Re: [PATCH 3/4] mac80211: document requirement for atomicity of callbacks

[Michael Buesch <mb@bu3sch.de>]

On Sunday 29 April 2007 12:49:29 Jiri Benc wrote:
> On Sun, 29 Apr 2007 11:15:43 +0200, Michael Buesch wrote:
> > On Sunday 29 April 2007 02:33:45 Jiri Benc wrote:
> > > +	 * Must be atomic. */
> > >  	int (*set_key)(struct ieee80211_hw *hw, set_key_cmd cmd,
> > >  		       u8 *addr, struct ieee80211_key_conf *key, int aid);
> > 
> > I think USB drivers and bcm43xx violate that. We can't easily workaround it
> > without doing the operation async.
> 
> I know. Unfortunately, it's not easy to fix that in mac80211. I really want
> this call not to be atomic and will try to change it later, but it's not a
> priority now.
> 
> For now, just always return success and complain to dmesg if something goes
> wrong during setting of the key in a workqueue.

Ok, I see. But I want to note that this might be more problematic
then just a lost error code. It might race.
For example it might race with the EAPOL KEY authentication stuff.
So I think we must also workaround that by disabling TX queues
until the key configuration request has actually happened
on the device.

-- 
Greetings Michael.