linux/kernel/time/tick-sched.c
<<
>>
Prefs
   1/*
   2 *  linux/kernel/time/tick-sched.c
   3 *
   4 *  Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
   5 *  Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
   6 *  Copyright(C) 2006-2007  Timesys Corp., Thomas Gleixner
   7 *
   8 *  No idle tick implementation for low and high resolution timers
   9 *
  10 *  Started by: Thomas Gleixner and Ingo Molnar
  11 *
  12 *  Distribute under GPLv2.
  13 */
  14#include <linux/cpu.h>
  15#include <linux/err.h>
  16#include <linux/hrtimer.h>
  17#include <linux/interrupt.h>
  18#include <linux/kernel_stat.h>
  19#include <linux/percpu.h>
  20#include <linux/profile.h>
  21#include <linux/sched.h>
  22#include <linux/module.h>
  23
  24#include <asm/irq_regs.h>
  25
  26#include "tick-internal.h"
  27
  28/*
  29 * Per cpu nohz control structure
  30 */
  31static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
  32
  33/*
  34 * The time, when the last jiffy update happened. Protected by xtime_lock.
  35 */
  36static ktime_t last_jiffies_update;
  37
  38struct tick_sched *tick_get_tick_sched(int cpu)
  39{
  40        return &per_cpu(tick_cpu_sched, cpu);
  41}
  42
  43/*
  44 * Must be called with interrupts disabled !
  45 */
  46static void tick_do_update_jiffies64(ktime_t now)
  47{
  48        unsigned long ticks = 0;
  49        ktime_t delta;
  50
  51        /*
  52         * Do a quick check without holding xtime_lock:
  53         */
  54        delta = ktime_sub(now, last_jiffies_update);
  55        if (delta.tv64 < tick_period.tv64)
  56                return;
  57
  58        /* Reevalute with xtime_lock held */
  59        write_seqlock(&xtime_lock);
  60
  61        delta = ktime_sub(now, last_jiffies_update);
  62        if (delta.tv64 >= tick_period.tv64) {
  63
  64                delta = ktime_sub(delta, tick_period);
  65                last_jiffies_update = ktime_add(last_jiffies_update,
  66                                                tick_period);
  67
  68                /* Slow path for long timeouts */
  69                if (unlikely(delta.tv64 >= tick_period.tv64)) {
  70                        s64 incr = ktime_to_ns(tick_period);
  71
  72                        ticks = ktime_divns(delta, incr);
  73
  74                        last_jiffies_update = ktime_add_ns(last_jiffies_update,
  75                                                           incr * ticks);
  76                }
  77                do_timer(++ticks);
  78
  79                /* Keep the tick_next_period variable up to date */
  80                tick_next_period = ktime_add(last_jiffies_update, tick_period);
  81        }
  82        write_sequnlock(&xtime_lock);
  83}
  84
  85/*
  86 * Initialize and return retrieve the jiffies update.
  87 */
  88static ktime_t tick_init_jiffy_update(void)
  89{
  90        ktime_t period;
  91
  92        write_seqlock(&xtime_lock);
  93        /* Did we start the jiffies update yet ? */
  94        if (last_jiffies_update.tv64 == 0)
  95                last_jiffies_update = tick_next_period;
  96        period = last_jiffies_update;
  97        write_sequnlock(&xtime_lock);
  98        return period;
  99}
 100
 101/*
 102 * NOHZ - aka dynamic tick functionality
 103 */
 104#ifdef CONFIG_NO_HZ
 105/*
 106 * NO HZ enabled ?
 107 */
 108int tick_nohz_enabled __read_mostly  = 1;
 109
 110/*
 111 * Enable / Disable tickless mode
 112 */
 113static int __init setup_tick_nohz(char *str)
 114{
 115        if (!strcmp(str, "off"))
 116                tick_nohz_enabled = 0;
 117        else if (!strcmp(str, "on"))
 118                tick_nohz_enabled = 1;
 119        else
 120                return 0;
 121        return 1;
 122}
 123
 124__setup("nohz=", setup_tick_nohz);
 125
 126/**
 127 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
 128 *
 129 * Called from interrupt entry when the CPU was idle
 130 *
 131 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
 132 * must be updated. Otherwise an interrupt handler could use a stale jiffy
 133 * value. We do this unconditionally on any cpu, as we don't know whether the
 134 * cpu, which has the update task assigned is in a long sleep.
 135 */
 136static void tick_nohz_update_jiffies(ktime_t now)
 137{
 138        int cpu = smp_processor_id();
 139        struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 140        unsigned long flags;
 141
 142        ts->idle_waketime = now;
 143
 144        local_irq_save(flags);
 145        tick_do_update_jiffies64(now);
 146        local_irq_restore(flags);
 147
 148        touch_softlockup_watchdog();
 149}
 150
 151/*
 152 * Updates the per cpu time idle statistics counters
 153 */
 154static void
 155update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time)
 156{
 157        ktime_t delta;
 158
 159        if (ts->idle_active) {
 160                delta = ktime_sub(now, ts->idle_entrytime);
 161                if (nr_iowait_cpu(cpu) > 0)
 162                        ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
 163                else
 164                        ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
 165                ts->idle_entrytime = now;
 166        }
 167
 168        if (last_update_time)
 169                *last_update_time = ktime_to_us(now);
 170
 171}
 172
 173static void tick_nohz_stop_idle(int cpu, ktime_t now)
 174{
 175        struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 176
 177        update_ts_time_stats(cpu, ts, now, NULL);
 178        ts->idle_active = 0;
 179
 180        sched_clock_idle_wakeup_event(0);
 181}
 182
 183static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
 184{
 185        ktime_t now = ktime_get();
 186
 187        ts->idle_entrytime = now;
 188        ts->idle_active = 1;
 189        sched_clock_idle_sleep_event();
 190        return now;
 191}
 192
 193/**
 194 * get_cpu_idle_time_us - get the total idle time of a cpu
 195 * @cpu: CPU number to query
 196 * @last_update_time: variable to store update time in. Do not update
 197 * counters if NULL.
 198 *
 199 * Return the cummulative idle time (since boot) for a given
 200 * CPU, in microseconds.
 201 *
 202 * This time is measured via accounting rather than sampling,
 203 * and is as accurate as ktime_get() is.
 204 *
 205 * This function returns -1 if NOHZ is not enabled.
 206 */
 207u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
 208{
 209        struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 210        ktime_t now, idle;
 211
 212        if (!tick_nohz_enabled)
 213                return -1;
 214
 215        now = ktime_get();
 216        if (last_update_time) {
 217                update_ts_time_stats(cpu, ts, now, last_update_time);
 218                idle = ts->idle_sleeptime;
 219        } else {
 220                if (ts->idle_active && !nr_iowait_cpu(cpu)) {
 221                        ktime_t delta = ktime_sub(now, ts->idle_entrytime);
 222
 223                        idle = ktime_add(ts->idle_sleeptime, delta);
 224                } else {
 225                        idle = ts->idle_sleeptime;
 226                }
 227        }
 228
 229        return ktime_to_us(idle);
 230
 231}
 232EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
 233
 234/**
 235 * get_cpu_iowait_time_us - get the total iowait time of a cpu
 236 * @cpu: CPU number to query
 237 * @last_update_time: variable to store update time in. Do not update
 238 * counters if NULL.
 239 *
 240 * Return the cummulative iowait time (since boot) for a given
 241 * CPU, in microseconds.
 242 *
 243 * This time is measured via accounting rather than sampling,
 244 * and is as accurate as ktime_get() is.
 245 *
 246 * This function returns -1 if NOHZ is not enabled.
 247 */
 248u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
 249{
 250        struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 251        ktime_t now, iowait;
 252
 253        if (!tick_nohz_enabled)
 254                return -1;
 255
 256        now = ktime_get();
 257        if (last_update_time) {
 258                update_ts_time_stats(cpu, ts, now, last_update_time);
 259                iowait = ts->iowait_sleeptime;
 260        } else {
 261                if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
 262                        ktime_t delta = ktime_sub(now, ts->idle_entrytime);
 263
 264                        iowait = ktime_add(ts->iowait_sleeptime, delta);
 265                } else {
 266                        iowait = ts->iowait_sleeptime;
 267                }
 268        }
 269
 270        return ktime_to_us(iowait);
 271}
 272EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
 273
 274static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
 275                                         ktime_t now, int cpu)
 276{
 277        unsigned long seq, last_jiffies, next_jiffies, delta_jiffies;
 278        ktime_t last_update, expires, ret = { .tv64 = 0 };
 279        unsigned long rcu_delta_jiffies;
 280        struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
 281        u64 time_delta;
 282
 283        /* Read jiffies and the time when jiffies were updated last */
 284        do {
 285                seq = read_seqbegin(&xtime_lock);
 286                last_update = last_jiffies_update;
 287                last_jiffies = jiffies;
 288                time_delta = timekeeping_max_deferment();
 289        } while (read_seqretry(&xtime_lock, seq));
 290
 291        if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || printk_needs_cpu(cpu) ||
 292            arch_needs_cpu(cpu)) {
 293                next_jiffies = last_jiffies + 1;
 294                delta_jiffies = 1;
 295        } else {
 296                /* Get the next timer wheel timer */
 297                next_jiffies = get_next_timer_interrupt(last_jiffies);
 298                delta_jiffies = next_jiffies - last_jiffies;
 299                if (rcu_delta_jiffies < delta_jiffies) {
 300                        next_jiffies = last_jiffies + rcu_delta_jiffies;
 301                        delta_jiffies = rcu_delta_jiffies;
 302                }
 303        }
 304        /*
 305         * Do not stop the tick, if we are only one off
 306         * or if the cpu is required for rcu
 307         */
 308        if (!ts->tick_stopped && delta_jiffies == 1)
 309                goto out;
 310
 311        /* Schedule the tick, if we are at least one jiffie off */
 312        if ((long)delta_jiffies >= 1) {
 313
 314                /*
 315                 * If this cpu is the one which updates jiffies, then
 316                 * give up the assignment and let it be taken by the
 317                 * cpu which runs the tick timer next, which might be
 318                 * this cpu as well. If we don't drop this here the
 319                 * jiffies might be stale and do_timer() never
 320                 * invoked. Keep track of the fact that it was the one
 321                 * which had the do_timer() duty last. If this cpu is
 322                 * the one which had the do_timer() duty last, we
 323                 * limit the sleep time to the timekeeping
 324                 * max_deferement value which we retrieved
 325                 * above. Otherwise we can sleep as long as we want.
 326                 */
 327                if (cpu == tick_do_timer_cpu) {
 328                        tick_do_timer_cpu = TICK_DO_TIMER_NONE;
 329                        ts->do_timer_last = 1;
 330                } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
 331                        time_delta = KTIME_MAX;
 332                        ts->do_timer_last = 0;
 333                } else if (!ts->do_timer_last) {
 334                        time_delta = KTIME_MAX;
 335                }
 336
 337                /*
 338                 * calculate the expiry time for the next timer wheel
 339                 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
 340                 * that there is no timer pending or at least extremely
 341                 * far into the future (12 days for HZ=1000). In this
 342                 * case we set the expiry to the end of time.
 343                 */
 344                if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
 345                        /*
 346                         * Calculate the time delta for the next timer event.
 347                         * If the time delta exceeds the maximum time delta
 348                         * permitted by the current clocksource then adjust
 349                         * the time delta accordingly to ensure the
 350                         * clocksource does not wrap.
 351                         */
 352                        time_delta = min_t(u64, time_delta,
 353                                           tick_period.tv64 * delta_jiffies);
 354                }
 355
 356                if (time_delta < KTIME_MAX)
 357                        expires = ktime_add_ns(last_update, time_delta);
 358                else
 359                        expires.tv64 = KTIME_MAX;
 360
 361                /* Skip reprogram of event if its not changed */
 362                if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
 363                        goto out;
 364
 365                ret = expires;
 366
 367                /*
 368                 * nohz_stop_sched_tick can be called several times before
 369                 * the nohz_restart_sched_tick is called. This happens when
 370                 * interrupts arrive which do not cause a reschedule. In the
 371                 * first call we save the current tick time, so we can restart
 372                 * the scheduler tick in nohz_restart_sched_tick.
 373                 */
 374                if (!ts->tick_stopped) {
 375                        select_nohz_load_balancer(1);
 376                        calc_load_enter_idle();
 377
 378                        ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
 379                        ts->tick_stopped = 1;
 380                }
 381
 382                /*
 383                 * If the expiration time == KTIME_MAX, then
 384                 * in this case we simply stop the tick timer.
 385                 */
 386                 if (unlikely(expires.tv64 == KTIME_MAX)) {
 387                        if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
 388                                hrtimer_cancel(&ts->sched_timer);
 389                        goto out;
 390                }
 391
 392                if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
 393                        hrtimer_start(&ts->sched_timer, expires,
 394                                      HRTIMER_MODE_ABS_PINNED);
 395                        /* Check, if the timer was already in the past */
 396                        if (hrtimer_active(&ts->sched_timer))
 397                                goto out;
 398                } else if (!tick_program_event(expires, 0))
 399                                goto out;
 400                /*
 401                 * We are past the event already. So we crossed a
 402                 * jiffie boundary. Update jiffies and raise the
 403                 * softirq.
 404                 */
 405                tick_do_update_jiffies64(ktime_get());
 406        }
 407        raise_softirq_irqoff(TIMER_SOFTIRQ);
 408out:
 409        ts->next_jiffies = next_jiffies;
 410        ts->last_jiffies = last_jiffies;
 411        ts->sleep_length = ktime_sub(dev->next_event, now);
 412
 413        return ret;
 414}
 415
 416static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
 417{
 418        /*
 419         * If this cpu is offline and it is the one which updates
 420         * jiffies, then give up the assignment and let it be taken by
 421         * the cpu which runs the tick timer next. If we don't drop
 422         * this here the jiffies might be stale and do_timer() never
 423         * invoked.
 424         */
 425        if (unlikely(!cpu_online(cpu))) {
 426                if (cpu == tick_do_timer_cpu)
 427                        tick_do_timer_cpu = TICK_DO_TIMER_NONE;
 428        }
 429
 430        if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
 431                return false;
 432
 433        if (need_resched())
 434                return false;
 435
 436        if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
 437                static int ratelimit;
 438
 439                if (ratelimit < 10) {
 440                        printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
 441                               (unsigned int) local_softirq_pending());
 442                        ratelimit++;
 443                }
 444                return false;
 445        }
 446
 447        return true;
 448}
 449
 450static void __tick_nohz_idle_enter(struct tick_sched *ts)
 451{
 452        ktime_t now, expires;
 453        int cpu = smp_processor_id();
 454
 455        now = tick_nohz_start_idle(cpu, ts);
 456
 457        if (can_stop_idle_tick(cpu, ts)) {
 458                int was_stopped = ts->tick_stopped;
 459
 460                ts->idle_calls++;
 461
 462                expires = tick_nohz_stop_sched_tick(ts, now, cpu);
 463                if (expires.tv64 > 0LL) {
 464                        ts->idle_sleeps++;
 465                        ts->idle_expires = expires;
 466                }
 467
 468                if (!was_stopped && ts->tick_stopped)
 469                        ts->idle_jiffies = ts->last_jiffies;
 470        }
 471}
 472
 473/**
 474 * tick_nohz_idle_enter - stop the idle tick from the idle task
 475 *
 476 * When the next event is more than a tick into the future, stop the idle tick
 477 * Called when we start the idle loop.
 478 *
 479 * The arch is responsible of calling:
 480 *
 481 * - rcu_idle_enter() after its last use of RCU before the CPU is put
 482 *  to sleep.
 483 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up.
 484 */
 485void tick_nohz_idle_enter(void)
 486{
 487        struct tick_sched *ts;
 488
 489        WARN_ON_ONCE(irqs_disabled());
 490
 491        /*
 492         * Update the idle state in the scheduler domain hierarchy
 493         * when tick_nohz_stop_sched_tick() is called from the idle loop.
 494         * State will be updated to busy during the first busy tick after
 495         * exiting idle.
 496         */
 497        set_cpu_sd_state_idle();
 498
 499        local_irq_disable();
 500
 501        ts = &__get_cpu_var(tick_cpu_sched);
 502        /*
 503         * set ts->inidle unconditionally. even if the system did not
 504         * switch to nohz mode the cpu frequency governers rely on the
 505         * update of the idle time accounting in tick_nohz_start_idle().
 506         */
 507        ts->inidle = 1;
 508        __tick_nohz_idle_enter(ts);
 509
 510        local_irq_enable();
 511}
 512
 513/**
 514 * tick_nohz_irq_exit - update next tick event from interrupt exit
 515 *
 516 * When an interrupt fires while we are idle and it doesn't cause
 517 * a reschedule, it may still add, modify or delete a timer, enqueue
 518 * an RCU callback, etc...
 519 * So we need to re-calculate and reprogram the next tick event.
 520 */
 521void tick_nohz_irq_exit(void)
 522{
 523        struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 524
 525        if (!ts->inidle)
 526                return;
 527
 528        __tick_nohz_idle_enter(ts);
 529}
 530
 531/**
 532 * tick_nohz_get_sleep_length - return the length of the current sleep
 533 *
 534 * Called from power state control code with interrupts disabled
 535 */
 536ktime_t tick_nohz_get_sleep_length(void)
 537{
 538        struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 539
 540        return ts->sleep_length;
 541}
 542
 543static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
 544{
 545        hrtimer_cancel(&ts->sched_timer);
 546        hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
 547
 548        while (1) {
 549                /* Forward the time to expire in the future */
 550                hrtimer_forward(&ts->sched_timer, now, tick_period);
 551
 552                if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
 553                        hrtimer_start_expires(&ts->sched_timer,
 554                                              HRTIMER_MODE_ABS_PINNED);
 555                        /* Check, if the timer was already in the past */
 556                        if (hrtimer_active(&ts->sched_timer))
 557                                break;
 558                } else {
 559                        if (!tick_program_event(
 560                                hrtimer_get_expires(&ts->sched_timer), 0))
 561                                break;
 562                }
 563                /* Reread time and update jiffies */
 564                now = ktime_get();
 565                tick_do_update_jiffies64(now);
 566        }
 567}
 568
 569static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
 570{
 571        /* Update jiffies first */
 572        select_nohz_load_balancer(0);
 573        tick_do_update_jiffies64(now);
 574        update_cpu_load_nohz();
 575
 576        calc_load_exit_idle();
 577        touch_softlockup_watchdog();
 578        /*
 579         * Cancel the scheduled timer and restore the tick
 580         */
 581        ts->tick_stopped  = 0;
 582        ts->idle_exittime = now;
 583
 584        tick_nohz_restart(ts, now);
 585}
 586
 587static void tick_nohz_account_idle_ticks(struct tick_sched *ts)
 588{
 589#ifndef CONFIG_VIRT_CPU_ACCOUNTING
 590        unsigned long ticks;
 591        /*
 592         * We stopped the tick in idle. Update process times would miss the
 593         * time we slept as update_process_times does only a 1 tick
 594         * accounting. Enforce that this is accounted to idle !
 595         */
 596        ticks = jiffies - ts->idle_jiffies;
 597        /*
 598         * We might be one off. Do not randomly account a huge number of ticks!
 599         */
 600        if (ticks && ticks < LONG_MAX)
 601                account_idle_ticks(ticks);
 602#endif
 603}
 604
 605/**
 606 * tick_nohz_idle_exit - restart the idle tick from the idle task
 607 *
 608 * Restart the idle tick when the CPU is woken up from idle
 609 * This also exit the RCU extended quiescent state. The CPU
 610 * can use RCU again after this function is called.
 611 */
 612void tick_nohz_idle_exit(void)
 613{
 614        int cpu = smp_processor_id();
 615        struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 616        ktime_t now;
 617
 618        local_irq_disable();
 619
 620        WARN_ON_ONCE(!ts->inidle);
 621
 622        ts->inidle = 0;
 623
 624        if (ts->idle_active || ts->tick_stopped)
 625                now = ktime_get();
 626
 627        if (ts->idle_active)
 628                tick_nohz_stop_idle(cpu, now);
 629
 630        if (ts->tick_stopped) {
 631                tick_nohz_restart_sched_tick(ts, now);
 632                tick_nohz_account_idle_ticks(ts);
 633        }
 634
 635        local_irq_enable();
 636}
 637
 638static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
 639{
 640        hrtimer_forward(&ts->sched_timer, now, tick_period);
 641        return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
 642}
 643
 644/*
 645 * The nohz low res interrupt handler
 646 */
 647static void tick_nohz_handler(struct clock_event_device *dev)
 648{
 649        struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 650        struct pt_regs *regs = get_irq_regs();
 651        int cpu = smp_processor_id();
 652        ktime_t now = ktime_get();
 653
 654        dev->next_event.tv64 = KTIME_MAX;
 655
 656        /*
 657         * Check if the do_timer duty was dropped. We don't care about
 658         * concurrency: This happens only when the cpu in charge went
 659         * into a long sleep. If two cpus happen to assign themself to
 660         * this duty, then the jiffies update is still serialized by
 661         * xtime_lock.
 662         */
 663        if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
 664                tick_do_timer_cpu = cpu;
 665
 666        /* Check, if the jiffies need an update */
 667        if (tick_do_timer_cpu == cpu)
 668                tick_do_update_jiffies64(now);
 669
 670        /*
 671         * When we are idle and the tick is stopped, we have to touch
 672         * the watchdog as we might not schedule for a really long
 673         * time. This happens on complete idle SMP systems while
 674         * waiting on the login prompt. We also increment the "start
 675         * of idle" jiffy stamp so the idle accounting adjustment we
 676         * do when we go busy again does not account too much ticks.
 677         */
 678        if (ts->tick_stopped) {
 679                touch_softlockup_watchdog();
 680                ts->idle_jiffies++;
 681        }
 682
 683        update_process_times(user_mode(regs));
 684        profile_tick(CPU_PROFILING);
 685
 686        while (tick_nohz_reprogram(ts, now)) {
 687                now = ktime_get();
 688                tick_do_update_jiffies64(now);
 689        }
 690}
 691
 692/**
 693 * tick_nohz_switch_to_nohz - switch to nohz mode
 694 */
 695static void tick_nohz_switch_to_nohz(void)
 696{
 697        struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 698        ktime_t next;
 699
 700        if (!tick_nohz_enabled)
 701                return;
 702
 703        local_irq_disable();
 704        if (tick_switch_to_oneshot(tick_nohz_handler)) {
 705                local_irq_enable();
 706                return;
 707        }
 708
 709        ts->nohz_mode = NOHZ_MODE_LOWRES;
 710
 711        /*
 712         * Recycle the hrtimer in ts, so we can share the
 713         * hrtimer_forward with the highres code.
 714         */
 715        hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
 716        /* Get the next period */
 717        next = tick_init_jiffy_update();
 718
 719        for (;;) {
 720                hrtimer_set_expires(&ts->sched_timer, next);
 721                if (!tick_program_event(next, 0))
 722                        break;
 723                next = ktime_add(next, tick_period);
 724        }
 725        local_irq_enable();
 726}
 727
 728/*
 729 * When NOHZ is enabled and the tick is stopped, we need to kick the
 730 * tick timer from irq_enter() so that the jiffies update is kept
 731 * alive during long running softirqs. That's ugly as hell, but
 732 * correctness is key even if we need to fix the offending softirq in
 733 * the first place.
 734 *
 735 * Note, this is different to tick_nohz_restart. We just kick the
 736 * timer and do not touch the other magic bits which need to be done
 737 * when idle is left.
 738 */
 739static void tick_nohz_kick_tick(int cpu, ktime_t now)
 740{
 741#if 0
 742        /* Switch back to 2.6.27 behaviour */
 743
 744        struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 745        ktime_t delta;
 746
 747        /*
 748         * Do not touch the tick device, when the next expiry is either
 749         * already reached or less/equal than the tick period.
 750         */
 751        delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
 752        if (delta.tv64 <= tick_period.tv64)
 753                return;
 754
 755        tick_nohz_restart(ts, now);
 756#endif
 757}
 758
 759static inline void tick_check_nohz(int cpu)
 760{
 761        struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 762        ktime_t now;
 763
 764        if (!ts->idle_active && !ts->tick_stopped)
 765                return;
 766        now = ktime_get();
 767        if (ts->idle_active)
 768                tick_nohz_stop_idle(cpu, now);
 769        if (ts->tick_stopped) {
 770                tick_nohz_update_jiffies(now);
 771                tick_nohz_kick_tick(cpu, now);
 772        }
 773}
 774
 775#else
 776
 777static inline void tick_nohz_switch_to_nohz(void) { }
 778static inline void tick_check_nohz(int cpu) { }
 779
 780#endif /* NO_HZ */
 781
 782/*
 783 * Called from irq_enter to notify about the possible interruption of idle()
 784 */
 785void tick_check_idle(int cpu)
 786{
 787        tick_check_oneshot_broadcast(cpu);
 788        tick_check_nohz(cpu);
 789}
 790
 791/*
 792 * High resolution timer specific code
 793 */
 794#ifdef CONFIG_HIGH_RES_TIMERS
 795/*
 796 * We rearm the timer until we get disabled by the idle code.
 797 * Called with interrupts disabled and timer->base->cpu_base->lock held.
 798 */
 799static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
 800{
 801        struct tick_sched *ts =
 802                container_of(timer, struct tick_sched, sched_timer);
 803        struct pt_regs *regs = get_irq_regs();
 804        ktime_t now = ktime_get();
 805        int cpu = smp_processor_id();
 806
 807#ifdef CONFIG_NO_HZ
 808        /*
 809         * Check if the do_timer duty was dropped. We don't care about
 810         * concurrency: This happens only when the cpu in charge went
 811         * into a long sleep. If two cpus happen to assign themself to
 812         * this duty, then the jiffies update is still serialized by
 813         * xtime_lock.
 814         */
 815        if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
 816                tick_do_timer_cpu = cpu;
 817#endif
 818
 819        /* Check, if the jiffies need an update */
 820        if (tick_do_timer_cpu == cpu)
 821                tick_do_update_jiffies64(now);
 822
 823        /*
 824         * Do not call, when we are not in irq context and have
 825         * no valid regs pointer
 826         */
 827        if (regs) {
 828                /*
 829                 * When we are idle and the tick is stopped, we have to touch
 830                 * the watchdog as we might not schedule for a really long
 831                 * time. This happens on complete idle SMP systems while
 832                 * waiting on the login prompt. We also increment the "start of
 833                 * idle" jiffy stamp so the idle accounting adjustment we do
 834                 * when we go busy again does not account too much ticks.
 835                 */
 836                if (ts->tick_stopped) {
 837                        touch_softlockup_watchdog();
 838                        if (is_idle_task(current))
 839                                ts->idle_jiffies++;
 840                }
 841                update_process_times(user_mode(regs));
 842                profile_tick(CPU_PROFILING);
 843        }
 844
 845        hrtimer_forward(timer, now, tick_period);
 846
 847        return HRTIMER_RESTART;
 848}
 849
 850static int sched_skew_tick;
 851
 852static int __init skew_tick(char *str)
 853{
 854        get_option(&str, &sched_skew_tick);
 855
 856        return 0;
 857}
 858early_param("skew_tick", skew_tick);
 859
 860/**
 861 * tick_setup_sched_timer - setup the tick emulation timer
 862 */
 863void tick_setup_sched_timer(void)
 864{
 865        struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 866        ktime_t now = ktime_get();
 867
 868        /*
 869         * Emulate tick processing via per-CPU hrtimers:
 870         */
 871        hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
 872        ts->sched_timer.function = tick_sched_timer;
 873
 874        /* Get the next period (per cpu) */
 875        hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
 876
 877        /* Offset the tick to avert xtime_lock contention. */
 878        if (sched_skew_tick) {
 879                u64 offset = ktime_to_ns(tick_period) >> 1;
 880                do_div(offset, num_possible_cpus());
 881                offset *= smp_processor_id();
 882                hrtimer_add_expires_ns(&ts->sched_timer, offset);
 883        }
 884
 885        for (;;) {
 886                hrtimer_forward(&ts->sched_timer, now, tick_period);
 887                hrtimer_start_expires(&ts->sched_timer,
 888                                      HRTIMER_MODE_ABS_PINNED);
 889                /* Check, if the timer was already in the past */
 890                if (hrtimer_active(&ts->sched_timer))
 891                        break;
 892                now = ktime_get();
 893        }
 894
 895#ifdef CONFIG_NO_HZ
 896        if (tick_nohz_enabled)
 897                ts->nohz_mode = NOHZ_MODE_HIGHRES;
 898#endif
 899}
 900#endif /* HIGH_RES_TIMERS */
 901
 902#if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
 903void tick_cancel_sched_timer(int cpu)
 904{
 905        struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 906
 907# ifdef CONFIG_HIGH_RES_TIMERS
 908        if (ts->sched_timer.base)
 909                hrtimer_cancel(&ts->sched_timer);
 910# endif
 911
 912        ts->nohz_mode = NOHZ_MODE_INACTIVE;
 913}
 914#endif
 915
 916/**
 917 * Async notification about clocksource changes
 918 */
 919void tick_clock_notify(void)
 920{
 921        int cpu;
 922
 923        for_each_possible_cpu(cpu)
 924                set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
 925}
 926
 927/*
 928 * Async notification about clock event changes
 929 */
 930void tick_oneshot_notify(void)
 931{
 932        struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 933
 934        set_bit(0, &ts->check_clocks);
 935}
 936
 937/**
 938 * Check, if a change happened, which makes oneshot possible.
 939 *
 940 * Called cyclic from the hrtimer softirq (driven by the timer
 941 * softirq) allow_nohz signals, that we can switch into low-res nohz
 942 * mode, because high resolution timers are disabled (either compile
 943 * or runtime).
 944 */
 945int tick_check_oneshot_change(int allow_nohz)
 946{
 947        struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 948
 949        if (!test_and_clear_bit(0, &ts->check_clocks))
 950                return 0;
 951
 952        if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
 953                return 0;
 954
 955        if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
 956                return 0;
 957
 958        if (!allow_nohz)
 959                return 1;
 960
 961        tick_nohz_switch_to_nohz();
 962        return 0;
 963}
 964
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.