> Now that I know the issue is only in TCP, I can speculate that all bytes are > > being reported, but done with fewer messages. It may warrant some > > investigation in case there is some kind of bug. > > This would definitely still be a bug and should not happen. We have > quite a bit of experience with TCP zerocopy and I have not run into > this in practice, so I do think that it is somehow a test artifact.
To be clear, I'm not saying that it is an artifact of your extensions. It's quite likely that the zerocopy benchmark was a bit flaky in that regard all along. No need to spend further time on that for this patchset.
