Should this table be const uint64_t ?

Should if_printf() be used here? Or should these prints go away?

Would it be possible to add a bit more wording about the measurement units here?

I don't have Chelsio/Mellanox docs handy but my expectation is that the hw rate is indeed bps.
In my understanding, what is calculated here is burst size for the goodput - e.g. actual data without accounting for encapsulation.

If that's correct, then in similar example where we have 31Mbps reported by both TCP stack and supported HW queue, we would get 40 bytes overhead for each packet, which will ends up queuing 1-second delay each 37.5 seconds.

Typo: pacets

Don't we need to provide some error notification for the caller that there is no suitable hw rate?

Given it returns the ideal burst size, would it be possible to name the function something like tcp_get_pacing_burst_size() or similar to avoid confusion around mss?

First both Chelsio and Mellanox use bytes per second for controlling the card. Second
my testing has shown that this rate is the complete rate not the goodput rate. I recently
changed the NF version of rack to add in the overhead for this very reason. So no it is
the complete size including the encap (E-H, IP and TCP Hdr).

I can add a bit more here if you like.

Thats should have been fixed by my second update... caught that as I was
re-reading the comments ;=_

I did not feel an error was appropriate for this return. In this last case we are forcing
the stack to fall back to its rate that it would be using without hardware assist.
But I can see if we are still growing a queue that may not be desirable, in this
case the stack has made an error and selected the wrong hardware rate for the
rate it wants to send at. Hmm I suppose I could add a return code via a
passed argument that can be checked.

sure thats fine :-)