Re: [PATCH 5/7] autofs4 - fix direct mount pending expire race

[Ian Kent <raven@themaw.net>]

On Sat, 2008-07-19 at 10:09 +0800, Ian Kent wrote:
> On Fri, 2008-07-18 at 16:08 -0400, Jeff Moyer wrote:
> > 
> > This makes me uneasy.  This should take d_mounted to zero.  Then, when
> > the daemon actually does the unmount, won't the d_mounted drop below
> > zero?  Following calls to d_mountpoint will return a negative value, but
> > everyone treats it as a boolean, so it will evaluate to true for a brief
> > time.  Or did I miss something?
> 
> Yes, I thought about doing exactly that.
> 
> But the thing that effects d_mounted is mounted on the dentry and so
> d_mounted may be decremented during the expire. So if we set it
> explicitly it would be incorrect at the end. While the
> decrement/increment isn't always correct throughout the expire we need
> to handle the mount following in ->follow_link() anyway and then the
> decrement/increment ends up with the correct value once the expire is
> complete.
> 
> One problem that has occurred to me is that user space could could
> manually umount it just when we change it. So a follow up patch to add a
> lock around the increment/decrement in fs/namespace.c and
> fs/autofs4/expire.c is in order. I'm having a look at that now.

DOH, this is rubbish.

I often wonder why I forget what I've done so quickly.

Of course user space can't umount this because every process except the
daemon is blocked in ->follow_link() during the expire. But we can't
rely on the the return from user space and cannot know if the mount was
actually umounted so the decrement/increment will maintain the
status-quo. If d_mounted does become negative during the expire then we
will still be sent to ->follow_link() as follow_down() will fail.

Ian