[PATCH v5 00/42] idmapped mounts

[Christian Brauner]

On Thu, Jan 14, 2021 at 09:12:41AM -0800, Darrick J. Wong wrote:
> On Tue, Jan 12, 2021 at 11:00:42PM +0100, Christian Brauner wrote:
> > Hey everyone,
> > 
> > The only major change is the inclusion of hch's patch to port XFS to
> > support idmapped mounts. Thanks to Christoph for doing that work.
> 
> Yay :)
> 
> > (For a full list of major changes between versions see the end of this
> >  cover letter.
> >  Please also note the large xfstests testsuite in patch 42 that has been
> >  kept as part of this series. It verifies correct vfs behavior with and
> >  without idmapped mounts including covering newer vfs features such as
> >  io_uring.
> >  I currently still plan to target the v5.12 merge window.)
> > 
> > With this patchset we make it possible to attach idmappings to mounts,
> > i.e. simply put different bind mounts can expose the same file or
> > directory with different ownership.
> > Shifting of ownership on a per-mount basis handles a wide range of
> > long standing use-cases. Here are just a few:
> > - Shifting of a subset of ownership-less filesystems (vfat) for use by
> >   multiple users, effectively allowing for DAC on such devices
> >   (systemd, Android, ...)
> > - Allow remapping uid/gid on external filesystems or paths (USB sticks,
> >   network filesystem, ...) to match the local system's user and groups.
> >   (David Howells intends to port AFS as a first candidate.)
> > - Shifting of a container rootfs or base image without having to mangle
> >   every file (runc, Docker, containerd, k8s, LXD, systemd ...)
> > - Sharing of data between host or privileged containers with
> >   unprivileged containers (runC, Docker, containerd, k8s, LXD, ...)
> > - Data sharing between multiple user namespaces with incompatible maps
> >   (LXD, k8s, ...)
> 
> That sounds neat.  AFAICT, the VFS passes the filesystem a mount userns
> structure, which is then carried down the call stack to whatever
> functions actually care about mapping kernel [ug]ids to their ondisk
> versions?

Yes. This requires not too many changes to the actual filesystems as you
can see from the xfs conversion that Christoph has done.

> 
> Does quota still work after this patchset is applied?  There isn't any
> mention of that in the cover letter and I don't see a code patch, so
> does that mean everything just works?  I'm particularly curious about

The most interesting quota codepaths I audited are dquot_transfer that
transfers quota from one inode to another one during setattr. That
happens via a struct iattr which will already contain correctly
translated ia_uid and ia_gid values according to the mount the caller is
coming from. I'll take another close look at that now and add tests for
that if I can find some in xfstests.

> whether there can exist processes with CAP_SYS_ADMIN and an idmapped
> mount?  Syscalls like bulkstat and quotactl present file [ug]ids to

Yes, that should be possible.

> programs, but afaict there won't be any translating going on?

quotactl operates on the superblock. So the caller would need a mapping
in the user namespace of the superblock. That doesn't need to change.
But we could in the future extend this to be on a per-mount basis if
this was a desired use-case. I don't think it needs to happen right now
though.

> 
> (To be fair, bulkstat is an xfs-only thing, but quota control isn't.)

I'm certain we'll find more things to cover after the first version has
landed. :)
We for sure won't cover it all in the first iteration.

> 
> I'll start skimming the patchset...

Thanks!
Christian