New and improved zero-copy xmit function!

1 files changed, 95 insertions, 45 deletions

diff --git a/quicktun.c b/quicktun.c
index 4881960..e745e48 100644
--- a/quicktun.c
+++ b/quicktun.c
@@ -9,6 +9,8 @@
 #include <linux/udp.h>
 #include <net/dst.h>
 #include <net/genetlink.h>
+#include <net/ip.h>
+#include <net/route.h>
 #include <net/sock.h>
 
 #include "quicktun.h"
@@ -54,73 +56,117 @@ struct send_work {
 };
 
 
-static void quicktun_send_bh(struct work_struct *work)
+static netdev_tx_t quicktun_net_xmit(struct sk_buff *skb, struct net_device *dev)
 {
-	struct send_work *send_work = container_of(work, struct send_work, work);
-	struct sk_buff *skb = send_work->skb;
-	struct quicktun_struct *tun = send_work->quicktun;
+	struct quicktun_struct *tun = netdev_priv(dev);
 	struct addr_struct *remote_addr;
-	struct msghdr msg;
-	struct kvec vec;
+	struct rtable *rt = NULL;
+	struct iphdr *iph;
+	struct udphdr *udph;
+	__wsum  csum;
+	unsigned int len;
 	int err;
-	
+
 	rcu_read_lock();
 
 	remote_addr = rcu_dereference(tun->remote_address);
 
 	if (likely(remote_addr->addr.sin_addr.s_addr)) {
-		err = skb_linearize(skb);
-		if (unlikely(err < 0))
+		struct flowi fl = {
+			.nl_u = {
+				.ip4_u = {
+					.saddr  = tun->local_address.sin_addr.s_addr,
+					.daddr  = remote_addr->addr.sin_addr.s_addr,
+					.tos    = 0,
+				}
+			},
+			.proto  = IPPROTO_UDP
+		};
+
+		if (ip_route_output_key(dev_net(dev), &rt, &fl)) {
+			tun->dev->stats.tx_carrier_errors++;
 			goto error;
+		}
 
-		vec.iov_base = skb->data;
-		vec.iov_len = skb->len;
+		if (rt->dst.dev == dev) {
+			tun->dev->stats.collisions++;
+			goto error;
+		}
 
-		msg.msg_name = &remote_addr->addr;
-		msg.msg_namelen = sizeof(remote_addr->addr);
-		msg.msg_control = NULL;
-		msg.msg_controllen = 0;
-		msg.msg_flags = MSG_DONTWAIT;
+		if (skb_headroom(skb) < (LL_RESERVED_SPACE(rt->dst.dev) + sizeof(struct iphdr) + sizeof(struct udphdr)) || skb_shared(skb) || (skb_cloned(skb) && !skb_clone_writable(skb, 0))) {
+			struct sk_buff *new_skb = skb_realloc_headroom(skb, sizeof(struct iphdr) + sizeof(struct udphdr));
+			if (!new_skb) {
+				tun->dev->stats.tx_dropped++;
+				goto error;
+			}
+
+			dev_kfree_skb(skb);
+			skb = new_skb;
+		}
 
-		kernel_sendmsg(tun->sock, &msg, &vec, 1, skb->len);
+		len = skb->len;
 
-		tun->dev->stats.tx_packets++;
-		tun->dev->stats.tx_bytes += skb->len;
+		skb_push(skb, sizeof(struct iphdr) + sizeof(struct udphdr));
+		skb_reset_network_header(skb);
+
+		skb_dst_drop(skb);
+		skb_dst_set(skb, &rt->dst);
+
+		iph           = (struct iphdr *)skb->data;
+		iph->version  = 4;
+		iph->ihl      = sizeof(struct iphdr)>>2;
+		iph->frag_off = 0;
+		iph->protocol = IPPROTO_UDP;
+		iph->tos      = 0;
+		iph->saddr    = rt->rt_src;
+		iph->daddr    = rt->rt_dst;
+		iph->tot_len  = htons(len + sizeof(struct iphdr) + sizeof(struct udphdr));
+		iph->ttl      = 64;
+		iph->check    = 0;
+
+		udph          = (struct udphdr *)(skb->data + sizeof(struct iphdr));
+		udph->source  = tun->local_address.sin_port;
+		udph->dest    = remote_addr->addr.sin_port;
+		udph->len     = htons(len + sizeof(struct udphdr));
+		udph->check   = 0;
+
+		ip_select_ident(ip_hdr(skb), &rt->dst, NULL);
+
+		csum = csum_partial(udph, ntohs(udph->len), 0);
+		udph->check = csum_tcpudp_magic(iph->saddr, iph->daddr, ntohs(udph->len), IPPROTO_UDP, csum);
+		skb->ip_summed = CHECKSUM_NONE;
+
+		nf_reset(skb);
+		skb->mark = 0;
+		err = ip_local_out(skb);
+
+		if (likely(net_xmit_eval(err) == 0)) {
+			tun->dev->stats.tx_bytes += len;
+			tun->dev->stats.tx_packets++;
+		} else {
+			tun->dev->stats.tx_errors++;
+			tun->dev->stats.tx_aborted_errors++;
+		}
 	}
-	else
+	else {
+		tun->dev->stats.tx_carrier_errors++;
 		goto error;
+	}
 
 	rcu_read_unlock();
 
-	consume_skb(skb);
-	kfree(send_work);
-
-	return;
+	return NETDEV_TX_OK;
 
 error:
 	rcu_read_unlock();
+	dev_kfree_skb(skb);
 
-	kfree_skb(skb);
-	kfree(send_work);
-
-	tun->dev->stats.tx_errors++;
-}
-
-
-static netdev_tx_t quicktun_net_xmit(struct sk_buff *skb, struct net_device *dev)
-{
-	struct send_work *work = kmalloc(sizeof(struct send_work), GFP_ATOMIC);
-	if (likely(work)) {
-		INIT_WORK(&work->work, quicktun_send_bh);
-		work->quicktun = netdev_priv(dev);
-		work->skb = skb;
-		schedule_work(&work->work);
-	}
+	if (rt)
+		ip_rt_put(rt);
 
 	return NETDEV_TX_OK;
 }
 
-
 static void free_addr_struct(struct rcu_head *rcu)
 {
 	struct addr_struct *addr = container_of(rcu, struct addr_struct, rcu);
@@ -156,6 +202,7 @@ static void quicktun_udp_data_ready(struct sock *sk, int bytes)
 	}
 
 	skb_orphan(skb);
+	nf_reset(skb);
 
 	err = skb_linearize(skb);
 	if (err < 0)
@@ -230,7 +277,7 @@ static void quicktun_udp_data_ready(struct sock *sk, int bytes)
 
 	read_unlock(&sk->sk_callback_lock);
 
-	netif_rx(skb);
+	netif_rx_ni(skb);
 
 	tun->dev->stats.rx_packets++;
 	tun->dev->stats.rx_bytes += len;
@@ -320,6 +367,9 @@ static void quicktun_net_init(struct net_device *dev)
 
 		ether_setup(dev);
 
+		dev->hard_header_len = LL_MAX_HEADER + sizeof(struct iphdr) + sizeof(struct udphdr) + ETH_HLEN;
+		dev->mtu = ETH_DATA_LEN - sizeof(struct iphdr) - sizeof(struct udphdr) - ETH_HLEN;
+
 		random_ether_addr(dev->dev_addr);
 		break;
 
@@ -327,9 +377,9 @@ static void quicktun_net_init(struct net_device *dev)
 		dev->netdev_ops = &quicktun_ip_netdev_ops;
 
 		dev->destructor = free_netdev;
-		dev->hard_header_len = 0;
-		dev->addr_len = 0;
-		dev->mtu = 1500;
+		dev->addr_len = 4;
+		dev->hard_header_len = LL_MAX_HEADER + sizeof(struct iphdr) + sizeof(struct udphdr);
+		dev->mtu = ETH_DATA_LEN - sizeof(struct iphdr) - sizeof(struct udphdr);
 
 		dev->type = ARPHRD_NONE;
 		dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST;