Consider using kqueue in hypervisor socket implementation

[djs55]

The core of hyperkit uses kqueue via mevent.c but pci_virtio_sock.c uses plain old select.

If hyperkit is used together with vpnkit in Docker for Mac and large numbers (> 1024) of connections are port forwarded then hyperkit becomes unable to process any more AF_VSOCK connections due to the accepted file descriptor being greater than FD_SETSIZE, see for example

hyperkit/src/lib/pci_virtio_sock.c

Line 1364 in 3ace985

if (fd >= FD_SETSIZE) {

This manifests as errors under load, for example

$ docker ps
Error response from daemon: Bad response from Docker engine

from docker/for-mac#2841

This scalability limit could be removed by switching from select to kqueue (or poll) in pci_virtio_sock.c.

[imavroukakis]

Hi team, any updates on this would be appreciated

[ijc]

At one point we were trying to avoid MacOS-isms in an attempt to keep the option of using Hyperkit on other platforms open in the future. I think we've basically given up on that so a PR to switch to kqueue should be fine.

I had thought that kqueue was callback based, making it a big rewrite to integrate with the current code structure, but it seems that I was mistaken and it should be possible to replace select with kqueue without a lot of non-localised changes.

Have we considered whether multiplexing in vpnkit might be easier though?

[imavroukakis]

Has there been any more discussion around this ? Would love to see it fixed 😄

[ijc]

AFAIK nobody has worked on this. PRs are welcome.

[imavroukakis]

@ijc thanks for the update. I would happily have given this a go but C is not something I'm anywhere near competent in.