[UPDATE: Rumprun aarch64 support has now been merged in upstream nabla.]

Nabla containers provide a new type of container designed for strong isolation on a host system. The foundation of nabla containers lies in three main components: rumpkernel, solo5, and runnc. The team that built nabla containers extended the rumprun unikernel framework to support solo5 (instead of hardware/baremetal or xen), so that a rumprun-baked unikernel application can be executed on top of a lightweight monitor such as solo5. In this post, we describe the steps we took in order to port Nabla containers to the ARMv8 architecture.

A bit of background Link to heading

Rumprun is a unikernel framework based on rumpkernel, a project that provides kernel-quality drivers for various components, e.g. file systems, network devices, POSIX system calls. Rump kernel exposes these drivers through the rump kernel hypercall interface to higher abstraction layers. While initially, rump kernels were designed to provide ‘userspace’ drivers, they evolved to become the base of the Rumprun unikernel.

Almost any POSIX-based application can be built into a Rumprun unikernel using the following workflow:

  1. Compile and link the application against the POSIX-y interface that Rumprun exposes.
  2. Bake the application, in order to add the bits and pieces needed to turn it into an image that is bootable on the targets that Rumprun runs on top of.

Upstream Rumprun defines the concept of target, i.e. the platforms on which a Rumprun unikernel can run on top of. At the moment the upstream Rumprun provides two targets:

  • The hw target provides support on raw hardware which includes most available hypervisors.
  • The Xen target is optimized for execution on top of the Xen hypervisor

The Nabla fork of Rumprun provides a new unikernel target, Solo5. Solo5 is, essentially, a hardware abstraction layer that provides a very thin interface, or else a minimal attach surface. Its purpose is to facilitate the port of libOS/unikernel frameworks on various hardware platforms, i.e. a unikernel that runs on top of Solo5, runs on top of all the hardware targets, or tenders, using the Solo5 terminology, supported by Solo5.

ARMv8 rumprun solo5 Link to heading

To port nabla containers to the ARMv8 architecture, one has to provide support for each of these components: rumprun, solo5 and runnc. We decided to tackle this challenge by separately porting each component and working on integrating as much code as possible from upstream repositories.

Solo5 Link to heading

For the solo5 port, most of the code was already in upstream solo5, although the seccomp tender (as it is now called) provided by nabla didn’t have support for aarch64 targets. To implement that, apart from adding the compilation target itself, we also had to provide the arch specific lds and add the hypercall-to-syscall mechanism used to aarch64. Finally, for everything to actually work we implemented reading the cpu timer frequency correctly for aarch64 and provided some missing seccomp definitions. As of the time of writing this post, an upstream solo5-seccomp tender for both x86 and aarch64 is under way and should be merged in really soon.

Rumprun Link to heading

For rumprun, things were a bit more complicated. Both upstream and nabla rumprun repos build necessary NetBSD parts from an old snapshot of the official sources. NetBSD has added aarch64 support fairly recently and with much work still being done we decided to base our build on the latest official sources instead of the stripped down version provided by rumprun. This presented a challenge: integrating the changes to a not actively maintained code base is not a walk in the park. First things first, we had to successfully compile the whole project: Rump parts of the NetBSD kernel are not actively tested with rumprun and so changes being made to related components are not always guaranteed to work. After adding the aarch64 platform to the required Makefiles for rump in NetBSD source and creating the relevant directories, we encountered linking errors pertaining to both double symbols between the provided libc and rump and also incorrectly linked for the rump case internal functions of the kernel itself. Thankfully similar problems have already been solved for the arm 32-bit architecture and so we could implement a solution based on existing code.

Having successfully compiled NetBSD’s source we then had to implement any missing rumpuser parts either for aarch64 or for functions introduced in upstream kernel’s rump.

The most challenging part was to understand how thread-level switching happens on aarch64 and implement that in the context of rumprun. Upon initial creation of “main” threads for the rump components, rumprun stores a “bouncer” function on top of their freshly allocated stacks and then switches between them using its scheduler. When a thread’s turn comes to run, the “bouncer” function is popped from the stack and the actual thread content is executed.

The mechanism doing the actual thread switching is implemented in arch-specific assembly. Although a basic implementation for the ARM 32-bit architecture is provided with upstream rumprun, porting to 64-bit ARMv8 was not trivial due to two main architectural differences of ARMv8: a) one cannot directly modify the program counter, and b) storing and retrieving registers to/from the stack has to be aligned if they are used for memory access. As a result, push and pop functions have to be implemented manually, as no specific instruction exists for that purpose. After implementing the actual thread switcher and bouncer, we also had to make sure the TLS (Thread-Local Storage) register was set correctly in the solo5 platform implementation.

Last but not least, with thread switching in place, the rest of the solo5 platform arch-specific bits needed to be created: the ldscript for linking, the relevant headers and the Makefile modifications for everything to build correctly.

NOTE: Stack protection in NetBSD source caused problems in our tests and so it’s globally disabled for rumprun pending further investigation.

runnc Link to heading

Since runnc is written in GO, the arch-specific bits were only build related, so porting it to aarch64 was as easy as defining the architecture (arm64) and pointing to the correct submodule repos.

We are currently working with the nabla containers team to merge in aarch64 support. In the meantime, you can browse through the code. Stay tuned for a tutorial on how to run a rumprun unikernel on a RPi3!