]> Gentwo Git Trees - linux/.git/commitdiff
netfilter: flowtable: Add IPIP rx sw acceleration
authorLorenzo Bianconi <lorenzo@kernel.org>
Fri, 7 Nov 2025 11:14:46 +0000 (12:14 +0100)
committerPablo Neira Ayuso <pablo@netfilter.org>
Fri, 28 Nov 2025 00:00:38 +0000 (00:00 +0000)
Introduce sw acceleration for rx path of IPIP tunnels relying on the
netfilter flowtable infrastructure. Subsequent patches will add sw
acceleration for IPIP tunnels tx path.
This series introduces basic infrastructure to accelerate other tunnel
types (e.g. IP6IP6).
IPIP rx sw acceleration can be tested running the following scenario where
the traffic is forwarded between two NICs (eth0 and eth1) and an IPIP
tunnel is used to access a remote site (using eth1 as the underlay device):

ETH0 -- TUN0 <==> ETH1 -- [IP network] -- TUN1 (192.168.100.2)

$ip addr show
6: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default qlen 1000
    link/ether 00:00:22:33:11:55 brd ff:ff:ff:ff:ff:ff
    inet 192.168.0.2/24 scope global eth0
       valid_lft forever preferred_lft forever
7: eth1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default qlen 1000
    link/ether 00:11:22:33:11:55 brd ff:ff:ff:ff:ff:ff
    inet 192.168.1.1/24 scope global eth1
       valid_lft forever preferred_lft forever
8: tun0@NONE: <POINTOPOINT,NOARP,UP,LOWER_UP> mtu 1480 qdisc noqueue state UNKNOWN group default qlen 1000
    link/ipip 192.168.1.1 peer 192.168.1.2
    inet 192.168.100.1/24 scope global tun0
       valid_lft forever preferred_lft forever

$ip route show
default via 192.168.100.2 dev tun0
192.168.0.0/24 dev eth0 proto kernel scope link src 192.168.0.2
192.168.1.0/24 dev eth1 proto kernel scope link src 192.168.1.1
192.168.100.0/24 dev tun0 proto kernel scope link src 192.168.100.1

$nft list ruleset
table inet filter {
        flowtable ft {
                hook ingress priority filter
                devices = { eth0, eth1 }
        }

        chain forward {
                type filter hook forward priority filter; policy accept;
                meta l4proto { tcp, udp } flow add @ft
        }
}

Reproducing the scenario described above using veths I got the following
results:
- TCP stream received from the IPIP tunnel:
  - net-next: (baseline) ~ 71Gbps
  - net-next + IPIP flowtbale support: ~101Gbps

Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
include/linux/netdevice.h
include/net/netfilter/nf_flow_table.h
net/ipv4/ipip.c
net/netfilter/nf_flow_table_core.c
net/netfilter/nf_flow_table_ip.c
net/netfilter/nf_flow_table_path.c

index e808071dbb7d3770decb3d8202c381836f1f1aa6..bf99fe8622da95b07701e3a224abef41667a0d6a 100644 (file)
@@ -877,6 +877,7 @@ enum net_device_path_type {
        DEV_PATH_PPPOE,
        DEV_PATH_DSA,
        DEV_PATH_MTK_WDMA,
+       DEV_PATH_TUN,
 };
 
 struct net_device_path {
@@ -888,6 +889,18 @@ struct net_device_path {
                        __be16          proto;
                        u8              h_dest[ETH_ALEN];
                } encap;
+               struct {
+                       union {
+                               struct in_addr  src_v4;
+                               struct in6_addr src_v6;
+                       };
+                       union {
+                               struct in_addr  dst_v4;
+                               struct in6_addr dst_v6;
+                       };
+
+                       u8      l3_proto;
+               } tun;
                struct {
                        enum {
                                DEV_PATH_BR_VLAN_KEEP,
index f7306276ece7141210d7928b05f2a56b43b2cd0b..b09c11c048d519d3254a80fbc0186c0049ad61f0 100644 (file)
@@ -107,6 +107,19 @@ enum flow_offload_xmit_type {
 
 #define NF_FLOW_TABLE_ENCAP_MAX                2
 
+struct flow_offload_tunnel {
+       union {
+               struct in_addr  src_v4;
+               struct in6_addr src_v6;
+       };
+       union {
+               struct in_addr  dst_v4;
+               struct in6_addr dst_v6;
+       };
+
+       u8      l3_proto;
+};
+
 struct flow_offload_tuple {
        union {
                struct in_addr          src_v4;
@@ -130,12 +143,15 @@ struct flow_offload_tuple {
                __be16                  proto;
        } encap[NF_FLOW_TABLE_ENCAP_MAX];
 
+       struct flow_offload_tunnel      tun;
+
        /* All members above are keys for lookups, see flow_offload_hash(). */
        struct { }                      __hash;
 
        u8                              dir:2,
                                        xmit_type:3,
                                        encap_num:2,
+                                       tun_num:2,
                                        in_vlan_ingress:2;
        u16                             mtu;
        union {
@@ -206,7 +222,9 @@ struct nf_flow_route {
                                u16             id;
                                __be16          proto;
                        } encap[NF_FLOW_TABLE_ENCAP_MAX];
+                       struct flow_offload_tunnel tun;
                        u8                      num_encaps:2,
+                                               num_tuns:2,
                                                ingress_vlans:2;
                } in;
                struct {
index 3e03af073a1ccc3d7597a998a515b6cfdded40b5..ff95b1b9908e9f4ba4bff207a5bd2c5d5670215a 100644 (file)
@@ -353,6 +353,30 @@ ipip_tunnel_ctl(struct net_device *dev, struct ip_tunnel_parm_kern *p, int cmd)
        return ip_tunnel_ctl(dev, p, cmd);
 }
 
+static int ipip_fill_forward_path(struct net_device_path_ctx *ctx,
+                                 struct net_device_path *path)
+{
+       struct ip_tunnel *tunnel = netdev_priv(ctx->dev);
+       const struct iphdr *tiph = &tunnel->parms.iph;
+       struct rtable *rt;
+
+       rt = ip_route_output(dev_net(ctx->dev), tiph->daddr, 0, 0, 0,
+                            RT_SCOPE_UNIVERSE);
+       if (IS_ERR(rt))
+               return PTR_ERR(rt);
+
+       path->type = DEV_PATH_TUN;
+       path->tun.src_v4.s_addr = tiph->saddr;
+       path->tun.dst_v4.s_addr = tiph->daddr;
+       path->tun.l3_proto = IPPROTO_IPIP;
+       path->dev = ctx->dev;
+
+       ctx->dev = rt->dst.dev;
+       ip_rt_put(rt);
+
+       return 0;
+}
+
 static const struct net_device_ops ipip_netdev_ops = {
        .ndo_init       = ipip_tunnel_init,
        .ndo_uninit     = ip_tunnel_uninit,
@@ -362,6 +386,7 @@ static const struct net_device_ops ipip_netdev_ops = {
        .ndo_get_stats64 = dev_get_tstats64,
        .ndo_get_iflink = ip_tunnel_get_iflink,
        .ndo_tunnel_ctl = ipip_tunnel_ctl,
+       .ndo_fill_forward_path = ipip_fill_forward_path,
 };
 
 #define IPIP_FEATURES (NETIF_F_SG |            \
index 6c6a5165f9933b130e772c096a01d3fcb1d61c94..06e8251a6644e2335117ae3720b9ebf159eb2264 100644 (file)
@@ -118,7 +118,10 @@ static int flow_offload_fill_route(struct flow_offload *flow,
                        flow_tuple->in_vlan_ingress |= BIT(j);
                j++;
        }
+
+       flow_tuple->tun = route->tuple[dir].in.tun;
        flow_tuple->encap_num = route->tuple[dir].in.num_encaps;
+       flow_tuple->tun_num = route->tuple[dir].in.num_tuns;
 
        switch (route->tuple[dir].xmit_type) {
        case FLOW_OFFLOAD_XMIT_DIRECT:
index 083ceb64ac17a88ecaa0dc782e1089f77e0210f3..d6a1f0cda189205678d47ed05441fef28638bfee 100644 (file)
@@ -145,8 +145,11 @@ static bool ip_has_options(unsigned int thoff)
 static void nf_flow_tuple_encap(struct sk_buff *skb,
                                struct flow_offload_tuple *tuple)
 {
+       __be16 inner_proto = skb->protocol;
        struct vlan_ethhdr *veth;
        struct pppoe_hdr *phdr;
+       struct iphdr *iph;
+       u16 offset = 0;
        int i = 0;
 
        if (skb_vlan_tag_present(skb)) {
@@ -159,13 +162,26 @@ static void nf_flow_tuple_encap(struct sk_buff *skb,
                veth = (struct vlan_ethhdr *)skb_mac_header(skb);
                tuple->encap[i].id = ntohs(veth->h_vlan_TCI);
                tuple->encap[i].proto = skb->protocol;
+               inner_proto = veth->h_vlan_encapsulated_proto;
+               offset += VLAN_HLEN;
                break;
        case htons(ETH_P_PPP_SES):
                phdr = (struct pppoe_hdr *)skb_network_header(skb);
                tuple->encap[i].id = ntohs(phdr->sid);
                tuple->encap[i].proto = skb->protocol;
+               inner_proto = *((__be16 *)(phdr + 1));
+               offset += PPPOE_SES_HLEN;
                break;
        }
+
+       if (inner_proto == htons(ETH_P_IP)) {
+               iph = (struct iphdr *)(skb_network_header(skb) + offset);
+               if (iph->protocol == IPPROTO_IPIP) {
+                       tuple->tun.dst_v4.s_addr = iph->daddr;
+                       tuple->tun.src_v4.s_addr = iph->saddr;
+                       tuple->tun.l3_proto = IPPROTO_IPIP;
+               }
+       }
 }
 
 struct nf_flowtable_ctx {
@@ -277,11 +293,46 @@ static unsigned int nf_flow_xmit_xfrm(struct sk_buff *skb,
        return NF_STOLEN;
 }
 
+static bool nf_flow_ip4_tunnel_proto(struct sk_buff *skb, u32 *psize)
+{
+       struct iphdr *iph;
+       u16 size;
+
+       if (!pskb_may_pull(skb, sizeof(*iph) + *psize))
+               return false;
+
+       iph = (struct iphdr *)(skb_network_header(skb) + *psize);
+       size = iph->ihl << 2;
+
+       if (ip_is_fragment(iph) || unlikely(ip_has_options(size)))
+               return false;
+
+       if (iph->ttl <= 1)
+               return false;
+
+       if (iph->protocol == IPPROTO_IPIP)
+               *psize += size;
+
+       return true;
+}
+
+static void nf_flow_ip4_tunnel_pop(struct sk_buff *skb)
+{
+       struct iphdr *iph = (struct iphdr *)skb_network_header(skb);
+
+       if (iph->protocol != IPPROTO_IPIP)
+               return;
+
+       skb_pull(skb, iph->ihl << 2);
+       skb_reset_network_header(skb);
+}
+
 static bool nf_flow_skb_encap_protocol(struct sk_buff *skb, __be16 proto,
                                       u32 *offset)
 {
+       __be16 inner_proto = skb->protocol;
        struct vlan_ethhdr *veth;
-       __be16 inner_proto;
+       bool ret = false;
 
        switch (skb->protocol) {
        case htons(ETH_P_8021Q):
@@ -291,19 +342,23 @@ static bool nf_flow_skb_encap_protocol(struct sk_buff *skb, __be16 proto,
                veth = (struct vlan_ethhdr *)skb_mac_header(skb);
                if (veth->h_vlan_encapsulated_proto == proto) {
                        *offset += VLAN_HLEN;
-                       return true;
+                       inner_proto = proto;
+                       ret = true;
                }
                break;
        case htons(ETH_P_PPP_SES):
                if (nf_flow_pppoe_proto(skb, &inner_proto) &&
                    inner_proto == proto) {
                        *offset += PPPOE_SES_HLEN;
-                       return true;
+                       ret = true;
                }
                break;
        }
 
-       return false;
+       if (inner_proto == htons(ETH_P_IP))
+               ret = nf_flow_ip4_tunnel_proto(skb, offset);
+
+       return ret;
 }
 
 static void nf_flow_encap_pop(struct sk_buff *skb,
@@ -331,6 +386,9 @@ static void nf_flow_encap_pop(struct sk_buff *skb,
                        break;
                }
        }
+
+       if (skb->protocol == htons(ETH_P_IP))
+               nf_flow_ip4_tunnel_pop(skb);
 }
 
 struct nf_flow_xmit {
@@ -356,8 +414,7 @@ nf_flow_offload_lookup(struct nf_flowtable_ctx *ctx,
 {
        struct flow_offload_tuple tuple = {};
 
-       if (skb->protocol != htons(ETH_P_IP) &&
-           !nf_flow_skb_encap_protocol(skb, htons(ETH_P_IP), &ctx->offset))
+       if (!nf_flow_skb_encap_protocol(skb, htons(ETH_P_IP), &ctx->offset))
                return NULL;
 
        if (nf_flow_tuple_ip(ctx, skb, &tuple) < 0)
index eb9b33a1873a0a6b32c0184716615f02a6d1af1c..73717cc32df5cb0415fc0f13f69d0af511db1f3b 100644 (file)
@@ -80,6 +80,8 @@ struct nft_forward_info {
                __be16  proto;
        } encap[NF_FLOW_TABLE_ENCAP_MAX];
        u8 num_encaps;
+       struct flow_offload_tunnel tun;
+       u8 num_tuns;
        u8 ingress_vlans;
        u8 h_source[ETH_ALEN];
        u8 h_dest[ETH_ALEN];
@@ -102,6 +104,7 @@ static void nft_dev_path_info(const struct net_device_path_stack *stack,
                case DEV_PATH_DSA:
                case DEV_PATH_VLAN:
                case DEV_PATH_PPPOE:
+               case DEV_PATH_TUN:
                        info->indev = path->dev;
                        if (is_zero_ether_addr(info->h_source))
                                memcpy(info->h_source, path->dev->dev_addr, ETH_ALEN);
@@ -113,14 +116,27 @@ static void nft_dev_path_info(const struct net_device_path_stack *stack,
                                break;
                        }
 
-                       /* DEV_PATH_VLAN and DEV_PATH_PPPOE */
-                       if (info->num_encaps >= NF_FLOW_TABLE_ENCAP_MAX) {
-                               info->indev = NULL;
-                               break;
+                       /* DEV_PATH_VLAN, DEV_PATH_PPPOE and DEV_PATH_TUN */
+                       if (path->type == DEV_PATH_TUN) {
+                               if (info->num_tuns) {
+                                       info->indev = NULL;
+                                       break;
+                               }
+                               info->tun.src_v6 = path->tun.src_v6;
+                               info->tun.dst_v6 = path->tun.dst_v6;
+                               info->tun.l3_proto = path->tun.l3_proto;
+                               info->num_tuns++;
+                       } else {
+                               if (info->num_encaps >= NF_FLOW_TABLE_ENCAP_MAX) {
+                                       info->indev = NULL;
+                                       break;
+                               }
+                               info->encap[info->num_encaps].id =
+                                       path->encap.id;
+                               info->encap[info->num_encaps].proto =
+                                       path->encap.proto;
+                               info->num_encaps++;
                        }
-                       info->encap[info->num_encaps].id = path->encap.id;
-                       info->encap[info->num_encaps].proto = path->encap.proto;
-                       info->num_encaps++;
                        if (path->type == DEV_PATH_PPPOE)
                                memcpy(info->h_dest, path->encap.h_dest, ETH_ALEN);
                        break;
@@ -203,6 +219,14 @@ static void nft_dev_forward_path(struct nf_flow_route *route,
                route->tuple[!dir].in.encap[i].id = info.encap[i].id;
                route->tuple[!dir].in.encap[i].proto = info.encap[i].proto;
        }
+
+       if (info.num_tuns) {
+               route->tuple[!dir].in.tun.src_v6 = info.tun.dst_v6;
+               route->tuple[!dir].in.tun.dst_v6 = info.tun.src_v6;
+               route->tuple[!dir].in.tun.l3_proto = info.tun.l3_proto;
+               route->tuple[!dir].in.num_tuns = info.num_tuns;
+       }
+
        route->tuple[!dir].in.num_encaps = info.num_encaps;
        route->tuple[!dir].in.ingress_vlans = info.ingress_vlans;
        route->tuple[dir].out.ifindex = info.outdev->ifindex;