TX A-mpdu (without the 'A')

[Bob Copeland]

Hi all,

I'm picking up where Chun-Yeow left off, looking at TX ampdu for
wcn36xx.  Has anyone had success with the latest driver?  Here's
what I see:

 - I'm testing with wcn36xx as a STA and ath9k as an AP with hostapd,
   running TCP iperf from the station to the AP, and capturing
   over the air with a third device.

 - after RX BA session established, wcn36xx will start the TX
   BA session, and the action frame exchange looks correct
   (though see below with regard to start seq num)

 - after a while wcn36xx will start sending a-mpdus, but there is
   only ever one mpdu per a-mpdu.  That is, there will be only
   one frame, then a BA, then another frame, another BA, etc.
   BA bitmap is all-ones and ssn is last_frame-63, and the A-mpdu
   reference numbers in radiotap change for each frame.

Conversely, with prima, I see aggregates of 12 or so frames in the
same setup.

What I have tried so far:

 - putting IEEE80211_TX_CTL_AMPDU frames on a queue and sending them
   from a workqueue later in order to send more to the hardware at
   once (printk shows I can queue 50+ frames this way)

 - futzing with only setting end-of-packet on the last frame in an
   aggregate; although I decided after more capturing and reading the
   code that prima and unmodified wcn36xx behave the same in this regard
   (one small difference: the valid bit on the BD DXE descriptor is only
   assigned after the SKB one, but I doubt this makes a difference).

 - on Pontus' suggestion, adding the following settings on bss_info_changed
   (with various values, I still need to go back to prima to see what values
   it uses).

+    wcn36xx_smd_update_cfg(wcn, WCN36XX_HAL_CFG_MAX_MPDUS_IN_AMPDU, 64);
+    wcn36xx_smd_update_cfg(wcn, WCN36XX_HAL_CFG_BA_THRESHOLD_HIGH, 10);
+    wcn36xx_smd_update_cfg(wcn, WCN36XX_HAL_CFG_MAX_BA_SESSIONS, 2);

 - fixing Start Sequence Number in wcn36xx -- I noticed when parsing out
   the FW response from add-ba-session that SSN didn't match what mac80211
   had sent.  ssn will be null by the time the TX ampdu is operational,
   so we would pass '0' to the hardware when mac80211 is really starting
   with, say, '2'.  This ugly hack got them in sync for at least the one
   stream test but changed nothing in the operation:

@@ -894,6 +899,7 @@ static int wcn36xx_ampdu_action(struct ieee80211_hw *hw,
 {
        struct wcn36xx *wcn = hw->priv;
        struct wcn36xx_sta *sta_priv = NULL;
+       static u16 tx_ssn = 0;

        wcn36xx_dbg(WCN36XX_DBG_MAC, "mac ampdu action action %d tid %d\n",
                    action, tid);
@@ -913,10 +919,11 @@ static int wcn36xx_ampdu_action(struct ieee80211_hw *hw,
                wcn36xx_smd_del_ba(wcn, tid, get_sta_index(vif, sta_priv));
                break;
        case IEEE80211_AMPDU_TX_START:
+               tx_ssn = *ssn;
                ieee80211_start_tx_ba_cb_irqsafe(vif, sta->addr, tid);
                break;
        case IEEE80211_AMPDU_TX_OPERATIONAL:
-               wcn36xx_smd_add_ba_session(wcn, sta, tid, ssn, 1,
+               wcn36xx_smd_add_ba_session(wcn, sta, tid, ssn ? ssn : &tx_ssn, 1,
                        get_sta_index(vif, sta_priv));
                break;
        case IEEE80211_AMPDU_TX_STOP_FLUSH:


The above tests were based on 7c0db3da "Merge pull request #134 from
 pof2/refactoring" from wcn36xx.  I'll test again with more recent changes and
do some more captures of prima for comparison purposes, but if anyone has any
other suggestions or has it working for them, I'd like to hear about it.
-- 
Bob Copeland %% www.bobcopeland.com