On Oct 22, 2021, at 7:55 AM, Stuart Henderson <s...@spacehopper.org> wrote: > > On 2021/10/22 11:25, Jan Klemkow wrote: >> this diff add hardware checksum offloading for the receive path of >> ixl(4) interfaces. > > Would be good to have this tested with NFS if anyone has a way to do so. > nics are probably better now but I'm pretty sure we have had problems > with NFS and offloading in the past.
Indeed, I'm working on restoring checksumming to the later em devices (the PCIe versions use a different descriptor format than the PCI-X ones, so offload is disabled on all of them; notably, apu2/4 boxes use a lot of i210/211s, which suffer from this). Hoping to get a chance to work it into a real patch sometime soon. The bright side is that the operations aren't *hugely* different, it's mostly just the descriptor format. Adding IPv6 offloads will be a bit more involved. I've also a fix for the "bug" which causes hardware Rx checksum stripping to be disabled on ones with multiple queues; it's not a bug so much as documentation which doesn't explicitly call out that the per-queue setting overrides the global setting, and it's disabled by default on the queues. Intel's driver at the time it entered OpenBSD had this issue as well, and in their git history it was fixed a few months later (they started explicitly initializing the queue setting), but that never made it back in here. Either way, I'll be sure to test it out with some NFS traffic. - Dave
