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_warning("NOHZ: 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_warning("NO_HZ: Can't run full dynticks because arch doesn't " 450 "support irq work self-IPIs\n"); 451 cpumask_clear(tick_nohz_full_mask); 452 cpumask_copy(housekeeping_mask, cpu_possible_mask); 453 tick_nohz_full_running = false; 454 return; 455 } 456 457 cpu = smp_processor_id(); 458 459 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 460 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", 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 int tick_nohz_enabled __read_mostly = 1; 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 if (!strcmp(str, "off")) 497 tick_nohz_enabled = 0; 498 else if (!strcmp(str, "on")) 499 tick_nohz_enabled = 1; 500 else 501 return 0; 502 return 1; 503 } 504 505 __setup("nohz=", setup_tick_nohz); 506 507 int tick_nohz_tick_stopped(void) 508 { 509 return __this_cpu_read(tick_cpu_sched.tick_stopped); 510 } 511 512 /** 513 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 514 * 515 * Called from interrupt entry when the CPU was idle 516 * 517 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 518 * must be updated. Otherwise an interrupt handler could use a stale jiffy 519 * value. We do this unconditionally on any cpu, as we don't know whether the 520 * cpu, which has the update task assigned is in a long sleep. 521 */ 522 static void tick_nohz_update_jiffies(ktime_t now) 523 { 524 unsigned long flags; 525 526 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 527 528 local_irq_save(flags); 529 tick_do_update_jiffies64(now); 530 local_irq_restore(flags); 531 532 touch_softlockup_watchdog_sched(); 533 } 534 535 /* 536 * Updates the per cpu time idle statistics counters 537 */ 538 static void 539 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 540 { 541 ktime_t delta; 542 543 if (ts->idle_active) { 544 delta = ktime_sub(now, ts->idle_entrytime); 545 if (nr_iowait_cpu(cpu) > 0) 546 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 547 else 548 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 549 ts->idle_entrytime = now; 550 } 551 552 if (last_update_time) 553 *last_update_time = ktime_to_us(now); 554 555 } 556 557 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 558 { 559 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 560 ts->idle_active = 0; 561 562 sched_clock_idle_wakeup_event(0); 563 } 564 565 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 566 { 567 ktime_t now = ktime_get(); 568 569 ts->idle_entrytime = now; 570 ts->idle_active = 1; 571 sched_clock_idle_sleep_event(); 572 return now; 573 } 574 575 /** 576 * get_cpu_idle_time_us - get the total idle time of a cpu 577 * @cpu: CPU number to query 578 * @last_update_time: variable to store update time in. Do not update 579 * counters if NULL. 580 * 581 * Return the cummulative idle time (since boot) for a given 582 * CPU, in microseconds. 583 * 584 * This time is measured via accounting rather than sampling, 585 * and is as accurate as ktime_get() is. 586 * 587 * This function returns -1 if NOHZ is not enabled. 588 */ 589 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 590 { 591 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 592 ktime_t now, idle; 593 594 if (!tick_nohz_active) 595 return -1; 596 597 now = ktime_get(); 598 if (last_update_time) { 599 update_ts_time_stats(cpu, ts, now, last_update_time); 600 idle = ts->idle_sleeptime; 601 } else { 602 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 603 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 604 605 idle = ktime_add(ts->idle_sleeptime, delta); 606 } else { 607 idle = ts->idle_sleeptime; 608 } 609 } 610 611 return ktime_to_us(idle); 612 613 } 614 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 615 616 /** 617 * get_cpu_iowait_time_us - get the total iowait time of a cpu 618 * @cpu: CPU number to query 619 * @last_update_time: variable to store update time in. Do not update 620 * counters if NULL. 621 * 622 * Return the cummulative iowait time (since boot) for a given 623 * CPU, in microseconds. 624 * 625 * This time is measured via accounting rather than sampling, 626 * and is as accurate as ktime_get() is. 627 * 628 * This function returns -1 if NOHZ is not enabled. 629 */ 630 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 631 { 632 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 633 ktime_t now, iowait; 634 635 if (!tick_nohz_active) 636 return -1; 637 638 now = ktime_get(); 639 if (last_update_time) { 640 update_ts_time_stats(cpu, ts, now, last_update_time); 641 iowait = ts->iowait_sleeptime; 642 } else { 643 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 644 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 645 646 iowait = ktime_add(ts->iowait_sleeptime, delta); 647 } else { 648 iowait = ts->iowait_sleeptime; 649 } 650 } 651 652 return ktime_to_us(iowait); 653 } 654 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 655 656 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 657 { 658 hrtimer_cancel(&ts->sched_timer); 659 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 660 661 /* Forward the time to expire in the future */ 662 hrtimer_forward(&ts->sched_timer, now, tick_period); 663 664 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 665 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 666 else 667 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 668 } 669 670 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 671 ktime_t now, int cpu) 672 { 673 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 674 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; 675 unsigned long seq, basejiff; 676 ktime_t tick; 677 678 /* Read jiffies and the time when jiffies were updated last */ 679 do { 680 seq = read_seqbegin(&jiffies_lock); 681 basemono = last_jiffies_update.tv64; 682 basejiff = jiffies; 683 } while (read_seqretry(&jiffies_lock, seq)); 684 ts->last_jiffies = basejiff; 685 686 if (rcu_needs_cpu(basemono, &next_rcu) || 687 arch_needs_cpu() || irq_work_needs_cpu()) { 688 next_tick = basemono + TICK_NSEC; 689 } else { 690 /* 691 * Get the next pending timer. If high resolution 692 * timers are enabled this only takes the timer wheel 693 * timers into account. If high resolution timers are 694 * disabled this also looks at the next expiring 695 * hrtimer. 696 */ 697 next_tmr = get_next_timer_interrupt(basejiff, basemono); 698 ts->next_timer = next_tmr; 699 /* Take the next rcu event into account */ 700 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; 701 } 702 703 /* 704 * If the tick is due in the next period, keep it ticking or 705 * force prod the timer. 706 */ 707 delta = next_tick - basemono; 708 if (delta <= (u64)TICK_NSEC) { 709 tick.tv64 = 0; 710 /* 711 * We've not stopped the tick yet, and there's a timer in the 712 * next period, so no point in stopping it either, bail. 713 */ 714 if (!ts->tick_stopped) 715 goto out; 716 717 /* 718 * If, OTOH, we did stop it, but there's a pending (expired) 719 * timer reprogram the timer hardware to fire now. 720 * 721 * We will not restart the tick proper, just prod the timer 722 * hardware into firing an interrupt to process the pending 723 * timers. Just like tick_irq_exit() will not restart the tick 724 * for 'normal' interrupts. 725 * 726 * Only once we exit the idle loop will we re-enable the tick, 727 * see tick_nohz_idle_exit(). 728 */ 729 if (delta == 0) { 730 tick_nohz_restart(ts, now); 731 goto out; 732 } 733 } 734 735 /* 736 * If this cpu is the one which updates jiffies, then give up 737 * the assignment and let it be taken by the cpu which runs 738 * the tick timer next, which might be this cpu as well. If we 739 * don't drop this here the jiffies might be stale and 740 * do_timer() never invoked. Keep track of the fact that it 741 * was the one which had the do_timer() duty last. If this cpu 742 * is the one which had the do_timer() duty last, we limit the 743 * sleep time to the timekeeping max_deferement value. 744 * Otherwise we can sleep as long as we want. 745 */ 746 delta = timekeeping_max_deferment(); 747 if (cpu == tick_do_timer_cpu) { 748 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 749 ts->do_timer_last = 1; 750 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 751 delta = KTIME_MAX; 752 ts->do_timer_last = 0; 753 } else if (!ts->do_timer_last) { 754 delta = KTIME_MAX; 755 } 756 757 #ifdef CONFIG_NO_HZ_FULL 758 /* Limit the tick delta to the maximum scheduler deferment */ 759 if (!ts->inidle) 760 delta = min(delta, scheduler_tick_max_deferment()); 761 #endif 762 763 /* Calculate the next expiry time */ 764 if (delta < (KTIME_MAX - basemono)) 765 expires = basemono + delta; 766 else 767 expires = KTIME_MAX; 768 769 expires = min_t(u64, expires, next_tick); 770 tick.tv64 = expires; 771 772 /* Skip reprogram of event if its not changed */ 773 if (ts->tick_stopped && (expires == dev->next_event.tv64)) 774 goto out; 775 776 /* 777 * nohz_stop_sched_tick can be called several times before 778 * the nohz_restart_sched_tick is called. This happens when 779 * interrupts arrive which do not cause a reschedule. In the 780 * first call we save the current tick time, so we can restart 781 * the scheduler tick in nohz_restart_sched_tick. 782 */ 783 if (!ts->tick_stopped) { 784 nohz_balance_enter_idle(cpu); 785 calc_load_enter_idle(); 786 787 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 788 ts->tick_stopped = 1; 789 trace_tick_stop(1, TICK_DEP_MASK_NONE); 790 } 791 792 /* 793 * If the expiration time == KTIME_MAX, then we simply stop 794 * the tick timer. 795 */ 796 if (unlikely(expires == KTIME_MAX)) { 797 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 798 hrtimer_cancel(&ts->sched_timer); 799 goto out; 800 } 801 802 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 803 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED); 804 else 805 tick_program_event(tick, 1); 806 out: 807 /* Update the estimated sleep length */ 808 ts->sleep_length = ktime_sub(dev->next_event, now); 809 return tick; 810 } 811 812 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now, int active) 813 { 814 /* Update jiffies first */ 815 tick_do_update_jiffies64(now); 816 update_cpu_load_nohz(active); 817 818 calc_load_exit_idle(); 819 touch_softlockup_watchdog_sched(); 820 /* 821 * Cancel the scheduled timer and restore the tick 822 */ 823 ts->tick_stopped = 0; 824 ts->idle_exittime = now; 825 826 tick_nohz_restart(ts, now); 827 } 828 829 static void tick_nohz_full_update_tick(struct tick_sched *ts) 830 { 831 #ifdef CONFIG_NO_HZ_FULL 832 int cpu = smp_processor_id(); 833 834 if (!tick_nohz_full_cpu(cpu)) 835 return; 836 837 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 838 return; 839 840 if (can_stop_full_tick(ts)) 841 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 842 else if (ts->tick_stopped) 843 tick_nohz_restart_sched_tick(ts, ktime_get(), 1); 844 #endif 845 } 846 847 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 848 { 849 /* 850 * If this cpu is offline and it is the one which updates 851 * jiffies, then give up the assignment and let it be taken by 852 * the cpu which runs the tick timer next. If we don't drop 853 * this here the jiffies might be stale and do_timer() never 854 * invoked. 855 */ 856 if (unlikely(!cpu_online(cpu))) { 857 if (cpu == tick_do_timer_cpu) 858 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 859 return false; 860 } 861 862 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 863 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ }; 864 return false; 865 } 866 867 if (need_resched()) 868 return false; 869 870 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 871 static int ratelimit; 872 873 if (ratelimit < 10 && 874 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 875 pr_warn("NOHZ: local_softirq_pending %02x\n", 876 (unsigned int) local_softirq_pending()); 877 ratelimit++; 878 } 879 return false; 880 } 881 882 if (tick_nohz_full_enabled()) { 883 /* 884 * Keep the tick alive to guarantee timekeeping progression 885 * if there are full dynticks CPUs around 886 */ 887 if (tick_do_timer_cpu == cpu) 888 return false; 889 /* 890 * Boot safety: make sure the timekeeping duty has been 891 * assigned before entering dyntick-idle mode, 892 */ 893 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 894 return false; 895 } 896 897 return true; 898 } 899 900 static void __tick_nohz_idle_enter(struct tick_sched *ts) 901 { 902 ktime_t now, expires; 903 int cpu = smp_processor_id(); 904 905 now = tick_nohz_start_idle(ts); 906 907 if (can_stop_idle_tick(cpu, ts)) { 908 int was_stopped = ts->tick_stopped; 909 910 ts->idle_calls++; 911 912 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 913 if (expires.tv64 > 0LL) { 914 ts->idle_sleeps++; 915 ts->idle_expires = expires; 916 } 917 918 if (!was_stopped && ts->tick_stopped) 919 ts->idle_jiffies = ts->last_jiffies; 920 } 921 } 922 923 /** 924 * tick_nohz_idle_enter - stop the idle tick from the idle task 925 * 926 * When the next event is more than a tick into the future, stop the idle tick 927 * Called when we start the idle loop. 928 * 929 * The arch is responsible of calling: 930 * 931 * - rcu_idle_enter() after its last use of RCU before the CPU is put 932 * to sleep. 933 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 934 */ 935 void tick_nohz_idle_enter(void) 936 { 937 struct tick_sched *ts; 938 939 WARN_ON_ONCE(irqs_disabled()); 940 941 /* 942 * Update the idle state in the scheduler domain hierarchy 943 * when tick_nohz_stop_sched_tick() is called from the idle loop. 944 * State will be updated to busy during the first busy tick after 945 * exiting idle. 946 */ 947 set_cpu_sd_state_idle(); 948 949 local_irq_disable(); 950 951 ts = this_cpu_ptr(&tick_cpu_sched); 952 ts->inidle = 1; 953 __tick_nohz_idle_enter(ts); 954 955 local_irq_enable(); 956 } 957 958 /** 959 * tick_nohz_irq_exit - update next tick event from interrupt exit 960 * 961 * When an interrupt fires while we are idle and it doesn't cause 962 * a reschedule, it may still add, modify or delete a timer, enqueue 963 * an RCU callback, etc... 964 * So we need to re-calculate and reprogram the next tick event. 965 */ 966 void tick_nohz_irq_exit(void) 967 { 968 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 969 970 if (ts->inidle) 971 __tick_nohz_idle_enter(ts); 972 else 973 tick_nohz_full_update_tick(ts); 974 } 975 976 /** 977 * tick_nohz_get_sleep_length - return the length of the current sleep 978 * 979 * Called from power state control code with interrupts disabled 980 */ 981 ktime_t tick_nohz_get_sleep_length(void) 982 { 983 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 984 985 return ts->sleep_length; 986 } 987 988 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 989 { 990 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 991 unsigned long ticks; 992 993 if (vtime_accounting_cpu_enabled()) 994 return; 995 /* 996 * We stopped the tick in idle. Update process times would miss the 997 * time we slept as update_process_times does only a 1 tick 998 * accounting. Enforce that this is accounted to idle ! 999 */ 1000 ticks = jiffies - ts->idle_jiffies; 1001 /* 1002 * We might be one off. Do not randomly account a huge number of ticks! 1003 */ 1004 if (ticks && ticks < LONG_MAX) 1005 account_idle_ticks(ticks); 1006 #endif 1007 } 1008 1009 /** 1010 * tick_nohz_idle_exit - restart the idle tick from the idle task 1011 * 1012 * Restart the idle tick when the CPU is woken up from idle 1013 * This also exit the RCU extended quiescent state. The CPU 1014 * can use RCU again after this function is called. 1015 */ 1016 void tick_nohz_idle_exit(void) 1017 { 1018 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1019 ktime_t now; 1020 1021 local_irq_disable(); 1022 1023 WARN_ON_ONCE(!ts->inidle); 1024 1025 ts->inidle = 0; 1026 1027 if (ts->idle_active || ts->tick_stopped) 1028 now = ktime_get(); 1029 1030 if (ts->idle_active) 1031 tick_nohz_stop_idle(ts, now); 1032 1033 if (ts->tick_stopped) { 1034 tick_nohz_restart_sched_tick(ts, now, 0); 1035 tick_nohz_account_idle_ticks(ts); 1036 } 1037 1038 local_irq_enable(); 1039 } 1040 1041 /* 1042 * The nohz low res interrupt handler 1043 */ 1044 static void tick_nohz_handler(struct clock_event_device *dev) 1045 { 1046 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1047 struct pt_regs *regs = get_irq_regs(); 1048 ktime_t now = ktime_get(); 1049 1050 dev->next_event.tv64 = KTIME_MAX; 1051 1052 tick_sched_do_timer(now); 1053 tick_sched_handle(ts, regs); 1054 1055 /* No need to reprogram if we are running tickless */ 1056 if (unlikely(ts->tick_stopped)) 1057 return; 1058 1059 hrtimer_forward(&ts->sched_timer, now, tick_period); 1060 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1061 } 1062 1063 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) 1064 { 1065 if (!tick_nohz_enabled) 1066 return; 1067 ts->nohz_mode = mode; 1068 /* One update is enough */ 1069 if (!test_and_set_bit(0, &tick_nohz_active)) 1070 timers_update_migration(true); 1071 } 1072 1073 /** 1074 * tick_nohz_switch_to_nohz - switch to nohz mode 1075 */ 1076 static void tick_nohz_switch_to_nohz(void) 1077 { 1078 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1079 ktime_t next; 1080 1081 if (!tick_nohz_enabled) 1082 return; 1083 1084 if (tick_switch_to_oneshot(tick_nohz_handler)) 1085 return; 1086 1087 /* 1088 * Recycle the hrtimer in ts, so we can share the 1089 * hrtimer_forward with the highres code. 1090 */ 1091 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1092 /* Get the next period */ 1093 next = tick_init_jiffy_update(); 1094 1095 hrtimer_set_expires(&ts->sched_timer, next); 1096 hrtimer_forward_now(&ts->sched_timer, tick_period); 1097 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1098 tick_nohz_activate(ts, NOHZ_MODE_LOWRES); 1099 } 1100 1101 /* 1102 * When NOHZ is enabled and the tick is stopped, we need to kick the 1103 * tick timer from irq_enter() so that the jiffies update is kept 1104 * alive during long running softirqs. That's ugly as hell, but 1105 * correctness is key even if we need to fix the offending softirq in 1106 * the first place. 1107 * 1108 * Note, this is different to tick_nohz_restart. We just kick the 1109 * timer and do not touch the other magic bits which need to be done 1110 * when idle is left. 1111 */ 1112 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now) 1113 { 1114 #if 0 1115 /* Switch back to 2.6.27 behaviour */ 1116 ktime_t delta; 1117 1118 /* 1119 * Do not touch the tick device, when the next expiry is either 1120 * already reached or less/equal than the tick period. 1121 */ 1122 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 1123 if (delta.tv64 <= tick_period.tv64) 1124 return; 1125 1126 tick_nohz_restart(ts, now); 1127 #endif 1128 } 1129 1130 static inline void tick_nohz_irq_enter(void) 1131 { 1132 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1133 ktime_t now; 1134 1135 if (!ts->idle_active && !ts->tick_stopped) 1136 return; 1137 now = ktime_get(); 1138 if (ts->idle_active) 1139 tick_nohz_stop_idle(ts, now); 1140 if (ts->tick_stopped) { 1141 tick_nohz_update_jiffies(now); 1142 tick_nohz_kick_tick(ts, now); 1143 } 1144 } 1145 1146 #else 1147 1148 static inline void tick_nohz_switch_to_nohz(void) { } 1149 static inline void tick_nohz_irq_enter(void) { } 1150 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } 1151 1152 #endif /* CONFIG_NO_HZ_COMMON */ 1153 1154 /* 1155 * Called from irq_enter to notify about the possible interruption of idle() 1156 */ 1157 void tick_irq_enter(void) 1158 { 1159 tick_check_oneshot_broadcast_this_cpu(); 1160 tick_nohz_irq_enter(); 1161 } 1162 1163 /* 1164 * High resolution timer specific code 1165 */ 1166 #ifdef CONFIG_HIGH_RES_TIMERS 1167 /* 1168 * We rearm the timer until we get disabled by the idle code. 1169 * Called with interrupts disabled. 1170 */ 1171 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1172 { 1173 struct tick_sched *ts = 1174 container_of(timer, struct tick_sched, sched_timer); 1175 struct pt_regs *regs = get_irq_regs(); 1176 ktime_t now = ktime_get(); 1177 1178 tick_sched_do_timer(now); 1179 1180 /* 1181 * Do not call, when we are not in irq context and have 1182 * no valid regs pointer 1183 */ 1184 if (regs) 1185 tick_sched_handle(ts, regs); 1186 1187 /* No need to reprogram if we are in idle or full dynticks mode */ 1188 if (unlikely(ts->tick_stopped)) 1189 return HRTIMER_NORESTART; 1190 1191 hrtimer_forward(timer, now, tick_period); 1192 1193 return HRTIMER_RESTART; 1194 } 1195 1196 static int sched_skew_tick; 1197 1198 static int __init skew_tick(char *str) 1199 { 1200 get_option(&str, &sched_skew_tick); 1201 1202 return 0; 1203 } 1204 early_param("skew_tick", skew_tick); 1205 1206 /** 1207 * tick_setup_sched_timer - setup the tick emulation timer 1208 */ 1209 void tick_setup_sched_timer(void) 1210 { 1211 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1212 ktime_t now = ktime_get(); 1213 1214 /* 1215 * Emulate tick processing via per-CPU hrtimers: 1216 */ 1217 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1218 ts->sched_timer.function = tick_sched_timer; 1219 1220 /* Get the next period (per cpu) */ 1221 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1222 1223 /* Offset the tick to avert jiffies_lock contention. */ 1224 if (sched_skew_tick) { 1225 u64 offset = ktime_to_ns(tick_period) >> 1; 1226 do_div(offset, num_possible_cpus()); 1227 offset *= smp_processor_id(); 1228 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1229 } 1230 1231 hrtimer_forward(&ts->sched_timer, now, tick_period); 1232 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 1233 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); 1234 } 1235 #endif /* HIGH_RES_TIMERS */ 1236 1237 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1238 void tick_cancel_sched_timer(int cpu) 1239 { 1240 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1241 1242 # ifdef CONFIG_HIGH_RES_TIMERS 1243 if (ts->sched_timer.base) 1244 hrtimer_cancel(&ts->sched_timer); 1245 # endif 1246 1247 memset(ts, 0, sizeof(*ts)); 1248 } 1249 #endif 1250 1251 /** 1252 * Async notification about clocksource changes 1253 */ 1254 void tick_clock_notify(void) 1255 { 1256 int cpu; 1257 1258 for_each_possible_cpu(cpu) 1259 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1260 } 1261 1262 /* 1263 * Async notification about clock event changes 1264 */ 1265 void tick_oneshot_notify(void) 1266 { 1267 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1268 1269 set_bit(0, &ts->check_clocks); 1270 } 1271 1272 /** 1273 * Check, if a change happened, which makes oneshot possible. 1274 * 1275 * Called cyclic from the hrtimer softirq (driven by the timer 1276 * softirq) allow_nohz signals, that we can switch into low-res nohz 1277 * mode, because high resolution timers are disabled (either compile 1278 * or runtime). Called with interrupts disabled. 1279 */ 1280 int tick_check_oneshot_change(int allow_nohz) 1281 { 1282 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1283 1284 if (!test_and_clear_bit(0, &ts->check_clocks)) 1285 return 0; 1286 1287 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1288 return 0; 1289 1290 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1291 return 0; 1292 1293 if (!allow_nohz) 1294 return 1; 1295 1296 tick_nohz_switch_to_nohz(); 1297 return 0; 1298 } 1299