Making io_uring pervasive in QEMU [pdf]

(vmsplice.net)

71 points | by ingve 13 hours ago ago

7 comments

  • pm215 10 hours ago

    If you want to watch the talk that these are the slides for, it's now up on youtube along with the other KVM Forum talks: https://youtu.be/gSB5sn3ZN3w

  • rictic 10 hours ago

    How is security looking with io_uring these days? I've been wary of it since https://security.googleblog.com/2023/06/learnings-from-kctf-...

    • seangrogg 8 hours ago

      I've only dabbled, so I'm happy to have people with more linux-side knowledge to call me out on any inaccuracies here, but...

      io_uring is effectively as "secure" as any other syscall unto itself. The issue is that the mechanism by which io_uring makes its syscalls as part of its submission/completion queues means that those underlying syscalls can't be filtered by seccomp. The real question is your security posture.

      If you're writing a hypervisor that's intended to partition resources between underlying users in a secure fashion, the ability for io_uring to bypass seccomp is largely a non-starter. But if you own the machine and you just want to run an application on it (i.e. an HTTP server that uses io_uring for file/network io) you should largely be in the clear.

      • linuxnewb99 5 hours ago

        Does it no longer suffer from TOCTOU?

    • JoshTriplett 7 hours ago

      Security with io_uring is great these days. Many years ago it moved away from the original architecture that led to several security issues; its current architecture is no more prone to security issues than any other part of the kernel.

      For context, the original architecture involved having privileged kernel-side offload processing that had to carefully drop privileges each time it did something on behalf of the userspace process. As you can imagine, that fail-insecure architecture was heavily prone to security holes.

      io_uring got rid of that architecture years ago, in favor of running with the permissions of the userspace process. With that change, there's no longer any reason to consider io_uring any less secure than the rest of the kernel.

      • 1oooqooq 5 hours ago

        wasn't the main issue about the asynchronous nature of the calls?

        • JoshTriplett 5 hours ago

          As far as I know, the new architecture still handles asynchronous offloading, it just uses a more secure-by-default means of doing so.