Re: cfg80211_disconnected memory leak

[Johannes Berg <johannes@sipsolutions.net>]

On Thu, 2012-08-02 at 10:11 -0600, Daniel Drake wrote:
> On Thu, Aug 2, 2012 at 2:02 AM, Johannes Berg <johannes@sipsolutions.net> wrote:
> > Oh, hm. I didn't think it could unregister before we give up our
> > reference, but I guess that makes sense after all.
> >
> > I'm not sure there's an easy way to fix it other than making the driver
> > not call cfg80211_disconnected() in case the disconnect was requested by
> > cfg80211 -- that call isn't needed and will not do anything at all, but
> > I'm not sure how easy that would be in the driver?
> 
> I guess you've considered clearing all the pending work before
> removing a netdev from the rdev's list?

Yes, but we can't just try to flush the workqueue because of locking
concerns in this area.

Hmm. Then again, I think we can call cfg80211_process_wdev_events() from
case NETDEV_UNREGISTER though, probably after removing from the list.
Maybe you could try that?

> I think a driver modification would be easy, if it is the right solution.
> 
> lbs_disconnect() is the function that calls cfg80211_disconnected().
> We only ever call this in 2 contexts:
> 
> 1. From our cfg80211_ops.disconnect handler - you say this isn't needed
> 2. From the netdev ndo_stop handler - I guess it is also not necessary
> to inform cfg80211 that we have disconnected at this point, it is kind
> of obvious..?

In fact, you won't get to ndo_stop without cfg80211 calling
disconnect(), because it does that from NETDEV_GOING_DOWN.

> So completely removing the call to cfg80211_disconnected() may be the
> right option here, is that what you recommend?

I'm not 100% sure about the API in this area right now though, it's been
a while and I never worked much with this API (rather than the mac80211
one with auth/assoc/disassoc/deauth.)

johannes