Andrew Morton wrote:
> John McCutchan <john@johnmccutchan.com> wrote:
> 
>> > > @@ -538,7 +537,7 @@
>> > >  	struct dentry *parent;
>> > >  	struct inode *inode;
>> > >  
>> > > -	if (!atomic_read (&inotify_watches))
>> > > +	if (!atomic_read (&dentry->d_sb->s_inotify_watches))
>> > >  		return;
>> > >  
>> > 
>> > What happens here if we're watching a mountpoint - the parent is on a
>> > different fs?
>>
>> There are four cases to consider here.
>>
>> Case 1: parent fs watched and child fs watched
>> 	correct results
>> Case 2: parent fs watched and child fs not watched
>> 	We may not deliver an event that should be delivered.
>> Case 3: parent fs not watched and child fs watched
>> 	We take d_lock when we don't need to
>> Case 4: parent fs not watched and child fs not watched
>> 	correct results
>>
>> Case 2 screws us. We have to take the lock to even look at the parent's
>> dentry->d_sb->s_inotify_watches. I don't know of a way around this one.
> 
> 
> Yeah.  There are a lot of "screw"s in this thread.
> 
> I wonder if RCU can save us - if we do an rcu_read_lock() we at least know
> that the dentries won't get deallocated.  Then we can take a look at
> d_parent (which might not be the parent any more).  Once in a million years
> we might send a false event or miss sending an event, depending on where
> our dentry suddenly got moved to.  Not very nice, but at least it won't
> oops.
> 
> (hopefully cc's Dipankar)

I saw this problem when testing my lockless pagecache a while back.

Attached is a first implementation of what was my idea then of how
to solve it... note it is pretty rough and I never got around to doing
much testing of it.

Basically: moves work out of inotify event time and to inotify attach
/detach time while staying out of the core VFS.

-- 
SUSE Labs, Novell Inc.