[pve-devel] [PATCH container] fix #5160: fix move_mount regression for mount point hotplug

[Fiona Ebner]

Am 08.01.24 um 14:54 schrieb Filip Schauer:
> Set up an Apparmor profile to allow moving mounts for mount point
> hotplug.
> 
> This fixes a regression caused by
> kernel commit 157a3537d6 ("apparmor: Fix regression in mount mediation")
> 
> The commit introduced move_mount mediation, which now requires
> move_mount to be allowed in the Apparmor profile. Although it is allowed
> for most paths in the /usr/bin/lxc-start profile, move_mount is called
> with a file descriptor instead of a path in mountpoint_insert_staged,
> thus it is not affected by the allow rules in
> /etc/apparmor.d/abstractions/lxc/container-base.
> 

Would it be difficult/impossible for us to do it path-based instead?

> To fix this, introduce a new Apparmor profile to allow move_mount on
> every mount, specifically for mount point hotplug.
> 
> Signed-off-by: Filip Schauer <f.schauer at proxmox.com>

Tested-by: Fiona Ebner <f.ebner at proxmox.com>

Fixes the issue and makes hotplug work again :)

However, I did run into one minor thing (number 1 below), but not sure
if it can be improved. Except the part about the undef warning which is
pre-existing. Another pre-existing issue I ran into is number 2 below,
should also be investigated, but is not caused by this patch.

1.

For an already running container, I got

> mp0: unable to hotplug mp0: failed to change apparmor profile:

after installing the patched pve-container package and attempting to
hotplug. Maybe that's expected and no big deal, because hotplug already
didn't work.

Checking in the journal, it's:

> AVC apparmor="DENIED" operation="change_profile" class="file" info="label not found" error=-2 profile="unconfined" name="pve-container-mounthotplug" pid=99919 comm=7076656461656D6F6E20776F726B65
> Use of uninitialized value in numeric ne (!=) at /usr/share/perl5/PVE/LXC.pm line 1978.

The latter warning seems to be pre-existing for the syswrite comparison
below:

> @@ -1974,7 +1974,7 @@ sub mountpoint_hotplug :prototype($$$$$) {
>  	my $dir = get_staging_mount_path($opt);
>  
>  	# Now switch our apparmor profile before mounting:
> -	my $data = 'changeprofile /usr/bin/lxc-start';
> +	my $data = 'changeprofile pve-container-mounthotplug';
>  	if (syswrite($aa_fd, $data, length($data)) != length($data)) {
>  	    die "failed to change apparmor profile: $!\n";
>  	}


But then $! should be set according to 'perldoc -f syswrite' which it
wasn't telling from my quoted error message above  ¯\_(ツ)_/¯

> Returns the number of bytes actually
>             written, or "undef" if there was an error (in this case the
>             errno variable $! is also set)

2.

After shutdown+start, the hotplug worked as expected. Played around a
bit more and something strange happened with LVM as the storage: While
it worked the first time, when I shut down the container, detached the
mount point, started again and then tried to hot-plug, I got:
mp1: unable to hotplug mp1: command 'mount /dev/lvm/vm-129-disk-1
/var/lib/lxc/.pve-staged-mounts/mp1' failed: exit code 32

Doesn't seem to happen with ZFS or directory storage and starting with
the mount point pre-attached also worked fine with previously written
files still there.

Also booted into 6.2.16-20-pve, installed unpatched pve-container and it
seems the issue was already present then.