linux/kernel/sched/core.c
<<
>>
Prefs
   1/*
   2 *  kernel/sched/core.c
   3 *
   4 *  Kernel scheduler and related syscalls
   5 *
   6 *  Copyright (C) 1991-2002  Linus Torvalds
   7 *
   8 *  1996-12-23  Modified by Dave Grothe to fix bugs in semaphores and
   9 *              make semaphores SMP safe
  10 *  1998-11-19  Implemented schedule_timeout() and related stuff
  11 *              by Andrea Arcangeli
  12 *  2002-01-04  New ultra-scalable O(1) scheduler by Ingo Molnar:
  13 *              hybrid priority-list and round-robin design with
  14 *              an array-switch method of distributing timeslices
  15 *              and per-CPU runqueues.  Cleanups and useful suggestions
  16 *              by Davide Libenzi, preemptible kernel bits by Robert Love.
  17 *  2003-09-03  Interactivity tuning by Con Kolivas.
  18 *  2004-04-02  Scheduler domains code by Nick Piggin
  19 *  2007-04-15  Work begun on replacing all interactivity tuning with a
  20 *              fair scheduling design by Con Kolivas.
  21 *  2007-05-05  Load balancing (smp-nice) and other improvements
  22 *              by Peter Williams
  23 *  2007-05-06  Interactivity improvements to CFS by Mike Galbraith
  24 *  2007-07-01  Group scheduling enhancements by Srivatsa Vaddagiri
  25 *  2007-11-29  RT balancing improvements by Steven Rostedt, Gregory Haskins,
  26 *              Thomas Gleixner, Mike Kravetz
  27 */
  28
  29#include <linux/mm.h>
  30#include <linux/module.h>
  31#include <linux/nmi.h>
  32#include <linux/init.h>
  33#include <linux/uaccess.h>
  34#include <linux/highmem.h>
  35#include <asm/mmu_context.h>
  36#include <linux/interrupt.h>
  37#include <linux/capability.h>
  38#include <linux/completion.h>
  39#include <linux/kernel_stat.h>
  40#include <linux/debug_locks.h>
  41#include <linux/perf_event.h>
  42#include <linux/security.h>
  43#include <linux/notifier.h>
  44#include <linux/profile.h>
  45#include <linux/freezer.h>
  46#include <linux/vmalloc.h>
  47#include <linux/blkdev.h>
  48#include <linux/delay.h>
  49#include <linux/pid_namespace.h>
  50#include <linux/smp.h>
  51#include <linux/threads.h>
  52#include <linux/timer.h>
  53#include <linux/rcupdate.h>
  54#include <linux/cpu.h>
  55#include <linux/cpuset.h>
  56#include <linux/percpu.h>
  57#include <linux/proc_fs.h>
  58#include <linux/seq_file.h>
  59#include <linux/sysctl.h>
  60#include <linux/syscalls.h>
  61#include <linux/times.h>
  62#include <linux/tsacct_kern.h>
  63#include <linux/kprobes.h>
  64#include <linux/delayacct.h>
  65#include <linux/unistd.h>
  66#include <linux/pagemap.h>
  67#include <linux/hrtimer.h>
  68#include <linux/tick.h>
  69#include <linux/debugfs.h>
  70#include <linux/ctype.h>
  71#include <linux/ftrace.h>
  72#include <linux/slab.h>
  73#include <linux/init_task.h>
  74#include <linux/binfmts.h>
  75
  76#include <asm/switch_to.h>
  77#include <asm/tlb.h>
  78#include <asm/irq_regs.h>
  79#include <asm/mutex.h>
  80#ifdef CONFIG_PARAVIRT
  81#include <asm/paravirt.h>
  82#endif
  83
  84#include "sched.h"
  85#include "../workqueue_sched.h"
  86#include "../smpboot.h"
  87
  88#define CREATE_TRACE_POINTS
  89#include <trace/events/sched.h>
  90
  91void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
  92{
  93        unsigned long delta;
  94        ktime_t soft, hard, now;
  95
  96        for (;;) {
  97                if (hrtimer_active(period_timer))
  98                        break;
  99
 100                now = hrtimer_cb_get_time(period_timer);
 101                hrtimer_forward(period_timer, now, period);
 102
 103                soft = hrtimer_get_softexpires(period_timer);
 104                hard = hrtimer_get_expires(period_timer);
 105                delta = ktime_to_ns(ktime_sub(hard, soft));
 106                __hrtimer_start_range_ns(period_timer, soft, delta,
 107                                         HRTIMER_MODE_ABS_PINNED, 0);
 108        }
 109}
 110
 111DEFINE_MUTEX(sched_domains_mutex);
 112DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
 113
 114static void update_rq_clock_task(struct rq *rq, s64 delta);
 115
 116void update_rq_clock(struct rq *rq)
 117{
 118        s64 delta;
 119
 120        if (rq->skip_clock_update > 0)
 121                return;
 122
 123        delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
 124        rq->clock += delta;
 125        update_rq_clock_task(rq, delta);
 126}
 127
 128/*
 129 * Debugging: various feature bits
 130 */
 131
 132#define SCHED_FEAT(name, enabled)       \
 133        (1UL << __SCHED_FEAT_##name) * enabled |
 134
 135const_debug unsigned int sysctl_sched_features =
 136#include "features.h"
 137        0;
 138
 139#undef SCHED_FEAT
 140
 141#ifdef CONFIG_SCHED_DEBUG
 142#define SCHED_FEAT(name, enabled)       \
 143        #name ,
 144
 145static const char * const sched_feat_names[] = {
 146#include "features.h"
 147};
 148
 149#undef SCHED_FEAT
 150
 151static int sched_feat_show(struct seq_file *m, void *v)
 152{
 153        int i;
 154
 155        for (i = 0; i < __SCHED_FEAT_NR; i++) {
 156                if (!(sysctl_sched_features & (1UL << i)))
 157                        seq_puts(m, "NO_");
 158                seq_printf(m, "%s ", sched_feat_names[i]);
 159        }
 160        seq_puts(m, "\n");
 161
 162        return 0;
 163}
 164
 165#ifdef HAVE_JUMP_LABEL
 166
 167#define jump_label_key__true  STATIC_KEY_INIT_TRUE
 168#define jump_label_key__false STATIC_KEY_INIT_FALSE
 169
 170#define SCHED_FEAT(name, enabled)       \
 171        jump_label_key__##enabled ,
 172
 173struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
 174#include "features.h"
 175};
 176
 177#undef SCHED_FEAT
 178
 179static void sched_feat_disable(int i)
 180{
 181        if (static_key_enabled(&sched_feat_keys[i]))
 182                static_key_slow_dec(&sched_feat_keys[i]);
 183}
 184
 185static void sched_feat_enable(int i)
 186{
 187        if (!static_key_enabled(&sched_feat_keys[i]))
 188                static_key_slow_inc(&sched_feat_keys[i]);
 189}
 190#else
 191static void sched_feat_disable(int i) { };
 192static void sched_feat_enable(int i) { };
 193#endif /* HAVE_JUMP_LABEL */
 194
 195static ssize_t
 196sched_feat_write(struct file *filp, const char __user *ubuf,
 197                size_t cnt, loff_t *ppos)
 198{
 199        char buf[64];
 200        char *cmp;
 201        int neg = 0;
 202        int i;
 203
 204        if (cnt > 63)
 205                cnt = 63;
 206
 207        if (copy_from_user(&buf, ubuf, cnt))
 208                return -EFAULT;
 209
 210        buf[cnt] = 0;
 211        cmp = strstrip(buf);
 212
 213        if (strncmp(cmp, "NO_", 3) == 0) {
 214                neg = 1;
 215                cmp += 3;
 216        }
 217
 218        for (i = 0; i < __SCHED_FEAT_NR; i++) {
 219                if (strcmp(cmp, sched_feat_names[i]) == 0) {
 220                        if (neg) {
 221                                sysctl_sched_features &= ~(1UL << i);
 222                                sched_feat_disable(i);
 223                        } else {
 224                                sysctl_sched_features |= (1UL << i);
 225                                sched_feat_enable(i);
 226                        }
 227                        break;
 228                }
 229        }
 230
 231        if (i == __SCHED_FEAT_NR)
 232                return -EINVAL;
 233
 234        *ppos += cnt;
 235
 236        return cnt;
 237}
 238
 239static int sched_feat_open(struct inode *inode, struct file *filp)
 240{
 241        return single_open(filp, sched_feat_show, NULL);
 242}
 243
 244static const struct file_operations sched_feat_fops = {
 245        .open           = sched_feat_open,
 246        .write          = sched_feat_write,
 247        .read           = seq_read,
 248        .llseek         = seq_lseek,
 249        .release        = single_release,
 250};
 251
 252static __init int sched_init_debug(void)
 253{
 254        debugfs_create_file("sched_features", 0644, NULL, NULL,
 255                        &sched_feat_fops);
 256
 257        return 0;
 258}
 259late_initcall(sched_init_debug);
 260#endif /* CONFIG_SCHED_DEBUG */
 261
 262/*
 263 * Number of tasks to iterate in a single balance run.
 264 * Limited because this is done with IRQs disabled.
 265 */
 266const_debug unsigned int sysctl_sched_nr_migrate = 32;
 267
 268/*
 269 * period over which we average the RT time consumption, measured
 270 * in ms.
 271 *
 272 * default: 1s
 273 */
 274const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
 275
 276/*
 277 * period over which we measure -rt task cpu usage in us.
 278 * default: 1s
 279 */
 280unsigned int sysctl_sched_rt_period = 1000000;
 281
 282__read_mostly int scheduler_running;
 283
 284/*
 285 * part of the period that we allow rt tasks to run in us.
 286 * default: 0.95s
 287 */
 288int sysctl_sched_rt_runtime = 950000;
 289
 290
 291
 292/*
 293 * __task_rq_lock - lock the rq @p resides on.
 294 */
 295static inline struct rq *__task_rq_lock(struct task_struct *p)
 296        __acquires(rq->lock)
 297{
 298        struct rq *rq;
 299
 300        lockdep_assert_held(&p->pi_lock);
 301
 302        for (;;) {
 303                rq = task_rq(p);
 304                raw_spin_lock(&rq->lock);
 305                if (likely(rq == task_rq(p)))
 306                        return rq;
 307                raw_spin_unlock(&rq->lock);
 308        }
 309}
 310
 311/*
 312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
 313 */
 314static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
 315        __acquires(p->pi_lock)
 316        __acquires(rq->lock)
 317{
 318        struct rq *rq;
 319
 320        for (;;) {
 321                raw_spin_lock_irqsave(&p->pi_lock, *flags);
 322                rq = task_rq(p);
 323                raw_spin_lock(&rq->lock);
 324                if (likely(rq == task_rq(p)))
 325                        return rq;
 326                raw_spin_unlock(&rq->lock);
 327                raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
 328        }
 329}
 330
 331static void __task_rq_unlock(struct rq *rq)
 332        __releases(rq->lock)
 333{
 334        raw_spin_unlock(&rq->lock);
 335}
 336
 337static inline void
 338task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
 339        __releases(rq->lock)
 340        __releases(p->pi_lock)
 341{
 342        raw_spin_unlock(&rq->lock);
 343        raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
 344}
 345
 346/*
 347 * this_rq_lock - lock this runqueue and disable interrupts.
 348 */
 349static struct rq *this_rq_lock(void)
 350        __acquires(rq->lock)
 351{
 352        struct rq *rq;
 353
 354        local_irq_disable();
 355        rq = this_rq();
 356        raw_spin_lock(&rq->lock);
 357
 358        return rq;
 359}
 360
 361#ifdef CONFIG_SCHED_HRTICK
 362/*
 363 * Use HR-timers to deliver accurate preemption points.
 364 *
 365 * Its all a bit involved since we cannot program an hrt while holding the
 366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
 367 * reschedule event.
 368 *
 369 * When we get rescheduled we reprogram the hrtick_timer outside of the
 370 * rq->lock.
 371 */
 372
 373static void hrtick_clear(struct rq *rq)
 374{
 375        if (hrtimer_active(&rq->hrtick_timer))
 376                hrtimer_cancel(&rq->hrtick_timer);
 377}
 378
 379/*
 380 * High-resolution timer tick.
 381 * Runs from hardirq context with interrupts disabled.
 382 */
 383static enum hrtimer_restart hrtick(struct hrtimer *timer)
 384{
 385        struct rq *rq = container_of(timer, struct rq, hrtick_timer);
 386
 387        WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
 388
 389        raw_spin_lock(&rq->lock);
 390        update_rq_clock(rq);
 391        rq->curr->sched_class->task_tick(rq, rq->curr, 1);
 392        raw_spin_unlock(&rq->lock);
 393
 394        return HRTIMER_NORESTART;
 395}
 396
 397#ifdef CONFIG_SMP
 398/*
 399 * called from hardirq (IPI) context
 400 */
 401static void __hrtick_start(void *arg)
 402{
 403        struct rq *rq = arg;
 404
 405        raw_spin_lock(&rq->lock);
 406        hrtimer_restart(&rq->hrtick_timer);
 407        rq->hrtick_csd_pending = 0;
 408        raw_spin_unlock(&rq->lock);
 409}
 410
 411/*
 412 * Called to set the hrtick timer state.
 413 *
 414 * called with rq->lock held and irqs disabled
 415 */
 416void hrtick_start(struct rq *rq, u64 delay)
 417{
 418        struct hrtimer *timer = &rq->hrtick_timer;
 419        ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
 420
 421        hrtimer_set_expires(timer, time);
 422
 423        if (rq == this_rq()) {
 424                hrtimer_restart(timer);
 425        } else if (!rq->hrtick_csd_pending) {
 426                __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
 427                rq->hrtick_csd_pending = 1;
 428        }
 429}
 430
 431static int
 432hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
 433{
 434        int cpu = (int)(long)hcpu;
 435
 436        switch (action) {
 437        case CPU_UP_CANCELED:
 438        case CPU_UP_CANCELED_FROZEN:
 439        case CPU_DOWN_PREPARE:
 440        case CPU_DOWN_PREPARE_FROZEN:
 441        case CPU_DEAD:
 442        case CPU_DEAD_FROZEN:
 443                hrtick_clear(cpu_rq(cpu));
 444                return NOTIFY_OK;
 445        }
 446
 447        return NOTIFY_DONE;
 448}
 449
 450static __init void init_hrtick(void)
 451{
 452        hotcpu_notifier(hotplug_hrtick, 0);
 453}
 454#else
 455/*
 456 * Called to set the hrtick timer state.
 457 *
 458 * called with rq->lock held and irqs disabled
 459 */
 460void hrtick_start(struct rq *rq, u64 delay)
 461{
 462        __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
 463                        HRTIMER_MODE_REL_PINNED, 0);
 464}
 465
 466static inline void init_hrtick(void)
 467{
 468}
 469#endif /* CONFIG_SMP */
 470
 471static void init_rq_hrtick(struct rq *rq)
 472{
 473#ifdef CONFIG_SMP
 474        rq->hrtick_csd_pending = 0;
 475
 476        rq->hrtick_csd.flags = 0;
 477        rq->hrtick_csd.func = __hrtick_start;
 478        rq->hrtick_csd.info = rq;
 479#endif
 480
 481        hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 482        rq->hrtick_timer.function = hrtick;
 483}
 484#else   /* CONFIG_SCHED_HRTICK */
 485static inline void hrtick_clear(struct rq *rq)
 486{
 487}
 488
 489static inline void init_rq_hrtick(struct rq *rq)
 490{
 491}
 492
 493static inline void init_hrtick(void)
 494{
 495}
 496#endif  /* CONFIG_SCHED_HRTICK */
 497
 498/*
 499 * resched_task - mark a task 'to be rescheduled now'.
 500 *
 501 * On UP this means the setting of the need_resched flag, on SMP it
 502 * might also involve a cross-CPU call to trigger the scheduler on
 503 * the target CPU.
 504 */
 505#ifdef CONFIG_SMP
 506
 507#ifndef tsk_is_polling
 508#define tsk_is_polling(t) 0
 509#endif
 510
 511void resched_task(struct task_struct *p)
 512{
 513        int cpu;
 514
 515        assert_raw_spin_locked(&task_rq(p)->lock);
 516
 517        if (test_tsk_need_resched(p))
 518                return;
 519
 520        set_tsk_need_resched(p);
 521
 522        cpu = task_cpu(p);
 523        if (cpu == smp_processor_id())
 524                return;
 525
 526        /* NEED_RESCHED must be visible before we test polling */
 527        smp_mb();
 528        if (!tsk_is_polling(p))
 529                smp_send_reschedule(cpu);
 530}
 531
 532void resched_cpu(int cpu)
 533{
 534        struct rq *rq = cpu_rq(cpu);
 535        unsigned long flags;
 536
 537        if (!raw_spin_trylock_irqsave(&rq->lock, flags))
 538                return;
 539        resched_task(cpu_curr(cpu));
 540        raw_spin_unlock_irqrestore(&rq->lock, flags);
 541}
 542
 543#ifdef CONFIG_NO_HZ
 544/*
 545 * In the semi idle case, use the nearest busy cpu for migrating timers
 546 * from an idle cpu.  This is good for power-savings.
 547 *
 548 * We don't do similar optimization for completely idle system, as
 549 * selecting an idle cpu will add more delays to the timers than intended
 550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
 551 */
 552int get_nohz_timer_target(void)
 553{
 554        int cpu = smp_processor_id();
 555        int i;
 556        struct sched_domain *sd;
 557
 558        rcu_read_lock();
 559        for_each_domain(cpu, sd) {
 560                for_each_cpu(i, sched_domain_span(sd)) {
 561                        if (!idle_cpu(i)) {
 562                                cpu = i;
 563                                goto unlock;
 564                        }
 565                }
 566        }
 567unlock:
 568        rcu_read_unlock();
 569        return cpu;
 570}
 571/*
 572 * When add_timer_on() enqueues a timer into the timer wheel of an
 573 * idle CPU then this timer might expire before the next timer event
 574 * which is scheduled to wake up that CPU. In case of a completely
 575 * idle system the next event might even be infinite time into the
 576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
 577 * leaves the inner idle loop so the newly added timer is taken into
 578 * account when the CPU goes back to idle and evaluates the timer
 579 * wheel for the next timer event.
 580 */
 581void wake_up_idle_cpu(int cpu)
 582{
 583        struct rq *rq = cpu_rq(cpu);
 584
 585        if (cpu == smp_processor_id())
 586                return;
 587
 588        /*
 589         * This is safe, as this function is called with the timer
 590         * wheel base lock of (cpu) held. When the CPU is on the way
 591         * to idle and has not yet set rq->curr to idle then it will
 592         * be serialized on the timer wheel base lock and take the new
 593         * timer into account automatically.
 594         */
 595        if (rq->curr != rq->idle)
 596                return;
 597
 598        /*
 599         * We can set TIF_RESCHED on the idle task of the other CPU
 600         * lockless. The worst case is that the other CPU runs the
 601         * idle task through an additional NOOP schedule()
 602         */
 603        set_tsk_need_resched(rq->idle);
 604
 605        /* NEED_RESCHED must be visible before we test polling */
 606        smp_mb();
 607        if (!tsk_is_polling(rq->idle))
 608                smp_send_reschedule(cpu);
 609}
 610
 611static inline bool got_nohz_idle_kick(void)
 612{
 613        int cpu = smp_processor_id();
 614        return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
 615}
 616
 617#else /* CONFIG_NO_HZ */
 618
 619static inline bool got_nohz_idle_kick(void)
 620{
 621        return false;
 622}
 623
 624#endif /* CONFIG_NO_HZ */
 625
 626void sched_avg_update(struct rq *rq)
 627{
 628        s64 period = sched_avg_period();
 629
 630        while ((s64)(rq->clock - rq->age_stamp) > period) {
 631                /*
 632                 * Inline assembly required to prevent the compiler
 633                 * optimising this loop into a divmod call.
 634                 * See __iter_div_u64_rem() for another example of this.
 635                 */
 636                asm("" : "+rm" (rq->age_stamp));
 637                rq->age_stamp += period;
 638                rq->rt_avg /= 2;
 639        }
 640}
 641
 642#else /* !CONFIG_SMP */
 643void resched_task(struct task_struct *p)
 644{
 645        assert_raw_spin_locked(&task_rq(p)->lock);
 646        set_tsk_need_resched(p);
 647}
 648#endif /* CONFIG_SMP */
 649
 650#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
 651                        (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
 652/*
 653 * Iterate task_group tree rooted at *from, calling @down when first entering a
 654 * node and @up when leaving it for the final time.
 655 *
 656 * Caller must hold rcu_lock or sufficient equivalent.
 657 */
 658int walk_tg_tree_from(struct task_group *from,
 659                             tg_visitor down, tg_visitor up, void *data)
 660{
 661        struct task_group *parent, *child;
 662        int ret;
 663
 664        parent = from;
 665
 666down:
 667        ret = (*down)(parent, data);
 668        if (ret)
 669                goto out;
 670        list_for_each_entry_rcu(child, &parent->children, siblings) {
 671                parent = child;
 672                goto down;
 673
 674up:
 675                continue;
 676        }
 677        ret = (*up)(parent, data);
 678        if (ret || parent == from)
 679                goto out;
 680
 681        child = parent;
 682        parent = parent->parent;
 683        if (parent)
 684                goto up;
 685out:
 686        return ret;
 687}
 688
 689int tg_nop(struct task_group *tg, void *data)
 690{
 691        return 0;
 692}
 693#endif
 694
 695static void set_load_weight(struct task_struct *p)
 696{
 697        int prio = p->static_prio - MAX_RT_PRIO;
 698        struct load_weight *load = &p->se.load;
 699
 700        /*
 701         * SCHED_IDLE tasks get minimal weight:
 702         */
 703        if (p->policy == SCHED_IDLE) {
 704                load->weight = scale_load(WEIGHT_IDLEPRIO);
 705                load->inv_weight = WMULT_IDLEPRIO;
 706                return;
 707        }
 708
 709        load->weight = scale_load(prio_to_weight[prio]);
 710        load->inv_weight = prio_to_wmult[prio];
 711}
 712
 713static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
 714{
 715        update_rq_clock(rq);
 716        sched_info_queued(p);
 717        p->sched_class->enqueue_task(rq, p, flags);
 718}
 719
 720static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
 721{
 722        update_rq_clock(rq);
 723        sched_info_dequeued(p);
 724        p->sched_class->dequeue_task(rq, p, flags);
 725}
 726
 727void activate_task(struct rq *rq, struct task_struct *p, int flags)
 728{
 729        if (task_contributes_to_load(p))
 730                rq->nr_uninterruptible--;
 731
 732        enqueue_task(rq, p, flags);
 733}
 734
 735void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
 736{
 737        if (task_contributes_to_load(p))
 738                rq->nr_uninterruptible++;
 739
 740        dequeue_task(rq, p, flags);
 741}
 742
 743static void update_rq_clock_task(struct rq *rq, s64 delta)
 744{
 745/*
 746 * In theory, the compile should just see 0 here, and optimize out the call
 747 * to sched_rt_avg_update. But I don't trust it...
 748 */
 749#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
 750        s64 steal = 0, irq_delta = 0;
 751#endif
 752#ifdef CONFIG_IRQ_TIME_ACCOUNTING
 753        irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
 754
 755        /*
 756         * Since irq_time is only updated on {soft,}irq_exit, we might run into
 757         * this case when a previous update_rq_clock() happened inside a
 758         * {soft,}irq region.
 759         *
 760         * When this happens, we stop ->clock_task and only update the
 761         * prev_irq_time stamp to account for the part that fit, so that a next
 762         * update will consume the rest. This ensures ->clock_task is
 763         * monotonic.
 764         *
 765         * It does however cause some slight miss-attribution of {soft,}irq
 766         * time, a more accurate solution would be to update the irq_time using
 767         * the current rq->clock timestamp, except that would require using
 768         * atomic ops.
 769         */
 770        if (irq_delta > delta)
 771                irq_delta = delta;
 772
 773        rq->prev_irq_time += irq_delta;
 774        delta -= irq_delta;
 775#endif
 776#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
 777        if (static_key_false((&paravirt_steal_rq_enabled))) {
 778                u64 st;
 779
 780                steal = paravirt_steal_clock(cpu_of(rq));
 781                steal -= rq->prev_steal_time_rq;
 782
 783                if (unlikely(steal > delta))
 784                        steal = delta;
 785
 786                st = steal_ticks(steal);
 787                steal = st * TICK_NSEC;
 788
 789                rq->prev_steal_time_rq += steal;
 790
 791                delta -= steal;
 792        }
 793#endif
 794
 795        rq->clock_task += delta;
 796
 797#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
 798        if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
 799                sched_rt_avg_update(rq, irq_delta + steal);
 800#endif
 801}
 802
 803void sched_set_stop_task(int cpu, struct task_struct *stop)
 804{
 805        struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
 806        struct task_struct *old_stop = cpu_rq(cpu)->stop;
 807
 808        if (stop) {
 809                /*
 810                 * Make it appear like a SCHED_FIFO task, its something
 811                 * userspace knows about and won't get confused about.
 812                 *
 813                 * Also, it will make PI more or less work without too
 814                 * much confusion -- but then, stop work should not
 815                 * rely on PI working anyway.
 816                 */
 817                sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
 818
 819                stop->sched_class = &stop_sched_class;
 820        }
 821
 822        cpu_rq(cpu)->stop = stop;
 823
 824        if (old_stop) {
 825                /*
 826                 * Reset it back to a normal scheduling class so that
 827                 * it can die in pieces.
 828                 */
 829                old_stop->sched_class = &rt_sched_class;
 830        }
 831}
 832
 833/*
 834 * __normal_prio - return the priority that is based on the static prio
 835 */
 836static inline int __normal_prio(struct task_struct *p)
 837{
 838        return p->static_prio;
 839}
 840
 841/*
 842 * Calculate the expected normal priority: i.e. priority
 843 * without taking RT-inheritance into account. Might be
 844 * boosted by interactivity modifiers. Changes upon fork,
 845 * setprio syscalls, and whenever the interactivity
 846 * estimator recalculates.
 847 */
 848static inline int normal_prio(struct task_struct *p)
 849{
 850        int prio;
 851
 852        if (task_has_rt_policy(p))
 853                prio = MAX_RT_PRIO-1 - p->rt_priority;
 854        else
 855                prio = __normal_prio(p);
 856        return prio;
 857}
 858
 859/*
 860 * Calculate the current priority, i.e. the priority
 861 * taken into account by the scheduler. This value might
 862 * be boosted by RT tasks, or might be boosted by
 863 * interactivity modifiers. Will be RT if the task got
 864 * RT-boosted. If not then it returns p->normal_prio.
 865 */
 866static int effective_prio(struct task_struct *p)
 867{
 868        p->normal_prio = normal_prio(p);
 869        /*
 870         * If we are RT tasks or we were boosted to RT priority,
 871         * keep the priority unchanged. Otherwise, update priority
 872         * to the normal priority:
 873         */
 874        if (!rt_prio(p->prio))
 875                return p->normal_prio;
 876        return p->prio;
 877}
 878
 879/**
 880 * task_curr - is this task currently executing on a CPU?
 881 * @p: the task in question.
 882 */
 883inline int task_curr(const struct task_struct *p)
 884{
 885        return cpu_curr(task_cpu(p)) == p;
 886}
 887
 888static inline void check_class_changed(struct rq *rq, struct task_struct *p,
 889                                       const struct sched_class *prev_class,
 890                                       int oldprio)
 891{
 892        if (prev_class != p->sched_class) {
 893                if (prev_class->switched_from)
 894                        prev_class->switched_from(rq, p);
 895                p->sched_class->switched_to(rq, p);
 896        } else if (oldprio != p->prio)
 897                p->sched_class->prio_changed(rq, p, oldprio);
 898}
 899
 900void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
 901{
 902        const struct sched_class *class;
 903
 904        if (p->sched_class == rq->curr->sched_class) {
 905                rq->curr->sched_class->check_preempt_curr(rq, p, flags);
 906        } else {
 907                for_each_class(class) {
 908                        if (class == rq->curr->sched_class)
 909                                break;
 910                        if (class == p->sched_class) {
 911                                resched_task(rq->curr);
 912                                break;
 913                        }
 914                }
 915        }
 916
 917        /*
 918         * A queue event has occurred, and we're going to schedule.  In
 919         * this case, we can save a useless back to back clock update.
 920         */
 921        if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
 922                rq->skip_clock_update = 1;
 923}
 924
 925#ifdef CONFIG_SMP
 926void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
 927{
 928#ifdef CONFIG_SCHED_DEBUG
 929        /*
 930         * We should never call set_task_cpu() on a blocked task,
 931         * ttwu() will sort out the placement.
 932         */
 933        WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
 934                        !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
 935
 936#ifdef CONFIG_LOCKDEP
 937        /*
 938         * The caller should hold either p->pi_lock or rq->lock, when changing
 939         * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
 940         *
 941         * sched_move_task() holds both and thus holding either pins the cgroup,
 942         * see task_group().
 943         *
 944         * Furthermore, all task_rq users should acquire both locks, see
 945         * task_rq_lock().
 946         */
 947        WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
 948                                      lockdep_is_held(&task_rq(p)->lock)));
 949#endif
 950#endif
 951
 952        trace_sched_migrate_task(p, new_cpu);
 953
 954        if (task_cpu(p) != new_cpu) {
 955                p->se.nr_migrations++;
 956                perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
 957        }
 958
 959        __set_task_cpu(p, new_cpu);
 960}
 961
 962struct migration_arg {
 963        struct task_struct *task;
 964        int dest_cpu;
 965};
 966
 967static int migration_cpu_stop(void *data);
 968
 969/*
 970 * wait_task_inactive - wait for a thread to unschedule.
 971 *
 972 * If @match_state is nonzero, it's the @p->state value just checked and
 973 * not expected to change.  If it changes, i.e. @p might have woken up,
 974 * then return zero.  When we succeed in waiting for @p to be off its CPU,
 975 * we return a positive number (its total switch count).  If a second call
 976 * a short while later returns the same number, the caller can be sure that
 977 * @p has remained unscheduled the whole time.
 978 *
 979 * The caller must ensure that the task *will* unschedule sometime soon,
 980 * else this function might spin for a *long* time. This function can't
 981 * be called with interrupts off, or it may introduce deadlock with
 982 * smp_call_function() if an IPI is sent by the same process we are
 983 * waiting to become inactive.
 984 */
 985unsigned long wait_task_inactive(struct task_struct *p, long match_state)
 986{
 987        unsigned long flags;
 988        int running, on_rq;
 989        unsigned long ncsw;
 990        struct rq *rq;
 991
 992        for (;;) {
 993                /*
 994                 * We do the initial early heuristics without holding
 995                 * any task-queue locks at all. We'll only try to get
 996                 * the runqueue lock when things look like they will
 997                 * work out!
 998                 */
 999                rq = task_rq(p);
1000
1001                /*
1002                 * If the task is actively running on another CPU
1003                 * still, just relax and busy-wait without holding
1004                 * any locks.
1005                 *
1006                 * NOTE! Since we don't hold any locks, it's not
1007                 * even sure that "rq" stays as the right runqueue!
1008                 * But we don't care, since "task_running()" will
1009                 * return false if the runqueue has changed and p
1010                 * is actually now running somewhere else!
1011                 */
1012                while (task_running(rq, p)) {
1013                        if (match_state && unlikely(p->state != match_state))
1014                                return 0;
1015                        cpu_relax();
1016                }
1017
1018                /*
1019                 * Ok, time to look more closely! We need the rq
1020                 * lock now, to be *sure*. If we're wrong, we'll
1021                 * just go back and repeat.
1022                 */
1023                rq = task_rq_lock(p, &flags);
1024                trace_sched_wait_task(p);
1025                running = task_running(rq, p);
1026                on_rq = p->on_rq;
1027                ncsw = 0;
1028                if (!match_state || p->state == match_state)
1029                        ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1030                task_rq_unlock(rq, p, &flags);
1031
1032                /*
1033                 * If it changed from the expected state, bail out now.
1034                 */
1035                if (unlikely(!ncsw))
1036                        break;
1037
1038                /*
1039                 * Was it really running after all now that we
1040                 * checked with the proper locks actually held?
1041                 *
1042                 * Oops. Go back and try again..
1043                 */
1044                if (unlikely(running)) {
1045                        cpu_relax();
1046                        continue;
1047                }
1048
1049                /*
1050                 * It's not enough that it's not actively running,
1051                 * it must be off the runqueue _entirely_, and not
1052                 * preempted!
1053                 *
1054                 * So if it was still runnable (but just not actively
1055                 * running right now), it's preempted, and we should
1056                 * yield - it could be a while.
1057                 */
1058                if (unlikely(on_rq)) {
1059                        ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1060
1061                        set_current_state(TASK_UNINTERRUPTIBLE);
1062                        schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1063                        continue;
1064                }
1065
1066                /*
1067                 * Ahh, all good. It wasn't running, and it wasn't
1068                 * runnable, which means that it will never become
1069                 * running in the future either. We're all done!
1070                 */
1071                break;
1072        }
1073
1074        return ncsw;
1075}
1076
1077/***
1078 * kick_process - kick a running thread to enter/exit the kernel
1079 * @p: the to-be-kicked thread
1080 *
1081 * Cause a process which is running on another CPU to enter
1082 * kernel-mode, without any delay. (to get signals handled.)
1083 *
1084 * NOTE: this function doesn't have to take the runqueue lock,
1085 * because all it wants to ensure is that the remote task enters
1086 * the kernel. If the IPI races and the task has been migrated
1087 * to another CPU then no harm is done and the purpose has been
1088 * achieved as well.
1089 */
1090void kick_process(struct task_struct *p)
1091{
1092        int cpu;
1093
1094        preempt_disable();
1095        cpu = task_cpu(p);
1096        if ((cpu != smp_processor_id()) && task_curr(p))
1097                smp_send_reschedule(cpu);
1098        preempt_enable();
1099}
1100EXPORT_SYMBOL_GPL(kick_process);
1101#endif /* CONFIG_SMP */
1102
1103#ifdef CONFIG_SMP
1104/*
1105 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1106 */
1107static int select_fallback_rq(int cpu, struct task_struct *p)
1108{
1109        const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1110        enum { cpuset, possible, fail } state = cpuset;
1111        int dest_cpu;
1112
1113        /* Look for allowed, online CPU in same node. */
1114        for_each_cpu(dest_cpu, nodemask) {
1115                if (!cpu_online(dest_cpu))
1116                        continue;
1117                if (!cpu_active(dest_cpu))
1118                        continue;
1119                if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1120                        return dest_cpu;
1121        }
1122
1123        for (;;) {
1124                /* Any allowed, online CPU? */
1125                for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1126                        if (!cpu_online(dest_cpu))
1127                                continue;
1128                        if (!cpu_active(dest_cpu))
1129                                continue;
1130                        goto out;
1131                }
1132
1133                switch (state) {
1134                case cpuset:
1135                        /* No more Mr. Nice Guy. */
1136                        cpuset_cpus_allowed_fallback(p);
1137                        state = possible;
1138                        break;
1139
1140                case possible:
1141                        do_set_cpus_allowed(p, cpu_possible_mask);
1142                        state = fail;
1143                        break;
1144
1145                case fail:
1146                        BUG();
1147                        break;
1148                }
1149        }
1150
1151out:
1152        if (state != cpuset) {
1153                /*
1154                 * Don't tell them about moving exiting tasks or
1155                 * kernel threads (both mm NULL), since they never
1156                 * leave kernel.
1157                 */
1158                if (p->mm && printk_ratelimit()) {
1159                        printk_sched("process %d (%s) no longer affine to cpu%d\n",
1160                                        task_pid_nr(p), p->comm, cpu);
1161                }
1162        }
1163
1164        return dest_cpu;
1165}
1166
1167/*
1168 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1169 */
1170static inline
1171int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1172{
1173        int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1174
1175        /*
1176         * In order not to call set_task_cpu() on a blocking task we need
1177         * to rely on ttwu() to place the task on a valid ->cpus_allowed
1178         * cpu.
1179         *
1180         * Since this is common to all placement strategies, this lives here.
1181         *
1182         * [ this allows ->select_task() to simply return task_cpu(p) and
1183         *   not worry about this generic constraint ]
1184         */
1185        if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1186                     !cpu_online(cpu)))
1187                cpu = select_fallback_rq(task_cpu(p), p);
1188
1189        return cpu;
1190}
1191
1192static void update_avg(u64 *avg, u64 sample)
1193{
1194        s64 diff = sample - *avg;
1195        *avg += diff >> 3;
1196}
1197#endif
1198
1199static void
1200ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1201{
1202#ifdef CONFIG_SCHEDSTATS
1203        struct rq *rq = this_rq();
1204
1205#ifdef CONFIG_SMP
1206        int this_cpu = smp_processor_id();
1207
1208        if (cpu == this_cpu) {
1209                schedstat_inc(rq, ttwu_local);
1210                schedstat_inc(p, se.statistics.nr_wakeups_local);
1211        } else {
1212                struct sched_domain *sd;
1213
1214                schedstat_inc(p, se.statistics.nr_wakeups_remote);
1215                rcu_read_lock();
1216                for_each_domain(this_cpu, sd) {
1217                        if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1218                                schedstat_inc(sd, ttwu_wake_remote);
1219                                break;
1220                        }
1221                }
1222                rcu_read_unlock();
1223        }
1224
1225        if (wake_flags & WF_MIGRATED)
1226                schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1227
1228#endif /* CONFIG_SMP */
1229
1230        schedstat_inc(rq, ttwu_count);
1231        schedstat_inc(p, se.statistics.nr_wakeups);
1232
1233        if (wake_flags & WF_SYNC)
1234                schedstat_inc(p, se.statistics.nr_wakeups_sync);
1235
1236#endif /* CONFIG_SCHEDSTATS */
1237}
1238
1239static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1240{
1241        activate_task(rq, p, en_flags);
1242        p->on_rq = 1;
1243
1244        /* if a worker is waking up, notify workqueue */
1245        if (p->flags & PF_WQ_WORKER)
1246                wq_worker_waking_up(p, cpu_of(rq));
1247}
1248
1249/*
1250 * Mark the task runnable and perform wakeup-preemption.
1251 */
1252static void
1253ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1254{
1255        trace_sched_wakeup(p, true);
1256        check_preempt_curr(rq, p, wake_flags);
1257
1258        p->state = TASK_RUNNING;
1259#ifdef CONFIG_SMP
1260        if (p->sched_class->task_woken)
1261                p->sched_class->task_woken(rq, p);
1262
1263        if (rq->idle_stamp) {
1264                u64 delta = rq->clock - rq->idle_stamp;
1265                u64 max = 2*sysctl_sched_migration_cost;
1266
1267                if (delta > max)
1268                        rq->avg_idle = max;
1269                else
1270                        update_avg(&rq->avg_idle, delta);
1271                rq->idle_stamp = 0;
1272        }
1273#endif
1274}
1275
1276static void
1277ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1278{
1279#ifdef CONFIG_SMP
1280        if (p->sched_contributes_to_load)
1281                rq->nr_uninterruptible--;
1282#endif
1283
1284        ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1285        ttwu_do_wakeup(rq, p, wake_flags);
1286}
1287
1288/*
1289 * Called in case the task @p isn't fully descheduled from its runqueue,
1290 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1291 * since all we need to do is flip p->state to TASK_RUNNING, since
1292 * the task is still ->on_rq.
1293 */
1294static int ttwu_remote(struct task_struct *p, int wake_flags)
1295{
1296        struct rq *rq;
1297        int ret = 0;
1298
1299        rq = __task_rq_lock(p);
1300        if (p->on_rq) {
1301                ttwu_do_wakeup(rq, p, wake_flags);
1302                ret = 1;
1303        }
1304        __task_rq_unlock(rq);
1305
1306        return ret;
1307}
1308
1309#ifdef CONFIG_SMP
1310static void sched_ttwu_pending(void)
1311{
1312        struct rq *rq = this_rq();
1313        struct llist_node *llist = llist_del_all(&rq->wake_list);
1314        struct task_struct *p;
1315
1316        raw_spin_lock(&rq->lock);
1317
1318        while (llist) {
1319                p = llist_entry(llist, struct task_struct, wake_entry);
1320                llist = llist_next(llist);
1321                ttwu_do_activate(rq, p, 0);
1322        }
1323
1324        raw_spin_unlock(&rq->lock);
1325}
1326
1327void scheduler_ipi(void)
1328{
1329        if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1330                return;
1331
1332        /*
1333         * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1334         * traditionally all their work was done from the interrupt return
1335         * path. Now that we actually do some work, we need to make sure
1336         * we do call them.
1337         *
1338         * Some archs already do call them, luckily irq_enter/exit nest
1339         * properly.
1340         *
1341         * Arguably we should visit all archs and update all handlers,
1342         * however a fair share of IPIs are still resched only so this would
1343         * somewhat pessimize the simple resched case.
1344         */
1345        irq_enter();
1346        sched_ttwu_pending();
1347
1348        /*
1349         * Check if someone kicked us for doing the nohz idle load balance.
1350         */
1351        if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1352                this_rq()->idle_balance = 1;
1353                raise_softirq_irqoff(SCHED_SOFTIRQ);
1354        }
1355        irq_exit();
1356}
1357
1358static void ttwu_queue_remote(struct task_struct *p, int cpu)
1359{
1360        if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1361                smp_send_reschedule(cpu);
1362}
1363
1364bool cpus_share_cache(int this_cpu, int that_cpu)
1365{
1366        return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1367}
1368#endif /* CONFIG_SMP */
1369
1370static void ttwu_queue(struct task_struct *p, int cpu)
1371{
1372        struct rq *rq = cpu_rq(cpu);
1373
1374#if defined(CONFIG_SMP)
1375        if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1376                sched_clock_cpu(cpu); /* sync clocks x-cpu */
1377                ttwu_queue_remote(p, cpu);
1378                return;
1379        }
1380#endif
1381
1382        raw_spin_lock(&rq->lock);
1383        ttwu_do_activate(rq, p, 0);
1384        raw_spin_unlock(&rq->lock);
1385}
1386
1387/**
1388 * try_to_wake_up - wake up a thread
1389 * @p: the thread to be awakened
1390 * @state: the mask of task states that can be woken
1391 * @wake_flags: wake modifier flags (WF_*)
1392 *
1393 * Put it on the run-queue if it's not already there. The "current"
1394 * thread is always on the run-queue (except when the actual
1395 * re-schedule is in progress), and as such you're allowed to do
1396 * the simpler "current->state = TASK_RUNNING" to mark yourself
1397 * runnable without the overhead of this.
1398 *
1399 * Returns %true if @p was woken up, %false if it was already running
1400 * or @state didn't match @p's state.
1401 */
1402static int
1403try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1404{
1405        unsigned long flags;
1406        int cpu, success = 0;
1407
1408        smp_wmb();
1409        raw_spin_lock_irqsave(&p->pi_lock, flags);
1410        if (!(p->state & state))
1411                goto out;
1412
1413        success = 1; /* we're going to change ->state */
1414        cpu = task_cpu(p);
1415
1416        if (p->on_rq && ttwu_remote(p, wake_flags))
1417                goto stat;
1418
1419#ifdef CONFIG_SMP
1420        /*
1421         * If the owning (remote) cpu is still in the middle of schedule() with
1422         * this task as prev, wait until its done referencing the task.
1423         */
1424        while (p->on_cpu)
1425                cpu_relax();
1426        /*
1427         * Pairs with the smp_wmb() in finish_lock_switch().
1428         */
1429        smp_rmb();
1430
1431        p->sched_contributes_to_load = !!task_contributes_to_load(p);
1432        p->state = TASK_WAKING;
1433
1434        if (p->sched_class->task_waking)
1435                p->sched_class->task_waking(p);
1436
1437        cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1438        if (task_cpu(p) != cpu) {
1439                wake_flags |= WF_MIGRATED;
1440                set_task_cpu(p, cpu);
1441        }
1442#endif /* CONFIG_SMP */
1443
1444        ttwu_queue(p, cpu);
1445stat:
1446        ttwu_stat(p, cpu, wake_flags);
1447out:
1448        raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1449
1450        return success;
1451}
1452
1453/**
1454 * try_to_wake_up_local - try to wake up a local task with rq lock held
1455 * @p: the thread to be awakened
1456 *
1457 * Put @p on the run-queue if it's not already there. The caller must
1458 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1459 * the current task.
1460 */
1461static void try_to_wake_up_local(struct task_struct *p)
1462{
1463        struct rq *rq = task_rq(p);
1464
1465        BUG_ON(rq != this_rq());
1466        BUG_ON(p == current);
1467        lockdep_assert_held(&rq->lock);
1468
1469        if (!raw_spin_trylock(&p->pi_lock)) {
1470                raw_spin_unlock(&rq->lock);
1471                raw_spin_lock(&p->pi_lock);
1472                raw_spin_lock(&rq->lock);
1473        }
1474
1475        if (!(p->state & TASK_NORMAL))
1476                goto out;
1477
1478        if (!p->on_rq)
1479                ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1480
1481        ttwu_do_wakeup(rq, p, 0);
1482        ttwu_stat(p, smp_processor_id(), 0);
1483out:
1484        raw_spin_unlock(&p->pi_lock);
1485}
1486
1487/**
1488 * wake_up_process - Wake up a specific process
1489 * @p: The process to be woken up.
1490 *
1491 * Attempt to wake up the nominated process and move it to the set of runnable
1492 * processes.  Returns 1 if the process was woken up, 0 if it was already
1493 * running.
1494 *
1495 * It may be assumed that this function implies a write memory barrier before
1496 * changing the task state if and only if any tasks are woken up.
1497 */
1498int wake_up_process(struct task_struct *p)
1499{
1500        WARN_ON(task_is_stopped_or_traced(p));
1501        return try_to_wake_up(p, TASK_NORMAL, 0);
1502}
1503EXPORT_SYMBOL(wake_up_process);
1504
1505int wake_up_state(struct task_struct *p, unsigned int state)
1506{
1507        return try_to_wake_up(p, state, 0);
1508}
1509
1510/*
1511 * Perform scheduler related setup for a newly forked process p.
1512 * p is forked by current.
1513 *
1514 * __sched_fork() is basic setup used by init_idle() too:
1515 */
1516static void __sched_fork(struct task_struct *p)
1517{
1518        p->on_rq                        = 0;
1519
1520        p->se.on_rq                     = 0;
1521        p->se.exec_start                = 0;
1522        p->se.sum_exec_runtime          = 0;
1523        p->se.prev_sum_exec_runtime     = 0;
1524        p->se.nr_migrations             = 0;
1525        p->se.vruntime                  = 0;
1526        INIT_LIST_HEAD(&p->se.group_node);
1527
1528#ifdef CONFIG_SCHEDSTATS
1529        memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1530#endif
1531
1532        INIT_LIST_HEAD(&p->rt.run_list);
1533
1534#ifdef CONFIG_PREEMPT_NOTIFIERS
1535        INIT_HLIST_HEAD(&p->preempt_notifiers);
1536#endif
1537}
1538
1539/*
1540 * fork()/clone()-time setup:
1541 */
1542void sched_fork(struct task_struct *p)
1543{
1544        unsigned long flags;
1545        int cpu = get_cpu();
1546
1547        __sched_fork(p);
1548        /*
1549         * We mark the process as running here. This guarantees that
1550         * nobody will actually run it, and a signal or other external
1551         * event cannot wake it up and insert it on the runqueue either.
1552         */
1553        p->state = TASK_RUNNING;
1554
1555        /*
1556         * Make sure we do not leak PI boosting priority to the child.
1557         */
1558        p->prio = current->normal_prio;
1559
1560        /*
1561         * Revert to default priority/policy on fork if requested.
1562         */
1563        if (unlikely(p->sched_reset_on_fork)) {
1564                if (task_has_rt_policy(p)) {
1565                        p->policy = SCHED_NORMAL;
1566                        p->static_prio = NICE_TO_PRIO(0);
1567                        p->rt_priority = 0;
1568                } else if (PRIO_TO_NICE(p->static_prio) < 0)
1569                        p->static_prio = NICE_TO_PRIO(0);
1570
1571                p->prio = p->normal_prio = __normal_prio(p);
1572                set_load_weight(p);
1573
1574                /*
1575                 * We don't need the reset flag anymore after the fork. It has
1576                 * fulfilled its duty:
1577                 */
1578                p->sched_reset_on_fork = 0;
1579        }
1580
1581        if (!rt_prio(p->prio))
1582                p->sched_class = &fair_sched_class;
1583
1584        if (p->sched_class->task_fork)
1585                p->sched_class->task_fork(p);
1586
1587        /*
1588         * The child is not yet in the pid-hash so no cgroup attach races,
1589         * and the cgroup is pinned to this child due to cgroup_fork()
1590         * is ran before sched_fork().
1591         *
1592         * Silence PROVE_RCU.
1593         */
1594        raw_spin_lock_irqsave(&p->pi_lock, flags);
1595        set_task_cpu(p, cpu);
1596        raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1597
1598#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1599        if (likely(sched_info_on()))
1600                memset(&p->sched_info, 0, sizeof(p->sched_info));
1601#endif
1602#if defined(CONFIG_SMP)
1603        p->on_cpu = 0;
1604#endif
1605#ifdef CONFIG_PREEMPT_COUNT
1606        /* Want to start with kernel preemption disabled. */
1607        task_thread_info(p)->preempt_count = 1;
1608#endif
1609#ifdef CONFIG_SMP
1610        plist_node_init(&p->pushable_tasks, MAX_PRIO);
1611#endif
1612
1613        put_cpu();
1614}
1615
1616/*
1617 * wake_up_new_task - wake up a newly created task for the first time.
1618 *
1619 * This function will do some initial scheduler statistics housekeeping
1620 * that must be done for every newly created context, then puts the task
1621 * on the runqueue and wakes it.
1622 */
1623void wake_up_new_task(struct task_struct *p)
1624{
1625        unsigned long flags;
1626        struct rq *rq;
1627
1628        raw_spin_lock_irqsave(&p->pi_lock, flags);
1629#ifdef CONFIG_SMP
1630        /*
1631         * Fork balancing, do it here and not earlier because:
1632         *  - cpus_allowed can change in the fork path
1633         *  - any previously selected cpu might disappear through hotplug
1634         */
1635        set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1636#endif
1637
1638        rq = __task_rq_lock(p);
1639        activate_task(rq, p, 0);
1640        p->on_rq = 1;
1641        trace_sched_wakeup_new(p, true);
1642        check_preempt_curr(rq, p, WF_FORK);
1643#ifdef CONFIG_SMP
1644        if (p->sched_class->task_woken)
1645                p->sched_class->task_woken(rq, p);
1646#endif
1647        task_rq_unlock(rq, p, &flags);
1648}
1649
1650#ifdef CONFIG_PREEMPT_NOTIFIERS
1651
1652/**
1653 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1654 * @notifier: notifier struct to register
1655 */
1656void preempt_notifier_register(struct preempt_notifier *notifier)
1657{
1658        hlist_add_head(&notifier->link, &current->preempt_notifiers);
1659}
1660EXPORT_SYMBOL_GPL(preempt_notifier_register);
1661
1662/**
1663 * preempt_notifier_unregister - no longer interested in preemption notifications
1664 * @notifier: notifier struct to unregister
1665 *
1666 * This is safe to call from within a preemption notifier.
1667 */
1668void preempt_notifier_unregister(struct preempt_notifier *notifier)
1669{
1670        hlist_del(&notifier->link);
1671}
1672EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1673
1674static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1675{
1676        struct preempt_notifier *notifier;
1677        struct hlist_node *node;
1678
1679        hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1680                notifier->ops->sched_in(notifier, raw_smp_processor_id());
1681}
1682
1683static void
1684fire_sched_out_preempt_notifiers(struct task_struct *curr,
1685                                 struct task_struct *next)
1686{
1687        struct preempt_notifier *notifier;
1688        struct hlist_node *node;
1689
1690        hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1691                notifier->ops->sched_out(notifier, next);
1692}
1693
1694#else /* !CONFIG_PREEMPT_NOTIFIERS */
1695
1696static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1697{
1698}
1699
1700static void
1701fire_sched_out_preempt_notifiers(struct task_struct *curr,
1702                                 struct task_struct *next)
1703{
1704}
1705
1706#endif /* CONFIG_PREEMPT_NOTIFIERS */
1707
1708/**
1709 * prepare_task_switch - prepare to switch tasks
1710 * @rq: the runqueue preparing to switch
1711 * @prev: the current task that is being switched out
1712 * @next: the task we are going to switch to.
1713 *
1714 * This is called with the rq lock held and interrupts off. It must
1715 * be paired with a subsequent finish_task_switch after the context
1716 * switch.
1717 *
1718 * prepare_task_switch sets up locking and calls architecture specific
1719 * hooks.
1720 */
1721static inline void
1722prepare_task_switch(struct rq *rq, struct task_struct *prev,
1723                    struct task_struct *next)
1724{
1725        trace_sched_switch(prev, next);
1726        sched_info_switch(prev, next);
1727        perf_event_task_sched_out(prev, next);
1728        fire_sched_out_preempt_notifiers(prev, next);
1729        prepare_lock_switch(rq, next);
1730        prepare_arch_switch(next);
1731}
1732
1733/**
1734 * finish_task_switch - clean up after a task-switch
1735 * @rq: runqueue associated with task-switch
1736 * @prev: the thread we just switched away from.
1737 *
1738 * finish_task_switch must be called after the context switch, paired
1739 * with a prepare_task_switch call before the context switch.
1740 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1741 * and do any other architecture-specific cleanup actions.
1742 *
1743 * Note that we may have delayed dropping an mm in context_switch(). If
1744 * so, we finish that here outside of the runqueue lock. (Doing it
1745 * with the lock held can cause deadlocks; see schedule() for
1746 * details.)
1747 */
1748static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1749        __releases(rq->lock)
1750{
1751        struct mm_struct *mm = rq->prev_mm;
1752        long prev_state;
1753
1754        rq->prev_mm = NULL;
1755
1756        /*
1757         * A task struct has one reference for the use as "current".
1758         * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1759         * schedule one last time. The schedule call will never return, and
1760         * the scheduled task must drop that reference.
1761         * The test for TASK_DEAD must occur while the runqueue locks are
1762         * still held, otherwise prev could be scheduled on another cpu, die
1763         * there before we look at prev->state, and then the reference would
1764         * be dropped twice.
1765         *              Manfred Spraul <manfred@colorfullife.com>
1766         */
1767        prev_state = prev->state;
1768        vtime_task_switch(prev);
1769        finish_arch_switch(prev);
1770        perf_event_task_sched_in(prev, current);
1771        finish_lock_switch(rq, prev);
1772        finish_arch_post_lock_switch();
1773
1774        fire_sched_in_preempt_notifiers(current);
1775        if (mm)
1776                mmdrop(mm);
1777        if (unlikely(prev_state == TASK_DEAD)) {
1778                /*
1779                 * Remove function-return probe instances associated with this
1780                 * task and put them back on the free list.
1781                 */
1782                kprobe_flush_task(prev);
1783                put_task_struct(prev);
1784        }
1785}
1786
1787#ifdef CONFIG_SMP
1788
1789/* assumes rq->lock is held */
1790static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1791{
1792        if (prev->sched_class->pre_schedule)
1793                prev->sched_class->pre_schedule(rq, prev);
1794}
1795
1796/* rq->lock is NOT held, but preemption is disabled */
1797static inline void post_schedule(struct rq *rq)
1798{
1799        if (rq->post_schedule) {
1800                unsigned long flags;
1801
1802                raw_spin_lock_irqsave(&rq->lock, flags);
1803                if (rq->curr->sched_class->post_schedule)
1804                        rq->curr->sched_class->post_schedule(rq);
1805                raw_spin_unlock_irqrestore(&rq->lock, flags);
1806
1807                rq->post_schedule = 0;
1808        }
1809}
1810
1811#else
1812
1813static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1814{
1815}
1816
1817static inline void post_schedule(struct rq *rq)
1818{
1819}
1820
1821#endif
1822
1823/**
1824 * schedule_tail - first thing a freshly forked thread must call.
1825 * @prev: the thread we just switched away from.
1826 */
1827asmlinkage void schedule_tail(struct task_struct *prev)
1828        __releases(rq->lock)
1829{
1830        struct rq *rq = this_rq();
1831
1832        finish_task_switch(rq, prev);
1833
1834        /*
1835         * FIXME: do we need to worry about rq being invalidated by the
1836         * task_switch?
1837         */
1838        post_schedule(rq);
1839
1840#ifdef __ARCH_WANT_UNLOCKED_CTXSW
1841        /* In this case, finish_task_switch does not reenable preemption */
1842        preempt_enable();
1843#endif
1844        if (current->set_child_tid)
1845                put_user(task_pid_vnr(current), current->set_child_tid);
1846}
1847
1848/*
1849 * context_switch - switch to the new MM and the new
1850 * thread's register state.
1851 */
1852static inline void
1853context_switch(struct rq *rq, struct task_struct *prev,
1854               struct task_struct *next)
1855{
1856        struct mm_struct *mm, *oldmm;
1857
1858        prepare_task_switch(rq, prev, next);
1859
1860        mm = next->mm;
1861        oldmm = prev->active_mm;
1862        /*
1863         * For paravirt, this is coupled with an exit in switch_to to
1864         * combine the page table reload and the switch backend into
1865         * one hypercall.
1866         */
1867        arch_start_context_switch(prev);
1868
1869        if (!mm) {
1870                next->active_mm = oldmm;
1871                atomic_inc(&oldmm->mm_count);
1872                enter_lazy_tlb(oldmm, next);
1873        } else
1874                switch_mm(oldmm, mm, next);
1875
1876        if (!prev->mm) {
1877                prev->active_mm = NULL;
1878                rq->prev_mm = oldmm;
1879        }
1880        /*
1881         * Since the runqueue lock will be released by the next
1882         * task (which is an invalid locking op but in the case
1883         * of the scheduler it's an obvious special-case), so we
1884         * do an early lockdep release here:
1885         */
1886#ifndef __ARCH_WANT_UNLOCKED_CTXSW
1887        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1888#endif
1889
1890        /* Here we just switch the register state and the stack. */
1891        rcu_switch(prev, next);
1892        switch_to(prev, next, prev);
1893
1894        barrier();
1895        /*
1896         * this_rq must be evaluated again because prev may have moved
1897         * CPUs since it called schedule(), thus the 'rq' on its stack
1898         * frame will be invalid.
1899         */
1900        finish_task_switch(this_rq(), prev);
1901}
1902
1903/*
1904 * nr_running, nr_uninterruptible and nr_context_switches:
1905 *
1906 * externally visible scheduler statistics: current number of runnable
1907 * threads, current number of uninterruptible-sleeping threads, total
1908 * number of context switches performed since bootup.
1909 */
1910unsigned long nr_running(void)
1911{
1912        unsigned long i, sum = 0;
1913
1914        for_each_online_cpu(i)
1915                sum += cpu_rq(i)->nr_running;
1916
1917        return sum;
1918}
1919
1920unsigned long nr_uninterruptible(void)
1921{
1922        unsigned long i, sum = 0;
1923
1924        for_each_possible_cpu(i)
1925                sum += cpu_rq(i)->nr_uninterruptible;
1926
1927        /*
1928         * Since we read the counters lockless, it might be slightly
1929         * inaccurate. Do not allow it to go below zero though:
1930         */
1931        if (unlikely((long)sum < 0))
1932                sum = 0;
1933
1934        return sum;
1935}
1936
1937unsigned long long nr_context_switches(void)
1938{
1939        int i;
1940        unsigned long long sum = 0;
1941
1942        for_each_possible_cpu(i)
1943                sum += cpu_rq(i)->nr_switches;
1944
1945        return sum;
1946}
1947
1948unsigned long nr_iowait(void)
1949{
1950        unsigned long i, sum = 0;
1951
1952        for_each_possible_cpu(i)
1953                sum += atomic_read(&cpu_rq(i)->nr_iowait);
1954
1955        return sum;
1956}
1957
1958unsigned long nr_iowait_cpu(int cpu)
1959{
1960        struct rq *this = cpu_rq(cpu);
1961        return atomic_read(&this->nr_iowait);
1962}
1963
1964unsigned long this_cpu_load(void)
1965{
1966        struct rq *this = this_rq();
1967        return this->cpu_load[0];
1968}
1969
1970
1971/*
1972 * Global load-average calculations
1973 *
1974 * We take a distributed and async approach to calculating the global load-avg
1975 * in order to minimize overhead.
1976 *
1977 * The global load average is an exponentially decaying average of nr_running +
1978 * nr_uninterruptible.
1979 *
1980 * Once every LOAD_FREQ:
1981 *
1982 *   nr_active = 0;
1983 *   for_each_possible_cpu(cpu)
1984 *      nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
1985 *
1986 *   avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
1987 *
1988 * Due to a number of reasons the above turns in the mess below:
1989 *
1990 *  - for_each_possible_cpu() is prohibitively expensive on machines with
1991 *    serious number of cpus, therefore we need to take a distributed approach
1992 *    to calculating nr_active.
1993 *
1994 *        \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
1995 *                      = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
1996 *
1997 *    So assuming nr_active := 0 when we start out -- true per definition, we
1998 *    can simply take per-cpu deltas and fold those into a global accumulate
1999 *    to obtain the same result. See calc_load_fold_active().
2000 *
2001 *    Furthermore, in order to avoid synchronizing all per-cpu delta folding
2002 *    across the machine, we assume 10 ticks is sufficient time for every
2003 *    cpu to have completed this task.
2004 *
2005 *    This places an upper-bound on the IRQ-off latency of the machine. Then
2006 *    again, being late doesn't loose the delta, just wrecks the sample.
2007 *
2008 *  - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2009 *    this would add another cross-cpu cacheline miss and atomic operation
2010 *    to the wakeup path. Instead we increment on whatever cpu the task ran
2011 *    when it went into uninterruptible state and decrement on whatever cpu
2012 *    did the wakeup. This means that only the sum of nr_uninterruptible over
2013 *    all cpus yields the correct result.
2014 *
2015 *  This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2016 */
2017
2018/* Variables and functions for calc_load */
2019static atomic_long_t calc_load_tasks;
2020static unsigned long calc_load_update;
2021unsigned long avenrun[3];
2022EXPORT_SYMBOL(avenrun); /* should be removed */
2023
2024/**
2025 * get_avenrun - get the load average array
2026 * @loads:      pointer to dest load array
2027 * @offset:     offset to add
2028 * @shift:      shift count to shift the result left
2029 *
2030 * These values are estimates at best, so no need for locking.
2031 */
2032void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2033{
2034        loads[0] = (avenrun[0] + offset) << shift;
2035        loads[1] = (avenrun[1] + offset) << shift;
2036        loads[2] = (avenrun[2] + offset) << shift;
2037}
2038
2039static long calc_load_fold_active(struct rq *this_rq)
2040{
2041        long nr_active, delta = 0;
2042
2043        nr_active = this_rq->nr_running;
2044        nr_active += (long) this_rq->nr_uninterruptible;
2045
2046        if (nr_active != this_rq->calc_load_active) {
2047                delta = nr_active - this_rq->calc_load_active;
2048                this_rq->calc_load_active = nr_active;
2049        }
2050
2051        return delta;
2052}
2053
2054/*
2055 * a1 = a0 * e + a * (1 - e)
2056 */
2057static unsigned long
2058calc_load(unsigned long load, unsigned long exp, unsigned long active)
2059{
2060        load *= exp;
2061        load += active * (FIXED_1 - exp);
2062        load += 1UL << (FSHIFT - 1);
2063        return load >> FSHIFT;
2064}
2065
2066#ifdef CONFIG_NO_HZ
2067/*
2068 * Handle NO_HZ for the global load-average.
2069 *
2070 * Since the above described distributed algorithm to compute the global
2071 * load-average relies on per-cpu sampling from the tick, it is affected by
2072 * NO_HZ.
2073 *
2074 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2075 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2076 * when we read the global state.
2077 *
2078 * Obviously reality has to ruin such a delightfully simple scheme:
2079 *
2080 *  - When we go NO_HZ idle during the window, we can negate our sample
2081 *    contribution, causing under-accounting.
2082 *
2083 *    We avoid this by keeping two idle-delta counters and flipping them
2084 *    when the window starts, thus separating old and new NO_HZ load.
2085 *
2086 *    The only trick is the slight shift in index flip for read vs write.
2087 *
2088 *        0s            5s            10s           15s
2089 *          +10           +10           +10           +10
2090 *        |-|-----------|-|-----------|-|-----------|-|
2091 *    r:0 0 1           1 0           0 1           1 0
2092 *    w:0 1 1           0 0           1 1           0 0
2093 *
2094 *    This ensures we'll fold the old idle contribution in this window while
2095 *    accumlating the new one.
2096 *
2097 *  - When we wake up from NO_HZ idle during the window, we push up our
2098 *    contribution, since we effectively move our sample point to a known
2099 *    busy state.
2100 *
2101 *    This is solved by pushing the window forward, and thus skipping the
2102 *    sample, for this cpu (effectively using the idle-delta for this cpu which
2103 *    was in effect at the time the window opened). This also solves the issue
2104 *    of having to deal with a cpu having been in NOHZ idle for multiple
2105 *    LOAD_FREQ intervals.
2106 *
2107 * When making the ILB scale, we should try to pull this in as well.
2108 */
2109static atomic_long_t calc_load_idle[2];
2110static int calc_load_idx;
2111
2112static inline int calc_load_write_idx(void)
2113{
2114        int idx = calc_load_idx;
2115
2116        /*
2117         * See calc_global_nohz(), if we observe the new index, we also
2118         * need to observe the new update time.
2119         */
2120        smp_rmb();
2121
2122        /*
2123         * If the folding window started, make sure we start writing in the
2124         * next idle-delta.
2125         */
2126        if (!time_before(jiffies, calc_load_update))
2127                idx++;
2128
2129        return idx & 1;
2130}
2131
2132static inline int calc_load_read_idx(void)
2133{
2134        return calc_load_idx & 1;
2135}
2136
2137void calc_load_enter_idle(void)
2138{
2139        struct rq *this_rq = this_rq();
2140        long delta;
2141
2142        /*
2143         * We're going into NOHZ mode, if there's any pending delta, fold it
2144         * into the pending idle delta.
2145         */
2146        delta = calc_load_fold_active(this_rq);
2147        if (delta) {
2148                int idx = calc_load_write_idx();
2149                atomic_long_add(delta, &calc_load_idle[idx]);
2150        }
2151}
2152
2153void calc_load_exit_idle(void)
2154{
2155        struct rq *this_rq = this_rq();
2156
2157        /*
2158         * If we're still before the sample window, we're done.
2159         */
2160        if (time_before(jiffies, this_rq->calc_load_update))
2161                return;
2162
2163        /*
2164         * We woke inside or after the sample window, this means we're already
2165         * accounted through the nohz accounting, so skip the entire deal and
2166         * sync up for the next window.
2167         */
2168        this_rq->calc_load_update = calc_load_update;
2169        if (time_before(jiffies, this_rq->calc_load_update + 10))
2170                this_rq->calc_load_update += LOAD_FREQ;
2171}
2172
2173static long calc_load_fold_idle(void)
2174{
2175        int idx = calc_load_read_idx();
2176        long delta = 0;
2177
2178        if (atomic_long_read(&calc_load_idle[idx]))
2179                delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2180
2181        return delta;
2182}
2183
2184/**
2185 * fixed_power_int - compute: x^n, in O(log n) time
2186 *
2187 * @x:         base of the power
2188 * @frac_bits: fractional bits of @x
2189 * @n:         power to raise @x to.
2190 *
2191 * By exploiting the relation between the definition of the natural power
2192 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2193 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2194 * (where: n_i \elem {0, 1}, the binary vector representing n),
2195 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2196 * of course trivially computable in O(log_2 n), the length of our binary
2197 * vector.
2198 */
2199static unsigned long
2200fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2201{
2202        unsigned long result = 1UL << frac_bits;
2203
2204        if (n) for (;;) {
2205                if (n & 1) {
2206                        result *= x;
2207                        result += 1UL << (frac_bits - 1);
2208                        result >>= frac_bits;
2209                }
2210                n >>= 1;
2211                if (!n)
2212                        break;
2213                x *= x;
2214                x += 1UL << (frac_bits - 1);
2215                x >>= frac_bits;
2216        }
2217
2218        return result;
2219}
2220
2221/*
2222 * a1 = a0 * e + a * (1 - e)
2223 *
2224 * a2 = a1 * e + a * (1 - e)
2225 *    = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2226 *    = a0 * e^2 + a * (1 - e) * (1 + e)
2227 *
2228 * a3 = a2 * e + a * (1 - e)
2229 *    = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2230 *    = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2231 *
2232 *  ...
2233 *
2234 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2235 *    = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2236 *    = a0 * e^n + a * (1 - e^n)
2237 *
2238 * [1] application of the geometric series:
2239 *
2240 *              n         1 - x^(n+1)
2241 *     S_n := \Sum x^i = -------------
2242 *             i=0          1 - x
2243 */
2244static unsigned long
2245calc_load_n(unsigned long load, unsigned long exp,
2246            unsigned long active, unsigned int n)
2247{
2248
2249        return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2250}
2251
2252/*
2253 * NO_HZ can leave us missing all per-cpu ticks calling
2254 * calc_load_account_active(), but since an idle CPU folds its delta into
2255 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2256 * in the pending idle delta if our idle period crossed a load cycle boundary.
2257 *
2258 * Once we've updated the global active value, we need to apply the exponential
2259 * weights adjusted to the number of cycles missed.
2260 */
2261static void calc_global_nohz(void)
2262{
2263        long delta, active, n;
2264
2265        if (!time_before(jiffies, calc_load_update + 10)) {
2266                /*
2267                 * Catch-up, fold however many we are behind still
2268                 */
2269                delta = jiffies - calc_load_update - 10;
2270                n = 1 + (delta / LOAD_FREQ);
2271
2272                active = atomic_long_read(&calc_load_tasks);
2273                active = active > 0 ? active * FIXED_1 : 0;
2274
2275                avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2276                avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2277                avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2278
2279                calc_load_update += n * LOAD_FREQ;
2280        }
2281
2282        /*
2283         * Flip the idle index...
2284         *
2285         * Make sure we first write the new time then flip the index, so that
2286         * calc_load_write_idx() will see the new time when it reads the new
2287         * index, this avoids a double flip messing things up.
2288         */
2289        smp_wmb();
2290        calc_load_idx++;
2291}
2292#else /* !CONFIG_NO_HZ */
2293
2294static inline long calc_load_fold_idle(void) { return 0; }
2295static inline void calc_global_nohz(void) { }
2296
2297#endif /* CONFIG_NO_HZ */
2298
2299/*
2300 * calc_load - update the avenrun load estimates 10 ticks after the
2301 * CPUs have updated calc_load_tasks.
2302 */
2303void calc_global_load(unsigned long ticks)
2304{
2305        long active, delta;
2306
2307        if (time_before(jiffies, calc_load_update + 10))
2308                return;
2309
2310        /*
2311         * Fold the 'old' idle-delta to include all NO_HZ cpus.
2312         */
2313        delta = calc_load_fold_idle();
2314        if (delta)
2315                atomic_long_add(delta, &calc_load_tasks);
2316
2317        active = atomic_long_read(&calc_load_tasks);
2318        active = active > 0 ? active * FIXED_1 : 0;
2319
2320        avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2321        avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2322        avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2323
2324        calc_load_update += LOAD_FREQ;
2325
2326        /*
2327         * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2328         */
2329        calc_global_nohz();
2330}
2331
2332/*
2333 * Called from update_cpu_load() to periodically update this CPU's
2334 * active count.
2335 */
2336static void calc_load_account_active(struct rq *this_rq)
2337{
2338        long delta;
2339
2340        if (time_before(jiffies, this_rq->calc_load_update))
2341                return;
2342
2343        delta  = calc_load_fold_active(this_rq);
2344        if (delta)
2345                atomic_long_add(delta, &calc_load_tasks);
2346
2347        this_rq->calc_load_update += LOAD_FREQ;
2348}
2349
2350/*
2351 * End of global load-average stuff
2352 */
2353
2354/*
2355 * The exact cpuload at various idx values, calculated at every tick would be
2356 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2357 *
2358 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2359 * on nth tick when cpu may be busy, then we have:
2360 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2361 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2362 *
2363 * decay_load_missed() below does efficient calculation of
2364 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2365 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2366 *
2367 * The calculation is approximated on a 128 point scale.
2368 * degrade_zero_ticks is the number of ticks after which load at any
2369 * particular idx is approximated to be zero.
2370 * degrade_factor is a precomputed table, a row for each load idx.
2371 * Each column corresponds to degradation factor for a power of two ticks,
2372 * based on 128 point scale.
2373 * Example:
2374 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2375 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2376 *
2377 * With this power of 2 load factors, we can degrade the load n times
2378 * by looking at 1 bits in n and doing as many mult/shift instead of
2379 * n mult/shifts needed by the exact degradation.
2380 */
2381#define DEGRADE_SHIFT           7
2382static const unsigned char
2383                degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2384static const unsigned char
2385                degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2386                                        {0, 0, 0, 0, 0, 0, 0, 0},
2387                                        {64, 32, 8, 0, 0, 0, 0, 0},
2388                                        {96, 72, 40, 12, 1, 0, 0},
2389                                        {112, 98, 75, 43, 15, 1, 0},
2390                                        {120, 112, 98, 76, 45, 16, 2} };
2391
2392/*
2393 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2394 * would be when CPU is idle and so we just decay the old load without
2395 * adding any new load.
2396 */
2397static unsigned long
2398decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2399{
2400        int j = 0;
2401
2402        if (!missed_updates)
2403                return load;
2404
2405        if (missed_updates >= degrade_zero_ticks[idx])
2406                return 0;
2407
2408        if (idx == 1)
2409                return load >> missed_updates;
2410
2411        while (missed_updates) {
2412                if (missed_updates % 2)
2413                        load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2414
2415                missed_updates >>= 1;
2416                j++;
2417        }
2418        return load;
2419}
2420
2421/*
2422 * Update rq->cpu_load[] statistics. This function is usually called every
2423 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2424 * every tick. We fix it up based on jiffies.
2425 */
2426static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2427                              unsigned long pending_updates)
2428{
2429        int i, scale;
2430
2431        this_rq->nr_load_updates++;
2432
2433        /* Update our load: */
2434        this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2435        for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2436                unsigned long old_load, new_load;
2437
2438                /* scale is effectively 1 << i now, and >> i divides by scale */
2439
2440                old_load = this_rq->cpu_load[i];
2441                old_load = decay_load_missed(old_load, pending_updates - 1, i);
2442                new_load = this_load;
2443                /*
2444                 * Round up the averaging division if load is increasing. This
2445                 * prevents us from getting stuck on 9 if the load is 10, for
2446                 * example.
2447                 */
2448                if (new_load > old_load)
2449                        new_load += scale - 1;
2450
2451                this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2452        }
2453
2454        sched_avg_update(this_rq);
2455}
2456
2457#ifdef CONFIG_NO_HZ
2458/*
2459 * There is no sane way to deal with nohz on smp when using jiffies because the
2460 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2461 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2462 *
2463 * Therefore we cannot use the delta approach from the regular tick since that
2464 * would seriously skew the load calculation. However we'll make do for those
2465 * updates happening while idle (nohz_idle_balance) or coming out of idle
2466 * (tick_nohz_idle_exit).
2467 *
2468 * This means we might still be one tick off for nohz periods.
2469 */
2470
2471/*
2472 * Called from nohz_idle_balance() to update the load ratings before doing the
2473 * idle balance.
2474 */
2475void update_idle_cpu_load(struct rq *this_rq)
2476{
2477        unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2478        unsigned long load = this_rq->load.weight;
2479        unsigned long pending_updates;
2480
2481        /*
2482         * bail if there's load or we're actually up-to-date.
2483         */
2484        if (load || curr_jiffies == this_rq->last_load_update_tick)
2485                return;
2486
2487        pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2488        this_rq->last_load_update_tick = curr_jiffies;
2489
2490        __update_cpu_load(this_rq, load, pending_updates);
2491}
2492
2493/*
2494 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2495 */
2496void update_cpu_load_nohz(void)
2497{
2498        struct rq *this_rq = this_rq();
2499        unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2500        unsigned long pending_updates;
2501
2502        if (curr_jiffies == this_rq->last_load_update_tick)
2503                return;
2504
2505        raw_spin_lock(&this_rq->lock);
2506        pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2507        if (pending_updates) {
2508                this_rq->last_load_update_tick = curr_jiffies;
2509                /*
2510                 * We were idle, this means load 0, the current load might be
2511                 * !0 due to remote wakeups and the sort.
2512                 */
2513                __update_cpu_load(this_rq, 0, pending_updates);
2514        }
2515        raw_spin_unlock(&this_rq->lock);
2516}
2517#endif /* CONFIG_NO_HZ */
2518
2519/*
2520 * Called from scheduler_tick()
2521 */
2522static void update_cpu_load_active(struct rq *this_rq)
2523{
2524        /*
2525         * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2526         */
2527        this_rq->last_load_update_tick = jiffies;
2528        __update_cpu_load(this_rq, this_rq->load.weight, 1);
2529
2530        calc_load_account_active(this_rq);
2531}
2532
2533#ifdef CONFIG_SMP
2534
2535/*
2536 * sched_exec - execve() is a valuable balancing opportunity, because at
2537 * this point the task has the smallest effective memory and cache footprint.
2538 */
2539void sched_exec(void)
2540{
2541        struct task_struct *p = current;
2542        unsigned long flags;
2543        int dest_cpu;
2544
2545        raw_spin_lock_irqsave(&p->pi_lock, flags);
2546        dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2547        if (dest_cpu == smp_processor_id())
2548                goto unlock;
2549
2550        if (likely(cpu_active(dest_cpu))) {
2551                struct migration_arg arg = { p, dest_cpu };
2552
2553                raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2554                stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2555                return;
2556        }
2557unlock:
2558        raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2559}
2560
2561#endif
2562
2563DEFINE_PER_CPU(struct kernel_stat, kstat);
2564DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2565
2566EXPORT_PER_CPU_SYMBOL(kstat);
2567EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2568
2569/*
2570 * Return any ns on the sched_clock that have not yet been accounted in
2571 * @p in case that task is currently running.
2572 *
2573 * Called with task_rq_lock() held on @rq.
2574 */
2575static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2576{
2577        u64 ns = 0;
2578
2579        if (task_current(rq, p)) {
2580                update_rq_clock(rq);
2581                ns = rq->clock_task - p->se.exec_start;
2582                if ((s64)ns < 0)
2583                        ns = 0;
2584        }
2585
2586        return ns;
2587}
2588
2589unsigned long long task_delta_exec(struct task_struct *p)
2590{
2591        unsigned long flags;
2592        struct rq *rq;
2593        u64 ns = 0;
2594
2595        rq = task_rq_lock(p, &flags);
2596        ns = do_task_delta_exec(p, rq);
2597        task_rq_unlock(rq, p, &flags);
2598
2599        return ns;
2600}
2601
2602/*
2603 * Return accounted runtime for the task.
2604 * In case the task is currently running, return the runtime plus current's
2605 * pending runtime that have not been accounted yet.
2606 */
2607unsigned long long task_sched_runtime(struct task_struct *p)
2608{
2609        unsigned long flags;
2610        struct rq *rq;
2611        u64 ns = 0;
2612
2613        rq = task_rq_lock(p, &flags);
2614        ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2615        task_rq_unlock(rq, p, &flags);
2616
2617        return ns;
2618}
2619
2620/*
2621 * This function gets called by the timer code, with HZ frequency.
2622 * We call it with interrupts disabled.
2623 */
2624void scheduler_tick(void)
2625{
2626        int cpu = smp_processor_id();
2627        struct rq *rq = cpu_rq(cpu);
2628        struct task_struct *curr = rq->curr;
2629
2630        sched_clock_tick();
2631
2632        raw_spin_lock(&rq->lock);
2633        update_rq_clock(rq);
2634        update_cpu_load_active(rq);
2635        curr->sched_class->task_tick(rq, curr, 0);
2636        raw_spin_unlock(&rq->lock);
2637
2638        perf_event_task_tick();
2639
2640#ifdef CONFIG_SMP
2641        rq->idle_balance = idle_cpu(cpu);
2642        trigger_load_balance(rq, cpu);
2643#endif
2644}
2645
2646notrace unsigned long get_parent_ip(unsigned long addr)
2647{
2648        if (in_lock_functions(addr)) {
2649                addr = CALLER_ADDR2;
2650                if (in_lock_functions(addr))
2651                        addr = CALLER_ADDR3;
2652        }
2653        return addr;
2654}
2655
2656#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2657                                defined(CONFIG_PREEMPT_TRACER))
2658
2659void __kprobes add_preempt_count(int val)
2660{
2661#ifdef CONFIG_DEBUG_PREEMPT
2662        /*
2663         * Underflow?
2664         */
2665        if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2666                return;
2667#endif
2668        preempt_count() += val;
2669#ifdef CONFIG_DEBUG_PREEMPT
2670        /*
2671         * Spinlock count overflowing soon?
2672         */
2673        DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2674                                PREEMPT_MASK - 10);
2675#endif
2676        if (preempt_count() == val)
2677                trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2678}
2679EXPORT_SYMBOL(add_preempt_count);
2680
2681void __kprobes sub_preempt_count(int val)
2682{
2683#ifdef CONFIG_DEBUG_PREEMPT
2684        /*
2685         * Underflow?
2686         */
2687        if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2688                return;
2689        /*
2690         * Is the spinlock portion underflowing?
2691         */
2692        if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2693                        !(preempt_count() & PREEMPT_MASK)))
2694                return;
2695#endif
2696
2697        if (preempt_count() == val)
2698                trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2699        preempt_count() -= val;
2700}
2701EXPORT_SYMBOL(sub_preempt_count);
2702
2703#endif
2704
2705/*
2706 * Print scheduling while atomic bug:
2707 */
2708static noinline void __schedule_bug(struct task_struct *prev)
2709{
2710        if (oops_in_progress)
2711                return;
2712
2713        printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2714                prev->comm, prev->pid, preempt_count());
2715
2716        debug_show_held_locks(prev);
2717        print_modules();
2718        if (irqs_disabled())
2719                print_irqtrace_events(prev);
2720        dump_stack();
2721        add_taint(TAINT_WARN);
2722}
2723
2724/*
2725 * Various schedule()-time debugging checks and statistics:
2726 */
2727static inline void schedule_debug(struct task_struct *prev)
2728{
2729        /*
2730         * Test if we are atomic. Since do_exit() needs to call into
2731         * schedule() atomically, we ignore that path for now.
2732         * Otherwise, whine if we are scheduling when we should not be.
2733         */
2734        if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2735                __schedule_bug(prev);
2736        rcu_sleep_check();
2737
2738        profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2739
2740        schedstat_inc(this_rq(), sched_count);
2741}
2742
2743static void put_prev_task(struct rq *rq, struct task_struct *prev)
2744{
2745        if (prev->on_rq || rq->skip_clock_update < 0)
2746                update_rq_clock(rq);
2747        prev->sched_class->put_prev_task(rq, prev);
2748}
2749
2750/*
2751 * Pick up the highest-prio task:
2752 */
2753static inline struct task_struct *
2754pick_next_task(struct rq *rq)
2755{
2756        const struct sched_class *class;
2757        struct task_struct *p;
2758
2759        /*
2760         * Optimization: we know that if all tasks are in
2761         * the fair class we can call that function directly:
2762         */
2763        if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2764                p = fair_sched_class.pick_next_task(rq);
2765                if (likely(p))
2766                        return p;
2767        }
2768
2769        for_each_class(class) {
2770                p = class->pick_next_task(rq);
2771                if (p)
2772                        return p;
2773        }
2774
2775        BUG(); /* the idle class will always have a runnable task */
2776}
2777
2778/*
2779 * __schedule() is the main scheduler function.
2780 *
2781 * The main means of driving the scheduler and thus entering this function are:
2782 *
2783 *   1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2784 *
2785 *   2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2786 *      paths. For example, see arch/x86/entry_64.S.
2787 *
2788 *      To drive preemption between tasks, the scheduler sets the flag in timer
2789 *      interrupt handler scheduler_tick().
2790 *
2791 *   3. Wakeups don't really cause entry into schedule(). They add a
2792 *      task to the run-queue and that's it.
2793 *
2794 *      Now, if the new task added to the run-queue preempts the current
2795 *      task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2796 *      called on the nearest possible occasion:
2797 *
2798 *       - If the kernel is preemptible (CONFIG_PREEMPT=y):
2799 *
2800 *         - in syscall or exception context, at the next outmost
2801 *           preempt_enable(). (this might be as soon as the wake_up()'s
2802 *           spin_unlock()!)
2803 *
2804 *         - in IRQ context, return from interrupt-handler to
2805 *           preemptible context
2806 *
2807 *       - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2808 *         then at the next:
2809 *
2810 *          - cond_resched() call
2811 *          - explicit schedule() call
2812 *          - return from syscall or exception to user-space
2813 *          - return from interrupt-handler to user-space
2814 */
2815static void __sched __schedule(void)
2816{
2817        struct task_struct *prev, *next;
2818        unsigned long *switch_count;
2819        struct rq *rq;
2820        int cpu;
2821
2822need_resched:
2823        preempt_disable();
2824        cpu = smp_processor_id();
2825        rq = cpu_rq(cpu);
2826        rcu_note_context_switch(cpu);
2827        prev = rq->curr;
2828
2829        schedule_debug(prev);
2830
2831        if (sched_feat(HRTICK))
2832                hrtick_clear(rq);
2833
2834        raw_spin_lock_irq(&rq->lock);
2835
2836        switch_count = &prev->nivcsw;
2837        if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2838                if (unlikely(signal_pending_state(prev->state, prev))) {
2839                        prev->state = TASK_RUNNING;
2840                } else {
2841                        deactivate_task(rq, prev, DEQUEUE_SLEEP);
2842                        prev->on_rq = 0;
2843
2844                        /*
2845                         * If a worker went to sleep, notify and ask workqueue
2846                         * whether it wants to wake up a task to maintain
2847                         * concurrency.
2848                         */
2849                        if (prev->flags & PF_WQ_WORKER) {
2850                                struct task_struct *to_wakeup;
2851
2852                                to_wakeup = wq_worker_sleeping(prev, cpu);
2853                                if (to_wakeup)
2854                                        try_to_wake_up_local(to_wakeup);
2855                        }
2856                }
2857                switch_count = &prev->nvcsw;
2858        }
2859
2860        pre_schedule(rq, prev);
2861
2862        if (unlikely(!rq->nr_running))
2863                idle_balance(cpu, rq);
2864
2865        put_prev_task(rq, prev);
2866        next = pick_next_task(rq);
2867        clear_tsk_need_resched(prev);
2868        rq->skip_clock_update = 0;
2869
2870        if (likely(prev != next)) {
2871                rq->nr_switches++;
2872                rq->curr = next;
2873                ++*switch_count;
2874
2875                context_switch(rq, prev, next); /* unlocks the rq */
2876                /*
2877                 * The context switch have flipped the stack from under us
2878                 * and restored the local variables which were saved when
2879                 * this task called schedule() in the past. prev == current
2880                 * is still correct, but it can be moved to another cpu/rq.
2881                 */
2882                cpu = smp_processor_id();
2883                rq = cpu_rq(cpu);
2884        } else
2885                raw_spin_unlock_irq(&rq->lock);
2886
2887        post_schedule(rq);
2888
2889        sched_preempt_enable_no_resched();
2890        if (need_resched())
2891                goto need_resched;
2892}
2893
2894static inline void sched_submit_work(struct task_struct *tsk)
2895{
2896        if (!tsk->state || tsk_is_pi_blocked(tsk))
2897                return;
2898        /*
2899         * If we are going to sleep and we have plugged IO queued,
2900         * make sure to submit it to avoid deadlocks.
2901         */
2902        if (blk_needs_flush_plug(tsk))
2903                blk_schedule_flush_plug(tsk);
2904}
2905
2906asmlinkage void __sched schedule(void)
2907{
2908        struct task_struct *tsk = current;
2909
2910        sched_submit_work(tsk);
2911        __schedule();
2912}
2913EXPORT_SYMBOL(schedule);
2914
2915#ifdef CONFIG_RCU_USER_QS
2916asmlinkage void __sched schedule_user(void)
2917{
2918        /*
2919         * If we come here after a random call to set_need_resched(),
2920         * or we have been woken up remotely but the IPI has not yet arrived,
2921         * we haven't yet exited the RCU idle mode. Do it here manually until
2922         * we find a better solution.
2923         */
2924        rcu_user_exit();
2925        schedule();
2926        rcu_user_enter();
2927}
2928#endif
2929
2930/**
2931 * schedule_preempt_disabled - called with preemption disabled
2932 *
2933 * Returns with preemption disabled. Note: preempt_count must be 1
2934 */
2935void __sched schedule_preempt_disabled(void)
2936{
2937        sched_preempt_enable_no_resched();
2938        schedule();
2939        preempt_disable();
2940}
2941
2942#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2943
2944static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2945{
2946        if (lock->owner != owner)
2947                return false;
2948
2949        /*
2950         * Ensure we emit the owner->on_cpu, dereference _after_ checking
2951         * lock->owner still matches owner, if that fails, owner might
2952         * point to free()d memory, if it still matches, the rcu_read_lock()
2953         * ensures the memory stays valid.
2954         */
2955        barrier();
2956
2957        return owner->on_cpu;
2958}
2959
2960/*
2961 * Look out! "owner" is an entirely speculative pointer
2962 * access and not reliable.
2963 */
2964int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
2965{
2966        if (!sched_feat(OWNER_SPIN))
2967                return 0;
2968
2969        rcu_read_lock();
2970        while (owner_running(lock, owner)) {
2971                if (need_resched())
2972                        break;
2973
2974                arch_mutex_cpu_relax();
2975        }
2976        rcu_read_unlock();
2977
2978        /*
2979         * We break out the loop above on need_resched() and when the
2980         * owner changed, which is a sign for heavy contention. Return
2981         * success only when lock->owner is NULL.
2982         */
2983        return lock->owner == NULL;
2984}
2985#endif
2986
2987#ifdef CONFIG_PREEMPT
2988/*
2989 * this is the entry point to schedule() from in-kernel preemption
2990 * off of preempt_enable. Kernel preemptions off return from interrupt
2991 * occur there and call schedule directly.
2992 */
2993asmlinkage void __sched notrace preempt_schedule(void)
2994{
2995        struct thread_info *ti = current_thread_info();
2996
2997        /*
2998         * If there is a non-zero preempt_count or interrupts are disabled,
2999         * we do not want to preempt the current task. Just return..
3000         */
3001        if (likely(ti->preempt_count || irqs_disabled()))
3002                return;
3003
3004        do {
3005                add_preempt_count_notrace(PREEMPT_ACTIVE);
3006                __schedule();
3007                sub_preempt_count_notrace(PREEMPT_ACTIVE);
3008
3009                /*
3010                 * Check again in case we missed a preemption opportunity
3011                 * between schedule and now.
3012                 */
3013                barrier();
3014        } while (need_resched());
3015}
3016EXPORT_SYMBOL(preempt_schedule);
3017
3018/*
3019 * this is the entry point to schedule() from kernel preemption
3020 * off of irq context.
3021 * Note, that this is called and return with irqs disabled. This will
3022 * protect us against recursive calling from irq.
3023 */
3024asmlinkage void __sched preempt_schedule_irq(void)
3025{
3026        struct thread_info *ti = current_thread_info();
3027
3028        /* Catch callers which need to be fixed */
3029        BUG_ON(ti->preempt_count || !irqs_disabled());
3030
3031        rcu_user_exit();
3032        do {
3033                add_preempt_count(PREEMPT_ACTIVE);
3034                local_irq_enable();
3035                __schedule();
3036                local_irq_disable();
3037                sub_preempt_count(PREEMPT_ACTIVE);
3038
3039                /*
3040                 * Check again in case we missed a preemption opportunity
3041                 * between schedule and now.
3042                 */
3043                barrier();
3044        } while (need_resched());
3045}
3046
3047#endif /* CONFIG_PREEMPT */
3048
3049int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3050                          void *key)
3051{
3052        return try_to_wake_up(curr->private, mode, wake_flags);
3053}
3054EXPORT_SYMBOL(default_wake_function);
3055
3056/*
3057 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3058 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3059 * number) then we wake all the non-exclusive tasks and one exclusive task.
3060 *
3061 * There are circumstances in which we can try to wake a task which has already
3062 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3063 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3064 */
3065static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3066                        int nr_exclusive, int wake_flags, void *key)
3067{
3068        wait_queue_t *curr, *next;
3069
3070        list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3071                unsigned flags = curr->flags;
3072
3073                if (curr->func(curr, mode, wake_flags, key) &&
3074                                (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3075                        break;
3076        }
3077}
3078
3079/**
3080 * __wake_up - wake up threads blocked on a waitqueue.
3081 * @q: the waitqueue
3082 * @mode: which threads
3083 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3084 * @key: is directly passed to the wakeup function
3085 *
3086 * It may be assumed that this function implies a write memory barrier before
3087 * changing the task state if and only if any tasks are woken up.
3088 */
3089void __wake_up(wait_queue_head_t *q, unsigned int mode,
3090                        int nr_exclusive, void *key)
3091{
3092        unsigned long flags;
3093
3094        spin_lock_irqsave(&q->lock, flags);
3095        __wake_up_common(q, mode, nr_exclusive, 0, key);
3096        spin_unlock_irqrestore(&q->lock, flags);
3097}
3098EXPORT_SYMBOL(__wake_up);
3099
3100/*
3101 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3102 */
3103void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3104{
3105        __wake_up_common(q, mode, nr, 0, NULL);
3106}
3107EXPORT_SYMBOL_GPL(__wake_up_locked);
3108
3109void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3110{
3111        __wake_up_common(q, mode, 1, 0, key);
3112}
3113EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3114
3115/**
3116 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3117 * @q: the waitqueue
3118 * @mode: which threads
3119 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3120 * @key: opaque value to be passed to wakeup targets
3121 *
3122 * The sync wakeup differs that the waker knows that it will schedule
3123 * away soon, so while the target thread will be woken up, it will not
3124 * be migrated to another CPU - ie. the two threads are 'synchronized'
3125 * with each other. This can prevent needless bouncing between CPUs.
3126 *
3127 * On UP it can prevent extra preemption.
3128 *
3129 * It may be assumed that this function implies a write memory barrier before
3130 * changing the task state if and only if any tasks are woken up.
3131 */
3132void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3133                        int nr_exclusive, void *key)
3134{
3135        unsigned long flags;
3136        int wake_flags = WF_SYNC;
3137
3138        if (unlikely(!q))
3139                return;
3140
3141        if (unlikely(!nr_exclusive))
3142                wake_flags = 0;
3143
3144        spin_lock_irqsave(&q->lock, flags);
3145        __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3146        spin_unlock_irqrestore(&q->lock, flags);
3147}
3148EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3149
3150/*
3151 * __wake_up_sync - see __wake_up_sync_key()
3152 */
3153void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3154{
3155        __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3156}
3157EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
3158
3159/**
3160 * complete: - signals a single thread waiting on this completion
3161 * @x:  holds the state of this particular completion
3162 *
3163 * This will wake up a single thread waiting on this completion. Threads will be
3164 * awakened in the same order in which they were queued.
3165 *
3166 * See also complete_all(), wait_for_completion() and related routines.
3167 *
3168 * It may be assumed that this function implies a write memory barrier before
3169 * changing the task state if and only if any tasks are woken up.
3170 */
3171void complete(struct completion *x)
3172{
3173        unsigned long flags;
3174
3175        spin_lock_irqsave(&x->wait.lock, flags);
3176        x->done++;
3177        __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3178        spin_unlock_irqrestore(&x->wait.lock, flags);
3179}
3180EXPORT_SYMBOL(complete);
3181
3182/**
3183 * complete_all: - signals all threads waiting on this completion
3184 * @x:  holds the state of this particular completion
3185 *
3186 * This will wake up all threads waiting on this particular completion event.
3187 *
3188 * It may be assumed that this function implies a write memory barrier before
3189 * changing the task state if and only if any tasks are woken up.
3190 */
3191void complete_all(struct completion *x)
3192{
3193        unsigned long flags;
3194
3195        spin_lock_irqsave(&x->wait.lock, flags);
3196        x->done += UINT_MAX/2;
3197        __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3198        spin_unlock_irqrestore(&x->wait.lock, flags);
3199}
3200EXPORT_SYMBOL(complete_all);
3201
3202static inline long __sched
3203do_wait_for_common(struct completion *x, long timeout, int state)
3204{
3205        if (!x->done) {
3206                DECLARE_WAITQUEUE(wait, current);
3207
3208                __add_wait_queue_tail_exclusive(&x->wait, &wait);
3209                do {
3210                        if (signal_pending_state(state, current)) {
3211                                timeout = -ERESTARTSYS;
3212                                break;
3213                        }
3214                        __set_current_state(state);
3215                        spin_unlock_irq(&x->wait.lock);
3216                        timeout = schedule_timeout(timeout);
3217                        spin_lock_irq(&x->wait.lock);
3218                } while (!x->done && timeout);
3219                __remove_wait_queue(&x->wait, &wait);
3220                if (!x->done)
3221                        return timeout;
3222        }
3223        x->done--;
3224        return timeout ?: 1;
3225}
3226
3227static long __sched
3228wait_for_common(struct completion *x, long timeout, int state)
3229{
3230        might_sleep();
3231
3232        spin_lock_irq(&x->wait.lock);
3233        timeout = do_wait_for_common(x, timeout, state);
3234        spin_unlock_irq(&x->wait.lock);
3235        return timeout;
3236}
3237
3238/**
3239 * wait_for_completion: - waits for completion of a task
3240 * @x:  holds the state of this particular completion
3241 *
3242 * This waits to be signaled for completion of a specific task. It is NOT
3243 * interruptible and there is no timeout.
3244 *
3245 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3246 * and interrupt capability. Also see complete().
3247 */
3248void __sched wait_for_completion(struct completion *x)
3249{
3250        wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3251}
3252EXPORT_SYMBOL(wait_for_completion);
3253
3254/**
3255 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3256 * @x:  holds the state of this particular completion
3257 * @timeout:  timeout value in jiffies
3258 *
3259 * This waits for either a completion of a specific task to be signaled or for a
3260 * specified timeout to expire. The timeout is in jiffies. It is not
3261 * interruptible.
3262 *
3263 * The return value is 0 if timed out, and positive (at least 1, or number of
3264 * jiffies left till timeout) if completed.
3265 */
3266unsigned long __sched
3267wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3268{
3269        return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3270}
3271EXPORT_SYMBOL(wait_for_completion_timeout);
3272
3273/**
3274 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3275 * @x:  holds the state of this particular completion
3276 *
3277 * This waits for completion of a specific task to be signaled. It is
3278 * interruptible.
3279 *
3280 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3281 */
3282int __sched wait_for_completion_interruptible(struct completion *x)
3283{
3284        long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3285        if (t == -ERESTARTSYS)
3286                return t;
3287        return 0;
3288}
3289EXPORT_SYMBOL(wait_for_completion_interruptible);
3290
3291/**
3292 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3293 * @x:  holds the state of this particular completion
3294 * @timeout:  timeout value in jiffies
3295 *
3296 * This waits for either a completion of a specific task to be signaled or for a
3297 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3298 *
3299 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3300 * positive (at least 1, or number of jiffies left till timeout) if completed.
3301 */
3302long __sched
3303wait_for_completion_interruptible_timeout(struct completion *x,
3304                                          unsigned long timeout)
3305{
3306        return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3307}
3308EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3309
3310/**
3311 * wait_for_completion_killable: - waits for completion of a task (killable)
3312 * @x:  holds the state of this particular completion
3313 *
3314 * This waits to be signaled for completion of a specific task. It can be
3315 * interrupted by a kill signal.
3316 *
3317 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3318 */
3319int __sched wait_for_completion_killable(struct completion *x)
3320{
3321        long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3322        if (t == -ERESTARTSYS)
3323                return t;
3324        return 0;
3325}
3326EXPORT_SYMBOL(wait_for_completion_killable);
3327
3328/**
3329 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3330 * @x:  holds the state of this particular completion
3331 * @timeout:  timeout value in jiffies
3332 *
3333 * This waits for either a completion of a specific task to be
3334 * signaled or for a specified timeout to expire. It can be
3335 * interrupted by a kill signal. The timeout is in jiffies.
3336 *
3337 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3338 * positive (at least 1, or number of jiffies left till timeout) if completed.
3339 */
3340long __sched
3341wait_for_completion_killable_timeout(struct completion *x,
3342                                     unsigned long timeout)
3343{
3344        return wait_for_common(x, timeout, TASK_KILLABLE);
3345}
3346EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3347
3348/**
3349 *      try_wait_for_completion - try to decrement a completion without blocking
3350 *      @x:     completion structure
3351 *
3352 *      Returns: 0 if a decrement cannot be done without blocking
3353 *               1 if a decrement succeeded.
3354 *
3355 *      If a completion is being used as a counting completion,
3356 *      attempt to decrement the counter without blocking. This
3357 *      enables us to avoid waiting if the resource the completion
3358 *      is protecting is not available.
3359 */
3360bool try_wait_for_completion(struct completion *x)
3361{
3362        unsigned long flags;
3363        int ret = 1;
3364
3365        spin_lock_irqsave(&x->wait.lock, flags);
3366        if (!x->done)
3367                ret = 0;
3368        else
3369                x->done--;
3370        spin_unlock_irqrestore(&x->wait.lock, flags);
3371        return ret;
3372}
3373EXPORT_SYMBOL(try_wait_for_completion);
3374
3375/**
3376 *      completion_done - Test to see if a completion has any waiters
3377 *      @x:     completion structure
3378 *
3379 *      Returns: 0 if there are waiters (wait_for_completion() in progress)
3380 *               1 if there are no waiters.
3381 *
3382 */
3383bool completion_done(struct completion *x)
3384{
3385        unsigned long flags;
3386        int ret = 1;
3387
3388        spin_lock_irqsave(&x->wait.lock, flags);
3389        if (!x->done)
3390                ret = 0;
3391        spin_unlock_irqrestore(&x->wait.lock, flags);
3392        return ret;
3393}
3394EXPORT_SYMBOL(completion_done);
3395
3396static long __sched
3397sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3398{
3399        unsigned long flags;
3400        wait_queue_t wait;
3401
3402        init_waitqueue_entry(&wait, current);
3403
3404        __set_current_state(state);
3405
3406        spin_lock_irqsave(&q->lock, flags);
3407        __add_wait_queue(q, &wait);
3408        spin_unlock(&q->lock);
3409        timeout = schedule_timeout(timeout);
3410        spin_lock_irq(&q->lock);
3411        __remove_wait_queue(q, &wait);
3412        spin_unlock_irqrestore(&q->lock, flags);
3413
3414        return timeout;
3415}
3416
3417void __sched interruptible_sleep_on(wait_queue_head_t *q)
3418{
3419        sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3420}
3421EXPORT_SYMBOL(interruptible_sleep_on);
3422
3423long __sched
3424interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3425{
3426        return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3427}
3428EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3429
3430void __sched sleep_on(wait_queue_head_t *q)
3431{
3432        sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3433}
3434EXPORT_SYMBOL(sleep_on);
3435
3436long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3437{
3438        return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3439}
3440EXPORT_SYMBOL(sleep_on_timeout);
3441
3442#ifdef CONFIG_RT_MUTEXES
3443
3444/*
3445 * rt_mutex_setprio - set the current priority of a task
3446 * @p: task
3447 * @prio: prio value (kernel-internal form)
3448 *
3449 * This function changes the 'effective' priority of a task. It does
3450 * not touch ->normal_prio like __setscheduler().
3451 *
3452 * Used by the rt_mutex code to implement priority inheritance logic.
3453 */
3454void rt_mutex_setprio(struct task_struct *p, int prio)
3455{
3456        int oldprio, on_rq, running;
3457        struct rq *rq;
3458        const struct sched_class *prev_class;
3459
3460        BUG_ON(prio < 0 || prio > MAX_PRIO);
3461
3462        rq = __task_rq_lock(p);
3463
3464        /*
3465         * Idle task boosting is a nono in general. There is one
3466         * exception, when PREEMPT_RT and NOHZ is active:
3467         *
3468         * The idle task calls get_next_timer_interrupt() and holds
3469         * the timer wheel base->lock on the CPU and another CPU wants
3470         * to access the timer (probably to cancel it). We can safely
3471         * ignore the boosting request, as the idle CPU runs this code
3472         * with interrupts disabled and will complete the lock
3473         * protected section without being interrupted. So there is no
3474         * real need to boost.
3475         */
3476        if (unlikely(p == rq->idle)) {
3477                WARN_ON(p != rq->curr);
3478                WARN_ON(p->pi_blocked_on);
3479                goto out_unlock;
3480        }
3481
3482        trace_sched_pi_setprio(p, prio);
3483        oldprio = p->prio;
3484        prev_class = p->sched_class;
3485        on_rq = p->on_rq;
3486        running = task_current(rq, p);
3487        if (on_rq)
3488                dequeue_task(rq, p, 0);
3489        if (running)
3490                p->sched_class->put_prev_task(rq, p);
3491
3492        if (rt_prio(prio))
3493                p->sched_class = &rt_sched_class;
3494        else
3495                p->sched_class = &fair_sched_class;
3496
3497        p->prio = prio;
3498
3499        if (running)
3500                p->sched_class->set_curr_task(rq);
3501        if (on_rq)
3502                enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3503
3504        check_class_changed(rq, p, prev_class, oldprio);
3505out_unlock:
3506        __task_rq_unlock(rq);
3507}
3508#endif
3509void set_user_nice(struct task_struct *p, long nice)
3510{
3511        int old_prio, delta, on_rq;
3512        unsigned long flags;
3513        struct rq *rq;
3514
3515        if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3516                return;
3517        /*
3518         * We have to be careful, if called from sys_setpriority(),
3519         * the task might be in the middle of scheduling on another CPU.
3520         */
3521        rq = task_rq_lock(p, &flags);
3522        /*
3523         * The RT priorities are set via sched_setscheduler(), but we still
3524         * allow the 'normal' nice value to be set - but as expected
3525         * it wont have any effect on scheduling until the task is
3526         * SCHED_FIFO/SCHED_RR:
3527         */
3528        if (task_has_rt_policy(p)) {
3529                p->static_prio = NICE_TO_PRIO(nice);
3530                goto out_unlock;
3531        }
3532        on_rq = p->on_rq;
3533        if (on_rq)
3534                dequeue_task(rq, p, 0);
3535
3536        p->static_prio = NICE_TO_PRIO(nice);
3537        set_load_weight(p);
3538        old_prio = p->prio;
3539        p->prio = effective_prio(p);
3540        delta = p->prio - old_prio;
3541
3542        if (on_rq) {
3543                enqueue_task(rq, p, 0);
3544                /*
3545                 * If the task increased its priority or is running and
3546                 * lowered its priority, then reschedule its CPU:
3547                 */
3548                if (delta < 0 || (delta > 0 && task_running(rq, p)))
3549                        resched_task(rq->curr);
3550        }
3551out_unlock:
3552        task_rq_unlock(rq, p, &flags);
3553}
3554EXPORT_SYMBOL(set_user_nice);
3555
3556/*
3557 * can_nice - check if a task can reduce its nice value
3558 * @p: task
3559 * @nice: nice value
3560 */
3561int can_nice(const struct task_struct *p, const int nice)
3562{
3563        /* convert nice value [19,-20] to rlimit style value [1,40] */
3564        int nice_rlim = 20 - nice;
3565
3566        return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3567                capable(CAP_SYS_NICE));
3568}
3569
3570#ifdef __ARCH_WANT_SYS_NICE
3571
3572/*
3573 * sys_nice - change the priority of the current process.
3574 * @increment: priority increment
3575 *
3576 * sys_setpriority is a more generic, but much slower function that
3577 * does similar things.
3578 */
3579SYSCALL_DEFINE1(nice, int, increment)
3580{
3581        long nice, retval;
3582
3583        /*
3584         * Setpriority might change our priority at the same moment.
3585         * We don't have to worry. Conceptually one call occurs first
3586         * and we have a single winner.
3587         */
3588        if (increment < -40)
3589                increment = -40;
3590        if (increment > 40)
3591                increment = 40;
3592
3593        nice = TASK_NICE(current) + increment;
3594        if (nice < -20)
3595                nice = -20;
3596        if (nice > 19)
3597                nice = 19;
3598
3599        if (increment < 0 && !can_nice(current, nice))
3600                return -EPERM;
3601
3602        retval = security_task_setnice(current, nice);
3603        if (retval)
3604                return retval;
3605
3606        set_user_nice(current, nice);
3607        return 0;
3608}
3609
3610#endif
3611
3612/**
3613 * task_prio - return the priority value of a given task.
3614 * @p: the task in question.
3615 *
3616 * This is the priority value as seen by users in /proc.
3617 * RT tasks are offset by -200. Normal tasks are centered
3618 * around 0, value goes from -16 to +15.
3619 */
3620int task_prio(const struct task_struct *p)
3621{
3622        return p->prio - MAX_RT_PRIO;
3623}
3624
3625/**
3626 * task_nice - return the nice value of a given task.
3627 * @p: the task in question.
3628 */
3629int task_nice(const struct task_struct *p)
3630{
3631        return TASK_NICE(p);
3632}
3633EXPORT_SYMBOL(task_nice);
3634
3635/**
3636 * idle_cpu - is a given cpu idle currently?
3637 * @cpu: the processor in question.
3638 */
3639int idle_cpu(int cpu)
3640{
3641        struct rq *rq = cpu_rq(cpu);
3642
3643        if (rq->curr != rq->idle)
3644                return 0;
3645
3646        if (rq->nr_running)
3647                return 0;
3648
3649#ifdef CONFIG_SMP
3650        if (!llist_empty(&rq->wake_list))
3651                return 0;
3652#endif
3653
3654        return 1;
3655}
3656
3657/**
3658 * idle_task - return the idle task for a given cpu.
3659 * @cpu: the processor in question.
3660 */
3661struct task_struct *idle_task(int cpu)
3662{
3663        return cpu_rq(cpu)->idle;
3664}
3665
3666/**
3667 * find_process_by_pid - find a process with a matching PID value.
3668 * @pid: the pid in question.
3669 */
3670static struct task_struct *find_process_by_pid(pid_t pid)
3671{
3672        return pid ? find_task_by_vpid(pid) : current;
3673}
3674
3675/* Actually do priority change: must hold rq lock. */
3676static void
3677__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3678{
3679        p->policy = policy;
3680        p->rt_priority = prio;
3681        p->normal_prio = normal_prio(p);
3682        /* we are holding p->pi_lock already */
3683        p->prio = rt_mutex_getprio(p);
3684        if (rt_prio(p->prio))
3685                p->sched_class = &rt_sched_class;
3686        else
3687                p->sched_class = &fair_sched_class;
3688        set_load_weight(p);
3689}
3690
3691/*
3692 * check the target process has a UID that matches the current process's
3693 */
3694static bool check_same_owner(struct task_struct *p)
3695{
3696        const struct cred *cred = current_cred(), *pcred;
3697        bool match;
3698
3699        rcu_read_lock();
3700        pcred = __task_cred(p);
3701        match = (uid_eq(cred->euid, pcred->euid) ||
3702                 uid_eq(cred->euid, pcred->uid));
3703        rcu_read_unlock();
3704        return match;
3705}
3706
3707static int __sched_setscheduler(struct task_struct *p, int policy,
3708                                const struct sched_param *param, bool user)
3709{
3710        int retval, oldprio, oldpolicy = -1, on_rq, running;
3711        unsigned long flags;
3712        const struct sched_class *prev_class;
3713        struct rq *rq;
3714        int reset_on_fork;
3715
3716        /* may grab non-irq protected spin_locks */
3717        BUG_ON(in_interrupt());
3718recheck:
3719        /* double check policy once rq lock held */
3720        if (policy < 0) {
3721                reset_on_fork = p->sched_reset_on_fork;
3722                policy = oldpolicy = p->policy;
3723        } else {
3724                reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3725                policy &= ~SCHED_RESET_ON_FORK;
3726
3727                if (policy != SCHED_FIFO && policy != SCHED_RR &&
3728                                policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3729                                policy != SCHED_IDLE)
3730                        return -EINVAL;
3731        }
3732
3733        /*
3734         * Valid priorities for SCHED_FIFO and SCHED_RR are
3735         * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3736         * SCHED_BATCH and SCHED_IDLE is 0.
3737         */
3738        if (param->sched_priority < 0 ||
3739            (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3740            (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3741                return -EINVAL;
3742        if (rt_policy(policy) != (param->sched_priority != 0))
3743                return -EINVAL;
3744
3745        /*
3746         * Allow unprivileged RT tasks to decrease priority:
3747         */
3748        if (user && !capable(CAP_SYS_NICE)) {
3749                if (rt_policy(policy)) {
3750                        unsigned long rlim_rtprio =
3751                                        task_rlimit(p, RLIMIT_RTPRIO);
3752
3753                        /* can't set/change the rt policy */
3754                        if (policy != p->policy && !rlim_rtprio)
3755                                return -EPERM;
3756
3757                        /* can't increase priority */
3758                        if (param->sched_priority > p->rt_priority &&
3759                            param->sched_priority > rlim_rtprio)
3760                                return -EPERM;
3761                }
3762
3763                /*
3764                 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3765                 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3766                 */
3767                if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3768                        if (!can_nice(p, TASK_NICE(p)))
3769                                return -EPERM;
3770                }
3771
3772                /* can't change other user's priorities */
3773                if (!check_same_owner(p))
3774                        return -EPERM;
3775
3776                /* Normal users shall not reset the sched_reset_on_fork flag */
3777                if (p->sched_reset_on_fork && !reset_on_fork)
3778                        return -EPERM;
3779        }
3780
3781        if (user) {
3782                retval = security_task_setscheduler(p);
3783                if (retval)
3784                        return retval;
3785        }
3786
3787        /*
3788         * make sure no PI-waiters arrive (or leave) while we are
3789         * changing the priority of the task:
3790         *
3791         * To be able to change p->policy safely, the appropriate
3792         * runqueue lock must be held.
3793         */
3794        rq = task_rq_lock(p, &flags);
3795
3796        /*
3797         * Changing the policy of the stop threads its a very bad idea
3798         */
3799        if (p == rq->stop) {
3800                task_rq_unlock(rq, p, &flags);
3801                return -EINVAL;
3802        }
3803
3804        /*
3805         * If not changing anything there's no need to proceed further:
3806         */
3807        if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3808                        param->sched_priority == p->rt_priority))) {
3809                task_rq_unlock(rq, p, &flags);
3810                return 0;
3811        }
3812
3813#ifdef CONFIG_RT_GROUP_SCHED
3814        if (user) {
3815                /*
3816                 * Do not allow realtime tasks into groups that have no runtime
3817                 * assigned.
3818                 */
3819                if (rt_bandwidth_enabled() && rt_policy(policy) &&
3820                                task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3821                                !task_group_is_autogroup(task_group(p))) {
3822                        task_rq_unlock(rq, p, &flags);
3823                        return -EPERM;
3824                }
3825        }
3826#endif
3827
3828        /* recheck policy now with rq lock held */
3829        if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3830                policy = oldpolicy = -1;
3831                task_rq_unlock(rq, p, &flags);
3832                goto recheck;
3833        }
3834        on_rq = p->on_rq;
3835        running = task_current(rq, p);
3836        if (on_rq)
3837                dequeue_task(rq, p, 0);
3838        if (running)
3839                p->sched_class->put_prev_task(rq, p);
3840
3841        p->sched_reset_on_fork = reset_on_fork;
3842
3843        oldprio = p->prio;
3844        prev_class = p->sched_class;
3845        __setscheduler(rq, p, policy, param->sched_priority);
3846
3847        if (running)
3848                p->sched_class->set_curr_task(rq);
3849        if (on_rq)
3850                enqueue_task(rq, p, 0);
3851
3852        check_class_changed(rq, p, prev_class, oldprio);
3853        task_rq_unlock(rq, p, &flags);
3854
3855        rt_mutex_adjust_pi(p);
3856
3857        return 0;
3858}
3859
3860/**
3861 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3862 * @p: the task in question.
3863 * @policy: new policy.
3864 * @param: structure containing the new RT priority.
3865 *
3866 * NOTE that the task may be already dead.
3867 */
3868int sched_setscheduler(struct task_struct *p, int policy,
3869                       const struct sched_param *param)
3870{
3871        return __sched_setscheduler(p, policy, param, true);
3872}
3873EXPORT_SYMBOL_GPL(sched_setscheduler);
3874
3875/**
3876 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3877 * @p: the task in question.
3878 * @policy: new policy.
3879 * @param: structure containing the new RT priority.
3880 *
3881 * Just like sched_setscheduler, only don't bother checking if the
3882 * current context has permission.  For example, this is needed in
3883 * stop_machine(): we create temporary high priority worker threads,
3884 * but our caller might not have that capability.
3885 */
3886int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3887                               const struct sched_param *param)
3888{
3889        return __sched_setscheduler(p, policy, param, false);
3890}
3891
3892static int
3893do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3894{
3895        struct sched_param lparam;
3896        struct task_struct *p;
3897        int retval;
3898
3899        if (!param || pid < 0)
3900                return -EINVAL;
3901        if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3902                return -EFAULT;
3903
3904        rcu_read_lock();
3905        retval = -ESRCH;
3906        p = find_process_by_pid(pid);
3907        if (p != NULL)
3908                retval = sched_setscheduler(p, policy, &lparam);
3909        rcu_read_unlock();
3910
3911        return retval;
3912}
3913
3914/**
3915 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3916 * @pid: the pid in question.
3917 * @policy: new policy.
3918 * @param: structure containing the new RT priority.
3919 */
3920SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3921                struct sched_param __user *, param)
3922{
3923        /* negative values for policy are not valid */
3924        if (policy < 0)
3925                return -EINVAL;
3926
3927        return do_sched_setscheduler(pid, policy, param);
3928}
3929
3930/**
3931 * sys_sched_setparam - set/change the RT priority of a thread
3932 * @pid: the pid in question.
3933 * @param: structure containing the new RT priority.
3934 */
3935SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3936{
3937        return do_sched_setscheduler(pid, -1, param);
3938}
3939
3940/**
3941 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3942 * @pid: the pid in question.
3943 */
3944SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3945{
3946        struct task_struct *p;
3947        int retval;
3948
3949        if (pid < 0)
3950                return -EINVAL;
3951
3952        retval = -ESRCH;
3953        rcu_read_lock();
3954        p = find_process_by_pid(pid);
3955        if (p) {
3956                retval = security_task_getscheduler(p);
3957                if (!retval)
3958                        retval = p->policy
3959                                | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3960        }
3961        rcu_read_unlock();
3962        return retval;
3963}
3964
3965/**
3966 * sys_sched_getparam - get the RT priority of a thread
3967 * @pid: the pid in question.
3968 * @param: structure containing the RT priority.
3969 */
3970SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3971{
3972        struct sched_param lp;
3973        struct task_struct *p;
3974        int retval;
3975
3976        if (!param || pid < 0)
3977                return -EINVAL;
3978
3979        rcu_read_lock();
3980        p = find_process_by_pid(pid);
3981        retval = -ESRCH;
3982        if (!p)
3983                goto out_unlock;
3984
3985        retval = security_task_getscheduler(p);
3986        if (retval)
3987                goto out_unlock;
3988
3989        lp.sched_priority = p->rt_priority;
3990        rcu_read_unlock();
3991
3992        /*
3993         * This one might sleep, we cannot do it with a spinlock held ...
3994         */
3995        retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3996
3997        return retval;
3998
3999out_unlock:
4000        rcu_read_unlock();
4001        return retval;
4002}
4003
4004long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4005{
4006        cpumask_var_t cpus_allowed, new_mask;
4007        struct task_struct *p;
4008        int retval;
4009
4010        get_online_cpus();
4011        rcu_read_lock();
4012
4013        p = find_process_by_pid(pid);
4014        if (!p) {
4015                rcu_read_unlock();
4016                put_online_cpus();
4017                return -ESRCH;
4018        }
4019
4020        /* Prevent p going away */
4021        get_task_struct(p);
4022        rcu_read_unlock();
4023
4024        if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4025                retval = -ENOMEM;
4026                goto out_put_task;
4027        }
4028        if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4029                retval = -ENOMEM;
4030                goto out_free_cpus_allowed;
4031        }
4032        retval = -EPERM;
4033        if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4034                goto out_unlock;
4035
4036        retval = security_task_setscheduler(p);
4037        if (retval)
4038                goto out_unlock;
4039
4040        cpuset_cpus_allowed(p, cpus_allowed);
4041        cpumask_and(new_mask, in_mask, cpus_allowed);
4042again:
4043        retval = set_cpus_allowed_ptr(p, new_mask);
4044
4045        if (!retval) {
4046                cpuset_cpus_allowed(p, cpus_allowed);
4047                if (!cpumask_subset(new_mask, cpus_allowed)) {
4048                        /*
4049                         * We must have raced with a concurrent cpuset
4050                         * update. Just reset the cpus_allowed to the
4051                         * cpuset's cpus_allowed
4052                         */
4053                        cpumask_copy(new_mask, cpus_allowed);
4054                        goto again;
4055                }
4056        }
4057out_unlock:
4058        free_cpumask_var(new_mask);
4059out_free_cpus_allowed:
4060        free_cpumask_var(cpus_allowed);
4061out_put_task:
4062        put_task_struct(p);
4063        put_online_cpus();
4064        return retval;
4065}
4066
4067static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4068                             struct cpumask *new_mask)
4069{
4070        if (len < cpumask_size())
4071                cpumask_clear(new_mask);
4072        else if (len > cpumask_size())
4073                len = cpumask_size();
4074
4075        return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4076}
4077
4078/**
4079 * sys_sched_setaffinity - set the cpu affinity of a process
4080 * @pid: pid of the process
4081 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4082 * @user_mask_ptr: user-space pointer to the new cpu mask
4083 */
4084SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4085                unsigned long __user *, user_mask_ptr)
4086{
4087        cpumask_var_t new_mask;
4088        int retval;
4089
4090        if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4091                return -ENOMEM;
4092
4093        retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4094        if (retval == 0)
4095                retval = sched_setaffinity(pid, new_mask);
4096        free_cpumask_var(new_mask);
4097        return retval;
4098}
4099
4100long sched_getaffinity(pid_t pid, struct cpumask *mask)
4101{
4102        struct task_struct *p;
4103        unsigned long flags;
4104        int retval;
4105
4106        get_online_cpus();
4107        rcu_read_lock();
4108
4109        retval = -ESRCH;
4110        p = find_process_by_pid(pid);
4111        if (!p)
4112                goto out_unlock;
4113
4114        retval = security_task_getscheduler(p);
4115        if (retval)
4116                goto out_unlock;
4117
4118        raw_spin_lock_irqsave(&p->pi_lock, flags);
4119        cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4120        raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4121
4122out_unlock:
4123        rcu_read_unlock();
4124        put_online_cpus();
4125
4126        return retval;
4127}
4128
4129/**
4130 * sys_sched_getaffinity - get the cpu affinity of a process
4131 * @pid: pid of the process
4132 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4133 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4134 */
4135SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4136                unsigned long __user *, user_mask_ptr)
4137{
4138        int ret;
4139        cpumask_var_t mask;
4140
4141        if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4142                return -EINVAL;
4143        if (len & (sizeof(unsigned long)-1))
4144                return -EINVAL;
4145
4146        if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4147                return -ENOMEM;
4148
4149        ret = sched_getaffinity(pid, mask);
4150        if (ret == 0) {
4151                size_t retlen = min_t(size_t, len, cpumask_size());
4152
4153                if (copy_to_user(user_mask_ptr, mask, retlen))
4154                        ret = -EFAULT;
4155                else
4156                        ret = retlen;
4157        }
4158        free_cpumask_var(mask);
4159
4160        return ret;
4161}
4162
4163/**
4164 * sys_sched_yield - yield the current processor to other threads.
4165 *
4166 * This function yields the current CPU to other tasks. If there are no
4167 * other threads running on this CPU then this function will return.
4168 */
4169SYSCALL_DEFINE0(sched_yield)
4170{
4171        struct rq *rq = this_rq_lock();
4172
4173        schedstat_inc(rq, yld_count);
4174        current->sched_class->yield_task(rq);
4175
4176        /*
4177         * Since we are going to call schedule() anyway, there's
4178         * no need to preempt or enable interrupts:
4179         */
4180        __release(rq->lock);
4181        spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4182        do_raw_spin_unlock(&rq->lock);
4183        sched_preempt_enable_no_resched();
4184
4185        schedule();
4186
4187        return 0;
4188}
4189
4190static inline int should_resched(void)
4191{
4192        return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4193}
4194
4195static void __cond_resched(void)
4196{
4197        add_preempt_count(PREEMPT_ACTIVE);
4198        __schedule();
4199        sub_preempt_count(PREEMPT_ACTIVE);
4200}
4201
4202int __sched _cond_resched(void)
4203{
4204        if (should_resched()) {
4205                __cond_resched();
4206                return 1;
4207        }
4208        return 0;
4209}
4210EXPORT_SYMBOL(_cond_resched);
4211
4212/*
4213 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4214 * call schedule, and on return reacquire the lock.
4215 *
4216 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4217 * operations here to prevent schedule() from being called twice (once via
4218 * spin_unlock(), once by hand).
4219 */
4220int __cond_resched_lock(spinlock_t *lock)
4221{
4222        int resched = should_resched();
4223        int ret = 0;
4224
4225        lockdep_assert_held(lock);
4226
4227        if (spin_needbreak(lock) || resched) {
4228                spin_unlock(lock);
4229                if (resched)
4230                        __cond_resched();
4231                else
4232                        cpu_relax();
4233                ret = 1;
4234                spin_lock(lock);
4235        }
4236        return ret;
4237}
4238EXPORT_SYMBOL(__cond_resched_lock);
4239
4240int __sched __cond_resched_softirq(void)
4241{
4242        BUG_ON(!in_softirq());
4243
4244        if (should_resched()) {
4245                local_bh_enable();
4246                __cond_resched();
4247                local_bh_disable();
4248                return 1;
4249        }
4250        return 0;
4251}
4252EXPORT_SYMBOL(__cond_resched_softirq);
4253
4254/**
4255 * yield - yield the current processor to other threads.
4256 *
4257 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4258 *
4259 * The scheduler is at all times free to pick the calling task as the most
4260 * eligible task to run, if removing the yield() call from your code breaks
4261 * it, its already broken.
4262 *
4263 * Typical broken usage is:
4264 *
4265 * while (!event)
4266 *      yield();
4267 *
4268 * where one assumes that yield() will let 'the other' process run that will
4269 * make event true. If the current task is a SCHED_FIFO task that will never
4270 * happen. Never use yield() as a progress guarantee!!
4271 *
4272 * If you want to use yield() to wait for something, use wait_event().
4273 * If you want to use yield() to be 'nice' for others, use cond_resched().
4274 * If you still want to use yield(), do not!
4275 */
4276void __sched yield(void)
4277{
4278        set_current_state(TASK_RUNNING);
4279        sys_sched_yield();
4280}
4281EXPORT_SYMBOL(yield);
4282
4283/**
4284 * yield_to - yield the current processor to another thread in
4285 * your thread group, or accelerate that thread toward the
4286 * processor it's on.
4287 * @p: target task
4288 * @preempt: whether task preemption is allowed or not
4289 *
4290 * It's the caller's job to ensure that the target task struct
4291 * can't go away on us before we can do any checks.
4292 *
4293 * Returns true if we indeed boosted the target task.
4294 */
4295bool __sched yield_to(struct task_struct *p, bool preempt)
4296{
4297        struct task_struct *curr = current;
4298        struct rq *rq, *p_rq;
4299        unsigned long flags;
4300        bool yielded = 0;
4301
4302        local_irq_save(flags);
4303        rq = this_rq();
4304
4305again:
4306        p_rq = task_rq(p);
4307        double_rq_lock(rq, p_rq);
4308        while (task_rq(p) != p_rq) {
4309                double_rq_unlock(rq, p_rq);
4310                goto again;
4311        }
4312
4313        if (!curr->sched_class->yield_to_task)
4314                goto out;
4315
4316        if (curr->sched_class != p->sched_class)
4317                goto out;
4318
4319        if (task_running(p_rq, p) || p->state)
4320                goto out;
4321
4322        yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4323        if (yielded) {
4324                schedstat_inc(rq, yld_count);
4325                /*
4326                 * Make p's CPU reschedule; pick_next_entity takes care of
4327                 * fairness.
4328                 */
4329                if (preempt && rq != p_rq)
4330                        resched_task(p_rq->curr);
4331        }
4332
4333out:
4334        double_rq_unlock(rq, p_rq);
4335        local_irq_restore(flags);
4336
4337        if (yielded)
4338                schedule();
4339
4340        return yielded;
4341}
4342EXPORT_SYMBOL_GPL(yield_to);
4343
4344/*
4345 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4346 * that process accounting knows that this is a task in IO wait state.
4347 */
4348void __sched io_schedule(void)
4349{
4350        struct rq *rq = raw_rq();
4351
4352        delayacct_blkio_start();
4353        atomic_inc(&rq->nr_iowait);
4354        blk_flush_plug(current);
4355        current->in_iowait = 1;
4356        schedule();
4357        current->in_iowait = 0;
4358        atomic_dec(&rq->nr_iowait);
4359        delayacct_blkio_end();
4360}
4361EXPORT_SYMBOL(io_schedule);
4362
4363long __sched io_schedule_timeout(long timeout)
4364{
4365        struct rq *rq = raw_rq();
4366        long ret;
4367
4368        delayacct_blkio_start();
4369        atomic_inc(&rq->nr_iowait);
4370        blk_flush_plug(current);
4371        current->in_iowait = 1;
4372        ret = schedule_timeout(timeout);
4373        current->in_iowait = 0;
4374        atomic_dec(&rq->nr_iowait);
4375        delayacct_blkio_end();
4376        return ret;
4377}
4378
4379/**
4380 * sys_sched_get_priority_max - return maximum RT priority.
4381 * @policy: scheduling class.
4382 *
4383 * this syscall returns the maximum rt_priority that can be used
4384 * by a given scheduling class.
4385 */
4386SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4387{
4388        int ret = -EINVAL;
4389
4390        switch (policy) {
4391        case SCHED_FIFO:
4392        case SCHED_RR:
4393                ret = MAX_USER_RT_PRIO-1;
4394                break;
4395        case SCHED_NORMAL:
4396        case SCHED_BATCH:
4397        case SCHED_IDLE:
4398                ret = 0;
4399                break;
4400        }
4401        return ret;
4402}
4403
4404/**
4405 * sys_sched_get_priority_min - return minimum RT priority.
4406 * @policy: scheduling class.
4407 *
4408 * this syscall returns the minimum rt_priority that can be used
4409 * by a given scheduling class.
4410 */
4411SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4412{
4413        int ret = -EINVAL;
4414
4415        switch (policy) {
4416        case SCHED_FIFO:
4417        case SCHED_RR:
4418                ret = 1;
4419                break;
4420        case SCHED_NORMAL:
4421        case SCHED_BATCH:
4422        case SCHED_IDLE:
4423                ret = 0;
4424        }
4425        return ret;
4426}
4427
4428/**
4429 * sys_sched_rr_get_interval - return the default timeslice of a process.
4430 * @pid: pid of the process.
4431 * @interval: userspace pointer to the timeslice value.
4432 *
4433 * this syscall writes the default timeslice value of a given process
4434 * into the user-space timespec buffer. A value of '0' means infinity.
4435 */
4436SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4437                struct timespec __user *, interval)
4438{
4439        struct task_struct *p;
4440        unsigned int time_slice;
4441        unsigned long flags;
4442        struct rq *rq;
4443        int retval;
4444        struct timespec t;
4445
4446        if (pid < 0)
4447                return -EINVAL;
4448
4449        retval = -ESRCH;
4450        rcu_read_lock();
4451        p = find_process_by_pid(pid);
4452        if (!p)
4453                goto out_unlock;
4454
4455        retval = security_task_getscheduler(p);
4456        if (retval)
4457                goto out_unlock;
4458
4459        rq = task_rq_lock(p, &flags);
4460        time_slice = p->sched_class->get_rr_interval(rq, p);
4461        task_rq_unlock(rq, p, &flags);
4462
4463        rcu_read_unlock();
4464        jiffies_to_timespec(time_slice, &t);
4465        retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4466        return retval;
4467
4468out_unlock:
4469        rcu_read_unlock();
4470        return retval;
4471}
4472
4473static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4474
4475void sched_show_task(struct task_struct *p)
4476{
4477        unsigned long free = 0;
4478        unsigned state;
4479
4480        state = p->state ? __ffs(p->state) + 1 : 0;
4481        printk(KERN_INFO "%-15.15s %c", p->comm,
4482                state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4483#if BITS_PER_LONG == 32
4484        if (state == TASK_RUNNING)
4485                printk(KERN_CONT " running  ");
4486        else
4487                printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4488#else
4489        if (state == TASK_RUNNING)
4490                printk(KERN_CONT "  running task    ");
4491        else
4492                printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4493#endif
4494#ifdef CONFIG_DEBUG_STACK_USAGE
4495        free = stack_not_used(p);
4496#endif
4497        printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4498                task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4499                (unsigned long)task_thread_info(p)->flags);
4500
4501        show_stack(p, NULL);
4502}
4503
4504void show_state_filter(unsigned long state_filter)
4505{
4506        struct task_struct *g, *p;
4507
4508#if BITS_PER_LONG == 32
4509        printk(KERN_INFO
4510                "  task                PC stack   pid father\n");
4511#else
4512        printk(KERN_INFO
4513                "  task                        PC stack   pid father\n");
4514#endif
4515        rcu_read_lock();
4516        do_each_thread(g, p) {
4517                /*
4518                 * reset the NMI-timeout, listing all files on a slow
4519                 * console might take a lot of time:
4520                 */
4521                touch_nmi_watchdog();
4522                if (!state_filter || (p->state & state_filter))
4523                        sched_show_task(p);
4524        } while_each_thread(g, p);
4525
4526        touch_all_softlockup_watchdogs();
4527
4528#ifdef CONFIG_SCHED_DEBUG
4529        sysrq_sched_debug_show();
4530#endif
4531        rcu_read_unlock();
4532        /*
4533         * Only show locks if all tasks are dumped:
4534         */
4535        if (!state_filter)
4536                debug_show_all_locks();
4537}
4538
4539void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4540{
4541        idle->sched_class = &idle_sched_class;
4542}
4543
4544/**
4545 * init_idle - set up an idle thread for a given CPU
4546 * @idle: task in question
4547 * @cpu: cpu the idle task belongs to
4548 *
4549 * NOTE: this function does not set the idle thread's NEED_RESCHED
4550 * flag, to make booting more robust.
4551 */
4552void __cpuinit init_idle(struct task_struct *idle, int cpu)
4553{
4554        struct rq *rq = cpu_rq(cpu);
4555        unsigned long flags;
4556
4557        raw_spin_lock_irqsave(&rq->lock, flags);
4558
4559        __sched_fork(idle);
4560        idle->state = TASK_RUNNING;
4561        idle->se.exec_start = sched_clock();
4562
4563        do_set_cpus_allowed(idle, cpumask_of(cpu));
4564        /*
4565         * We're having a chicken and egg problem, even though we are
4566         * holding rq->lock, the cpu isn't yet set to this cpu so the
4567         * lockdep check in task_group() will fail.
4568         *
4569         * Similar case to sched_fork(). / Alternatively we could
4570         * use task_rq_lock() here and obtain the other rq->lock.
4571         *
4572         * Silence PROVE_RCU
4573         */
4574        rcu_read_lock();
4575        __set_task_cpu(idle, cpu);
4576        rcu_read_unlock();
4577
4578        rq->curr = rq->idle = idle;
4579#if defined(CONFIG_SMP)
4580        idle->on_cpu = 1;
4581#endif
4582        raw_spin_unlock_irqrestore(&rq->lock, flags);
4583
4584        /* Set the preempt count _outside_ the spinlocks! */
4585        task_thread_info(idle)->preempt_count = 0;
4586
4587        /*
4588         * The idle tasks have their own, simple scheduling class:
4589         */
4590        idle->sched_class = &idle_sched_class;
4591        ftrace_graph_init_idle_task(idle, cpu);
4592#if defined(CONFIG_SMP)
4593        sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4594#endif
4595}
4596
4597#ifdef CONFIG_SMP
4598void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4599{
4600        if (p->sched_class && p->sched_class->set_cpus_allowed)
4601                p->sched_class->set_cpus_allowed(p, new_mask);
4602
4603        cpumask_copy(&p->cpus_allowed, new_mask);
4604        p->nr_cpus_allowed = cpumask_weight(new_mask);
4605}
4606
4607/*
4608 * This is how migration works:
4609 *
4610 * 1) we invoke migration_cpu_stop() on the target CPU using
4611 *    stop_one_cpu().
4612 * 2) stopper starts to run (implicitly forcing the migrated thread
4613 *    off the CPU)
4614 * 3) it checks whether the migrated task is still in the wrong runqueue.
4615 * 4) if it's in the wrong runqueue then the migration thread removes
4616 *    it and puts it into the right queue.
4617 * 5) stopper completes and stop_one_cpu() returns and the migration
4618 *    is done.
4619 */
4620
4621/*
4622 * Change a given task's CPU affinity. Migrate the thread to a
4623 * proper CPU and schedule it away if the CPU it's executing on
4624 * is removed from the allowed bitmask.
4625 *
4626 * NOTE: the caller must have a valid reference to the task, the
4627 * task must not exit() & deallocate itself prematurely. The
4628 * call is not atomic; no spinlocks may be held.
4629 */
4630int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4631{
4632        unsigned long flags;
4633        struct rq *rq;
4634        unsigned int dest_cpu;
4635        int ret = 0;
4636
4637        rq = task_rq_lock(p, &flags);
4638
4639        if (cpumask_equal(&p->cpus_allowed, new_mask))
4640                goto out;
4641
4642        if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4643                ret = -EINVAL;
4644                goto out;
4645        }
4646
4647        if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4648                ret = -EINVAL;
4649                goto out;
4650        }
4651
4652        do_set_cpus_allowed(p, new_mask);
4653
4654        /* Can the task run on the task's current CPU? If so, we're done */
4655        if (cpumask_test_cpu(task_cpu(p), new_mask))
4656                goto out;
4657
4658        dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4659        if (p->on_rq) {
4660                struct migration_arg arg = { p, dest_cpu };
4661                /* Need help from migration thread: drop lock and wait. */
4662                task_rq_unlock(rq, p, &flags);
4663                stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4664                tlb_migrate_finish(p->mm);
4665                return 0;
4666        }
4667out:
4668        task_rq_unlock(rq, p, &flags);
4669
4670        return ret;
4671}
4672EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4673
4674/*
4675 * Move (not current) task off this cpu, onto dest cpu. We're doing
4676 * this because either it can't run here any more (set_cpus_allowed()
4677 * away from this CPU, or CPU going down), or because we're
4678 * attempting to rebalance this task on exec (sched_exec).
4679 *
4680 * So we race with normal scheduler movements, but that's OK, as long
4681 * as the task is no longer on this CPU.
4682 *
4683 * Returns non-zero if task was successfully migrated.
4684 */
4685static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4686{
4687        struct rq *rq_dest, *rq_src;
4688        int ret = 0;
4689
4690        if (unlikely(!cpu_active(dest_cpu)))
4691                return ret;
4692
4693        rq_src = cpu_rq(src_cpu);
4694        rq_dest = cpu_rq(dest_cpu);
4695
4696        raw_spin_lock(&p->pi_lock);
4697        double_rq_lock(rq_src, rq_dest);
4698        /* Already moved. */
4699        if (task_cpu(p) != src_cpu)
4700                goto done;
4701        /* Affinity changed (again). */
4702        if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4703                goto fail;
4704
4705        /*
4706         * If we're not on a rq, the next wake-up will ensure we're
4707         * placed properly.
4708         */
4709        if (p->on_rq) {
4710                dequeue_task(rq_src, p, 0);
4711                set_task_cpu(p, dest_cpu);
4712                enqueue_task(rq_dest, p, 0);
4713                check_preempt_curr(rq_dest, p, 0);
4714        }
4715done:
4716        ret = 1;
4717fail:
4718        double_rq_unlock(rq_src, rq_dest);
4719        raw_spin_unlock(&p->pi_lock);
4720        return ret;
4721}
4722
4723/*
4724 * migration_cpu_stop - this will be executed by a highprio stopper thread
4725 * and performs thread migration by bumping thread off CPU then
4726 * 'pushing' onto another runqueue.
4727 */
4728static int migration_cpu_stop(void *data)
4729{
4730        struct migration_arg *arg = data;
4731
4732        /*
4733         * The original target cpu might have gone down and we might
4734         * be on another cpu but it doesn't matter.
4735         */
4736        local_irq_disable();
4737        __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4738        local_irq_enable();
4739        return 0;
4740}
4741
4742#ifdef CONFIG_HOTPLUG_CPU
4743
4744/*
4745 * Ensures that the idle task is using init_mm right before its cpu goes
4746 * offline.
4747 */
4748void idle_task_exit(void)
4749{
4750        struct mm_struct *mm = current->active_mm;
4751
4752        BUG_ON(cpu_online(smp_processor_id()));
4753
4754        if (mm != &init_mm)
4755                switch_mm(mm, &init_mm, current);
4756        mmdrop(mm);
4757}
4758
4759/*
4760 * Since this CPU is going 'away' for a while, fold any nr_active delta
4761 * we might have. Assumes we're called after migrate_tasks() so that the
4762 * nr_active count is stable.
4763 *
4764 * Also see the comment "Global load-average calculations".
4765 */
4766static void calc_load_migrate(struct rq *rq)
4767{
4768        long delta = calc_load_fold_active(rq);
4769        if (delta)
4770                atomic_long_add(delta, &calc_load_tasks);
4771}
4772
4773/*
4774 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4775 * try_to_wake_up()->select_task_rq().
4776 *
4777 * Called with rq->lock held even though we'er in stop_machine() and
4778 * there's no concurrency possible, we hold the required locks anyway
4779 * because of lock validation efforts.
4780 */
4781static void migrate_tasks(unsigned int dead_cpu)
4782{
4783        struct rq *rq = cpu_rq(dead_cpu);
4784        struct task_struct *next, *stop = rq->stop;
4785        int dest_cpu;
4786
4787        /*
4788         * Fudge the rq selection such that the below task selection loop
4789         * doesn't get stuck on the currently eligible stop task.
4790         *
4791         * We're currently inside stop_machine() and the rq is either stuck
4792         * in the stop_machine_cpu_stop() loop, or we're executing this code,
4793         * either way we should never end up calling schedule() until we're
4794         * done here.
4795         */
4796        rq->stop = NULL;
4797
4798        for ( ; ; ) {
4799                /*
4800                 * There's this thread running, bail when that's the only
4801                 * remaining thread.
4802                 */
4803                if (rq->nr_running == 1)
4804                        break;
4805
4806                next = pick_next_task(rq);
4807                BUG_ON(!next);
4808                next->sched_class->put_prev_task(rq, next);
4809
4810                /* Find suitable destination for @next, with force if needed. */
4811                dest_cpu = select_fallback_rq(dead_cpu, next);
4812                raw_spin_unlock(&rq->lock);
4813
4814                __migrate_task(next, dead_cpu, dest_cpu);
4815
4816                raw_spin_lock(&rq->lock);
4817        }
4818
4819        rq->stop = stop;
4820}
4821
4822#endif /* CONFIG_HOTPLUG_CPU */
4823
4824#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4825
4826static struct ctl_table sd_ctl_dir[] = {
4827        {
4828                .procname       = "sched_domain",
4829                .mode           = 0555,
4830        },
4831        {}
4832};
4833
4834static struct ctl_table sd_ctl_root[] = {
4835        {
4836                .procname       = "kernel",
4837                .mode           = 0555,
4838                .child          = sd_ctl_dir,
4839        },
4840        {}
4841};
4842
4843static struct ctl_table *sd_alloc_ctl_entry(int n)
4844{
4845        struct ctl_table *entry =
4846                kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4847
4848        return entry;
4849}
4850
4851static void sd_free_ctl_entry(struct ctl_table **tablep)
4852{
4853        struct ctl_table *entry;
4854
4855        /*
4856         * In the intermediate directories, both the child directory and
4857         * procname are dynamically allocated and could fail but the mode
4858         * will always be set. In the lowest directory the names are
4859         * static strings and all have proc handlers.
4860         */
4861        for (entry = *tablep; entry->mode; entry++) {
4862                if (entry->child)
4863                        sd_free_ctl_entry(&entry->child);
4864                if (entry->proc_handler == NULL)
4865                        kfree(entry->procname);
4866        }
4867
4868        kfree(*tablep);
4869        *tablep = NULL;
4870}
4871
4872static int min_load_idx = 0;
4873static int max_load_idx = CPU_LOAD_IDX_MAX;
4874
4875static void
4876set_table_entry(struct ctl_table *entry,
4877                const char *procname, void *data, int maxlen,
4878                umode_t mode, proc_handler *proc_handler,
4879                bool load_idx)
4880{
4881        entry->procname = procname;
4882        entry->data = data;
4883        entry->maxlen = maxlen;
4884        entry->mode = mode;
4885        entry->proc_handler = proc_handler;
4886
4887        if (load_idx) {
4888                entry->extra1 = &min_load_idx;
4889                entry->extra2 = &max_load_idx;
4890        }
4891}
4892
4893static struct ctl_table *
4894sd_alloc_ctl_domain_table(struct sched_domain *sd)
4895{
4896        struct ctl_table *table = sd_alloc_ctl_entry(13);
4897
4898        if (table == NULL)
4899                return NULL;
4900
4901        set_table_entry(&table[0], "min_interval", &sd->min_interval,
4902                sizeof(long), 0644, proc_doulongvec_minmax, false);
4903        set_table_entry(&table[1], "max_interval", &sd->max_interval,
4904                sizeof(long), 0644, proc_doulongvec_minmax, false);
4905        set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4906                sizeof(int), 0644, proc_dointvec_minmax, true);
4907        set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4908                sizeof(int), 0644, proc_dointvec_minmax, true);
4909        set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4910                sizeof(int), 0644, proc_dointvec_minmax, true);
4911        set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4912                sizeof(int), 0644, proc_dointvec_minmax, true);
4913        set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4914                sizeof(int), 0644, proc_dointvec_minmax, true);
4915        set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4916                sizeof(int), 0644, proc_dointvec_minmax, false);
4917        set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4918                sizeof(int), 0644, proc_dointvec_minmax, false);
4919        set_table_entry(&table[9], "cache_nice_tries",
4920                &sd->cache_nice_tries,
4921                sizeof(int), 0644, proc_dointvec_minmax, false);
4922        set_table_entry(&table[10], "flags", &sd->flags,
4923                sizeof(int), 0644, proc_dointvec_minmax, false);
4924        set_table_entry(&table[11], "name", sd->name,
4925                CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4926        /* &table[12] is terminator */
4927
4928        return table;
4929}
4930
4931static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4932{
4933        struct ctl_table *entry, *table;
4934        struct sched_domain *sd;
4935        int domain_num = 0, i;
4936        char buf[32];
4937
4938        for_each_domain(cpu, sd)
4939                domain_num++;
4940        entry = table = sd_alloc_ctl_entry(domain_num + 1);
4941        if (table == NULL)
4942                return NULL;
4943
4944        i = 0;
4945        for_each_domain(cpu, sd) {
4946                snprintf(buf, 32, "domain%d", i);
4947                entry->procname = kstrdup(buf, GFP_KERNEL);
4948                entry->mode = 0555;
4949                entry->child = sd_alloc_ctl_domain_table(sd);
4950                entry++;
4951                i++;
4952        }
4953        return table;
4954}
4955
4956static struct ctl_table_header *sd_sysctl_header;
4957static void register_sched_domain_sysctl(void)
4958{
4959        int i, cpu_num = num_possible_cpus();
4960        struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4961        char buf[32];
4962
4963        WARN_ON(sd_ctl_dir[0].child);
4964        sd_ctl_dir[0].child = entry;
4965
4966        if (entry == NULL)
4967                return;
4968
4969        for_each_possible_cpu(i) {
4970                snprintf(buf, 32, "cpu%d", i);
4971                entry->procname = kstrdup(buf, GFP_KERNEL);
4972                entry->mode = 0555;
4973                entry->child = sd_alloc_ctl_cpu_table(i);
4974                entry++;
4975        }
4976
4977        WARN_ON(sd_sysctl_header);
4978        sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4979}
4980
4981/* may be called multiple times per register */
4982static void unregister_sched_domain_sysctl(void)
4983{
4984        if (sd_sysctl_header)
4985                unregister_sysctl_table(sd_sysctl_header);
4986        sd_sysctl_header = NULL;
4987        if (sd_ctl_dir[0].child)
4988                sd_free_ctl_entry(&sd_ctl_dir[0].child);
4989}
4990#else
4991static void register_sched_domain_sysctl(void)
4992{
4993}
4994static void unregister_sched_domain_sysctl(void)
4995{
4996}
4997#endif
4998
4999static void set_rq_online(struct rq *rq)
5000{
5001        if (!rq->online) {
5002                const struct sched_class *class;
5003
5004                cpumask_set_cpu(rq->cpu, rq->rd->online);
5005                rq->online = 1;
5006
5007                for_each_class(class) {
5008                        if (class->rq_online)
5009                                class->rq_online(rq);
5010                }
5011        }
5012}
5013
5014static void set_rq_offline(struct rq *rq)
5015{
5016        if (rq->online) {
5017                const struct sched_class *class;
5018
5019                for_each_class(class) {
5020                        if (class->rq_offline)
5021                                class->rq_offline(rq);
5022                }
5023
5024                cpumask_clear_cpu(rq->cpu, rq->rd->online);
5025                rq->online = 0;
5026        }
5027}
5028
5029/*
5030 * migration_call - callback that gets triggered when a CPU is added.
5031 * Here we can start up the necessary migration thread for the new CPU.
5032 */
5033static int __cpuinit
5034migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5035{
5036        int cpu = (long)hcpu;
5037        unsigned long flags;
5038        struct rq *rq = cpu_rq(cpu);
5039
5040        switch (action & ~CPU_TASKS_FROZEN) {
5041
5042        case CPU_UP_PREPARE:
5043                rq->calc_load_update = calc_load_update;
5044                break;
5045
5046        case CPU_ONLINE:
5047                /* Update our root-domain */
5048                raw_spin_lock_irqsave(&rq->lock, flags);
5049                if (rq->rd) {
5050                        BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5051
5052                        set_rq_online(rq);
5053                }
5054                raw_spin_unlock_irqrestore(&rq->lock, flags);
5055                break;
5056
5057#ifdef CONFIG_HOTPLUG_CPU
5058        case CPU_DYING:
5059                sched_ttwu_pending();
5060                /* Update our root-domain */
5061                raw_spin_lock_irqsave(&rq->lock, flags);
5062                if (rq->rd) {
5063                        BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5064                        set_rq_offline(rq);
5065                }
5066                migrate_tasks(cpu);
5067                BUG_ON(rq->nr_running != 1); /* the migration thread */
5068                raw_spin_unlock_irqrestore(&rq->lock, flags);
5069                break;
5070
5071        case CPU_DEAD:
5072                calc_load_migrate(rq);
5073                break;
5074#endif
5075        }
5076
5077        update_max_interval();
5078
5079        return NOTIFY_OK;
5080}
5081
5082/*
5083 * Register at high priority so that task migration (migrate_all_tasks)
5084 * happens before everything else.  This has to be lower priority than
5085 * the notifier in the perf_event subsystem, though.
5086 */
5087static struct notifier_block __cpuinitdata migration_notifier = {
5088        .notifier_call = migration_call,
5089        .priority = CPU_PRI_MIGRATION,
5090};
5091
5092static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5093                                      unsigned long action, void *hcpu)
5094{
5095        switch (action & ~CPU_TASKS_FROZEN) {
5096        case CPU_STARTING:
5097        case CPU_DOWN_FAILED:
5098                set_cpu_active((long)hcpu, true);
5099                return NOTIFY_OK;
5100        default:
5101                return NOTIFY_DONE;
5102        }
5103}
5104
5105static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5106                                        unsigned long action, void *hcpu)
5107{
5108        switch (action & ~CPU_TASKS_FROZEN) {
5109        case CPU_DOWN_PREPARE:
5110                set_cpu_active((long)hcpu, false);
5111                return NOTIFY_OK;
5112        default:
5113                return NOTIFY_DONE;
5114        }
5115}
5116
5117static int __init migration_init(void)
5118{
5119        void *cpu = (void *)(long)smp_processor_id();
5120        int err;
5121
5122        /* Initialize migration for the boot CPU */
5123        err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5124        BUG_ON(err == NOTIFY_BAD);
5125        migration_call(&migration_notifier, CPU_ONLINE, cpu);
5126        register_cpu_notifier(&migration_notifier);
5127
5128        /* Register cpu active notifiers */
5129        cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5130        cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5131
5132        return 0;
5133}
5134early_initcall(migration_init);
5135#endif
5136
5137#ifdef CONFIG_SMP
5138
5139static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5140
5141#ifdef CONFIG_SCHED_DEBUG
5142
5143static __read_mostly int sched_debug_enabled;
5144
5145static int __init sched_debug_setup(char *str)
5146{
5147        sched_debug_enabled = 1;
5148
5149        return 0;
5150}
5151early_param("sched_debug", sched_debug_setup);
5152
5153static inline bool sched_debug(void)
5154{
5155        return sched_debug_enabled;
5156}
5157
5158static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5159                                  struct cpumask *groupmask)
5160{
5161        struct sched_group *group = sd->groups;
5162        char str[256];
5163
5164        cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5165        cpumask_clear(groupmask);
5166
5167        printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5168
5169        if (!(sd->flags & SD_LOAD_BALANCE)) {
5170                printk("does not load-balance\n");
5171                if (sd->parent)
5172                        printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5173                                        " has parent");
5174                return -1;
5175        }
5176
5177        printk(KERN_CONT "span %s level %s\n", str, sd->name);
5178
5179        if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5180                printk(KERN_ERR "ERROR: domain->span does not contain "
5181                                "CPU%d\n", cpu);
5182        }
5183        if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5184                printk(KERN_ERR "ERROR: domain->groups does not contain"
5185                                " CPU%d\n", cpu);
5186        }
5187
5188        printk(KERN_DEBUG "%*s groups:", level + 1, "");
5189        do {
5190                if (!group) {
5191                        printk("\n");
5192                        printk(KERN_ERR "ERROR: group is NULL\n");
5193                        break;
5194                }
5195
5196                /*
5197                 * Even though we initialize ->power to something semi-sane,
5198                 * we leave power_orig unset. This allows us to detect if
5199                 * domain iteration is still funny without causing /0 traps.
5200                 */
5201                if (!group->sgp->power_orig) {
5202                        printk(KERN_CONT "\n");
5203                        printk(KERN_ERR "ERROR: domain->cpu_power not "
5204                                        "set\n");
5205                        break;
5206                }
5207
5208                if (!cpumask_weight(sched_group_cpus(group))) {
5209                        printk(KERN_CONT "\n");
5210                        printk(KERN_ERR "ERROR: empty group\n");
5211                        break;
5212                }
5213
5214                if (!(sd->flags & SD_OVERLAP) &&
5215                    cpumask_intersects(groupmask, sched_group_cpus(group))) {
5216                        printk(KERN_CONT "\n");
5217                        printk(KERN_ERR "ERROR: repeated CPUs\n");
5218                        break;
5219                }
5220
5221                cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5222
5223                cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5224
5225                printk(KERN_CONT " %s", str);
5226                if (group->sgp->power != SCHED_POWER_SCALE) {
5227                        printk(KERN_CONT " (cpu_power = %d)",
5228                                group->sgp->power);
5229                }
5230
5231                group = group->next;
5232        } while (group != sd->groups);
5233        printk(KERN_CONT "\n");
5234
5235        if (!cpumask_equal(sched_domain_span(sd), groupmask))
5236                printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5237
5238        if (sd->parent &&
5239            !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5240                printk(KERN_ERR "ERROR: parent span is not a superset "
5241                        "of domain->span\n");
5242        return 0;
5243}
5244
5245static void sched_domain_debug(struct sched_domain *sd, int cpu)
5246{
5247        int level = 0;
5248
5249        if (!sched_debug_enabled)
5250                return;
5251
5252        if (!sd) {
5253                printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5254                return;
5255        }
5256
5257        printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5258
5259        for (;;) {
5260                if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5261                        break;
5262                level++;
5263                sd = sd->parent;
5264                if (!sd)
5265                        break;
5266        }
5267}
5268#else /* !CONFIG_SCHED_DEBUG */
5269# define sched_domain_debug(sd, cpu) do { } while (0)
5270static inline bool sched_debug(void)
5271{
5272        return false;
5273}
5274#endif /* CONFIG_SCHED_DEBUG */
5275
5276static int sd_degenerate(struct sched_domain *sd)
5277{
5278        if (cpumask_weight(sched_domain_span(sd)) == 1)
5279                return 1;
5280
5281        /* Following flags need at least 2 groups */
5282        if (sd->flags & (SD_LOAD_BALANCE |
5283                         SD_BALANCE_NEWIDLE |
5284                         SD_BALANCE_FORK |
5285                         SD_BALANCE_EXEC |
5286                         SD_SHARE_CPUPOWER |
5287                         SD_SHARE_PKG_RESOURCES)) {
5288                if (sd->groups != sd->groups->next)
5289                        return 0;
5290        }
5291
5292        /* Following flags don't use groups */
5293        if (sd->flags & (SD_WAKE_AFFINE))
5294                return 0;
5295
5296        return 1;
5297}
5298
5299static int
5300sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5301{
5302        unsigned long cflags = sd->flags, pflags = parent->flags;
5303
5304        if (sd_degenerate(parent))
5305                return 1;
5306
5307        if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5308                return 0;
5309
5310        /* Flags needing groups don't count if only 1 group in parent */
5311        if (parent->groups == parent->groups->next) {
5312                pflags &= ~(SD_LOAD_BALANCE |
5313                                SD_BALANCE_NEWIDLE |
5314                                SD_BALANCE_FORK |
5315                                SD_BALANCE_EXEC |
5316                                SD_SHARE_CPUPOWER |
5317                                SD_SHARE_PKG_RESOURCES);
5318                if (nr_node_ids == 1)
5319                        pflags &= ~SD_SERIALIZE;
5320        }
5321        if (~cflags & pflags)
5322                return 0;
5323
5324        return 1;
5325}
5326
5327static void free_rootdomain(struct rcu_head *rcu)
5328{
5329        struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5330
5331        cpupri_cleanup(&rd->cpupri);
5332        free_cpumask_var(rd->rto_mask);
5333        free_cpumask_var(rd->online);
5334        free_cpumask_var(rd->span);
5335        kfree(rd);
5336}
5337
5338static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5339{
5340        struct root_domain *old_rd = NULL;
5341        unsigned long flags;
5342
5343        raw_spin_lock_irqsave(&rq->lock, flags);
5344
5345        if (rq->rd) {
5346                old_rd = rq->rd;
5347
5348                if (cpumask_test_cpu(rq->cpu, old_rd->online))
5349                        set_rq_offline(rq);
5350
5351                cpumask_clear_cpu(rq->cpu, old_rd->span);
5352
5353                /*
5354                 * If we dont want to free the old_rt yet then
5355                 * set old_rd to NULL to skip the freeing later
5356                 * in this function:
5357                 */
5358                if (!atomic_dec_and_test(&old_rd->refcount))
5359                        old_rd = NULL;
5360        }
5361
5362        atomic_inc(&rd->refcount);
5363        rq->rd = rd;
5364
5365        cpumask_set_cpu(rq->cpu, rd->span);
5366        if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5367                set_rq_online(rq);
5368
5369        raw_spin_unlock_irqrestore(&rq->lock, flags);
5370
5371        if (old_rd)
5372                call_rcu_sched(&old_rd->rcu, free_rootdomain);
5373}
5374
5375static int init_rootdomain(struct root_domain *rd)
5376{
5377        memset(rd, 0, sizeof(*rd));
5378
5379        if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5380                goto out;
5381        if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5382                goto free_span;
5383        if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5384                goto free_online;
5385
5386        if (cpupri_init(&rd->cpupri) != 0)
5387                goto free_rto_mask;
5388        return 0;
5389
5390free_rto_mask:
5391        free_cpumask_var(rd->rto_mask);
5392free_online:
5393        free_cpumask_var(rd->online);
5394free_span:
5395        free_cpumask_var(rd->span);
5396out:
5397        return -ENOMEM;
5398}
5399
5400/*
5401 * By default the system creates a single root-domain with all cpus as
5402 * members (mimicking the global state we have today).
5403 */
5404struct root_domain def_root_domain;
5405
5406static void init_defrootdomain(void)
5407{
5408        init_rootdomain(&def_root_domain);
5409
5410        atomic_set(&def_root_domain.refcount, 1);
5411}
5412
5413static struct root_domain *alloc_rootdomain(void)
5414{
5415        struct root_domain *rd;
5416
5417        rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5418        if (!rd)
5419                return NULL;
5420
5421        if (init_rootdomain(rd) != 0) {
5422                kfree(rd);
5423                return NULL;
5424        }
5425
5426        return rd;
5427}
5428
5429static void free_sched_groups(struct sched_group *sg, int free_sgp)
5430{
5431        struct sched_group *tmp, *first;
5432
5433        if (!sg)
5434                return;
5435
5436        first = sg;
5437        do {
5438                tmp = sg->next;
5439
5440                if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5441                        kfree(sg->sgp);
5442
5443                kfree(sg);
5444                sg = tmp;
5445        } while (sg != first);
5446}
5447
5448static void free_sched_domain(struct rcu_head *rcu)
5449{
5450        struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5451
5452        /*
5453         * If its an overlapping domain it has private groups, iterate and
5454         * nuke them all.
5455         */
5456        if (sd->flags & SD_OVERLAP) {
5457                free_sched_groups(sd->groups, 1);
5458        } else if (atomic_dec_and_test(&sd->groups->ref)) {
5459                kfree(sd->groups->sgp);
5460                kfree(sd->groups);
5461        }
5462        kfree(sd);
5463}
5464
5465static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5466{
5467        call_rcu(&sd->rcu, free_sched_domain);
5468}
5469
5470static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5471{
5472        for (; sd; sd = sd->parent)
5473                destroy_sched_domain(sd, cpu);
5474}
5475
5476/*
5477 * Keep a special pointer to the highest sched_domain that has
5478 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5479 * allows us to avoid some pointer chasing select_idle_sibling().
5480 *
5481 * Also keep a unique ID per domain (we use the first cpu number in
5482 * the cpumask of the domain), this allows us to quickly tell if
5483 * two cpus are in the same cache domain, see cpus_share_cache().
5484 */
5485DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5486DEFINE_PER_CPU(int, sd_llc_id);
5487
5488static void update_top_cache_domain(int cpu)
5489{
5490        struct sched_domain *sd;
5491        int id = cpu;
5492
5493        sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5494        if (sd)
5495                id = cpumask_first(sched_domain_span(sd));
5496
5497        rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5498        per_cpu(sd_llc_id, cpu) = id;
5499}
5500
5501/*
5502 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5503 * hold the hotplug lock.
5504 */
5505static void
5506cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5507{
5508        struct rq *rq = cpu_rq(cpu);
5509        struct sched_domain *tmp;
5510
5511        /* Remove the sched domains which do not contribute to scheduling. */
5512        for (tmp = sd; tmp; ) {
5513                struct sched_domain *parent = tmp->parent;
5514                if (!parent)
5515                        break;
5516
5517                if (sd_parent_degenerate(tmp, parent)) {
5518                        tmp->parent = parent->parent;
5519                        if (parent->parent)
5520                                parent->parent->child = tmp;
5521                        destroy_sched_domain(parent, cpu);
5522                } else
5523                        tmp = tmp->parent;
5524        }
5525
5526        if (sd && sd_degenerate(sd)) {
5527                tmp = sd;
5528                sd = sd->parent;
5529                destroy_sched_domain(tmp, cpu);
5530                if (sd)
5531                        sd->child = NULL;
5532        }
5533
5534        sched_domain_debug(sd, cpu);
5535
5536        rq_attach_root(rq, rd);
5537        tmp = rq->sd;
5538        rcu_assign_pointer(rq->sd, sd);
5539        destroy_sched_domains(tmp, cpu);
5540
5541        update_top_cache_domain(cpu);
5542}
5543
5544/* cpus with isolated domains */
5545static cpumask_var_t cpu_isolated_map;
5546
5547/* Setup the mask of cpus configured for isolated domains */
5548static int __init isolated_cpu_setup(char *str)
5549{
5550        alloc_bootmem_cpumask_var(&cpu_isolated_map);
5551        cpulist_parse(str, cpu_isolated_map);
5552        return 1;
5553}
5554
5555__setup("isolcpus=", isolated_cpu_setup);
5556
5557static const struct cpumask *cpu_cpu_mask(int cpu)
5558{
5559        return cpumask_of_node(cpu_to_node(cpu));
5560}
5561
5562struct sd_data {
5563        struct sched_domain **__percpu sd;
5564        struct sched_group **__percpu sg;
5565        struct sched_group_power **__percpu sgp;
5566};
5567
5568struct s_data {
5569        struct sched_domain ** __percpu sd;
5570        struct root_domain      *rd;
5571};
5572
5573enum s_alloc {
5574        sa_rootdomain,
5575        sa_sd,
5576        sa_sd_storage,
5577        sa_none,
5578};
5579
5580struct sched_domain_topology_level;
5581
5582typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5583typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5584
5585#define SDTL_OVERLAP    0x01
5586
5587struct sched_domain_topology_level {
5588        sched_domain_init_f init;
5589        sched_domain_mask_f mask;
5590        int                 flags;
5591        int                 numa_level;
5592        struct sd_data      data;
5593};
5594
5595/*
5596 * Build an iteration mask that can exclude certain CPUs from the upwards
5597 * domain traversal.
5598 *
5599 * Asymmetric node setups can result in situations where the domain tree is of
5600 * unequal depth, make sure to skip domains that already cover the entire
5601 * range.
5602 *
5603 * In that case build_sched_domains() will have terminated the iteration early
5604 * and our sibling sd spans will be empty. Domains should always include the
5605 * cpu they're built on, so check that.
5606 *
5607 */
5608static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5609{
5610        const struct cpumask *span = sched_domain_span(sd);
5611        struct sd_data *sdd = sd->private;
5612        struct sched_domain *sibling;
5613        int i;
5614
5615        for_each_cpu(i, span) {
5616                sibling = *per_cpu_ptr(sdd->sd, i);
5617                if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5618                        continue;
5619
5620                cpumask_set_cpu(i, sched_group_mask(sg));
5621        }
5622}
5623
5624/*
5625 * Return the canonical balance cpu for this group, this is the first cpu
5626 * of this group that's also in the iteration mask.
5627 */
5628int group_balance_cpu(struct sched_group *sg)
5629{
5630        return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5631}
5632
5633static int
5634build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5635{
5636        struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5637        const struct cpumask *span = sched_domain_span(sd);
5638        struct cpumask *covered = sched_domains_tmpmask;
5639        struct sd_data *sdd = sd->private;
5640        struct sched_domain *child;
5641        int i;
5642
5643        cpumask_clear(covered);
5644
5645        for_each_cpu(i, span) {
5646                struct cpumask *sg_span;
5647
5648                if (cpumask_test_cpu(i, covered))
5649                        continue;
5650
5651                child = *per_cpu_ptr(sdd->sd, i);
5652
5653                /* See the comment near build_group_mask(). */
5654                if (!cpumask_test_cpu(i, sched_domain_span(child)))
5655                        continue;
5656
5657                sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5658                                GFP_KERNEL, cpu_to_node(cpu));
5659
5660                if (!sg)
5661                        goto fail;
5662
5663                sg_span = sched_group_cpus(sg);
5664                if (child->child) {
5665                        child = child->child;
5666                        cpumask_copy(sg_span, sched_domain_span(child));
5667                } else
5668                        cpumask_set_cpu(i, sg_span);
5669
5670                cpumask_or(covered, covered, sg_span);
5671
5672                sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5673                if (atomic_inc_return(&sg->sgp->ref) == 1)
5674                        build_group_mask(sd, sg);
5675
5676                /*
5677                 * Initialize sgp->power such that even if we mess up the
5678                 * domains and no possible iteration will get us here, we won't
5679                 * die on a /0 trap.
5680                 */
5681                sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5682
5683                /*
5684                 * Make sure the first group of this domain contains the
5685                 * canonical balance cpu. Otherwise the sched_domain iteration
5686                 * breaks. See update_sg_lb_stats().
5687                 */
5688                if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5689                    group_balance_cpu(sg) == cpu)
5690                        groups = sg;
5691
5692                if (!first)
5693                        first = sg;
5694                if (last)
5695                        last->next = sg;
5696                last = sg;
5697                last->next = first;
5698        }
5699        sd->groups = groups;
5700
5701        return 0;
5702
5703fail:
5704        free_sched_groups(first, 0);
5705
5706        return -ENOMEM;
5707}
5708
5709static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5710{
5711        struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5712        struct sched_domain *child = sd->child;
5713
5714        if (child)
5715                cpu = cpumask_first(sched_domain_span(child));
5716
5717        if (sg) {
5718                *sg = *per_cpu_ptr(sdd->sg, cpu);
5719                (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5720                atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5721        }
5722
5723        return cpu;
5724}
5725
5726/*
5727 * build_sched_groups will build a circular linked list of the groups
5728 * covered by the given span, and will set each group's ->cpumask correctly,
5729 * and ->cpu_power to 0.
5730 *
5731 * Assumes the sched_domain tree is fully constructed
5732 */
5733static int
5734build_sched_groups(struct sched_domain *sd, int cpu)
5735{
5736        struct sched_group *first = NULL, *last = NULL;
5737        struct sd_data *sdd = sd->private;
5738        const struct cpumask *span = sched_domain_span(sd);
5739        struct cpumask *covered;
5740        int i;
5741
5742        get_group(cpu, sdd, &sd->groups);
5743        atomic_inc(&sd->groups->ref);
5744
5745        if (cpu != cpumask_first(sched_domain_span(sd)))
5746                return 0;
5747
5748        lockdep_assert_held(&sched_domains_mutex);
5749        covered = sched_domains_tmpmask;
5750
5751        cpumask_clear(covered);
5752
5753        for_each_cpu(i, span) {
5754                struct sched_group *sg;
5755                int group = get_group(i, sdd, &sg);
5756                int j;
5757
5758                if (cpumask_test_cpu(i, covered))
5759                        continue;
5760
5761                cpumask_clear(sched_group_cpus(sg));
5762                sg->sgp->power = 0;
5763                cpumask_setall(sched_group_mask(sg));
5764
5765                for_each_cpu(j, span) {
5766                        if (get_group(j, sdd, NULL) != group)
5767                                continue;
5768
5769                        cpumask_set_cpu(j, covered);
5770                        cpumask_set_cpu(j, sched_group_cpus(sg));
5771                }
5772
5773                if (!first)
5774                        first = sg;
5775                if (last)
5776                        last->next = sg;
5777                last = sg;
5778        }
5779        last->next = first;
5780
5781        return 0;
5782}
5783
5784/*
5785 * Initialize sched groups cpu_power.
5786 *
5787 * cpu_power indicates the capacity of sched group, which is used while
5788 * distributing the load between different sched groups in a sched domain.
5789 * Typically cpu_power for all the groups in a sched domain will be same unless
5790 * there are asymmetries in the topology. If there are asymmetries, group
5791 * having more cpu_power will pickup more load compared to the group having
5792 * less cpu_power.
5793 */
5794static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5795{
5796        struct sched_group *sg = sd->groups;
5797
5798        WARN_ON(!sd || !sg);
5799
5800        do {
5801                sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5802                sg = sg->next;
5803        } while (sg != sd->groups);
5804
5805        if (cpu != group_balance_cpu(sg))
5806                return;
5807
5808        update_group_power(sd, cpu);
5809        atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5810}
5811
5812int __weak arch_sd_sibling_asym_packing(void)
5813{
5814       return 0*SD_ASYM_PACKING;
5815}
5816
5817/*
5818 * Initializers for schedule domains
5819 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5820 */
5821
5822#ifdef CONFIG_SCHED_DEBUG
5823# define SD_INIT_NAME(sd, type)         sd->name = #type
5824#else
5825# define SD_INIT_NAME(sd, type)         do { } while (0)
5826#endif
5827
5828#define SD_INIT_FUNC(type)                                              \
5829static noinline struct sched_domain *                                   \
5830sd_init_##type(struct sched_domain_topology_level *tl, int cpu)         \
5831{                                                                       \
5832        struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);       \
5833        *sd = SD_##type##_INIT;                                         \
5834        SD_INIT_NAME(sd, type);                                         \
5835        sd->private = &tl->data;                                        \
5836        return sd;                                                      \
5837}
5838
5839SD_INIT_FUNC(CPU)
5840#ifdef CONFIG_SCHED_SMT
5841 SD_INIT_FUNC(SIBLING)
5842#endif
5843#ifdef CONFIG_SCHED_MC
5844 SD_INIT_FUNC(MC)
5845#endif
5846#ifdef CONFIG_SCHED_BOOK
5847 SD_INIT_FUNC(BOOK)
5848#endif
5849
5850static int default_relax_domain_level = -1;
5851int sched_domain_level_max;
5852
5853static int __init setup_relax_domain_level(char *str)
5854{
5855        if (kstrtoint(str, 0, &default_relax_domain_level))
5856                pr_warn("Unable to set relax_domain_level\n");
5857
5858        return 1;
5859}
5860__setup("relax_domain_level=", setup_relax_domain_level);
5861
5862static void set_domain_attribute(struct sched_domain *sd,
5863                                 struct sched_domain_attr *attr)
5864{
5865        int request;
5866