Re: Linux 3.0.92

[Johannes Berg]

[too many email threads, so + a few folks]

> > This one turns out to be buggy, see thread called "3.11-rc6 genetlink
> > locking fix offends lockdep".
> 
> Yeah, I messed up in keeping it here, I'll go revert it and push out a
> new 3.0 release.

Sorry everyone, I thought I tested that code but clearly I didn't test
it well enough. I can easily reproduce the issues now so not sure why I
didn't see it before.

> I thought that there was a fix already for this...  Ah, no, that was for
> another reported regression in an older 3.10-stable release, my bad.
> 
> Johannes, what do you want to do here?  Just revert it in Linus's tree
> for now, given the late -rc cycle?

I think that'd be the best course of action for now. I just tried a few
other approaches but I can't come up with a dead-lock free way to
actually add locking here, short of providing some way for dump to
actually always have locking from "outside" (netlink code).

I think the better way would be to convert it to just use RCU. This
isn't very efficient, but then again we don't unregister generic netlink
families all the time. Below patch works without lockdep complaining,
but that's obvious since it can't check RCU ... I'm not sure I want to
do this so close to a release?

johannes