Re: [PATCH net-next 1/3] net: gro: set {inner_,}network_header in receive phase

[Richard Gobert]

Eric Dumazet wrote:
> 
> My intuition is that this patch has a high cost for normal GRO processing.
> SW-GRO is already a bottleneck on ARM cores in smart NICS.
> 
> I would suggest instead using parameters to give both the nhoff and thoff values
> this would avoid many conditionals in the fast path.
> 
> ->
> 
> INDIRECT_CALLABLE_SCOPE int udp6_gro_complete(struct sk_buff *skb, int
> nhoff, int thoff)
> {
>  const struct ipv6hdr *ipv6h = (const struct ipv6hdr *)(skb->data + nhoff);
>  struct udphdr *uh = (struct udphdr *)(skb->data + thoff);
> ...
> }
> 
> INDIRECT_CALLABLE_SCOPE int tcp6_gro_complete(struct sk_buff *skb, int
> nhoff, int thoff)
> {
>        const struct ipv6hdr *iph =  (const struct ipv6hdr *)(skb->data + nhoff);
>        struct tcphdr *th = (struct tcphdr *)(skb->data + thoff);
> 
> Why storing in skb fields things that really could be propagated more
> efficiently as function parameters ?

Hi Eric,
Thanks for the review!
 
I agree, the conditionals could be a problem and are actually not needed.
The third commit in this patch series introduces an optimisation for
ipv6/ipv4 using the correct {inner_}network_header. We can remove the
conditionals; I thought about multiple ways to do so. First, remove the
conditional in skb_gro_network_offset:
 
    static inline int skb_gro_network_offset(const struct sk_buff *skb)
    {
        const u32 mask = NAPI_GRO_CB(skb)->encap_mark - 1;
        return (skb_network_offset(skb) & mask) | (skb_inner_network_offset(skb) & ~mask);
    }
 
And for the conditionals in {inet,ipv6}_gro_receive I thought about two
ideas. The first is to move set_inner_network_header to encapsulation gro
functions like ipip_gro_receive, this way there's one less write (in
comparison to main) in these functions:

    static struct sk_buff *ipip_gro_receive(struct list_head *head,
                        struct sk_buff *skb)
    {
        ...
 
        NAPI_GRO_CB(skb)->encap_mark = 1;
        skb_set_inner_network_header(skb, skb_gro_offset(skb));
 
The second way is to always write to inner_network_header:

    INDIRECT_CALLABLE_SCOPE struct sk_buff *ipv6_gro_receive(struct list_head *head,
                                struct sk_buff *skb)
    {
        ...
        skb_set_inner_network_header(skb, off);
        ...
 
What do you think is better? I think the 1st is more beneficial for the
fast path.

We could then use the {inner_}network_header separation to optimise the
receive path, such as in the 3rd commit in this patch series.
 
Regards,
Richard