[llvm-dev] call_once and TSan

[Kuba Brecka via llvm-dev]

> On 2 Sep 2016, at 12:11, Dmitry Vyukov <dvyukov at google.com> wrote:
> 
> On Fri, Sep 2, 2016 at 12:09 PM, Kuba Brecka <kuba.brecka at gmail.com> wrote:
>> 
>>> On 2 Sep 2016, at 11:18, Dmitry Vyukov via llvm-dev <llvm-dev at lists.llvm.org> wrote:
>>> 
>>> On Thu, Sep 1, 2016 at 2:30 PM, Kuba Brecka <kuba.brecka at gmail.com> wrote:
>>>> Hi,
>>>> 
>>>> I'm trying to write a TSan interceptor for the C++11 call_once function.  There are currently false positive reports, because the inner __call_once function is located in the (non-instrumented) libcxx library, and on macOS we can't expect the users to build their own instrumented libcxx.
>>>> 
>>>> TSan already supports pthread_once and dispatch_once by having interceptors that re-implement the logic.  However, doing the same for call_once/__call_once doesn't work, because call_once is explicitly supposed to be exception-safe, but the sanitizer runtime libraries disallow exception handling.
>>>> 
>>>> Any ideas how to handle call_once in TSan?
>>> 
>>> Does anybody remember exact reasons we disable exceptions in sanitizer
>>> runtimes? One is that it won't link with C programs. Are there any
>>> other?
>>> If C is the only reason: there is already a part of tsan runtime is
>>> linked only to C++ programs (it contains operator new/delete
>>> interceptors). We could add additional files to the cxx part of
>>> runtime and build them with exceptions.
>>> 
>>> Alternatively, the interceptor can handle only synchronization but
>>> forward actual logic to the real function. Along the lines of:
>>> 
>>> INTERCEPTOR(call_once, o) {
>>> __tsan_acquire_release(o);
>>> REAL(call_once)(o);
>>> }
>>> 
>>> That will have some performance impact. If we hardcode the "fully
>>> initialized" value, then we can eliminate the additional overhead:
>>> 
>>> INTERCEPTOR(call_once, o) {
>>> if (__atomic_load(o, acquire) == FULLY_INITIALIZED) {
>>>    __tsan_acquire(o);
>>>    return;
>>> }
>>> __tsan_acquire_release(o);
>>> REAL(call_once)(o);
>>> }
>> 
>> Unfortunately, the first fast-path check is inlined and cannot be intercepted.  We can only intercept the inner call to __call_once.  But how would __tsan_acquire_release help here?  The issue is that we need to perform the release *after* user code has run, but before the "o" flag is changed.  Otherwise, TSan will still see a false positive where one thread has already run user code, and another thread already sees that call_once is finished, but the acquire has no release to pair with.
> 
> 
> 
> Will then the following work?
> 
> INTERCEPTOR(call_once, o) {
>  REAL(call_once)(o);
>  __tsan_release_merge(o);
>  __tsan_acquire(o);
> }

Still racy.  Suppose thread A is still inside REAL(call_once), but already after it has run user code and updated "o" to ~0.  Thread B load-acquires "o", finds ~0, assumes it's fully initialized, keeps going, but user code stores hasn't been properly published.  

Kuba