On 04/05/2025 11:35, Yannick Le Pennec wrote:
Dear Maintainer,
On CephFS, creating a snapshot is done suchly:
$ mkdir .snap/snapshot-name
This creates a directory which contains a view of the current directory
at the time the snapshot was created, and is therefore non-empty if the
current directory was non-empty at that time. Its contents are immutable.
Removing a snapshot is done suchly:
$ rmdir .snap/snapshot-name
This works fine even if .snap/snapshot-name is non-empty, as rmdir calls
rmdir(2) directly without trying to be smart, and the Ceph filesystem
understands what rmdir on a snapshot directory means: remove the snapshot.
One would naturally expect rm -d to behave *exactly the same* as rmdir,
but this isn't the case, because rm -d first does a getdents on the
directory, observes it is non-empty, and refuses to perform the removal.
From what I understand, this is non-compliant behavior per XCU rm 2.a. and 4.
and XSH remove():
https://pubs.opengroup.org/onlinepubs/9799919799/utilities/rm.html
2. If file is of type directory, the following steps shall be taken:
a. If neither the -R option nor the -r option is specified, but -d
is specified, rm shall proceed with step 3 for the current file.
4. rm shall perform actions equivalent to the remove() function defined
in the System Interfaces volume of POSIX.1-2024 called with a
pathname of the current file used as the path argument.
https://pubs.opengroup.org/onlinepubs/9799919799/functions/remove.html
If path names a directory, remove(path) shall be equivalent to rmdir(path)
In short:
rm -d .snap/snapshot-name
shall be equivalent to:
remove(".snap/snapshot-name")
which in turn shall be equivalent to:
rmdir(".snap/snapshot-name")
Moreover other coreutils rm implementations (such as uutils or voreutils)
do not exhibit this bug and correctly remove non-empty directories.
Regards,
Yannick Le Pennec
For reference this was also reported at:
https://bugs.debian.org/1104300
