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 ktime_t now, int cpu) 276 { 277 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; 278 ktime_t last_update, expires; 279 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; 280 u64 time_delta; 281 282 283 /* Read jiffies and the time when jiffies were updated last */ 284 do { 285 seq = read_seqbegin(&xtime_lock); 286 last_update = last_jiffies_update; 287 last_jiffies = jiffies; 288 time_delta = timekeeping_max_deferment(); 289 } while (read_seqretry(&xtime_lock, seq)); 290 291 if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) || 292 arch_needs_cpu(cpu)) { 293 next_jiffies = last_jiffies + 1; 294 delta_jiffies = 1; 295 } else { 296 /* Get the next timer wheel timer */ 297 next_jiffies = get_next_timer_interrupt(last_jiffies); 298 delta_jiffies = next_jiffies - last_jiffies; 299 } 300 /* 301 * Do not stop the tick, if we are only one off 302 * or if the cpu is required for rcu 303 */ 304 if (!ts->tick_stopped && delta_jiffies == 1) 305 goto out; 306 307 /* Schedule the tick, if we are at least one jiffie off */ 308 if ((long)delta_jiffies >= 1) { 309 310 /* 311 * If this cpu is the one which updates jiffies, then 312 * give up the assignment and let it be taken by the 313 * cpu which runs the tick timer next, which might be 314 * this cpu as well. If we don't drop this here the 315 * jiffies might be stale and do_timer() never 316 * invoked. Keep track of the fact that it was the one 317 * which had the do_timer() duty last. If this cpu is 318 * the one which had the do_timer() duty last, we 319 * limit the sleep time to the timekeeping 320 * max_deferement value which we retrieved 321 * above. Otherwise we can sleep as long as we want. 322 */ 323 if (cpu == tick_do_timer_cpu) { 324 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 325 ts->do_timer_last = 1; 326 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 327 time_delta = KTIME_MAX; 328 ts->do_timer_last = 0; 329 } else if (!ts->do_timer_last) { 330 time_delta = KTIME_MAX; 331 } 332 333 /* 334 * calculate the expiry time for the next timer wheel 335 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals 336 * that there is no timer pending or at least extremely 337 * far into the future (12 days for HZ=1000). In this 338 * case we set the expiry to the end of time. 339 */ 340 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { 341 /* 342 * Calculate the time delta for the next timer event. 343 * If the time delta exceeds the maximum time delta 344 * permitted by the current clocksource then adjust 345 * the time delta accordingly to ensure the 346 * clocksource does not wrap. 347 */ 348 time_delta = min_t(u64, time_delta, 349 tick_period.tv64 * delta_jiffies); 350 } 351 352 if (time_delta < KTIME_MAX) 353 expires = ktime_add_ns(last_update, time_delta); 354 else 355 expires.tv64 = KTIME_MAX; 356 357 /* Skip reprogram of event if its not changed */ 358 if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) 359 goto out; 360 361 /* 362 * nohz_stop_sched_tick can be called several times before 363 * the nohz_restart_sched_tick is called. This happens when 364 * interrupts arrive which do not cause a reschedule. In the 365 * first call we save the current tick time, so we can restart 366 * the scheduler tick in nohz_restart_sched_tick. 367 */ 368 if (!ts->tick_stopped) { 369 select_nohz_load_balancer(1); 370 371 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 372 ts->tick_stopped = 1; 373 } 374 375 ts->idle_sleeps++; 376 377 /* Mark expires */ 378 ts->idle_expires = expires; 379 380 /* 381 * If the expiration time == KTIME_MAX, then 382 * in this case we simply stop the tick timer. 383 */ 384 if (unlikely(expires.tv64 == KTIME_MAX)) { 385 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 386 hrtimer_cancel(&ts->sched_timer); 387 goto out; 388 } 389 390 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 391 hrtimer_start(&ts->sched_timer, expires, 392 HRTIMER_MODE_ABS_PINNED); 393 /* Check, if the timer was already in the past */ 394 if (hrtimer_active(&ts->sched_timer)) 395 goto out; 396 } else if (!tick_program_event(expires, 0)) 397 goto out; 398 /* 399 * We are past the event already. So we crossed a 400 * jiffie boundary. Update jiffies and raise the 401 * softirq. 402 */ 403 tick_do_update_jiffies64(ktime_get()); 404 } 405 raise_softirq_irqoff(TIMER_SOFTIRQ); 406 out: 407 ts->next_jiffies = next_jiffies; 408 ts->last_jiffies = last_jiffies; 409 ts->sleep_length = ktime_sub(dev->next_event, now); 410 } 411 412 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 413 { 414 /* 415 * If this cpu is offline and it is the one which updates 416 * jiffies, then give up the assignment and let it be taken by 417 * the cpu which runs the tick timer next. If we don't drop 418 * this here the jiffies might be stale and do_timer() never 419 * invoked. 420 */ 421 if (unlikely(!cpu_online(cpu))) { 422 if (cpu == tick_do_timer_cpu) 423 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 424 } 425 426 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) 427 return false; 428 429 if (need_resched()) 430 return false; 431 432 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 433 static int ratelimit; 434 435 if (ratelimit < 10) { 436 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", 437 (unsigned int) local_softirq_pending()); 438 ratelimit++; 439 } 440 return false; 441 } 442 443 return true; 444 } 445 446 static void __tick_nohz_idle_enter(struct tick_sched *ts) 447 { 448 ktime_t now; 449 int cpu = smp_processor_id(); 450 451 now = tick_nohz_start_idle(cpu, ts); 452 453 if (can_stop_idle_tick(cpu, ts)) { 454 int was_stopped = ts->tick_stopped; 455 456 ts->idle_calls++; 457 tick_nohz_stop_sched_tick(ts, now, cpu); 458 459 if (!was_stopped && ts->tick_stopped) 460 ts->idle_jiffies = ts->last_jiffies; 461 } 462 } 463 464 /** 465 * tick_nohz_idle_enter - stop the idle tick from the idle task 466 * 467 * When the next event is more than a tick into the future, stop the idle tick 468 * Called when we start the idle loop. 469 * 470 * The arch is responsible of calling: 471 * 472 * - rcu_idle_enter() after its last use of RCU before the CPU is put 473 * to sleep. 474 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 475 */ 476 void tick_nohz_idle_enter(void) 477 { 478 struct tick_sched *ts; 479 480 WARN_ON_ONCE(irqs_disabled()); 481 482 /* 483 * Update the idle state in the scheduler domain hierarchy 484 * when tick_nohz_stop_sched_tick() is called from the idle loop. 485 * State will be updated to busy during the first busy tick after 486 * exiting idle. 487 */ 488 set_cpu_sd_state_idle(); 489 490 local_irq_disable(); 491 492 ts = &__get_cpu_var(tick_cpu_sched); 493 /* 494 * set ts->inidle unconditionally. even if the system did not 495 * switch to nohz mode the cpu frequency governers rely on the 496 * update of the idle time accounting in tick_nohz_start_idle(). 497 */ 498 ts->inidle = 1; 499 __tick_nohz_idle_enter(ts); 500 501 local_irq_enable(); 502 } 503 504 /** 505 * tick_nohz_irq_exit - update next tick event from interrupt exit 506 * 507 * When an interrupt fires while we are idle and it doesn't cause 508 * a reschedule, it may still add, modify or delete a timer, enqueue 509 * an RCU callback, etc... 510 * So we need to re-calculate and reprogram the next tick event. 511 */ 512 void tick_nohz_irq_exit(void) 513 { 514 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 515 516 if (!ts->inidle) 517 return; 518 519 __tick_nohz_idle_enter(ts); 520 } 521 522 /** 523 * tick_nohz_get_sleep_length - return the length of the current sleep 524 * 525 * Called from power state control code with interrupts disabled 526 */ 527 ktime_t tick_nohz_get_sleep_length(void) 528 { 529 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 530 531 return ts->sleep_length; 532 } 533 534 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 535 { 536 hrtimer_cancel(&ts->sched_timer); 537 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 538 539 while (1) { 540 /* Forward the time to expire in the future */ 541 hrtimer_forward(&ts->sched_timer, now, tick_period); 542 543 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 544 hrtimer_start_expires(&ts->sched_timer, 545 HRTIMER_MODE_ABS_PINNED); 546 /* Check, if the timer was already in the past */ 547 if (hrtimer_active(&ts->sched_timer)) 548 break; 549 } else { 550 if (!tick_program_event( 551 hrtimer_get_expires(&ts->sched_timer), 0)) 552 break; 553 } 554 /* Reread time and update jiffies */ 555 now = ktime_get(); 556 tick_do_update_jiffies64(now); 557 } 558 } 559 560 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 561 { 562 /* Update jiffies first */ 563 select_nohz_load_balancer(0); 564 tick_do_update_jiffies64(now); 565 update_cpu_load_nohz(); 566 567 touch_softlockup_watchdog(); 568 /* 569 * Cancel the scheduled timer and restore the tick 570 */ 571 ts->tick_stopped = 0; 572 ts->idle_exittime = now; 573 574 tick_nohz_restart(ts, now); 575 } 576 577 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 578 { 579 #ifndef CONFIG_VIRT_CPU_ACCOUNTING 580 unsigned long ticks; 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 } 594 595 /** 596 * tick_nohz_idle_exit - restart the idle tick from the idle task 597 * 598 * Restart the idle tick when the CPU is woken up from idle 599 * This also exit the RCU extended quiescent state. The CPU 600 * can use RCU again after this function is called. 601 */ 602 void tick_nohz_idle_exit(void) 603 { 604 int cpu = smp_processor_id(); 605 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 606 ktime_t now; 607 608 local_irq_disable(); 609 610 WARN_ON_ONCE(!ts->inidle); 611 612 ts->inidle = 0; 613 614 if (ts->idle_active || ts->tick_stopped) 615 now = ktime_get(); 616 617 if (ts->idle_active) 618 tick_nohz_stop_idle(cpu, now); 619 620 if (ts->tick_stopped) { 621 tick_nohz_restart_sched_tick(ts, now); 622 tick_nohz_account_idle_ticks(ts); 623 } 624 625 local_irq_enable(); 626 } 627 628 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 629 { 630 hrtimer_forward(&ts->sched_timer, now, tick_period); 631 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); 632 } 633 634 /* 635 * The nohz low res interrupt handler 636 */ 637 static void tick_nohz_handler(struct clock_event_device *dev) 638 { 639 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 640 struct pt_regs *regs = get_irq_regs(); 641 int cpu = smp_processor_id(); 642 ktime_t now = ktime_get(); 643 644 dev->next_event.tv64 = KTIME_MAX; 645 646 /* 647 * Check if the do_timer duty was dropped. We don't care about 648 * concurrency: This happens only when the cpu in charge went 649 * into a long sleep. If two cpus happen to assign themself to 650 * this duty, then the jiffies update is still serialized by 651 * xtime_lock. 652 */ 653 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) 654 tick_do_timer_cpu = cpu; 655 656 /* Check, if the jiffies need an update */ 657 if (tick_do_timer_cpu == cpu) 658 tick_do_update_jiffies64(now); 659 660 /* 661 * When we are idle and the tick is stopped, we have to touch 662 * the watchdog as we might not schedule for a really long 663 * time. This happens on complete idle SMP systems while 664 * waiting on the login prompt. We also increment the "start 665 * of idle" jiffy stamp so the idle accounting adjustment we 666 * do when we go busy again does not account too much ticks. 667 */ 668 if (ts->tick_stopped) { 669 touch_softlockup_watchdog(); 670 ts->idle_jiffies++; 671 } 672 673 update_process_times(user_mode(regs)); 674 profile_tick(CPU_PROFILING); 675 676 while (tick_nohz_reprogram(ts, now)) { 677 now = ktime_get(); 678 tick_do_update_jiffies64(now); 679 } 680 } 681 682 /** 683 * tick_nohz_switch_to_nohz - switch to nohz mode 684 */ 685 static void tick_nohz_switch_to_nohz(void) 686 { 687 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 688 ktime_t next; 689 690 if (!tick_nohz_enabled) 691 return; 692 693 local_irq_disable(); 694 if (tick_switch_to_oneshot(tick_nohz_handler)) { 695 local_irq_enable(); 696 return; 697 } 698 699 ts->nohz_mode = NOHZ_MODE_LOWRES; 700 701 /* 702 * Recycle the hrtimer in ts, so we can share the 703 * hrtimer_forward with the highres code. 704 */ 705 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 706 /* Get the next period */ 707 next = tick_init_jiffy_update(); 708 709 for (;;) { 710 hrtimer_set_expires(&ts->sched_timer, next); 711 if (!tick_program_event(next, 0)) 712 break; 713 next = ktime_add(next, tick_period); 714 } 715 local_irq_enable(); 716 } 717 718 /* 719 * When NOHZ is enabled and the tick is stopped, we need to kick the 720 * tick timer from irq_enter() so that the jiffies update is kept 721 * alive during long running softirqs. That's ugly as hell, but 722 * correctness is key even if we need to fix the offending softirq in 723 * the first place. 724 * 725 * Note, this is different to tick_nohz_restart. We just kick the 726 * timer and do not touch the other magic bits which need to be done 727 * when idle is left. 728 */ 729 static void tick_nohz_kick_tick(int cpu, ktime_t now) 730 { 731 #if 0 732 /* Switch back to 2.6.27 behaviour */ 733 734 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 735 ktime_t delta; 736 737 /* 738 * Do not touch the tick device, when the next expiry is either 739 * already reached or less/equal than the tick period. 740 */ 741 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 742 if (delta.tv64 <= tick_period.tv64) 743 return; 744 745 tick_nohz_restart(ts, now); 746 #endif 747 } 748 749 static inline void tick_check_nohz(int cpu) 750 { 751 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 752 ktime_t now; 753 754 if (!ts->idle_active && !ts->tick_stopped) 755 return; 756 now = ktime_get(); 757 if (ts->idle_active) 758 tick_nohz_stop_idle(cpu, now); 759 if (ts->tick_stopped) { 760 tick_nohz_update_jiffies(now); 761 tick_nohz_kick_tick(cpu, now); 762 } 763 } 764 765 #else 766 767 static inline void tick_nohz_switch_to_nohz(void) { } 768 static inline void tick_check_nohz(int cpu) { } 769 770 #endif /* NO_HZ */ 771 772 /* 773 * Called from irq_enter to notify about the possible interruption of idle() 774 */ 775 void tick_check_idle(int cpu) 776 { 777 tick_check_oneshot_broadcast(cpu); 778 tick_check_nohz(cpu); 779 } 780 781 /* 782 * High resolution timer specific code 783 */ 784 #ifdef CONFIG_HIGH_RES_TIMERS 785 /* 786 * We rearm the timer until we get disabled by the idle code. 787 * Called with interrupts disabled and timer->base->cpu_base->lock held. 788 */ 789 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 790 { 791 struct tick_sched *ts = 792 container_of(timer, struct tick_sched, sched_timer); 793 struct pt_regs *regs = get_irq_regs(); 794 ktime_t now = ktime_get(); 795 int cpu = smp_processor_id(); 796 797 #ifdef CONFIG_NO_HZ 798 /* 799 * Check if the do_timer duty was dropped. We don't care about 800 * concurrency: This happens only when the cpu in charge went 801 * into a long sleep. If two cpus happen to assign themself to 802 * this duty, then the jiffies update is still serialized by 803 * xtime_lock. 804 */ 805 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) 806 tick_do_timer_cpu = cpu; 807 #endif 808 809 /* Check, if the jiffies need an update */ 810 if (tick_do_timer_cpu == cpu) 811 tick_do_update_jiffies64(now); 812 813 /* 814 * Do not call, when we are not in irq context and have 815 * no valid regs pointer 816 */ 817 if (regs) { 818 /* 819 * When we are idle and the tick is stopped, we have to touch 820 * the watchdog as we might not schedule for a really long 821 * time. This happens on complete idle SMP systems while 822 * waiting on the login prompt. We also increment the "start of 823 * idle" jiffy stamp so the idle accounting adjustment we do 824 * when we go busy again does not account too much ticks. 825 */ 826 if (ts->tick_stopped) { 827 touch_softlockup_watchdog(); 828 if (idle_cpu(cpu)) 829 ts->idle_jiffies++; 830 } 831 update_process_times(user_mode(regs)); 832 profile_tick(CPU_PROFILING); 833 } 834 835 hrtimer_forward(timer, now, tick_period); 836 837 return HRTIMER_RESTART; 838 } 839 840 static int sched_skew_tick; 841 842 static int __init skew_tick(char *str) 843 { 844 get_option(&str, &sched_skew_tick); 845 846 return 0; 847 } 848 early_param("skew_tick", skew_tick); 849 850 /** 851 * tick_setup_sched_timer - setup the tick emulation timer 852 */ 853 void tick_setup_sched_timer(void) 854 { 855 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 856 ktime_t now = ktime_get(); 857 858 /* 859 * Emulate tick processing via per-CPU hrtimers: 860 */ 861 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 862 ts->sched_timer.function = tick_sched_timer; 863 864 /* Get the next period (per cpu) */ 865 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 866 867 /* Offset the tick to avert xtime_lock contention. */ 868 if (sched_skew_tick) { 869 u64 offset = ktime_to_ns(tick_period) >> 1; 870 do_div(offset, num_possible_cpus()); 871 offset *= smp_processor_id(); 872 hrtimer_add_expires_ns(&ts->sched_timer, offset); 873 } 874 875 for (;;) { 876 hrtimer_forward(&ts->sched_timer, now, tick_period); 877 hrtimer_start_expires(&ts->sched_timer, 878 HRTIMER_MODE_ABS_PINNED); 879 /* Check, if the timer was already in the past */ 880 if (hrtimer_active(&ts->sched_timer)) 881 break; 882 now = ktime_get(); 883 } 884 885 #ifdef CONFIG_NO_HZ 886 if (tick_nohz_enabled) 887 ts->nohz_mode = NOHZ_MODE_HIGHRES; 888 #endif 889 } 890 #endif /* HIGH_RES_TIMERS */ 891 892 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS 893 void tick_cancel_sched_timer(int cpu) 894 { 895 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 896 897 # ifdef CONFIG_HIGH_RES_TIMERS 898 if (ts->sched_timer.base) 899 hrtimer_cancel(&ts->sched_timer); 900 # endif 901 902 ts->nohz_mode = NOHZ_MODE_INACTIVE; 903 } 904 #endif 905 906 /** 907 * Async notification about clocksource changes 908 */ 909 void tick_clock_notify(void) 910 { 911 int cpu; 912 913 for_each_possible_cpu(cpu) 914 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 915 } 916 917 /* 918 * Async notification about clock event changes 919 */ 920 void tick_oneshot_notify(void) 921 { 922 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 923 924 set_bit(0, &ts->check_clocks); 925 } 926 927 /** 928 * Check, if a change happened, which makes oneshot possible. 929 * 930 * Called cyclic from the hrtimer softirq (driven by the timer 931 * softirq) allow_nohz signals, that we can switch into low-res nohz 932 * mode, because high resolution timers are disabled (either compile 933 * or runtime). 934 */ 935 int tick_check_oneshot_change(int allow_nohz) 936 { 937 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 938 939 if (!test_and_clear_bit(0, &ts->check_clocks)) 940 return 0; 941 942 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 943 return 0; 944 945 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 946 return 0; 947 948 if (!allow_nohz) 949 return 1; 950 951 tick_nohz_switch_to_nohz(); 952 return 0; 953 } 954