Subj : Re: Did pthread_once init_routine succeed?
To   : comp.programming.threads
From : gottlobfrege
Date : Wed Feb 23 2005 07:33 pm


David Schwartz wrote:
> "Ian Pilcher" <i.pilcher@comcast.net> wrote in message
> news:ksidnY-GyNgDkoDfRVn-2w@comcast.com...
>
> > Is it safe to use a file-scope variable to report the result of an
> > init_routine called via pthread_once?  I believe that it is, if the
> > following conditions are met:
>
>     Yes. What would be the use of pthread_once if you couldn't modify
memory
> in the initialization routine and access the results after
pthread_once
> returns? (Think about it, that's about all pthread_once is good for.)
>
> >   * The variable is only set by the init_routine.
> >
> >   * The variable is only read immediately following a call to
> >     pthread_once (with the appropriate parameters).
> >
> > It seems to me that one of three things can happen when a thread
hits a
> > call to pthread_once:
> >
> >   1. The thread is the "first" to call pthread_once (with these
> >      parameters).  It will execute the init_routine, which sets the
> >      value of the variable, and then return from pthread_once.  The
> >      value of the variable is determinate when the thread reads it.
> >
> >   2. The thread is not the "first" to call pthread_once, but the
> >      init_routine has not yet returned (in whatever thread it is
> >      executing).  The thread will block until the other thread
> >      completes the init_routine (setting the variable).  The value
> >      of the variable is determinate when the thread reads it.
> >
> >   3. The thread is not the "first" to call pthread_once, and
> >      init_routine has already completed (in this or another
thread).
> >      The variable was set by init_routine, so its value is
determinate
> >      when the thread reads it.
>
>     Correct, this is the semantics of pthread_once and the reason it
is
> useful at all.
>
>     DS

I agree completely, but one question:

Should we be concerned about the memory-visibility of the global
variable - ie maybe the second thread, on a second processor, will
still be seeing the old value of the global variable, because the CPU's
cache is not in sync with the other thread's CPU's cache.

ie do we need a memory barrier on the read of the global variable?

.