Barco: Linux Containers from Scratch in C
github.com
github.com
Also I find it interesting that in Portuguese as well as in Spanish the word exists in both genders: el barco, la barca, o barco, a barca.
Just wondering the main reason you're C since most of the container project now seems to be using Go or Rust?
One example that immediately comes to mind from Rust is a bug with O_PATH file descriptors I found a while ago[1], which would've made certain code we use in runc not work. And from Go, here is a bug I just found in their code for handling file descriptors for ForkExec[2] which is causing issues in a runc patch I'm working on. Neither of these issues exist in C programs. Though of course, C programs have their own issues. For better or worse, the Linux kernel APIs are easiest to use from C.
In runc we actually implement the core container setup code in C because Go doesn't allow you to do everything we need for setting up a container (it has gotten better though, in the past it was completely impossible to set up a container properly in pure Go -- now you can set one up but there are still certain configurations that are not possible to implement in pure Go, such as "docker exec"). You also cannot run Go in single-threaded mode, which means that certain kernel APIs (unshare(CLONE_NEWUSER) for instance) simply cannot be used from regular Go code.
[1]: https://github.com/rust-lang/rust/issues/62314 [2]: https://github.com/golang/go/issues/61751
There is no way to make an entire Go program run as a single threaded program without using CGo the way we do in runc. Even GOMAXPROCS=1 doesn't work. CLONE_NEWUSER will always fail in a multi-threaded program.
Maybe, but:
1. It's irrelevant to this product (no one is writing portable UNIX code when they are writing some Linux-specific software, like container implementations).
and
2. It's irrelevant to the author's goals (learning Linux kernel stuff using the language that the interface to the kernel uses is a better idea than using a different language and hacking shims for all the stuff you want to do).
and
3. The cost to switch to a new language is substantial, and only makes sense if you're either joining a team and project that uses that new language, or if the goal is to learn that new language.
2. Any languge able to call into Linux API surface is usable. And if we go down the UNIX native languages route born at Bell Labs, C++ also counts.
3. That depends on how much someone knows C (properly), versus other alternatives
Just like HP-UX vaults, and Solaris Zones aren't using namespaces, cgroups and seccomp, rather their own UNIX flavours.
This issue doesn't intrinsically affect Rust as a language (when compared to C), because you can just do exactly the same thing as you'd have done in C:
let fd = libc::open(b"/path\0".as_ptr().cast(), libc::libc::O_CLOEXEC | libc::O_PATH);
let fd = libc::syscall(libc::SYS_open, b"/path\0".as_ptr(), libc::O_CLOEXEC | libc::O_PATH);
And for setting up containers you'll have to do this anyway because Rust's standard library doesn't expose all of the necessary functionality anyway.
In general, C programs do not require workarounds for dealing with kernel APIs for the simple reason that the vast majority of kernel APIs are developed with test programs written in C, so kernel developers will usually not design an API that is awful to use in C.
Another thing that surprised me when I first started programming in Rust is that:
let fd = File::open("foo")?.as_raw_fd();
let f = File::open("foo")?;
let fd = f.as_raw_fd();
This does happen elsewhere in Rust, but often when you have methods on &self that return something you use later, the method returns something with the same lifetime (fn foo(&'a self) -> Foo<&'a>) and thus the original object will be kept alive until the end of the scope. It just so happens that file descriptors are tied to the lifetime of the File in a way that Rust cannot express nor detect.
I don't know if clippy has a warning for this particular case. It might be useful to add it.
[0] https://doc.rust-lang.org/std/os/fd/struct.BorrowedFd.html
And `open` is not a kernel API either. It's a libc API. If you want to directly access the API provided by the kernel you're supposed to make a syscall, which essentially is exactly the same in Rust and in C.
And to make the point of `open` in C *not* being a kernel API more clear, in glibc the `open` function *doesn't* actually call the `open` syscall, but `openat` with `AT_FDCWD`. Glibc doesn't guarantee that a given function will actually call a given syscall, and new versions of glibc often change which syscalls are called by a given function. This is important if you're also doing e.g. seccomp sandboxing, because suddenly your program might stop working if glibc is updated. For example, glibc 2.34 started using the `clone3` syscall under the hood, which broke Chromium Embedded Framework's sandbox.
So, again, your argument that a language like Rust "makes it harder to understand what exactly your program is doing" compared to C in this particular case isn't really valid, because C has exactly the same problem if you use libc functions, and the only way to guarantee that the program is doing exactly what you want is to use syscalls, which is the same both in C and Rust.
> Another thing that surprised me when I first started programming in Rust
Yep. That's one of the Rust's few badly designed APIs.
the head-space that adopting a new language for a specific project takes is immense compared to tackling it in a familiar language that you know you're already able in; there is rarely a benefit to doing so outside of team environments where a certain level of on-boarding is expected, or because you have a really niche language requirement/feature that your project is begging for.
Their explanation:
"While most of the tools used in the Linux containers ecosystem are written in Go, I believe C is a better fit for a lower level tool like a container runtime. runc, the most used implementation of the OCI runtime specs written in Go, re-execs itself and use a module written in C for setting up the environment before the container process starts.
crun aims to be also usable as a library that can be easily included in programs without requiring an external process for managing OCI containers."
Aren't containers never suitable for running untrusted code? You need AppArmor, bwrap, or similar AFAIK.
https://cloud.google.com/blog/products/containers-kubernetes...
They are suitable provided the kernel is secure.
The more nuanced answer is that containers have several layers of protections (seccomp, LSMs, user namespaces, namespaces, cgroups, capabilities, and standard process permissions by running as an unprivileged user) which all act together to help protect against container attacks. It's not perfect, but most container breakout attacks we've had so far are related to when container runtimes have to operate on a container during process setup (IMHO because the process for creating a container process is far from atomic) -- some of these attacks were enabled by kernel bugs which we went and fixed as well. It is very difficult to break out of a container once it has been configured and left alone.
Also while user namespaces help in theory, in practice expanding the attack surface of the kernel exposed to unprivileged users has consequences in allowing container breakout (e.g. CVE-2022-0185) also more recently CVE-2023-3390
> Also while user namespaces help in theory
You should use user namespaces to contain untrusted code, you absolutely should not enable CLONE_NEWUSER inside a container. I was referring to the former, you're talking about the latter.
I'd agree that unprivileged user namespaces should not be available inside containers, however the default stances of some Linux distros (enable unprivileged user namespaces) and Kubernetes (disable the CRIs seccomp filter by default) mean that an awful lot of envirionments will end up in a situation where this is possible.
Jessica Frazelle ran a public bug bounty to break out of a container image that is properly secured, and as far as I know nobody collected the bounty. The website isn't up at the moment, maybe she took it down. https://contained.af/
Since in reality most attacks will be against your app itself before the attacker has direct access to syscalls, I see namespaces/containers as extra protection.
However, all this code is quite complex, and the kernel and the software ecosystem are lacking in having a layered approach to security that goes all the way down to the low-level nitty gritty stuff. For example, kernel memory structures are not robustly protected against the usual memory exploits, and there isn't as strong W^X protection as desired. Windows, in contrast, is able to provide layered security through a variety of approaches, including running the entire operating system in a virtual machine, with the host ensuring integrity of kernel memory. These sorts of layered approaches to security are desirable because there will always be defects in any complex software.
Side note: AppArmor and bwrap are distinct. Bubblewrap is a relatively simple userspace program that makes use of existing kernel containerization features (the same ones that Docker/Podman use), whereas AppArmor and SELinux are security features that are patched into the kernel itself. AppArmor and SELinux have made some progress in adding layered low-level security to the kernel, but it's not particularly impressive. Bubblewrap has done great work in exposing the kernel's existing tech to users, but they are not fundamental improvements to the kernel itself.
Containers with own PID namespace can't spy on process IDs on the host though? Not sure what you mean here.
> and there isn't as strong W^X protection as desired
What level is desired? Bootup warnings for W^X got merged a while ago. Changes that try to include anything violating it are rejected (see bcachefs).
> Windows, in contrast, is able to provide layered security through a variety of approaches, including running the entire operating system in a virtual machine, with the host ensuring integrity of kernel memory.
What? Xen existed for years, that's not "in contrast". Secureboot and lockdown exists on Linux too. There's also per-service firecracker microvm.
> whereas AppArmor and SELinux are new security features that are patched into the kernel itself
That's very misleading. They're not new - selinux is over 2 decades old. They're also not "patched in" - LSMs have been integrated into Linux for a very long time with multiple implementations available. Selinux had multilabel security created for gov use. It's quite impressive actually.
I misspoke on SELinux and AppArmor being "new". What I was getting at is that they are distinct kernel features, compared to bwrap which is just a user of kernel features already familiar in this discussion. So "new", as in, "additional", e.g. "We turned up some new evidence from the old files".
And yes, SELinux is included as a first-class kernel feature. AppArmor is a bit different because it still has a lot of hurdles before all its features can be upstreamed. However, upstreaming is not the end-all-be-all so it's not necessarily a bad thing that parts of AppArmor are patched in, so I'm not emphasizing this point at all.
> What? Xen existed for years, that's not "in contrast".
Does the Xen or KVM ecosystems provide anything comparable to Windows hypervisor-enforced code integrity? That is, the host is aware of what kernel memory needs to be set to read-only or checked regularly for corruption or irregularities, in a system that is impossible to interfere with without a VM break. (https://learn.microsoft.com/en-us/windows-hardware/design/de... )
Secure VMs are great, VMs that are actually monitoring and enhancing the security of the code running inside are even better.
Xen can do it, KVM had some attempts at memory enforcement patches, but not sure where it ended up. I'm not sure how well utilised it is though from the guest side ootb. (I think poorly)
A lot of the signing/module loading issues are prevented on the guest side though with https://man7.org/linux/man-pages/man7/kernel_lockdown.7.html
So it's... not the same, but the facilities are available. (Unless someone wants to correct me and it's already used)
Properly configured WASI runtime are great for security but they're worse than containers on most other fronts. I don't think the downsides make sense unless you're building a business that lets random customers upload WASM files you execute.