Unikernels

At LCA I attended a talk about Unikernels. Here are the reasons why I think that they are a bad idea:

Single Address Space

According to the Unikernel Wikipedia page [1] a significant criteria for a Unikernel system is that it has a single address space. This gives performance benefits as there is no need to change CPU memory mappings when making system calls. But the disadvantage is that any code in the application/kernel can access any other code directly.

In a typical modern OS (Linux, BSD, Windows, etc) every application has a separate address space and there are separate memory regions for code and data. While an application can request the ability to modify it’s own executable code in some situations (if the OS is configured to allow that) it won’t happen by default. In MS-DOS and in a Unikernel system all code has read/write/execute access to all memory. MS-DOS was the least reliable OS that I ever used. It was unreliable because it performed tasks that were more complex than CP/M but had no memory protection so any bug in any code was likely to cause a system crash. The crash could be delayed by some time (EG corrupting data structures that are only rarely accessed) which would make it very difficult to fix. It would be possible to have a Unikernel system with non-modifyable executable areas and non-executable data areas and it is conceivable that a virtual machine system like Xen could enforce that. But that still wouldn’t solve the problem of all code being able to write to all data.

On a Linux system when an application writes to the wrong address there is a reasonable probability that it will not have write access and you will immediately get a SEGV which is logged and informs the sysadmin of the address of the crash.

When Linux applications have bugs that are difficult to diagnose (EG buffer overruns that happen in production and can’t be reproduced in a test environment) there are a variety of ways of debugging them. Tools such as Valgrind can analyse memory access and tell the developers which code had a bug and what the bug does. It’s theoretically possible to link something like Valgrind into a Unikernel, but the lack of multiple processes would make it difficult to manage.

Debugging

A full Unix environment has a rich array of debugging tools, strace, ltrace, gdb, valgrind and more. If there are performance problems then tools like sysstat, sar, iostat, top, iotop, and more. I don’t know which of those tools I might need to debug problems at some future time.

I don’t think that any Internet facing service can be expected to be reliable enough that it will never need any sort of debugging.

Service Complexity

It’s very rare for a server to have only a single process performing the essential tasks. It’s not uncommon to have a web server running CGI-BIN scripts or calling shell scripts from PHP code as part of the essential service. Also many Unix daemons are not written to run as a single process, at least threading is required and many daemons require multiple processes.

It’s also very common for the design of a daemon to rely on a cron job to clean up temporary files etc. It is possible to build the functionality of cron into a Unikernel, but that means more potential bugs and more time spent not actually developing the core application.

One could argue that there are design benefits to writing simple servers that don’t require multiple programs. But most programmers aren’t used to doing that and in many cases it would result in a less efficient result.

One can also argue that a Finite State Machine design is the best way to deal with many problems that are usually solved by multi-threading or multiple processes. But most programmers are better at writing threaded code so forcing programmers to use a FSM design doesn’t seem like a good idea for security.

Management

The typical server programs rely on cron jobs to rotate log files and monitoring software to inspect the state of the system for the purposes of graphing performance and flagging potential problems.

It would be possible to compile the functionality of something like the Nagios NRPE into a Unikernel if you want to have your monitoring code running in the kernel. I’ve seen something very similar implemented in the past, the CA Unicenter monitoring system on Solaris used to have a kernel module for monitoring (I don’t know why). My experience was that Unicenter caused many kernel panics and more downtime than all other problems combined. It would not be difficult to write better code than the typical CA employee, but writing code that is good enough to have a monitoring system running in the kernel on a single-threaded system is asking a lot.

One of the claimed benefits of a Unikernel was that it’s supposedly risky to allow ssh access. The recent ssh security issue was an attack against the ssh client if it connected to a hostile server. If you had a ssh server only accepting connections from management workstations (a reasonably common configuration for running servers) and only allowed the ssh clients to connect to servers related to work (an uncommon configuration that’s not difficult to implement) then there wouldn’t be any problems in this regard.

I think that I’m a good programmer, but I don’t think that I can write server code that’s likely to be more secure than sshd.

On Designing It Yourself

One thing that everyone who has any experience in security has witnessed is that people who design their own encryption inevitably do it badly. The people who are experts in cryptology don’t design their own custom algorithm because they know that encryption algorithms need significant review before they can be trusted. The people who know how to do it well know that they can’t do it well on their own. The people who know little just go ahead and do it.

I think that the same thing applies to operating systems. I’ve contributed a few patches to the Linux kernel and spent a lot of time working on SE Linux (including maintaining out of tree kernel patches) and know how hard it is to do it properly. Even though I’m a good programmer I know better than to think I could just build my own kernel and expect it to be secure.

I think that the Unikernel people haven’t learned this.

• [1] https://en.wikipedia.org/wiki/Unikernel