Date: Sun, 15 Aug 2004 15:53:32 +0100 From: Doug Rabson <dfr@nlsystems.com> To: Daniel Eischen <deischen@freebsd.org> Cc: freebsd-current@freebsd.org Subject: Re: New nvidia drivers available Message-ID: <200408151553.32463.dfr@nlsystems.com> In-Reply-To: <Pine.GSO.4.43.0408151001050.15254-100000@sea.ntplx.net> References: <Pine.GSO.4.43.0408151001050.15254-100000@sea.ntplx.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sunday 15 August 2004 15:03, Daniel Eischen wrote:
> On Sun, 15 Aug 2004, Doug Rabson wrote:
> > On Sunday 15 August 2004 14:47, Daniel Eischen wrote:
> > > On Sun, 15 Aug 2004, Doug Rabson wrote:
> > > > > Same error message here when starting neverball. If I map
> > > > > libpthread->libc_r it's working again.
> > > > >
> > > > > xawtv also stopped working. Fortunately setting -xvport
> > > > > manually did fix that.
> > > >
> > > > This might be because libGL calls libpthread's version of
> > > > open() before libpthread has initialised properly. This patch
> > > > might fix it - it fixes neverball's map compiler for me but I
> > > > haven't actually run neverball itself.
> > > >
> > > > Index: thr_open.c
> > > > ===============================================================
> > > >==== RCS file: /home/ncvs/src/lib/libpthread/thread/thr_open.c,v
> > > > retrieving revision 1.16
> > > > diff -u -r1.16 thr_open.c
> > > > --- thr_open.c 9 Dec 2003 02:20:56 -0000 1.16
> > > > +++ thr_open.c 15 Aug 2004 09:19:42 -0000
> > > > @@ -45,11 +45,15 @@
> > > > int
> > > > __open(const char *path, int flags,...)
> > > > {
> > > > - struct pthread *curthread = _get_curthread();
> > > > + struct pthread *curthread;
> > > > int ret;
> > > > int mode = 0;
> > > > va_list ap;
> > > >
> > > > + if (_thr_initial == NULL)
> > > > + _libpthread_init(NULL);
> > > > +
> > > > + curthread = _get_curthread();
> > >
> > > I thought the C++ style constructor in thr_autoinit.c is supposed
> > > to take care of things like this?
> >
> > The problem is that there is no particular ordering for
> > constructors. In the case with neverball, the C++ constructor in
> > libGL which initialises OpenGL ran first and quite reasonably tried
> > to call open(2). This was intercepted by libpthread, which hadn't
> > yet had its constructor called.
>
> Is there something else we can do so that libpthread gets
> initialized first? Use _init()? I'm not sure how to
> not add that when building static libpthread though (let's
> kill static libpthread!).
There is no way other than putting a special case for pthread
initialisation into rtld. The C++ constructor trick is implemented via
_init anyway so that won't work either.
I have the same problem initialising TLS in static programs. My TLS
initialisation must run before pthread initialises so that I can
allocate the correct sized blocks for pthread's main thread. I ended up
hooking the TLS initialisation directly from crt1.c before calling
_init.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200408151553.32463.dfr>