/sysroot private mount and /home -> /sysroot/home

[cgwalters]

Related: #2085

Reported on IRC, today gnome-continuous uses /home ➡️ /sysroot/home but this conflicts with #1438 if the user creates mounts in their home directory.

I think basically everyone should use /var/home, but probably ostree should either disable the private mount for /sysroot if we detect these symlinks.

[valentindavid]

I have noticed that if I remove totally /home ➡️ var/home, then Flatpak applications with --filesystem=host do not have access to home directory, unless they also have --filesystem=home.

So it seems that home should be defined within /var/home in /etc/passwd, but the symlink should still exist for flatpak application to work correctly. (I still have to test that).

[jlebon]

So it seems that home should be defined within /var/home in /etc/passwd

Yes, note rpm-ostree for example does this: coreos/rpm-ostree#1726.

[valentindavid]

I can confirm that for Flatpak applications like org.flatpak.Builder need to have both the /home ➡️ var/home AND the home directory defined in /etc/passwd should be using /var/home. I have just tested it.

It is very peculiar that removing the symlink breaks flatpak.

[dbnicholson]

We hit a very related issue in Endless OS recently that's caused by /sysroot's private mount propagation. We have a test mode that mounts an overlay on /sysroot (and other disk backed mounts) so that you can make broad changes to the
system and have them discarded when the system is shut down. If a process is started with a separate mount namespace before test mode is invoked, then that process will not inherit the /sysroot overlay. This was causing new user home directories to persist in /sysroot/home since accountsservice started using systemd sandboxing features a couple years ago, but it just hit Endless OS early this year and apparently no one has run the test mode since.

As it turns out, I think private /sysroot is solving the problem the wrong way. That was added to prevent submounts in /var from propagating back to /sysroot/deploy/.../var. That's very useful, but private mount propagation is a very large hammer for it since it stops all mount propagation. Instead, I think it's better if /sysroot is the default shared and /var is made a slave so that it receives mount events from /sysroot but not vice versa. Then you retain the behavior that /var submounts don't propagate back to /sysroot, but submounts of /sysroot propagate in the completely normal way. This is how systemd and toolbox setup mount namespaces.