deferred hardlink failure

[cgwalters]

We're seeing a case in bootc where the deferred hardlink processing is failing on a particular layer.

The layer in question is over 10G ($reasons) so that was really unwieldly to debug. I ended up writing this https://gist.github.com/cgwalters/f9e9afcc8d94572554184c3b2d5b054e which trims down the layer to a mere 200M. I'm going to try minimizing it here further, but I now have it at least so

$ ostree commit --fsync=false -b test --tree=tar=meta.tar

reliably fails. Will update this issue.

[cgwalters]

OK this is related to containers/buildah#6297. The default for most tar serialization tools (including buildah prior to that issue) is to not create "hardlink chains" but have "direct" links.

This is clearly something we want to fix in ostree. I'm looking at it. But the buildah change is on track to be reverted, and so anyone hit by this should update buildah.