Re: [PATCH net-next v9 2/3] net: gro: move L3 flush checks to tcp_gro_receive and udp_gro_receive_segment

[Eric Dumazet]

On Thu, May 9, 2024 at 8:58 PM Richard Gobert <richardbgobert@gmailcom> wrote:
>

>
> Interesting, I think that is indeed a bug, that exists also in the current
> implementation.
> NAPI_GRO_CB(p)->ip_fixedid (is_atomic before we renamed it in this commit)
> is cleared as being part of NAPI_GRO_CB(skb)->zeroed in dev_gro_receive.

And the code there seems wrong.

A compiler can absolutely reorder things, I have seen this many times.

I would play safe here, to make sure NAPI_GRO_CB(skb)->is_atomic = 1;
can not be lost.

diff --git a/net/core/gro.c b/net/core/gro.c
index c7901253a1a8fc1e9425add77014e15b363a1623..6e4203ea4d54b8955a504e42633f7667740b796e
100644
--- a/net/core/gro.c
+++ b/net/core/gro.c
@@ -470,6 +470,7 @@ static enum gro_result dev_gro_receive(struct
napi_struct *napi, struct sk_buff
        BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct napi_gro_cb, zeroed),
                                        sizeof(u32))); /* Avoid slow
unaligned acc */
        *(u32 *)&NAPI_GRO_CB(skb)->zeroed = 0;
+       barrier();
        NAPI_GRO_CB(skb)->flush = skb_has_frag_list(skb);
        NAPI_GRO_CB(skb)->is_atomic = 1;
        NAPI_GRO_CB(skb)->count = 1;