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/module.h>
  23
  24 #include <asm/irq_regs.h>
  25
  26 #include "tick-internal.h"
  27
  28 /*
  29  * Per cpu nohz control structure
  30  */
  31 static 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  */
  36 static ktime_t last_jiffies_update;
  37
  38 struct 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  */
  46 static 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  */
  88 static 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  */
 108 static int tick_nohz_enabled __read_mostly  = 1;
 109
 110 /*
 111  * Enable / Disable tickless mode
 112  */
 113 static 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  */
 136 static 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  */
 154 static void
 155 update_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
 173 static 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
 183 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts)
 184 {
 185         ktime_t now;
 186
 187         now = ktime_get();
 188
 189         update_ts_time_stats(cpu, ts, now, NULL);
 190
 191         ts->idle_entrytime = now;
 192         ts->idle_active = 1;
 193         sched_clock_idle_sleep_event();
 194         return now;
 195 }
 196
 197 /**
 198  * get_cpu_idle_time_us - get the total idle time of a cpu
 199  * @cpu: CPU number to query
 200  * @last_update_time: variable to store update time in. Do not update
 201  * counters if NULL.
 202  *
 203  * Return the cummulative idle time (since boot) for a given
 204  * CPU, in microseconds.
 205  *
 206  * This time is measured via accounting rather than sampling,
 207  * and is as accurate as ktime_get() is.
 208  *
 209  * This function returns -1 if NOHZ is not enabled.
 210  */
 211 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
 212 {
 213         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 214         ktime_t now, idle;
 215
 216         if (!tick_nohz_enabled)
 217                 return -1;
 218
 219         now = ktime_get();
 220         if (last_update_time) {
 221                 update_ts_time_stats(cpu, ts, now, last_update_time);
 222                 idle = ts->idle_sleeptime;
 223         } else {
 224                 if (ts->idle_active && !nr_iowait_cpu(cpu)) {
 225                         ktime_t delta = ktime_sub(now, ts->idle_entrytime);
 226
 227                         idle = ktime_add(ts->idle_sleeptime, delta);
 228                 } else {
 229                         idle = ts->idle_sleeptime;
 230                 }
 231         }
 232
 233         return ktime_to_us(idle);
 234
 235 }
 236 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
 237
 238 /**
 239  * get_cpu_iowait_time_us - get the total iowait time of a cpu
 240  * @cpu: CPU number to query
 241  * @last_update_time: variable to store update time in. Do not update
 242  * counters if NULL.
 243  *
 244  * Return the cummulative iowait time (since boot) for a given
 245  * CPU, in microseconds.
 246  *
 247  * This time is measured via accounting rather than sampling,
 248  * and is as accurate as ktime_get() is.
 249  *
 250  * This function returns -1 if NOHZ is not enabled.
 251  */
 252 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
 253 {
 254         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 255         ktime_t now, iowait;
 256
 257         if (!tick_nohz_enabled)
 258                 return -1;
 259
 260         now = ktime_get();
 261         if (last_update_time) {
 262                 update_ts_time_stats(cpu, ts, now, last_update_time);
 263                 iowait = ts->iowait_sleeptime;
 264         } else {
 265                 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) {
 266                         ktime_t delta = ktime_sub(now, ts->idle_entrytime);
 267
 268                         iowait = ktime_add(ts->iowait_sleeptime, delta);
 269                 } else {
 270                         iowait = ts->iowait_sleeptime;
 271                 }
 272         }
 273
 274         return ktime_to_us(iowait);
 275 }
 276 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
 277
 278 /**
 279  * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
 280  *
 281  * When the next event is more than a tick into the future, stop the idle tick
 282  * Called either from the idle loop or from irq_exit() when an idle period was
 283  * just interrupted by an interrupt which did not cause a reschedule.
 284  */
 285 void tick_nohz_stop_sched_tick(int inidle)
 286 {
 287         unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags;
 288         struct tick_sched *ts;
 289         ktime_t last_update, expires, now;
 290         struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev;
 291         u64 time_delta;
 292         int cpu;
 293
 294         local_irq_save(flags);
 295
 296         cpu = smp_processor_id();
 297         ts = &per_cpu(tick_cpu_sched, cpu);
 298
 299         /*
 300          * Call to tick_nohz_start_idle stops the last_update_time from being
 301          * updated. Thus, it must not be called in the event we are called from
 302          * irq_exit() with the prior state different than idle.
 303          */
 304         if (!inidle && !ts->inidle)
 305                 goto end;
 306
 307         /*
 308          * Set ts->inidle unconditionally. Even if the system did not
 309          * switch to NOHZ mode the cpu frequency governers rely on the
 310          * update of the idle time accounting in tick_nohz_start_idle().
 311          */
 312         ts->inidle = 1;
 313
 314         now = tick_nohz_start_idle(cpu, ts);
 315
 316         /*
 317          * If this cpu is offline and it is the one which updates
 318          * jiffies, then give up the assignment and let it be taken by
 319          * the cpu which runs the tick timer next. If we don't drop
 320          * this here the jiffies might be stale and do_timer() never
 321          * invoked.
 322          */
 323         if (unlikely(!cpu_online(cpu))) {
 324                 if (cpu == tick_do_timer_cpu)
 325                         tick_do_timer_cpu = TICK_DO_TIMER_NONE;
 326         }
 327
 328         if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
 329                 goto end;
 330
 331         if (need_resched())
 332                 goto end;
 333
 334         if (unlikely(local_softirq_pending() && cpu_online(cpu))) {
 335                 static int ratelimit;
 336
 337                 if (ratelimit < 10) {
 338                         printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n",
 339                                (unsigned int) local_softirq_pending());
 340                         ratelimit++;
 341                 }
 342                 goto end;
 343         }
 344
 345         ts->idle_calls++;
 346         /* Read jiffies and the time when jiffies were updated last */
 347         do {
 348                 seq = read_seqbegin(&xtime_lock);
 349                 last_update = last_jiffies_update;
 350                 last_jiffies = jiffies;
 351                 time_delta = timekeeping_max_deferment();
 352         } while (read_seqretry(&xtime_lock, seq));
 353
 354         if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) ||
 355             arch_needs_cpu(cpu)) {
 356                 next_jiffies = last_jiffies + 1;
 357                 delta_jiffies = 1;
 358         } else {
 359                 /* Get the next timer wheel timer */
 360                 next_jiffies = get_next_timer_interrupt(last_jiffies);
 361                 delta_jiffies = next_jiffies - last_jiffies;
 362         }
 363         /*
 364          * Do not stop the tick, if we are only one off
 365          * or if the cpu is required for rcu
 366          */
 367         if (!ts->tick_stopped && delta_jiffies == 1)
 368                 goto out;
 369
 370         /* Schedule the tick, if we are at least one jiffie off */
 371         if ((long)delta_jiffies >= 1) {
 372
 373                 /*
 374                  * If this cpu is the one which updates jiffies, then
 375                  * give up the assignment and let it be taken by the
 376                  * cpu which runs the tick timer next, which might be
 377                  * this cpu as well. If we don't drop this here the
 378                  * jiffies might be stale and do_timer() never
 379                  * invoked. Keep track of the fact that it was the one
 380                  * which had the do_timer() duty last. If this cpu is
 381                  * the one which had the do_timer() duty last, we
 382                  * limit the sleep time to the timekeeping
 383                  * max_deferement value which we retrieved
 384                  * above. Otherwise we can sleep as long as we want.
 385                  */
 386                 if (cpu == tick_do_timer_cpu) {
 387                         tick_do_timer_cpu = TICK_DO_TIMER_NONE;
 388                         ts->do_timer_last = 1;
 389                 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) {
 390                         time_delta = KTIME_MAX;
 391                         ts->do_timer_last = 0;
 392                 } else if (!ts->do_timer_last) {
 393                         time_delta = KTIME_MAX;
 394                 }
 395
 396                 /*
 397                  * calculate the expiry time for the next timer wheel
 398                  * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals
 399                  * that there is no timer pending or at least extremely
 400                  * far into the future (12 days for HZ=1000). In this
 401                  * case we set the expiry to the end of time.
 402                  */
 403                 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) {
 404                         /*
 405                          * Calculate the time delta for the next timer event.
 406                          * If the time delta exceeds the maximum time delta
 407                          * permitted by the current clocksource then adjust
 408                          * the time delta accordingly to ensure the
 409                          * clocksource does not wrap.
 410                          */
 411                         time_delta = min_t(u64, time_delta,
 412                                            tick_period.tv64 * delta_jiffies);
 413                 }
 414
 415                 if (time_delta < KTIME_MAX)
 416                         expires = ktime_add_ns(last_update, time_delta);
 417                 else
 418                         expires.tv64 = KTIME_MAX;
 419
 420                 /* Skip reprogram of event if its not changed */
 421                 if (ts->tick_stopped && ktime_equal(expires, dev->next_event))
 422                         goto out;
 423
 424                 /*
 425                  * nohz_stop_sched_tick can be called several times before
 426                  * the nohz_restart_sched_tick is called. This happens when
 427                  * interrupts arrive which do not cause a reschedule. In the
 428                  * first call we save the current tick time, so we can restart
 429                  * the scheduler tick in nohz_restart_sched_tick.
 430                  */
 431                 if (!ts->tick_stopped) {
 432                         select_nohz_load_balancer(1);
 433
 434                         ts->idle_tick = hrtimer_get_expires(&ts->sched_timer);
 435                         ts->tick_stopped = 1;
 436                         ts->idle_jiffies = last_jiffies;
 437                         rcu_enter_nohz();
 438                 }
 439
 440                 ts->idle_sleeps++;
 441
 442                 /* Mark expires */
 443                 ts->idle_expires = expires;
 444
 445                 /*
 446                  * If the expiration time == KTIME_MAX, then
 447                  * in this case we simply stop the tick timer.
 448                  */
 449                  if (unlikely(expires.tv64 == KTIME_MAX)) {
 450                         if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
 451                                 hrtimer_cancel(&ts->sched_timer);
 452                         goto out;
 453                 }
 454
 455                 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
 456                         hrtimer_start(&ts->sched_timer, expires,
 457                                       HRTIMER_MODE_ABS_PINNED);
 458                         /* Check, if the timer was already in the past */
 459                         if (hrtimer_active(&ts->sched_timer))
 460                                 goto out;
 461                 } else if (!tick_program_event(expires, 0))
 462                                 goto out;
 463                 /*
 464                  * We are past the event already. So we crossed a
 465                  * jiffie boundary. Update jiffies and raise the
 466                  * softirq.
 467                  */
 468                 tick_do_update_jiffies64(ktime_get());
 469         }
 470         raise_softirq_irqoff(TIMER_SOFTIRQ);
 471 out:
 472         ts->next_jiffies = next_jiffies;
 473         ts->last_jiffies = last_jiffies;
 474         ts->sleep_length = ktime_sub(dev->next_event, now);
 475 end:
 476         local_irq_restore(flags);
 477 }
 478
 479 /**
 480  * tick_nohz_get_sleep_length - return the length of the current sleep
 481  *
 482  * Called from power state control code with interrupts disabled
 483  */
 484 ktime_t tick_nohz_get_sleep_length(void)
 485 {
 486         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 487
 488         return ts->sleep_length;
 489 }
 490
 491 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
 492 {
 493         hrtimer_cancel(&ts->sched_timer);
 494         hrtimer_set_expires(&ts->sched_timer, ts->idle_tick);
 495
 496         while (1) {
 497                 /* Forward the time to expire in the future */
 498                 hrtimer_forward(&ts->sched_timer, now, tick_period);
 499
 500                 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) {
 501                         hrtimer_start_expires(&ts->sched_timer,
 502                                               HRTIMER_MODE_ABS_PINNED);
 503                         /* Check, if the timer was already in the past */
 504                         if (hrtimer_active(&ts->sched_timer))
 505                                 break;
 506                 } else {
 507                         if (!tick_program_event(
 508                                 hrtimer_get_expires(&ts->sched_timer), 0))
 509                                 break;
 510                 }
 511                 /* Update jiffies and reread time */
 512                 tick_do_update_jiffies64(now);
 513                 now = ktime_get();
 514         }
 515 }
 516
 517 /**
 518  * tick_nohz_restart_sched_tick - restart the idle tick from the idle task
 519  *
 520  * Restart the idle tick when the CPU is woken up from idle
 521  */
 522 void tick_nohz_restart_sched_tick(void)
 523 {
 524         int cpu = smp_processor_id();
 525         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 526 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
 527         unsigned long ticks;
 528 #endif
 529         ktime_t now;
 530
 531         local_irq_disable();
 532         if (ts->idle_active || (ts->inidle && ts->tick_stopped))
 533                 now = ktime_get();
 534
 535         if (ts->idle_active)
 536                 tick_nohz_stop_idle(cpu, now);
 537
 538         if (!ts->inidle || !ts->tick_stopped) {
 539                 ts->inidle = 0;
 540                 local_irq_enable();
 541                 return;
 542         }
 543
 544         ts->inidle = 0;
 545
 546         rcu_exit_nohz();
 547
 548         /* Update jiffies first */
 549         select_nohz_load_balancer(0);
 550         tick_do_update_jiffies64(now);
 551
 552 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
 553         /*
 554          * We stopped the tick in idle. Update process times would miss the
 555          * time we slept as update_process_times does only a 1 tick
 556          * accounting. Enforce that this is accounted to idle !
 557          */
 558         ticks = jiffies - ts->idle_jiffies;
 559         /*
 560          * We might be one off. Do not randomly account a huge number of ticks!
 561          */
 562         if (ticks && ticks < LONG_MAX)
 563                 account_idle_ticks(ticks);
 564 #endif
 565
 566         touch_softlockup_watchdog();
 567         /*
 568          * Cancel the scheduled timer and restore the tick
 569          */
 570         ts->tick_stopped  = 0;
 571         ts->idle_exittime = now;
 572
 573         tick_nohz_restart(ts, now);
 574
 575         local_irq_enable();
 576 }
 577
 578 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now)
 579 {
 580         hrtimer_forward(&ts->sched_timer, now, tick_period);
 581         return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0);
 582 }
 583
 584 /*
 585  * The nohz low res interrupt handler
 586  */
 587 static void tick_nohz_handler(struct clock_event_device *dev)
 588 {
 589         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 590         struct pt_regs *regs = get_irq_regs();
 591         int cpu = smp_processor_id();
 592         ktime_t now = ktime_get();
 593
 594         dev->next_event.tv64 = KTIME_MAX;
 595
 596         /*
 597          * Check if the do_timer duty was dropped. We don't care about
 598          * concurrency: This happens only when the cpu in charge went
 599          * into a long sleep. If two cpus happen to assign themself to
 600          * this duty, then the jiffies update is still serialized by
 601          * xtime_lock.
 602          */
 603         if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
 604                 tick_do_timer_cpu = cpu;
 605
 606         /* Check, if the jiffies need an update */
 607         if (tick_do_timer_cpu == cpu)
 608                 tick_do_update_jiffies64(now);
 609
 610         /*
 611          * When we are idle and the tick is stopped, we have to touch
 612          * the watchdog as we might not schedule for a really long
 613          * time. This happens on complete idle SMP systems while
 614          * waiting on the login prompt. We also increment the "start
 615          * of idle" jiffy stamp so the idle accounting adjustment we
 616          * do when we go busy again does not account too much ticks.
 617          */
 618         if (ts->tick_stopped) {
 619                 touch_softlockup_watchdog();
 620                 ts->idle_jiffies++;
 621         }
 622
 623         update_process_times(user_mode(regs));
 624         profile_tick(CPU_PROFILING);
 625
 626         while (tick_nohz_reprogram(ts, now)) {
 627                 now = ktime_get();
 628                 tick_do_update_jiffies64(now);
 629         }
 630 }
 631
 632 /**
 633  * tick_nohz_switch_to_nohz - switch to nohz mode
 634  */
 635 static void tick_nohz_switch_to_nohz(void)
 636 {
 637         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 638         ktime_t next;
 639
 640         if (!tick_nohz_enabled)
 641                 return;
 642
 643         local_irq_disable();
 644         if (tick_switch_to_oneshot(tick_nohz_handler)) {
 645                 local_irq_enable();
 646                 return;
 647         }
 648
 649         ts->nohz_mode = NOHZ_MODE_LOWRES;
 650
 651         /*
 652          * Recycle the hrtimer in ts, so we can share the
 653          * hrtimer_forward with the highres code.
 654          */
 655         hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
 656         /* Get the next period */
 657         next = tick_init_jiffy_update();
 658
 659         for (;;) {
 660                 hrtimer_set_expires(&ts->sched_timer, next);
 661                 if (!tick_program_event(next, 0))
 662                         break;
 663                 next = ktime_add(next, tick_period);
 664         }
 665         local_irq_enable();
 666 }
 667
 668 /*
 669  * When NOHZ is enabled and the tick is stopped, we need to kick the
 670  * tick timer from irq_enter() so that the jiffies update is kept
 671  * alive during long running softirqs. That's ugly as hell, but
 672  * correctness is key even if we need to fix the offending softirq in
 673  * the first place.
 674  *
 675  * Note, this is different to tick_nohz_restart. We just kick the
 676  * timer and do not touch the other magic bits which need to be done
 677  * when idle is left.
 678  */
 679 static void tick_nohz_kick_tick(int cpu, ktime_t now)
 680 {
 681 #if 0
 682         /* Switch back to 2.6.27 behaviour */
 683
 684         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 685         ktime_t delta;
 686
 687         /*
 688          * Do not touch the tick device, when the next expiry is either
 689          * already reached or less/equal than the tick period.
 690          */
 691         delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now);
 692         if (delta.tv64 <= tick_period.tv64)
 693                 return;
 694
 695         tick_nohz_restart(ts, now);
 696 #endif
 697 }
 698
 699 static inline void tick_check_nohz(int cpu)
 700 {
 701         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 702         ktime_t now;
 703
 704         if (!ts->idle_active && !ts->tick_stopped)
 705                 return;
 706         now = ktime_get();
 707         if (ts->idle_active)
 708                 tick_nohz_stop_idle(cpu, now);
 709         if (ts->tick_stopped) {
 710                 tick_nohz_update_jiffies(now);
 711                 tick_nohz_kick_tick(cpu, now);
 712         }
 713 }
 714
 715 #else
 716
 717 static inline void tick_nohz_switch_to_nohz(void) { }
 718 static inline void tick_check_nohz(int cpu) { }
 719
 720 #endif /* NO_HZ */
 721
 722 /*
 723  * Called from irq_enter to notify about the possible interruption of idle()
 724  */
 725 void tick_check_idle(int cpu)
 726 {
 727         tick_check_oneshot_broadcast(cpu);
 728         tick_check_nohz(cpu);
 729 }
 730
 731 /*
 732  * High resolution timer specific code
 733  */
 734 #ifdef CONFIG_HIGH_RES_TIMERS
 735 /*
 736  * We rearm the timer until we get disabled by the idle code.
 737  * Called with interrupts disabled and timer->base->cpu_base->lock held.
 738  */
 739 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
 740 {
 741         struct tick_sched *ts =
 742                 container_of(timer, struct tick_sched, sched_timer);
 743         struct pt_regs *regs = get_irq_regs();
 744         ktime_t now = ktime_get();
 745         int cpu = smp_processor_id();
 746
 747 #ifdef CONFIG_NO_HZ
 748         /*
 749          * Check if the do_timer duty was dropped. We don't care about
 750          * concurrency: This happens only when the cpu in charge went
 751          * into a long sleep. If two cpus happen to assign themself to
 752          * this duty, then the jiffies update is still serialized by
 753          * xtime_lock.
 754          */
 755         if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE))
 756                 tick_do_timer_cpu = cpu;
 757 #endif
 758
 759         /* Check, if the jiffies need an update */
 760         if (tick_do_timer_cpu == cpu)
 761                 tick_do_update_jiffies64(now);
 762
 763         /*
 764          * Do not call, when we are not in irq context and have
 765          * no valid regs pointer
 766          */
 767         if (regs) {
 768                 /*
 769                  * When we are idle and the tick is stopped, we have to touch
 770                  * the watchdog as we might not schedule for a really long
 771                  * time. This happens on complete idle SMP systems while
 772                  * waiting on the login prompt. We also increment the "start of
 773                  * idle" jiffy stamp so the idle accounting adjustment we do
 774                  * when we go busy again does not account too much ticks.
 775                  */
 776                 if (ts->tick_stopped) {
 777                         touch_softlockup_watchdog();
 778                         ts->idle_jiffies++;
 779                 }
 780                 update_process_times(user_mode(regs));
 781                 profile_tick(CPU_PROFILING);
 782         }
 783
 784         hrtimer_forward(timer, now, tick_period);
 785
 786         return HRTIMER_RESTART;
 787 }
 788
 789 /**
 790  * tick_setup_sched_timer - setup the tick emulation timer
 791  */
 792 void tick_setup_sched_timer(void)
 793 {
 794         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 795         ktime_t now = ktime_get();
 796
 797         /*
 798          * Emulate tick processing via per-CPU hrtimers:
 799          */
 800         hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
 801         ts->sched_timer.function = tick_sched_timer;
 802
 803         /* Get the next period (per cpu) */
 804         hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
 805
 806         for (;;) {
 807                 hrtimer_forward(&ts->sched_timer, now, tick_period);
 808                 hrtimer_start_expires(&ts->sched_timer,
 809                                       HRTIMER_MODE_ABS_PINNED);
 810                 /* Check, if the timer was already in the past */
 811                 if (hrtimer_active(&ts->sched_timer))
 812                         break;
 813                 now = ktime_get();
 814         }
 815
 816 #ifdef CONFIG_NO_HZ
 817         if (tick_nohz_enabled)
 818                 ts->nohz_mode = NOHZ_MODE_HIGHRES;
 819 #endif
 820 }
 821 #endif /* HIGH_RES_TIMERS */
 822
 823 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS
 824 void tick_cancel_sched_timer(int cpu)
 825 {
 826         struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 827
 828 # ifdef CONFIG_HIGH_RES_TIMERS
 829         if (ts->sched_timer.base)
 830                 hrtimer_cancel(&ts->sched_timer);
 831 # endif
 832
 833         ts->nohz_mode = NOHZ_MODE_INACTIVE;
 834 }
 835 #endif
 836
 837 /**
 838  * Async notification about clocksource changes
 839  */
 840 void tick_clock_notify(void)
 841 {
 842         int cpu;
 843
 844         for_each_possible_cpu(cpu)
 845                 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
 846 }
 847
 848 /*
 849  * Async notification about clock event changes
 850  */
 851 void tick_oneshot_notify(void)
 852 {
 853         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 854
 855         set_bit(0, &ts->check_clocks);
 856 }
 857
 858 /**
 859  * Check, if a change happened, which makes oneshot possible.
 860  *
 861  * Called cyclic from the hrtimer softirq (driven by the timer
 862  * softirq) allow_nohz signals, that we can switch into low-res nohz
 863  * mode, because high resolution timers are disabled (either compile
 864  * or runtime).
 865  */
 866 int tick_check_oneshot_change(int allow_nohz)
 867 {
 868         struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched);
 869
 870         if (!test_and_clear_bit(0, &ts->check_clocks))
 871                 return 0;
 872
 873         if (ts->nohz_mode != NOHZ_MODE_INACTIVE)
 874                 return 0;
 875
 876         if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
 877                 return 0;
 878
 879         if (!allow_nohz)
 880                 return 1;
 881
 882         tick_nohz_switch_to_nohz();
 883         return 0;
 884 }