K 10
svn:author
V 6
adrian
K 8
svn:date
V 27
2011-09-20T04:20:55.598997Z
K 7
svn:log
V 1879
Undo the BAR TX stuff (as it's broken); correct what the new LH edge
of the BAW should be.

The BAR TX stuff I committed earlier is just plain wrong.
What should happen is this (in ADDBA mode):

* If a software retransmit completely fails, all TX for that
  given TID should be paused until all existing TXes for said
  TID are completed (failed or not);
* Once all the hardware-queued transmissions are completed,
  the left-hand edge of the BAW (from the TX point of view)
  should be correct. Ie, there may be further packets queued
  in the software TX queue, but the left-hand edge of the BAW
  will point to one frame past the last hardware-queued frame
  (again, successful or not, it doesn't matter.)
* At that point, a BAR should be sent reflecting what the new
  BAW left hand edge is.
* Once that's been received, TX on the given TID should resume.
* If it fails, net80211 will cancel the addba session for us; we
  just need to resume TX'ing on the TID anyway (newly downgraded
  to non-aggregation.)

Now, the things to keep in mind when implementing the correct-er
solution:

* There may be plenty of outstanding frames on the hardware TX
  queue, so you can't just send one BAR for each of those.

* Whilst the BAR is being TXed, the node may be reset; this
  SHOULD correctly be handled (ie, cancelling the frame will
  cause net80211 to reschedule another one)

* You can't just call pause() and resume() like I was doing,
  as multiple outstanding frames may be on the hardware queue
  (and multiples ones may fail.) A new method of marking
  a TID as "BAR pending" needs to be added first.
  That way the TID can be paused and unpaused a total of once,
  no matter how many outstanding frames fail before the TID
  has no more frames on the hardware queue.

* Finally, this all only works if tid->hwq_depth is completely
  accurate. This needs to be verified!

END