make mountpoints shared

Hello,

I'd be interested in having / mounted as shared, either by default or as an option.
(Feel free to tell me to go open an issue in openrc instead)

The rationale for this is that I selfishly need to have some mountpoints mounted as shared for some container usages (when alpine is the host OS), in order to backpropagate mounts made by a privileged container in a volume marked as 'shared'.
In my case the volumes shouldn't be made directly off /, so just marking the FS where volumes should be as shared is probably going to be enough but it's not something I have direct control of so I'd be more comfortable with / marked shared as systemd has been doing (not a great argument admittedly... But it's what users are used to)

FWIW, it looks like podman 4 in rootless mode also seems to depend on it: https://github.com/containers/podman/discussions/12721 and aports#13541 (closed) (the hang might not be related but it illustrates the warning message being shown), so we might get other requests at this point.

With all that said, the question would be where to do it -- I think openrc's init.d/root might be the place, if a user adds shared to fstab, fstabinfo -o / will print it and we can mount --make-(r)shared / at this point (or other way around if it's to be made default -- do it here unless there's "private" in the fstab options), but perhaps there is a better way I'm not aware of?
Or should I just do it in my container starting service, and not bother the rest of alpine with it?

Thanks!

moved from aports#13565

cc @mpolanski (for the podman 4 rootless container requirement), @ncopa (openrc)

Please tell me if I should just bring it to openrc upstream first, but as this is linux-specific I feel we ought to reach an agreement on what to do here first.

Having the root mount point marked as shared is indeed convenient for container usage. But this is a big change that can break users' workflows. On the other hand most users are indeed accustomed to this behavior in other distributions.

Mount namespaces are not trivial so here are some links worth reading:

• Kernel documentation: https://www.kernel.org/doc/Documentation/filesystems/sharedsubtree.txt
• systemd's rationale: https://cgit.freedesktop.org/systemd/systemd/commit/?id=b3ac5f8cb98757416d8660023d6564a7c411f0a0
• LWN article "Mount namespaces and shared subtrees": https://lwn.net/Articles/689856/

@kdaudt could you move this issue to the alpine/tsc repository for further consideration?

Thanks for the quick reply!

I also thought it could break things and tested it extensively before making it default for our system (my work ships an alpine-based system on embedded devices, so it has to work for users), and tools appear to do the right thing:

• unshare -m (or -Um) will keep working as intended as it makes all mounts private, unless the user explicitly passes --propagation shared/slave/unchanged -- so it'd affect people who explicitly make some (but not all) mount points shared on the host and use this option. I think these people could make the non-shared filesystems non-shared again manually if required.
• podman/crun will also mark all container mounts private explicitly (both on host when it prepares its directories, inside container, and for --volume bind mounts unless explicitely required with :shared or :slave)
• I couldn't think of any other command that uses mount namespaces and would be impacted

Of course, the 'root' init service should allow choosing either. My proposal here wasn't even changing the default, just making it possible to add a shared option and mount --make-rshared there if present in the fstab for / entry like ro -- but if the default change it's even more convenient for me ;)

added agenda:proposal label

added type:system-change label

added agenda:scheduled label and removed agenda:proposal label

I think this is an issue for upstream openrc.

Well ultimately it depends on what we want to do for alpine first: if there's no interest in doing this except for myself, there's no point in implementing it in openrc.

That being said, it doesn't hurt to ask anyway, so I'll open an issue there.

https://github.com/OpenRC/openrc/issues/525 opened.

thanks!

I mean, I think this is nice to have configurable regardless what the default is, but that should be done in openrc /etc/init.d/root imho.

added agenda:follow-up label and removed agenda:scheduled label