Re: [PATCH 12/18] shared mount handling: bind and rbind

[Al Viro <viro@ftp.linux.org.uk>]

On Wed, Nov 09, 2005 at 04:22:23PM +0100, Miklos Szeredi wrote:
> > On Wed, Nov 09, 2005 at 11:54:36AM +0100, Miklos Szeredi wrote:
> > > Shouldn't this check go before copy_tree()?  Not much point in copying
> > > the tree if we are sure it won't be used.
> > 
> > Incorrect.  Propagation nodes further down the tree can very well be
> > mountable.
> 
> Can you please give an example.  I'm feeling thick.
> 
> What I see in the code is that the the mount tree is copied, then put
> on the tmp_list, and at the end the newly copied tree is freed with
> umount_tree().

Before it gets freed it may end up being copied.  Example: vfsmounts
A and B are peers, C is a slave of that peer group.  It happens to be
on slave list of B.  B has root deeper than A, which, in turn is deeper
than that of C (e.g. A and B had been created by binding subtrees of
C, which had been made slave afterwards).  We bind on something in A,
outside of the subtree mapped by B.

Alternatively, have A -> (B, D) -> C, with C on slave list of B.  Mountpoint
is within subtrees for A, C and D, but not B.  And no, we can't say "skip B,
just make a slave of tree on A and slap it on C" - correct result is to
have T_A -> T_D -> T_C (i.e. tree on C gets propagation from tree on D).
Which kills the variants with not creating that copy and making subsequent
ones directly from the original tree.