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