hrtimer.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 *  High-resolution kernel timers
   8 *
   9 *  In contrast to the low-resolution timeout API, aka timer wheel,
  10 *  hrtimers provide finer resolution and accuracy depending on system
  11 *  configuration and capabilities.
  12 *
  13 *  Started by: Thomas Gleixner and Ingo Molnar
  14 *
  15 *  Credits:
  16 *	Based on the original timer wheel code
  17 *
  18 *	Help, testing, suggestions, bugfixes, improvements were
  19 *	provided by:
  20 *
  21 *	George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
  22 *	et. al.
  23 */
  24
  25#include <linux/cpu.h>
  26#include <linux/export.h>
  27#include <linux/percpu.h>
  28#include <linux/hrtimer.h>
  29#include <linux/notifier.h>
  30#include <linux/syscalls.h>
  31#include <linux/interrupt.h>
  32#include <linux/tick.h>
  33#include <linux/err.h>
  34#include <linux/debugobjects.h>
  35#include <linux/sched/signal.h>
  36#include <linux/sched/sysctl.h>
  37#include <linux/sched/rt.h>
  38#include <linux/sched/deadline.h>
  39#include <linux/sched/nohz.h>
  40#include <linux/sched/debug.h>
  41#include <linux/sched/isolation.h>
  42#include <linux/timer.h>
  43#include <linux/freezer.h>
  44#include <linux/compat.h>
  45
  46#include <linux/uaccess.h>
  47
  48#include <trace/events/timer.h>
  49
  50#include "tick-internal.h"
  51
  52/*
  53 * Masks for selecting the soft and hard context timers from
  54 * cpu_base->active
  55 */
  56#define MASK_SHIFT		(HRTIMER_BASE_MONOTONIC_SOFT)
  57#define HRTIMER_ACTIVE_HARD	((1U << MASK_SHIFT) - 1)
  58#define HRTIMER_ACTIVE_SOFT	(HRTIMER_ACTIVE_HARD << MASK_SHIFT)
  59#define HRTIMER_ACTIVE_ALL	(HRTIMER_ACTIVE_SOFT | HRTIMER_ACTIVE_HARD)
  60
  61static void retrigger_next_event(void *arg);
  62static ktime_t __hrtimer_cb_get_time(clockid_t clock_id);
  63
  64/*
  65 * The timer bases:
  66 *
  67 * There are more clockids than hrtimer bases. Thus, we index
  68 * into the timer bases by the hrtimer_base_type enum. When trying
  69 * to reach a base using a clockid, hrtimer_clockid_to_base()
  70 * is used to convert from clockid to the proper hrtimer_base_type.
  71 */
  72DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
  73{
  74	.lock = __RAW_SPIN_LOCK_UNLOCKED(hrtimer_bases.lock),
  75	.clock_base =
  76	{
  77		{
  78			.index = HRTIMER_BASE_MONOTONIC,
  79			.clockid = CLOCK_MONOTONIC,
  80		},
  81		{
  82			.index = HRTIMER_BASE_REALTIME,
  83			.clockid = CLOCK_REALTIME,
  84		},
  85		{
  86			.index = HRTIMER_BASE_BOOTTIME,
  87			.clockid = CLOCK_BOOTTIME,
  88		},
  89		{
  90			.index = HRTIMER_BASE_TAI,
  91			.clockid = CLOCK_TAI,
  92		},
  93		{
  94			.index = HRTIMER_BASE_MONOTONIC_SOFT,
  95			.clockid = CLOCK_MONOTONIC,
  96		},
  97		{
  98			.index = HRTIMER_BASE_REALTIME_SOFT,
  99			.clockid = CLOCK_REALTIME,
 100		},
 101		{
 102			.index = HRTIMER_BASE_BOOTTIME_SOFT,
 103			.clockid = CLOCK_BOOTTIME,
 104		},
 105		{
 106			.index = HRTIMER_BASE_TAI_SOFT,
 107			.clockid = CLOCK_TAI,
 108		},
 109	},
 110	.csd = CSD_INIT(retrigger_next_event, NULL)
 111};
 112
 113static inline bool hrtimer_base_is_online(struct hrtimer_cpu_base *base)
 114{
 115	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
 116		return true;
 117	else
 118		return likely(base->online);
 119}
 120
 121/*
 122 * Functions and macros which are different for UP/SMP systems are kept in a
 123 * single place
 124 */
 125#ifdef CONFIG_SMP
 126
 127/*
 128 * We require the migration_base for lock_hrtimer_base()/switch_hrtimer_base()
 129 * such that hrtimer_callback_running() can unconditionally dereference
 130 * timer->base->cpu_base
 131 */
 132static struct hrtimer_cpu_base migration_cpu_base = {
 133	.clock_base = { {
 134		.cpu_base = &migration_cpu_base,
 135		.seq      = SEQCNT_RAW_SPINLOCK_ZERO(migration_cpu_base.seq,
 136						     &migration_cpu_base.lock),
 137	}, },
 138};
 139
 140#define migration_base	migration_cpu_base.clock_base[0]
 141
 142/*
 143 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
 144 * means that all timers which are tied to this base via timer->base are
 145 * locked, and the base itself is locked too.
 146 *
 147 * So __run_timers/migrate_timers can safely modify all timers which could
 148 * be found on the lists/queues.
 149 *
 150 * When the timer's base is locked, and the timer removed from list, it is
 151 * possible to set timer->base = &migration_base and drop the lock: the timer
 152 * remains locked.
 153 */
 154static
 155struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
 156					     unsigned long *flags)
 157	__acquires(&timer->base->lock)
 158{
 159	struct hrtimer_clock_base *base;
 160
 161	for (;;) {
 162		base = READ_ONCE(timer->base);
 163		if (likely(base != &migration_base)) {
 164			raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
 165			if (likely(base == timer->base))
 166				return base;
 167			/* The timer has migrated to another CPU: */
 168			raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
 169		}
 170		cpu_relax();
 171	}
 172}
 173
 174/*
 175 * Check if the elected target is suitable considering its next
 176 * event and the hotplug state of the current CPU.
 177 *
 178 * If the elected target is remote and its next event is after the timer
 179 * to queue, then a remote reprogram is necessary. However there is no
 180 * guarantee the IPI handling the operation would arrive in time to meet
 181 * the high resolution deadline. In this case the local CPU becomes a
 182 * preferred target, unless it is offline.
 183 *
 184 * High and low resolution modes are handled the same way for simplicity.
 185 *
 186 * Called with cpu_base->lock of target cpu held.
 187 */
 188static bool hrtimer_suitable_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base,
 189				    struct hrtimer_cpu_base *new_cpu_base,
 190				    struct hrtimer_cpu_base *this_cpu_base)
 191{
 192	ktime_t expires;
 193
 194	/*
 195	 * The local CPU clockevent can be reprogrammed. Also get_target_base()
 196	 * guarantees it is online.
 197	 */
 198	if (new_cpu_base == this_cpu_base)
 199		return true;
 200
 201	/*
 202	 * The offline local CPU can't be the default target if the
 203	 * next remote target event is after this timer. Keep the
 204	 * elected new base. An IPI will be issued to reprogram
 205	 * it as a last resort.
 206	 */
 207	if (!hrtimer_base_is_online(this_cpu_base))
 208		return true;
 209
 210	expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
 211
 212	return expires >= new_base->cpu_base->expires_next;
 213}
 214
 215static inline struct hrtimer_cpu_base *get_target_base(struct hrtimer_cpu_base *base, int pinned)
 216{
 217	if (!hrtimer_base_is_online(base)) {
 218		int cpu = cpumask_any_and(cpu_online_mask, housekeeping_cpumask(HK_TYPE_TIMER));
 219
 220		return &per_cpu(hrtimer_bases, cpu);
 221	}
 222
 223#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
 224	if (static_branch_likely(&timers_migration_enabled) && !pinned)
 225		return &per_cpu(hrtimer_bases, get_nohz_timer_target());
 226#endif
 227	return base;
 228}
 229
 230/*
 231 * We switch the timer base to a power-optimized selected CPU target,
 232 * if:
 233 *	- NO_HZ_COMMON is enabled
 234 *	- timer migration is enabled
 235 *	- the timer callback is not running
 236 *	- the timer is not the first expiring timer on the new target
 237 *
 238 * If one of the above requirements is not fulfilled we move the timer
 239 * to the current CPU or leave it on the previously assigned CPU if
 240 * the timer callback is currently running.
 241 */
 242static inline struct hrtimer_clock_base *
 243switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
 244		    int pinned)
 245{
 246	struct hrtimer_cpu_base *new_cpu_base, *this_cpu_base;
 247	struct hrtimer_clock_base *new_base;
 248	int basenum = base->index;
 249
 250	this_cpu_base = this_cpu_ptr(&hrtimer_bases);
 251	new_cpu_base = get_target_base(this_cpu_base, pinned);
 252again:
 253	new_base = &new_cpu_base->clock_base[basenum];
 254
 255	if (base != new_base) {
 256		/*
 257		 * We are trying to move timer to new_base.
 258		 * However we can't change timer's base while it is running,
 259		 * so we keep it on the same CPU. No hassle vs. reprogramming
 260		 * the event source in the high resolution case. The softirq
 261		 * code will take care of this when the timer function has
 262		 * completed. There is no conflict as we hold the lock until
 263		 * the timer is enqueued.
 264		 */
 265		if (unlikely(hrtimer_callback_running(timer)))
 266			return base;
 267
 268		/* See the comment in lock_hrtimer_base() */
 269		WRITE_ONCE(timer->base, &migration_base);
 270		raw_spin_unlock(&base->cpu_base->lock);
 271		raw_spin_lock(&new_base->cpu_base->lock);
 272
 273		if (!hrtimer_suitable_target(timer, new_base, new_cpu_base,
 274					     this_cpu_base)) {
 275			raw_spin_unlock(&new_base->cpu_base->lock);
 276			raw_spin_lock(&base->cpu_base->lock);
 277			new_cpu_base = this_cpu_base;
 278			WRITE_ONCE(timer->base, base);
 279			goto again;
 280		}
 281		WRITE_ONCE(timer->base, new_base);
 282	} else {
 283		if (!hrtimer_suitable_target(timer, new_base,  new_cpu_base, this_cpu_base)) {
 284			new_cpu_base = this_cpu_base;
 285			goto again;
 286		}
 287	}
 288	return new_base;
 289}
 290
 291#else /* CONFIG_SMP */
 292
 293static inline struct hrtimer_clock_base *
 294lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
 295	__acquires(&timer->base->cpu_base->lock)
 296{
 297	struct hrtimer_clock_base *base = timer->base;
 298
 299	raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
 300
 301	return base;
 302}
 303
 304# define switch_hrtimer_base(t, b, p)	(b)
 305
 306#endif	/* !CONFIG_SMP */
 307
 308/*
 309 * Functions for the union type storage format of ktime_t which are
 310 * too large for inlining:
 311 */
 312#if BITS_PER_LONG < 64
 313/*
 314 * Divide a ktime value by a nanosecond value
 315 */
 316s64 __ktime_divns(const ktime_t kt, s64 div)
 317{
 318	int sft = 0;
 319	s64 dclc;
 320	u64 tmp;
 321
 322	dclc = ktime_to_ns(kt);
 323	tmp = dclc < 0 ? -dclc : dclc;
 324
 325	/* Make sure the divisor is less than 2^32: */
 326	while (div >> 32) {
 327		sft++;
 328		div >>= 1;
 329	}
 330	tmp >>= sft;
 331	do_div(tmp, (u32) div);
 332	return dclc < 0 ? -tmp : tmp;
 333}
 334EXPORT_SYMBOL_GPL(__ktime_divns);
 335#endif /* BITS_PER_LONG >= 64 */
 336
 337/*
 338 * Add two ktime values and do a safety check for overflow:
 339 */
 340ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
 341{
 342	ktime_t res = ktime_add_unsafe(lhs, rhs);
 343
 344	/*
 345	 * We use KTIME_SEC_MAX here, the maximum timeout which we can
 346	 * return to user space in a timespec:
 347	 */
 348	if (res < 0 || res < lhs || res < rhs)
 349		res = ktime_set(KTIME_SEC_MAX, 0);
 350
 351	return res;
 352}
 353
 354EXPORT_SYMBOL_GPL(ktime_add_safe);
 355
 356#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
 357
 358static const struct debug_obj_descr hrtimer_debug_descr;
 359
 360static void *hrtimer_debug_hint(void *addr)
 361{
 362	return ACCESS_PRIVATE((struct hrtimer *)addr, function);
 363}
 364
 365/*
 366 * fixup_init is called when:
 367 * - an active object is initialized
 368 */
 369static bool hrtimer_fixup_init(void *addr, enum debug_obj_state state)
 370{
 371	struct hrtimer *timer = addr;
 372
 373	switch (state) {
 374	case ODEBUG_STATE_ACTIVE:
 375		hrtimer_cancel(timer);
 376		debug_object_init(timer, &hrtimer_debug_descr);
 377		return true;
 378	default:
 379		return false;
 380	}
 381}
 382
 383/*
 384 * fixup_activate is called when:
 385 * - an active object is activated
 386 * - an unknown non-static object is activated
 387 */
 388static bool hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
 389{
 390	switch (state) {
 391	case ODEBUG_STATE_ACTIVE:
 392		WARN_ON(1);
 393		fallthrough;
 394	default:
 395		return false;
 396	}
 397}
 398
 399/*
 400 * fixup_free is called when:
 401 * - an active object is freed
 402 */
 403static bool hrtimer_fixup_free(void *addr, enum debug_obj_state state)
 404{
 405	struct hrtimer *timer = addr;
 406
 407	switch (state) {
 408	case ODEBUG_STATE_ACTIVE:
 409		hrtimer_cancel(timer);
 410		debug_object_free(timer, &hrtimer_debug_descr);
 411		return true;
 412	default:
 413		return false;
 414	}
 415}
 416
 417static const struct debug_obj_descr hrtimer_debug_descr = {
 418	.name		= "hrtimer",
 419	.debug_hint	= hrtimer_debug_hint,
 420	.fixup_init	= hrtimer_fixup_init,
 421	.fixup_activate	= hrtimer_fixup_activate,
 422	.fixup_free	= hrtimer_fixup_free,
 423};
 424
 425static inline void debug_hrtimer_init(struct hrtimer *timer)
 426{
 427	debug_object_init(timer, &hrtimer_debug_descr);
 428}
 429
 430static inline void debug_hrtimer_init_on_stack(struct hrtimer *timer)
 431{
 432	debug_object_init_on_stack(timer, &hrtimer_debug_descr);
 433}
 434
 435static inline void debug_hrtimer_activate(struct hrtimer *timer,
 436					  enum hrtimer_mode mode)
 437{
 438	debug_object_activate(timer, &hrtimer_debug_descr);
 439}
 440
 441static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
 442{
 443	debug_object_deactivate(timer, &hrtimer_debug_descr);
 444}
 445
 446void destroy_hrtimer_on_stack(struct hrtimer *timer)
 447{
 448	debug_object_free(timer, &hrtimer_debug_descr);
 449}
 450EXPORT_SYMBOL_GPL(destroy_hrtimer_on_stack);
 451
 452#else
 453
 454static inline void debug_hrtimer_init(struct hrtimer *timer) { }
 455static inline void debug_hrtimer_init_on_stack(struct hrtimer *timer) { }
 456static inline void debug_hrtimer_activate(struct hrtimer *timer,
 457					  enum hrtimer_mode mode) { }
 458static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
 459#endif
 460
 461static inline void debug_setup(struct hrtimer *timer, clockid_t clockid, enum hrtimer_mode mode)
 462{
 463	debug_hrtimer_init(timer);
 464	trace_hrtimer_setup(timer, clockid, mode);
 465}
 466
 467static inline void debug_setup_on_stack(struct hrtimer *timer, clockid_t clockid,
 468					enum hrtimer_mode mode)
 469{
 470	debug_hrtimer_init_on_stack(timer);
 471	trace_hrtimer_setup(timer, clockid, mode);
 472}
 473
 474static inline void debug_activate(struct hrtimer *timer,
 475				  enum hrtimer_mode mode)
 476{
 477	debug_hrtimer_activate(timer, mode);
 478	trace_hrtimer_start(timer, mode);
 479}
 480
 481static inline void debug_deactivate(struct hrtimer *timer)
 482{
 483	debug_hrtimer_deactivate(timer);
 484	trace_hrtimer_cancel(timer);
 485}
 486
 487static struct hrtimer_clock_base *
 488__next_base(struct hrtimer_cpu_base *cpu_base, unsigned int *active)
 489{
 490	unsigned int idx;
 491
 492	if (!*active)
 493		return NULL;
 494
 495	idx = __ffs(*active);
 496	*active &= ~(1U << idx);
 497
 498	return &cpu_base->clock_base[idx];
 499}
 500
 501#define for_each_active_base(base, cpu_base, active)	\
 502	while ((base = __next_base((cpu_base), &(active))))
 503
 504static ktime_t __hrtimer_next_event_base(struct hrtimer_cpu_base *cpu_base,
 505					 const struct hrtimer *exclude,
 506					 unsigned int active,
 507					 ktime_t expires_next)
 508{
 509	struct hrtimer_clock_base *base;
 510	ktime_t expires;
 511
 512	for_each_active_base(base, cpu_base, active) {
 513		struct timerqueue_node *next;
 514		struct hrtimer *timer;
 515
 516		next = timerqueue_getnext(&base->active);
 517		timer = container_of(next, struct hrtimer, node);
 518		if (timer == exclude) {
 519			/* Get to the next timer in the queue. */
 520			next = timerqueue_iterate_next(next);
 521			if (!next)
 522				continue;
 523
 524			timer = container_of(next, struct hrtimer, node);
 525		}
 526		expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
 527		if (expires < expires_next) {
 528			expires_next = expires;
 529
 530			/* Skip cpu_base update if a timer is being excluded. */
 531			if (exclude)
 532				continue;
 533
 534			if (timer->is_soft)
 535				cpu_base->softirq_next_timer = timer;
 536			else
 537				cpu_base->next_timer = timer;
 538		}
 539	}
 540	/*
 541	 * clock_was_set() might have changed base->offset of any of
 542	 * the clock bases so the result might be negative. Fix it up
 543	 * to prevent a false positive in clockevents_program_event().
 544	 */
 545	if (expires_next < 0)
 546		expires_next = 0;
 547	return expires_next;
 548}
 549
 550/*
 551 * Recomputes cpu_base::*next_timer and returns the earliest expires_next
 552 * but does not set cpu_base::*expires_next, that is done by
 553 * hrtimer[_force]_reprogram and hrtimer_interrupt only. When updating
 554 * cpu_base::*expires_next right away, reprogramming logic would no longer
 555 * work.
 556 *
 557 * When a softirq is pending, we can ignore the HRTIMER_ACTIVE_SOFT bases,
 558 * those timers will get run whenever the softirq gets handled, at the end of
 559 * hrtimer_run_softirq(), hrtimer_update_softirq_timer() will re-add these bases.
 560 *
 561 * Therefore softirq values are those from the HRTIMER_ACTIVE_SOFT clock bases.
 562 * The !softirq values are the minima across HRTIMER_ACTIVE_ALL, unless an actual
 563 * softirq is pending, in which case they're the minima of HRTIMER_ACTIVE_HARD.
 564 *
 565 * @active_mask must be one of:
 566 *  - HRTIMER_ACTIVE_ALL,
 567 *  - HRTIMER_ACTIVE_SOFT, or
 568 *  - HRTIMER_ACTIVE_HARD.
 569 */
 570static ktime_t
 571__hrtimer_get_next_event(struct hrtimer_cpu_base *cpu_base, unsigned int active_mask)
 572{
 573	unsigned int active;
 574	struct hrtimer *next_timer = NULL;
 575	ktime_t expires_next = KTIME_MAX;
 576
 577	if (!cpu_base->softirq_activated && (active_mask & HRTIMER_ACTIVE_SOFT)) {
 578		active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
 579		cpu_base->softirq_next_timer = NULL;
 580		expires_next = __hrtimer_next_event_base(cpu_base, NULL,
 581							 active, KTIME_MAX);
 582
 583		next_timer = cpu_base->softirq_next_timer;
 584	}
 585
 586	if (active_mask & HRTIMER_ACTIVE_HARD) {
 587		active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
 588		cpu_base->next_timer = next_timer;
 589		expires_next = __hrtimer_next_event_base(cpu_base, NULL, active,
 590							 expires_next);
 591	}
 592
 593	return expires_next;
 594}
 595
 596static ktime_t hrtimer_update_next_event(struct hrtimer_cpu_base *cpu_base)
 597{
 598	ktime_t expires_next, soft = KTIME_MAX;
 599
 600	/*
 601	 * If the soft interrupt has already been activated, ignore the
 602	 * soft bases. They will be handled in the already raised soft
 603	 * interrupt.
 604	 */
 605	if (!cpu_base->softirq_activated) {
 606		soft = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
 607		/*
 608		 * Update the soft expiry time. clock_settime() might have
 609		 * affected it.
 610		 */
 611		cpu_base->softirq_expires_next = soft;
 612	}
 613
 614	expires_next = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_HARD);
 615	/*
 616	 * If a softirq timer is expiring first, update cpu_base->next_timer
 617	 * and program the hardware with the soft expiry time.
 618	 */
 619	if (expires_next > soft) {
 620		cpu_base->next_timer = cpu_base->softirq_next_timer;
 621		expires_next = soft;
 622	}
 623
 624	return expires_next;
 625}
 626
 627static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
 628{
 629	ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
 630	ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
 631	ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
 632
 633	ktime_t now = ktime_get_update_offsets_now(&base->clock_was_set_seq,
 634					    offs_real, offs_boot, offs_tai);
 635
 636	base->clock_base[HRTIMER_BASE_REALTIME_SOFT].offset = *offs_real;
 637	base->clock_base[HRTIMER_BASE_BOOTTIME_SOFT].offset = *offs_boot;
 638	base->clock_base[HRTIMER_BASE_TAI_SOFT].offset = *offs_tai;
 639
 640	return now;
 641}
 642
 643/*
 644 * Is the high resolution mode active ?
 645 */
 646static inline int hrtimer_hres_active(struct hrtimer_cpu_base *cpu_base)
 647{
 648	return IS_ENABLED(CONFIG_HIGH_RES_TIMERS) ?
 649		cpu_base->hres_active : 0;
 650}
 651
 652static void __hrtimer_reprogram(struct hrtimer_cpu_base *cpu_base,
 653				struct hrtimer *next_timer,
 654				ktime_t expires_next)
 655{
 656	cpu_base->expires_next = expires_next;
 657
 658	/*
 659	 * If hres is not active, hardware does not have to be
 660	 * reprogrammed yet.
 661	 *
 662	 * If a hang was detected in the last timer interrupt then we
 663	 * leave the hang delay active in the hardware. We want the
 664	 * system to make progress. That also prevents the following
 665	 * scenario:
 666	 * T1 expires 50ms from now
 667	 * T2 expires 5s from now
 668	 *
 669	 * T1 is removed, so this code is called and would reprogram
 670	 * the hardware to 5s from now. Any hrtimer_start after that
 671	 * will not reprogram the hardware due to hang_detected being
 672	 * set. So we'd effectively block all timers until the T2 event
 673	 * fires.
 674	 */
 675	if (!hrtimer_hres_active(cpu_base) || cpu_base->hang_detected)
 676		return;
 677
 678	tick_program_event(expires_next, 1);
 679}
 680
 681/*
 682 * Reprogram the event source with checking both queues for the
 683 * next event
 684 * Called with interrupts disabled and base->lock held
 685 */
 686static void
 687hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
 688{
 689	ktime_t expires_next;
 690
 691	expires_next = hrtimer_update_next_event(cpu_base);
 692
 693	if (skip_equal && expires_next == cpu_base->expires_next)
 694		return;
 695
 696	__hrtimer_reprogram(cpu_base, cpu_base->next_timer, expires_next);
 697}
 698
 699/* High resolution timer related functions */
 700#ifdef CONFIG_HIGH_RES_TIMERS
 701
 702/*
 703 * High resolution timer enabled ?
 704 */
 705static bool hrtimer_hres_enabled __read_mostly  = true;
 706unsigned int hrtimer_resolution __read_mostly = LOW_RES_NSEC;
 707EXPORT_SYMBOL_GPL(hrtimer_resolution);
 708
 709/*
 710 * Enable / Disable high resolution mode
 711 */
 712static int __init setup_hrtimer_hres(char *str)
 713{
 714	return (kstrtobool(str, &hrtimer_hres_enabled) == 0);
 715}
 716
 717__setup("highres=", setup_hrtimer_hres);
 718
 719/*
 720 * hrtimer_high_res_enabled - query, if the highres mode is enabled
 721 */
 722static inline int hrtimer_is_hres_enabled(void)
 723{
 724	return hrtimer_hres_enabled;
 725}
 726
 727/*
 728 * Switch to high resolution mode
 729 */
 730static void hrtimer_switch_to_hres(void)
 731{
 732	struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
 733
 734	if (tick_init_highres()) {
 735		pr_warn("Could not switch to high resolution mode on CPU %u\n",
 736			base->cpu);
 737		return;
 738	}
 739	base->hres_active = 1;
 740	hrtimer_resolution = HIGH_RES_NSEC;
 741
 742	tick_setup_sched_timer(true);
 743	/* "Retrigger" the interrupt to get things going */
 744	retrigger_next_event(NULL);
 745}
 746
 747#else
 748
 749static inline int hrtimer_is_hres_enabled(void) { return 0; }
 750static inline void hrtimer_switch_to_hres(void) { }
 751
 752#endif /* CONFIG_HIGH_RES_TIMERS */
 753/*
 754 * Retrigger next event is called after clock was set with interrupts
 755 * disabled through an SMP function call or directly from low level
 756 * resume code.
 757 *
 758 * This is only invoked when:
 759 *	- CONFIG_HIGH_RES_TIMERS is enabled.
 760 *	- CONFIG_NOHZ_COMMON is enabled
 761 *
 762 * For the other cases this function is empty and because the call sites
 763 * are optimized out it vanishes as well, i.e. no need for lots of
 764 * #ifdeffery.
 765 */
 766static void retrigger_next_event(void *arg)
 767{
 768	struct hrtimer_cpu_base *base = this_cpu_ptr(&hrtimer_bases);
 769
 770	/*
 771	 * When high resolution mode or nohz is active, then the offsets of
 772	 * CLOCK_REALTIME/TAI/BOOTTIME have to be updated. Otherwise the
 773	 * next tick will take care of that.
 774	 *
 775	 * If high resolution mode is active then the next expiring timer
 776	 * must be reevaluated and the clock event device reprogrammed if
 777	 * necessary.
 778	 *
 779	 * In the NOHZ case the update of the offset and the reevaluation
 780	 * of the next expiring timer is enough. The return from the SMP
 781	 * function call will take care of the reprogramming in case the
 782	 * CPU was in a NOHZ idle sleep.
 783	 *
 784	 * In periodic low resolution mode, the next softirq expiration
 785	 * must also be updated.
 786	 */
 787	raw_spin_lock(&base->lock);
 788	hrtimer_update_base(base);
 789	if (hrtimer_hres_active(base))
 790		hrtimer_force_reprogram(base, 0);
 791	else
 792		hrtimer_update_next_event(base);
 793	raw_spin_unlock(&base->lock);
 794}
 795
 796/*
 797 * When a timer is enqueued and expires earlier than the already enqueued
 798 * timers, we have to check, whether it expires earlier than the timer for
 799 * which the clock event device was armed.
 800 *
 801 * Called with interrupts disabled and base->cpu_base.lock held
 802 */
 803static void hrtimer_reprogram(struct hrtimer *timer, bool reprogram)
 804{
 805	struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
 806	struct hrtimer_clock_base *base = timer->base;
 807	ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
 808
 809	WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
 810
 811	/*
 812	 * CLOCK_REALTIME timer might be requested with an absolute
 813	 * expiry time which is less than base->offset. Set it to 0.
 814	 */
 815	if (expires < 0)
 816		expires = 0;
 817
 818	if (timer->is_soft) {
 819		/*
 820		 * soft hrtimer could be started on a remote CPU. In this
 821		 * case softirq_expires_next needs to be updated on the
 822		 * remote CPU. The soft hrtimer will not expire before the
 823		 * first hard hrtimer on the remote CPU -
 824		 * hrtimer_check_target() prevents this case.
 825		 */
 826		struct hrtimer_cpu_base *timer_cpu_base = base->cpu_base;
 827
 828		if (timer_cpu_base->softirq_activated)
 829			return;
 830
 831		if (!ktime_before(expires, timer_cpu_base->softirq_expires_next))
 832			return;
 833
 834		timer_cpu_base->softirq_next_timer = timer;
 835		timer_cpu_base->softirq_expires_next = expires;
 836
 837		if (!ktime_before(expires, timer_cpu_base->expires_next) ||
 838		    !reprogram)
 839			return;
 840	}
 841
 842	/*
 843	 * If the timer is not on the current cpu, we cannot reprogram
 844	 * the other cpus clock event device.
 845	 */
 846	if (base->cpu_base != cpu_base)
 847		return;
 848
 849	if (expires >= cpu_base->expires_next)
 850		return;
 851
 852	/*
 853	 * If the hrtimer interrupt is running, then it will reevaluate the
 854	 * clock bases and reprogram the clock event device.
 855	 */
 856	if (cpu_base->in_hrtirq)
 857		return;
 858
 859	cpu_base->next_timer = timer;
 860
 861	__hrtimer_reprogram(cpu_base, timer, expires);
 862}
 863
 864static bool update_needs_ipi(struct hrtimer_cpu_base *cpu_base,
 865			     unsigned int active)
 866{
 867	struct hrtimer_clock_base *base;
 868	unsigned int seq;
 869	ktime_t expires;
 870
 871	/*
 872	 * Update the base offsets unconditionally so the following
 873	 * checks whether the SMP function call is required works.
 874	 *
 875	 * The update is safe even when the remote CPU is in the hrtimer
 876	 * interrupt or the hrtimer soft interrupt and expiring affected
 877	 * bases. Either it will see the update before handling a base or
 878	 * it will see it when it finishes the processing and reevaluates
 879	 * the next expiring timer.
 880	 */
 881	seq = cpu_base->clock_was_set_seq;
 882	hrtimer_update_base(cpu_base);
 883
 884	/*
 885	 * If the sequence did not change over the update then the
 886	 * remote CPU already handled it.
 887	 */
 888	if (seq == cpu_base->clock_was_set_seq)
 889		return false;
 890
 891	/*
 892	 * If the remote CPU is currently handling an hrtimer interrupt, it
 893	 * will reevaluate the first expiring timer of all clock bases
 894	 * before reprogramming. Nothing to do here.
 895	 */
 896	if (cpu_base->in_hrtirq)
 897		return false;
 898
 899	/*
 900	 * Walk the affected clock bases and check whether the first expiring
 901	 * timer in a clock base is moving ahead of the first expiring timer of
 902	 * @cpu_base. If so, the IPI must be invoked because per CPU clock
 903	 * event devices cannot be remotely reprogrammed.
 904	 */
 905	active &= cpu_base->active_bases;
 906
 907	for_each_active_base(base, cpu_base, active) {
 908		struct timerqueue_node *next;
 909
 910		next = timerqueue_getnext(&base->active);
 911		expires = ktime_sub(next->expires, base->offset);
 912		if (expires < cpu_base->expires_next)
 913			return true;
 914
 915		/* Extra check for softirq clock bases */
 916		if (base->index < HRTIMER_BASE_MONOTONIC_SOFT)
 917			continue;
 918		if (cpu_base->softirq_activated)
 919			continue;
 920		if (expires < cpu_base->softirq_expires_next)
 921			return true;
 922	}
 923	return false;
 924}
 925
 926/*
 927 * Clock was set. This might affect CLOCK_REALTIME, CLOCK_TAI and
 928 * CLOCK_BOOTTIME (for late sleep time injection).
 929 *
 930 * This requires to update the offsets for these clocks
 931 * vs. CLOCK_MONOTONIC. When high resolution timers are enabled, then this
 932 * also requires to eventually reprogram the per CPU clock event devices
 933 * when the change moves an affected timer ahead of the first expiring
 934 * timer on that CPU. Obviously remote per CPU clock event devices cannot
 935 * be reprogrammed. The other reason why an IPI has to be sent is when the
 936 * system is in !HIGH_RES and NOHZ mode. The NOHZ mode updates the offsets
 937 * in the tick, which obviously might be stopped, so this has to bring out
 938 * the remote CPU which might sleep in idle to get this sorted.
 939 */
 940void clock_was_set(unsigned int bases)
 941{
 942	struct hrtimer_cpu_base *cpu_base = raw_cpu_ptr(&hrtimer_bases);
 943	cpumask_var_t mask;
 944	int cpu;
 945
 946	if (!hrtimer_hres_active(cpu_base) && !tick_nohz_active)
 947		goto out_timerfd;
 948
 949	if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
 950		on_each_cpu(retrigger_next_event, NULL, 1);
 951		goto out_timerfd;
 952	}
 953
 954	/* Avoid interrupting CPUs if possible */
 955	cpus_read_lock();
 956	for_each_online_cpu(cpu) {
 957		unsigned long flags;
 958
 959		cpu_base = &per_cpu(hrtimer_bases, cpu);
 960		raw_spin_lock_irqsave(&cpu_base->lock, flags);
 961
 962		if (update_needs_ipi(cpu_base, bases))
 963			cpumask_set_cpu(cpu, mask);
 964
 965		raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
 966	}
 967
 968	preempt_disable();
 969	smp_call_function_many(mask, retrigger_next_event, NULL, 1);
 970	preempt_enable();
 971	cpus_read_unlock();
 972	free_cpumask_var(mask);
 973
 974out_timerfd:
 975	timerfd_clock_was_set();
 976}
 977
 978static void clock_was_set_work(struct work_struct *work)
 979{
 980	clock_was_set(CLOCK_SET_WALL);
 981}
 982
 983static DECLARE_WORK(hrtimer_work, clock_was_set_work);
 984
 985/*
 986 * Called from timekeeping code to reprogram the hrtimer interrupt device
 987 * on all cpus and to notify timerfd.
 988 */
 989void clock_was_set_delayed(void)
 990{
 991	schedule_work(&hrtimer_work);
 992}
 993
 994/*
 995 * Called during resume either directly from via timekeeping_resume()
 996 * or in the case of s2idle from tick_unfreeze() to ensure that the
 997 * hrtimers are up to date.
 998 */
 999void hrtimers_resume_local(void)
1000{
1001	lockdep_assert_irqs_disabled();
1002	/* Retrigger on the local CPU */
1003	retrigger_next_event(NULL);
1004}
1005
1006/*
1007 * Counterpart to lock_hrtimer_base above:
1008 */
1009static inline
1010void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
1011	__releases(&timer->base->cpu_base->lock)
1012{
1013	raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
1014}
1015
1016/**
1017 * hrtimer_forward() - forward the timer expiry
1018 * @timer:	hrtimer to forward
1019 * @now:	forward past this time
1020 * @interval:	the interval to forward
1021 *
1022 * Forward the timer expiry so it will expire in the future.
1023 *
1024 * .. note::
1025 *  This only updates the timer expiry value and does not requeue the timer.
1026 *
1027 * There is also a variant of the function hrtimer_forward_now().
1028 *
1029 * Context: Can be safely called from the callback function of @timer. If called
1030 *          from other contexts @timer must neither be enqueued nor running the
1031 *          callback and the caller needs to take care of serialization.
1032 *
1033 * Return: The number of overruns are returned.
1034 */
1035u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
1036{
1037	u64 orun = 1;
1038	ktime_t delta;
1039
1040	delta = ktime_sub(now, hrtimer_get_expires(timer));
1041
1042	if (delta < 0)
1043		return 0;
1044
1045	if (WARN_ON(timer->state & HRTIMER_STATE_ENQUEUED))
1046		return 0;
1047
1048	if (interval < hrtimer_resolution)
1049		interval = hrtimer_resolution;
1050
1051	if (unlikely(delta >= interval)) {
1052		s64 incr = ktime_to_ns(interval);
1053
1054		orun = ktime_divns(delta, incr);
1055		hrtimer_add_expires_ns(timer, incr * orun);
1056		if (hrtimer_get_expires_tv64(timer) > now)
1057			return orun;
1058		/*
1059		 * This (and the ktime_add() below) is the
1060		 * correction for exact:
1061		 */
1062		orun++;
1063	}
1064	hrtimer_add_expires(timer, interval);
1065
1066	return orun;
1067}
1068EXPORT_SYMBOL_GPL(hrtimer_forward);
1069
1070/*
1071 * enqueue_hrtimer - internal function to (re)start a timer
1072 *
1073 * The timer is inserted in expiry order. Insertion into the
1074 * red black tree is O(log(n)). Must hold the base lock.
1075 *
1076 * Returns true when the new timer is the leftmost timer in the tree.
1077 */
1078static bool enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base,
1079			    enum hrtimer_mode mode)
1080{
1081	debug_activate(timer, mode);
1082	WARN_ON_ONCE(!base->cpu_base->online);
1083
1084	base->cpu_base->active_bases |= 1 << base->index;
1085
1086	/* Pairs with the lockless read in hrtimer_is_queued() */
1087	WRITE_ONCE(timer->state, HRTIMER_STATE_ENQUEUED);
1088
1089	return timerqueue_add(&base->active, &timer->node);
1090}
1091
1092/*
1093 * __remove_hrtimer - internal function to remove a timer
1094 *
1095 * Caller must hold the base lock.
1096 *
1097 * High resolution timer mode reprograms the clock event device when the
1098 * timer is the one which expires next. The caller can disable this by setting
1099 * reprogram to zero. This is useful, when the context does a reprogramming
1100 * anyway (e.g. timer interrupt)
1101 */
1102static void __remove_hrtimer(struct hrtimer *timer,
1103			     struct hrtimer_clock_base *base,
1104			     u8 newstate, int reprogram)
1105{
1106	struct hrtimer_cpu_base *cpu_base = base->cpu_base;
1107	u8 state = timer->state;
1108
1109	/* Pairs with the lockless read in hrtimer_is_queued() */
1110	WRITE_ONCE(timer->state, newstate);
1111	if (!(state & HRTIMER_STATE_ENQUEUED))
1112		return;
1113
1114	if (!timerqueue_del(&base->active, &timer->node))
1115		cpu_base->active_bases &= ~(1 << base->index);
1116
1117	/*
1118	 * Note: If reprogram is false we do not update
1119	 * cpu_base->next_timer. This happens when we remove the first
1120	 * timer on a remote cpu. No harm as we never dereference
1121	 * cpu_base->next_timer. So the worst thing what can happen is
1122	 * an superfluous call to hrtimer_force_reprogram() on the
1123	 * remote cpu later on if the same timer gets enqueued again.
1124	 */
1125	if (reprogram && timer == cpu_base->next_timer)
1126		hrtimer_force_reprogram(cpu_base, 1);
1127}
1128
1129/*
1130 * remove hrtimer, called with base lock held
1131 */
1132static inline int
1133remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base,
1134	       bool restart, bool keep_local)
1135{
1136	u8 state = timer->state;
1137
1138	if (state & HRTIMER_STATE_ENQUEUED) {
1139		bool reprogram;
1140
1141		/*
1142		 * Remove the timer and force reprogramming when high
1143		 * resolution mode is active and the timer is on the current
1144		 * CPU. If we remove a timer on another CPU, reprogramming is
1145		 * skipped. The interrupt event on this CPU is fired and
1146		 * reprogramming happens in the interrupt handler. This is a
1147		 * rare case and less expensive than a smp call.
1148		 */
1149		debug_deactivate(timer);
1150		reprogram = base->cpu_base == this_cpu_ptr(&hrtimer_bases);
1151
1152		/*
1153		 * If the timer is not restarted then reprogramming is
1154		 * required if the timer is local. If it is local and about
1155		 * to be restarted, avoid programming it twice (on removal
1156		 * and a moment later when it's requeued).
1157		 */
1158		if (!restart)
1159			state = HRTIMER_STATE_INACTIVE;
1160		else
1161			reprogram &= !keep_local;
1162
1163		__remove_hrtimer(timer, base, state, reprogram);
1164		return 1;
1165	}
1166	return 0;
1167}
1168
1169static inline ktime_t hrtimer_update_lowres(struct hrtimer *timer, ktime_t tim,
1170					    const enum hrtimer_mode mode)
1171{
1172#ifdef CONFIG_TIME_LOW_RES
1173	/*
1174	 * CONFIG_TIME_LOW_RES indicates that the system has no way to return
1175	 * granular time values. For relative timers we add hrtimer_resolution
1176	 * (i.e. one jiffy) to prevent short timeouts.
1177	 */
1178	timer->is_rel = mode & HRTIMER_MODE_REL;
1179	if (timer->is_rel)
1180		tim = ktime_add_safe(tim, hrtimer_resolution);
1181#endif
1182	return tim;
1183}
1184
1185static void
1186hrtimer_update_softirq_timer(struct hrtimer_cpu_base *cpu_base, bool reprogram)
1187{
1188	ktime_t expires;
1189
1190	/*
1191	 * Find the next SOFT expiration.
1192	 */
1193	expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_SOFT);
1194
1195	/*
1196	 * reprogramming needs to be triggered, even if the next soft
1197	 * hrtimer expires at the same time than the next hard
1198	 * hrtimer. cpu_base->softirq_expires_next needs to be updated!
1199	 */
1200	if (expires == KTIME_MAX)
1201		return;
1202
1203	/*
1204	 * cpu_base->*next_timer is recomputed by __hrtimer_get_next_event()
1205	 * cpu_base->*expires_next is only set by hrtimer_reprogram()
1206	 */
1207	hrtimer_reprogram(cpu_base->softirq_next_timer, reprogram);
1208}
1209
1210static int __hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1211				    u64 delta_ns, const enum hrtimer_mode mode,
1212				    struct hrtimer_clock_base *base)
1213{
1214	struct hrtimer_cpu_base *this_cpu_base = this_cpu_ptr(&hrtimer_bases);
1215	struct hrtimer_clock_base *new_base;
1216	bool force_local, first;
1217
1218	/*
1219	 * If the timer is on the local cpu base and is the first expiring
1220	 * timer then this might end up reprogramming the hardware twice
1221	 * (on removal and on enqueue). To avoid that by prevent the
1222	 * reprogram on removal, keep the timer local to the current CPU
1223	 * and enforce reprogramming after it is queued no matter whether
1224	 * it is the new first expiring timer again or not.
1225	 */
1226	force_local = base->cpu_base == this_cpu_base;
1227	force_local &= base->cpu_base->next_timer == timer;
1228
1229	/*
1230	 * Don't force local queuing if this enqueue happens on a unplugged
1231	 * CPU after hrtimer_cpu_dying() has been invoked.
1232	 */
1233	force_local &= this_cpu_base->online;
1234
1235	/*
1236	 * Remove an active timer from the queue. In case it is not queued
1237	 * on the current CPU, make sure that remove_hrtimer() updates the
1238	 * remote data correctly.
1239	 *
1240	 * If it's on the current CPU and the first expiring timer, then
1241	 * skip reprogramming, keep the timer local and enforce
1242	 * reprogramming later if it was the first expiring timer.  This
1243	 * avoids programming the underlying clock event twice (once at
1244	 * removal and once after enqueue).
1245	 */
1246	remove_hrtimer(timer, base, true, force_local);
1247
1248	if (mode & HRTIMER_MODE_REL)
1249		tim = ktime_add_safe(tim, __hrtimer_cb_get_time(base->clockid));
1250
1251	tim = hrtimer_update_lowres(timer, tim, mode);
1252
1253	hrtimer_set_expires_range_ns(timer, tim, delta_ns);
1254
1255	/* Switch the timer base, if necessary: */
1256	if (!force_local) {
1257		new_base = switch_hrtimer_base(timer, base,
1258					       mode & HRTIMER_MODE_PINNED);
1259	} else {
1260		new_base = base;
1261	}
1262
1263	first = enqueue_hrtimer(timer, new_base, mode);
1264	if (!force_local) {
1265		/*
1266		 * If the current CPU base is online, then the timer is
1267		 * never queued on a remote CPU if it would be the first
1268		 * expiring timer there.
1269		 */
1270		if (hrtimer_base_is_online(this_cpu_base))
1271			return first;
1272
1273		/*
1274		 * Timer was enqueued remote because the current base is
1275		 * already offline. If the timer is the first to expire,
1276		 * kick the remote CPU to reprogram the clock event.
1277		 */
1278		if (first) {
1279			struct hrtimer_cpu_base *new_cpu_base = new_base->cpu_base;
1280
1281			smp_call_function_single_async(new_cpu_base->cpu, &new_cpu_base->csd);
1282		}
1283		return 0;
1284	}
1285
1286	/*
1287	 * Timer was forced to stay on the current CPU to avoid
1288	 * reprogramming on removal and enqueue. Force reprogram the
1289	 * hardware by evaluating the new first expiring timer.
1290	 */
1291	hrtimer_force_reprogram(new_base->cpu_base, 1);
1292	return 0;
1293}
1294
1295/**
1296 * hrtimer_start_range_ns - (re)start an hrtimer
1297 * @timer:	the timer to be added
1298 * @tim:	expiry time
1299 * @delta_ns:	"slack" range for the timer
1300 * @mode:	timer mode: absolute (HRTIMER_MODE_ABS) or
1301 *		relative (HRTIMER_MODE_REL), and pinned (HRTIMER_MODE_PINNED);
1302 *		softirq based mode is considered for debug purpose only!
1303 */
1304void hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1305			    u64 delta_ns, const enum hrtimer_mode mode)
1306{
1307	struct hrtimer_clock_base *base;
1308	unsigned long flags;
1309
1310	/*
1311	 * Check whether the HRTIMER_MODE_SOFT bit and hrtimer.is_soft
1312	 * match on CONFIG_PREEMPT_RT = n. With PREEMPT_RT check the hard
1313	 * expiry mode because unmarked timers are moved to softirq expiry.
1314	 */
1315	if (!IS_ENABLED(CONFIG_PREEMPT_RT))
1316		WARN_ON_ONCE(!(mode & HRTIMER_MODE_SOFT) ^ !timer->is_soft);
1317	else
1318		WARN_ON_ONCE(!(mode & HRTIMER_MODE_HARD) ^ !timer->is_hard);
1319
1320	base = lock_hrtimer_base(timer, &flags);
1321
1322	if (__hrtimer_start_range_ns(timer, tim, delta_ns, mode, base))
1323		hrtimer_reprogram(timer, true);
1324
1325	unlock_hrtimer_base(timer, &flags);
1326}
1327EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1328
1329/**
1330 * hrtimer_try_to_cancel - try to deactivate a timer
1331 * @timer:	hrtimer to stop
1332 *
1333 * Returns:
1334 *
1335 *  *  0 when the timer was not active
1336 *  *  1 when the timer was active
1337 *  * -1 when the timer is currently executing the callback function and
1338 *    cannot be stopped
1339 */
1340int hrtimer_try_to_cancel(struct hrtimer *timer)
1341{
1342	struct hrtimer_clock_base *base;
1343	unsigned long flags;
1344	int ret = -1;
1345
1346	/*
1347	 * Check lockless first. If the timer is not active (neither
1348	 * enqueued nor running the callback, nothing to do here.  The
1349	 * base lock does not serialize against a concurrent enqueue,
1350	 * so we can avoid taking it.
1351	 */
1352	if (!hrtimer_active(timer))
1353		return 0;
1354
1355	base = lock_hrtimer_base(timer, &flags);
1356
1357	if (!hrtimer_callback_running(timer))
1358		ret = remove_hrtimer(timer, base, false, false);
1359
1360	unlock_hrtimer_base(timer, &flags);
1361
1362	return ret;
1363
1364}
1365EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1366
1367#ifdef CONFIG_PREEMPT_RT
1368static void hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base)
1369{
1370	spin_lock_init(&base->softirq_expiry_lock);
1371}
1372
1373static void hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base)
1374	__acquires(&base->softirq_expiry_lock)
1375{
1376	spin_lock(&base->softirq_expiry_lock);
1377}
1378
1379static void hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base)
1380	__releases(&base->softirq_expiry_lock)
1381{
1382	spin_unlock(&base->softirq_expiry_lock);
1383}
1384
1385/*
1386 * The counterpart to hrtimer_cancel_wait_running().
1387 *
1388 * If there is a waiter for cpu_base->expiry_lock, then it was waiting for
1389 * the timer callback to finish. Drop expiry_lock and reacquire it. That
1390 * allows the waiter to acquire the lock and make progress.
1391 */
1392static void hrtimer_sync_wait_running(struct hrtimer_cpu_base *cpu_base,
1393				      unsigned long flags)
1394{
1395	if (atomic_read(&cpu_base->timer_waiters)) {
1396		raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1397		spin_unlock(&cpu_base->softirq_expiry_lock);
1398		spin_lock(&cpu_base->softirq_expiry_lock);
1399		raw_spin_lock_irq(&cpu_base->lock);
1400	}
1401}
1402
1403#ifdef CONFIG_SMP
1404static __always_inline bool is_migration_base(struct hrtimer_clock_base *base)
1405{
1406	return base == &migration_base;
1407}
1408#else
1409static __always_inline bool is_migration_base(struct hrtimer_clock_base *base)
1410{
1411	return false;
1412}
1413#endif
1414
1415/*
1416 * This function is called on PREEMPT_RT kernels when the fast path
1417 * deletion of a timer failed because the timer callback function was
1418 * running.
1419 *
1420 * This prevents priority inversion: if the soft irq thread is preempted
1421 * in the middle of a timer callback, then calling hrtimer_cancel() can
1422 * lead to two issues:
1423 *
1424 *  - If the caller is on a remote CPU then it has to spin wait for the timer
1425 *    handler to complete. This can result in unbound priority inversion.
1426 *
1427 *  - If the caller originates from the task which preempted the timer
1428 *    handler on the same CPU, then spin waiting for the timer handler to
1429 *    complete is never going to end.
1430 */
1431void hrtimer_cancel_wait_running(const struct hrtimer *timer)
1432{
1433	/* Lockless read. Prevent the compiler from reloading it below */
1434	struct hrtimer_clock_base *base = READ_ONCE(timer->base);
1435
1436	/*
1437	 * Just relax if the timer expires in hard interrupt context or if
1438	 * it is currently on the migration base.
1439	 */
1440	if (!timer->is_soft || is_migration_base(base)) {
1441		cpu_relax();
1442		return;
1443	}
1444
1445	/*
1446	 * Mark the base as contended and grab the expiry lock, which is
1447	 * held by the softirq across the timer callback. Drop the lock
1448	 * immediately so the softirq can expire the next timer. In theory
1449	 * the timer could already be running again, but that's more than
1450	 * unlikely and just causes another wait loop.
1451	 */
1452	atomic_inc(&base->cpu_base->timer_waiters);
1453	spin_lock_bh(&base->cpu_base->softirq_expiry_lock);
1454	atomic_dec(&base->cpu_base->timer_waiters);
1455	spin_unlock_bh(&base->cpu_base->softirq_expiry_lock);
1456}
1457#else
1458static inline void
1459hrtimer_cpu_base_init_expiry_lock(struct hrtimer_cpu_base *base) { }
1460static inline void
1461hrtimer_cpu_base_lock_expiry(struct hrtimer_cpu_base *base) { }
1462static inline void
1463hrtimer_cpu_base_unlock_expiry(struct hrtimer_cpu_base *base) { }
1464static inline void hrtimer_sync_wait_running(struct hrtimer_cpu_base *base,
1465					     unsigned long flags) { }
1466#endif
1467
1468/**
1469 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1470 * @timer:	the timer to be cancelled
1471 *
1472 * Returns:
1473 *  0 when the timer was not active
1474 *  1 when the timer was active
1475 */
1476int hrtimer_cancel(struct hrtimer *timer)
1477{
1478	int ret;
1479
1480	do {
1481		ret = hrtimer_try_to_cancel(timer);
1482
1483		if (ret < 0)
1484			hrtimer_cancel_wait_running(timer);
1485	} while (ret < 0);
1486	return ret;
1487}
1488EXPORT_SYMBOL_GPL(hrtimer_cancel);
1489
1490/**
1491 * __hrtimer_get_remaining - get remaining time for the timer
1492 * @timer:	the timer to read
1493 * @adjust:	adjust relative timers when CONFIG_TIME_LOW_RES=y
1494 */
1495ktime_t __hrtimer_get_remaining(const struct hrtimer *timer, bool adjust)
1496{
1497	unsigned long flags;
1498	ktime_t rem;
1499
1500	lock_hrtimer_base(timer, &flags);
1501	if (IS_ENABLED(CONFIG_TIME_LOW_RES) && adjust)
1502		rem = hrtimer_expires_remaining_adjusted(timer);
1503	else
1504		rem = hrtimer_expires_remaining(timer);
1505	unlock_hrtimer_base(timer, &flags);
1506
1507	return rem;
1508}
1509EXPORT_SYMBOL_GPL(__hrtimer_get_remaining);
1510
1511#ifdef CONFIG_NO_HZ_COMMON
1512/**
1513 * hrtimer_get_next_event - get the time until next expiry event
1514 *
1515 * Returns the next expiry time or KTIME_MAX if no timer is pending.
1516 */
1517u64 hrtimer_get_next_event(void)
1518{
1519	struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1520	u64 expires = KTIME_MAX;
1521	unsigned long flags;
1522
1523	raw_spin_lock_irqsave(&cpu_base->lock, flags);
1524
1525	if (!hrtimer_hres_active(cpu_base))
1526		expires = __hrtimer_get_next_event(cpu_base, HRTIMER_ACTIVE_ALL);
1527
1528	raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1529
1530	return expires;
1531}
1532
1533/**
1534 * hrtimer_next_event_without - time until next expiry event w/o one timer
1535 * @exclude:	timer to exclude
1536 *
1537 * Returns the next expiry time over all timers except for the @exclude one or
1538 * KTIME_MAX if none of them is pending.
1539 */
1540u64 hrtimer_next_event_without(const struct hrtimer *exclude)
1541{
1542	struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1543	u64 expires = KTIME_MAX;
1544	unsigned long flags;
1545
1546	raw_spin_lock_irqsave(&cpu_base->lock, flags);
1547
1548	if (hrtimer_hres_active(cpu_base)) {
1549		unsigned int active;
1550
1551		if (!cpu_base->softirq_activated) {
1552			active = cpu_base->active_bases & HRTIMER_ACTIVE_SOFT;
1553			expires = __hrtimer_next_event_base(cpu_base, exclude,
1554							    active, KTIME_MAX);
1555		}
1556		active = cpu_base->active_bases & HRTIMER_ACTIVE_HARD;
1557		expires = __hrtimer_next_event_base(cpu_base, exclude, active,
1558						    expires);
1559	}
1560
1561	raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1562
1563	return expires;
1564}
1565#endif
1566
1567static inline int hrtimer_clockid_to_base(clockid_t clock_id)
1568{
1569	switch (clock_id) {
1570	case CLOCK_MONOTONIC:
1571		return HRTIMER_BASE_MONOTONIC;
1572	case CLOCK_REALTIME:
1573		return HRTIMER_BASE_REALTIME;
1574	case CLOCK_BOOTTIME:
1575		return HRTIMER_BASE_BOOTTIME;
1576	case CLOCK_TAI:
1577		return HRTIMER_BASE_TAI;
1578	default:
1579		WARN(1, "Invalid clockid %d. Using MONOTONIC\n", clock_id);
1580		return HRTIMER_BASE_MONOTONIC;
1581	}
1582}
1583
1584static ktime_t __hrtimer_cb_get_time(clockid_t clock_id)
1585{
1586	switch (clock_id) {
1587	case CLOCK_MONOTONIC:
1588		return ktime_get();
1589	case CLOCK_REALTIME:
1590		return ktime_get_real();
1591	case CLOCK_BOOTTIME:
1592		return ktime_get_boottime();
1593	case CLOCK_TAI:
1594		return ktime_get_clocktai();
1595	default:
1596		WARN(1, "Invalid clockid %d. Using MONOTONIC\n", clock_id);
1597		return ktime_get();
1598	}
1599}
1600
1601ktime_t hrtimer_cb_get_time(const struct hrtimer *timer)
1602{
1603	return __hrtimer_cb_get_time(timer->base->clockid);
1604}
1605EXPORT_SYMBOL_GPL(hrtimer_cb_get_time);
1606
1607static void __hrtimer_setup(struct hrtimer *timer,
1608			    enum hrtimer_restart (*function)(struct hrtimer *),
1609			    clockid_t clock_id, enum hrtimer_mode mode)
1610{
1611	bool softtimer = !!(mode & HRTIMER_MODE_SOFT);
1612	struct hrtimer_cpu_base *cpu_base;
1613	int base;
1614
1615	/*
1616	 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
1617	 * marked for hard interrupt expiry mode are moved into soft
1618	 * interrupt context for latency reasons and because the callbacks
1619	 * can invoke functions which might sleep on RT, e.g. spin_lock().
1620	 */
1621	if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(mode & HRTIMER_MODE_HARD))
1622		softtimer = true;
1623
1624	memset(timer, 0, sizeof(struct hrtimer));
1625
1626	cpu_base = raw_cpu_ptr(&hrtimer_bases);
1627
1628	/*
1629	 * POSIX magic: Relative CLOCK_REALTIME timers are not affected by
1630	 * clock modifications, so they needs to become CLOCK_MONOTONIC to
1631	 * ensure POSIX compliance.
1632	 */
1633	if (clock_id == CLOCK_REALTIME && mode & HRTIMER_MODE_REL)
1634		clock_id = CLOCK_MONOTONIC;
1635
1636	base = softtimer ? HRTIMER_MAX_CLOCK_BASES / 2 : 0;
1637	base += hrtimer_clockid_to_base(clock_id);
1638	timer->is_soft = softtimer;
1639	timer->is_hard = !!(mode & HRTIMER_MODE_HARD);
1640	timer->base = &cpu_base->clock_base[base];
1641	timerqueue_init(&timer->node);
1642
1643	if (WARN_ON_ONCE(!function))
1644		ACCESS_PRIVATE(timer, function) = hrtimer_dummy_timeout;
1645	else
1646		ACCESS_PRIVATE(timer, function) = function;
1647}
1648
1649/**
1650 * hrtimer_setup - initialize a timer to the given clock
1651 * @timer:	the timer to be initialized
1652 * @function:	the callback function
1653 * @clock_id:	the clock to be used
1654 * @mode:       The modes which are relevant for initialization:
1655 *              HRTIMER_MODE_ABS, HRTIMER_MODE_REL, HRTIMER_MODE_ABS_SOFT,
1656 *              HRTIMER_MODE_REL_SOFT
1657 *
1658 *              The PINNED variants of the above can be handed in,
1659 *              but the PINNED bit is ignored as pinning happens
1660 *              when the hrtimer is started
1661 */
1662void hrtimer_setup(struct hrtimer *timer, enum hrtimer_restart (*function)(struct hrtimer *),
1663		   clockid_t clock_id, enum hrtimer_mode mode)
1664{
1665	debug_setup(timer, clock_id, mode);
1666	__hrtimer_setup(timer, function, clock_id, mode);
1667}
1668EXPORT_SYMBOL_GPL(hrtimer_setup);
1669
1670/**
1671 * hrtimer_setup_on_stack - initialize a timer on stack memory
1672 * @timer:	The timer to be initialized
1673 * @function:	the callback function
1674 * @clock_id:	The clock to be used
1675 * @mode:       The timer mode
1676 *
1677 * Similar to hrtimer_setup(), except that this one must be used if struct hrtimer is in stack
1678 * memory.
1679 */
1680void hrtimer_setup_on_stack(struct hrtimer *timer,
1681			    enum hrtimer_restart (*function)(struct hrtimer *),
1682			    clockid_t clock_id, enum hrtimer_mode mode)
1683{
1684	debug_setup_on_stack(timer, clock_id, mode);
1685	__hrtimer_setup(timer, function, clock_id, mode);
1686}
1687EXPORT_SYMBOL_GPL(hrtimer_setup_on_stack);
1688
1689/*
1690 * A timer is active, when it is enqueued into the rbtree or the
1691 * callback function is running or it's in the state of being migrated
1692 * to another cpu.
1693 *
1694 * It is important for this function to not return a false negative.
1695 */
1696bool hrtimer_active(const struct hrtimer *timer)
1697{
1698	struct hrtimer_clock_base *base;
1699	unsigned int seq;
1700
1701	do {
1702		base = READ_ONCE(timer->base);
1703		seq = raw_read_seqcount_begin(&base->seq);
1704
1705		if (timer->state != HRTIMER_STATE_INACTIVE ||
1706		    base->running == timer)
1707			return true;
1708
1709	} while (read_seqcount_retry(&base->seq, seq) ||
1710		 base != READ_ONCE(timer->base));
1711
1712	return false;
1713}
1714EXPORT_SYMBOL_GPL(hrtimer_active);
1715
1716/*
1717 * The write_seqcount_barrier()s in __run_hrtimer() split the thing into 3
1718 * distinct sections:
1719 *
1720 *  - queued:	the timer is queued
1721 *  - callback:	the timer is being ran
1722 *  - post:	the timer is inactive or (re)queued
1723 *
1724 * On the read side we ensure we observe timer->state and cpu_base->running
1725 * from the same section, if anything changed while we looked at it, we retry.
1726 * This includes timer->base changing because sequence numbers alone are
1727 * insufficient for that.
1728 *
1729 * The sequence numbers are required because otherwise we could still observe
1730 * a false negative if the read side got smeared over multiple consecutive
1731 * __run_hrtimer() invocations.
1732 */
1733
1734static void __run_hrtimer(struct hrtimer_cpu_base *cpu_base,
1735			  struct hrtimer_clock_base *base,
1736			  struct hrtimer *timer, ktime_t *now,
1737			  unsigned long flags) __must_hold(&cpu_base->lock)
1738{
1739	enum hrtimer_restart (*fn)(struct hrtimer *);
1740	bool expires_in_hardirq;
1741	int restart;
1742
1743	lockdep_assert_held(&cpu_base->lock);
1744
1745	debug_hrtimer_deactivate(timer);
1746	base->running = timer;
1747
1748	/*
1749	 * Separate the ->running assignment from the ->state assignment.
1750	 *
1751	 * As with a regular write barrier, this ensures the read side in
1752	 * hrtimer_active() cannot observe base->running == NULL &&
1753	 * timer->state == INACTIVE.
1754	 */
1755	raw_write_seqcount_barrier(&base->seq);
1756
1757	__remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
1758	fn = ACCESS_PRIVATE(timer, function);
1759
1760	/*
1761	 * Clear the 'is relative' flag for the TIME_LOW_RES case. If the
1762	 * timer is restarted with a period then it becomes an absolute
1763	 * timer. If its not restarted it does not matter.
1764	 */
1765	if (IS_ENABLED(CONFIG_TIME_LOW_RES))
1766		timer->is_rel = false;
1767
1768	/*
1769	 * The timer is marked as running in the CPU base, so it is
1770	 * protected against migration to a different CPU even if the lock
1771	 * is dropped.
1772	 */
1773	raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1774	trace_hrtimer_expire_entry(timer, now);
1775	expires_in_hardirq = lockdep_hrtimer_enter(timer);
1776
1777	restart = fn(timer);
1778
1779	lockdep_hrtimer_exit(expires_in_hardirq);
1780	trace_hrtimer_expire_exit(timer);
1781	raw_spin_lock_irq(&cpu_base->lock);
1782
1783	/*
1784	 * Note: We clear the running state after enqueue_hrtimer and
1785	 * we do not reprogram the event hardware. Happens either in
1786	 * hrtimer_start_range_ns() or in hrtimer_interrupt()
1787	 *
1788	 * Note: Because we dropped the cpu_base->lock above,
1789	 * hrtimer_start_range_ns() can have popped in and enqueued the timer
1790	 * for us already.
1791	 */
1792	if (restart != HRTIMER_NORESTART &&
1793	    !(timer->state & HRTIMER_STATE_ENQUEUED))
1794		enqueue_hrtimer(timer, base, HRTIMER_MODE_ABS);
1795
1796	/*
1797	 * Separate the ->running assignment from the ->state assignment.
1798	 *
1799	 * As with a regular write barrier, this ensures the read side in
1800	 * hrtimer_active() cannot observe base->running.timer == NULL &&
1801	 * timer->state == INACTIVE.
1802	 */
1803	raw_write_seqcount_barrier(&base->seq);
1804
1805	WARN_ON_ONCE(base->running != timer);
1806	base->running = NULL;
1807}
1808
1809static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now,
1810				 unsigned long flags, unsigned int active_mask)
1811{
1812	struct hrtimer_clock_base *base;
1813	unsigned int active = cpu_base->active_bases & active_mask;
1814
1815	for_each_active_base(base, cpu_base, active) {
1816		struct timerqueue_node *node;
1817		ktime_t basenow;
1818
1819		basenow = ktime_add(now, base->offset);
1820
1821		while ((node = timerqueue_getnext(&base->active))) {
1822			struct hrtimer *timer;
1823
1824			timer = container_of(node, struct hrtimer, node);
1825
1826			/*
1827			 * The immediate goal for using the softexpires is
1828			 * minimizing wakeups, not running timers at the
1829			 * earliest interrupt after their soft expiration.
1830			 * This allows us to avoid using a Priority Search
1831			 * Tree, which can answer a stabbing query for
1832			 * overlapping intervals and instead use the simple
1833			 * BST we already have.
1834			 * We don't add extra wakeups by delaying timers that
1835			 * are right-of a not yet expired timer, because that
1836			 * timer will have to trigger a wakeup anyway.
1837			 */
1838			if (basenow < hrtimer_get_softexpires_tv64(timer))
1839				break;
1840
1841			__run_hrtimer(cpu_base, base, timer, &basenow, flags);
1842			if (active_mask == HRTIMER_ACTIVE_SOFT)
1843				hrtimer_sync_wait_running(cpu_base, flags);
1844		}
1845	}
1846}
1847
1848static __latent_entropy void hrtimer_run_softirq(void)
1849{
1850	struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1851	unsigned long flags;
1852	ktime_t now;
1853
1854	hrtimer_cpu_base_lock_expiry(cpu_base);
1855	raw_spin_lock_irqsave(&cpu_base->lock, flags);
1856
1857	now = hrtimer_update_base(cpu_base);
1858	__hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_SOFT);
1859
1860	cpu_base->softirq_activated = 0;
1861	hrtimer_update_softirq_timer(cpu_base, true);
1862
1863	raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1864	hrtimer_cpu_base_unlock_expiry(cpu_base);
1865}
1866
1867#ifdef CONFIG_HIGH_RES_TIMERS
1868
1869/*
1870 * High resolution timer interrupt
1871 * Called with interrupts disabled
1872 */
1873void hrtimer_interrupt(struct clock_event_device *dev)
1874{
1875	struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1876	ktime_t expires_next, now, entry_time, delta;
1877	unsigned long flags;
1878	int retries = 0;
1879
1880	BUG_ON(!cpu_base->hres_active);
1881	cpu_base->nr_events++;
1882	dev->next_event = KTIME_MAX;
1883
1884	raw_spin_lock_irqsave(&cpu_base->lock, flags);
1885	entry_time = now = hrtimer_update_base(cpu_base);
1886retry:
1887	cpu_base->in_hrtirq = 1;
1888	/*
1889	 * We set expires_next to KTIME_MAX here with cpu_base->lock
1890	 * held to prevent that a timer is enqueued in our queue via
1891	 * the migration code. This does not affect enqueueing of
1892	 * timers which run their callback and need to be requeued on
1893	 * this CPU.
1894	 */
1895	cpu_base->expires_next = KTIME_MAX;
1896
1897	if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1898		cpu_base->softirq_expires_next = KTIME_MAX;
1899		cpu_base->softirq_activated = 1;
1900		raise_timer_softirq(HRTIMER_SOFTIRQ);
1901	}
1902
1903	__hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1904
1905	/* Reevaluate the clock bases for the [soft] next expiry */
1906	expires_next = hrtimer_update_next_event(cpu_base);
1907	/*
1908	 * Store the new expiry value so the migration code can verify
1909	 * against it.
1910	 */
1911	cpu_base->expires_next = expires_next;
1912	cpu_base->in_hrtirq = 0;
1913	raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1914
1915	/* Reprogramming necessary ? */
1916	if (!tick_program_event(expires_next, 0)) {
1917		cpu_base->hang_detected = 0;
1918		return;
1919	}
1920
1921	/*
1922	 * The next timer was already expired due to:
1923	 * - tracing
1924	 * - long lasting callbacks
1925	 * - being scheduled away when running in a VM
1926	 *
1927	 * We need to prevent that we loop forever in the hrtimer
1928	 * interrupt routine. We give it 3 attempts to avoid
1929	 * overreacting on some spurious event.
1930	 *
1931	 * Acquire base lock for updating the offsets and retrieving
1932	 * the current time.
1933	 */
1934	raw_spin_lock_irqsave(&cpu_base->lock, flags);
1935	now = hrtimer_update_base(cpu_base);
1936	cpu_base->nr_retries++;
1937	if (++retries < 3)
1938		goto retry;
1939	/*
1940	 * Give the system a chance to do something else than looping
1941	 * here. We stored the entry time, so we know exactly how long
1942	 * we spent here. We schedule the next event this amount of
1943	 * time away.
1944	 */
1945	cpu_base->nr_hangs++;
1946	cpu_base->hang_detected = 1;
1947	raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1948
1949	delta = ktime_sub(now, entry_time);
1950	if ((unsigned int)delta > cpu_base->max_hang_time)
1951		cpu_base->max_hang_time = (unsigned int) delta;
1952	/*
1953	 * Limit it to a sensible value as we enforce a longer
1954	 * delay. Give the CPU at least 100ms to catch up.
1955	 */
1956	if (delta > 100 * NSEC_PER_MSEC)
1957		expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1958	else
1959		expires_next = ktime_add(now, delta);
1960	tick_program_event(expires_next, 1);
1961	pr_warn_once("hrtimer: interrupt took %llu ns\n", ktime_to_ns(delta));
1962}
1963#endif /* !CONFIG_HIGH_RES_TIMERS */
1964
1965/*
1966 * Called from run_local_timers in hardirq context every jiffy
1967 */
1968void hrtimer_run_queues(void)
1969{
1970	struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
1971	unsigned long flags;
1972	ktime_t now;
1973
1974	if (hrtimer_hres_active(cpu_base))
1975		return;
1976
1977	/*
1978	 * This _is_ ugly: We have to check periodically, whether we
1979	 * can switch to highres and / or nohz mode. The clocksource
1980	 * switch happens with xtime_lock held. Notification from
1981	 * there only sets the check bit in the tick_oneshot code,
1982	 * otherwise we might deadlock vs. xtime_lock.
1983	 */
1984	if (tick_check_oneshot_change(!hrtimer_is_hres_enabled())) {
1985		hrtimer_switch_to_hres();
1986		return;
1987	}
1988
1989	raw_spin_lock_irqsave(&cpu_base->lock, flags);
1990	now = hrtimer_update_base(cpu_base);
1991
1992	if (!ktime_before(now, cpu_base->softirq_expires_next)) {
1993		cpu_base->softirq_expires_next = KTIME_MAX;
1994		cpu_base->softirq_activated = 1;
1995		raise_timer_softirq(HRTIMER_SOFTIRQ);
1996	}
1997
1998	__hrtimer_run_queues(cpu_base, now, flags, HRTIMER_ACTIVE_HARD);
1999	raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
2000}
2001
2002/*
2003 * Sleep related functions:
2004 */
2005static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
2006{
2007	struct hrtimer_sleeper *t =
2008		container_of(timer, struct hrtimer_sleeper, timer);
2009	struct task_struct *task = t->task;
2010
2011	t->task = NULL;
2012	if (task)
2013		wake_up_process(task);
2014
2015	return HRTIMER_NORESTART;
2016}
2017
2018/**
2019 * hrtimer_sleeper_start_expires - Start a hrtimer sleeper timer
2020 * @sl:		sleeper to be started
2021 * @mode:	timer mode abs/rel
2022 *
2023 * Wrapper around hrtimer_start_expires() for hrtimer_sleeper based timers
2024 * to allow PREEMPT_RT to tweak the delivery mode (soft/hardirq context)
2025 */
2026void hrtimer_sleeper_start_expires(struct hrtimer_sleeper *sl,
2027				   enum hrtimer_mode mode)
2028{
2029	/*
2030	 * Make the enqueue delivery mode check work on RT. If the sleeper
2031	 * was initialized for hard interrupt delivery, force the mode bit.
2032	 * This is a special case for hrtimer_sleepers because
2033	 * __hrtimer_setup_sleeper() determines the delivery mode on RT so the
2034	 * fiddling with this decision is avoided at the call sites.
2035	 */
2036	if (IS_ENABLED(CONFIG_PREEMPT_RT) && sl->timer.is_hard)
2037		mode |= HRTIMER_MODE_HARD;
2038
2039	hrtimer_start_expires(&sl->timer, mode);
2040}
2041EXPORT_SYMBOL_GPL(hrtimer_sleeper_start_expires);
2042
2043static void __hrtimer_setup_sleeper(struct hrtimer_sleeper *sl,
2044				    clockid_t clock_id, enum hrtimer_mode mode)
2045{
2046	/*
2047	 * On PREEMPT_RT enabled kernels hrtimers which are not explicitly
2048	 * marked for hard interrupt expiry mode are moved into soft
2049	 * interrupt context either for latency reasons or because the
2050	 * hrtimer callback takes regular spinlocks or invokes other
2051	 * functions which are not suitable for hard interrupt context on
2052	 * PREEMPT_RT.
2053	 *
2054	 * The hrtimer_sleeper callback is RT compatible in hard interrupt
2055	 * context, but there is a latency concern: Untrusted userspace can
2056	 * spawn many threads which arm timers for the same expiry time on
2057	 * the same CPU. That causes a latency spike due to the wakeup of
2058	 * a gazillion threads.
2059	 *
2060	 * OTOH, privileged real-time user space applications rely on the
2061	 * low latency of hard interrupt wakeups. If the current task is in
2062	 * a real-time scheduling class, mark the mode for hard interrupt
2063	 * expiry.
2064	 */
2065	if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
2066		if (rt_or_dl_task_policy(current) && !(mode & HRTIMER_MODE_SOFT))
2067			mode |= HRTIMER_MODE_HARD;
2068	}
2069
2070	__hrtimer_setup(&sl->timer, hrtimer_wakeup, clock_id, mode);
2071	sl->task = current;
2072}
2073
2074/**
2075 * hrtimer_setup_sleeper_on_stack - initialize a sleeper in stack memory
2076 * @sl:		sleeper to be initialized
2077 * @clock_id:	the clock to be used
2078 * @mode:	timer mode abs/rel
2079 */
2080void hrtimer_setup_sleeper_on_stack(struct hrtimer_sleeper *sl,
2081				    clockid_t clock_id, enum hrtimer_mode mode)
2082{
2083	debug_setup_on_stack(&sl->timer, clock_id, mode);
2084	__hrtimer_setup_sleeper(sl, clock_id, mode);
2085}
2086EXPORT_SYMBOL_GPL(hrtimer_setup_sleeper_on_stack);
2087
2088int nanosleep_copyout(struct restart_block *restart, struct timespec64 *ts)
2089{
2090	switch(restart->nanosleep.type) {
2091#ifdef CONFIG_COMPAT_32BIT_TIME
2092	case TT_COMPAT:
2093		if (put_old_timespec32(ts, restart->nanosleep.compat_rmtp))
2094			return -EFAULT;
2095		break;
2096#endif
2097	case TT_NATIVE:
2098		if (put_timespec64(ts, restart->nanosleep.rmtp))
2099			return -EFAULT;
2100		break;
2101	default:
2102		BUG();
2103	}
2104	return -ERESTART_RESTARTBLOCK;
2105}
2106
2107static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
2108{
2109	struct restart_block *restart;
2110
2111	do {
2112		set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
2113		hrtimer_sleeper_start_expires(t, mode);
2114
2115		if (likely(t->task))
2116			schedule();
2117
2118		hrtimer_cancel(&t->timer);
2119		mode = HRTIMER_MODE_ABS;
2120
2121	} while (t->task && !signal_pending(current));
2122
2123	__set_current_state(TASK_RUNNING);
2124
2125	if (!t->task)
2126		return 0;
2127
2128	restart = &current->restart_block;
2129	if (restart->nanosleep.type != TT_NONE) {
2130		ktime_t rem = hrtimer_expires_remaining(&t->timer);
2131		struct timespec64 rmt;
2132
2133		if (rem <= 0)
2134			return 0;
2135		rmt = ktime_to_timespec64(rem);
2136
2137		return nanosleep_copyout(restart, &rmt);
2138	}
2139	return -ERESTART_RESTARTBLOCK;
2140}
2141
2142static long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
2143{
2144	struct hrtimer_sleeper t;
2145	int ret;
2146
2147	hrtimer_setup_sleeper_on_stack(&t, restart->nanosleep.clockid, HRTIMER_MODE_ABS);
2148	hrtimer_set_expires(&t.timer, restart->nanosleep.expires);
2149	ret = do_nanosleep(&t, HRTIMER_MODE_ABS);
2150	destroy_hrtimer_on_stack(&t.timer);
2151	return ret;
2152}
2153
2154long hrtimer_nanosleep(ktime_t rqtp, const enum hrtimer_mode mode,
2155		       const clockid_t clockid)
2156{
2157	struct restart_block *restart;
2158	struct hrtimer_sleeper t;
2159	int ret = 0;
2160
2161	hrtimer_setup_sleeper_on_stack(&t, clockid, mode);
2162	hrtimer_set_expires_range_ns(&t.timer, rqtp, current->timer_slack_ns);
2163	ret = do_nanosleep(&t, mode);
2164	if (ret != -ERESTART_RESTARTBLOCK)
2165		goto out;
2166
2167	/* Absolute timers do not update the rmtp value and restart: */
2168	if (mode == HRTIMER_MODE_ABS) {
2169		ret = -ERESTARTNOHAND;
2170		goto out;
2171	}
2172
2173	restart = &current->restart_block;
2174	restart->nanosleep.clockid = t.timer.base->clockid;
2175	restart->nanosleep.expires = hrtimer_get_expires(&t.timer);
2176	set_restart_fn(restart, hrtimer_nanosleep_restart);
2177out:
2178	destroy_hrtimer_on_stack(&t.timer);
2179	return ret;
2180}
2181
2182#ifdef CONFIG_64BIT
2183
2184SYSCALL_DEFINE2(nanosleep, struct __kernel_timespec __user *, rqtp,
2185		struct __kernel_timespec __user *, rmtp)
2186{
2187	struct timespec64 tu;
2188
2189	if (get_timespec64(&tu, rqtp))
2190		return -EFAULT;
2191
2192	if (!timespec64_valid(&tu))
2193		return -EINVAL;
2194
2195	current->restart_block.fn = do_no_restart_syscall;
2196	current->restart_block.nanosleep.type = rmtp ? TT_NATIVE : TT_NONE;
2197	current->restart_block.nanosleep.rmtp = rmtp;
2198	return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2199				 CLOCK_MONOTONIC);
2200}
2201
2202#endif
2203
2204#ifdef CONFIG_COMPAT_32BIT_TIME
2205
2206SYSCALL_DEFINE2(nanosleep_time32, struct old_timespec32 __user *, rqtp,
2207		       struct old_timespec32 __user *, rmtp)
2208{
2209	struct timespec64 tu;
2210
2211	if (get_old_timespec32(&tu, rqtp))
2212		return -EFAULT;
2213
2214	if (!timespec64_valid(&tu))
2215		return -EINVAL;
2216
2217	current->restart_block.fn = do_no_restart_syscall;
2218	current->restart_block.nanosleep.type = rmtp ? TT_COMPAT : TT_NONE;
2219	current->restart_block.nanosleep.compat_rmtp = rmtp;
2220	return hrtimer_nanosleep(timespec64_to_ktime(tu), HRTIMER_MODE_REL,
2221				 CLOCK_MONOTONIC);
2222}
2223#endif
2224
2225/*
2226 * Functions related to boot-time initialization:
2227 */
2228int hrtimers_prepare_cpu(unsigned int cpu)
2229{
2230	struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
2231	int i;
2232
2233	for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2234		struct hrtimer_clock_base *clock_b = &cpu_base->clock_base[i];
2235
2236		clock_b->cpu_base = cpu_base;
2237		seqcount_raw_spinlock_init(&clock_b->seq, &cpu_base->lock);
2238		timerqueue_init_head(&clock_b->active);
2239	}
2240
2241	cpu_base->cpu = cpu;
2242	hrtimer_cpu_base_init_expiry_lock(cpu_base);
2243	return 0;
2244}
2245
2246int hrtimers_cpu_starting(unsigned int cpu)
2247{
2248	struct hrtimer_cpu_base *cpu_base = this_cpu_ptr(&hrtimer_bases);
2249
2250	/* Clear out any left over state from a CPU down operation */
2251	cpu_base->active_bases = 0;
2252	cpu_base->hres_active = 0;
2253	cpu_base->hang_detected = 0;
2254	cpu_base->next_timer = NULL;
2255	cpu_base->softirq_next_timer = NULL;
2256	cpu_base->expires_next = KTIME_MAX;
2257	cpu_base->softirq_expires_next = KTIME_MAX;
2258	cpu_base->online = 1;
2259	return 0;
2260}
2261
2262#ifdef CONFIG_HOTPLUG_CPU
2263
2264static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
2265				struct hrtimer_clock_base *new_base)
2266{
2267	struct hrtimer *timer;
2268	struct timerqueue_node *node;
2269
2270	while ((node = timerqueue_getnext(&old_base->active))) {
2271		timer = container_of(node, struct hrtimer, node);
2272		BUG_ON(hrtimer_callback_running(timer));
2273		debug_deactivate(timer);
2274
2275		/*
2276		 * Mark it as ENQUEUED not INACTIVE otherwise the
2277		 * timer could be seen as !active and just vanish away
2278		 * under us on another CPU
2279		 */
2280		__remove_hrtimer(timer, old_base, HRTIMER_STATE_ENQUEUED, 0);
2281		timer->base = new_base;
2282		/*
2283		 * Enqueue the timers on the new cpu. This does not
2284		 * reprogram the event device in case the timer
2285		 * expires before the earliest on this CPU, but we run
2286		 * hrtimer_interrupt after we migrated everything to
2287		 * sort out already expired timers and reprogram the
2288		 * event device.
2289		 */
2290		enqueue_hrtimer(timer, new_base, HRTIMER_MODE_ABS);
2291	}
2292}
2293
2294int hrtimers_cpu_dying(unsigned int dying_cpu)
2295{
2296	int i, ncpu = cpumask_any_and(cpu_active_mask, housekeeping_cpumask(HK_TYPE_TIMER));
2297	struct hrtimer_cpu_base *old_base, *new_base;
2298
2299	old_base = this_cpu_ptr(&hrtimer_bases);
2300	new_base = &per_cpu(hrtimer_bases, ncpu);
2301
2302	/*
2303	 * The caller is globally serialized and nobody else
2304	 * takes two locks at once, deadlock is not possible.
2305	 */
2306	raw_spin_lock(&old_base->lock);
2307	raw_spin_lock_nested(&new_base->lock, SINGLE_DEPTH_NESTING);
2308
2309	for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
2310		migrate_hrtimer_list(&old_base->clock_base[i],
2311				     &new_base->clock_base[i]);
2312	}
2313
2314	/* Tell the other CPU to retrigger the next event */
2315	smp_call_function_single(ncpu, retrigger_next_event, NULL, 0);
2316
2317	raw_spin_unlock(&new_base->lock);
2318	old_base->online = 0;
2319	raw_spin_unlock(&old_base->lock);
2320
2321	return 0;
2322}
2323
2324#endif /* CONFIG_HOTPLUG_CPU */
2325
2326void __init hrtimers_init(void)
2327{
2328	hrtimers_prepare_cpu(smp_processor_id());
2329	hrtimers_cpu_starting(smp_processor_id());
2330	open_softirq(HRTIMER_SOFTIRQ, hrtimer_run_softirq);
2331}