Re: [PATCH net-next] selftests: net: local_termination: annotate the expected failures

[Petr Machata <petrm@nvidia.com>]

Jakub Kicinski <kuba@kernel.org> writes:

> On Mon, 13 May 2024 15:25:38 +0200 Petr Machata wrote:
>> For veth specifically there is xfail_on_veth:
>> 
>> xfail_on_veth $rcv_if_name \
>> 	check_rcv $rcv_if_name "Unicast IPv4 to unknown MAC address" \
>> 		  "$smac > $UNKNOWN_UC_ADDR1, ethertype IPv4 (0x0800)" \
>> 		  false
>> 
>> Which is IMHO clearer than passing an extra boolean.
>
> The extra boolean is because we want to only fail particular subcases.
> I think that's legit. If the other cases regress we want to know.
> Otherwise running the test on SW bridge is only useful for catching
> crashes (so less useful).

Likewise you only annotate with xfail_on_* the testcases that you want
to xfail. The FAIL_TO_XFAIL ought to only be set for that one subcase
and then revert to its former value. (That's the intention anyway.)

> So we probably need to reset FAIL_TO_XFAIL in this case.
> Any preference on how to go about that? I'm tempted to:
>
> diff --git a/tools/testing/selftests/net/forwarding/lib.sh
> b/tools/testing/selftests/net/forwarding/lib.sh index
> 112c85c35092..79aa3c8bc288 100644 ---
> a/tools/testing/selftests/net/forwarding/lib.sh +++
> b/tools/testing/selftests/net/forwarding/lib.sh @@ -473,6 +473,13 @@
> ret_set_ksft_status() # Whether FAILs should be interpreted as XFAILs.
> Internal. FAIL_TO_XFAIL=
>  
> +# Clear internal failure tracking for the next test case
> +begin_test()
> +{
> +    RET=0
> +    FAIL_TO_XFAIL=
> +}
> +
>  check_err()
>  {
>  	local err=$1
> diff --git
> a/tools/testing/selftests/net/forwarding/local_termination.sh
> b/tools/testing/selftests/net/forwarding/local_termination.sh index
> 65694cdf2dbb..0781ceba1348 100755 ---
> a/tools/testing/selftests/net/forwarding/local_termination.sh +++
> b/tools/testing/selftests/net/forwarding/local_termination.sh @@ -76,7
> +76,7 @@ check_rcv() local xfail_sw=$5 
>  	[ $should_receive = true ] && should_fail=0 || should_fail=1
> -	RET=0
> +	begin_test
>  
>  	tcpdump_show $if_name | grep -q "$pattern"
>  
>
>> Not sure what to do about the bridge bit though. In principle the
>> various xfail_on_'s can be chained, so e.g. this should work:
>> 
>> xfail_on_bridge $rcv_if_name \
>> xfail_on_veth $rcv_if_name \
>> 	check_rcv $rcv_if_name "Unicast IPv4 to unknown MAC address" \
>> 		  "$smac > $UNKNOWN_UC_ADDR1, ethertype IPv4 (0x0800)" \
>> 		  false
>> 
>> I find this preferable to adding these ad-hoc tweaks to each test
>> individually. Maybe it would make sense to have:
>> 
>> xfail_on_kind $rcv_if_name veth bridge \
>> 	check_rcv $rcv_if_name "Unicast IPv4 to unknown MAC address" \
>> 		  "$smac > $UNKNOWN_UC_ADDR1, ethertype IPv4 (0x0800)" \
>> 		  false
>> 
>> And then either replace the existing xfail_on_veth's (there are just a
>> handful) or convert xfail_on_veth to a wrapper around xfail_on_kind.
>
> I think the bridge thing we can workaround by just checking
> if ${NETIFS[p1]} is veth, rather than $rcv_if_name.
> Since the two behave the same.

I don't follow. The test has two legs, one creates a VRF and attaches
p2, the other creates a bridge and attaches p2. Whether p1 and p2 are
veth or HW seems orthogonal to whether $rcv_if_name is a bridge or a
veth.