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