xfs
[Top] [All Lists]

Re: [PATCH 08/12] superblock: introduce per-sb cache shrinker infrastruc

To: Christoph Hellwig <hch@xxxxxxxxxxxxx>
Subject: Re: [PATCH 08/12] superblock: introduce per-sb cache shrinker infrastructure
From: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
Date: Sat, 4 Jun 2011 15:24:48 +0100
Cc: Dave Chinner <david@xxxxxxxxxxxxx>, linux-fsdevel@xxxxxxxxxxxxxxx, linux-kernel@xxxxxxxxxxxxxxx, linux-mm@xxxxxxxxx, xfs@xxxxxxxxxxx
In-reply-to: <20110604141940.GW11521@xxxxxxxxxxxxxxxxxx>
References: <1306998067-27659-1-git-send-email-david@xxxxxxxxxxxxx> <1306998067-27659-9-git-send-email-david@xxxxxxxxxxxxx> <20110604004231.GV11521@xxxxxxxxxxxxxxxxxx> <20110604015212.GD561@dastard> <20110604140848.GA20404@xxxxxxxxxxxxx> <20110604141940.GW11521@xxxxxxxxxxxxxxxxxx>
Sender: Al Viro <viro@xxxxxxxxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Sat, Jun 04, 2011 at 03:19:40PM +0100, Al Viro wrote:
> > The iprune_sem removal is fine as soon as you have a per-sb shrinker
> > for the inodes which keeps an active reference on the superblock until
> > all the inodes are evicted.
> 
> I really don't like that.  Stuff keeping active refs, worse yet doing that
> asynchronously...  Shrinkers should *not* do that.  Just grab a passive
> ref (i.e. bump s_count), try grab s_umount (shared) and if that thing still
> has ->s_root while we hold s_umount, go ahead.  Unregister either at the
> end of generic_shutdown_super() or from deactivate_locked_super(), between
> the calls of ->kill_sb() and put_filesystem().

PS: shrinkers should not acquire active refs; more specifically, they should
not _drop_ active refs, lest they end up dropping the last active one and
trigger unregistering a shrinker for superblock in question.  From inside of
->shrink(), with shrinker_rwsem held by caller.  Deadlock...

<Prev in Thread] Current Thread [Next in Thread>