Hi, On 03/20/2017 11:27 PM, Volth wrote:
> 9P is used by NixOS to share host's nix store with Qemu virtual > machines. Such technique is used in the build process, in the > test-driver, so to say in the critical places. > Recently few bugs in 9P were found (#23957 #23020 #22695) which > reveals that 9P code is not very mature and perhaps NixOS is the first > team which uses 9P heavily and relies on it in production. I wouldn't call using it in tests "production" :-) > Shouldn't we replace 9P with something battle-tested like NFS or Samba? We used to use Samba/CIFS for the VM tests, but it was slower than 9p [1], required TCP/IP at an early stage in the initrd, and required babysitting a Samba process, whereas 9p is built into QEMU. It was also not super-reliable (e.g. CIFS mounting had a tendency to time out if the host was under high load). A while ago I played with Ganesha NFS but couldn't get it to work. It's probably doable but much more complicated than passing a couple of flags to QEMU. [1] https://github.com/NixOS/nixpkgs/commit/0e386d0c1373809d3459d09532ad36be0faf6dfb > It may also improve the performance because 9P server works in qemu > process, in user mode and there are as many servers as virtual > machines running. It would probably be slower, requiring more context switches between the VM, QEMU and the FS server. Also, 9p uses the virtio transport which is ideal for VM <-> host communication. Vladimir wrote: >> In terms of performance getting rid of QEMU where possible is probably >> the better option. Containers are fairly mature these days, and then >> sharing file-systems is a matter of bind-mounting. > > For some tests containers certainly aren't enough, as we do want to test > booting, the kernel, etc. Also, running a NixOS system requires a full UID namespace, whereas currently Nix builds only get a single UID. For tests that don't require testing the kernel or initrd, running them as a container would be great though. -- Eelco Dolstra | LogicBlox, Inc. | http://nixos.org/~eelco/ _______________________________________________ nix-dev mailing list nix-dev@lists.science.uu.nl http://lists.science.uu.nl/mailman/listinfo/nix-dev
