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/sched/nohz.h> 26 #include <linux/module.h> 27 #include <linux/irq_work.h> 28 #include <linux/posix-timers.h> 29 #include <linux/context_tracking.h> 30 31 #include <asm/irq_regs.h> 32 33 #include "tick-internal.h" 34 35 #include <trace/events/timer.h> 36 37 /* 38 * Per-CPU nohz control structure 39 */ 40 static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched); 41 42 struct tick_sched *tick_get_tick_sched(int cpu) 43 { 44 return &per_cpu(tick_cpu_sched, cpu); 45 } 46 47 #if defined(CONFIG_NO_HZ_COMMON) || defined(CONFIG_HIGH_RES_TIMERS) 48 /* 49 * The time, when the last jiffy update happened. Protected by jiffies_lock. 50 */ 51 static ktime_t last_jiffies_update; 52 53 /* 54 * Must be called with interrupts disabled ! 55 */ 56 static void tick_do_update_jiffies64(ktime_t now) 57 { 58 unsigned long ticks = 0; 59 ktime_t delta; 60 61 /* 62 * Do a quick check without holding jiffies_lock: 63 */ 64 delta = ktime_sub(now, last_jiffies_update); 65 if (delta < tick_period) 66 return; 67 68 /* Reevaluate with jiffies_lock held */ 69 write_seqlock(&jiffies_lock); 70 71 delta = ktime_sub(now, last_jiffies_update); 72 if (delta >= tick_period) { 73 74 delta = ktime_sub(delta, tick_period); 75 last_jiffies_update = ktime_add(last_jiffies_update, 76 tick_period); 77 78 /* Slow path for long timeouts */ 79 if (unlikely(delta >= tick_period)) { 80 s64 incr = ktime_to_ns(tick_period); 81 82 ticks = ktime_divns(delta, incr); 83 84 last_jiffies_update = ktime_add_ns(last_jiffies_update, 85 incr * ticks); 86 } 87 do_timer(++ticks); 88 89 /* Keep the tick_next_period variable up to date */ 90 tick_next_period = ktime_add(last_jiffies_update, tick_period); 91 } else { 92 write_sequnlock(&jiffies_lock); 93 return; 94 } 95 write_sequnlock(&jiffies_lock); 96 update_wall_time(); 97 } 98 99 /* 100 * Initialize and return retrieve the jiffies update. 101 */ 102 static ktime_t tick_init_jiffy_update(void) 103 { 104 ktime_t period; 105 106 write_seqlock(&jiffies_lock); 107 /* Did we start the jiffies update yet ? */ 108 if (last_jiffies_update == 0) 109 last_jiffies_update = tick_next_period; 110 period = last_jiffies_update; 111 write_sequnlock(&jiffies_lock); 112 return period; 113 } 114 115 116 static void tick_sched_do_timer(ktime_t now) 117 { 118 int cpu = smp_processor_id(); 119 120 #ifdef CONFIG_NO_HZ_COMMON 121 /* 122 * Check if the do_timer duty was dropped. We don't care about 123 * concurrency: This happens only when the CPU in charge went 124 * into a long sleep. If two CPUs happen to assign themselves to 125 * this duty, then the jiffies update is still serialized by 126 * jiffies_lock. 127 */ 128 if (unlikely(tick_do_timer_cpu == TICK_DO_TIMER_NONE) 129 && !tick_nohz_full_cpu(cpu)) 130 tick_do_timer_cpu = cpu; 131 #endif 132 133 /* Check, if the jiffies need an update */ 134 if (tick_do_timer_cpu == cpu) 135 tick_do_update_jiffies64(now); 136 } 137 138 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) 139 { 140 #ifdef CONFIG_NO_HZ_COMMON 141 /* 142 * When we are idle and the tick is stopped, we have to touch 143 * the watchdog as we might not schedule for a really long 144 * time. This happens on complete idle SMP systems while 145 * waiting on the login prompt. We also increment the "start of 146 * idle" jiffy stamp so the idle accounting adjustment we do 147 * when we go busy again does not account too much ticks. 148 */ 149 if (ts->tick_stopped) { 150 touch_softlockup_watchdog_sched(); 151 if (is_idle_task(current)) 152 ts->idle_jiffies++; 153 /* 154 * In case the current tick fired too early past its expected 155 * expiration, make sure we don't bypass the next clock reprogramming 156 * to the same deadline. 157 */ 158 ts->next_tick = 0; 159 } 160 #endif 161 update_process_times(user_mode(regs)); 162 profile_tick(CPU_PROFILING); 163 } 164 #endif 165 166 #ifdef CONFIG_NO_HZ_FULL 167 cpumask_var_t tick_nohz_full_mask; 168 cpumask_var_t housekeeping_mask; 169 bool tick_nohz_full_running; 170 static atomic_t tick_dep_mask; 171 172 static bool check_tick_dependency(atomic_t *dep) 173 { 174 int val = atomic_read(dep); 175 176 if (val & TICK_DEP_MASK_POSIX_TIMER) { 177 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER); 178 return true; 179 } 180 181 if (val & TICK_DEP_MASK_PERF_EVENTS) { 182 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS); 183 return true; 184 } 185 186 if (val & TICK_DEP_MASK_SCHED) { 187 trace_tick_stop(0, TICK_DEP_MASK_SCHED); 188 return true; 189 } 190 191 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) { 192 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE); 193 return true; 194 } 195 196 return false; 197 } 198 199 static bool can_stop_full_tick(int cpu, struct tick_sched *ts) 200 { 201 WARN_ON_ONCE(!irqs_disabled()); 202 203 if (unlikely(!cpu_online(cpu))) 204 return false; 205 206 if (check_tick_dependency(&tick_dep_mask)) 207 return false; 208 209 if (check_tick_dependency(&ts->tick_dep_mask)) 210 return false; 211 212 if (check_tick_dependency(¤t->tick_dep_mask)) 213 return false; 214 215 if (check_tick_dependency(¤t->signal->tick_dep_mask)) 216 return false; 217 218 return true; 219 } 220 221 static void nohz_full_kick_func(struct irq_work *work) 222 { 223 /* Empty, the tick restart happens on tick_nohz_irq_exit() */ 224 } 225 226 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = { 227 .func = nohz_full_kick_func, 228 }; 229 230 /* 231 * Kick this CPU if it's full dynticks in order to force it to 232 * re-evaluate its dependency on the tick and restart it if necessary. 233 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), 234 * is NMI safe. 235 */ 236 static void tick_nohz_full_kick(void) 237 { 238 if (!tick_nohz_full_cpu(smp_processor_id())) 239 return; 240 241 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work)); 242 } 243 244 /* 245 * Kick the CPU if it's full dynticks in order to force it to 246 * re-evaluate its dependency on the tick and restart it if necessary. 247 */ 248 void tick_nohz_full_kick_cpu(int cpu) 249 { 250 if (!tick_nohz_full_cpu(cpu)) 251 return; 252 253 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); 254 } 255 256 /* 257 * Kick all full dynticks CPUs in order to force these to re-evaluate 258 * their dependency on the tick and restart it if necessary. 259 */ 260 static void tick_nohz_full_kick_all(void) 261 { 262 int cpu; 263 264 if (!tick_nohz_full_running) 265 return; 266 267 preempt_disable(); 268 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask) 269 tick_nohz_full_kick_cpu(cpu); 270 preempt_enable(); 271 } 272 273 static void tick_nohz_dep_set_all(atomic_t *dep, 274 enum tick_dep_bits bit) 275 { 276 int prev; 277 278 prev = atomic_fetch_or(BIT(bit), dep); 279 if (!prev) 280 tick_nohz_full_kick_all(); 281 } 282 283 /* 284 * Set a global tick dependency. Used by perf events that rely on freq and 285 * by unstable clock. 286 */ 287 void tick_nohz_dep_set(enum tick_dep_bits bit) 288 { 289 tick_nohz_dep_set_all(&tick_dep_mask, bit); 290 } 291 292 void tick_nohz_dep_clear(enum tick_dep_bits bit) 293 { 294 atomic_andnot(BIT(bit), &tick_dep_mask); 295 } 296 297 /* 298 * Set per-CPU tick dependency. Used by scheduler and perf events in order to 299 * manage events throttling. 300 */ 301 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit) 302 { 303 int prev; 304 struct tick_sched *ts; 305 306 ts = per_cpu_ptr(&tick_cpu_sched, cpu); 307 308 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask); 309 if (!prev) { 310 preempt_disable(); 311 /* Perf needs local kick that is NMI safe */ 312 if (cpu == smp_processor_id()) { 313 tick_nohz_full_kick(); 314 } else { 315 /* Remote irq work not NMI-safe */ 316 if (!WARN_ON_ONCE(in_nmi())) 317 tick_nohz_full_kick_cpu(cpu); 318 } 319 preempt_enable(); 320 } 321 } 322 323 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit) 324 { 325 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); 326 327 atomic_andnot(BIT(bit), &ts->tick_dep_mask); 328 } 329 330 /* 331 * Set a per-task tick dependency. Posix CPU timers need this in order to elapse 332 * per task timers. 333 */ 334 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit) 335 { 336 /* 337 * We could optimize this with just kicking the target running the task 338 * if that noise matters for nohz full users. 339 */ 340 tick_nohz_dep_set_all(&tsk->tick_dep_mask, bit); 341 } 342 343 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit) 344 { 345 atomic_andnot(BIT(bit), &tsk->tick_dep_mask); 346 } 347 348 /* 349 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse 350 * per process timers. 351 */ 352 void tick_nohz_dep_set_signal(struct signal_struct *sig, enum tick_dep_bits bit) 353 { 354 tick_nohz_dep_set_all(&sig->tick_dep_mask, bit); 355 } 356 357 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit) 358 { 359 atomic_andnot(BIT(bit), &sig->tick_dep_mask); 360 } 361 362 /* 363 * Re-evaluate the need for the tick as we switch the current task. 364 * It might need the tick due to per task/process properties: 365 * perf events, posix CPU timers, ... 366 */ 367 void __tick_nohz_task_switch(void) 368 { 369 unsigned long flags; 370 struct tick_sched *ts; 371 372 local_irq_save(flags); 373 374 if (!tick_nohz_full_cpu(smp_processor_id())) 375 goto out; 376 377 ts = this_cpu_ptr(&tick_cpu_sched); 378 379 if (ts->tick_stopped) { 380 if (atomic_read(¤t->tick_dep_mask) || 381 atomic_read(¤t->signal->tick_dep_mask)) 382 tick_nohz_full_kick(); 383 } 384 out: 385 local_irq_restore(flags); 386 } 387 388 /* Parse the boot-time nohz CPU list from the kernel parameters. */ 389 static int __init tick_nohz_full_setup(char *str) 390 { 391 alloc_bootmem_cpumask_var(&tick_nohz_full_mask); 392 if (cpulist_parse(str, tick_nohz_full_mask) < 0) { 393 pr_warn("NO_HZ: Incorrect nohz_full cpumask\n"); 394 free_bootmem_cpumask_var(tick_nohz_full_mask); 395 return 1; 396 } 397 tick_nohz_full_running = true; 398 399 return 1; 400 } 401 __setup("nohz_full=", tick_nohz_full_setup); 402 403 static int tick_nohz_cpu_down(unsigned int cpu) 404 { 405 /* 406 * The boot CPU handles housekeeping duty (unbound timers, 407 * workqueues, timekeeping, ...) on behalf of full dynticks 408 * CPUs. It must remain online when nohz full is enabled. 409 */ 410 if (tick_nohz_full_running && tick_do_timer_cpu == cpu) 411 return -EBUSY; 412 return 0; 413 } 414 415 static int tick_nohz_init_all(void) 416 { 417 int err = -1; 418 419 #ifdef CONFIG_NO_HZ_FULL_ALL 420 if (!alloc_cpumask_var(&tick_nohz_full_mask, GFP_KERNEL)) { 421 WARN(1, "NO_HZ: Can't allocate full dynticks cpumask\n"); 422 return err; 423 } 424 err = 0; 425 cpumask_setall(tick_nohz_full_mask); 426 tick_nohz_full_running = true; 427 #endif 428 return err; 429 } 430 431 void __init tick_nohz_init(void) 432 { 433 int cpu, ret; 434 435 if (!tick_nohz_full_running) { 436 if (tick_nohz_init_all() < 0) 437 return; 438 } 439 440 if (!alloc_cpumask_var(&housekeeping_mask, GFP_KERNEL)) { 441 WARN(1, "NO_HZ: Can't allocate not-full dynticks cpumask\n"); 442 cpumask_clear(tick_nohz_full_mask); 443 tick_nohz_full_running = false; 444 return; 445 } 446 447 /* 448 * Full dynticks uses irq work to drive the tick rescheduling on safe 449 * locking contexts. But then we need irq work to raise its own 450 * interrupts to avoid circular dependency on the tick 451 */ 452 if (!arch_irq_work_has_interrupt()) { 453 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n"); 454 cpumask_clear(tick_nohz_full_mask); 455 cpumask_copy(housekeeping_mask, cpu_possible_mask); 456 tick_nohz_full_running = false; 457 return; 458 } 459 460 cpu = smp_processor_id(); 461 462 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 463 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", 464 cpu); 465 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 466 } 467 468 cpumask_andnot(housekeeping_mask, 469 cpu_possible_mask, tick_nohz_full_mask); 470 471 for_each_cpu(cpu, tick_nohz_full_mask) 472 context_tracking_cpu_set(cpu); 473 474 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 475 "kernel/nohz:predown", NULL, 476 tick_nohz_cpu_down); 477 WARN_ON(ret < 0); 478 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", 479 cpumask_pr_args(tick_nohz_full_mask)); 480 481 /* 482 * We need at least one CPU to handle housekeeping work such 483 * as timekeeping, unbound timers, workqueues, ... 484 */ 485 WARN_ON_ONCE(cpumask_empty(housekeeping_mask)); 486 } 487 #endif 488 489 /* 490 * NOHZ - aka dynamic tick functionality 491 */ 492 #ifdef CONFIG_NO_HZ_COMMON 493 /* 494 * NO HZ enabled ? 495 */ 496 bool tick_nohz_enabled __read_mostly = true; 497 unsigned long tick_nohz_active __read_mostly; 498 /* 499 * Enable / Disable tickless mode 500 */ 501 static int __init setup_tick_nohz(char *str) 502 { 503 return (kstrtobool(str, &tick_nohz_enabled) == 0); 504 } 505 506 __setup("nohz=", setup_tick_nohz); 507 508 int tick_nohz_tick_stopped(void) 509 { 510 return __this_cpu_read(tick_cpu_sched.tick_stopped); 511 } 512 513 /** 514 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 515 * 516 * Called from interrupt entry when the CPU was idle 517 * 518 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 519 * must be updated. Otherwise an interrupt handler could use a stale jiffy 520 * value. We do this unconditionally on any CPU, as we don't know whether the 521 * CPU, which has the update task assigned is in a long sleep. 522 */ 523 static void tick_nohz_update_jiffies(ktime_t now) 524 { 525 unsigned long flags; 526 527 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 528 529 local_irq_save(flags); 530 tick_do_update_jiffies64(now); 531 local_irq_restore(flags); 532 533 touch_softlockup_watchdog_sched(); 534 } 535 536 /* 537 * Updates the per-CPU time idle statistics counters 538 */ 539 static void 540 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 541 { 542 ktime_t delta; 543 544 if (ts->idle_active) { 545 delta = ktime_sub(now, ts->idle_entrytime); 546 if (nr_iowait_cpu(cpu) > 0) 547 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 548 else 549 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 550 ts->idle_entrytime = now; 551 } 552 553 if (last_update_time) 554 *last_update_time = ktime_to_us(now); 555 556 } 557 558 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 559 { 560 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 561 ts->idle_active = 0; 562 563 sched_clock_idle_wakeup_event(0); 564 } 565 566 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 567 { 568 ktime_t now = ktime_get(); 569 570 ts->idle_entrytime = now; 571 ts->idle_active = 1; 572 sched_clock_idle_sleep_event(); 573 return now; 574 } 575 576 /** 577 * get_cpu_idle_time_us - get the total idle time of a CPU 578 * @cpu: CPU number to query 579 * @last_update_time: variable to store update time in. Do not update 580 * counters if NULL. 581 * 582 * Return the cumulative idle time (since boot) for a given 583 * CPU, in microseconds. 584 * 585 * This time is measured via accounting rather than sampling, 586 * and is as accurate as ktime_get() is. 587 * 588 * This function returns -1 if NOHZ is not enabled. 589 */ 590 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 591 { 592 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 593 ktime_t now, idle; 594 595 if (!tick_nohz_active) 596 return -1; 597 598 now = ktime_get(); 599 if (last_update_time) { 600 update_ts_time_stats(cpu, ts, now, last_update_time); 601 idle = ts->idle_sleeptime; 602 } else { 603 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 604 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 605 606 idle = ktime_add(ts->idle_sleeptime, delta); 607 } else { 608 idle = ts->idle_sleeptime; 609 } 610 } 611 612 return ktime_to_us(idle); 613 614 } 615 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 616 617 /** 618 * get_cpu_iowait_time_us - get the total iowait time of a CPU 619 * @cpu: CPU number to query 620 * @last_update_time: variable to store update time in. Do not update 621 * counters if NULL. 622 * 623 * Return the cumulative iowait time (since boot) for a given 624 * CPU, in microseconds. 625 * 626 * This time is measured via accounting rather than sampling, 627 * and is as accurate as ktime_get() is. 628 * 629 * This function returns -1 if NOHZ is not enabled. 630 */ 631 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 632 { 633 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 634 ktime_t now, iowait; 635 636 if (!tick_nohz_active) 637 return -1; 638 639 now = ktime_get(); 640 if (last_update_time) { 641 update_ts_time_stats(cpu, ts, now, last_update_time); 642 iowait = ts->iowait_sleeptime; 643 } else { 644 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 645 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 646 647 iowait = ktime_add(ts->iowait_sleeptime, delta); 648 } else { 649 iowait = ts->iowait_sleeptime; 650 } 651 } 652 653 return ktime_to_us(iowait); 654 } 655 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 656 657 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 658 { 659 hrtimer_cancel(&ts->sched_timer); 660 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 661 662 /* Forward the time to expire in the future */ 663 hrtimer_forward(&ts->sched_timer, now, tick_period); 664 665 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 666 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 667 else 668 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 669 670 /* 671 * Reset to make sure next tick stop doesn't get fooled by past 672 * cached clock deadline. 673 */ 674 ts->next_tick = 0; 675 } 676 677 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 678 ktime_t now, int cpu) 679 { 680 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 681 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; 682 unsigned long seq, basejiff; 683 ktime_t tick; 684 685 /* Read jiffies and the time when jiffies were updated last */ 686 do { 687 seq = read_seqbegin(&jiffies_lock); 688 basemono = last_jiffies_update; 689 basejiff = jiffies; 690 } while (read_seqretry(&jiffies_lock, seq)); 691 ts->last_jiffies = basejiff; 692 693 if (rcu_needs_cpu(basemono, &next_rcu) || 694 arch_needs_cpu() || irq_work_needs_cpu()) { 695 next_tick = basemono + TICK_NSEC; 696 } else { 697 /* 698 * Get the next pending timer. If high resolution 699 * timers are enabled this only takes the timer wheel 700 * timers into account. If high resolution timers are 701 * disabled this also looks at the next expiring 702 * hrtimer. 703 */ 704 next_tmr = get_next_timer_interrupt(basejiff, basemono); 705 ts->next_timer = next_tmr; 706 /* Take the next rcu event into account */ 707 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; 708 } 709 710 /* 711 * If the tick is due in the next period, keep it ticking or 712 * force prod the timer. 713 */ 714 delta = next_tick - basemono; 715 if (delta <= (u64)TICK_NSEC) { 716 tick = 0; 717 718 /* 719 * Tell the timer code that the base is not idle, i.e. undo 720 * the effect of get_next_timer_interrupt(): 721 */ 722 timer_clear_idle(); 723 /* 724 * We've not stopped the tick yet, and there's a timer in the 725 * next period, so no point in stopping it either, bail. 726 */ 727 if (!ts->tick_stopped) 728 goto out; 729 730 /* 731 * If, OTOH, we did stop it, but there's a pending (expired) 732 * timer reprogram the timer hardware to fire now. 733 * 734 * We will not restart the tick proper, just prod the timer 735 * hardware into firing an interrupt to process the pending 736 * timers. Just like tick_irq_exit() will not restart the tick 737 * for 'normal' interrupts. 738 * 739 * Only once we exit the idle loop will we re-enable the tick, 740 * see tick_nohz_idle_exit(). 741 */ 742 if (delta == 0) { 743 tick_nohz_restart(ts, now); 744 goto out; 745 } 746 } 747 748 /* 749 * If this CPU is the one which updates jiffies, then give up 750 * the assignment and let it be taken by the CPU which runs 751 * the tick timer next, which might be this CPU as well. If we 752 * don't drop this here the jiffies might be stale and 753 * do_timer() never invoked. Keep track of the fact that it 754 * was the one which had the do_timer() duty last. If this CPU 755 * is the one which had the do_timer() duty last, we limit the 756 * sleep time to the timekeeping max_deferment value. 757 * Otherwise we can sleep as long as we want. 758 */ 759 delta = timekeeping_max_deferment(); 760 if (cpu == tick_do_timer_cpu) { 761 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 762 ts->do_timer_last = 1; 763 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 764 delta = KTIME_MAX; 765 ts->do_timer_last = 0; 766 } else if (!ts->do_timer_last) { 767 delta = KTIME_MAX; 768 } 769 770 #ifdef CONFIG_NO_HZ_FULL 771 /* Limit the tick delta to the maximum scheduler deferment */ 772 if (!ts->inidle) 773 delta = min(delta, scheduler_tick_max_deferment()); 774 #endif 775 776 /* Calculate the next expiry time */ 777 if (delta < (KTIME_MAX - basemono)) 778 expires = basemono + delta; 779 else 780 expires = KTIME_MAX; 781 782 expires = min_t(u64, expires, next_tick); 783 tick = expires; 784 785 /* Skip reprogram of event if its not changed */ 786 if (ts->tick_stopped && (expires == ts->next_tick)) { 787 /* Sanity check: make sure clockevent is actually programmed */ 788 if (likely(dev->next_event <= ts->next_tick)) 789 goto out; 790 791 WARN_ON_ONCE(1); 792 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n", 793 basemono, ts->next_tick, dev->next_event, 794 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer)); 795 } 796 797 /* 798 * nohz_stop_sched_tick can be called several times before 799 * the nohz_restart_sched_tick is called. This happens when 800 * interrupts arrive which do not cause a reschedule. In the 801 * first call we save the current tick time, so we can restart 802 * the scheduler tick in nohz_restart_sched_tick. 803 */ 804 if (!ts->tick_stopped) { 805 nohz_balance_enter_idle(cpu); 806 calc_load_enter_idle(); 807 cpu_load_update_nohz_start(); 808 809 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 810 ts->tick_stopped = 1; 811 trace_tick_stop(1, TICK_DEP_MASK_NONE); 812 } 813 814 ts->next_tick = tick; 815 816 /* 817 * If the expiration time == KTIME_MAX, then we simply stop 818 * the tick timer. 819 */ 820 if (unlikely(expires == KTIME_MAX)) { 821 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 822 hrtimer_cancel(&ts->sched_timer); 823 goto out; 824 } 825 826 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 827 hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED); 828 else 829 tick_program_event(tick, 1); 830 out: 831 /* 832 * Update the estimated sleep length until the next timer 833 * (not only the tick). 834 */ 835 ts->sleep_length = ktime_sub(dev->next_event, now); 836 return tick; 837 } 838 839 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 840 { 841 /* Update jiffies first */ 842 tick_do_update_jiffies64(now); 843 cpu_load_update_nohz_stop(); 844 845 /* 846 * Clear the timer idle flag, so we avoid IPIs on remote queueing and 847 * the clock forward checks in the enqueue path: 848 */ 849 timer_clear_idle(); 850 851 calc_load_exit_idle(); 852 touch_softlockup_watchdog_sched(); 853 /* 854 * Cancel the scheduled timer and restore the tick 855 */ 856 ts->tick_stopped = 0; 857 ts->idle_exittime = now; 858 859 tick_nohz_restart(ts, now); 860 } 861 862 static void tick_nohz_full_update_tick(struct tick_sched *ts) 863 { 864 #ifdef CONFIG_NO_HZ_FULL 865 int cpu = smp_processor_id(); 866 867 if (!tick_nohz_full_cpu(cpu)) 868 return; 869 870 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 871 return; 872 873 if (can_stop_full_tick(cpu, ts)) 874 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 875 else if (ts->tick_stopped) 876 tick_nohz_restart_sched_tick(ts, ktime_get()); 877 #endif 878 } 879 880 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 881 { 882 /* 883 * If this CPU is offline and it is the one which updates 884 * jiffies, then give up the assignment and let it be taken by 885 * the CPU which runs the tick timer next. If we don't drop 886 * this here the jiffies might be stale and do_timer() never 887 * invoked. 888 */ 889 if (unlikely(!cpu_online(cpu))) { 890 if (cpu == tick_do_timer_cpu) 891 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 892 /* 893 * Make sure the CPU doesn't get fooled by obsolete tick 894 * deadline if it comes back online later. 895 */ 896 ts->next_tick = 0; 897 return false; 898 } 899 900 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 901 ts->sleep_length = NSEC_PER_SEC / HZ; 902 return false; 903 } 904 905 if (need_resched()) 906 return false; 907 908 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 909 static int ratelimit; 910 911 if (ratelimit < 10 && 912 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 913 pr_warn("NOHZ: local_softirq_pending %02x\n", 914 (unsigned int) local_softirq_pending()); 915 ratelimit++; 916 } 917 return false; 918 } 919 920 if (tick_nohz_full_enabled()) { 921 /* 922 * Keep the tick alive to guarantee timekeeping progression 923 * if there are full dynticks CPUs around 924 */ 925 if (tick_do_timer_cpu == cpu) 926 return false; 927 /* 928 * Boot safety: make sure the timekeeping duty has been 929 * assigned before entering dyntick-idle mode, 930 */ 931 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 932 return false; 933 } 934 935 return true; 936 } 937 938 static void __tick_nohz_idle_enter(struct tick_sched *ts) 939 { 940 ktime_t now, expires; 941 int cpu = smp_processor_id(); 942 943 now = tick_nohz_start_idle(ts); 944 945 if (can_stop_idle_tick(cpu, ts)) { 946 int was_stopped = ts->tick_stopped; 947 948 ts->idle_calls++; 949 950 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 951 if (expires > 0LL) { 952 ts->idle_sleeps++; 953 ts->idle_expires = expires; 954 } 955 956 if (!was_stopped && ts->tick_stopped) 957 ts->idle_jiffies = ts->last_jiffies; 958 } 959 } 960 961 /** 962 * tick_nohz_idle_enter - stop the idle tick from the idle task 963 * 964 * When the next event is more than a tick into the future, stop the idle tick 965 * Called when we start the idle loop. 966 * 967 * The arch is responsible of calling: 968 * 969 * - rcu_idle_enter() after its last use of RCU before the CPU is put 970 * to sleep. 971 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 972 */ 973 void tick_nohz_idle_enter(void) 974 { 975 struct tick_sched *ts; 976 977 WARN_ON_ONCE(irqs_disabled()); 978 979 /* 980 * Update the idle state in the scheduler domain hierarchy 981 * when tick_nohz_stop_sched_tick() is called from the idle loop. 982 * State will be updated to busy during the first busy tick after 983 * exiting idle. 984 */ 985 set_cpu_sd_state_idle(); 986 987 local_irq_disable(); 988 989 ts = this_cpu_ptr(&tick_cpu_sched); 990 ts->inidle = 1; 991 __tick_nohz_idle_enter(ts); 992 993 local_irq_enable(); 994 } 995 996 /** 997 * tick_nohz_irq_exit - update next tick event from interrupt exit 998 * 999 * When an interrupt fires while we are idle and it doesn't cause 1000 * a reschedule, it may still add, modify or delete a timer, enqueue 1001 * an RCU callback, etc... 1002 * So we need to re-calculate and reprogram the next tick event. 1003 */ 1004 void tick_nohz_irq_exit(void) 1005 { 1006 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1007 1008 if (ts->inidle) 1009 __tick_nohz_idle_enter(ts); 1010 else 1011 tick_nohz_full_update_tick(ts); 1012 } 1013 1014 /** 1015 * tick_nohz_get_sleep_length - return the length of the current sleep 1016 * 1017 * Called from power state control code with interrupts disabled 1018 */ 1019 ktime_t tick_nohz_get_sleep_length(void) 1020 { 1021 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1022 1023 return ts->sleep_length; 1024 } 1025 1026 /** 1027 * tick_nohz_get_idle_calls - return the current idle calls counter value 1028 * 1029 * Called from the schedutil frequency scaling governor in scheduler context. 1030 */ 1031 unsigned long tick_nohz_get_idle_calls(void) 1032 { 1033 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1034 1035 return ts->idle_calls; 1036 } 1037 1038 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 1039 { 1040 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 1041 unsigned long ticks; 1042 1043 if (vtime_accounting_cpu_enabled()) 1044 return; 1045 /* 1046 * We stopped the tick in idle. Update process times would miss the 1047 * time we slept as update_process_times does only a 1 tick 1048 * accounting. Enforce that this is accounted to idle ! 1049 */ 1050 ticks = jiffies - ts->idle_jiffies; 1051 /* 1052 * We might be one off. Do not randomly account a huge number of ticks! 1053 */ 1054 if (ticks && ticks < LONG_MAX) 1055 account_idle_ticks(ticks); 1056 #endif 1057 } 1058 1059 /** 1060 * tick_nohz_idle_exit - restart the idle tick from the idle task 1061 * 1062 * Restart the idle tick when the CPU is woken up from idle 1063 * This also exit the RCU extended quiescent state. The CPU 1064 * can use RCU again after this function is called. 1065 */ 1066 void tick_nohz_idle_exit(void) 1067 { 1068 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1069 ktime_t now; 1070 1071 local_irq_disable(); 1072 1073 WARN_ON_ONCE(!ts->inidle); 1074 1075 ts->inidle = 0; 1076 1077 if (ts->idle_active || ts->tick_stopped) 1078 now = ktime_get(); 1079 1080 if (ts->idle_active) 1081 tick_nohz_stop_idle(ts, now); 1082 1083 if (ts->tick_stopped) { 1084 tick_nohz_restart_sched_tick(ts, now); 1085 tick_nohz_account_idle_ticks(ts); 1086 } 1087 1088 local_irq_enable(); 1089 } 1090 1091 /* 1092 * The nohz low res interrupt handler 1093 */ 1094 static void tick_nohz_handler(struct clock_event_device *dev) 1095 { 1096 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1097 struct pt_regs *regs = get_irq_regs(); 1098 ktime_t now = ktime_get(); 1099 1100 dev->next_event = KTIME_MAX; 1101 1102 tick_sched_do_timer(now); 1103 tick_sched_handle(ts, regs); 1104 1105 /* No need to reprogram if we are running tickless */ 1106 if (unlikely(ts->tick_stopped)) 1107 return; 1108 1109 hrtimer_forward(&ts->sched_timer, now, tick_period); 1110 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1111 } 1112 1113 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) 1114 { 1115 if (!tick_nohz_enabled) 1116 return; 1117 ts->nohz_mode = mode; 1118 /* One update is enough */ 1119 if (!test_and_set_bit(0, &tick_nohz_active)) 1120 timers_update_migration(true); 1121 } 1122 1123 /** 1124 * tick_nohz_switch_to_nohz - switch to nohz mode 1125 */ 1126 static void tick_nohz_switch_to_nohz(void) 1127 { 1128 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1129 ktime_t next; 1130 1131 if (!tick_nohz_enabled) 1132 return; 1133 1134 if (tick_switch_to_oneshot(tick_nohz_handler)) 1135 return; 1136 1137 /* 1138 * Recycle the hrtimer in ts, so we can share the 1139 * hrtimer_forward with the highres code. 1140 */ 1141 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1142 /* Get the next period */ 1143 next = tick_init_jiffy_update(); 1144 1145 hrtimer_set_expires(&ts->sched_timer, next); 1146 hrtimer_forward_now(&ts->sched_timer, tick_period); 1147 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1148 tick_nohz_activate(ts, NOHZ_MODE_LOWRES); 1149 } 1150 1151 static inline void tick_nohz_irq_enter(void) 1152 { 1153 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1154 ktime_t now; 1155 1156 if (!ts->idle_active && !ts->tick_stopped) 1157 return; 1158 now = ktime_get(); 1159 if (ts->idle_active) 1160 tick_nohz_stop_idle(ts, now); 1161 if (ts->tick_stopped) 1162 tick_nohz_update_jiffies(now); 1163 } 1164 1165 #else 1166 1167 static inline void tick_nohz_switch_to_nohz(void) { } 1168 static inline void tick_nohz_irq_enter(void) { } 1169 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } 1170 1171 #endif /* CONFIG_NO_HZ_COMMON */ 1172 1173 /* 1174 * Called from irq_enter to notify about the possible interruption of idle() 1175 */ 1176 void tick_irq_enter(void) 1177 { 1178 tick_check_oneshot_broadcast_this_cpu(); 1179 tick_nohz_irq_enter(); 1180 } 1181 1182 /* 1183 * High resolution timer specific code 1184 */ 1185 #ifdef CONFIG_HIGH_RES_TIMERS 1186 /* 1187 * We rearm the timer until we get disabled by the idle code. 1188 * Called with interrupts disabled. 1189 */ 1190 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1191 { 1192 struct tick_sched *ts = 1193 container_of(timer, struct tick_sched, sched_timer); 1194 struct pt_regs *regs = get_irq_regs(); 1195 ktime_t now = ktime_get(); 1196 1197 tick_sched_do_timer(now); 1198 1199 /* 1200 * Do not call, when we are not in irq context and have 1201 * no valid regs pointer 1202 */ 1203 if (regs) 1204 tick_sched_handle(ts, regs); 1205 1206 /* No need to reprogram if we are in idle or full dynticks mode */ 1207 if (unlikely(ts->tick_stopped)) 1208 return HRTIMER_NORESTART; 1209 1210 hrtimer_forward(timer, now, tick_period); 1211 1212 return HRTIMER_RESTART; 1213 } 1214 1215 static int sched_skew_tick; 1216 1217 static int __init skew_tick(char *str) 1218 { 1219 get_option(&str, &sched_skew_tick); 1220 1221 return 0; 1222 } 1223 early_param("skew_tick", skew_tick); 1224 1225 /** 1226 * tick_setup_sched_timer - setup the tick emulation timer 1227 */ 1228 void tick_setup_sched_timer(void) 1229 { 1230 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1231 ktime_t now = ktime_get(); 1232 1233 /* 1234 * Emulate tick processing via per-CPU hrtimers: 1235 */ 1236 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1237 ts->sched_timer.function = tick_sched_timer; 1238 1239 /* Get the next period (per-CPU) */ 1240 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1241 1242 /* Offset the tick to avert jiffies_lock contention. */ 1243 if (sched_skew_tick) { 1244 u64 offset = ktime_to_ns(tick_period) >> 1; 1245 do_div(offset, num_possible_cpus()); 1246 offset *= smp_processor_id(); 1247 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1248 } 1249 1250 hrtimer_forward(&ts->sched_timer, now, tick_period); 1251 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 1252 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); 1253 } 1254 #endif /* HIGH_RES_TIMERS */ 1255 1256 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1257 void tick_cancel_sched_timer(int cpu) 1258 { 1259 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1260 1261 # ifdef CONFIG_HIGH_RES_TIMERS 1262 if (ts->sched_timer.base) 1263 hrtimer_cancel(&ts->sched_timer); 1264 # endif 1265 1266 memset(ts, 0, sizeof(*ts)); 1267 } 1268 #endif 1269 1270 /** 1271 * Async notification about clocksource changes 1272 */ 1273 void tick_clock_notify(void) 1274 { 1275 int cpu; 1276 1277 for_each_possible_cpu(cpu) 1278 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1279 } 1280 1281 /* 1282 * Async notification about clock event changes 1283 */ 1284 void tick_oneshot_notify(void) 1285 { 1286 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1287 1288 set_bit(0, &ts->check_clocks); 1289 } 1290 1291 /** 1292 * Check, if a change happened, which makes oneshot possible. 1293 * 1294 * Called cyclic from the hrtimer softirq (driven by the timer 1295 * softirq) allow_nohz signals, that we can switch into low-res nohz 1296 * mode, because high resolution timers are disabled (either compile 1297 * or runtime). Called with interrupts disabled. 1298 */ 1299 int tick_check_oneshot_change(int allow_nohz) 1300 { 1301 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1302 1303 if (!test_and_clear_bit(0, &ts->check_clocks)) 1304 return 0; 1305 1306 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1307 return 0; 1308 1309 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1310 return 0; 1311 1312 if (!allow_nohz) 1313 return 1; 1314 1315 tick_nohz_switch_to_nohz(); 1316 return 0; 1317 } 1318