Re: d_splice_alias() problem.

[Greg Banks <gnb@melbourne.sgi.com>]

Neil Brown wrote:
> 
> > Of course I now have an issue with the misleading name ;-)
> 
> Maybe:   DCACHE_NOT_KNOWN_TO_BE_CONNECTED.
> Unfortunately the absence of the flags is stronger information that
> it's presence and that makes it hard to name...

;-)

> > What I'm wondering is, do we still need DCACHE_DISCONNECTED at all?
> > Perhaps the uses of it could be replaced with combinations of checks
> > of IS_ROOT() and (d == d->d_sb->s_root) ?
> 
> It is still needed.
> Suppose one thread creates a disconnected dentry, and then starts building
> the path from the bottom up.
> When it is half way up another request comes in for the same
> filehandle.  The same dentry is found.  It is now not IS_ROOT, but
> still DCACHE_DISCONNECTED.  Until it is fully connected the second
> request shouldn't proceed, and without the DCACHE_DISCONNECTED flag it
> is expensive to test for connected-ness.  !IS_ROOT certainly isn't
> enough.

Sure, IS_ROOT() only tells you whether a dentry is connected to its parent
and to tell if its connected to the fs root you need to traverse all the
parents.  But nfsd_acceptable() already does this to do permission checks
in the (default) subtree_check case.

So it seems that what the flag gives you is that when its absent and
nosubtree_check is in effect, you know you can short-circuit the connection
walk.  Fair enough.

Greg.
-- 
Greg Banks, R&D Software Engineer, SGI Australian Software Group.
I don't speak for SGI.