All my amd64 problems appear to be KSE

[Tim Robbins]

On Sat, Jun 05, 2004 at 12:06:01PM -0400, Daniel Eischen wrote:
> On Sun, 6 Jun 2004, Tim Robbins wrote:
> 
> > On Sat, Jun 05, 2004 at 10:22:58AM -0400, Daniel Eischen wrote:
> > > On Fri, 4 Jun 2004, Sean McNeil wrote:
> > > 
> > > > On Fri, 2004-06-04 at 21:06, Daniel Eischen wrote:
> > > > > On Fri, 4 Jun 2004, Sean McNeil wrote:
> > > > > 
> > > > > > I just switched over to libc_r via. libmap.conf and I no longer get any
> > > > > > mysterious crashes of applications in gnome.  My gnome-terminals and
> > > > > > everything else are just fine now.
> > > > > 
> > > > > How were you using libpthread?  It's best to use libmap.conf
> > > > > to map everything to libpthread so that libc_r isn't pulled
> > > > > in by libraries still linked to it.
> > > > 
> > > > This is a completely rebuilt machine from a working i386 install that
> > > > has tracked -current forever.  There is absolutely nothing that is
> > > > compiled and pointing to libc_r.
> > > 
> > > Sorry, unless you have tried using libmap to map libc_r to
>                                        ^^^^^^^^^^^^^^^^^^^^^^^
> > > libpthread, I don't trust that.
>     ^^^^^^^^^^
> 
> ???
> 
> This is the first thing that I would ask anyone to try, and it's
> very simple to do.  Noone (or do we now spell it as 'no_one' ;-))
> has said, "yes, I've done that and it doesn't help".
> 
> > I can corroborate what Sean is saying. There is definitely something wrong
> > with KSE on amd64 (I haven't tried it on i386.) I've been chasing it down,
> > on-and-off, for about a month, but have come up with nothing so far.
> > gnome-terminal and xmms are two examples of applications that have gone
> > from unusable to usable after changing from libpthread to libc_r. Mozilla
> > may also be affected, but I'm less sure of that. The applications in
> > question just mysteriously vanish, sometimes with signal 10 or 11, sometimes
> > with no signal in the kernel log at all.
> 
> Is it something recent that broke?  Last I knew, libpthread was working
> on amd64 with mozilla & kde.

KDE works flawlessly. Mozilla almost works, but suffers from the "mysterious
vanishing" I mentioned.

If you want to reproduce the problems, either:
(a) Fire up gnome-terminal, open a few new tabs, close them again, repeat
    until it crashes. It often takes only 2-3 invocations of "Open Tab"
    to make it happen.
(b) Load a handful of MP3 files into XMMS, click the next/previous
    track buttons once every few seconds until it crashes.

Everything seems to work fine if I build libpthread with SYSTEM_SCOPE_ONLY
(at least it hasn't crashed so far.) My current guess is that there's a
bug in context manipulation or signals. I initially thought we weren't
saving enough FPU context in _amd64_save_context, but adding an fxsave
in there didn't help.


Tim