Re: [Bug #17361] Watchdog detected hard LOCKUP in jbd2_journal_get_write_access

From: Vegard Nossum <vegard.nossum@gmail.com>
To: "Ted Ts'o" <tytso@mit.edu>, "Rafael J. Wysocki" <rjw@sisk.pl>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	Kernel Testers List <kernel-testers@vger.kernel.org>,
	Maciej Rutecki <maciej.rutecki@gmail.com>,
	Florian Mickler <florian@mickler.org>,
	Christian Casteyde <casteyde.christian@free.fr>
Cc: Frederic Weisbecker <fweisbec@gmail.com>,
	Ingo Molnar <mingo@elte.hu>,
	Mathieu Desnoyers <compudj@krystal.dyndns.org>
Subject: Re: [Bug #17361] Watchdog detected hard LOCKUP in jbd2_journal_get_write_access
Date: Sun, 3 Oct 2010 13:21:03 +0200	[thread overview]
Message-ID: <AANLkTimPjz85+N8aiVV3yxz808LMG1-aZKAaO8CwUK=m@mail.gmail.com> (raw)
In-Reply-To: <20101002165215.GK21129@thunk.org>

On 2 October 2010 18:52, Ted Ts'o <tytso@mit.edu> wrote:
> On Sun, Sep 26, 2010 at 10:04:13PM +0200, Rafael J. Wysocki wrote:
>>
>> Bug-Entry     : http://bugzilla.kernel.org/show_bug.cgi?id=17361
>> Subject               : Watchdog detected hard LOCKUP in jbd2_journal_get_write_access
>> Submitter     : Christian Casteyde <casteyde.christian@free.fr>
>> Date          : 2010-08-29 19:59 (29 days old)
>
> See my latest comment here:
>
>    https://bugzilla.kernel.org/show_bug.cgi?id=17361#c14
>
> This subject line is highly misleading, since after -rc4, the stack
> traces are in places all over the kernel, in other places other than
> ext4/jbd2.  So I fear no one is looking at this bug report given the
> highly misleading subject line.
>
> It looks like you have spinlock debugging, and yet there wan't any
> spinlocks listed on the initial ext4 might_sleep() warning.  So
> something looks highly confused.
>
> The fact that you closed other bugs as duplicates of this one that
> relate to kmemcheck makes me wonder if this is really a kmemcheck bug.
> (If so, the subject line here is doubly, doubly misleading.)
>
> Do you see any symptoms if you turn off kmemcheck?  Are you sure this
> isn't just only a kmemcheck bug?

I just had a quick glance at the report, and here's my gut feeling: I
see perf symbols in the stack trace. I don't think kmemcheck and perf
play nicely together (for example if perf uses NMIs to write data to
its buffers, it could get a page fault inside the NMI handler, which
is not so nice, I think).

Isn't this exactly what Frederic Weisbecker tried to detect and warn
about in a patch that I saw recently?

Please do as Ted suggested and try to turn kmemcheck off.

Vegard