With modern Xen and Linux kernels domains seem to be able to use MTU larger
than 1500 (up to 16K) on xennet. I had a brief look at the code, but I do
not see specifically how this is supported. Does anyone know?
I'm currently looking at implementing "TSO" between the xen net frontend and the backend. On a real card, TSO exploits scatter-gather dma to accumulate mbuf chain data into the card buffer. On the shared memory ring buffer, we indicate "TSO" by using multiple slots in the shared ring which enables the other end to collate multiple slots into the same packet.
I'm not aware of a multi-page cross-domain copy optimisation using the shared ring.
I'd be keen to know if someone understands the "16k" MTU option, above.