1 /* 2 * linux/kernel/time/tick-sched.c 3 * 4 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de> 5 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar 6 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner 7 * 8 * No idle tick implementation for low and high resolution timers 9 * 10 * Started by: Thomas Gleixner and Ingo Molnar 11 * 12 * Distribute under GPLv2. 13 */ 14 #include <linux/cpu.h> 15 #include <linux/err.h> 16 #include <linux/hrtimer.h> 17 #include <linux/interrupt.h> 18 #include <linux/kernel_stat.h> 19 #include <linux/percpu.h> 20 #include <linux/profile.h> 21 #include <linux/sched.h> 22 #include <linux/module.h> 23 #include <linux/irq_work.h> 24 #include <linux/posix-timers.h> 25 #include <linux/perf_event.h> 26 27 #include <asm/irq_regs.h> 28 29 #include "tick-internal.h" 30 31 #include <trace/events/timer.h> 32 33 /* 34 * Per cpu nohz control structure 35 */ 36 DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); 37 38 /* 39 * The time, when the last jiffy update happened. Protected by jiffies_lock. 40 */ 41 static ktime_t last_jiffies_update; 42 43 struct tick_sched *tick_get_tick_sched(int cpu) 44 { 45 return &per_cpu(tick_cpu_sched, cpu); 46 } 47 48 /* 49 * Must be called with interrupts disabled ! 50 */ 51 static void tick_do_update_jiffies64(ktime_t now) 52 { 53 unsigned long ticks = 0; 54 ktime_t delta; 55 56 /* 57 * Do a quick check without holding jiffies_lock: 58 */ 59 delta = ktime_sub(now, last_jiffies_update); 60 if (delta.tv64 < tick_period.tv64) 61 return; 62 63 /* Reevalute with jiffies_lock held */ 64 write_seqlock(&jiffies_lock); 65 66 delta = ktime_sub(now, last_jiffies_update); 67 if (delta.tv64 >= tick_period.tv64) { 68 69 delta = ktime_sub(delta, tick_period); 70 last_jiffies_update = ktime_add(last_jiffies_update, 71 tick_period); 72 73 /* Slow path for long timeouts */ 74 if (unlikely(delta.tv64 >= tick_period.tv64)) { 75 s64 incr = ktime_to_ns(tick_period); 76 77 ticks = ktime_divns(delta, incr); 78 79 last_jiffies_update = ktime_add_ns(last_jiffies_update, 80 incr * ticks); 81 } 82 do_timer(++ticks); 83 84 /* Keep the tick_next_period variable up to date */ 85 tick_next_period = ktime_add(last_jiffies_update, tick_period); 86 } 87 write_sequnlock(&jiffies_lock); 88 } 89 90 /* 91 * Initialize and return retrieve the jiffies update. 92 */ 93 static ktime_t tick_init_jiffy_update(void) 94 { 95 ktime_t period; 96 97 write_seqlock(&jiffies_lock); 98 /* Did we start the jiffies update yet ? */ 99 if (last_jiffies_update.tv64 == 0) 100 last_jiffies_update = tick_next_period; 101 period = last_jiffies_update; 102 write_sequnlock(&jiffies_lock); 103 return period; 104 } 105 106 107 static void tick_sched_do_timer(ktime_t now) 108 { 109 int cpu = smp_processor_id(); 110 111 #ifdef CONFIG_NO_HZ_COMMON 112 /* 113 * Check if the do_timer duty was dropped. We don't care about 114 * concurrency: This happens only when the cpu in charge went 115 * into a long sleep. If two cpus happen to assign themself to 116 * this duty, then the jiffies update is still serialized by 117 * jiffies_lock. 118 */ 119 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE) 120 && !tick_nohz_full_cpu(cpu)) 121 tick_do_timer_cpu = cpu; 122 #endif 123 124 /* Check, if the jiffies need an update */ 125 if (tick_do_timer_cpu == cpu) 126 tick_do_update_jiffies64(now); 127 } 128 129 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) 130 { 131 #ifdef CONFIG_NO_HZ_COMMON 132 /* 133 * When we are idle and the tick is stopped, we have to touch 134 * the watchdog as we might not schedule for a really long 135 * time. This happens on complete idle SMP systems while 136 * waiting on the login prompt. We also increment the "start of 137 * idle" jiffy stamp so the idle accounting adjustment we do 138 * when we go busy again does not account too much ticks. 139 */ 140 if (ts->tick_stopped) { 141 touch_softlockup_watchdog(); 142 if (is_idle_task(current)) 143 ts->idle_jiffies++; 144 } 145 #endif 146 update_process_times(user_mode(regs)); 147 profile_tick(CPU_PROFILING); 148 } 149 150 #ifdef CONFIG_NO_HZ_FULL 151 static cpumask_var_t nohz_full_mask; 152 bool have_nohz_full_mask; 153 154 static bool can_stop_full_tick(void) 155 { 156 WARN_ON_ONCE(!irqs_disabled()); 157 158 if (!sched_can_stop_tick()) { 159 trace_tick_stop(0, "more than 1 task in runqueue\n"); 160 return false; 161 } 162 163 if (!posix_cpu_timers_can_stop_tick(current)) { 164 trace_tick_stop(0, "posix timers running\n"); 165 return false; 166 } 167 168 if (!perf_event_can_stop_tick()) { 169 trace_tick_stop(0, "perf events running\n"); 170 return false; 171 } 172 173 /* sched_clock_tick() needs us? */ 174 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK 175 /* 176 * TODO: kick full dynticks CPUs when 177 * sched_clock_stable is set. 178 */ 179 if (!sched_clock_stable) { 180 trace_tick_stop(0, "unstable sched clock\n"); 181 return false; 182 } 183 #endif 184 185 return true; 186 } 187 188 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now); 189 190 /* 191 * Re-evaluate the need for the tick on the current CPU 192 * and restart it if necessary. 193 */ 194 void tick_nohz_full_check(void) 195 { 196 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 197 198 if (tick_nohz_full_cpu(smp_processor_id())) { 199 if (ts->tick_stopped && !is_idle_task(current)) { 200 if (!can_stop_full_tick()) 201 tick_nohz_restart_sched_tick(ts, ktime_get()); 202 } 203 } 204 } 205 206 static void nohz_full_kick_work_func(struct irq_work *work) 207 { 208 tick_nohz_full_check(); 209 } 210 211 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 212 .func = nohz_full_kick_work_func, 213 }; 214 215 /* 216 * Kick the current CPU if it's full dynticks in order to force it to 217 * re-evaluate its dependency on the tick and restart it if necessary. 218 */ 219 void tick_nohz_full_kick(void) 220 { 221 if (tick_nohz_full_cpu(smp_processor_id())) 222 irq_work_queue(&__get_cpu_var(nohz_full_kick_work)); 223 } 224 225 static void nohz_full_kick_ipi(void *info) 226 { 227 tick_nohz_full_check(); 228 } 229 230 /* 231 * Kick all full dynticks CPUs in order to force these to re-evaluate 232 * their dependency on the tick and restart it if necessary. 233 */ 234 void tick_nohz_full_kick_all(void) 235 { 236 if (!have_nohz_full_mask) 237 return; 238 239 preempt_disable(); 240 smp_call_function_many(nohz_full_mask, 241 nohz_full_kick_ipi, NULL, false); 242 preempt_enable(); 243 } 244 245 /* 246 * Re-evaluate the need for the tick as we switch the current task. 247 * It might need the tick due to per task/process properties: 248 * perf events, posix cpu timers, ... 249 */ 250 void tick_nohz_task_switch(struct task_struct *tsk) 251 { 252 unsigned long flags; 253 254 if (!tick_nohz_full_cpu(smp_processor_id())) 255 return; 256 257 local_irq_save(flags); 258 259 if (tick_nohz_tick_stopped() && !can_stop_full_tick()) 260 tick_nohz_full_kick(); 261 262 local_irq_restore(flags); 263 } 264 265 int tick_nohz_full_cpu(int cpu) 266 { 267 if (!have_nohz_full_mask) 268 return 0; 269 270 return cpumask_test_cpu(cpu, nohz_full_mask); 271 } 272 273 /* Parse the boot-time nohz CPU list from the kernel parameters. */ 274 static int __init tick_nohz_full_setup(char *str) 275 { 276 int cpu; 277 278 alloc_bootmem_cpumask_var(&nohz_full_mask); 279 if (cpulist_parse(str, nohz_full_mask) < 0) { 280 pr_warning("NOHZ: Incorrect nohz_full cpumask\n"); 281 return 1; 282 } 283 284 cpu = smp_processor_id(); 285 if (cpumask_test_cpu(cpu, nohz_full_mask)) { 286 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu); 287 cpumask_clear_cpu(cpu, nohz_full_mask); 288 } 289 have_nohz_full_mask = true; 290 291 return 1; 292 } 293 __setup("nohz_full=", tick_nohz_full_setup); 294 295 static int __cpuinit tick_nohz_cpu_down_callback(struct notifier_block *nfb, 296 unsigned long action, 297 void *hcpu) 298 { 299 unsigned int cpu = (unsigned long)hcpu; 300 301 switch (action & ~CPU_TASKS_FROZEN) { 302 case CPU_DOWN_PREPARE: 303 /* 304 * If we handle the timekeeping duty for full dynticks CPUs, 305 * we can't safely shutdown that CPU. 306 */ 307 if (have_nohz_full_mask && tick_do_timer_cpu == cpu) 308 return -EINVAL; 309 break; 310 } 311 return NOTIFY_OK; 312 } 313 314 /* 315 * Worst case string length in chunks of CPU range seems 2 steps 316 * separations: 0,2,4,6,... 317 * This is NR_CPUS + sizeof('\0') 318 */ 319 static char __initdata nohz_full_buf[NR_CPUS + 1]; 320 321 static int tick_nohz_init_all(void) 322 { 323 int err = -1; 324 325 #ifdef CONFIG_NO_HZ_FULL_ALL 326 if (!alloc_cpumask_var(&nohz_full_mask, GFP_KERNEL)) { 327 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n"); 328 return err; 329 } 330 err = 0; 331 cpumask_setall(nohz_full_mask); 332 cpumask_clear_cpu(smp_processor_id(), nohz_full_mask); 333 have_nohz_full_mask = true; 334 #endif 335 return err; 336 } 337 338 void __init tick_nohz_init(void) 339 { 340 int cpu; 341 342 if (!have_nohz_full_mask) { 343 if (tick_nohz_init_all() < 0) 344 return; 345 } 346 347 cpu_notifier(tick_nohz_cpu_down_callback, 0); 348 349 /* Make sure full dynticks CPU are also RCU nocbs */ 350 for_each_cpu(cpu, nohz_full_mask) { 351 if (!rcu_is_nocb_cpu(cpu)) { 352 pr_warning("NO_HZ: CPU %d is not RCU nocb: " 353 "cleared from nohz_full range", cpu); 354 cpumask_clear_cpu(cpu, nohz_full_mask); 355 } 356 } 357 358 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), nohz_full_mask); 359 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf); 360 } 361 #else 362 #define have_nohz_full_mask (0) 363 #endif 364 365 /* 366 * NOHZ - aka dynamic tick functionality 367 */ 368 #ifdef CONFIG_NO_HZ_COMMON 369 /* 370 * NO HZ enabled ? 371 */ 372 int tick_nohz_enabled __read_mostly = 1; 373 374 /* 375 * Enable / Disable tickless mode 376 */ 377 static int __init setup_tick_nohz(char *str) 378 { 379 if (!strcmp(str, "off")) 380 tick_nohz_enabled = 0; 381 else if (!strcmp(str, "on")) 382 tick_nohz_enabled = 1; 383 else 384 return 0; 385 return 1; 386 } 387 388 __setup("nohz=", setup_tick_nohz); 389 390 /** 391 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 392 * 393 * Called from interrupt entry when the CPU was idle 394 * 395 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 396 * must be updated. Otherwise an interrupt handler could use a stale jiffy 397 * value. We do this unconditionally on any cpu, as we don't know whether the 398 * cpu, which has the update task assigned is in a long sleep. 399 */ 400 static void tick_nohz_update_jiffies(ktime_t now) 401 { 402 int cpu = smp_processor_id(); 403 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 404 unsigned long flags; 405 406 ts->idle_waketime = now; 407 408 local_irq_save(flags); 409 tick_do_update_jiffies64(now); 410 local_irq_restore(flags); 411 412 touch_softlockup_watchdog(); 413 } 414 415 /* 416 * Updates the per cpu time idle statistics counters 417 */ 418 static void 419 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 420 { 421 ktime_t delta; 422 423 if (ts->idle_active) { 424 delta = ktime_sub(now, ts->idle_entrytime); 425 if (nr_iowait_cpu(cpu) > 0) 426 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 427 else 428 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 429 ts->idle_entrytime = now; 430 } 431 432 if (last_update_time) 433 *last_update_time = ktime_to_us(now); 434 435 } 436 437 static void tick_nohz_stop_idle(int cpu, ktime_t now) 438 { 439 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 440 441 update_ts_time_stats(cpu, ts, now, NULL); 442 ts->idle_active = 0; 443 444 sched_clock_idle_wakeup_event(0); 445 } 446 447 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts) 448 { 449 ktime_t now = ktime_get(); 450 451 ts->idle_entrytime = now; 452 ts->idle_active = 1; 453 sched_clock_idle_sleep_event(); 454 return now; 455 } 456 457 /** 458 * get_cpu_idle_time_us - get the total idle time of a cpu 459 * @cpu: CPU number to query 460 * @last_update_time: variable to store update time in. Do not update 461 * counters if NULL. 462 * 463 * Return the cummulative idle time (since boot) for a given 464 * CPU, in microseconds. 465 * 466 * This time is measured via accounting rather than sampling, 467 * and is as accurate as ktime_get() is. 468 * 469 * This function returns -1 if NOHZ is not enabled. 470 */ 471 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 472 { 473 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 474 ktime_t now, idle; 475 476 if (!tick_nohz_enabled) 477 return -1; 478 479 now = ktime_get(); 480 if (last_update_time) { 481 update_ts_time_stats(cpu, ts, now, last_update_time); 482 idle = ts->idle_sleeptime; 483 } else { 484 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 485 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 486 487 idle = ktime_add(ts->idle_sleeptime, delta); 488 } else { 489 idle = ts->idle_sleeptime; 490 } 491 } 492 493 return ktime_to_us(idle); 494 495 } 496 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 497 498 /** 499 * get_cpu_iowait_time_us - get the total iowait time of a cpu 500 * @cpu: CPU number to query 501 * @last_update_time: variable to store update time in. Do not update 502 * counters if NULL. 503 * 504 * Return the cummulative iowait time (since boot) for a given 505 * CPU, in microseconds. 506 * 507 * This time is measured via accounting rather than sampling, 508 * and is as accurate as ktime_get() is. 509 * 510 * This function returns -1 if NOHZ is not enabled. 511 */ 512 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 513 { 514 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 515 ktime_t now, iowait; 516 517 if (!tick_nohz_enabled) 518 return -1; 519 520 now = ktime_get(); 521 if (last_update_time) { 522 update_ts_time_stats(cpu, ts, now, last_update_time); 523 iowait = ts->iowait_sleeptime; 524 } else { 525 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 526 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 527 528 iowait = ktime_add(ts->iowait_sleeptime, delta); 529 } else { 530 iowait = ts->iowait_sleeptime; 531 } 532 } 533 534 return ktime_to_us(iowait); 535 } 536 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 537 538 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 539 ktime_t now, int cpu) 540 { 541 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; 542 ktime_t last_update, expires, ret = { .tv64 = 0 }; 543 unsigned long rcu_delta_jiffies; 544 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; 545 u64 time_delta; 546 547 /* Read jiffies and the time when jiffies were updated last */ 548 do { 549 seq = read_seqbegin(&jiffies_lock); 550 last_update = last_jiffies_update; 551 last_jiffies = jiffies; 552 time_delta = timekeeping_max_deferment(); 553 } while (read_seqretry(&jiffies_lock, seq)); 554 555 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || 556 arch_needs_cpu(cpu) || irq_work_needs_cpu()) { 557 next_jiffies = last_jiffies + 1; 558 delta_jiffies = 1; 559 } else { 560 /* Get the next timer wheel timer */ 561 next_jiffies = get_next_timer_interrupt(last_jiffies); 562 delta_jiffies = next_jiffies - last_jiffies; 563 if (rcu_delta_jiffies < delta_jiffies) { 564 next_jiffies = last_jiffies + rcu_delta_jiffies; 565 delta_jiffies = rcu_delta_jiffies; 566 } 567 } 568 /* 569 * Do not stop the tick, if we are only one off 570 * or if the cpu is required for rcu 571 */ 572 if (!ts->tick_stopped && delta_jiffies == 1) 573 goto out; 574 575 /* Schedule the tick, if we are at least one jiffie off */ 576 if ((long)delta_jiffies >= 1) { 577 578 /* 579 * If this cpu is the one which updates jiffies, then 580 * give up the assignment and let it be taken by the 581 * cpu which runs the tick timer next, which might be 582 * this cpu as well. If we don't drop this here the 583 * jiffies might be stale and do_timer() never 584 * invoked. Keep track of the fact that it was the one 585 * which had the do_timer() duty last. If this cpu is 586 * the one which had the do_timer() duty last, we 587 * limit the sleep time to the timekeeping 588 * max_deferement value which we retrieved 589 * above. Otherwise we can sleep as long as we want. 590 */ 591 if (cpu == tick_do_timer_cpu) { 592 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 593 ts->do_timer_last = 1; 594 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 595 time_delta = KTIME_MAX; 596 ts->do_timer_last = 0; 597 } else if (!ts->do_timer_last) { 598 time_delta = KTIME_MAX; 599 } 600 601 /* 602 * calculate the expiry time for the next timer wheel 603 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals 604 * that there is no timer pending or at least extremely 605 * far into the future (12 days for HZ=1000). In this 606 * case we set the expiry to the end of time. 607 */ 608 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { 609 /* 610 * Calculate the time delta for the next timer event. 611 * If the time delta exceeds the maximum time delta 612 * permitted by the current clocksource then adjust 613 * the time delta accordingly to ensure the 614 * clocksource does not wrap. 615 */ 616 time_delta = min_t(u64, time_delta, 617 tick_period.tv64 * delta_jiffies); 618 } 619 620 if (time_delta < KTIME_MAX) 621 expires = ktime_add_ns(last_update, time_delta); 622 else 623 expires.tv64 = KTIME_MAX; 624 625 /* Skip reprogram of event if its not changed */ 626 if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) 627 goto out; 628 629 ret = expires; 630 631 /* 632 * nohz_stop_sched_tick can be called several times before 633 * the nohz_restart_sched_tick is called. This happens when 634 * interrupts arrive which do not cause a reschedule. In the 635 * first call we save the current tick time, so we can restart 636 * the scheduler tick in nohz_restart_sched_tick. 637 */ 638 if (!ts->tick_stopped) { 639 nohz_balance_enter_idle(cpu); 640 calc_load_enter_idle(); 641 642 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 643 ts->tick_stopped = 1; 644 trace_tick_stop(1, " "); 645 } 646 647 /* 648 * If the expiration time == KTIME_MAX, then 649 * in this case we simply stop the tick timer. 650 */ 651 if (unlikely(expires.tv64 == KTIME_MAX)) { 652 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 653 hrtimer_cancel(&ts->sched_timer); 654 goto out; 655 } 656 657 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 658 hrtimer_start(&ts->sched_timer, expires, 659 HRTIMER_MODE_ABS_PINNED); 660 /* Check, if the timer was already in the past */ 661 if (hrtimer_active(&ts->sched_timer)) 662 goto out; 663 } else if (!tick_program_event(expires, 0)) 664 goto out; 665 /* 666 * We are past the event already. So we crossed a 667 * jiffie boundary. Update jiffies and raise the 668 * softirq. 669 */ 670 tick_do_update_jiffies64(ktime_get()); 671 } 672 raise_softirq_irqoff(TIMER_SOFTIRQ); 673 out: 674 ts->next_jiffies = next_jiffies; 675 ts->last_jiffies = last_jiffies; 676 ts->sleep_length = ktime_sub(dev->next_event, now); 677 678 return ret; 679 } 680 681 static void tick_nohz_full_stop_tick(struct tick_sched *ts) 682 { 683 #ifdef CONFIG_NO_HZ_FULL 684 int cpu = smp_processor_id(); 685 686 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current)) 687 return; 688 689 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 690 return; 691 692 if (!can_stop_full_tick()) 693 return; 694 695 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 696 #endif 697 } 698 699 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 700 { 701 /* 702 * If this cpu is offline and it is the one which updates 703 * jiffies, then give up the assignment and let it be taken by 704 * the cpu which runs the tick timer next. If we don't drop 705 * this here the jiffies might be stale and do_timer() never 706 * invoked. 707 */ 708 if (unlikely(!cpu_online(cpu))) { 709 if (cpu == tick_do_timer_cpu) 710 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 711 } 712 713 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) 714 return false; 715 716 if (need_resched()) 717 return false; 718 719 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 720 static int ratelimit; 721 722 if (ratelimit < 10 && 723 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 724 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", 725 (unsigned int) local_softirq_pending()); 726 ratelimit++; 727 } 728 return false; 729 } 730 731 if (have_nohz_full_mask) { 732 /* 733 * Keep the tick alive to guarantee timekeeping progression 734 * if there are full dynticks CPUs around 735 */ 736 if (tick_do_timer_cpu == cpu) 737 return false; 738 /* 739 * Boot safety: make sure the timekeeping duty has been 740 * assigned before entering dyntick-idle mode, 741 */ 742 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 743 return false; 744 } 745 746 return true; 747 } 748 749 static void __tick_nohz_idle_enter(struct tick_sched *ts) 750 { 751 ktime_t now, expires; 752 int cpu = smp_processor_id(); 753 754 now = tick_nohz_start_idle(cpu, ts); 755 756 if (can_stop_idle_tick(cpu, ts)) { 757 int was_stopped = ts->tick_stopped; 758 759 ts->idle_calls++; 760 761 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 762 if (expires.tv64 > 0LL) { 763 ts->idle_sleeps++; 764 ts->idle_expires = expires; 765 } 766 767 if (!was_stopped && ts->tick_stopped) 768 ts->idle_jiffies = ts->last_jiffies; 769 } 770 } 771 772 /** 773 * tick_nohz_idle_enter - stop the idle tick from the idle task 774 * 775 * When the next event is more than a tick into the future, stop the idle tick 776 * Called when we start the idle loop. 777 * 778 * The arch is responsible of calling: 779 * 780 * - rcu_idle_enter() after its last use of RCU before the CPU is put 781 * to sleep. 782 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 783 */ 784 void tick_nohz_idle_enter(void) 785 { 786 struct tick_sched *ts; 787 788 WARN_ON_ONCE(irqs_disabled()); 789 790 /* 791 * Update the idle state in the scheduler domain hierarchy 792 * when tick_nohz_stop_sched_tick() is called from the idle loop. 793 * State will be updated to busy during the first busy tick after 794 * exiting idle. 795 */ 796 set_cpu_sd_state_idle(); 797 798 local_irq_disable(); 799 800 ts = &__get_cpu_var(tick_cpu_sched); 801 /* 802 * set ts->inidle unconditionally. even if the system did not 803 * switch to nohz mode the cpu frequency governers rely on the 804 * update of the idle time accounting in tick_nohz_start_idle(). 805 */ 806 ts->inidle = 1; 807 __tick_nohz_idle_enter(ts); 808 809 local_irq_enable(); 810 } 811 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter); 812 813 /** 814 * tick_nohz_irq_exit - update next tick event from interrupt exit 815 * 816 * When an interrupt fires while we are idle and it doesn't cause 817 * a reschedule, it may still add, modify or delete a timer, enqueue 818 * an RCU callback, etc... 819 * So we need to re-calculate and reprogram the next tick event. 820 */ 821 void tick_nohz_irq_exit(void) 822 { 823 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 824 825 if (ts->inidle) { 826 /* Cancel the timer because CPU already waken up from the C-states*/ 827 menu_hrtimer_cancel(); 828 __tick_nohz_idle_enter(ts); 829 } else { 830 tick_nohz_full_stop_tick(ts); 831 } 832 } 833 834 /** 835 * tick_nohz_get_sleep_length - return the length of the current sleep 836 * 837 * Called from power state control code with interrupts disabled 838 */ 839 ktime_t tick_nohz_get_sleep_length(void) 840 { 841 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 842 843 return ts->sleep_length; 844 } 845 846 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 847 { 848 hrtimer_cancel(&ts->sched_timer); 849 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 850 851 while (1) { 852 /* Forward the time to expire in the future */ 853 hrtimer_forward(&ts->sched_timer, now, tick_period); 854 855 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 856 hrtimer_start_expires(&ts->sched_timer, 857 HRTIMER_MODE_ABS_PINNED); 858 /* Check, if the timer was already in the past */ 859 if (hrtimer_active(&ts->sched_timer)) 860 break; 861 } else { 862 if (!tick_program_event( 863 hrtimer_get_expires(&ts->sched_timer), 0)) 864 break; 865 } 866 /* Reread time and update jiffies */ 867 now = ktime_get(); 868 tick_do_update_jiffies64(now); 869 } 870 } 871 872 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 873 { 874 /* Update jiffies first */ 875 tick_do_update_jiffies64(now); 876 update_cpu_load_nohz(); 877 878 calc_load_exit_idle(); 879 touch_softlockup_watchdog(); 880 /* 881 * Cancel the scheduled timer and restore the tick 882 */ 883 ts->tick_stopped = 0; 884 ts->idle_exittime = now; 885 886 tick_nohz_restart(ts, now); 887 } 888 889 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 890 { 891 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 892 unsigned long ticks; 893 894 if (vtime_accounting_enabled()) 895 return; 896 /* 897 * We stopped the tick in idle. Update process times would miss the 898 * time we slept as update_process_times does only a 1 tick 899 * accounting. Enforce that this is accounted to idle ! 900 */ 901 ticks = jiffies - ts->idle_jiffies; 902 /* 903 * We might be one off. Do not randomly account a huge number of ticks! 904 */ 905 if (ticks && ticks < LONG_MAX) 906 account_idle_ticks(ticks); 907 #endif 908 } 909 910 /** 911 * tick_nohz_idle_exit - restart the idle tick from the idle task 912 * 913 * Restart the idle tick when the CPU is woken up from idle 914 * This also exit the RCU extended quiescent state. The CPU 915 * can use RCU again after this function is called. 916 */ 917 void tick_nohz_idle_exit(void) 918 { 919 int cpu = smp_processor_id(); 920 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 921 ktime_t now; 922 923 local_irq_disable(); 924 925 WARN_ON_ONCE(!ts->inidle); 926 927 ts->inidle = 0; 928 929 /* Cancel the timer because CPU already waken up from the C-states*/ 930 menu_hrtimer_cancel(); 931 if (ts->idle_active || ts->tick_stopped) 932 now = ktime_get(); 933 934 if (ts->idle_active) 935 tick_nohz_stop_idle(cpu, now); 936 937 if (ts->tick_stopped) { 938 tick_nohz_restart_sched_tick(ts, now); 939 tick_nohz_account_idle_ticks(ts); 940 } 941 942 local_irq_enable(); 943 } 944 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit); 945 946 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 947 { 948 hrtimer_forward(&ts->sched_timer, now, tick_period); 949 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); 950 } 951 952 /* 953 * The nohz low res interrupt handler 954 */ 955 static void tick_nohz_handler(struct clock_event_device *dev) 956 { 957 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 958 struct pt_regs *regs = get_irq_regs(); 959 ktime_t now = ktime_get(); 960 961 dev->next_event.tv64 = KTIME_MAX; 962 963 tick_sched_do_timer(now); 964 tick_sched_handle(ts, regs); 965 966 while (tick_nohz_reprogram(ts, now)) { 967 now = ktime_get(); 968 tick_do_update_jiffies64(now); 969 } 970 } 971 972 /** 973 * tick_nohz_switch_to_nohz - switch to nohz mode 974 */ 975 static void tick_nohz_switch_to_nohz(void) 976 { 977 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 978 ktime_t next; 979 980 if (!tick_nohz_enabled) 981 return; 982 983 local_irq_disable(); 984 if (tick_switch_to_oneshot(tick_nohz_handler)) { 985 local_irq_enable(); 986 return; 987 } 988 989 ts->nohz_mode = NOHZ_MODE_LOWRES; 990 991 /* 992 * Recycle the hrtimer in ts, so we can share the 993 * hrtimer_forward with the highres code. 994 */ 995 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 996 /* Get the next period */ 997 next = tick_init_jiffy_update(); 998 999 for (;;) { 1000 hrtimer_set_expires(&ts->sched_timer, next); 1001 if (!tick_program_event(next, 0)) 1002 break; 1003 next = ktime_add(next, tick_period); 1004 } 1005 local_irq_enable(); 1006 } 1007 1008 /* 1009 * When NOHZ is enabled and the tick is stopped, we need to kick the 1010 * tick timer from irq_enter() so that the jiffies update is kept 1011 * alive during long running softirqs. That's ugly as hell, but 1012 * correctness is key even if we need to fix the offending softirq in 1013 * the first place. 1014 * 1015 * Note, this is different to tick_nohz_restart. We just kick the 1016 * timer and do not touch the other magic bits which need to be done 1017 * when idle is left. 1018 */ 1019 static void tick_nohz_kick_tick(int cpu, ktime_t now) 1020 { 1021 #if 0 1022 /* Switch back to 2.6.27 behaviour */ 1023 1024 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1025 ktime_t delta; 1026 1027 /* 1028 * Do not touch the tick device, when the next expiry is either 1029 * already reached or less/equal than the tick period. 1030 */ 1031 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 1032 if (delta.tv64 <= tick_period.tv64) 1033 return; 1034 1035 tick_nohz_restart(ts, now); 1036 #endif 1037 } 1038 1039 static inline void tick_check_nohz(int cpu) 1040 { 1041 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1042 ktime_t now; 1043 1044 if (!ts->idle_active && !ts->tick_stopped) 1045 return; 1046 now = ktime_get(); 1047 if (ts->idle_active) 1048 tick_nohz_stop_idle(cpu, now); 1049 if (ts->tick_stopped) { 1050 tick_nohz_update_jiffies(now); 1051 tick_nohz_kick_tick(cpu, now); 1052 } 1053 } 1054 1055 #else 1056 1057 static inline void tick_nohz_switch_to_nohz(void) { } 1058 static inline void tick_check_nohz(int cpu) { } 1059 1060 #endif /* CONFIG_NO_HZ_COMMON */ 1061 1062 /* 1063 * Called from irq_enter to notify about the possible interruption of idle() 1064 */ 1065 void tick_check_idle(int cpu) 1066 { 1067 tick_check_oneshot_broadcast(cpu); 1068 tick_check_nohz(cpu); 1069 } 1070 1071 /* 1072 * High resolution timer specific code 1073 */ 1074 #ifdef CONFIG_HIGH_RES_TIMERS 1075 /* 1076 * We rearm the timer until we get disabled by the idle code. 1077 * Called with interrupts disabled. 1078 */ 1079 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1080 { 1081 struct tick_sched *ts = 1082 container_of(timer, struct tick_sched, sched_timer); 1083 struct pt_regs *regs = get_irq_regs(); 1084 ktime_t now = ktime_get(); 1085 1086 tick_sched_do_timer(now); 1087 1088 /* 1089 * Do not call, when we are not in irq context and have 1090 * no valid regs pointer 1091 */ 1092 if (regs) 1093 tick_sched_handle(ts, regs); 1094 1095 hrtimer_forward(timer, now, tick_period); 1096 1097 return HRTIMER_RESTART; 1098 } 1099 1100 static int sched_skew_tick; 1101 1102 static int __init skew_tick(char *str) 1103 { 1104 get_option(&str, &sched_skew_tick); 1105 1106 return 0; 1107 } 1108 early_param("skew_tick", skew_tick); 1109 1110 /** 1111 * tick_setup_sched_timer - setup the tick emulation timer 1112 */ 1113 void tick_setup_sched_timer(void) 1114 { 1115 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 1116 ktime_t now = ktime_get(); 1117 1118 /* 1119 * Emulate tick processing via per-CPU hrtimers: 1120 */ 1121 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1122 ts->sched_timer.function = tick_sched_timer; 1123 1124 /* Get the next period (per cpu) */ 1125 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1126 1127 /* Offset the tick to avert jiffies_lock contention. */ 1128 if (sched_skew_tick) { 1129 u64 offset = ktime_to_ns(tick_period) >> 1; 1130 do_div(offset, num_possible_cpus()); 1131 offset *= smp_processor_id(); 1132 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1133 } 1134 1135 for (;;) { 1136 hrtimer_forward(&ts->sched_timer, now, tick_period); 1137 hrtimer_start_expires(&ts->sched_timer, 1138 HRTIMER_MODE_ABS_PINNED); 1139 /* Check, if the timer was already in the past */ 1140 if (hrtimer_active(&ts->sched_timer)) 1141 break; 1142 now = ktime_get(); 1143 } 1144 1145 #ifdef CONFIG_NO_HZ_COMMON 1146 if (tick_nohz_enabled) 1147 ts->nohz_mode = NOHZ_MODE_HIGHRES; 1148 #endif 1149 } 1150 #endif /* HIGH_RES_TIMERS */ 1151 1152 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1153 void tick_cancel_sched_timer(int cpu) 1154 { 1155 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1156 1157 # ifdef CONFIG_HIGH_RES_TIMERS 1158 if (ts->sched_timer.base) 1159 hrtimer_cancel(&ts->sched_timer); 1160 # endif 1161 1162 ts->nohz_mode = NOHZ_MODE_INACTIVE; 1163 } 1164 #endif 1165 1166 /** 1167 * Async notification about clocksource changes 1168 */ 1169 void tick_clock_notify(void) 1170 { 1171 int cpu; 1172 1173 for_each_possible_cpu(cpu) 1174 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1175 } 1176 1177 /* 1178 * Async notification about clock event changes 1179 */ 1180 void tick_oneshot_notify(void) 1181 { 1182 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 1183 1184 set_bit(0, &ts->check_clocks); 1185 } 1186 1187 /** 1188 * Check, if a change happened, which makes oneshot possible. 1189 * 1190 * Called cyclic from the hrtimer softirq (driven by the timer 1191 * softirq) allow_nohz signals, that we can switch into low-res nohz 1192 * mode, because high resolution timers are disabled (either compile 1193 * or runtime). 1194 */ 1195 int tick_check_oneshot_change(int allow_nohz) 1196 { 1197 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 1198 1199 if (!test_and_clear_bit(0, &ts->check_clocks)) 1200 return 0; 1201 1202 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1203 return 0; 1204 1205 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1206 return 0; 1207 1208 if (!allow_nohz) 1209 return 1; 1210 1211 tick_nohz_switch_to_nohz(); 1212 return 0; 1213 } 1214