Re: [PATCH 00/20] Mount writer count and read-only bind mounts (v3)

[Dave Hansen <haveblue@us.ibm.com>]

On Tue, 2006-06-27 at 18:38 -0700, Andrew Morton wrote:
> On Tue, 27 Jun 2006 15:14:36 -0700
> > One note: the previous patches all worked this way:
> > 
> > 	mount --bind -o ro /source /dest
> > 
> > These patches have changed that behavior.  It now requires two steps:
> > 
> > 	mount --bind /source /dest
> > 	mount -o remount,ro  /dest
> 
> That seems a step backwards.

It is, in a way.  But, it keeps the bind mounting process itself a much
simpler operation in the kernel.  The --bind operation itself stays just
a matter of copying a vfsmount.  Otherwise, you end up trying to manage
a bunch of state transitions if, for instance, the source vfsmount is
r/w and the bind is requested to be r/o.

Plus, the previous behavior was only established by the original
out-of-tree patches from vserver.  Herbert, this doesn't cause you too
much of a headache, right?

> > Since the last revision, the locking in faccessat() and
> > mnt_is_readonly() has been changed to fix a race which might have
> > caused a false-negative mount-is-readonly return when faccessat()
> > is called while another two processes are racing to make a mount
> > readonly.
> > 
> umm, what's it all for?

Mostly for vserver, for now.  They allow a filesystem to be r/w, but
have r/o views into it.  This is really handy so that every vserver can
use a common install but still allow the administrator to update it.

-- Dave