fs: mount namespace within runsc doesn't seem to work

[yoshiat]

From @AkihiroSuda here.
The context is that he has been looking into running BuildKit on UML/gVisor on Docker on real Linux kernel. And BuiltKit needs to invoke runc or gVisor. As running runc may require more syscalls (and options) to be supported, he thought BuildKit could run by replacing runc with unshare -m chroot if gVisor supports it.

$ docker run -it --rm --runtime=runsc alpine \
/ # unshare --uts --ipc --net --pid --user echo hi \
hi
/ # unshare --mount echo hi
unshare: can't mount none on / (flags:0x44000): Bad address
/ #

Before going down this path, he got stuck with the lack of MS_BIND|MS_REC support which should be tracked in #179.

[amscanne]

Yes, most other namespaces are implemented but not mount namespaces. It will follow after the major file system refactor in progress.

[pkit]

vfs2 is the default. Can it be implemented now? I can poke it, if needed.

[kevinGC]

@nixprime WDYT?

[pkit]

@kevinGC it seems like it has some TODO here

gvisor/pkg/sentry/kernel/task_clone.go

Line 128 in 5ffcc1f

mntnsVFS2 := t.mountNamespaceVFS2

[pkit]

Ah, bind mounts are not there too. Probably needs bind mounts first...

[avagin]

Fixed by 41bb04c.