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/context_tracking.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 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); 37 38 struct tick_sched *tick_get_tick_sched(int cpu) 39 { 40 return &per_cpu(tick_cpu_sched, cpu); 41 } 42 43 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS) 44 /* 45 * The time, when the last jiffy update happened. Protected by jiffies_lock. 46 */ 47 static ktime_t last_jiffies_update; 48 49 /* 50 * Must be called with interrupts disabled ! 51 */ 52 static void tick_do_update_jiffies64(ktime_t now) 53 { 54 unsigned long ticks = 0; 55 ktime_t delta; 56 57 /* 58 * Do a quick check without holding jiffies_lock: 59 */ 60 delta = ktime_sub(now, last_jiffies_update); 61 if (delta.tv64 < tick_period.tv64) 62 return; 63 64 /* Reevalute with jiffies_lock held */ 65 write_seqlock(&jiffies_lock); 66 67 delta = ktime_sub(now, last_jiffies_update); 68 if (delta.tv64 >= tick_period.tv64) { 69 70 delta = ktime_sub(delta, tick_period); 71 last_jiffies_update = ktime_add(last_jiffies_update, 72 tick_period); 73 74 /* Slow path for long timeouts */ 75 if (unlikely(delta.tv64 >= tick_period.tv64)) { 76 s64 incr = ktime_to_ns(tick_period); 77 78 ticks = ktime_divns(delta, incr); 79 80 last_jiffies_update = ktime_add_ns(last_jiffies_update, 81 incr * ticks); 82 } 83 do_timer(++ticks); 84 85 /* Keep the tick_next_period variable up to date */ 86 tick_next_period = ktime_add(last_jiffies_update, tick_period); 87 } else { 88 write_sequnlock(&jiffies_lock); 89 return; 90 } 91 write_sequnlock(&jiffies_lock); 92 update_wall_time(); 93 } 94 95 /* 96 * Initialize and return retrieve the jiffies update. 97 */ 98 static ktime_t tick_init_jiffy_update(void) 99 { 100 ktime_t period; 101 102 write_seqlock(&jiffies_lock); 103 /* Did we start the jiffies update yet ? */ 104 if (last_jiffies_update.tv64 == 0) 105 last_jiffies_update = tick_next_period; 106 period = last_jiffies_update; 107 write_sequnlock(&jiffies_lock); 108 return period; 109 } 110 111 112 static void tick_sched_do_timer(ktime_t now) 113 { 114 int cpu = smp_processor_id(); 115 116 #ifdef CONFIG_NO_HZ_COMMON 117 /* 118 * Check if the do_timer duty was dropped. We don't care about 119 * concurrency: This happens only when the cpu in charge went 120 * into a long sleep. If two cpus happen to assign themself to 121 * this duty, then the jiffies update is still serialized by 122 * jiffies_lock. 123 */ 124 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE) 125 && !tick_nohz_full_cpu(cpu)) 126 tick_do_timer_cpu = cpu; 127 #endif 128 129 /* Check, if the jiffies need an update */ 130 if (tick_do_timer_cpu == cpu) 131 tick_do_update_jiffies64(now); 132 } 133 134 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) 135 { 136 #ifdef CONFIG_NO_HZ_COMMON 137 /* 138 * When we are idle and the tick is stopped, we have to touch 139 * the watchdog as we might not schedule for a really long 140 * time. This happens on complete idle SMP systems while 141 * waiting on the login prompt. We also increment the "start of 142 * idle" jiffy stamp so the idle accounting adjustment we do 143 * when we go busy again does not account too much ticks. 144 */ 145 if (ts->tick_stopped) { 146 touch_softlockup_watchdog_sched(); 147 if (is_idle_task(current)) 148 ts->idle_jiffies++; 149 } 150 #endif 151 update_process_times(user_mode(regs)); 152 profile_tick(CPU_PROFILING); 153 } 154 #endif 155 156 #ifdef CONFIG_NO_HZ_FULL 157 cpumask_var_t tick_nohz_full_mask; 158 cpumask_var_t housekeeping_mask; 159 bool tick_nohz_full_running; 160 static unsigned long tick_dep_mask; 161 162 static void trace_tick_dependency(unsigned long dep) 163 { 164 if (dep & TICK_DEP_MASK_POSIX_TIMER) { 165 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER); 166 return; 167 } 168 169 if (dep & TICK_DEP_MASK_PERF_EVENTS) { 170 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS); 171 return; 172 } 173 174 if (dep & TICK_DEP_MASK_SCHED) { 175 trace_tick_stop(0, TICK_DEP_MASK_SCHED); 176 return; 177 } 178 179 if (dep & TICK_DEP_MASK_CLOCK_UNSTABLE) 180 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE); 181 } 182 183 static bool can_stop_full_tick(struct tick_sched *ts) 184 { 185 WARN_ON_ONCE(!irqs_disabled()); 186 187 if (tick_dep_mask) { 188 trace_tick_dependency(tick_dep_mask); 189 return false; 190 } 191 192 if (ts->tick_dep_mask) { 193 trace_tick_dependency(ts->tick_dep_mask); 194 return false; 195 } 196 197 if (current->tick_dep_mask) { 198 trace_tick_dependency(current->tick_dep_mask); 199 return false; 200 } 201 202 if (current->signal->tick_dep_mask) { 203 trace_tick_dependency(current->signal->tick_dep_mask); 204 return false; 205 } 206 207 return true; 208 } 209 210 static void nohz_full_kick_func(struct irq_work *work) 211 { 212 /* Empty, the tick restart happens on tick_nohz_irq_exit() */ 213 } 214 215 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 216 .func = nohz_full_kick_func, 217 }; 218 219 /* 220 * Kick this CPU if it's full dynticks in order to force it to 221 * re-evaluate its dependency on the tick and restart it if necessary. 222 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), 223 * is NMI safe. 224 */ 225 static void tick_nohz_full_kick(void) 226 { 227 if (!tick_nohz_full_cpu(smp_processor_id())) 228 return; 229 230 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work)); 231 } 232 233 /* 234 * Kick the CPU if it's full dynticks in order to force it to 235 * re-evaluate its dependency on the tick and restart it if necessary. 236 */ 237 void tick_nohz_full_kick_cpu(int cpu) 238 { 239 if (!tick_nohz_full_cpu(cpu)) 240 return; 241 242 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); 243 } 244 245 /* 246 * Kick all full dynticks CPUs in order to force these to re-evaluate 247 * their dependency on the tick and restart it if necessary. 248 */ 249 static void tick_nohz_full_kick_all(void) 250 { 251 int cpu; 252 253 if (!tick_nohz_full_running) 254 return; 255 256 preempt_disable(); 257 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask) 258 tick_nohz_full_kick_cpu(cpu); 259 preempt_enable(); 260 } 261 262 static void tick_nohz_dep_set_all(unsigned long *dep, 263 enum tick_dep_bits bit) 264 { 265 unsigned long prev; 266 267 prev = fetch_or(dep, BIT_MASK(bit)); 268 if (!prev) 269 tick_nohz_full_kick_all(); 270 } 271 272 /* 273 * Set a global tick dependency. Used by perf events that rely on freq and 274 * by unstable clock. 275 */ 276 void tick_nohz_dep_set(enum tick_dep_bits bit) 277 { 278 tick_nohz_dep_set_all(&tick_dep_mask, bit); 279 } 280 281 void tick_nohz_dep_clear(enum tick_dep_bits bit) 282 { 283 clear_bit(bit, &tick_dep_mask); 284 } 285 286 /* 287 * Set per-CPU tick dependency. Used by scheduler and perf events in order to 288 * manage events throttling. 289 */ 290 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit) 291 { 292 unsigned long prev; 293 struct tick_sched *ts; 294 295 ts = per_cpu_ptr(&tick_cpu_sched, cpu); 296 297 prev = fetch_or(&ts->tick_dep_mask, BIT_MASK(bit)); 298 if (!prev) { 299 preempt_disable(); 300 /* Perf needs local kick that is NMI safe */ 301 if (cpu == smp_processor_id()) { 302 tick_nohz_full_kick(); 303 } else { 304 /* Remote irq work not NMI-safe */ 305 if (!WARN_ON_ONCE(in_nmi())) 306 tick_nohz_full_kick_cpu(cpu); 307 } 308 preempt_enable(); 309 } 310 } 311 312 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit) 313 { 314 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); 315 316 clear_bit(bit, &ts->tick_dep_mask); 317 } 318 319 /* 320 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse 321 * per task timers. 322 */ 323 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit) 324 { 325 /* 326 * We could optimize this with just kicking the target running the task 327 * if that noise matters for nohz full users. 328 */ 329 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit); 330 } 331 332 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit) 333 { 334 clear_bit(bit, &tsk->tick_dep_mask); 335 } 336 337 /* 338 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse 339 * per process timers. 340 */ 341 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit) 342 { 343 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit); 344 } 345 346 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit) 347 { 348 clear_bit(bit, &sig->tick_dep_mask); 349 } 350 351 /* 352 * Re-evaluate the need for the tick as we switch the current task. 353 * It might need the tick due to per task/process properties: 354 * perf events, posix cpu timers, ... 355 */ 356 void __tick_nohz_task_switch(void) 357 { 358 unsigned long flags; 359 struct tick_sched *ts; 360 361 local_irq_save(flags); 362 363 if (!tick_nohz_full_cpu(smp_processor_id())) 364 goto out; 365 366 ts = this_cpu_ptr(&tick_cpu_sched); 367 368 if (ts->tick_stopped) { 369 if (current->tick_dep_mask || current->signal->tick_dep_mask) 370 tick_nohz_full_kick(); 371 } 372 out: 373 local_irq_restore(flags); 374 } 375 376 /* Parse the boot-time nohz CPU list from the kernel parameters. */ 377 static int __init tick_nohz_full_setup(char *str) 378 { 379 alloc_bootmem_cpumask_var(&tick_nohz_full_mask); 380 if (cpulist_parse(str, tick_nohz_full_mask) < 0) { 381 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n"); 382 free_bootmem_cpumask_var(tick_nohz_full_mask); 383 return 1; 384 } 385 tick_nohz_full_running = true; 386 387 return 1; 388 } 389 __setup("nohz_full=", tick_nohz_full_setup); 390 391 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb, 392 unsigned long action, 393 void *hcpu) 394 { 395 unsigned int cpu = (unsigned long)hcpu; 396 397 switch (action & ~CPU_TASKS_FROZEN) { 398 case CPU_DOWN_PREPARE: 399 /* 400 * The boot CPU handles housekeeping duty (unbound timers, 401 * workqueues, timekeeping, ...) on behalf of full dynticks 402 * CPUs. It must remain online when nohz full is enabled. 403 */ 404 if (tick_nohz_full_running && tick_do_timer_cpu == cpu) 405 return NOTIFY_BAD; 406 break; 407 } 408 return NOTIFY_OK; 409 } 410 411 static int tick_nohz_init_all(void) 412 { 413 int err = -1; 414 415 #ifdef CONFIG_NO_HZ_FULL_ALL 416 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { 417 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n"); 418 return err; 419 } 420 err = 0; 421 cpumask_setall(tick_nohz_full_mask); 422 tick_nohz_full_running = true; 423 #endif 424 return err; 425 } 426 427 void __init tick_nohz_init(void) 428 { 429 int cpu; 430 431 if (!tick_nohz_full_running) { 432 if (tick_nohz_init_all() < 0) 433 return; 434 } 435 436 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) { 437 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n"); 438 cpumask_clear(tick_nohz_full_mask); 439 tick_nohz_full_running = false; 440 return; 441 } 442 443 /* 444 * Full dynticks uses irq work to drive the tick rescheduling on safe 445 * locking contexts. But then we need irq work to raise its own 446 * interrupts to avoid circular dependency on the tick 447 */ 448 if (!arch_irq_work_has_interrupt()) { 449 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n"); 450 cpumask_clear(tick_nohz_full_mask); 451 cpumask_copy(housekeeping_mask, cpu_possible_mask); 452 tick_nohz_full_running = false; 453 return; 454 } 455 456 cpu = smp_processor_id(); 457 458 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 459 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", 460 cpu); 461 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 462 } 463 464 cpumask_andnot(housekeeping_mask, 465 cpu_possible_mask, tick_nohz_full_mask); 466 467 for_each_cpu(cpu, tick_nohz_full_mask) 468 context_tracking_cpu_set(cpu); 469 470 cpu_notifier(tick_nohz_cpu_down_callback, 0); 471 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", 472 cpumask_pr_args(tick_nohz_full_mask)); 473 474 /* 475 * We need at least one CPU to handle housekeeping work such 476 * as timekeeping, unbound timers, workqueues, ... 477 */ 478 WARN_ON_ONCE(cpumask_empty(housekeeping_mask)); 479 } 480 #endif 481 482 /* 483 * NOHZ - aka dynamic tick functionality 484 */ 485 #ifdef CONFIG_NO_HZ_COMMON 486 /* 487 * NO HZ enabled ? 488 */ 489 bool tick_nohz_enabled __read_mostly = true; 490 unsigned long tick_nohz_active __read_mostly; 491 /* 492 * Enable / Disable tickless mode 493 */ 494 static int __init setup_tick_nohz(char *str) 495 { 496 return (kstrtobool(str, &tick_nohz_enabled) == 0); 497 } 498 499 __setup("nohz=", setup_tick_nohz); 500 501 int tick_nohz_tick_stopped(void) 502 { 503 return __this_cpu_read(tick_cpu_sched.tick_stopped); 504 } 505 506 /** 507 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 508 * 509 * Called from interrupt entry when the CPU was idle 510 * 511 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 512 * must be updated. Otherwise an interrupt handler could use a stale jiffy 513 * value. We do this unconditionally on any cpu, as we don't know whether the 514 * cpu, which has the update task assigned is in a long sleep. 515 */ 516 static void tick_nohz_update_jiffies(ktime_t now) 517 { 518 unsigned long flags; 519 520 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 521 522 local_irq_save(flags); 523 tick_do_update_jiffies64(now); 524 local_irq_restore(flags); 525 526 touch_softlockup_watchdog_sched(); 527 } 528 529 /* 530 * Updates the per cpu time idle statistics counters 531 */ 532 static void 533 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 534 { 535 ktime_t delta; 536 537 if (ts->idle_active) { 538 delta = ktime_sub(now, ts->idle_entrytime); 539 if (nr_iowait_cpu(cpu) > 0) 540 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 541 else 542 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 543 ts->idle_entrytime = now; 544 } 545 546 if (last_update_time) 547 *last_update_time = ktime_to_us(now); 548 549 } 550 551 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 552 { 553 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 554 ts->idle_active = 0; 555 556 sched_clock_idle_wakeup_event(0); 557 } 558 559 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 560 { 561 ktime_t now = ktime_get(); 562 563 ts->idle_entrytime = now; 564 ts->idle_active = 1; 565 sched_clock_idle_sleep_event(); 566 return now; 567 } 568 569 /** 570 * get_cpu_idle_time_us - get the total idle time of a cpu 571 * @cpu: CPU number to query 572 * @last_update_time: variable to store update time in. Do not update 573 * counters if NULL. 574 * 575 * Return the cummulative idle time (since boot) for a given 576 * CPU, in microseconds. 577 * 578 * This time is measured via accounting rather than sampling, 579 * and is as accurate as ktime_get() is. 580 * 581 * This function returns -1 if NOHZ is not enabled. 582 */ 583 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 584 { 585 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 586 ktime_t now, idle; 587 588 if (!tick_nohz_active) 589 return -1; 590 591 now = ktime_get(); 592 if (last_update_time) { 593 update_ts_time_stats(cpu, ts, now, last_update_time); 594 idle = ts->idle_sleeptime; 595 } else { 596 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 597 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 598 599 idle = ktime_add(ts->idle_sleeptime, delta); 600 } else { 601 idle = ts->idle_sleeptime; 602 } 603 } 604 605 return ktime_to_us(idle); 606 607 } 608 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 609 610 /** 611 * get_cpu_iowait_time_us - get the total iowait time of a cpu 612 * @cpu: CPU number to query 613 * @last_update_time: variable to store update time in. Do not update 614 * counters if NULL. 615 * 616 * Return the cummulative iowait time (since boot) for a given 617 * CPU, in microseconds. 618 * 619 * This time is measured via accounting rather than sampling, 620 * and is as accurate as ktime_get() is. 621 * 622 * This function returns -1 if NOHZ is not enabled. 623 */ 624 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 625 { 626 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 627 ktime_t now, iowait; 628 629 if (!tick_nohz_active) 630 return -1; 631 632 now = ktime_get(); 633 if (last_update_time) { 634 update_ts_time_stats(cpu, ts, now, last_update_time); 635 iowait = ts->iowait_sleeptime; 636 } else { 637 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 638 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 639 640 iowait = ktime_add(ts->iowait_sleeptime, delta); 641 } else { 642 iowait = ts->iowait_sleeptime; 643 } 644 } 645 646 return ktime_to_us(iowait); 647 } 648 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 649 650 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 651 { 652 hrtimer_cancel(&ts->sched_timer); 653 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 654 655 /* Forward the time to expire in the future */ 656 hrtimer_forward(&ts->sched_timer, now, tick_period); 657 658 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 659 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 660 else 661 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 662 } 663 664 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 665 ktime_t now, int cpu) 666 { 667 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 668 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; 669 unsigned long seq, basejiff; 670 ktime_t tick; 671 672 /* Read jiffies and the time when jiffies were updated last */ 673 do { 674 seq = read_seqbegin(&jiffies_lock); 675 basemono = last_jiffies_update.tv64; 676 basejiff = jiffies; 677 } while (read_seqretry(&jiffies_lock, seq)); 678 ts->last_jiffies = basejiff; 679 680 if (rcu_needs_cpu(basemono, &next_rcu) || 681 arch_needs_cpu() || irq_work_needs_cpu()) { 682 next_tick = basemono + TICK_NSEC; 683 } else { 684 /* 685 * Get the next pending timer. If high resolution 686 * timers are enabled this only takes the timer wheel 687 * timers into account. If high resolution timers are 688 * disabled this also looks at the next expiring 689 * hrtimer. 690 */ 691 next_tmr = get_next_timer_interrupt(basejiff, basemono); 692 ts->next_timer = next_tmr; 693 /* Take the next rcu event into account */ 694 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; 695 } 696 697 /* 698 * If the tick is due in the next period, keep it ticking or 699 * force prod the timer. 700 */ 701 delta = next_tick - basemono; 702 if (delta <= (u64)TICK_NSEC) { 703 tick.tv64 = 0; 704 /* 705 * We've not stopped the tick yet, and there's a timer in the 706 * next period, so no point in stopping it either, bail. 707 */ 708 if (!ts->tick_stopped) 709 goto out; 710 711 /* 712 * If, OTOH, we did stop it, but there's a pending (expired) 713 * timer reprogram the timer hardware to fire now. 714 * 715 * We will not restart the tick proper, just prod the timer 716 * hardware into firing an interrupt to process the pending 717 * timers. Just like tick_irq_exit() will not restart the tick 718 * for 'normal' interrupts. 719 * 720 * Only once we exit the idle loop will we re-enable the tick, 721 * see tick_nohz_idle_exit(). 722 */ 723 if (delta == 0) { 724 tick_nohz_restart(ts, now); 725 goto out; 726 } 727 } 728 729 /* 730 * If this cpu is the one which updates jiffies, then give up 731 * the assignment and let it be taken by the cpu which runs 732 * the tick timer next, which might be this cpu as well. If we 733 * don't drop this here the jiffies might be stale and 734 * do_timer() never invoked. Keep track of the fact that it 735 * was the one which had the do_timer() duty last. If this cpu 736 * is the one which had the do_timer() duty last, we limit the 737 * sleep time to the timekeeping max_deferement value. 738 * Otherwise we can sleep as long as we want. 739 */ 740 delta = timekeeping_max_deferment(); 741 if (cpu == tick_do_timer_cpu) { 742 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 743 ts->do_timer_last = 1; 744 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 745 delta = KTIME_MAX; 746 ts->do_timer_last = 0; 747 } else if (!ts->do_timer_last) { 748 delta = KTIME_MAX; 749 } 750 751 #ifdef CONFIG_NO_HZ_FULL 752 /* Limit the tick delta to the maximum scheduler deferment */ 753 if (!ts->inidle) 754 delta = min(delta, scheduler_tick_max_deferment()); 755 #endif 756 757 /* Calculate the next expiry time */ 758 if (delta < (KTIME_MAX - basemono)) 759 expires = basemono + delta; 760 else 761 expires = KTIME_MAX; 762 763 expires = min_t(u64, expires, next_tick); 764 tick.tv64 = expires; 765 766 /* Skip reprogram of event if its not changed */ 767 if (ts->tick_stopped && (expires == dev->next_event.tv64)) 768 goto out; 769 770 /* 771 * nohz_stop_sched_tick can be called several times before 772 * the nohz_restart_sched_tick is called. This happens when 773 * interrupts arrive which do not cause a reschedule. In the 774 * first call we save the current tick time, so we can restart 775 * the scheduler tick in nohz_restart_sched_tick. 776 */ 777 if (!ts->tick_stopped) { 778 nohz_balance_enter_idle(cpu); 779 calc_load_enter_idle(); 780 781 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 782 ts->tick_stopped = 1; 783 trace_tick_stop(1, TICK_DEP_MASK_NONE); 784 } 785 786 /* 787 * If the expiration time == KTIME_MAX, then we simply stop 788 * the tick timer. 789 */ 790 if (unlikely(expires == KTIME_MAX)) { 791 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 792 hrtimer_cancel(&ts->sched_timer); 793 goto out; 794 } 795 796 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 797 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED); 798 else 799 tick_program_event(tick, 1); 800 out: 801 /* Update the estimated sleep length */ 802 ts->sleep_length = ktime_sub(dev->next_event, now); 803 return tick; 804 } 805 806 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now, int active) 807 { 808 /* Update jiffies first */ 809 tick_do_update_jiffies64(now); 810 update_cpu_load_nohz(active); 811 812 calc_load_exit_idle(); 813 touch_softlockup_watchdog_sched(); 814 /* 815 * Cancel the scheduled timer and restore the tick 816 */ 817 ts->tick_stopped = 0; 818 ts->idle_exittime = now; 819 820 tick_nohz_restart(ts, now); 821 } 822 823 static void tick_nohz_full_update_tick(struct tick_sched *ts) 824 { 825 #ifdef CONFIG_NO_HZ_FULL 826 int cpu = smp_processor_id(); 827 828 if (!tick_nohz_full_cpu(cpu)) 829 return; 830 831 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 832 return; 833 834 if (can_stop_full_tick(ts)) 835 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 836 else if (ts->tick_stopped) 837 tick_nohz_restart_sched_tick(ts, ktime_get(), 1); 838 #endif 839 } 840 841 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 842 { 843 /* 844 * If this cpu is offline and it is the one which updates 845 * jiffies, then give up the assignment and let it be taken by 846 * the cpu which runs the tick timer next. If we don't drop 847 * this here the jiffies might be stale and do_timer() never 848 * invoked. 849 */ 850 if (unlikely(!cpu_online(cpu))) { 851 if (cpu == tick_do_timer_cpu) 852 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 853 return false; 854 } 855 856 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 857 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ }; 858 return false; 859 } 860 861 if (need_resched()) 862 return false; 863 864 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 865 static int ratelimit; 866 867 if (ratelimit < 10 && 868 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 869 pr_warn("NOHZ: local_softirq_pending %02x\n", 870 (unsigned int) local_softirq_pending()); 871 ratelimit++; 872 } 873 return false; 874 } 875 876 if (tick_nohz_full_enabled()) { 877 /* 878 * Keep the tick alive to guarantee timekeeping progression 879 * if there are full dynticks CPUs around 880 */ 881 if (tick_do_timer_cpu == cpu) 882 return false; 883 /* 884 * Boot safety: make sure the timekeeping duty has been 885 * assigned before entering dyntick-idle mode, 886 */ 887 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 888 return false; 889 } 890 891 return true; 892 } 893 894 static void __tick_nohz_idle_enter(struct tick_sched *ts) 895 { 896 ktime_t now, expires; 897 int cpu = smp_processor_id(); 898 899 now = tick_nohz_start_idle(ts); 900 901 if (can_stop_idle_tick(cpu, ts)) { 902 int was_stopped = ts->tick_stopped; 903 904 ts->idle_calls++; 905 906 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 907 if (expires.tv64 > 0LL) { 908 ts->idle_sleeps++; 909 ts->idle_expires = expires; 910 } 911 912 if (!was_stopped && ts->tick_stopped) 913 ts->idle_jiffies = ts->last_jiffies; 914 } 915 } 916 917 /** 918 * tick_nohz_idle_enter - stop the idle tick from the idle task 919 * 920 * When the next event is more than a tick into the future, stop the idle tick 921 * Called when we start the idle loop. 922 * 923 * The arch is responsible of calling: 924 * 925 * - rcu_idle_enter() after its last use of RCU before the CPU is put 926 * to sleep. 927 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 928 */ 929 void tick_nohz_idle_enter(void) 930 { 931 struct tick_sched *ts; 932 933 WARN_ON_ONCE(irqs_disabled()); 934 935 /* 936 * Update the idle state in the scheduler domain hierarchy 937 * when tick_nohz_stop_sched_tick() is called from the idle loop. 938 * State will be updated to busy during the first busy tick after 939 * exiting idle. 940 */ 941 set_cpu_sd_state_idle(); 942 943 local_irq_disable(); 944 945 ts = this_cpu_ptr(&tick_cpu_sched); 946 ts->inidle = 1; 947 __tick_nohz_idle_enter(ts); 948 949 local_irq_enable(); 950 } 951 952 /** 953 * tick_nohz_irq_exit - update next tick event from interrupt exit 954 * 955 * When an interrupt fires while we are idle and it doesn't cause 956 * a reschedule, it may still add, modify or delete a timer, enqueue 957 * an RCU callback, etc... 958 * So we need to re-calculate and reprogram the next tick event. 959 */ 960 void tick_nohz_irq_exit(void) 961 { 962 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 963 964 if (ts->inidle) 965 __tick_nohz_idle_enter(ts); 966 else 967 tick_nohz_full_update_tick(ts); 968 } 969 970 /** 971 * tick_nohz_get_sleep_length - return the length of the current sleep 972 * 973 * Called from power state control code with interrupts disabled 974 */ 975 ktime_t tick_nohz_get_sleep_length(void) 976 { 977 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 978 979 return ts->sleep_length; 980 } 981 982 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 983 { 984 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 985 unsigned long ticks; 986 987 if (vtime_accounting_cpu_enabled()) 988 return; 989 /* 990 * We stopped the tick in idle. Update process times would miss the 991 * time we slept as update_process_times does only a 1 tick 992 * accounting. Enforce that this is accounted to idle ! 993 */ 994 ticks = jiffies - ts->idle_jiffies; 995 /* 996 * We might be one off. Do not randomly account a huge number of ticks! 997 */ 998 if (ticks && ticks < LONG_MAX) 999 account_idle_ticks(ticks); 1000 #endif 1001 } 1002 1003 /** 1004 * tick_nohz_idle_exit - restart the idle tick from the idle task 1005 * 1006 * Restart the idle tick when the CPU is woken up from idle 1007 * This also exit the RCU extended quiescent state. The CPU 1008 * can use RCU again after this function is called. 1009 */ 1010 void tick_nohz_idle_exit(void) 1011 { 1012 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1013 ktime_t now; 1014 1015 local_irq_disable(); 1016 1017 WARN_ON_ONCE(!ts->inidle); 1018 1019 ts->inidle = 0; 1020 1021 if (ts->idle_active || ts->tick_stopped) 1022 now = ktime_get(); 1023 1024 if (ts->idle_active) 1025 tick_nohz_stop_idle(ts, now); 1026 1027 if (ts->tick_stopped) { 1028 tick_nohz_restart_sched_tick(ts, now, 0); 1029 tick_nohz_account_idle_ticks(ts); 1030 } 1031 1032 local_irq_enable(); 1033 } 1034 1035 /* 1036 * The nohz low res interrupt handler 1037 */ 1038 static void tick_nohz_handler(struct clock_event_device *dev) 1039 { 1040 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1041 struct pt_regs *regs = get_irq_regs(); 1042 ktime_t now = ktime_get(); 1043 1044 dev->next_event.tv64 = KTIME_MAX; 1045 1046 tick_sched_do_timer(now); 1047 tick_sched_handle(ts, regs); 1048 1049 /* No need to reprogram if we are running tickless */ 1050 if (unlikely(ts->tick_stopped)) 1051 return; 1052 1053 hrtimer_forward(&ts->sched_timer, now, tick_period); 1054 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1055 } 1056 1057 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) 1058 { 1059 if (!tick_nohz_enabled) 1060 return; 1061 ts->nohz_mode = mode; 1062 /* One update is enough */ 1063 if (!test_and_set_bit(0, &tick_nohz_active)) 1064 timers_update_migration(true); 1065 } 1066 1067 /** 1068 * tick_nohz_switch_to_nohz - switch to nohz mode 1069 */ 1070 static void tick_nohz_switch_to_nohz(void) 1071 { 1072 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1073 ktime_t next; 1074 1075 if (!tick_nohz_enabled) 1076 return; 1077 1078 if (tick_switch_to_oneshot(tick_nohz_handler)) 1079 return; 1080 1081 /* 1082 * Recycle the hrtimer in ts, so we can share the 1083 * hrtimer_forward with the highres code. 1084 */ 1085 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1086 /* Get the next period */ 1087 next = tick_init_jiffy_update(); 1088 1089 hrtimer_set_expires(&ts->sched_timer, next); 1090 hrtimer_forward_now(&ts->sched_timer, tick_period); 1091 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1092 tick_nohz_activate(ts, NOHZ_MODE_LOWRES); 1093 } 1094 1095 /* 1096 * When NOHZ is enabled and the tick is stopped, we need to kick the 1097 * tick timer from irq_enter() so that the jiffies update is kept 1098 * alive during long running softirqs. That's ugly as hell, but 1099 * correctness is key even if we need to fix the offending softirq in 1100 * the first place. 1101 * 1102 * Note, this is different to tick_nohz_restart. We just kick the 1103 * timer and do not touch the other magic bits which need to be done 1104 * when idle is left. 1105 */ 1106 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now) 1107 { 1108 #if 0 1109 /* Switch back to 2.6.27 behaviour */ 1110 ktime_t delta; 1111 1112 /* 1113 * Do not touch the tick device, when the next expiry is either 1114 * already reached or less/equal than the tick period. 1115 */ 1116 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 1117 if (delta.tv64 <= tick_period.tv64) 1118 return; 1119 1120 tick_nohz_restart(ts, now); 1121 #endif 1122 } 1123 1124 static inline void tick_nohz_irq_enter(void) 1125 { 1126 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1127 ktime_t now; 1128 1129 if (!ts->idle_active && !ts->tick_stopped) 1130 return; 1131 now = ktime_get(); 1132 if (ts->idle_active) 1133 tick_nohz_stop_idle(ts, now); 1134 if (ts->tick_stopped) { 1135 tick_nohz_update_jiffies(now); 1136 tick_nohz_kick_tick(ts, now); 1137 } 1138 } 1139 1140 #else 1141 1142 static inline void tick_nohz_switch_to_nohz(void) { } 1143 static inline void tick_nohz_irq_enter(void) { } 1144 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } 1145 1146 #endif /* CONFIG_NO_HZ_COMMON */ 1147 1148 /* 1149 * Called from irq_enter to notify about the possible interruption of idle() 1150 */ 1151 void tick_irq_enter(void) 1152 { 1153 tick_check_oneshot_broadcast_this_cpu(); 1154 tick_nohz_irq_enter(); 1155 } 1156 1157 /* 1158 * High resolution timer specific code 1159 */ 1160 #ifdef CONFIG_HIGH_RES_TIMERS 1161 /* 1162 * We rearm the timer until we get disabled by the idle code. 1163 * Called with interrupts disabled. 1164 */ 1165 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1166 { 1167 struct tick_sched *ts = 1168 container_of(timer, struct tick_sched, sched_timer); 1169 struct pt_regs *regs = get_irq_regs(); 1170 ktime_t now = ktime_get(); 1171 1172 tick_sched_do_timer(now); 1173 1174 /* 1175 * Do not call, when we are not in irq context and have 1176 * no valid regs pointer 1177 */ 1178 if (regs) 1179 tick_sched_handle(ts, regs); 1180 1181 /* No need to reprogram if we are in idle or full dynticks mode */ 1182 if (unlikely(ts->tick_stopped)) 1183 return HRTIMER_NORESTART; 1184 1185 hrtimer_forward(timer, now, tick_period); 1186 1187 return HRTIMER_RESTART; 1188 } 1189 1190 static int sched_skew_tick; 1191 1192 static int __init skew_tick(char *str) 1193 { 1194 get_option(&str, &sched_skew_tick); 1195 1196 return 0; 1197 } 1198 early_param("skew_tick", skew_tick); 1199 1200 /** 1201 * tick_setup_sched_timer - setup the tick emulation timer 1202 */ 1203 void tick_setup_sched_timer(void) 1204 { 1205 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1206 ktime_t now = ktime_get(); 1207 1208 /* 1209 * Emulate tick processing via per-CPU hrtimers: 1210 */ 1211 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1212 ts->sched_timer.function = tick_sched_timer; 1213 1214 /* Get the next period (per cpu) */ 1215 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1216 1217 /* Offset the tick to avert jiffies_lock contention. */ 1218 if (sched_skew_tick) { 1219 u64 offset = ktime_to_ns(tick_period) >> 1; 1220 do_div(offset, num_possible_cpus()); 1221 offset *= smp_processor_id(); 1222 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1223 } 1224 1225 hrtimer_forward(&ts->sched_timer, now, tick_period); 1226 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 1227 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); 1228 } 1229 #endif /* HIGH_RES_TIMERS */ 1230 1231 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1232 void tick_cancel_sched_timer(int cpu) 1233 { 1234 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1235 1236 # ifdef CONFIG_HIGH_RES_TIMERS 1237 if (ts->sched_timer.base) 1238 hrtimer_cancel(&ts->sched_timer); 1239 # endif 1240 1241 memset(ts, 0, sizeof(*ts)); 1242 } 1243 #endif 1244 1245 /** 1246 * Async notification about clocksource changes 1247 */ 1248 void tick_clock_notify(void) 1249 { 1250 int cpu; 1251 1252 for_each_possible_cpu(cpu) 1253 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1254 } 1255 1256 /* 1257 * Async notification about clock event changes 1258 */ 1259 void tick_oneshot_notify(void) 1260 { 1261 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1262 1263 set_bit(0, &ts->check_clocks); 1264 } 1265 1266 /** 1267 * Check, if a change happened, which makes oneshot possible. 1268 * 1269 * Called cyclic from the hrtimer softirq (driven by the timer 1270 * softirq) allow_nohz signals, that we can switch into low-res nohz 1271 * mode, because high resolution timers are disabled (either compile 1272 * or runtime). Called with interrupts disabled. 1273 */ 1274 int tick_check_oneshot_change(int allow_nohz) 1275 { 1276 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1277 1278 if (!test_and_clear_bit(0, &ts->check_clocks)) 1279 return 0; 1280 1281 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1282 return 0; 1283 1284 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1285 return 0; 1286 1287 if (!allow_nohz) 1288 return 1; 1289 1290 tick_nohz_switch_to_nohz(); 1291 return 0; 1292 } 1293