redesigning authentication/association flows

redesigning authentication/association flows

[Johannes Berg]

Hi,

I've been growing unhappy with the way I designed authentication and
association in cfg80211 and I'm ready to admit that the way it works now
was a design mistake. Just see the tx_sync as an example, and if we
explicitly do remain-on-channel for the authentication or association
then we'll have to start doing GO NoA sync within r-o-c which will be a
nightmare.

So I think we should just revisit all this. Here's what I currently
think I will do:

== cfg80211
 - get rid of the entire auth_bss/auth_try_bss, etc. Allow
   authentication and association at any time (even association without
   authentication, AP will reject anyway)
 - keep track only of current_bss with authenticated/associated flag to
   list it in scan results
 - entirely handle local state change for authentication in cfg80211
   since drivers can no longer expect auth/assoc sequencing

(note that in current code all of this only impacts mac80211 since no
other driver supports authenticate/associate commands)

== mac80211
When asked to auth/assoc change sequence to do:
 - if associated to BSSID != new BSSID:
   - send null data PM=1 to AP (?)
 - if authenticated/associated to BSSID != new BSSID:
   - bss_info_changed: zero bssid, not associated, zero aid etc
   - remove AP station entry
 - if not already configured as such:
   - clear IDLE for interface
   - set channel (in the future: for interface)
   - bss_info_changed: set correct bssid, not associated, ...
   - add AP station entry (even to driver, not dummy as today)
 - send frame
 - check response, if failure clear configuration


Note that this gets rid of the stupid tx_sync, the attempt at doing this
off-channel, and we never have to deal with tx_sync in remain-on-channel
which would be really really strange.

The only difference really is that today, if auth fails, we can stay
connected. HOWEVER, wpa_supplicant isn't actually able to deal with
that, gets confused and recovers with clear_state_mismatch().


I think this will help all drivers, from at76c50x which is broken
completely today, wl1251 which has a dummy join, wl12xx which has a work
around with remain-on-channel but this doesn't work well with a GO (NoA)
and of course iwlwifi (which has tx_sync for this) and iwlegacy which
has a bunch of RXON ping-pong things going on.

I don't think ath5k, ath9k and most other drivers will really be
impacted, the ones I'm not sure about are Ralink and Realtek drivers.

Anyone have major issues with this?

johannes

Re: redesigning authentication/association flows

[Luciano Coelho]

On Wed, 2011-11-16 at 12:13 +0100, Johannes Berg wrote: 
> I've been growing unhappy with the way I designed authentication and
> association in cfg80211 and I'm ready to admit that the way it works now
> was a design mistake. Just see the tx_sync as an example, and if we
> explicitly do remain-on-channel for the authentication or association
> then we'll have to start doing GO NoA sync within r-o-c which will be a
> nightmare.

[...]

> Note that this gets rid of the stupid tx_sync, the attempt at doing this
> off-channel, and we never have to deal with tx_sync in remain-on-channel
> which would be really really strange.
> 
> The only difference really is that today, if auth fails, we can stay
> connected. HOWEVER, wpa_supplicant isn't actually able to deal with
> that, gets confused and recovers with clear_state_mismatch().
> 
> 
> I think this will help all drivers, from at76c50x which is broken
> completely today, wl1251 which has a dummy join, wl12xx which has a work
> around with remain-on-channel but this doesn't work well with a GO (NoA)
> and of course iwlwifi (which has tx_sync for this) and iwlegacy which
> has a bunch of RXON ping-pong things going on.

As discussed on the phone, I think this should be okay with wl12xx.  We
will have to study it a bit further to be sure, but my first impression
is that it's should be fine.

One small correction from my mistake, wl1251 doesn't have a dummy join,
it's the old wl12xx which did (with older firmwares that didn't support
roc/croc).  Sorry, I got confused.

-- 
Cheers,
Luca.