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 static void tick_nohz_update_jiffies(ktime_t now) 138 { 139 int cpu = smp_processor_id(); 140 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 141 unsigned long flags; 142 143 cpumask_clear_cpu(cpu, nohz_cpu_mask); 144 ts->idle_waketime = now; 145 146 local_irq_save(flags); 147 tick_do_update_jiffies64(now); 148 local_irq_restore(flags); 149 150 touch_softlockup_watchdog(); 151 } 152 153 /* 154 * Updates the per cpu time idle statistics counters 155 */ 156 static void 157 update_ts_time_stats(struct tick_sched *ts, ktime_t now, u64 *last_update_time) 158 { 159 ktime_t delta; 160 161 if (ts->idle_active) { 162 delta = ktime_sub(now, ts->idle_entrytime); 163 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 164 if (nr_iowait_cpu() > 0) 165 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 166 ts->idle_entrytime = now; 167 } 168 169 if (last_update_time) 170 *last_update_time = ktime_to_us(now); 171 172 } 173 174 static void tick_nohz_stop_idle(int cpu, ktime_t now) 175 { 176 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 177 178 update_ts_time_stats(ts, now, NULL); 179 ts->idle_active = 0; 180 181 sched_clock_idle_wakeup_event(0); 182 } 183 184 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 185 { 186 ktime_t now; 187 188 now = ktime_get(); 189 190 update_ts_time_stats(ts, now, NULL); 191 192 ts->idle_entrytime = now; 193 ts->idle_active = 1; 194 sched_clock_idle_sleep_event(); 195 return now; 196 } 197 198 /** 199 * get_cpu_idle_time_us - get the total idle time of a cpu 200 * @cpu: CPU number to query 201 * @last_update_time: variable to store update time in 202 * 203 * Return the cummulative idle time (since boot) for a given 204 * CPU, in microseconds. The idle time returned includes 205 * the iowait time (unlike what "top" and co report). 206 * 207 * This time is measured via accounting rather than sampling, 208 * and is as accurate as ktime_get() is. 209 * 210 * This function returns -1 if NOHZ is not enabled. 211 */ 212 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 213 { 214 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 215 216 if (!tick_nohz_enabled) 217 return -1; 218 219 update_ts_time_stats(ts, ktime_get(), last_update_time); 220 221 return ktime_to_us(ts->idle_sleeptime); 222 } 223 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 224 225 /* 226 * get_cpu_iowait_time_us - get the total iowait time of a cpu 227 * @cpu: CPU number to query 228 * @last_update_time: variable to store update time in 229 * 230 * Return the cummulative iowait time (since boot) for a given 231 * CPU, in microseconds. 232 * 233 * This time is measured via accounting rather than sampling, 234 * and is as accurate as ktime_get() is. 235 * 236 * This function returns -1 if NOHZ is not enabled. 237 */ 238 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 239 { 240 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 241 242 if (!tick_nohz_enabled) 243 return -1; 244 245 update_ts_time_stats(ts, ktime_get(), last_update_time); 246 247 return ktime_to_us(ts->iowait_sleeptime); 248 } 249 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 250 251 /** 252 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task 253 * 254 * When the next event is more than a tick into the future, stop the idle tick 255 * Called either from the idle loop or from irq_exit() when an idle period was 256 * just interrupted by an interrupt which did not cause a reschedule. 257 */ 258 void tick_nohz_stop_sched_tick(int inidle) 259 { 260 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; 261 struct tick_sched *ts; 262 ktime_t last_update, expires, now; 263 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; 264 u64 time_delta; 265 int cpu; 266 267 local_irq_save(flags); 268 269 cpu = smp_processor_id(); 270 ts = &per_cpu(tick_cpu_sched, cpu); 271 272 /* 273 * Call to tick_nohz_start_idle stops the last_update_time from being 274 * updated. Thus, it must not be called in the event we are called from 275 * irq_exit() with the prior state different than idle. 276 */ 277 if (!inidle && !ts->inidle) 278 goto end; 279 280 /* 281 * Set ts->inidle unconditionally. Even if the system did not 282 * switch to NOHZ mode the cpu frequency governers rely on the 283 * update of the idle time accounting in tick_nohz_start_idle(). 284 */ 285 ts->inidle = 1; 286 287 now = tick_nohz_start_idle(ts); 288 289 /* 290 * If this cpu is offline and it is the one which updates 291 * jiffies, then give up the assignment and let it be taken by 292 * the cpu which runs the tick timer next. If we don't drop 293 * this here the jiffies might be stale and do_timer() never 294 * invoked. 295 */ 296 if (unlikely(!cpu_online(cpu))) { 297 if (cpu == tick_do_timer_cpu) 298 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 299 } 300 301 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) 302 goto end; 303 304 if (need_resched()) 305 goto end; 306 307 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 308 static int ratelimit; 309 310 if (ratelimit < 10) { 311 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", 312 (unsigned int) local_softirq_pending()); 313 ratelimit++; 314 } 315 goto end; 316 } 317 318 ts->idle_calls++; 319 /* Read jiffies and the time when jiffies were updated last */ 320 do { 321 seq = read_seqbegin(&xtime_lock); 322 last_update = last_jiffies_update; 323 last_jiffies = jiffies; 324 time_delta = timekeeping_max_deferment(); 325 } while (read_seqretry(&xtime_lock, seq)); 326 327 if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) || 328 arch_needs_cpu(cpu) || nohz_ratelimit(cpu)) { 329 next_jiffies = last_jiffies + 1; 330 delta_jiffies = 1; 331 } else { 332 /* Get the next timer wheel timer */ 333 next_jiffies = get_next_timer_interrupt(last_jiffies); 334 delta_jiffies = next_jiffies - last_jiffies; 335 } 336 /* 337 * Do not stop the tick, if we are only one off 338 * or if the cpu is required for rcu 339 */ 340 if (!ts->tick_stopped && delta_jiffies == 1) 341 goto out; 342 343 /* Schedule the tick, if we are at least one jiffie off */ 344 if ((long)delta_jiffies >= 1) { 345 346 /* 347 * If this cpu is the one which updates jiffies, then 348 * give up the assignment and let it be taken by the 349 * cpu which runs the tick timer next, which might be 350 * this cpu as well. If we don't drop this here the 351 * jiffies might be stale and do_timer() never 352 * invoked. Keep track of the fact that it was the one 353 * which had the do_timer() duty last. If this cpu is 354 * the one which had the do_timer() duty last, we 355 * limit the sleep time to the timekeeping 356 * max_deferement value which we retrieved 357 * above. Otherwise we can sleep as long as we want. 358 */ 359 if (cpu == tick_do_timer_cpu) { 360 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 361 ts->do_timer_last = 1; 362 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 363 time_delta = KTIME_MAX; 364 ts->do_timer_last = 0; 365 } else if (!ts->do_timer_last) { 366 time_delta = KTIME_MAX; 367 } 368 369 /* 370 * calculate the expiry time for the next timer wheel 371 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals 372 * that there is no timer pending or at least extremely 373 * far into the future (12 days for HZ=1000). In this 374 * case we set the expiry to the end of time. 375 */ 376 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { 377 /* 378 * Calculate the time delta for the next timer event. 379 * If the time delta exceeds the maximum time delta 380 * permitted by the current clocksource then adjust 381 * the time delta accordingly to ensure the 382 * clocksource does not wrap. 383 */ 384 time_delta = min_t(u64, time_delta, 385 tick_period.tv64 * delta_jiffies); 386 } 387 388 if (time_delta < KTIME_MAX) 389 expires = ktime_add_ns(last_update, time_delta); 390 else 391 expires.tv64 = KTIME_MAX; 392 393 if (delta_jiffies > 1) 394 cpumask_set_cpu(cpu, nohz_cpu_mask); 395 396 /* Skip reprogram of event if its not changed */ 397 if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) 398 goto out; 399 400 /* 401 * nohz_stop_sched_tick can be called several times before 402 * the nohz_restart_sched_tick is called. This happens when 403 * interrupts arrive which do not cause a reschedule. In the 404 * first call we save the current tick time, so we can restart 405 * the scheduler tick in nohz_restart_sched_tick. 406 */ 407 if (!ts->tick_stopped) { 408 if (select_nohz_load_balancer(1)) { 409 /* 410 * sched tick not stopped! 411 */ 412 cpumask_clear_cpu(cpu, nohz_cpu_mask); 413 goto out; 414 } 415 416 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer); 417 ts->tick_stopped = 1; 418 ts->idle_jiffies = last_jiffies; 419 rcu_enter_nohz(); 420 } 421 422 ts->idle_sleeps++; 423 424 /* Mark expires */ 425 ts->idle_expires = expires; 426 427 /* 428 * If the expiration time == KTIME_MAX, then 429 * in this case we simply stop the tick timer. 430 */ 431 if (unlikely(expires.tv64 == KTIME_MAX)) { 432 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 433 hrtimer_cancel(&ts->sched_timer); 434 goto out; 435 } 436 437 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 438 hrtimer_start(&ts->sched_timer, expires, 439 HRTIMER_MODE_ABS_PINNED); 440 /* Check, if the timer was already in the past */ 441 if (hrtimer_active(&ts->sched_timer)) 442 goto out; 443 } else if (!tick_program_event(expires, 0)) 444 goto out; 445 /* 446 * We are past the event already. So we crossed a 447 * jiffie boundary. Update jiffies and raise the 448 * softirq. 449 */ 450 tick_do_update_jiffies64(ktime_get()); 451 cpumask_clear_cpu(cpu, nohz_cpu_mask); 452 } 453 raise_softirq_irqoff(TIMER_SOFTIRQ); 454 out: 455 ts->next_jiffies = next_jiffies; 456 ts->last_jiffies = last_jiffies; 457 ts->sleep_length = ktime_sub(dev->next_event, now); 458 end: 459 local_irq_restore(flags); 460 } 461 462 /** 463 * tick_nohz_get_sleep_length - return the length of the current sleep 464 * 465 * Called from power state control code with interrupts disabled 466 */ 467 ktime_t tick_nohz_get_sleep_length(void) 468 { 469 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 470 471 return ts->sleep_length; 472 } 473 474 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 475 { 476 hrtimer_cancel(&ts->sched_timer); 477 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick); 478 479 while (1) { 480 /* Forward the time to expire in the future */ 481 hrtimer_forward(&ts->sched_timer, now, tick_period); 482 483 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 484 hrtimer_start_expires(&ts->sched_timer, 485 HRTIMER_MODE_ABS_PINNED); 486 /* Check, if the timer was already in the past */ 487 if (hrtimer_active(&ts->sched_timer)) 488 break; 489 } else { 490 if (!tick_program_event( 491 hrtimer_get_expires(&ts->sched_timer), 0)) 492 break; 493 } 494 /* Update jiffies and reread time */ 495 tick_do_update_jiffies64(now); 496 now = ktime_get(); 497 } 498 } 499 500 /** 501 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task 502 * 503 * Restart the idle tick when the CPU is woken up from idle 504 */ 505 void tick_nohz_restart_sched_tick(void) 506 { 507 int cpu = smp_processor_id(); 508 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 509 #ifndef CONFIG_VIRT_CPU_ACCOUNTING 510 unsigned long ticks; 511 #endif 512 ktime_t now; 513 514 local_irq_disable(); 515 if (ts->idle_active || (ts->inidle && ts->tick_stopped)) 516 now = ktime_get(); 517 518 if (ts->idle_active) 519 tick_nohz_stop_idle(cpu, now); 520 521 if (!ts->inidle || !ts->tick_stopped) { 522 ts->inidle = 0; 523 local_irq_enable(); 524 return; 525 } 526 527 ts->inidle = 0; 528 529 rcu_exit_nohz(); 530 531 /* Update jiffies first */ 532 select_nohz_load_balancer(0); 533 tick_do_update_jiffies64(now); 534 cpumask_clear_cpu(cpu, nohz_cpu_mask); 535 536 #ifndef CONFIG_VIRT_CPU_ACCOUNTING 537 /* 538 * We stopped the tick in idle. Update process times would miss the 539 * time we slept as update_process_times does only a 1 tick 540 * accounting. Enforce that this is accounted to idle ! 541 */ 542 ticks = jiffies - ts->idle_jiffies; 543 /* 544 * We might be one off. Do not randomly account a huge number of ticks! 545 */ 546 if (ticks && ticks < LONG_MAX) 547 account_idle_ticks(ticks); 548 #endif 549 550 touch_softlockup_watchdog(); 551 /* 552 * Cancel the scheduled timer and restore the tick 553 */ 554 ts->tick_stopped = 0; 555 ts->idle_exittime = now; 556 557 tick_nohz_restart(ts, now); 558 559 local_irq_enable(); 560 } 561 562 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 563 { 564 hrtimer_forward(&ts->sched_timer, now, tick_period); 565 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); 566 } 567 568 /* 569 * The nohz low res interrupt handler 570 */ 571 static void tick_nohz_handler(struct clock_event_device *dev) 572 { 573 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 574 struct pt_regs *regs = get_irq_regs(); 575 int cpu = smp_processor_id(); 576 ktime_t now = ktime_get(); 577 578 dev->next_event.tv64 = KTIME_MAX; 579 580 /* 581 * Check if the do_timer duty was dropped. We don't care about 582 * concurrency: This happens only when the cpu in charge went 583 * into a long sleep. If two cpus happen to assign themself to 584 * this duty, then the jiffies update is still serialized by 585 * xtime_lock. 586 */ 587 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) 588 tick_do_timer_cpu = cpu; 589 590 /* Check, if the jiffies need an update */ 591 if (tick_do_timer_cpu == cpu) 592 tick_do_update_jiffies64(now); 593 594 /* 595 * When we are idle and the tick is stopped, we have to touch 596 * the watchdog as we might not schedule for a really long 597 * time. This happens on complete idle SMP systems while 598 * waiting on the login prompt. We also increment the "start 599 * of idle" jiffy stamp so the idle accounting adjustment we 600 * do when we go busy again does not account too much ticks. 601 */ 602 if (ts->tick_stopped) { 603 touch_softlockup_watchdog(); 604 ts->idle_jiffies++; 605 } 606 607 update_process_times(user_mode(regs)); 608 profile_tick(CPU_PROFILING); 609 610 while (tick_nohz_reprogram(ts, now)) { 611 now = ktime_get(); 612 tick_do_update_jiffies64(now); 613 } 614 } 615 616 /** 617 * tick_nohz_switch_to_nohz - switch to nohz mode 618 */ 619 static void tick_nohz_switch_to_nohz(void) 620 { 621 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 622 ktime_t next; 623 624 if (!tick_nohz_enabled) 625 return; 626 627 local_irq_disable(); 628 if (tick_switch_to_oneshot(tick_nohz_handler)) { 629 local_irq_enable(); 630 return; 631 } 632 633 ts->nohz_mode = NOHZ_MODE_LOWRES; 634 635 /* 636 * Recycle the hrtimer in ts, so we can share the 637 * hrtimer_forward with the highres code. 638 */ 639 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 640 /* Get the next period */ 641 next = tick_init_jiffy_update(); 642 643 for (;;) { 644 hrtimer_set_expires(&ts->sched_timer, next); 645 if (!tick_program_event(next, 0)) 646 break; 647 next = ktime_add(next, tick_period); 648 } 649 local_irq_enable(); 650 651 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", 652 smp_processor_id()); 653 } 654 655 /* 656 * When NOHZ is enabled and the tick is stopped, we need to kick the 657 * tick timer from irq_enter() so that the jiffies update is kept 658 * alive during long running softirqs. That's ugly as hell, but 659 * correctness is key even if we need to fix the offending softirq in 660 * the first place. 661 * 662 * Note, this is different to tick_nohz_restart. We just kick the 663 * timer and do not touch the other magic bits which need to be done 664 * when idle is left. 665 */ 666 static void tick_nohz_kick_tick(int cpu, ktime_t now) 667 { 668 #if 0 669 /* Switch back to 2.6.27 behaviour */ 670 671 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 672 ktime_t delta; 673 674 /* 675 * Do not touch the tick device, when the next expiry is either 676 * already reached or less/equal than the tick period. 677 */ 678 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 679 if (delta.tv64 <= tick_period.tv64) 680 return; 681 682 tick_nohz_restart(ts, now); 683 #endif 684 } 685 686 static inline void tick_check_nohz(int cpu) 687 { 688 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 689 ktime_t now; 690 691 if (!ts->idle_active && !ts->tick_stopped) 692 return; 693 now = ktime_get(); 694 if (ts->idle_active) 695 tick_nohz_stop_idle(cpu, now); 696 if (ts->tick_stopped) { 697 tick_nohz_update_jiffies(now); 698 tick_nohz_kick_tick(cpu, now); 699 } 700 } 701 702 #else 703 704 static inline void tick_nohz_switch_to_nohz(void) { } 705 static inline void tick_check_nohz(int cpu) { } 706 707 #endif /* NO_HZ */ 708 709 /* 710 * Called from irq_enter to notify about the possible interruption of idle() 711 */ 712 void tick_check_idle(int cpu) 713 { 714 tick_check_oneshot_broadcast(cpu); 715 tick_check_nohz(cpu); 716 } 717 718 /* 719 * High resolution timer specific code 720 */ 721 #ifdef CONFIG_HIGH_RES_TIMERS 722 /* 723 * We rearm the timer until we get disabled by the idle code. 724 * Called with interrupts disabled and timer->base->cpu_base->lock held. 725 */ 726 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 727 { 728 struct tick_sched *ts = 729 container_of(timer, struct tick_sched, sched_timer); 730 struct pt_regs *regs = get_irq_regs(); 731 ktime_t now = ktime_get(); 732 int cpu = smp_processor_id(); 733 734 #ifdef CONFIG_NO_HZ 735 /* 736 * Check if the do_timer duty was dropped. We don't care about 737 * concurrency: This happens only when the cpu in charge went 738 * into a long sleep. If two cpus happen to assign themself to 739 * this duty, then the jiffies update is still serialized by 740 * xtime_lock. 741 */ 742 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) 743 tick_do_timer_cpu = cpu; 744 #endif 745 746 /* Check, if the jiffies need an update */ 747 if (tick_do_timer_cpu == cpu) 748 tick_do_update_jiffies64(now); 749 750 /* 751 * Do not call, when we are not in irq context and have 752 * no valid regs pointer 753 */ 754 if (regs) { 755 /* 756 * When we are idle and the tick is stopped, we have to touch 757 * the watchdog as we might not schedule for a really long 758 * time. This happens on complete idle SMP systems while 759 * waiting on the login prompt. We also increment the "start of 760 * idle" jiffy stamp so the idle accounting adjustment we do 761 * when we go busy again does not account too much ticks. 762 */ 763 if (ts->tick_stopped) { 764 touch_softlockup_watchdog(); 765 ts->idle_jiffies++; 766 } 767 update_process_times(user_mode(regs)); 768 profile_tick(CPU_PROFILING); 769 } 770 771 hrtimer_forward(timer, now, tick_period); 772 773 return HRTIMER_RESTART; 774 } 775 776 /** 777 * tick_setup_sched_timer - setup the tick emulation timer 778 */ 779 void tick_setup_sched_timer(void) 780 { 781 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 782 ktime_t now = ktime_get(); 783 u64 offset; 784 785 /* 786 * Emulate tick processing via per-CPU hrtimers: 787 */ 788 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 789 ts->sched_timer.function = tick_sched_timer; 790 791 /* Get the next period (per cpu) */ 792 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 793 offset = ktime_to_ns(tick_period) >> 1; 794 do_div(offset, num_possible_cpus()); 795 offset *= smp_processor_id(); 796 hrtimer_add_expires_ns(&ts->sched_timer, offset); 797 798 for (;;) { 799 hrtimer_forward(&ts->sched_timer, now, tick_period); 800 hrtimer_start_expires(&ts->sched_timer, 801 HRTIMER_MODE_ABS_PINNED); 802 /* Check, if the timer was already in the past */ 803 if (hrtimer_active(&ts->sched_timer)) 804 break; 805 now = ktime_get(); 806 } 807 808 #ifdef CONFIG_NO_HZ 809 if (tick_nohz_enabled) 810 ts->nohz_mode = NOHZ_MODE_HIGHRES; 811 #endif 812 } 813 #endif /* HIGH_RES_TIMERS */ 814 815 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS 816 void tick_cancel_sched_timer(int cpu) 817 { 818 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 819 820 # ifdef CONFIG_HIGH_RES_TIMERS 821 if (ts->sched_timer.base) 822 hrtimer_cancel(&ts->sched_timer); 823 # endif 824 825 ts->nohz_mode = NOHZ_MODE_INACTIVE; 826 } 827 #endif 828 829 /** 830 * Async notification about clocksource changes 831 */ 832 void tick_clock_notify(void) 833 { 834 int cpu; 835 836 for_each_possible_cpu(cpu) 837 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 838 } 839 840 /* 841 * Async notification about clock event changes 842 */ 843 void tick_oneshot_notify(void) 844 { 845 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 846 847 set_bit(0, &ts->check_clocks); 848 } 849 850 /** 851 * Check, if a change happened, which makes oneshot possible. 852 * 853 * Called cyclic from the hrtimer softirq (driven by the timer 854 * softirq) allow_nohz signals, that we can switch into low-res nohz 855 * mode, because high resolution timers are disabled (either compile 856 * or runtime). 857 */ 858 int tick_check_oneshot_change(int allow_nohz) 859 { 860 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 861 862 if (!test_and_clear_bit(0, &ts->check_clocks)) 863 return 0; 864 865 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 866 return 0; 867 868 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 869 return 0; 870 871 if (!allow_nohz) 872 return 1; 873 874 tick_nohz_switch_to_nohz(); 875 return 0; 876 } 877