starvation in pthread_once?

[Alexander Terekhov]

[... pthread_once() and cancelation ...]

> Should one or more waiting threads get woken up and made to
> re-compete to run the init routine?

Yes. As for the rest, what you need here is DCSI-TLS or DCSI-MBR
(acquire/release memory barriers are needed on both Itanic and
XBOX NEXT... and, apart from hardware memory model, compiler
shall respect acquire/release semantics on IA-32 as well). For
serialization simply use a named mutex associated with address
of control variable and process id (to minimize contention).

DCSI-TLS: (__declspec(thread) for control variable; DLL issues
aside for a moment)

  if (!once_control) {
    named_mutex::guard guard(&once_control);
    if (!once_control) {
      <init>
      once_control = true;
    }
  }

DCSI-MBR: (atomic<> for control variable)

  if (!once_control.load(msync::acq)) {
    named_mutex::guard guard(&once_control);
    if (!once_control.load(msync::none)) {
      <init>
      once_control.store(true, msync::rel);
    }
  }

regards,
alexander.