When unmarshalling a hand-written string it is possible to break the
invariants of longs:

Python 2.6.2 (release26-maint, Apr 19 2009, 01:56:41)
[GCC 4.3.3] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> x = marshal.loads('l\x02\x00\x00\x00\x00\x00\x00\x00')
>>> print x
00000
>>> x == 0
False
>>> bool(x)
True
>>> x + 1
1L

I would expect this to raise an error instead.

Thanks for the report;  I'll take a look.

Though in general, I think this comes under the heading of 'Don't do
that, then.'  I suspect there are many ways to produce odd behaviour by
feeding malformed input to marshal.  See the big pink boxes at:

http://docs.python.org/library/marshal.html

:-)

Yes, I know :-). I thought I'd report it anyway, as it does more than
"just" give you random behavior, but actually produces a broken object.

So there are two options:  (1) raise an exception, or (2) strip leading 
zero digits and produce a properly normalized Python long.  Do you have 
any strong preferences?  I'm inclined towards (2).

How did you encounter this in the first place?

On second thoughts, I think Carl was right in the first place: it's better 
to raise an exception.  If marshal.loads is getting input that doesn't 
precisely match the marshal.dumps format, then something's wrong 
somewhere.

Looks like we missed the window for 2.6.3.  :-(

Here's a patch for Python 2.6.

And a patch for trunk.  (Significantly different, because marshal for 
Python longs was reworked when 30-bit long digits were added.)

Changing versions:  3.1 and 3.2 also need fixing (the trunk patch should 
be easily mergeable);  it's not a security issue (AFAIK), so won't be 
fixed in 2.5.

Fixed in r75141 (trunk), r75145 (py3k) and r75146 (release31-maint).  I'll 
apply the 2.6 fix once the release26-maint branch is unfrozen.

[...]
> How did you encounter this in the first place?

I was working on PyPy's marshaler, broke it in such a way that it was 
producing the bad input that I gave above. Then I fixed it, but didn't 
kill my .pyc files. So a few days later I got a long zero with the 
properties above in PyPy (which was quite annoying to debug). Then I 
figured that CPython likely has the same problem.

Thanks a lot for fixing this!

Carl Friedrich

2.6 fix applied in r75203.