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/nmi.h> 21 #include <linux/profile.h> 22 #include <linux/sched/signal.h> 23 #include <linux/sched/clock.h> 24 #include <linux/sched/stat.h> 25 #include <linux/module.h> 26 #include <linux/irq_work.h> 27 #include <linux/posix-timers.h> 28 #include <linux/context_tracking.h> 29 30 #include <asm/irq_regs.h> 31 32 #include "tick-internal.h" 33 34 #include <trace/events/timer.h> 35 36 /* 37 * Per-CPU nohz control structure 38 */ 39 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); 40 41 struct tick_sched *tick_get_tick_sched(int cpu) 42 { 43 return &per_cpu(tick_cpu_sched, cpu); 44 } 45 46 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS) 47 /* 48 * The time, when the last jiffy update happened. Protected by jiffies_lock. 49 */ 50 static ktime_t last_jiffies_update; 51 52 /* 53 * Must be called with interrupts disabled ! 54 */ 55 static void tick_do_update_jiffies64(ktime_t now) 56 { 57 unsigned long ticks = 0; 58 ktime_t delta; 59 60 /* 61 * Do a quick check without holding jiffies_lock: 62 */ 63 delta = ktime_sub(now, last_jiffies_update); 64 if (delta < tick_period) 65 return; 66 67 /* Reevaluate with jiffies_lock held */ 68 write_seqlock(&jiffies_lock); 69 70 delta = ktime_sub(now, last_jiffies_update); 71 if (delta >= tick_period) { 72 73 delta = ktime_sub(delta, tick_period); 74 last_jiffies_update = ktime_add(last_jiffies_update, 75 tick_period); 76 77 /* Slow path for long timeouts */ 78 if (unlikely(delta >= tick_period)) { 79 s64 incr = ktime_to_ns(tick_period); 80 81 ticks = ktime_divns(delta, incr); 82 83 last_jiffies_update = ktime_add_ns(last_jiffies_update, 84 incr * ticks); 85 } 86 do_timer(++ticks); 87 88 /* Keep the tick_next_period variable up to date */ 89 tick_next_period = ktime_add(last_jiffies_update, tick_period); 90 } else { 91 write_sequnlock(&jiffies_lock); 92 return; 93 } 94 write_sequnlock(&jiffies_lock); 95 update_wall_time(); 96 } 97 98 /* 99 * Initialize and return retrieve the jiffies update. 100 */ 101 static ktime_t tick_init_jiffy_update(void) 102 { 103 ktime_t period; 104 105 write_seqlock(&jiffies_lock); 106 /* Did we start the jiffies update yet ? */ 107 if (last_jiffies_update == 0) 108 last_jiffies_update = tick_next_period; 109 period = last_jiffies_update; 110 write_sequnlock(&jiffies_lock); 111 return period; 112 } 113 114 115 static void tick_sched_do_timer(ktime_t now) 116 { 117 int cpu = smp_processor_id(); 118 119 #ifdef CONFIG_NO_HZ_COMMON 120 /* 121 * Check if the do_timer duty was dropped. We don't care about 122 * concurrency: This happens only when the CPU in charge went 123 * into a long sleep. If two CPUs happen to assign themselves to 124 * this duty, then the jiffies update is still serialized by 125 * jiffies_lock. 126 */ 127 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE) 128 && !tick_nohz_full_cpu(cpu)) 129 tick_do_timer_cpu = cpu; 130 #endif 131 132 /* Check, if the jiffies need an update */ 133 if (tick_do_timer_cpu == cpu) 134 tick_do_update_jiffies64(now); 135 } 136 137 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) 138 { 139 #ifdef CONFIG_NO_HZ_COMMON 140 /* 141 * When we are idle and the tick is stopped, we have to touch 142 * the watchdog as we might not schedule for a really long 143 * time. This happens on complete idle SMP systems while 144 * waiting on the login prompt. We also increment the "start of 145 * idle" jiffy stamp so the idle accounting adjustment we do 146 * when we go busy again does not account too much ticks. 147 */ 148 if (ts->tick_stopped) { 149 touch_softlockup_watchdog_sched(); 150 if (is_idle_task(current)) 151 ts->idle_jiffies++; 152 } 153 #endif 154 update_process_times(user_mode(regs)); 155 profile_tick(CPU_PROFILING); 156 } 157 #endif 158 159 #ifdef CONFIG_NO_HZ_FULL 160 cpumask_var_t tick_nohz_full_mask; 161 cpumask_var_t housekeeping_mask; 162 bool tick_nohz_full_running; 163 static atomic_t tick_dep_mask; 164 165 static bool check_tick_dependency(atomic_t *dep) 166 { 167 int val = atomic_read(dep); 168 169 if (val & TICK_DEP_MASK_POSIX_TIMER) { 170 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER); 171 return true; 172 } 173 174 if (val & TICK_DEP_MASK_PERF_EVENTS) { 175 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS); 176 return true; 177 } 178 179 if (val & TICK_DEP_MASK_SCHED) { 180 trace_tick_stop(0, TICK_DEP_MASK_SCHED); 181 return true; 182 } 183 184 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) { 185 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE); 186 return true; 187 } 188 189 return false; 190 } 191 192 static bool can_stop_full_tick(int cpu, struct tick_sched *ts) 193 { 194 WARN_ON_ONCE(!irqs_disabled()); 195 196 if (unlikely(!cpu_online(cpu))) 197 return false; 198 199 if (check_tick_dependency(&tick_dep_mask)) 200 return false; 201 202 if (check_tick_dependency(&ts->tick_dep_mask)) 203 return false; 204 205 if (check_tick_dependency(¤t->tick_dep_mask)) 206 return false; 207 208 if (check_tick_dependency(¤t->signal->tick_dep_mask)) 209 return false; 210 211 return true; 212 } 213 214 static void nohz_full_kick_func(struct irq_work *work) 215 { 216 /* Empty, the tick restart happens on tick_nohz_irq_exit() */ 217 } 218 219 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 220 .func = nohz_full_kick_func, 221 }; 222 223 /* 224 * Kick this CPU if it's full dynticks in order to force it to 225 * re-evaluate its dependency on the tick and restart it if necessary. 226 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), 227 * is NMI safe. 228 */ 229 static void tick_nohz_full_kick(void) 230 { 231 if (!tick_nohz_full_cpu(smp_processor_id())) 232 return; 233 234 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work)); 235 } 236 237 /* 238 * Kick the CPU if it's full dynticks in order to force it to 239 * re-evaluate its dependency on the tick and restart it if necessary. 240 */ 241 void tick_nohz_full_kick_cpu(int cpu) 242 { 243 if (!tick_nohz_full_cpu(cpu)) 244 return; 245 246 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); 247 } 248 249 /* 250 * Kick all full dynticks CPUs in order to force these to re-evaluate 251 * their dependency on the tick and restart it if necessary. 252 */ 253 static void tick_nohz_full_kick_all(void) 254 { 255 int cpu; 256 257 if (!tick_nohz_full_running) 258 return; 259 260 preempt_disable(); 261 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask) 262 tick_nohz_full_kick_cpu(cpu); 263 preempt_enable(); 264 } 265 266 static void tick_nohz_dep_set_all(atomic_t *dep, 267 enum tick_dep_bits bit) 268 { 269 int prev; 270 271 prev = atomic_fetch_or(BIT(bit), dep); 272 if (!prev) 273 tick_nohz_full_kick_all(); 274 } 275 276 /* 277 * Set a global tick dependency. Used by perf events that rely on freq and 278 * by unstable clock. 279 */ 280 void tick_nohz_dep_set(enum tick_dep_bits bit) 281 { 282 tick_nohz_dep_set_all(&tick_dep_mask, bit); 283 } 284 285 void tick_nohz_dep_clear(enum tick_dep_bits bit) 286 { 287 atomic_andnot(BIT(bit), &tick_dep_mask); 288 } 289 290 /* 291 * Set per-CPU tick dependency. Used by scheduler and perf events in order to 292 * manage events throttling. 293 */ 294 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit) 295 { 296 int prev; 297 struct tick_sched *ts; 298 299 ts = per_cpu_ptr(&tick_cpu_sched, cpu); 300 301 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask); 302 if (!prev) { 303 preempt_disable(); 304 /* Perf needs local kick that is NMI safe */ 305 if (cpu == smp_processor_id()) { 306 tick_nohz_full_kick(); 307 } else { 308 /* Remote irq work not NMI-safe */ 309 if (!WARN_ON_ONCE(in_nmi())) 310 tick_nohz_full_kick_cpu(cpu); 311 } 312 preempt_enable(); 313 } 314 } 315 316 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit) 317 { 318 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); 319 320 atomic_andnot(BIT(bit), &ts->tick_dep_mask); 321 } 322 323 /* 324 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse 325 * per task timers. 326 */ 327 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit) 328 { 329 /* 330 * We could optimize this with just kicking the target running the task 331 * if that noise matters for nohz full users. 332 */ 333 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit); 334 } 335 336 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit) 337 { 338 atomic_andnot(BIT(bit), &tsk->tick_dep_mask); 339 } 340 341 /* 342 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse 343 * per process timers. 344 */ 345 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit) 346 { 347 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit); 348 } 349 350 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit) 351 { 352 atomic_andnot(BIT(bit), &sig->tick_dep_mask); 353 } 354 355 /* 356 * Re-evaluate the need for the tick as we switch the current task. 357 * It might need the tick due to per task/process properties: 358 * perf events, posix CPU timers, ... 359 */ 360 void __tick_nohz_task_switch(void) 361 { 362 unsigned long flags; 363 struct tick_sched *ts; 364 365 local_irq_save(flags); 366 367 if (!tick_nohz_full_cpu(smp_processor_id())) 368 goto out; 369 370 ts = this_cpu_ptr(&tick_cpu_sched); 371 372 if (ts->tick_stopped) { 373 if (atomic_read(¤t->tick_dep_mask) || 374 atomic_read(¤t->signal->tick_dep_mask)) 375 tick_nohz_full_kick(); 376 } 377 out: 378 local_irq_restore(flags); 379 } 380 381 /* Parse the boot-time nohz CPU list from the kernel parameters. */ 382 static int __init tick_nohz_full_setup(char *str) 383 { 384 alloc_bootmem_cpumask_var(&tick_nohz_full_mask); 385 if (cpulist_parse(str, tick_nohz_full_mask) < 0) { 386 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n"); 387 free_bootmem_cpumask_var(tick_nohz_full_mask); 388 return 1; 389 } 390 tick_nohz_full_running = true; 391 392 return 1; 393 } 394 __setup("nohz_full=", tick_nohz_full_setup); 395 396 static int tick_nohz_cpu_down(unsigned int cpu) 397 { 398 /* 399 * The boot CPU handles housekeeping duty (unbound timers, 400 * workqueues, timekeeping, ...) on behalf of full dynticks 401 * CPUs. It must remain online when nohz full is enabled. 402 */ 403 if (tick_nohz_full_running && tick_do_timer_cpu == cpu) 404 return -EBUSY; 405 return 0; 406 } 407 408 static int tick_nohz_init_all(void) 409 { 410 int err = -1; 411 412 #ifdef CONFIG_NO_HZ_FULL_ALL 413 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { 414 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n"); 415 return err; 416 } 417 err = 0; 418 cpumask_setall(tick_nohz_full_mask); 419 tick_nohz_full_running = true; 420 #endif 421 return err; 422 } 423 424 void __init tick_nohz_init(void) 425 { 426 int cpu, ret; 427 428 if (!tick_nohz_full_running) { 429 if (tick_nohz_init_all() < 0) 430 return; 431 } 432 433 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) { 434 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n"); 435 cpumask_clear(tick_nohz_full_mask); 436 tick_nohz_full_running = false; 437 return; 438 } 439 440 /* 441 * Full dynticks uses irq work to drive the tick rescheduling on safe 442 * locking contexts. But then we need irq work to raise its own 443 * interrupts to avoid circular dependency on the tick 444 */ 445 if (!arch_irq_work_has_interrupt()) { 446 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n"); 447 cpumask_clear(tick_nohz_full_mask); 448 cpumask_copy(housekeeping_mask, cpu_possible_mask); 449 tick_nohz_full_running = false; 450 return; 451 } 452 453 cpu = smp_processor_id(); 454 455 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 456 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", 457 cpu); 458 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 459 } 460 461 cpumask_andnot(housekeeping_mask, 462 cpu_possible_mask, tick_nohz_full_mask); 463 464 for_each_cpu(cpu, tick_nohz_full_mask) 465 context_tracking_cpu_set(cpu); 466 467 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 468 "kernel/nohz:predown", NULL, 469 tick_nohz_cpu_down); 470 WARN_ON(ret < 0); 471 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", 472 cpumask_pr_args(tick_nohz_full_mask)); 473 474 /* 475 * We need at least one CPU to handle housekeeping work such 476 * as timekeeping, unbound timers, workqueues, ... 477 */ 478 WARN_ON_ONCE(cpumask_empty(housekeeping_mask)); 479 } 480 #endif 481 482 /* 483 * NOHZ - aka dynamic tick functionality 484 */ 485 #ifdef CONFIG_NO_HZ_COMMON 486 /* 487 * NO HZ enabled ? 488 */ 489 bool tick_nohz_enabled __read_mostly = true; 490 unsigned long tick_nohz_active __read_mostly; 491 /* 492 * Enable / Disable tickless mode 493 */ 494 static int __init setup_tick_nohz(char *str) 495 { 496 return (kstrtobool(str, &tick_nohz_enabled) == 0); 497 } 498 499 __setup("nohz=", setup_tick_nohz); 500 501 int tick_nohz_tick_stopped(void) 502 { 503 return __this_cpu_read(tick_cpu_sched.tick_stopped); 504 } 505 506 /** 507 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 508 * 509 * Called from interrupt entry when the CPU was idle 510 * 511 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 512 * must be updated. Otherwise an interrupt handler could use a stale jiffy 513 * value. We do this unconditionally on any CPU, as we don't know whether the 514 * CPU, which has the update task assigned is in a long sleep. 515 */ 516 static void tick_nohz_update_jiffies(ktime_t now) 517 { 518 unsigned long flags; 519 520 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 521 522 local_irq_save(flags); 523 tick_do_update_jiffies64(now); 524 local_irq_restore(flags); 525 526 touch_softlockup_watchdog_sched(); 527 } 528 529 /* 530 * Updates the per-CPU time idle statistics counters 531 */ 532 static void 533 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 534 { 535 ktime_t delta; 536 537 if (ts->idle_active) { 538 delta = ktime_sub(now, ts->idle_entrytime); 539 if (nr_iowait_cpu(cpu) > 0) 540 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 541 else 542 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 543 ts->idle_entrytime = now; 544 } 545 546 if (last_update_time) 547 *last_update_time = ktime_to_us(now); 548 549 } 550 551 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 552 { 553 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 554 ts->idle_active = 0; 555 556 sched_clock_idle_wakeup_event(0); 557 } 558 559 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 560 { 561 ktime_t now = ktime_get(); 562 563 ts->idle_entrytime = now; 564 ts->idle_active = 1; 565 sched_clock_idle_sleep_event(); 566 return now; 567 } 568 569 /** 570 * get_cpu_idle_time_us - get the total idle time of a CPU 571 * @cpu: CPU number to query 572 * @last_update_time: variable to store update time in. Do not update 573 * counters if NULL. 574 * 575 * Return the cumulative idle time (since boot) for a given 576 * CPU, in microseconds. 577 * 578 * This time is measured via accounting rather than sampling, 579 * and is as accurate as ktime_get() is. 580 * 581 * This function returns -1 if NOHZ is not enabled. 582 */ 583 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 584 { 585 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 586 ktime_t now, idle; 587 588 if (!tick_nohz_active) 589 return -1; 590 591 now = ktime_get(); 592 if (last_update_time) { 593 update_ts_time_stats(cpu, ts, now, last_update_time); 594 idle = ts->idle_sleeptime; 595 } else { 596 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 597 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 598 599 idle = ktime_add(ts->idle_sleeptime, delta); 600 } else { 601 idle = ts->idle_sleeptime; 602 } 603 } 604 605 return ktime_to_us(idle); 606 607 } 608 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 609 610 /** 611 * get_cpu_iowait_time_us - get the total iowait time of a CPU 612 * @cpu: CPU number to query 613 * @last_update_time: variable to store update time in. Do not update 614 * counters if NULL. 615 * 616 * Return the cumulative iowait time (since boot) for a given 617 * CPU, in microseconds. 618 * 619 * This time is measured via accounting rather than sampling, 620 * and is as accurate as ktime_get() is. 621 * 622 * This function returns -1 if NOHZ is not enabled. 623 */ 624 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 625 { 626 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 627 ktime_t now, iowait; 628 629 if (!tick_nohz_active) 630 return -1; 631 632 now = ktime_get(); 633 if (last_update_time) { 634 update_ts_time_stats(cpu, ts, now, last_update_time); 635 iowait = ts->iowait_sleeptime; 636 } else { 637 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 638 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 639 640 iowait = ktime_add(ts->iowait_sleeptime, delta); 641 } else { 642 iowait = ts->iowait_sleeptime; 643 } 644 } 645 646 return ktime_to_us(iowait); 647 } 648 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 649 650 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 651 { 652 hrtimer_cancel(&ts->sched_timer); 653 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 654 655 /* Forward the time to expire in the future */ 656 hrtimer_forward(&ts->sched_timer, now, tick_period); 657 658 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 659 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 660 else 661 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 662 } 663 664 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 665 ktime_t now, int cpu) 666 { 667 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 668 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; 669 unsigned long seq, basejiff; 670 ktime_t tick; 671 672 /* Read jiffies and the time when jiffies were updated last */ 673 do { 674 seq = read_seqbegin(&jiffies_lock); 675 basemono = last_jiffies_update; 676 basejiff = jiffies; 677 } while (read_seqretry(&jiffies_lock, seq)); 678 ts->last_jiffies = basejiff; 679 680 if (rcu_needs_cpu(basemono, &next_rcu) || 681 arch_needs_cpu() || irq_work_needs_cpu()) { 682 next_tick = basemono + TICK_NSEC; 683 } else { 684 /* 685 * Get the next pending timer. If high resolution 686 * timers are enabled this only takes the timer wheel 687 * timers into account. If high resolution timers are 688 * disabled this also looks at the next expiring 689 * hrtimer. 690 */ 691 next_tmr = get_next_timer_interrupt(basejiff, basemono); 692 ts->next_timer = next_tmr; 693 /* Take the next rcu event into account */ 694 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; 695 } 696 697 /* 698 * If the tick is due in the next period, keep it ticking or 699 * force prod the timer. 700 */ 701 delta = next_tick - basemono; 702 if (delta <= (u64)TICK_NSEC) { 703 tick = 0; 704 705 /* 706 * Tell the timer code that the base is not idle, i.e. undo 707 * the effect of get_next_timer_interrupt(): 708 */ 709 timer_clear_idle(); 710 /* 711 * We've not stopped the tick yet, and there's a timer in the 712 * next period, so no point in stopping it either, bail. 713 */ 714 if (!ts->tick_stopped) 715 goto out; 716 717 /* 718 * If, OTOH, we did stop it, but there's a pending (expired) 719 * timer reprogram the timer hardware to fire now. 720 * 721 * We will not restart the tick proper, just prod the timer 722 * hardware into firing an interrupt to process the pending 723 * timers. Just like tick_irq_exit() will not restart the tick 724 * for 'normal' interrupts. 725 * 726 * Only once we exit the idle loop will we re-enable the tick, 727 * see tick_nohz_idle_exit(). 728 */ 729 if (delta == 0) { 730 tick_nohz_restart(ts, now); 731 goto out; 732 } 733 } 734 735 /* 736 * If this CPU is the one which updates jiffies, then give up 737 * the assignment and let it be taken by the CPU which runs 738 * the tick timer next, which might be this CPU as well. If we 739 * don't drop this here the jiffies might be stale and 740 * do_timer() never invoked. Keep track of the fact that it 741 * was the one which had the do_timer() duty last. If this CPU 742 * is the one which had the do_timer() duty last, we limit the 743 * sleep time to the timekeeping max_deferment value. 744 * Otherwise we can sleep as long as we want. 745 */ 746 delta = timekeeping_max_deferment(); 747 if (cpu == tick_do_timer_cpu) { 748 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 749 ts->do_timer_last = 1; 750 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 751 delta = KTIME_MAX; 752 ts->do_timer_last = 0; 753 } else if (!ts->do_timer_last) { 754 delta = KTIME_MAX; 755 } 756 757 #ifdef CONFIG_NO_HZ_FULL 758 /* Limit the tick delta to the maximum scheduler deferment */ 759 if (!ts->inidle) 760 delta = min(delta, scheduler_tick_max_deferment()); 761 #endif 762 763 /* Calculate the next expiry time */ 764 if (delta < (KTIME_MAX - basemono)) 765 expires = basemono + delta; 766 else 767 expires = KTIME_MAX; 768 769 expires = min_t(u64, expires, next_tick); 770 tick = expires; 771 772 /* Skip reprogram of event if its not changed */ 773 if (ts->tick_stopped && (expires == dev->next_event)) 774 goto out; 775 776 /* 777 * nohz_stop_sched_tick can be called several times before 778 * the nohz_restart_sched_tick is called. This happens when 779 * interrupts arrive which do not cause a reschedule. In the 780 * first call we save the current tick time, so we can restart 781 * the scheduler tick in nohz_restart_sched_tick. 782 */ 783 if (!ts->tick_stopped) { 784 nohz_balance_enter_idle(cpu); 785 calc_load_enter_idle(); 786 cpu_load_update_nohz_start(); 787 788 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 789 ts->tick_stopped = 1; 790 trace_tick_stop(1, TICK_DEP_MASK_NONE); 791 } 792 793 /* 794 * If the expiration time == KTIME_MAX, then we simply stop 795 * the tick timer. 796 */ 797 if (unlikely(expires == KTIME_MAX)) { 798 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 799 hrtimer_cancel(&ts->sched_timer); 800 goto out; 801 } 802 803 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 804 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED); 805 else 806 tick_program_event(tick, 1); 807 out: 808 /* Update the estimated sleep length */ 809 ts->sleep_length = ktime_sub(dev->next_event, now); 810 return tick; 811 } 812 813 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 814 { 815 /* Update jiffies first */ 816 tick_do_update_jiffies64(now); 817 cpu_load_update_nohz_stop(); 818 819 /* 820 * Clear the timer idle flag, so we avoid IPIs on remote queueing and 821 * the clock forward checks in the enqueue path: 822 */ 823 timer_clear_idle(); 824 825 calc_load_exit_idle(); 826 touch_softlockup_watchdog_sched(); 827 /* 828 * Cancel the scheduled timer and restore the tick 829 */ 830 ts->tick_stopped = 0; 831 ts->idle_exittime = now; 832 833 tick_nohz_restart(ts, now); 834 } 835 836 static void tick_nohz_full_update_tick(struct tick_sched *ts) 837 { 838 #ifdef CONFIG_NO_HZ_FULL 839 int cpu = smp_processor_id(); 840 841 if (!tick_nohz_full_cpu(cpu)) 842 return; 843 844 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 845 return; 846 847 if (can_stop_full_tick(cpu, ts)) 848 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 849 else if (ts->tick_stopped) 850 tick_nohz_restart_sched_tick(ts, ktime_get()); 851 #endif 852 } 853 854 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 855 { 856 /* 857 * If this CPU is offline and it is the one which updates 858 * jiffies, then give up the assignment and let it be taken by 859 * the CPU which runs the tick timer next. If we don't drop 860 * this here the jiffies might be stale and do_timer() never 861 * invoked. 862 */ 863 if (unlikely(!cpu_online(cpu))) { 864 if (cpu == tick_do_timer_cpu) 865 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 866 return false; 867 } 868 869 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 870 ts->sleep_length = NSEC_PER_SEC / HZ; 871 return false; 872 } 873 874 if (need_resched()) 875 return false; 876 877 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 878 static int ratelimit; 879 880 if (ratelimit < 10 && 881 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 882 pr_warn("NOHZ: local_softirq_pending %02x\n", 883 (unsigned int) local_softirq_pending()); 884 ratelimit++; 885 } 886 return false; 887 } 888 889 if (tick_nohz_full_enabled()) { 890 /* 891 * Keep the tick alive to guarantee timekeeping progression 892 * if there are full dynticks CPUs around 893 */ 894 if (tick_do_timer_cpu == cpu) 895 return false; 896 /* 897 * Boot safety: make sure the timekeeping duty has been 898 * assigned before entering dyntick-idle mode, 899 */ 900 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 901 return false; 902 } 903 904 return true; 905 } 906 907 static void __tick_nohz_idle_enter(struct tick_sched *ts) 908 { 909 ktime_t now, expires; 910 int cpu = smp_processor_id(); 911 912 now = tick_nohz_start_idle(ts); 913 914 if (can_stop_idle_tick(cpu, ts)) { 915 int was_stopped = ts->tick_stopped; 916 917 ts->idle_calls++; 918 919 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 920 if (expires > 0LL) { 921 ts->idle_sleeps++; 922 ts->idle_expires = expires; 923 } 924 925 if (!was_stopped && ts->tick_stopped) 926 ts->idle_jiffies = ts->last_jiffies; 927 } 928 } 929 930 /** 931 * tick_nohz_idle_enter - stop the idle tick from the idle task 932 * 933 * When the next event is more than a tick into the future, stop the idle tick 934 * Called when we start the idle loop. 935 * 936 * The arch is responsible of calling: 937 * 938 * - rcu_idle_enter() after its last use of RCU before the CPU is put 939 * to sleep. 940 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 941 */ 942 void tick_nohz_idle_enter(void) 943 { 944 struct tick_sched *ts; 945 946 WARN_ON_ONCE(irqs_disabled()); 947 948 /* 949 * Update the idle state in the scheduler domain hierarchy 950 * when tick_nohz_stop_sched_tick() is called from the idle loop. 951 * State will be updated to busy during the first busy tick after 952 * exiting idle. 953 */ 954 set_cpu_sd_state_idle(); 955 956 local_irq_disable(); 957 958 ts = this_cpu_ptr(&tick_cpu_sched); 959 ts->inidle = 1; 960 __tick_nohz_idle_enter(ts); 961 962 local_irq_enable(); 963 } 964 965 /** 966 * tick_nohz_irq_exit - update next tick event from interrupt exit 967 * 968 * When an interrupt fires while we are idle and it doesn't cause 969 * a reschedule, it may still add, modify or delete a timer, enqueue 970 * an RCU callback, etc... 971 * So we need to re-calculate and reprogram the next tick event. 972 */ 973 void tick_nohz_irq_exit(void) 974 { 975 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 976 977 if (ts->inidle) 978 __tick_nohz_idle_enter(ts); 979 else 980 tick_nohz_full_update_tick(ts); 981 } 982 983 /** 984 * tick_nohz_get_sleep_length - return the length of the current sleep 985 * 986 * Called from power state control code with interrupts disabled 987 */ 988 ktime_t tick_nohz_get_sleep_length(void) 989 { 990 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 991 992 return ts->sleep_length; 993 } 994 995 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 996 { 997 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 998 unsigned long ticks; 999 1000 if (vtime_accounting_cpu_enabled()) 1001 return; 1002 /* 1003 * We stopped the tick in idle. Update process times would miss the 1004 * time we slept as update_process_times does only a 1 tick 1005 * accounting. Enforce that this is accounted to idle ! 1006 */ 1007 ticks = jiffies - ts->idle_jiffies; 1008 /* 1009 * We might be one off. Do not randomly account a huge number of ticks! 1010 */ 1011 if (ticks && ticks < LONG_MAX) 1012 account_idle_ticks(ticks); 1013 #endif 1014 } 1015 1016 /** 1017 * tick_nohz_idle_exit - restart the idle tick from the idle task 1018 * 1019 * Restart the idle tick when the CPU is woken up from idle 1020 * This also exit the RCU extended quiescent state. The CPU 1021 * can use RCU again after this function is called. 1022 */ 1023 void tick_nohz_idle_exit(void) 1024 { 1025 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1026 ktime_t now; 1027 1028 local_irq_disable(); 1029 1030 WARN_ON_ONCE(!ts->inidle); 1031 1032 ts->inidle = 0; 1033 1034 if (ts->idle_active || ts->tick_stopped) 1035 now = ktime_get(); 1036 1037 if (ts->idle_active) 1038 tick_nohz_stop_idle(ts, now); 1039 1040 if (ts->tick_stopped) { 1041 tick_nohz_restart_sched_tick(ts, now); 1042 tick_nohz_account_idle_ticks(ts); 1043 } 1044 1045 local_irq_enable(); 1046 } 1047 1048 /* 1049 * The nohz low res interrupt handler 1050 */ 1051 static void tick_nohz_handler(struct clock_event_device *dev) 1052 { 1053 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1054 struct pt_regs *regs = get_irq_regs(); 1055 ktime_t now = ktime_get(); 1056 1057 dev->next_event = KTIME_MAX; 1058 1059 tick_sched_do_timer(now); 1060 tick_sched_handle(ts, regs); 1061 1062 /* No need to reprogram if we are running tickless */ 1063 if (unlikely(ts->tick_stopped)) 1064 return; 1065 1066 hrtimer_forward(&ts->sched_timer, now, tick_period); 1067 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1068 } 1069 1070 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) 1071 { 1072 if (!tick_nohz_enabled) 1073 return; 1074 ts->nohz_mode = mode; 1075 /* One update is enough */ 1076 if (!test_and_set_bit(0, &tick_nohz_active)) 1077 timers_update_migration(true); 1078 } 1079 1080 /** 1081 * tick_nohz_switch_to_nohz - switch to nohz mode 1082 */ 1083 static void tick_nohz_switch_to_nohz(void) 1084 { 1085 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1086 ktime_t next; 1087 1088 if (!tick_nohz_enabled) 1089 return; 1090 1091 if (tick_switch_to_oneshot(tick_nohz_handler)) 1092 return; 1093 1094 /* 1095 * Recycle the hrtimer in ts, so we can share the 1096 * hrtimer_forward with the highres code. 1097 */ 1098 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1099 /* Get the next period */ 1100 next = tick_init_jiffy_update(); 1101 1102 hrtimer_set_expires(&ts->sched_timer, next); 1103 hrtimer_forward_now(&ts->sched_timer, tick_period); 1104 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1105 tick_nohz_activate(ts, NOHZ_MODE_LOWRES); 1106 } 1107 1108 static inline void tick_nohz_irq_enter(void) 1109 { 1110 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1111 ktime_t now; 1112 1113 if (!ts->idle_active && !ts->tick_stopped) 1114 return; 1115 now = ktime_get(); 1116 if (ts->idle_active) 1117 tick_nohz_stop_idle(ts, now); 1118 if (ts->tick_stopped) 1119 tick_nohz_update_jiffies(now); 1120 } 1121 1122 #else 1123 1124 static inline void tick_nohz_switch_to_nohz(void) { } 1125 static inline void tick_nohz_irq_enter(void) { } 1126 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } 1127 1128 #endif /* CONFIG_NO_HZ_COMMON */ 1129 1130 /* 1131 * Called from irq_enter to notify about the possible interruption of idle() 1132 */ 1133 void tick_irq_enter(void) 1134 { 1135 tick_check_oneshot_broadcast_this_cpu(); 1136 tick_nohz_irq_enter(); 1137 } 1138 1139 /* 1140 * High resolution timer specific code 1141 */ 1142 #ifdef CONFIG_HIGH_RES_TIMERS 1143 /* 1144 * We rearm the timer until we get disabled by the idle code. 1145 * Called with interrupts disabled. 1146 */ 1147 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1148 { 1149 struct tick_sched *ts = 1150 container_of(timer, struct tick_sched, sched_timer); 1151 struct pt_regs *regs = get_irq_regs(); 1152 ktime_t now = ktime_get(); 1153 1154 tick_sched_do_timer(now); 1155 1156 /* 1157 * Do not call, when we are not in irq context and have 1158 * no valid regs pointer 1159 */ 1160 if (regs) 1161 tick_sched_handle(ts, regs); 1162 1163 /* No need to reprogram if we are in idle or full dynticks mode */ 1164 if (unlikely(ts->tick_stopped)) 1165 return HRTIMER_NORESTART; 1166 1167 hrtimer_forward(timer, now, tick_period); 1168 1169 return HRTIMER_RESTART; 1170 } 1171 1172 static int sched_skew_tick; 1173 1174 static int __init skew_tick(char *str) 1175 { 1176 get_option(&str, &sched_skew_tick); 1177 1178 return 0; 1179 } 1180 early_param("skew_tick", skew_tick); 1181 1182 /** 1183 * tick_setup_sched_timer - setup the tick emulation timer 1184 */ 1185 void tick_setup_sched_timer(void) 1186 { 1187 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1188 ktime_t now = ktime_get(); 1189 1190 /* 1191 * Emulate tick processing via per-CPU hrtimers: 1192 */ 1193 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1194 ts->sched_timer.function = tick_sched_timer; 1195 1196 /* Get the next period (per-CPU) */ 1197 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1198 1199 /* Offset the tick to avert jiffies_lock contention. */ 1200 if (sched_skew_tick) { 1201 u64 offset = ktime_to_ns(tick_period) >> 1; 1202 do_div(offset, num_possible_cpus()); 1203 offset *= smp_processor_id(); 1204 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1205 } 1206 1207 hrtimer_forward(&ts->sched_timer, now, tick_period); 1208 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 1209 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); 1210 } 1211 #endif /* HIGH_RES_TIMERS */ 1212 1213 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1214 void tick_cancel_sched_timer(int cpu) 1215 { 1216 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1217 1218 # ifdef CONFIG_HIGH_RES_TIMERS 1219 if (ts->sched_timer.base) 1220 hrtimer_cancel(&ts->sched_timer); 1221 # endif 1222 1223 memset(ts, 0, sizeof(*ts)); 1224 } 1225 #endif 1226 1227 /** 1228 * Async notification about clocksource changes 1229 */ 1230 void tick_clock_notify(void) 1231 { 1232 int cpu; 1233 1234 for_each_possible_cpu(cpu) 1235 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1236 } 1237 1238 /* 1239 * Async notification about clock event changes 1240 */ 1241 void tick_oneshot_notify(void) 1242 { 1243 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1244 1245 set_bit(0, &ts->check_clocks); 1246 } 1247 1248 /** 1249 * Check, if a change happened, which makes oneshot possible. 1250 * 1251 * Called cyclic from the hrtimer softirq (driven by the timer 1252 * softirq) allow_nohz signals, that we can switch into low-res nohz 1253 * mode, because high resolution timers are disabled (either compile 1254 * or runtime). Called with interrupts disabled. 1255 */ 1256 int tick_check_oneshot_change(int allow_nohz) 1257 { 1258 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1259 1260 if (!test_and_clear_bit(0, &ts->check_clocks)) 1261 return 0; 1262 1263 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1264 return 0; 1265 1266 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1267 return 0; 1268 1269 if (!allow_nohz) 1270 return 1; 1271 1272 tick_nohz_switch_to_nohz(); 1273 return 0; 1274 } 1275