Re: Design hole in nptl (and more general in POSIX threads)

<[email protected]> writes:
> ...
> > 
> > When a thread registers itself with the collector, it adds its own
> > __thread structure to a global linked list.  The collecting thread
> > finds other threads' structures via that global linked list.  So
> > there's no problem there.
> 
>   What I see as "problem" here is: to know how to find whatever thread
> specific data, runtime must (minimally) use pthread_self() to find for
> which thread it must lookup; or some data in TLS for same thing. There
> is no way __thread can pull this out of thin air - so it's no more async
> safe than safer one of these two methods.

Okay, I see what you were getting at now.

I think the real answer is that, since no standard specifies the
semantics of __thread, it's only meaningful to ask whether a specific
implementation of __thread is async-safe.  (Thus the form of Roland's
answer.)