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