== A NFS mount accident on Linux

Something that you can do on a modern Linux system by accident:

> _mount -t nfs -o hard,intr,rw localhost:/ /_

(instead of mounting it on _/mnt_, as I had intended.)

It is surprisingly hard to recover from this. In fact, I don't think I
succeeded, and I wound up having to reboot. At the same time, I don't
think anything broke, so I theoretically could have kept on running the
machine like that.

(Why you might want such a loopback mount is covered [[here
../sysadmin/PeekingUnderMountsWithNFS]].)

On Linux specifically, another way of achieving the same goal is to
use a bind mount:

> _mount --bind / /mnt_

This doesn't need NFS daemons, but I'm an old dog and sometimes we
automatically reach for our old tricks, no matter how complicated
they are.

=== Things I tried that don't work, so you can skip them

- _umount localhost:/_: Did nothing.
- _umount /_ and _umount -t nfs /_: Either did nothing or complained
  that _/_ was busy.
- _umount -l -t nfs /_: Unmounted everything except for the rootfs
  mount of _/_, which leaves the system pretty much unrecoverable
  afterwards since udev's _/dev_ is gone, so you don't have any
  devices to mount stuff from.

From this it looks like lazy unmounts detach subtrees of the mount that
you're unmounting, as well as the mount itself, which I suppose is not
too surprising.

=== Sidebar: how to keep access to _/proc/mounts_

When _/proc_ becomes inaccessible, how do you find out what is
and isn't mounted?

It's a fairly core principle of Linux that while you may get
detached from the filesystem tree, your current directory
doesn't actually go away. So I did:

>  ; cd /proc
>  ; python
>  >>> import os, sys
>  >>> def cat(fn):
>  ...   fp = open(fn, "r")
>  ...   fd = fp.read()
>  ...   sys.stdout.write(fd)
>  >>> cat("mounts")

(The Python transcript has been slightly simplified.)

Using Python and importing stuff ahead of time meant that I wasn't
counting on things like _/bin/cat_ remaining accessible; everything
I needed to monitor _/proc/mounts_ was live in a running process.
This turned out to be a good thing.