On Mon, Oct 4, 2010 at 11:21 AM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> On Mon, Oct 04, 2010 at 11:01:19AM +0200, Tejun Heo wrote:
>> On 09/25/2010 03:08 PM, Torsten Kaiser wrote:
>> > [adding Tejun Heo, because I'm partly blaming his workqueues]
>> :-) Sorry about the delay. I was away the last week.
>> > After that I enabled lockdep and retried the same update: This time no
>> > hang, only the same lockdep report that Yang Ruirui had.
>> > So I'm currently at a loss how I should continue from here. The XFS
>> > false positive drowns any other lockdep problems, but XFS had some
>> > hang problems with the new workqueues. (Personally I had no other hang
>> > with earlier 2.6.36-rcs)
> What XFS hang problem? I've seen temporary livelocks from the new
> workqueue code, but thay have been fixed. I haven't come across any
> deadlocks or hangs. Do you have a pointer to a bug report in case my
> memory is failing me?
I was referring to the hang that prompted to increase of the priority
of the xfslogd thread. Its commit
51749e47e191db8e588ad5cebea731caf7b705d7 in Linus' tree. But it seems
I have misremembered that one as a hang.
The hang I have seen is describe here:
As 2.6.36-rc5 already includes this priority fix, I believe this hang
is something different and wrote that mail.
>> If workqueue lockdep annotations are triggering spuriously, just nuke
>> all lockdep annotations in workqueue.c. Currently the annotations are
>> the same as before and trigger even when an actual deadlock won't
>> happen anymore around flushes (but then again all the current users
>> should be operating under the same constraints as before). I'm
>> looking into how to improve the lockdep annotations but in the end it
>> might be best to leave it as it is. I'll try to analyze the warnings.
> It looks like there's been some discussion that has been private so
> I'm clueless as to the context being discussed here. Can someone
> fill me in on whatever is going on?
I only send the above message both to the lkml and the XFS mailing list.
But it could have been wrong to make it a reply in this thread,
because now there are three different things mixed up.
1) the old livelock, that I was referring too -> That is already fixed
in mainline, I just wanted to note, that this is unrelated to the hang
I have seen.
2) the hang with 2.6.35-rc5 I have seen twice (triggered probably by
high load while building KOffice on an tmpfs)
3) the lockdep issue that started this thread and that I am seeing in
mainline 2.6.36-rc5 and -rc6 and that seems to be rather easy to
trigger for me.
Because 3) is regarded as a false positive, it should not be the cause
of 2). But because 3) triggers on my system I can't see if lockdep
might give better pointers to what might be the real cause of 2).
So, I would like to test the patch Christoph was testing in
http://marc.info/?l=linux-kernel&m=128501006509743&w=2 to see, if that
helps to debug my hang. (And because I did not see that patch on the
lkml, I replied to Christophs mail.)
I CC'ed Tejun, because I believe the primary suspect for my hang was
the stuck kworker/u:8:845 and I was hoping he might see something
obvious wrong in that stacktrace.