From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <linux-btrfs-owner@vger.kernel.org>
Received: from mx0a-00082601.pphosted.com ([67.231.145.42]:30177 "EHLO
	mx0a-00082601.pphosted.com" rhost-flags-OK-OK-OK-OK)
	by vger.kernel.org with ESMTP id S932213AbaFSQqT (ORCPT
	<rfc822;linux-btrfs@vger.kernel.org>);
	Thu, 19 Jun 2014 12:46:19 -0400
Message-ID: <53A31514.8030308@fb.com>
Date: Thu, 19 Jun 2014 12:51:32 -0400
From: Chris Mason <clm@fb.com>
MIME-Version: 1.0
To: Waiman Long <waiman.long@hp.com>
CC: Marc Dionne <marc.c.dionne@gmail.com>, Josef Bacik <jbacik@fb.com>,
        <linux-btrfs@vger.kernel.org>, <t-itoh@jp.fujitsu.com>
Subject: Re: Lockups with btrfs on 3.16-rc1 - bisected
References: <CAB9dFdsyB40V2BGwdQKzSSz7YUWMh-C2AWdDNK=PMG9h+BO57w@mail.gmail.com>	<53A20FFF.3010807@hp.com>	<53A2125B.3050701@fb.com>	<53A21702.8090109@hp.com>	<53A21C78.1040809@fb.com>	<53A21E84.2050103@hp.com>	<53A22064.7080400@fb.com>	<53A2212E.7090907@hp.com>	<53A2268F.3080807@fb.com>	<CAB9dFds5J8Ti3CgtxHrgqM3j1TiwqYCnhP+A3Nb0q-FNgmQbXA@mail.gmail.com>	<53A22A01.7080505@hp.com>	<CAB9dFdvD9QMZ_rONhhP5ZD2E7+A7qxnsY978Snc3zHEbdPRszw@mail.gmail.com> <CAB9dFdtTFN+uQ8=WXVa0Y5GbBSYM5iVLgTaEeh_FqHt=XFKNvQ@mail.gmail.com> <53A246C6.5050408@fb.com> <53A2573B.1060901@hp.com>
In-Reply-To: <53A2573B.1060901@hp.com>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-btrfs-owner@vger.kernel.org
List-ID: <linux-btrfs.vger.kernel.org>

On 06/18/2014 11:21 PM, Waiman Long wrote:
> On 06/18/2014 10:11 PM, Chris Mason wrote:
>> On 06/18/2014 10:03 PM, Marc Dionne wrote:
>>> On Wed, Jun 18, 2014 at 8:41 PM, Marc
>>> Dionne<marc.c.dionne@gmail.com>  wrote:
>>>> On Wed, Jun 18, 2014 at 8:08 PM, Waiman Long<waiman.long@hp.com> 
>>>> wrote:
>>>>> On 06/18/2014 08:03 PM, Marc Dionne wrote:
>>> And for an additional data point, just removing those
>>> CONFIG_DEBUG_LOCK_ALLOC ifdefs looks like it's sufficient to prevent
>>> the symptoms when lockdep is not enabled.
>> Ok, somehow we've added a lock inversion here that wasn't here before.
>> Thanks for confirming, I'll nail it down.
>>
>> -chris
>>
> 
> I am pretty sure that the hangup is caused by the following kind of code
> fragment in the locking.c file:
> 
>  if (eb->lock_nested) {
>                 read_lock(&eb->lock);
>                 if (eb->lock_nested && current->pid == eb->lock_owner) {
> 
> Is it possible to do the check without taking the read_lock?

I think you're right, we haven't added any new recursive takers of the
lock.  The path where we are deadlocking has an extent buffer that isn't
in the path yet locked.  I think we're taking the read lock while that
one is write locked.

Reworking the nesting a big here.

-chris