Re: [PATCH] Cygwin: pthread: Fix a race issue introduced by the commit 2c5433e5da82

Re: [PATCH] Cygwin: pthread: Fix a race issue introduced by the commit 2c5433e5da82

[Bruno Haible]

Takashi Yano wrote in cygwin-patches:
> To avoid race issues, pthread::once() uses pthread_mutex. This caused
> the handle leak which was fixed by the commit 2c5433e5da82. However,
> this fix introduced another race issue, i.e., the mutex may be used
> after it is destroyed. With this patch, do not use pthread_mutex in
> pthread::once() to avoid both issues. Instead, InterlockedExchage()
> is used.

This patch is bogus as well, because it allows one thread to return
from a pthread_once call while the other thread is currently
executing the init_routine and not yet done with it.

> +  if (!InterlockedExchange (&once_control->state, 1))
> +    init_routine ();
> return 0;
> }

There is no code after the init_routine () call here. This means
that other threads are not notified when the init_routine () call
is complete. Therefore this implementation *cannot* be correct.

See: Assume thread1 and thread2 call pthread_once on the same
once_control.

            thread1                      thread2
            -------                      -------

         enters pthread_once       enters pthread_once

         sets state to 1

                                   sees that state == 1

                                   returns from pthread_once

                                   executes code that assumes
                                   init_routine has completed

         starts executing
         init_routine

         finished executing
         init_routine

         returns from pthread_once

Bruno