On 2/14/19 9:01 AM, David Miller wrote:
> From: Jose Abreu <jose.ab...@synopsys.com>
> Date: Wed, 13 Feb 2019 18:00:43 +0100
>
>> Commit 8fce33317023 introduced the concept of NAPI per-channel and
>> independent cleaning of TX path.
>>
>> This is currently breaking performance in some cases. The scenario
>> happens when all packets are being received in Queue 0 but the TX is
>> performed in Queue != 0.
>>
>> I didn't look very deep but it seems that NAPI for Queue 0 will clean
>> the RX path but as TX is in different NAPI, this last one is called at a
>> slower rate which kills performance in TX. I suspect this is due to TX
>> cleaning takes much longer than RX and because NAPI will get canceled
>> once we return with 0 budget consumed (e.g. when TX is still not done it
>> will return 0 budget).
>>
>> Fix this by looking at all TX channels in NAPI poll function.
>>
>> Signed-off-by: Jose Abreu <joab...@synopsys.com>
>> Fixes: 8fce33317023 ("net: stmmac: Rework coalesce timer and fix multi-queue
>> races")
>
> No this isn't right.
>
> The TX interrupt events for Queue != 0 should clean up the TX packets
> on those queues.
>
> Furthermore you are breaking the locality of the TX processing.
>
> I'm not applying this, sorry.
Agreed, why don't you create per-queue NAPI instances such that they are
all independent and can complete their TX completion/RX processing
entirely separately?
--
Florian
