Not addressed here: the AF_XDP zero-copy Rx path.
igb_clean_rx_irq_zc() has the same unguarded strip, and with 2 KiB
UMEM chunks the hardware buffer is rounded down to 1 KiB, so frames
span descriptors there. The igc counterpart is Maciej's buffer-size
patch in [2]. I do not use AF_XDP and cannot test it, so I would
prefer to leave that to someone who can.
I assume review bots will give you a hard time if you don't address
this within your series. Seems you have HW handy plus you do use LLM
so I don't see why you couldn't ask it to address AF_XDP zc rx ?
FWIW blessed user space tool has always been xdpsock:
https://github.com/xdp-project/bpf-examples/tree/main/AF_XDP-example
so you can ask llm to run this in rxdrop mode on receiver side and
feed it with some traffic from peer side and then see that frames
were not fragmented.
OK, fair point. I will look into it, and I will hold v5 until that is done and tested.
Thanks,
Tjerk