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 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK 208 /* 209 * sched_clock_tick() needs us? 210 * 211 * TODO: kick full dynticks CPUs when 212 * sched_clock_stable is set. 213 */ 214 if (!sched_clock_stable()) { 215 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE); 216 /* 217 * Don't allow the user to think they can get 218 * full NO_HZ with this machine. 219 */ 220 WARN_ONCE(tick_nohz_full_running, 221 "NO_HZ FULL will not work with unstable sched clock"); 222 return false; 223 } 224 #endif 225 226 return true; 227 } 228 229 static void nohz_full_kick_func(struct irq_work *work) 230 { 231 /* Empty, the tick restart happens on tick_nohz_irq_exit() */ 232 } 233 234 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 235 .func = nohz_full_kick_func, 236 }; 237 238 /* 239 * Kick this CPU if it's full dynticks in order to force it to 240 * re-evaluate its dependency on the tick and restart it if necessary. 241 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), 242 * is NMI safe. 243 */ 244 static void tick_nohz_full_kick(void) 245 { 246 if (!tick_nohz_full_cpu(smp_processor_id())) 247 return; 248 249 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work)); 250 } 251 252 /* 253 * Kick the CPU if it's full dynticks in order to force it to 254 * re-evaluate its dependency on the tick and restart it if necessary. 255 */ 256 void tick_nohz_full_kick_cpu(int cpu) 257 { 258 if (!tick_nohz_full_cpu(cpu)) 259 return; 260 261 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); 262 } 263 264 /* 265 * Kick all full dynticks CPUs in order to force these to re-evaluate 266 * their dependency on the tick and restart it if necessary. 267 */ 268 static void tick_nohz_full_kick_all(void) 269 { 270 int cpu; 271 272 if (!tick_nohz_full_running) 273 return; 274 275 preempt_disable(); 276 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask) 277 tick_nohz_full_kick_cpu(cpu); 278 preempt_enable(); 279 } 280 281 static void tick_nohz_dep_set_all(unsigned long *dep, 282 enum tick_dep_bits bit) 283 { 284 unsigned long prev; 285 286 prev = fetch_or(dep, BIT_MASK(bit)); 287 if (!prev) 288 tick_nohz_full_kick_all(); 289 } 290 291 /* 292 * Set a global tick dependency. Used by perf events that rely on freq and 293 * by unstable clock. 294 */ 295 void tick_nohz_dep_set(enum tick_dep_bits bit) 296 { 297 tick_nohz_dep_set_all(&tick_dep_mask, bit); 298 } 299 300 void tick_nohz_dep_clear(enum tick_dep_bits bit) 301 { 302 clear_bit(bit, &tick_dep_mask); 303 } 304 305 /* 306 * Set per-CPU tick dependency. Used by scheduler and perf events in order to 307 * manage events throttling. 308 */ 309 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit) 310 { 311 unsigned long prev; 312 struct tick_sched *ts; 313 314 ts = per_cpu_ptr(&tick_cpu_sched, cpu); 315 316 prev = fetch_or(&ts->tick_dep_mask, BIT_MASK(bit)); 317 if (!prev) { 318 preempt_disable(); 319 /* Perf needs local kick that is NMI safe */ 320 if (cpu == smp_processor_id()) { 321 tick_nohz_full_kick(); 322 } else { 323 /* Remote irq work not NMI-safe */ 324 if (!WARN_ON_ONCE(in_nmi())) 325 tick_nohz_full_kick_cpu(cpu); 326 } 327 preempt_enable(); 328 } 329 } 330 331 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit) 332 { 333 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); 334 335 clear_bit(bit, &ts->tick_dep_mask); 336 } 337 338 /* 339 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse 340 * per task timers. 341 */ 342 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit) 343 { 344 /* 345 * We could optimize this with just kicking the target running the task 346 * if that noise matters for nohz full users. 347 */ 348 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit); 349 } 350 351 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit) 352 { 353 clear_bit(bit, &tsk->tick_dep_mask); 354 } 355 356 /* 357 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse 358 * per process timers. 359 */ 360 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit) 361 { 362 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit); 363 } 364 365 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit) 366 { 367 clear_bit(bit, &sig->tick_dep_mask); 368 } 369 370 /* 371 * Re-evaluate the need for the tick as we switch the current task. 372 * It might need the tick due to per task/process properties: 373 * perf events, posix cpu timers, ... 374 */ 375 void __tick_nohz_task_switch(void) 376 { 377 unsigned long flags; 378 struct tick_sched *ts; 379 380 local_irq_save(flags); 381 382 if (!tick_nohz_full_cpu(smp_processor_id())) 383 goto out; 384 385 ts = this_cpu_ptr(&tick_cpu_sched); 386 387 if (ts->tick_stopped) { 388 if (current->tick_dep_mask || current->signal->tick_dep_mask) 389 tick_nohz_full_kick(); 390 } 391 out: 392 local_irq_restore(flags); 393 } 394 395 /* Parse the boot-time nohz CPU list from the kernel parameters. */ 396 static int __init tick_nohz_full_setup(char *str) 397 { 398 alloc_bootmem_cpumask_var(&tick_nohz_full_mask); 399 if (cpulist_parse(str, tick_nohz_full_mask) < 0) { 400 pr_warning("NOHZ: Incorrect nohz_full cpumask\n"); 401 free_bootmem_cpumask_var(tick_nohz_full_mask); 402 return 1; 403 } 404 tick_nohz_full_running = true; 405 406 return 1; 407 } 408 __setup("nohz_full=", tick_nohz_full_setup); 409 410 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb, 411 unsigned long action, 412 void *hcpu) 413 { 414 unsigned int cpu = (unsigned long)hcpu; 415 416 switch (action & ~CPU_TASKS_FROZEN) { 417 case CPU_DOWN_PREPARE: 418 /* 419 * The boot CPU handles housekeeping duty (unbound timers, 420 * workqueues, timekeeping, ...) on behalf of full dynticks 421 * CPUs. It must remain online when nohz full is enabled. 422 */ 423 if (tick_nohz_full_running && tick_do_timer_cpu == cpu) 424 return NOTIFY_BAD; 425 break; 426 } 427 return NOTIFY_OK; 428 } 429 430 static int tick_nohz_init_all(void) 431 { 432 int err = -1; 433 434 #ifdef CONFIG_NO_HZ_FULL_ALL 435 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { 436 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n"); 437 return err; 438 } 439 err = 0; 440 cpumask_setall(tick_nohz_full_mask); 441 tick_nohz_full_running = true; 442 #endif 443 return err; 444 } 445 446 void __init tick_nohz_init(void) 447 { 448 int cpu; 449 450 if (!tick_nohz_full_running) { 451 if (tick_nohz_init_all() < 0) 452 return; 453 } 454 455 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) { 456 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n"); 457 cpumask_clear(tick_nohz_full_mask); 458 tick_nohz_full_running = false; 459 return; 460 } 461 462 /* 463 * Full dynticks uses irq work to drive the tick rescheduling on safe 464 * locking contexts. But then we need irq work to raise its own 465 * interrupts to avoid circular dependency on the tick 466 */ 467 if (!arch_irq_work_has_interrupt()) { 468 pr_warning("NO_HZ: Can't run full dynticks because arch doesn't " 469 "support irq work self-IPIs\n"); 470 cpumask_clear(tick_nohz_full_mask); 471 cpumask_copy(housekeeping_mask, cpu_possible_mask); 472 tick_nohz_full_running = false; 473 return; 474 } 475 476 cpu = smp_processor_id(); 477 478 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 479 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu); 480 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 481 } 482 483 cpumask_andnot(housekeeping_mask, 484 cpu_possible_mask, tick_nohz_full_mask); 485 486 for_each_cpu(cpu, tick_nohz_full_mask) 487 context_tracking_cpu_set(cpu); 488 489 cpu_notifier(tick_nohz_cpu_down_callback, 0); 490 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", 491 cpumask_pr_args(tick_nohz_full_mask)); 492 493 /* 494 * We need at least one CPU to handle housekeeping work such 495 * as timekeeping, unbound timers, workqueues, ... 496 */ 497 WARN_ON_ONCE(cpumask_empty(housekeeping_mask)); 498 } 499 #endif 500 501 /* 502 * NOHZ - aka dynamic tick functionality 503 */ 504 #ifdef CONFIG_NO_HZ_COMMON 505 /* 506 * NO HZ enabled ? 507 */ 508 int tick_nohz_enabled __read_mostly = 1; 509 unsigned long tick_nohz_active __read_mostly; 510 /* 511 * Enable / Disable tickless mode 512 */ 513 static int __init setup_tick_nohz(char *str) 514 { 515 if (!strcmp(str, "off")) 516 tick_nohz_enabled = 0; 517 else if (!strcmp(str, "on")) 518 tick_nohz_enabled = 1; 519 else 520 return 0; 521 return 1; 522 } 523 524 __setup("nohz=", setup_tick_nohz); 525 526 int tick_nohz_tick_stopped(void) 527 { 528 return __this_cpu_read(tick_cpu_sched.tick_stopped); 529 } 530 531 /** 532 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 533 * 534 * Called from interrupt entry when the CPU was idle 535 * 536 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 537 * must be updated. Otherwise an interrupt handler could use a stale jiffy 538 * value. We do this unconditionally on any cpu, as we don't know whether the 539 * cpu, which has the update task assigned is in a long sleep. 540 */ 541 static void tick_nohz_update_jiffies(ktime_t now) 542 { 543 unsigned long flags; 544 545 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 546 547 local_irq_save(flags); 548 tick_do_update_jiffies64(now); 549 local_irq_restore(flags); 550 551 touch_softlockup_watchdog_sched(); 552 } 553 554 /* 555 * Updates the per cpu time idle statistics counters 556 */ 557 static void 558 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 559 { 560 ktime_t delta; 561 562 if (ts->idle_active) { 563 delta = ktime_sub(now, ts->idle_entrytime); 564 if (nr_iowait_cpu(cpu) > 0) 565 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 566 else 567 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 568 ts->idle_entrytime = now; 569 } 570 571 if (last_update_time) 572 *last_update_time = ktime_to_us(now); 573 574 } 575 576 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 577 { 578 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 579 ts->idle_active = 0; 580 581 sched_clock_idle_wakeup_event(0); 582 } 583 584 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 585 { 586 ktime_t now = ktime_get(); 587 588 ts->idle_entrytime = now; 589 ts->idle_active = 1; 590 sched_clock_idle_sleep_event(); 591 return now; 592 } 593 594 /** 595 * get_cpu_idle_time_us - get the total idle time of a cpu 596 * @cpu: CPU number to query 597 * @last_update_time: variable to store update time in. Do not update 598 * counters if NULL. 599 * 600 * Return the cummulative idle time (since boot) for a given 601 * CPU, in microseconds. 602 * 603 * This time is measured via accounting rather than sampling, 604 * and is as accurate as ktime_get() is. 605 * 606 * This function returns -1 if NOHZ is not enabled. 607 */ 608 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 609 { 610 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 611 ktime_t now, idle; 612 613 if (!tick_nohz_active) 614 return -1; 615 616 now = ktime_get(); 617 if (last_update_time) { 618 update_ts_time_stats(cpu, ts, now, last_update_time); 619 idle = ts->idle_sleeptime; 620 } else { 621 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 622 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 623 624 idle = ktime_add(ts->idle_sleeptime, delta); 625 } else { 626 idle = ts->idle_sleeptime; 627 } 628 } 629 630 return ktime_to_us(idle); 631 632 } 633 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 634 635 /** 636 * get_cpu_iowait_time_us - get the total iowait time of a cpu 637 * @cpu: CPU number to query 638 * @last_update_time: variable to store update time in. Do not update 639 * counters if NULL. 640 * 641 * Return the cummulative iowait time (since boot) for a given 642 * CPU, in microseconds. 643 * 644 * This time is measured via accounting rather than sampling, 645 * and is as accurate as ktime_get() is. 646 * 647 * This function returns -1 if NOHZ is not enabled. 648 */ 649 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 650 { 651 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 652 ktime_t now, iowait; 653 654 if (!tick_nohz_active) 655 return -1; 656 657 now = ktime_get(); 658 if (last_update_time) { 659 update_ts_time_stats(cpu, ts, now, last_update_time); 660 iowait = ts->iowait_sleeptime; 661 } else { 662 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 663 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 664 665 iowait = ktime_add(ts->iowait_sleeptime, delta); 666 } else { 667 iowait = ts->iowait_sleeptime; 668 } 669 } 670 671 return ktime_to_us(iowait); 672 } 673 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 674 675 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 676 { 677 hrtimer_cancel(&ts->sched_timer); 678 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 679 680 /* Forward the time to expire in the future */ 681 hrtimer_forward(&ts->sched_timer, now, tick_period); 682 683 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 684 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 685 else 686 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 687 } 688 689 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 690 ktime_t now, int cpu) 691 { 692 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 693 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; 694 unsigned long seq, basejiff; 695 ktime_t tick; 696 697 /* Read jiffies and the time when jiffies were updated last */ 698 do { 699 seq = read_seqbegin(&jiffies_lock); 700 basemono = last_jiffies_update.tv64; 701 basejiff = jiffies; 702 } while (read_seqretry(&jiffies_lock, seq)); 703 ts->last_jiffies = basejiff; 704 705 if (rcu_needs_cpu(basemono, &next_rcu) || 706 arch_needs_cpu() || irq_work_needs_cpu()) { 707 next_tick = basemono + TICK_NSEC; 708 } else { 709 /* 710 * Get the next pending timer. If high resolution 711 * timers are enabled this only takes the timer wheel 712 * timers into account. If high resolution timers are 713 * disabled this also looks at the next expiring 714 * hrtimer. 715 */ 716 next_tmr = get_next_timer_interrupt(basejiff, basemono); 717 ts->next_timer = next_tmr; 718 /* Take the next rcu event into account */ 719 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; 720 } 721 722 /* 723 * If the tick is due in the next period, keep it ticking or 724 * force prod the timer. 725 */ 726 delta = next_tick - basemono; 727 if (delta <= (u64)TICK_NSEC) { 728 tick.tv64 = 0; 729 /* 730 * We've not stopped the tick yet, and there's a timer in the 731 * next period, so no point in stopping it either, bail. 732 */ 733 if (!ts->tick_stopped) 734 goto out; 735 736 /* 737 * If, OTOH, we did stop it, but there's a pending (expired) 738 * timer reprogram the timer hardware to fire now. 739 * 740 * We will not restart the tick proper, just prod the timer 741 * hardware into firing an interrupt to process the pending 742 * timers. Just like tick_irq_exit() will not restart the tick 743 * for 'normal' interrupts. 744 * 745 * Only once we exit the idle loop will we re-enable the tick, 746 * see tick_nohz_idle_exit(). 747 */ 748 if (delta == 0) { 749 tick_nohz_restart(ts, now); 750 goto out; 751 } 752 } 753 754 /* 755 * If this cpu is the one which updates jiffies, then give up 756 * the assignment and let it be taken by the cpu which runs 757 * the tick timer next, which might be this cpu as well. If we 758 * don't drop this here the jiffies might be stale and 759 * do_timer() never invoked. Keep track of the fact that it 760 * was the one which had the do_timer() duty last. If this cpu 761 * is the one which had the do_timer() duty last, we limit the 762 * sleep time to the timekeeping max_deferement value. 763 * Otherwise we can sleep as long as we want. 764 */ 765 delta = timekeeping_max_deferment(); 766 if (cpu == tick_do_timer_cpu) { 767 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 768 ts->do_timer_last = 1; 769 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 770 delta = KTIME_MAX; 771 ts->do_timer_last = 0; 772 } else if (!ts->do_timer_last) { 773 delta = KTIME_MAX; 774 } 775 776 #ifdef CONFIG_NO_HZ_FULL 777 /* Limit the tick delta to the maximum scheduler deferment */ 778 if (!ts->inidle) 779 delta = min(delta, scheduler_tick_max_deferment()); 780 #endif 781 782 /* Calculate the next expiry time */ 783 if (delta < (KTIME_MAX - basemono)) 784 expires = basemono + delta; 785 else 786 expires = KTIME_MAX; 787 788 expires = min_t(u64, expires, next_tick); 789 tick.tv64 = expires; 790 791 /* Skip reprogram of event if its not changed */ 792 if (ts->tick_stopped && (expires == dev->next_event.tv64)) 793 goto out; 794 795 /* 796 * nohz_stop_sched_tick can be called several times before 797 * the nohz_restart_sched_tick is called. This happens when 798 * interrupts arrive which do not cause a reschedule. In the 799 * first call we save the current tick time, so we can restart 800 * the scheduler tick in nohz_restart_sched_tick. 801 */ 802 if (!ts->tick_stopped) { 803 nohz_balance_enter_idle(cpu); 804 calc_load_enter_idle(); 805 806 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 807 ts->tick_stopped = 1; 808 trace_tick_stop(1, TICK_DEP_MASK_NONE); 809 } 810 811 /* 812 * If the expiration time == KTIME_MAX, then we simply stop 813 * the tick timer. 814 */ 815 if (unlikely(expires == KTIME_MAX)) { 816 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 817 hrtimer_cancel(&ts->sched_timer); 818 goto out; 819 } 820 821 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 822 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED); 823 else 824 tick_program_event(tick, 1); 825 out: 826 /* Update the estimated sleep length */ 827 ts->sleep_length = ktime_sub(dev->next_event, now); 828 return tick; 829 } 830 831 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now, int active) 832 { 833 /* Update jiffies first */ 834 tick_do_update_jiffies64(now); 835 update_cpu_load_nohz(active); 836 837 calc_load_exit_idle(); 838 touch_softlockup_watchdog_sched(); 839 /* 840 * Cancel the scheduled timer and restore the tick 841 */ 842 ts->tick_stopped = 0; 843 ts->idle_exittime = now; 844 845 tick_nohz_restart(ts, now); 846 } 847 848 static void tick_nohz_full_update_tick(struct tick_sched *ts) 849 { 850 #ifdef CONFIG_NO_HZ_FULL 851 int cpu = smp_processor_id(); 852 853 if (!tick_nohz_full_cpu(cpu)) 854 return; 855 856 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 857 return; 858 859 if (can_stop_full_tick(ts)) 860 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 861 else if (ts->tick_stopped) 862 tick_nohz_restart_sched_tick(ts, ktime_get(), 1); 863 #endif 864 } 865 866 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 867 { 868 /* 869 * If this cpu is offline and it is the one which updates 870 * jiffies, then give up the assignment and let it be taken by 871 * the cpu which runs the tick timer next. If we don't drop 872 * this here the jiffies might be stale and do_timer() never 873 * invoked. 874 */ 875 if (unlikely(!cpu_online(cpu))) { 876 if (cpu == tick_do_timer_cpu) 877 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 878 return false; 879 } 880 881 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 882 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ }; 883 return false; 884 } 885 886 if (need_resched()) 887 return false; 888 889 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 890 static int ratelimit; 891 892 if (ratelimit < 10 && 893 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 894 pr_warn("NOHZ: local_softirq_pending %02x\n", 895 (unsigned int) local_softirq_pending()); 896 ratelimit++; 897 } 898 return false; 899 } 900 901 if (tick_nohz_full_enabled()) { 902 /* 903 * Keep the tick alive to guarantee timekeeping progression 904 * if there are full dynticks CPUs around 905 */ 906 if (tick_do_timer_cpu == cpu) 907 return false; 908 /* 909 * Boot safety: make sure the timekeeping duty has been 910 * assigned before entering dyntick-idle mode, 911 */ 912 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 913 return false; 914 } 915 916 return true; 917 } 918 919 static void __tick_nohz_idle_enter(struct tick_sched *ts) 920 { 921 ktime_t now, expires; 922 int cpu = smp_processor_id(); 923 924 now = tick_nohz_start_idle(ts); 925 926 if (can_stop_idle_tick(cpu, ts)) { 927 int was_stopped = ts->tick_stopped; 928 929 ts->idle_calls++; 930 931 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 932 if (expires.tv64 > 0LL) { 933 ts->idle_sleeps++; 934 ts->idle_expires = expires; 935 } 936 937 if (!was_stopped && ts->tick_stopped) 938 ts->idle_jiffies = ts->last_jiffies; 939 } 940 } 941 942 /** 943 * tick_nohz_idle_enter - stop the idle tick from the idle task 944 * 945 * When the next event is more than a tick into the future, stop the idle tick 946 * Called when we start the idle loop. 947 * 948 * The arch is responsible of calling: 949 * 950 * - rcu_idle_enter() after its last use of RCU before the CPU is put 951 * to sleep. 952 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 953 */ 954 void tick_nohz_idle_enter(void) 955 { 956 struct tick_sched *ts; 957 958 WARN_ON_ONCE(irqs_disabled()); 959 960 /* 961 * Update the idle state in the scheduler domain hierarchy 962 * when tick_nohz_stop_sched_tick() is called from the idle loop. 963 * State will be updated to busy during the first busy tick after 964 * exiting idle. 965 */ 966 set_cpu_sd_state_idle(); 967 968 local_irq_disable(); 969 970 ts = this_cpu_ptr(&tick_cpu_sched); 971 ts->inidle = 1; 972 __tick_nohz_idle_enter(ts); 973 974 local_irq_enable(); 975 } 976 977 /** 978 * tick_nohz_irq_exit - update next tick event from interrupt exit 979 * 980 * When an interrupt fires while we are idle and it doesn't cause 981 * a reschedule, it may still add, modify or delete a timer, enqueue 982 * an RCU callback, etc... 983 * So we need to re-calculate and reprogram the next tick event. 984 */ 985 void tick_nohz_irq_exit(void) 986 { 987 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 988 989 if (ts->inidle) 990 __tick_nohz_idle_enter(ts); 991 else 992 tick_nohz_full_update_tick(ts); 993 } 994 995 /** 996 * tick_nohz_get_sleep_length - return the length of the current sleep 997 * 998 * Called from power state control code with interrupts disabled 999 */ 1000 ktime_t tick_nohz_get_sleep_length(void) 1001 { 1002 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1003 1004 return ts->sleep_length; 1005 } 1006 1007 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 1008 { 1009 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 1010 unsigned long ticks; 1011 1012 if (vtime_accounting_cpu_enabled()) 1013 return; 1014 /* 1015 * We stopped the tick in idle. Update process times would miss the 1016 * time we slept as update_process_times does only a 1 tick 1017 * accounting. Enforce that this is accounted to idle ! 1018 */ 1019 ticks = jiffies - ts->idle_jiffies; 1020 /* 1021 * We might be one off. Do not randomly account a huge number of ticks! 1022 */ 1023 if (ticks && ticks < LONG_MAX) 1024 account_idle_ticks(ticks); 1025 #endif 1026 } 1027 1028 /** 1029 * tick_nohz_idle_exit - restart the idle tick from the idle task 1030 * 1031 * Restart the idle tick when the CPU is woken up from idle 1032 * This also exit the RCU extended quiescent state. The CPU 1033 * can use RCU again after this function is called. 1034 */ 1035 void tick_nohz_idle_exit(void) 1036 { 1037 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1038 ktime_t now; 1039 1040 local_irq_disable(); 1041 1042 WARN_ON_ONCE(!ts->inidle); 1043 1044 ts->inidle = 0; 1045 1046 if (ts->idle_active || ts->tick_stopped) 1047 now = ktime_get(); 1048 1049 if (ts->idle_active) 1050 tick_nohz_stop_idle(ts, now); 1051 1052 if (ts->tick_stopped) { 1053 tick_nohz_restart_sched_tick(ts, now, 0); 1054 tick_nohz_account_idle_ticks(ts); 1055 } 1056 1057 local_irq_enable(); 1058 } 1059 1060 /* 1061 * The nohz low res interrupt handler 1062 */ 1063 static void tick_nohz_handler(struct clock_event_device *dev) 1064 { 1065 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1066 struct pt_regs *regs = get_irq_regs(); 1067 ktime_t now = ktime_get(); 1068 1069 dev->next_event.tv64 = KTIME_MAX; 1070 1071 tick_sched_do_timer(now); 1072 tick_sched_handle(ts, regs); 1073 1074 /* No need to reprogram if we are running tickless */ 1075 if (unlikely(ts->tick_stopped)) 1076 return; 1077 1078 hrtimer_forward(&ts->sched_timer, now, tick_period); 1079 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1080 } 1081 1082 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) 1083 { 1084 if (!tick_nohz_enabled) 1085 return; 1086 ts->nohz_mode = mode; 1087 /* One update is enough */ 1088 if (!test_and_set_bit(0, &tick_nohz_active)) 1089 timers_update_migration(true); 1090 } 1091 1092 /** 1093 * tick_nohz_switch_to_nohz - switch to nohz mode 1094 */ 1095 static void tick_nohz_switch_to_nohz(void) 1096 { 1097 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1098 ktime_t next; 1099 1100 if (!tick_nohz_enabled) 1101 return; 1102 1103 if (tick_switch_to_oneshot(tick_nohz_handler)) 1104 return; 1105 1106 /* 1107 * Recycle the hrtimer in ts, so we can share the 1108 * hrtimer_forward with the highres code. 1109 */ 1110 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1111 /* Get the next period */ 1112 next = tick_init_jiffy_update(); 1113 1114 hrtimer_set_expires(&ts->sched_timer, next); 1115 hrtimer_forward_now(&ts->sched_timer, tick_period); 1116 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1117 tick_nohz_activate(ts, NOHZ_MODE_LOWRES); 1118 } 1119 1120 /* 1121 * When NOHZ is enabled and the tick is stopped, we need to kick the 1122 * tick timer from irq_enter() so that the jiffies update is kept 1123 * alive during long running softirqs. That's ugly as hell, but 1124 * correctness is key even if we need to fix the offending softirq in 1125 * the first place. 1126 * 1127 * Note, this is different to tick_nohz_restart. We just kick the 1128 * timer and do not touch the other magic bits which need to be done 1129 * when idle is left. 1130 */ 1131 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now) 1132 { 1133 #if 0 1134 /* Switch back to 2.6.27 behaviour */ 1135 ktime_t delta; 1136 1137 /* 1138 * Do not touch the tick device, when the next expiry is either 1139 * already reached or less/equal than the tick period. 1140 */ 1141 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 1142 if (delta.tv64 <= tick_period.tv64) 1143 return; 1144 1145 tick_nohz_restart(ts, now); 1146 #endif 1147 } 1148 1149 static inline void tick_nohz_irq_enter(void) 1150 { 1151 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1152 ktime_t now; 1153 1154 if (!ts->idle_active && !ts->tick_stopped) 1155 return; 1156 now = ktime_get(); 1157 if (ts->idle_active) 1158 tick_nohz_stop_idle(ts, now); 1159 if (ts->tick_stopped) { 1160 tick_nohz_update_jiffies(now); 1161 tick_nohz_kick_tick(ts, now); 1162 } 1163 } 1164 1165 #else 1166 1167 static inline void tick_nohz_switch_to_nohz(void) { } 1168 static inline void tick_nohz_irq_enter(void) { } 1169 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } 1170 1171 #endif /* CONFIG_NO_HZ_COMMON */ 1172 1173 /* 1174 * Called from irq_enter to notify about the possible interruption of idle() 1175 */ 1176 void tick_irq_enter(void) 1177 { 1178 tick_check_oneshot_broadcast_this_cpu(); 1179 tick_nohz_irq_enter(); 1180 } 1181 1182 /* 1183 * High resolution timer specific code 1184 */ 1185 #ifdef CONFIG_HIGH_RES_TIMERS 1186 /* 1187 * We rearm the timer until we get disabled by the idle code. 1188 * Called with interrupts disabled. 1189 */ 1190 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1191 { 1192 struct tick_sched *ts = 1193 container_of(timer, struct tick_sched, sched_timer); 1194 struct pt_regs *regs = get_irq_regs(); 1195 ktime_t now = ktime_get(); 1196 1197 tick_sched_do_timer(now); 1198 1199 /* 1200 * Do not call, when we are not in irq context and have 1201 * no valid regs pointer 1202 */ 1203 if (regs) 1204 tick_sched_handle(ts, regs); 1205 1206 /* No need to reprogram if we are in idle or full dynticks mode */ 1207 if (unlikely(ts->tick_stopped)) 1208 return HRTIMER_NORESTART; 1209 1210 hrtimer_forward(timer, now, tick_period); 1211 1212 return HRTIMER_RESTART; 1213 } 1214 1215 static int sched_skew_tick; 1216 1217 static int __init skew_tick(char *str) 1218 { 1219 get_option(&str, &sched_skew_tick); 1220 1221 return 0; 1222 } 1223 early_param("skew_tick", skew_tick); 1224 1225 /** 1226 * tick_setup_sched_timer - setup the tick emulation timer 1227 */ 1228 void tick_setup_sched_timer(void) 1229 { 1230 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1231 ktime_t now = ktime_get(); 1232 1233 /* 1234 * Emulate tick processing via per-CPU hrtimers: 1235 */ 1236 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1237 ts->sched_timer.function = tick_sched_timer; 1238 1239 /* Get the next period (per cpu) */ 1240 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1241 1242 /* Offset the tick to avert jiffies_lock contention. */ 1243 if (sched_skew_tick) { 1244 u64 offset = ktime_to_ns(tick_period) >> 1; 1245 do_div(offset, num_possible_cpus()); 1246 offset *= smp_processor_id(); 1247 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1248 } 1249 1250 hrtimer_forward(&ts->sched_timer, now, tick_period); 1251 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 1252 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); 1253 } 1254 #endif /* HIGH_RES_TIMERS */ 1255 1256 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1257 void tick_cancel_sched_timer(int cpu) 1258 { 1259 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1260 1261 # ifdef CONFIG_HIGH_RES_TIMERS 1262 if (ts->sched_timer.base) 1263 hrtimer_cancel(&ts->sched_timer); 1264 # endif 1265 1266 memset(ts, 0, sizeof(*ts)); 1267 } 1268 #endif 1269 1270 /** 1271 * Async notification about clocksource changes 1272 */ 1273 void tick_clock_notify(void) 1274 { 1275 int cpu; 1276 1277 for_each_possible_cpu(cpu) 1278 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1279 } 1280 1281 /* 1282 * Async notification about clock event changes 1283 */ 1284 void tick_oneshot_notify(void) 1285 { 1286 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1287 1288 set_bit(0, &ts->check_clocks); 1289 } 1290 1291 /** 1292 * Check, if a change happened, which makes oneshot possible. 1293 * 1294 * Called cyclic from the hrtimer softirq (driven by the timer 1295 * softirq) allow_nohz signals, that we can switch into low-res nohz 1296 * mode, because high resolution timers are disabled (either compile 1297 * or runtime). Called with interrupts disabled. 1298 */ 1299 int tick_check_oneshot_change(int allow_nohz) 1300 { 1301 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1302 1303 if (!test_and_clear_bit(0, &ts->check_clocks)) 1304 return 0; 1305 1306 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1307 return 0; 1308 1309 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1310 return 0; 1311 1312 if (!allow_nohz) 1313 return 1; 1314 1315 tick_nohz_switch_to_nohz(); 1316 return 0; 1317 } 1318