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