[Sbcl-devel] The OpenBSD port finally works again.

[William H. N. <wil...@ai...>]

(The original message I'm replying to was on sbcl-help, so I'm posting
this there. But I've also talked about this problem on sbcl-devel, so
I'm posting it there too.)

On Tue, May 22, 2001 at 11:34:16AM -0500, William Harold Newman wrote:
> On Thu, May 17, 2001 at 07:57:50AM -0500, William Harold Newman wrote:
> > On Wed, May 16, 2001 at 04:02:34PM +0200, Hannah Schroeter wrote:
> > > Yes, but as "compensation" it seems that it doesn't work on BSD,
> > > especially OpenBSD, anymore.
> [..]
> > After that, OpenBSD remains a reasonably high priority. Unfortunately,
> > the problems begun last week have been a rather humbling experience --
> > it's remarkable how confused I've gotten.. -- and I don't know whether
> > I'll get tripped up by more OpenBSD kernel bugs, so I don't have much
> > of a guess whether the OpenBSD fixes will be done today or this
> > weekend or what.
[..] 
> I'll probably continue investigating the problem by trying to learn
> why the garbage collector is complaining after warm init saves
> sbcl.core. With luck, I can localize that problem in a manageable
> region by commenting out more and more of warm init until it goes
> away, then poke around with gdb or assertions or something to see
> what's going on. But I don't know how long that will take, or how many
> hours per day I'll devote to it. OS panics are discouraging..

sbcl-0.6.12.17, now committed to CVS, works on OpenBSD again, for the
first time since the Alpha patches were committed.

The necessary fix was to change the memory map, moving the control
stack out of its previous address range.

The bug was in fact found as suggested above, by figuring out why the
GC was complaining after warm init saves sbcl.core. The process was
complicated by my attempts to use gdb to do so, which never worked.
Eventually I just found it by bisection, rebuilding smaller and
smaller subsets of the program in order to find what functionality was
required to cause the corruption to occur.

(Judging from the behavior of the new trymap utility, the same address
space conflict has existed at least since OpenBSD 2.7 (the oldest
version I can test conveniently). So it's unclear to me how the
pre-Alpha-patches versions of SBCL managed to work, since they all
used the same address space mapping.)

-- 
William Harold Newman <wil...@ai...>
"The beatings will continue until morale improves." -- ??
PGP key fingerprint 85 CE 1C BA 79 8D 51 8C  B9 25 FB EE E0 C3 E5 7C