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