Question on ath10k_mgmt_tx_flush

[Michal Kazior]

On 14 April 2014 17:45, Ben Greear <greearb at candelatech.com> wrote:
> On 04/13/2014 11:31 PM, Michal Kazior wrote:
>> On 11 April 2014 15:25, Ben Greear <greearb at candelatech.com> wrote:
>>> On 04/10/2014 10:21 PM, Michal Kazior wrote:
>>>>
>>>> On 10 April 2014 15:41, Ben Greear <greearb at candelatech.com> wrote:
>>>>>
>>>>> Can we optimize this method to return early if the tx-credits
>>>>> are fully replenished (ie, == 2) instead of just sleeping the
>>>>> 2 x beacon-interval?  That would indicate all messages
>>>>> have been flushed, right?
>>>>
>>>>
>>>> Yeah. You're _almost_ right. Every odd mgmt frame will trigger tx
>>>> credit replenishment, even if you set NEEDS_CREDITS htc tx flag for
>>>> all packets. It seems that tx credits aren't replenished until you
>>>> submit an even number of mgmt tx:
>>>>
>>>>   [tx credits =2]
>>>>   vdev create [-1, =1]
>>>>   [replenish +1, =2]
>>>>   mgmt tx [-1, =1]
>>>>   [frame is seen on air, means it left tx queue, but no replenishment]
>>>>   vdev set param [-1, =0]
>>>>   [replenish +1, =1]
>>>>   mgmt tx [-1, =0]
>>>>   [frame seen on air]
>>>>   [replenish +2, =2]
>>>>
>>>> However once you flush peer tids you get the tx credit immediately.
>>>> This means you don't ever reach having 2 mgmt tx consuming 2 tx
>>>> credits (unless things go terribly terribly wrong at which point it's
>>>> probably already beyond help).
>>>>
>>>> A very ugly hack would be to try and send out mgmt tx in pairs - a
>>>> requested frame and a dummy frame (such that firmware will not buffer
>>>> it) so that you use tx credit replenishment as tx completion
>>>> indication.
>>>
>>>
>>> If you ignore how firmware replenishes the tx-credits for now,
>>> is it safe to assume that if you have 2 tx-credits while in
>>> the flush routine, then everything is indeed flushed and
>>> we can skip the sleep?
>>
>> You want to always skip the sleep between mgmt tx and flush commands?
>> I'm afraid this won't work because tx flush command can end up with
>> queued frame being dropped or transmitted regardless of destination
>> station powersave state if submitted too soon.
>
> I just want to know if I can skip the sleep if we currently have 2 tx-credits.

No, not really. At least not in any sane way I can think of.

Problem 1:
If you submit a frame that gets stuck for 10 seconds and you don't
flush it (and you have no idea when it is safe to flush other than
timing this with arbitrary numbers like beacon intervals) and you try
to queue another frame that gets stuck you end up with 0 tx credits.

Problem 2:
If you submit a frame and immediately call tx flush you can end up not
sending out the frame at all or ignoring client power save state I'm
afraid.

I was thinking of a more clever hack that would involve sending dummy
frames (i.e. spurious probe responses) to tickle firmware and force it
replenish tx credit after sending the actual requested mgmt frame.
This is ugly, intrusive, wrong and I'm not even sure if this can even
work.


> From your answer below, that seems to be a yes.  So, 1/2 of the time (on avg)
> is worth the hack to me, and if/when I fix the firmware (and/or if QCA fixes
> the firmware), then we should have 2 tx-credits more often and only sleep when
> there really is work to do.

The problem is you don't have a way of knowing if mgmt tx is stuck or
not both because there is no tx completion and tx credits aren't
replenished in a sane way. If any of those were ok, we could have a
wait_for_completion() instead of a msleep() and thus cut down wait
times for frames that don't get stuck.


>>> I imagine both I and the QCA firmware guys can make the
>>> firmware properly replenish tx-credits one at a time instead
>>> of the only-on-even/odd behaviour you found.
>>
>> That would be nice.
>>
>>
>>> But even without that, we might be able to speed up the
>>> flush 1/2 of the time?
>>
>> Yes, although I'm not very fond of the idea of applying a hack on top
>> of another hack.
>
> What is your preferred fix for this?

Fix the firmware, obviously.

999.999.0.636 uses htt for mgmt tx so the problem doesn't exist there.
That's solely a 10.1 branch issue and I doubt this will be ever fixed
in this branch...


Michał