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#include <linux/irq_work.h> 24#include <linux/posix-timers.h> 25#include <linux/perf_event.h> 26#include <linux/context_tracking.h> 27 28#include <asm/irq_regs.h> 29 30#include "tick-internal.h" 31 32#include <trace/events/timer.h> 33 34/* 35 * Per cpu nohz control structure 36 */ 37DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); 38 39/* 40 * The time, when the last jiffy update happened. Protected by jiffies_lock. 41 */ 42static ktime_t last_jiffies_update; 43 44struct tick_sched *tick_get_tick_sched(int cpu) 45{ 46 return &per_cpu(tick_cpu_sched, cpu); 47} 48 49/* 50 * Must be called with interrupts disabled ! 51 */ 52static void tick_do_update_jiffies64(ktime_t now) 53{ 54 unsigned long ticks = 0; 55 ktime_t delta; 56 57 /* 58 * Do a quick check without holding jiffies_lock: 59 */ 60 delta = ktime_sub(now, last_jiffies_update); 61 if (delta.tv64 < tick_period.tv64) 62 return; 63 64 /* Reevalute with jiffies_lock held */ 65 write_seqlock(&jiffies_lock); 66 67 delta = ktime_sub(now, last_jiffies_update); 68 if (delta.tv64 >= tick_period.tv64) { 69 70 delta = ktime_sub(delta, tick_period); 71 last_jiffies_update = ktime_add(last_jiffies_update, 72 tick_period); 73 74 /* Slow path for long timeouts */ 75 if (unlikely(delta.tv64 >= tick_period.tv64)) { 76 s64 incr = ktime_to_ns(tick_period); 77 78 ticks = ktime_divns(delta, incr); 79 80 last_jiffies_update = ktime_add_ns(last_jiffies_update, 81 incr * ticks); 82 } 83 do_timer(++ticks); 84 85 /* Keep the tick_next_period variable up to date */ 86 tick_next_period = ktime_add(last_jiffies_update, tick_period); 87 } else { 88 write_sequnlock(&jiffies_lock); 89 return; 90 } 91 write_sequnlock(&jiffies_lock); 92 update_wall_time(); 93} 94 95/* 96 * Initialize and return retrieve the jiffies update. 97 */ 98static ktime_t tick_init_jiffy_update(void) 99{ 100 ktime_t period; 101 102 write_seqlock(&jiffies_lock); 103 /* Did we start the jiffies update yet ? */ 104 if (last_jiffies_update.tv64 == 0) 105 last_jiffies_update = tick_next_period; 106 period = last_jiffies_update; 107 write_sequnlock(&jiffies_lock); 108 return period; 109} 110 111 112static void tick_sched_do_timer(ktime_t now) 113{ 114 int cpu = smp_processor_id(); 115 116#ifdef CONFIG_NO_HZ_COMMON 117 /* 118 * Check if the do_timer duty was dropped. We don't care about 119 * concurrency: This happens only when the cpu in charge went 120 * into a long sleep. If two cpus happen to assign themself to 121 * this duty, then the jiffies update is still serialized by 122 * jiffies_lock. 123 */ 124 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE) 125 && !tick_nohz_full_cpu(cpu)) 126 tick_do_timer_cpu = cpu; 127#endif 128 129 /* Check, if the jiffies need an update */ 130 if (tick_do_timer_cpu == cpu) 131 tick_do_update_jiffies64(now); 132} 133 134static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) 135{ 136#ifdef CONFIG_NO_HZ_COMMON 137 /* 138 * When we are idle and the tick is stopped, we have to touch 139 * the watchdog as we might not schedule for a really long 140 * time. This happens on complete idle SMP systems while 141 * waiting on the login prompt. We also increment the "start of 142 * idle" jiffy stamp so the idle accounting adjustment we do 143 * when we go busy again does not account too much ticks. 144 */ 145 if (ts->tick_stopped) { 146 touch_softlockup_watchdog(); 147 if (is_idle_task(current)) 148 ts->idle_jiffies++; 149 } 150#endif 151 update_process_times(user_mode(regs)); 152 profile_tick(CPU_PROFILING); 153} 154 155#ifdef CONFIG_NO_HZ_FULL 156cpumask_var_t tick_nohz_full_mask; 157cpumask_var_t housekeeping_mask; 158bool tick_nohz_full_running; 159 160static bool can_stop_full_tick(void) 161{ 162 WARN_ON_ONCE(!irqs_disabled()); 163 164 if (!sched_can_stop_tick()) { 165 trace_tick_stop(0, "more than 1 task in runqueue\n"); 166 return false; 167 } 168 169 if (!posix_cpu_timers_can_stop_tick(current)) { 170 trace_tick_stop(0, "posix timers running\n"); 171 return false; 172 } 173 174 if (!perf_event_can_stop_tick()) { 175 trace_tick_stop(0, "perf events running\n"); 176 return false; 177 } 178 179 /* sched_clock_tick() needs us? */ 180#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK 181 /* 182 * TODO: kick full dynticks CPUs when 183 * sched_clock_stable is set. 184 */ 185 if (!sched_clock_stable()) { 186 trace_tick_stop(0, "unstable sched clock\n"); 187 /* 188 * Don't allow the user to think they can get 189 * full NO_HZ with this machine. 190 */ 191 WARN_ONCE(tick_nohz_full_running, 192 "NO_HZ FULL will not work with unstable sched clock"); 193 return false; 194 } 195#endif 196 197 return true; 198} 199 200static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now); 201 202/* 203 * Re-evaluate the need for the tick on the current CPU 204 * and restart it if necessary. 205 */ 206void __tick_nohz_full_check(void) 207{ 208 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 209 210 if (tick_nohz_full_cpu(smp_processor_id())) { 211 if (ts->tick_stopped && !is_idle_task(current)) { 212 if (!can_stop_full_tick()) 213 tick_nohz_restart_sched_tick(ts, ktime_get()); 214 } 215 } 216} 217 218static void nohz_full_kick_work_func(struct irq_work *work) 219{ 220 __tick_nohz_full_check(); 221} 222 223static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 224 .func = nohz_full_kick_work_func, 225}; 226 227/* 228 * Kick this CPU if it's full dynticks in order to force it to 229 * re-evaluate its dependency on the tick and restart it if necessary. 230 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), 231 * is NMI safe. 232 */ 233void tick_nohz_full_kick(void) 234{ 235 if (!tick_nohz_full_cpu(smp_processor_id())) 236 return; 237 238 irq_work_queue(&__get_cpu_var(nohz_full_kick_work)); 239} 240 241/* 242 * Kick the CPU if it's full dynticks in order to force it to 243 * re-evaluate its dependency on the tick and restart it if necessary. 244 */ 245void tick_nohz_full_kick_cpu(int cpu) 246{ 247 if (!tick_nohz_full_cpu(cpu)) 248 return; 249 250 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); 251} 252 253static void nohz_full_kick_ipi(void *info) 254{ 255 __tick_nohz_full_check(); 256} 257 258/* 259 * Kick all full dynticks CPUs in order to force these to re-evaluate 260 * their dependency on the tick and restart it if necessary. 261 */ 262void tick_nohz_full_kick_all(void) 263{ 264 if (!tick_nohz_full_running) 265 return; 266 267 preempt_disable(); 268 smp_call_function_many(tick_nohz_full_mask, 269 nohz_full_kick_ipi, NULL, false); 270 tick_nohz_full_kick(); 271 preempt_enable(); 272} 273 274/* 275 * Re-evaluate the need for the tick as we switch the current task. 276 * It might need the tick due to per task/process properties: 277 * perf events, posix cpu timers, ... 278 */ 279void __tick_nohz_task_switch(struct task_struct *tsk) 280{ 281 unsigned long flags; 282 283 local_irq_save(flags); 284 285 if (!tick_nohz_full_cpu(smp_processor_id())) 286 goto out; 287 288 if (tick_nohz_tick_stopped() && !can_stop_full_tick()) 289 tick_nohz_full_kick(); 290 291out: 292 local_irq_restore(flags); 293} 294 295/* Parse the boot-time nohz CPU list from the kernel parameters. */ 296static int __init tick_nohz_full_setup(char *str) 297{ 298 alloc_bootmem_cpumask_var(&tick_nohz_full_mask); 299 if (cpulist_parse(str, tick_nohz_full_mask) < 0) { 300 pr_warning("NOHZ: Incorrect nohz_full cpumask\n"); 301 free_bootmem_cpumask_var(tick_nohz_full_mask); 302 return 1; 303 } 304 tick_nohz_full_running = true; 305 306 return 1; 307} 308__setup("nohz_full=", tick_nohz_full_setup); 309 310static int tick_nohz_cpu_down_callback(struct notifier_block *nfb, 311 unsigned long action, 312 void *hcpu) 313{ 314 unsigned int cpu = (unsigned long)hcpu; 315 316 switch (action & ~CPU_TASKS_FROZEN) { 317 case CPU_DOWN_PREPARE: 318 /* 319 * If we handle the timekeeping duty for full dynticks CPUs, 320 * we can't safely shutdown that CPU. 321 */ 322 if (tick_nohz_full_running && tick_do_timer_cpu == cpu) 323 return NOTIFY_BAD; 324 break; 325 } 326 return NOTIFY_OK; 327} 328 329/* 330 * Worst case string length in chunks of CPU range seems 2 steps 331 * separations: 0,2,4,6,... 332 * This is NR_CPUS + sizeof('\0') 333 */ 334static char __initdata nohz_full_buf[NR_CPUS + 1]; 335 336static int tick_nohz_init_all(void) 337{ 338 int err = -1; 339 340#ifdef CONFIG_NO_HZ_FULL_ALL 341 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { 342 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n"); 343 return err; 344 } 345 err = 0; 346 cpumask_setall(tick_nohz_full_mask); 347 tick_nohz_full_running = true; 348#endif 349 return err; 350} 351 352void __init tick_nohz_init(void) 353{ 354 int cpu; 355 356 if (!tick_nohz_full_running) { 357 if (tick_nohz_init_all() < 0) 358 return; 359 } 360 361 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) { 362 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n"); 363 cpumask_clear(tick_nohz_full_mask); 364 tick_nohz_full_running = false; 365 return; 366 } 367 368 /* 369 * Full dynticks uses irq work to drive the tick rescheduling on safe 370 * locking contexts. But then we need irq work to raise its own 371 * interrupts to avoid circular dependency on the tick 372 */ 373 if (!arch_irq_work_has_interrupt()) { 374 pr_warning("NO_HZ: Can't run full dynticks because arch doesn't " 375 "support irq work self-IPIs\n"); 376 cpumask_clear(tick_nohz_full_mask); 377 cpumask_copy(housekeeping_mask, cpu_possible_mask); 378 tick_nohz_full_running = false; 379 return; 380 } 381 382 cpu = smp_processor_id(); 383 384 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 385 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu); 386 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 387 } 388 389 cpumask_andnot(housekeeping_mask, 390 cpu_possible_mask, tick_nohz_full_mask); 391 392 for_each_cpu(cpu, tick_nohz_full_mask) 393 context_tracking_cpu_set(cpu); 394 395 cpu_notifier(tick_nohz_cpu_down_callback, 0); 396 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), tick_nohz_full_mask); 397 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf); 398} 399#endif 400 401/* 402 * NOHZ - aka dynamic tick functionality 403 */ 404#ifdef CONFIG_NO_HZ_COMMON 405/* 406 * NO HZ enabled ? 407 */ 408static int tick_nohz_enabled __read_mostly = 1; 409int tick_nohz_active __read_mostly; 410/* 411 * Enable / Disable tickless mode 412 */ 413static int __init setup_tick_nohz(char *str) 414{ 415 if (!strcmp(str, "off")) 416 tick_nohz_enabled = 0; 417 else if (!strcmp(str, "on")) 418 tick_nohz_enabled = 1; 419 else 420 return 0; 421 return 1; 422} 423 424__setup("nohz=", setup_tick_nohz); 425 426/** 427 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 428 * 429 * Called from interrupt entry when the CPU was idle 430 * 431 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 432 * must be updated. Otherwise an interrupt handler could use a stale jiffy 433 * value. We do this unconditionally on any cpu, as we don't know whether the 434 * cpu, which has the update task assigned is in a long sleep. 435 */ 436static void tick_nohz_update_jiffies(ktime_t now) 437{ 438 unsigned long flags; 439 440 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 441 442 local_irq_save(flags); 443 tick_do_update_jiffies64(now); 444 local_irq_restore(flags); 445 446 touch_softlockup_watchdog(); 447} 448 449/* 450 * Updates the per cpu time idle statistics counters 451 */ 452static void 453update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 454{ 455 ktime_t delta; 456 457 if (ts->idle_active) { 458 delta = ktime_sub(now, ts->idle_entrytime); 459 if (nr_iowait_cpu(cpu) > 0) 460 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 461 else 462 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 463 ts->idle_entrytime = now; 464 } 465 466 if (last_update_time) 467 *last_update_time = ktime_to_us(now); 468 469} 470 471static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 472{ 473 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 474 ts->idle_active = 0; 475 476 sched_clock_idle_wakeup_event(0); 477} 478 479static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 480{ 481 ktime_t now = ktime_get(); 482 483 ts->idle_entrytime = now; 484 ts->idle_active = 1; 485 sched_clock_idle_sleep_event(); 486 return now; 487} 488 489/** 490 * get_cpu_idle_time_us - get the total idle time of a cpu 491 * @cpu: CPU number to query 492 * @last_update_time: variable to store update time in. Do not update 493 * counters if NULL. 494 * 495 * Return the cummulative idle time (since boot) for a given 496 * CPU, in microseconds. 497 * 498 * This time is measured via accounting rather than sampling, 499 * and is as accurate as ktime_get() is. 500 * 501 * This function returns -1 if NOHZ is not enabled. 502 */ 503u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 504{ 505 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 506 ktime_t now, idle; 507 508 if (!tick_nohz_active) 509 return -1; 510 511 now = ktime_get(); 512 if (last_update_time) { 513 update_ts_time_stats(cpu, ts, now, last_update_time); 514 idle = ts->idle_sleeptime; 515 } else { 516 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 517 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 518 519 idle = ktime_add(ts->idle_sleeptime, delta); 520 } else { 521 idle = ts->idle_sleeptime; 522 } 523 } 524 525 return ktime_to_us(idle); 526 527} 528EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 529 530/** 531 * get_cpu_iowait_time_us - get the total iowait time of a cpu 532 * @cpu: CPU number to query 533 * @last_update_time: variable to store update time in. Do not update 534 * counters if NULL. 535 * 536 * Return the cummulative iowait time (since boot) for a given 537 * CPU, in microseconds. 538 * 539 * This time is measured via accounting rather than sampling, 540 * and is as accurate as ktime_get() is. 541 * 542 * This function returns -1 if NOHZ is not enabled. 543 */ 544u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 545{ 546 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 547 ktime_t now, iowait; 548 549 if (!tick_nohz_active) 550 return -1; 551 552 now = ktime_get(); 553 if (last_update_time) { 554 update_ts_time_stats(cpu, ts, now, last_update_time); 555 iowait = ts->iowait_sleeptime; 556 } else { 557 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 558 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 559 560 iowait = ktime_add(ts->iowait_sleeptime, delta); 561 } else { 562 iowait = ts->iowait_sleeptime; 563 } 564 } 565 566 return ktime_to_us(iowait); 567} 568EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 569 570static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 571 ktime_t now, int cpu) 572{ 573 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; 574 ktime_t last_update, expires, ret = { .tv64 = 0 }; 575 unsigned long rcu_delta_jiffies; 576 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 577 u64 time_delta; 578 579 time_delta = timekeeping_max_deferment(); 580 581 /* Read jiffies and the time when jiffies were updated last */ 582 do { 583 seq = read_seqbegin(&jiffies_lock); 584 last_update = last_jiffies_update; 585 last_jiffies = jiffies; 586 } while (read_seqretry(&jiffies_lock, seq)); 587 588 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || 589 arch_needs_cpu() || irq_work_needs_cpu()) { 590 next_jiffies = last_jiffies + 1; 591 delta_jiffies = 1; 592 } else { 593 /* Get the next timer wheel timer */ 594 next_jiffies = get_next_timer_interrupt(last_jiffies); 595 delta_jiffies = next_jiffies - last_jiffies; 596 if (rcu_delta_jiffies < delta_jiffies) { 597 next_jiffies = last_jiffies + rcu_delta_jiffies; 598 delta_jiffies = rcu_delta_jiffies; 599 } 600 } 601 602 /* 603 * Do not stop the tick, if we are only one off (or less) 604 * or if the cpu is required for RCU: 605 */ 606 if (!ts->tick_stopped && delta_jiffies <= 1) 607 goto out; 608 609 /* Schedule the tick, if we are at least one jiffie off */ 610 if ((long)delta_jiffies >= 1) { 611 612 /* 613 * If this cpu is the one which updates jiffies, then 614 * give up the assignment and let it be taken by the 615 * cpu which runs the tick timer next, which might be 616 * this cpu as well. If we don't drop this here the 617 * jiffies might be stale and do_timer() never 618 * invoked. Keep track of the fact that it was the one 619 * which had the do_timer() duty last. If this cpu is 620 * the one which had the do_timer() duty last, we 621 * limit the sleep time to the timekeeping 622 * max_deferement value which we retrieved 623 * above. Otherwise we can sleep as long as we want. 624 */ 625 if (cpu == tick_do_timer_cpu) { 626 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 627 ts->do_timer_last = 1; 628 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 629 time_delta = KTIME_MAX; 630 ts->do_timer_last = 0; 631 } else if (!ts->do_timer_last) { 632 time_delta = KTIME_MAX; 633 } 634 635#ifdef CONFIG_NO_HZ_FULL 636 if (!ts->inidle) { 637 time_delta = min(time_delta, 638 scheduler_tick_max_deferment()); 639 } 640#endif 641 642 /* 643 * calculate the expiry time for the next timer wheel 644 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals 645 * that there is no timer pending or at least extremely 646 * far into the future (12 days for HZ=1000). In this 647 * case we set the expiry to the end of time. 648 */ 649 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { 650 /* 651 * Calculate the time delta for the next timer event. 652 * If the time delta exceeds the maximum time delta 653 * permitted by the current clocksource then adjust 654 * the time delta accordingly to ensure the 655 * clocksource does not wrap. 656 */ 657 time_delta = min_t(u64, time_delta, 658 tick_period.tv64 * delta_jiffies); 659 } 660 661 if (time_delta < KTIME_MAX) 662 expires = ktime_add_ns(last_update, time_delta); 663 else 664 expires.tv64 = KTIME_MAX; 665 666 /* Skip reprogram of event if its not changed */ 667 if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) 668 goto out; 669 670 ret = expires; 671 672 /* 673 * nohz_stop_sched_tick can be called several times before 674 * the nohz_restart_sched_tick is called. This happens when 675 * interrupts arrive which do not cause a reschedule. In the 676 * first call we save the current tick time, so we can restart 677 * the scheduler tick in nohz_restart_sched_tick. 678 */ 679 if (!ts->tick_stopped) { 680 nohz_balance_enter_idle(cpu); 681 calc_load_enter_idle(); 682 683 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 684 ts->tick_stopped = 1; 685 trace_tick_stop(1, " "); 686 } 687 688 /* 689 * If the expiration time == KTIME_MAX, then 690 * in this case we simply stop the tick timer. 691 */ 692 if (unlikely(expires.tv64 == KTIME_MAX)) { 693 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 694 hrtimer_cancel(&ts->sched_timer); 695 goto out; 696 } 697 698 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 699 hrtimer_start(&ts->sched_timer, expires, 700 HRTIMER_MODE_ABS_PINNED); 701 /* Check, if the timer was already in the past */ 702 if (hrtimer_active(&ts->sched_timer)) 703 goto out; 704 } else if (!tick_program_event(expires, 0)) 705 goto out; 706 /* 707 * We are past the event already. So we crossed a 708 * jiffie boundary. Update jiffies and raise the 709 * softirq. 710 */ 711 tick_do_update_jiffies64(ktime_get()); 712 } 713 raise_softirq_irqoff(TIMER_SOFTIRQ); 714out: 715 ts->next_jiffies = next_jiffies; 716 ts->last_jiffies = last_jiffies; 717 ts->sleep_length = ktime_sub(dev->next_event, now); 718 719 return ret; 720} 721 722static void tick_nohz_full_stop_tick(struct tick_sched *ts) 723{ 724#ifdef CONFIG_NO_HZ_FULL 725 int cpu = smp_processor_id(); 726 727 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current)) 728 return; 729 730 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 731 return; 732 733 if (!can_stop_full_tick()) 734 return; 735 736 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 737#endif 738} 739 740static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 741{ 742 /* 743 * If this cpu is offline and it is the one which updates 744 * jiffies, then give up the assignment and let it be taken by 745 * the cpu which runs the tick timer next. If we don't drop 746 * this here the jiffies might be stale and do_timer() never 747 * invoked. 748 */ 749 if (unlikely(!cpu_online(cpu))) { 750 if (cpu == tick_do_timer_cpu) 751 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 752 return false; 753 } 754 755 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 756 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ }; 757 return false; 758 } 759 760 if (need_resched()) 761 return false; 762 763 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 764 static int ratelimit; 765 766 if (ratelimit < 10 && 767 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 768 pr_warn("NOHZ: local_softirq_pending %02x\n", 769 (unsigned int) local_softirq_pending()); 770 ratelimit++; 771 } 772 return false; 773 } 774 775 if (tick_nohz_full_enabled()) { 776 /* 777 * Keep the tick alive to guarantee timekeeping progression 778 * if there are full dynticks CPUs around 779 */ 780 if (tick_do_timer_cpu == cpu) 781 return false; 782 /* 783 * Boot safety: make sure the timekeeping duty has been 784 * assigned before entering dyntick-idle mode, 785 */ 786 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 787 return false; 788 } 789 790 return true; 791} 792 793static void __tick_nohz_idle_enter(struct tick_sched *ts) 794{ 795 ktime_t now, expires; 796 int cpu = smp_processor_id(); 797 798 now = tick_nohz_start_idle(ts); 799 800 if (can_stop_idle_tick(cpu, ts)) { 801 int was_stopped = ts->tick_stopped; 802 803 ts->idle_calls++; 804 805 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 806 if (expires.tv64 > 0LL) { 807 ts->idle_sleeps++; 808 ts->idle_expires = expires; 809 } 810 811 if (!was_stopped && ts->tick_stopped) 812 ts->idle_jiffies = ts->last_jiffies; 813 } 814} 815 816/** 817 * tick_nohz_idle_enter - stop the idle tick from the idle task 818 * 819 * When the next event is more than a tick into the future, stop the idle tick 820 * Called when we start the idle loop. 821 * 822 * The arch is responsible of calling: 823 * 824 * - rcu_idle_enter() after its last use of RCU before the CPU is put 825 * to sleep. 826 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 827 */ 828void tick_nohz_idle_enter(void) 829{ 830 struct tick_sched *ts; 831 832 WARN_ON_ONCE(irqs_disabled()); 833 834 /* 835 * Update the idle state in the scheduler domain hierarchy 836 * when tick_nohz_stop_sched_tick() is called from the idle loop. 837 * State will be updated to busy during the first busy tick after 838 * exiting idle. 839 */ 840 set_cpu_sd_state_idle(); 841 842 local_irq_disable(); 843 844 ts = this_cpu_ptr(&tick_cpu_sched); 845 ts->inidle = 1; 846 __tick_nohz_idle_enter(ts); 847 848 local_irq_enable(); 849} 850EXPORT_SYMBOL_GPL(tick_nohz_idle_enter); 851 852/** 853 * tick_nohz_irq_exit - update next tick event from interrupt exit 854 * 855 * When an interrupt fires while we are idle and it doesn't cause 856 * a reschedule, it may still add, modify or delete a timer, enqueue 857 * an RCU callback, etc... 858 * So we need to re-calculate and reprogram the next tick event. 859 */ 860void tick_nohz_irq_exit(void) 861{ 862 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 863 864 if (ts->inidle) 865 __tick_nohz_idle_enter(ts); 866 else 867 tick_nohz_full_stop_tick(ts); 868} 869 870/** 871 * tick_nohz_get_sleep_length - return the length of the current sleep 872 * 873 * Called from power state control code with interrupts disabled 874 */ 875ktime_t tick_nohz_get_sleep_length(void) 876{ 877 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 878 879 return ts->sleep_length; 880} 881 882static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 883{ 884 hrtimer_cancel(&ts->sched_timer); 885 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 886 887 while (1) { 888 /* Forward the time to expire in the future */ 889 hrtimer_forward(&ts->sched_timer, now, tick_period); 890 891 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 892 hrtimer_start_expires(&ts->sched_timer, 893 HRTIMER_MODE_ABS_PINNED); 894 /* Check, if the timer was already in the past */ 895 if (hrtimer_active(&ts->sched_timer)) 896 break; 897 } else { 898 if (!tick_program_event( 899 hrtimer_get_expires(&ts->sched_timer), 0)) 900 break; 901 } 902 /* Reread time and update jiffies */ 903 now = ktime_get(); 904 tick_do_update_jiffies64(now); 905 } 906} 907 908static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 909{ 910 /* Update jiffies first */ 911 tick_do_update_jiffies64(now); 912 update_cpu_load_nohz(); 913 914 calc_load_exit_idle(); 915 touch_softlockup_watchdog(); 916 /* 917 * Cancel the scheduled timer and restore the tick 918 */ 919 ts->tick_stopped = 0; 920 ts->idle_exittime = now; 921 922 tick_nohz_restart(ts, now); 923} 924 925static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 926{ 927#ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 928 unsigned long ticks; 929 930 if (vtime_accounting_enabled()) 931 return; 932 /* 933 * We stopped the tick in idle. Update process times would miss the 934 * time we slept as update_process_times does only a 1 tick 935 * accounting. Enforce that this is accounted to idle ! 936 */ 937 ticks = jiffies - ts->idle_jiffies; 938 /* 939 * We might be one off. Do not randomly account a huge number of ticks! 940 */ 941 if (ticks && ticks < LONG_MAX) 942 account_idle_ticks(ticks); 943#endif 944} 945 946/** 947 * tick_nohz_idle_exit - restart the idle tick from the idle task 948 * 949 * Restart the idle tick when the CPU is woken up from idle 950 * This also exit the RCU extended quiescent state. The CPU 951 * can use RCU again after this function is called. 952 */ 953void tick_nohz_idle_exit(void) 954{ 955 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 956 ktime_t now; 957 958 local_irq_disable(); 959 960 WARN_ON_ONCE(!ts->inidle); 961 962 ts->inidle = 0; 963 964 if (ts->idle_active || ts->tick_stopped) 965 now = ktime_get(); 966 967 if (ts->idle_active) 968 tick_nohz_stop_idle(ts, now); 969 970 if (ts->tick_stopped) { 971 tick_nohz_restart_sched_tick(ts, now); 972 tick_nohz_account_idle_ticks(ts); 973 } 974 975 local_irq_enable(); 976} 977EXPORT_SYMBOL_GPL(tick_nohz_idle_exit); 978 979static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 980{ 981 hrtimer_forward(&ts->sched_timer, now, tick_period); 982 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); 983} 984 985/* 986 * The nohz low res interrupt handler 987 */ 988static void tick_nohz_handler(struct clock_event_device *dev) 989{ 990 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 991 struct pt_regs *regs = get_irq_regs(); 992 ktime_t now = ktime_get(); 993 994 dev->next_event.tv64 = KTIME_MAX; 995 996 tick_sched_do_timer(now); 997 tick_sched_handle(ts, regs); 998 999 /* No need to reprogram if we are running tickless */ 1000 if (unlikely(ts->tick_stopped)) 1001 return; 1002 1003 while (tick_nohz_reprogram(ts, now)) { 1004 now = ktime_get(); 1005 tick_do_update_jiffies64(now); 1006 } 1007} 1008 1009/** 1010 * tick_nohz_switch_to_nohz - switch to nohz mode 1011 */ 1012static void tick_nohz_switch_to_nohz(void) 1013{ 1014 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1015 ktime_t next; 1016 1017 if (!tick_nohz_enabled) 1018 return; 1019 1020 local_irq_disable(); 1021 if (tick_switch_to_oneshot(tick_nohz_handler)) { 1022 local_irq_enable(); 1023 return; 1024 } 1025 tick_nohz_active = 1; 1026 ts->nohz_mode = NOHZ_MODE_LOWRES; 1027 1028 /* 1029 * Recycle the hrtimer in ts, so we can share the 1030 * hrtimer_forward with the highres code. 1031 */ 1032 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1033 /* Get the next period */ 1034 next = tick_init_jiffy_update(); 1035 1036 for (;;) { 1037 hrtimer_set_expires(&ts->sched_timer, next); 1038 if (!tick_program_event(next, 0)) 1039 break; 1040 next = ktime_add(next, tick_period); 1041 } 1042 local_irq_enable(); 1043} 1044 1045/* 1046 * When NOHZ is enabled and the tick is stopped, we need to kick the 1047 * tick timer from irq_enter() so that the jiffies update is kept 1048 * alive during long running softirqs. That's ugly as hell, but 1049 * correctness is key even if we need to fix the offending softirq in 1050 * the first place. 1051 * 1052 * Note, this is different to tick_nohz_restart. We just kick the 1053 * timer and do not touch the other magic bits which need to be done 1054 * when idle is left. 1055 */ 1056static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now) 1057{ 1058#if 0 1059 /* Switch back to 2.6.27 behaviour */ 1060 ktime_t delta; 1061 1062 /* 1063 * Do not touch the tick device, when the next expiry is either 1064 * already reached or less/equal than the tick period. 1065 */ 1066 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 1067 if (delta.tv64 <= tick_period.tv64) 1068 return; 1069 1070 tick_nohz_restart(ts, now); 1071#endif 1072} 1073 1074static inline void tick_nohz_irq_enter(void) 1075{ 1076 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1077 ktime_t now; 1078 1079 if (!ts->idle_active && !ts->tick_stopped) 1080 return; 1081 now = ktime_get(); 1082 if (ts->idle_active) 1083 tick_nohz_stop_idle(ts, now); 1084 if (ts->tick_stopped) { 1085 tick_nohz_update_jiffies(now); 1086 tick_nohz_kick_tick(ts, now); 1087 } 1088} 1089 1090#else 1091 1092static inline void tick_nohz_switch_to_nohz(void) { } 1093static inline void tick_nohz_irq_enter(void) { } 1094 1095#endif /* CONFIG_NO_HZ_COMMON */ 1096 1097/* 1098 * Called from irq_enter to notify about the possible interruption of idle() 1099 */ 1100void tick_irq_enter(void) 1101{ 1102 tick_check_oneshot_broadcast_this_cpu(); 1103 tick_nohz_irq_enter(); 1104} 1105 1106/* 1107 * High resolution timer specific code 1108 */ 1109#ifdef CONFIG_HIGH_RES_TIMERS 1110/* 1111 * We rearm the timer until we get disabled by the idle code. 1112 * Called with interrupts disabled. 1113 */ 1114static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1115{ 1116 struct tick_sched *ts = 1117 container_of(timer, struct tick_sched, sched_timer); 1118 struct pt_regs *regs = get_irq_regs(); 1119 ktime_t now = ktime_get(); 1120 1121 tick_sched_do_timer(now); 1122 1123 /* 1124 * Do not call, when we are not in irq context and have 1125 * no valid regs pointer 1126 */ 1127 if (regs) 1128 tick_sched_handle(ts, regs); 1129 1130 /* No need to reprogram if we are in idle or full dynticks mode */ 1131 if (unlikely(ts->tick_stopped)) 1132 return HRTIMER_NORESTART; 1133 1134 hrtimer_forward(timer, now, tick_period); 1135 1136 return HRTIMER_RESTART; 1137} 1138 1139static int sched_skew_tick; 1140 1141static int __init skew_tick(char *str) 1142{ 1143 get_option(&str, &sched_skew_tick); 1144 1145 return 0; 1146} 1147early_param("skew_tick", skew_tick); 1148 1149/** 1150 * tick_setup_sched_timer - setup the tick emulation timer 1151 */ 1152void tick_setup_sched_timer(void) 1153{ 1154 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1155 ktime_t now = ktime_get(); 1156 1157 /* 1158 * Emulate tick processing via per-CPU hrtimers: 1159 */ 1160 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1161 ts->sched_timer.function = tick_sched_timer; 1162 1163 /* Get the next period (per cpu) */ 1164 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1165 1166 /* Offset the tick to avert jiffies_lock contention. */ 1167 if (sched_skew_tick) { 1168 u64 offset = ktime_to_ns(tick_period) >> 1; 1169 do_div(offset, num_possible_cpus()); 1170 offset *= smp_processor_id(); 1171 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1172 } 1173 1174 for (;;) { 1175 hrtimer_forward(&ts->sched_timer, now, tick_period); 1176 hrtimer_start_expires(&ts->sched_timer, 1177 HRTIMER_MODE_ABS_PINNED); 1178 /* Check, if the timer was already in the past */ 1179 if (hrtimer_active(&ts->sched_timer)) 1180 break; 1181 now = ktime_get(); 1182 } 1183 1184#ifdef CONFIG_NO_HZ_COMMON 1185 if (tick_nohz_enabled) { 1186 ts->nohz_mode = NOHZ_MODE_HIGHRES; 1187 tick_nohz_active = 1; 1188 } 1189#endif 1190} 1191#endif /* HIGH_RES_TIMERS */ 1192 1193#if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1194void tick_cancel_sched_timer(int cpu) 1195{ 1196 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1197 1198# ifdef CONFIG_HIGH_RES_TIMERS 1199 if (ts->sched_timer.base) 1200 hrtimer_cancel(&ts->sched_timer); 1201# endif 1202 1203 memset(ts, 0, sizeof(*ts)); 1204} 1205#endif 1206 1207/** 1208 * Async notification about clocksource changes 1209 */ 1210void tick_clock_notify(void) 1211{ 1212 int cpu; 1213 1214 for_each_possible_cpu(cpu) 1215 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1216} 1217 1218/* 1219 * Async notification about clock event changes 1220 */ 1221void tick_oneshot_notify(void) 1222{ 1223 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1224 1225 set_bit(0, &ts->check_clocks); 1226} 1227 1228/** 1229 * Check, if a change happened, which makes oneshot possible. 1230 * 1231 * Called cyclic from the hrtimer softirq (driven by the timer 1232 * softirq) allow_nohz signals, that we can switch into low-res nohz 1233 * mode, because high resolution timers are disabled (either compile 1234 * or runtime). 1235 */ 1236int tick_check_oneshot_change(int allow_nohz) 1237{ 1238 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1239 1240 if (!test_and_clear_bit(0, &ts->check_clocks)) 1241 return 0; 1242 1243 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1244 return 0; 1245 1246 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1247 return 0; 1248 1249 if (!allow_nohz) 1250 return 1; 1251 1252 tick_nohz_switch_to_nohz(); 1253 return 0; 1254} 1255