Re: [PATCH net-next 4/4] net: add heuristic for enabling TCP fraglist GRO

[Eric Dumazet]

On Wed, Apr 24, 2024 at 8:05 PM Felix Fietkau <nbd@xxxxxxxx> wrote:
>
> When forwarding TCP after GRO, software segmentation is very expensive,
> especially when the checksum needs to be recalculated.
> One case where that's currently unavoidable is when routing packets over
> PPPoE. Performance improves significantly when using fraglist GRO
> implemented in the same way as for UDP.
>
> When NETIF_F_GRO_FRAGLIST is enabled, perform a lookup for an established
> socket in the same netns as the receiving device. While this may not
> cover all relevant use cases in multi-netns configurations, it should be
> good enough for most configurations that need this.
>
> Here's a measurement of running 2 TCP streams through a MediaTek MT7622
> device (2-core Cortex-A53), which runs NAT with flow offload enabled from
> one ethernet port to PPPoE on another ethernet port + cake qdisc set to
> 1Gbps.
>
> rx-gro-list off: 630 Mbit/s, CPU 35% idle
> rx-gro-list on:  770 Mbit/s, CPU 40% idle
>
> Signe-off-by: Felix Fietkau <nbd@xxxxxxxx>
> ---
>  net/ipv4/tcp_offload.c   | 45 ++++++++++++++++++++++++++++++++++++++-
>  net/ipv6/tcpv6_offload.c | 46 +++++++++++++++++++++++++++++++++++++++-
>  2 files changed, 89 insertions(+), 2 deletions(-)
>
> diff --git a/net/ipv4/tcp_offload.c b/net/ipv4/tcp_offload.c
> index 6294e7a5c099..f987e2d8423a 100644
> --- a/net/ipv4/tcp_offload.c
> +++ b/net/ipv4/tcp_offload.c
> @@ -404,6 +404,49 @@ void tcp_gro_complete(struct sk_buff *skb)
>  }
>  EXPORT_SYMBOL(tcp_gro_complete);
>
> +static bool tcp4_check_fraglist_gro(struct sk_buff *skb)
> +{
> +       const struct iphdr *iph = skb_gro_network_header(skb);
> +       struct net *net = dev_net(skb->dev);
> +       unsigned int off, hlen, thlen;
> +       struct tcphdr *th;
> +       struct sock *sk;
> +       int iif, sdif;
> +
> +       if (!(skb->dev->features & NETIF_F_GRO_FRAGLIST))
> +               return false;
> +
> +       inet_get_iif_sdif(skb, &iif, &sdif);
> +
> +       off = skb_gro_offset(skb);
> +       hlen = off + sizeof(*th);
> +       th = skb_gro_header(skb, hlen, off);
> +       if (unlikely(!th))
> +               return false;
> +
> +       thlen = th->doff * 4;
> +       if (thlen < sizeof(*th))
> +               return false;
> +
> +       hlen = off + thlen;
> +       if (!skb_gro_may_pull(skb, hlen)) {
> +               th = skb_gro_header_slow(skb, hlen, off);
> +               if (unlikely(!th))
> +                       return false;
> +       }
> +
> +       sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
> +                                      iph->saddr, th->source,
> +                                      iph->daddr, ntohs(th->dest),
> +                                      iif, sdif);

Presumably all this could be done only for the first skb/segment of a GRO train.

We could store the fraglist in a single bit in NAPI_GRO_CB(skb) ?

GRO does a full tuple evaluation, we can trust it.