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