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 static void tick_sched_do_timer(ktime_t now) 103 { 104 int cpu = smp_processor_id(); 105 106 #ifdef CONFIG_NO_HZ 107 /* 108 * Check if the do_timer duty was dropped. We don't care about 109 * concurrency: This happens only when the cpu in charge went 110 * into a long sleep. If two cpus happen to assign themself to 111 * this duty, then the jiffies update is still serialized by 112 * xtime_lock. 113 */ 114 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) 115 tick_do_timer_cpu = cpu; 116 #endif 117 118 /* Check, if the jiffies need an update */ 119 if (tick_do_timer_cpu == cpu) 120 tick_do_update_jiffies64(now); 121 } 122 123 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) 124 { 125 /* 126 * When we are idle and the tick is stopped, we have to touch 127 * the watchdog as we might not schedule for a really long 128 * time. This happens on complete idle SMP systems while 129 * waiting on the login prompt. We also increment the "start of 130 * idle" jiffy stamp so the idle accounting adjustment we do 131 * when we go busy again does not account too much ticks. 132 */ 133 if (ts->tick_stopped) { 134 touch_softlockup_watchdog(); 135 if (is_idle_task(current)) 136 ts->idle_jiffies++; 137 } 138 update_process_times(user_mode(regs)); 139 profile_tick(CPU_PROFILING); 140 } 141 142 /* 143 * NOHZ - aka dynamic tick functionality 144 */ 145 #ifdef CONFIG_NO_HZ 146 /* 147 * NO HZ enabled ? 148 */ 149 int tick_nohz_enabled __read_mostly = 1; 150 151 /* 152 * Enable / Disable tickless mode 153 */ 154 static int __init setup_tick_nohz(char *str) 155 { 156 if (!strcmp(str, "off")) 157 tick_nohz_enabled = 0; 158 else if (!strcmp(str, "on")) 159 tick_nohz_enabled = 1; 160 else 161 return 0; 162 return 1; 163 } 164 165 __setup("nohz=", setup_tick_nohz); 166 167 /** 168 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 169 * 170 * Called from interrupt entry when the CPU was idle 171 * 172 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 173 * must be updated. Otherwise an interrupt handler could use a stale jiffy 174 * value. We do this unconditionally on any cpu, as we don't know whether the 175 * cpu, which has the update task assigned is in a long sleep. 176 */ 177 static void tick_nohz_update_jiffies(ktime_t now) 178 { 179 int cpu = smp_processor_id(); 180 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 181 unsigned long flags; 182 183 ts->idle_waketime = now; 184 185 local_irq_save(flags); 186 tick_do_update_jiffies64(now); 187 local_irq_restore(flags); 188 189 touch_softlockup_watchdog(); 190 } 191 192 /* 193 * Updates the per cpu time idle statistics counters 194 */ 195 static void 196 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 197 { 198 ktime_t delta; 199 200 if (ts->idle_active) { 201 delta = ktime_sub(now, ts->idle_entrytime); 202 if (nr_iowait_cpu(cpu) > 0) 203 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 204 else 205 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 206 ts->idle_entrytime = now; 207 } 208 209 if (last_update_time) 210 *last_update_time = ktime_to_us(now); 211 212 } 213 214 static void tick_nohz_stop_idle(int cpu, ktime_t now) 215 { 216 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 217 218 update_ts_time_stats(cpu, ts, now, NULL); 219 ts->idle_active = 0; 220 221 sched_clock_idle_wakeup_event(0); 222 } 223 224 static ktime_t tick_nohz_start_idle(int cpu, struct tick_sched *ts) 225 { 226 ktime_t now = ktime_get(); 227 228 ts->idle_entrytime = now; 229 ts->idle_active = 1; 230 sched_clock_idle_sleep_event(); 231 return now; 232 } 233 234 /** 235 * get_cpu_idle_time_us - get the total idle 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 idle 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_idle_time_us(int cpu, u64 *last_update_time) 249 { 250 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 251 ktime_t now, idle; 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 idle = ts->idle_sleeptime; 260 } else { 261 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 262 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 263 264 idle = ktime_add(ts->idle_sleeptime, delta); 265 } else { 266 idle = ts->idle_sleeptime; 267 } 268 } 269 270 return ktime_to_us(idle); 271 272 } 273 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 274 275 /** 276 * get_cpu_iowait_time_us - get the total iowait time of a cpu 277 * @cpu: CPU number to query 278 * @last_update_time: variable to store update time in. Do not update 279 * counters if NULL. 280 * 281 * Return the cummulative iowait time (since boot) for a given 282 * CPU, in microseconds. 283 * 284 * This time is measured via accounting rather than sampling, 285 * and is as accurate as ktime_get() is. 286 * 287 * This function returns -1 if NOHZ is not enabled. 288 */ 289 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 290 { 291 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 292 ktime_t now, iowait; 293 294 if (!tick_nohz_enabled) 295 return -1; 296 297 now = ktime_get(); 298 if (last_update_time) { 299 update_ts_time_stats(cpu, ts, now, last_update_time); 300 iowait = ts->iowait_sleeptime; 301 } else { 302 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 303 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 304 305 iowait = ktime_add(ts->iowait_sleeptime, delta); 306 } else { 307 iowait = ts->iowait_sleeptime; 308 } 309 } 310 311 return ktime_to_us(iowait); 312 } 313 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 314 315 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 316 ktime_t now, int cpu) 317 { 318 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; 319 ktime_t last_update, expires, ret = { .tv64 = 0 }; 320 unsigned long rcu_delta_jiffies; 321 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; 322 u64 time_delta; 323 324 /* Read jiffies and the time when jiffies were updated last */ 325 do { 326 seq = read_seqbegin(&xtime_lock); 327 last_update = last_jiffies_update; 328 last_jiffies = jiffies; 329 time_delta = timekeeping_max_deferment(); 330 } while (read_seqretry(&xtime_lock, seq)); 331 332 if (rcu_needs_cpu(cpu, &rcu_delta_jiffies) || printk_needs_cpu(cpu) || 333 arch_needs_cpu(cpu)) { 334 next_jiffies = last_jiffies + 1; 335 delta_jiffies = 1; 336 } else { 337 /* Get the next timer wheel timer */ 338 next_jiffies = get_next_timer_interrupt(last_jiffies); 339 delta_jiffies = next_jiffies - last_jiffies; 340 if (rcu_delta_jiffies < delta_jiffies) { 341 next_jiffies = last_jiffies + rcu_delta_jiffies; 342 delta_jiffies = rcu_delta_jiffies; 343 } 344 } 345 /* 346 * Do not stop the tick, if we are only one off 347 * or if the cpu is required for rcu 348 */ 349 if (!ts->tick_stopped && delta_jiffies == 1) 350 goto out; 351 352 /* Schedule the tick, if we are at least one jiffie off */ 353 if ((long)delta_jiffies >= 1) { 354 355 /* 356 * If this cpu is the one which updates jiffies, then 357 * give up the assignment and let it be taken by the 358 * cpu which runs the tick timer next, which might be 359 * this cpu as well. If we don't drop this here the 360 * jiffies might be stale and do_timer() never 361 * invoked. Keep track of the fact that it was the one 362 * which had the do_timer() duty last. If this cpu is 363 * the one which had the do_timer() duty last, we 364 * limit the sleep time to the timekeeping 365 * max_deferement value which we retrieved 366 * above. Otherwise we can sleep as long as we want. 367 */ 368 if (cpu == tick_do_timer_cpu) { 369 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 370 ts->do_timer_last = 1; 371 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 372 time_delta = KTIME_MAX; 373 ts->do_timer_last = 0; 374 } else if (!ts->do_timer_last) { 375 time_delta = KTIME_MAX; 376 } 377 378 /* 379 * calculate the expiry time for the next timer wheel 380 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals 381 * that there is no timer pending or at least extremely 382 * far into the future (12 days for HZ=1000). In this 383 * case we set the expiry to the end of time. 384 */ 385 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { 386 /* 387 * Calculate the time delta for the next timer event. 388 * If the time delta exceeds the maximum time delta 389 * permitted by the current clocksource then adjust 390 * the time delta accordingly to ensure the 391 * clocksource does not wrap. 392 */ 393 time_delta = min_t(u64, time_delta, 394 tick_period.tv64 * delta_jiffies); 395 } 396 397 if (time_delta < KTIME_MAX) 398 expires = ktime_add_ns(last_update, time_delta); 399 else 400 expires.tv64 = KTIME_MAX; 401 402 /* Skip reprogram of event if its not changed */ 403 if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) 404 goto out; 405 406 ret = expires; 407 408 /* 409 * nohz_stop_sched_tick can be called several times before 410 * the nohz_restart_sched_tick is called. This happens when 411 * interrupts arrive which do not cause a reschedule. In the 412 * first call we save the current tick time, so we can restart 413 * the scheduler tick in nohz_restart_sched_tick. 414 */ 415 if (!ts->tick_stopped) { 416 nohz_balance_enter_idle(cpu); 417 calc_load_enter_idle(); 418 419 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 420 ts->tick_stopped = 1; 421 } 422 423 /* 424 * If the expiration time == KTIME_MAX, then 425 * in this case we simply stop the tick timer. 426 */ 427 if (unlikely(expires.tv64 == KTIME_MAX)) { 428 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 429 hrtimer_cancel(&ts->sched_timer); 430 goto out; 431 } 432 433 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 434 hrtimer_start(&ts->sched_timer, expires, 435 HRTIMER_MODE_ABS_PINNED); 436 /* Check, if the timer was already in the past */ 437 if (hrtimer_active(&ts->sched_timer)) 438 goto out; 439 } else if (!tick_program_event(expires, 0)) 440 goto out; 441 /* 442 * We are past the event already. So we crossed a 443 * jiffie boundary. Update jiffies and raise the 444 * softirq. 445 */ 446 tick_do_update_jiffies64(ktime_get()); 447 } 448 raise_softirq_irqoff(TIMER_SOFTIRQ); 449 out: 450 ts->next_jiffies = next_jiffies; 451 ts->last_jiffies = last_jiffies; 452 ts->sleep_length = ktime_sub(dev->next_event, now); 453 454 return ret; 455 } 456 457 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 458 { 459 /* 460 * If this cpu is offline and it is the one which updates 461 * jiffies, then give up the assignment and let it be taken by 462 * the cpu which runs the tick timer next. If we don't drop 463 * this here the jiffies might be stale and do_timer() never 464 * invoked. 465 */ 466 if (unlikely(!cpu_online(cpu))) { 467 if (cpu == tick_do_timer_cpu) 468 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 469 } 470 471 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) 472 return false; 473 474 if (need_resched()) 475 return false; 476 477 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 478 static int ratelimit; 479 480 if (ratelimit < 10 && 481 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 482 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", 483 (unsigned int) local_softirq_pending()); 484 ratelimit++; 485 } 486 return false; 487 } 488 489 return true; 490 } 491 492 static void __tick_nohz_idle_enter(struct tick_sched *ts) 493 { 494 ktime_t now, expires; 495 int cpu = smp_processor_id(); 496 497 now = tick_nohz_start_idle(cpu, ts); 498 499 if (can_stop_idle_tick(cpu, ts)) { 500 int was_stopped = ts->tick_stopped; 501 502 ts->idle_calls++; 503 504 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 505 if (expires.tv64 > 0LL) { 506 ts->idle_sleeps++; 507 ts->idle_expires = expires; 508 } 509 510 if (!was_stopped && ts->tick_stopped) 511 ts->idle_jiffies = ts->last_jiffies; 512 } 513 } 514 515 /** 516 * tick_nohz_idle_enter - stop the idle tick from the idle task 517 * 518 * When the next event is more than a tick into the future, stop the idle tick 519 * Called when we start the idle loop. 520 * 521 * The arch is responsible of calling: 522 * 523 * - rcu_idle_enter() after its last use of RCU before the CPU is put 524 * to sleep. 525 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 526 */ 527 void tick_nohz_idle_enter(void) 528 { 529 struct tick_sched *ts; 530 531 WARN_ON_ONCE(irqs_disabled()); 532 533 /* 534 * Update the idle state in the scheduler domain hierarchy 535 * when tick_nohz_stop_sched_tick() is called from the idle loop. 536 * State will be updated to busy during the first busy tick after 537 * exiting idle. 538 */ 539 set_cpu_sd_state_idle(); 540 541 local_irq_disable(); 542 543 ts = &__get_cpu_var(tick_cpu_sched); 544 /* 545 * set ts->inidle unconditionally. even if the system did not 546 * switch to nohz mode the cpu frequency governers rely on the 547 * update of the idle time accounting in tick_nohz_start_idle(). 548 */ 549 ts->inidle = 1; 550 __tick_nohz_idle_enter(ts); 551 552 local_irq_enable(); 553 } 554 555 /** 556 * tick_nohz_irq_exit - update next tick event from interrupt exit 557 * 558 * When an interrupt fires while we are idle and it doesn't cause 559 * a reschedule, it may still add, modify or delete a timer, enqueue 560 * an RCU callback, etc... 561 * So we need to re-calculate and reprogram the next tick event. 562 */ 563 void tick_nohz_irq_exit(void) 564 { 565 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 566 567 if (!ts->inidle) 568 return; 569 570 __tick_nohz_idle_enter(ts); 571 } 572 573 /** 574 * tick_nohz_get_sleep_length - return the length of the current sleep 575 * 576 * Called from power state control code with interrupts disabled 577 */ 578 ktime_t tick_nohz_get_sleep_length(void) 579 { 580 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 581 582 return ts->sleep_length; 583 } 584 585 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 586 { 587 hrtimer_cancel(&ts->sched_timer); 588 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 589 590 while (1) { 591 /* Forward the time to expire in the future */ 592 hrtimer_forward(&ts->sched_timer, now, tick_period); 593 594 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 595 hrtimer_start_expires(&ts->sched_timer, 596 HRTIMER_MODE_ABS_PINNED); 597 /* Check, if the timer was already in the past */ 598 if (hrtimer_active(&ts->sched_timer)) 599 break; 600 } else { 601 if (!tick_program_event( 602 hrtimer_get_expires(&ts->sched_timer), 0)) 603 break; 604 } 605 /* Reread time and update jiffies */ 606 now = ktime_get(); 607 tick_do_update_jiffies64(now); 608 } 609 } 610 611 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 612 { 613 /* Update jiffies first */ 614 tick_do_update_jiffies64(now); 615 update_cpu_load_nohz(); 616 617 calc_load_exit_idle(); 618 touch_softlockup_watchdog(); 619 /* 620 * Cancel the scheduled timer and restore the tick 621 */ 622 ts->tick_stopped = 0; 623 ts->idle_exittime = now; 624 625 tick_nohz_restart(ts, now); 626 } 627 628 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 629 { 630 #ifndef CONFIG_VIRT_CPU_ACCOUNTING 631 unsigned long ticks; 632 /* 633 * We stopped the tick in idle. Update process times would miss the 634 * time we slept as update_process_times does only a 1 tick 635 * accounting. Enforce that this is accounted to idle ! 636 */ 637 ticks = jiffies - ts->idle_jiffies; 638 /* 639 * We might be one off. Do not randomly account a huge number of ticks! 640 */ 641 if (ticks && ticks < LONG_MAX) 642 account_idle_ticks(ticks); 643 #endif 644 } 645 646 /** 647 * tick_nohz_idle_exit - restart the idle tick from the idle task 648 * 649 * Restart the idle tick when the CPU is woken up from idle 650 * This also exit the RCU extended quiescent state. The CPU 651 * can use RCU again after this function is called. 652 */ 653 void tick_nohz_idle_exit(void) 654 { 655 int cpu = smp_processor_id(); 656 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 657 ktime_t now; 658 659 local_irq_disable(); 660 661 WARN_ON_ONCE(!ts->inidle); 662 663 ts->inidle = 0; 664 665 if (ts->idle_active || ts->tick_stopped) 666 now = ktime_get(); 667 668 if (ts->idle_active) 669 tick_nohz_stop_idle(cpu, now); 670 671 if (ts->tick_stopped) { 672 tick_nohz_restart_sched_tick(ts, now); 673 tick_nohz_account_idle_ticks(ts); 674 } 675 676 local_irq_enable(); 677 } 678 679 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 680 { 681 hrtimer_forward(&ts->sched_timer, now, tick_period); 682 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); 683 } 684 685 /* 686 * The nohz low res interrupt handler 687 */ 688 static void tick_nohz_handler(struct clock_event_device *dev) 689 { 690 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 691 struct pt_regs *regs = get_irq_regs(); 692 ktime_t now = ktime_get(); 693 694 dev->next_event.tv64 = KTIME_MAX; 695 696 tick_sched_do_timer(now); 697 tick_sched_handle(ts, regs); 698 699 while (tick_nohz_reprogram(ts, now)) { 700 now = ktime_get(); 701 tick_do_update_jiffies64(now); 702 } 703 } 704 705 /** 706 * tick_nohz_switch_to_nohz - switch to nohz mode 707 */ 708 static void tick_nohz_switch_to_nohz(void) 709 { 710 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 711 ktime_t next; 712 713 if (!tick_nohz_enabled) 714 return; 715 716 local_irq_disable(); 717 if (tick_switch_to_oneshot(tick_nohz_handler)) { 718 local_irq_enable(); 719 return; 720 } 721 722 ts->nohz_mode = NOHZ_MODE_LOWRES; 723 724 /* 725 * Recycle the hrtimer in ts, so we can share the 726 * hrtimer_forward with the highres code. 727 */ 728 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 729 /* Get the next period */ 730 next = tick_init_jiffy_update(); 731 732 for (;;) { 733 hrtimer_set_expires(&ts->sched_timer, next); 734 if (!tick_program_event(next, 0)) 735 break; 736 next = ktime_add(next, tick_period); 737 } 738 local_irq_enable(); 739 } 740 741 /* 742 * When NOHZ is enabled and the tick is stopped, we need to kick the 743 * tick timer from irq_enter() so that the jiffies update is kept 744 * alive during long running softirqs. That's ugly as hell, but 745 * correctness is key even if we need to fix the offending softirq in 746 * the first place. 747 * 748 * Note, this is different to tick_nohz_restart. We just kick the 749 * timer and do not touch the other magic bits which need to be done 750 * when idle is left. 751 */ 752 static void tick_nohz_kick_tick(int cpu, ktime_t now) 753 { 754 #if 0 755 /* Switch back to 2.6.27 behaviour */ 756 757 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 758 ktime_t delta; 759 760 /* 761 * Do not touch the tick device, when the next expiry is either 762 * already reached or less/equal than the tick period. 763 */ 764 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 765 if (delta.tv64 <= tick_period.tv64) 766 return; 767 768 tick_nohz_restart(ts, now); 769 #endif 770 } 771 772 static inline void tick_check_nohz(int cpu) 773 { 774 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 775 ktime_t now; 776 777 if (!ts->idle_active && !ts->tick_stopped) 778 return; 779 now = ktime_get(); 780 if (ts->idle_active) 781 tick_nohz_stop_idle(cpu, now); 782 if (ts->tick_stopped) { 783 tick_nohz_update_jiffies(now); 784 tick_nohz_kick_tick(cpu, now); 785 } 786 } 787 788 #else 789 790 static inline void tick_nohz_switch_to_nohz(void) { } 791 static inline void tick_check_nohz(int cpu) { } 792 793 #endif /* NO_HZ */ 794 795 /* 796 * Called from irq_enter to notify about the possible interruption of idle() 797 */ 798 void tick_check_idle(int cpu) 799 { 800 tick_check_oneshot_broadcast(cpu); 801 tick_check_nohz(cpu); 802 } 803 804 /* 805 * High resolution timer specific code 806 */ 807 #ifdef CONFIG_HIGH_RES_TIMERS 808 /* 809 * We rearm the timer until we get disabled by the idle code. 810 * Called with interrupts disabled and timer->base->cpu_base->lock held. 811 */ 812 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 813 { 814 struct tick_sched *ts = 815 container_of(timer, struct tick_sched, sched_timer); 816 struct pt_regs *regs = get_irq_regs(); 817 ktime_t now = ktime_get(); 818 819 tick_sched_do_timer(now); 820 821 /* 822 * Do not call, when we are not in irq context and have 823 * no valid regs pointer 824 */ 825 if (regs) 826 tick_sched_handle(ts, regs); 827 828 hrtimer_forward(timer, now, tick_period); 829 830 return HRTIMER_RESTART; 831 } 832 833 static int sched_skew_tick; 834 835 static int __init skew_tick(char *str) 836 { 837 get_option(&str, &sched_skew_tick); 838 839 return 0; 840 } 841 early_param("skew_tick", skew_tick); 842 843 /** 844 * tick_setup_sched_timer - setup the tick emulation timer 845 */ 846 void tick_setup_sched_timer(void) 847 { 848 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 849 ktime_t now = ktime_get(); 850 851 /* 852 * Emulate tick processing via per-CPU hrtimers: 853 */ 854 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 855 ts->sched_timer.function = tick_sched_timer; 856 857 /* Get the next period (per cpu) */ 858 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 859 860 /* Offset the tick to avert xtime_lock contention. */ 861 if (sched_skew_tick) { 862 u64 offset = ktime_to_ns(tick_period) >> 1; 863 do_div(offset, num_possible_cpus()); 864 offset *= smp_processor_id(); 865 hrtimer_add_expires_ns(&ts->sched_timer, offset); 866 } 867 868 for (;;) { 869 hrtimer_forward(&ts->sched_timer, now, tick_period); 870 hrtimer_start_expires(&ts->sched_timer, 871 HRTIMER_MODE_ABS_PINNED); 872 /* Check, if the timer was already in the past */ 873 if (hrtimer_active(&ts->sched_timer)) 874 break; 875 now = ktime_get(); 876 } 877 878 #ifdef CONFIG_NO_HZ 879 if (tick_nohz_enabled) 880 ts->nohz_mode = NOHZ_MODE_HIGHRES; 881 #endif 882 } 883 #endif /* HIGH_RES_TIMERS */ 884 885 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS 886 void tick_cancel_sched_timer(int cpu) 887 { 888 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 889 890 # ifdef CONFIG_HIGH_RES_TIMERS 891 if (ts->sched_timer.base) 892 hrtimer_cancel(&ts->sched_timer); 893 # endif 894 895 ts->nohz_mode = NOHZ_MODE_INACTIVE; 896 } 897 #endif 898 899 /** 900 * Async notification about clocksource changes 901 */ 902 void tick_clock_notify(void) 903 { 904 int cpu; 905 906 for_each_possible_cpu(cpu) 907 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 908 } 909 910 /* 911 * Async notification about clock event changes 912 */ 913 void tick_oneshot_notify(void) 914 { 915 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 916 917 set_bit(0, &ts->check_clocks); 918 } 919 920 /** 921 * Check, if a change happened, which makes oneshot possible. 922 * 923 * Called cyclic from the hrtimer softirq (driven by the timer 924 * softirq) allow_nohz signals, that we can switch into low-res nohz 925 * mode, because high resolution timers are disabled (either compile 926 * or runtime). 927 */ 928 int tick_check_oneshot_change(int allow_nohz) 929 { 930 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 931 932 if (!test_and_clear_bit(0, &ts->check_clocks)) 933 return 0; 934 935 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 936 return 0; 937 938 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 939 return 0; 940 941 if (!allow_nohz) 942 return 1; 943 944 tick_nohz_switch_to_nohz(); 945 return 0; 946 } 947