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 } 293 294 /* 295 * If this cpu is the one which updates jiffies, then 296 * give up the assignment and let it be taken by the 297 * cpu which runs the tick timer next, which might be 298 * this cpu as well. If we don't drop this here the 299 * jiffies might be stale and do_timer() never 300 * invoked. 301 */ 302 if (cpu == tick_do_timer_cpu) 303 tick_do_timer_cpu = -1; 304 305 ts->idle_sleeps++; 306 307 /* 308 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that 309 * there is no timer pending or at least extremly far 310 * into the future (12 days for HZ=1000). In this case 311 * we simply stop the tick timer: 312 */ 313 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) { 314 ts->idle_expires.tv64 = KTIME_MAX; 315 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 316 hrtimer_cancel(&ts->sched_timer); 317 goto out; 318 } 319 320 /* 321 * calculate the expiry time for the next timer wheel 322 * timer 323 */ 324 expires = ktime_add_ns(last_update, tick_period.tv64 * 325 delta_jiffies); 326 ts->idle_expires = expires; 327 328 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 329 hrtimer_start(&ts->sched_timer, expires, 330 HRTIMER_MODE_ABS); 331 /* Check, if the timer was already in the past */ 332 if (hrtimer_active(&ts->sched_timer)) 333 goto out; 334 } else if (!tick_program_event(expires, 0)) 335 goto out; 336 /* 337 * We are past the event already. So we crossed a 338 * jiffie boundary. Update jiffies and raise the 339 * softirq. 340 */ 341 tick_do_update_jiffies64(ktime_get()); 342 cpu_clear(cpu, nohz_cpu_mask); 343 } 344 raise_softirq_irqoff(TIMER_SOFTIRQ); 345 out: 346 ts->next_jiffies = next_jiffies; 347 ts->last_jiffies = last_jiffies; 348 ts->sleep_length = ktime_sub(dev->next_event, now); 349 end: 350 local_irq_restore(flags); 351 } 352 353 /** 354 * tick_nohz_get_sleep_length - return the length of the current sleep 355 * 356 * Called from power state control code with interrupts disabled 357 */ 358 ktime_t tick_nohz_get_sleep_length(void) 359 { 360 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 361 362 return ts->sleep_length; 363 } 364 365 /** 366 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task 367 * 368 * Restart the idle tick when the CPU is woken up from idle 369 */ 370 void tick_nohz_restart_sched_tick(void) 371 { 372 int cpu = smp_processor_id(); 373 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 374 unsigned long ticks; 375 ktime_t now; 376 377 local_irq_disable(); 378 tick_nohz_stop_idle(cpu); 379 380 if (!ts->inidle || !ts->tick_stopped) { 381 ts->inidle = 0; 382 local_irq_enable(); 383 return; 384 } 385 386 ts->inidle = 0; 387 388 rcu_exit_nohz(); 389 390 /* Update jiffies first */ 391 select_nohz_load_balancer(0); 392 now = ktime_get(); 393 tick_do_update_jiffies64(now); 394 cpu_clear(cpu, nohz_cpu_mask); 395 396 /* 397 * We stopped the tick in idle. Update process times would miss the 398 * time we slept as update_process_times does only a 1 tick 399 * accounting. Enforce that this is accounted to idle ! 400 */ 401 ticks = jiffies - ts->idle_jiffies; 402 /* 403 * We might be one off. Do not randomly account a huge number of ticks! 404 */ 405 if (ticks && ticks < LONG_MAX) { 406 add_preempt_count(HARDIRQ_OFFSET); 407 account_system_time(current, HARDIRQ_OFFSET, 408 jiffies_to_cputime(ticks)); 409 sub_preempt_count(HARDIRQ_OFFSET); 410 } 411 412 touch_softlockup_watchdog(); 413 /* 414 * Cancel the scheduled timer and restore the tick 415 */ 416 ts->tick_stopped = 0; 417 ts->idle_exittime = now; 418 hrtimer_cancel(&ts->sched_timer); 419 ts->sched_timer.expires = ts->idle_tick; 420 421 while (1) { 422 /* Forward the time to expire in the future */ 423 hrtimer_forward(&ts->sched_timer, now, tick_period); 424 425 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 426 hrtimer_start(&ts->sched_timer, 427 ts->sched_timer.expires, 428 HRTIMER_MODE_ABS); 429 /* Check, if the timer was already in the past */ 430 if (hrtimer_active(&ts->sched_timer)) 431 break; 432 } else { 433 if (!tick_program_event(ts->sched_timer.expires, 0)) 434 break; 435 } 436 /* Update jiffies and reread time */ 437 tick_do_update_jiffies64(now); 438 now = ktime_get(); 439 } 440 local_irq_enable(); 441 } 442 443 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 444 { 445 hrtimer_forward(&ts->sched_timer, now, tick_period); 446 return tick_program_event(ts->sched_timer.expires, 0); 447 } 448 449 /* 450 * The nohz low res interrupt handler 451 */ 452 static void tick_nohz_handler(struct clock_event_device *dev) 453 { 454 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 455 struct pt_regs *regs = get_irq_regs(); 456 int cpu = smp_processor_id(); 457 ktime_t now = ktime_get(); 458 459 dev->next_event.tv64 = KTIME_MAX; 460 461 /* 462 * Check if the do_timer duty was dropped. We don't care about 463 * concurrency: This happens only when the cpu in charge went 464 * into a long sleep. If two cpus happen to assign themself to 465 * this duty, then the jiffies update is still serialized by 466 * xtime_lock. 467 */ 468 if (unlikely(tick_do_timer_cpu == -1)) 469 tick_do_timer_cpu = cpu; 470 471 /* Check, if the jiffies need an update */ 472 if (tick_do_timer_cpu == cpu) 473 tick_do_update_jiffies64(now); 474 475 /* 476 * When we are idle and the tick is stopped, we have to touch 477 * the watchdog as we might not schedule for a really long 478 * time. This happens on complete idle SMP systems while 479 * waiting on the login prompt. We also increment the "start 480 * of idle" jiffy stamp so the idle accounting adjustment we 481 * do when we go busy again does not account too much ticks. 482 */ 483 if (ts->tick_stopped) { 484 touch_softlockup_watchdog(); 485 ts->idle_jiffies++; 486 } 487 488 update_process_times(user_mode(regs)); 489 profile_tick(CPU_PROFILING); 490 491 /* Do not restart, when we are in the idle loop */ 492 if (ts->tick_stopped) 493 return; 494 495 while (tick_nohz_reprogram(ts, now)) { 496 now = ktime_get(); 497 tick_do_update_jiffies64(now); 498 } 499 } 500 501 /** 502 * tick_nohz_switch_to_nohz - switch to nohz mode 503 */ 504 static void tick_nohz_switch_to_nohz(void) 505 { 506 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 507 ktime_t next; 508 509 if (!tick_nohz_enabled) 510 return; 511 512 local_irq_disable(); 513 if (tick_switch_to_oneshot(tick_nohz_handler)) { 514 local_irq_enable(); 515 return; 516 } 517 518 ts->nohz_mode = NOHZ_MODE_LOWRES; 519 520 /* 521 * Recycle the hrtimer in ts, so we can share the 522 * hrtimer_forward with the highres code. 523 */ 524 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 525 /* Get the next period */ 526 next = tick_init_jiffy_update(); 527 528 for (;;) { 529 ts->sched_timer.expires = next; 530 if (!tick_program_event(next, 0)) 531 break; 532 next = ktime_add(next, tick_period); 533 } 534 local_irq_enable(); 535 536 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", 537 smp_processor_id()); 538 } 539 540 #else 541 542 static inline void tick_nohz_switch_to_nohz(void) { } 543 544 #endif /* NO_HZ */ 545 546 /* 547 * High resolution timer specific code 548 */ 549 #ifdef CONFIG_HIGH_RES_TIMERS 550 /* 551 * We rearm the timer until we get disabled by the idle code. 552 * Called with interrupts disabled and timer->base->cpu_base->lock held. 553 */ 554 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 555 { 556 struct tick_sched *ts = 557 container_of(timer, struct tick_sched, sched_timer); 558 struct pt_regs *regs = get_irq_regs(); 559 ktime_t now = ktime_get(); 560 int cpu = smp_processor_id(); 561 562 #ifdef CONFIG_NO_HZ 563 /* 564 * Check if the do_timer duty was dropped. We don't care about 565 * concurrency: This happens only when the cpu in charge went 566 * into a long sleep. If two cpus happen to assign themself to 567 * this duty, then the jiffies update is still serialized by 568 * xtime_lock. 569 */ 570 if (unlikely(tick_do_timer_cpu == -1)) 571 tick_do_timer_cpu = cpu; 572 #endif 573 574 /* Check, if the jiffies need an update */ 575 if (tick_do_timer_cpu == cpu) 576 tick_do_update_jiffies64(now); 577 578 /* 579 * Do not call, when we are not in irq context and have 580 * no valid regs pointer 581 */ 582 if (regs) { 583 /* 584 * When we are idle and the tick is stopped, we have to touch 585 * the watchdog as we might not schedule for a really long 586 * time. This happens on complete idle SMP systems while 587 * waiting on the login prompt. We also increment the "start of 588 * idle" jiffy stamp so the idle accounting adjustment we do 589 * when we go busy again does not account too much ticks. 590 */ 591 if (ts->tick_stopped) { 592 touch_softlockup_watchdog(); 593 ts->idle_jiffies++; 594 } 595 update_process_times(user_mode(regs)); 596 profile_tick(CPU_PROFILING); 597 } 598 599 /* Do not restart, when we are in the idle loop */ 600 if (ts->tick_stopped) 601 return HRTIMER_NORESTART; 602 603 hrtimer_forward(timer, now, tick_period); 604 605 return HRTIMER_RESTART; 606 } 607 608 /** 609 * tick_setup_sched_timer - setup the tick emulation timer 610 */ 611 void tick_setup_sched_timer(void) 612 { 613 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 614 ktime_t now = ktime_get(); 615 u64 offset; 616 617 /* 618 * Emulate tick processing via per-CPU hrtimers: 619 */ 620 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 621 ts->sched_timer.function = tick_sched_timer; 622 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; 623 624 /* Get the next period (per cpu) */ 625 ts->sched_timer.expires = tick_init_jiffy_update(); 626 offset = ktime_to_ns(tick_period) >> 1; 627 do_div(offset, num_possible_cpus()); 628 offset *= smp_processor_id(); 629 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset); 630 631 for (;;) { 632 hrtimer_forward(&ts->sched_timer, now, tick_period); 633 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires, 634 HRTIMER_MODE_ABS); 635 /* Check, if the timer was already in the past */ 636 if (hrtimer_active(&ts->sched_timer)) 637 break; 638 now = ktime_get(); 639 } 640 641 #ifdef CONFIG_NO_HZ 642 if (tick_nohz_enabled) 643 ts->nohz_mode = NOHZ_MODE_HIGHRES; 644 #endif 645 } 646 #endif /* HIGH_RES_TIMERS */ 647 648 #if defined CONFIG_NO_HZ || defined CONFIG_HIGH_RES_TIMERS 649 void tick_cancel_sched_timer(int cpu) 650 { 651 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 652 653 # ifdef CONFIG_HIGH_RES_TIMERS 654 if (ts->sched_timer.base) 655 hrtimer_cancel(&ts->sched_timer); 656 # endif 657 658 ts->nohz_mode = NOHZ_MODE_INACTIVE; 659 } 660 #endif 661 662 /** 663 * Async notification about clocksource changes 664 */ 665 void tick_clock_notify(void) 666 { 667 int cpu; 668 669 for_each_possible_cpu(cpu) 670 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 671 } 672 673 /* 674 * Async notification about clock event changes 675 */ 676 void tick_oneshot_notify(void) 677 { 678 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 679 680 set_bit(0, &ts->check_clocks); 681 } 682 683 /** 684 * Check, if a change happened, which makes oneshot possible. 685 * 686 * Called cyclic from the hrtimer softirq (driven by the timer 687 * softirq) allow_nohz signals, that we can switch into low-res nohz 688 * mode, because high resolution timers are disabled (either compile 689 * or runtime). 690 */ 691 int tick_check_oneshot_change(int allow_nohz) 692 { 693 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 694 695 if (!test_and_clear_bit(0, &ts->check_clocks)) 696 return 0; 697 698 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 699 return 0; 700 701 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 702 return 0; 703 704 if (!allow_nohz) 705 return 1; 706 707 tick_nohz_switch_to_nohz(); 708 return 0; 709 } 710