AF_XDP (i40e) behavior change in 5.11?

AF_XDP (i40e) behavior change in 5.11?

[Dan Siemon]

./xsk_fwd -i ens1f0 -q 2 -i ens1f1 -q 2 -c 5 -i ens1f0 -q 3 -i ens1f1 -
q 3 -c 6

On 5.10, the cores (2,3) assigned interrupts for the queues (2,3) are
idle when there is no traffic.

On 5.11, the cores immediately go to 100% system time when the program
starts (as viewed in htop). There is no network traffic.

Intel 710/i40e.

I tried this with xsk_fwd built from 5.10 and 5.11 trees with the same
result.

Is this behavior change expected?

Re: AF_XDP (i40e) behavior change in 5.11?

[Magnus Karlsson]

On Wed, Mar 24, 2021 at 10:22 AM Dan Siemon <dan@coverfire.com> wrote:
>
> ./xsk_fwd -i ens1f0 -q 2 -i ens1f1 -q 2 -c 5 -i ens1f0 -q 3 -i ens1f1 -
> q 3 -c 6
>
> On 5.10, the cores (2,3) assigned interrupts for the queues (2,3) are
> idle when there is no traffic.
>
> On 5.11, the cores immediately go to 100% system time when the program
> starts (as viewed in htop). There is no network traffic.
>
> Intel 710/i40e.
>
> I tried this with xsk_fwd built from 5.10 and 5.11 trees with the same
> result.
>
> Is this behavior change expected?

Thanks for reporting this. No, this is not intentional. I did send up
a fix for a bug in the i40e driver that was introduced in 5.11. It
might be the culprit. Could you please try the patch below and see if
it works? It is not yet in net.

https://www.spinics.net/lists/netdev/msg729128.html

/Magnus

Re: AF_XDP (i40e) behavior change in 5.11?

[Dan Siemon]

On Wed, 2021-03-24 at 11:00 +0100, Magnus Karlsson wrote:
> On Wed, Mar 24, 2021 at 10:22 AM Dan Siemon <dan@coverfire.com>
> wrote:
> > 
> > ./xsk_fwd -i ens1f0 -q 2 -i ens1f1 -q 2 -c 5 -i ens1f0 -q 3 -i
> > ens1f1 -
> > q 3 -c 6
> > 
> > On 5.10, the cores (2,3) assigned interrupts for the queues (2,3)
> > are
> > idle when there is no traffic.
> > 
> > On 5.11, the cores immediately go to 100% system time when the
> > program
> > starts (as viewed in htop). There is no network traffic.
> > 
> > Intel 710/i40e.
> > 
> > I tried this with xsk_fwd built from 5.10 and 5.11 trees with the
> > same
> > result.
> > 
> > Is this behavior change expected?
> 
> Thanks for reporting this. No, this is not intentional. I did send up
> a fix for a bug in the i40e driver that was introduced in 5.11. It
> might be the culprit. Could you please try the patch below and see if
> it works? It is not yet in net.
> 
> https://www.spinics.net/lists/netdev/msg729128.htm

Hi Magnus,

It appears (quick testing) that the patch fixes the problem.

Thanks.

Re: AF_XDP (i40e) behavior change in 5.11?

[Magnus Karlsson]

On Thu, Mar 25, 2021 at 11:17 PM Dan Siemon <dan@coverfire.com> wrote:
>
> On Wed, 2021-03-24 at 11:00 +0100, Magnus Karlsson wrote:
> > On Wed, Mar 24, 2021 at 10:22 AM Dan Siemon <dan@coverfire.com>
> > wrote:
> > >
> > > ./xsk_fwd -i ens1f0 -q 2 -i ens1f1 -q 2 -c 5 -i ens1f0 -q 3 -i
> > > ens1f1 -
> > > q 3 -c 6
> > >
> > > On 5.10, the cores (2,3) assigned interrupts for the queues (2,3)
> > > are
> > > idle when there is no traffic.
> > >
> > > On 5.11, the cores immediately go to 100% system time when the
> > > program
> > > starts (as viewed in htop). There is no network traffic.
> > >
> > > Intel 710/i40e.
> > >
> > > I tried this with xsk_fwd built from 5.10 and 5.11 trees with the
> > > same
> > > result.
> > >
> > > Is this behavior change expected?
> >
> > Thanks for reporting this. No, this is not intentional. I did send up
> > a fix for a bug in the i40e driver that was introduced in 5.11. It
> > might be the culprit. Could you please try the patch below and see if
> > it works? It is not yet in net.
> >
> > https://www.spinics.net/lists/netdev/msg729128.htm
>
> Hi Magnus,
>
> It appears (quick testing) that the patch fixes the problem.

Perfect! I will send a mail and try to get it through the internal
testing as soon as possible so we can get it up on the list and
merged.

Thank you: Magnus

> Thanks.
>