Re: [LOCKDEP BUG][2.6.36-rc1] xt_info_wrlock?

[Eric Dumazet <eric.dumazet@gmail.com>]

Le lundi 16 août 2010 à 22:01 +0200, Peter Zijlstra a écrit :
> On Mon, 2010-08-16 at 21:35 +0200, Eric Dumazet wrote:
> > Le lundi 16 août 2010 à 21:16 +0200, Peter Zijlstra a écrit :
> > 
> > > Adding lockdep_off() is just plain wrong, if you cannot describe the
> > > locking there's a fair chance its wrong anyway.
> > > 
> > 
> > I see.
> > 
> > I described the fine locking after Steven comment, adding a long
> > Changelog.
> > 
> > http://patchwork.ozlabs.org/patch/61827/
> > 
> > If someone thinks this locking is buggy, please speak now ;)
> 
> Urgh,.. I think it might be correct, but wtf! Wasn't this originally RCU
> code, why not go back to using RCU now that we have
> synchronize_rcu_expedited()?
> 
> As to the original issue, why not keep that bh stuff disabled for
> CONFIG_PROVE_LOCKING instead, that will at least let you keep lock
> coverage, adding lockdep_off() will hide any cycles that would involve
> this lock (even though its currently a leaf lock, you never know what
> creative things the future brings).
> 
> This fancy open coded lock looks like utter fail for -rt though.. please
> use RCU if at all possible.
> 
> 

Please read history of why RCU failed in this context.

I believe I did most of RCU conversions in kernel, you dont need to
shout on me.

And its a bit late in 2.6.36 to even think about it.

I am happy that you volunteer for next RCU conversion, thanks Peter !

In the mean time, we just are going to disable BH again, I'll post a
patch.