Sockets don't work through mount about go-nfs HOT 11 CLOSED

https://superuser.com/questions/1608873/does-a-unix-domain-socket-in-an-nfs-exported-fs-magically-work-across-hosts

I don't know if there's a way to communicate the internal kernel state about how the socket is supposed to work through the nfs protocol

from go-nfs.

Interestingly, if I try to start nc using the socket in the mount (so, nc -lU mount/test.sock instead of nc -lU test.sock), I get [ERROR] call to 0x11ae70 failed: Operation not supported. If the file already exists, Remove is called.

from go-nfs.

It's probably because onMknod has no support.

from go-nfs.

good catch. it should be reasonable straight forward to add that - i think the trick will be that the underlying billy fs interface also doesn't expose anything around this, so we'll need to add another optional interface for handlers to implement if they want to expose that on their filesystem implementation.

from go-nfs.

Yes --- we could just copy the syscall argument list --- path, mode, device number.

from go-nfs.

Do you have a sense of what the underlying system call should be when Mknod for a socket is sent over NFS?

• the rfc indicates that Mknod can take a file type of NF3SOCK indicating creation of a socket. When that is requested, the only passed data is the sattr3 pipe_attributes; - which does not include a major or minor number.

from go-nfs.

I think i have a sketch of the expected backing behavior - working through it in https://github.com/willscott/go-nfs/tree/feat/mknod and will PR once ready.

The rfc is pretty thin on details on the expected semantics of a created socket in particular, but this may be close enough

from go-nfs.

the branch has a first pass at what seems like plausible for this code, and the osnfs example supporting mknod operations.

if you can point me to concrete repro steps for things not working ("on ubuntu, make an nfs mount, run these lines, get error") i'll be able to get to fixing potential bugs in here faster

from go-nfs.

Sorry, didn't see this till now: interestingly, you can't create just "hanging" sockets, but it looks like you figured something out.

There are two reproducers: one, create a mount exposing a local directory --- mount 127.0.0.1:dir mount. Run nc -lkU dir/test.sock. Try to do nc mount/test.sock. You should get a "Connection refused". However, this does not seem possible to solve.

Another reproducer: Do nc -lkU mount/test.sock. Then try to do nc mount/test.sock. You should get a Remove call and an "Operation not permitted". This is something that should be fixable.

from go-nfs.

Updated the branch linked above. I was able to create a socket and a fifo over the mount locally at least.

can you see if that gives you better behavior?

from go-nfs.

It does seem to work now --- the second reproducer now leads to the expected behavior (I can connect and send messages back and forth). #91 is still an issue though, so sockets wasn't the issue. Thanks!

from go-nfs.