Well, it does not really work and we are getting retransmits.
To replicate, set nfs.read_size large enough.
What needs to happen is, we read ethernet packet, if it has
IPv4 payload and that payload is fragment, we create reassembly
queue (sorted by growing fragment offset) and on last
fragment, we can build complete packet. Once done properly,
the network load can utilize larger read sizes.
While there, move ARP (and other) processing out of readipv4().