kernel/time/tick-sched.c

   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/tick.h>
  23 #include <linux/module.h>
  24
  25 #include <asm/irq_regs.h>
  26
  27 #include "tick-internal.h"
  28
  29 /*
  30  * Per cpu nohz control structure
  31  */
  32 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
  33
  34 /*
  35  * The time, when the last jiffy update happened. Protected by xtime_lock.
  36  */
  37 static ktime_t last_jiffies_update;
  38
  39 struct tick_sched *tick_get_tick_sched(int cpu)
  40 {
  41         return &per_cpu(tick_cpu_sched, cpu);
  42 }
  43
  44 /*
  45  * Must be called with interrupts disabled !
  46  */
  47 static void tick_do_update_jiffies64(ktime_t now)
  48 {
  49         unsigned long ticks = 0;
  50         ktime_t delta;
  51
  52         /*
  53          * Do a quick check without holding xtime_lock:
  54          */
  55         delta = ktime_sub(now, last_jiffies_update);
  56         if (delta.tv64 < tick_period.tv64)
  57                 return;
  58
  59         /* Reevalute with xtime_lock held */
  60         write_seqlock(&xtime_lock);
  61
  62         delta = ktime_sub(now, last_jiffies_update);
  63         if (delta.tv64 >= tick_period.tv64) {
  64
  65                 delta = ktime_sub(delta, tick_period);
  66                 last_jiffies_update = ktime_add(last_jiffies_update,
  67                                                 tick_period);
  68
  69                 /* Slow path for long timeouts */
  70                 if (unlikely(delta.tv64 >= tick_period.tv64)) {
  71                         s64 incr = ktime_to_ns(tick_period);
  72
  73                         ticks = ktime_divns(delta, incr);
  74
  75                         last_jiffies_update = ktime_add_ns(last_jiffies_update,
  76                                                            incr * ticks);
  77                 }
  78                 do_timer(++ticks);
  79
  80                 /* Keep the tick_next_period variable up to date */
  81                 tick_next_period = ktime_add(last_jiffies_update, tick_period);
  82         }
  83         write_sequnlock(&xtime_lock);
  84 }
  85
  86 /*
  87  * Initialize and return retrieve the jiffies update.
  88  */
  89 static ktime_t tick_init_jiffy_update(void)
  90 {
  91         ktime_t period;
  92
  93         write_seqlock(&xtime_lock);
  94         /* Did we start the jiffies update yet ? */
  95         if (last_jiffies_update.tv64 == 0)
  96                 last_jiffies_update = tick_next_period;
  97         period = last_jiffies_update;
  98         write_sequnlock(&xtime_lock);
  99         return period;
 100 }
 101
 102 /*
 103  * NOHZ - aka dynamic tick functionality
 104  */
 105 #ifdef CONFIG_NO_HZ
 106 /*
 107  * NO HZ enabled ?
 108  */
 109 static int tick_nohz_enabled __read_mostly  = 1;
 110
 111 /*
 112  * Enable / Disable tickless mode
 113  */
 114 static int __init setup_tick_nohz(char *str)
 115 {
 116         if (!strcmp(str, "off"))
 117                 tick_nohz_enabled = 0;
 118         else if (!strcmp(str, "on"))
 119                 tick_nohz_enabled = 1;
 120         else
 121                 return 0;
 122         return 1;
 123 }
 124
 125 __setup("nohz=", setup_tick_nohz);
 126
 127 /**
 128  * tick_nohz_update_jiffies - update jiffies when idle was interrupted
 129  *
 130  * Called from interrupt entry when the CPU was idle
 131  *
 132  * In case the sched_tick was stopped on this CPU, we have to check if jiffies
 133  * must be updated. Otherwise an interrupt handler could use a stale jiffy
 134  * value. We do this unconditionally on any cpu, as we don't know whether the
 135  * cpu, which has the update task assigned is in a long sleep.
 136  */
 137 void tick_nohz_update_jiffies(void)
 138 {
 139         int cpu = smp_processor_id();
 140         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 141         unsigned long flags;
 142         ktime_t now;
 143
 144         if (!ts->tick_stopped)
 145                 return;
 146
 147         cpu_clear(cpu, nohz_cpu_mask);
 148         now = ktime_get();
 149         ts->idle_waketime = now;
 150
 151         local_irq_save(flags);
 152         tick_do_update_jiffies64(now);
 153         local_irq_restore(flags);
 154
 155         touch_softlockup_watchdog();
 156 }
 157
 158 static void tick_nohz_stop_idle(int cpu)
 159 {
 160         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 161
 162         if (ts->idle_active) {
 163                 ktime_t now, delta;
 164                 now = ktime_get();
 165                 delta = ktime_sub(now, ts->idle_entrytime);
 166                 ts->idle_lastupdate = now;
 167                 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
 168                 ts->idle_active = 0;
 169
 170                 sched_clock_idle_wakeup_event(0);
 171         }
 172 }
 173
 174 static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
 175 {
 176         ktime_t now, delta;
 177
 178         now = ktime_get();
 179         if (ts->idle_active) {
 180                 delta = ktime_sub(now, ts->idle_entrytime);
 181                 ts->idle_lastupdate = now;
 182                 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
 183         }
 184         ts->idle_entrytime = now;
 185         ts->idle_active = 1;
 186         sched_clock_idle_sleep_event();
 187         return now;
 188 }
 189
 190 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
 191 {
 192         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 193
 194         if (!tick_nohz_enabled)
 195                 return -1;
 196
 197         if (ts->idle_active)
 198                 *last_update_time = ktime_to_us(ts->idle_lastupdate);
 199         else
 200                 *last_update_time = ktime_to_us(ktime_get());
 201
 202         return ktime_to_us(ts->idle_sleeptime);
 203 }
 204 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
 205
 206 /**
 207  * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
 208  *
 209  * When the next event is more than a tick into the future, stop the idle tick
 210  * Called either from the idle loop or from irq_exit() when an idle period was
 211  * just interrupted by an interrupt which did not cause a reschedule.
 212  */
 213 void tick_nohz_stop_sched_tick(int inidle)
 214 {
 215         unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
 216         struct tick_sched *ts;
 217         ktime_t last_update, expires, now;
 218         struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
 219         int cpu;
 220
 221         local_irq_save(flags);
 222
 223         cpu = smp_processor_id();
 224         ts = &per_cpu(tick_cpu_sched, cpu);
 225         now = tick_nohz_start_idle(ts);
 226
 227         /*
 228          * If this cpu is offline and it is the one which updates
 229          * jiffies, then give up the assignment and let it be taken by
 230          * the cpu which runs the tick timer next. If we don't drop
 231          * this here the jiffies might be stale and do_timer() never
 232          * invoked.
 233          */
 234         if (unlikely(!cpu_online(cpu))) {
 235                 if (cpu == tick_do_timer_cpu)
 236                         tick_do_timer_cpu = TICK_DO_TIMER_NONE;
 237         }
 238
 239         if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
 240                 goto end;
 241
 242         if (!inidle && !ts->inidle)
 243                 goto end;
 244
 245         ts->inidle = 1;
 246
 247         if (need_resched())
 248                 goto end;
 249
 250         if (unlikely(local_softirq_pending())) {
 251                 static int ratelimit;
 252
 253                 if (ratelimit < 10) {
 254                         printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
 255                                local_softirq_pending());
 256                         ratelimit++;
 257                 }
 258                 goto end;
 259         }
 260
 261         ts->idle_calls++;
 262         /* Read jiffies and the time when jiffies were updated last */
 263         do {
 264                 seq = read_seqbegin(&xtime_lock);
 265                 last_update = last_jiffies_update;
 266                 last_jiffies = jiffies;
 267         } while (read_seqretry(&xtime_lock, seq));
 268
 269         /* Get the next timer wheel timer */
 270         next_jiffies = get_next_timer_interrupt(last_jiffies);
 271         delta_jiffies = next_jiffies - last_jiffies;
 272
 273         if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu))
 274                 delta_jiffies = 1;
 275         /*
 276          * Do not stop the tick, if we are only one off
 277          * or if the cpu is required for rcu
 278          */
 279         if (!ts->tick_stopped && delta_jiffies == 1)
 280                 goto out;
 281
 282         /* Schedule the tick, if we are at least one jiffie off */
 283         if ((long)delta_jiffies >= 1) {
 284
 285                 if (delta_jiffies > 1)
 286                         cpu_set(cpu, nohz_cpu_mask);
 287                 /*
 288                  * nohz_stop_sched_tick can be called several times before
 289                  * the nohz_restart_sched_tick is called. This happens when
 290                  * interrupts arrive which do not cause a reschedule. In the
 291                  * first call we save the current tick time, so we can restart
 292                  * the scheduler tick in nohz_restart_sched_tick.
 293                  */
 294                 if (!ts->tick_stopped) {
 295                         if (select_nohz_load_balancer(1)) {
 296                                 /*
 297                                  * sched tick not stopped!
 298                                  */
 299                                 cpu_clear(cpu, nohz_cpu_mask);
 300                                 goto out;
 301                         }
 302
 303                         ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
 304                         ts->tick_stopped = 1;
 305                         ts->idle_jiffies = last_jiffies;
 306                         rcu_enter_nohz();
 307                 }
 308
 309                 /*
 310                  * If this cpu is the one which updates jiffies, then
 311                  * give up the assignment and let it be taken by the
 312                  * cpu which runs the tick timer next, which might be
 313                  * this cpu as well. If we don't drop this here the
 314                  * jiffies might be stale and do_timer() never
 315                  * invoked.
 316                  */
 317                 if (cpu == tick_do_timer_cpu)
 318                         tick_do_timer_cpu = TICK_DO_TIMER_NONE;
 319
 320                 ts->idle_sleeps++;
 321
 322                 /*
 323                  * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that
 324                  * there is no timer pending or at least extremly far
 325                  * into the future (12 days for HZ=1000). In this case
 326                  * we simply stop the tick timer:
 327                  */
 328                 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) {
 329                         ts->idle_expires.tv64 = KTIME_MAX;
 330                         if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
 331                                 hrtimer_cancel(&ts->sched_timer);
 332                         goto out;
 333                 }
 334
 335                 /*
 336                  * calculate the expiry time for the next timer wheel
 337                  * timer
 338                  */
 339                 expires = ktime_add_ns(last_update, tick_period.tv64 *
 340                                        delta_jiffies);
 341                 ts->idle_expires = expires;
 342
 343                 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
 344                         hrtimer_start(&ts->sched_timer, expires,
 345                                       HRTIMER_MODE_ABS);
 346                         /* Check, if the timer was already in the past */
 347                         if (hrtimer_active(&ts->sched_timer))
 348                                 goto out;
 349                 } else if (!tick_program_event(expires, 0))
 350                                 goto out;
 351                 /*
 352                  * We are past the event already. So we crossed a
 353                  * jiffie boundary. Update jiffies and raise the
 354                  * softirq.
 355                  */
 356                 tick_do_update_jiffies64(ktime_get());
 357                 cpu_clear(cpu, nohz_cpu_mask);
 358         }
 359         raise_softirq_irqoff(TIMER_SOFTIRQ);
 360 out:
 361         ts->next_jiffies = next_jiffies;
 362         ts->last_jiffies = last_jiffies;
 363         ts->sleep_length = ktime_sub(dev->next_event, now);
 364 end:
 365         local_irq_restore(flags);
 366 }
 367
 368 /**
 369  * tick_nohz_get_sleep_length - return the length of the current sleep
 370  *
 371  * Called from power state control code with interrupts disabled
 372  */
 373 ktime_t tick_nohz_get_sleep_length(void)
 374 {
 375         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 376
 377         return ts->sleep_length;
 378 }
 379
 380 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
 381 {
 382         hrtimer_cancel(&ts->sched_timer);
 383         hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
 384
 385         while (1) {
 386                 /* Forward the time to expire in the future */
 387                 hrtimer_forward(&ts->sched_timer, now, tick_period);
 388
 389                 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
 390                         hrtimer_start_expires(&ts->sched_timer,
 391                                       HRTIMER_MODE_ABS);
 392                         /* Check, if the timer was already in the past */
 393                         if (hrtimer_active(&ts->sched_timer))
 394                                 break;
 395                 } else {
 396                         if (!tick_program_event(
 397                                 hrtimer_get_expires(&ts->sched_timer), 0))
 398                                 break;
 399                 }
 400                 /* Update jiffies and reread time */
 401                 tick_do_update_jiffies64(now);
 402                 now = ktime_get();
 403         }
 404 }
 405
 406 /**
 407  * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
 408  *
 409  * Restart the idle tick when the CPU is woken up from idle
 410  */
 411 void tick_nohz_restart_sched_tick(void)
 412 {
 413         int cpu = smp_processor_id();
 414         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 415         unsigned long ticks;
 416         ktime_t now;
 417
 418         local_irq_disable();
 419         tick_nohz_stop_idle(cpu);
 420
 421         if (!ts->inidle || !ts->tick_stopped) {
 422                 ts->inidle = 0;
 423                 local_irq_enable();
 424                 return;
 425         }
 426
 427         ts->inidle = 0;
 428
 429         rcu_exit_nohz();
 430
 431         /* Update jiffies first */
 432         select_nohz_load_balancer(0);
 433         now = ktime_get();
 434         tick_do_update_jiffies64(now);
 435         cpu_clear(cpu, nohz_cpu_mask);
 436
 437         /*
 438          * We stopped the tick in idle. Update process times would miss the
 439          * time we slept as update_process_times does only a 1 tick
 440          * accounting. Enforce that this is accounted to idle !
 441          */
 442         ticks = jiffies - ts->idle_jiffies;
 443         /*
 444          * We might be one off. Do not randomly account a huge number of ticks!
 445          */
 446         if (ticks && ticks < LONG_MAX) {
 447                 add_preempt_count(HARDIRQ_OFFSET);
 448                 account_system_time(current, HARDIRQ_OFFSET,
 449                                     jiffies_to_cputime(ticks));
 450                 sub_preempt_count(HARDIRQ_OFFSET);
 451         }
 452
 453         touch_softlockup_watchdog();
 454         /*
 455          * Cancel the scheduled timer and restore the tick
 456          */
 457         ts->tick_stopped  = 0;
 458         ts->idle_exittime = now;
 459
 460         tick_nohz_restart(ts, now);
 461
 462         local_irq_enable();
 463 }
 464
 465 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
 466 {
 467         hrtimer_forward(&ts->sched_timer, now, tick_period);
 468         return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
 469 }
 470
 471 /*
 472  * The nohz low res interrupt handler
 473  */
 474 static void tick_nohz_handler(struct clock_event_device *dev)
 475 {
 476         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 477         struct pt_regs *regs = get_irq_regs();
 478         int cpu = smp_processor_id();
 479         ktime_t now = ktime_get();
 480
 481         dev->next_event.tv64 = KTIME_MAX;
 482
 483         /*
 484          * Check if the do_timer duty was dropped. We don't care about
 485          * concurrency: This happens only when the cpu in charge went
 486          * into a long sleep. If two cpus happen to assign themself to
 487          * this duty, then the jiffies update is still serialized by
 488          * xtime_lock.
 489          */
 490         if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
 491                 tick_do_timer_cpu = cpu;
 492
 493         /* Check, if the jiffies need an update */
 494         if (tick_do_timer_cpu == cpu)
 495                 tick_do_update_jiffies64(now);
 496
 497         /*
 498          * When we are idle and the tick is stopped, we have to touch
 499          * the watchdog as we might not schedule for a really long
 500          * time. This happens on complete idle SMP systems while
 501          * waiting on the login prompt. We also increment the "start
 502          * of idle" jiffy stamp so the idle accounting adjustment we
 503          * do when we go busy again does not account too much ticks.
 504          */
 505         if (ts->tick_stopped) {
 506                 touch_softlockup_watchdog();
 507                 ts->idle_jiffies++;
 508         }
 509
 510         update_process_times(user_mode(regs));
 511         profile_tick(CPU_PROFILING);
 512
 513         while (tick_nohz_reprogram(ts, now)) {
 514                 now = ktime_get();
 515                 tick_do_update_jiffies64(now);
 516         }
 517 }
 518
 519 /**
 520  * tick_nohz_switch_to_nohz - switch to nohz mode
 521  */
 522 static void tick_nohz_switch_to_nohz(void)
 523 {
 524         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 525         ktime_t next;
 526
 527         if (!tick_nohz_enabled)
 528                 return;
 529
 530         local_irq_disable();
 531         if (tick_switch_to_oneshot(tick_nohz_handler)) {
 532                 local_irq_enable();
 533                 return;
 534         }
 535
 536         ts->nohz_mode = NOHZ_MODE_LOWRES;
 537
 538         /*
 539          * Recycle the hrtimer in ts, so we can share the
 540          * hrtimer_forward with the highres code.
 541          */
 542         hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
 543         /* Get the next period */
 544         next = tick_init_jiffy_update();
 545
 546         for (;;) {
 547                 hrtimer_set_expires(&ts->sched_timer, next);
 548                 if (!tick_program_event(next, 0))
 549                         break;
 550                 next = ktime_add(next, tick_period);
 551         }
 552         local_irq_enable();
 553
 554         printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n",
 555                smp_processor_id());
 556 }
 557
 558 /*
 559  * When NOHZ is enabled and the tick is stopped, we need to kick the
 560  * tick timer from irq_enter() so that the jiffies update is kept
 561  * alive during long running softirqs. That's ugly as hell, but
 562  * correctness is key even if we need to fix the offending softirq in
 563  * the first place.
 564  *
 565  * Note, this is different to tick_nohz_restart. We just kick the
 566  * timer and do not touch the other magic bits which need to be done
 567  * when idle is left.
 568  */
 569 static void tick_nohz_kick_tick(int cpu)
 570 {
 571         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 572         ktime_t delta, now;
 573
 574         if (!ts->tick_stopped)
 575                 return;
 576
 577         /*
 578          * Do not touch the tick device, when the next expiry is either
 579          * already reached or less/equal than the tick period.
 580          */
 581         now = ktime_get();
 582         delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
 583         if (delta.tv64 <= tick_period.tv64)
 584                 return;
 585
 586         tick_nohz_restart(ts, now);
 587 }
 588
 589 #else
 590
 591 static inline void tick_nohz_switch_to_nohz(void) { }
 592
 593 #endif /* NO_HZ */
 594
 595 /*
 596  * Called from irq_enter to notify about the possible interruption of idle()
 597  */
 598 void tick_check_idle(int cpu)
 599 {
 600         tick_check_oneshot_broadcast(cpu);
 601 #ifdef CONFIG_NO_HZ
 602         tick_nohz_stop_idle(cpu);
 603         tick_nohz_update_jiffies();
 604         tick_nohz_kick_tick(cpu);
 605 #endif
 606 }
 607
 608 /*
 609  * High resolution timer specific code
 610  */
 611 #ifdef CONFIG_HIGH_RES_TIMERS
 612 /*
 613  * We rearm the timer until we get disabled by the idle code.
 614  * Called with interrupts disabled and timer->base->cpu_base->lock held.
 615  */
 616 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
 617 {
 618         struct tick_sched *ts =
 619                 container_of(timer, struct tick_sched, sched_timer);
 620         struct pt_regs *regs = get_irq_regs();
 621         ktime_t now = ktime_get();
 622         int cpu = smp_processor_id();
 623
 624 #ifdef CONFIG_NO_HZ
 625         /*
 626          * Check if the do_timer duty was dropped. We don't care about
 627          * concurrency: This happens only when the cpu in charge went
 628          * into a long sleep. If two cpus happen to assign themself to
 629          * this duty, then the jiffies update is still serialized by
 630          * xtime_lock.
 631          */
 632         if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
 633                 tick_do_timer_cpu = cpu;
 634 #endif
 635
 636         /* Check, if the jiffies need an update */
 637         if (tick_do_timer_cpu == cpu)
 638                 tick_do_update_jiffies64(now);
 639
 640         /*
 641          * Do not call, when we are not in irq context and have
 642          * no valid regs pointer
 643          */
 644         if (regs) {
 645                 /*
 646                  * When we are idle and the tick is stopped, we have to touch
 647                  * the watchdog as we might not schedule for a really long
 648                  * time. This happens on complete idle SMP systems while
 649                  * waiting on the login prompt. We also increment the "start of
 650                  * idle" jiffy stamp so the idle accounting adjustment we do
 651                  * when we go busy again does not account too much ticks.
 652                  */
 653                 if (ts->tick_stopped) {
 654                         touch_softlockup_watchdog();
 655                         ts->idle_jiffies++;
 656                 }
 657                 update_process_times(user_mode(regs));
 658                 profile_tick(CPU_PROFILING);
 659         }
 660
 661         hrtimer_forward(timer, now, tick_period);
 662
 663         return HRTIMER_RESTART;
 664 }
 665
 666 /**
 667  * tick_setup_sched_timer - setup the tick emulation timer
 668  */
 669 void tick_setup_sched_timer(void)
 670 {
 671         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 672         ktime_t now = ktime_get();
 673         u64 offset;
 674
 675         /*
 676          * Emulate tick processing via per-CPU hrtimers:
 677          */
 678         hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
 679         ts->sched_timer.function = tick_sched_timer;
 680         ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
 681
 682         /* Get the next period (per cpu) */
 683         hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
 684         offset = ktime_to_ns(tick_period) >> 1;
 685         do_div(offset, num_possible_cpus());
 686         offset *= smp_processor_id();
 687         hrtimer_add_expires_ns(&ts->sched_timer, offset);
 688
 689         for (;;) {
 690                 hrtimer_forward(&ts->sched_timer, now, tick_period);
 691                 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS);
 692                 /* Check, if the timer was already in the past */
 693                 if (hrtimer_active(&ts->sched_timer))
 694                         break;
 695                 now = ktime_get();
 696         }
 697
 698 #ifdef CONFIG_NO_HZ
 699         if (tick_nohz_enabled)
 700                 ts->nohz_mode = NOHZ_MODE_HIGHRES;
 701 #endif
 702 }
 703 #endif /* HIGH_RES_TIMERS */
 704
 705 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
 706 void tick_cancel_sched_timer(int cpu)
 707 {
 708         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 709
 710 # ifdef CONFIG_HIGH_RES_TIMERS
 711         if (ts->sched_timer.base)
 712                 hrtimer_cancel(&ts->sched_timer);
 713 # endif
 714
 715         ts->nohz_mode = NOHZ_MODE_INACTIVE;
 716 }
 717 #endif
 718
 719 /**
 720  * Async notification about clocksource changes
 721  */
 722 void tick_clock_notify(void)
 723 {
 724         int cpu;
 725
 726         for_each_possible_cpu(cpu)
 727                 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
 728 }
 729
 730 /*
 731  * Async notification about clock event changes
 732  */
 733 void tick_oneshot_notify(void)
 734 {
 735         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 736
 737         set_bit(0, &ts->check_clocks);
 738 }
 739
 740 /**
 741  * Check, if a change happened, which makes oneshot possible.
 742  *
 743  * Called cyclic from the hrtimer softirq (driven by the timer
 744  * softirq) allow_nohz signals, that we can switch into low-res nohz
 745  * mode, because high resolution timers are disabled (either compile
 746  * or runtime).
 747  */
 748 int tick_check_oneshot_change(int allow_nohz)
 749 {
 750         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 751
 752         if (!test_and_clear_bit(0, &ts->check_clocks))
 753                 return 0;
 754
 755         if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
 756                 return 0;
 757
 758         if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
 759                 return 0;
 760
 761         if (!allow_nohz)
 762                 return 1;
 763
 764         tick_nohz_switch_to_nohz();
 765         return 0;
 766 }