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