Re: GDB cannot access memory after Emacs abort

[Michael Snyder <msnyder at specifix dot com>]

On Mon, 2007-11-12 at 00:01 +0100, Stephen Berman wrote:
> On Sat, 10 Nov 2007 22:38:14 -0800 Michael Snyder <msnyder@specifix.com> wrote:
> 
> > Hi Stephen, 
> >
> > See questions inline:
> >
> > On Sun, 2007-11-11 at 00:42 +0100, Stephen Berman wrote:
> >> I recently experienced an abort in CVS Emacs and was unable to get a
> >> backtrace from GDB.  The Emacs bug causing the abort was fixed, but
> >> Richard Stallman responded to the lack of a backtrace with this comment:
> >> 
> >> > That could be a serious problem in GDB.  It would be good to talk
> >> > with GDB developers about how to investigate it.  Since the bug's
> >> > cause is known, they could focus on figuring out why GDB fails
> >> > to give a backtrace.
> >
> > Out of curiosity, and because RMS seems to think it's
> > relevant -- what is the bug's cause?
> 
> It had to do with the handling of named icons in GTK+ tool bars.  I
> don't know the code well enough to say more, but the diff is here:
> http://cvs.savannah.gnu.org/viewvc/emacs/emacs/src/gtkutil.c?r1=1.120&r2=1.121&pathrev=MAIN 

Fair enough -- I think that's out of my depth.  ;-)

> >> Here is the GDB-relevant part of my bug report about the abort (Emacs
> >> was built using the GTK+ toolkit):
> >
> > What's your host architecture?  OS?  How is gdb configured (host-target
> > tuple)?
> 
> Linux escher 2.6.22.12-0.1-default i686 athlon i386 GNU/Linux (openSUSE
> 10.3)
> 
> GNU gdb 6.6.50.20070726-cvs
> This GDB was configured as "i586-suse-linux".

Thanks, that's all very helpful.



> 
> > Making sure that I understand -- you ran emacs under gdb, 
> > you set a breakpoint at abort, you hit the breakpoint -- 
> > and your desktop is locked up?
> 
> Yes (as Eli Zaretskii pointed out, Emacs set a breakpoint at abort).
> 
> > That seems unusual -- do you have any idea of the cause?
> 
> No!  I'm hoping someone here might have some insight.

OK -- just hoping for more info.  

> > Is it possible that emacs is in an infinite recursion
> > and has consumed all of virtual memory, or something
> > of the sort?
> 
> This has happened on (rare) occasion, but it never locked up the
> desktop, I could always at least kill -9 the emacs process from within
> the X window system (in the case under discussion, I was able to switch
> to a virtual tty and kill -9 the emacs process from there, but X was
> locked up solid).

Yeah, that's pretty rare in my experience -- but I'm not
a GUI or X hacker.

Seems like something that shouldn't happen, unless thru a
bug in X.  No client app should be able to take control of
the entire windowing system and prevent anything else from
getting access.

Please note, I'm not at all trying to say "this isn't 
our (gdb's) problem".  

> >> >   Cannot access memory at address 0x8321b6c
> >
> > Is that a valid address for your architecture?
> 
> How can I determine that?

Just thought you might know.  Forget about it.

> Anyway, it sound like you don't suspect a bug in GDB that prevented
> getting a backtrace, or is that still a possibility?

Oh, I didn't mean to imply that at all.
Just asking for more information.

So just to define the problem ...

1) emacs calls abort.  This is apparently due to a bug
   in emacs, for which you already have a patch.

2) While emacs is held at the abort by gdb, your X system
   is frozen and you can't get any other X window client to
   work.  We don't know the cause of this, but it probably
   isn't gdb, so let's forget about it in this context.

3) Within gdb, when you're at the breakpoint at abort, 
   backtrace doesn't work.  This part is within our domain.

Now 3 could well be a bug in gdb, but there are other
possibilities.  Something could have corrupted the stack, 
so badly that gdb can't unwind it.

Personally I don't see how to decide that question, based
on the information we have so far.  Maybe Daniel and/or Eli
might have an idea?

To pursue it further, we can go one of two ways:

A) maybe you can provide us with enough information and
context to reproduce the problem ourselves?  This seems 
unlikely, but maybe, for instance, you know that with 
a certain released version of emacs and a certain released
version of linux, you can give a fixed sequence of commands
and reliably reproduce the crash?

or 

B) we can keep asking you for more information, question
and answer style.

For instance, I'd like to know the output that you get
from the following gdb commands when you're at the breakpoint:

i) info registers
ii) info target
iii) x /64x $esp