Re: 5.3-BETA5 AMD64 Python problems

Matt,

I finally got some time to attempt to reproduce this ...

I can trigger the bus error on the last test of bug 418626, which goes
into a recursion loop.  This is the call:

re.search('(a|b)*?c', 10000*'ab'+'cd')

I've also been able to confirm that compiling Python without threads
avoids the crash.

Debugging it from here will be tricky, however; gdb in -current has
problems with stack unwinding and threads, and if the stack is getting
corrupted (which is highly likely) then it'll make things harder than they
already are.

I'll try to poke at this a bit more.

On Sun, 3 Oct 2004, Matt Hamilton wrote:

> Hi All,
>    We have just bought ourselves an AMD64 machine (HP DL145) and I have
> put 5.3-BETA5 on it.  This is our first foray into 5.x, so far we have
> kept all our servers on 4.x.  The server is meant to be a Zope server,
> but I am having trouble with python.
>
> Python 2.3.4 as built from ports cannot pass its own unit tests on this
> platform, specifically it fails test_re.py in about the 4th test
> (something to do with catching infinite recursion on a regular
> expression).  The test just core dumps (as does trying to start Zope).
> I thought this was the well known 'FreeBSD has a tiny pthread stack
> size' issue.  This is fixed in recent pythons (and the ports) by calling
> pthread_setstacksize to increase the stack to 1MB.
>
> I've tried this on 5.3-BETA5/AMD64, but to no avail.  If I compile
> python *without* thread support, then the unit test passes.  So I think
> it is related to threads.  I've not managed to keep my finger on the
> -current development pulse, but wasn't threads re-written in 5.x?  Or
> was it just the scheduler (I am using whichever scheduler 5.3-BETA5 uses
> by default).
>
> I've even tried raising the stack size to 64MB in python, but still it
> core dumps.
>
> Anyone else using python on 5.x?  Can you test it passes its regex unit
> tests?  As currently I'm unsure if the problem is related to:
>
> 1) AMD64 support
> 2) 5.x's new threads and/or scheduler
> 3) something else :)
>
> I'm pretty sure I've heard of people running Zope on 5.1 (I'm guessing
> it was i386) so I'm thinking it is a 64-bit issue.
>
> -Matt
>
>

-- 
Doug White                    |  FreeBSD: The Power to Serve
dwhite_at_gumbysoft.com          |  www.FreeBSD.org