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 cpu_clear(cpu, nohz_cpu_mask); 137 now = ktime_get(); 138 ts->idle_waketime = now; 139 140 local_irq_save(flags); 141 tick_do_update_jiffies64(now); 142 local_irq_restore(flags); 143 144 touch_softlockup_watchdog(); 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 } 280 281 /* 282 * If this cpu is the one which updates jiffies, then 283 * give up the assignment and let it be taken by the 284 * cpu which runs the tick timer next, which might be 285 * this cpu as well. If we don't drop this here the 286 * jiffies might be stale and do_timer() never 287 * invoked. 288 */ 289 if (cpu == tick_do_timer_cpu) 290 tick_do_timer_cpu = -1; 291 292 ts->idle_sleeps++; 293 294 /* 295 * delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that 296 * there is no timer pending or at least extremly far 297 * into the future (12 days for HZ=1000). In this case 298 * we simply stop the tick timer: 299 */ 300 if (unlikely(delta_jiffies >= NEXT_TIMER_MAX_DELTA)) { 301 ts->idle_expires.tv64 = KTIME_MAX; 302 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 303 hrtimer_cancel(&ts->sched_timer); 304 goto out; 305 } 306 307 /* 308 * calculate the expiry time for the next timer wheel 309 * timer 310 */ 311 expires = ktime_add_ns(last_update, tick_period.tv64 * 312 delta_jiffies); 313 ts->idle_expires = expires; 314 315 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 316 hrtimer_start(&ts->sched_timer, expires, 317 HRTIMER_MODE_ABS); 318 /* Check, if the timer was already in the past */ 319 if (hrtimer_active(&ts->sched_timer)) 320 goto out; 321 } else if (!tick_program_event(expires, 0)) 322 goto out; 323 /* 324 * We are past the event already. So we crossed a 325 * jiffie boundary. Update jiffies and raise the 326 * softirq. 327 */ 328 tick_do_update_jiffies64(ktime_get()); 329 cpu_clear(cpu, nohz_cpu_mask); 330 } 331 raise_softirq_irqoff(TIMER_SOFTIRQ); 332 out: 333 ts->next_jiffies = next_jiffies; 334 ts->last_jiffies = last_jiffies; 335 ts->sleep_length = ktime_sub(dev->next_event, now); 336 end: 337 local_irq_restore(flags); 338 } 339 340 /** 341 * tick_nohz_get_sleep_length - return the length of the current sleep 342 * 343 * Called from power state control code with interrupts disabled 344 */ 345 ktime_t tick_nohz_get_sleep_length(void) 346 { 347 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 348 349 return ts->sleep_length; 350 } 351 352 /** 353 * tick_nohz_restart_sched_tick - restart the idle tick from the idle task 354 * 355 * Restart the idle tick when the CPU is woken up from idle 356 */ 357 void tick_nohz_restart_sched_tick(void) 358 { 359 int cpu = smp_processor_id(); 360 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 361 unsigned long ticks; 362 ktime_t now; 363 364 local_irq_disable(); 365 tick_nohz_stop_idle(cpu); 366 367 if (!ts->tick_stopped) { 368 local_irq_enable(); 369 return; 370 } 371 372 rcu_exit_nohz(); 373 374 /* Update jiffies first */ 375 select_nohz_load_balancer(0); 376 now = ktime_get(); 377 tick_do_update_jiffies64(now); 378 cpu_clear(cpu, nohz_cpu_mask); 379 380 /* 381 * We stopped the tick in idle. Update process times would miss the 382 * time we slept as update_process_times does only a 1 tick 383 * accounting. Enforce that this is accounted to idle ! 384 */ 385 ticks = jiffies - ts->idle_jiffies; 386 /* 387 * We might be one off. Do not randomly account a huge number of ticks! 388 */ 389 if (ticks && ticks < LONG_MAX) { 390 add_preempt_count(HARDIRQ_OFFSET); 391 account_system_time(current, HARDIRQ_OFFSET, 392 jiffies_to_cputime(ticks)); 393 sub_preempt_count(HARDIRQ_OFFSET); 394 } 395 396 touch_softlockup_watchdog(); 397 /* 398 * Cancel the scheduled timer and restore the tick 399 */ 400 ts->tick_stopped = 0; 401 ts->idle_exittime = now; 402 hrtimer_cancel(&ts->sched_timer); 403 ts->sched_timer.expires = ts->idle_tick; 404 405 while (1) { 406 /* Forward the time to expire in the future */ 407 hrtimer_forward(&ts->sched_timer, now, tick_period); 408 409 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) { 410 hrtimer_start(&ts->sched_timer, 411 ts->sched_timer.expires, 412 HRTIMER_MODE_ABS); 413 /* Check, if the timer was already in the past */ 414 if (hrtimer_active(&ts->sched_timer)) 415 break; 416 } else { 417 if (!tick_program_event(ts->sched_timer.expires, 0)) 418 break; 419 } 420 /* Update jiffies and reread time */ 421 tick_do_update_jiffies64(now); 422 now = ktime_get(); 423 } 424 local_irq_enable(); 425 } 426 427 static int tick_nohz_reprogram(struct tick_sched *ts, ktime_t now) 428 { 429 hrtimer_forward(&ts->sched_timer, now, tick_period); 430 return tick_program_event(ts->sched_timer.expires, 0); 431 } 432 433 /* 434 * The nohz low res interrupt handler 435 */ 436 static void tick_nohz_handler(struct clock_event_device *dev) 437 { 438 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 439 struct pt_regs *regs = get_irq_regs(); 440 int cpu = smp_processor_id(); 441 ktime_t now = ktime_get(); 442 443 dev->next_event.tv64 = KTIME_MAX; 444 445 /* 446 * Check if the do_timer duty was dropped. We don't care about 447 * concurrency: This happens only when the cpu in charge went 448 * into a long sleep. If two cpus happen to assign themself to 449 * this duty, then the jiffies update is still serialized by 450 * xtime_lock. 451 */ 452 if (unlikely(tick_do_timer_cpu == -1)) 453 tick_do_timer_cpu = cpu; 454 455 /* Check, if the jiffies need an update */ 456 if (tick_do_timer_cpu == cpu) 457 tick_do_update_jiffies64(now); 458 459 /* 460 * When we are idle and the tick is stopped, we have to touch 461 * the watchdog as we might not schedule for a really long 462 * time. This happens on complete idle SMP systems while 463 * waiting on the login prompt. We also increment the "start 464 * of idle" jiffy stamp so the idle accounting adjustment we 465 * do when we go busy again does not account too much ticks. 466 */ 467 if (ts->tick_stopped) { 468 touch_softlockup_watchdog(); 469 ts->idle_jiffies++; 470 } 471 472 update_process_times(user_mode(regs)); 473 profile_tick(CPU_PROFILING); 474 475 /* Do not restart, when we are in the idle loop */ 476 if (ts->tick_stopped) 477 return; 478 479 while (tick_nohz_reprogram(ts, now)) { 480 now = ktime_get(); 481 tick_do_update_jiffies64(now); 482 } 483 } 484 485 /** 486 * tick_nohz_switch_to_nohz - switch to nohz mode 487 */ 488 static void tick_nohz_switch_to_nohz(void) 489 { 490 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 491 ktime_t next; 492 493 if (!tick_nohz_enabled) 494 return; 495 496 local_irq_disable(); 497 if (tick_switch_to_oneshot(tick_nohz_handler)) { 498 local_irq_enable(); 499 return; 500 } 501 502 ts->nohz_mode = NOHZ_MODE_LOWRES; 503 504 /* 505 * Recycle the hrtimer in ts, so we can share the 506 * hrtimer_forward with the highres code. 507 */ 508 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 509 /* Get the next period */ 510 next = tick_init_jiffy_update(); 511 512 for (;;) { 513 ts->sched_timer.expires = next; 514 if (!tick_program_event(next, 0)) 515 break; 516 next = ktime_add(next, tick_period); 517 } 518 local_irq_enable(); 519 520 printk(KERN_INFO "Switched to NOHz mode on CPU #%d\n", 521 smp_processor_id()); 522 } 523 524 #else 525 526 static inline void tick_nohz_switch_to_nohz(void) { } 527 528 #endif /* NO_HZ */ 529 530 /* 531 * High resolution timer specific code 532 */ 533 #ifdef CONFIG_HIGH_RES_TIMERS 534 /* 535 * We rearm the timer until we get disabled by the idle code. 536 * Called with interrupts disabled and timer->base->cpu_base->lock held. 537 */ 538 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 539 { 540 struct tick_sched *ts = 541 container_of(timer, struct tick_sched, sched_timer); 542 struct pt_regs *regs = get_irq_regs(); 543 ktime_t now = ktime_get(); 544 int cpu = smp_processor_id(); 545 546 #ifdef CONFIG_NO_HZ 547 /* 548 * Check if the do_timer duty was dropped. We don't care about 549 * concurrency: This happens only when the cpu in charge went 550 * into a long sleep. If two cpus happen to assign themself to 551 * this duty, then the jiffies update is still serialized by 552 * xtime_lock. 553 */ 554 if (unlikely(tick_do_timer_cpu == -1)) 555 tick_do_timer_cpu = cpu; 556 #endif 557 558 /* Check, if the jiffies need an update */ 559 if (tick_do_timer_cpu == cpu) 560 tick_do_update_jiffies64(now); 561 562 /* 563 * Do not call, when we are not in irq context and have 564 * no valid regs pointer 565 */ 566 if (regs) { 567 /* 568 * When we are idle and the tick is stopped, we have to touch 569 * the watchdog as we might not schedule for a really long 570 * time. This happens on complete idle SMP systems while 571 * waiting on the login prompt. We also increment the "start of 572 * idle" jiffy stamp so the idle accounting adjustment we do 573 * when we go busy again does not account too much ticks. 574 */ 575 if (ts->tick_stopped) { 576 touch_softlockup_watchdog(); 577 ts->idle_jiffies++; 578 } 579 update_process_times(user_mode(regs)); 580 profile_tick(CPU_PROFILING); 581 } 582 583 /* Do not restart, when we are in the idle loop */ 584 if (ts->tick_stopped) 585 return HRTIMER_NORESTART; 586 587 hrtimer_forward(timer, now, tick_period); 588 589 return HRTIMER_RESTART; 590 } 591 592 /** 593 * tick_setup_sched_timer - setup the tick emulation timer 594 */ 595 void tick_setup_sched_timer(void) 596 { 597 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 598 ktime_t now = ktime_get(); 599 u64 offset; 600 601 /* 602 * Emulate tick processing via per-CPU hrtimers: 603 */ 604 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 605 ts->sched_timer.function = tick_sched_timer; 606 ts->sched_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; 607 608 /* Get the next period (per cpu) */ 609 ts->sched_timer.expires = tick_init_jiffy_update(); 610 offset = ktime_to_ns(tick_period) >> 1; 611 do_div(offset, num_possible_cpus()); 612 offset *= smp_processor_id(); 613 ts->sched_timer.expires = ktime_add_ns(ts->sched_timer.expires, offset); 614 615 for (;;) { 616 hrtimer_forward(&ts->sched_timer, now, tick_period); 617 hrtimer_start(&ts->sched_timer, ts->sched_timer.expires, 618 HRTIMER_MODE_ABS); 619 /* Check, if the timer was already in the past */ 620 if (hrtimer_active(&ts->sched_timer)) 621 break; 622 now = ktime_get(); 623 } 624 625 #ifdef CONFIG_NO_HZ 626 if (tick_nohz_enabled) 627 ts->nohz_mode = NOHZ_MODE_HIGHRES; 628 #endif 629 } 630 631 void tick_cancel_sched_timer(int cpu) 632 { 633 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 634 635 if (ts->sched_timer.base) 636 hrtimer_cancel(&ts->sched_timer); 637 638 ts->nohz_mode = NOHZ_MODE_INACTIVE; 639 } 640 #endif /* HIGH_RES_TIMERS */ 641 642 /** 643 * Async notification about clocksource changes 644 */ 645 void tick_clock_notify(void) 646 { 647 int cpu; 648 649 for_each_possible_cpu(cpu) 650 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 651 } 652 653 /* 654 * Async notification about clock event changes 655 */ 656 void tick_oneshot_notify(void) 657 { 658 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 659 660 set_bit(0, &ts->check_clocks); 661 } 662 663 /** 664 * Check, if a change happened, which makes oneshot possible. 665 * 666 * Called cyclic from the hrtimer softirq (driven by the timer 667 * softirq) allow_nohz signals, that we can switch into low-res nohz 668 * mode, because high resolution timers are disabled (either compile 669 * or runtime). 670 */ 671 int tick_check_oneshot_change(int allow_nohz) 672 { 673 struct tick_sched *ts = &__get_cpu_var(tick_cpu_sched); 674 675 if (!test_and_clear_bit(0, &ts->check_clocks)) 676 return 0; 677 678 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 679 return 0; 680 681 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 682 return 0; 683 684 if (!allow_nohz) 685 return 1; 686 687 tick_nohz_switch_to_nohz(); 688 return 0; 689 } 690