[llvm-dev] call_once and TSan

[Dmitry Vyukov via llvm-dev]

On Fri, Sep 2, 2016 at 2:21 PM, Kuba Brecka <kuba.brecka at gmail.com> wrote:
> Same problem exists, thread A can still be within REAL(call_once), but after it ran user code and set the flag to ~0.  Roughly, call_once does:
>
> __call_once(flag, arg, func) {
>   mutex_lock(mut);
>   if (flag == BEING_INITIALIZED) { wait }
>   else if (flag == NOT_INITIALIZED_AT_ALL) {
>     flag = BEING_INITIALIZED;
>     mutex_unlock(mut);
>     func(arg);    // <=== user code callback
>     mutex_lock(mut);
>     atomic_store(&flag, FULLY_INITIALIZED, mo_release);   // "release" store, but within a compiled dylib, thus invisible to TSan
>   }
>   mutex_unlock(mut);
> }
>
> If thread A is just after the release store, which is invisible to TSan, __tsan_acquire in thread B will have no effect, and the stores from the callback to func(arg) will not be synchronized to thread B.

right

> Anyway, I just realized that we can wrap "func" into our own callback, which will perform the (extra) __tsan_release...  Do you think that would work?  E.g.:
>
> void call_once_callback_wrapper(...) {
>   orig_func(orig_arg);
>   __tsan_release(flag);
> }
>
> INTERCEPTOR(call_once, o, func, arg) {
>   REAL(call_once)(flag, ..., call_once_callback_wrapper);
> }

this should work
but I am obviously bad at thinking today