Re: [GIT PULL] fs fix for v5.19-rc3

From: Christian Brauner
Date: Thu Jun 16 2022 - 03:59:28 EST


On Wed, Jun 15, 2022 at 09:31:32AM -0700, Linus Torvalds wrote:
> On Wed, Jun 15, 2022 at 3:39 AM Christian Brauner <brauner@xxxxxxxxxx> wrote:
> >
> > This fixes an issue where we fail to change the group of a file when the caller
> > owns the file and is a member of the group to change to. This is only relevant
> > on idmapped mounts.
>
> Pulled, but this code makes me go "Eww".
>
> I do get the feeling that maybe we should have some kind of static
> type help here, with "unmapped" and "filesystem mapped" uid/gid values
> having different types, the same way we have kuid_t vs uid_t for
> user-namespace mapping.
>
> Because it feels like the problem here is that "kgid_t" has different
> meanings depending on context.
>
> In fact, just within that one function, we have *three* very different kgid_t's:
>
> - the argument to the function is a "kgid_t gid", which has *not*
> been mapped into the filesystem mapping (why not?)
>
> - the function itself declares a "kgid_t kgid", which is the inode
> group ID mapped into the mount
>
> - it how additionally has that thied "kgid_t mapped_gid", which is
> the first 'gid' mapped into the mount
>
> and honestly, I find that hugely confusing. The naming doesn't exactly
> clarify things either (ie "gid" is a "kgid_t", not a "git_t", and so
> is "kgid" - not helpful).
>
> And being confusing with no type system support isn't a great thing.
>
> So I'm wondering if <linux/mnt_idmapping.h> might be able to do the
> same kind of things we do in <linux/uidgid.h>.
>
> NOTE: I did not look into just how hugely painful that would be. Maybe
> the above is the rantings of a madman.

I'm actually looking into this and have been for a while. I had thought
about this from the start but just now have time to play with this!