1 /* 2 * linux/kernel/time/tick-sched.c 3 * 4 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de> 5 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar 6 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner 7 * 8 * No idle tick implementation for low and high resolution timers 9 * 10 * Started by: Thomas Gleixner and Ingo Molnar 11 * 12 * Distribute under GPLv2. 13 */ 14 #include <linux/cpu.h> 15 #include <linux/err.h> 16 #include <linux/hrtimer.h> 17 #include <linux/interrupt.h> 18 #include <linux/kernel_stat.h> 19 #include <linux/percpu.h> 20 #include <linux/profile.h> 21 #include <linux/sched.h> 22 #include <linux/module.h> 23 #include <linux/irq_work.h> 24 #include <linux/posix-timers.h> 25 #include <linux/perf_event.h> 26 #include <linux/context_tracking.h> 27 28 #include <asm/irq_regs.h> 29 30 #include "tick-internal.h" 31 32 #include <trace/events/timer.h> 33 34 /* 35 * Per cpu nohz control structure 36 */ 37 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); 38 39 /* 40 * The time, when the last jiffy update happened. Protected by jiffies_lock. 41 */ 42 static ktime_t last_jiffies_update; 43 44 struct tick_sched *tick_get_tick_sched(int cpu) 45 { 46 return &per_cpu(tick_cpu_sched, cpu); 47 } 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 } 154 155 #ifdef CONFIG_NO_HZ_FULL 156 cpumask_var_t tick_nohz_full_mask; 157 cpumask_var_t housekeeping_mask; 158 bool tick_nohz_full_running; 159 160 static bool can_stop_full_tick(void) 161 { 162 WARN_ON_ONCE(!irqs_disabled()); 163 164 if (!sched_can_stop_tick()) { 165 trace_tick_stop(0, "more than 1 task in runqueue\n"); 166 return false; 167 } 168 169 if (!posix_cpu_timers_can_stop_tick(current)) { 170 trace_tick_stop(0, "posix timers running\n"); 171 return false; 172 } 173 174 if (!perf_event_can_stop_tick()) { 175 trace_tick_stop(0, "perf events running\n"); 176 return false; 177 } 178 179 /* sched_clock_tick() needs us? */ 180 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK 181 /* 182 * TODO: kick full dynticks CPUs when 183 * sched_clock_stable is set. 184 */ 185 if (!sched_clock_stable()) { 186 trace_tick_stop(0, "unstable sched clock\n"); 187 /* 188 * Don't allow the user to think they can get 189 * full NO_HZ with this machine. 190 */ 191 WARN_ONCE(tick_nohz_full_running, 192 "NO_HZ FULL will not work with unstable sched clock"); 193 return false; 194 } 195 #endif 196 197 return true; 198 } 199 200 static void nohz_full_kick_work_func(struct irq_work *work) 201 { 202 /* Empty, the tick restart happens on tick_nohz_irq_exit() */ 203 } 204 205 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 206 .func = nohz_full_kick_work_func, 207 }; 208 209 /* 210 * Kick this CPU if it's full dynticks in order to force it to 211 * re-evaluate its dependency on the tick and restart it if necessary. 212 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), 213 * is NMI safe. 214 */ 215 void tick_nohz_full_kick(void) 216 { 217 if (!tick_nohz_full_cpu(smp_processor_id())) 218 return; 219 220 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work)); 221 } 222 223 /* 224 * Kick the CPU if it's full dynticks in order to force it to 225 * re-evaluate its dependency on the tick and restart it if necessary. 226 */ 227 void tick_nohz_full_kick_cpu(int cpu) 228 { 229 if (!tick_nohz_full_cpu(cpu)) 230 return; 231 232 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); 233 } 234 235 static void nohz_full_kick_ipi(void *info) 236 { 237 /* Empty, the tick restart happens on tick_nohz_irq_exit() */ 238 } 239 240 /* 241 * Kick all full dynticks CPUs in order to force these to re-evaluate 242 * their dependency on the tick and restart it if necessary. 243 */ 244 void tick_nohz_full_kick_all(void) 245 { 246 if (!tick_nohz_full_running) 247 return; 248 249 preempt_disable(); 250 smp_call_function_many(tick_nohz_full_mask, 251 nohz_full_kick_ipi, NULL, false); 252 tick_nohz_full_kick(); 253 preempt_enable(); 254 } 255 256 /* 257 * Re-evaluate the need for the tick as we switch the current task. 258 * It might need the tick due to per task/process properties: 259 * perf events, posix cpu timers, ... 260 */ 261 void __tick_nohz_task_switch(void) 262 { 263 unsigned long flags; 264 265 local_irq_save(flags); 266 267 if (!tick_nohz_full_cpu(smp_processor_id())) 268 goto out; 269 270 if (tick_nohz_tick_stopped() && !can_stop_full_tick()) 271 tick_nohz_full_kick(); 272 273 out: 274 local_irq_restore(flags); 275 } 276 277 /* Parse the boot-time nohz CPU list from the kernel parameters. */ 278 static int __init tick_nohz_full_setup(char *str) 279 { 280 alloc_bootmem_cpumask_var(&tick_nohz_full_mask); 281 if (cpulist_parse(str, tick_nohz_full_mask) < 0) { 282 pr_warning("NOHZ: Incorrect nohz_full cpumask\n"); 283 free_bootmem_cpumask_var(tick_nohz_full_mask); 284 return 1; 285 } 286 tick_nohz_full_running = true; 287 288 return 1; 289 } 290 __setup("nohz_full=", tick_nohz_full_setup); 291 292 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb, 293 unsigned long action, 294 void *hcpu) 295 { 296 unsigned int cpu = (unsigned long)hcpu; 297 298 switch (action & ~CPU_TASKS_FROZEN) { 299 case CPU_DOWN_PREPARE: 300 /* 301 * The boot CPU handles housekeeping duty (unbound timers, 302 * workqueues, timekeeping, ...) on behalf of full dynticks 303 * CPUs. It must remain online when nohz full is enabled. 304 */ 305 if (tick_nohz_full_running && tick_do_timer_cpu == cpu) 306 return NOTIFY_BAD; 307 break; 308 } 309 return NOTIFY_OK; 310 } 311 312 static int tick_nohz_init_all(void) 313 { 314 int err = -1; 315 316 #ifdef CONFIG_NO_HZ_FULL_ALL 317 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { 318 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n"); 319 return err; 320 } 321 err = 0; 322 cpumask_setall(tick_nohz_full_mask); 323 tick_nohz_full_running = true; 324 #endif 325 return err; 326 } 327 328 void __init tick_nohz_init(void) 329 { 330 int cpu; 331 332 if (!tick_nohz_full_running) { 333 if (tick_nohz_init_all() < 0) 334 return; 335 } 336 337 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) { 338 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n"); 339 cpumask_clear(tick_nohz_full_mask); 340 tick_nohz_full_running = false; 341 return; 342 } 343 344 /* 345 * Full dynticks uses irq work to drive the tick rescheduling on safe 346 * locking contexts. But then we need irq work to raise its own 347 * interrupts to avoid circular dependency on the tick 348 */ 349 if (!arch_irq_work_has_interrupt()) { 350 pr_warning("NO_HZ: Can't run full dynticks because arch doesn't " 351 "support irq work self-IPIs\n"); 352 cpumask_clear(tick_nohz_full_mask); 353 cpumask_copy(housekeeping_mask, cpu_possible_mask); 354 tick_nohz_full_running = false; 355 return; 356 } 357 358 cpu = smp_processor_id(); 359 360 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 361 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu); 362 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 363 } 364 365 cpumask_andnot(housekeeping_mask, 366 cpu_possible_mask, tick_nohz_full_mask); 367 368 for_each_cpu(cpu, tick_nohz_full_mask) 369 context_tracking_cpu_set(cpu); 370 371 cpu_notifier(tick_nohz_cpu_down_callback, 0); 372 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", 373 cpumask_pr_args(tick_nohz_full_mask)); 374 375 /* 376 * We need at least one CPU to handle housekeeping work such 377 * as timekeeping, unbound timers, workqueues, ... 378 */ 379 WARN_ON_ONCE(cpumask_empty(housekeeping_mask)); 380 } 381 #endif 382 383 /* 384 * NOHZ - aka dynamic tick functionality 385 */ 386 #ifdef CONFIG_NO_HZ_COMMON 387 /* 388 * NO HZ enabled ? 389 */ 390 int tick_nohz_enabled __read_mostly = 1; 391 unsigned long tick_nohz_active __read_mostly; 392 /* 393 * Enable / Disable tickless mode 394 */ 395 static int __init setup_tick_nohz(char *str) 396 { 397 if (!strcmp(str, "off")) 398 tick_nohz_enabled = 0; 399 else if (!strcmp(str, "on")) 400 tick_nohz_enabled = 1; 401 else 402 return 0; 403 return 1; 404 } 405 406 __setup("nohz=", setup_tick_nohz); 407 408 int tick_nohz_tick_stopped(void) 409 { 410 return __this_cpu_read(tick_cpu_sched.tick_stopped); 411 } 412 413 /** 414 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 415 * 416 * Called from interrupt entry when the CPU was idle 417 * 418 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 419 * must be updated. Otherwise an interrupt handler could use a stale jiffy 420 * value. We do this unconditionally on any cpu, as we don't know whether the 421 * cpu, which has the update task assigned is in a long sleep. 422 */ 423 static void tick_nohz_update_jiffies(ktime_t now) 424 { 425 unsigned long flags; 426 427 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 428 429 local_irq_save(flags); 430 tick_do_update_jiffies64(now); 431 local_irq_restore(flags); 432 433 touch_softlockup_watchdog_sched(); 434 } 435 436 /* 437 * Updates the per cpu time idle statistics counters 438 */ 439 static void 440 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 441 { 442 ktime_t delta; 443 444 if (ts->idle_active) { 445 delta = ktime_sub(now, ts->idle_entrytime); 446 if (nr_iowait_cpu(cpu) > 0) 447 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 448 else 449 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 450 ts->idle_entrytime = now; 451 } 452 453 if (last_update_time) 454 *last_update_time = ktime_to_us(now); 455 456 } 457 458 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 459 { 460 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 461 ts->idle_active = 0; 462 463 sched_clock_idle_wakeup_event(0); 464 } 465 466 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 467 { 468 ktime_t now = ktime_get(); 469 470 ts->idle_entrytime = now; 471 ts->idle_active = 1; 472 sched_clock_idle_sleep_event(); 473 return now; 474 } 475 476 /** 477 * get_cpu_idle_time_us - get the total idle time of a cpu 478 * @cpu: CPU number to query 479 * @last_update_time: variable to store update time in. Do not update 480 * counters if NULL. 481 * 482 * Return the cummulative idle time (since boot) for a given 483 * CPU, in microseconds. 484 * 485 * This time is measured via accounting rather than sampling, 486 * and is as accurate as ktime_get() is. 487 * 488 * This function returns -1 if NOHZ is not enabled. 489 */ 490 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 491 { 492 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 493 ktime_t now, idle; 494 495 if (!tick_nohz_active) 496 return -1; 497 498 now = ktime_get(); 499 if (last_update_time) { 500 update_ts_time_stats(cpu, ts, now, last_update_time); 501 idle = ts->idle_sleeptime; 502 } else { 503 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 504 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 505 506 idle = ktime_add(ts->idle_sleeptime, delta); 507 } else { 508 idle = ts->idle_sleeptime; 509 } 510 } 511 512 return ktime_to_us(idle); 513 514 } 515 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 516 517 /** 518 * get_cpu_iowait_time_us - get the total iowait time of a cpu 519 * @cpu: CPU number to query 520 * @last_update_time: variable to store update time in. Do not update 521 * counters if NULL. 522 * 523 * Return the cummulative iowait time (since boot) for a given 524 * CPU, in microseconds. 525 * 526 * This time is measured via accounting rather than sampling, 527 * and is as accurate as ktime_get() is. 528 * 529 * This function returns -1 if NOHZ is not enabled. 530 */ 531 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 532 { 533 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 534 ktime_t now, iowait; 535 536 if (!tick_nohz_active) 537 return -1; 538 539 now = ktime_get(); 540 if (last_update_time) { 541 update_ts_time_stats(cpu, ts, now, last_update_time); 542 iowait = ts->iowait_sleeptime; 543 } else { 544 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 545 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 546 547 iowait = ktime_add(ts->iowait_sleeptime, delta); 548 } else { 549 iowait = ts->iowait_sleeptime; 550 } 551 } 552 553 return ktime_to_us(iowait); 554 } 555 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 556 557 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 558 { 559 hrtimer_cancel(&ts->sched_timer); 560 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 561 562 /* Forward the time to expire in the future */ 563 hrtimer_forward(&ts->sched_timer, now, tick_period); 564 565 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 566 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 567 else 568 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 569 } 570 571 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 572 ktime_t now, int cpu) 573 { 574 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 575 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; 576 unsigned long seq, basejiff; 577 ktime_t tick; 578 579 /* Read jiffies and the time when jiffies were updated last */ 580 do { 581 seq = read_seqbegin(&jiffies_lock); 582 basemono = last_jiffies_update.tv64; 583 basejiff = jiffies; 584 } while (read_seqretry(&jiffies_lock, seq)); 585 ts->last_jiffies = basejiff; 586 587 if (rcu_needs_cpu(basemono, &next_rcu) || 588 arch_needs_cpu() || irq_work_needs_cpu()) { 589 next_tick = basemono + TICK_NSEC; 590 } else { 591 /* 592 * Get the next pending timer. If high resolution 593 * timers are enabled this only takes the timer wheel 594 * timers into account. If high resolution timers are 595 * disabled this also looks at the next expiring 596 * hrtimer. 597 */ 598 next_tmr = get_next_timer_interrupt(basejiff, basemono); 599 ts->next_timer = next_tmr; 600 /* Take the next rcu event into account */ 601 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; 602 } 603 604 /* 605 * If the tick is due in the next period, keep it ticking or 606 * force prod the timer. 607 */ 608 delta = next_tick - basemono; 609 if (delta <= (u64)TICK_NSEC) { 610 tick.tv64 = 0; 611 /* 612 * We've not stopped the tick yet, and there's a timer in the 613 * next period, so no point in stopping it either, bail. 614 */ 615 if (!ts->tick_stopped) 616 goto out; 617 618 /* 619 * If, OTOH, we did stop it, but there's a pending (expired) 620 * timer reprogram the timer hardware to fire now. 621 * 622 * We will not restart the tick proper, just prod the timer 623 * hardware into firing an interrupt to process the pending 624 * timers. Just like tick_irq_exit() will not restart the tick 625 * for 'normal' interrupts. 626 * 627 * Only once we exit the idle loop will we re-enable the tick, 628 * see tick_nohz_idle_exit(). 629 */ 630 if (delta == 0) { 631 tick_nohz_restart(ts, now); 632 goto out; 633 } 634 } 635 636 /* 637 * If this cpu is the one which updates jiffies, then give up 638 * the assignment and let it be taken by the cpu which runs 639 * the tick timer next, which might be this cpu as well. If we 640 * don't drop this here the jiffies might be stale and 641 * do_timer() never invoked. Keep track of the fact that it 642 * was the one which had the do_timer() duty last. If this cpu 643 * is the one which had the do_timer() duty last, we limit the 644 * sleep time to the timekeeping max_deferement value. 645 * Otherwise we can sleep as long as we want. 646 */ 647 delta = timekeeping_max_deferment(); 648 if (cpu == tick_do_timer_cpu) { 649 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 650 ts->do_timer_last = 1; 651 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 652 delta = KTIME_MAX; 653 ts->do_timer_last = 0; 654 } else if (!ts->do_timer_last) { 655 delta = KTIME_MAX; 656 } 657 658 #ifdef CONFIG_NO_HZ_FULL 659 /* Limit the tick delta to the maximum scheduler deferment */ 660 if (!ts->inidle) 661 delta = min(delta, scheduler_tick_max_deferment()); 662 #endif 663 664 /* Calculate the next expiry time */ 665 if (delta < (KTIME_MAX - basemono)) 666 expires = basemono + delta; 667 else 668 expires = KTIME_MAX; 669 670 expires = min_t(u64, expires, next_tick); 671 tick.tv64 = expires; 672 673 /* Skip reprogram of event if its not changed */ 674 if (ts->tick_stopped && (expires == dev->next_event.tv64)) 675 goto out; 676 677 /* 678 * nohz_stop_sched_tick can be called several times before 679 * the nohz_restart_sched_tick is called. This happens when 680 * interrupts arrive which do not cause a reschedule. In the 681 * first call we save the current tick time, so we can restart 682 * the scheduler tick in nohz_restart_sched_tick. 683 */ 684 if (!ts->tick_stopped) { 685 nohz_balance_enter_idle(cpu); 686 calc_load_enter_idle(); 687 688 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 689 ts->tick_stopped = 1; 690 trace_tick_stop(1, " "); 691 } 692 693 /* 694 * If the expiration time == KTIME_MAX, then we simply stop 695 * the tick timer. 696 */ 697 if (unlikely(expires == KTIME_MAX)) { 698 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 699 hrtimer_cancel(&ts->sched_timer); 700 goto out; 701 } 702 703 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 704 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED); 705 else 706 tick_program_event(tick, 1); 707 out: 708 /* Update the estimated sleep length */ 709 ts->sleep_length = ktime_sub(dev->next_event, now); 710 return tick; 711 } 712 713 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now, int active) 714 { 715 /* Update jiffies first */ 716 tick_do_update_jiffies64(now); 717 update_cpu_load_nohz(active); 718 719 calc_load_exit_idle(); 720 touch_softlockup_watchdog_sched(); 721 /* 722 * Cancel the scheduled timer and restore the tick 723 */ 724 ts->tick_stopped = 0; 725 ts->idle_exittime = now; 726 727 tick_nohz_restart(ts, now); 728 } 729 730 static void tick_nohz_full_update_tick(struct tick_sched *ts) 731 { 732 #ifdef CONFIG_NO_HZ_FULL 733 int cpu = smp_processor_id(); 734 735 if (!tick_nohz_full_cpu(cpu)) 736 return; 737 738 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 739 return; 740 741 if (can_stop_full_tick()) 742 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 743 else if (ts->tick_stopped) 744 tick_nohz_restart_sched_tick(ts, ktime_get(), 1); 745 #endif 746 } 747 748 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 749 { 750 /* 751 * If this cpu is offline and it is the one which updates 752 * jiffies, then give up the assignment and let it be taken by 753 * the cpu which runs the tick timer next. If we don't drop 754 * this here the jiffies might be stale and do_timer() never 755 * invoked. 756 */ 757 if (unlikely(!cpu_online(cpu))) { 758 if (cpu == tick_do_timer_cpu) 759 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 760 return false; 761 } 762 763 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 764 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ }; 765 return false; 766 } 767 768 if (need_resched()) 769 return false; 770 771 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 772 static int ratelimit; 773 774 if (ratelimit < 10 && 775 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 776 pr_warn("NOHZ: local_softirq_pending %02x\n", 777 (unsigned int) local_softirq_pending()); 778 ratelimit++; 779 } 780 return false; 781 } 782 783 if (tick_nohz_full_enabled()) { 784 /* 785 * Keep the tick alive to guarantee timekeeping progression 786 * if there are full dynticks CPUs around 787 */ 788 if (tick_do_timer_cpu == cpu) 789 return false; 790 /* 791 * Boot safety: make sure the timekeeping duty has been 792 * assigned before entering dyntick-idle mode, 793 */ 794 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 795 return false; 796 } 797 798 return true; 799 } 800 801 static void __tick_nohz_idle_enter(struct tick_sched *ts) 802 { 803 ktime_t now, expires; 804 int cpu = smp_processor_id(); 805 806 now = tick_nohz_start_idle(ts); 807 808 if (can_stop_idle_tick(cpu, ts)) { 809 int was_stopped = ts->tick_stopped; 810 811 ts->idle_calls++; 812 813 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 814 if (expires.tv64 > 0LL) { 815 ts->idle_sleeps++; 816 ts->idle_expires = expires; 817 } 818 819 if (!was_stopped && ts->tick_stopped) 820 ts->idle_jiffies = ts->last_jiffies; 821 } 822 } 823 824 /** 825 * tick_nohz_idle_enter - stop the idle tick from the idle task 826 * 827 * When the next event is more than a tick into the future, stop the idle tick 828 * Called when we start the idle loop. 829 * 830 * The arch is responsible of calling: 831 * 832 * - rcu_idle_enter() after its last use of RCU before the CPU is put 833 * to sleep. 834 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 835 */ 836 void tick_nohz_idle_enter(void) 837 { 838 struct tick_sched *ts; 839 840 WARN_ON_ONCE(irqs_disabled()); 841 842 /* 843 * Update the idle state in the scheduler domain hierarchy 844 * when tick_nohz_stop_sched_tick() is called from the idle loop. 845 * State will be updated to busy during the first busy tick after 846 * exiting idle. 847 */ 848 set_cpu_sd_state_idle(); 849 850 local_irq_disable(); 851 852 ts = this_cpu_ptr(&tick_cpu_sched); 853 ts->inidle = 1; 854 __tick_nohz_idle_enter(ts); 855 856 local_irq_enable(); 857 } 858 859 /** 860 * tick_nohz_irq_exit - update next tick event from interrupt exit 861 * 862 * When an interrupt fires while we are idle and it doesn't cause 863 * a reschedule, it may still add, modify or delete a timer, enqueue 864 * an RCU callback, etc... 865 * So we need to re-calculate and reprogram the next tick event. 866 */ 867 void tick_nohz_irq_exit(void) 868 { 869 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 870 871 if (ts->inidle) 872 __tick_nohz_idle_enter(ts); 873 else 874 tick_nohz_full_update_tick(ts); 875 } 876 877 /** 878 * tick_nohz_get_sleep_length - return the length of the current sleep 879 * 880 * Called from power state control code with interrupts disabled 881 */ 882 ktime_t tick_nohz_get_sleep_length(void) 883 { 884 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 885 886 return ts->sleep_length; 887 } 888 889 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 890 { 891 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 892 unsigned long ticks; 893 894 if (vtime_accounting_cpu_enabled()) 895 return; 896 /* 897 * We stopped the tick in idle. Update process times would miss the 898 * time we slept as update_process_times does only a 1 tick 899 * accounting. Enforce that this is accounted to idle ! 900 */ 901 ticks = jiffies - ts->idle_jiffies; 902 /* 903 * We might be one off. Do not randomly account a huge number of ticks! 904 */ 905 if (ticks && ticks < LONG_MAX) 906 account_idle_ticks(ticks); 907 #endif 908 } 909 910 /** 911 * tick_nohz_idle_exit - restart the idle tick from the idle task 912 * 913 * Restart the idle tick when the CPU is woken up from idle 914 * This also exit the RCU extended quiescent state. The CPU 915 * can use RCU again after this function is called. 916 */ 917 void tick_nohz_idle_exit(void) 918 { 919 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 920 ktime_t now; 921 922 local_irq_disable(); 923 924 WARN_ON_ONCE(!ts->inidle); 925 926 ts->inidle = 0; 927 928 if (ts->idle_active || ts->tick_stopped) 929 now = ktime_get(); 930 931 if (ts->idle_active) 932 tick_nohz_stop_idle(ts, now); 933 934 if (ts->tick_stopped) { 935 tick_nohz_restart_sched_tick(ts, now, 0); 936 tick_nohz_account_idle_ticks(ts); 937 } 938 939 local_irq_enable(); 940 } 941 942 /* 943 * The nohz low res interrupt handler 944 */ 945 static void tick_nohz_handler(struct clock_event_device *dev) 946 { 947 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 948 struct pt_regs *regs = get_irq_regs(); 949 ktime_t now = ktime_get(); 950 951 dev->next_event.tv64 = KTIME_MAX; 952 953 tick_sched_do_timer(now); 954 tick_sched_handle(ts, regs); 955 956 /* No need to reprogram if we are running tickless */ 957 if (unlikely(ts->tick_stopped)) 958 return; 959 960 hrtimer_forward(&ts->sched_timer, now, tick_period); 961 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 962 } 963 964 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) 965 { 966 if (!tick_nohz_enabled) 967 return; 968 ts->nohz_mode = mode; 969 /* One update is enough */ 970 if (!test_and_set_bit(0, &tick_nohz_active)) 971 timers_update_migration(true); 972 } 973 974 /** 975 * tick_nohz_switch_to_nohz - switch to nohz mode 976 */ 977 static void tick_nohz_switch_to_nohz(void) 978 { 979 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 980 ktime_t next; 981 982 if (!tick_nohz_enabled) 983 return; 984 985 if (tick_switch_to_oneshot(tick_nohz_handler)) 986 return; 987 988 /* 989 * Recycle the hrtimer in ts, so we can share the 990 * hrtimer_forward with the highres code. 991 */ 992 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 993 /* Get the next period */ 994 next = tick_init_jiffy_update(); 995 996 hrtimer_forward_now(&ts->sched_timer, tick_period); 997 hrtimer_set_expires(&ts->sched_timer, next); 998 tick_program_event(next, 1); 999 tick_nohz_activate(ts, NOHZ_MODE_LOWRES); 1000 } 1001 1002 /* 1003 * When NOHZ is enabled and the tick is stopped, we need to kick the 1004 * tick timer from irq_enter() so that the jiffies update is kept 1005 * alive during long running softirqs. That's ugly as hell, but 1006 * correctness is key even if we need to fix the offending softirq in 1007 * the first place. 1008 * 1009 * Note, this is different to tick_nohz_restart. We just kick the 1010 * timer and do not touch the other magic bits which need to be done 1011 * when idle is left. 1012 */ 1013 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now) 1014 { 1015 #if 0 1016 /* Switch back to 2.6.27 behaviour */ 1017 ktime_t delta; 1018 1019 /* 1020 * Do not touch the tick device, when the next expiry is either 1021 * already reached or less/equal than the tick period. 1022 */ 1023 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 1024 if (delta.tv64 <= tick_period.tv64) 1025 return; 1026 1027 tick_nohz_restart(ts, now); 1028 #endif 1029 } 1030 1031 static inline void tick_nohz_irq_enter(void) 1032 { 1033 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1034 ktime_t now; 1035 1036 if (!ts->idle_active && !ts->tick_stopped) 1037 return; 1038 now = ktime_get(); 1039 if (ts->idle_active) 1040 tick_nohz_stop_idle(ts, now); 1041 if (ts->tick_stopped) { 1042 tick_nohz_update_jiffies(now); 1043 tick_nohz_kick_tick(ts, now); 1044 } 1045 } 1046 1047 #else 1048 1049 static inline void tick_nohz_switch_to_nohz(void) { } 1050 static inline void tick_nohz_irq_enter(void) { } 1051 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } 1052 1053 #endif /* CONFIG_NO_HZ_COMMON */ 1054 1055 /* 1056 * Called from irq_enter to notify about the possible interruption of idle() 1057 */ 1058 void tick_irq_enter(void) 1059 { 1060 tick_check_oneshot_broadcast_this_cpu(); 1061 tick_nohz_irq_enter(); 1062 } 1063 1064 /* 1065 * High resolution timer specific code 1066 */ 1067 #ifdef CONFIG_HIGH_RES_TIMERS 1068 /* 1069 * We rearm the timer until we get disabled by the idle code. 1070 * Called with interrupts disabled. 1071 */ 1072 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1073 { 1074 struct tick_sched *ts = 1075 container_of(timer, struct tick_sched, sched_timer); 1076 struct pt_regs *regs = get_irq_regs(); 1077 ktime_t now = ktime_get(); 1078 1079 tick_sched_do_timer(now); 1080 1081 /* 1082 * Do not call, when we are not in irq context and have 1083 * no valid regs pointer 1084 */ 1085 if (regs) 1086 tick_sched_handle(ts, regs); 1087 1088 /* No need to reprogram if we are in idle or full dynticks mode */ 1089 if (unlikely(ts->tick_stopped)) 1090 return HRTIMER_NORESTART; 1091 1092 hrtimer_forward(timer, now, tick_period); 1093 1094 return HRTIMER_RESTART; 1095 } 1096 1097 static int sched_skew_tick; 1098 1099 static int __init skew_tick(char *str) 1100 { 1101 get_option(&str, &sched_skew_tick); 1102 1103 return 0; 1104 } 1105 early_param("skew_tick", skew_tick); 1106 1107 /** 1108 * tick_setup_sched_timer - setup the tick emulation timer 1109 */ 1110 void tick_setup_sched_timer(void) 1111 { 1112 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1113 ktime_t now = ktime_get(); 1114 1115 /* 1116 * Emulate tick processing via per-CPU hrtimers: 1117 */ 1118 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1119 ts->sched_timer.function = tick_sched_timer; 1120 1121 /* Get the next period (per cpu) */ 1122 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1123 1124 /* Offset the tick to avert jiffies_lock contention. */ 1125 if (sched_skew_tick) { 1126 u64 offset = ktime_to_ns(tick_period) >> 1; 1127 do_div(offset, num_possible_cpus()); 1128 offset *= smp_processor_id(); 1129 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1130 } 1131 1132 hrtimer_forward(&ts->sched_timer, now, tick_period); 1133 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 1134 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); 1135 } 1136 #endif /* HIGH_RES_TIMERS */ 1137 1138 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1139 void tick_cancel_sched_timer(int cpu) 1140 { 1141 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1142 1143 # ifdef CONFIG_HIGH_RES_TIMERS 1144 if (ts->sched_timer.base) 1145 hrtimer_cancel(&ts->sched_timer); 1146 # endif 1147 1148 memset(ts, 0, sizeof(*ts)); 1149 } 1150 #endif 1151 1152 /** 1153 * Async notification about clocksource changes 1154 */ 1155 void tick_clock_notify(void) 1156 { 1157 int cpu; 1158 1159 for_each_possible_cpu(cpu) 1160 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1161 } 1162 1163 /* 1164 * Async notification about clock event changes 1165 */ 1166 void tick_oneshot_notify(void) 1167 { 1168 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1169 1170 set_bit(0, &ts->check_clocks); 1171 } 1172 1173 /** 1174 * Check, if a change happened, which makes oneshot possible. 1175 * 1176 * Called cyclic from the hrtimer softirq (driven by the timer 1177 * softirq) allow_nohz signals, that we can switch into low-res nohz 1178 * mode, because high resolution timers are disabled (either compile 1179 * or runtime). Called with interrupts disabled. 1180 */ 1181 int tick_check_oneshot_change(int allow_nohz) 1182 { 1183 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1184 1185 if (!test_and_clear_bit(0, &ts->check_clocks)) 1186 return 0; 1187 1188 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1189 return 0; 1190 1191 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1192 return 0; 1193 1194 if (!allow_nohz) 1195 return 1; 1196 1197 tick_nohz_switch_to_nohz(); 1198 return 0; 1199 } 1200