dev_iavf: Unable to use more queues than offered initially

[vrpolakatcisco]

25.02 marks the first CSIT release that uses dev_iavf, and I see some issues with 710 series NICs. This one is related to the difference between the initial and maximal number of queue pairs. Specifically, dev_attach trickles down to iavf_init wich first asks [0] for a resource and uses the value found there to set a maximum [1], which for Intel-xxv710 means 4 queue pairs, so that is how many queues are created [2]. Now, when a test uses 4 cores (with hyperthreading, so 8 workers), CSIT uses dev_create_port_if asking for 8 RX queues, which is more than the previous maximum so it gets rejected [3], causing the test to fail [4].

Compare that behavior to plugins/avf which does not consult any default values and just asks [5] directly for the requested number of queues, driver obliges and test passes, proving NICs like Intel-xxv710 can indeed support more than 4 AVF queue pairs. Not sure what would be a correct fix for plugins/dev_iavf.

[0] https://github.com/FDio/vpp/blob/d9c2d94e80444f0eec5dddf86a7a98d335761f48/src/plugins/dev_iavf/iavf.c#L145 [1] https://github.com/FDio/vpp/blob/d9c2d94e80444f0eec5dddf86a7a98d335761f48/src/plugins/dev_iavf/iavf.c#L175-L176 [2] https://github.com/FDio/vpp/blob/d9c2d94e80444f0eec5dddf86a7a98d335761f48/src/vnet/dev/port.c#L610-L612 [3] https://github.com/FDio/vpp/blob/d9c2d94e80444f0eec5dddf86a7a98d335761f48/src/vnet/dev/api.c#L204-L205 [4] https://s3-logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-report-iterative-2502-2n-zn2/42/log.html.gz#s1-s1-s1-s2-s10-t3-k2-k6-k1-k1-k1-k1-k1-k13-k1-k1-k1-k1 [5] https://github.com/FDio/vpp/blob/d9c2d94e80444f0eec5dddf86a7a98d335761f48/src/plugins/avf/device.c#L912-L918