On 17/05/18 09:35 AM, Vlad Buslov wrote:

On Wed 16 May 2018 at 21:51, Jiri Pirko [off-list ref] wrote:

quoted

Wed, May 16, 2018 at 11:23:41PM CEST, vladbu@mellanox.com wrote:

quoted

quoted

quoted

quoted

Please make sure you have these in your kernel config:

CONFIG_NET_ACT_IFE=y
CONFIG_NET_IFE_SKBMARK=m
CONFIG_NET_IFE_SKBPRIO=m
CONFIG_NET_IFE_SKBTCINDEX=m

Roman, could you please add this to some file? Something similar to:
tools/testing/selftests/net/forwarding/config

How would putting the file there help?

quoted

Thanks!

quoted

quoted

For tdc to run all the tests, it is assumed that all the supported tc
actions/filters are enabled and compiled.

Enabling these options allowed all ife tests to pass. Thanks!

Error in u32 test still appears however:

Test e9a3: Add u32 with source match

-----> prepare stage *** Could not execute: "$TC qdisc add dev $DEV1 ingress"

-----> prepare stage *** Error message: "Cannot find device "v0p1"

I investigated and was able to fix u32 problems.

First of all, u32 test requires having veth interfaces that are not
created by test infrastructure by default. Following command fixes the
issue:

sudo ip link add v0p0 type veth peer name v0p1

That is documented on the README i believe - however, we should
be able to detect that a test needs the device and create it via
a plugin. Lucas?

After executing this command test passes, however looking at test
definition itself it seems meaningless. It creates filter with match
source IP 127.0.0.1, then tests if filter with source IP 127.0.0.2
exists, but passes successfully because it actually expects to match
zero filters with such IP :)

I fixed it and it passed properly matching single filter with source IP
127.0.0.2.

Please send a patch.

After this flower test failed. The flower test expects that user
explicitly provide "-d" option with interface to use. With -d it failed
again. This time because it expects action to have 1m references, but
actual value was 1000001. I investigated it and found out that test
passed, if executed without running other tests first. So it seemed that
some other test was leaking reference to gact action. It turned out that
culprit was mirred test 6fb4, which created pipe action but didn't flush
it afterward.

Hopefully the last patch from Roman fixes that? Otherwise send something
on top.

With all tests passing on that particular version of net-next, I will
now rebase my changes on top of it and run them again.

Thank you Vlad!

cheers,
jamal