[PATCH] Fix NetworkManager/wpa_supplicant race condition

[Tim Gardner]

Dan Williams wrote:
> On Wed, 2007-04-18 at 10:40 -0600, Tim Gardner wrote:
>> This patch fixes an assumption made by wpa_supplicant. Any time
>> wpa_supplicant requests to set an ESSID (e.g., associate), it expects an
>> event notifying that association has completed. If the Networkmanager
>> has already setup an association, such as for an open auth AP, then the
>> request to associate by wpa_supplicant will be ignored.
>>
>> If Networkmanager is requested to restart the connection, such as by
>> clicking on the SSID, then wpa_supplicant is allowed to build the
>> association from scratch, which always works.
>>
>> https://bugs.launchpad.net/ubuntu/+source/linux-source-2.6.20/+bug/103768
>>
>> By always emitting this event, am I causing any unintended side effects?
> 
> Interesting.  Setting the SSID should _always_ trigger a reassociation,
> and therefore eventually trigger an SIOCGIWAP event to userspace.  So
> this patch looks right from the behavioral point of view.
> 
> But it doesn't look right from the technical perspective.  Why isn't
> softmac trying to reassociate?  Does it automatically reassociate if
> parameters, like auth mode, keys, WPA, etc are set and so therefore,
> when it comes around to the SSID being set it doesn't really matter?  If
> so, we should still be sending the event, and this patch is OK.  But can
> somebody _guarantee_ that if I authenticate to an AP, then later call
> SIOCSIWENCODE, and then SIOCSIWESSID to the same SSID, that the new WEP
> keys have been applied and a reassociation has occurred?  If the auth
> mode is shared key, a reassociation attempt needs to happen.
> 
> WEXT convention is that setting the SSID or BSSID triggers reassociation
> with the current parameters.  I'd argue that softmac should be starting
> the association process over again when either SIOCSIWESSID or
> SIOCSIWBSSID is called.
> 
> Dan
> 
>> rtg
>> plain text document attachment (bug_103768)
>> diff --git a/net/ieee80211/softmac/ieee80211softmac_wx.c b/net/ieee80211/softmac/ieee80211softmac_wx.c
>> index fa2f7da..cc2e8ba 100644
>> --- a/net/ieee80211/softmac/ieee80211softmac_wx.c
>> +++ b/net/ieee80211/softmac/ieee80211softmac_wx.c
>> @@ -88,6 +88,13 @@ ieee80211softmac_wx_set_essid(struct net_device *net_dev,
>>  		   !memcmp(n->essid.data, extra, n->essid.len)) {
>>  			dprintk(KERN_INFO PFX "Already associating or associated to "MAC_FMT"\n",
>>  				MAC_ARG(sm->associnfo.bssid));
>> +			/* wpa_supplicant expects an association event, regardless of prior
>> +			 * association state. If associating, then the associnfo.work task
>> +			 * will send the appropriate event.
>> +			 */
>> +			if (sm->associnfo.associated)
>> +				ieee80211softmac_call_events_locked(sm,
>> +					IEEE80211SOFTMAC_EVENT_ASSOCIATED, n);
>>  			goto out;
>>  		} else {
>>  			dprintk(KERN_INFO PFX "Canceling existing associate request!\n");
> 

I don't think simply reassociating is sufficient. Consider what happens
if the authentication algorithm is changed. ieee80211_wx_set_auth() does
nothing more then set some values. A subsequent attempt to reassociate
would get all hosed up.

A quick look at ieee80211_ioctl_siwessid() shows that it always forces
an authentication cycle.

I'll experiment a little to see how the Softmac can be driven through an
authentication/association cycle on calls to SIOCSIWESSID or SIOCSIWAP.

rtg
-- 
Tim Gardner tim.gardner at ubuntu.com