Questions about checksum hacks

[Jimmy-Z]

I'm reading mimic docs, since I'm curious, and I've got several questions around "When is Checksum Hack Not Necessary?":

• when the driver doesn't use csum_offset, my understanding is that the NIC simply calculates the checksum from scratch, so mimic doesn't need to fix them, but does mimic disable the hack when this is the case? I should probably read the code, but I'm not there yet.
• "You can check if your NIC's driver source code contains csum_offset", maybe my understanding is incorrect, but shouldn't a proper hardware checksum offloading being implemented in hardware? not the driver? if the driver is handling it, what's the purpose of "offloading" then, it still runs on the CPU?
• when "checksum offload is maunally disabled", wouldn't that require mimic to calculate the entire checksum? a much bigger hassle? a typo there BTW.
• why are userspace udp sockets not affected? the document simply states that without any explanation, not that you're obligated to have one, but I want it :D

I tried to feed the document to Gemini, but I didn't get any sensible answers from it.

Thank you for this amazing piece of software, being using it for several months without any issue.

[hack3ric]

Hi, sorry for the late reply:

• No, Mimic still performs the hack if it's enabled at compile-time. It essentially becomes no-op in such cases. And it should be a trivial enough operation (at least in the default kfunc path) so it won't hurt performance too much.

• Yes, it's the hardware that does the final checksum calculation. But from my understanding some hardware implementations don't discover packet structure on their own and need the driver to pass this information, since the OS/driver already parses the packet (result stored in sk_buff), and all the hardware need for calculating plain RFC 1071 checksum is just csum_{start,offset}. Therefore it's convenient to pass some info from driver to hardware and avoid duplicated work in both CPU and NIC.

• Mimic calculates the delta of the checksum using bpf_csum_diff and bpf_l4_csum_replace (and a bunch of custom helpers). So if an incoming packet's checksum is full (not partial) and correct, it should stay corrected without the need of calculating from scratch. Checksum hack guides the OS/driver/NIC to find the correct the place of checksum field, rather than affecting the calculation of it.

• If checksum offloading is turned off, UDP userspace sockets just send packets with full checksums instead of partial ones. It's based on my observation in Wireshark, haven't read the kernel code on this one.