process.uptime() not correct when adjust time

[yjhjstz]

Now , iojs use relative time to compute uptime[1], but when call settimeofday to change time,
process.time() isn't correct. So, my PR #2572 is to slove this, using time that refer to the kernel boot time.
That's always ok at most platforms.

refer [1]: https://en.wikipedia.org/wiki/Uptime

[bnoordhuis]

but when call settimeofday to change time, process.time isn't correct

process.time is a typo of process.uptime?

On what platform are you seeing this? I'm not sure I understand, libuv uses monotonic clocks - i.e. clocks that are unaffected by settimeofday - whenever possible.

[yjhjstz]

I mean js-function process.time() implemented in static void Uptime(...).

diff --git a/src/node.cc b/src/node.cc
index b900388..1751f09 100644
--- a/src/node.cc
+++ b/src/node.cc
@@ -1972,10 +1972,10 @@ static void Uptime(const FunctionCallbackInfo<Value>& args) {
   Environment* env = Environment::GetCurrent(args);
   double uptime;

-  uv_update_time(env->event_loop());
-  uptime = uv_now(env->event_loop()) - prog_start_time;
+  uv_uptime(&uptime);
+  uptime -= prog_start_time;

-  args.GetReturnValue().Set(Number::New(env->isolate(), uptime / 1000));
+  args.GetReturnValue().Set(Number::New(env->isolate(), uptime));
 }

@@ -3606,8 +3606,7 @@ void Init(int* argc,
           int* exec_argc,
           const char*** exec_argv) {
   // Initialize prog_start_time to get relative uptime.
-  prog_start_time = static_cast<double>(uv_now(uv_default_loop()));
-
+  uv_uptime(&prog_start_time);
   // Make inherited handles noninheritable.

uv_now() can be affected by settimeofday, so I use uv_uptime(). @bnoordhuis

[bnoordhuis]

I mean js-function process.time() implemented in static void Uptime(...).

There is no process.time() but I understand that you mean process.uptime(). I ask because there is also process.hrtime() which does something else.

uv_now() can be affected by settimeofday, so I use uv_uptime().

Can you explain how? Like I said, libuv uses monotonic clocks (think clock_gettime(CLOCK_MONOTONIC)) on most platforms. You should never see the uptime jump backwards, although it is possible for a monotonic clock to slow down in the presence of time adjustments.

On some platforms it doesn't matter if you use uv_now() or uv_uptime() because they both call clock_gettime(). With that in mind, can you explain more about the exact behavior you're seeing and on what platform?

[yjhjstz]

in netbsd,

int uv_uptime(double* uptime) {
  time_t now;
  struct timeval info;
  size_t size = sizeof(info);
  static int which[] = {CTL_KERN, KERN_BOOTTIME};

  if (sysctl(which, 2, &info, &size, NULL, 0))
    return -errno;

  now = time(NULL);

  *uptime = (double)(now - info.tv_sec);
  return 0;
}
uint64_t uv__hrtime(uv_clocktype_t type) {
  struct timespec ts;
  clock_gettime(CLOCK_MONOTONIC, &ts);
  return (((uint64_t) ts.tv_sec) * NANOSEC + ts.tv_nsec);
}

I want to know, why uv_uptime not use clock_gettime, but to use sysctl?

[bnoordhuis]

I don't know the exact reason. Someone (@shigeki?) contributed that code a long time ago.

[mscdex]

All of the BSDs use the same uv_uptime() implementation in libuv. process.uptime() uses a monotonic clock via uv_hrtime() whereas require('os').uptime() uses uv_uptime(). So are you referring to os.uptime() instead of process.uptime()?

[shigeki]

I forgot the detail when I submitted netBSD patch in 3 years ago. Probably I referred FreeBSD implementation of uv_uptime().

I guess it comes from an old implementation of uptime(1) in FreeBSD and it was changed and has been using clock_gettime since 2005 as in freebsd/freebsd-src@74d3dde. We can change it in libuv now.

[yjhjstz]

I think , we can update libuv implementation. And I was wondering why os.uptime() implementation is different from process.uptime() .

[saghul]

FWIW, this landed in libuv, thanks @yjhjstz!

[evanlucas]

This was fixed by the upgrade to libuv 1.7.5. Thanks!