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/context_tracking.h> 26 27 #include <asm/irq_regs.h> 28 29 #include "tick-internal.h" 30 31 #include <trace/events/timer.h> 32 33 /* 34 * Per cpu nohz control structure 35 */ 36 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); 37 38 struct tick_sched *tick_get_tick_sched(int cpu) 39 { 40 return &per_cpu(tick_cpu_sched, cpu); 41 } 42 43 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS) 44 /* 45 * The time, when the last jiffy update happened. Protected by jiffies_lock. 46 */ 47 static ktime_t last_jiffies_update; 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_sched(); 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 #endif 155 156 #ifdef CONFIG_NO_HZ_FULL 157 cpumask_var_t tick_nohz_full_mask; 158 cpumask_var_t housekeeping_mask; 159 bool tick_nohz_full_running; 160 static unsigned long tick_dep_mask; 161 162 static void trace_tick_dependency(unsigned long dep) 163 { 164 if (dep & TICK_DEP_MASK_POSIX_TIMER) { 165 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER); 166 return; 167 } 168 169 if (dep & TICK_DEP_MASK_PERF_EVENTS) { 170 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS); 171 return; 172 } 173 174 if (dep & TICK_DEP_MASK_SCHED) { 175 trace_tick_stop(0, TICK_DEP_MASK_SCHED); 176 return; 177 } 178 179 if (dep & TICK_DEP_MASK_CLOCK_UNSTABLE) 180 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE); 181 } 182 183 static bool can_stop_full_tick(struct tick_sched *ts) 184 { 185 WARN_ON_ONCE(!irqs_disabled()); 186 187 if (tick_dep_mask) { 188 trace_tick_dependency(tick_dep_mask); 189 return false; 190 } 191 192 if (ts->tick_dep_mask) { 193 trace_tick_dependency(ts->tick_dep_mask); 194 return false; 195 } 196 197 if (current->tick_dep_mask) { 198 trace_tick_dependency(current->tick_dep_mask); 199 return false; 200 } 201 202 if (current->signal->tick_dep_mask) { 203 trace_tick_dependency(current->signal->tick_dep_mask); 204 return false; 205 } 206 207 if (!sched_can_stop_tick()) { 208 trace_tick_stop(0, TICK_DEP_MASK_SCHED); 209 return false; 210 } 211 212 if (!posix_cpu_timers_can_stop_tick(current)) { 213 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER); 214 return false; 215 } 216 217 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK 218 /* 219 * sched_clock_tick() needs us? 220 * 221 * TODO: kick full dynticks CPUs when 222 * sched_clock_stable is set. 223 */ 224 if (!sched_clock_stable()) { 225 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE); 226 /* 227 * Don't allow the user to think they can get 228 * full NO_HZ with this machine. 229 */ 230 WARN_ONCE(tick_nohz_full_running, 231 "NO_HZ FULL will not work with unstable sched clock"); 232 return false; 233 } 234 #endif 235 236 return true; 237 } 238 239 static void nohz_full_kick_func(struct irq_work *work) 240 { 241 /* Empty, the tick restart happens on tick_nohz_irq_exit() */ 242 } 243 244 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 245 .func = nohz_full_kick_func, 246 }; 247 248 /* 249 * Kick this CPU if it's full dynticks in order to force it to 250 * re-evaluate its dependency on the tick and restart it if necessary. 251 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), 252 * is NMI safe. 253 */ 254 static void tick_nohz_full_kick(void) 255 { 256 if (!tick_nohz_full_cpu(smp_processor_id())) 257 return; 258 259 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work)); 260 } 261 262 /* 263 * Kick the CPU if it's full dynticks in order to force it to 264 * re-evaluate its dependency on the tick and restart it if necessary. 265 */ 266 void tick_nohz_full_kick_cpu(int cpu) 267 { 268 if (!tick_nohz_full_cpu(cpu)) 269 return; 270 271 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); 272 } 273 274 /* 275 * Kick all full dynticks CPUs in order to force these to re-evaluate 276 * their dependency on the tick and restart it if necessary. 277 */ 278 void tick_nohz_full_kick_all(void) 279 { 280 int cpu; 281 282 if (!tick_nohz_full_running) 283 return; 284 285 preempt_disable(); 286 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask) 287 tick_nohz_full_kick_cpu(cpu); 288 preempt_enable(); 289 } 290 291 static void tick_nohz_dep_set_all(unsigned long *dep, 292 enum tick_dep_bits bit) 293 { 294 unsigned long prev; 295 296 prev = fetch_or(dep, BIT_MASK(bit)); 297 if (!prev) 298 tick_nohz_full_kick_all(); 299 } 300 301 /* 302 * Set a global tick dependency. Used by perf events that rely on freq and 303 * by unstable clock. 304 */ 305 void tick_nohz_dep_set(enum tick_dep_bits bit) 306 { 307 tick_nohz_dep_set_all(&tick_dep_mask, bit); 308 } 309 310 void tick_nohz_dep_clear(enum tick_dep_bits bit) 311 { 312 clear_bit(bit, &tick_dep_mask); 313 } 314 315 /* 316 * Set per-CPU tick dependency. Used by scheduler and perf events in order to 317 * manage events throttling. 318 */ 319 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit) 320 { 321 unsigned long prev; 322 struct tick_sched *ts; 323 324 ts = per_cpu_ptr(&tick_cpu_sched, cpu); 325 326 prev = fetch_or(&ts->tick_dep_mask, BIT_MASK(bit)); 327 if (!prev) { 328 preempt_disable(); 329 /* Perf needs local kick that is NMI safe */ 330 if (cpu == smp_processor_id()) { 331 tick_nohz_full_kick(); 332 } else { 333 /* Remote irq work not NMI-safe */ 334 if (!WARN_ON_ONCE(in_nmi())) 335 tick_nohz_full_kick_cpu(cpu); 336 } 337 preempt_enable(); 338 } 339 } 340 341 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit) 342 { 343 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); 344 345 clear_bit(bit, &ts->tick_dep_mask); 346 } 347 348 /* 349 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse 350 * per task timers. 351 */ 352 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit) 353 { 354 /* 355 * We could optimize this with just kicking the target running the task 356 * if that noise matters for nohz full users. 357 */ 358 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit); 359 } 360 361 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit) 362 { 363 clear_bit(bit, &tsk->tick_dep_mask); 364 } 365 366 /* 367 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse 368 * per process timers. 369 */ 370 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit) 371 { 372 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit); 373 } 374 375 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit) 376 { 377 clear_bit(bit, &sig->tick_dep_mask); 378 } 379 380 /* 381 * Re-evaluate the need for the tick as we switch the current task. 382 * It might need the tick due to per task/process properties: 383 * perf events, posix cpu timers, ... 384 */ 385 void __tick_nohz_task_switch(void) 386 { 387 unsigned long flags; 388 struct tick_sched *ts; 389 390 local_irq_save(flags); 391 392 if (!tick_nohz_full_cpu(smp_processor_id())) 393 goto out; 394 395 ts = this_cpu_ptr(&tick_cpu_sched); 396 397 if (ts->tick_stopped) { 398 if (current->tick_dep_mask || current->signal->tick_dep_mask) 399 tick_nohz_full_kick(); 400 } 401 out: 402 local_irq_restore(flags); 403 } 404 405 /* Parse the boot-time nohz CPU list from the kernel parameters. */ 406 static int __init tick_nohz_full_setup(char *str) 407 { 408 alloc_bootmem_cpumask_var(&tick_nohz_full_mask); 409 if (cpulist_parse(str, tick_nohz_full_mask) < 0) { 410 pr_warning("NOHZ: Incorrect nohz_full cpumask\n"); 411 free_bootmem_cpumask_var(tick_nohz_full_mask); 412 return 1; 413 } 414 tick_nohz_full_running = true; 415 416 return 1; 417 } 418 __setup("nohz_full=", tick_nohz_full_setup); 419 420 static int tick_nohz_cpu_down_callback(struct notifier_block *nfb, 421 unsigned long action, 422 void *hcpu) 423 { 424 unsigned int cpu = (unsigned long)hcpu; 425 426 switch (action & ~CPU_TASKS_FROZEN) { 427 case CPU_DOWN_PREPARE: 428 /* 429 * The boot CPU handles housekeeping duty (unbound timers, 430 * workqueues, timekeeping, ...) on behalf of full dynticks 431 * CPUs. It must remain online when nohz full is enabled. 432 */ 433 if (tick_nohz_full_running && tick_do_timer_cpu == cpu) 434 return NOTIFY_BAD; 435 break; 436 } 437 return NOTIFY_OK; 438 } 439 440 static int tick_nohz_init_all(void) 441 { 442 int err = -1; 443 444 #ifdef CONFIG_NO_HZ_FULL_ALL 445 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { 446 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n"); 447 return err; 448 } 449 err = 0; 450 cpumask_setall(tick_nohz_full_mask); 451 tick_nohz_full_running = true; 452 #endif 453 return err; 454 } 455 456 void __init tick_nohz_init(void) 457 { 458 int cpu; 459 460 if (!tick_nohz_full_running) { 461 if (tick_nohz_init_all() < 0) 462 return; 463 } 464 465 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) { 466 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n"); 467 cpumask_clear(tick_nohz_full_mask); 468 tick_nohz_full_running = false; 469 return; 470 } 471 472 /* 473 * Full dynticks uses irq work to drive the tick rescheduling on safe 474 * locking contexts. But then we need irq work to raise its own 475 * interrupts to avoid circular dependency on the tick 476 */ 477 if (!arch_irq_work_has_interrupt()) { 478 pr_warning("NO_HZ: Can't run full dynticks because arch doesn't " 479 "support irq work self-IPIs\n"); 480 cpumask_clear(tick_nohz_full_mask); 481 cpumask_copy(housekeeping_mask, cpu_possible_mask); 482 tick_nohz_full_running = false; 483 return; 484 } 485 486 cpu = smp_processor_id(); 487 488 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 489 pr_warning("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", cpu); 490 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 491 } 492 493 cpumask_andnot(housekeeping_mask, 494 cpu_possible_mask, tick_nohz_full_mask); 495 496 for_each_cpu(cpu, tick_nohz_full_mask) 497 context_tracking_cpu_set(cpu); 498 499 cpu_notifier(tick_nohz_cpu_down_callback, 0); 500 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", 501 cpumask_pr_args(tick_nohz_full_mask)); 502 503 /* 504 * We need at least one CPU to handle housekeeping work such 505 * as timekeeping, unbound timers, workqueues, ... 506 */ 507 WARN_ON_ONCE(cpumask_empty(housekeeping_mask)); 508 } 509 #endif 510 511 /* 512 * NOHZ - aka dynamic tick functionality 513 */ 514 #ifdef CONFIG_NO_HZ_COMMON 515 /* 516 * NO HZ enabled ? 517 */ 518 int tick_nohz_enabled __read_mostly = 1; 519 unsigned long tick_nohz_active __read_mostly; 520 /* 521 * Enable / Disable tickless mode 522 */ 523 static int __init setup_tick_nohz(char *str) 524 { 525 if (!strcmp(str, "off")) 526 tick_nohz_enabled = 0; 527 else if (!strcmp(str, "on")) 528 tick_nohz_enabled = 1; 529 else 530 return 0; 531 return 1; 532 } 533 534 __setup("nohz=", setup_tick_nohz); 535 536 int tick_nohz_tick_stopped(void) 537 { 538 return __this_cpu_read(tick_cpu_sched.tick_stopped); 539 } 540 541 /** 542 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 543 * 544 * Called from interrupt entry when the CPU was idle 545 * 546 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 547 * must be updated. Otherwise an interrupt handler could use a stale jiffy 548 * value. We do this unconditionally on any cpu, as we don't know whether the 549 * cpu, which has the update task assigned is in a long sleep. 550 */ 551 static void tick_nohz_update_jiffies(ktime_t now) 552 { 553 unsigned long flags; 554 555 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 556 557 local_irq_save(flags); 558 tick_do_update_jiffies64(now); 559 local_irq_restore(flags); 560 561 touch_softlockup_watchdog_sched(); 562 } 563 564 /* 565 * Updates the per cpu time idle statistics counters 566 */ 567 static void 568 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 569 { 570 ktime_t delta; 571 572 if (ts->idle_active) { 573 delta = ktime_sub(now, ts->idle_entrytime); 574 if (nr_iowait_cpu(cpu) > 0) 575 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 576 else 577 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 578 ts->idle_entrytime = now; 579 } 580 581 if (last_update_time) 582 *last_update_time = ktime_to_us(now); 583 584 } 585 586 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 587 { 588 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 589 ts->idle_active = 0; 590 591 sched_clock_idle_wakeup_event(0); 592 } 593 594 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 595 { 596 ktime_t now = ktime_get(); 597 598 ts->idle_entrytime = now; 599 ts->idle_active = 1; 600 sched_clock_idle_sleep_event(); 601 return now; 602 } 603 604 /** 605 * get_cpu_idle_time_us - get the total idle time of a cpu 606 * @cpu: CPU number to query 607 * @last_update_time: variable to store update time in. Do not update 608 * counters if NULL. 609 * 610 * Return the cummulative idle time (since boot) for a given 611 * CPU, in microseconds. 612 * 613 * This time is measured via accounting rather than sampling, 614 * and is as accurate as ktime_get() is. 615 * 616 * This function returns -1 if NOHZ is not enabled. 617 */ 618 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 619 { 620 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 621 ktime_t now, idle; 622 623 if (!tick_nohz_active) 624 return -1; 625 626 now = ktime_get(); 627 if (last_update_time) { 628 update_ts_time_stats(cpu, ts, now, last_update_time); 629 idle = ts->idle_sleeptime; 630 } else { 631 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 632 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 633 634 idle = ktime_add(ts->idle_sleeptime, delta); 635 } else { 636 idle = ts->idle_sleeptime; 637 } 638 } 639 640 return ktime_to_us(idle); 641 642 } 643 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 644 645 /** 646 * get_cpu_iowait_time_us - get the total iowait time of a cpu 647 * @cpu: CPU number to query 648 * @last_update_time: variable to store update time in. Do not update 649 * counters if NULL. 650 * 651 * Return the cummulative iowait time (since boot) for a given 652 * CPU, in microseconds. 653 * 654 * This time is measured via accounting rather than sampling, 655 * and is as accurate as ktime_get() is. 656 * 657 * This function returns -1 if NOHZ is not enabled. 658 */ 659 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 660 { 661 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 662 ktime_t now, iowait; 663 664 if (!tick_nohz_active) 665 return -1; 666 667 now = ktime_get(); 668 if (last_update_time) { 669 update_ts_time_stats(cpu, ts, now, last_update_time); 670 iowait = ts->iowait_sleeptime; 671 } else { 672 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 673 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 674 675 iowait = ktime_add(ts->iowait_sleeptime, delta); 676 } else { 677 iowait = ts->iowait_sleeptime; 678 } 679 } 680 681 return ktime_to_us(iowait); 682 } 683 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 684 685 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 686 { 687 hrtimer_cancel(&ts->sched_timer); 688 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 689 690 /* Forward the time to expire in the future */ 691 hrtimer_forward(&ts->sched_timer, now, tick_period); 692 693 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 694 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 695 else 696 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 697 } 698 699 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 700 ktime_t now, int cpu) 701 { 702 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 703 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; 704 unsigned long seq, basejiff; 705 ktime_t tick; 706 707 /* Read jiffies and the time when jiffies were updated last */ 708 do { 709 seq = read_seqbegin(&jiffies_lock); 710 basemono = last_jiffies_update.tv64; 711 basejiff = jiffies; 712 } while (read_seqretry(&jiffies_lock, seq)); 713 ts->last_jiffies = basejiff; 714 715 if (rcu_needs_cpu(basemono, &next_rcu) || 716 arch_needs_cpu() || irq_work_needs_cpu()) { 717 next_tick = basemono + TICK_NSEC; 718 } else { 719 /* 720 * Get the next pending timer. If high resolution 721 * timers are enabled this only takes the timer wheel 722 * timers into account. If high resolution timers are 723 * disabled this also looks at the next expiring 724 * hrtimer. 725 */ 726 next_tmr = get_next_timer_interrupt(basejiff, basemono); 727 ts->next_timer = next_tmr; 728 /* Take the next rcu event into account */ 729 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; 730 } 731 732 /* 733 * If the tick is due in the next period, keep it ticking or 734 * force prod the timer. 735 */ 736 delta = next_tick - basemono; 737 if (delta <= (u64)TICK_NSEC) { 738 tick.tv64 = 0; 739 /* 740 * We've not stopped the tick yet, and there's a timer in the 741 * next period, so no point in stopping it either, bail. 742 */ 743 if (!ts->tick_stopped) 744 goto out; 745 746 /* 747 * If, OTOH, we did stop it, but there's a pending (expired) 748 * timer reprogram the timer hardware to fire now. 749 * 750 * We will not restart the tick proper, just prod the timer 751 * hardware into firing an interrupt to process the pending 752 * timers. Just like tick_irq_exit() will not restart the tick 753 * for 'normal' interrupts. 754 * 755 * Only once we exit the idle loop will we re-enable the tick, 756 * see tick_nohz_idle_exit(). 757 */ 758 if (delta == 0) { 759 tick_nohz_restart(ts, now); 760 goto out; 761 } 762 } 763 764 /* 765 * If this cpu is the one which updates jiffies, then give up 766 * the assignment and let it be taken by the cpu which runs 767 * the tick timer next, which might be this cpu as well. If we 768 * don't drop this here the jiffies might be stale and 769 * do_timer() never invoked. Keep track of the fact that it 770 * was the one which had the do_timer() duty last. If this cpu 771 * is the one which had the do_timer() duty last, we limit the 772 * sleep time to the timekeeping max_deferement value. 773 * Otherwise we can sleep as long as we want. 774 */ 775 delta = timekeeping_max_deferment(); 776 if (cpu == tick_do_timer_cpu) { 777 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 778 ts->do_timer_last = 1; 779 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 780 delta = KTIME_MAX; 781 ts->do_timer_last = 0; 782 } else if (!ts->do_timer_last) { 783 delta = KTIME_MAX; 784 } 785 786 #ifdef CONFIG_NO_HZ_FULL 787 /* Limit the tick delta to the maximum scheduler deferment */ 788 if (!ts->inidle) 789 delta = min(delta, scheduler_tick_max_deferment()); 790 #endif 791 792 /* Calculate the next expiry time */ 793 if (delta < (KTIME_MAX - basemono)) 794 expires = basemono + delta; 795 else 796 expires = KTIME_MAX; 797 798 expires = min_t(u64, expires, next_tick); 799 tick.tv64 = expires; 800 801 /* Skip reprogram of event if its not changed */ 802 if (ts->tick_stopped && (expires == dev->next_event.tv64)) 803 goto out; 804 805 /* 806 * nohz_stop_sched_tick can be called several times before 807 * the nohz_restart_sched_tick is called. This happens when 808 * interrupts arrive which do not cause a reschedule. In the 809 * first call we save the current tick time, so we can restart 810 * the scheduler tick in nohz_restart_sched_tick. 811 */ 812 if (!ts->tick_stopped) { 813 nohz_balance_enter_idle(cpu); 814 calc_load_enter_idle(); 815 816 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 817 ts->tick_stopped = 1; 818 trace_tick_stop(1, TICK_DEP_MASK_NONE); 819 } 820 821 /* 822 * If the expiration time == KTIME_MAX, then we simply stop 823 * the tick timer. 824 */ 825 if (unlikely(expires == KTIME_MAX)) { 826 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 827 hrtimer_cancel(&ts->sched_timer); 828 goto out; 829 } 830 831 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 832 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED); 833 else 834 tick_program_event(tick, 1); 835 out: 836 /* Update the estimated sleep length */ 837 ts->sleep_length = ktime_sub(dev->next_event, now); 838 return tick; 839 } 840 841 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now, int active) 842 { 843 /* Update jiffies first */ 844 tick_do_update_jiffies64(now); 845 update_cpu_load_nohz(active); 846 847 calc_load_exit_idle(); 848 touch_softlockup_watchdog_sched(); 849 /* 850 * Cancel the scheduled timer and restore the tick 851 */ 852 ts->tick_stopped = 0; 853 ts->idle_exittime = now; 854 855 tick_nohz_restart(ts, now); 856 } 857 858 static void tick_nohz_full_update_tick(struct tick_sched *ts) 859 { 860 #ifdef CONFIG_NO_HZ_FULL 861 int cpu = smp_processor_id(); 862 863 if (!tick_nohz_full_cpu(cpu)) 864 return; 865 866 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 867 return; 868 869 if (can_stop_full_tick(ts)) 870 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 871 else if (ts->tick_stopped) 872 tick_nohz_restart_sched_tick(ts, ktime_get(), 1); 873 #endif 874 } 875 876 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 877 { 878 /* 879 * If this cpu is offline and it is the one which updates 880 * jiffies, then give up the assignment and let it be taken by 881 * the cpu which runs the tick timer next. If we don't drop 882 * this here the jiffies might be stale and do_timer() never 883 * invoked. 884 */ 885 if (unlikely(!cpu_online(cpu))) { 886 if (cpu == tick_do_timer_cpu) 887 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 888 return false; 889 } 890 891 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 892 ts->sleep_length = (ktime_t) { .tv64 = NSEC_PER_SEC/HZ }; 893 return false; 894 } 895 896 if (need_resched()) 897 return false; 898 899 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 900 static int ratelimit; 901 902 if (ratelimit < 10 && 903 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 904 pr_warn("NOHZ: local_softirq_pending %02x\n", 905 (unsigned int) local_softirq_pending()); 906 ratelimit++; 907 } 908 return false; 909 } 910 911 if (tick_nohz_full_enabled()) { 912 /* 913 * Keep the tick alive to guarantee timekeeping progression 914 * if there are full dynticks CPUs around 915 */ 916 if (tick_do_timer_cpu == cpu) 917 return false; 918 /* 919 * Boot safety: make sure the timekeeping duty has been 920 * assigned before entering dyntick-idle mode, 921 */ 922 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 923 return false; 924 } 925 926 return true; 927 } 928 929 static void __tick_nohz_idle_enter(struct tick_sched *ts) 930 { 931 ktime_t now, expires; 932 int cpu = smp_processor_id(); 933 934 now = tick_nohz_start_idle(ts); 935 936 if (can_stop_idle_tick(cpu, ts)) { 937 int was_stopped = ts->tick_stopped; 938 939 ts->idle_calls++; 940 941 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 942 if (expires.tv64 > 0LL) { 943 ts->idle_sleeps++; 944 ts->idle_expires = expires; 945 } 946 947 if (!was_stopped && ts->tick_stopped) 948 ts->idle_jiffies = ts->last_jiffies; 949 } 950 } 951 952 /** 953 * tick_nohz_idle_enter - stop the idle tick from the idle task 954 * 955 * When the next event is more than a tick into the future, stop the idle tick 956 * Called when we start the idle loop. 957 * 958 * The arch is responsible of calling: 959 * 960 * - rcu_idle_enter() after its last use of RCU before the CPU is put 961 * to sleep. 962 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 963 */ 964 void tick_nohz_idle_enter(void) 965 { 966 struct tick_sched *ts; 967 968 WARN_ON_ONCE(irqs_disabled()); 969 970 /* 971 * Update the idle state in the scheduler domain hierarchy 972 * when tick_nohz_stop_sched_tick() is called from the idle loop. 973 * State will be updated to busy during the first busy tick after 974 * exiting idle. 975 */ 976 set_cpu_sd_state_idle(); 977 978 local_irq_disable(); 979 980 ts = this_cpu_ptr(&tick_cpu_sched); 981 ts->inidle = 1; 982 __tick_nohz_idle_enter(ts); 983 984 local_irq_enable(); 985 } 986 987 /** 988 * tick_nohz_irq_exit - update next tick event from interrupt exit 989 * 990 * When an interrupt fires while we are idle and it doesn't cause 991 * a reschedule, it may still add, modify or delete a timer, enqueue 992 * an RCU callback, etc... 993 * So we need to re-calculate and reprogram the next tick event. 994 */ 995 void tick_nohz_irq_exit(void) 996 { 997 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 998 999 if (ts->inidle) 1000 __tick_nohz_idle_enter(ts); 1001 else 1002 tick_nohz_full_update_tick(ts); 1003 } 1004 1005 /** 1006 * tick_nohz_get_sleep_length - return the length of the current sleep 1007 * 1008 * Called from power state control code with interrupts disabled 1009 */ 1010 ktime_t tick_nohz_get_sleep_length(void) 1011 { 1012 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1013 1014 return ts->sleep_length; 1015 } 1016 1017 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 1018 { 1019 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 1020 unsigned long ticks; 1021 1022 if (vtime_accounting_cpu_enabled()) 1023 return; 1024 /* 1025 * We stopped the tick in idle. Update process times would miss the 1026 * time we slept as update_process_times does only a 1 tick 1027 * accounting. Enforce that this is accounted to idle ! 1028 */ 1029 ticks = jiffies - ts->idle_jiffies; 1030 /* 1031 * We might be one off. Do not randomly account a huge number of ticks! 1032 */ 1033 if (ticks && ticks < LONG_MAX) 1034 account_idle_ticks(ticks); 1035 #endif 1036 } 1037 1038 /** 1039 * tick_nohz_idle_exit - restart the idle tick from the idle task 1040 * 1041 * Restart the idle tick when the CPU is woken up from idle 1042 * This also exit the RCU extended quiescent state. The CPU 1043 * can use RCU again after this function is called. 1044 */ 1045 void tick_nohz_idle_exit(void) 1046 { 1047 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1048 ktime_t now; 1049 1050 local_irq_disable(); 1051 1052 WARN_ON_ONCE(!ts->inidle); 1053 1054 ts->inidle = 0; 1055 1056 if (ts->idle_active || ts->tick_stopped) 1057 now = ktime_get(); 1058 1059 if (ts->idle_active) 1060 tick_nohz_stop_idle(ts, now); 1061 1062 if (ts->tick_stopped) { 1063 tick_nohz_restart_sched_tick(ts, now, 0); 1064 tick_nohz_account_idle_ticks(ts); 1065 } 1066 1067 local_irq_enable(); 1068 } 1069 1070 /* 1071 * The nohz low res interrupt handler 1072 */ 1073 static void tick_nohz_handler(struct clock_event_device *dev) 1074 { 1075 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1076 struct pt_regs *regs = get_irq_regs(); 1077 ktime_t now = ktime_get(); 1078 1079 dev->next_event.tv64 = KTIME_MAX; 1080 1081 tick_sched_do_timer(now); 1082 tick_sched_handle(ts, regs); 1083 1084 /* No need to reprogram if we are running tickless */ 1085 if (unlikely(ts->tick_stopped)) 1086 return; 1087 1088 hrtimer_forward(&ts->sched_timer, now, tick_period); 1089 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1090 } 1091 1092 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) 1093 { 1094 if (!tick_nohz_enabled) 1095 return; 1096 ts->nohz_mode = mode; 1097 /* One update is enough */ 1098 if (!test_and_set_bit(0, &tick_nohz_active)) 1099 timers_update_migration(true); 1100 } 1101 1102 /** 1103 * tick_nohz_switch_to_nohz - switch to nohz mode 1104 */ 1105 static void tick_nohz_switch_to_nohz(void) 1106 { 1107 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1108 ktime_t next; 1109 1110 if (!tick_nohz_enabled) 1111 return; 1112 1113 if (tick_switch_to_oneshot(tick_nohz_handler)) 1114 return; 1115 1116 /* 1117 * Recycle the hrtimer in ts, so we can share the 1118 * hrtimer_forward with the highres code. 1119 */ 1120 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1121 /* Get the next period */ 1122 next = tick_init_jiffy_update(); 1123 1124 hrtimer_set_expires(&ts->sched_timer, next); 1125 hrtimer_forward_now(&ts->sched_timer, tick_period); 1126 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1127 tick_nohz_activate(ts, NOHZ_MODE_LOWRES); 1128 } 1129 1130 /* 1131 * When NOHZ is enabled and the tick is stopped, we need to kick the 1132 * tick timer from irq_enter() so that the jiffies update is kept 1133 * alive during long running softirqs. That's ugly as hell, but 1134 * correctness is key even if we need to fix the offending softirq in 1135 * the first place. 1136 * 1137 * Note, this is different to tick_nohz_restart. We just kick the 1138 * timer and do not touch the other magic bits which need to be done 1139 * when idle is left. 1140 */ 1141 static void tick_nohz_kick_tick(struct tick_sched *ts, ktime_t now) 1142 { 1143 #if 0 1144 /* Switch back to 2.6.27 behaviour */ 1145 ktime_t delta; 1146 1147 /* 1148 * Do not touch the tick device, when the next expiry is either 1149 * already reached or less/equal than the tick period. 1150 */ 1151 delta = ktime_sub(hrtimer_get_expires(&ts->sched_timer), now); 1152 if (delta.tv64 <= tick_period.tv64) 1153 return; 1154 1155 tick_nohz_restart(ts, now); 1156 #endif 1157 } 1158 1159 static inline void tick_nohz_irq_enter(void) 1160 { 1161 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1162 ktime_t now; 1163 1164 if (!ts->idle_active && !ts->tick_stopped) 1165 return; 1166 now = ktime_get(); 1167 if (ts->idle_active) 1168 tick_nohz_stop_idle(ts, now); 1169 if (ts->tick_stopped) { 1170 tick_nohz_update_jiffies(now); 1171 tick_nohz_kick_tick(ts, now); 1172 } 1173 } 1174 1175 #else 1176 1177 static inline void tick_nohz_switch_to_nohz(void) { } 1178 static inline void tick_nohz_irq_enter(void) { } 1179 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } 1180 1181 #endif /* CONFIG_NO_HZ_COMMON */ 1182 1183 /* 1184 * Called from irq_enter to notify about the possible interruption of idle() 1185 */ 1186 void tick_irq_enter(void) 1187 { 1188 tick_check_oneshot_broadcast_this_cpu(); 1189 tick_nohz_irq_enter(); 1190 } 1191 1192 /* 1193 * High resolution timer specific code 1194 */ 1195 #ifdef CONFIG_HIGH_RES_TIMERS 1196 /* 1197 * We rearm the timer until we get disabled by the idle code. 1198 * Called with interrupts disabled. 1199 */ 1200 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1201 { 1202 struct tick_sched *ts = 1203 container_of(timer, struct tick_sched, sched_timer); 1204 struct pt_regs *regs = get_irq_regs(); 1205 ktime_t now = ktime_get(); 1206 1207 tick_sched_do_timer(now); 1208 1209 /* 1210 * Do not call, when we are not in irq context and have 1211 * no valid regs pointer 1212 */ 1213 if (regs) 1214 tick_sched_handle(ts, regs); 1215 1216 /* No need to reprogram if we are in idle or full dynticks mode */ 1217 if (unlikely(ts->tick_stopped)) 1218 return HRTIMER_NORESTART; 1219 1220 hrtimer_forward(timer, now, tick_period); 1221 1222 return HRTIMER_RESTART; 1223 } 1224 1225 static int sched_skew_tick; 1226 1227 static int __init skew_tick(char *str) 1228 { 1229 get_option(&str, &sched_skew_tick); 1230 1231 return 0; 1232 } 1233 early_param("skew_tick", skew_tick); 1234 1235 /** 1236 * tick_setup_sched_timer - setup the tick emulation timer 1237 */ 1238 void tick_setup_sched_timer(void) 1239 { 1240 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1241 ktime_t now = ktime_get(); 1242 1243 /* 1244 * Emulate tick processing via per-CPU hrtimers: 1245 */ 1246 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1247 ts->sched_timer.function = tick_sched_timer; 1248 1249 /* Get the next period (per cpu) */ 1250 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1251 1252 /* Offset the tick to avert jiffies_lock contention. */ 1253 if (sched_skew_tick) { 1254 u64 offset = ktime_to_ns(tick_period) >> 1; 1255 do_div(offset, num_possible_cpus()); 1256 offset *= smp_processor_id(); 1257 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1258 } 1259 1260 hrtimer_forward(&ts->sched_timer, now, tick_period); 1261 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 1262 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); 1263 } 1264 #endif /* HIGH_RES_TIMERS */ 1265 1266 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1267 void tick_cancel_sched_timer(int cpu) 1268 { 1269 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1270 1271 # ifdef CONFIG_HIGH_RES_TIMERS 1272 if (ts->sched_timer.base) 1273 hrtimer_cancel(&ts->sched_timer); 1274 # endif 1275 1276 memset(ts, 0, sizeof(*ts)); 1277 } 1278 #endif 1279 1280 /** 1281 * Async notification about clocksource changes 1282 */ 1283 void tick_clock_notify(void) 1284 { 1285 int cpu; 1286 1287 for_each_possible_cpu(cpu) 1288 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1289 } 1290 1291 /* 1292 * Async notification about clock event changes 1293 */ 1294 void tick_oneshot_notify(void) 1295 { 1296 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1297 1298 set_bit(0, &ts->check_clocks); 1299 } 1300 1301 /** 1302 * Check, if a change happened, which makes oneshot possible. 1303 * 1304 * Called cyclic from the hrtimer softirq (driven by the timer 1305 * softirq) allow_nohz signals, that we can switch into low-res nohz 1306 * mode, because high resolution timers are disabled (either compile 1307 * or runtime). Called with interrupts disabled. 1308 */ 1309 int tick_check_oneshot_change(int allow_nohz) 1310 { 1311 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1312 1313 if (!test_and_clear_bit(0, &ts->check_clocks)) 1314 return 0; 1315 1316 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1317 return 0; 1318 1319 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1320 return 0; 1321 1322 if (!allow_nohz) 1323 return 1; 1324 1325 tick_nohz_switch_to_nohz(); 1326 return 0; 1327 } 1328