On Sun, 30 Nov 2025 09:56:24 -0500 Willem de Bruijn wrote: > Jakub Kicinski wrote: > > On Fri, 28 Nov 2025 15:42:40 -0500 Willem de Bruijn wrote: > > > So GRO off disables HW_GRO, but not LRO? That difference is behavior > > > is confusing. Could we still see this as a regression and make the > > > ethtool HW_GRO feature equally independent from SW_GRO? > > > > I couldn't convince myself that it's justified. Of course it would have > > made testing a lot easier. But apart from that - what's your reading of > > the status quo? Working backwards from were we ended up (and I > > haven't dug into the git history) I'm guessing that LRO disable is used > > to prevent changing geometry of the packets. GRO would presumably be > > disabled when user knows that it will be ineffective, to save the cost. > > Or when some portion of the stack (XDP?) can't deal with super frames. > > > > If those are the reasons, practically, I don't see why user would want > > HW GRO without SW. Ever since we allowed SW GRO to re-GRO HW GRO'ed > > frames it's always better to leave SW enabled. HW leaves a lot of > > aggregation opportunities on the table. > > > > I concluded that changing the current behavior would not help any real > > life scenario, just testing. LMK if you see one or the inconsistency > > is a big enough reason. > > I think that's fair. > > But from reading the code I don't see how disabling NETIF_F_GRO also > disables NETIF_F_GRO_HW. And indeed I just tested on one (admittedly > not latest upstream) IDPF driver and it does not.
Looks like you're right. Broadcom drivers where GRO_HW originates do it locally, so does qede. I guess somewhere along the way drives started treating GRO_HW as a separate feature rather than a GRO offload. I don't think it changes the reasoning in any major way?
