[PATCH] bcm53xx: disable GRO support at kernel level

[Rafał Miłecki]

On 23.06.2022 05:49, Rafał Miłecki wrote:
> On 21.06.2022 01:13, Rosen Penev wrote:
>> On Mon, Jun 20, 2022 at 2:33 AM Rafał Miłecki <zajec5 at gmail.com> wrote:
>>>
>>> On 10.06.2022 15:16, Rafał Miłecki wrote:
>>>> This improves NAT masquarade network performance.
>>>>
>>>> An alternative to kernel change would be runtime setup but that requires
>>>> ethtool and identifying relevant network interface and all related
>>>> switch ports interfaces.
>>>
>>> I sent another patch that should supersede this one:
>>> [PATCH] bcm53xx: disable GRO by default at kernel level
>>> https://patchwork.ozlabs.org/project/openwrt/patch/20220620082120.11714-1-zajec5@gmail.com/
>>>
>>> It's probably a better idea to disable GRO *by default* (and allow
>>> re-enabling it).
>> Reminds me of this patch:
>> https://github.com/openwrt/openwrt/commit/13e5e473699b92f171205e0f5c57c9ebe7922492
>> which was subsequently reverted because of slower routing performance
>>
>> Is this patch the same as using napi_gro_receive and passing NETIF_F_GRO ?
> 
> Kind of. Reverting above patch removes ar71xx support for GRO totally.
> It's not the best option as some people / cases / devices may want it.
> It's much better to just disable it.
> 
> GRO can be runtime disabled using ethtool like:
> ethtool -K eth0 gro off
> ethtool -K lan1 gro off
> ethtool -K lan2 gro off
> ethtool -K lan3 gro off
> ethtool -K lan4 gro off
> ethtool -K wan gro off
> (or enabled with "on").
> 
> It can also be disabled by default with patch like the one I sent. In
> that case you can still enable it with:
> ethtool -K eth0 gro on
> ethtool -K lan1 gro on
> ethtool -K lan2 gro on
> ethtool -K lan3 gro on
> ethtool -K lan4 gro on
> ethtool -K wan gro on

I was refering to the
[PATCH] bcm53xx: disable GRO by default at kernel level
https://patchwork.ozlabs.org/project/openwrt/patch/20220620082120.11714-1-zajec5@gmail.com/

With *this* patch (the one we're commenting on) it isn't possible to
enable GRO with ethtool as it's permanently disabled on kernel level.
That's why I decided to send patch to disable GRO only in default
(initial) setup.


> So for ar71xx I'd suggest to:
> 1. Re-implement GRO support
> 2. Disable GRO by default if it worsens performance for most devices