Re: [PATCH] Revert "__d_unalias() should refuse to move mountpoints"

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

On Tue, Sep 25, 2012 at 04:29:58AM -0700, Eric W. Biederman wrote:
> Maarten Lankhorst <maarten.lankhorst@canonical.com> writes:
> 
> >> Could you try the following patch?  This should report what directories
> >> cannot be renamed because one of them is a mount point and it gives some
> >> real insight into what is going on.
> >
> > ls /
> > __d_unalias: /dev -> /dev
> > __d_unalias: /proc -> /proc
> > __d_unalias: /sys -> /sys
> 
> Ok.  That is what I thought was going on.  For some reason nfs is
> attempting to recreate an existing dentry.
> 
> Does this fix the nfs problem for you?
> 
> Eric
> 
> diff --git a/fs/dcache.c b/fs/dcache.c
> index 8086636..6390f0f 100644
> --- a/fs/dcache.c
> +++ b/fs/dcache.c
> @@ -2404,6 +2404,9 @@ out_unalias:
>  	if (likely(!d_mountpoint(alias))) {
>  		__d_move(alias, dentry);
>  		ret = alias;
> +	} else if ((alias->d_parent == dentry->d_parent) &&
> +		   !dentry_cmp(alias, dentry->d_name.name, dentry->d_name.len))
> +		ret = alias;
>  	}

The interesting question is why the hell had it decided that preexisting
dentry was not good enough for it?  Note that we have arrived to nfs_lookup()
after we'd decided *not* to use the damn alias.  The trace posted upthread
went __lookup_hash() -> lookup_real().  It means that lookup_dcache()
has not produced this one.  And no, even if ->d_revalidate() decided it
was no good, the logics in d_invalidate() would've said "busy" and we'd
gone with that dentry anyway.  So it means that d_lookup() has not
found it at all.

IOW, something out there is blindly unhashing mountpoint dentries; that's
where the real root of the problem seems to be.  Could you slap
WARN_ON(d_mountpoint(dentry)) in __d_drop() and see what it catches?