Bug 22360 – WTF current time code uses unprotected static variables

• Component: Web Template Framework
• Status: NEW

[Alexey Proskuryakov]

Recently, a number of static variables were added to JSC::getCurrentUTCTimeWithMicroseconds(). These include doubles, which are not written to memory atomically on common platforms. Likely, this also created race conditions in this function, although I don't have any to pinpoint.

I'm fairly sure that this will make JavaScriptCore date/time functions misbehave when used from multiple concurrent threads.

[Darin Adler]

Since this is Windows-specific code, we can just use Windows thread-specific globals for this. Should be very quick to fix that way.

[Darin Adler]

I mean __declspec(thread).

[Darin Adler]

Steve mentioned that __declspec(thread) doesn't work for delay-loaded libraries (at least on XP) so we probably need to stay away from it.

[Darin Adler]

Generally speaking, for our uses it seems fine for each thread to have its own separate "drift-free" mechanism, so we don't need to introduce locking here -- per-thread data should suffice.

I'm contemplating adding JSGlobalData* arguments into these calls and moving the data there.

[Alexey Proskuryakov]

Someone just mentioned on IRC that they wanted to move this function to WTF, in order to replace gettimeofday in ThreadingPthreads (I don't know why exactly).

[Kevin Watters]

(In reply to comment #5)
> Someone just mentioned on IRC that they wanted to move this function to WTF, in
> order to replace gettimeofday in ThreadingPthreads (I don't know why exactly).
> 

The reason is that the wx port is switching from ThreadingNone.cpp to ThreadingPthreads.cpp, in which the ThreadCondition::timedWait function needs a high resolution timer to calculate a target time to give to pthread_cond_timedwait.

My temporary solution was to implement gettimeofday in that file using the Win32 function GetSystemTimeAsFileTime--but reusing getCurrentUTCTimeWithMicroseconds would work as well if it makes sense to have some time functionality in WTF.

[Gavin Barraclough]

Naturally aligned memory accesses are atomic on X86, so if this is windows specific code, it this an issue?

Alexey, I can't find a method called getCurrentUTCTimeWithMicroseconds - do you know if this has been renamed?

cheers,
G.

[Darin Adler]

It’s the currentTime() functions in CurrentTime.cpp.

[Alexey Proskuryakov]

> Naturally aligned memory accesses are atomic on X86

I didn't know that double accesses were also atomic. I believe x86 also has some guarantees for in-order writes, but code like below still makes me nervous:
    static bool init = false;
    static double lastTime;
    static DWORD lastTickCount;
    if (!init) {
        // ..
    }

On most platforms, monotonicallyIncreasingTime() has a whole static mach_timebase_info_data_t variable, e.g.

    static mach_timebase_info_data_t timebaseInfo;
    if (!timebaseInfo.denom) {

[Alexey Proskuryakov]

> On most platforms

s/most platforms/the Mac/