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 25 #include <asm/irq_regs.h> 26 27 #include "tick-internal.h" 28 29 /* 30 * Per cpu nohz control structure 31 */ 32 DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); 33 34 /* 35 * The time, when the last jiffy update happened. Protected by jiffies_lock. 36 */ 37 static ktime_t last_jiffies_update; 38 39 struct tick_sched *tick_get_tick_sched(int cpu) 40 { 41 return &per_cpu(tick_cpu_sched, cpu); 42 } 43 44 /* 45 * Must be called with interrupts disabled ! 46 */ 47 static void tick_do_update_jiffies64(ktime_t now) 48 { 49 unsigned long ticks = 0; 50 ktime_t delta; 51 52 /* 53 * Do a quick check without holding jiffies_lock: 54 */ 55 delta = ktime_sub(now, last_jiffies_update); 56 if (delta.tv64 < tick_period.tv64) 57 return; 58 59 /* Reevalute with jiffies_lock held */ 60 write_seqlock(&jiffies_lock); 61 62 delta = ktime_sub(now, last_jiffies_update); 63 if (delta.tv64 >= tick_period.tv64) { 64 65 delta = ktime_sub(delta, tick_period); 66 last_jiffies_update = ktime_add(last_jiffies_update, 67 tick_period); 68 69 /* Slow path for long timeouts */ 70 if (unlikely(delta.tv64 >= tick_period.tv64)) { 71 s64 incr = ktime_to_ns(tick_period); 72 73 ticks = ktime_divns(delta, incr); 74 75 last_jiffies_update = ktime_add_ns(last_jiffies_update, 76 incr * ticks); 77 } 78 do_timer(++ticks); 79 80 /* Keep the tick_next_period variable up to date */ 81 tick_next_period = ktime_add(last_jiffies_update, tick_period); 82 } 83 write_sequnlock(&jiffies_lock); 84 } 85 86 /* 87 * Initialize and return retrieve the jiffies update. 88 */ 89 static ktime_t tick_init_jiffy_update(void) 90 { 91 ktime_t period; 92 93 write_seqlock(&jiffies_lock); 94 /* Did we start the jiffies update yet ? */ 95 if (last_jiffies_update.tv64 == 0) 96 last_jiffies_update = tick_next_period; 97 period = last_jiffies_update; 98 write_sequnlock(&jiffies_lock); 99 return period; 100 } 101 102 103 static void tick_sched_do_timer(ktime_t now) 104 { 105 int cpu = smp_processor_id(); 106 107 #ifdef CONFIG_NO_HZ_COMMON 108 /* 109 * Check if the do_timer duty was dropped. We don't care about 110 * concurrency: This happens only when the cpu in charge went 111 * into a long sleep. If two cpus happen to assign themself to 112 * this duty, then the jiffies update is still serialized by 113 * jiffies_lock. 114 */ 115 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE) 116 && !tick_nohz_full_cpu(cpu)) 117 tick_do_timer_cpu = cpu; 118 #endif 119 120 /* Check, if the jiffies need an update */ 121 if (tick_do_timer_cpu == cpu) 122 tick_do_update_jiffies64(now); 123 } 124 125 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) 126 { 127 #ifdef CONFIG_NO_HZ_COMMON 128 /* 129 * When we are idle and the tick is stopped, we have to touch 130 * the watchdog as we might not schedule for a really long 131 * time. This happens on complete idle SMP systems while 132 * waiting on the login prompt. We also increment the "start of 133 * idle" jiffy stamp so the idle accounting adjustment we do 134 * when we go busy again does not account too much ticks. 135 */ 136 if (ts->tick_stopped) { 137 touch_softlockup_watchdog(); 138 if (is_idle_task(current)) 139 ts->idle_jiffies++; 140 } 141 #endif 142 update_process_times(user_mode(regs)); 143 profile_tick(CPU_PROFILING); 144 } 145 146 #ifdef CONFIG_NO_HZ_FULL 147 static cpumask_var_t nohz_full_mask; 148 bool have_nohz_full_mask; 149 150 /* 151 * Re-evaluate the need for the tick on the current CPU 152 * and restart it if necessary. 153 */ 154 void tick_nohz_full_check(void) 155 { 156 /* 157 * STUB for now, will be filled with the full tick stop/restart 158 * infrastructure patches 159 */ 160 } 161 162 static void nohz_full_kick_work_func(struct irq_work *work) 163 { 164 tick_nohz_full_check(); 165 } 166 167 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 168 .func = nohz_full_kick_work_func, 169 }; 170 171 /* 172 * Kick the current CPU if it's full dynticks in order to force it to 173 * re-evaluate its dependency on the tick and restart it if necessary. 174 */ 175 void tick_nohz_full_kick(void) 176 { 177 if (tick_nohz_full_cpu(smp_processor_id())) 178 irq_work_queue(&__get_cpu_var(nohz_full_kick_work)); 179 } 180 181 static void nohz_full_kick_ipi(void *info) 182 { 183 tick_nohz_full_check(); 184 } 185 186 /* 187 * Kick all full dynticks CPUs in order to force these to re-evaluate 188 * their dependency on the tick and restart it if necessary. 189 */ 190 void tick_nohz_full_kick_all(void) 191 { 192 if (!have_nohz_full_mask) 193 return; 194 195 preempt_disable(); 196 smp_call_function_many(nohz_full_mask, 197 nohz_full_kick_ipi, NULL, false); 198 preempt_enable(); 199 } 200 201 int tick_nohz_full_cpu(int cpu) 202 { 203 if (!have_nohz_full_mask) 204 return 0; 205 206 return cpumask_test_cpu(cpu, nohz_full_mask); 207 } 208 209 /* Parse the boot-time nohz CPU list from the kernel parameters. */ 210 static int __init tick_nohz_full_setup(char *str) 211 { 212 int cpu; 213 214 alloc_bootmem_cpumask_var(&nohz_full_mask); 215 if (cpulist_parse(str, nohz_full_mask) < 0) { 216 pr_warning("NOHZ: Incorrect nohz_full cpumask\n"); 217 return 1; 218 } 219 220 cpu = smp_processor_id(); 221 if (cpumask_test_cpu(cpu, nohz_full_mask)) { 222 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu); 223 cpumask_clear_cpu(cpu, nohz_full_mask); 224 } 225 have_nohz_full_mask = true; 226 227 return 1; 228 } 229 __setup("nohz_full=", tick_nohz_full_setup); 230 231 static int __cpuinit tick_nohz_cpu_down_callback(struct notifier_block *nfb, 232 unsigned long action, 233 void *hcpu) 234 { 235 unsigned int cpu = (unsigned long)hcpu; 236 237 switch (action & ~CPU_TASKS_FROZEN) { 238 case CPU_DOWN_PREPARE: 239 /* 240 * If we handle the timekeeping duty for full dynticks CPUs, 241 * we can't safely shutdown that CPU. 242 */ 243 if (have_nohz_full_mask && tick_do_timer_cpu == cpu) 244 return -EINVAL; 245 break; 246 } 247 return NOTIFY_OK; 248 } 249 250 /* 251 * Worst case string length in chunks of CPU range seems 2 steps 252 * separations: 0,2,4,6,... 253 * This is NR_CPUS + sizeof('\0') 254 */ 255 static char __initdata nohz_full_buf[NR_CPUS + 1]; 256 257 static int tick_nohz_init_all(void) 258 { 259 int err = -1; 260 261 #ifdef CONFIG_NO_HZ_FULL_ALL 262 if (!alloc_cpumask_var(&nohz_full_mask, GFP_KERNEL)) { 263 pr_err("NO_HZ: Can't allocate full dynticks cpumask\n"); 264 return err; 265 } 266 err = 0; 267 cpumask_setall(nohz_full_mask); 268 cpumask_clear_cpu(smp_processor_id(), nohz_full_mask); 269 have_nohz_full_mask = true; 270 #endif 271 return err; 272 } 273 274 void __init tick_nohz_init(void) 275 { 276 int cpu; 277 278 if (!have_nohz_full_mask) { 279 if (tick_nohz_init_all() < 0) 280 return; 281 } 282 283 cpu_notifier(tick_nohz_cpu_down_callback, 0); 284 285 /* Make sure full dynticks CPU are also RCU nocbs */ 286 for_each_cpu(cpu, nohz_full_mask) { 287 if (!rcu_is_nocb_cpu(cpu)) { 288 pr_warning("NO_HZ: CPU %d is not RCU nocb: " 289 "cleared from nohz_full range", cpu); 290 cpumask_clear_cpu(cpu, nohz_full_mask); 291 } 292 } 293 294 cpulist_scnprintf(nohz_full_buf, sizeof(nohz_full_buf), nohz_full_mask); 295 pr_info("NO_HZ: Full dynticks CPUs: %s.\n", nohz_full_buf); 296 } 297 #else 298 #define have_nohz_full_mask (0) 299 #endif 300 301 /* 302 * NOHZ - aka dynamic tick functionality 303 */ 304 #ifdef CONFIG_NO_HZ_COMMON 305 /* 306 * NO HZ enabled ? 307 */ 308 int tick_nohz_enabled __read_mostly = 1; 309 310 /* 311 * Enable / Disable tickless mode 312 */ 313 static int __init setup_tick_nohz(char *str) 314 { 315 if (!strcmp(str, "off")) 316 tick_nohz_enabled = 0; 317 else if (!strcmp(str, "on")) 318 tick_nohz_enabled = 1; 319 else 320 return 0; 321 return 1; 322 } 323 324 __setup("nohz=", setup_tick_nohz); 325 326 /** 327 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 328 * 329 * Called from interrupt entry when the CPU was idle 330 * 331 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 332 * must be updated. Otherwise an interrupt handler could use a stale jiffy 333 * value. We do this unconditionally on any cpu, as we don't know whether the 334 * cpu, which has the update task assigned is in a long sleep. 335 */ 336 static void tick_nohz_update_jiffies(ktime_t now) 337 { 338 int cpu = smp_processor_id(); 339 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 340 unsigned long flags; 341 342 ts->idle_waketime = now; 343 344 local_irq_save(flags); 345 tick_do_update_jiffies64(now); 346 local_irq_restore(flags); 347 348 touch_softlockup_watchdog(); 349 } 350 351 /* 352 * Updates the per cpu time idle statistics counters 353 */ 354 static void 355 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 356 { 357 ktime_t delta; 358 359 if (ts->idle_active) { 360 delta = ktime_sub(now, ts->idle_entrytime); 361 if (nr_iowait_cpu(cpu) > 0) 362 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 363 else 364 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 365 ts->idle_entrytime = now; 366 } 367 368 if (last_update_time) 369 *last_update_time = ktime_to_us(now); 370 371 } 372 373 static void tick_nohz_stop_idle(int cpu, ktime_t now) 374 { 375 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 376 377 update_ts_time_stats(cpu, ts, now, NULL); 378 ts->idle_active = 0; 379 380 sched_clock_idle_wakeup_event(0); 381 } 382 383 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts) 384 { 385 ktime_t now = ktime_get(); 386 387 ts->idle_entrytime = now; 388 ts->idle_active = 1; 389 sched_clock_idle_sleep_event(); 390 return now; 391 } 392 393 /** 394 * get_cpu_idle_time_us - get the total idle time of a cpu 395 * @cpu: CPU number to query 396 * @last_update_time: variable to store update time in. Do not update 397 * counters if NULL. 398 * 399 * Return the cummulative idle time (since boot) for a given 400 * CPU, in microseconds. 401 * 402 * This time is measured via accounting rather than sampling, 403 * and is as accurate as ktime_get() is. 404 * 405 * This function returns -1 if NOHZ is not enabled. 406 */ 407 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 408 { 409 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 410 ktime_t now, idle; 411 412 if (!tick_nohz_enabled) 413 return -1; 414 415 now = ktime_get(); 416 if (last_update_time) { 417 update_ts_time_stats(cpu, ts, now, last_update_time); 418 idle = ts->idle_sleeptime; 419 } else { 420 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 421 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 422 423 idle = ktime_add(ts->idle_sleeptime, delta); 424 } else { 425 idle = ts->idle_sleeptime; 426 } 427 } 428 429 return ktime_to_us(idle); 430 431 } 432 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 433 434 /** 435 * get_cpu_iowait_time_us - get the total iowait time of a cpu 436 * @cpu: CPU number to query 437 * @last_update_time: variable to store update time in. Do not update 438 * counters if NULL. 439 * 440 * Return the cummulative iowait time (since boot) for a given 441 * CPU, in microseconds. 442 * 443 * This time is measured via accounting rather than sampling, 444 * and is as accurate as ktime_get() is. 445 * 446 * This function returns -1 if NOHZ is not enabled. 447 */ 448 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 449 { 450 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 451 ktime_t now, iowait; 452 453 if (!tick_nohz_enabled) 454 return -1; 455 456 now = ktime_get(); 457 if (last_update_time) { 458 update_ts_time_stats(cpu, ts, now, last_update_time); 459 iowait = ts->iowait_sleeptime; 460 } else { 461 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 462 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 463 464 iowait = ktime_add(ts->iowait_sleeptime, delta); 465 } else { 466 iowait = ts->iowait_sleeptime; 467 } 468 } 469 470 return ktime_to_us(iowait); 471 } 472 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 473 474 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 475 ktime_t now, int cpu) 476 { 477 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; 478 ktime_t last_update, expires, ret = { .tv64 = 0 }; 479 unsigned long rcu_delta_jiffies; 480 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; 481 u64 time_delta; 482 483 /* Read jiffies and the time when jiffies were updated last */ 484 do { 485 seq = read_seqbegin(&jiffies_lock); 486 last_update = last_jiffies_update; 487 last_jiffies = jiffies; 488 time_delta = timekeeping_max_deferment(); 489 } while (read_seqretry(&jiffies_lock, seq)); 490 491 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || 492 arch_needs_cpu(cpu) || irq_work_needs_cpu()) { 493 next_jiffies = last_jiffies + 1; 494 delta_jiffies = 1; 495 } else { 496 /* Get the next timer wheel timer */ 497 next_jiffies = get_next_timer_interrupt(last_jiffies); 498 delta_jiffies = next_jiffies - last_jiffies; 499 if (rcu_delta_jiffies < delta_jiffies) { 500 next_jiffies = last_jiffies + rcu_delta_jiffies; 501 delta_jiffies = rcu_delta_jiffies; 502 } 503 } 504 /* 505 * Do not stop the tick, if we are only one off 506 * or if the cpu is required for rcu 507 */ 508 if (!ts->tick_stopped && delta_jiffies == 1) 509 goto out; 510 511 /* Schedule the tick, if we are at least one jiffie off */ 512 if ((long)delta_jiffies >= 1) { 513 514 /* 515 * If this cpu is the one which updates jiffies, then 516 * give up the assignment and let it be taken by the 517 * cpu which runs the tick timer next, which might be 518 * this cpu as well. If we don't drop this here the 519 * jiffies might be stale and do_timer() never 520 * invoked. Keep track of the fact that it was the one 521 * which had the do_timer() duty last. If this cpu is 522 * the one which had the do_timer() duty last, we 523 * limit the sleep time to the timekeeping 524 * max_deferement value which we retrieved 525 * above. Otherwise we can sleep as long as we want. 526 */ 527 if (cpu == tick_do_timer_cpu) { 528 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 529 ts->do_timer_last = 1; 530 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 531 time_delta = KTIME_MAX; 532 ts->do_timer_last = 0; 533 } else if (!ts->do_timer_last) { 534 time_delta = KTIME_MAX; 535 } 536 537 /* 538 * calculate the expiry time for the next timer wheel 539 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals 540 * that there is no timer pending or at least extremely 541 * far into the future (12 days for HZ=1000). In this 542 * case we set the expiry to the end of time. 543 */ 544 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { 545 /* 546 * Calculate the time delta for the next timer event. 547 * If the time delta exceeds the maximum time delta 548 * permitted by the current clocksource then adjust 549 * the time delta accordingly to ensure the 550 * clocksource does not wrap. 551 */ 552 time_delta = min_t(u64, time_delta, 553 tick_period.tv64 * delta_jiffies); 554 } 555 556 if (time_delta < KTIME_MAX) 557 expires = ktime_add_ns(last_update, time_delta); 558 else 559 expires.tv64 = KTIME_MAX; 560 561 /* Skip reprogram of event if its not changed */ 562 if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) 563 goto out; 564 565 ret = expires; 566 567 /* 568 * nohz_stop_sched_tick can be called several times before 569 * the nohz_restart_sched_tick is called. This happens when 570 * interrupts arrive which do not cause a reschedule. In the 571 * first call we save the current tick time, so we can restart 572 * the scheduler tick in nohz_restart_sched_tick. 573 */ 574 if (!ts->tick_stopped) { 575 nohz_balance_enter_idle(cpu); 576 calc_load_enter_idle(); 577 578 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 579 ts->tick_stopped = 1; 580 } 581 582 /* 583 * If the expiration time == KTIME_MAX, then 584 * in this case we simply stop the tick timer. 585 */ 586 if (unlikely(expires.tv64 == KTIME_MAX)) { 587 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 588 hrtimer_cancel(&ts->sched_timer); 589 goto out; 590 } 591 592 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 593 hrtimer_start(&ts->sched_timer, expires, 594 HRTIMER_MODE_ABS_PINNED); 595 /* Check, if the timer was already in the past */ 596 if (hrtimer_active(&ts->sched_timer)) 597 goto out; 598 } else if (!tick_program_event(expires, 0)) 599 goto out; 600 /* 601 * We are past the event already. So we crossed a 602 * jiffie boundary. Update jiffies and raise the 603 * softirq. 604 */ 605 tick_do_update_jiffies64(ktime_get()); 606 } 607 raise_softirq_irqoff(TIMER_SOFTIRQ); 608 out: 609 ts->next_jiffies = next_jiffies; 610 ts->last_jiffies = last_jiffies; 611 ts->sleep_length = ktime_sub(dev->next_event, now); 612 613 return ret; 614 } 615 616 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 617 { 618 /* 619 * If this cpu is offline and it is the one which updates 620 * jiffies, then give up the assignment and let it be taken by 621 * the cpu which runs the tick timer next. If we don't drop 622 * this here the jiffies might be stale and do_timer() never 623 * invoked. 624 */ 625 if (unlikely(!cpu_online(cpu))) { 626 if (cpu == tick_do_timer_cpu) 627 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 628 } 629 630 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) 631 return false; 632 633 if (need_resched()) 634 return false; 635 636 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 637 static int ratelimit; 638 639 if (ratelimit < 10 && 640 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 641 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", 642 (unsigned int) local_softirq_pending()); 643 ratelimit++; 644 } 645 return false; 646 } 647 648 if (have_nohz_full_mask) { 649 /* 650 * Keep the tick alive to guarantee timekeeping progression 651 * if there are full dynticks CPUs around 652 */ 653 if (tick_do_timer_cpu == cpu) 654 return false; 655 /* 656 * Boot safety: make sure the timekeeping duty has been 657 * assigned before entering dyntick-idle mode, 658 */ 659 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 660 return false; 661 } 662 663 return true; 664 } 665 666 static void __tick_nohz_idle_enter(struct tick_sched *ts) 667 { 668 ktime_t now, expires; 669 int cpu = smp_processor_id(); 670 671 now = tick_nohz_start_idle(cpu, ts); 672 673 if (can_stop_idle_tick(cpu, ts)) { 674 int was_stopped = ts->tick_stopped; 675 676 ts->idle_calls++; 677 678 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 679 if (expires.tv64 > 0LL) { 680 ts->idle_sleeps++; 681 ts->idle_expires = expires; 682 } 683 684 if (!was_stopped && ts->tick_stopped) 685 ts->idle_jiffies = ts->last_jiffies; 686 } 687 } 688 689 /** 690 * tick_nohz_idle_enter - stop the idle tick from the idle task 691 * 692 * When the next event is more than a tick into the future, stop the idle tick 693 * Called when we start the idle loop. 694 * 695 * The arch is responsible of calling: 696 * 697 * - rcu_idle_enter() after its last use of RCU before the CPU is put 698 * to sleep. 699 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 700 */ 701 void tick_nohz_idle_enter(void) 702 { 703 struct tick_sched *ts; 704 705 WARN_ON_ONCE(irqs_disabled()); 706 707 /* 708 * Update the idle state in the scheduler domain hierarchy 709 * when tick_nohz_stop_sched_tick() is called from the idle loop. 710 * State will be updated to busy during the first busy tick after 711 * exiting idle. 712 */ 713 set_cpu_sd_state_idle(); 714 715 local_irq_disable(); 716 717 ts = &__get_cpu_var(tick_cpu_sched); 718 /* 719 * set ts->inidle unconditionally. even if the system did not 720 * switch to nohz mode the cpu frequency governers rely on the 721 * update of the idle time accounting in tick_nohz_start_idle(). 722 */ 723 ts->inidle = 1; 724 __tick_nohz_idle_enter(ts); 725 726 local_irq_enable(); 727 } 728 EXPORT_SYMBOL_GPL(tick_nohz_idle_enter); 729 730 /** 731 * tick_nohz_irq_exit - update next tick event from interrupt exit 732 * 733 * When an interrupt fires while we are idle and it doesn't cause 734 * a reschedule, it may still add, modify or delete a timer, enqueue 735 * an RCU callback, etc... 736 * So we need to re-calculate and reprogram the next tick event. 737 */ 738 void tick_nohz_irq_exit(void) 739 { 740 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 741 742 if (!ts->inidle) 743 return; 744 745 /* Cancel the timer because CPU already waken up from the C-states*/ 746 menu_hrtimer_cancel(); 747 __tick_nohz_idle_enter(ts); 748 } 749 750 /** 751 * tick_nohz_get_sleep_length - return the length of the current sleep 752 * 753 * Called from power state control code with interrupts disabled 754 */ 755 ktime_t tick_nohz_get_sleep_length(void) 756 { 757 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 758 759 return ts->sleep_length; 760 } 761 762 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 763 { 764 hrtimer_cancel(&ts->sched_timer); 765 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 766 767 while (1) { 768 /* Forward the time to expire in the future */ 769 hrtimer_forward(&ts->sched_timer, now, tick_period); 770 771 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 772 hrtimer_start_expires(&ts->sched_timer, 773 HRTIMER_MODE_ABS_PINNED); 774 /* Check, if the timer was already in the past */ 775 if (hrtimer_active(&ts->sched_timer)) 776 break; 777 } else { 778 if (!tick_program_event( 779 hrtimer_get_expires(&ts->sched_timer), 0)) 780 break; 781 } 782 /* Reread time and update jiffies */ 783 now = ktime_get(); 784 tick_do_update_jiffies64(now); 785 } 786 } 787 788 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 789 { 790 /* Update jiffies first */ 791 tick_do_update_jiffies64(now); 792 update_cpu_load_nohz(); 793 794 calc_load_exit_idle(); 795 touch_softlockup_watchdog(); 796 /* 797 * Cancel the scheduled timer and restore the tick 798 */ 799 ts->tick_stopped = 0; 800 ts->idle_exittime = now; 801 802 tick_nohz_restart(ts, now); 803 } 804 805 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 806 { 807 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 808 unsigned long ticks; 809 810 if (vtime_accounting_enabled()) 811 return; 812 /* 813 * We stopped the tick in idle. Update process times would miss the 814 * time we slept as update_process_times does only a 1 tick 815 * accounting. Enforce that this is accounted to idle ! 816 */ 817 ticks = jiffies - ts->idle_jiffies; 818 /* 819 * We might be one off. Do not randomly account a huge number of ticks! 820 */ 821 if (ticks && ticks < LONG_MAX) 822 account_idle_ticks(ticks); 823 #endif 824 } 825 826 /** 827 * tick_nohz_idle_exit - restart the idle tick from the idle task 828 * 829 * Restart the idle tick when the CPU is woken up from idle 830 * This also exit the RCU extended quiescent state. The CPU 831 * can use RCU again after this function is called. 832 */ 833 void tick_nohz_idle_exit(void) 834 { 835 int cpu = smp_processor_id(); 836 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 837 ktime_t now; 838 839 local_irq_disable(); 840 841 WARN_ON_ONCE(!ts->inidle); 842 843 ts->inidle = 0; 844 845 /* Cancel the timer because CPU already waken up from the C-states*/ 846 menu_hrtimer_cancel(); 847 if (ts->idle_active || ts->tick_stopped) 848 now = ktime_get(); 849 850 if (ts->idle_active) 851 tick_nohz_stop_idle(cpu, now); 852 853 if (ts->tick_stopped) { 854 tick_nohz_restart_sched_tick(ts, now); 855 tick_nohz_account_idle_ticks(ts); 856 } 857 858 local_irq_enable(); 859 } 860 EXPORT_SYMBOL_GPL(tick_nohz_idle_exit); 861 862 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 863 { 864 hrtimer_forward(&ts->sched_timer, now, tick_period); 865 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); 866 } 867 868 /* 869 * The nohz low res interrupt handler 870 */ 871 static void tick_nohz_handler(struct clock_event_device *dev) 872 { 873 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 874 struct pt_regs *regs = get_irq_regs(); 875 ktime_t now = ktime_get(); 876 877 dev->next_event.tv64 = KTIME_MAX; 878 879 tick_sched_do_timer(now); 880 tick_sched_handle(ts, regs); 881 882 while (tick_nohz_reprogram(ts, now)) { 883 now = ktime_get(); 884 tick_do_update_jiffies64(now); 885 } 886 } 887 888 /** 889 * tick_nohz_switch_to_nohz - switch to nohz mode 890 */ 891 static void tick_nohz_switch_to_nohz(void) 892 { 893 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 894 ktime_t next; 895 896 if (!tick_nohz_enabled) 897 return; 898 899 local_irq_disable(); 900 if (tick_switch_to_oneshot(tick_nohz_handler)) { 901 local_irq_enable(); 902 return; 903 } 904 905 ts->nohz_mode = NOHZ_MODE_LOWRES; 906 907 /* 908 * Recycle the hrtimer in ts, so we can share the 909 * hrtimer_forward with the highres code. 910 */ 911 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 912 /* Get the next period */ 913 next = tick_init_jiffy_update(); 914 915 for (;;) { 916 hrtimer_set_expires(&ts->sched_timer, next); 917 if (!tick_program_event(next, 0)) 918 break; 919 next = ktime_add(next, tick_period); 920 } 921 local_irq_enable(); 922 } 923 924 /* 925 * When NOHZ is enabled and the tick is stopped, we need to kick the 926 * tick timer from irq_enter() so that the jiffies update is kept 927 * alive during long running softirqs. That's ugly as hell, but 928 * correctness is key even if we need to fix the offending softirq in 929 * the first place. 930 * 931 * Note, this is different to tick_nohz_restart. We just kick the 932 * timer and do not touch the other magic bits which need to be done 933 * when idle is left. 934 */ 935 static void tick_nohz_kick_tick(int cpu, ktime_t now) 936 { 937 #if 0 938 /* Switch back to 2.6.27 behaviour */ 939 940 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 941 ktime_t delta; 942 943 /* 944 * Do not touch the tick device, when the next expiry is either 945 * already reached or less/equal than the tick period. 946 */ 947 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 948 if (delta.tv64 <= tick_period.tv64) 949 return; 950 951 tick_nohz_restart(ts, now); 952 #endif 953 } 954 955 static inline void tick_check_nohz(int cpu) 956 { 957 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 958 ktime_t now; 959 960 if (!ts->idle_active && !ts->tick_stopped) 961 return; 962 now = ktime_get(); 963 if (ts->idle_active) 964 tick_nohz_stop_idle(cpu, now); 965 if (ts->tick_stopped) { 966 tick_nohz_update_jiffies(now); 967 tick_nohz_kick_tick(cpu, now); 968 } 969 } 970 971 #else 972 973 static inline void tick_nohz_switch_to_nohz(void) { } 974 static inline void tick_check_nohz(int cpu) { } 975 976 #endif /* CONFIG_NO_HZ_COMMON */ 977 978 /* 979 * Called from irq_enter to notify about the possible interruption of idle() 980 */ 981 void tick_check_idle(int cpu) 982 { 983 tick_check_oneshot_broadcast(cpu); 984 tick_check_nohz(cpu); 985 } 986 987 /* 988 * High resolution timer specific code 989 */ 990 #ifdef CONFIG_HIGH_RES_TIMERS 991 /* 992 * We rearm the timer until we get disabled by the idle code. 993 * Called with interrupts disabled. 994 */ 995 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 996 { 997 struct tick_sched *ts = 998 container_of(timer, struct tick_sched, sched_timer); 999 struct pt_regs *regs = get_irq_regs(); 1000 ktime_t now = ktime_get(); 1001 1002 tick_sched_do_timer(now); 1003 1004 /* 1005 * Do not call, when we are not in irq context and have 1006 * no valid regs pointer 1007 */ 1008 if (regs) 1009 tick_sched_handle(ts, regs); 1010 1011 hrtimer_forward(timer, now, tick_period); 1012 1013 return HRTIMER_RESTART; 1014 } 1015 1016 static int sched_skew_tick; 1017 1018 static int __init skew_tick(char *str) 1019 { 1020 get_option(&str, &sched_skew_tick); 1021 1022 return 0; 1023 } 1024 early_param("skew_tick", skew_tick); 1025 1026 /** 1027 * tick_setup_sched_timer - setup the tick emulation timer 1028 */ 1029 void tick_setup_sched_timer(void) 1030 { 1031 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 1032 ktime_t now = ktime_get(); 1033 1034 /* 1035 * Emulate tick processing via per-CPU hrtimers: 1036 */ 1037 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1038 ts->sched_timer.function = tick_sched_timer; 1039 1040 /* Get the next period (per cpu) */ 1041 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1042 1043 /* Offset the tick to avert jiffies_lock contention. */ 1044 if (sched_skew_tick) { 1045 u64 offset = ktime_to_ns(tick_period) >> 1; 1046 do_div(offset, num_possible_cpus()); 1047 offset *= smp_processor_id(); 1048 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1049 } 1050 1051 for (;;) { 1052 hrtimer_forward(&ts->sched_timer, now, tick_period); 1053 hrtimer_start_expires(&ts->sched_timer, 1054 HRTIMER_MODE_ABS_PINNED); 1055 /* Check, if the timer was already in the past */ 1056 if (hrtimer_active(&ts->sched_timer)) 1057 break; 1058 now = ktime_get(); 1059 } 1060 1061 #ifdef CONFIG_NO_HZ_COMMON 1062 if (tick_nohz_enabled) 1063 ts->nohz_mode = NOHZ_MODE_HIGHRES; 1064 #endif 1065 } 1066 #endif /* HIGH_RES_TIMERS */ 1067 1068 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1069 void tick_cancel_sched_timer(int cpu) 1070 { 1071 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1072 1073 # ifdef CONFIG_HIGH_RES_TIMERS 1074 if (ts->sched_timer.base) 1075 hrtimer_cancel(&ts->sched_timer); 1076 # endif 1077 1078 ts->nohz_mode = NOHZ_MODE_INACTIVE; 1079 } 1080 #endif 1081 1082 /** 1083 * Async notification about clocksource changes 1084 */ 1085 void tick_clock_notify(void) 1086 { 1087 int cpu; 1088 1089 for_each_possible_cpu(cpu) 1090 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1091 } 1092 1093 /* 1094 * Async notification about clock event changes 1095 */ 1096 void tick_oneshot_notify(void) 1097 { 1098 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 1099 1100 set_bit(0, &ts->check_clocks); 1101 } 1102 1103 /** 1104 * Check, if a change happened, which makes oneshot possible. 1105 * 1106 * Called cyclic from the hrtimer softirq (driven by the timer 1107 * softirq) allow_nohz signals, that we can switch into low-res nohz 1108 * mode, because high resolution timers are disabled (either compile 1109 * or runtime). 1110 */ 1111 int tick_check_oneshot_change(int allow_nohz) 1112 { 1113 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 1114 1115 if (!test_and_clear_bit(0, &ts->check_clocks)) 1116 return 0; 1117 1118 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1119 return 0; 1120 1121 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1122 return 0; 1123 1124 if (!allow_nohz) 1125 return 1; 1126 1127 tick_nohz_switch_to_nohz(); 1128 return 0; 1129 } 1130