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