bug#49261: Segfault during loadup

[Paul Eggert]

On 7/7/21 5:09 PM, Lars Ingebrigtsen wrote:
> There seems to be a corruption in the hash tables somewhere.  It's
> totally reproducible with the recipe

Thanks for the recipe; it let me reproduce the problem on Fedora 34 x86-64.

The problem comes from the fact that mark_maybe_pointer works 
differently on pdumper objects than it works on ordinary objects. On 
ordinary objects, roots can point anywhere into an object (because this 
sort of thing has happened on real machines), whereas on pdumper 
objects, roots had to point to the start of the object.

I worked around this particular problem changing mark_maybe_pointer so 
that pdumper roots can also be tagged (see first attached patch). 
However, I suspect this is not a complete fix, as it doesn't cover the 
case where a root points to some part of a pdumper object that is not at 
the object's start. I added a FIXME about this. Perhaps Daniel can take 
a look at it sometime. I think the remaining bug will be hit only rarely 
(if ever).

The second attached patch is in the same area, but is not part of the 
fix. It causes the GC to be a bit more accurate (i.e., less 
conservative) for roots, which can help avoid some leaks.

0001-Fix-pdumper-related-GC-bug.patch
Description: Text Data

0002-Make-pdumper-marking-pickier.patch
Description: Text Data