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