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