> Hum, I don't know it may depend on TCP options. When I wrote the tl driver
> padding I remember seeing much more short packets than that.

i don't think tcp with timestamp option involves such small packets.

> Maybe it depends on UDP traffic too (like NIS or DNS).

esp. the last ip fragment of udp datagram can be very small.
eg. using nfs over udp with files of the particular sizes.

btw, i vaguely remember the bug in question is only triggered with
very small packets like ip header + a few octets.
ie. the threshold is not ETHER_MIN_LEN.

YAMAMOTO Takashi