I don't recall what you said earlier about congestion control for VFs? Is it that they are effectively forced to do the equivalent of 'cong_drop=1'? Will they ever send PAUSE frames?

This is a recent change that fixed a panic with the UDP_STREAM netperf test on the VF driver. More details in the git commit, but that test generated packets where the first mbuf was the Ethernet header, and the second mbuf contained IP + UDP + payload. This loop terminated after walking past the first mbuf due to len being 0 (when looking for the L3 header) without setting p. As a result it returned a NULL pointer which panic'd. The forever loop ensures it sets p to the start of the second mbuf before breaking out of the loop.

I think the Linux VF driver has a module parameter for this. I could add a tunable/sysctl to enforce the same thing.

Not sure if we want to try to be this fancy or if I should just remove the comment or remove the #ifdef around it?

The second "Terminator" is redundant, can just say "Terminator 4 and 5 ASICs"

Is "Note that" needed again on this second sentence?

That aside should not be there. The sentence is introducing a list of supported cards, and then suddenly veers into a tree.
It should at least use .Ql for the "cxlv" part. Maybe move that to a separate sentence after the list of cards?

(card list)

Ports created for the T5 cards are named
.Ql cxlv .

(Not part of your patch, I know.)

Lists like these generally read better without the redundant introductory article on every item. So just "Number of tx queues..."

"Note that" is probably not needed here, or maybe could be said differently.

This is actually in the source manpage (cxgbe.4). Note that "Terminator 4" is the product name of the ASIC. I'm not sure if Chelsio would prefer that the names be explicit vs combined.

It is repetitive, but I wanted a way to call out the sharing since it might be a potential "gotcha" for users who may not expect the same variable to affect two different (albeit related) drivers. Do you have a wording suggestion?

this is also copied from the cxgbe.4 page, though I can fix that in both places. We do already mention cxlv vs cxgbe earlier around line 68. I think the desire in cxgbe(4) was to note the difference close to the list of cards. However, it might be that we can just drop the aside entirely?

I'd say either option is preferable to leaving it where it is now, but don't know which is better. Your choice. :)

I'm not sure that makes it clear that hw.cxgbe.foo affects both cxgbeX devices (T4 PF driver) and cxgbevX (T4 VF driver). Maybe just s/Both/The/ works though. (I think with "both" there is the chance that it could be misread as just saying that the two PF drivers share the tunables since "both" implies exactly 2, whereas in this case it is really 4 things sharing {T4,T5}{VF,PF}) Actually, the previous wording makes it clear that even though there are multiple names, there is a shared driver (which is true, kldload if_cxgbe gives you both cxgbe and cxl). I've gone with:

The Physical Function driver for T4 and T5 adapters shares these tunables.