Some drivers/devices would like to

• use a separate MAC address
• use a separate control path for P2P usage. This could even help mac80211-based drivers like iwlagn since currently, iwlagn needs to enable P2P in the device when a remain-on-channel is done, and disable it after a timeout or when a P2P interface is used.

A separate netdev would be the most obvious choice, but can be confusing:

• to the user – new interface is there, what does it do?
• to the developer – no data traffic on this interface

Better: use dedicated API in nl80211:

• start-P2P → returns cookie
• stop-P2P → uses cookie

(or maybe don't have “stop-P2P” but simply stop when socket is closed like mgmt frame subscriptions)

The only issue with this is that things like scan, mgmt-tx etc. need a netdev index now. However, this can be changed, idea:

• use cookie to identify the P2P device interface
• internally, create a

  struct wireless_dev

  but without a netdev

• modify cfg80211 API (e.g.

  scan

  ,

  remain_on_channel

  ) to take

  struct wireless_dev

  instead of netdev, driver can check what the type is etc.

• this needs separate P2P-device iftype that can't really be used as an iftype, which is fine Questions:
  • lifetime: does the P2P-device interface become the P2P-group/client interface like in wpa_supplicant, which means that it is removed before/when the real netdev is added? (personally I prefer it would stay around I think since I think discovery/public action things would still be done with it, not the real interface – Johannes)
  • pure software implementation of this in mac80211 for drivers that don't care, to unify API? but wpa_s needs old code anyway for backward compatibility

This could also be a good framework for additional features that we'll need to add:

• device-based P2P listen/search timing (soon)
• maybe some more P2P offloads (WoP2P anyone? )