Date: Thu, 13 Jun 2013 16:26:48 -0400
From: Jeff Layton <jlayton@redhat.com>
To: "J. Bruce Fields" <bfields@fieldses.org>
Cc: viro@zeniv.linux.org.uk, matthew@wil.cx, dhowells@redhat.com,
        sage@inktank.com, smfrench@gmail.com, swhiteho@redhat.com,
        Trond.Myklebust@netapp.com, akpm@linux-foundation.org,
        linux-kernel@vger.kernel.org, linux-afs@lists.infradead.org,
        ceph-devel@vger.kernel.org, linux-cifs@vger.kernel.org,
        samba-technical@lists.samba.org, cluster-devel@redhat.com,
        linux-nfs@vger.kernel.org, linux-fsdevel@vger.kernel.org,
        piastryyy@gmail.com
Subject: Re: [PATCH v2 06/14] locks: don't walk inode->i_flock list in
 locks_show
Message-ID: <20130613162648.176979bc@tlielax.poochiereds.net>
In-Reply-To: <20130613194545.GC19218@fieldses.org>
References: <1370948948-31784-1-git-send-email-jlayton@redhat.com>
	<1370948948-31784-7-git-send-email-jlayton@redhat.com>
	<20130613194545.GC19218@fieldses.org>
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Sender: linux-nfs-owner@vger.kernel.org

On Thu, 13 Jun 2013 15:45:46 -0400
"J. Bruce Fields" <bfields@fieldses.org> wrote:

> On Tue, Jun 11, 2013 at 07:09:00AM -0400, Jeff Layton wrote:
> > When we convert over to using the i_lock to protect the i_flock list,
> > that will introduce a potential lock inversion problem in locks_show.
> > When we want to walk the i_flock list, we'll need to take the i_lock.
> > 
> > Rather than do that, just walk the global blocked_locks list and print
> > out any that are blocked on the given lock.
> 
> I'm OK with this as obviously /proc/locks shouldn't be the common case,
> but it still bugs me a bit that we're suddenly making it something like
> 
> 	O(number of held locks * number of waiters)
> 
> where it used to be
> 
> 	O(number of held lock + number of waiters)
> 
> I wonder if there's any solution that's just as easy and avoids scanning
> the blocked list each time.
> 
> --b.
> 
> > 
> > Signed-off-by: Jeff Layton <jlayton@redhat.com>
> > ---
> >  fs/locks.c |    6 ++++--
> >  1 files changed, 4 insertions(+), 2 deletions(-)
> > 
> > diff --git a/fs/locks.c b/fs/locks.c
> > index e451d18..3fd27f0 100644
> > --- a/fs/locks.c
> > +++ b/fs/locks.c
> > @@ -2249,8 +2249,10 @@ static int locks_show(struct seq_file *f, void *v)
> >  
> >  	lock_get_status(f, fl, *((loff_t *)f->private), "");
> >  
> > -	list_for_each_entry(bfl, &fl->fl_block, fl_block)
> > -		lock_get_status(f, bfl, *((loff_t *)f->private), " ->");
> > +	list_for_each_entry(bfl, &blocked_list, fl_link) {
> > +		if (bfl->fl_next == fl)
> > +			lock_get_status(f, bfl, *((loff_t *)f->private), " ->");
> > +	}
> >  
> >  	return 0;
> >  }
> > -- 
> > 1.7.1
> > 

Yeah, it's ugly, but I don't see a real alternative. We could try to
use RCU for this, but that slows down the list manipulation at the cost
of optimizing a rarely read procfile.

Now that I look though...it occurs to me that we have a problem here
anyway. Only blocked POSIX requests go onto that list currently, so
this misses seeing any blocked flock requests.

The only real solution I can think of is to put flock locks into the
blocked_list/blocked_hash too, or maybe giving them a simple hlist to
sit on.

I'll fix that up in the next iteration. It'll probably make flock()
tests run slower, but such is the cost of preserving this procfile...

-- 
Jeff Layton <jlayton@redhat.com>