Re: [PATCH 10/13] selftests/bpf: test_xsk: Split xskxceiver

[Bastien Curutchet <bastien.curutchet@bootlin.com>]

Hi Maciej,

On 3/18/25 2:16 PM, Maciej Fijalkowski wrote:
> On Thu, Mar 13, 2025 at 11:48:08AM +0100, Bastien Curutchet (eBPF Foundation) wrote:
>> AF_XDP features are tested by the test_xsk.sh script but not by the
>> test_progs framework. The tests used by the script are defined in
>> xksxceiver.c which can't be integrated in the test_progs framework as is.
>>
>> Extract these test definitions from xskxceiver{.c/.h} to put them in new
>> test_xsk{.c/.h} files.
>> Keep the main() function and its unshared dependencies in xksxceiver to
>> avoid impacting the test_xsk.sh script which is often used to test real
>> hardware.
>>
>> Signed-off-by: Bastien Curutchet (eBPF Foundation) <bastien.curutchet@bootlin.com>
>> ---
>>   tools/testing/selftests/bpf/Makefile     |    2 +-
>>   tools/testing/selftests/bpf/test_xsk.c   | 2416 ++++++++++++++++++++++++++++
>>   tools/testing/selftests/bpf/test_xsk.h   |  279 ++++
>>   tools/testing/selftests/bpf/xskxceiver.c | 2503 +-----------------------------
>>   tools/testing/selftests/bpf/xskxceiver.h |  156 --
>>   5 files changed, 2727 insertions(+), 2629 deletions(-)
>> +
> 
> (...)
> 
>> +int testapp_hw_sw_max_ring_size(struct test_spec *test)
>> +{
>> +	u32 max_descs = XSK_RING_PROD__DEFAULT_NUM_DESCS * 4;
>> +	int ret;
>> +
>> +	test->set_ring = true;
>> +	test->total_steps = 2;
>> +	test->ifobj_tx->ring.tx_pending = test->ifobj_tx->ring.tx_max_pending;
>> +	test->ifobj_tx->ring.rx_pending  = test->ifobj_tx->ring.rx_max_pending;
>> +	test->ifobj_rx->umem->num_frames = max_descs;
>> +	test->ifobj_rx->umem->fill_size = max_descs;
>> +	test->ifobj_rx->umem->comp_size = max_descs;
>> +	test->ifobj_tx->xsk->batch_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
>> +	test->ifobj_rx->xsk->batch_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
>> +
>> +	ret = testapp_validate_traffic(test);
>> +	if (ret)
>> +		return ret;
>> +
>> +	/* Set batch_size to 8152 for testing, as the ice HW ignores the 3 lowest bits when
>> +	 * updating the Rx HW tail register.
>> +	 */
>> +	test->ifobj_tx->xsk->batch_size = test->ifobj_tx->ring.tx_max_pending - 8;
>> +	test->ifobj_rx->xsk->batch_size = test->ifobj_tx->ring.tx_max_pending - 8;
>> +	if (!pkt_stream_replace(test, max_descs, MIN_PKT_SIZE))
> 
> Here's the victim of test failures that i reported last week. This
> function succeds with 0 and you interpret it as failure:)

Oops ... Thanks for the feedback.

> One sign wrong caused two days of debugging, but it was kinda fun.
> 

I should have ordered patches in an other way to split the xskxceiver.c 
sooner, the bisection would have narrowed the bug more precisely. I'll 
do this in next iteration.

> What was happening was due to the failure here one of the sockets was not
> deleted and later on whole test suite could not attach the socket for
> every other test case which caused the ever going failures. Which makes me
> think that since you changed the logic from exits to returning failures
> probably we need to take care of state cleanup in a better way so case
> like this one described here wouldn't take place.
> 

I hadn't notice the cleanup done in __testapp_validate_traffic(), I'll 
handle this in a better way in next iteration.

> This test is only executed for hw tests that's why you haven't seen this
> problem as you were focused on veth case.
> 
> If you want to proceed with that then i would like to ask you to grab the
> hw and check you're not introducing regressions. FWIW i have been working
> with i40e and ice drivers.
> 

I don't think I have such hardware available but I'll try to find some 
equivalent to test the next iteration on real HW before submitting it.

> One last note is that verbose mode seemed to be broken for me.
> 

I'll take a look at this, thank you.


Best regards,
Bastien