Erik Nordmark wrote:
That might be true of the currently implementation of TCP/IP. But a driver shouldn't assume that this will remain true. Drivers should do the checksum offload operation based on how the checksum offload interface is specified. And that specification doesn't say that the checksum offload will never apply to packets with IPv4 options or IPv6 extension headers.
Yes, you are correct that the comments are based on the implementation artifacts. But on the other hand, the original question was specific to partial (not full) checksum offload, and in such a case we encounter hardware limitations -- i.e. IIRC, ce's start/stuff offset is 6-bit wide and I'm quite sure the NIC Andrew is working with has a limitation similar to that. It seems unlikely that future hardware will "fix" this issue as the trend is moving towards full or more intelligent checksum offload cases. Perhaps all of this can be clarified when Sun finally publishes the hardware checksum interface. I recall there were efforts to do this, though I'm not sure where they stand now. Adi _______________________________________________ networking-discuss mailing list [email protected]
