On Tue, Dec 12, 2023 at 05:50:55PM +0100, Paul Menzel wrote:
> Dear Michal, dear Joshua,
>
>
> Thank you for your patch.
>
> On 12/12/23 15:55, Michal Kubiak wrote:
> > From: Joshua Hay <[email protected]>
> >
> > Tell hardware to writeback completed descriptors even when interrupts
>
> Should you resend, the verb is spelled with a space: write back.
Sure, I will fix it.
>
> > are disabled. Otherwise, descriptors might not be written back until
> > the hardware can flush a full cacheline of descriptors. This can cause
> > unnecessary delays when traffic is light (or even trigger Tx queue
> > timeout).
>
> How can the problem be reproduced and the patch be verified?
>
>
> Kind regards,
>
> Paul
>
>
Hi Paul,
To be honest, I have noticed the problem during the implementation of
AF_XDP feature for IDPF driver. In my scenario, I had 2 Tx queues:
- regular LAN Tx queue
- and XDP Tx queue
added to the same q_vector attached to the same NAPI, so those 2 Tx
queues were handled in the same NAPI poll loop.
Then, when I started a huge Tx zero-copy trafic using AF_XDP (on the XDP
queue), and, at the same time, tried to xmit a few packets using the second
(non-XDP) queue (e.g. with scapy), I was getting the Tx timeout on that regular
LAN Tx queue.
That is why I decided to upstream this fix. With disabled writebacks,
there is no chance to get the completion descriptor for the queue where
the traffic is much lighter.
I have never tried to reproduce the scenario described by Joshua
in his original patch ("unnecessary delays when traffic is light").
Thanks,
Michal
_______________________________________________
Intel-wired-lan mailing list
[email protected]
https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
