[arch-projects] [initscrupts][RFC] umount tmpfs and ramfs on shutdown

[Tom Gundersen]

On Mon, Apr 25, 2011 at 2:42 AM, Seblu <seblu at seblu.net> wrote:
> On Sun, Apr 24, 2011 at 11:55 PM, Tom Gundersen <teg at jklm.no> wrote:
>> a process X (some userspace filesystem daemon) cannot be killed as
>> long as a mount Y is mounted
>> process X has a file opened rw on mount Z (maybe a logfile on /var)
>> a tmpfs/ramfs is a submount of Y
>
> Tricky case. Unmouting tmpfs will not solve the case.
> Mount Z cannot be unmounted until X process stop and process cannot
> unmount until it stop to write his log.
> It's a bad designed mounting.

Solutions would be:
unmount -a (this unmounts Y and its submounts)
killall (this kills X)
unmount -a (this unmounts Z)

>> I have been looking at what systemd is doing and it seems quite
>> robust. Their shutdown binary doesn't seem to work for us though.
> And what's your conclusion?

That we should look to it for inspiration (and that we sadly cannot use it...).

> Another question why we don't unmount _netdev marked fs? Network fs
> should not be unmounted?

I don't know why this is not done, I would have thought that a forced
unmount of _netdev would be needed just as unmounting tmpfs would.

I think a proper solution would be to write a tool that would do
something like the following:

1) remount everything ro
2) if it fail kill everything and try again
3) if it still fails unmount everything
4) if it still fails kill everything and try again
 (with we should loop the retries until they time out where it makes sense)

It would probably make sense to do the unmounting/killing manually by
parsing /proc/self/mountinfo and /proc, which would allow us to
blacklist mounts (/dev, /run, /proc, /sys,...) and processes (as
requested on FS).

Alternatively, we could look into what it would take to be able to use
the systemd shutdown binary which does something like this already
(and should in principle work independently of systemd).

Until someone writes something like this, I suggest we keep the commit
I made (with your fixes added), as the "least wrong" solution for now.

-t