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/tick.h> 23 #include <linux/module.h> 24 25 #include <asm/irq_regs.h> 26 27 #include "tick-internal.h" 28 29 /* 30 * Per cpu nohz control structure 31 */ 32 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); 33 34 /* 35 * The time, when the last jiffy update happened. Protected by xtime_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 xtime_lock: 54 */ 55 delta = ktime_sub(now, last_jiffies_update); 56 if (delta.tv64 < tick_period.tv64) 57 return; 58 59 /* Reevalute with xtime_lock held */ 60 write_seqlock(&xtime_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(&xtime_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(&xtime_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(&xtime_lock); 99 return period; 100 } 101 102 /* 103 * NOHZ - aka dynamic tick functionality 104 */ 105 #ifdef CONFIG_NO_HZ 106 /* 107 * NO HZ enabled ? 108 */ 109 static int tick_nohz_enabled __read_mostly = 1; 110 111 /* 112 * Enable / Disable tickless mode 113 */ 114 static int __init setup_tick_nohz(char *str) 115 { 116 if (!strcmp(str, "off")) 117 tick_nohz_enabled = 0; 118 else if (!strcmp(str, "on")) 119 tick_nohz_enabled = 1; 120 else 121 return 0; 122 return 1; 123 } 124 125 __setup("nohz=", setup_tick_nohz); 126 127 /** 128 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 129 * 130 * Called from interrupt entry when the CPU was idle 131 * 132 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 133 * must be updated. Otherwise an interrupt handler could use a stale jiffy 134 * value. We do this unconditionally on any cpu, as we don't know whether the 135 * cpu, which has the update task assigned is in a long sleep. 136 */ 137 static void tick_nohz_update_jiffies(ktime_t now) 138 { 139 int cpu = smp_processor_id(); 140 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 141 unsigned long flags; 142 143 cpumask_clear_cpu(cpu, nohz_cpu_mask); 144 ts->idle_waketime = now; 145 146 local_irq_save(flags); 147 tick_do_update_jiffies64(now); 148 local_irq_restore(flags); 149 150 touch_softlockup_watchdog(); 151 } 152 153 static void tick_nohz_stop_idle(int cpu, ktime_t now) 154 { 155 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 156 ktime_t delta; 157 158 delta = ktime_sub(now, ts->idle_entrytime); 159 ts->idle_lastupdate = now; 160 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 161 ts->idle_active = 0; 162 163 sched_clock_idle_wakeup_event(0); 164 } 165 166 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 167 { 168 ktime_t now, delta; 169 170 now = ktime_get(); 171 if (ts->idle_active) { 172 delta = ktime_sub(now, ts->idle_entrytime); 173 ts->idle_lastupdate = now; 174 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 175 } 176 ts->idle_entrytime = now; 177 ts->idle_active = 1; 178 sched_clock_idle_sleep_event(); 179 return now; 180 } 181 182 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 183 { 184 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 185 186 if (!tick_nohz_enabled) 187 return -1; 188 189 if (ts->idle_active) 190 *last_update_time = ktime_to_us(ts->idle_lastupdate); 191 else 192 *last_update_time = ktime_to_us(ktime_get()); 193 194 return ktime_to_us(ts->idle_sleeptime); 195 } 196 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 197 198 /** 199 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task 200 * 201 * When the next event is more than a tick into the future, stop the idle tick 202 * Called either from the idle loop or from irq_exit() when an idle period was 203 * just interrupted by an interrupt which did not cause a reschedule. 204 */ 205 void tick_nohz_stop_sched_tick(int inidle) 206 { 207 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; 208 struct tick_sched *ts; 209 ktime_t last_update, expires, now; 210 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; 211 u64 time_delta; 212 int cpu; 213 214 local_irq_save(flags); 215 216 cpu = smp_processor_id(); 217 ts = &per_cpu(tick_cpu_sched, cpu); 218 219 /* 220 * Call to tick_nohz_start_idle stops the last_update_time from being 221 * updated. Thus, it must not be called in the event we are called from 222 * irq_exit() with the prior state different than idle. 223 */ 224 if (!inidle && !ts->inidle) 225 goto end; 226 227 /* 228 * Set ts->inidle unconditionally. Even if the system did not 229 * switch to NOHZ mode the cpu frequency governers rely on the 230 * update of the idle time accounting in tick_nohz_start_idle(). 231 */ 232 ts->inidle = 1; 233 234 now = tick_nohz_start_idle(ts); 235 236 /* 237 * If this cpu is offline and it is the one which updates 238 * jiffies, then give up the assignment and let it be taken by 239 * the cpu which runs the tick timer next. If we don't drop 240 * this here the jiffies might be stale and do_timer() never 241 * invoked. 242 */ 243 if (unlikely(!cpu_online(cpu))) { 244 if (cpu == tick_do_timer_cpu) 245 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 246 } 247 248 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) 249 goto end; 250 251 if (need_resched()) 252 goto end; 253 254 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 255 static int ratelimit; 256 257 if (ratelimit < 10) { 258 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", 259 (unsigned int) local_softirq_pending()); 260 ratelimit++; 261 } 262 goto end; 263 } 264 265 ts->idle_calls++; 266 /* Read jiffies and the time when jiffies were updated last */ 267 do { 268 seq = read_seqbegin(&xtime_lock); 269 last_update = last_jiffies_update; 270 last_jiffies = jiffies; 271 time_delta = timekeeping_max_deferment(); 272 } while (read_seqretry(&xtime_lock, seq)); 273 274 if (rcu_needs_cpu(cpu) || printk_needs_cpu(cpu) || 275 arch_needs_cpu(cpu)) { 276 next_jiffies = last_jiffies + 1; 277 delta_jiffies = 1; 278 } else { 279 /* Get the next timer wheel timer */ 280 next_jiffies = get_next_timer_interrupt(last_jiffies); 281 delta_jiffies = next_jiffies - last_jiffies; 282 } 283 /* 284 * Do not stop the tick, if we are only one off 285 * or if the cpu is required for rcu 286 */ 287 if (!ts->tick_stopped && delta_jiffies == 1) 288 goto out; 289 290 /* Schedule the tick, if we are at least one jiffie off */ 291 if ((long)delta_jiffies >= 1) { 292 293 /* 294 * If this cpu is the one which updates jiffies, then 295 * give up the assignment and let it be taken by the 296 * cpu which runs the tick timer next, which might be 297 * this cpu as well. If we don't drop this here the 298 * jiffies might be stale and do_timer() never 299 * invoked. Keep track of the fact that it was the one 300 * which had the do_timer() duty last. If this cpu is 301 * the one which had the do_timer() duty last, we 302 * limit the sleep time to the timekeeping 303 * max_deferement value which we retrieved 304 * above. Otherwise we can sleep as long as we want. 305 */ 306 if (cpu == tick_do_timer_cpu) { 307 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 308 ts->do_timer_last = 1; 309 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 310 time_delta = KTIME_MAX; 311 ts->do_timer_last = 0; 312 } else if (!ts->do_timer_last) { 313 time_delta = KTIME_MAX; 314 } 315 316 /* 317 * calculate the expiry time for the next timer wheel 318 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals 319 * that there is no timer pending or at least extremely 320 * far into the future (12 days for HZ=1000). In this 321 * case we set the expiry to the end of time. 322 */ 323 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { 324 /* 325 * Calculate the time delta for the next timer event. 326 * If the time delta exceeds the maximum time delta 327 * permitted by the current clocksource then adjust 328 * the time delta accordingly to ensure the 329 * clocksource does not wrap. 330 */ 331 time_delta = min_t(u64, time_delta, 332 tick_period.tv64 * delta_jiffies); 333 } 334 335 if (time_delta < KTIME_MAX) 336 expires = ktime_add_ns(last_update, time_delta); 337 else 338 expires.tv64 = KTIME_MAX; 339 340 if (delta_jiffies > 1) 341 cpumask_set_cpu(cpu, nohz_cpu_mask); 342 343 /* Skip reprogram of event if its not changed */ 344 if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) 345 goto out; 346 347 /* 348 * nohz_stop_sched_tick can be called several times before 349 * the nohz_restart_sched_tick is called. This happens when 350 * interrupts arrive which do not cause a reschedule. In the 351 * first call we save the current tick time, so we can restart 352 * the scheduler tick in nohz_restart_sched_tick. 353 */ 354 if (!ts->tick_stopped) { 355 if (select_nohz_load_balancer(1)) { 356 /* 357 * sched tick not stopped! 358 */ 359 cpumask_clear_cpu(cpu, nohz_cpu_mask); 360 goto out; 361 } 362 363 ts->idle_tick = hrtimer_get_expires(&ts->sched_timer); 364 ts->tick_stopped = 1; 365 ts->idle_jiffies = last_jiffies; 366 rcu_enter_nohz(); 367 } 368 369 ts->idle_sleeps++; 370 371 /* Mark expires */ 372 ts->idle_expires = expires; 373 374 /* 375 * If the expiration time == KTIME_MAX, then 376 * in this case we simply stop the tick timer. 377 */ 378 if (unlikely(expires.tv64 == KTIME_MAX)) { 379 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 380 hrtimer_cancel(&ts->sched_timer); 381 goto out; 382 } 383 384 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 385 hrtimer_start(&ts->sched_timer, expires, 386 HRTIMER_MODE_ABS_PINNED); 387 /* Check, if the timer was already in the past */ 388 if (hrtimer_active(&ts->sched_timer)) 389 goto out; 390 } else if (!tick_program_event(expires, 0)) 391 goto out; 392 /* 393 * We are past the event already. So we crossed a 394 * jiffie boundary. Update jiffies and raise the 395 * softirq. 396 */ 397 tick_do_update_jiffies64(ktime_get()); 398 cpumask_clear_cpu(cpu, nohz_cpu_mask); 399 } 400 raise_softirq_irqoff(TIMER_SOFTIRQ); 401 out: 402 ts->next_jiffies = next_jiffies; 403 ts->last_jiffies = last_jiffies; 404 ts->sleep_length = ktime_sub(dev->next_event, now); 405 end: 406 local_irq_restore(flags); 407 } 408 409 /** 410 * tick_nohz_get_sleep_length - return the length of the current sleep 411 * 412 * Called from power state control code with interrupts disabled 413 */ 414 ktime_t tick_nohz_get_sleep_length(void) 415 { 416 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 417 418 return ts->sleep_length; 419 } 420 421 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 422 { 423 hrtimer_cancel(&ts->sched_timer); 424 hrtimer_set_expires(&ts->sched_timer, ts->idle_tick); 425 426 while (1) { 427 /* Forward the time to expire in the future */ 428 hrtimer_forward(&ts->sched_timer, now, tick_period); 429 430 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 431 hrtimer_start_expires(&ts->sched_timer, 432 HRTIMER_MODE_ABS_PINNED); 433 /* Check, if the timer was already in the past */ 434 if (hrtimer_active(&ts->sched_timer)) 435 break; 436 } else { 437 if (!tick_program_event( 438 hrtimer_get_expires(&ts->sched_timer), 0)) 439 break; 440 } 441 /* Update jiffies and reread time */ 442 tick_do_update_jiffies64(now); 443 now = ktime_get(); 444 } 445 } 446 447 /** 448 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task 449 * 450 * Restart the idle tick when the CPU is woken up from idle 451 */ 452 void tick_nohz_restart_sched_tick(void) 453 { 454 int cpu = smp_processor_id(); 455 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 456 #ifndef CONFIG_VIRT_CPU_ACCOUNTING 457 unsigned long ticks; 458 #endif 459 ktime_t now; 460 461 local_irq_disable(); 462 if (ts->idle_active || (ts->inidle && ts->tick_stopped)) 463 now = ktime_get(); 464 465 if (ts->idle_active) 466 tick_nohz_stop_idle(cpu, now); 467 468 if (!ts->inidle || !ts->tick_stopped) { 469 ts->inidle = 0; 470 local_irq_enable(); 471 return; 472 } 473 474 ts->inidle = 0; 475 476 rcu_exit_nohz(); 477 478 /* Update jiffies first */ 479 select_nohz_load_balancer(0); 480 tick_do_update_jiffies64(now); 481 cpumask_clear_cpu(cpu, nohz_cpu_mask); 482 483 #ifndef CONFIG_VIRT_CPU_ACCOUNTING 484 /* 485 * We stopped the tick in idle. Update process times would miss the 486 * time we slept as update_process_times does only a 1 tick 487 * accounting. Enforce that this is accounted to idle ! 488 */ 489 ticks = jiffies - ts->idle_jiffies; 490 /* 491 * We might be one off. Do not randomly account a huge number of ticks! 492 */ 493 if (ticks && ticks < LONG_MAX) 494 account_idle_ticks(ticks); 495 #endif 496 497 touch_softlockup_watchdog(); 498 /* 499 * Cancel the scheduled timer and restore the tick 500 */ 501 ts->tick_stopped = 0; 502 ts->idle_exittime = now; 503 504 tick_nohz_restart(ts, now); 505 506 local_irq_enable(); 507 } 508 509 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 510 { 511 hrtimer_forward(&ts->sched_timer, now, tick_period); 512 return tick_program_event(hrtimer_get_expires(&ts->sched_timer), 0); 513 } 514 515 /* 516 * The nohz low res interrupt handler 517 */ 518 static void tick_nohz_handler(struct clock_event_device *dev) 519 { 520 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 521 struct pt_regs *regs = get_irq_regs(); 522 int cpu = smp_processor_id(); 523 ktime_t now = ktime_get(); 524 525 dev->next_event.tv64 = KTIME_MAX; 526 527 /* 528 * Check if the do_timer duty was dropped. We don't care about 529 * concurrency: This happens only when the cpu in charge went 530 * into a long sleep. If two cpus happen to assign themself to 531 * this duty, then the jiffies update is still serialized by 532 * xtime_lock. 533 */ 534 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) 535 tick_do_timer_cpu = cpu; 536 537 /* Check, if the jiffies need an update */ 538 if (tick_do_timer_cpu == cpu) 539 tick_do_update_jiffies64(now); 540 541 /* 542 * When we are idle and the tick is stopped, we have to touch 543 * the watchdog as we might not schedule for a really long 544 * time. This happens on complete idle SMP systems while 545 * waiting on the login prompt. We also increment the "start 546 * of idle" jiffy stamp so the idle accounting adjustment we 547 * do when we go busy again does not account too much ticks. 548 */ 549 if (ts->tick_stopped) { 550 touch_softlockup_watchdog(); 551 ts->idle_jiffies++; 552 } 553 554 update_process_times(user_mode(regs)); 555 profile_tick(CPU_PROFILING); 556 557 while (tick_nohz_reprogram(ts, now)) { 558 now = ktime_get(); 559 tick_do_update_jiffies64(now); 560 } 561 } 562 563 /** 564 * tick_nohz_switch_to_nohz - switch to nohz mode 565 */ 566 static void tick_nohz_switch_to_nohz(void) 567 { 568 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 569 ktime_t next; 570 571 if (!tick_nohz_enabled) 572 return; 573 574 local_irq_disable(); 575 if (tick_switch_to_oneshot(tick_nohz_handler)) { 576 local_irq_enable(); 577 return; 578 } 579 580 ts->nohz_mode = NOHZ_MODE_LOWRES; 581 582 /* 583 * Recycle the hrtimer in ts, so we can share the 584 * hrtimer_forward with the highres code. 585 */ 586 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 587 /* Get the next period */ 588 next = tick_init_jiffy_update(); 589 590 for (;;) { 591 hrtimer_set_expires(&ts->sched_timer, next); 592 if (!tick_program_event(next, 0)) 593 break; 594 next = ktime_add(next, tick_period); 595 } 596 local_irq_enable(); 597 598 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", 599 smp_processor_id()); 600 } 601 602 /* 603 * When NOHZ is enabled and the tick is stopped, we need to kick the 604 * tick timer from irq_enter() so that the jiffies update is kept 605 * alive during long running softirqs. That's ugly as hell, but 606 * correctness is key even if we need to fix the offending softirq in 607 * the first place. 608 * 609 * Note, this is different to tick_nohz_restart. We just kick the 610 * timer and do not touch the other magic bits which need to be done 611 * when idle is left. 612 */ 613 static void tick_nohz_kick_tick(int cpu, ktime_t now) 614 { 615 #if 0 616 /* Switch back to 2.6.27 behaviour */ 617 618 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 619 ktime_t delta; 620 621 /* 622 * Do not touch the tick device, when the next expiry is either 623 * already reached or less/equal than the tick period. 624 */ 625 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 626 if (delta.tv64 <= tick_period.tv64) 627 return; 628 629 tick_nohz_restart(ts, now); 630 #endif 631 } 632 633 static inline void tick_check_nohz(int cpu) 634 { 635 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 636 ktime_t now; 637 638 if (!ts->idle_active && !ts->tick_stopped) 639 return; 640 now = ktime_get(); 641 if (ts->idle_active) 642 tick_nohz_stop_idle(cpu, now); 643 if (ts->tick_stopped) { 644 tick_nohz_update_jiffies(now); 645 tick_nohz_kick_tick(cpu, now); 646 } 647 } 648 649 #else 650 651 static inline void tick_nohz_switch_to_nohz(void) { } 652 static inline void tick_check_nohz(int cpu) { } 653 654 #endif /* NO_HZ */ 655 656 /* 657 * Called from irq_enter to notify about the possible interruption of idle() 658 */ 659 void tick_check_idle(int cpu) 660 { 661 tick_check_oneshot_broadcast(cpu); 662 tick_check_nohz(cpu); 663 } 664 665 /* 666 * High resolution timer specific code 667 */ 668 #ifdef CONFIG_HIGH_RES_TIMERS 669 /* 670 * We rearm the timer until we get disabled by the idle code. 671 * Called with interrupts disabled and timer->base->cpu_base->lock held. 672 */ 673 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 674 { 675 struct tick_sched *ts = 676 container_of(timer, struct tick_sched, sched_timer); 677 struct pt_regs *regs = get_irq_regs(); 678 ktime_t now = ktime_get(); 679 int cpu = smp_processor_id(); 680 681 #ifdef CONFIG_NO_HZ 682 /* 683 * Check if the do_timer duty was dropped. We don't care about 684 * concurrency: This happens only when the cpu in charge went 685 * into a long sleep. If two cpus happen to assign themself to 686 * this duty, then the jiffies update is still serialized by 687 * xtime_lock. 688 */ 689 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE)) 690 tick_do_timer_cpu = cpu; 691 #endif 692 693 /* Check, if the jiffies need an update */ 694 if (tick_do_timer_cpu == cpu) 695 tick_do_update_jiffies64(now); 696 697 /* 698 * Do not call, when we are not in irq context and have 699 * no valid regs pointer 700 */ 701 if (regs) { 702 /* 703 * When we are idle and the tick is stopped, we have to touch 704 * the watchdog as we might not schedule for a really long 705 * time. This happens on complete idle SMP systems while 706 * waiting on the login prompt. We also increment the "start of 707 * idle" jiffy stamp so the idle accounting adjustment we do 708 * when we go busy again does not account too much ticks. 709 */ 710 if (ts->tick_stopped) { 711 touch_softlockup_watchdog(); 712 ts->idle_jiffies++; 713 } 714 update_process_times(user_mode(regs)); 715 profile_tick(CPU_PROFILING); 716 } 717 718 hrtimer_forward(timer, now, tick_period); 719 720 return HRTIMER_RESTART; 721 } 722 723 /** 724 * tick_setup_sched_timer - setup the tick emulation timer 725 */ 726 void tick_setup_sched_timer(void) 727 { 728 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 729 ktime_t now = ktime_get(); 730 u64 offset; 731 732 /* 733 * Emulate tick processing via per-CPU hrtimers: 734 */ 735 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 736 ts->sched_timer.function = tick_sched_timer; 737 738 /* Get the next period (per cpu) */ 739 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 740 offset = ktime_to_ns(tick_period) >> 1; 741 do_div(offset, num_possible_cpus()); 742 offset *= smp_processor_id(); 743 hrtimer_add_expires_ns(&ts->sched_timer, offset); 744 745 for (;;) { 746 hrtimer_forward(&ts->sched_timer, now, tick_period); 747 hrtimer_start_expires(&ts->sched_timer, 748 HRTIMER_MODE_ABS_PINNED); 749 /* Check, if the timer was already in the past */ 750 if (hrtimer_active(&ts->sched_timer)) 751 break; 752 now = ktime_get(); 753 } 754 755 #ifdef CONFIG_NO_HZ 756 if (tick_nohz_enabled) 757 ts->nohz_mode = NOHZ_MODE_HIGHRES; 758 #endif 759 } 760 #endif /* HIGH_RES_TIMERS */ 761 762 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS 763 void tick_cancel_sched_timer(int cpu) 764 { 765 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 766 767 # ifdef CONFIG_HIGH_RES_TIMERS 768 if (ts->sched_timer.base) 769 hrtimer_cancel(&ts->sched_timer); 770 # endif 771 772 ts->nohz_mode = NOHZ_MODE_INACTIVE; 773 } 774 #endif 775 776 /** 777 * Async notification about clocksource changes 778 */ 779 void tick_clock_notify(void) 780 { 781 int cpu; 782 783 for_each_possible_cpu(cpu) 784 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 785 } 786 787 /* 788 * Async notification about clock event changes 789 */ 790 void tick_oneshot_notify(void) 791 { 792 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 793 794 set_bit(0, &ts->check_clocks); 795 } 796 797 /** 798 * Check, if a change happened, which makes oneshot possible. 799 * 800 * Called cyclic from the hrtimer softirq (driven by the timer 801 * softirq) allow_nohz signals, that we can switch into low-res nohz 802 * mode, because high resolution timers are disabled (either compile 803 * or runtime). 804 */ 805 int tick_check_oneshot_change(int allow_nohz) 806 { 807 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 808 809 if (!test_and_clear_bit(0, &ts->check_clocks)) 810 return 0; 811 812 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 813 return 0; 814 815 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 816 return 0; 817 818 if (!allow_nohz) 819 return 1; 820 821 tick_nohz_switch_to_nohz(); 822 return 0; 823 } 824