net: gro: remove obsolete code from skb_gro_receive() (58025e46) · Commits · 戴 / test

Some drivers use copybreak to copy tiny frames into smaller skb, and this smaller skb might not have skb->head_frag set for various reasons. skb_gro_receive() currently doesn't allow to aggregate the smaller skb into the previous GRO packet if this GRO packet has at least 2 MSS in it. Following workload easily demonstrates the problem. netperf -t TCP_RR -H target -- -r 3000,3000 (tcpdump shows one GRO packet with 2 MSS, plus one additional packet of 104 bytes that should have been appended.) It turns out that we can remove code from skb_gro_receive(), because commit ("net: gro: allow to build full sized skb") and its followups removed the assumption that a GRO packet with a frag_list had to have an empty head. Removing this code allows the aggregation of the last (incomplete) frame in some RPC workloads. Note that tcp_gro_receive() already takes care of forcing a flush if necessary, including this case. If we want to avoid using frag_list in the first place (in forwarding workloads for example, as the outgoing NIC is generally not able to cope with skbs having a frag_list), we need to address this separately. Signed-off-by:

Eric Dumazet <edumazet@google.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by:

David S. Miller <davem@davemloft.net>

net/core/skbuff.c

+1 −44

Original line number	Diff line number	Diff line
		@@ -3206,10 +3206,9 @@ int skb_gro_receive(struct sk_buff *head, struct sk_buff skb)
		struct skb_shared_info pinfo, skbinfo = skb_shinfo(skb);
		unsigned int offset = skb_gro_offset(skb);
		unsigned int headlen = skb_headlen(skb);
		struct sk_buff nskb, lp, p = head;
		unsigned int len = skb_gro_len(skb);
		struct sk_buff lp, p = *head;
		unsigned int delta_truesize;
		unsigned int headroom;

		if (unlikely(p->len + len >= 65536))
		return -E2BIG;
		@@ -3276,48 +3275,6 @@ int skb_gro_receive(struct sk_buff *head, struct sk_buff skb)
		NAPI_GRO_CB(skb)->free = NAPI_GRO_FREE_STOLEN_HEAD;
		goto done;
		}
		/* switch back to head shinfo */
		pinfo = skb_shinfo(p);

		if (pinfo->frag_list)
		goto merge;
		if (skb_gro_len(p) != pinfo->gso_size)
		return -E2BIG;

		headroom = skb_headroom(p);
		nskb = alloc_skb(headroom + skb_gro_offset(p), GFP_ATOMIC);
		if (unlikely(!nskb))
		return -ENOMEM;

		__copy_skb_header(nskb, p);
		nskb->mac_len = p->mac_len;

		skb_reserve(nskb, headroom);
		__skb_put(nskb, skb_gro_offset(p));

		skb_set_mac_header(nskb, skb_mac_header(p) - p->data);
		skb_set_network_header(nskb, skb_network_offset(p));
		skb_set_transport_header(nskb, skb_transport_offset(p));

		__skb_pull(p, skb_gro_offset(p));
		memcpy(skb_mac_header(nskb), skb_mac_header(p),
		p->data - skb_mac_header(p));

		skb_shinfo(nskb)->frag_list = p;
		skb_shinfo(nskb)->gso_size = pinfo->gso_size;
		pinfo->gso_size = 0;
		__skb_header_release(p);
		NAPI_GRO_CB(nskb)->last = p;

		nskb->data_len += p->len;
		nskb->truesize += p->truesize;
		nskb->len += p->len;

		*head = nskb;
		nskb->next = p->next;
		p->next = NULL;

		p = nskb;

		merge:
		delta_truesize = skb->truesize;

Admin message