[PATCH 00/34] fs: idmapped mounts

[Sargun Dhillon]

On Thu, Oct 29, 2020 at 05:05:02PM +0100, Lennart Poettering wrote:
> On Do, 29.10.20 10:47, Eric W. Biederman (ebiederm at xmission.com) wrote:
> 
> > Is that the use case you are looking at removing the need for
> > systemd-homed to avoid chowning after lugging encrypted home directories
> > from one system to another?  Why would it be desirable to avoid the
> > chown?
> 
> Yes, I am very interested in seeing Christian's work succeed, for the
> usecase in systemd-homed. In systemd-homed each user gets their own
> private file system, and these fs shall be owned by the user's local
> UID, regardless in which system it is used. The UID should be an
> artifact of the local, individual system in this model, and thus
> the UID on of the same user/home on system A might be picked as 1010
> and on another as 1543, and on a third as 1323, and it shouldn't
> matter. This way, home directories become migratable without having to
> universially sync UID assignments: it doesn't matter anymore what the
> local UID is.
> 
> Right now we do a recursive chown() at login time to ensure the home
> dir is properly owned. This has two disadvantages:
> 
> 1. It's slow. In particular on large home dirs, it takes a while to go
>    through the whole user's homedir tree and chown/adjust ACLs for
>    everything.
> 
> 2. Because it is so slow we take a shortcut right now: if the
>    top-level home dir inode itself is owned by the correct user, we
>    skip the recursive chowning. This means in the typical case where a
>    user uses the same system most of the time, and thus the UID is
>    stable we can avoid the slowness. But this comes at a drawback: if
>    the user for some reason ends up with files in their homedir owned
>    by an unrelated user, then we'll never notice or readjust.
> 
> > If the goal is to solve fragmented administration of uid assignment I
> > suggest that it might be better to solve the administration problem so
> > that all of the uids of interest get assigned the same way on all of the
> > systems of interest.
> 
> Well, the goal is to make things simple and be able to use the home
> dir everywhere without any prior preparation, without central UID
> assignment authority.
> 
> The goal is to have a scheme that requires no administration, by
> making the UID management problem go away. Hence, if you suggest
> solving this by having a central administrative authority: this is
> exactly what the model wants to get away from.
> 
> Or to say this differently: just because I personally use three
> different computers, I certainly don't want to set up LDAP or sync
> UIDs manually.
> 
> Lennart
> 
> --
> Lennart Poettering, Berlin

Can you help me understand systemd-homed a little bit?

In the man page it says:

systemd-homed is a system service that may be used to create, remove, change or 
inspect home areas (directories and network mounts and real or loopback block 
devices with a filesystem, optionally encrypted).

It seems that the "underlay?" (If you'll call it that, maybe there is a better 
term) can either be a standalone block device (this sounds close to systemd 
machined?), a btrfs subvolume (which receives its own superblock (IIRC?, I might 
be wrong. It's been a while since I've used btrfs), or just be a directory 
that's mapped?

What decides whether it's just a directory and bind-mounted (or a similar 
vfsmount), or an actual superblock?

How is the mapping of "real UIDs" to "namespace UIDs" works when it's just a 
bind mount? From the perspective of multiple user namespaces, are all 
"underlying" UIDs mapped through, or if I try to look at another user's
home directory will they not show up?

Is there a reason you can't / don't / wont use overlayfs instead of bind mounts?