On Tue, Nov 29, 2011 at 10:31, Johannes Berg [off-list ref] wrote:

I noticed that the logs are a bit odd wrt. timing.

quoted

quoted

Intersperesed I see some other messages that are new to me:
[ 4019.443129] Open BA session requested for 00:0a:79:eb:56:10 tid 0
[ 4019.500149] activated addBA response timer on tid 0
[ 4020.500033] addBA response timer expired on tid 0

I guess the delay here is due to the synchronize_net()? That can take a
while, 57ms seems a lot but I suppose it's possible.

quoted

quoted

[ 4020.501626] Tx BA session stop requested for 00:0a:79:eb:56:10 tid 0
[ 4023.740570] switched off addBA timer for tid 0
[ 4023.740578] got addBA resp for tid 0 but we already gave up

Here is the AP is finally replying

It's kinda hard to believe that the AP took 4 seconds (!) to reply to
the frame. Where could the frame get stuck? I don't see any other work
processing happening etc. either. It's also curious that in those 3
seconds between these messages, we didn't actually get around to
stopping the session, that only happens just after:

Yeah you are right, didn't look at the timestamps. Not sure you would
see work being processed though.

quoted

quoted

[ 4023.740619] Stopping Tx BA session for 00:0a:79:eb:56:10 tid 0

(here)

quoted

quoted

[ 4023.768544] Open BA session requested for 00:0a:79:eb:56:10 tid 0

Here we are trying again

quoted

[ 4023.784292] activated addBA response timer on tid 0
[ 4023.786294] switched off addBA timer for tid 0

20ms response time here, that's much more reasonable.


Could something be hogging the workqueues?

Frankly, I am seeing issues that seem to point to workqueues too.
Sometimes mac80211 seems just not responsive.
Sometimes I come back to mac80211 for the AGG callback (start or
stop), and it takes ages (5 seconds !) until it actually move to
operationnal / stopped state.

It might that we are holding the mac80211 workqueue in the driver too...
I guess we could try to enable MAC80211 debug flag with timestamps to check.

johannes