Re: [PATCH net-next 2/2] gre: add GSO support

[Eric Dumazet <eric.dumazet@gmail.com>]

On Fri, 2012-09-14 at 10:12 -0700, Maciej Żenczykowski wrote:
> Thanks!
> 
> So I'm guessing it should be easy to add support like this to IPIP and
> SIT now as well
> (not sure if they're missing, but I'd assume so)?

Yes, absolutely.

I was also playing adding GRO support to tunnels.

Here the prototype patch I currently have, it gives very nice speedups

diff --git a/net/ipv4/ip_gre.c b/net/ipv4/ip_gre.c
index b062a98..ca6ab66 100644
--- a/net/ipv4/ip_gre.c
+++ b/net/ipv4/ip_gre.c
@@ -589,6 +589,65 @@ ipgre_ecn_encapsulate(u8 tos, const struct iphdr *old_iph, struct sk_buff *skb)
 	return INET_ECN_encapsulate(tos, inner);
 }
 
+static void tunnel_prequeue(struct ip_tunnel *tunnel, struct sk_buff *skb)
+{
+	unsigned long flags;
+	const struct iphdr *iph;
+
+	if (skb_cloned(skb)) {
+		netif_rx(skb);
+		return;
+	}
+	if (!pskb_may_pull(skb, sizeof(*iph)) ||
+	    skb_queue_len(&tunnel->napi_skbs) > 1000) {
+drop:
+		atomic_long_inc(&tunnel->dev->rx_dropped);
+		kfree_skb(skb);
+		return;
+	}
+	iph = ip_hdr(skb);
+	if (*(u8 *)iph == 0x45 &&
+	    iph->protocol == IPPROTO_TCP &&
+	    skb->ip_summed == CHECKSUM_NONE) {
+		__skb_pull(skb, sizeof(*iph));
+
+		skb->csum = csum_tcpudp_nofold(iph->saddr, iph->daddr,
+					       skb->len, IPPROTO_TCP, 0);
+		if (__skb_checksum_complete(skb))
+			goto drop;
+
+		__skb_push(skb, sizeof(*iph));
+	}
+
+	spin_lock_irqsave(&tunnel->napi_skbs.lock, flags);
+
+	__skb_queue_tail(&tunnel->napi_skbs, skb);
+	if (skb_queue_len(&tunnel->napi_skbs) == 1)
+		napi_schedule(&tunnel->napi);
+
+	spin_unlock_irqrestore(&tunnel->napi_skbs.lock, flags);
+}
+
+static int tunnel_napi_poll(struct napi_struct *napi, int budget)
+{
+	struct ip_tunnel *tunnel = container_of(napi, struct ip_tunnel, napi);
+	struct sk_buff *skb;
+	int work_done = 0;
+
+	while (work_done < budget) {
+		skb = skb_dequeue(&tunnel->napi_skbs);
+		if (!skb)
+			break;
+		skb->next = NULL;
+		napi_gro_receive(napi, skb);
+		work_done++;
+	}
+
+	if (work_done < budget)
+		napi_complete(napi);
+	return work_done;
+}
+
 static int ipgre_rcv(struct sk_buff *skb)
 {
 	const struct iphdr *iph;
@@ -714,8 +773,7 @@ static int ipgre_rcv(struct sk_buff *skb)
 		skb_reset_network_header(skb);
 		ipgre_ecn_decapsulate(iph, skb);
 
-		netif_rx(skb);
-
+		tunnel_prequeue(tunnel, skb);
 		rcu_read_unlock();
 		return 0;
 	}
@@ -745,6 +803,10 @@ static netdev_tx_t ipgre_tunnel_xmit(struct sk_buff *skb, struct net_device *dev
 	__be32 dst;
 	int    mtu;
 
+	if (skb->ip_summed == CHECKSUM_PARTIAL &&
+	    skb_checksum_help(skb))
+		goto tx_error;
+
 	if (dev->type == ARPHRD_ETHER)
 		IPCB(skb)->flags = 0;
 
@@ -1292,10 +1354,19 @@ static const struct net_device_ops ipgre_netdev_ops = {
 
 static void ipgre_dev_free(struct net_device *dev)
 {
+	struct ip_tunnel *tunnel = netdev_priv(dev);
+
+	netif_napi_del(&tunnel->napi);
+	skb_queue_purge(&tunnel->napi_skbs);
 	free_percpu(dev->tstats);
 	free_netdev(dev);
 }
 
+#define GRE_FEATURES (NETIF_F_SG |		\
+		      NETIF_F_FRAGLIST |	\
+		      NETIF_F_HIGHDMA |		\
+		      NETIF_F_HW_CSUM)
+
 static void ipgre_tunnel_setup(struct net_device *dev)
 {
 	dev->netdev_ops		= &ipgre_netdev_ops;
@@ -1309,6 +1380,9 @@ static void ipgre_tunnel_setup(struct net_device *dev)
 	dev->addr_len		= 4;
 	dev->features		|= NETIF_F_NETNS_LOCAL;
 	dev->priv_flags		&= ~IFF_XMIT_DST_RELEASE;
+
+	dev->features		|= GRE_FEATURES;
+	dev->hw_features	|= GRE_FEATURES;
 }
 
 static int ipgre_tunnel_init(struct net_device *dev)
@@ -1340,7 +1414,9 @@ static int ipgre_tunnel_init(struct net_device *dev)
 	dev->tstats = alloc_percpu(struct pcpu_tstats);
 	if (!dev->tstats)
 		return -ENOMEM;
-
+	skb_queue_head_init(&tunnel->napi_skbs);
+	netif_napi_add(dev, &tunnel->napi, tunnel_napi_poll, 64);
+	napi_enable(&tunnel->napi);
 	return 0;
 }