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 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 /* 434 * Full dynticks uses irq work to drive the tick rescheduling on safe 435 * locking contexts. But then we need irq work to raise its own 436 * interrupts to avoid circular dependency on the tick 437 */ 438 if (!arch_irq_work_has_interrupt()) { 439 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support irq work self-IPIs\n"); 440 cpumask_clear(tick_nohz_full_mask); 441 tick_nohz_full_running = false; 442 return; 443 } 444 445 cpu = smp_processor_id(); 446 447 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 448 pr_warn("NO_HZ: Clearing %d from nohz_full range for timekeeping\n", 449 cpu); 450 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 451 } 452 453 for_each_cpu(cpu, tick_nohz_full_mask) 454 context_tracking_cpu_set(cpu); 455 456 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 457 "kernel/nohz:predown", NULL, 458 tick_nohz_cpu_down); 459 WARN_ON(ret < 0); 460 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", 461 cpumask_pr_args(tick_nohz_full_mask)); 462 } 463 #endif 464 465 /* 466 * NOHZ - aka dynamic tick functionality 467 */ 468 #ifdef CONFIG_NO_HZ_COMMON 469 /* 470 * NO HZ enabled ? 471 */ 472 bool tick_nohz_enabled __read_mostly = true; 473 unsigned long tick_nohz_active __read_mostly; 474 /* 475 * Enable / Disable tickless mode 476 */ 477 static int __init setup_tick_nohz(char *str) 478 { 479 return (kstrtobool(str, &tick_nohz_enabled) == 0); 480 } 481 482 __setup("nohz=", setup_tick_nohz); 483 484 bool tick_nohz_tick_stopped(void) 485 { 486 return __this_cpu_read(tick_cpu_sched.tick_stopped); 487 } 488 489 bool tick_nohz_tick_stopped_cpu(int cpu) 490 { 491 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); 492 493 return ts->tick_stopped; 494 } 495 496 /** 497 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 498 * 499 * Called from interrupt entry when the CPU was idle 500 * 501 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 502 * must be updated. Otherwise an interrupt handler could use a stale jiffy 503 * value. We do this unconditionally on any CPU, as we don't know whether the 504 * CPU, which has the update task assigned is in a long sleep. 505 */ 506 static void tick_nohz_update_jiffies(ktime_t now) 507 { 508 unsigned long flags; 509 510 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 511 512 local_irq_save(flags); 513 tick_do_update_jiffies64(now); 514 local_irq_restore(flags); 515 516 touch_softlockup_watchdog_sched(); 517 } 518 519 /* 520 * Updates the per-CPU time idle statistics counters 521 */ 522 static void 523 update_ts_time_stats(int cpu, struct tick_sched *ts, ktime_t now, u64 *last_update_time) 524 { 525 ktime_t delta; 526 527 if (ts->idle_active) { 528 delta = ktime_sub(now, ts->idle_entrytime); 529 if (nr_iowait_cpu(cpu) > 0) 530 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 531 else 532 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 533 ts->idle_entrytime = now; 534 } 535 536 if (last_update_time) 537 *last_update_time = ktime_to_us(now); 538 539 } 540 541 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 542 { 543 update_ts_time_stats(smp_processor_id(), ts, now, NULL); 544 ts->idle_active = 0; 545 546 sched_clock_idle_wakeup_event(); 547 } 548 549 static ktime_t tick_nohz_start_idle(struct tick_sched *ts) 550 { 551 ktime_t now = ktime_get(); 552 553 ts->idle_entrytime = now; 554 ts->idle_active = 1; 555 sched_clock_idle_sleep_event(); 556 return now; 557 } 558 559 /** 560 * get_cpu_idle_time_us - get the total idle time of a CPU 561 * @cpu: CPU number to query 562 * @last_update_time: variable to store update time in. Do not update 563 * counters if NULL. 564 * 565 * Return the cumulative idle time (since boot) for a given 566 * CPU, in microseconds. 567 * 568 * This time is measured via accounting rather than sampling, 569 * and is as accurate as ktime_get() is. 570 * 571 * This function returns -1 if NOHZ is not enabled. 572 */ 573 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 574 { 575 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 576 ktime_t now, idle; 577 578 if (!tick_nohz_active) 579 return -1; 580 581 now = ktime_get(); 582 if (last_update_time) { 583 update_ts_time_stats(cpu, ts, now, last_update_time); 584 idle = ts->idle_sleeptime; 585 } else { 586 if (ts->idle_active && !nr_iowait_cpu(cpu)) { 587 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 588 589 idle = ktime_add(ts->idle_sleeptime, delta); 590 } else { 591 idle = ts->idle_sleeptime; 592 } 593 } 594 595 return ktime_to_us(idle); 596 597 } 598 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 599 600 /** 601 * get_cpu_iowait_time_us - get the total iowait time of a CPU 602 * @cpu: CPU number to query 603 * @last_update_time: variable to store update time in. Do not update 604 * counters if NULL. 605 * 606 * Return the cumulative iowait time (since boot) for a given 607 * CPU, in microseconds. 608 * 609 * This time is measured via accounting rather than sampling, 610 * and is as accurate as ktime_get() is. 611 * 612 * This function returns -1 if NOHZ is not enabled. 613 */ 614 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 615 { 616 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 617 ktime_t now, iowait; 618 619 if (!tick_nohz_active) 620 return -1; 621 622 now = ktime_get(); 623 if (last_update_time) { 624 update_ts_time_stats(cpu, ts, now, last_update_time); 625 iowait = ts->iowait_sleeptime; 626 } else { 627 if (ts->idle_active && nr_iowait_cpu(cpu) > 0) { 628 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 629 630 iowait = ktime_add(ts->iowait_sleeptime, delta); 631 } else { 632 iowait = ts->iowait_sleeptime; 633 } 634 } 635 636 return ktime_to_us(iowait); 637 } 638 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 639 640 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 641 { 642 hrtimer_cancel(&ts->sched_timer); 643 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 644 645 /* Forward the time to expire in the future */ 646 hrtimer_forward(&ts->sched_timer, now, tick_period); 647 648 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 649 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 650 else 651 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 652 653 /* 654 * Reset to make sure next tick stop doesn't get fooled by past 655 * cached clock deadline. 656 */ 657 ts->next_tick = 0; 658 } 659 660 static inline bool local_timer_softirq_pending(void) 661 { 662 return local_softirq_pending() & TIMER_SOFTIRQ; 663 } 664 665 static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts, 666 ktime_t now, int cpu) 667 { 668 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 669 u64 basemono, next_tick, next_tmr, next_rcu, delta, expires; 670 unsigned long seq, basejiff; 671 ktime_t tick; 672 673 /* Read jiffies and the time when jiffies were updated last */ 674 do { 675 seq = read_seqbegin(&jiffies_lock); 676 basemono = last_jiffies_update; 677 basejiff = jiffies; 678 } while (read_seqretry(&jiffies_lock, seq)); 679 ts->last_jiffies = basejiff; 680 681 /* 682 * Keep the periodic tick, when RCU, architecture or irq_work 683 * requests it. 684 * Aside of that check whether the local timer softirq is 685 * pending. If so its a bad idea to call get_next_timer_interrupt() 686 * because there is an already expired timer, so it will request 687 * immeditate expiry, which rearms the hardware timer with a 688 * minimal delta which brings us back to this place 689 * immediately. Lather, rinse and repeat... 690 */ 691 if (rcu_needs_cpu(basemono, &next_rcu) || arch_needs_cpu() || 692 irq_work_needs_cpu() || local_timer_softirq_pending()) { 693 next_tick = basemono + TICK_NSEC; 694 } else { 695 /* 696 * Get the next pending timer. If high resolution 697 * timers are enabled this only takes the timer wheel 698 * timers into account. If high resolution timers are 699 * disabled this also looks at the next expiring 700 * hrtimer. 701 */ 702 next_tmr = get_next_timer_interrupt(basejiff, basemono); 703 ts->next_timer = next_tmr; 704 /* Take the next rcu event into account */ 705 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr; 706 } 707 708 /* 709 * If the tick is due in the next period, keep it ticking or 710 * force prod the timer. 711 */ 712 delta = next_tick - basemono; 713 if (delta <= (u64)TICK_NSEC) { 714 /* 715 * Tell the timer code that the base is not idle, i.e. undo 716 * the effect of get_next_timer_interrupt(): 717 */ 718 timer_clear_idle(); 719 /* 720 * We've not stopped the tick yet, and there's a timer in the 721 * next period, so no point in stopping it either, bail. 722 */ 723 if (!ts->tick_stopped) { 724 tick = 0; 725 goto out; 726 } 727 } 728 729 /* 730 * If this CPU is the one which updates jiffies, then give up 731 * the assignment and let it be taken by the CPU which runs 732 * the tick timer next, which might be this CPU as well. If we 733 * don't drop this here the jiffies might be stale and 734 * do_timer() never invoked. Keep track of the fact that it 735 * was the one which had the do_timer() duty last. If this CPU 736 * is the one which had the do_timer() duty last, we limit the 737 * sleep time to the timekeeping max_deferment value. 738 * Otherwise we can sleep as long as we want. 739 */ 740 delta = timekeeping_max_deferment(); 741 if (cpu == tick_do_timer_cpu) { 742 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 743 ts->do_timer_last = 1; 744 } else if (tick_do_timer_cpu != TICK_DO_TIMER_NONE) { 745 delta = KTIME_MAX; 746 ts->do_timer_last = 0; 747 } else if (!ts->do_timer_last) { 748 delta = KTIME_MAX; 749 } 750 751 #ifdef CONFIG_NO_HZ_FULL 752 /* Limit the tick delta to the maximum scheduler deferment */ 753 if (!ts->inidle) 754 delta = min(delta, scheduler_tick_max_deferment()); 755 #endif 756 757 /* Calculate the next expiry time */ 758 if (delta < (KTIME_MAX - basemono)) 759 expires = basemono + delta; 760 else 761 expires = KTIME_MAX; 762 763 expires = min_t(u64, expires, next_tick); 764 tick = expires; 765 766 /* Skip reprogram of event if its not changed */ 767 if (ts->tick_stopped && (expires == ts->next_tick)) { 768 /* Sanity check: make sure clockevent is actually programmed */ 769 if (tick == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer)) 770 goto out; 771 772 WARN_ON_ONCE(1); 773 printk_once("basemono: %llu ts->next_tick: %llu dev->next_event: %llu timer->active: %d timer->expires: %llu\n", 774 basemono, ts->next_tick, dev->next_event, 775 hrtimer_active(&ts->sched_timer), hrtimer_get_expires(&ts->sched_timer)); 776 } 777 778 /* 779 * nohz_stop_sched_tick can be called several times before 780 * the nohz_restart_sched_tick is called. This happens when 781 * interrupts arrive which do not cause a reschedule. In the 782 * first call we save the current tick time, so we can restart 783 * the scheduler tick in nohz_restart_sched_tick. 784 */ 785 if (!ts->tick_stopped) { 786 calc_load_nohz_start(); 787 cpu_load_update_nohz_start(); 788 quiet_vmstat(); 789 790 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 791 ts->tick_stopped = 1; 792 trace_tick_stop(1, TICK_DEP_MASK_NONE); 793 } 794 795 ts->next_tick = tick; 796 797 /* 798 * If the expiration time == KTIME_MAX, then we simply stop 799 * the tick timer. 800 */ 801 if (unlikely(expires == KTIME_MAX)) { 802 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 803 hrtimer_cancel(&ts->sched_timer); 804 goto out; 805 } 806 807 hrtimer_set_expires(&ts->sched_timer, tick); 808 809 if (ts->nohz_mode == NOHZ_MODE_HIGHRES) 810 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 811 else 812 tick_program_event(tick, 1); 813 out: 814 /* 815 * Update the estimated sleep length until the next timer 816 * (not only the tick). 817 */ 818 ts->sleep_length = ktime_sub(dev->next_event, now); 819 return tick; 820 } 821 822 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 823 { 824 /* Update jiffies first */ 825 tick_do_update_jiffies64(now); 826 cpu_load_update_nohz_stop(); 827 828 /* 829 * Clear the timer idle flag, so we avoid IPIs on remote queueing and 830 * the clock forward checks in the enqueue path: 831 */ 832 timer_clear_idle(); 833 834 calc_load_nohz_stop(); 835 touch_softlockup_watchdog_sched(); 836 /* 837 * Cancel the scheduled timer and restore the tick 838 */ 839 ts->tick_stopped = 0; 840 ts->idle_exittime = now; 841 842 tick_nohz_restart(ts, now); 843 } 844 845 static void tick_nohz_full_update_tick(struct tick_sched *ts) 846 { 847 #ifdef CONFIG_NO_HZ_FULL 848 int cpu = smp_processor_id(); 849 850 if (!tick_nohz_full_cpu(cpu)) 851 return; 852 853 if (!ts->tick_stopped && ts->nohz_mode == NOHZ_MODE_INACTIVE) 854 return; 855 856 if (can_stop_full_tick(cpu, ts)) 857 tick_nohz_stop_sched_tick(ts, ktime_get(), cpu); 858 else if (ts->tick_stopped) 859 tick_nohz_restart_sched_tick(ts, ktime_get()); 860 #endif 861 } 862 863 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 864 { 865 /* 866 * If this CPU is offline and it is the one which updates 867 * jiffies, then give up the assignment and let it be taken by 868 * the CPU which runs the tick timer next. If we don't drop 869 * this here the jiffies might be stale and do_timer() never 870 * invoked. 871 */ 872 if (unlikely(!cpu_online(cpu))) { 873 if (cpu == tick_do_timer_cpu) 874 tick_do_timer_cpu = TICK_DO_TIMER_NONE; 875 /* 876 * Make sure the CPU doesn't get fooled by obsolete tick 877 * deadline if it comes back online later. 878 */ 879 ts->next_tick = 0; 880 return false; 881 } 882 883 if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) { 884 ts->sleep_length = NSEC_PER_SEC / HZ; 885 return false; 886 } 887 888 if (need_resched()) 889 return false; 890 891 if (unlikely(local_softirq_pending() && cpu_online(cpu))) { 892 static int ratelimit; 893 894 if (ratelimit < 10 && 895 (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) { 896 pr_warn("NOHZ: local_softirq_pending %02x\n", 897 (unsigned int) local_softirq_pending()); 898 ratelimit++; 899 } 900 return false; 901 } 902 903 if (tick_nohz_full_enabled()) { 904 /* 905 * Keep the tick alive to guarantee timekeeping progression 906 * if there are full dynticks CPUs around 907 */ 908 if (tick_do_timer_cpu == cpu) 909 return false; 910 /* 911 * Boot safety: make sure the timekeeping duty has been 912 * assigned before entering dyntick-idle mode, 913 */ 914 if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) 915 return false; 916 } 917 918 return true; 919 } 920 921 static void __tick_nohz_idle_enter(struct tick_sched *ts) 922 { 923 ktime_t now, expires; 924 int cpu = smp_processor_id(); 925 926 now = tick_nohz_start_idle(ts); 927 928 if (can_stop_idle_tick(cpu, ts)) { 929 int was_stopped = ts->tick_stopped; 930 931 ts->idle_calls++; 932 933 expires = tick_nohz_stop_sched_tick(ts, now, cpu); 934 if (expires > 0LL) { 935 ts->idle_sleeps++; 936 ts->idle_expires = expires; 937 } 938 939 if (!was_stopped && ts->tick_stopped) { 940 ts->idle_jiffies = ts->last_jiffies; 941 nohz_balance_enter_idle(cpu); 942 } 943 } 944 } 945 946 /** 947 * tick_nohz_idle_enter - stop the idle tick from the idle task 948 * 949 * When the next event is more than a tick into the future, stop the idle tick 950 * Called when we start the idle loop. 951 * 952 * The arch is responsible of calling: 953 * 954 * - rcu_idle_enter() after its last use of RCU before the CPU is put 955 * to sleep. 956 * - rcu_idle_exit() before the first use of RCU after the CPU is woken up. 957 */ 958 void tick_nohz_idle_enter(void) 959 { 960 struct tick_sched *ts; 961 962 lockdep_assert_irqs_enabled(); 963 /* 964 * Update the idle state in the scheduler domain hierarchy 965 * when tick_nohz_stop_sched_tick() is called from the idle loop. 966 * State will be updated to busy during the first busy tick after 967 * exiting idle. 968 */ 969 set_cpu_sd_state_idle(); 970 971 local_irq_disable(); 972 973 ts = this_cpu_ptr(&tick_cpu_sched); 974 ts->inidle = 1; 975 __tick_nohz_idle_enter(ts); 976 977 local_irq_enable(); 978 } 979 980 /** 981 * tick_nohz_irq_exit - update next tick event from interrupt exit 982 * 983 * When an interrupt fires while we are idle and it doesn't cause 984 * a reschedule, it may still add, modify or delete a timer, enqueue 985 * an RCU callback, etc... 986 * So we need to re-calculate and reprogram the next tick event. 987 */ 988 void tick_nohz_irq_exit(void) 989 { 990 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 991 992 if (ts->inidle) 993 __tick_nohz_idle_enter(ts); 994 else 995 tick_nohz_full_update_tick(ts); 996 } 997 998 /** 999 * tick_nohz_get_sleep_length - return the length of the current sleep 1000 * 1001 * Called from power state control code with interrupts disabled 1002 */ 1003 ktime_t tick_nohz_get_sleep_length(void) 1004 { 1005 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1006 1007 return ts->sleep_length; 1008 } 1009 1010 /** 1011 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value 1012 * for a particular CPU. 1013 * 1014 * Called from the schedutil frequency scaling governor in scheduler context. 1015 */ 1016 unsigned long tick_nohz_get_idle_calls_cpu(int cpu) 1017 { 1018 struct tick_sched *ts = tick_get_tick_sched(cpu); 1019 1020 return ts->idle_calls; 1021 } 1022 1023 /** 1024 * tick_nohz_get_idle_calls - return the current idle calls counter value 1025 * 1026 * Called from the schedutil frequency scaling governor in scheduler context. 1027 */ 1028 unsigned long tick_nohz_get_idle_calls(void) 1029 { 1030 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1031 1032 return ts->idle_calls; 1033 } 1034 1035 static void tick_nohz_account_idle_ticks(struct tick_sched *ts) 1036 { 1037 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE 1038 unsigned long ticks; 1039 1040 if (vtime_accounting_cpu_enabled()) 1041 return; 1042 /* 1043 * We stopped the tick in idle. Update process times would miss the 1044 * time we slept as update_process_times does only a 1 tick 1045 * accounting. Enforce that this is accounted to idle ! 1046 */ 1047 ticks = jiffies - ts->idle_jiffies; 1048 /* 1049 * We might be one off. Do not randomly account a huge number of ticks! 1050 */ 1051 if (ticks && ticks < LONG_MAX) 1052 account_idle_ticks(ticks); 1053 #endif 1054 } 1055 1056 /** 1057 * tick_nohz_idle_exit - restart the idle tick from the idle task 1058 * 1059 * Restart the idle tick when the CPU is woken up from idle 1060 * This also exit the RCU extended quiescent state. The CPU 1061 * can use RCU again after this function is called. 1062 */ 1063 void tick_nohz_idle_exit(void) 1064 { 1065 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1066 ktime_t now; 1067 1068 local_irq_disable(); 1069 1070 WARN_ON_ONCE(!ts->inidle); 1071 1072 ts->inidle = 0; 1073 1074 if (ts->idle_active || ts->tick_stopped) 1075 now = ktime_get(); 1076 1077 if (ts->idle_active) 1078 tick_nohz_stop_idle(ts, now); 1079 1080 if (ts->tick_stopped) { 1081 tick_nohz_restart_sched_tick(ts, now); 1082 tick_nohz_account_idle_ticks(ts); 1083 } 1084 1085 local_irq_enable(); 1086 } 1087 1088 /* 1089 * The nohz low res interrupt handler 1090 */ 1091 static void tick_nohz_handler(struct clock_event_device *dev) 1092 { 1093 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1094 struct pt_regs *regs = get_irq_regs(); 1095 ktime_t now = ktime_get(); 1096 1097 dev->next_event = KTIME_MAX; 1098 1099 tick_sched_do_timer(now); 1100 tick_sched_handle(ts, regs); 1101 1102 /* No need to reprogram if we are running tickless */ 1103 if (unlikely(ts->tick_stopped)) 1104 return; 1105 1106 hrtimer_forward(&ts->sched_timer, now, tick_period); 1107 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1108 } 1109 1110 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) 1111 { 1112 if (!tick_nohz_enabled) 1113 return; 1114 ts->nohz_mode = mode; 1115 /* One update is enough */ 1116 if (!test_and_set_bit(0, &tick_nohz_active)) 1117 timers_update_nohz(); 1118 } 1119 1120 /** 1121 * tick_nohz_switch_to_nohz - switch to nohz mode 1122 */ 1123 static void tick_nohz_switch_to_nohz(void) 1124 { 1125 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1126 ktime_t next; 1127 1128 if (!tick_nohz_enabled) 1129 return; 1130 1131 if (tick_switch_to_oneshot(tick_nohz_handler)) 1132 return; 1133 1134 /* 1135 * Recycle the hrtimer in ts, so we can share the 1136 * hrtimer_forward with the highres code. 1137 */ 1138 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1139 /* Get the next period */ 1140 next = tick_init_jiffy_update(); 1141 1142 hrtimer_set_expires(&ts->sched_timer, next); 1143 hrtimer_forward_now(&ts->sched_timer, tick_period); 1144 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1145 tick_nohz_activate(ts, NOHZ_MODE_LOWRES); 1146 } 1147 1148 static inline void tick_nohz_irq_enter(void) 1149 { 1150 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1151 ktime_t now; 1152 1153 if (!ts->idle_active && !ts->tick_stopped) 1154 return; 1155 now = ktime_get(); 1156 if (ts->idle_active) 1157 tick_nohz_stop_idle(ts, now); 1158 if (ts->tick_stopped) 1159 tick_nohz_update_jiffies(now); 1160 } 1161 1162 #else 1163 1164 static inline void tick_nohz_switch_to_nohz(void) { } 1165 static inline void tick_nohz_irq_enter(void) { } 1166 static inline void tick_nohz_activate(struct tick_sched *ts, int mode) { } 1167 1168 #endif /* CONFIG_NO_HZ_COMMON */ 1169 1170 /* 1171 * Called from irq_enter to notify about the possible interruption of idle() 1172 */ 1173 void tick_irq_enter(void) 1174 { 1175 tick_check_oneshot_broadcast_this_cpu(); 1176 tick_nohz_irq_enter(); 1177 } 1178 1179 /* 1180 * High resolution timer specific code 1181 */ 1182 #ifdef CONFIG_HIGH_RES_TIMERS 1183 /* 1184 * We rearm the timer until we get disabled by the idle code. 1185 * Called with interrupts disabled. 1186 */ 1187 static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer) 1188 { 1189 struct tick_sched *ts = 1190 container_of(timer, struct tick_sched, sched_timer); 1191 struct pt_regs *regs = get_irq_regs(); 1192 ktime_t now = ktime_get(); 1193 1194 tick_sched_do_timer(now); 1195 1196 /* 1197 * Do not call, when we are not in irq context and have 1198 * no valid regs pointer 1199 */ 1200 if (regs) 1201 tick_sched_handle(ts, regs); 1202 else 1203 ts->next_tick = 0; 1204 1205 /* No need to reprogram if we are in idle or full dynticks mode */ 1206 if (unlikely(ts->tick_stopped)) 1207 return HRTIMER_NORESTART; 1208 1209 hrtimer_forward(timer, now, tick_period); 1210 1211 return HRTIMER_RESTART; 1212 } 1213 1214 static int sched_skew_tick; 1215 1216 static int __init skew_tick(char *str) 1217 { 1218 get_option(&str, &sched_skew_tick); 1219 1220 return 0; 1221 } 1222 early_param("skew_tick", skew_tick); 1223 1224 /** 1225 * tick_setup_sched_timer - setup the tick emulation timer 1226 */ 1227 void tick_setup_sched_timer(void) 1228 { 1229 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1230 ktime_t now = ktime_get(); 1231 1232 /* 1233 * Emulate tick processing via per-CPU hrtimers: 1234 */ 1235 hrtimer_init(&ts->sched_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); 1236 ts->sched_timer.function = tick_sched_timer; 1237 1238 /* Get the next period (per-CPU) */ 1239 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1240 1241 /* Offset the tick to avert jiffies_lock contention. */ 1242 if (sched_skew_tick) { 1243 u64 offset = ktime_to_ns(tick_period) >> 1; 1244 do_div(offset, num_possible_cpus()); 1245 offset *= smp_processor_id(); 1246 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1247 } 1248 1249 hrtimer_forward(&ts->sched_timer, now, tick_period); 1250 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED); 1251 tick_nohz_activate(ts, NOHZ_MODE_HIGHRES); 1252 } 1253 #endif /* HIGH_RES_TIMERS */ 1254 1255 #if defined CONFIG_NO_HZ_COMMON || defined CONFIG_HIGH_RES_TIMERS 1256 void tick_cancel_sched_timer(int cpu) 1257 { 1258 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1259 1260 # ifdef CONFIG_HIGH_RES_TIMERS 1261 if (ts->sched_timer.base) 1262 hrtimer_cancel(&ts->sched_timer); 1263 # endif 1264 1265 memset(ts, 0, sizeof(*ts)); 1266 } 1267 #endif 1268 1269 /** 1270 * Async notification about clocksource changes 1271 */ 1272 void tick_clock_notify(void) 1273 { 1274 int cpu; 1275 1276 for_each_possible_cpu(cpu) 1277 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1278 } 1279 1280 /* 1281 * Async notification about clock event changes 1282 */ 1283 void tick_oneshot_notify(void) 1284 { 1285 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1286 1287 set_bit(0, &ts->check_clocks); 1288 } 1289 1290 /** 1291 * Check, if a change happened, which makes oneshot possible. 1292 * 1293 * Called cyclic from the hrtimer softirq (driven by the timer 1294 * softirq) allow_nohz signals, that we can switch into low-res nohz 1295 * mode, because high resolution timers are disabled (either compile 1296 * or runtime). Called with interrupts disabled. 1297 */ 1298 int tick_check_oneshot_change(int allow_nohz) 1299 { 1300 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1301 1302 if (!test_and_clear_bit(0, &ts->check_clocks)) 1303 return 0; 1304 1305 if (ts->nohz_mode != NOHZ_MODE_INACTIVE) 1306 return 0; 1307 1308 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1309 return 0; 1310 1311 if (!allow_nohz) 1312 return 1; 1313 1314 tick_nohz_switch_to_nohz(); 1315 return 0; 1316 } 1317