Using pthread_once() in libc

[John Baldwin]

On Thursday 19 November 2009 3:00:19 pm Daniel Eischen wrote:
> On Thu, 19 Nov 2009, John Baldwin wrote:
> 
> > On Thursday 19 November 2009 12:09:33 pm Daniel Eischen wrote:
> >> On Thu, 19 Nov 2009, John Baldwin wrote:
> >>
> >>> On Thursday 19 November 2009 11:48:54 am Daniel Eischen wrote:
> >>>> On Thu, 19 Nov 2009, John Baldwin wrote:
> >>>>
> >>>>> I would like to provide a pthread_once()-like facility in libc that library
> >>>>> bits can use to initialize data safely rather than trying to home-roll their
> >>>>> own variants (see the recent commit to stdtime in libc).  Ideally what I
> >>>>> would like to do is have libc use the "real" pthread_once() when libthr is
> >>>>> linked in and fall back to a simple stub without libthr linked in.  I know we
> >>>>> already do something like this for _spinlock() and friends.  My question is
> >>>>> what is the most correct way to do this?  Should libc grow a new _once()
> >>>>> symbol ala _spinlock() that is a weak symbol to a stub version and
> >>>>> pthread_once() in thr_once.c would override that, or should there be a
> >>>>> _pthread_once() in libc that is a stub in place of the current stub_zero?  I
> >>>>> noticed a comment in thr_spinlock.c saying the spinlock stuff is kept for
> >>>>> backwards compat.  Does this mean that for the future we would like to expose
> >>>>> pthread symbols directly in libc?  Meaning would we rather have libc export a
> >>>>> pthread_once() and that ideally libc would be using pthread_mutex_lock/unlock
> >>>>> instead of _spinlock/unlock?
> >>>>
> >>>> pthread_once() is already a stub in libc that gets overloaded with the
> >>>> real thing when libthr is linked.  See libc/gen/_pthread_stubs.c.
> >>>> Isn't that what you want or does it not serve your purpose?
> >>>
> >>> Hmm, the libc stub will never run the init routine.  I would like to do
> >>> something like this:
> >>
> >> Well, I suppose you could do that.  But what happens if libthr gets
> >> dlopen()'d and your once function needs to initialize a mutex or
> >> something that can only be properly done by a real threads library?
> >> Can we envision a scenario where that would be a problem?
> >
> > Hmmm, so I guess __is_threaded is how the dlopen() case is handled now for
> > mutex lock/unlock as that avoids resolving the symbol until pthread_create()
> > has been invoked?  I guess then we could take an approach that works
> > something like this:
> >
> > /* libc-internal function */
> > int
> > _once(pthread_once_t *once_control, void (*init_routine)(void))
> > {
> >
> > 	if (__is_threaded)
> > 		return (_pthread_once(once_control, init_routine));
> >
> > 	return (_stub_once(once_control, init_routine));
> > }
> >
> > It is probably still a good idea to have the stub_once() patch I think so
> > that pthread_once() DTRT in a single-threaded program.
> 
> I guess that works.

Ok, after a few rounds with the compiler the attached patch actually finished
a buildworld.  Do you have any feedback on it?

-- 
John Baldwin
-------------- next part --------------
A non-text attachment was scrubbed...
Name: stdtime.patch
Type: text/x-diff
Size: 4684 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-threads/attachments/20091119/b640be99/stdtime.bin