[zenoprax]

> "kernel escape doesn’t land you on a 0 UID"

I'm not sure I agree/understand. If you've somehow bypassed AppArmor and cgroup mechanisms then any UID/GID remapping is irrelevant. At this point you're in a position to directly manage memory.

What do you mean by "kernel escape"?

[beardedwizard]

I'm enjoying how nobody in this thread seems to know what a container actually is, and folks may be surprised to learn kernel namespace underpins both docker and lxc.

[cyphar]

> If you've somehow bypassed AppArmor and cgroup mechanisms then any UID/GID remapping is irrelevant. At this point you're in a position to directly manage memory.

Not really, user namespaces (despite all of the issues that unprivileged user namespaces have caused) provide an additional layer of protection as lots of privilege checks are either based on kuid/kgid or are userns-aware. These are some of the deepest security models that Linux has (in the sense that every codepath that involves operating on kernel objects involves a kuid/kgid check and possibly a capability check), so making full use of them is fairly prudent. The vast majority of container breakouts reported over the past decade were complete non-issues or very limited impact if you used user namespaces.

AppArmor is not a particularly strong security boundary (it's better than nothing, but there are all sorts of issues with having path-based policies and so they mostly act as final layer of defence against dumb attacks). cgroups are mostly just resource limits, but the devices cgroup (and devices eBPF filter) are are security barrier that prevent obviously bad stuff like direct write access to your host drive. However, those are not the only kinds of protections we need or use in containers, and breaking just those is not enough to "directly manage memory" (but /dev/kmem is inaccessible to user namespaced processes so if that is something you're worried about, user namespaces are another good layer of defence ;)).

It should also be noted that LXC is not the only runtime to support this, the OCI ecosystem supports this too and has for quite a long time now (and the latest release of Kubernetes officially supports isolated user namespaces). Most of my container runtime talks in the past decade have had a slide telling people to use user namespaces but sadly they are not widely used yet.

On the topic of whether containers are a security boundary, I consider them to be fairly secure these days if you use reasonable defaults (user namespaces, reasonable seccomp rules, ideally non-root inside the container). The main reason we struggle in ways that BSD Jails and Solaris Zones do not is because containers on Linux require putting together a lot of disparate components and while this does mean that you can harden non-container programs using them, it opens the door to more bugs. If we had a way to consolidate more operations and information in-kernel things would be much easier to secure (one perennial issue is that of the inatomicity of switching to the container security zone).

(Disclaimer: I am a maintainer of runc.)

[zenoprax]

Awesome, thanks for the explanation. I didn't know that kuid/kgid existed! That also explains why Proxmox manages the re-mapping in the `LXC.conf` rather than the AppArmor profile. The cascade of AppArmor configs seemed to focus quite a bit on access to `/proc` and `/sys` so I think I mixed that up cgroups with my comments about memory access.

I've been reading up on them (https://www.kernel.org/doc/html/latest/filesystems/idmapping...) seeing some of the notation for user IDs (e.g. `u20000`) reminded me that my Hetzner ZFS storage was accessed using a similar UID format for the username.

[cyphar]

> The cascade of AppArmor configs seemed to focus quite a bit on access to `/proc` and `/sys` so I think I mixed that up cgroups with my comments about memory access.

Funnily enough that is a good example of how fickle AppArmor's protections are -- if you give containers mount privileges (needed to enable container nesting and most system container usecases) you can bypass most (if not all) AppArmor path rules because you can create alternative mounts that don't match the ones in the rules. With the fsopen(2) and open_tree(2) mount APIs, it's even easier -- AppArmor uses d_path to compute the path for policy purposes but detached mounts for procfs do not have a /proc prefix in their d_path form!

My general impression is SELinux is better as it applies to objects directly, but we've had security issues with it too.