--- title: Technology author: someone abstract: some abstract --- We develop digital infrastructure with a minimal footprint. Where other approaches try to patch general purpose operating systems by adding more layers of indirection, we strive to build a secure system from the ground up. Each piece of digital infrastructure or service is written in a high-level memory-safe programming language and tailored to only contain the required functionality at compilation time. This reduces the attack vectors and the attack surface. The resulting service is executed as a virtual machine on a modern hypervisor. Its size is usually around 1-10 MB, much smaller than a UNIX / Linux system, and boots within milliseconds. ## MirageOS - bespoke operating systems Our work is based on MirageOS, a suite to build operating systems. It has been developed since 2009 at University of Cambridge, UK and is written in the programming language OCaml (see [Why OCaml](#Why-OCaml)). Most libraries are developed as open source (MIT/ISC/BSD2/Apache2). MirageOS is a library operating system. It composes OCaml libraries into a bespoke operating system, called a unikernel. A unikernel can be a compiled as a UNIX binary, or a standalone virtual machine image. To build the right unikernel for your custom business logic, we can pick from hundreds of libraries which implement network protocols, storage on block devices, or interfaces to network devices via the hypervisor. On top of the hypervisor, a small layer of C code unifies the interface on which OCaml runs. OCaml is a functional programming language that minimizes side effects and mutable state. Its functional programming concepts give us a list of security advantages for MirageOS. ## Running unikernel, system security Aside from automated memory management to avoid memory corruption, and type checking to avoid many common programming errors, the major advantage of functional programming is localized reasoning about program code. All inputs, outputs and effects of a function are known. Immutable datastructures and cooperative multitasking allow us to reason about the state of the entire system, even if we use parallelism and complex distributed systems. ### Simple config management model with localized reasoning There are three ways to feed a virtual machine with configuration data like network configuration or TLS certificate and key: - compile the information into the virtual machine image, which requires recompilation on configuration change - pass the information as boot parameters, which requires reboot on configuration change - store this information in a virtual block device attached to the virtual machine. For example, logs can be written from the unikernel to a syslog collector with UDP, TCP, or TLS as transport. The transport needs to be chosen at compile time because TLS requires the TLS library to be linked into the kernel image, but the log destination is passed as boot parameter. ### Simple concurrency model with localized reasoning MirageOS is an event based operating system with asynchronous tasks. A task yields the CPU once its execution is finished, or if it has to wait for IO. This concurrency model leads to a cooperative multitasking programming style, rather than the error prone preemptive multitasking, where each code block needs to make sure to use appropriate locking strategies to avoid reentrant execution errors. A recent example for code which is not safe under reentrant execution [in Ethereum](http://hackingdistributed.com/2016/06/18/analysis-of-the-dao-exploit/) lead to a huge amount of ether being transferred. Established software like the [Firefox JavaScript engine](http://www.nist.org/news.php?extend.175), or [PHP](https://bugs.php.net/bug.php?id=74308) shows similar problems on a regular basis. ### Simple process memory model with localized reasoning The virtual memory subsystem in contemporary operating systems provides an address mapping for each process. Since a unikernel is only a single service, it uses a single address space, avoiding the need for complex address mapping code altogether. An example for corrupting the page table is [Xen's XSA-182](http://xenbits.xen.org/xsa/advisory-182.html). ### Simple library model with localized reasoning A MirageOS unikernel is much smaller than a comparable UNIX virtual machine. By avoiding superfluous code we decrease the attack surface immensly. Consider the breakdown of the code of the example system [Bitcoin Piñata](/Projects/Pinata) compared to a virtual machine using Linux and OpenSSL, measured in thousands of lines of code:
LinuxMirageOS
Kernel160048
Runtime68925
Crypto23023
TLS416
Total2560102
### Secure updates If a security flaw is discovered in a library, and there is a security update, all unikernels depending on this library need to be updated as well. This can be done with the [OCaml package manager](https://opam.ocaml.org). It resolves dependencies and lets [authors sign their releases](https://github.com/hannesm/conex), so there is no need for a central package repository server. Central repository servers are known targets for attackers and have been breached in the past, amongst them the [Linux kernel](https://lwn.net/Articles/57135/), [FreeBSD infrastructure](https://www.freebsd.org/news/2012-compromise.html), [Debian](https://www.debian.org/News/2003/20031202) and [PHP](http://php.net/archive/2013.php#id2013-10-24-2). ## Why OCaml ### Functional programming style As discussed in the system security paragraph, many security advantages of MirageOS are based on the choice of a programming language in which problems can be solved in a functional style. This style allows us to reason about the possible states of a system. ### Performance OCaml code compiles to native code, which is as performant as compiled C++ code. As an example, our [TLS library](https://usenix15.nqsb.io) has up to 85% of the bulk throughput of OpenSSL (using AES128-CBC). The TLS handshake performance is comparable with OpenSSL. ### Dependency management MirageOS leverages OCaml's module system to adapt the unikernel to the compilation target. Each operating system service in MirageOS is a module, for example the console, the network stack, the random number generator. Each of the services has multiple implementations that are chosen based on the target. On UNIX, the sockets API of the host is used as networking stack. On a unikernel, the TCP/IP stack natively implemented in OCaml is being used. A MirageOS developer does not need to reason about compilation targets, just about the module interface. ### Verification A large subset of the OCaml semantics has been [mechanically proven sound](http://www.cl.cam.ac.uk/~so294/ocaml/) in a theorem prover. OCaml is the implementation language of the well-known proof assistant [Coq](https://coq.inria.fr). Development in Coq can be extracted to OCaml code, as demonstrated by [compcert](http://compcert.inria.fr/), a formally verified optimizing C compiler, in order to be compiled and executed. The other direction is also possible: OCaml code can be translated into Coq definitions (using [Coq of OCaml](https://github.com/clarus/coq-of-ocaml/)). The National Cybersecurity Agency of France reviewed the implementation of the OCaml runtime system, [their report](http://www.ssi.gouv.fr/agence/publication/lafosec-securite-et-langages-fonctionnels/) prompted some language modifications, such as that strings are no longer mutable. ### Modern dialects and compile targets OCaml is a mature programming language that is used both in industry (Facebook, Jane Street Capital, Docker, ahrefs, simcorp, lexifi) and academia. The OCaml compiler generates native code for x86, arm, etc., and has a bytecode backend, which can target microcontrollers (PIC18 family in the [OcaPIC project](http://www.algo-prog.info/ocapic/web/?id=OCAPIC:OCAPIC)). OCaml can also be compiled to JavaScript, so both client and server side of a web application can be developed in the same language with shared interface code (more details at the [ocsigen project](http://ocsigen.org/)). In 2016, Facebook developed [reason](https://reasonml.github.io/), a dialect of OCaml which syntax is closer to JavaScript, and easier to comprehend for beginners. Reason and OCaml code can be easily combined into a single application, since they use the same compiler. More literature on why OCaml is a good choice has been written by Yaron Minsky (Jane Street) in the article [OCaml for the masses](http://queue.acm.org/detail.cfm?id=2038036), and more recently by the crypto-ledger [tezos](https://www.tezos.com/static/papers/position_paper.pdf).