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