Re: [PATCH] fs/dcache: allow __d_obtain_alias() to return unhashed dentries

["J. Bruce Fields" <bfields@fieldses.org>]

On Thu, Jun 28, 2012 at 09:59:27AM -0400, J. Bruce Fields wrote:
> Coming back to this now, just trying to review the
> filehandle-lookup/dcache interactions:
> 
> On Fri, Mar 11, 2011 at 03:07:49PM +1100, NeilBrown wrote:
> > 1/ Originally DCACHE_DISCONNECTED didn't really mean much - it's presence
> >    was only a hint, its absence was a strong statement.
> >    If the flag is set, the dentry might not be linked to the root.
> >    If it is clear, it definitely is link through to the root.
> >    However I think it was used with stronger intent than that.
> > 
> >    Now it seems to mean a little bit more:  If it is set and the dentry
> >    is hashed, then it must be on the sb->s_anon list.
> 
> The code that makes that assumption is __d_shrink (which does the work
> of d_drop)--it uses DCACHE_DISCONECTED to decide which hash chain to
> lock.
> 
> I can't find any basis for that assumption.  The only code that clears
> DCACHE_DISCONNECTED is in expfs.c, and it isn't done at the same time as
> hashing.  Am I missing something?
> 
> >    This is a significant
> >    which I never noticed (I haven't been watching).  Originally a
> >    disconnected dentry would be attached (and hashed) to its parent.  Then
> >    that parent would get its own parent and so on until it was attached all
> >    the way to the root.  Only then would be start clearing
> >    DCACHE_DISCONNECTED.  It seems we must clear it sooner now... I wonder if
> >    that is correct.
> 
> It looks wrong to me:
> 
> If we clear DCACHE_DISCONNECTED too early, then we risk a filehandle
> lookup thinking the dentry is OK to use.  That could mean for example
> trying to rename across directories that don't have any ancestor
> relationship to each other in the dcache yet.
> 
> So we need to wait to clear DCACHE_DISCONNECTED until we *know* the
> dentry's parents go all the way back to the root.  As you say, that's
> what the current code does.
> 
> But that means DCACHE_DISCONNECTED dentries can be hashed to their
> parents, and __d_shrink can be handed such dentries and then get the
> locking wrong.
> 
> It looks like this bug might originate with Nick Piggin's ceb5bdc2d246
> "fs: dcache per-bucket dcache hash locking"?  There's no discussion in
> the changelog, so probably it was just based on an unexamined assumption
> about DCACHE_DISCONNECTED.
> 
> I wonder if an IS_ROOT() test could replace the DCACHE_DISCONNECTED test
> in __d_shrink(), or if we need another flag, or ?

Bah, sorry, and I only just noticed that you already said as much later
and did the IS_ROOT() thing in your patch.

Anyway, here's just that one change with a slightly more painstaking
changelog.

--b.

commit b1fa644355122627424fe2240a9fc60cbef4c349
Author: J. Bruce Fields <bfields@redhat.com>
Date:   Thu Jun 28 12:10:55 2012 -0400

    dcache: use IS_ROOT to decide where dentry is hashed
    
    Every hashed dentry is either hashed in the dentry_hashtable, or a
    superblock's s_anon list.
    
    __d_shrink assumes it can determine which is the case by checking
    DCACHE_DISCONNECTED; this is not true.
    
    It is true that when DCACHE_DISCONNECTED is cleared, the dentry is not
    only hashed on dentry_hashtable, but is fully connected to its parents
    back to the root.
    
    But the converse is *not* true: fs/exportfs/expfs.c:reconnect_path()
    attempts to connect a directory (found by filehandle lookup) back to
    root by ascending to parents and performing lookups one at a time.  It
    does not clear DCACHE_DISCONNECTED until its done, and that is not at
    all an atomic process.
    
    In particular, it is possible for DCACHE_DISCONNECTED to be set on a
    dentry which is hashed on the dentry_hashtable.
    
    Instead, use IS_ROOT() to check which hash chain a dentry is on.  This
    *does* work:
    
    Dentries are hashed only by:
    
    	- d_obtain_alias, which adds an IS_ROOT() dentry to sb_anon.
    
    	- __d_rehash, called by _d_rehash: hashes to the dentry's
    	  parent, and all callers of _d_rehash appear to have d_parent
    	  set to a "real" parent.
    	- __d_rehash, called by __d_move: rehashes the moved dentry to
    	  hash chain determined by target, and assigns target's d_parent
    	  to its d_parent, before dropping the dentry's d_lock.
    
    Therefore I believe it's safe for a holder of a dentry's d_lock to
    assume that it is hashed on sb_anon if and only if IS_ROOT(dentry) is
    true.
    
    I believe the incorrect assumption about DCACHE_DISCONNECTED was
    originally introduced by ceb5bdc2d246 "fs: dcache per-bucket dcache hash
    locking".
    
    Cc: Neil Brown <neilb@suse.de>
    Cc: Nick Piggin <npiggin@kernel.dk>
    Signed-off-by: J. Bruce Fields <bfields@redhat.com>

diff --git a/fs/dcache.c b/fs/dcache.c
index 87c2da7..b2b382c 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -410,7 +410,7 @@ static void __d_shrink(struct dentry *dentry)
 {
 	if (!d_unhashed(dentry)) {
 		struct hlist_bl_head *b;
-		if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
+		if (unlikely(IS_ROOT(dentry->d_flags)))
 			b = &dentry->d_sb->s_anon;
 		else
 			b = d_hash(dentry->d_parent, dentry->d_name.hash);