On Fedora Rawhide, rpm --erase leaves empty directories on disk

[omajid]

Describe the bug

After I remove a package via dnf remove or rpm --erase, some empty directories on left on disk

To Reproduce
Steps to reproduce the behavior:

1. podman run --pull=newer -it fedora:rawhide /bin/bash
2. dnf install -y koji
3. koji download-build -a $(uname -m) dotnet9.0-9.0.106-1.fc43
4. dnf install *rpm
5. rpm --erase $(rpm -qa '*dotnet*' 'aspnetcore*' 'netstandard*')
6. find /usr/lib64/dotnet/

Expected behavior

No files should have been left on disk. The output of find should be an error about that path not existing.

Output

# find /usr/lib64/dotnet/
/usr/lib64/dotnet/
/usr/lib64/dotnet/packs

Environment

• OS / Distribution: Fedora Rawhide
• Version: RPM version 5.99.90

Additional context

This is the package: https://src.fedoraproject.org/rpms/dotnet9.0/c/c2c2d69fb0381a0dc987c86d9a2d0f902b302918?branch=rawhide

I built this exact commit on all Fedora branches. This issue of stale files left on disk after package is uninstalled only happened on Fedora Rawhide, not on 42 or 41. The order or number of uninstalling the package might matter: I removed the subpackages that own /usr/lib64/dotnet/packs one-by-one, and when I uninstalled dotnet-apphost-pack-9.0 last, the directory got removed correctly.

[ffesti]

OK, turns out this is caused somewhere between c00d8a0 (last good) and 08a6a5e (know broken). So during converting the finger printing code to C++. Some intermediate patches don't build and so the whole set needs to be examine more closely.

[pmatilai]

The intermediate commits fail to build because of a missing semicolon in cd89baf. With that fixed, it bisects down to 08a6a5e. No big surprises there, it was by far the trickiest of these STL'fications of our hashes, and hence the last one.

[pmatilai]

This is the relevant difference between a good and a bad run:

+D: erase      040755  6 (   0,   0)  4096 /usr/lib64/dotnet/packs
[...]
-D: erase      040755  2 (   0,   0)  4096 /usr/lib64/dotnet/packs
[...]
-D: erase      040755  2 (   0,   0)  4096 /usr/lib64/dotnet
+D: erase      040755  3 (   0,   0)  4096 /usr/lib64/dotnet

Note the link count difference: it attempts to remove /usr/lib64/dotnet/packs way too early in context of dotnet-sdk-9.0 where it's not empty, whereas it should be removed from dotnet-apphost-pack-9.0 towards the end of the transaction. That failure causes /usr/lib64/dotnet removal to fail as well.

From transaction.c:

       * If both this and the other package are being removed, then each
        * file removal from preceding packages needs to be skipped so that
        * the file removal occurs only on the last occurrence of an overlapped
        * file in the transaction set.

That's the thing that is failing here, and a bit of extra debug logging shows that it's indeed completely broken by that commit. It could have some more grave consequences than just files left behind. I'll look into it some more, but this really is a blocker for beta2 and if I can't find a fix in a reasonable time, we'll revert the damn thing.

Thanks a lot @omajid for reporting this!