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 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(); 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 tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now); 201 202 /* 203 * Re-evaluate the need for the tick on the current CPU 204 * and restart it if necessary. 205 */ 206 void __tick_nohz_full_check(void) 207 { 208 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 209 210 if (tick_nohz_full_cpu(smp_processor_id())) { 211 if (ts->tick_stopped && !is_idle_task(current)) { 212 if (!can_stop_full_tick()) 213 tick_nohz_restart_sched_tick(ts, ktime_get()); 214 } 215 } 216 } 217 218 static void nohz_full_kick_work_func(struct irq_work *work) 219 { 220 __tick_nohz_full_check(); 221 } 222 223 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 224 .func = nohz_full_kick_work_func, 225 }; 226 227 /* 228 * Kick this CPU if it's full dynticks in order to force it to 229 * re-evaluate its dependency on the tick and restart it if necessary. 230 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), 231 * is NMI safe. 232 */ 233 void tick_nohz_full_kick(void) 234 { 235 if (!tick_nohz_full_cpu(smp_processor_id())) 236 return; 237 238 irq_work_queue(&__get_cpu_var(nohz_full_kick_work)); 239 } 240 241 /* 242 * Kick the CPU if it's full dynticks in order to force it to 243 * re-evaluate its dependency on the tick and restart it if necessary. 244 */ 245 void tick_nohz_full_kick_cpu(int cpu) 246 { 247 if (!tick_nohz_full_cpu(cpu)) 248 return; 249 250 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); 251 } 252 253 static void nohz_full_kick_ipi(void *info) 254 { 255 __tick_nohz_full_check(); 256 } 257 258 /* 259 * Kick all full dynticks CPUs in order to force these to re-evaluate 260 * their dependency on the tick and restart it if necessary. 261 */ 262 void tick_nohz_full_kick_all(void) 263 { 264 if (!tick_nohz_full_running) 265 return; 266 267 preempt_disable(); 268 smp_call_function_many(tick_nohz_full_mask, 269 nohz_full_kick_ipi, NULL, false); 270 tick_nohz_full_kick(); 271 preempt_enable(); 272 } 273 274 /* 275 * Re-evaluate the need for the tick as we switch the current task. 276 * It might need the tick due to per task/process properties: 277 * perf events, posix cpu timers, ... 278 */ 279 void __tick_nohz_task_switch(struct task_struct *tsk) 280 { 281 unsigned long flags; 282 283 local_irq_save(flags); 284 285 if (!tick_nohz_full_cpu(smp_processor_id())) 286 goto out; 287 288 if (tick_nohz_tick_stopped() && !can_stop_full_tick()) 289 tick_nohz_full_kick(); 290 291 out: 292 local_irq_restore(flags); 293 } 294 295 /* Parse the boot-time nohz CPU list from the kernel parameters. */ 296 static int __init tick_nohz_full_setup(char *str) 297 { 298 int cpu; 299 300 alloc_bootmem_cpumask_var(&tick_nohz_full_mask); 301 alloc_bootmem_cpumask_var(&housekeeping_mask); 302 if (cpulist_parse(str, tick_nohz_full_mask) < 0) { 303 pr_warning("NOHZ: Incorrect nohz_full cpumask\n"); 304 return 1; 305 } 306 307 cpu = smp_processor_id(); 308 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 309 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu); 310 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 311 } 312 cpumask_andnot(housekeeping_mask, 313 cpu_possible_mask, tick_nohz_full_mask); 314 tick_nohz_full_running = true; 315 316 return 1; 317 } 318 __setup("nohz_full=", tick_nohz_full_setup); 319 320 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb, 321 unsigned long action, 322 void *hcpu) 323 { 324 unsigned int cpu = (unsigned long)hcpu; 325 326 switch (action & ~CPU_TASKS_FROZEN) { 327 case CPU_DOWN_PREPARE: 328 /* 329 * If we handle the timekeeping duty for full dynticks CPUs, 330 * we can't safely shutdown that CPU. 331 */ 332 if (tick_nohz_full_running && tick_do_timer_cpu == cpu) 333 return NOTIFY_BAD; 334 break; 335 } 336 return NOTIFY_OK; 337 } 338 339 /* 340 * Worst case string length in chunks of CPU range seems 2 steps 341 * separations: 0,2,4,6,... 342 * This is NR_CPUS + sizeof('\0') 343 */ 344 static char __initdata nohz_full_buf[NR_CPUS + 1]; 345 346 static int tick_nohz_init_all(void) 347 { 348 int err = -1; 349 350 #ifdef CONFIG_NO_HZ_FULL_ALL 351 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { 352 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n"); 353 return err; 354 } 355 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) { 356 pr_err("NO_HZ: Can't allocate not-full dynticks cpumask\n"); 357 return err; 358 } 359 err = 0; 360 cpumask_setall(tick_nohz_full_mask); 361 cpumask_clear_cpu(smp_processor_id(), tick_nohz_full_mask); 362 cpumask_clear(housekeeping_mask); 363 cpumask_set_cpu(smp_processor_id(), housekeeping_mask); 364 tick_nohz_full_running = true; 365 #endif 366 return err; 367 } 368 369 void __init tick_nohz_init(void) 370 { 371 int cpu; 372 373 if (!tick_nohz_full_running) { 374 if (tick_nohz_init_all() < 0) 375 return; 376 } 377 378 for_each_cpu(cpu, tick_nohz_full_mask) 379 context_tracking_cpu_set(cpu); 380 381 cpu_notifier(tick_nohz_cpu_down_callback, 0); 382 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), tick_nohz_full_mask); 383 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf); 384 } 385 #endif 386 387 /* 388 * NOHZ - aka dynamic tick functionality 389 */ 390 #ifdef CONFIG_NO_HZ_COMMON 391 /* 392 * NO HZ enabled ? 393 */ 394 static int tick_nohz_enabled __read_mostly = 1; 395 int tick_nohz_active __read_mostly; 396 /* 397 * Enable / Disable tickless mode 398 */ 399 static int __init setup_tick_nohz(char *str) 400 { 401 if (!strcmp(str, "off")) 402 tick_nohz_enabled = 0; 403 else if (!strcmp(str, "on")) 404 tick_nohz_enabled = 1; 405 else 406 return 0; 407 return 1; 408 } 409 410 __setup("nohz=", setup_tick_nohz); 411 412 /** 413 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 414 * 415 * Called from interrupt entry when the CPU was idle 416 * 417 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 418 * must be updated. Otherwise an interrupt handler could use a stale jiffy 419 * value. We do this unconditionally on any cpu, as we don't know whether the 420 * cpu, which has the update task assigned is in a long sleep. 421 */ 422 static void tick_nohz_update_jiffies(ktime_t now) 423 { 424 unsigned long flags; 425 426 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 427 428 local_irq_save(flags); 429 tick_do_update_jiffies64(now); 430 local_irq_restore(flags); 431 432 touch_softlockup_watchdog(); 433 } 434 435 /* 436 * Updates the per cpu time idle statistics counters 437 */ 438 static void 439 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 440 { 441 ktime_t delta; 442 443 if (ts->idle_active) { 444 delta = ktime_sub(now, ts->idle_entrytime); 445 if (nr_iowait_cpu(cpu) > 0) 446 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 447 else 448 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 449 ts->idle_entrytime = now; 450 } 451 452 if (last_update_time) 453 *last_update_time = ktime_to_us(now); 454 455 } 456 457 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 458 { 459 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 460 ts->idle_active = 0; 461 462 sched_clock_idle_wakeup_event(0); 463 } 464 465 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 466 { 467 ktime_t now = ktime_get(); 468 469 ts->idle_entrytime = now; 470 ts->idle_active = 1; 471 sched_clock_idle_sleep_event(); 472 return now; 473 } 474 475 /** 476 * get_cpu_idle_time_us - get the total idle time of a cpu 477 * @cpu: CPU number to query 478 * @last_update_time: variable to store update time in. Do not update 479 * counters if NULL. 480 * 481 * Return the cummulative idle time (since boot) for a given 482 * CPU, in microseconds. 483 * 484 * This time is measured via accounting rather than sampling, 485 * and is as accurate as ktime_get() is. 486 * 487 * This function returns -1 if NOHZ is not enabled. 488 */ 489 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 490 { 491 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 492 ktime_t now, idle; 493 494 if (!tick_nohz_active) 495 return -1; 496 497 now = ktime_get(); 498 if (last_update_time) { 499 update_ts_time_stats(cpu, ts, now, last_update_time); 500 idle = ts->idle_sleeptime; 501 } else { 502 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 503 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 504 505 idle = ktime_add(ts->idle_sleeptime, delta); 506 } else { 507 idle = ts->idle_sleeptime; 508 } 509 } 510 511 return ktime_to_us(idle); 512 513 } 514 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 515 516 /** 517 * get_cpu_iowait_time_us - get the total iowait time of a cpu 518 * @cpu: CPU number to query 519 * @last_update_time: variable to store update time in. Do not update 520 * counters if NULL. 521 * 522 * Return the cummulative iowait time (since boot) for a given 523 * CPU, in microseconds. 524 * 525 * This time is measured via accounting rather than sampling, 526 * and is as accurate as ktime_get() is. 527 * 528 * This function returns -1 if NOHZ is not enabled. 529 */ 530 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 531 { 532 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 533 ktime_t now, iowait; 534 535 if (!tick_nohz_active) 536 return -1; 537 538 now = ktime_get(); 539 if (last_update_time) { 540 update_ts_time_stats(cpu, ts, now, last_update_time); 541 iowait = ts->iowait_sleeptime; 542 } else { 543 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 544 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 545 546 iowait = ktime_add(ts->iowait_sleeptime, delta); 547 } else { 548 iowait = ts->iowait_sleeptime; 549 } 550 } 551 552 return ktime_to_us(iowait); 553 } 554 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 555 556 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 557 ktime_t now, int cpu) 558 { 559 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; 560 ktime_t last_update, expires, ret = { .tv64 = 0 }; 561 unsigned long rcu_delta_jiffies; 562 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; 563 u64 time_delta; 564 565 time_delta = timekeeping_max_deferment(); 566 567 /* Read jiffies and the time when jiffies were updated last */ 568 do { 569 seq = read_seqbegin(&jiffies_lock); 570 last_update = last_jiffies_update; 571 last_jiffies = jiffies; 572 } while (read_seqretry(&jiffies_lock, seq)); 573 574 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || 575 arch_needs_cpu() || irq_work_needs_cpu()) { 576 next_jiffies = last_jiffies + 1; 577 delta_jiffies = 1; 578 } else { 579 /* Get the next timer wheel timer */ 580 next_jiffies = get_next_timer_interrupt(last_jiffies); 581 delta_jiffies = next_jiffies - last_jiffies; 582 if (rcu_delta_jiffies < delta_jiffies) { 583 next_jiffies = last_jiffies + rcu_delta_jiffies; 584 delta_jiffies = rcu_delta_jiffies; 585 } 586 } 587 588 /* 589 * Do not stop the tick, if we are only one off (or less) 590 * or if the cpu is required for RCU: 591 */ 592 if (!ts->tick_stopped && delta_jiffies <= 1) 593 goto out; 594 595 /* Schedule the tick, if we are at least one jiffie off */ 596 if ((long)delta_jiffies >= 1) { 597 598 /* 599 * If this cpu is the one which updates jiffies, then 600 * give up the assignment and let it be taken by the 601 * cpu which runs the tick timer next, which might be 602 * this cpu as well. If we don't drop this here the 603 * jiffies might be stale and do_timer() never 604 * invoked. Keep track of the fact that it was the one 605 * which had the do_timer() duty last. If this cpu is 606 * the one which had the do_timer() duty last, we 607 * limit the sleep time to the timekeeping 608 * max_deferement value which we retrieved 609 * above. Otherwise we can sleep as long as we want. 610 */ 611 if (cpu == tick_do_timer_cpu) { 612 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 613 ts->do_timer_last = 1; 614 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 615 time_delta = KTIME_MAX; 616 ts->do_timer_last = 0; 617 } else if (!ts->do_timer_last) { 618 time_delta = KTIME_MAX; 619 } 620 621 #ifdef CONFIG_NO_HZ_FULL 622 if (!ts->inidle) { 623 time_delta = min(time_delta, 624 scheduler_tick_max_deferment()); 625 } 626 #endif 627 628 /* 629 * calculate the expiry time for the next timer wheel 630 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals 631 * that there is no timer pending or at least extremely 632 * far into the future (12 days for HZ=1000). In this 633 * case we set the expiry to the end of time. 634 */ 635 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { 636 /* 637 * Calculate the time delta for the next timer event. 638 * If the time delta exceeds the maximum time delta 639 * permitted by the current clocksource then adjust 640 * the time delta accordingly to ensure the 641 * clocksource does not wrap. 642 */ 643 time_delta = min_t(u64, time_delta, 644 tick_period.tv64 * delta_jiffies); 645 } 646 647 if (time_delta < KTIME_MAX) 648 expires = ktime_add_ns(last_update, time_delta); 649 else 650 expires.tv64 = KTIME_MAX; 651 652 /* Skip reprogram of event if its not changed */ 653 if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) 654 goto out; 655 656 ret = expires; 657 658 /* 659 * nohz_stop_sched_tick can be called several times before 660 * the nohz_restart_sched_tick is called. This happens when 661 * interrupts arrive which do not cause a reschedule. In the 662 * first call we save the current tick time, so we can restart 663 * the scheduler tick in nohz_restart_sched_tick. 664 */ 665 if (!ts->tick_stopped) { 666 nohz_balance_enter_idle(cpu); 667 calc_load_enter_idle(); 668 669 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 670 ts->tick_stopped = 1; 671 trace_tick_stop(1, " "); 672 } 673 674 /* 675 * If the expiration time == KTIME_MAX, then 676 * in this case we simply stop the tick timer. 677 */ 678 if (unlikely(expires.tv64 == KTIME_MAX)) { 679 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 680 hrtimer_cancel(&ts->sched_timer); 681 goto out; 682 } 683 684 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 685 hrtimer_start(&ts->sched_timer, expires, 686 HRTIMER_MODE_ABS_PINNED); 687 /* Check, if the timer was already in the past */ 688 if (hrtimer_active(&ts->sched_timer)) 689 goto out; 690 } else if (!tick_program_event(expires, 0)) 691 goto out; 692 /* 693 * We are past the event already. So we crossed a 694 * jiffie boundary. Update jiffies and raise the 695 * softirq. 696 */ 697 tick_do_update_jiffies64(ktime_get()); 698 } 699 raise_softirq_irqoff(TIMER_SOFTIRQ); 700 out: 701 ts->next_jiffies = next_jiffies; 702 ts->last_jiffies = last_jiffies; 703 ts->sleep_length = ktime_sub(dev->next_event, now); 704 705 return ret; 706 } 707 708 static void tick_nohz_full_stop_tick(struct tick_sched *ts) 709 { 710 #ifdef CONFIG_NO_HZ_FULL 711 int cpu = smp_processor_id(); 712 713 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current)) 714 return; 715 716 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 717 return; 718 719 if (!can_stop_full_tick()) 720 return; 721 722 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 723 #endif 724 } 725 726 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 727 { 728 /* 729 * If this cpu is offline and it is the one which updates 730 * jiffies, then give up the assignment and let it be taken by 731 * the cpu which runs the tick timer next. If we don't drop 732 * this here the jiffies might be stale and do_timer() never 733 * invoked. 734 */ 735 if (unlikely(!cpu_online(cpu))) { 736 if (cpu == tick_do_timer_cpu) 737 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 738 return false; 739 } 740 741 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 742 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ }; 743 return false; 744 } 745 746 if (need_resched()) 747 return false; 748 749 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 750 static int ratelimit; 751 752 if (ratelimit < 10 && 753 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 754 pr_warn("NOHZ: local_softirq_pending %02x\n", 755 (unsigned int) local_softirq_pending()); 756 ratelimit++; 757 } 758 return false; 759 } 760 761 if (tick_nohz_full_enabled()) { 762 /* 763 * Keep the tick alive to guarantee timekeeping progression 764 * if there are full dynticks CPUs around 765 */ 766 if (tick_do_timer_cpu == cpu) 767 return false; 768 /* 769 * Boot safety: make sure the timekeeping duty has been 770 * assigned before entering dyntick-idle mode, 771 */ 772 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 773 return false; 774 } 775 776 return true; 777 } 778 779 static void __tick_nohz_idle_enter(struct tick_sched *ts) 780 { 781 ktime_t now, expires; 782 int cpu = smp_processor_id(); 783 784 now = tick_nohz_start_idle(ts); 785 786 if (can_stop_idle_tick(cpu, ts)) { 787 int was_stopped = ts->tick_stopped; 788 789 ts->idle_calls++; 790 791 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 792 if (expires.tv64 > 0LL) { 793 ts->idle_sleeps++; 794 ts->idle_expires = expires; 795 } 796 797 if (!was_stopped && ts->tick_stopped) 798 ts->idle_jiffies = ts->last_jiffies; 799 } 800 } 801 802 /** 803 * tick_nohz_idle_enter - stop the idle tick from the idle task 804 * 805 * When the next event is more than a tick into the future, stop the idle tick 806 * Called when we start the idle loop. 807 * 808 * The arch is responsible of calling: 809 * 810 * - rcu_idle_enter() after its last use of RCU before the CPU is put 811 * to sleep. 812 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 813 */ 814 void tick_nohz_idle_enter(void) 815 { 816 struct tick_sched *ts; 817 818 WARN_ON_ONCE(irqs_disabled()); 819 820 /* 821 * Update the idle state in the scheduler domain hierarchy 822 * when tick_nohz_stop_sched_tick() is called from the idle loop. 823 * State will be updated to busy during the first busy tick after 824 * exiting idle. 825 */ 826 set_cpu_sd_state_idle(); 827 828 local_irq_disable(); 829 830 ts = &__get_cpu_var(tick_cpu_sched); 831 ts->inidle = 1; 832 __tick_nohz_idle_enter(ts); 833 834 local_irq_enable(); 835 } 836 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter); 837 838 /** 839 * tick_nohz_irq_exit - update next tick event from interrupt exit 840 * 841 * When an interrupt fires while we are idle and it doesn't cause 842 * a reschedule, it may still add, modify or delete a timer, enqueue 843 * an RCU callback, etc... 844 * So we need to re-calculate and reprogram the next tick event. 845 */ 846 void tick_nohz_irq_exit(void) 847 { 848 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 849 850 if (ts->inidle) 851 __tick_nohz_idle_enter(ts); 852 else 853 tick_nohz_full_stop_tick(ts); 854 } 855 856 /** 857 * tick_nohz_get_sleep_length - return the length of the current sleep 858 * 859 * Called from power state control code with interrupts disabled 860 */ 861 ktime_t tick_nohz_get_sleep_length(void) 862 { 863 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 864 865 return ts->sleep_length; 866 } 867 868 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 869 { 870 hrtimer_cancel(&ts->sched_timer); 871 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 872 873 while (1) { 874 /* Forward the time to expire in the future */ 875 hrtimer_forward(&ts->sched_timer, now, tick_period); 876 877 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 878 hrtimer_start_expires(&ts->sched_timer, 879 HRTIMER_MODE_ABS_PINNED); 880 /* Check, if the timer was already in the past */ 881 if (hrtimer_active(&ts->sched_timer)) 882 break; 883 } else { 884 if (!tick_program_event( 885 hrtimer_get_expires(&ts->sched_timer), 0)) 886 break; 887 } 888 /* Reread time and update jiffies */ 889 now = ktime_get(); 890 tick_do_update_jiffies64(now); 891 } 892 } 893 894 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 895 { 896 /* Update jiffies first */ 897 tick_do_update_jiffies64(now); 898 update_cpu_load_nohz(); 899 900 calc_load_exit_idle(); 901 touch_softlockup_watchdog(); 902 /* 903 * Cancel the scheduled timer and restore the tick 904 */ 905 ts->tick_stopped = 0; 906 ts->idle_exittime = now; 907 908 tick_nohz_restart(ts, now); 909 } 910 911 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 912 { 913 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 914 unsigned long ticks; 915 916 if (vtime_accounting_enabled()) 917 return; 918 /* 919 * We stopped the tick in idle. Update process times would miss the 920 * time we slept as update_process_times does only a 1 tick 921 * accounting. Enforce that this is accounted to idle ! 922 */ 923 ticks = jiffies - ts->idle_jiffies; 924 /* 925 * We might be one off. Do not randomly account a huge number of ticks! 926 */ 927 if (ticks && ticks < LONG_MAX) 928 account_idle_ticks(ticks); 929 #endif 930 } 931 932 /** 933 * tick_nohz_idle_exit - restart the idle tick from the idle task 934 * 935 * Restart the idle tick when the CPU is woken up from idle 936 * This also exit the RCU extended quiescent state. The CPU 937 * can use RCU again after this function is called. 938 */ 939 void tick_nohz_idle_exit(void) 940 { 941 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 942 ktime_t now; 943 944 local_irq_disable(); 945 946 WARN_ON_ONCE(!ts->inidle); 947 948 ts->inidle = 0; 949 950 if (ts->idle_active || ts->tick_stopped) 951 now = ktime_get(); 952 953 if (ts->idle_active) 954 tick_nohz_stop_idle(ts, now); 955 956 if (ts->tick_stopped) { 957 tick_nohz_restart_sched_tick(ts, now); 958 tick_nohz_account_idle_ticks(ts); 959 } 960 961 local_irq_enable(); 962 } 963 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit); 964 965 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 966 { 967 hrtimer_forward(&ts->sched_timer, now, tick_period); 968 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); 969 } 970 971 /* 972 * The nohz low res interrupt handler 973 */ 974 static void tick_nohz_handler(struct clock_event_device *dev) 975 { 976 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 977 struct pt_regs *regs = get_irq_regs(); 978 ktime_t now = ktime_get(); 979 980 dev->next_event.tv64 = KTIME_MAX; 981 982 tick_sched_do_timer(now); 983 tick_sched_handle(ts, regs); 984 985 while (tick_nohz_reprogram(ts, now)) { 986 now = ktime_get(); 987 tick_do_update_jiffies64(now); 988 } 989 } 990 991 /** 992 * tick_nohz_switch_to_nohz - switch to nohz mode 993 */ 994 static void tick_nohz_switch_to_nohz(void) 995 { 996 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 997 ktime_t next; 998 999 if (!tick_nohz_enabled) 1000 return; 1001 1002 local_irq_disable(); 1003 if (tick_switch_to_oneshot(tick_nohz_handler)) { 1004 local_irq_enable(); 1005 return; 1006 } 1007 tick_nohz_active = 1; 1008 ts->nohz_mode = NOHZ_MODE_LOWRES; 1009 1010 /* 1011 * Recycle the hrtimer in ts, so we can share the 1012 * hrtimer_forward with the highres code. 1013 */ 1014 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1015 /* Get the next period */ 1016 next = tick_init_jiffy_update(); 1017 1018 for (;;) { 1019 hrtimer_set_expires(&ts->sched_timer, next); 1020 if (!tick_program_event(next, 0)) 1021 break; 1022 next = ktime_add(next, tick_period); 1023 } 1024 local_irq_enable(); 1025 } 1026 1027 /* 1028 * When NOHZ is enabled and the tick is stopped, we need to kick the 1029 * tick timer from irq_enter() so that the jiffies update is kept 1030 * alive during long running softirqs. That's ugly as hell, but 1031 * correctness is key even if we need to fix the offending softirq in 1032 * the first place. 1033 * 1034 * Note, this is different to tick_nohz_restart. We just kick the 1035 * timer and do not touch the other magic bits which need to be done 1036 * when idle is left. 1037 */ 1038 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now) 1039 { 1040 #if 0 1041 /* Switch back to 2.6.27 behaviour */ 1042 ktime_t delta; 1043 1044 /* 1045 * Do not touch the tick device, when the next expiry is either 1046 * already reached or less/equal than the tick period. 1047 */ 1048 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 1049 if (delta.tv64 <= tick_period.tv64) 1050 return; 1051 1052 tick_nohz_restart(ts, now); 1053 #endif 1054 } 1055 1056 static inline void tick_nohz_irq_enter(void) 1057 { 1058 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 1059 ktime_t now; 1060 1061 if (!ts->idle_active && !ts->tick_stopped) 1062 return; 1063 now = ktime_get(); 1064 if (ts->idle_active) 1065 tick_nohz_stop_idle(ts, now); 1066 if (ts->tick_stopped) { 1067 tick_nohz_update_jiffies(now); 1068 tick_nohz_kick_tick(ts, now); 1069 } 1070 } 1071 1072 #else 1073 1074 static inline void tick_nohz_switch_to_nohz(void) { } 1075 static inline void tick_nohz_irq_enter(void) { } 1076 1077 #endif /* CONFIG_NO_HZ_COMMON */ 1078 1079 /* 1080 * Called from irq_enter to notify about the possible interruption of idle() 1081 */ 1082 void tick_irq_enter(void) 1083 { 1084 tick_check_oneshot_broadcast_this_cpu(); 1085 tick_nohz_irq_enter(); 1086 } 1087 1088 /* 1089 * High resolution timer specific code 1090 */ 1091 #ifdef CONFIG_HIGH_RES_TIMERS 1092 /* 1093 * We rearm the timer until we get disabled by the idle code. 1094 * Called with interrupts disabled. 1095 */ 1096 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1097 { 1098 struct tick_sched *ts = 1099 container_of(timer, struct tick_sched, sched_timer); 1100 struct pt_regs *regs = get_irq_regs(); 1101 ktime_t now = ktime_get(); 1102 1103 tick_sched_do_timer(now); 1104 1105 /* 1106 * Do not call, when we are not in irq context and have 1107 * no valid regs pointer 1108 */ 1109 if (regs) 1110 tick_sched_handle(ts, regs); 1111 1112 hrtimer_forward(timer, now, tick_period); 1113 1114 return HRTIMER_RESTART; 1115 } 1116 1117 static int sched_skew_tick; 1118 1119 static int __init skew_tick(char *str) 1120 { 1121 get_option(&str, &sched_skew_tick); 1122 1123 return 0; 1124 } 1125 early_param("skew_tick", skew_tick); 1126 1127 /** 1128 * tick_setup_sched_timer - setup the tick emulation timer 1129 */ 1130 void tick_setup_sched_timer(void) 1131 { 1132 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 1133 ktime_t now = ktime_get(); 1134 1135 /* 1136 * Emulate tick processing via per-CPU hrtimers: 1137 */ 1138 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1139 ts->sched_timer.function = tick_sched_timer; 1140 1141 /* Get the next period (per cpu) */ 1142 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1143 1144 /* Offset the tick to avert jiffies_lock contention. */ 1145 if (sched_skew_tick) { 1146 u64 offset = ktime_to_ns(tick_period) >> 1; 1147 do_div(offset, num_possible_cpus()); 1148 offset *= smp_processor_id(); 1149 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1150 } 1151 1152 for (;;) { 1153 hrtimer_forward(&ts->sched_timer, now, tick_period); 1154 hrtimer_start_expires(&ts->sched_timer, 1155 HRTIMER_MODE_ABS_PINNED); 1156 /* Check, if the timer was already in the past */ 1157 if (hrtimer_active(&ts->sched_timer)) 1158 break; 1159 now = ktime_get(); 1160 } 1161 1162 #ifdef CONFIG_NO_HZ_COMMON 1163 if (tick_nohz_enabled) { 1164 ts->nohz_mode = NOHZ_MODE_HIGHRES; 1165 tick_nohz_active = 1; 1166 } 1167 #endif 1168 } 1169 #endif /* HIGH_RES_TIMERS */ 1170 1171 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1172 void tick_cancel_sched_timer(int cpu) 1173 { 1174 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1175 1176 # ifdef CONFIG_HIGH_RES_TIMERS 1177 if (ts->sched_timer.base) 1178 hrtimer_cancel(&ts->sched_timer); 1179 # endif 1180 1181 memset(ts, 0, sizeof(*ts)); 1182 } 1183 #endif 1184 1185 /** 1186 * Async notification about clocksource changes 1187 */ 1188 void tick_clock_notify(void) 1189 { 1190 int cpu; 1191 1192 for_each_possible_cpu(cpu) 1193 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1194 } 1195 1196 /* 1197 * Async notification about clock event changes 1198 */ 1199 void tick_oneshot_notify(void) 1200 { 1201 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 1202 1203 set_bit(0, &ts->check_clocks); 1204 } 1205 1206 /** 1207 * Check, if a change happened, which makes oneshot possible. 1208 * 1209 * Called cyclic from the hrtimer softirq (driven by the timer 1210 * softirq) allow_nohz signals, that we can switch into low-res nohz 1211 * mode, because high resolution timers are disabled (either compile 1212 * or runtime). 1213 */ 1214 int tick_check_oneshot_change(int allow_nohz) 1215 { 1216 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 1217 1218 if (!test_and_clear_bit(0, &ts->check_clocks)) 1219 return 0; 1220 1221 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1222 return 0; 1223 1224 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1225 return 0; 1226 1227 if (!allow_nohz) 1228 return 1; 1229 1230 tick_nohz_switch_to_nohz(); 1231 return 0; 1232 } 1233