linux/kernel/hrtimer.c
<<
>>
Prefs
   1/*
   2 *  linux/kernel/hrtimer.c
   3 *
   4 *  Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
   5 *  Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
   6 *  Copyright(C) 2006-2007  Timesys Corp., Thomas Gleixner
   7 *
   8 *  High-resolution kernel timers
   9 *
  10 *  In contrast to the low-resolution timeout API implemented in
  11 *  kernel/timer.c, hrtimers provide finer resolution and accuracy
  12 *  depending on system configuration and capabilities.
  13 *
  14 *  These timers are currently used for:
  15 *   - itimers
  16 *   - POSIX timers
  17 *   - nanosleep
  18 *   - precise in-kernel timing
  19 *
  20 *  Started by: Thomas Gleixner and Ingo Molnar
  21 *
  22 *  Credits:
  23 *      based on kernel/timer.c
  24 *
  25 *      Help, testing, suggestions, bugfixes, improvements were
  26 *      provided by:
  27 *
  28 *      George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
  29 *      et. al.
  30 *
  31 *  For licencing details see kernel-base/COPYING
  32 */
  33
  34#include <linux/cpu.h>
  35#include <linux/export.h>
  36#include <linux/percpu.h>
  37#include <linux/hrtimer.h>
  38#include <linux/notifier.h>
  39#include <linux/syscalls.h>
  40#include <linux/kallsyms.h>
  41#include <linux/interrupt.h>
  42#include <linux/tick.h>
  43#include <linux/seq_file.h>
  44#include <linux/err.h>
  45#include <linux/debugobjects.h>
  46#include <linux/sched.h>
  47#include <linux/sched/sysctl.h>
  48#include <linux/sched/rt.h>
  49#include <linux/timer.h>
  50#include <linux/freezer.h>
  51
  52#include <asm/uaccess.h>
  53
  54#include <trace/events/timer.h>
  55
  56/*
  57 * The timer bases:
  58 *
  59 * There are more clockids then hrtimer bases. Thus, we index
  60 * into the timer bases by the hrtimer_base_type enum. When trying
  61 * to reach a base using a clockid, hrtimer_clockid_to_base()
  62 * is used to convert from clockid to the proper hrtimer_base_type.
  63 */
  64DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
  65{
  66
  67        .lock = __RAW_SPIN_LOCK_UNLOCKED(hrtimer_bases.lock),
  68        .clock_base =
  69        {
  70                {
  71                        .index = HRTIMER_BASE_MONOTONIC,
  72                        .clockid = CLOCK_MONOTONIC,
  73                        .get_time = &ktime_get,
  74                        .resolution = KTIME_LOW_RES,
  75                },
  76                {
  77                        .index = HRTIMER_BASE_REALTIME,
  78                        .clockid = CLOCK_REALTIME,
  79                        .get_time = &ktime_get_real,
  80                        .resolution = KTIME_LOW_RES,
  81                },
  82                {
  83                        .index = HRTIMER_BASE_BOOTTIME,
  84                        .clockid = CLOCK_BOOTTIME,
  85                        .get_time = &ktime_get_boottime,
  86                        .resolution = KTIME_LOW_RES,
  87                },
  88                {
  89                        .index = HRTIMER_BASE_TAI,
  90                        .clockid = CLOCK_TAI,
  91                        .get_time = &ktime_get_clocktai,
  92                        .resolution = KTIME_LOW_RES,
  93                },
  94        }
  95};
  96
  97static const int hrtimer_clock_to_base_table[MAX_CLOCKS] = {
  98        [CLOCK_REALTIME]        = HRTIMER_BASE_REALTIME,
  99        [CLOCK_MONOTONIC]       = HRTIMER_BASE_MONOTONIC,
 100        [CLOCK_BOOTTIME]        = HRTIMER_BASE_BOOTTIME,
 101        [CLOCK_TAI]             = HRTIMER_BASE_TAI,
 102};
 103
 104static inline int hrtimer_clockid_to_base(clockid_t clock_id)
 105{
 106        return hrtimer_clock_to_base_table[clock_id];
 107}
 108
 109
 110/*
 111 * Get the coarse grained time at the softirq based on xtime and
 112 * wall_to_monotonic.
 113 */
 114static void hrtimer_get_softirq_time(struct hrtimer_cpu_base *base)
 115{
 116        ktime_t xtim, mono, boot;
 117        struct timespec xts, tom, slp;
 118        s32 tai_offset;
 119
 120        get_xtime_and_monotonic_and_sleep_offset(&xts, &tom, &slp);
 121        tai_offset = timekeeping_get_tai_offset();
 122
 123        xtim = timespec_to_ktime(xts);
 124        mono = ktime_add(xtim, timespec_to_ktime(tom));
 125        boot = ktime_add(mono, timespec_to_ktime(slp));
 126        base->clock_base[HRTIMER_BASE_REALTIME].softirq_time = xtim;
 127        base->clock_base[HRTIMER_BASE_MONOTONIC].softirq_time = mono;
 128        base->clock_base[HRTIMER_BASE_BOOTTIME].softirq_time = boot;
 129        base->clock_base[HRTIMER_BASE_TAI].softirq_time =
 130                                ktime_add(xtim, ktime_set(tai_offset, 0));
 131}
 132
 133/*
 134 * Functions and macros which are different for UP/SMP systems are kept in a
 135 * single place
 136 */
 137#ifdef CONFIG_SMP
 138
 139/*
 140 * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
 141 * means that all timers which are tied to this base via timer->base are
 142 * locked, and the base itself is locked too.
 143 *
 144 * So __run_timers/migrate_timers can safely modify all timers which could
 145 * be found on the lists/queues.
 146 *
 147 * When the timer's base is locked, and the timer removed from list, it is
 148 * possible to set timer->base = NULL and drop the lock: the timer remains
 149 * locked.
 150 */
 151static
 152struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
 153                                             unsigned long *flags)
 154{
 155        struct hrtimer_clock_base *base;
 156
 157        for (;;) {
 158                base = timer->base;
 159                if (likely(base != NULL)) {
 160                        raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
 161                        if (likely(base == timer->base))
 162                                return base;
 163                        /* The timer has migrated to another CPU: */
 164                        raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
 165                }
 166                cpu_relax();
 167        }
 168}
 169
 170
 171/*
 172 * Get the preferred target CPU for NOHZ
 173 */
 174static int hrtimer_get_target(int this_cpu, int pinned)
 175{
 176#ifdef CONFIG_NO_HZ_COMMON
 177        if (!pinned && get_sysctl_timer_migration() && idle_cpu(this_cpu))
 178                return get_nohz_timer_target();
 179#endif
 180        return this_cpu;
 181}
 182
 183/*
 184 * With HIGHRES=y we do not migrate the timer when it is expiring
 185 * before the next event on the target cpu because we cannot reprogram
 186 * the target cpu hardware and we would cause it to fire late.
 187 *
 188 * Called with cpu_base->lock of target cpu held.
 189 */
 190static int
 191hrtimer_check_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base)
 192{
 193#ifdef CONFIG_HIGH_RES_TIMERS
 194        ktime_t expires;
 195
 196        if (!new_base->cpu_base->hres_active)
 197                return 0;
 198
 199        expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
 200        return expires.tv64 <= new_base->cpu_base->expires_next.tv64;
 201#else
 202        return 0;
 203#endif
 204}
 205
 206/*
 207 * Switch the timer base to the current CPU when possible.
 208 */
 209static inline struct hrtimer_clock_base *
 210switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
 211                    int pinned)
 212{
 213        struct hrtimer_clock_base *new_base;
 214        struct hrtimer_cpu_base *new_cpu_base;
 215        int this_cpu = smp_processor_id();
 216        int cpu = hrtimer_get_target(this_cpu, pinned);
 217        int basenum = base->index;
 218
 219again:
 220        new_cpu_base = &per_cpu(hrtimer_bases, cpu);
 221        new_base = &new_cpu_base->clock_base[basenum];
 222
 223        if (base != new_base) {
 224                /*
 225                 * We are trying to move timer to new_base.
 226                 * However we can't change timer's base while it is running,
 227                 * so we keep it on the same CPU. No hassle vs. reprogramming
 228                 * the event source in the high resolution case. The softirq
 229                 * code will take care of this when the timer function has
 230                 * completed. There is no conflict as we hold the lock until
 231                 * the timer is enqueued.
 232                 */
 233                if (unlikely(hrtimer_callback_running(timer)))
 234                        return base;
 235
 236                /* See the comment in lock_timer_base() */
 237                timer->base = NULL;
 238                raw_spin_unlock(&base->cpu_base->lock);
 239                raw_spin_lock(&new_base->cpu_base->lock);
 240
 241                if (cpu != this_cpu && hrtimer_check_target(timer, new_base)) {
 242                        cpu = this_cpu;
 243                        raw_spin_unlock(&new_base->cpu_base->lock);
 244                        raw_spin_lock(&base->cpu_base->lock);
 245                        timer->base = base;
 246                        goto again;
 247                }
 248                timer->base = new_base;
 249        }
 250        return new_base;
 251}
 252
 253#else /* CONFIG_SMP */
 254
 255static inline struct hrtimer_clock_base *
 256lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
 257{
 258        struct hrtimer_clock_base *base = timer->base;
 259
 260        raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
 261
 262        return base;
 263}
 264
 265# define switch_hrtimer_base(t, b, p)   (b)
 266
 267#endif  /* !CONFIG_SMP */
 268
 269/*
 270 * Functions for the union type storage format of ktime_t which are
 271 * too large for inlining:
 272 */
 273#if BITS_PER_LONG < 64
 274# ifndef CONFIG_KTIME_SCALAR
 275/**
 276 * ktime_add_ns - Add a scalar nanoseconds value to a ktime_t variable
 277 * @kt:         addend
 278 * @nsec:       the scalar nsec value to add
 279 *
 280 * Returns the sum of kt and nsec in ktime_t format
 281 */
 282ktime_t ktime_add_ns(const ktime_t kt, u64 nsec)
 283{
 284        ktime_t tmp;
 285
 286        if (likely(nsec < NSEC_PER_SEC)) {
 287                tmp.tv64 = nsec;
 288        } else {
 289                unsigned long rem = do_div(nsec, NSEC_PER_SEC);
 290
 291                /* Make sure nsec fits into long */
 292                if (unlikely(nsec > KTIME_SEC_MAX))
 293                        return (ktime_t){ .tv64 = KTIME_MAX };
 294
 295                tmp = ktime_set((long)nsec, rem);
 296        }
 297
 298        return ktime_add(kt, tmp);
 299}
 300
 301EXPORT_SYMBOL_GPL(ktime_add_ns);
 302
 303/**
 304 * ktime_sub_ns - Subtract a scalar nanoseconds value from a ktime_t variable
 305 * @kt:         minuend
 306 * @nsec:       the scalar nsec value to subtract
 307 *
 308 * Returns the subtraction of @nsec from @kt in ktime_t format
 309 */
 310ktime_t ktime_sub_ns(const ktime_t kt, u64 nsec)
 311{
 312        ktime_t tmp;
 313
 314        if (likely(nsec < NSEC_PER_SEC)) {
 315                tmp.tv64 = nsec;
 316        } else {
 317                unsigned long rem = do_div(nsec, NSEC_PER_SEC);
 318
 319                tmp = ktime_set((long)nsec, rem);
 320        }
 321
 322        return ktime_sub(kt, tmp);
 323}
 324
 325EXPORT_SYMBOL_GPL(ktime_sub_ns);
 326# endif /* !CONFIG_KTIME_SCALAR */
 327
 328/*
 329 * Divide a ktime value by a nanosecond value
 330 */
 331u64 ktime_divns(const ktime_t kt, s64 div)
 332{
 333        u64 dclc;
 334        int sft = 0;
 335
 336        dclc = ktime_to_ns(kt);
 337        /* Make sure the divisor is less than 2^32: */
 338        while (div >> 32) {
 339                sft++;
 340                div >>= 1;
 341        }
 342        dclc >>= sft;
 343        do_div(dclc, (unsigned long) div);
 344
 345        return dclc;
 346}
 347#endif /* BITS_PER_LONG >= 64 */
 348
 349/*
 350 * Add two ktime values and do a safety check for overflow:
 351 */
 352ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
 353{
 354        ktime_t res = ktime_add(lhs, rhs);
 355
 356        /*
 357         * We use KTIME_SEC_MAX here, the maximum timeout which we can
 358         * return to user space in a timespec:
 359         */
 360        if (res.tv64 < 0 || res.tv64 < lhs.tv64 || res.tv64 < rhs.tv64)
 361                res = ktime_set(KTIME_SEC_MAX, 0);
 362
 363        return res;
 364}
 365
 366EXPORT_SYMBOL_GPL(ktime_add_safe);
 367
 368#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
 369
 370static struct debug_obj_descr hrtimer_debug_descr;
 371
 372static void *hrtimer_debug_hint(void *addr)
 373{
 374        return ((struct hrtimer *) addr)->function;
 375}
 376
 377/*
 378 * fixup_init is called when:
 379 * - an active object is initialized
 380 */
 381static int hrtimer_fixup_init(void *addr, enum debug_obj_state state)
 382{
 383        struct hrtimer *timer = addr;
 384
 385        switch (state) {
 386        case ODEBUG_STATE_ACTIVE:
 387                hrtimer_cancel(timer);
 388                debug_object_init(timer, &hrtimer_debug_descr);
 389                return 1;
 390        default:
 391                return 0;
 392        }
 393}
 394
 395/*
 396 * fixup_activate is called when:
 397 * - an active object is activated
 398 * - an unknown object is activated (might be a statically initialized object)
 399 */
 400static int hrtimer_fixup_activate(void *addr, enum debug_obj_state state)
 401{
 402        switch (state) {
 403
 404        case ODEBUG_STATE_NOTAVAILABLE:
 405                WARN_ON_ONCE(1);
 406                return 0;
 407
 408        case ODEBUG_STATE_ACTIVE:
 409                WARN_ON(1);
 410
 411        default:
 412                return 0;
 413        }
 414}
 415
 416/*
 417 * fixup_free is called when:
 418 * - an active object is freed
 419 */
 420static int hrtimer_fixup_free(void *addr, enum debug_obj_state state)
 421{
 422        struct hrtimer *timer = addr;
 423
 424        switch (state) {
 425        case ODEBUG_STATE_ACTIVE:
 426                hrtimer_cancel(timer);
 427                debug_object_free(timer, &hrtimer_debug_descr);
 428                return 1;
 429        default:
 430                return 0;
 431        }
 432}
 433
 434static struct debug_obj_descr hrtimer_debug_descr = {
 435        .name           = "hrtimer",
 436        .debug_hint     = hrtimer_debug_hint,
 437        .fixup_init     = hrtimer_fixup_init,
 438        .fixup_activate = hrtimer_fixup_activate,
 439        .fixup_free     = hrtimer_fixup_free,
 440};
 441
 442static inline void debug_hrtimer_init(struct hrtimer *timer)
 443{
 444        debug_object_init(timer, &hrtimer_debug_descr);
 445}
 446
 447static inline void debug_hrtimer_activate(struct hrtimer *timer)
 448{
 449        debug_object_activate(timer, &hrtimer_debug_descr);
 450}
 451
 452static inline void debug_hrtimer_deactivate(struct hrtimer *timer)
 453{
 454        debug_object_deactivate(timer, &hrtimer_debug_descr);
 455}
 456
 457static inline void debug_hrtimer_free(struct hrtimer *timer)
 458{
 459        debug_object_free(timer, &hrtimer_debug_descr);
 460}
 461
 462static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
 463                           enum hrtimer_mode mode);
 464
 465void hrtimer_init_on_stack(struct hrtimer *timer, clockid_t clock_id,
 466                           enum hrtimer_mode mode)
 467{
 468        debug_object_init_on_stack(timer, &hrtimer_debug_descr);
 469        __hrtimer_init(timer, clock_id, mode);
 470}
 471EXPORT_SYMBOL_GPL(hrtimer_init_on_stack);
 472
 473void destroy_hrtimer_on_stack(struct hrtimer *timer)
 474{
 475        debug_object_free(timer, &hrtimer_debug_descr);
 476}
 477
 478#else
 479static inline void debug_hrtimer_init(struct hrtimer *timer) { }
 480static inline void debug_hrtimer_activate(struct hrtimer *timer) { }
 481static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
 482#endif
 483
 484static inline void
 485debug_init(struct hrtimer *timer, clockid_t clockid,
 486           enum hrtimer_mode mode)
 487{
 488        debug_hrtimer_init(timer);
 489        trace_hrtimer_init(timer, clockid, mode);
 490}
 491
 492static inline void debug_activate(struct hrtimer *timer)
 493{
 494        debug_hrtimer_activate(timer);
 495        trace_hrtimer_start(timer);
 496}
 497
 498static inline void debug_deactivate(struct hrtimer *timer)
 499{
 500        debug_hrtimer_deactivate(timer);
 501        trace_hrtimer_cancel(timer);
 502}
 503
 504/* High resolution timer related functions */
 505#ifdef CONFIG_HIGH_RES_TIMERS
 506
 507/*
 508 * High resolution timer enabled ?
 509 */
 510static int hrtimer_hres_enabled __read_mostly  = 1;
 511
 512/*
 513 * Enable / Disable high resolution mode
 514 */
 515static int __init setup_hrtimer_hres(char *str)
 516{
 517        if (!strcmp(str, "off"))
 518                hrtimer_hres_enabled = 0;
 519        else if (!strcmp(str, "on"))
 520                hrtimer_hres_enabled = 1;
 521        else
 522                return 0;
 523        return 1;
 524}
 525
 526__setup("highres=", setup_hrtimer_hres);
 527
 528/*
 529 * hrtimer_high_res_enabled - query, if the highres mode is enabled
 530 */
 531static inline int hrtimer_is_hres_enabled(void)
 532{
 533        return hrtimer_hres_enabled;
 534}
 535
 536/*
 537 * Is the high resolution mode active ?
 538 */
 539static inline int hrtimer_hres_active(void)
 540{
 541        return __this_cpu_read(hrtimer_bases.hres_active);
 542}
 543
 544/*
 545 * Reprogram the event source with checking both queues for the
 546 * next event
 547 * Called with interrupts disabled and base->lock held
 548 */
 549static void
 550hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
 551{
 552        int i;
 553        struct hrtimer_clock_base *base = cpu_base->clock_base;
 554        ktime_t expires, expires_next;
 555
 556        expires_next.tv64 = KTIME_MAX;
 557
 558        for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++, base++) {
 559                struct hrtimer *timer;
 560                struct timerqueue_node *next;
 561
 562                next = timerqueue_getnext(&base->active);
 563                if (!next)
 564                        continue;
 565                timer = container_of(next, struct hrtimer, node);
 566
 567                expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
 568                /*
 569                 * clock_was_set() has changed base->offset so the
 570                 * result might be negative. Fix it up to prevent a
 571                 * false positive in clockevents_program_event()
 572                 */
 573                if (expires.tv64 < 0)
 574                        expires.tv64 = 0;
 575                if (expires.tv64 < expires_next.tv64)
 576                        expires_next = expires;
 577        }
 578
 579        if (skip_equal && expires_next.tv64 == cpu_base->expires_next.tv64)
 580                return;
 581
 582        cpu_base->expires_next.tv64 = expires_next.tv64;
 583
 584        if (cpu_base->expires_next.tv64 != KTIME_MAX)
 585                tick_program_event(cpu_base->expires_next, 1);
 586}
 587
 588/*
 589 * Shared reprogramming for clock_realtime and clock_monotonic
 590 *
 591 * When a timer is enqueued and expires earlier than the already enqueued
 592 * timers, we have to check, whether it expires earlier than the timer for
 593 * which the clock event device was armed.
 594 *
 595 * Called with interrupts disabled and base->cpu_base.lock held
 596 */
 597static int hrtimer_reprogram(struct hrtimer *timer,
 598                             struct hrtimer_clock_base *base)
 599{
 600        struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
 601        ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
 602        int res;
 603
 604        WARN_ON_ONCE(hrtimer_get_expires_tv64(timer) < 0);
 605
 606        /*
 607         * When the callback is running, we do not reprogram the clock event
 608         * device. The timer callback is either running on a different CPU or
 609         * the callback is executed in the hrtimer_interrupt context. The
 610         * reprogramming is handled either by the softirq, which called the
 611         * callback or at the end of the hrtimer_interrupt.
 612         */
 613        if (hrtimer_callback_running(timer))
 614                return 0;
 615
 616        /*
 617         * CLOCK_REALTIME timer might be requested with an absolute
 618         * expiry time which is less than base->offset. Nothing wrong
 619         * about that, just avoid to call into the tick code, which
 620         * has now objections against negative expiry values.
 621         */
 622        if (expires.tv64 < 0)
 623                return -ETIME;
 624
 625        if (expires.tv64 >= cpu_base->expires_next.tv64)
 626                return 0;
 627
 628        /*
 629         * If a hang was detected in the last timer interrupt then we
 630         * do not schedule a timer which is earlier than the expiry
 631         * which we enforced in the hang detection. We want the system
 632         * to make progress.
 633         */
 634        if (cpu_base->hang_detected)
 635                return 0;
 636
 637        /*
 638         * Clockevents returns -ETIME, when the event was in the past.
 639         */
 640        res = tick_program_event(expires, 0);
 641        if (!IS_ERR_VALUE(res))
 642                cpu_base->expires_next = expires;
 643        return res;
 644}
 645
 646/*
 647 * Initialize the high resolution related parts of cpu_base
 648 */
 649static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base)
 650{
 651        base->expires_next.tv64 = KTIME_MAX;
 652        base->hres_active = 0;
 653}
 654
 655/*
 656 * When High resolution timers are active, try to reprogram. Note, that in case
 657 * the state has HRTIMER_STATE_CALLBACK set, no reprogramming and no expiry
 658 * check happens. The timer gets enqueued into the rbtree. The reprogramming
 659 * and expiry check is done in the hrtimer_interrupt or in the softirq.
 660 */
 661static inline int hrtimer_enqueue_reprogram(struct hrtimer *timer,
 662                                            struct hrtimer_clock_base *base)
 663{
 664        return base->cpu_base->hres_active && hrtimer_reprogram(timer, base);
 665}
 666
 667static inline ktime_t hrtimer_update_base(struct hrtimer_cpu_base *base)
 668{
 669        ktime_t *offs_real = &base->clock_base[HRTIMER_BASE_REALTIME].offset;
 670        ktime_t *offs_boot = &base->clock_base[HRTIMER_BASE_BOOTTIME].offset;
 671        ktime_t *offs_tai = &base->clock_base[HRTIMER_BASE_TAI].offset;
 672
 673        return ktime_get_update_offsets(offs_real, offs_boot, offs_tai);
 674}
 675
 676/*
 677 * Retrigger next event is called after clock was set
 678 *
 679 * Called with interrupts disabled via on_each_cpu()
 680 */
 681static void retrigger_next_event(void *arg)
 682{
 683        struct hrtimer_cpu_base *base = &__get_cpu_var(hrtimer_bases);
 684
 685        if (!hrtimer_hres_active())
 686                return;
 687
 688        raw_spin_lock(&base->lock);
 689        hrtimer_update_base(base);
 690        hrtimer_force_reprogram(base, 0);
 691        raw_spin_unlock(&base->lock);
 692}
 693
 694/*
 695 * Switch to high resolution mode
 696 */
 697static int hrtimer_switch_to_hres(void)
 698{
 699        int i, cpu = smp_processor_id();
 700        struct hrtimer_cpu_base *base = &per_cpu(hrtimer_bases, cpu);
 701        unsigned long flags;
 702
 703        if (base->hres_active)
 704                return 1;
 705
 706        local_irq_save(flags);
 707
 708        if (tick_init_highres()) {
 709                local_irq_restore(flags);
 710                printk(KERN_WARNING "Could not switch to high resolution "
 711                                    "mode on CPU %d\n", cpu);
 712                return 0;
 713        }
 714        base->hres_active = 1;
 715        for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++)
 716                base->clock_base[i].resolution = KTIME_HIGH_RES;
 717
 718        tick_setup_sched_timer();
 719        /* "Retrigger" the interrupt to get things going */
 720        retrigger_next_event(NULL);
 721        local_irq_restore(flags);
 722        return 1;
 723}
 724
 725static void clock_was_set_work(struct work_struct *work)
 726{
 727        clock_was_set();
 728}
 729
 730static DECLARE_WORK(hrtimer_work, clock_was_set_work);
 731
 732/*
 733 * Called from timekeeping and resume code to reprogramm the hrtimer
 734 * interrupt device on all cpus.
 735 */
 736void clock_was_set_delayed(void)
 737{
 738        schedule_work(&hrtimer_work);
 739}
 740
 741#else
 742
 743static inline int hrtimer_hres_active(void) { return 0; }
 744static inline int hrtimer_is_hres_enabled(void) { return 0; }
 745static inline int hrtimer_switch_to_hres(void) { return 0; }
 746static inline void
 747hrtimer_force_reprogram(struct hrtimer_cpu_base *base, int skip_equal) { }
 748static inline int hrtimer_enqueue_reprogram(struct hrtimer *timer,
 749                                            struct hrtimer_clock_base *base)
 750{
 751        return 0;
 752}
 753static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base) { }
 754static inline void retrigger_next_event(void *arg) { }
 755
 756#endif /* CONFIG_HIGH_RES_TIMERS */
 757
 758/*
 759 * Clock realtime was set
 760 *
 761 * Change the offset of the realtime clock vs. the monotonic
 762 * clock.
 763 *
 764 * We might have to reprogram the high resolution timer interrupt. On
 765 * SMP we call the architecture specific code to retrigger _all_ high
 766 * resolution timer interrupts. On UP we just disable interrupts and
 767 * call the high resolution interrupt code.
 768 */
 769void clock_was_set(void)
 770{
 771#ifdef CONFIG_HIGH_RES_TIMERS
 772        /* Retrigger the CPU local events everywhere */
 773        on_each_cpu(retrigger_next_event, NULL, 1);
 774#endif
 775        timerfd_clock_was_set();
 776}
 777
 778/*
 779 * During resume we might have to reprogram the high resolution timer
 780 * interrupt on all online CPUs.  However, all other CPUs will be
 781 * stopped with IRQs interrupts disabled so the clock_was_set() call
 782 * must be deferred.
 783 */
 784void hrtimers_resume(void)
 785{
 786        WARN_ONCE(!irqs_disabled(),
 787                  KERN_INFO "hrtimers_resume() called with IRQs enabled!");
 788
 789        /* Retrigger on the local CPU */
 790        retrigger_next_event(NULL);
 791        /* And schedule a retrigger for all others */
 792        clock_was_set_delayed();
 793}
 794
 795static inline void timer_stats_hrtimer_set_start_info(struct hrtimer *timer)
 796{
 797#ifdef CONFIG_TIMER_STATS
 798        if (timer->start_site)
 799                return;
 800        timer->start_site = __builtin_return_address(0);
 801        memcpy(timer->start_comm, current->comm, TASK_COMM_LEN);
 802        timer->start_pid = current->pid;
 803#endif
 804}
 805
 806static inline void timer_stats_hrtimer_clear_start_info(struct hrtimer *timer)
 807{
 808#ifdef CONFIG_TIMER_STATS
 809        timer->start_site = NULL;
 810#endif
 811}
 812
 813static inline void timer_stats_account_hrtimer(struct hrtimer *timer)
 814{
 815#ifdef CONFIG_TIMER_STATS
 816        if (likely(!timer_stats_active))
 817                return;
 818        timer_stats_update_stats(timer, timer->start_pid, timer->start_site,
 819                                 timer->function, timer->start_comm, 0);
 820#endif
 821}
 822
 823/*
 824 * Counterpart to lock_hrtimer_base above:
 825 */
 826static inline
 827void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
 828{
 829        raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
 830}
 831
 832/**
 833 * hrtimer_forward - forward the timer expiry
 834 * @timer:      hrtimer to forward
 835 * @now:        forward past this time
 836 * @interval:   the interval to forward
 837 *
 838 * Forward the timer expiry so it will expire in the future.
 839 * Returns the number of overruns.
 840 */
 841u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
 842{
 843        u64 orun = 1;
 844        ktime_t delta;
 845
 846        delta = ktime_sub(now, hrtimer_get_expires(timer));
 847
 848        if (delta.tv64 < 0)
 849                return 0;
 850
 851        if (interval.tv64 < timer->base->resolution.tv64)
 852                interval.tv64 = timer->base->resolution.tv64;
 853
 854        if (unlikely(delta.tv64 >= interval.tv64)) {
 855                s64 incr = ktime_to_ns(interval);
 856
 857                orun = ktime_divns(delta, incr);
 858                hrtimer_add_expires_ns(timer, incr * orun);
 859                if (hrtimer_get_expires_tv64(timer) > now.tv64)
 860                        return orun;
 861                /*
 862                 * This (and the ktime_add() below) is the
 863                 * correction for exact:
 864                 */
 865                orun++;
 866        }
 867        hrtimer_add_expires(timer, interval);
 868
 869        return orun;
 870}
 871EXPORT_SYMBOL_GPL(hrtimer_forward);
 872
 873/*
 874 * enqueue_hrtimer - internal function to (re)start a timer
 875 *
 876 * The timer is inserted in expiry order. Insertion into the
 877 * red black tree is O(log(n)). Must hold the base lock.
 878 *
 879 * Returns 1 when the new timer is the leftmost timer in the tree.
 880 */
 881static int enqueue_hrtimer(struct hrtimer *timer,
 882                           struct hrtimer_clock_base *base)
 883{
 884        debug_activate(timer);
 885
 886        timerqueue_add(&base->active, &timer->node);
 887        base->cpu_base->active_bases |= 1 << base->index;
 888
 889        /*
 890         * HRTIMER_STATE_ENQUEUED is or'ed to the current state to preserve the
 891         * state of a possibly running callback.
 892         */
 893        timer->state |= HRTIMER_STATE_ENQUEUED;
 894
 895        return (&timer->node == base->active.next);
 896}
 897
 898/*
 899 * __remove_hrtimer - internal function to remove a timer
 900 *
 901 * Caller must hold the base lock.
 902 *
 903 * High resolution timer mode reprograms the clock event device when the
 904 * timer is the one which expires next. The caller can disable this by setting
 905 * reprogram to zero. This is useful, when the context does a reprogramming
 906 * anyway (e.g. timer interrupt)
 907 */
 908static void __remove_hrtimer(struct hrtimer *timer,
 909                             struct hrtimer_clock_base *base,
 910                             unsigned long newstate, int reprogram)
 911{
 912        struct timerqueue_node *next_timer;
 913        if (!(timer->state & HRTIMER_STATE_ENQUEUED))
 914                goto out;
 915
 916        next_timer = timerqueue_getnext(&base->active);
 917        timerqueue_del(&base->active, &timer->node);
 918        if (&timer->node == next_timer) {
 919#ifdef CONFIG_HIGH_RES_TIMERS
 920                /* Reprogram the clock event device. if enabled */
 921                if (reprogram && hrtimer_hres_active()) {
 922                        ktime_t expires;
 923
 924                        expires = ktime_sub(hrtimer_get_expires(timer),
 925                                            base->offset);
 926                        if (base->cpu_base->expires_next.tv64 == expires.tv64)
 927                                hrtimer_force_reprogram(base->cpu_base, 1);
 928                }
 929#endif
 930        }
 931        if (!timerqueue_getnext(&base->active))
 932                base->cpu_base->active_bases &= ~(1 << base->index);
 933out:
 934        timer->state = newstate;
 935}
 936
 937/*
 938 * remove hrtimer, called with base lock held
 939 */
 940static inline int
 941remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base)
 942{
 943        if (hrtimer_is_queued(timer)) {
 944                unsigned long state;
 945                int reprogram;
 946
 947                /*
 948                 * Remove the timer and force reprogramming when high
 949                 * resolution mode is active and the timer is on the current
 950                 * CPU. If we remove a timer on another CPU, reprogramming is
 951                 * skipped. The interrupt event on this CPU is fired and
 952                 * reprogramming happens in the interrupt handler. This is a
 953                 * rare case and less expensive than a smp call.
 954                 */
 955                debug_deactivate(timer);
 956                timer_stats_hrtimer_clear_start_info(timer);
 957                reprogram = base->cpu_base == &__get_cpu_var(hrtimer_bases);
 958                /*
 959                 * We must preserve the CALLBACK state flag here,
 960                 * otherwise we could move the timer base in
 961                 * switch_hrtimer_base.
 962                 */
 963                state = timer->state & HRTIMER_STATE_CALLBACK;
 964                __remove_hrtimer(timer, base, state, reprogram);
 965                return 1;
 966        }
 967        return 0;
 968}
 969
 970int __hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
 971                unsigned long delta_ns, const enum hrtimer_mode mode,
 972                int wakeup)
 973{
 974        struct hrtimer_clock_base *base, *new_base;
 975        unsigned long flags;
 976        int ret, leftmost;
 977
 978        base = lock_hrtimer_base(timer, &flags);
 979
 980        /* Remove an active timer from the queue: */
 981        ret = remove_hrtimer(timer, base);
 982
 983        /* Switch the timer base, if necessary: */
 984        new_base = switch_hrtimer_base(timer, base, mode & HRTIMER_MODE_PINNED);
 985
 986        if (mode & HRTIMER_MODE_REL) {
 987                tim = ktime_add_safe(tim, new_base->get_time());
 988                /*
 989                 * CONFIG_TIME_LOW_RES is a temporary way for architectures
 990                 * to signal that they simply return xtime in
 991                 * do_gettimeoffset(). In this case we want to round up by
 992                 * resolution when starting a relative timer, to avoid short
 993                 * timeouts. This will go away with the GTOD framework.
 994                 */
 995#ifdef CONFIG_TIME_LOW_RES
 996                tim = ktime_add_safe(tim, base->resolution);
 997#endif
 998        }
 999
1000        hrtimer_set_expires_range_ns(timer, tim, delta_ns);
1001
1002        timer_stats_hrtimer_set_start_info(timer);
1003
1004        leftmost = enqueue_hrtimer(timer, new_base);
1005
1006        /*
1007         * Only allow reprogramming if the new base is on this CPU.
1008         * (it might still be on another CPU if the timer was pending)
1009         *
1010         * XXX send_remote_softirq() ?
1011         */
1012        if (leftmost && new_base->cpu_base == &__get_cpu_var(hrtimer_bases)
1013                && hrtimer_enqueue_reprogram(timer, new_base)) {
1014                if (wakeup) {
1015                        /*
1016                         * We need to drop cpu_base->lock to avoid a
1017                         * lock ordering issue vs. rq->lock.
1018                         */
1019                        raw_spin_unlock(&new_base->cpu_base->lock);
1020                        raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1021                        local_irq_restore(flags);
1022                        return ret;
1023                } else {
1024                        __raise_softirq_irqoff(HRTIMER_SOFTIRQ);
1025                }
1026        }
1027
1028        unlock_hrtimer_base(timer, &flags);
1029
1030        return ret;
1031}
1032
1033/**
1034 * hrtimer_start_range_ns - (re)start an hrtimer on the current CPU
1035 * @timer:      the timer to be added
1036 * @tim:        expiry time
1037 * @delta_ns:   "slack" range for the timer
1038 * @mode:       expiry mode: absolute (HRTIMER_MODE_ABS) or
1039 *              relative (HRTIMER_MODE_REL)
1040 *
1041 * Returns:
1042 *  0 on success
1043 *  1 when the timer was active
1044 */
1045int hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
1046                unsigned long delta_ns, const enum hrtimer_mode mode)
1047{
1048        return __hrtimer_start_range_ns(timer, tim, delta_ns, mode, 1);
1049}
1050EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
1051
1052/**
1053 * hrtimer_start - (re)start an hrtimer on the current CPU
1054 * @timer:      the timer to be added
1055 * @tim:        expiry time
1056 * @mode:       expiry mode: absolute (HRTIMER_MODE_ABS) or
1057 *              relative (HRTIMER_MODE_REL)
1058 *
1059 * Returns:
1060 *  0 on success
1061 *  1 when the timer was active
1062 */
1063int
1064hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
1065{
1066        return __hrtimer_start_range_ns(timer, tim, 0, mode, 1);
1067}
1068EXPORT_SYMBOL_GPL(hrtimer_start);
1069
1070
1071/**
1072 * hrtimer_try_to_cancel - try to deactivate a timer
1073 * @timer:      hrtimer to stop
1074 *
1075 * Returns:
1076 *  0 when the timer was not active
1077 *  1 when the timer was active
1078 * -1 when the timer is currently excuting the callback function and
1079 *    cannot be stopped
1080 */
1081int hrtimer_try_to_cancel(struct hrtimer *timer)
1082{
1083        struct hrtimer_clock_base *base;
1084        unsigned long flags;
1085        int ret = -1;
1086
1087        base = lock_hrtimer_base(timer, &flags);
1088
1089        if (!hrtimer_callback_running(timer))
1090                ret = remove_hrtimer(timer, base);
1091
1092        unlock_hrtimer_base(timer, &flags);
1093
1094        return ret;
1095
1096}
1097EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
1098
1099/**
1100 * hrtimer_cancel - cancel a timer and wait for the handler to finish.
1101 * @timer:      the timer to be cancelled
1102 *
1103 * Returns:
1104 *  0 when the timer was not active
1105 *  1 when the timer was active
1106 */
1107int hrtimer_cancel(struct hrtimer *timer)
1108{
1109        for (;;) {
1110                int ret = hrtimer_try_to_cancel(timer);
1111
1112                if (ret >= 0)
1113                        return ret;
1114                cpu_relax();
1115        }
1116}
1117EXPORT_SYMBOL_GPL(hrtimer_cancel);
1118
1119/**
1120 * hrtimer_get_remaining - get remaining time for the timer
1121 * @timer:      the timer to read
1122 */
1123ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
1124{
1125        unsigned long flags;
1126        ktime_t rem;
1127
1128        lock_hrtimer_base(timer, &flags);
1129        rem = hrtimer_expires_remaining(timer);
1130        unlock_hrtimer_base(timer, &flags);
1131
1132        return rem;
1133}
1134EXPORT_SYMBOL_GPL(hrtimer_get_remaining);
1135
1136#ifdef CONFIG_NO_HZ_COMMON
1137/**
1138 * hrtimer_get_next_event - get the time until next expiry event
1139 *
1140 * Returns the delta to the next expiry event or KTIME_MAX if no timer
1141 * is pending.
1142 */
1143ktime_t hrtimer_get_next_event(void)
1144{
1145        struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
1146        struct hrtimer_clock_base *base = cpu_base->clock_base;
1147        ktime_t delta, mindelta = { .tv64 = KTIME_MAX };
1148        unsigned long flags;
1149        int i;
1150
1151        raw_spin_lock_irqsave(&cpu_base->lock, flags);
1152
1153        if (!hrtimer_hres_active()) {
1154                for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++, base++) {
1155                        struct hrtimer *timer;
1156                        struct timerqueue_node *next;
1157
1158                        next = timerqueue_getnext(&base->active);
1159                        if (!next)
1160                                continue;
1161
1162                        timer = container_of(next, struct hrtimer, node);
1163                        delta.tv64 = hrtimer_get_expires_tv64(timer);
1164                        delta = ktime_sub(delta, base->get_time());
1165                        if (delta.tv64 < mindelta.tv64)
1166                                mindelta.tv64 = delta.tv64;
1167                }
1168        }
1169
1170        raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
1171
1172        if (mindelta.tv64 < 0)
1173                mindelta.tv64 = 0;
1174        return mindelta;
1175}
1176#endif
1177
1178static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1179                           enum hrtimer_mode mode)
1180{
1181        struct hrtimer_cpu_base *cpu_base;
1182        int base;
1183
1184        memset(timer, 0, sizeof(struct hrtimer));
1185
1186        cpu_base = &__raw_get_cpu_var(hrtimer_bases);
1187
1188        if (clock_id == CLOCK_REALTIME && mode != HRTIMER_MODE_ABS)
1189                clock_id = CLOCK_MONOTONIC;
1190
1191        base = hrtimer_clockid_to_base(clock_id);
1192        timer->base = &cpu_base->clock_base[base];
1193        timerqueue_init(&timer->node);
1194
1195#ifdef CONFIG_TIMER_STATS
1196        timer->start_site = NULL;
1197        timer->start_pid = -1;
1198        memset(timer->start_comm, 0, TASK_COMM_LEN);
1199#endif
1200}
1201
1202/**
1203 * hrtimer_init - initialize a timer to the given clock
1204 * @timer:      the timer to be initialized
1205 * @clock_id:   the clock to be used
1206 * @mode:       timer mode abs/rel
1207 */
1208void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
1209                  enum hrtimer_mode mode)
1210{
1211        debug_init(timer, clock_id, mode);
1212        __hrtimer_init(timer, clock_id, mode);
1213}
1214EXPORT_SYMBOL_GPL(hrtimer_init);
1215
1216/**
1217 * hrtimer_get_res - get the timer resolution for a clock
1218 * @which_clock: which clock to query
1219 * @tp:          pointer to timespec variable to store the resolution
1220 *
1221 * Store the resolution of the clock selected by @which_clock in the
1222 * variable pointed to by @tp.
1223 */
1224int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
1225{
1226        struct hrtimer_cpu_base *cpu_base;
1227        int base = hrtimer_clockid_to_base(which_clock);
1228
1229        cpu_base = &__raw_get_cpu_var(hrtimer_bases);
1230        *tp = ktime_to_timespec(cpu_base->clock_base[base].resolution);
1231
1232        return 0;
1233}
1234EXPORT_SYMBOL_GPL(hrtimer_get_res);
1235
1236static void __run_hrtimer(struct hrtimer *timer, ktime_t *now)
1237{
1238        struct hrtimer_clock_base *base = timer->base;
1239        struct hrtimer_cpu_base *cpu_base = base->cpu_base;
1240        enum hrtimer_restart (*fn)(struct hrtimer *);
1241        int restart;
1242
1243        WARN_ON(!irqs_disabled());
1244
1245        debug_deactivate(timer);
1246        __remove_hrtimer(timer, base, HRTIMER_STATE_CALLBACK, 0);
1247        timer_stats_account_hrtimer(timer);
1248        fn = timer->function;
1249
1250        /*
1251         * Because we run timers from hardirq context, there is no chance
1252         * they get migrated to another cpu, therefore its safe to unlock
1253         * the timer base.
1254         */
1255        raw_spin_unlock(&cpu_base->lock);
1256        trace_hrtimer_expire_entry(timer, now);
1257        restart = fn(timer);
1258        trace_hrtimer_expire_exit(timer);
1259        raw_spin_lock(&cpu_base->lock);
1260
1261        /*
1262         * Note: We clear the CALLBACK bit after enqueue_hrtimer and
1263         * we do not reprogramm the event hardware. Happens either in
1264         * hrtimer_start_range_ns() or in hrtimer_interrupt()
1265         */
1266        if (restart != HRTIMER_NORESTART) {
1267                BUG_ON(timer->state != HRTIMER_STATE_CALLBACK);
1268                enqueue_hrtimer(timer, base);
1269        }
1270
1271        WARN_ON_ONCE(!(timer->state & HRTIMER_STATE_CALLBACK));
1272
1273        timer->state &= ~HRTIMER_STATE_CALLBACK;
1274}
1275
1276#ifdef CONFIG_HIGH_RES_TIMERS
1277
1278/*
1279 * High resolution timer interrupt
1280 * Called with interrupts disabled
1281 */
1282void hrtimer_interrupt(struct clock_event_device *dev)
1283{
1284        struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
1285        ktime_t expires_next, now, entry_time, delta;
1286        int i, retries = 0;
1287
1288        BUG_ON(!cpu_base->hres_active);
1289        cpu_base->nr_events++;
1290        dev->next_event.tv64 = KTIME_MAX;
1291
1292        raw_spin_lock(&cpu_base->lock);
1293        entry_time = now = hrtimer_update_base(cpu_base);
1294retry:
1295        expires_next.tv64 = KTIME_MAX;
1296        /*
1297         * We set expires_next to KTIME_MAX here with cpu_base->lock
1298         * held to prevent that a timer is enqueued in our queue via
1299         * the migration code. This does not affect enqueueing of
1300         * timers which run their callback and need to be requeued on
1301         * this CPU.
1302         */
1303        cpu_base->expires_next.tv64 = KTIME_MAX;
1304
1305        for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1306                struct hrtimer_clock_base *base;
1307                struct timerqueue_node *node;
1308                ktime_t basenow;
1309
1310                if (!(cpu_base->active_bases & (1 << i)))
1311                        continue;
1312
1313                base = cpu_base->clock_base + i;
1314                basenow = ktime_add(now, base->offset);
1315
1316                while ((node = timerqueue_getnext(&base->active))) {
1317                        struct hrtimer *timer;
1318
1319                        timer = container_of(node, struct hrtimer, node);
1320
1321                        /*
1322                         * The immediate goal for using the softexpires is
1323                         * minimizing wakeups, not running timers at the
1324                         * earliest interrupt after their soft expiration.
1325                         * This allows us to avoid using a Priority Search
1326                         * Tree, which can answer a stabbing querry for
1327                         * overlapping intervals and instead use the simple
1328                         * BST we already have.
1329                         * We don't add extra wakeups by delaying timers that
1330                         * are right-of a not yet expired timer, because that
1331                         * timer will have to trigger a wakeup anyway.
1332                         */
1333
1334                        if (basenow.tv64 < hrtimer_get_softexpires_tv64(timer)) {
1335                                ktime_t expires;
1336
1337                                expires = ktime_sub(hrtimer_get_expires(timer),
1338                                                    base->offset);
1339                                if (expires.tv64 < 0)
1340                                        expires.tv64 = KTIME_MAX;
1341                                if (expires.tv64 < expires_next.tv64)
1342                                        expires_next = expires;
1343                                break;
1344                        }
1345
1346                        __run_hrtimer(timer, &basenow);
1347                }
1348        }
1349
1350        /*
1351         * Store the new expiry value so the migration code can verify
1352         * against it.
1353         */
1354        cpu_base->expires_next = expires_next;
1355        raw_spin_unlock(&cpu_base->lock);
1356
1357        /* Reprogramming necessary ? */
1358        if (expires_next.tv64 == KTIME_MAX ||
1359            !tick_program_event(expires_next, 0)) {
1360                cpu_base->hang_detected = 0;
1361                return;
1362        }
1363
1364        /*
1365         * The next timer was already expired due to:
1366         * - tracing
1367         * - long lasting callbacks
1368         * - being scheduled away when running in a VM
1369         *
1370         * We need to prevent that we loop forever in the hrtimer
1371         * interrupt routine. We give it 3 attempts to avoid
1372         * overreacting on some spurious event.
1373         *
1374         * Acquire base lock for updating the offsets and retrieving
1375         * the current time.
1376         */
1377        raw_spin_lock(&cpu_base->lock);
1378        now = hrtimer_update_base(cpu_base);
1379        cpu_base->nr_retries++;
1380        if (++retries < 3)
1381                goto retry;
1382        /*
1383         * Give the system a chance to do something else than looping
1384         * here. We stored the entry time, so we know exactly how long
1385         * we spent here. We schedule the next event this amount of
1386         * time away.
1387         */
1388        cpu_base->nr_hangs++;
1389        cpu_base->hang_detected = 1;
1390        raw_spin_unlock(&cpu_base->lock);
1391        delta = ktime_sub(now, entry_time);
1392        if (delta.tv64 > cpu_base->max_hang_time.tv64)
1393                cpu_base->max_hang_time = delta;
1394        /*
1395         * Limit it to a sensible value as we enforce a longer
1396         * delay. Give the CPU at least 100ms to catch up.
1397         */
1398        if (delta.tv64 > 100 * NSEC_PER_MSEC)
1399                expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
1400        else
1401                expires_next = ktime_add(now, delta);
1402        tick_program_event(expires_next, 1);
1403        printk_once(KERN_WARNING "hrtimer: interrupt took %llu ns\n",
1404                    ktime_to_ns(delta));
1405}
1406
1407/*
1408 * local version of hrtimer_peek_ahead_timers() called with interrupts
1409 * disabled.
1410 */
1411static void __hrtimer_peek_ahead_timers(void)
1412{
1413        struct tick_device *td;
1414
1415        if (!hrtimer_hres_active())
1416                return;
1417
1418        td = &__get_cpu_var(tick_cpu_device);
1419        if (td && td->evtdev)
1420                hrtimer_interrupt(td->evtdev);
1421}
1422
1423/**
1424 * hrtimer_peek_ahead_timers -- run soft-expired timers now
1425 *
1426 * hrtimer_peek_ahead_timers will peek at the timer queue of
1427 * the current cpu and check if there are any timers for which
1428 * the soft expires time has passed. If any such timers exist,
1429 * they are run immediately and then removed from the timer queue.
1430 *
1431 */
1432void hrtimer_peek_ahead_timers(void)
1433{
1434        unsigned long flags;
1435
1436        local_irq_save(flags);
1437        __hrtimer_peek_ahead_timers();
1438        local_irq_restore(flags);
1439}
1440
1441static void run_hrtimer_softirq(struct softirq_action *h)
1442{
1443        hrtimer_peek_ahead_timers();
1444}
1445
1446#else /* CONFIG_HIGH_RES_TIMERS */
1447
1448static inline void __hrtimer_peek_ahead_timers(void) { }
1449
1450#endif  /* !CONFIG_HIGH_RES_TIMERS */
1451
1452/*
1453 * Called from timer softirq every jiffy, expire hrtimers:
1454 *
1455 * For HRT its the fall back code to run the softirq in the timer
1456 * softirq context in case the hrtimer initialization failed or has
1457 * not been done yet.
1458 */
1459void hrtimer_run_pending(void)
1460{
1461        if (hrtimer_hres_active())
1462                return;
1463
1464        /*
1465         * This _is_ ugly: We have to check in the softirq context,
1466         * whether we can switch to highres and / or nohz mode. The
1467         * clocksource switch happens in the timer interrupt with
1468         * xtime_lock held. Notification from there only sets the
1469         * check bit in the tick_oneshot code, otherwise we might
1470         * deadlock vs. xtime_lock.
1471         */
1472        if (tick_check_oneshot_change(!hrtimer_is_hres_enabled()))
1473                hrtimer_switch_to_hres();
1474}
1475
1476/*
1477 * Called from hardirq context every jiffy
1478 */
1479void hrtimer_run_queues(void)
1480{
1481        struct timerqueue_node *node;
1482        struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
1483        struct hrtimer_clock_base *base;
1484        int index, gettime = 1;
1485
1486        if (hrtimer_hres_active())
1487                return;
1488
1489        for (index = 0; index < HRTIMER_MAX_CLOCK_BASES; index++) {
1490                base = &cpu_base->clock_base[index];
1491                if (!timerqueue_getnext(&base->active))
1492                        continue;
1493
1494                if (gettime) {
1495                        hrtimer_get_softirq_time(cpu_base);
1496                        gettime = 0;
1497                }
1498
1499                raw_spin_lock(&cpu_base->lock);
1500
1501                while ((node = timerqueue_getnext(&base->active))) {
1502                        struct hrtimer *timer;
1503
1504                        timer = container_of(node, struct hrtimer, node);
1505                        if (base->softirq_time.tv64 <=
1506                                        hrtimer_get_expires_tv64(timer))
1507                                break;
1508
1509                        __run_hrtimer(timer, &base->softirq_time);
1510                }
1511                raw_spin_unlock(&cpu_base->lock);
1512        }
1513}
1514
1515/*
1516 * Sleep related functions:
1517 */
1518static enum hrtimer_restart hrtimer_wakeup(struct hrtimer *timer)
1519{
1520        struct hrtimer_sleeper *t =
1521                container_of(timer, struct hrtimer_sleeper, timer);
1522        struct task_struct *task = t->task;
1523
1524        t->task = NULL;
1525        if (task)
1526                wake_up_process(task);
1527
1528        return HRTIMER_NORESTART;
1529}
1530
1531void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, struct task_struct *task)
1532{
1533        sl->timer.function = hrtimer_wakeup;
1534        sl->task = task;
1535}
1536EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
1537
1538static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
1539{
1540        hrtimer_init_sleeper(t, current);
1541
1542        do {
1543                set_current_state(TASK_INTERRUPTIBLE);
1544                hrtimer_start_expires(&t->timer, mode);
1545                if (!hrtimer_active(&t->timer))
1546                        t->task = NULL;
1547
1548                if (likely(t->task))
1549                        freezable_schedule();
1550
1551                hrtimer_cancel(&t->timer);
1552                mode = HRTIMER_MODE_ABS;
1553
1554        } while (t->task && !signal_pending(current));
1555
1556        __set_current_state(TASK_RUNNING);
1557
1558        return t->task == NULL;
1559}
1560
1561static int update_rmtp(struct hrtimer *timer, struct timespec __user *rmtp)
1562{
1563        struct timespec rmt;
1564        ktime_t rem;
1565
1566        rem = hrtimer_expires_remaining(timer);
1567        if (rem.tv64 <= 0)
1568                return 0;
1569        rmt = ktime_to_timespec(rem);
1570
1571        if (copy_to_user(rmtp, &rmt, sizeof(*rmtp)))
1572                return -EFAULT;
1573
1574        return 1;
1575}
1576
1577long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
1578{
1579        struct hrtimer_sleeper t;
1580        struct timespec __user  *rmtp;
1581        int ret = 0;
1582
1583        hrtimer_init_on_stack(&t.timer, restart->nanosleep.clockid,
1584                                HRTIMER_MODE_ABS);
1585        hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
1586
1587        if (do_nanosleep(&t, HRTIMER_MODE_ABS))
1588                goto out;
1589
1590        rmtp = restart->nanosleep.rmtp;
1591        if (rmtp) {
1592                ret = update_rmtp(&t.timer, rmtp);
1593                if (ret <= 0)
1594                        goto out;
1595        }
1596
1597        /* The other values in restart are already filled in */
1598        ret = -ERESTART_RESTARTBLOCK;
1599out:
1600        destroy_hrtimer_on_stack(&t.timer);
1601        return ret;
1602}
1603
1604long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
1605                       const enum hrtimer_mode mode, const clockid_t clockid)
1606{
1607        struct restart_block *restart;
1608        struct hrtimer_sleeper t;
1609        int ret = 0;
1610        unsigned long slack;
1611
1612        slack = current->timer_slack_ns;
1613        if (rt_task(current))
1614                slack = 0;
1615
1616        hrtimer_init_on_stack(&t.timer, clockid, mode);
1617        hrtimer_set_expires_range_ns(&t.timer, timespec_to_ktime(*rqtp), slack);
1618        if (do_nanosleep(&t, mode))
1619                goto out;
1620
1621        /* Absolute timers do not update the rmtp value and restart: */
1622        if (mode == HRTIMER_MODE_ABS) {
1623                ret = -ERESTARTNOHAND;
1624                goto out;
1625        }
1626
1627        if (rmtp) {
1628                ret = update_rmtp(&t.timer, rmtp);
1629                if (ret <= 0)
1630                        goto out;
1631        }
1632
1633        restart = &current_thread_info()->restart_block;
1634        restart->fn = hrtimer_nanosleep_restart;
1635        restart->nanosleep.clockid = t.timer.base->clockid;
1636        restart->nanosleep.rmtp = rmtp;
1637        restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
1638
1639        ret = -ERESTART_RESTARTBLOCK;
1640out:
1641        destroy_hrtimer_on_stack(&t.timer);
1642        return ret;
1643}
1644
1645SYSCALL_DEFINE2(nanosleep, struct timespec __user *, rqtp,
1646                struct timespec __user *, rmtp)
1647{
1648        struct timespec tu;
1649
1650        if (copy_from_user(&tu, rqtp, sizeof(tu)))
1651                return -EFAULT;
1652
1653        if (!timespec_valid(&tu))
1654                return -EINVAL;
1655
1656        return hrtimer_nanosleep(&tu, rmtp, HRTIMER_MODE_REL, CLOCK_MONOTONIC);
1657}
1658
1659/*
1660 * Functions related to boot-time initialization:
1661 */
1662static void init_hrtimers_cpu(int cpu)
1663{
1664        struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
1665        int i;
1666
1667        for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1668                cpu_base->clock_base[i].cpu_base = cpu_base;
1669                timerqueue_init_head(&cpu_base->clock_base[i].active);
1670        }
1671
1672        hrtimer_init_hres(cpu_base);
1673}
1674
1675#ifdef CONFIG_HOTPLUG_CPU
1676
1677static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
1678                                struct hrtimer_clock_base *new_base)
1679{
1680        struct hrtimer *timer;
1681        struct timerqueue_node *node;
1682
1683        while ((node = timerqueue_getnext(&old_base->active))) {
1684                timer = container_of(node, struct hrtimer, node);
1685                BUG_ON(hrtimer_callback_running(timer));
1686                debug_deactivate(timer);
1687
1688                /*
1689                 * Mark it as STATE_MIGRATE not INACTIVE otherwise the
1690                 * timer could be seen as !active and just vanish away
1691                 * under us on another CPU
1692                 */
1693                __remove_hrtimer(timer, old_base, HRTIMER_STATE_MIGRATE, 0);
1694                timer->base = new_base;
1695                /*
1696                 * Enqueue the timers on the new cpu. This does not
1697                 * reprogram the event device in case the timer
1698                 * expires before the earliest on this CPU, but we run
1699                 * hrtimer_interrupt after we migrated everything to
1700                 * sort out already expired timers and reprogram the
1701                 * event device.
1702                 */
1703                enqueue_hrtimer(timer, new_base);
1704
1705                /* Clear the migration state bit */
1706                timer->state &= ~HRTIMER_STATE_MIGRATE;
1707        }
1708}
1709
1710static void migrate_hrtimers(int scpu)
1711{
1712        struct hrtimer_cpu_base *old_base, *new_base;
1713        int i;
1714
1715        BUG_ON(cpu_online(scpu));
1716        tick_cancel_sched_timer(scpu);
1717
1718        local_irq_disable();
1719        old_base = &per_cpu(hrtimer_bases, scpu);
1720        new_base = &__get_cpu_var(hrtimer_bases);
1721        /*
1722         * The caller is globally serialized and nobody else
1723         * takes two locks at once, deadlock is not possible.
1724         */
1725        raw_spin_lock(&new_base->lock);
1726        raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
1727
1728        for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
1729                migrate_hrtimer_list(&old_base->clock_base[i],
1730                                     &new_base->clock_base[i]);
1731        }
1732
1733        raw_spin_unlock(&old_base->lock);
1734        raw_spin_unlock(&new_base->lock);
1735
1736        /* Check, if we got expired work to do */
1737        __hrtimer_peek_ahead_timers();
1738        local_irq_enable();
1739}
1740
1741#endif /* CONFIG_HOTPLUG_CPU */
1742
1743static int hrtimer_cpu_notify(struct notifier_block *self,
1744                                        unsigned long action, void *hcpu)
1745{
1746        int scpu = (long)hcpu;
1747
1748        switch (action) {
1749
1750        case CPU_UP_PREPARE:
1751        case CPU_UP_PREPARE_FROZEN:
1752                init_hrtimers_cpu(scpu);
1753                break;
1754
1755#ifdef CONFIG_HOTPLUG_CPU
1756        case CPU_DYING:
1757        case CPU_DYING_FROZEN:
1758                clockevents_notify(CLOCK_EVT_NOTIFY_CPU_DYING, &scpu);
1759                break;
1760        case CPU_DEAD:
1761        case CPU_DEAD_FROZEN:
1762        {
1763                clockevents_notify(CLOCK_EVT_NOTIFY_CPU_DEAD, &scpu);
1764                migrate_hrtimers(scpu);
1765                break;
1766        }
1767#endif
1768
1769        default:
1770                break;
1771        }
1772
1773        return NOTIFY_OK;
1774}
1775
1776static struct notifier_block hrtimers_nb = {
1777        .notifier_call = hrtimer_cpu_notify,
1778};
1779
1780void __init hrtimers_init(void)
1781{
1782        hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
1783                          (void *)(long)smp_processor_id());
1784        register_cpu_notifier(&hrtimers_nb);
1785#ifdef CONFIG_HIGH_RES_TIMERS
1786        open_softirq(HRTIMER_SOFTIRQ, run_hrtimer_softirq);
1787#endif
1788}
1789
1790/**
1791 * schedule_hrtimeout_range_clock - sleep until timeout
1792 * @expires:    timeout value (ktime_t)
1793 * @delta:      slack in expires timeout (ktime_t)
1794 * @mode:       timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
1795 * @clock:      timer clock, CLOCK_MONOTONIC or CLOCK_REALTIME
1796 */
1797int __sched
1798schedule_hrtimeout_range_clock(ktime_t *expires, unsigned long delta,
1799                               const enum hrtimer_mode mode, int clock)
1800{
1801        struct hrtimer_sleeper t;
1802
1803        /*
1804         * Optimize when a zero timeout value is given. It does not
1805         * matter whether this is an absolute or a relative time.
1806         */
1807        if (expires && !expires->tv64) {
1808                __set_current_state(TASK_RUNNING);
1809                return 0;
1810        }
1811
1812        /*
1813         * A NULL parameter means "infinite"
1814         */
1815        if (!expires) {
1816                schedule();
1817                __set_current_state(TASK_RUNNING);
1818                return -EINTR;
1819        }
1820
1821        hrtimer_init_on_stack(&t.timer, clock, mode);
1822        hrtimer_set_expires_range_ns(&t.timer, *expires, delta);
1823
1824        hrtimer_init_sleeper(&t, current);
1825
1826        hrtimer_start_expires(&t.timer, mode);
1827        if (!hrtimer_active(&t.timer))
1828                t.task = NULL;
1829
1830        if (likely(t.task))
1831                schedule();
1832
1833        hrtimer_cancel(&t.timer);
1834        destroy_hrtimer_on_stack(&t.timer);
1835
1836        __set_current_state(TASK_RUNNING);
1837
1838        return !t.task ? 0 : -EINTR;
1839}
1840
1841/**
1842 * schedule_hrtimeout_range - sleep until timeout
1843 * @expires:    timeout value (ktime_t)
1844 * @delta:      slack in expires timeout (ktime_t)
1845 * @mode:       timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
1846 *
1847 * Make the current task sleep until the given expiry time has
1848 * elapsed. The routine will return immediately unless
1849 * the current task state has been set (see set_current_state()).
1850 *
1851 * The @delta argument gives the kernel the freedom to schedule the
1852 * actual wakeup to a time that is both power and performance friendly.
1853 * The kernel give the normal best effort behavior for "@expires+@delta",
1854 * but may decide to fire the timer earlier, but no earlier than @expires.
1855 *
1856 * You can set the task state as follows -
1857 *
1858 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
1859 * pass before the routine returns.
1860 *
1861 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
1862 * delivered to the current task.
1863 *
1864 * The current task state is guaranteed to be TASK_RUNNING when this
1865 * routine returns.
1866 *
1867 * Returns 0 when the timer has expired otherwise -EINTR
1868 */
1869int __sched schedule_hrtimeout_range(ktime_t *expires, unsigned long delta,
1870                                     const enum hrtimer_mode mode)
1871{
1872        return schedule_hrtimeout_range_clock(expires, delta, mode,
1873                                              CLOCK_MONOTONIC);
1874}
1875EXPORT_SYMBOL_GPL(schedule_hrtimeout_range);
1876
1877/**
1878 * schedule_hrtimeout - sleep until timeout
1879 * @expires:    timeout value (ktime_t)
1880 * @mode:       timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
1881 *
1882 * Make the current task sleep until the given expiry time has
1883 * elapsed. The routine will return immediately unless
1884 * the current task state has been set (see set_current_state()).
1885 *
1886 * You can set the task state as follows -
1887 *
1888 * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
1889 * pass before the routine returns.
1890 *
1891 * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
1892 * delivered to the current task.
1893 *
1894 * The current task state is guaranteed to be TASK_RUNNING when this
1895 * routine returns.
1896 *
1897 * Returns 0 when the timer has expired otherwise -EINTR
1898 */
1899int __sched schedule_hrtimeout(ktime_t *expires,
1900                               const enum hrtimer_mode mode)
1901{
1902        return schedule_hrtimeout_range(expires, 0, mode);
1903}
1904EXPORT_SYMBOL_GPL(schedule_hrtimeout);
1905
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.