JVM源码分析之System.currentTimeMillis及nanoTime原理详解

Posted 2021-01-07 jpfss

JDK7和JDK8下的System.nanoTime()输出完全不一样，而且差距还非常大，是不是两个版本里的实现不一样，之前我也没注意过这个细节，觉得非常奇怪，于是自己也在本地mac机器上马上测试了一下，得到如下输出：

~/Documents/workspace/Test/src ? /Library/Java/JavaVirtualMachines/jdk1.7.0_79.jdk/Contents/Home/bin/java NanosTest

1480265318432558000

~/Documents/workspace/Test/src ? /Library/Java/JavaVirtualMachines/jdk1.8.0_101.jdk/Contents/Home/bin/java NanosTest

1188453233877

还真不一样，于是我再到linux下跑了一把，发现两个版本下的值基本上差不多的，也就是主要是mac下的实现可能不一样

于是我又调用System.currentTimeMillis()，发现其输出结果和System.nanoTime()也完全不是1000000倍的比例

~/Documents/workspace/Test/src ? /Library/Java/JavaVirtualMachines/jdk1.8.0_101.jdk/Contents/Home/bin/java NanosTest

1563115443175

1480265707257

另外System.nanoTime()输出的到底是什么东西，这个数字好奇怪

这三个小细节平时没有留意，好奇心作祟，于是马上想一查究竟

再列下主要想理清楚的三个问题

·         在mac下发现System.nanoTime()在JDK7和JDK8下输出的值怎么完全不一样

·         System.nanoTime()的值很奇怪，究竟是怎么算出来的

·         System.currentTimeMillis()为何不是System.nanoTime()的1000000倍

MAC不同JDK版本下nanoTime实现异同

在mac下，首先看JDK7的nanoTime实现

jlong os::javaTimeNanos() {

  if (Bsd::supports_monotonic_clock()) {

    struct timespec tp;

    int status = Bsd::clock_gettime(CLOCK_MONOTONIC, &tp);

    assert(status == 0, "gettime error");

    jlong result = jlong(tp.tv_sec) * (1000 * 1000 * 1000) + jlong(tp.tv_nsec);

    return result;

  } else {

    timeval time;

    int status = gettimeofday(&time, NULL);

    assert(status != -1, "bsd error");

    jlong usecs = jlong(time.tv_sec) * (1000 * 1000) + jlong(time.tv_usec);

    return 1000 * usecs;

  }

}

 

再来看JDK8下的实现

#ifdef __APPLE__

 

jlong os::javaTimeNanos() {

    const uint64_t tm = mach_absolute_time();

    const uint64_t now = (tm * Bsd::_timebase_info.numer) / Bsd::_timebase_info.denom;

    const uint64_t prev = Bsd::_max_abstime;

    if (now <= prev) {

      return prev;   // same or retrograde time;

    }

    const uint64_t obsv = Atomic::cmpxchg(now, (volatile jlong*)&Bsd::_max_abstime, prev);

    assert(obsv >= prev, "invariant");   // Monotonicity

    // If the CAS succeeded then we‘re done and return "now".

    // If the CAS failed and the observed value "obsv" is >= now then

    // we should return "obsv".  If the CAS failed and now > obsv > prv then

    // some other thread raced this thread and installed a new value, in which case

    // we could either (a) retry the entire operation, (b) retry trying to install now

    // or (c) just return obsv.  We use (c).   No loop is required although in some cases

    // we might discard a higher "now" value in deference to a slightly lower but freshly

    // installed obsv value.   That‘s entirely benign -- it admits no new orderings compared

    // to (a) or (b) -- and greatly reduces coherence traffic.

    // We might also condition (c) on the magnitude of the delta between obsv and now.

    // Avoiding excessive CAS operations to hot RW locations is critical.

    // See https://blogs.oracle.com/dave/entry/cas_and_cache_trivia_invalidate

    return (prev == obsv) ? now : obsv;

}

 

#else // __APPLE__

果然发现JDK8下多了一个__APPLE__宏下定义的实现，和JDK7及之前的版本的实现是不一样的，不过其他BSD系统是一样的，只是macos有点不一样，因为平时咱们主要使用的环境还是Linux为主，因此对于macos下具体异同就不做过多解释了，有兴趣的自己去研究一下。

Linux下nanoTime的实现

在linux下JDK7和JDK8的实现都是一样的

jlong os::javaTimeNanos() {

  if (Linux::supports_monotonic_clock()) {

    struct timespec tp;

    int status = Linux::clock_gettime(CLOCK_MONOTONIC, &tp);

    assert(status == 0, "gettime error");

    jlong result = jlong(tp.tv_sec) * (1000 * 1000 * 1000) + jlong(tp.tv_nsec);

    return result;

  } else {

    timeval time;

    int status = gettimeofday(&time, NULL);

    assert(status != -1, "linux error");

    jlong usecs = jlong(time.tv_sec) * (1000 * 1000) + jlong(time.tv_usec);

    return 1000 * usecs;

  }

}

而Linux::supports_monotonic_clock决定了走哪个具体的分支

static inline bool supports_monotonic_clock() {

    return _clock_gettime != NULL;

}

_clock_gettime的定义在

void os::Linux::clock_init() {

  // we do dlopen‘s in this particular order due to bug in linux

  // dynamical loader (see 6348968) leading to crash on exit

  void* handle = dlopen("librt.so.1", RTLD_LAZY);

  if (handle == NULL) {

    handle = dlopen("librt.so", RTLD_LAZY);

  }

 

  if (handle) {

    int (*clock_getres_func)(clockid_t, struct timespec*) =

           (int(*)(clockid_t, struct timespec*))dlsym(handle, "clock_getres");

    int (*clock_gettime_func)(clockid_t, struct timespec*) =

           (int(*)(clockid_t, struct timespec*))dlsym(handle, "clock_gettime");

    if (clock_getres_func && clock_gettime_func) {

      // See if monotonic clock is supported by the kernel. Note that some

      // early implementations simply return kernel jiffies (updated every

      // 1/100 or 1/1000 second). It would be bad to use such a low res clock

      // for nano time (though the monotonic property is still nice to have).

      // It‘s fixed in newer kernels, however clock_getres() still returns

      // 1/HZ. We check if clock_getres() works, but will ignore its reported

      // resolution for now. Hopefully as people move to new kernels, this

      // won‘t be a problem.

      struct timespec res;

      struct timespec tp;

      if (clock_getres_func (CLOCK_MONOTONIC, &res) == 0 &&

          clock_gettime_func(CLOCK_MONOTONIC, &tp)  == 0) {

        // yes, monotonic clock is supported

        _clock_gettime = clock_gettime_func;

        return;

      } else {

        // close librt if there is no monotonic clock

        dlclose(handle);

      }

    }

  }

  warning("No monotonic clock was available - timed services may "

          "be adversely affected if the time-of-day clock changes");

}

说白了，其实就是看librt.so.1或者librt.so中是否定义了clock_gettime函数，如果定义了，就直接调用这个函数来获取时间，注意下上面的传给clock_gettime的一个参数是CLOCK_MONOTONIC，至于这个参数的作用后面会说，这个函数在glibc中有定义

/* Get current value of CLOCK and store it in TP.  */

int

__clock_gettime (clockid_t clock_id, struct timespec *tp)

{

  int retval = -1;

 

  switch (clock_id)

    {

#ifdef SYSDEP_GETTIME

      SYSDEP_GETTIME;

#endif

 

#ifndef HANDLED_REALTIME

    case CLOCK_REALTIME:

      {

    struct timeval tv;

    retval = gettimeofday (&tv, NULL);

    if (retval == 0)

      TIMEVAL_TO_TIMESPEC (&tv, tp);

      }

      break;

#endif

 

    default:

#ifdef SYSDEP_GETTIME_CPU

      SYSDEP_GETTIME_CPU (clock_id, tp);

#endif

#if HP_TIMING_AVAIL

      if ((clock_id & ((1 << CLOCK_IDFIELD_SIZE) - 1))

      == CLOCK_THREAD_CPUTIME_ID)

    retval = hp_timing_gettime (clock_id, tp);

      else

#endif

    __set_errno (EINVAL);

      break;

 

#if HP_TIMING_AVAIL && !defined HANDLED_CPUTIME

    case CLOCK_PROCESS_CPUTIME_ID:

      retval = hp_timing_gettime (clock_id, tp);

      break;

#endif

    }

 

  return retval;

}

weak_alias (__clock_gettime, clock_gettime)

libc_hidden_def (__clock_gettime)

而对应的宏SYSDEP_GETTIME定义如下：

#define SYSDEP_GETTIME

  SYSDEP_GETTIME_CPUTIME;                             

  case CLOCK_REALTIME:                                

  case CLOCK_MONOTONIC:                               

    retval = INLINE_VSYSCALL (clock_gettime, 2, clock_id, tp);           

    break

 

/* We handled the REALTIME clock here.  */

#define HANDLED_REALTIME    1

#define HANDLED_CPUTIME 1

 

#define SYSDEP_GETTIME_CPU(clock_id, tp)

  retval = INLINE_VSYSCALL (clock_gettime, 2, clock_id, tp);

  break

#define SYSDEP_GETTIME_CPUTIME  /* Default catches them too.  */

最终是调用的clock_gettime系统调用：

int clock_gettime(clockid_t, struct timespec *)

    __attribute__((weak, alias("__vdso_clock_gettime")));

 

 

notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)

{

    if (likely(gtod->sysctl_enabled))

        switch (clock) {

        case CLOCK_REALTIME:

            if (likely(gtod->clock.vread))

                return do_realtime(ts);

            break;

        case CLOCK_MONOTONIC:

            if (likely(gtod->clock.vread))

                return do_monotonic(ts);

            break;

        case CLOCK_REALTIME_COARSE:

            return do_realtime_coarse(ts);

        case CLOCK_MONOTONIC_COARSE:

            return do_monotonic_coarse(ts);

        }

    return vdso_fallback_gettime(clock, ts);

}  

而我们JVM里取纳秒数时传入的是CLOCK_MONOTONIC这个参数，因此会调用如下的方法

notrace static noinline int do_monotonic(struct timespec *ts)

{

    unsigned long seq, ns, secs;

    do {

        seq = read_seqbegin(&gtod->lock);

        secs = gtod->wall_time_sec;

        ns = gtod->wall_time_nsec + vgetns();

        secs += gtod->wall_to_monotonic.tv_sec;

        ns += gtod->wall_to_monotonic.tv_nsec;

    } while (unlikely(read_seqretry(&gtod->lock, seq)));

    vset_normalized_timespec(ts, secs, ns);

    return 0;

}

上面的wall_to_monotonic的tv_sec以及tv_nsec都是负数，在系统启动初始化的时候设置，记录了启动的时间

void __init timekeeping_init(void)

{

    struct clocksource *clock;

    unsigned long flags;

    struct timespec now, boot;

 

    read_persistent_clock(&now);

    read_boot_clock(&boot);

 

    write_seqlock_irqsave(&xtime_lock, flags);

 

    ntp_init();

 

    clock = clocksource_default_clock();

    if (clock->enable)

        clock->enable(clock);

    timekeeper_setup_internals(clock);

 

    xtime.tv_sec = now.tv_sec;

    xtime.tv_nsec = now.tv_nsec;

    raw_time.tv_sec = 0;

    raw_time.tv_nsec = 0;

    if (boot.tv_sec == 0 && boot.tv_nsec == 0) {

        boot.tv_sec = xtime.tv_sec;

        boot.tv_nsec = xtime.tv_nsec;

    }

    set_normalized_timespec(&wall_to_monotonic,

                -boot.tv_sec, -boot.tv_nsec);

    total_sleep_time.tv_sec = 0;

    total_sleep_time.tv_nsec = 0;

    write_sequnlock_irqrestore(&xtime_lock, flags);

}

因此nanoTime其实算出来的是一个相对的时间，相对于系统启动的时候的时间

Java里currentTimeMillis的实现

我们其实可以写一个简单的例子从侧面来验证currentTimeMillis返回的到底是什么值

    public static void main(String args[]) {

        System.out.println(new Date().getTime()-new Date(0).getTime());

        System.out.println(System.currentTimeMillis());

    }

你将看到输出结果会是两个一样的值，这说明了什么？另外new Date(0).getTime()其实就是1970/01/01 08:00:00,而new Date().getTime()是返回的当前时间，两个日期一减，其实就是当前时间距离1970/01/01 08:00:00有多少毫秒，而System.currentTimeMillis()返回的正好是这个值，也就是说System.currentTimeMillis()就是返回的当前时间距离1970/01/01 08:00:00的毫秒数。

就实现上来说，currentTimeMillis其实是通过gettimeofday来实现的

jlong os::javaTimeMillis() {

  timeval time;

  int status = gettimeofday(&time, NULL);

  assert(status != -1, "linux error");

  return jlong(time.tv_sec) * 1000  +  jlong(time.tv_usec / 1000);

}

至此应该大家也清楚了，为什么currentTimeMillis返回的值并不是nanoTime返回的值的1000000倍左右了，因为两个值的参照不一样，所以没有可比性