Re: [PATCH net-next] net: avoid unneeded UDP L4 and fraglist GSO resegmentation

[Alexander Lobakin]

From: Willem de Bruijn <willemdebruijn.kernel@xxxxxxxxx>
Date: Sat, 31 Oct 2020 11:26:24 -0400

>>>> I think it is fine to reenable this again, now that UDP sockets will
>>>> segment unexpected UDP GSO packets that may have looped. We previously
>>>> added general software support in commit 83aa025f535f ("udp: add gso
>>>> support to virtual devices"). Then reduced its scope to egress only in
>>>> 8eea1ca82be9 ("gso: limit udp gso to egress-only virtual devices") to
>>>> handle that edge case.
>>
>> Regarding bonding and teaming: I think they should also use
>> NETIF_F_GSO_SOFTWARE mask, not NETIF_F_ALL_TSO, as SCTP also has
>> a software fallback. This way we could also remove a separate
>> advertising of NETIF_F_GSO_UDP_L4, as it will be included in the first.
>>
>> So, if this one:
>> 1. Add NETIF_F_GSO_UDP_L4 and NETIF_F_GSO_FRAGLIST to
>>    NETIF_F_GSO_SOFTWARE;
>> 2. Change bonding and teaming features mask from NETIF_F_ALL_TSO |
>>    NETIF_F_GSO_UDP_L4 to NETIF_F_GSO_SOFTWARE;
>> 3. Check that every virtual netdev has NETIF_F_GSO_SOFTWARE _or_
>>    NETIF_F_GSO_MASK in its advertising.
>>
>> is fine for everyone, I'll publish more appropriate and polished v2 soon.
>
> I think we can revert 8eea1ca82be9. Except for the part where it
> defines the feature in NETIF_F_GSO_ENCAP_ALL instead of
> NETIF_F_GSO_SOFTWARE. That appears to have been a peculiar choice. I
> can't recall exactly why I chose that. Most likely because that was
> (at the time) the only macro that covered all the devices I wanted to
> capture.
>
> As for SCTP: that has the same concern that prompted that commit for
> UDP: is it safe to forward those packets to the ingress path today?

Oh well. I just looked up into net/sctp/offload.c and see no GRO
receiving callbacks, only GSO ones. On the other hand,
NETIF_F_GSO_SOFTWARE includes GSO_SCTP and is used in almost every
virtual netdev driver, including macvlan and veth mentioned earlier,
so that seems to be fine.

> I had missed that there may be non-mainline drivers that do ;)
> Great to see these features getting offload support.

It will be mainlined sooner or later depending on my workload :)
UDP fraglists *really* boosted the things up for me, so I don't quite
understand why not a single mainline driver has a support for them.

Al