In praise of qemu September 2, 2022 on Drew DeVault's blog

qemu is another in a long line of great software started by Fabrice Bellard. It provides virtual machines for a wide variety of software architectures. Combined with KVM, it forms the foundation of nearly all cloud services, and it runs SourceHut in our self-hosted datacenters. Much like Bellard’s ffmpeg revolutionized the multimedia software industry, qemu revolutionized virtualisation.

qemu comes with a large variety of studiously implemented virtual devices, from standard real-world hardware like e1000 network interfaces to accelerated virtual hardware like virtio drives. One can, with the right combination of command line arguments, produce a virtual machine of essentially any configuration, either for testing novel configurations or for running production-ready virtual machines. Network adapters, mouse & keyboard, IDE or SCSI or SATA drives, sound cards, graphics cards, serial ports — the works. Lower level, often arch-specific features, such as AHCI devices, SMP, NUMA, and so on, are also available and invaluable for testing any conceivable system configurations. And these configurations work, and work reliably.

I have relied on this testing quite a bit when working on kernels, particularly on my own Helios kernel. With a little bit of command line magic, I can run a fully virtualised system with a serial driver connected to the parent terminal, with a hardware configuration appropriate to whatever I happen to be testing, in a manner such that running and testing my kernel is no different from running any other program. With -gdb I can set up gdb remote debugging and even debug my kernel as if it were a typical program. Anyone who remembers osdev in the Bochs days — or even earlier — understands the unprecedented luxury of such a development environment. Should I ever find myself working on a hardware configuration which is unsupported by qemu, my very first step will be patching qemu to support it. In my reckoning, qemu support is nearly as important for bringing up a new system as a C compiler is.

And qemu’s implementation in C is simple, robust, and comprehensive. On the several occasions when I’ve had to read the code, it has been a pleasure. Furthermore, the comprehensive approach allows you to build out a virtualisation environment tuned precisely to your needs, whatever they may be, and it is accommodating of many needs. Sure, it’s low level — running a qemu command line is certainly more intimidating than, say, VirtualBox — but the trade-off in power afforded to the user opens up innumerable use-cases that are simply not available on any other virtualisation platform.

One of my favorite, lesser-known features of qemu is qemu-user, which allows you to register a binfmt handler to run executables compiled for an arbitrary architecture on Linux. Combined with a little chroot, this has made cross-arch development easier than it has ever been before, something I frequently rely on when working on Hare. If you do cross-architecture work and you haven’t set up qemu-user yet, you’re missing out.

$ uname -a
Linux taiga 5.15.63-0-lts #1-Alpine SMP Fri, 26 Aug 2022 07:02:59 +0000 x86_64 GNU/Linux
$ doas chroot roots/alpine-riscv64/ /bin/sh
# uname -a
Linux taiga 5.15.63-0-lts #1-Alpine SMP Fri, 26 Aug 2022 07:02:59 +0000 riscv64 Linux

This is amazing.

qemu also holds a special place in my heart as one of the first projects I contributed to over email 🙂 And they still use email today, and even recommend SourceHut to make the process easier for novices.

So, to Fabrice, and the thousands of other contributors to qemu, I offer my thanks. qemu is one of my favorite pieces of software.