tarfile.data_filter wrongly rejects some tarballs with symlinks

[encukou]

My implementation of PEP-706 has a bug: it wrongly determines the target of symlinks, and thus wrongly raises LinkOutsideDestinationError on some valid tarballs.

I didn't pay enough attention to this quirk of the format (which I'd like to add to TarInfolinkname docs):

For symbolic links (SYMTYPE), the linkname is relative to the directory that contains the link.
For hard links (LNKTYPE), the linkname is relative to the root of the archive.

So, in a tarball that contains the following, the links point to dir/target:

• dir/target
• other_dir/symlink -> ../dir/target
• other_dir/hardlink -> dir/target

But data_filter thinks that other_dir/symlink will point to ../dir target outside the destination directory.

I have a fix but would like to test it more next week, before merging.
Sorry for the extra work this'll cause for a lot of people :(

Linked PRs

• gh-107845: Fix symlink handling for tarfile.data_filter #107846
• [3.11] gh-107845: Fix symlink handling for tarfile.data_filter (GH-107846) #108209
• [3.10] gh-107845: Fix symlink handling for tarfile.data_filter (GH-107846) #108210
• [3.12] gh-107845: Fix symlink handling for tarfile.data_filter (GH-107846) #108211
• [3.9] gh-107845: Fix symlink handling for tarfile.data_filter (GH-107846) #108274
• [3.8] gh-107845: Fix symlink handling for tarfile.data_filter (GH-107846) #108279

[encukou]

3.13 PR is merged.
3.12 PR is merged (thanks Thomas)!
Buildbots are running on the 3.11 PR: #108209
If the bots don't find any issues I'll continue with the older releases (3.10 needs adjusting tests, 3.9-3.8 doesn't apply cleanly in Git).

[hugovk]

Ready to close this issue?

[encukou]

Yes, thank you for the ping.
For reference, the affected versions, which can raise unwanted tarfile.LinkOutsideDestinationError on valid input, are:

sys.version_info[:3] in {(3, 8, 17), (3, 9, 17), (3, 10, 12), (3, 11, 4)}