tick-sched.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  Copyright(C) 2005-2006, Linutronix GmbH, Thomas Gleixner <tglx@kernel.org>
   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 */
  40static DEFINE_PER_CPU(struct tick_sched, tick_cpu_sched);
  41
  42struct 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 */
  52static ktime_t last_jiffies_update;
  53
  54/*
  55 * Must be called with interrupts disabled !
  56 */
  57static 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 */
 155static 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
 184static inline int tick_sched_flag_test(struct tick_sched *ts,
 185				       unsigned long flag)
 186{
 187	return !!(ts->flags & flag);
 188}
 189
 190static 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
 197static 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/*
 205 * Allow only one non-timekeeper CPU at a time update jiffies from
 206 * the timer tick.
 207 *
 208 * Returns true if update was run.
 209 */
 210static bool tick_limited_update_jiffies64(struct tick_sched *ts, ktime_t now)
 211{
 212	static atomic_t in_progress;
 213	int inp;
 214
 215	inp = atomic_read(&in_progress);
 216	if (inp || !atomic_try_cmpxchg(&in_progress, &inp, 1))
 217		return false;
 218
 219	if (ts->last_tick_jiffies == jiffies)
 220		tick_do_update_jiffies64(now);
 221	atomic_set(&in_progress, 0);
 222	return true;
 223}
 224
 225#define MAX_STALLED_JIFFIES 5
 226
 227static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
 228{
 229	int tick_cpu, cpu = smp_processor_id();
 230
 231	/*
 232	 * Check if the do_timer duty was dropped. We don't care about
 233	 * concurrency: This happens only when the CPU in charge went
 234	 * into a long sleep. If two CPUs happen to assign themselves to
 235	 * this duty, then the jiffies update is still serialized by
 236	 * 'jiffies_lock'.
 237	 *
 238	 * If nohz_full is enabled, this should not happen because the
 239	 * 'tick_do_timer_cpu' CPU never relinquishes.
 240	 */
 241	tick_cpu = READ_ONCE(tick_do_timer_cpu);
 242
 243	if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && unlikely(tick_cpu == TICK_DO_TIMER_NONE)) {
 244#ifdef CONFIG_NO_HZ_FULL
 245		WARN_ON_ONCE(tick_nohz_full_running);
 246#endif
 247		WRITE_ONCE(tick_do_timer_cpu, cpu);
 248		tick_cpu = cpu;
 249	}
 250
 251	/* Check if jiffies need an update */
 252	if (tick_cpu == cpu)
 253		tick_do_update_jiffies64(now);
 254
 255	/*
 256	 * If the jiffies update stalled for too long (timekeeper in stop_machine()
 257	 * or VMEXIT'ed for several msecs), force an update.
 258	 */
 259	if (ts->last_tick_jiffies != jiffies) {
 260		ts->stalled_jiffies = 0;
 261		ts->last_tick_jiffies = READ_ONCE(jiffies);
 262	} else {
 263		if (++ts->stalled_jiffies >= MAX_STALLED_JIFFIES) {
 264			if (tick_limited_update_jiffies64(ts, now)) {
 265				ts->stalled_jiffies = 0;
 266				ts->last_tick_jiffies = READ_ONCE(jiffies);
 267			}
 268		}
 269	}
 270
 271	if (tick_sched_flag_test(ts, TS_FLAG_INIDLE))
 272		ts->got_idle_tick = 1;
 273}
 274
 275static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
 276{
 277	/*
 278	 * When we are idle and the tick is stopped, we have to touch
 279	 * the watchdog as we might not schedule for a really long
 280	 * time. This happens on completely idle SMP systems while
 281	 * waiting on the login prompt. We also increment the "start of
 282	 * idle" jiffy stamp so the idle accounting adjustment we do
 283	 * when we go busy again does not account too many ticks.
 284	 */
 285	if (IS_ENABLED(CONFIG_NO_HZ_COMMON) &&
 286	    tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
 287		touch_softlockup_watchdog_sched();
 288		if (is_idle_task(current))
 289			ts->idle_jiffies++;
 290		/*
 291		 * In case the current tick fired too early past its expected
 292		 * expiration, make sure we don't bypass the next clock reprogramming
 293		 * to the same deadline.
 294		 */
 295		ts->next_tick = 0;
 296	}
 297
 298	update_process_times(user_mode(regs));
 299	profile_tick(CPU_PROFILING);
 300}
 301
 302/*
 303 * We rearm the timer until we get disabled by the idle code.
 304 * Called with interrupts disabled.
 305 */
 306static enum hrtimer_restart tick_nohz_handler(struct hrtimer *timer)
 307{
 308	struct tick_sched *ts =	container_of(timer, struct tick_sched, sched_timer);
 309	struct pt_regs *regs = get_irq_regs();
 310	ktime_t now = ktime_get();
 311
 312	tick_sched_do_timer(ts, now);
 313
 314	/*
 315	 * Do not call when we are not in IRQ context and have
 316	 * no valid 'regs' pointer
 317	 */
 318	if (regs)
 319		tick_sched_handle(ts, regs);
 320	else
 321		ts->next_tick = 0;
 322
 323	/*
 324	 * In dynticks mode, tick reprogram is deferred:
 325	 * - to the idle task if in dynticks-idle
 326	 * - to IRQ exit if in full-dynticks.
 327	 */
 328	if (unlikely(tick_sched_flag_test(ts, TS_FLAG_STOPPED)))
 329		return HRTIMER_NORESTART;
 330
 331	hrtimer_forward(timer, now, TICK_NSEC);
 332
 333	return HRTIMER_RESTART;
 334}
 335
 336#ifdef CONFIG_NO_HZ_FULL
 337cpumask_var_t tick_nohz_full_mask;
 338EXPORT_SYMBOL_GPL(tick_nohz_full_mask);
 339bool tick_nohz_full_running;
 340EXPORT_SYMBOL_GPL(tick_nohz_full_running);
 341static atomic_t tick_dep_mask;
 342
 343static bool check_tick_dependency(atomic_t *dep)
 344{
 345	int val = atomic_read(dep);
 346
 347	if (val & TICK_DEP_MASK_POSIX_TIMER) {
 348		trace_tick_stop(0, TICK_DEP_MASK_POSIX_TIMER);
 349		return true;
 350	}
 351
 352	if (val & TICK_DEP_MASK_PERF_EVENTS) {
 353		trace_tick_stop(0, TICK_DEP_MASK_PERF_EVENTS);
 354		return true;
 355	}
 356
 357	if (val & TICK_DEP_MASK_SCHED) {
 358		trace_tick_stop(0, TICK_DEP_MASK_SCHED);
 359		return true;
 360	}
 361
 362	if (val & TICK_DEP_MASK_CLOCK_UNSTABLE) {
 363		trace_tick_stop(0, TICK_DEP_MASK_CLOCK_UNSTABLE);
 364		return true;
 365	}
 366
 367	if (val & TICK_DEP_MASK_RCU) {
 368		trace_tick_stop(0, TICK_DEP_MASK_RCU);
 369		return true;
 370	}
 371
 372	if (val & TICK_DEP_MASK_RCU_EXP) {
 373		trace_tick_stop(0, TICK_DEP_MASK_RCU_EXP);
 374		return true;
 375	}
 376
 377	return false;
 378}
 379
 380static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
 381{
 382	lockdep_assert_irqs_disabled();
 383
 384	if (unlikely(!cpu_online(cpu)))
 385		return false;
 386
 387	if (check_tick_dependency(&tick_dep_mask))
 388		return false;
 389
 390	if (check_tick_dependency(&ts->tick_dep_mask))
 391		return false;
 392
 393	if (check_tick_dependency(&current->tick_dep_mask))
 394		return false;
 395
 396	if (check_tick_dependency(&current->signal->tick_dep_mask))
 397		return false;
 398
 399	return true;
 400}
 401
 402static void nohz_full_kick_func(struct irq_work *work)
 403{
 404	/* Empty, the tick restart happens on tick_nohz_irq_exit() */
 405}
 406
 407static DEFINE_PER_CPU(struct irq_work, nohz_full_kick_work) =
 408	IRQ_WORK_INIT_HARD(nohz_full_kick_func);
 409
 410/*
 411 * Kick this CPU if it's full dynticks in order to force it to
 412 * re-evaluate its dependency on the tick and restart it if necessary.
 413 * This kick, unlike tick_nohz_full_kick_cpu() and tick_nohz_full_kick_all(),
 414 * is NMI safe.
 415 */
 416static void tick_nohz_full_kick(void)
 417{
 418	if (!tick_nohz_full_cpu(smp_processor_id()))
 419		return;
 420
 421	irq_work_queue(this_cpu_ptr(&nohz_full_kick_work));
 422}
 423
 424/*
 425 * Kick the CPU if it's full dynticks in order to force it to
 426 * re-evaluate its dependency on the tick and restart it if necessary.
 427 */
 428void tick_nohz_full_kick_cpu(int cpu)
 429{
 430	if (!tick_nohz_full_cpu(cpu))
 431		return;
 432
 433	irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
 434}
 435
 436static void tick_nohz_kick_task(struct task_struct *tsk)
 437{
 438	int cpu;
 439
 440	/*
 441	 * If the task is not running, run_posix_cpu_timers()
 442	 * has nothing to elapse, and an IPI can then be optimized out.
 443	 *
 444	 * activate_task()                      STORE p->tick_dep_mask
 445	 *   STORE p->on_rq
 446	 * __schedule() (switch to task 'p')    smp_mb() (atomic_fetch_or())
 447	 *   LOCK rq->lock                      LOAD p->on_rq
 448	 *   smp_mb__after_spin_lock()
 449	 *   tick_nohz_task_switch()
 450	 *     LOAD p->tick_dep_mask
 451	 *
 452	 * XXX given a task picks up the dependency on schedule(), should we
 453	 * only care about tasks that are currently on the CPU instead of all
 454	 * that are on the runqueue?
 455	 *
 456	 * That is, does this want to be: task_on_cpu() / task_curr()?
 457	 */
 458	if (!sched_task_on_rq(tsk))
 459		return;
 460
 461	/*
 462	 * If the task concurrently migrates to another CPU,
 463	 * we guarantee it sees the new tick dependency upon
 464	 * schedule.
 465	 *
 466	 * set_task_cpu(p, cpu);
 467	 *   STORE p->cpu = @cpu
 468	 * __schedule() (switch to task 'p')
 469	 *   LOCK rq->lock
 470	 *   smp_mb__after_spin_lock()          STORE p->tick_dep_mask
 471	 *   tick_nohz_task_switch()            smp_mb() (atomic_fetch_or())
 472	 *      LOAD p->tick_dep_mask           LOAD p->cpu
 473	 */
 474	cpu = task_cpu(tsk);
 475
 476	preempt_disable();
 477	if (cpu_online(cpu))
 478		tick_nohz_full_kick_cpu(cpu);
 479	preempt_enable();
 480}
 481
 482/*
 483 * Kick all full dynticks CPUs in order to force these to re-evaluate
 484 * their dependency on the tick and restart it if necessary.
 485 */
 486static void tick_nohz_full_kick_all(void)
 487{
 488	int cpu;
 489
 490	if (!tick_nohz_full_running)
 491		return;
 492
 493	preempt_disable();
 494	for_each_cpu_and(cpu, tick_nohz_full_mask, cpu_online_mask)
 495		tick_nohz_full_kick_cpu(cpu);
 496	preempt_enable();
 497}
 498
 499static void tick_nohz_dep_set_all(atomic_t *dep,
 500				  enum tick_dep_bits bit)
 501{
 502	int prev;
 503
 504	prev = atomic_fetch_or(BIT(bit), dep);
 505	if (!prev)
 506		tick_nohz_full_kick_all();
 507}
 508
 509/*
 510 * Set a global tick dependency. Used by perf events that rely on freq and
 511 * unstable clocks.
 512 */
 513void tick_nohz_dep_set(enum tick_dep_bits bit)
 514{
 515	tick_nohz_dep_set_all(&tick_dep_mask, bit);
 516}
 517
 518void tick_nohz_dep_clear(enum tick_dep_bits bit)
 519{
 520	atomic_andnot(BIT(bit), &tick_dep_mask);
 521}
 522
 523/*
 524 * Set per-CPU tick dependency. Used by scheduler and perf events in order to
 525 * manage event-throttling.
 526 */
 527void tick_nohz_dep_set_cpu(int cpu, enum tick_dep_bits bit)
 528{
 529	int prev;
 530	struct tick_sched *ts;
 531
 532	ts = per_cpu_ptr(&tick_cpu_sched, cpu);
 533
 534	prev = atomic_fetch_or(BIT(bit), &ts->tick_dep_mask);
 535	if (!prev) {
 536		preempt_disable();
 537		/* Perf needs local kick that is NMI safe */
 538		if (cpu == smp_processor_id()) {
 539			tick_nohz_full_kick();
 540		} else {
 541			/* Remote IRQ work not NMI-safe */
 542			if (!WARN_ON_ONCE(in_nmi()))
 543				tick_nohz_full_kick_cpu(cpu);
 544		}
 545		preempt_enable();
 546	}
 547}
 548EXPORT_SYMBOL_GPL(tick_nohz_dep_set_cpu);
 549
 550void tick_nohz_dep_clear_cpu(int cpu, enum tick_dep_bits bit)
 551{
 552	struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
 553
 554	atomic_andnot(BIT(bit), &ts->tick_dep_mask);
 555}
 556EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_cpu);
 557
 558/*
 559 * Set a per-task tick dependency. RCU needs this. Also posix CPU timers
 560 * in order to elapse per task timers.
 561 */
 562void tick_nohz_dep_set_task(struct task_struct *tsk, enum tick_dep_bits bit)
 563{
 564	if (!atomic_fetch_or(BIT(bit), &tsk->tick_dep_mask))
 565		tick_nohz_kick_task(tsk);
 566}
 567EXPORT_SYMBOL_GPL(tick_nohz_dep_set_task);
 568
 569void tick_nohz_dep_clear_task(struct task_struct *tsk, enum tick_dep_bits bit)
 570{
 571	atomic_andnot(BIT(bit), &tsk->tick_dep_mask);
 572}
 573EXPORT_SYMBOL_GPL(tick_nohz_dep_clear_task);
 574
 575/*
 576 * Set a per-taskgroup tick dependency. Posix CPU timers need this in order to elapse
 577 * per process timers.
 578 */
 579void tick_nohz_dep_set_signal(struct task_struct *tsk,
 580			      enum tick_dep_bits bit)
 581{
 582	int prev;
 583	struct signal_struct *sig = tsk->signal;
 584
 585	prev = atomic_fetch_or(BIT(bit), &sig->tick_dep_mask);
 586	if (!prev) {
 587		struct task_struct *t;
 588
 589		lockdep_assert_held(&tsk->sighand->siglock);
 590		__for_each_thread(sig, t)
 591			tick_nohz_kick_task(t);
 592	}
 593}
 594
 595void tick_nohz_dep_clear_signal(struct signal_struct *sig, enum tick_dep_bits bit)
 596{
 597	atomic_andnot(BIT(bit), &sig->tick_dep_mask);
 598}
 599
 600/*
 601 * Re-evaluate the need for the tick as we switch the current task.
 602 * It might need the tick due to per task/process properties:
 603 * perf events, posix CPU timers, ...
 604 */
 605void __tick_nohz_task_switch(void)
 606{
 607	struct tick_sched *ts;
 608
 609	if (!tick_nohz_full_cpu(smp_processor_id()))
 610		return;
 611
 612	ts = this_cpu_ptr(&tick_cpu_sched);
 613
 614	if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
 615		if (atomic_read(&current->tick_dep_mask) ||
 616		    atomic_read(&current->signal->tick_dep_mask))
 617			tick_nohz_full_kick();
 618	}
 619}
 620
 621/* Get the boot-time nohz CPU list from the kernel parameters. */
 622void __init tick_nohz_full_setup(cpumask_var_t cpumask)
 623{
 624	alloc_bootmem_cpumask_var(&tick_nohz_full_mask);
 625	cpumask_copy(tick_nohz_full_mask, cpumask);
 626	tick_nohz_full_running = true;
 627}
 628
 629bool tick_nohz_cpu_hotpluggable(unsigned int cpu)
 630{
 631	/*
 632	 * The 'tick_do_timer_cpu' CPU handles housekeeping duty (unbound
 633	 * timers, workqueues, timekeeping, ...) on behalf of full dynticks
 634	 * CPUs. It must remain online when nohz full is enabled.
 635	 */
 636	if (tick_nohz_full_running && READ_ONCE(tick_do_timer_cpu) == cpu)
 637		return false;
 638	return true;
 639}
 640
 641static int tick_nohz_cpu_down(unsigned int cpu)
 642{
 643	return tick_nohz_cpu_hotpluggable(cpu) ? 0 : -EBUSY;
 644}
 645
 646void __init tick_nohz_init(void)
 647{
 648	int cpu, ret;
 649
 650	if (!tick_nohz_full_running)
 651		return;
 652
 653	/*
 654	 * Full dynticks uses IRQ work to drive the tick rescheduling on safe
 655	 * locking contexts. But then we need IRQ work to raise its own
 656	 * interrupts to avoid circular dependency on the tick.
 657	 */
 658	if (!arch_irq_work_has_interrupt()) {
 659		pr_warn("NO_HZ: Can't run full dynticks because arch doesn't support IRQ work self-IPIs\n");
 660		cpumask_clear(tick_nohz_full_mask);
 661		tick_nohz_full_running = false;
 662		return;
 663	}
 664
 665	if (IS_ENABLED(CONFIG_PM_SLEEP_SMP) &&
 666			!IS_ENABLED(CONFIG_PM_SLEEP_SMP_NONZERO_CPU)) {
 667		cpu = smp_processor_id();
 668
 669		if (cpumask_test_cpu(cpu, tick_nohz_full_mask)) {
 670			pr_warn("NO_HZ: Clearing %d from nohz_full range "
 671				"for timekeeping\n", cpu);
 672			cpumask_clear_cpu(cpu, tick_nohz_full_mask);
 673		}
 674	}
 675
 676	for_each_cpu(cpu, tick_nohz_full_mask)
 677		ct_cpu_track_user(cpu);
 678
 679	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 680					"kernel/nohz:predown", NULL,
 681					tick_nohz_cpu_down);
 682	WARN_ON(ret < 0);
 683	pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
 684		cpumask_pr_args(tick_nohz_full_mask));
 685}
 686#endif /* #ifdef CONFIG_NO_HZ_FULL */
 687
 688/*
 689 * NOHZ - aka dynamic tick functionality
 690 */
 691#ifdef CONFIG_NO_HZ_COMMON
 692/*
 693 * NO HZ enabled ?
 694 */
 695bool tick_nohz_enabled __read_mostly  = true;
 696unsigned long tick_nohz_active  __read_mostly;
 697/*
 698 * Enable / Disable tickless mode
 699 */
 700static int __init setup_tick_nohz(char *str)
 701{
 702	return (kstrtobool(str, &tick_nohz_enabled) == 0);
 703}
 704
 705__setup("nohz=", setup_tick_nohz);
 706
 707bool tick_nohz_tick_stopped(void)
 708{
 709	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
 710
 711	return tick_sched_flag_test(ts, TS_FLAG_STOPPED);
 712}
 713
 714bool tick_nohz_tick_stopped_cpu(int cpu)
 715{
 716	struct tick_sched *ts = per_cpu_ptr(&tick_cpu_sched, cpu);
 717
 718	return tick_sched_flag_test(ts, TS_FLAG_STOPPED);
 719}
 720
 721/**
 722 * tick_nohz_update_jiffies - update jiffies when idle was interrupted
 723 * @now: current ktime_t
 724 *
 725 * Called from interrupt entry when the CPU was idle
 726 *
 727 * In case the sched_tick was stopped on this CPU, we have to check if jiffies
 728 * must be updated. Otherwise an interrupt handler could use a stale jiffy
 729 * value. We do this unconditionally on any CPU, as we don't know whether the
 730 * CPU, which has the update task assigned, is in a long sleep.
 731 */
 732static void tick_nohz_update_jiffies(ktime_t now)
 733{
 734	unsigned long flags;
 735
 736	__this_cpu_write(tick_cpu_sched.idle_waketime, now);
 737
 738	local_irq_save(flags);
 739	tick_do_update_jiffies64(now);
 740	local_irq_restore(flags);
 741
 742	touch_softlockup_watchdog_sched();
 743}
 744
 745static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
 746{
 747	ktime_t delta;
 748
 749	if (WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE)))
 750		return;
 751
 752	delta = ktime_sub(now, ts->idle_entrytime);
 753
 754	write_seqcount_begin(&ts->idle_sleeptime_seq);
 755	if (nr_iowait_cpu(smp_processor_id()) > 0)
 756		ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
 757	else
 758		ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
 759
 760	ts->idle_entrytime = now;
 761	tick_sched_flag_clear(ts, TS_FLAG_IDLE_ACTIVE);
 762	write_seqcount_end(&ts->idle_sleeptime_seq);
 763
 764	sched_clock_idle_wakeup_event();
 765}
 766
 767static void tick_nohz_start_idle(struct tick_sched *ts)
 768{
 769	write_seqcount_begin(&ts->idle_sleeptime_seq);
 770	ts->idle_entrytime = ktime_get();
 771	tick_sched_flag_set(ts, TS_FLAG_IDLE_ACTIVE);
 772	write_seqcount_end(&ts->idle_sleeptime_seq);
 773
 774	sched_clock_idle_sleep_event();
 775}
 776
 777static u64 get_cpu_sleep_time_us(struct tick_sched *ts, ktime_t *sleeptime,
 778				 bool compute_delta, u64 *last_update_time)
 779{
 780	ktime_t now, idle;
 781	unsigned int seq;
 782
 783	if (!tick_nohz_active)
 784		return -1;
 785
 786	now = ktime_get();
 787	if (last_update_time)
 788		*last_update_time = ktime_to_us(now);
 789
 790	do {
 791		seq = read_seqcount_begin(&ts->idle_sleeptime_seq);
 792
 793		if (tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE) && compute_delta) {
 794			ktime_t delta = ktime_sub(now, ts->idle_entrytime);
 795
 796			idle = ktime_add(*sleeptime, delta);
 797		} else {
 798			idle = *sleeptime;
 799		}
 800	} while (read_seqcount_retry(&ts->idle_sleeptime_seq, seq));
 801
 802	return ktime_to_us(idle);
 803
 804}
 805
 806/**
 807 * get_cpu_idle_time_us - get the total idle time of a CPU
 808 * @cpu: CPU number to query
 809 * @last_update_time: variable to store update time in. Do not update
 810 * counters if NULL.
 811 *
 812 * Return the cumulative idle time (since boot) for a given
 813 * CPU, in microseconds. Note that this is partially broken due to
 814 * the counter of iowait tasks that can be remotely updated without
 815 * any synchronization. Therefore it is possible to observe backward
 816 * values within two consecutive reads.
 817 *
 818 * This time is measured via accounting rather than sampling,
 819 * and is as accurate as ktime_get() is.
 820 *
 821 * Return: -1 if NOHZ is not enabled, else total idle time of the @cpu
 822 */
 823u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
 824{
 825	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 826
 827	return get_cpu_sleep_time_us(ts, &ts->idle_sleeptime,
 828				     !nr_iowait_cpu(cpu), last_update_time);
 829}
 830EXPORT_SYMBOL_GPL(get_cpu_idle_time_us);
 831
 832/**
 833 * get_cpu_iowait_time_us - get the total iowait time of a CPU
 834 * @cpu: CPU number to query
 835 * @last_update_time: variable to store update time in. Do not update
 836 * counters if NULL.
 837 *
 838 * Return the cumulative iowait time (since boot) for a given
 839 * CPU, in microseconds. Note this is partially broken due to
 840 * the counter of iowait tasks that can be remotely updated without
 841 * any synchronization. Therefore it is possible to observe backward
 842 * values within two consecutive reads.
 843 *
 844 * This time is measured via accounting rather than sampling,
 845 * and is as accurate as ktime_get() is.
 846 *
 847 * Return: -1 if NOHZ is not enabled, else total iowait time of @cpu
 848 */
 849u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
 850{
 851	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
 852
 853	return get_cpu_sleep_time_us(ts, &ts->iowait_sleeptime,
 854				     nr_iowait_cpu(cpu), last_update_time);
 855}
 856EXPORT_SYMBOL_GPL(get_cpu_iowait_time_us);
 857
 858static void tick_nohz_restart(struct tick_sched *ts, ktime_t now)
 859{
 860	hrtimer_cancel(&ts->sched_timer);
 861	hrtimer_set_expires(&ts->sched_timer, ts->last_tick);
 862
 863	/* Forward the time to expire in the future */
 864	hrtimer_forward(&ts->sched_timer, now, TICK_NSEC);
 865
 866	if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES)) {
 867		hrtimer_start_expires(&ts->sched_timer,
 868				      HRTIMER_MODE_ABS_PINNED_HARD);
 869	} else {
 870		tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
 871	}
 872
 873	/*
 874	 * Reset to make sure the next tick stop doesn't get fooled by past
 875	 * cached clock deadline.
 876	 */
 877	ts->next_tick = 0;
 878}
 879
 880static inline bool local_timer_softirq_pending(void)
 881{
 882	return local_timers_pending() & BIT(TIMER_SOFTIRQ);
 883}
 884
 885/*
 886 * Read jiffies and the time when jiffies were updated last
 887 */
 888u64 get_jiffies_update(unsigned long *basej)
 889{
 890	unsigned long basejiff;
 891	unsigned int seq;
 892	u64 basemono;
 893
 894	do {
 895		seq = read_seqcount_begin(&jiffies_seq);
 896		basemono = last_jiffies_update;
 897		basejiff = jiffies;
 898	} while (read_seqcount_retry(&jiffies_seq, seq));
 899	*basej = basejiff;
 900	return basemono;
 901}
 902
 903/**
 904 * tick_nohz_next_event() - return the clock monotonic based next event
 905 * @ts:		pointer to tick_sched struct
 906 * @cpu:	CPU number
 907 *
 908 * Return:
 909 * *%0		- When the next event is a maximum of TICK_NSEC in the future
 910 *		  and the tick is not stopped yet
 911 * *%next_event	- Next event based on clock monotonic
 912 */
 913static ktime_t tick_nohz_next_event(struct tick_sched *ts, int cpu)
 914{
 915	u64 basemono, next_tick, delta, expires;
 916	unsigned long basejiff;
 917	int tick_cpu;
 918
 919	basemono = get_jiffies_update(&basejiff);
 920	ts->last_jiffies = basejiff;
 921	ts->timer_expires_base = basemono;
 922
 923	/*
 924	 * Keep the periodic tick, when RCU, architecture or irq_work
 925	 * requests it.
 926	 * Aside of that, check whether the local timer softirq is
 927	 * pending. If so, its a bad idea to call get_next_timer_interrupt(),
 928	 * because there is an already expired timer, so it will request
 929	 * immediate expiry, which rearms the hardware timer with a
 930	 * minimal delta, which brings us back to this place
 931	 * immediately. Lather, rinse and repeat...
 932	 */
 933	if (rcu_needs_cpu() || arch_needs_cpu() ||
 934	    irq_work_needs_cpu() || local_timer_softirq_pending()) {
 935		next_tick = basemono + TICK_NSEC;
 936	} else {
 937		/*
 938		 * Get the next pending timer. If high resolution
 939		 * timers are enabled this only takes the timer wheel
 940		 * timers into account. If high resolution timers are
 941		 * disabled this also looks at the next expiring
 942		 * hrtimer.
 943		 */
 944		next_tick = get_next_timer_interrupt(basejiff, basemono);
 945		ts->next_timer = next_tick;
 946	}
 947
 948	/* Make sure next_tick is never before basemono! */
 949	if (WARN_ON_ONCE(basemono > next_tick))
 950		next_tick = basemono;
 951
 952	/*
 953	 * If the tick is due in the next period, keep it ticking or
 954	 * force prod the timer.
 955	 */
 956	delta = next_tick - basemono;
 957	if (delta <= (u64)TICK_NSEC) {
 958		/*
 959		 * We've not stopped the tick yet, and there's a timer in the
 960		 * next period, so no point in stopping it either, bail.
 961		 */
 962		if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
 963			ts->timer_expires = 0;
 964			goto out;
 965		}
 966	}
 967
 968	/*
 969	 * If this CPU is the one which had the do_timer() duty last, we limit
 970	 * the sleep time to the timekeeping 'max_deferment' value.
 971	 * Otherwise we can sleep as long as we want.
 972	 */
 973	delta = timekeeping_max_deferment();
 974	tick_cpu = READ_ONCE(tick_do_timer_cpu);
 975	if (tick_cpu != cpu &&
 976	    (tick_cpu != TICK_DO_TIMER_NONE || !tick_sched_flag_test(ts, TS_FLAG_DO_TIMER_LAST)))
 977		delta = KTIME_MAX;
 978
 979	/* Calculate the next expiry time */
 980	if (delta < (KTIME_MAX - basemono))
 981		expires = basemono + delta;
 982	else
 983		expires = KTIME_MAX;
 984
 985	ts->timer_expires = min_t(u64, expires, next_tick);
 986
 987out:
 988	return ts->timer_expires;
 989}
 990
 991static void tick_nohz_stop_tick(struct tick_sched *ts, int cpu)
 992{
 993	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
 994	unsigned long basejiff = ts->last_jiffies;
 995	u64 basemono = ts->timer_expires_base;
 996	bool timer_idle = tick_sched_flag_test(ts, TS_FLAG_STOPPED);
 997	int tick_cpu;
 998	u64 expires;
 999
1000	/* Make sure we won't be trying to stop it twice in a row. */
1001	ts->timer_expires_base = 0;
1002
1003	/*
1004	 * Now the tick should be stopped definitely - so the timer base needs
1005	 * to be marked idle as well to not miss a newly queued timer.
1006	 */
1007	expires = timer_base_try_to_set_idle(basejiff, basemono, &timer_idle);
1008	if (expires > ts->timer_expires) {
1009		/*
1010		 * This path could only happen when the first timer was removed
1011		 * between calculating the possible sleep length and now (when
1012		 * high resolution mode is not active, timer could also be a
1013		 * hrtimer).
1014		 *
1015		 * We have to stick to the original calculated expiry value to
1016		 * not stop the tick for too long with a shallow C-state (which
1017		 * was programmed by cpuidle because of an early next expiration
1018		 * value).
1019		 */
1020		expires = ts->timer_expires;
1021	}
1022
1023	/* If the timer base is not idle, retain the not yet stopped tick. */
1024	if (!timer_idle)
1025		return;
1026
1027	/*
1028	 * If this CPU is the one which updates jiffies, then give up
1029	 * the assignment and let it be taken by the CPU which runs
1030	 * the tick timer next, which might be this CPU as well. If we
1031	 * don't drop this here, the jiffies might be stale and
1032	 * do_timer() never gets invoked. Keep track of the fact that it
1033	 * was the one which had the do_timer() duty last.
1034	 */
1035	tick_cpu = READ_ONCE(tick_do_timer_cpu);
1036	if (tick_cpu == cpu) {
1037		WRITE_ONCE(tick_do_timer_cpu, TICK_DO_TIMER_NONE);
1038		tick_sched_flag_set(ts, TS_FLAG_DO_TIMER_LAST);
1039	} else if (tick_cpu != TICK_DO_TIMER_NONE) {
1040		tick_sched_flag_clear(ts, TS_FLAG_DO_TIMER_LAST);
1041	}
1042
1043	/* Skip reprogram of event if it's not changed */
1044	if (tick_sched_flag_test(ts, TS_FLAG_STOPPED) && (expires == ts->next_tick)) {
1045		/* Sanity check: make sure clockevent is actually programmed */
1046		if (expires == KTIME_MAX || ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
1047			return;
1048
1049		WARN_ONCE(1, "basemono: %llu ts->next_tick: %llu dev->next_event: %llu "
1050			  "timer->active: %d timer->expires: %llu\n", basemono, ts->next_tick,
1051			  dev->next_event, hrtimer_active(&ts->sched_timer),
1052			  hrtimer_get_expires(&ts->sched_timer));
1053	}
1054
1055	/*
1056	 * tick_nohz_stop_tick() can be called several times before
1057	 * tick_nohz_restart_sched_tick() is called. This happens when
1058	 * interrupts arrive which do not cause a reschedule. In the first
1059	 * call we save the current tick time, so we can restart the
1060	 * scheduler tick in tick_nohz_restart_sched_tick().
1061	 */
1062	if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
1063		calc_load_nohz_start();
1064		quiet_vmstat();
1065
1066		ts->last_tick = hrtimer_get_expires(&ts->sched_timer);
1067		tick_sched_flag_set(ts, TS_FLAG_STOPPED);
1068		trace_tick_stop(1, TICK_DEP_MASK_NONE);
1069	}
1070
1071	ts->next_tick = expires;
1072
1073	/*
1074	 * If the expiration time == KTIME_MAX, then we simply stop
1075	 * the tick timer.
1076	 */
1077	if (unlikely(expires == KTIME_MAX)) {
1078		if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES))
1079			hrtimer_cancel(&ts->sched_timer);
1080		else
1081			tick_program_event(KTIME_MAX, 1);
1082		return;
1083	}
1084
1085	if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES)) {
1086		hrtimer_start(&ts->sched_timer, expires,
1087			      HRTIMER_MODE_ABS_PINNED_HARD);
1088	} else {
1089		hrtimer_set_expires(&ts->sched_timer, expires);
1090		tick_program_event(expires, 1);
1091	}
1092}
1093
1094static void tick_nohz_retain_tick(struct tick_sched *ts)
1095{
1096	ts->timer_expires_base = 0;
1097}
1098
1099#ifdef CONFIG_NO_HZ_FULL
1100static void tick_nohz_full_stop_tick(struct tick_sched *ts, int cpu)
1101{
1102	if (tick_nohz_next_event(ts, cpu))
1103		tick_nohz_stop_tick(ts, cpu);
1104	else
1105		tick_nohz_retain_tick(ts);
1106}
1107#endif /* CONFIG_NO_HZ_FULL */
1108
1109static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
1110{
1111	/* Update jiffies first */
1112	tick_do_update_jiffies64(now);
1113
1114	/*
1115	 * Clear the timer idle flag, so we avoid IPIs on remote queueing and
1116	 * the clock forward checks in the enqueue path:
1117	 */
1118	timer_clear_idle();
1119
1120	calc_load_nohz_stop();
1121	touch_softlockup_watchdog_sched();
1122
1123	/* Cancel the scheduled timer and restore the tick: */
1124	tick_sched_flag_clear(ts, TS_FLAG_STOPPED);
1125	tick_nohz_restart(ts, now);
1126}
1127
1128static void __tick_nohz_full_update_tick(struct tick_sched *ts,
1129					 ktime_t now)
1130{
1131#ifdef CONFIG_NO_HZ_FULL
1132	int cpu = smp_processor_id();
1133
1134	if (can_stop_full_tick(cpu, ts))
1135		tick_nohz_full_stop_tick(ts, cpu);
1136	else if (tick_sched_flag_test(ts, TS_FLAG_STOPPED))
1137		tick_nohz_restart_sched_tick(ts, now);
1138#endif
1139}
1140
1141static void tick_nohz_full_update_tick(struct tick_sched *ts)
1142{
1143	if (!tick_nohz_full_cpu(smp_processor_id()))
1144		return;
1145
1146	if (!tick_sched_flag_test(ts, TS_FLAG_NOHZ))
1147		return;
1148
1149	__tick_nohz_full_update_tick(ts, ktime_get());
1150}
1151
1152/*
1153 * A pending softirq outside an IRQ (or softirq disabled section) context
1154 * should be waiting for ksoftirqd to handle it. Therefore we shouldn't
1155 * reach this code due to the need_resched() early check in can_stop_idle_tick().
1156 *
1157 * However if we are between CPUHP_AP_SMPBOOT_THREADS and CPU_TEARDOWN_CPU on the
1158 * cpu_down() process, softirqs can still be raised while ksoftirqd is parked,
1159 * triggering the code below, since wakep_softirqd() is ignored.
1160 *
1161 */
1162static bool report_idle_softirq(void)
1163{
1164	static int ratelimit;
1165	unsigned int pending = local_softirq_pending();
1166
1167	if (likely(!pending))
1168		return false;
1169
1170	/* Some softirqs claim to be safe against hotplug and ksoftirqd parking */
1171	if (!cpu_active(smp_processor_id())) {
1172		pending &= ~SOFTIRQ_HOTPLUG_SAFE_MASK;
1173		if (!pending)
1174			return false;
1175	}
1176
1177	/* On RT, softirq handling may be waiting on some lock */
1178	if (local_bh_blocked())
1179		return false;
1180
1181	if (ratelimit < 10) {
1182		pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",
1183			pending);
1184		ratelimit++;
1185	}
1186
1187	return true;
1188}
1189
1190static bool can_stop_idle_tick(int cpu, struct tick_sched *ts)
1191{
1192	WARN_ON_ONCE(cpu_is_offline(cpu));
1193
1194	if (unlikely(!tick_sched_flag_test(ts, TS_FLAG_NOHZ)))
1195		return false;
1196
1197	if (need_resched())
1198		return false;
1199
1200	if (unlikely(report_idle_softirq()))
1201		return false;
1202
1203	if (tick_nohz_full_enabled()) {
1204		int tick_cpu = READ_ONCE(tick_do_timer_cpu);
1205
1206		/*
1207		 * Keep the tick alive to guarantee timekeeping progression
1208		 * if there are full dynticks CPUs around
1209		 */
1210		if (tick_cpu == cpu)
1211			return false;
1212
1213		/* Should not happen for nohz-full */
1214		if (WARN_ON_ONCE(tick_cpu == TICK_DO_TIMER_NONE))
1215			return false;
1216	}
1217
1218	return true;
1219}
1220
1221/**
1222 * tick_nohz_idle_stop_tick - stop the idle tick from the idle task
1223 *
1224 * When the next event is more than a tick into the future, stop the idle tick
1225 */
1226void tick_nohz_idle_stop_tick(void)
1227{
1228	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1229	int cpu = smp_processor_id();
1230	ktime_t expires;
1231
1232	/*
1233	 * If tick_nohz_get_sleep_length() ran tick_nohz_next_event(), the
1234	 * tick timer expiration time is known already.
1235	 */
1236	if (ts->timer_expires_base)
1237		expires = ts->timer_expires;
1238	else if (can_stop_idle_tick(cpu, ts))
1239		expires = tick_nohz_next_event(ts, cpu);
1240	else
1241		return;
1242
1243	ts->idle_calls++;
1244
1245	if (expires > 0LL) {
1246		int was_stopped = tick_sched_flag_test(ts, TS_FLAG_STOPPED);
1247
1248		tick_nohz_stop_tick(ts, cpu);
1249
1250		ts->idle_sleeps++;
1251		ts->idle_expires = expires;
1252
1253		if (!was_stopped && tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
1254			ts->idle_jiffies = ts->last_jiffies;
1255			nohz_balance_enter_idle(cpu);
1256		}
1257	} else {
1258		tick_nohz_retain_tick(ts);
1259	}
1260}
1261
1262void tick_nohz_idle_retain_tick(void)
1263{
1264	tick_nohz_retain_tick(this_cpu_ptr(&tick_cpu_sched));
1265}
1266
1267/**
1268 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
1269 *
1270 * Called when we start the idle loop.
1271 */
1272void tick_nohz_idle_enter(void)
1273{
1274	struct tick_sched *ts;
1275
1276	lockdep_assert_irqs_enabled();
1277
1278	local_irq_disable();
1279
1280	ts = this_cpu_ptr(&tick_cpu_sched);
1281
1282	WARN_ON_ONCE(ts->timer_expires_base);
1283
1284	tick_sched_flag_set(ts, TS_FLAG_INIDLE);
1285	tick_nohz_start_idle(ts);
1286
1287	local_irq_enable();
1288}
1289
1290/**
1291 * tick_nohz_irq_exit - Notify the tick about IRQ exit
1292 *
1293 * A timer may have been added/modified/deleted either by the current IRQ,
1294 * or by another place using this IRQ as a notification. This IRQ may have
1295 * also updated the RCU callback list. These events may require a
1296 * re-evaluation of the next tick. Depending on the context:
1297 *
1298 * 1) If the CPU is idle and no resched is pending, just proceed with idle
1299 *    time accounting. The next tick will be re-evaluated on the next idle
1300 *    loop iteration.
1301 *
1302 * 2) If the CPU is nohz_full:
1303 *
1304 *    2.1) If there is any tick dependency, restart the tick if stopped.
1305 *
1306 *    2.2) If there is no tick dependency, (re-)evaluate the next tick and
1307 *         stop/update it accordingly.
1308 */
1309void tick_nohz_irq_exit(void)
1310{
1311	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1312
1313	if (tick_sched_flag_test(ts, TS_FLAG_INIDLE))
1314		tick_nohz_start_idle(ts);
1315	else
1316		tick_nohz_full_update_tick(ts);
1317}
1318
1319/**
1320 * tick_nohz_idle_got_tick - Check whether or not the tick handler has run
1321 *
1322 * Return: %true if the tick handler has run, otherwise %false
1323 */
1324bool tick_nohz_idle_got_tick(void)
1325{
1326	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1327
1328	if (ts->got_idle_tick) {
1329		ts->got_idle_tick = 0;
1330		return true;
1331	}
1332	return false;
1333}
1334
1335/**
1336 * tick_nohz_get_next_hrtimer - return the next expiration time for the hrtimer
1337 * or the tick, whichever expires first. Note that, if the tick has been
1338 * stopped, it returns the next hrtimer.
1339 *
1340 * Called from power state control code with interrupts disabled
1341 *
1342 * Return: the next expiration time
1343 */
1344ktime_t tick_nohz_get_next_hrtimer(void)
1345{
1346	return __this_cpu_read(tick_cpu_device.evtdev)->next_event;
1347}
1348
1349/**
1350 * tick_nohz_get_sleep_length - return the expected length of the current sleep
1351 * @delta_next: duration until the next event if the tick cannot be stopped
1352 *
1353 * Called from power state control code with interrupts disabled.
1354 *
1355 * The return value of this function and/or the value returned by it through the
1356 * @delta_next pointer can be negative which must be taken into account by its
1357 * callers.
1358 *
1359 * Return: the expected length of the current sleep
1360 */
1361ktime_t tick_nohz_get_sleep_length(ktime_t *delta_next)
1362{
1363	struct clock_event_device *dev = __this_cpu_read(tick_cpu_device.evtdev);
1364	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1365	int cpu = smp_processor_id();
1366	/*
1367	 * The idle entry time is expected to be a sufficient approximation of
1368	 * the current time at this point.
1369	 */
1370	ktime_t now = ts->idle_entrytime;
1371	ktime_t next_event;
1372
1373	WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_INIDLE));
1374
1375	*delta_next = ktime_sub(dev->next_event, now);
1376
1377	if (!can_stop_idle_tick(cpu, ts))
1378		return *delta_next;
1379
1380	next_event = tick_nohz_next_event(ts, cpu);
1381	if (!next_event)
1382		return *delta_next;
1383
1384	/*
1385	 * If the next highres timer to expire is earlier than 'next_event', the
1386	 * idle governor needs to know that.
1387	 */
1388	next_event = min_t(u64, next_event,
1389			   hrtimer_next_event_without(&ts->sched_timer));
1390
1391	return ktime_sub(next_event, now);
1392}
1393
1394/**
1395 * tick_nohz_get_idle_calls_cpu - return the current idle calls counter value
1396 * for a particular CPU.
1397 * @cpu: target CPU number
1398 *
1399 * Called from the schedutil frequency scaling governor in scheduler context.
1400 *
1401 * Return: the current idle calls counter value for @cpu
1402 */
1403unsigned long tick_nohz_get_idle_calls_cpu(int cpu)
1404{
1405	struct tick_sched *ts = tick_get_tick_sched(cpu);
1406
1407	return ts->idle_calls;
1408}
1409
1410static void tick_nohz_account_idle_time(struct tick_sched *ts,
1411					ktime_t now)
1412{
1413	unsigned long ticks;
1414
1415	ts->idle_exittime = now;
1416
1417	if (vtime_accounting_enabled_this_cpu())
1418		return;
1419	/*
1420	 * We stopped the tick in idle. update_process_times() would miss the
1421	 * time we slept, as it does only a 1 tick accounting.
1422	 * Enforce that this is accounted to idle !
1423	 */
1424	ticks = jiffies - ts->idle_jiffies;
1425	/*
1426	 * We might be one off. Do not randomly account a huge number of ticks!
1427	 */
1428	if (ticks && ticks < LONG_MAX)
1429		account_idle_ticks(ticks);
1430}
1431
1432void tick_nohz_idle_restart_tick(void)
1433{
1434	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1435
1436	if (tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
1437		ktime_t now = ktime_get();
1438		tick_nohz_restart_sched_tick(ts, now);
1439		tick_nohz_account_idle_time(ts, now);
1440	}
1441}
1442
1443static void tick_nohz_idle_update_tick(struct tick_sched *ts, ktime_t now)
1444{
1445	if (tick_nohz_full_cpu(smp_processor_id()))
1446		__tick_nohz_full_update_tick(ts, now);
1447	else
1448		tick_nohz_restart_sched_tick(ts, now);
1449
1450	tick_nohz_account_idle_time(ts, now);
1451}
1452
1453/**
1454 * tick_nohz_idle_exit - Update the tick upon idle task exit
1455 *
1456 * When the idle task exits, update the tick depending on the
1457 * following situations:
1458 *
1459 * 1) If the CPU is not in nohz_full mode (most cases), then
1460 *    restart the tick.
1461 *
1462 * 2) If the CPU is in nohz_full mode (corner case):
1463 *   2.1) If the tick can be kept stopped (no tick dependencies)
1464 *        then re-evaluate the next tick and try to keep it stopped
1465 *        as long as possible.
1466 *   2.2) If the tick has dependencies, restart the tick.
1467 *
1468 */
1469void tick_nohz_idle_exit(void)
1470{
1471	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1472	bool idle_active, tick_stopped;
1473	ktime_t now;
1474
1475	local_irq_disable();
1476
1477	WARN_ON_ONCE(!tick_sched_flag_test(ts, TS_FLAG_INIDLE));
1478	WARN_ON_ONCE(ts->timer_expires_base);
1479
1480	tick_sched_flag_clear(ts, TS_FLAG_INIDLE);
1481	idle_active = tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE);
1482	tick_stopped = tick_sched_flag_test(ts, TS_FLAG_STOPPED);
1483
1484	if (idle_active || tick_stopped)
1485		now = ktime_get();
1486
1487	if (idle_active)
1488		tick_nohz_stop_idle(ts, now);
1489
1490	if (tick_stopped)
1491		tick_nohz_idle_update_tick(ts, now);
1492
1493	local_irq_enable();
1494}
1495
1496/*
1497 * In low-resolution mode, the tick handler must be implemented directly
1498 * at the clockevent level. hrtimer can't be used instead, because its
1499 * infrastructure actually relies on the tick itself as a backend in
1500 * low-resolution mode (see hrtimer_run_queues()).
1501 */
1502static void tick_nohz_lowres_handler(struct clock_event_device *dev)
1503{
1504	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1505
1506	dev->next_event = KTIME_MAX;
1507
1508	if (likely(tick_nohz_handler(&ts->sched_timer) == HRTIMER_RESTART))
1509		tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1510}
1511
1512static inline void tick_nohz_activate(struct tick_sched *ts)
1513{
1514	if (!tick_nohz_enabled)
1515		return;
1516	tick_sched_flag_set(ts, TS_FLAG_NOHZ);
1517	/* One update is enough */
1518	if (!test_and_set_bit(0, &tick_nohz_active))
1519		timers_update_nohz();
1520}
1521
1522/**
1523 * tick_nohz_switch_to_nohz - switch to NOHZ mode
1524 */
1525static void tick_nohz_switch_to_nohz(void)
1526{
1527	if (!tick_nohz_enabled)
1528		return;
1529
1530	if (tick_switch_to_oneshot(tick_nohz_lowres_handler))
1531		return;
1532
1533	/*
1534	 * Recycle the hrtimer in 'ts', so we can share the
1535	 * highres code.
1536	 */
1537	tick_setup_sched_timer(false);
1538}
1539
1540static inline void tick_nohz_irq_enter(void)
1541{
1542	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1543	ktime_t now;
1544
1545	if (!tick_sched_flag_test(ts, TS_FLAG_STOPPED | TS_FLAG_IDLE_ACTIVE))
1546		return;
1547	now = ktime_get();
1548	if (tick_sched_flag_test(ts, TS_FLAG_IDLE_ACTIVE))
1549		tick_nohz_stop_idle(ts, now);
1550	/*
1551	 * If all CPUs are idle we may need to update a stale jiffies value.
1552	 * Note nohz_full is a special case: a timekeeper is guaranteed to stay
1553	 * alive but it might be busy looping with interrupts disabled in some
1554	 * rare case (typically stop machine). So we must make sure we have a
1555	 * last resort.
1556	 */
1557	if (tick_sched_flag_test(ts, TS_FLAG_STOPPED))
1558		tick_nohz_update_jiffies(now);
1559}
1560
1561#else
1562
1563static inline void tick_nohz_switch_to_nohz(void) { }
1564static inline void tick_nohz_irq_enter(void) { }
1565static inline void tick_nohz_activate(struct tick_sched *ts) { }
1566
1567#endif /* CONFIG_NO_HZ_COMMON */
1568
1569/*
1570 * Called from irq_enter() to notify about the possible interruption of idle()
1571 */
1572void tick_irq_enter(void)
1573{
1574	tick_check_oneshot_broadcast_this_cpu();
1575	tick_nohz_irq_enter();
1576}
1577
1578static int sched_skew_tick;
1579
1580static int __init skew_tick(char *str)
1581{
1582	get_option(&str, &sched_skew_tick);
1583
1584	return 0;
1585}
1586early_param("skew_tick", skew_tick);
1587
1588/**
1589 * tick_setup_sched_timer - setup the tick emulation timer
1590 * @hrtimer: whether to use the hrtimer or not
1591 */
1592void tick_setup_sched_timer(bool hrtimer)
1593{
1594	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1595
1596	/* Emulate tick processing via per-CPU hrtimers: */
1597	hrtimer_setup(&ts->sched_timer, tick_nohz_handler, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_HARD);
1598
1599	if (IS_ENABLED(CONFIG_HIGH_RES_TIMERS) && hrtimer)
1600		tick_sched_flag_set(ts, TS_FLAG_HIGHRES);
1601
1602	/* Get the next period (per-CPU) */
1603	hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
1604
1605	/* Offset the tick to avert 'jiffies_lock' contention. */
1606	if (sched_skew_tick) {
1607		u64 offset = TICK_NSEC >> 1;
1608		do_div(offset, num_possible_cpus());
1609		offset *= smp_processor_id();
1610		hrtimer_add_expires_ns(&ts->sched_timer, offset);
1611	}
1612
1613	hrtimer_forward_now(&ts->sched_timer, TICK_NSEC);
1614	if (IS_ENABLED(CONFIG_HIGH_RES_TIMERS) && hrtimer)
1615		hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED_HARD);
1616	else
1617		tick_program_event(hrtimer_get_expires(&ts->sched_timer), 1);
1618	tick_nohz_activate(ts);
1619}
1620
1621/*
1622 * Shut down the tick and make sure the CPU won't try to retake the timekeeping
1623 * duty before disabling IRQs in idle for the last time.
1624 */
1625void tick_sched_timer_dying(int cpu)
1626{
1627	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
1628	ktime_t idle_sleeptime, iowait_sleeptime;
1629	unsigned long idle_calls, idle_sleeps;
1630
1631	/* This must happen before hrtimers are migrated! */
1632	if (tick_sched_flag_test(ts, TS_FLAG_HIGHRES))
1633		hrtimer_cancel(&ts->sched_timer);
1634
1635	idle_sleeptime = ts->idle_sleeptime;
1636	iowait_sleeptime = ts->iowait_sleeptime;
1637	idle_calls = ts->idle_calls;
1638	idle_sleeps = ts->idle_sleeps;
1639	memset(ts, 0, sizeof(*ts));
1640	ts->idle_sleeptime = idle_sleeptime;
1641	ts->iowait_sleeptime = iowait_sleeptime;
1642	ts->idle_calls = idle_calls;
1643	ts->idle_sleeps = idle_sleeps;
1644}
1645
1646/*
1647 * Async notification about clocksource changes
1648 */
1649void tick_clock_notify(void)
1650{
1651	int cpu;
1652
1653	for_each_possible_cpu(cpu)
1654		set_bit(0, &per_cpu(tick_cpu_sched, cpu).check_clocks);
1655}
1656
1657/*
1658 * Async notification about clock event changes
1659 */
1660void tick_oneshot_notify(void)
1661{
1662	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1663
1664	set_bit(0, &ts->check_clocks);
1665}
1666
1667/*
1668 * Check if a change happened, which makes oneshot possible.
1669 *
1670 * Called cyclically from the hrtimer softirq (driven by the timer
1671 * softirq). 'allow_nohz' signals that we can switch into low-res NOHZ
1672 * mode, because high resolution timers are disabled (either compile
1673 * or runtime). Called with interrupts disabled.
1674 */
1675int tick_check_oneshot_change(int allow_nohz)
1676{
1677	struct tick_sched *ts = this_cpu_ptr(&tick_cpu_sched);
1678
1679	if (!test_and_clear_bit(0, &ts->check_clocks))
1680		return 0;
1681
1682	if (tick_sched_flag_test(ts, TS_FLAG_NOHZ))
1683		return 0;
1684
1685	if (!timekeeping_valid_for_hres() || !tick_is_oneshot_available())
1686		return 0;
1687
1688	if (!allow_nohz)
1689		return 1;
1690
1691	tick_nohz_switch_to_nohz();
1692	return 0;
1693}