I also experienced that issue during my research. It is important to note that only TX needs to be reduced to resolve the original issue. Further more the change fails to consider that the RX buffer is later reduced by IPHETH_IP_ALIGN (2 bytes) resulting in 1512 bytes effective. Hence packets of size 1513-1514 result in overflow in ipheth_rcvbulk_callback(), after which the driver malfunctions and all communication stops. My patch corrects that by increasing the RX buffer size by adding IPHETH_IP_ALIGN to compensate because that the same padding is subtracted later.
Anyone experiencing disconnects after the original patch, please help me find a reviewer for the updated patch below!