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/module.h> 23 #include <linux/irq_work.h> 24 #include <linux/posix-timers.h> 25 #include <linux/perf_event.h> 26 #include <linux/context_tracking.h> 27 28 #include <asm/irq_regs.h> 29 30 #include "tick-internal.h" 31 32 #include <trace/events/timer.h> 33 34 /* 35 * Per cpu nohz control structure 36 */ 37 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); 38 39 /* 40 * The time, when the last jiffy update happened. Protected by jiffies_lock. 41 */ 42 static ktime_t last_jiffies_update; 43 44 struct tick_sched *tick_get_tick_sched(int cpu) 45 { 46 return &per_cpu(tick_cpu_sched, cpu); 47 } 48 49 /* 50 * Must be called with interrupts disabled ! 51 */ 52 static void tick_do_update_jiffies64(ktime_t now) 53 { 54 unsigned long ticks = 0; 55 ktime_t delta; 56 57 /* 58 * Do a quick check without holding jiffies_lock: 59 */ 60 delta = ktime_sub(now, last_jiffies_update); 61 if (delta.tv64 < tick_period.tv64) 62 return; 63 64 /* Reevalute with jiffies_lock held */ 65 write_seqlock(&jiffies_lock); 66 67 delta = ktime_sub(now, last_jiffies_update); 68 if (delta.tv64 >= tick_period.tv64) { 69 70 delta = ktime_sub(delta, tick_period); 71 last_jiffies_update = ktime_add(last_jiffies_update, 72 tick_period); 73 74 /* Slow path for long timeouts */ 75 if (unlikely(delta.tv64 >= tick_period.tv64)) { 76 s64 incr = ktime_to_ns(tick_period); 77 78 ticks = ktime_divns(delta, incr); 79 80 last_jiffies_update = ktime_add_ns(last_jiffies_update, 81 incr * ticks); 82 } 83 do_timer(++ticks); 84 85 /* Keep the tick_next_period variable up to date */ 86 tick_next_period = ktime_add(last_jiffies_update, tick_period); 87 } else { 88 write_sequnlock(&jiffies_lock); 89 return; 90 } 91 write_sequnlock(&jiffies_lock); 92 update_wall_time(); 93 } 94 95 /* 96 * Initialize and return retrieve the jiffies update. 97 */ 98 static ktime_t tick_init_jiffy_update(void) 99 { 100 ktime_t period; 101 102 write_seqlock(&jiffies_lock); 103 /* Did we start the jiffies update yet ? */ 104 if (last_jiffies_update.tv64 == 0) 105 last_jiffies_update = tick_next_period; 106 period = last_jiffies_update; 107 write_sequnlock(&jiffies_lock); 108 return period; 109 } 110 111 112 static void tick_sched_do_timer(ktime_t now) 113 { 114 int cpu = smp_processor_id(); 115 116 #ifdef CONFIG_NO_HZ_COMMON 117 /* 118 * Check if the do_timer duty was dropped. We don't care about 119 * concurrency: This happens only when the cpu in charge went 120 * into a long sleep. If two cpus happen to assign themself to 121 * this duty, then the jiffies update is still serialized by 122 * jiffies_lock. 123 */ 124 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE) 125 && !tick_nohz_full_cpu(cpu)) 126 tick_do_timer_cpu = cpu; 127 #endif 128 129 /* Check, if the jiffies need an update */ 130 if (tick_do_timer_cpu == cpu) 131 tick_do_update_jiffies64(now); 132 } 133 134 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) 135 { 136 #ifdef CONFIG_NO_HZ_COMMON 137 /* 138 * When we are idle and the tick is stopped, we have to touch 139 * the watchdog as we might not schedule for a really long 140 * time. This happens on complete idle SMP systems while 141 * waiting on the login prompt. We also increment the "start of 142 * idle" jiffy stamp so the idle accounting adjustment we do 143 * when we go busy again does not account too much ticks. 144 */ 145 if (ts->tick_stopped) { 146 touch_softlockup_watchdog(); 147 if (is_idle_task(current)) 148 ts->idle_jiffies++; 149 } 150 #endif 151 update_process_times(user_mode(regs)); 152 profile_tick(CPU_PROFILING); 153 } 154 155 #ifdef CONFIG_NO_HZ_FULL 156 cpumask_var_t tick_nohz_full_mask; 157 cpumask_var_t housekeeping_mask; 158 bool tick_nohz_full_running; 159 160 static bool can_stop_full_tick(void) 161 { 162 WARN_ON_ONCE(!irqs_disabled()); 163 164 if (!sched_can_stop_tick()) { 165 trace_tick_stop(0, "more than 1 task in runqueue\n"); 166 return false; 167 } 168 169 if (!posix_cpu_timers_can_stop_tick(current)) { 170 trace_tick_stop(0, "posix timers running\n"); 171 return false; 172 } 173 174 if (!perf_event_can_stop_tick()) { 175 trace_tick_stop(0, "perf events running\n"); 176 return false; 177 } 178 179 /* sched_clock_tick() needs us? */ 180 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK 181 /* 182 * TODO: kick full dynticks CPUs when 183 * sched_clock_stable is set. 184 */ 185 if (!sched_clock_stable()) { 186 trace_tick_stop(0, "unstable sched clock\n"); 187 /* 188 * Don't allow the user to think they can get 189 * full NO_HZ with this machine. 190 */ 191 WARN_ONCE(tick_nohz_full_running, 192 "NO_HZ FULL will not work with unstable sched clock"); 193 return false; 194 } 195 #endif 196 197 return true; 198 } 199 200 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now); 201 202 /* 203 * Re-evaluate the need for the tick on the current CPU 204 * and restart it if necessary. 205 */ 206 void __tick_nohz_full_check(void) 207 { 208 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 209 210 if (tick_nohz_full_cpu(smp_processor_id())) { 211 if (ts->tick_stopped && !is_idle_task(current)) { 212 if (!can_stop_full_tick()) 213 tick_nohz_restart_sched_tick(ts, ktime_get()); 214 } 215 } 216 } 217 218 static void nohz_full_kick_work_func(struct irq_work *work) 219 { 220 __tick_nohz_full_check(); 221 } 222 223 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 224 .func = nohz_full_kick_work_func, 225 }; 226 227 /* 228 * Kick this CPU if it's full dynticks in order to force it to 229 * re-evaluate its dependency on the tick and restart it if necessary. 230 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), 231 * is NMI safe. 232 */ 233 void tick_nohz_full_kick(void) 234 { 235 if (!tick_nohz_full_cpu(smp_processor_id())) 236 return; 237 238 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work)); 239 } 240 241 /* 242 * Kick the CPU if it's full dynticks in order to force it to 243 * re-evaluate its dependency on the tick and restart it if necessary. 244 */ 245 void tick_nohz_full_kick_cpu(int cpu) 246 { 247 if (!tick_nohz_full_cpu(cpu)) 248 return; 249 250 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); 251 } 252 253 static void nohz_full_kick_ipi(void *info) 254 { 255 __tick_nohz_full_check(); 256 } 257 258 /* 259 * Kick all full dynticks CPUs in order to force these to re-evaluate 260 * their dependency on the tick and restart it if necessary. 261 */ 262 void tick_nohz_full_kick_all(void) 263 { 264 if (!tick_nohz_full_running) 265 return; 266 267 preempt_disable(); 268 smp_call_function_many(tick_nohz_full_mask, 269 nohz_full_kick_ipi, NULL, false); 270 tick_nohz_full_kick(); 271 preempt_enable(); 272 } 273 274 /* 275 * Re-evaluate the need for the tick as we switch the current task. 276 * It might need the tick due to per task/process properties: 277 * perf events, posix cpu timers, ... 278 */ 279 void __tick_nohz_task_switch(struct task_struct *tsk) 280 { 281 unsigned long flags; 282 283 local_irq_save(flags); 284 285 if (!tick_nohz_full_cpu(smp_processor_id())) 286 goto out; 287 288 if (tick_nohz_tick_stopped() && !can_stop_full_tick()) 289 tick_nohz_full_kick(); 290 291 out: 292 local_irq_restore(flags); 293 } 294 295 /* Parse the boot-time nohz CPU list from the kernel parameters. */ 296 static int __init tick_nohz_full_setup(char *str) 297 { 298 alloc_bootmem_cpumask_var(&tick_nohz_full_mask); 299 if (cpulist_parse(str, tick_nohz_full_mask) < 0) { 300 pr_warning("NOHZ: Incorrect nohz_full cpumask\n"); 301 free_bootmem_cpumask_var(tick_nohz_full_mask); 302 return 1; 303 } 304 tick_nohz_full_running = true; 305 306 return 1; 307 } 308 __setup("nohz_full=", tick_nohz_full_setup); 309 310 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb, 311 unsigned long action, 312 void *hcpu) 313 { 314 unsigned int cpu = (unsigned long)hcpu; 315 316 switch (action & ~CPU_TASKS_FROZEN) { 317 case CPU_DOWN_PREPARE: 318 /* 319 * If we handle the timekeeping duty for full dynticks CPUs, 320 * we can't safely shutdown that CPU. 321 */ 322 if (tick_nohz_full_running && tick_do_timer_cpu == cpu) 323 return NOTIFY_BAD; 324 break; 325 } 326 return NOTIFY_OK; 327 } 328 329 static int tick_nohz_init_all(void) 330 { 331 int err = -1; 332 333 #ifdef CONFIG_NO_HZ_FULL_ALL 334 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { 335 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n"); 336 return err; 337 } 338 err = 0; 339 cpumask_setall(tick_nohz_full_mask); 340 tick_nohz_full_running = true; 341 #endif 342 return err; 343 } 344 345 void __init tick_nohz_init(void) 346 { 347 int cpu; 348 349 if (!tick_nohz_full_running) { 350 if (tick_nohz_init_all() < 0) 351 return; 352 } 353 354 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) { 355 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n"); 356 cpumask_clear(tick_nohz_full_mask); 357 tick_nohz_full_running = false; 358 return; 359 } 360 361 /* 362 * Full dynticks uses irq work to drive the tick rescheduling on safe 363 * locking contexts. But then we need irq work to raise its own 364 * interrupts to avoid circular dependency on the tick 365 */ 366 if (!arch_irq_work_has_interrupt()) { 367 pr_warning("NO_HZ: Can't run full dynticks because arch doesn't " 368 "support irq work self-IPIs\n"); 369 cpumask_clear(tick_nohz_full_mask); 370 cpumask_copy(housekeeping_mask, cpu_possible_mask); 371 tick_nohz_full_running = false; 372 return; 373 } 374 375 cpu = smp_processor_id(); 376 377 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 378 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu); 379 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 380 } 381 382 cpumask_andnot(housekeeping_mask, 383 cpu_possible_mask, tick_nohz_full_mask); 384 385 for_each_cpu(cpu, tick_nohz_full_mask) 386 context_tracking_cpu_set(cpu); 387 388 cpu_notifier(tick_nohz_cpu_down_callback, 0); 389 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", 390 cpumask_pr_args(tick_nohz_full_mask)); 391 } 392 #endif 393 394 /* 395 * NOHZ - aka dynamic tick functionality 396 */ 397 #ifdef CONFIG_NO_HZ_COMMON 398 /* 399 * NO HZ enabled ? 400 */ 401 static int tick_nohz_enabled __read_mostly = 1; 402 int tick_nohz_active __read_mostly; 403 /* 404 * Enable / Disable tickless mode 405 */ 406 static int __init setup_tick_nohz(char *str) 407 { 408 if (!strcmp(str, "off")) 409 tick_nohz_enabled = 0; 410 else if (!strcmp(str, "on")) 411 tick_nohz_enabled = 1; 412 else 413 return 0; 414 return 1; 415 } 416 417 __setup("nohz=", setup_tick_nohz); 418 419 int tick_nohz_tick_stopped(void) 420 { 421 return __this_cpu_read(tick_cpu_sched.tick_stopped); 422 } 423 424 /** 425 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 426 * 427 * Called from interrupt entry when the CPU was idle 428 * 429 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 430 * must be updated. Otherwise an interrupt handler could use a stale jiffy 431 * value. We do this unconditionally on any cpu, as we don't know whether the 432 * cpu, which has the update task assigned is in a long sleep. 433 */ 434 static void tick_nohz_update_jiffies(ktime_t now) 435 { 436 unsigned long flags; 437 438 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 439 440 local_irq_save(flags); 441 tick_do_update_jiffies64(now); 442 local_irq_restore(flags); 443 444 touch_softlockup_watchdog(); 445 } 446 447 /* 448 * Updates the per cpu time idle statistics counters 449 */ 450 static void 451 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 452 { 453 ktime_t delta; 454 455 if (ts->idle_active) { 456 delta = ktime_sub(now, ts->idle_entrytime); 457 if (nr_iowait_cpu(cpu) > 0) 458 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 459 else 460 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 461 ts->idle_entrytime = now; 462 } 463 464 if (last_update_time) 465 *last_update_time = ktime_to_us(now); 466 467 } 468 469 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 470 { 471 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 472 ts->idle_active = 0; 473 474 sched_clock_idle_wakeup_event(0); 475 } 476 477 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 478 { 479 ktime_t now = ktime_get(); 480 481 ts->idle_entrytime = now; 482 ts->idle_active = 1; 483 sched_clock_idle_sleep_event(); 484 return now; 485 } 486 487 /** 488 * get_cpu_idle_time_us - get the total idle time of a cpu 489 * @cpu: CPU number to query 490 * @last_update_time: variable to store update time in. Do not update 491 * counters if NULL. 492 * 493 * Return the cummulative idle time (since boot) for a given 494 * CPU, in microseconds. 495 * 496 * This time is measured via accounting rather than sampling, 497 * and is as accurate as ktime_get() is. 498 * 499 * This function returns -1 if NOHZ is not enabled. 500 */ 501 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 502 { 503 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 504 ktime_t now, idle; 505 506 if (!tick_nohz_active) 507 return -1; 508 509 now = ktime_get(); 510 if (last_update_time) { 511 update_ts_time_stats(cpu, ts, now, last_update_time); 512 idle = ts->idle_sleeptime; 513 } else { 514 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 515 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 516 517 idle = ktime_add(ts->idle_sleeptime, delta); 518 } else { 519 idle = ts->idle_sleeptime; 520 } 521 } 522 523 return ktime_to_us(idle); 524 525 } 526 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 527 528 /** 529 * get_cpu_iowait_time_us - get the total iowait time of a cpu 530 * @cpu: CPU number to query 531 * @last_update_time: variable to store update time in. Do not update 532 * counters if NULL. 533 * 534 * Return the cummulative iowait time (since boot) for a given 535 * CPU, in microseconds. 536 * 537 * This time is measured via accounting rather than sampling, 538 * and is as accurate as ktime_get() is. 539 * 540 * This function returns -1 if NOHZ is not enabled. 541 */ 542 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 543 { 544 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 545 ktime_t now, iowait; 546 547 if (!tick_nohz_active) 548 return -1; 549 550 now = ktime_get(); 551 if (last_update_time) { 552 update_ts_time_stats(cpu, ts, now, last_update_time); 553 iowait = ts->iowait_sleeptime; 554 } else { 555 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 556 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 557 558 iowait = ktime_add(ts->iowait_sleeptime, delta); 559 } else { 560 iowait = ts->iowait_sleeptime; 561 } 562 } 563 564 return ktime_to_us(iowait); 565 } 566 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 567 568 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 569 { 570 hrtimer_cancel(&ts->sched_timer); 571 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 572 573 /* Forward the time to expire in the future */ 574 hrtimer_forward(&ts->sched_timer, now, tick_period); 575 576 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 577 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 578 else 579 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 580 } 581 582 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 583 ktime_t now, int cpu) 584 { 585 unsigned long seq, last_jiffies, next_jiffies, delta_jiffies; 586 ktime_t last_update, expires, ret = { .tv64 = 0 }; 587 unsigned long rcu_delta_jiffies; 588 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 589 u64 time_delta; 590 591 time_delta = timekeeping_max_deferment(); 592 593 /* Read jiffies and the time when jiffies were updated last */ 594 do { 595 seq = read_seqbegin(&jiffies_lock); 596 last_update = last_jiffies_update; 597 last_jiffies = jiffies; 598 } while (read_seqretry(&jiffies_lock, seq)); 599 600 if (rcu_needs_cpu(&rcu_delta_jiffies) || 601 arch_needs_cpu() || irq_work_needs_cpu()) { 602 next_jiffies = last_jiffies + 1; 603 delta_jiffies = 1; 604 } else { 605 /* Get the next timer wheel timer */ 606 next_jiffies = get_next_timer_interrupt(last_jiffies); 607 delta_jiffies = next_jiffies - last_jiffies; 608 if (rcu_delta_jiffies < delta_jiffies) { 609 next_jiffies = last_jiffies + rcu_delta_jiffies; 610 delta_jiffies = rcu_delta_jiffies; 611 } 612 } 613 614 if ((long)delta_jiffies <= 1) { 615 if (!ts->tick_stopped) 616 goto out; 617 if (delta_jiffies == 0) { 618 /* Tick is stopped, but required now. Enforce it */ 619 tick_nohz_restart(ts, now); 620 goto out; 621 } 622 } 623 624 /* 625 * If this cpu is the one which updates jiffies, then give up 626 * the assignment and let it be taken by the cpu which runs 627 * the tick timer next, which might be this cpu as well. If we 628 * don't drop this here the jiffies might be stale and 629 * do_timer() never invoked. Keep track of the fact that it 630 * was the one which had the do_timer() duty last. If this cpu 631 * is the one which had the do_timer() duty last, we limit the 632 * sleep time to the timekeeping max_deferement value which we 633 * retrieved above. Otherwise we can sleep as long as we want. 634 */ 635 if (cpu == tick_do_timer_cpu) { 636 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 637 ts->do_timer_last = 1; 638 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 639 time_delta = KTIME_MAX; 640 ts->do_timer_last = 0; 641 } else if (!ts->do_timer_last) { 642 time_delta = KTIME_MAX; 643 } 644 645 #ifdef CONFIG_NO_HZ_FULL 646 if (!ts->inidle) 647 time_delta = min(time_delta, scheduler_tick_max_deferment()); 648 #endif 649 650 /* 651 * calculate the expiry time for the next timer wheel 652 * timer. delta_jiffies >= NEXT_TIMER_MAX_DELTA signals that 653 * there is no timer pending or at least extremely far into 654 * the future (12 days for HZ=1000). In this case we set the 655 * expiry to the end of time. 656 */ 657 if (likely(delta_jiffies < NEXT_TIMER_MAX_DELTA)) { 658 /* 659 * Calculate the time delta for the next timer event. 660 * If the time delta exceeds the maximum time delta 661 * permitted by the current clocksource then adjust 662 * the time delta accordingly to ensure the 663 * clocksource does not wrap. 664 */ 665 time_delta = min_t(u64, time_delta, 666 tick_period.tv64 * delta_jiffies); 667 } 668 669 if (time_delta < KTIME_MAX) 670 expires = ktime_add_ns(last_update, time_delta); 671 else 672 expires.tv64 = KTIME_MAX; 673 674 /* Skip reprogram of event if its not changed */ 675 if (ts->tick_stopped && ktime_equal(expires, dev->next_event)) 676 goto out; 677 678 ret = expires; 679 680 /* 681 * nohz_stop_sched_tick can be called several times before 682 * the nohz_restart_sched_tick is called. This happens when 683 * interrupts arrive which do not cause a reschedule. In the 684 * first call we save the current tick time, so we can restart 685 * the scheduler tick in nohz_restart_sched_tick. 686 */ 687 if (!ts->tick_stopped) { 688 nohz_balance_enter_idle(cpu); 689 calc_load_enter_idle(); 690 691 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 692 ts->tick_stopped = 1; 693 trace_tick_stop(1, " "); 694 } 695 696 /* 697 * If the expiration time == KTIME_MAX, then 698 * in this case we simply stop the tick timer. 699 */ 700 if (unlikely(expires.tv64 == KTIME_MAX)) { 701 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 702 hrtimer_cancel(&ts->sched_timer); 703 goto out; 704 } 705 706 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 707 hrtimer_start(&ts->sched_timer, expires, 708 HRTIMER_MODE_ABS_PINNED); 709 else 710 tick_program_event(expires, 1); 711 out: 712 ts->next_jiffies = next_jiffies; 713 ts->last_jiffies = last_jiffies; 714 ts->sleep_length = ktime_sub(dev->next_event, now); 715 716 return ret; 717 } 718 719 static void tick_nohz_full_stop_tick(struct tick_sched *ts) 720 { 721 #ifdef CONFIG_NO_HZ_FULL 722 int cpu = smp_processor_id(); 723 724 if (!tick_nohz_full_cpu(cpu) || is_idle_task(current)) 725 return; 726 727 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 728 return; 729 730 if (!can_stop_full_tick()) 731 return; 732 733 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 734 #endif 735 } 736 737 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 738 { 739 /* 740 * If this cpu is offline and it is the one which updates 741 * jiffies, then give up the assignment and let it be taken by 742 * the cpu which runs the tick timer next. If we don't drop 743 * this here the jiffies might be stale and do_timer() never 744 * invoked. 745 */ 746 if (unlikely(!cpu_online(cpu))) { 747 if (cpu == tick_do_timer_cpu) 748 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 749 return false; 750 } 751 752 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 753 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ }; 754 return false; 755 } 756 757 if (need_resched()) 758 return false; 759 760 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 761 static int ratelimit; 762 763 if (ratelimit < 10 && 764 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 765 pr_warn("NOHZ: local_softirq_pending %02x\n", 766 (unsigned int) local_softirq_pending()); 767 ratelimit++; 768 } 769 return false; 770 } 771 772 if (tick_nohz_full_enabled()) { 773 /* 774 * Keep the tick alive to guarantee timekeeping progression 775 * if there are full dynticks CPUs around 776 */ 777 if (tick_do_timer_cpu == cpu) 778 return false; 779 /* 780 * Boot safety: make sure the timekeeping duty has been 781 * assigned before entering dyntick-idle mode, 782 */ 783 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 784 return false; 785 } 786 787 return true; 788 } 789 790 static void __tick_nohz_idle_enter(struct tick_sched *ts) 791 { 792 ktime_t now, expires; 793 int cpu = smp_processor_id(); 794 795 now = tick_nohz_start_idle(ts); 796 797 if (can_stop_idle_tick(cpu, ts)) { 798 int was_stopped = ts->tick_stopped; 799 800 ts->idle_calls++; 801 802 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 803 if (expires.tv64 > 0LL) { 804 ts->idle_sleeps++; 805 ts->idle_expires = expires; 806 } 807 808 if (!was_stopped && ts->tick_stopped) 809 ts->idle_jiffies = ts->last_jiffies; 810 } 811 } 812 813 /** 814 * tick_nohz_idle_enter - stop the idle tick from the idle task 815 * 816 * When the next event is more than a tick into the future, stop the idle tick 817 * Called when we start the idle loop. 818 * 819 * The arch is responsible of calling: 820 * 821 * - rcu_idle_enter() after its last use of RCU before the CPU is put 822 * to sleep. 823 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 824 */ 825 void tick_nohz_idle_enter(void) 826 { 827 struct tick_sched *ts; 828 829 WARN_ON_ONCE(irqs_disabled()); 830 831 /* 832 * Update the idle state in the scheduler domain hierarchy 833 * when tick_nohz_stop_sched_tick() is called from the idle loop. 834 * State will be updated to busy during the first busy tick after 835 * exiting idle. 836 */ 837 set_cpu_sd_state_idle(); 838 839 local_irq_disable(); 840 841 ts = this_cpu_ptr(&tick_cpu_sched); 842 ts->inidle = 1; 843 __tick_nohz_idle_enter(ts); 844 845 local_irq_enable(); 846 } 847 848 /** 849 * tick_nohz_irq_exit - update next tick event from interrupt exit 850 * 851 * When an interrupt fires while we are idle and it doesn't cause 852 * a reschedule, it may still add, modify or delete a timer, enqueue 853 * an RCU callback, etc... 854 * So we need to re-calculate and reprogram the next tick event. 855 */ 856 void tick_nohz_irq_exit(void) 857 { 858 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 859 860 if (ts->inidle) 861 __tick_nohz_idle_enter(ts); 862 else 863 tick_nohz_full_stop_tick(ts); 864 } 865 866 /** 867 * tick_nohz_get_sleep_length - return the length of the current sleep 868 * 869 * Called from power state control code with interrupts disabled 870 */ 871 ktime_t tick_nohz_get_sleep_length(void) 872 { 873 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 874 875 return ts->sleep_length; 876 } 877 878 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 879 { 880 /* Update jiffies first */ 881 tick_do_update_jiffies64(now); 882 update_cpu_load_nohz(); 883 884 calc_load_exit_idle(); 885 touch_softlockup_watchdog(); 886 /* 887 * Cancel the scheduled timer and restore the tick 888 */ 889 ts->tick_stopped = 0; 890 ts->idle_exittime = now; 891 892 tick_nohz_restart(ts, now); 893 } 894 895 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 896 { 897 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 898 unsigned long ticks; 899 900 if (vtime_accounting_enabled()) 901 return; 902 /* 903 * We stopped the tick in idle. Update process times would miss the 904 * time we slept as update_process_times does only a 1 tick 905 * accounting. Enforce that this is accounted to idle ! 906 */ 907 ticks = jiffies - ts->idle_jiffies; 908 /* 909 * We might be one off. Do not randomly account a huge number of ticks! 910 */ 911 if (ticks && ticks < LONG_MAX) 912 account_idle_ticks(ticks); 913 #endif 914 } 915 916 /** 917 * tick_nohz_idle_exit - restart the idle tick from the idle task 918 * 919 * Restart the idle tick when the CPU is woken up from idle 920 * This also exit the RCU extended quiescent state. The CPU 921 * can use RCU again after this function is called. 922 */ 923 void tick_nohz_idle_exit(void) 924 { 925 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 926 ktime_t now; 927 928 local_irq_disable(); 929 930 WARN_ON_ONCE(!ts->inidle); 931 932 ts->inidle = 0; 933 934 if (ts->idle_active || ts->tick_stopped) 935 now = ktime_get(); 936 937 if (ts->idle_active) 938 tick_nohz_stop_idle(ts, now); 939 940 if (ts->tick_stopped) { 941 tick_nohz_restart_sched_tick(ts, now); 942 tick_nohz_account_idle_ticks(ts); 943 } 944 945 local_irq_enable(); 946 } 947 948 /* 949 * The nohz low res interrupt handler 950 */ 951 static void tick_nohz_handler(struct clock_event_device *dev) 952 { 953 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 954 struct pt_regs *regs = get_irq_regs(); 955 ktime_t now = ktime_get(); 956 957 dev->next_event.tv64 = KTIME_MAX; 958 959 tick_sched_do_timer(now); 960 tick_sched_handle(ts, regs); 961 962 /* No need to reprogram if we are running tickless */ 963 if (unlikely(ts->tick_stopped)) 964 return; 965 966 hrtimer_forward(&ts->sched_timer, now, tick_period); 967 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 968 } 969 970 /** 971 * tick_nohz_switch_to_nohz - switch to nohz mode 972 */ 973 static void tick_nohz_switch_to_nohz(void) 974 { 975 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 976 ktime_t next; 977 978 if (!tick_nohz_enabled) 979 return; 980 981 local_irq_disable(); 982 if (tick_switch_to_oneshot(tick_nohz_handler)) { 983 local_irq_enable(); 984 return; 985 } 986 tick_nohz_active = 1; 987 ts->nohz_mode = NOHZ_MODE_LOWRES; 988 989 /* 990 * Recycle the hrtimer in ts, so we can share the 991 * hrtimer_forward with the highres code. 992 */ 993 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 994 /* Get the next period */ 995 next = tick_init_jiffy_update(); 996 997 hrtimer_forward_now(&ts->sched_timer, tick_period); 998 hrtimer_set_expires(&ts->sched_timer, next); 999 tick_program_event(next, 1); 1000 local_irq_enable(); 1001 } 1002 1003 /* 1004 * When NOHZ is enabled and the tick is stopped, we need to kick the 1005 * tick timer from irq_enter() so that the jiffies update is kept 1006 * alive during long running softirqs. That's ugly as hell, but 1007 * correctness is key even if we need to fix the offending softirq in 1008 * the first place. 1009 * 1010 * Note, this is different to tick_nohz_restart. We just kick the 1011 * timer and do not touch the other magic bits which need to be done 1012 * when idle is left. 1013 */ 1014 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now) 1015 { 1016 #if 0 1017 /* Switch back to 2.6.27 behaviour */ 1018 ktime_t delta; 1019 1020 /* 1021 * Do not touch the tick device, when the next expiry is either 1022 * already reached or less/equal than the tick period. 1023 */ 1024 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 1025 if (delta.tv64 <= tick_period.tv64) 1026 return; 1027 1028 tick_nohz_restart(ts, now); 1029 #endif 1030 } 1031 1032 static inline void tick_nohz_irq_enter(void) 1033 { 1034 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1035 ktime_t now; 1036 1037 if (!ts->idle_active && !ts->tick_stopped) 1038 return; 1039 now = ktime_get(); 1040 if (ts->idle_active) 1041 tick_nohz_stop_idle(ts, now); 1042 if (ts->tick_stopped) { 1043 tick_nohz_update_jiffies(now); 1044 tick_nohz_kick_tick(ts, now); 1045 } 1046 } 1047 1048 #else 1049 1050 static inline void tick_nohz_switch_to_nohz(void) { } 1051 static inline void tick_nohz_irq_enter(void) { } 1052 1053 #endif /* CONFIG_NO_HZ_COMMON */ 1054 1055 /* 1056 * Called from irq_enter to notify about the possible interruption of idle() 1057 */ 1058 void tick_irq_enter(void) 1059 { 1060 tick_check_oneshot_broadcast_this_cpu(); 1061 tick_nohz_irq_enter(); 1062 } 1063 1064 /* 1065 * High resolution timer specific code 1066 */ 1067 #ifdef CONFIG_HIGH_RES_TIMERS 1068 /* 1069 * We rearm the timer until we get disabled by the idle code. 1070 * Called with interrupts disabled. 1071 */ 1072 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1073 { 1074 struct tick_sched *ts = 1075 container_of(timer, struct tick_sched, sched_timer); 1076 struct pt_regs *regs = get_irq_regs(); 1077 ktime_t now = ktime_get(); 1078 1079 tick_sched_do_timer(now); 1080 1081 /* 1082 * Do not call, when we are not in irq context and have 1083 * no valid regs pointer 1084 */ 1085 if (regs) 1086 tick_sched_handle(ts, regs); 1087 1088 /* No need to reprogram if we are in idle or full dynticks mode */ 1089 if (unlikely(ts->tick_stopped)) 1090 return HRTIMER_NORESTART; 1091 1092 hrtimer_forward(timer, now, tick_period); 1093 1094 return HRTIMER_RESTART; 1095 } 1096 1097 static int sched_skew_tick; 1098 1099 static int __init skew_tick(char *str) 1100 { 1101 get_option(&str, &sched_skew_tick); 1102 1103 return 0; 1104 } 1105 early_param("skew_tick", skew_tick); 1106 1107 /** 1108 * tick_setup_sched_timer - setup the tick emulation timer 1109 */ 1110 void tick_setup_sched_timer(void) 1111 { 1112 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1113 ktime_t now = ktime_get(); 1114 1115 /* 1116 * Emulate tick processing via per-CPU hrtimers: 1117 */ 1118 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1119 ts->sched_timer.function = tick_sched_timer; 1120 1121 /* Get the next period (per cpu) */ 1122 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1123 1124 /* Offset the tick to avert jiffies_lock contention. */ 1125 if (sched_skew_tick) { 1126 u64 offset = ktime_to_ns(tick_period) >> 1; 1127 do_div(offset, num_possible_cpus()); 1128 offset *= smp_processor_id(); 1129 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1130 } 1131 1132 hrtimer_forward(&ts->sched_timer, now, tick_period); 1133 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 1134 1135 #ifdef CONFIG_NO_HZ_COMMON 1136 if (tick_nohz_enabled) { 1137 ts->nohz_mode = NOHZ_MODE_HIGHRES; 1138 tick_nohz_active = 1; 1139 } 1140 #endif 1141 } 1142 #endif /* HIGH_RES_TIMERS */ 1143 1144 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1145 void tick_cancel_sched_timer(int cpu) 1146 { 1147 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1148 1149 # ifdef CONFIG_HIGH_RES_TIMERS 1150 if (ts->sched_timer.base) 1151 hrtimer_cancel(&ts->sched_timer); 1152 # endif 1153 1154 memset(ts, 0, sizeof(*ts)); 1155 } 1156 #endif 1157 1158 /** 1159 * Async notification about clocksource changes 1160 */ 1161 void tick_clock_notify(void) 1162 { 1163 int cpu; 1164 1165 for_each_possible_cpu(cpu) 1166 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1167 } 1168 1169 /* 1170 * Async notification about clock event changes 1171 */ 1172 void tick_oneshot_notify(void) 1173 { 1174 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1175 1176 set_bit(0, &ts->check_clocks); 1177 } 1178 1179 /** 1180 * Check, if a change happened, which makes oneshot possible. 1181 * 1182 * Called cyclic from the hrtimer softirq (driven by the timer 1183 * softirq) allow_nohz signals, that we can switch into low-res nohz 1184 * mode, because high resolution timers are disabled (either compile 1185 * or runtime). 1186 */ 1187 int tick_check_oneshot_change(int allow_nohz) 1188 { 1189 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1190 1191 if (!test_and_clear_bit(0, &ts->check_clocks)) 1192 return 0; 1193 1194 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1195 return 0; 1196 1197 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1198 return 0; 1199 1200 if (!allow_nohz) 1201 return 1; 1202 1203 tick_nohz_switch_to_nohz(); 1204 return 0; 1205 } 1206