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 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 write_sequnlock(&xtime_lock); 80 } 81 82 /* 83 * Initialize and return retrieve the jiffies update. 84 */ 85 static ktime_t tick_init_jiffy_update(void) 86 { 87 ktime_t period; 88 89 write_seqlock(&xtime_lock); 90 /* Did we start the jiffies update yet ? */ 91 if (last_jiffies_update.tv64 == 0) 92 last_jiffies_update = tick_next_period; 93 period = last_jiffies_update; 94 write_sequnlock(&xtime_lock); 95 return period; 96 } 97 98 /* 99 * NOHZ - aka dynamic tick functionality 100 */ 101 #ifdef CONFIG_NO_HZ 102 /* 103 * NO HZ enabled ? 104 */ 105 static int tick_nohz_enabled __read_mostly = 1; 106 107 /* 108 * Enable / Disable tickless mode 109 */ 110 static int __init setup_tick_nohz(char *str) 111 { 112 if (!strcmp(str, "off")) 113 tick_nohz_enabled = 0; 114 else if (!strcmp(str, "on")) 115 tick_nohz_enabled = 1; 116 else 117 return 0; 118 return 1; 119 } 120 121 __setup("nohz=", setup_tick_nohz); 122 123 /** 124 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 125 * 126 * Called from interrupt entry when the CPU was idle 127 * 128 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 129 * must be updated. Otherwise an interrupt handler could use a stale jiffy 130 * value. We do this unconditionally on any cpu, as we don't know whether the 131 * cpu, which has the update task assigned is in a long sleep. 132 */ 133 void tick_nohz_update_jiffies(void) 134 { 135 int cpu = smp_processor_id(); 136 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 137 unsigned long flags; 138 ktime_t now; 139 140 if (!ts->tick_stopped) 141 return; 142 143 cpu_clear(cpu, nohz_cpu_mask); 144 now = ktime_get(); 145 ts->idle_waketime = now; 146 147 local_irq_save(flags); 148 tick_do_update_jiffies64(now); 149 local_irq_restore(flags); 150 151 touch_softlockup_watchdog(); 152 } 153 154 void tick_nohz_stop_idle(int cpu) 155 { 156 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 157 158 if (ts->idle_active) { 159 ktime_t now, delta; 160 now = ktime_get(); 161 delta = ktime_sub(now, ts->idle_entrytime); 162 ts->idle_lastupdate = now; 163 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 164 ts->idle_active = 0; 165 166 sched_clock_idle_wakeup_event(0); 167 } 168 } 169 170 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 171 { 172 ktime_t now, delta; 173 174 now = ktime_get(); 175 if (ts->idle_active) { 176 delta = ktime_sub(now, ts->idle_entrytime); 177 ts->idle_lastupdate = now; 178 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 179 } 180 ts->idle_entrytime = now; 181 ts->idle_active = 1; 182 sched_clock_idle_sleep_event(); 183 return now; 184 } 185 186 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 187 { 188 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 189 190 *last_update_time = ktime_to_us(ts->idle_lastupdate); 191 return ktime_to_us(ts->idle_sleeptime); 192 } 193 194 /** 195 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task 196 * 197 * When the next event is more than a tick into the future, stop the idle tick 198 * Called either from the idle loop or from irq_exit() when an idle period was 199 * just interrupted by an interrupt which did not cause a reschedule. 200 */ 201 void tick_nohz_stop_sched_tick(int inidle) 202 { 203 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; 204 struct tick_sched *ts; 205 ktime_t last_update, expires, now; 206 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; 207 int cpu; 208 209 local_irq_save(flags); 210 211 cpu = smp_processor_id(); 212 ts = &per_cpu(tick_cpu_sched, cpu); 213 now = tick_nohz_start_idle(ts); 214 215 /* 216 * If this cpu is offline and it is the one which updates 217 * jiffies, then give up the assignment and let it be taken by 218 * the cpu which runs the tick timer next. If we don't drop 219 * this here the jiffies might be stale and do_timer() never 220 * invoked. 221 */ 222 if (unlikely(!cpu_online(cpu))) { 223 if (cpu == tick_do_timer_cpu) 224 tick_do_timer_cpu = -1; 225 } 226 227 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) 228 goto end; 229 230 if (!inidle && !ts->inidle) 231 goto end; 232 233 ts->inidle = 1; 234 235 if (need_resched()) 236 goto end; 237 238 if (unlikely(local_softirq_pending())) { 239 static int ratelimit; 240 241 if (ratelimit < 10) { 242 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", 243 local_softirq_pending()); 244 ratelimit++; 245 } 246 goto end; 247 } 248 249 ts->idle_calls++; 250 /* Read jiffies and the time when jiffies were updated last */ 251 do { 252 seq = read_seqbegin(&xtime_lock); 253 last_update = last_jiffies_update; 254 last_jiffies = jiffies; 255 } while (read_seqretry(&xtime_lock, seq)); 256 257 /* Get the next timer wheel timer */ 258 next_jiffies = get_next_timer_interrupt(last_jiffies); 259 delta_jiffies = next_jiffies - last_jiffies; 260 261 if (rcu_needs_cpu(cpu)) 262 delta_jiffies = 1; 263 /* 264 * Do not stop the tick, if we are only one off 265 * or if the cpu is required for rcu 266 */ 267 if (!ts->tick_stopped && delta_jiffies == 1) 268 goto out; 269 270 /* Schedule the tick, if we are at least one jiffie off */ 271 if ((long)delta_jiffies >= 1) { 272 273 if (delta_jiffies > 1) 274 cpu_set(cpu, nohz_cpu_mask); 275 /* 276 * nohz_stop_sched_tick can be called several times before 277 * the nohz_restart_sched_tick is called. This happens when 278 * interrupts arrive which do not cause a reschedule. In the 279 * first call we save the current tick time, so we can restart 280 * the scheduler tick in nohz_restart_sched_tick. 281 */ 282 if (!ts->tick_stopped) { 283 if (select_nohz_load_balancer(1)) { 284 /* 285 * sched tick not stopped! 286 */ 287 cpu_clear(cpu, nohz_cpu_mask); 288 goto out; 289 } 290 291 ts->idle_tick = ts->sched_timer.expires; 292 ts->tick_stopped = 1; 293 ts->idle_jiffies = last_jiffies; 294 rcu_enter_nohz(); 295 } 296 297 /* 298 * If this cpu is the one which updates jiffies, then 299 * give up the assignment and let it be taken by the 300 * cpu which runs the tick timer next, which might be 301 * this cpu as well. If we don't drop this here the 302 * jiffies might be stale and do_timer() never 303 * invoked. 304 */ 305 if (cpu == tick_do_timer_cpu) 306 tick_do_timer_cpu = -1; 307 308 ts->idle_sleeps++; 309 310 /* 311 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that 312 * there is no timer pending or at least extremly far 313 * into the future (12 days for HZ=1000). In this case 314 * we simply stop the tick timer: 315 */ 316 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) { 317 ts->idle_expires.tv64 = KTIME_MAX; 318 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 319 hrtimer_cancel(&ts->sched_timer); 320 goto out; 321 } 322 323 /* 324 * calculate the expiry time for the next timer wheel 325 * timer 326 */ 327 expires = ktime_add_ns(last_update, tick_period.tv64 * 328 delta_jiffies); 329 ts->idle_expires = expires; 330 331 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 332 hrtimer_start(&ts->sched_timer, expires, 333 HRTIMER_MODE_ABS); 334 /* Check, if the timer was already in the past */ 335 if (hrtimer_active(&ts->sched_timer)) 336 goto out; 337 } else if (!tick_program_event(expires, 0)) 338 goto out; 339 /* 340 * We are past the event already. So we crossed a 341 * jiffie boundary. Update jiffies and raise the 342 * softirq. 343 */ 344 tick_do_update_jiffies64(ktime_get()); 345 cpu_clear(cpu, nohz_cpu_mask); 346 } 347 raise_softirq_irqoff(TIMER_SOFTIRQ); 348 out: 349 ts->next_jiffies = next_jiffies; 350 ts->last_jiffies = last_jiffies; 351 ts->sleep_length = ktime_sub(dev->next_event, now); 352 end: 353 local_irq_restore(flags); 354 } 355 356 /** 357 * tick_nohz_get_sleep_length - return the length of the current sleep 358 * 359 * Called from power state control code with interrupts disabled 360 */ 361 ktime_t tick_nohz_get_sleep_length(void) 362 { 363 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 364 365 return ts->sleep_length; 366 } 367 368 /** 369 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task 370 * 371 * Restart the idle tick when the CPU is woken up from idle 372 */ 373 void tick_nohz_restart_sched_tick(void) 374 { 375 int cpu = smp_processor_id(); 376 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 377 unsigned long ticks; 378 ktime_t now; 379 380 local_irq_disable(); 381 tick_nohz_stop_idle(cpu); 382 383 if (!ts->inidle || !ts->tick_stopped) { 384 ts->inidle = 0; 385 local_irq_enable(); 386 return; 387 } 388 389 ts->inidle = 0; 390 391 rcu_exit_nohz(); 392 393 /* Update jiffies first */ 394 select_nohz_load_balancer(0); 395 now = ktime_get(); 396 tick_do_update_jiffies64(now); 397 cpu_clear(cpu, nohz_cpu_mask); 398 399 /* 400 * We stopped the tick in idle. Update process times would miss the 401 * time we slept as update_process_times does only a 1 tick 402 * accounting. Enforce that this is accounted to idle ! 403 */ 404 ticks = jiffies - ts->idle_jiffies; 405 /* 406 * We might be one off. Do not randomly account a huge number of ticks! 407 */ 408 if (ticks && ticks < LONG_MAX) { 409 add_preempt_count(HARDIRQ_OFFSET); 410 account_system_time(current, HARDIRQ_OFFSET, 411 jiffies_to_cputime(ticks)); 412 sub_preempt_count(HARDIRQ_OFFSET); 413 } 414 415 touch_softlockup_watchdog(); 416 /* 417 * Cancel the scheduled timer and restore the tick 418 */ 419 ts->tick_stopped = 0; 420 ts->idle_exittime = now; 421 hrtimer_cancel(&ts->sched_timer); 422 ts->sched_timer.expires = ts->idle_tick; 423 424 while (1) { 425 /* Forward the time to expire in the future */ 426 hrtimer_forward(&ts->sched_timer, now, tick_period); 427 428 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 429 hrtimer_start(&ts->sched_timer, 430 ts->sched_timer.expires, 431 HRTIMER_MODE_ABS); 432 /* Check, if the timer was already in the past */ 433 if (hrtimer_active(&ts->sched_timer)) 434 break; 435 } else { 436 if (!tick_program_event(ts->sched_timer.expires, 0)) 437 break; 438 } 439 /* Update jiffies and reread time */ 440 tick_do_update_jiffies64(now); 441 now = ktime_get(); 442 } 443 local_irq_enable(); 444 } 445 446 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 447 { 448 hrtimer_forward(&ts->sched_timer, now, tick_period); 449 return tick_program_event(ts->sched_timer.expires, 0); 450 } 451 452 /* 453 * The nohz low res interrupt handler 454 */ 455 static void tick_nohz_handler(struct clock_event_device *dev) 456 { 457 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 458 struct pt_regs *regs = get_irq_regs(); 459 int cpu = smp_processor_id(); 460 ktime_t now = ktime_get(); 461 462 dev->next_event.tv64 = KTIME_MAX; 463 464 /* 465 * Check if the do_timer duty was dropped. We don't care about 466 * concurrency: This happens only when the cpu in charge went 467 * into a long sleep. If two cpus happen to assign themself to 468 * this duty, then the jiffies update is still serialized by 469 * xtime_lock. 470 */ 471 if (unlikely(tick_do_timer_cpu == -1)) 472 tick_do_timer_cpu = cpu; 473 474 /* Check, if the jiffies need an update */ 475 if (tick_do_timer_cpu == cpu) 476 tick_do_update_jiffies64(now); 477 478 /* 479 * When we are idle and the tick is stopped, we have to touch 480 * the watchdog as we might not schedule for a really long 481 * time. This happens on complete idle SMP systems while 482 * waiting on the login prompt. We also increment the "start 483 * of idle" jiffy stamp so the idle accounting adjustment we 484 * do when we go busy again does not account too much ticks. 485 */ 486 if (ts->tick_stopped) { 487 touch_softlockup_watchdog(); 488 ts->idle_jiffies++; 489 } 490 491 update_process_times(user_mode(regs)); 492 profile_tick(CPU_PROFILING); 493 494 /* Do not restart, when we are in the idle loop */ 495 if (ts->tick_stopped) 496 return; 497 498 while (tick_nohz_reprogram(ts, now)) { 499 now = ktime_get(); 500 tick_do_update_jiffies64(now); 501 } 502 } 503 504 /** 505 * tick_nohz_switch_to_nohz - switch to nohz mode 506 */ 507 static void tick_nohz_switch_to_nohz(void) 508 { 509 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 510 ktime_t next; 511 512 if (!tick_nohz_enabled) 513 return; 514 515 local_irq_disable(); 516 if (tick_switch_to_oneshot(tick_nohz_handler)) { 517 local_irq_enable(); 518 return; 519 } 520 521 ts->nohz_mode = NOHZ_MODE_LOWRES; 522 523 /* 524 * Recycle the hrtimer in ts, so we can share the 525 * hrtimer_forward with the highres code. 526 */ 527 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 528 /* Get the next period */ 529 next = tick_init_jiffy_update(); 530 531 for (;;) { 532 ts->sched_timer.expires = next; 533 if (!tick_program_event(next, 0)) 534 break; 535 next = ktime_add(next, tick_period); 536 } 537 local_irq_enable(); 538 539 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", 540 smp_processor_id()); 541 } 542 543 #else 544 545 static inline void tick_nohz_switch_to_nohz(void) { } 546 547 #endif /* NO_HZ */ 548 549 /* 550 * High resolution timer specific code 551 */ 552 #ifdef CONFIG_HIGH_RES_TIMERS 553 /* 554 * We rearm the timer until we get disabled by the idle code. 555 * Called with interrupts disabled and timer->base->cpu_base->lock held. 556 */ 557 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 558 { 559 struct tick_sched *ts = 560 container_of(timer, struct tick_sched, sched_timer); 561 struct pt_regs *regs = get_irq_regs(); 562 ktime_t now = ktime_get(); 563 int cpu = smp_processor_id(); 564 565 #ifdef CONFIG_NO_HZ 566 /* 567 * Check if the do_timer duty was dropped. We don't care about 568 * concurrency: This happens only when the cpu in charge went 569 * into a long sleep. If two cpus happen to assign themself to 570 * this duty, then the jiffies update is still serialized by 571 * xtime_lock. 572 */ 573 if (unlikely(tick_do_timer_cpu == -1)) 574 tick_do_timer_cpu = cpu; 575 #endif 576 577 /* Check, if the jiffies need an update */ 578 if (tick_do_timer_cpu == cpu) 579 tick_do_update_jiffies64(now); 580 581 /* 582 * Do not call, when we are not in irq context and have 583 * no valid regs pointer 584 */ 585 if (regs) { 586 /* 587 * When we are idle and the tick is stopped, we have to touch 588 * the watchdog as we might not schedule for a really long 589 * time. This happens on complete idle SMP systems while 590 * waiting on the login prompt. We also increment the "start of 591 * idle" jiffy stamp so the idle accounting adjustment we do 592 * when we go busy again does not account too much ticks. 593 */ 594 if (ts->tick_stopped) { 595 touch_softlockup_watchdog(); 596 ts->idle_jiffies++; 597 } 598 update_process_times(user_mode(regs)); 599 profile_tick(CPU_PROFILING); 600 } 601 602 /* Do not restart, when we are in the idle loop */ 603 if (ts->tick_stopped) 604 return HRTIMER_NORESTART; 605 606 hrtimer_forward(timer, now, tick_period); 607 608 return HRTIMER_RESTART; 609 } 610 611 /** 612 * tick_setup_sched_timer - setup the tick emulation timer 613 */ 614 void tick_setup_sched_timer(void) 615 { 616 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 617 ktime_t now = ktime_get(); 618 u64 offset; 619 620 /* 621 * Emulate tick processing via per-CPU hrtimers: 622 */ 623 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 624 ts->sched_timer.function = tick_sched_timer; 625 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; 626 627 /* Get the next period (per cpu) */ 628 ts->sched_timer.expires = tick_init_jiffy_update(); 629 offset = ktime_to_ns(tick_period) >> 1; 630 do_div(offset, num_possible_cpus()); 631 offset *= smp_processor_id(); 632 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset); 633 634 for (;;) { 635 hrtimer_forward(&ts->sched_timer, now, tick_period); 636 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires, 637 HRTIMER_MODE_ABS); 638 /* Check, if the timer was already in the past */ 639 if (hrtimer_active(&ts->sched_timer)) 640 break; 641 now = ktime_get(); 642 } 643 644 #ifdef CONFIG_NO_HZ 645 if (tick_nohz_enabled) 646 ts->nohz_mode = NOHZ_MODE_HIGHRES; 647 #endif 648 } 649 #endif /* HIGH_RES_TIMERS */ 650 651 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS 652 void tick_cancel_sched_timer(int cpu) 653 { 654 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 655 656 # ifdef CONFIG_HIGH_RES_TIMERS 657 if (ts->sched_timer.base) 658 hrtimer_cancel(&ts->sched_timer); 659 # endif 660 661 ts->nohz_mode = NOHZ_MODE_INACTIVE; 662 } 663 #endif 664 665 /** 666 * Async notification about clocksource changes 667 */ 668 void tick_clock_notify(void) 669 { 670 int cpu; 671 672 for_each_possible_cpu(cpu) 673 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 674 } 675 676 /* 677 * Async notification about clock event changes 678 */ 679 void tick_oneshot_notify(void) 680 { 681 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 682 683 set_bit(0, &ts->check_clocks); 684 } 685 686 /** 687 * Check, if a change happened, which makes oneshot possible. 688 * 689 * Called cyclic from the hrtimer softirq (driven by the timer 690 * softirq) allow_nohz signals, that we can switch into low-res nohz 691 * mode, because high resolution timers are disabled (either compile 692 * or runtime). 693 */ 694 int tick_check_oneshot_change(int allow_nohz) 695 { 696 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 697 698 if (!test_and_clear_bit(0, &ts->check_clocks)) 699 return 0; 700 701 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 702 return 0; 703 704 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 705 return 0; 706 707 if (!allow_nohz) 708 return 1; 709 710 tick_nohz_switch_to_nohz(); 711 return 0; 712 } 713