Dear TI staff,
it seems like a fix for the CPSW Linux driver from the end of last year introduced a regression which causes short Ethernet frames to be padded to 68B in total instead of the standard defined 64B.
We confirmed the behavior on the Beaglebone Black with an AM3358 processor by modifying the sent frame length incrementally and observing the traffic with Wireshark. We also compared with a few different USB-Ethernet-dongles which all showed the expected traffic. While for instance the ARP protocol does not seem to have a problem with additional padding in the payload, some industrial automation protocols seem to be strict in this regard.
Although we saw it on a Beaglebone Black, it seems to be a general driver problem which should affect everyone who uses CPSW in VLAN-unaware mode as the problem can be fixed for us by reverting this commit. My guess would be that the driver would have to track the configuration of the switch hardware to correctly pad the frame.
Are you aware of such a bug or should this maybe be avoided by using a different CPSW configuration?