This patchset implements a new TX lock, covering both the per-VAP (and
thus per-node) TX locking and the serialisation through to the underlying
physical device.
This implements the hard requirement that frames to the underlying physical
device are scheduled to the underlying device in the same order that they
are processed at the VAP layer. This includes adding extra encapsulation
state (such as sequence numbers and CCMP IV numbers.) Any order mismatch
here will result in dropped packets at the receiver.
There are multiple transmit contexts from the upper protocol layers as well
as the "raw" interface via the management and BPF transmit paths.
All of these need to be correctly serialised or bad behaviour will result
under load.
The specifics:
* add a new TX IC lock - it will eventually just be used for serialisation
to the underlying physical device but for now it's used for both the
VAP encapsulation/serialisation and the physical device dispatch.
This lock is specifically non-recursive.
* Methodize the parent transmit, vap transmit and ic_raw_xmit function
pointers; use lock assertions in the parent/vap transmit routines.
* Add a lock assertion in ieee80211_encap() - the TX lock must be held
here to guarantee sensible behaviour.
* Refactor out the packet sending code from ieee80211_start() - now
ieee80211_start() is just a loop over the ifnet queue and it dispatches
each VAP packet send through ieee80211_start_pkt().
Yes, I will likely rename ieee80211_start_pkt() to something that
better reflects its status as a VAP packet transmit path. More on
that later.
* Add locking around the management and BAR TX sending - to ensure that
encapsulation and TX are done hand-in-hand.
* Add locking in the mesh code - again, to ensure that encapsulation
and mesh transmit are done hand-in-hand.
* Add locking around the power save queue and ageq handling, when
dispatching to the parent interface.
* Add locking around the WDS handoff.
* Add a note in the mesh dispatch code that the TX path needs to be
re-thought-out - right now it's doing a direct parent device transmit
rather than going via the vap layer. It may "work", but it's likely
incorrect (as it bypasses any possible per-node power save and
aggregation handling.)
Why not a per-VAP or per-node lock?
Because in order to ensure per-VAP ordering, we'd have to hold the
VAP lock across parent->if_transmit(). There are a few problems
with this:
* There's some state being setup during each driver transmit - specifically,
the encryption encap / CCMP IV setup. That should eventually be dragged
back into the encapsulation phase but for now it lives in the driver TX path.
This should be locked.
* Two drivers (ath, iwn) re-use the node->ni_txseqs array in order to
allocate sequence numbers when doing transmit aggregation. This should
also be locked.
* Drivers may have multiple frames queued already - so when one calls
if_transmit(), it may end up dispatching multiple frames for different
VAPs/nodes, each needing a different lock when handling that particular
end destination.
So to be "correct" locking-wise, we'd end up needing to grab a VAP or
node lock inside the driver TX path when setting up crypto / AMPDU sequence
numbers, and we may already _have_ a TX lock held - mostly for the same
destination vap/node, but sometimes it'll be for others. That could lead
to LORs and thus deadlocks.
So for now, I'm sticking with an IC TX lock. It has the advantage of
papering over the above and it also has the added advantage that I can
assert that it's being held when doing a parent device transmit.
I'll look at splitting the locks out a bit more later on.
General outstanding net80211 TX path issues / TODO:
* Look into separating out the VAP serialisation and the IC handoff.
It's going to be tricky as parent->if_transmit() doesn't give me the
opportunity to split queuing from driver dispatch. See above.
* Work with monthadar to fix up the mesh transmit path so it doesn't go via
the parent interface when retransmitting frames.
* Push the encryption handling back into the driver, if it's at all
architectually sane to do so. I know it's possible - it's what mac80211
in Linux does.
* Make ieee80211_raw_xmit() queue a frame into VAP or parent queue rather
than doing a short-cut direct into the driver. There are QoS issues
here - you do want your management frames to be encapsulated and pushed
onto the stack sooner than the (large, bursty) amount of data frames
that are queued. But there has to be a saner way to do this.
* Fragments are still broken - drivers need to be upgraded to an if_transmit()
implementation and then fragmentation handling needs to be properly fixed.
Tested:
* STA - AR5416, AR9280, Intel 5300 abgn wifi
* Hostap - AR5416, AR9160, AR9280
* Mesh - some testing by monthadar@, more to come.
291b1e2469