Re: [PATCH] Reserve static TLS for dynamically loaded initial-exec TLS only [BZ #25051]

[Florian Weimer <fweimer at redhat dot com>]

* Szabolcs Nagy:

> it has correctness implications too, e.g. in case of
> errno or soft-float fenv emulation (ppc only?) tls
> access may happen in signal handlers and i think
> currently only initial (and local) exec tls is
> guaranteed to be as-safe.

I think you are probably right.  (I don't have the performance numbers
yet.)

> libc also has struct pthread, i wonder if any of that
> needs to be duplicated for dlmopen (is it valid to
> dlmopen a different version of libc.so.6 ?)

Not a different version, but you get a different copy, including a fresh
set of TLS variables (hence the dlmopen limitation).  But obviously not
of struct pthread and the TCB.  Now that Hurd supports THREAD_GETMEM, we
can make a choice whether we want to share per-thread data across the
dlmopen boundary or not.

Thanks,
Florian