net80211: move the storing of ni in the mbuf further down

[bz]

Summary

We store the node (ni) in the mbuf meta data already after
ieee80211_dwds_mcast() (see comment). However ieee80211_encap()
further down can possibly allocate a new mbuf calling into
ieee80211_mbuf_adjust() so move storing the node information right
in front of the call which will rely on it to avoid any further
problems.

Sponsored by: The FreeBSD Foundation
MFC after: 3 days

[adrian]

Is that true though? like, the superg/amsdu transmit path is also queuing the packet and it needs the node reference for it to do so.

[bz]

In D49818#1135726, @adrian wrote:

Is that true though? like, the superg/amsdu transmit path is also queuing the packet and it needs the node reference for it to do so.

Sugar. And ff_transmit() has the same problem that it might lose it and then we go kaboom in ieee80211_parent_xmitpkt error case; okay; there's also an improvement to be had there. I'll do the other one first and revise this. I think the answer is to audit ieee80211_encap callers and if all do the same make sure ieee80211_encap preserves the meta information.

[adrian]

Yeah, let's both do that (some auditing/review) first.

My earlier diff on documenting the transmit paths at least document the node ref stuff, I can also go document the encap behaviour too in another diff.