of a given task.
4024 * @p: the task in question.
4025 */
4026int task_nice(const struct task_struct *p)
4027{
4028        return TASK_NICE(p);
4029}
4030EXPORT_SYMBOL_GPL(task_nice);
4031
4032/**
4033 * idle_cpu - is a given cpu idle currently?
4034 * @cpu: the processor in question.
4035 */
4036int idle_cpu(int cpu)
4037{
4038        return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4039}
4040
4041/**
4042 * idle_task - return the idle task for a given cpu.
4043 * @cpu: the processor in question.
4044 */
4045struct task_struct *idle_task(int cpu)
4046{
4047        return cpu_rq(cpu)->idle;
4048}
4049
4050/**
4051 * find_process_by_pid - find a process with a matching PID value.
4052 * @pid: the pid in question.
4053 */
4054static inline struct task_struct *find_process_by_pid(pid_t pid)
4055{
4056        return pid ? find_task_by_pid(pid) : current;
4057}
4058
4059/* Actually do priority change: must hold rq lock. */
4060static void __setscheduler(struct task_struct *p, int policy, int prio)
4061{
4062        BUG_ON(p->array);
4063
4064        p->policy = policy;
4065        p->rt_priority = prio;
4066        p->normal_prio = normal_prio(p);
4067        /* we are holding p->pi_lock already */
4068        p->prio = rt_mutex_getprio(p);
4069        /*
4070         * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4071         */
4072        if (policy == SCHED_BATCH)
4073                p->sleep_avg = 0;
4074        set_load_weight(p);
4075}
4076
4077/**
4078 * sched_setscheduler - change the scheduling policy and/or RT priority of
4079 * a thread.
4080 * @p: the task in question.
4081 * @policy: new policy.
4082 * @param: structure containing the new RT priority.
4083 */
4084int sched_setscheduler(struct task_struct *p, int policy,
4085                       struct sched_param *param)
4086{
4087        int retval, oldprio, oldpolicy = -1;
4088        struct prio_array *array;
4089        unsigned long flags;
4090        struct rq *rq;
4091
4092        /* may grab non-irq protected spin_locks */
4093        BUG_ON(in_interrupt());
4094recheck:
4095        /* double check policy once rq lock held */
4096        if (policy < 0)
4097                policy = oldpolicy = p->policy;
4098        else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4099                        policy != SCHED_NORMAL && policy != SCHED_BATCH)
4100                return -EINVAL;
4101        /*
4102         * Valid priorities for SCHED_FIFO and SCHED_RR are
4103         * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4104         * SCHED_BATCH is 0.
4105         */
4106        if (param->sched_priority < 0 ||
4107            (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4108            (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4109                return -EINVAL;
4110        if ((policy == SCHED_NORMAL || policy == SCHED_BATCH)
4111                                        != (param->sched_priority == 0))
4112                return -EINVAL;
4113
4114        /*
4115         * Allow unprivileged RT tasks to decrease priority:
4116         */
4117        if (!capable(CAP_SYS_NICE)) {
4118                /*
4119                 * can't change policy, except between SCHED_NORMAL
4120                 * and SCHED_BATCH:
4121                 */
4122                if (((policy != SCHED_NORMAL && p->policy != SCHED_BATCH) &&
4123                        (policy != SCHED_BATCH && p->policy != SCHED_NORMAL)) &&
4124                                !p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
4125                        return -EPERM;
4126                /* can't increase priority */
4127                if ((policy != SCHED_NORMAL && policy != SCHED_BATCH) &&
4128                    param->sched_priority > p->rt_priority &&
4129                    param->sched_priority >
4130                                p->signal->rlim[RLIMIT_RTPRIO].rlim_cur)
4131                        return -EPERM;
4132                /* can't change other user's priorities */
4133                if ((current->euid != p->euid) &&
4134                    (current->euid != p->uid))
4135                        return -EPERM;
4136        }
4137
4138        retval = security_task_setscheduler(p, policy, param);
4139        if (retval)
4140                return retval;
4141        /*
4142         * make sure no PI-waiters arrive (or leave) while we are
4143         * changing the priority of the task:
4144         */
4145        spin_lock_irqsave(&p->pi_lock, flags);
4146        /*
4147         * To be able to change p->policy safely, the apropriate
4148         * runqueue lock must be held.
4149         */
4150        rq = __task_rq_lock(p);
4151        /* recheck policy now with rq lock held */
4152        if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4153                policy = oldpolicy = -1;
4154                __task_rq_unlock(rq);
4155                spin_unlock_irqrestore(&p->pi_lock, flags);
4156                goto recheck;
4157        }
4158        array = p->array;
4159        if (array)
4160                deactivate_task(p, rq);
4161        oldprio = p->prio;
4162        __setscheduler(p, policy, param->sched_priority);
4163        if (array) {
4164                __activate_task(p, rq);
4165                /*
4166                 * Reschedule if we are currently running on this runqueue and
4167                 * our priority decreased, or if we are not currently running on
4168                 * this runqueue and our priority is higher than the current's
4169                 */
4170                if (task_running(rq, p)) {
4171                        if (p->prio > oldprio)
4172                                resched_task(rq->curr);
4173                } else if (TASK_PREEMPTS_CURR(p, rq))
4174                        resched_task(rq->curr);
4175        }
4176        __task_rq_unlock(rq);
4177        spin_unlock_irqrestore(&p->pi_lock, flags);
4178
4179        rt_mutex_adjust_pi(p);
4180
4181        return 0;
4182}
4183EXPORT_SYMBOL_GPL(sched_setscheduler);
4184
4185static int
4186do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4187{
4188        struct sched_param lparam;
4189        struct task_struct *p;
4190        int retval;
4191
4192        if (!param || pid < 0)
4193                return -EINVAL;
4194        if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4195                return -EFAULT;
4196        read_lock_irq(&tasklist_lock);
4197        p = find_process_by_pid(pid);
4198        if (!p) {
4199                read_unlock_irq(&tasklist_lock);
4200                return -ESRCH;
4201        }
4202        retval = sched_setscheduler(p, policy, &lparam);
4203        read_unlock_irq(&tasklist_lock);
4204
4205        return retval;
4206}
4207
4208/**
4209 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4210 * @pid: the pid in question.
4211 * @policy: new policy.
4212 * @param: structure containing the new RT priority.
4213 */
4214asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
4215                                       struct sched_param __user *param)
4216{
4217        /* negative values for policy are not valid */
4218        if (policy < 0)
4219                return -EINVAL;
4220
4221        return do_sched_setscheduler(pid, policy, param);
4222}
4223
4224/**
4225 * sys_sched_setparam - set/change the RT priority of a thread
4226 * @pid: the pid in question.
4227 * @param: structure containing the new RT priority.
4228 */
4229asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4230{
4231        return do_sched_setscheduler(pid, -1, param);
4232}
4233
4234/**
4235 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4236 * @pid: the pid in question.
4237 */
4238asmlinkage long sys_sched_getscheduler(pid_t pid)
4239{
4240        struct task_struct *p;
4241        int retval = -EINVAL;
4242
4243        if (pid < 0)
4244                goto out_nounlock;
4245
4246        retval = -ESRCH;
4247        read_lock(&tasklist_lock);
4248        p = find_process_by_pid(pid);
4249        if (p) {
4250                retval = security_task_getscheduler(p);
4251                if (!retval)
4252                        retval = p->policy;
4253        }
4254        read_unlock(&tasklist_lock);
4255
4256out_nounlock:
4257        return retval;
4258}
4259
4260/**
4261 * sys_sched_getscheduler - get the RT priority of a thread
4262 * @pid: the pid in question.
4263 * @param: structure containing the RT priority.
4264 */
4265asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4266{
4267        struct sched_param lp;
4268        struct task_struct *p;
4269        int retval = -EINVAL;
4270
4271        if (!param || pid < 0)
4272                goto out_nounlock;
4273
4274        read_lock(&tasklist_lock);
4275        p = find_process_by_pid(pid);
4276        retval = -ESRCH;
4277        if (!p)
4278                goto out_unlock;
4279
4280        retval = security_task_getscheduler(p);
4281        if (retval)
4282                goto out_unlock;
4283
4284        lp.sched_priority = p->rt_priority;
4285        read_unlock(&tasklist_lock);
4286
4287        /*
4288         * This one might sleep, we cannot do it with a spinlock held ...
4289         */
4290        retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4291
4292out_nounlock:
4293        return retval;
4294
4295out_unlock:
4296        read_unlock(&tasklist_lock);
4297        return retval;
4298}
4299
4300long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4301{
4302        cpumask_t cpus_allowed;
4303        struct task_struct *p;
4304        int retval;
4305
4306        lock_cpu_hotplug();
4307        read_lock(&tasklist_lock);
4308
4309        p = find_process_by_pid(pid);
4310        if (!p) {
4311                read_unlock(&tasklist_lock);
4312                unlock_cpu_hotplug();
4313                return -ESRCH;
4314        }
4315
4316        /*
4317         * It is not safe to call set_cpus_allowed with the
4318         * tasklist_lock held.  We will bump the task_struct's
4319         * usage count and then drop tasklist_lock.
4320         */
4321        get_task_struct(p);
4322        read_unlock(&tasklist_lock);
4323
4324        retval = -EPERM;
4325        if ((current->euid != p->euid) && (current->euid != p->uid) &&
4326                        !capable(CAP_SYS_NICE))
4327                goto out_unlock;
4328
4329        retval = security_task_setscheduler(p, 0, NULL);
4330        if (retval)
4331                goto out_unlock;
4332
4333        cpus_allowed = cpuset_cpus_allowed(p);
4334        cpus_and(new_mask, new_mask, cpus_allowed);
4335        retval = set_cpus_allowed(p, new_mask);
4336
4337out_unlock:
4338        put_task_struct(p);
4339        unlock_cpu_hotplug();
4340        return retval;
4341}
4342
4343static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4344                             cpumask_t *new_mask)
4345{
4346        if (len < sizeof(cpumask_t)) {
4347                memset(new_mask, 0, sizeof(cpumask_t));
4348        } else if (len > sizeof(cpumask_t)) {
4349                len = sizeof(cpumask_t);
4350        }
4351        return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4352}
4353
4354/**
4355 * sys_sched_setaffinity - set the cpu affinity of a process
4356 * @pid: pid of the process
4357 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4358 * @user_mask_ptr: user-space pointer to the new cpu mask
4359 */
4360asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4361                                      unsigned long __user *user_mask_ptr)
4362{
4363        cpumask_t new_mask;
4364        int retval;
4365
4366        retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4367        if (retval)
4368                return retval;
4369
4370        return sched_setaffinity(pid, new_mask);
4371}
4372
4373/*
4374 * Represents all cpu's present in the system
4375 * In systems capable of hotplug, this map could dynamically grow
4376 * as new cpu's are detected in the system via any platform specific
4377 * method, such as ACPI for e.g.
4378 */
4379
4380cpumask_t cpu_present_map __read_mostly;
4381EXPORT_SYMBOL(cpu_present_map);
4382
4383#ifndef CONFIG_SMP
4384cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
4385cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
4386#endif
4387
4388long sched_getaffinity(pid_t pid, cpumask_t *mask)
4389{
4390        struct task_struct *p;
4391        int retval;
4392
4393        lock_cpu_hotplug();
4394        read_lock(&tasklist_lock);
4395
4396        retval = -ESRCH;
4397        p = find_process_by_pid(pid);
4398        if (!p)
4399                goto out_unlock;
4400
4401        retval = security_task_getscheduler(p);
4402        if (retval)
4403                goto out_unlock;
4404
4405        cpus_and(*mask, p->cpus_allowed, cpu_online_map);
4406
4407out_unlock:
4408        read_unlock(&tasklist_lock);
4409        unlock_cpu_hotplug();
4410        if (retval)
4411                return retval;
4412
4413        return 0;
4414}
4415
4416/**
4417 * sys_sched_getaffinity - get the cpu affinity of a process
4418 * @pid: pid of the process
4419 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4420 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4421 */
4422asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4423                                      unsigned long __user *user_mask_ptr)
4424{
4425        int ret;
4426        cpumask_t mask;
4427
4428        if (len < sizeof(cpumask_t))
4429                return -EINVAL;
4430
4431        ret = sched_getaffinity(pid, &mask);
4432        if (ret < 0)
4433                return ret;
4434
4435        if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4436                return -EFAULT;
4437
4438        return sizeof(cpumask_t);
4439}
4440
4441/**
4442 * sys_sched_yield - yield the current processor to other threads.
4443 *
4444 * this function yields the current CPU by moving the calling thread
4445 * to the expired array. If there are no other threads running on this
4446 * CPU then this function will return.
4447 */
4448asmlinkage long sys_sched_yield(void)
4449{
4450        struct rq *rq = this_rq_lock();
4451        struct prio_array *array = current->array, *target = rq->expired;
4452
4453        schedstat_inc(rq, yld_cnt);
4454        /*
4455         * We implement yielding by moving the task into the expired
4456         * queue.
4457         *
4458         * (special rule: RT tasks will just roundrobin in the active
4459         *  array.)
4460         */
4461        if (rt_task(current))
4462                target = rq->active;
4463
4464        if (array->nr_active == 1) {
4465                schedstat_inc(rq, yld_act_empty);
4466                if (!rq->expired->nr_active)
4467                        schedstat_inc(rq, yld_both_empty);
4468        } else if (!rq->expired->nr_active)
4469                schedstat_inc(rq, yld_exp_empty);
4470
4471        if (array != target) {
4472                dequeue_task(current, array);
4473                enqueue_task(current, target);
4474        } else
4475                /*
4476                 * requeue_task is cheaper so perform that if possible.
4477                 */
4478                requeue_task(current, array);
4479
4480        /*
4481         * Since we are going to call schedule() anyway, there's
4482         * no need to preempt or enable interrupts:
4483         */
4484        __release(rq->lock);
4485        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4486        _raw_spin_unlock(&rq->lock);
4487        preempt_enable_no_resched();
4488
4489        schedule();
4490
4491        return 0;
4492}
4493
4494static inline int __resched_legal(int expected_preempt_count)
4495{
4496        if (unlikely(preempt_count() != expected_preempt_count))
4497                return 0;
4498        if (unlikely(system_state != SYSTEM_RUNNING))
4499                return 0;
4500        return 1;
4501}
4502
4503static void __cond_resched(void)
4504{
4505#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4506        __might_sleep(__FILE__, __LINE__);
4507#endif
4508        /*
4509         * The BKS might be reacquired before we have dropped
4510         * PREEMPT_ACTIVE, which could trigger a second
4511         * cond_resched() call.
4512         */
4513        do {
4514                add_preempt_count(PREEMPT_ACTIVE);
4515                schedule();
4516                sub_preempt_count(PREEMPT_ACTIVE);
4517        } while (need_resched());
4518}
4519
4520int __sched cond_resched(void)
4521{
4522        if (need_resched() && __resched_legal(0)) {
4523                __cond_resched();
4524                return 1;
4525        }
4526        return 0;
4527}
4528EXPORT_SYMBOL(cond_resched);
4529
4530/*
4531 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4532 * call schedule, and on return reacquire the lock.
4533 *
4534 * This works OK both with and without CONFIG_PREEMPT.  We do strange low-level
4535 * operations here to prevent schedule() from being called twice (once via
4536 * spin_unlock(), once by hand).
4537 */
4538int cond_resched_lock(spinlock_t *lock)
4539{
4540        int ret = 0;
4541
4542        if (need_lockbreak(lock)) {
4543                spin_unlock(lock);
4544                cpu_relax();
4545                ret = 1;
4546                spin_lock(lock);
4547        }
4548        if (need_resched() && __resched_legal(1)) {
4549                spin_release(&lock->dep_map, 1, _THIS_IP_);
4550                _raw_spin_unlock(lock);
4551                preempt_enable_no_resched();
4552                __cond_resched();
4553                ret = 1;
4554                spin_lock(lock);
4555        }
4556        return ret;
4557}
4558EXPORT_SYMBOL(cond_resched_lock);
4559
4560int __sched cond_resched_softirq(void)
4561{
4562        BUG_ON(!in_softirq());
4563
4564        if (need_resched() && __resched_legal(0)) {
4565                raw_local_irq_disable();
4566                _local_bh_enable();
4567                raw_local_irq_enable();
4568                __cond_resched();
4569                local_bh_disable();
4570                return 1;
4571        }
4572        return 0;
4573}
4574EXPORT_SYMBOL(cond_resched_softirq);
4575
4576/**
4577 * yield - yield the current processor to other threads.
4578 *
4579 * this is a shortcut for kernel-space yielding - it marks the
4580 * thread runnable and calls sys_sched_yield().
4581 */
4582void __sched yield(void)
4583{
4584        set_current_state(TASK_RUNNING);
4585        sys_sched_yield();
4586}
4587EXPORT_SYMBOL(yield);
4588
4589/*
4590 * This task is about to go to sleep on IO.  Increment rq->nr_iowait so
4591 * that process accounting knows that this is a task in IO wait state.
4592 *
4593 * But don't do that if it is a deliberate, throttling IO wait (this task
4594 * has set its backing_dev_info: the queue against which it should throttle)
4595 */
4596void __sched io_schedule(void)
4597{
4598        struct rq *rq = &__raw_get_cpu_var(runqueues);
4599
4600        delayacct_blkio_start();
4601        atomic_inc(&rq->nr_iowait);
4602        schedule();
4603        atomic_dec(&rq->nr_iowait);
4604        delayacct_blkio_end();
4605}
4606EXPORT_SYMBOL(io_schedule);
4607
4608long __sched io_schedule_timeout(long timeout)
4609{
4610        struct rq *rq = &__raw_get_cpu_var(runqueues);
4611        long ret;
4612
4613        delayacct_blkio_start();
4614        atomic_inc(&rq->nr_iowait);
4615        ret = schedule_timeout(timeout);
4616        atomic_dec(&rq->nr_iowait);
4617        delayacct_blkio_end();
4618        return ret;
4619}
4620
4621/**
4622 * sys_sched_get_priority_max - return maximum RT priority.
4623 * @policy: scheduling class.
4624 *
4625 * this syscall returns the maximum rt_priority that can be used
4626 * by a given scheduling class.
4627 */
4628asmlinkage long sys_sched_get_priority_max(int policy)
4629{
4630        int ret = -EINVAL;
4631
4632        switch (policy) {
4633        case SCHED_FIFO:
4634        case SCHED_RR:
4635                ret = MAX_USER_RT_PRIO-1;
4636                break;
4637        case SCHED_NORMAL:
4638        case SCHED_BATCH:
4639                ret = 0;
4640                break;
4641        }
4642        return ret;
4643}
4644
4645/**
4646 * sys_sched_get_priority_min - return minimum RT priority.
4647 * @policy: scheduling class.
4648 *
4649 * this syscall returns the minimum rt_priority that can be used
4650 * by a given scheduling class.
4651 */
4652asmlinkage long sys_sched_get_priority_min(int policy)
4653{
4654        int ret = -EINVAL;
4655
4656        switch (policy) {
4657        case SCHED_FIFO:
4658        case SCHED_RR:
4659                ret = 1;
4660                break;
4661        case SCHED_NORMAL:
4662        case SCHED_BATCH:
4663                ret = 0;
4664        }
4665        return ret;
4666}
4667
4668/**
4669 * sys_sched_rr_get_interval - return the default timeslice of a process.
4670 * @pid: pid of the process.
4671 * @interval: userspace pointer to the timeslice value.
4672 *
4673 * this syscall writes the default timeslice value of a given process
4674 * into the user-space timespec buffer. A value of '0' means infinity.
4675 */
4676asmlinkage
4677long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4678{
4679        struct task_struct *p;
4680        int retval = -EINVAL;
4681        struct timespec t;
4682
4683        if (pid < 0)
4684                goto out_nounlock;
4685
4686        retval = -ESRCH;
4687        read_lock(&tasklist_lock);
4688        p = find_process_by_pid(pid);
4689        if (!p)
4690                goto out_unlock;
4691
4692        retval = security_task_getscheduler(p);
4693        if (retval)
4694                goto out_unlock;
4695
4696        jiffies_to_timespec(p->policy == SCHED_FIFO ?
4697                                0 : task_timeslice(p), &t);
4698        read_unlock(&tasklist_lock);
4699        retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4700out_nounlock:
4701        return retval;
4702out_unlock:
4703        read_unlock(&tasklist_lock);
4704        return retval;
4705}
4706
4707static inline struct task_struct *eldest_child(struct task_struct *p)
4708{
4709        if (list_empty(&p->children))
4710                return NULL;
4711        return list_entry(p->children.next,struct task_struct,sibling);
4712}
4713
4714static inline struct task_struct *older_sibling(struct task_struct *p)
4715{
4716        if (p->sibling.prev==&p->parent->children)
4717                return NULL;
4718        return list_entry(p->sibling.prev,struct task_struct,sibling);
4719}
4720
4721static inline struct task_struct *younger_sibling(struct task_struct *p)
4722{
4723        if (p->sibling.next==&p->parent->children)
4724                return NULL;
4725        return list_entry(p->sibling.next,struct task_struct,sibling);
4726}
4727
4728static const char stat_nam[] = "RSDTtZX";
4729
4730static void show_task(struct task_struct *p)
4731{
4732        struct task_struct *relative;
4733        unsigned long free = 0;
4734        unsigned state;
4735
4736        state = p->state ? __ffs(p->state) + 1 : 0;
4737        printk("%-13.13s %c", p->comm,
4738                state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4739#if (BITS_PER_LONG == 32)
4740        if (state == TASK_RUNNING)
4741                printk(" running ");
4742        else
4743                printk(" %08lX ", thread_saved_pc(p));
4744#else
4745        if (state == TASK_RUNNING)
4746                printk("  running task   ");
4747        else
4748                printk(" %016lx ", thread_saved_pc(p));
4749#endif
4750#ifdef CONFIG_DEBUG_STACK_USAGE
4751        {
4752                unsigned long *n = end_of_stack(p);
4753                while (!*n)
4754                        n++;
4755                free = (unsigned long)n - (unsigned long)end_of_stack(p);
4756        }
4757#endif
4758        printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
4759        if ((relative = eldest_child(p)))
4760                printk("%5d ", relative->pid);
4761        else
4762                printk("      ");
4763        if ((relative = younger_sibling(p)))
4764                printk("%7d", relative->pid);
4765        else
4766                printk("       ");
4767        if ((relative = older_sibling(p)))
4768                printk(" %5d", relative->pid);
4769        else
4770                printk("      ");
4771        if (!p->mm)
4772                printk(" (L-TLB)\n");
4773        else
4774                printk(" (NOTLB)\n");
4775
4776        if (state != TASK_RUNNING)
4777                show_stack(p, NULL);
4778}
4779
4780void show_state(void)
4781{
4782        struct task_struct *g, *p;
4783
4784#if (BITS_PER_LONG == 32)
4785        printk("\n"
4786               "                                               sibling\n");
4787        printk("  task             PC      pid father child younger older\n");
4788#else
4789        printk("\n"
4790               "                                                       sibling\n");
4791        printk("  task                 PC          pid father child younger older\n");
4792#endif
4793        read_lock(&tasklist_lock);
4794        do_each_thread(g, p) {
4795                /*
4796                 * reset the NMI-timeout, listing all files on a slow
4797                 * console might take alot of time:
4798                 */
4799                touch_nmi_watchdog();
4800                show_task(p);
4801        } while_each_thread(g, p);
4802
4803        read_unlock(&tasklist_lock);
4804        debug_show_all_locks();
4805}
4806
4807/**
4808 * init_idle - set up an idle thread for a given CPU
4809 * @idle: task in question
4810 * @cpu: cpu the idle task belongs to
4811 *
4812 * NOTE: this function does not set the idle thread's NEED_RESCHED
4813 * flag, to make booting more robust.
4814 */
4815void __devinit init_idle(struct task_struct *idle, int cpu)
4816{
4817        struct rq *rq = cpu_rq(cpu);
4818        unsigned long flags;
4819
4820        idle->timestamp = sched_clock();
4821        idle->sleep_avg = 0;
4822        idle->array = NULL;
4823        idle->prio = idle->normal_prio = MAX_PRIO;
4824        idle->state = TASK_RUNNING;
4825        idle->cpus_allowed = cpumask_of_cpu(cpu);
4826        set_task_cpu(idle, cpu);
4827
4828        spin_lock_irqsave(&rq->lock, flags);
4829        rq->curr = rq->idle = idle;
4830#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4831        idle->oncpu = 1;
4832#endif
4833        spin_unlock_irqrestore(&rq->lock, flags);
4834
4835        /* Set the preempt count _outside_ the spinlocks! */
4836#if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4837        task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
4838#else
4839        task_thread_info(idle)->preempt_count = 0;
4840#endif
4841}
4842
4843/*
4844 * In a system that switches off the HZ timer nohz_cpu_mask
4845 * indicates which cpus entered this state. This is used
4846 * in the rcu update to wait only for active cpus. For system
4847 * which do not switch off the HZ timer nohz_cpu_mask should
4848 * always be CPU_MASK_NONE.
4849 */
4850cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
4851
4852#ifdef CONFIG_SMP
4853/*
4854 * This is how migration works:
4855 *
4856 * 1) we queue a struct migration_req structure in the source CPU's
4857 *    runqueue and wake up that CPU's migration thread.
4858 * 2) we down() the locked semaphore => thread blocks.
4859 * 3) migration thread wakes up (implicitly it forces the migrated
4860 *    thread off the CPU)
4861 * 4) it gets the migration request and checks whether the migrated
4862 *    task is still in the wrong runqueue.
4863 * 5) if it's in the wrong runqueue then the migration thread removes
4864 *    it and puts it into the right queue.
4865 * 6) migration thread up()s the semaphore.
4866 * 7) we wake up and the migration is done.
4867 */
4868
4869/*
4870 * Change a given task's CPU affinity. Migrate the thread to a
4871 * proper CPU and schedule it away if the CPU it's executing on
4872 * is removed from the allowed bitmask.
4873 *
4874 * NOTE: the caller must have a valid reference to the task, the
4875 * task must not exit() & deallocate itself prematurely.  The
4876 * call is not atomic; no spinlocks may be held.
4877 */
4878int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
4879{
4880        struct migration_req req;
4881        unsigned long flags;
4882        struct rq *rq;
4883        int ret = 0;
4884
4885        rq = task_rq_lock(p, &flags);
4886        if (!cpus_intersects(new_mask, cpu_online_map)) {
4887                ret = -EINVAL;
4888                goto out;
4889        }
4890
4891        p->cpus_allowed = new_mask;
4892        /* Can the task run on the task's current CPU? If so, we're done */
4893        if (cpu_isset(task_cpu(p), new_mask))
4894                goto out;
4895
4896        if (migrate_task(p, any_online_cpu(new_mask), &req)) {
4897                /* Need help from migration thread: drop lock and wait. */
4898                task_rq_unlock(rq, &flags);
4899                wake_up_process(rq->migration_thread);
4900                wait_for_completion(&req.done);
4901                tlb_migrate_finish(p->mm);
4902                return 0;
4903        }
4904out:
4905        task_rq_unlock(rq, &flags);
4906
4907        return ret;
4908}
4909EXPORT_SYMBOL_GPL(set_cpus_allowed);
4910
4911/*
4912 * Move (not current) task off this cpu, onto dest cpu.  We're doing
4913 * this because either it can't run here any more (set_cpus_allowed()
4914 * away from this CPU, or CPU going down), or because we're
4915 * attempting to rebalance this task on exec (sched_exec).
4916 *
4917 * So we race with normal scheduler movements, but that's OK, as long
4918 * as the task is no longer on this CPU.
4919 *
4920 * Returns non-zero if task was successfully migrated.
4921 */
4922static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4923{
4924        struct rq *rq_dest, *rq_src;
4925        int ret = 0;
4926
4927        if (unlikely(cpu_is_offline(dest_cpu)))
4928                return ret;
4929
4930        rq_src = cpu_rq(src_cpu);
4931        rq_dest = cpu_rq(dest_cpu);
4932
4933        double_rq_lock(rq_src, rq_dest);
4934        /* Already moved. */
4935        if (task_cpu(p) != src_cpu)
4936                goto out;
4937        /* Affinity changed (again). */
4938        if (!cpu_isset(dest_cpu, p->cpus_allowed))
4939                goto out;
4940
4941        set_task_cpu(p, dest_cpu);
4942        if (p->array) {
4943                /*
4944                 * Sync timestamp with rq_dest's before activating.
4945                 * The same thing could be achieved by doing this step
4946                 * afterwards, and pretending it was a local activate.
4947                 * This way is cleaner and logically correct.
4948                 */
4949                p->timestamp = p->timestamp - rq_src->timestamp_last_tick
4950                                + rq_dest->timestamp_last_tick;
4951                deactivate_task(p, rq_src);
4952                __activate_task(p, rq_dest);
4953                if (TASK_PREEMPTS_CURR(p, rq_dest))
4954                        resched_task(rq_dest->curr);
4955        }
4956        ret = 1;
4957out:
4958        double_rq_unlock(rq_src, rq_dest);
4959        return ret;
4960}
4961
4962/*
4963 * migration_thread - this is a highprio system thread that performs
4964 * thread migration by bumping thread off CPU then 'pushing' onto
4965 * another runqueue.
4966 */
4967static int migration_thread(void *data)
4968{
4969        int cpu = (long)data;
4970        struct rq *rq;
4971
4972        rq = cpu_rq(cpu);
4973        BUG_ON(rq->migration_thread != current);
4974
4975        set_current_state(TASK_INTERRUPTIBLE);
4976        while (!kthread_should_stop()) {
4977                struct migration_req *req;
4978                struct list_head *head;
4979
4980                try_to_freeze();
4981
4982                spin_lock_irq(&rq->lock);
4983
4984                if (cpu_is_offline(cpu)) {
4985                        spin_unlock_irq(&rq->lock);
4986                        goto wait_to_die;
4987                }
4988
4989                if (rq->active_balance) {
4990                        active_load_balance(rq, cpu);
4991                        rq->active_balance = 0;
4992                }
4993
4994                head = &rq->migration_queue;
4995
4996                if (list_empty(head)) {
4997                        spin_unlock_irq(&rq->lock);
4998                        schedule();
4999                        set_current_state(TASK_INTERRUPTIBLE);
5000                        continue;
5001                }
5002                req = list_entry(head->next, struct migration_req, list);
5003                list_del_init(head->next);
5004
5005                spin_unlock(&rq->lock);
5006                __migrate_task(req->task, cpu, req->dest_cpu);
5007                local_irq_enable();
5008
5009                complete(&req->done);
5010        }
5011        __set_current_state(TASK_RUNNING);
5012        return 0;
5013
5014wait_to_die:
5015        /* Wait for kthread_stop */
5016        set_current_state(TASK_INTERRUPTIBLE);
5017        while (!kthread_should_stop()) {
5018                schedule();
5019                set_current_state(TASK_INTERRUPTIBLE);
5020        }
5021        __set_current_state(TASK_RUNNING);
5022        return 0;
5023}
5024
5025#ifdef CONFIG_HOTPLUG_CPU
5026/* Figure out where task on dead CPU should go, use force if neccessary. */
5027static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5028{
5029        unsigned long flags;
5030        cpumask_t mask;
5031        struct rq *rq;
5032        int dest_cpu;
5033
5034restart:
5035        /* On same node? */
5036        mask = node_to_cpumask(cpu_to_node(dead_cpu));
5037        cpus_and(mask, mask, p->cpus_allowed);
5038        dest_cpu = any_online_cpu(mask);
5039
5040        /* On any allowed CPU? */
5041        if (dest_cpu == NR_CPUS)
5042                dest_cpu = any_online_cpu(p->cpus_allowed);
5043
5044        /* No more Mr. Nice Guy. */
5045        if (dest_cpu == NR_CPUS) {
5046                rq = task_rq_lock(p, &flags);
5047                cpus_setall(p->cpus_allowed);
5048                dest_cpu = any_online_cpu(p->cpus_allowed);
5049                task_rq_unlock(rq, &flags);
5050
5051                /*
5052                 * Don't tell them about moving exiting tasks or
5053                 * kernel threads (both mm NULL), since they never
5054                 * leave kernel.
5055                 */
5056                if (p->mm && printk_ratelimit())
5057                        printk(KERN_INFO "process %d (%s) no "
5058                               "longer affine to cpu%d\n",
5059                               p->pid, p->comm, dead_cpu);
5060        }
5061        if (!__migrate_task(p, dead_cpu, dest_cpu))
5062                goto restart;
5063}
5064
5065/*
5066 * While a dead CPU has no uninterruptible tasks queued at this point,
5067 * it might still have a nonzero ->nr_uninterruptible counter, because
5068 * for performance reasons the counter is not stricly tracking tasks to
5069 * their home CPUs. So we just add the counter to another CPU's counter,
5070 * to keep the global sum constant after CPU-down:
5071 */
5072static void migrate_nr_uninterruptible(struct rq *rq_src)
5073{
5074        struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
5075        unsigned long flags;
5076
5077        local_irq_save(flags);
5078        double_rq_lock(rq_src, rq_dest);
5079        rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5080        rq_src->nr_uninterruptible = 0;
5081        double_rq_unlock(rq_src, rq_dest);
5082        local_irq_restore(flags);
5083}
5084
5085/* Run through task list and migrate tasks from the dead cpu. */
5086static void migrate_live_tasks(int src_cpu)
5087{
5088        struct task_struct *p, *t;
5089
5090        write_lock_irq(&tasklist_lock);
5091
5092        do_each_thread(t, p) {
5093                if (p == current)
5094                        continue;
5095
5096                if (task_cpu(p) == src_cpu)
5097                        move_task_off_dead_cpu(src_cpu, p);
5098        } while_each_thread(t, p);
5099
5100        write_unlock_irq(&tasklist_lock);
5101}
5102
5103/* Schedules idle task to be the next runnable task on current CPU.
5104 * It does so by boosting its priority to highest possible and adding it to
5105 * the _front_ of the runqueue. Used by CPU offline code.
5106 */
5107void sched_idle_next(void)
5108{
5109        int this_cpu = smp_processor_id();
5110        struct rq *rq = cpu_rq(this_cpu);
5111        struct task_struct *p = rq->idle;
5112        unsigned long flags;
5113
5114        /* cpu has to be offline */
5115        BUG_ON(cpu_online(this_cpu));
5116
5117        /*