christos@zoulas.com (Christos Zoulas) writes:

> I don't think so, that is why I did not want to mask the bug. I think
> we should put more DIAGNOSTIC stuff somewhere. Maybe the bug is in
> checkalias() as you suggested? what happens if your nfs mounted partition
> has device nodes? I guess you could not trigger the problem because
> your nfs mount did not have devices?

Still, I couldn't reproduce it by myself.  checkalias() is one
candidate since it may remove vnode from per mount vnode list, but I
can't find such code path yet.  Another interesting point is that
v_size of failed vnode is -1 in the original PR.  Such vnode may
exists if VOP_GETATTR() failed in nfs_nget() (I believe it should
vgone() instead of vput()), but it is unclear why it is left after
unmount.

enami.