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