개념 설명 전체 · v6.18.37 / kernel/time/tick-sched.c
1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de> 4 * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar 5 * Copyright(C) 2006-2007 Timesys Corp., Thomas Gleixner 6 * 7 * NOHZ implementation for low and high resolution timers 8 * 9 * Started by: Thomas Gleixner and Ingo Molnar 10 */ 11 #include <linux/compiler.h> 12 #include <linux/cpu.h> 13 #include <linux/err.h> 14 #include <linux/hrtimer.h> 15 #include <linux/interrupt.h> 16 #include <linux/kernel_stat.h> 17 #include <linux/percpu.h> 18 #include <linux/nmi.h> 19 #include <linux/profile.h> 20 #include <linux/sched/signal.h> 21 #include <linux/sched/clock.h> 22 #include <linux/sched/stat.h> 23 #include <linux/sched/nohz.h> 24 #include <linux/sched/loadavg.h> 25 #include <linux/module.h> 26 #include <linux/irq_work.h> 27 #include <linux/posix-timers.h> 28 #include <linux/context_tracking.h> 29 #include <linux/mm.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 /* 48 * The time when the last jiffy update happened. Write access must hold 49 * jiffies_lock and jiffies_seq. tick_nohz_next_event() needs to get a 50 * consistent view of jiffies and last_jiffies_update. 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 = 1; 60 ktime_t delta, nextp; 61 62 /* 63 * 64-bit can do a quick check without holding the jiffies lock and 64 * without looking at the sequence count. The smp_load_acquire() 65 * pairs with the update done later in this function. 66 * 67 * 32-bit cannot do that because the store of 'tick_next_period' 68 * consists of two 32-bit stores, and the first store could be 69 * moved by the CPU to a random point in the future. 70 */ 71 if (IS_ENABLED(CONFIG_64BIT)) { 72 if (ktime_before(now, smp_load_acquire(&tick_next_period))) 73 return; 74 } else { 75 unsigned int seq; 76 77 /* 78 * Avoid contention on 'jiffies_lock' and protect the quick 79 * check with the sequence count. 80 */ 81 do { 82 seq = read_seqcount_begin(&jiffies_seq); 83 nextp = tick_next_period; 84 } while (read_seqcount_retry(&jiffies_seq, seq)); 85 86 if (ktime_before(now, nextp)) 87 return; 88 } 89 90 /* Quick check failed, i.e. update is required. */ 91 raw_spin_lock(&jiffies_lock); 92 /* 93 * Re-evaluate with the lock held. Another CPU might have done the 94 * update already. 95 */ 96 if (ktime_before(now, tick_next_period)) { 97 raw_spin_unlock(&jiffies_lock); 98 return; 99 } 100 101 write_seqcount_begin(&jiffies_seq); 102 103 delta = ktime_sub(now, tick_next_period); 104 if (unlikely(delta >= TICK_NSEC)) { 105 /* Slow path for long idle sleep times */ 106 s64 incr = TICK_NSEC; 107 108 ticks += ktime_divns(delta, incr); 109 110 last_jiffies_update = ktime_add_ns(last_jiffies_update, 111 incr * ticks); 112 } else { 113 last_jiffies_update = ktime_add_ns(last_jiffies_update, 114 TICK_NSEC); 115 } 116 117 /* Advance jiffies to complete the 'jiffies_seq' protected job */ 118 jiffies_64 += ticks; 119 120 /* Keep the tick_next_period variable up to date */ 121 nextp = ktime_add_ns(last_jiffies_update, TICK_NSEC); 122 123 if (IS_ENABLED(CONFIG_64BIT)) { 124 /* 125 * Pairs with smp_load_acquire() in the lockless quick 126 * check above, and ensures that the update to 'jiffies_64' is 127 * not reordered vs. the store to 'tick_next_period', neither 128 * by the compiler nor by the CPU. 129 */ 130 smp_store_release(&tick_next_period, nextp); 131 } else { 132 /* 133 * A plain store is good enough on 32-bit, as the quick check 134 * above is protected by the sequence count. 135 */ 136 tick_next_period = nextp; 137 } 138 139 /* 140 * Release the sequence count. calc_global_load() below is not 141 * protected by it, but 'jiffies_lock' needs to be held to prevent 142 * concurrent invocations. 143 */ 144 write_seqcount_end(&jiffies_seq); 145 146 calc_global_load(); 147 148 raw_spin_unlock(&jiffies_lock); 149 update_wall_time(); 150 } 151 152 /* 153 * Initialize and return retrieve the jiffies update. 154 */ 155 static ktime_t tick_init_jiffy_update(void) 156 { 157 ktime_t period; 158 159 raw_spin_lock(&jiffies_lock); 160 write_seqcount_begin(&jiffies_seq); 161 162 /* Have we started the jiffies update yet ? */ 163 if (last_jiffies_update == 0) { 164 u32 rem; 165 166 /* 167 * Ensure that the tick is aligned to a multiple of 168 * TICK_NSEC. 169 */ 170 div_u64_rem(tick_next_period, TICK_NSEC, &rem); 171 if (rem) 172 tick_next_period += TICK_NSEC - rem; 173 174 last_jiffies_update = tick_next_period; 175 } 176 period = last_jiffies_update; 177 178 write_seqcount_end(&jiffies_seq); 179 raw_spin_unlock(&jiffies_lock); 180 181 return period; 182 } 183 184 static inline int tick_sched_flag_test(struct tick_sched *ts, 185 unsigned long flag) 186 { 187 return !!(ts->flags & flag); 188 } 189 190 static inline void tick_sched_flag_set(struct tick_sched *ts, 191 unsigned long flag) 192 { 193 lockdep_assert_irqs_disabled(); 194 ts->flags |= flag; 195 } 196 197 static inline void tick_sched_flag_clear(struct tick_sched *ts, 198 unsigned long flag) 199 { 200 lockdep_assert_irqs_disabled(); 201 ts->flags &= ~flag; 202 } 203 204 #define MAX_STALLED_JIFFIES 5 205 206 static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now) 207 { 208 int tick_cpu, cpu = smp_processor_id(); 209 210 /* 211 * Check if the do_timer duty was dropped. We don't care about 212 * concurrency: This happens only when the CPU in charge went 213 * into a long sleep. If two CPUs happen to assign themselves to 214 * this duty, then the jiffies update is still serialized by 215 * 'jiffies_lock'. 216 * 217 * If nohz_full is enabled, this should not happen because the 218 * 'tick_do_timer_cpu' CPU never relinquishes. 219 */ 220 tick_cpu = READ_ONCE(tick_do_timer_cpu); 221 222 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && unlikely(tick_cpu == TICK_DO_TIMER_NONE)) { 223 #ifdef CONFIG_NO_HZ_FULL 224 WARN_ON_ONCE(tick_nohz_full_running); 225 #endif 226 WRITE_ONCE(tick_do_timer_cpu, cpu); 227 tick_cpu = cpu; 228 } 229 230 /* Check if jiffies need an update */ 231 if (tick_cpu == cpu) 232 tick_do_update_jiffies64(now); 233 234 /* 235 * If the jiffies update stalled for too long (timekeeper in stop_machine() 236 * or VMEXIT'ed for several msecs), force an update. 237 */ 238 if (ts->last_tick_jiffies != jiffies) { 239 ts->stalled_jiffies = 0; 240 ts->last_tick_jiffies = READ_ONCE(jiffies); 241 } else { 242 if (++ts->stalled_jiffies == MAX_STALLED_JIFFIES) { 243 tick_do_update_jiffies64(now); 244 ts->stalled_jiffies = 0; 245 ts->last_tick_jiffies = READ_ONCE(jiffies); 246 } 247 } 248 249 if (tick_sched_flag_test(ts, TS_FLAG_INIDLE)) 250 ts->got_idle_tick = 1; 251 } 252 253 static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs) 254 { 255 /* 256 * When we are idle and the tick is stopped, we have to touch 257 * the watchdog as we might not schedule for a really long 258 * time. This happens on completely idle SMP systems while 259 * waiting on the login prompt. We also increment the "start of 260 * idle" jiffy stamp so the idle accounting adjustment we do 261 * when we go busy again does not account too many ticks. 262 */ 263 if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && 264 tick_sched_flag_test(ts, TS_FLAG_STOPPED)) { 265 touch_softlockup_watchdog_sched(); 266 if (is_idle_task(current)) 267 ts->idle_jiffies++; 268 /* 269 * In case the current tick fired too early past its expected 270 * expiration, make sure we don't bypass the next clock reprogramming 271 * to the same deadline. 272 */ 273 ts->next_tick = 0; 274 } 275 276 update_process_times(user_mode(regs)); 277 profile_tick(CPU_PROFILING); 278 } 279 280 /* 281 * We rearm the timer until we get disabled by the idle code. 282 * Called with interrupts disabled. 283 */ 284 static enum hrtimer_restart tick_nohz_handler(struct hrtimer *timer) 285 { 286 struct tick_sched *ts = container_of(timer, struct tick_sched, sched_timer); 287 struct pt_regs *regs = get_irq_regs(); 288 ktime_t now = ktime_get(); 289 290 tick_sched_do_timer(ts, now); 291 292 /* 293 * Do not call when we are not in IRQ context and have 294 * no valid 'regs' pointer 295 */ 296 if (regs) 297 tick_sched_handle(ts, regs); 298 else 299 ts->next_tick = 0; 300 301 /* 302 * In dynticks mode, tick reprogram is deferred: 303 * - to the idle task if in dynticks-idle 304 * - to IRQ exit if in full-dynticks. 305 */ 306 if (unlikely(tick_sched_flag_test(ts, TS_FLAG_STOPPED))) 307 return HRTIMER_NORESTART; 308 309 hrtimer_forward(timer, now, TICK_NSEC); 310 311 return HRTIMER_RESTART; 312 } 313 314 #ifdef CONFIG_NO_HZ_FULL 315 cpumask_var_t tick_nohz_full_mask; 316 EXPORT_SYMBOL_GPL(tick_nohz_full_mask); 317 bool tick_nohz_full_running; 318 EXPORT_SYMBOL_GPL(tick_nohz_full_running); 319 static atomic_t tick_dep_mask; 320 321 static bool check_tick_dependency(atomic_t *dep) 322 { 323 int val = atomic_read(dep); 324 325 if (val & TICK_DEP_MASK_POSIX_TIMER) { 326 trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER); 327 return true; 328 } 329 330 if (val & TICK_DEP_MASK_PERF_EVENTS) { 331 trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS); 332 return true; 333 } 334 335 if (val & TICK_DEP_MASK_SCHED) { 336 trace_tick_stop(0, TICK_DEP_MASK_SCHED); 337 return true; 338 } 339 340 if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) { 341 trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE); 342 return true; 343 } 344 345 if (val & TICK_DEP_MASK_RCU) { 346 trace_tick_stop(0, TICK_DEP_MASK_RCU); 347 return true; 348 } 349 350 if (val & TICK_DEP_MASK_RCU_EXP) { 351 trace_tick_stop(0, TICK_DEP_MASK_RCU_EXP); 352 return true; 353 } 354 355 return false; 356 } 357 358 static bool can_stop_full_tick(int cpu, struct tick_sched *ts) 359 { 360 lockdep_assert_irqs_disabled(); 361 362 if (unlikely(!cpu_online(cpu))) 363 return false; 364 365 if (check_tick_dependency(&tick_dep_mask)) 366 return false; 367 368 if (check_tick_dependency(&ts->tick_dep_mask)) 369 return false; 370 371 if (check_tick_dependency(¤t->tick_dep_mask)) 372 return false; 373 374 if (check_tick_dependency(¤t->signal->tick_dep_mask)) 375 return false; 376 377 return true; 378 } 379 380 static void nohz_full_kick_func(struct irq_work *work) 381 { 382 /* Empty, the tick restart happens on tick_nohz_irq_exit() */ 383 } 384 385 static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) = 386 IRQ_WORK_INIT_HARD(nohz_full_kick_func); 387 388 /* 389 * Kick this CPU if it's full dynticks in order to force it to 390 * re-evaluate its dependency on the tick and restart it if necessary. 391 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(), 392 * is NMI safe. 393 */ 394 static void tick_nohz_full_kick(void) 395 { 396 if (!tick_nohz_full_cpu(smp_processor_id())) 397 return; 398 399 irq_work_queue(this_cpu_ptr(&nohz_full_kick_work)); 400 } 401 402 /* 403 * Kick the CPU if it's full dynticks in order to force it to 404 * re-evaluate its dependency on the tick and restart it if necessary. 405 */ 406 void tick_nohz_full_kick_cpu(int cpu) 407 { 408 if (!tick_nohz_full_cpu(cpu)) 409 return; 410 411 irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu); 412 } 413 414 static void tick_nohz_kick_task(struct task_struct *tsk) 415 { 416 int cpu; 417 418 /* 419 * If the task is not running, run_posix_cpu_timers() 420 * has nothing to elapse, and an IPI can then be optimized out. 421 * 422 * activate_task() STORE p->tick_dep_mask 423 * STORE p->on_rq 424 * __schedule() (switch to task 'p') smp_mb() (atomic_fetch_or()) 425 * LOCK rq->lock LOAD p->on_rq 426 * smp_mb__after_spin_lock() 427 * tick_nohz_task_switch() 428 * LOAD p->tick_dep_mask 429 * 430 * XXX given a task picks up the dependency on schedule(), should we 431 * only care about tasks that are currently on the CPU instead of all 432 * that are on the runqueue? 433 * 434 * That is, does this want to be: task_on_cpu() / task_curr()? 435 */ 436 if (!sched_task_on_rq(tsk)) 437 return; 438 439 /* 440 * If the task concurrently migrates to another CPU, 441 * we guarantee it sees the new tick dependency upon 442 * schedule. 443 * 444 * set_task_cpu(p, cpu); 445 * STORE p->cpu = @cpu 446 * __schedule() (switch to task 'p') 447 * LOCK rq->lock 448 * smp_mb__after_spin_lock() STORE p->tick_dep_mask 449 * tick_nohz_task_switch() smp_mb() (atomic_fetch_or()) 450 * LOAD p->tick_dep_mask LOAD p->cpu 451 */ 452 cpu = task_cpu(tsk); 453 454 preempt_disable(); 455 if (cpu_online(cpu)) 456 tick_nohz_full_kick_cpu(cpu); 457 preempt_enable(); 458 } 459 460 /* 461 * Kick all full dynticks CPUs in order to force these to re-evaluate 462 * their dependency on the tick and restart it if necessary. 463 */ 464 static void tick_nohz_full_kick_all(void) 465 { 466 int cpu; 467 468 if (!tick_nohz_full_running) 469 return; 470 471 preempt_disable(); 472 for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask) 473 tick_nohz_full_kick_cpu(cpu); 474 preempt_enable(); 475 } 476 477 static void tick_nohz_dep_set_all(atomic_t *dep, 478 enum tick_dep_bits bit) 479 { 480 int prev; 481 482 prev = atomic_fetch_or(BIT(bit), dep); 483 if (!prev) 484 tick_nohz_full_kick_all(); 485 } 486 487 /* 488 * Set a global tick dependency. Used by perf events that rely on freq and 489 * unstable clocks. 490 */ 491 void tick_nohz_dep_set(enum tick_dep_bits bit) 492 { 493 tick_nohz_dep_set_all(&tick_dep_mask, bit); 494 } 495 496 void tick_nohz_dep_clear(enum tick_dep_bits bit) 497 { 498 atomic_andnot(BIT(bit), &tick_dep_mask); 499 } 500 501 /* 502 * Set per-CPU tick dependency. Used by scheduler and perf events in order to 503 * manage event-throttling. 504 */ 505 void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit) 506 { 507 int prev; 508 struct tick_sched *ts; 509 510 ts = per_cpu_ptr(&tick_cpu_sched, cpu); 511 512 prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask); 513 if (!prev) { 514 preempt_disable(); 515 /* Perf needs local kick that is NMI safe */ 516 if (cpu == smp_processor_id()) { 517 tick_nohz_full_kick(); 518 } else { 519 /* Remote IRQ work not NMI-safe */ 520 if (!WARN_ON_ONCE(in_nmi())) 521 tick_nohz_full_kick_cpu(cpu); 522 } 523 preempt_enable(); 524 } 525 } 526 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu); 527 528 void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit) 529 { 530 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); 531 532 atomic_andnot(BIT(bit), &ts->tick_dep_mask); 533 } 534 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu); 535 536 /* 537 * Set a per-task tick dependency. RCU needs this. Also posix CPU timers 538 * in order to elapse per task timers. 539 */ 540 void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit) 541 { 542 if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask)) 543 tick_nohz_kick_task(tsk); 544 } 545 EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task); 546 547 void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit) 548 { 549 atomic_andnot(BIT(bit), &tsk->tick_dep_mask); 550 } 551 EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task); 552 553 /* 554 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse 555 * per process timers. 556 */ 557 void tick_nohz_dep_set_signal(struct task_struct *tsk, 558 enum tick_dep_bits bit) 559 { 560 int prev; 561 struct signal_struct *sig = tsk->signal; 562 563 prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask); 564 if (!prev) { 565 struct task_struct *t; 566 567 lockdep_assert_held(&tsk->sighand->siglock); 568 __for_each_thread(sig, t) 569 tick_nohz_kick_task(t); 570 } 571 } 572 573 void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit) 574 { 575 atomic_andnot(BIT(bit), &sig->tick_dep_mask); 576 } 577 578 /* 579 * Re-evaluate the need for the tick as we switch the current task. 580 * It might need the tick due to per task/process properties: 581 * perf events, posix CPU timers, ... 582 */ 583 void __tick_nohz_task_switch(void) 584 { 585 struct tick_sched *ts; 586 587 if (!tick_nohz_full_cpu(smp_processor_id())) 588 return; 589 590 ts = this_cpu_ptr(&tick_cpu_sched); 591 592 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) { 593 if (atomic_read(¤t->tick_dep_mask) || 594 atomic_read(¤t->signal->tick_dep_mask)) 595 tick_nohz_full_kick(); 596 } 597 } 598 599 /* Get the boot-time nohz CPU list from the kernel parameters. */ 600 void __init tick_nohz_full_setup(cpumask_var_t cpumask) 601 { 602 alloc_bootmem_cpumask_var(&tick_nohz_full_mask); 603 cpumask_copy(tick_nohz_full_mask, cpumask); 604 tick_nohz_full_running = true; 605 } 606 607 bool tick_nohz_cpu_hotpluggable(unsigned int cpu) 608 { 609 /* 610 * The 'tick_do_timer_cpu' CPU handles housekeeping duty (unbound 611 * timers, workqueues, timekeeping, ...) on behalf of full dynticks 612 * CPUs. It must remain online when nohz full is enabled. 613 */ 614 if (tick_nohz_full_running && READ_ONCE(tick_do_timer_cpu) == cpu) 615 return false; 616 return true; 617 } 618 619 static int tick_nohz_cpu_down(unsigned int cpu) 620 { 621 return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY; 622 } 623 624 void __init tick_nohz_init(void) 625 { 626 int cpu, ret; 627 628 if (!tick_nohz_full_running) 629 return; 630 631 /* 632 * Full dynticks uses IRQ work to drive the tick rescheduling on safe 633 * locking contexts. But then we need IRQ work to raise its own 634 * interrupts to avoid circular dependency on the tick. 635 */ 636 if (!arch_irq_work_has_interrupt()) { 637 pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support IRQ work self-IPIs\n"); 638 cpumask_clear(tick_nohz_full_mask); 639 tick_nohz_full_running = false; 640 return; 641 } 642 643 if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) && 644 !IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) { 645 cpu = smp_processor_id(); 646 647 if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) { 648 pr_warn("NO_HZ: Clearing %d from nohz_full range " 649 "for timekeeping\n", cpu); 650 cpumask_clear_cpu(cpu, tick_nohz_full_mask); 651 } 652 } 653 654 for_each_cpu(cpu, tick_nohz_full_mask) 655 ct_cpu_track_user(cpu); 656 657 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 658 "kernel/nohz:predown", NULL, 659 tick_nohz_cpu_down); 660 WARN_ON(ret < 0); 661 pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n", 662 cpumask_pr_args(tick_nohz_full_mask)); 663 } 664 #endif /* #ifdef CONFIG_NO_HZ_FULL */ 665 666 /* 667 * NOHZ - aka dynamic tick functionality 668 */ 669 #ifdef CONFIG_NO_HZ_COMMON 670 /* 671 * NO HZ enabled ? 672 */ 673 bool tick_nohz_enabled __read_mostly = true; 674 unsigned long tick_nohz_active __read_mostly; 675 /* 676 * Enable / Disable tickless mode 677 */ 678 static int __init setup_tick_nohz(char *str) 679 { 680 return (kstrtobool(str, &tick_nohz_enabled) == 0); 681 } 682 683 __setup("nohz=", setup_tick_nohz); 684 685 bool tick_nohz_tick_stopped(void) 686 { 687 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 688 689 return tick_sched_flag_test(ts, TS_FLAG_STOPPED); 690 } 691 692 bool tick_nohz_tick_stopped_cpu(int cpu) 693 { 694 struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu); 695 696 return tick_sched_flag_test(ts, TS_FLAG_STOPPED); 697 } 698 699 /** 700 * tick_nohz_update_jiffies - update jiffies when idle was interrupted 701 * @now: current ktime_t 702 * 703 * Called from interrupt entry when the CPU was idle 704 * 705 * In case the sched_tick was stopped on this CPU, we have to check if jiffies 706 * must be updated. Otherwise an interrupt handler could use a stale jiffy 707 * value. We do this unconditionally on any CPU, as we don't know whether the 708 * CPU, which has the update task assigned, is in a long sleep. 709 */ 710 static void tick_nohz_update_jiffies(ktime_t now) 711 { 712 unsigned long flags; 713 714 __this_cpu_write(tick_cpu_sched.idle_waketime, now); 715 716 local_irq_save(flags); 717 tick_do_update_jiffies64(now); 718 local_irq_restore(flags); 719 720 touch_softlockup_watchdog_sched(); 721 } 722 723 static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now) 724 { 725 ktime_t delta; 726 727 if (WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE))) 728 return; 729 730 delta = ktime_sub(now, ts->idle_entrytime); 731 732 write_seqcount_begin(&ts->idle_sleeptime_seq); 733 if (nr_iowait_cpu(smp_processor_id()) > 0) 734 ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta); 735 else 736 ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta); 737 738 ts->idle_entrytime = now; 739 tick_sched_flag_clear(ts, TS_FLAG_IDLE_ACTIVE); 740 write_seqcount_end(&ts->idle_sleeptime_seq); 741 742 sched_clock_idle_wakeup_event(); 743 } 744 745 static void tick_nohz_start_idle(struct tick_sched *ts) 746 { 747 write_seqcount_begin(&ts->idle_sleeptime_seq); 748 ts->idle_entrytime = ktime_get(); 749 tick_sched_flag_set(ts, TS_FLAG_IDLE_ACTIVE); 750 write_seqcount_end(&ts->idle_sleeptime_seq); 751 752 sched_clock_idle_sleep_event(); 753 } 754 755 static u64 get_cpu_sleep_time_us(struct tick_sched *ts, ktime_t *sleeptime, 756 bool compute_delta, u64 *last_update_time) 757 { 758 ktime_t now, idle; 759 unsigned int seq; 760 761 if (!tick_nohz_active) 762 return -1; 763 764 now = ktime_get(); 765 if (last_update_time) 766 *last_update_time = ktime_to_us(now); 767 768 do { 769 seq = read_seqcount_begin(&ts->idle_sleeptime_seq); 770 771 if (tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE) && compute_delta) { 772 ktime_t delta = ktime_sub(now, ts->idle_entrytime); 773 774 idle = ktime_add(*sleeptime, delta); 775 } else { 776 idle = *sleeptime; 777 } 778 } while (read_seqcount_retry(&ts->idle_sleeptime_seq, seq)); 779 780 return ktime_to_us(idle); 781 782 } 783 784 /** 785 * get_cpu_idle_time_us - get the total idle time of a CPU 786 * @cpu: CPU number to query 787 * @last_update_time: variable to store update time in. Do not update 788 * counters if NULL. 789 * 790 * Return the cumulative idle time (since boot) for a given 791 * CPU, in microseconds. Note that this is partially broken due to 792 * the counter of iowait tasks that can be remotely updated without 793 * any synchronization. Therefore it is possible to observe backward 794 * values within two consecutive reads. 795 * 796 * This time is measured via accounting rather than sampling, 797 * and is as accurate as ktime_get() is. 798 * 799 * Return: -1 if NOHZ is not enabled, else total idle time of the @cpu 800 */ 801 u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time) 802 { 803 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 804 805 return get_cpu_sleep_time_us(ts, &ts->idle_sleeptime, 806 !nr_iowait_cpu(cpu), last_update_time); 807 } 808 EXPORT_SYMBOL_GPL(get_cpu_idle_time_us); 809 810 /** 811 * get_cpu_iowait_time_us - get the total iowait time of a CPU 812 * @cpu: CPU number to query 813 * @last_update_time: variable to store update time in. Do not update 814 * counters if NULL. 815 * 816 * Return the cumulative iowait time (since boot) for a given 817 * CPU, in microseconds. Note this is partially broken due to 818 * the counter of iowait tasks that can be remotely updated without 819 * any synchronization. Therefore it is possible to observe backward 820 * values within two consecutive reads. 821 * 822 * This time is measured via accounting rather than sampling, 823 * and is as accurate as ktime_get() is. 824 * 825 * Return: -1 if NOHZ is not enabled, else total iowait time of @cpu 826 */ 827 u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time) 828 { 829 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 830 831 return get_cpu_sleep_time_us(ts, &ts->iowait_sleeptime, 832 nr_iowait_cpu(cpu), last_update_time); 833 } 834 EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us); 835 836 static void tick_nohz_restart(struct tick_sched *ts, ktime_t now) 837 { 838 hrtimer_cancel(&ts->sched_timer); 839 hrtimer_set_expires(&ts->sched_timer, ts->last_tick); 840 841 /* Forward the time to expire in the future */ 842 hrtimer_forward(&ts->sched_timer, now, TICK_NSEC); 843 844 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES)) { 845 hrtimer_start_expires(&ts->sched_timer, 846 HRTIMER_MODE_ABS_PINNED_HARD); 847 } else { 848 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 849 } 850 851 /* 852 * Reset to make sure the next tick stop doesn't get fooled by past 853 * cached clock deadline. 854 */ 855 ts->next_tick = 0; 856 } 857 858 static inline bool local_timer_softirq_pending(void) 859 { 860 return local_timers_pending() & BIT(TIMER_SOFTIRQ); 861 } 862 863 /* 864 * Read jiffies and the time when jiffies were updated last 865 */ 866 u64 get_jiffies_update(unsigned long *basej) 867 { 868 unsigned long basejiff; 869 unsigned int seq; 870 u64 basemono; 871 872 do { 873 seq = read_seqcount_begin(&jiffies_seq); 874 basemono = last_jiffies_update; 875 basejiff = jiffies; 876 } while (read_seqcount_retry(&jiffies_seq, seq)); 877 *basej = basejiff; 878 return basemono; 879 } 880 881 /** 882 * tick_nohz_next_event() - return the clock monotonic based next event 883 * @ts: pointer to tick_sched struct 884 * @cpu: CPU number 885 * 886 * Return: 887 * *%0 - When the next event is a maximum of TICK_NSEC in the future 888 * and the tick is not stopped yet 889 * *%next_event - Next event based on clock monotonic 890 */ 891 static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu) 892 { 893 u64 basemono, next_tick, delta, expires; 894 unsigned long basejiff; 895 int tick_cpu; 896 897 basemono = get_jiffies_update(&basejiff); 898 ts->last_jiffies = basejiff; 899 ts->timer_expires_base = basemono; 900 901 /* 902 * Keep the periodic tick, when RCU, architecture or irq_work 903 * requests it. 904 * Aside of that, check whether the local timer softirq is 905 * pending. If so, its a bad idea to call get_next_timer_interrupt(), 906 * because there is an already expired timer, so it will request 907 * immediate expiry, which rearms the hardware timer with a 908 * minimal delta, which brings us back to this place 909 * immediately. Lather, rinse and repeat... 910 */ 911 if (rcu_needs_cpu() || arch_needs_cpu() || 912 irq_work_needs_cpu() || local_timer_softirq_pending()) { 913 next_tick = basemono + TICK_NSEC; 914 } else { 915 /* 916 * Get the next pending timer. If high resolution 917 * timers are enabled this only takes the timer wheel 918 * timers into account. If high resolution timers are 919 * disabled this also looks at the next expiring 920 * hrtimer. 921 */ 922 next_tick = get_next_timer_interrupt(basejiff, basemono); 923 ts->next_timer = next_tick; 924 } 925 926 /* Make sure next_tick is never before basemono! */ 927 if (WARN_ON_ONCE(basemono > next_tick)) 928 next_tick = basemono; 929 930 /* 931 * If the tick is due in the next period, keep it ticking or 932 * force prod the timer. 933 */ 934 delta = next_tick - basemono; 935 if (delta <= (u64)TICK_NSEC) { 936 /* 937 * We've not stopped the tick yet, and there's a timer in the 938 * next period, so no point in stopping it either, bail. 939 */ 940 if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED)) { 941 ts->timer_expires = 0; 942 goto out; 943 } 944 } 945 946 /* 947 * If this CPU is the one which had the do_timer() duty last, we limit 948 * the sleep time to the timekeeping 'max_deferment' value. 949 * Otherwise we can sleep as long as we want. 950 */ 951 delta = timekeeping_max_deferment(); 952 tick_cpu = READ_ONCE(tick_do_timer_cpu); 953 if (tick_cpu != cpu && 954 (tick_cpu != TICK_DO_TIMER_NONE || !tick_sched_flag_test(ts, TS_FLAG_DO_TIMER_LAST))) 955 delta = KTIME_MAX; 956 957 /* Calculate the next expiry time */ 958 if (delta < (KTIME_MAX - basemono)) 959 expires = basemono + delta; 960 else 961 expires = KTIME_MAX; 962 963 ts->timer_expires = min_t(u64, expires, next_tick); 964 965 out: 966 return ts->timer_expires; 967 } 968 969 static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu) 970 { 971 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 972 unsigned long basejiff = ts->last_jiffies; 973 u64 basemono = ts->timer_expires_base; 974 bool timer_idle = tick_sched_flag_test(ts, TS_FLAG_STOPPED); 975 int tick_cpu; 976 u64 expires; 977 978 /* Make sure we won't be trying to stop it twice in a row. */ 979 ts->timer_expires_base = 0; 980 981 /* 982 * Now the tick should be stopped definitely - so the timer base needs 983 * to be marked idle as well to not miss a newly queued timer. 984 */ 985 expires = timer_base_try_to_set_idle(basejiff, basemono, &timer_idle); 986 if (expires > ts->timer_expires) { 987 /* 988 * This path could only happen when the first timer was removed 989 * between calculating the possible sleep length and now (when 990 * high resolution mode is not active, timer could also be a 991 * hrtimer). 992 * 993 * We have to stick to the original calculated expiry value to 994 * not stop the tick for too long with a shallow C-state (which 995 * was programmed by cpuidle because of an early next expiration 996 * value). 997 */ 998 expires = ts->timer_expires; 999 } 1000 1001 /* If the timer base is not idle, retain the not yet stopped tick. */ 1002 if (!timer_idle) 1003 return; 1004 1005 /* 1006 * If this CPU is the one which updates jiffies, then give up 1007 * the assignment and let it be taken by the CPU which runs 1008 * the tick timer next, which might be this CPU as well. If we 1009 * don't drop this here, the jiffies might be stale and 1010 * do_timer() never gets invoked. Keep track of the fact that it 1011 * was the one which had the do_timer() duty last. 1012 */ 1013 tick_cpu = READ_ONCE(tick_do_timer_cpu); 1014 if (tick_cpu == cpu) { 1015 WRITE_ONCE(tick_do_timer_cpu, TICK_DO_TIMER_NONE); 1016 tick_sched_flag_set(ts, TS_FLAG_DO_TIMER_LAST); 1017 } else if (tick_cpu != TICK_DO_TIMER_NONE) { 1018 tick_sched_flag_clear(ts, TS_FLAG_DO_TIMER_LAST); 1019 } 1020 1021 /* Skip reprogram of event if it's not changed */ 1022 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED) && (expires == ts->next_tick)) { 1023 /* Sanity check: make sure clockevent is actually programmed */ 1024 if (expires == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer)) 1025 return; 1026 1027 WARN_ONCE(1, "basemono: %llu ts->next_tick: %llu dev->next_event: %llu " 1028 "timer->active: %d timer->expires: %llu\n", basemono, ts->next_tick, 1029 dev->next_event, hrtimer_active(&ts->sched_timer), 1030 hrtimer_get_expires(&ts->sched_timer)); 1031 } 1032 1033 /* 1034 * tick_nohz_stop_tick() can be called several times before 1035 * tick_nohz_restart_sched_tick() is called. This happens when 1036 * interrupts arrive which do not cause a reschedule. In the first 1037 * call we save the current tick time, so we can restart the 1038 * scheduler tick in tick_nohz_restart_sched_tick(). 1039 */ 1040 if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED)) { 1041 calc_load_nohz_start(); 1042 quiet_vmstat(); 1043 1044 ts->last_tick = hrtimer_get_expires(&ts->sched_timer); 1045 tick_sched_flag_set(ts, TS_FLAG_STOPPED); 1046 trace_tick_stop(1, TICK_DEP_MASK_NONE); 1047 } 1048 1049 ts->next_tick = expires; 1050 1051 /* 1052 * If the expiration time == KTIME_MAX, then we simply stop 1053 * the tick timer. 1054 */ 1055 if (unlikely(expires == KTIME_MAX)) { 1056 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES)) 1057 hrtimer_cancel(&ts->sched_timer); 1058 else 1059 tick_program_event(KTIME_MAX, 1); 1060 return; 1061 } 1062 1063 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES)) { 1064 hrtimer_start(&ts->sched_timer, expires, 1065 HRTIMER_MODE_ABS_PINNED_HARD); 1066 } else { 1067 hrtimer_set_expires(&ts->sched_timer, expires); 1068 tick_program_event(expires, 1); 1069 } 1070 } 1071 1072 static void tick_nohz_retain_tick(struct tick_sched *ts) 1073 { 1074 ts->timer_expires_base = 0; 1075 } 1076 1077 #ifdef CONFIG_NO_HZ_FULL 1078 static void tick_nohz_full_stop_tick(struct tick_sched *ts, int cpu) 1079 { 1080 if (tick_nohz_next_event(ts, cpu)) 1081 tick_nohz_stop_tick(ts, cpu); 1082 else 1083 tick_nohz_retain_tick(ts); 1084 } 1085 #endif /* CONFIG_NO_HZ_FULL */ 1086 1087 static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now) 1088 { 1089 /* Update jiffies first */ 1090 tick_do_update_jiffies64(now); 1091 1092 /* 1093 * Clear the timer idle flag, so we avoid IPIs on remote queueing and 1094 * the clock forward checks in the enqueue path: 1095 */ 1096 timer_clear_idle(); 1097 1098 calc_load_nohz_stop(); 1099 touch_softlockup_watchdog_sched(); 1100 1101 /* Cancel the scheduled timer and restore the tick: */ 1102 tick_sched_flag_clear(ts, TS_FLAG_STOPPED); 1103 tick_nohz_restart(ts, now); 1104 } 1105 1106 static void __tick_nohz_full_update_tick(struct tick_sched *ts, 1107 ktime_t now) 1108 { 1109 #ifdef CONFIG_NO_HZ_FULL 1110 int cpu = smp_processor_id(); 1111 1112 if (can_stop_full_tick(cpu, ts)) 1113 tick_nohz_full_stop_tick(ts, cpu); 1114 else if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) 1115 tick_nohz_restart_sched_tick(ts, now); 1116 #endif 1117 } 1118 1119 static void tick_nohz_full_update_tick(struct tick_sched *ts) 1120 { 1121 if (!tick_nohz_full_cpu(smp_processor_id())) 1122 return; 1123 1124 if (!tick_sched_flag_test(ts, TS_FLAG_NOHZ)) 1125 return; 1126 1127 __tick_nohz_full_update_tick(ts, ktime_get()); 1128 } 1129 1130 /* 1131 * A pending softirq outside an IRQ (or softirq disabled section) context 1132 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't 1133 * reach this code due to the need_resched() early check in can_stop_idle_tick(). 1134 * 1135 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the 1136 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked, 1137 * triggering the code below, since wakep_softirqd() is ignored. 1138 * 1139 */ 1140 static bool report_idle_softirq(void) 1141 { 1142 static int ratelimit; 1143 unsigned int pending = local_softirq_pending(); 1144 1145 if (likely(!pending)) 1146 return false; 1147 1148 /* Some softirqs claim to be safe against hotplug and ksoftirqd parking */ 1149 if (!cpu_active(smp_processor_id())) { 1150 pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK; 1151 if (!pending) 1152 return false; 1153 } 1154 1155 /* On RT, softirq handling may be waiting on some lock */ 1156 if (local_bh_blocked()) 1157 return false; 1158 1159 if (ratelimit < 10) { 1160 pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n", 1161 pending); 1162 ratelimit++; 1163 } 1164 1165 return true; 1166 } 1167 1168 static bool can_stop_idle_tick(int cpu, struct tick_sched *ts) 1169 { 1170 WARN_ON_ONCE(cpu_is_offline(cpu)); 1171 1172 if (unlikely(!tick_sched_flag_test(ts, TS_FLAG_NOHZ))) 1173 return false; 1174 1175 if (need_resched()) 1176 return false; 1177 1178 if (unlikely(report_idle_softirq())) 1179 return false; 1180 1181 if (tick_nohz_full_enabled()) { 1182 int tick_cpu = READ_ONCE(tick_do_timer_cpu); 1183 1184 /* 1185 * Keep the tick alive to guarantee timekeeping progression 1186 * if there are full dynticks CPUs around 1187 */ 1188 if (tick_cpu == cpu) 1189 return false; 1190 1191 /* Should not happen for nohz-full */ 1192 if (WARN_ON_ONCE(tick_cpu == TICK_DO_TIMER_NONE)) 1193 return false; 1194 } 1195 1196 return true; 1197 } 1198 1199 /** 1200 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task 1201 * 1202 * When the next event is more than a tick into the future, stop the idle tick 1203 */ 1204 void tick_nohz_idle_stop_tick(void) 1205 { 1206 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1207 int cpu = smp_processor_id(); 1208 ktime_t expires; 1209 1210 /* 1211 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the 1212 * tick timer expiration time is known already. 1213 */ 1214 if (ts->timer_expires_base) 1215 expires = ts->timer_expires; 1216 else if (can_stop_idle_tick(cpu, ts)) 1217 expires = tick_nohz_next_event(ts, cpu); 1218 else 1219 return; 1220 1221 ts->idle_calls++; 1222 1223 if (expires > 0LL) { 1224 int was_stopped = tick_sched_flag_test(ts, TS_FLAG_STOPPED); 1225 1226 tick_nohz_stop_tick(ts, cpu); 1227 1228 ts->idle_sleeps++; 1229 ts->idle_expires = expires; 1230 1231 if (!was_stopped && tick_sched_flag_test(ts, TS_FLAG_STOPPED)) { 1232 ts->idle_jiffies = ts->last_jiffies; 1233 nohz_balance_enter_idle(cpu); 1234 } 1235 } else { 1236 tick_nohz_retain_tick(ts); 1237 } 1238 } 1239 1240 void tick_nohz_idle_retain_tick(void) 1241 { 1242 tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched)); 1243 } 1244 1245 /** 1246 * tick_nohz_idle_enter - prepare for entering idle on the current CPU 1247 * 1248 * Called when we start the idle loop. 1249 */ 1250 void tick_nohz_idle_enter(void) 1251 { 1252 struct tick_sched *ts; 1253 1254 lockdep_assert_irqs_enabled(); 1255 1256 local_irq_disable(); 1257 1258 ts = this_cpu_ptr(&tick_cpu_sched); 1259 1260 WARN_ON_ONCE(ts->timer_expires_base); 1261 1262 tick_sched_flag_set(ts, TS_FLAG_INIDLE); 1263 tick_nohz_start_idle(ts); 1264 1265 local_irq_enable(); 1266 } 1267 1268 /** 1269 * tick_nohz_irq_exit - Notify the tick about IRQ exit 1270 * 1271 * A timer may have been added/modified/deleted either by the current IRQ, 1272 * or by another place using this IRQ as a notification. This IRQ may have 1273 * also updated the RCU callback list. These events may require a 1274 * re-evaluation of the next tick. Depending on the context: 1275 * 1276 * 1) If the CPU is idle and no resched is pending, just proceed with idle 1277 * time accounting. The next tick will be re-evaluated on the next idle 1278 * loop iteration. 1279 * 1280 * 2) If the CPU is nohz_full: 1281 * 1282 * 2.1) If there is any tick dependency, restart the tick if stopped. 1283 * 1284 * 2.2) If there is no tick dependency, (re-)evaluate the next tick and 1285 * stop/update it accordingly. 1286 */ 1287 void tick_nohz_irq_exit(void) 1288 { 1289 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1290 1291 if (tick_sched_flag_test(ts, TS_FLAG_INIDLE)) 1292 tick_nohz_start_idle(ts); 1293 else 1294 tick_nohz_full_update_tick(ts); 1295 } 1296 1297 /** 1298 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run 1299 * 1300 * Return: %true if the tick handler has run, otherwise %false 1301 */ 1302 bool tick_nohz_idle_got_tick(void) 1303 { 1304 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1305 1306 if (ts->got_idle_tick) { 1307 ts->got_idle_tick = 0; 1308 return true; 1309 } 1310 return false; 1311 } 1312 1313 /** 1314 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer 1315 * or the tick, whichever expires first. Note that, if the tick has been 1316 * stopped, it returns the next hrtimer. 1317 * 1318 * Called from power state control code with interrupts disabled 1319 * 1320 * Return: the next expiration time 1321 */ 1322 ktime_t tick_nohz_get_next_hrtimer(void) 1323 { 1324 return __this_cpu_read(tick_cpu_device.evtdev)->next_event; 1325 } 1326 1327 /** 1328 * tick_nohz_get_sleep_length - return the expected length of the current sleep 1329 * @delta_next: duration until the next event if the tick cannot be stopped 1330 * 1331 * Called from power state control code with interrupts disabled. 1332 * 1333 * The return value of this function and/or the value returned by it through the 1334 * @delta_next pointer can be negative which must be taken into account by its 1335 * callers. 1336 * 1337 * Return: the expected length of the current sleep 1338 */ 1339 ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next) 1340 { 1341 struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev); 1342 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1343 int cpu = smp_processor_id(); 1344 /* 1345 * The idle entry time is expected to be a sufficient approximation of 1346 * the current time at this point. 1347 */ 1348 ktime_t now = ts->idle_entrytime; 1349 ktime_t next_event; 1350 1351 WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_INIDLE)); 1352 1353 *delta_next = ktime_sub(dev->next_event, now); 1354 1355 if (!can_stop_idle_tick(cpu, ts)) 1356 return *delta_next; 1357 1358 next_event = tick_nohz_next_event(ts, cpu); 1359 if (!next_event) 1360 return *delta_next; 1361 1362 /* 1363 * If the next highres timer to expire is earlier than 'next_event', the 1364 * idle governor needs to know that. 1365 */ 1366 next_event = min_t(u64, next_event, 1367 hrtimer_next_event_without(&ts->sched_timer)); 1368 1369 return ktime_sub(next_event, now); 1370 } 1371 1372 /** 1373 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value 1374 * for a particular CPU. 1375 * @cpu: target CPU number 1376 * 1377 * Called from the schedutil frequency scaling governor in scheduler context. 1378 * 1379 * Return: the current idle calls counter value for @cpu 1380 */ 1381 unsigned long tick_nohz_get_idle_calls_cpu(int cpu) 1382 { 1383 struct tick_sched *ts = tick_get_tick_sched(cpu); 1384 1385 return ts->idle_calls; 1386 } 1387 1388 static void tick_nohz_account_idle_time(struct tick_sched *ts, 1389 ktime_t now) 1390 { 1391 unsigned long ticks; 1392 1393 ts->idle_exittime = now; 1394 1395 if (vtime_accounting_enabled_this_cpu()) 1396 return; 1397 /* 1398 * We stopped the tick in idle. update_process_times() would miss the 1399 * time we slept, as it does only a 1 tick accounting. 1400 * Enforce that this is accounted to idle ! 1401 */ 1402 ticks = jiffies - ts->idle_jiffies; 1403 /* 1404 * We might be one off. Do not randomly account a huge number of ticks! 1405 */ 1406 if (ticks && ticks < LONG_MAX) 1407 account_idle_ticks(ticks); 1408 } 1409 1410 void tick_nohz_idle_restart_tick(void) 1411 { 1412 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1413 1414 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) { 1415 ktime_t now = ktime_get(); 1416 tick_nohz_restart_sched_tick(ts, now); 1417 tick_nohz_account_idle_time(ts, now); 1418 } 1419 } 1420 1421 static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now) 1422 { 1423 if (tick_nohz_full_cpu(smp_processor_id())) 1424 __tick_nohz_full_update_tick(ts, now); 1425 else 1426 tick_nohz_restart_sched_tick(ts, now); 1427 1428 tick_nohz_account_idle_time(ts, now); 1429 } 1430 1431 /** 1432 * tick_nohz_idle_exit - Update the tick upon idle task exit 1433 * 1434 * When the idle task exits, update the tick depending on the 1435 * following situations: 1436 * 1437 * 1) If the CPU is not in nohz_full mode (most cases), then 1438 * restart the tick. 1439 * 1440 * 2) If the CPU is in nohz_full mode (corner case): 1441 * 2.1) If the tick can be kept stopped (no tick dependencies) 1442 * then re-evaluate the next tick and try to keep it stopped 1443 * as long as possible. 1444 * 2.2) If the tick has dependencies, restart the tick. 1445 * 1446 */ 1447 void tick_nohz_idle_exit(void) 1448 { 1449 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1450 bool idle_active, tick_stopped; 1451 ktime_t now; 1452 1453 local_irq_disable(); 1454 1455 WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_INIDLE)); 1456 WARN_ON_ONCE(ts->timer_expires_base); 1457 1458 tick_sched_flag_clear(ts, TS_FLAG_INIDLE); 1459 idle_active = tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE); 1460 tick_stopped = tick_sched_flag_test(ts, TS_FLAG_STOPPED); 1461 1462 if (idle_active || tick_stopped) 1463 now = ktime_get(); 1464 1465 if (idle_active) 1466 tick_nohz_stop_idle(ts, now); 1467 1468 if (tick_stopped) 1469 tick_nohz_idle_update_tick(ts, now); 1470 1471 local_irq_enable(); 1472 } 1473 1474 /* 1475 * In low-resolution mode, the tick handler must be implemented directly 1476 * at the clockevent level. hrtimer can't be used instead, because its 1477 * infrastructure actually relies on the tick itself as a backend in 1478 * low-resolution mode (see hrtimer_run_queues()). 1479 */ 1480 static void tick_nohz_lowres_handler(struct clock_event_device *dev) 1481 { 1482 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1483 1484 dev->next_event = KTIME_MAX; 1485 1486 if (likely(tick_nohz_handler(&ts->sched_timer) == HRTIMER_RESTART)) 1487 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1488 } 1489 1490 static inline void tick_nohz_activate(struct tick_sched *ts) 1491 { 1492 if (!tick_nohz_enabled) 1493 return; 1494 tick_sched_flag_set(ts, TS_FLAG_NOHZ); 1495 /* One update is enough */ 1496 if (!test_and_set_bit(0, &tick_nohz_active)) 1497 timers_update_nohz(); 1498 } 1499 1500 /** 1501 * tick_nohz_switch_to_nohz - switch to NOHZ mode 1502 */ 1503 static void tick_nohz_switch_to_nohz(void) 1504 { 1505 if (!tick_nohz_enabled) 1506 return; 1507 1508 if (tick_switch_to_oneshot(tick_nohz_lowres_handler)) 1509 return; 1510 1511 /* 1512 * Recycle the hrtimer in 'ts', so we can share the 1513 * highres code. 1514 */ 1515 tick_setup_sched_timer(false); 1516 } 1517 1518 static inline void tick_nohz_irq_enter(void) 1519 { 1520 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1521 ktime_t now; 1522 1523 if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED | TS_FLAG_IDLE_ACTIVE)) 1524 return; 1525 now = ktime_get(); 1526 if (tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE)) 1527 tick_nohz_stop_idle(ts, now); 1528 /* 1529 * If all CPUs are idle we may need to update a stale jiffies value. 1530 * Note nohz_full is a special case: a timekeeper is guaranteed to stay 1531 * alive but it might be busy looping with interrupts disabled in some 1532 * rare case (typically stop machine). So we must make sure we have a 1533 * last resort. 1534 */ 1535 if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) 1536 tick_nohz_update_jiffies(now); 1537 } 1538 1539 #else 1540 1541 static inline void tick_nohz_switch_to_nohz(void) { } 1542 static inline void tick_nohz_irq_enter(void) { } 1543 static inline void tick_nohz_activate(struct tick_sched *ts) { } 1544 1545 #endif /* CONFIG_NO_HZ_COMMON */ 1546 1547 /* 1548 * Called from irq_enter() to notify about the possible interruption of idle() 1549 */ 1550 void tick_irq_enter(void) 1551 { 1552 tick_check_oneshot_broadcast_this_cpu(); 1553 tick_nohz_irq_enter(); 1554 } 1555 1556 static int sched_skew_tick; 1557 1558 static int __init skew_tick(char *str) 1559 { 1560 get_option(&str, &sched_skew_tick); 1561 1562 return 0; 1563 } 1564 early_param("skew_tick", skew_tick); 1565 1566 /** 1567 * tick_setup_sched_timer - setup the tick emulation timer 1568 * @hrtimer: whether to use the hrtimer or not 1569 */ 1570 void tick_setup_sched_timer(bool hrtimer) 1571 { 1572 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1573 1574 /* Emulate tick processing via per-CPU hrtimers: */ 1575 hrtimer_setup(&ts->sched_timer, tick_nohz_handler, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD); 1576 1577 if (IS_ENABLED(CONFIG_HIGH_RES_TIMERS) && hrtimer) 1578 tick_sched_flag_set(ts, TS_FLAG_HIGHRES); 1579 1580 /* Get the next period (per-CPU) */ 1581 hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update()); 1582 1583 /* Offset the tick to avert 'jiffies_lock' contention. */ 1584 if (sched_skew_tick) { 1585 u64 offset = TICK_NSEC >> 1; 1586 do_div(offset, num_possible_cpus()); 1587 offset *= smp_processor_id(); 1588 hrtimer_add_expires_ns(&ts->sched_timer, offset); 1589 } 1590 1591 hrtimer_forward_now(&ts->sched_timer, TICK_NSEC); 1592 if (IS_ENABLED(CONFIG_HIGH_RES_TIMERS) && hrtimer) 1593 hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD); 1594 else 1595 tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1); 1596 tick_nohz_activate(ts); 1597 } 1598 1599 /* 1600 * Shut down the tick and make sure the CPU won't try to retake the timekeeping 1601 * duty before disabling IRQs in idle for the last time. 1602 */ 1603 void tick_sched_timer_dying(int cpu) 1604 { 1605 struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu); 1606 ktime_t idle_sleeptime, iowait_sleeptime; 1607 unsigned long idle_calls, idle_sleeps; 1608 1609 /* This must happen before hrtimers are migrated! */ 1610 if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES)) 1611 hrtimer_cancel(&ts->sched_timer); 1612 1613 idle_sleeptime = ts->idle_sleeptime; 1614 iowait_sleeptime = ts->iowait_sleeptime; 1615 idle_calls = ts->idle_calls; 1616 idle_sleeps = ts->idle_sleeps; 1617 memset(ts, 0, sizeof(*ts)); 1618 ts->idle_sleeptime = idle_sleeptime; 1619 ts->iowait_sleeptime = iowait_sleeptime; 1620 ts->idle_calls = idle_calls; 1621 ts->idle_sleeps = idle_sleeps; 1622 } 1623 1624 /* 1625 * Async notification about clocksource changes 1626 */ 1627 void tick_clock_notify(void) 1628 { 1629 int cpu; 1630 1631 for_each_possible_cpu(cpu) 1632 set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks); 1633 } 1634 1635 /* 1636 * Async notification about clock event changes 1637 */ 1638 void tick_oneshot_notify(void) 1639 { 1640 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1641 1642 set_bit(0, &ts->check_clocks); 1643 } 1644 1645 /* 1646 * Check if a change happened, which makes oneshot possible. 1647 * 1648 * Called cyclically from the hrtimer softirq (driven by the timer 1649 * softirq). 'allow_nohz' signals that we can switch into low-res NOHZ 1650 * mode, because high resolution timers are disabled (either compile 1651 * or runtime). Called with interrupts disabled. 1652 */ 1653 int tick_check_oneshot_change(int allow_nohz) 1654 { 1655 struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched); 1656 1657 if (!test_and_clear_bit(0, &ts->check_clocks)) 1658 return 0; 1659 1660 if (tick_sched_flag_test(ts, TS_FLAG_NOHZ)) 1661 return 0; 1662 1663 if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available()) 1664 return 0; 1665 1666 if (!allow_nohz) 1667 return 1; 1668 1669 tick_nohz_switch_to_nohz(); 1670 return 0; 1671 }