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