On 1 September 2016 at 22:52, Ben Greear <greearb@xxxxxxxxxxxxxxx> wrote: > On 09/01/2016 11:53 AM, Johannes Berg wrote: >> On Thu, 2016-09-01 at 11:23 -0700, Ben Greear wrote: >>> >>> Could easily be that others are corrupted too, but since probe resp >>> is bad, the association will not proceed. >> >> makes sense. >> >>> Heh, I spent 4 days tracking this down, so I wanted to be precise in >>> my bug report :) >> >> Ahrg, ouch. Sorry about that. I really didn't think the flag would >> cause any issues for anyone. >> >>> The result I see is that there is an extra 10 bytes at the end of the >>> frame on air. But, it looks like the exact same pkt is sent to the >>> firmware both with and without this patch. Maybe the firmware is >>> using the wrong tid or something like that due to how the station is >>> created differently with this patch. >> >> That makes no sense though, unless this only happens on say the second >> station that connects? Until the first station sends an authentication >> frame, that patch really should have no impact whatsoever. > > Ok, I found the problem. > > In the 10.1 firmware (at least), it will force the TID to be NON-QOS-TID > if the peer object does not have the qos_enabled flag set. This is probably > a work-around for some other thing lost in antiquity. > > When using my firmware that puts mgt frames over HTT, the TID for mgt > frames was being over-ridden and set to non-qos TID. Due to other > hackery and work-arounds, mgt frames cannot be sent on the non-qos > TID because they will be put on-air 10 bytes too long. They must be > sent on the mgt-tid or non-pause tid. Sounds like 802.11 header vs 802.3 header length problem (24 - 14 = 10). You can hit this if you start playing around tx encap mode as well. You could try fooling firmware into thinking the frame has a different length either in htt TX_FRM or tx fragment list (or both) but since this seems to be related to RA/DA peer state at xmit time it's probably not going to be reliable unless you introduce extra tx flushing barriers in the driver. Michał
