On Tue, Apr 15, 2025 at 02:27:24PM +0200, Daniel Kral wrote:
> On 2/20/25 13:15, Fiona Ebner wrote:
> > I also noticed that we have no check against starting a container with
> > volumes on a storage that does not support 'rootdir'. We have such a
> > check for VMs IIRC. Prohibiting that would also be good, but maybe
> > something for PVE 9 where we can also check for misconfigured
> > containers/storages via the pve8to9 script up front so users can adapt.
>
> I'm preparing the v3 for this now, but I just noticed there actually is a
> assertion for this since e6da5357cc ("fix #3421: allow custom storage
> plugins to support rootfs") if I'm not missing something here in
> __mountpoint_mount(...).
>
> What I don't yet understand is why there is no similar check for this in
> __mountpoint_mount for subvolumes, e.g. I can't start the container if I
> have a mountpoint on a directory storage without 'rootdir' support, but I
> can do so if the mountpoint is on a zfs pool without 'rootdir' support.
>
> Since starting the container results in
>
> run_buffer: 571 Script exited with status 25
> lxc_init: 845 Failed to run lxc.hook.pre-start for container "101"
> __lxc_start: 2034 Failed to initialize container "101"
> TASK ERROR: startup for container '101' failed
>
> for the WebGUI, I'll try to squeeze in a patch to make the error message a
> little more readable if there's something going wrong when mounting.
>
> ---
>
> On another note, I've also noticed that if the root disk / mountpoint is
> already on a storage which does not support 'rootdir', the user is unable to
> move it to another storage... Shouldn't we allow users to do that so they
> can easily move out error states? Either way, this can be a follow-up anway,
> so no need to make this patch series any longer.
The `rootdir` content type is generally a bit wonky currently.
The problem is we're mixing content types and "allowed contents"
together:
For ZFS and btrfs we use subvolumes which are their own *content type*:
"subvol".
For *other* directory storages, we have a special case for `size=0`
where we have a directory of an "unlimited" size, which is also
considered to be of *content type* "subvol".
In the other case we actually allocate the content type *image*,
*BUT*(!!!) the Plugin.pm's default `list_volumes` implementation will
artificially *name* it "rootdir" *if* the *associated VMID* is from a
container by querying the VM list via `PVE::Cluster::get_vmlist()`.
This is not something we can ask external storage plugin devs to do,
IMO.
*rootdir* as an actual *content type* does not *really* exist in
pve-storage, other than as a remnant from openvz times for directories
under the `rootdir/` directory on a directory storage, which I'm fairly
certain we don't support in pve-container...
This means that currently what is referred to as the "rootdir" content
type is actually just the *permission* to put containers on the storage.
This is something we really need to fix up with PVE9 one way or another.
Personally I'd argue the content type should disappear entirely.
`list_volumes` should call use the correct type (images or subvol), and
the rootdir content permission should (and always / indepenent from the
storage and what *content type* we create) in pve-container's disk
allocation code.
_______________________________________________
pve-devel mailing list
pve-devel@lists.proxmox.com
https://lists.proxmox.com/cgi-bin/mailman/listinfo/pve-devel
