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 /* Reevalute with xtime_lock held */ 52 write_seqlock(&xtime_lock); 53 54 delta = ktime_sub(now, last_jiffies_update); 55 if (delta.tv64 >= tick_period.tv64) { 56 57 delta = ktime_sub(delta, tick_period); 58 last_jiffies_update = ktime_add(last_jiffies_update, 59 tick_period); 60 61 /* Slow path for long timeouts */ 62 if (unlikely(delta.tv64 >= tick_period.tv64)) { 63 s64 incr = ktime_to_ns(tick_period); 64 65 ticks = ktime_divns(delta, incr); 66 67 last_jiffies_update = ktime_add_ns(last_jiffies_update, 68 incr * ticks); 69 } 70 do_timer(++ticks); 71 } 72 write_sequnlock(&xtime_lock); 73 } 74 75 /* 76 * Initialize and return retrieve the jiffies update. 77 */ 78 static ktime_t tick_init_jiffy_update(void) 79 { 80 ktime_t period; 81 82 write_seqlock(&xtime_lock); 83 /* Did we start the jiffies update yet ? */ 84 if (last_jiffies_update.tv64 == 0) 85 last_jiffies_update = tick_next_period; 86 period = last_jiffies_update; 87 write_sequnlock(&xtime_lock); 88 return period; 89 } 90 91 /* 92 * NOHZ - aka dynamic tick functionality 93 */ 94 #ifdef CONFIG_NO_HZ 95 /* 96 * NO HZ enabled ? 97 */ 98 static int tick_nohz_enabled __read_mostly = 1; 99 100 /* 101 * Enable / Disable tickless mode 102 */ 103 static int __init setup_tick_nohz(char *str) 104 { 105 if (!strcmp(str, "off")) 106 tick_nohz_enabled = 0; 107 else if (!strcmp(str, "on")) 108 tick_nohz_enabled = 1; 109 else 110 return 0; 111 return 1; 112 } 113 114 __setup("nohz=", setup_tick_nohz); 115 116 /** 117 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 118 * 119 * Called from interrupt entry when the CPU was idle 120 * 121 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 122 * must be updated. Otherwise an interrupt handler could use a stale jiffy 123 * value. We do this unconditionally on any cpu, as we don't know whether the 124 * cpu, which has the update task assigned is in a long sleep. 125 */ 126 void tick_nohz_update_jiffies(void) 127 { 128 int cpu = smp_processor_id(); 129 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 130 unsigned long flags; 131 ktime_t now; 132 133 if (!ts->tick_stopped) 134 return; 135 136 touch_softlockup_watchdog(); 137 138 cpu_clear(cpu, nohz_cpu_mask); 139 now = ktime_get(); 140 ts->idle_waketime = now; 141 142 local_irq_save(flags); 143 tick_do_update_jiffies64(now); 144 local_irq_restore(flags); 145 } 146 147 void tick_nohz_stop_idle(int cpu) 148 { 149 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 150 151 if (ts->idle_active) { 152 ktime_t now, delta; 153 now = ktime_get(); 154 delta = ktime_sub(now, ts->idle_entrytime); 155 ts->idle_lastupdate = now; 156 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 157 ts->idle_active = 0; 158 } 159 } 160 161 static ktime_t tick_nohz_start_idle(int cpu) 162 { 163 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 164 ktime_t now, delta; 165 166 now = ktime_get(); 167 if (ts->idle_active) { 168 delta = ktime_sub(now, ts->idle_entrytime); 169 ts->idle_lastupdate = now; 170 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 171 } 172 ts->idle_entrytime = now; 173 ts->idle_active = 1; 174 return now; 175 } 176 177 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 178 { 179 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 180 181 *last_update_time = ktime_to_us(ts->idle_lastupdate); 182 return ktime_to_us(ts->idle_sleeptime); 183 } 184 185 /** 186 * tick_nohz_stop_sched_tick - stop the idle tick from the idle task 187 * 188 * When the next event is more than a tick into the future, stop the idle tick 189 * Called either from the idle loop or from irq_exit() when an idle period was 190 * just interrupted by an interrupt which did not cause a reschedule. 191 */ 192 void tick_nohz_stop_sched_tick(void) 193 { 194 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies, flags; 195 unsigned long rt_jiffies; 196 struct tick_sched *ts; 197 ktime_t last_update, expires, now; 198 struct clock_event_device *dev = __get_cpu_var(tick_cpu_device).evtdev; 199 int cpu; 200 201 local_irq_save(flags); 202 203 cpu = smp_processor_id(); 204 now = tick_nohz_start_idle(cpu); 205 ts = &per_cpu(tick_cpu_sched, cpu); 206 207 /* 208 * If this cpu is offline and it is the one which updates 209 * jiffies, then give up the assignment and let it be taken by 210 * the cpu which runs the tick timer next. If we don't drop 211 * this here the jiffies might be stale and do_timer() never 212 * invoked. 213 */ 214 if (unlikely(!cpu_online(cpu))) { 215 if (cpu == tick_do_timer_cpu) 216 tick_do_timer_cpu = -1; 217 } 218 219 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) 220 goto end; 221 222 if (need_resched()) 223 goto end; 224 225 cpu = smp_processor_id(); 226 if (unlikely(local_softirq_pending())) { 227 static int ratelimit; 228 229 if (ratelimit < 10) { 230 printk(KERN_ERR "NOHZ: local_softirq_pending %02x\n", 231 local_softirq_pending()); 232 ratelimit++; 233 } 234 } 235 236 ts->idle_calls++; 237 /* Read jiffies and the time when jiffies were updated last */ 238 do { 239 seq = read_seqbegin(&xtime_lock); 240 last_update = last_jiffies_update; 241 last_jiffies = jiffies; 242 } while (read_seqretry(&xtime_lock, seq)); 243 244 /* Get the next timer wheel timer */ 245 next_jiffies = get_next_timer_interrupt(last_jiffies); 246 delta_jiffies = next_jiffies - last_jiffies; 247 248 rt_jiffies = rt_needs_cpu(cpu); 249 if (rt_jiffies && rt_jiffies < delta_jiffies) 250 delta_jiffies = rt_jiffies; 251 252 if (rcu_needs_cpu(cpu)) 253 delta_jiffies = 1; 254 /* 255 * Do not stop the tick, if we are only one off 256 * or if the cpu is required for rcu 257 */ 258 if (!ts->tick_stopped && delta_jiffies == 1) 259 goto out; 260 261 /* Schedule the tick, if we are at least one jiffie off */ 262 if ((long)delta_jiffies >= 1) { 263 264 if (delta_jiffies > 1) 265 cpu_set(cpu, nohz_cpu_mask); 266 /* 267 * nohz_stop_sched_tick can be called several times before 268 * the nohz_restart_sched_tick is called. This happens when 269 * interrupts arrive which do not cause a reschedule. In the 270 * first call we save the current tick time, so we can restart 271 * the scheduler tick in nohz_restart_sched_tick. 272 */ 273 if (!ts->tick_stopped) { 274 if (select_nohz_load_balancer(1)) { 275 /* 276 * sched tick not stopped! 277 */ 278 cpu_clear(cpu, nohz_cpu_mask); 279 goto out; 280 } 281 282 ts->idle_tick = ts->sched_timer.expires; 283 ts->tick_stopped = 1; 284 ts->idle_jiffies = last_jiffies; 285 rcu_enter_nohz(); 286 } 287 288 /* 289 * If this cpu is the one which updates jiffies, then 290 * give up the assignment and let it be taken by the 291 * cpu which runs the tick timer next, which might be 292 * this cpu as well. If we don't drop this here the 293 * jiffies might be stale and do_timer() never 294 * invoked. 295 */ 296 if (cpu == tick_do_timer_cpu) 297 tick_do_timer_cpu = -1; 298 299 ts->idle_sleeps++; 300 301 /* 302 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that 303 * there is no timer pending or at least extremly far 304 * into the future (12 days for HZ=1000). In this case 305 * we simply stop the tick timer: 306 */ 307 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) { 308 ts->idle_expires.tv64 = KTIME_MAX; 309 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 310 hrtimer_cancel(&ts->sched_timer); 311 goto out; 312 } 313 314 /* 315 * calculate the expiry time for the next timer wheel 316 * timer 317 */ 318 expires = ktime_add_ns(last_update, tick_period.tv64 * 319 delta_jiffies); 320 ts->idle_expires = expires; 321 322 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 323 hrtimer_start(&ts->sched_timer, expires, 324 HRTIMER_MODE_ABS); 325 /* Check, if the timer was already in the past */ 326 if (hrtimer_active(&ts->sched_timer)) 327 goto out; 328 } else if (!tick_program_event(expires, 0)) 329 goto out; 330 /* 331 * We are past the event already. So we crossed a 332 * jiffie boundary. Update jiffies and raise the 333 * softirq. 334 */ 335 tick_do_update_jiffies64(ktime_get()); 336 cpu_clear(cpu, nohz_cpu_mask); 337 } 338 raise_softirq_irqoff(TIMER_SOFTIRQ); 339 out: 340 ts->next_jiffies = next_jiffies; 341 ts->last_jiffies = last_jiffies; 342 ts->sleep_length = ktime_sub(dev->next_event, now); 343 end: 344 local_irq_restore(flags); 345 } 346 347 /** 348 * tick_nohz_get_sleep_length - return the length of the current sleep 349 * 350 * Called from power state control code with interrupts disabled 351 */ 352 ktime_t tick_nohz_get_sleep_length(void) 353 { 354 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 355 356 return ts->sleep_length; 357 } 358 359 /** 360 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task 361 * 362 * Restart the idle tick when the CPU is woken up from idle 363 */ 364 void tick_nohz_restart_sched_tick(void) 365 { 366 int cpu = smp_processor_id(); 367 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 368 unsigned long ticks; 369 ktime_t now; 370 371 local_irq_disable(); 372 tick_nohz_stop_idle(cpu); 373 374 if (!ts->tick_stopped) { 375 local_irq_enable(); 376 return; 377 } 378 379 rcu_exit_nohz(); 380 381 /* Update jiffies first */ 382 select_nohz_load_balancer(0); 383 now = ktime_get(); 384 tick_do_update_jiffies64(now); 385 cpu_clear(cpu, nohz_cpu_mask); 386 387 /* 388 * We stopped the tick in idle. Update process times would miss the 389 * time we slept as update_process_times does only a 1 tick 390 * accounting. Enforce that this is accounted to idle ! 391 */ 392 ticks = jiffies - ts->idle_jiffies; 393 /* 394 * We might be one off. Do not randomly account a huge number of ticks! 395 */ 396 if (ticks && ticks < LONG_MAX) { 397 add_preempt_count(HARDIRQ_OFFSET); 398 account_system_time(current, HARDIRQ_OFFSET, 399 jiffies_to_cputime(ticks)); 400 sub_preempt_count(HARDIRQ_OFFSET); 401 } 402 403 /* 404 * Cancel the scheduled timer and restore the tick 405 */ 406 ts->tick_stopped = 0; 407 ts->idle_exittime = now; 408 hrtimer_cancel(&ts->sched_timer); 409 ts->sched_timer.expires = ts->idle_tick; 410 411 while (1) { 412 /* Forward the time to expire in the future */ 413 hrtimer_forward(&ts->sched_timer, now, tick_period); 414 415 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 416 hrtimer_start(&ts->sched_timer, 417 ts->sched_timer.expires, 418 HRTIMER_MODE_ABS); 419 /* Check, if the timer was already in the past */ 420 if (hrtimer_active(&ts->sched_timer)) 421 break; 422 } else { 423 if (!tick_program_event(ts->sched_timer.expires, 0)) 424 break; 425 } 426 /* Update jiffies and reread time */ 427 tick_do_update_jiffies64(now); 428 now = ktime_get(); 429 } 430 local_irq_enable(); 431 } 432 433 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 434 { 435 hrtimer_forward(&ts->sched_timer, now, tick_period); 436 return tick_program_event(ts->sched_timer.expires, 0); 437 } 438 439 /* 440 * The nohz low res interrupt handler 441 */ 442 static void tick_nohz_handler(struct clock_event_device *dev) 443 { 444 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 445 struct pt_regs *regs = get_irq_regs(); 446 int cpu = smp_processor_id(); 447 ktime_t now = ktime_get(); 448 449 dev->next_event.tv64 = KTIME_MAX; 450 451 /* 452 * Check if the do_timer duty was dropped. We don't care about 453 * concurrency: This happens only when the cpu in charge went 454 * into a long sleep. If two cpus happen to assign themself to 455 * this duty, then the jiffies update is still serialized by 456 * xtime_lock. 457 */ 458 if (unlikely(tick_do_timer_cpu == -1)) 459 tick_do_timer_cpu = cpu; 460 461 /* Check, if the jiffies need an update */ 462 if (tick_do_timer_cpu == cpu) 463 tick_do_update_jiffies64(now); 464 465 /* 466 * When we are idle and the tick is stopped, we have to touch 467 * the watchdog as we might not schedule for a really long 468 * time. This happens on complete idle SMP systems while 469 * waiting on the login prompt. We also increment the "start 470 * of idle" jiffy stamp so the idle accounting adjustment we 471 * do when we go busy again does not account too much ticks. 472 */ 473 if (ts->tick_stopped) { 474 touch_softlockup_watchdog(); 475 ts->idle_jiffies++; 476 } 477 478 update_process_times(user_mode(regs)); 479 profile_tick(CPU_PROFILING); 480 481 /* Do not restart, when we are in the idle loop */ 482 if (ts->tick_stopped) 483 return; 484 485 while (tick_nohz_reprogram(ts, now)) { 486 now = ktime_get(); 487 tick_do_update_jiffies64(now); 488 } 489 } 490 491 /** 492 * tick_nohz_switch_to_nohz - switch to nohz mode 493 */ 494 static void tick_nohz_switch_to_nohz(void) 495 { 496 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 497 ktime_t next; 498 499 if (!tick_nohz_enabled) 500 return; 501 502 local_irq_disable(); 503 if (tick_switch_to_oneshot(tick_nohz_handler)) { 504 local_irq_enable(); 505 return; 506 } 507 508 ts->nohz_mode = NOHZ_MODE_LOWRES; 509 510 /* 511 * Recycle the hrtimer in ts, so we can share the 512 * hrtimer_forward with the highres code. 513 */ 514 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 515 /* Get the next period */ 516 next = tick_init_jiffy_update(); 517 518 for (;;) { 519 ts->sched_timer.expires = next; 520 if (!tick_program_event(next, 0)) 521 break; 522 next = ktime_add(next, tick_period); 523 } 524 local_irq_enable(); 525 526 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", 527 smp_processor_id()); 528 } 529 530 #else 531 532 static inline void tick_nohz_switch_to_nohz(void) { } 533 534 #endif /* NO_HZ */ 535 536 /* 537 * High resolution timer specific code 538 */ 539 #ifdef CONFIG_HIGH_RES_TIMERS 540 /* 541 * We rearm the timer until we get disabled by the idle code. 542 * Called with interrupts disabled and timer->base->cpu_base->lock held. 543 */ 544 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 545 { 546 struct tick_sched *ts = 547 container_of(timer, struct tick_sched, sched_timer); 548 struct pt_regs *regs = get_irq_regs(); 549 ktime_t now = ktime_get(); 550 int cpu = smp_processor_id(); 551 552 #ifdef CONFIG_NO_HZ 553 /* 554 * Check if the do_timer duty was dropped. We don't care about 555 * concurrency: This happens only when the cpu in charge went 556 * into a long sleep. If two cpus happen to assign themself to 557 * this duty, then the jiffies update is still serialized by 558 * xtime_lock. 559 */ 560 if (unlikely(tick_do_timer_cpu == -1)) 561 tick_do_timer_cpu = cpu; 562 #endif 563 564 /* Check, if the jiffies need an update */ 565 if (tick_do_timer_cpu == cpu) 566 tick_do_update_jiffies64(now); 567 568 /* 569 * Do not call, when we are not in irq context and have 570 * no valid regs pointer 571 */ 572 if (regs) { 573 /* 574 * When we are idle and the tick is stopped, we have to touch 575 * the watchdog as we might not schedule for a really long 576 * time. This happens on complete idle SMP systems while 577 * waiting on the login prompt. We also increment the "start of 578 * idle" jiffy stamp so the idle accounting adjustment we do 579 * when we go busy again does not account too much ticks. 580 */ 581 if (ts->tick_stopped) { 582 touch_softlockup_watchdog(); 583 ts->idle_jiffies++; 584 } 585 update_process_times(user_mode(regs)); 586 profile_tick(CPU_PROFILING); 587 } 588 589 /* Do not restart, when we are in the idle loop */ 590 if (ts->tick_stopped) 591 return HRTIMER_NORESTART; 592 593 hrtimer_forward(timer, now, tick_period); 594 595 return HRTIMER_RESTART; 596 } 597 598 /** 599 * tick_setup_sched_timer - setup the tick emulation timer 600 */ 601 void tick_setup_sched_timer(void) 602 { 603 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 604 ktime_t now = ktime_get(); 605 u64 offset; 606 607 /* 608 * Emulate tick processing via per-CPU hrtimers: 609 */ 610 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 611 ts->sched_timer.function = tick_sched_timer; 612 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; 613 614 /* Get the next period (per cpu) */ 615 ts->sched_timer.expires = tick_init_jiffy_update(); 616 offset = ktime_to_ns(tick_period) >> 1; 617 do_div(offset, num_possible_cpus()); 618 offset *= smp_processor_id(); 619 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset); 620 621 for (;;) { 622 hrtimer_forward(&ts->sched_timer, now, tick_period); 623 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires, 624 HRTIMER_MODE_ABS); 625 /* Check, if the timer was already in the past */ 626 if (hrtimer_active(&ts->sched_timer)) 627 break; 628 now = ktime_get(); 629 } 630 631 #ifdef CONFIG_NO_HZ 632 if (tick_nohz_enabled) 633 ts->nohz_mode = NOHZ_MODE_HIGHRES; 634 #endif 635 } 636 637 void tick_cancel_sched_timer(int cpu) 638 { 639 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 640 641 if (ts->sched_timer.base) 642 hrtimer_cancel(&ts->sched_timer); 643 644 ts->nohz_mode = NOHZ_MODE_INACTIVE; 645 } 646 #endif /* HIGH_RES_TIMERS */ 647 648 /** 649 * Async notification about clocksource changes 650 */ 651 void tick_clock_notify(void) 652 { 653 int cpu; 654 655 for_each_possible_cpu(cpu) 656 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 657 } 658 659 /* 660 * Async notification about clock event changes 661 */ 662 void tick_oneshot_notify(void) 663 { 664 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 665 666 set_bit(0, &ts->check_clocks); 667 } 668 669 /** 670 * Check, if a change happened, which makes oneshot possible. 671 * 672 * Called cyclic from the hrtimer softirq (driven by the timer 673 * softirq) allow_nohz signals, that we can switch into low-res nohz 674 * mode, because high resolution timers are disabled (either compile 675 * or runtime). 676 */ 677 int tick_check_oneshot_change(int allow_nohz) 678 { 679 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 680 681 if (!test_and_clear_bit(0, &ts->check_clocks)) 682 return 0; 683 684 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 685 return 0; 686 687 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 688 return 0; 689 690 if (!allow_nohz) 691 return 1; 692 693 tick_nohz_switch_to_nohz(); 694 return 0; 695 } 696