On Thu, Jan 25, 2024 at 12:38 PM Paolo Abeni <pabeni@xxxxxxxxxx> wrote:
>
> On Thu, 2024-01-25 at 09:48 +0100, Eric Dumazet wrote:
> > On Thu, Jan 25, 2024 at 9:23 AM Paolo Abeni <pabeni@xxxxxxxxxx> wrote:
> > >
> > > the udpgro_fraglist self-test uses the BPF classifiers, but the
> > > current net self-test configuration does not include it, causing
> > > CI failures:
> > >
> > > # selftests: net: udpgro_frglist.sh
> > > # ipv6
> > > # tcp - over veth touching data
> > > # -l 4 -6 -D 2001:db8::1 -t rx -4 -t
> > > # Error: TC classifier not found.
> > > # We have an error talking to the kernel
> > > # Error: TC classifier not found.
> > > # We have an error talking to the kernel
> > >
> > > Add the missing knob.
> > >
> > > Fixes: edae34a3ed92 ("selftests net: add UDP GRO fraglist + bpf self-tests")
> > > Signed-off-by: Paolo Abeni <pabeni@xxxxxxxxxx>
> >
> > FYI, while looking at the gro test, I found that using strace was
> > making it failing as well.
>
> It looks like the gro.sh (large) tests send the to-be-aggregate
> segments individually and relay on the gro flush timeout being large
> enough to fit all the relevant write operations. I suspect/hope
> something alike:
>
> ---
> diff --git a/tools/testing/selftests/net/setup_veth.sh b/tools/testing/selftests/net/setup_veth.sh
> index a9a1759e035c..1f78a87f6f37 100644
> --- a/tools/testing/selftests/net/setup_veth.sh
> +++ b/tools/testing/selftests/net/setup_veth.sh
> @@ -11,7 +11,7 @@ setup_veth_ns() {
> local -r ns_mac="$4"
>
> [[ -e /var/run/netns/"${ns_name}" ]] || ip netns add "${ns_name}"
> - echo 100000 > "/sys/class/net/${ns_dev}/gro_flush_timeout"
> + echo 1000000 > "/sys/class/net/${ns_dev}/gro_flush_timeout"
> ip link set dev "${ns_dev}" netns "${ns_name}" mtu 65535
> ip -netns "${ns_name}" link set dev "${ns_dev}" up
> ---
> should solve the sporadic issues.
I think you are right.
I tried multiple values, and found 600,000 was not enough in some cases.
With 1,000,000, I was able to run the test (with the strace overhead)
100 times without a single failure.
>
> > Not sure about this one...
>
> All the udpgro* test should write a single UDP GSO packet and let the
> veth segmenting it, hopefully slowing down either ends should not
> impact them - but I did not check yet!
>
> Perhaps even the gro.sh tests could be modified alike?
>
> Cheers,
>
> Paolo
>
