1c9422999SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 20e6fbc5bSPravin B Shelar /* 30e6fbc5bSPravin B Shelar * Copyright (c) 2013 Nicira, Inc. 40e6fbc5bSPravin B Shelar */ 50e6fbc5bSPravin B Shelar 60e6fbc5bSPravin B Shelar #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 70e6fbc5bSPravin B Shelar 80e6fbc5bSPravin B Shelar #include <linux/types.h> 90e6fbc5bSPravin B Shelar #include <linux/kernel.h> 100e6fbc5bSPravin B Shelar #include <linux/skbuff.h> 110e6fbc5bSPravin B Shelar #include <linux/netdevice.h> 120e6fbc5bSPravin B Shelar #include <linux/in.h> 130e6fbc5bSPravin B Shelar #include <linux/if_arp.h> 140e6fbc5bSPravin B Shelar #include <linux/init.h> 150e6fbc5bSPravin B Shelar #include <linux/in6.h> 160e6fbc5bSPravin B Shelar #include <linux/inetdevice.h> 170e6fbc5bSPravin B Shelar #include <linux/netfilter_ipv4.h> 180e6fbc5bSPravin B Shelar #include <linux/etherdevice.h> 190e6fbc5bSPravin B Shelar #include <linux/if_ether.h> 200e6fbc5bSPravin B Shelar #include <linux/if_vlan.h> 21e7030878SThomas Graf #include <linux/static_key.h> 220e6fbc5bSPravin B Shelar 230e6fbc5bSPravin B Shelar #include <net/ip.h> 240e6fbc5bSPravin B Shelar #include <net/icmp.h> 250e6fbc5bSPravin B Shelar #include <net/protocol.h> 260e6fbc5bSPravin B Shelar #include <net/ip_tunnels.h> 27058214a4STom Herbert #include <net/ip6_tunnel.h> 280e6fbc5bSPravin B Shelar #include <net/arp.h> 290e6fbc5bSPravin B Shelar #include <net/checksum.h> 300e6fbc5bSPravin B Shelar #include <net/dsfield.h> 310e6fbc5bSPravin B Shelar #include <net/inet_ecn.h> 320e6fbc5bSPravin B Shelar #include <net/xfrm.h> 330e6fbc5bSPravin B Shelar #include <net/net_namespace.h> 340e6fbc5bSPravin B Shelar #include <net/netns/generic.h> 350e6fbc5bSPravin B Shelar #include <net/rtnetlink.h> 3663d008a4SJiri Benc #include <net/dst_metadata.h> 374ece4778SXin Long #include <net/geneve.h> 38edf31cbbSXin Long #include <net/vxlan.h> 39b0a21810SXin Long #include <net/erspan.h> 400e6fbc5bSPravin B Shelar 4155c2bc14STom Herbert const struct ip_tunnel_encap_ops __rcu * 4255c2bc14STom Herbert iptun_encaps[MAX_IPTUN_ENCAP_OPS] __read_mostly; 4355c2bc14STom Herbert EXPORT_SYMBOL(iptun_encaps); 4455c2bc14STom Herbert 45058214a4STom Herbert const struct ip6_tnl_encap_ops __rcu * 46058214a4STom Herbert ip6tun_encaps[MAX_IPTUN_ENCAP_OPS] __read_mostly; 47058214a4STom Herbert EXPORT_SYMBOL(ip6tun_encaps); 48058214a4STom Herbert 49039f5062SPravin B Shelar void iptunnel_xmit(struct sock *sk, struct rtable *rt, struct sk_buff *skb, 500e6fbc5bSPravin B Shelar __be32 src, __be32 dst, __u8 proto, 51963a88b3SNicolas Dichtel __u8 tos, __u8 ttl, __be16 df, bool xnet) 520e6fbc5bSPravin B Shelar { 53bc22a0e2SNicolas Dichtel int pkt_len = skb->len - skb_inner_network_offset(skb); 54f859b0f6SEric W. Biederman struct net *net = dev_net(rt->dst.dev); 55039f5062SPravin B Shelar struct net_device *dev = skb->dev; 560e6fbc5bSPravin B Shelar struct iphdr *iph; 570e6fbc5bSPravin B Shelar int err; 580e6fbc5bSPravin B Shelar 59963a88b3SNicolas Dichtel skb_scrub_packet(skb, xnet); 60963a88b3SNicolas Dichtel 61bf8d85d4SEric Dumazet skb_clear_hash_if_not_l4(skb); 620e6fbc5bSPravin B Shelar skb_dst_set(skb, &rt->dst); 630e6fbc5bSPravin B Shelar memset(IPCB(skb), 0, sizeof(*IPCB(skb))); 640e6fbc5bSPravin B Shelar 650e6fbc5bSPravin B Shelar /* Push down and install the IP header. */ 6678a3694dSSteffen Klassert skb_push(skb, sizeof(struct iphdr)); 670e6fbc5bSPravin B Shelar skb_reset_network_header(skb); 680e6fbc5bSPravin B Shelar 690e6fbc5bSPravin B Shelar iph = ip_hdr(skb); 700e6fbc5bSPravin B Shelar 710e6fbc5bSPravin B Shelar iph->version = 4; 720e6fbc5bSPravin B Shelar iph->ihl = sizeof(struct iphdr) >> 2; 7316f7eb2bSSabrina Dubroca iph->frag_off = ip_mtu_locked(&rt->dst) ? 0 : df; 740e6fbc5bSPravin B Shelar iph->protocol = proto; 750e6fbc5bSPravin B Shelar iph->tos = tos; 760e6fbc5bSPravin B Shelar iph->daddr = dst; 770e6fbc5bSPravin B Shelar iph->saddr = src; 780e6fbc5bSPravin B Shelar iph->ttl = ttl; 79f859b0f6SEric W. Biederman __ip_select_ident(net, iph, skb_shinfo(skb)->gso_segs ?: 1); 800e6fbc5bSPravin B Shelar 8133224b16SEric W. Biederman err = ip_local_out(net, sk, skb); 825684abf7SXin Long 835684abf7SXin Long if (dev) { 840e6fbc5bSPravin B Shelar if (unlikely(net_xmit_eval(err))) 850e6fbc5bSPravin B Shelar pkt_len = 0; 86039f5062SPravin B Shelar iptunnel_xmit_stats(dev, pkt_len); 870e6fbc5bSPravin B Shelar } 885684abf7SXin Long } 890e6fbc5bSPravin B Shelar EXPORT_SYMBOL_GPL(iptunnel_xmit); 903d7b46cdSPravin B Shelar 91a6d5bbf3SJiri Benc int __iptunnel_pull_header(struct sk_buff *skb, int hdr_len, 92a6d5bbf3SJiri Benc __be16 inner_proto, bool raw_proto, bool xnet) 933d7b46cdSPravin B Shelar { 943d7b46cdSPravin B Shelar if (unlikely(!pskb_may_pull(skb, hdr_len))) 953d7b46cdSPravin B Shelar return -ENOMEM; 963d7b46cdSPravin B Shelar 973d7b46cdSPravin B Shelar skb_pull_rcsum(skb, hdr_len); 983d7b46cdSPravin B Shelar 99a6d5bbf3SJiri Benc if (!raw_proto && inner_proto == htons(ETH_P_TEB)) { 1001245dfc8SLi RongQing struct ethhdr *eh; 1013d7b46cdSPravin B Shelar 1023d7b46cdSPravin B Shelar if (unlikely(!pskb_may_pull(skb, ETH_HLEN))) 1033d7b46cdSPravin B Shelar return -ENOMEM; 1043d7b46cdSPravin B Shelar 1051245dfc8SLi RongQing eh = (struct ethhdr *)skb->data; 106d181ddcaSAlexander Duyck if (likely(eth_proto_is_802_3(eh->h_proto))) 1073d7b46cdSPravin B Shelar skb->protocol = eh->h_proto; 1083d7b46cdSPravin B Shelar else 1093d7b46cdSPravin B Shelar skb->protocol = htons(ETH_P_802_2); 1103d7b46cdSPravin B Shelar 1113d7b46cdSPravin B Shelar } else { 1123d7b46cdSPravin B Shelar skb->protocol = inner_proto; 1133d7b46cdSPravin B Shelar } 1143d7b46cdSPravin B Shelar 1157539fadcSTom Herbert skb_clear_hash_if_not_l4(skb); 1163e2ed0c2SMichał Mirosław __vlan_hwaccel_clear_tag(skb); 1173d7b46cdSPravin B Shelar skb_set_queue_mapping(skb, 0); 1187f290c94SJiri Benc skb_scrub_packet(skb, xnet); 119a09a4c8dSJesse Gross 120a09a4c8dSJesse Gross return iptunnel_pull_offloads(skb); 1213d7b46cdSPravin B Shelar } 122a6d5bbf3SJiri Benc EXPORT_SYMBOL_GPL(__iptunnel_pull_header); 1232d26f0a3SEric Dumazet 12463d008a4SJiri Benc struct metadata_dst *iptunnel_metadata_reply(struct metadata_dst *md, 12563d008a4SJiri Benc gfp_t flags) 12663d008a4SJiri Benc { 12763d008a4SJiri Benc struct metadata_dst *res; 12863d008a4SJiri Benc struct ip_tunnel_info *dst, *src; 12963d008a4SJiri Benc 1303fcece12SJakub Kicinski if (!md || md->type != METADATA_IP_TUNNEL || 1313fcece12SJakub Kicinski md->u.tun_info.mode & IP_TUNNEL_INFO_TX) 13263d008a4SJiri Benc return NULL; 13363d008a4SJiri Benc 134f52f11ecSXin Long src = &md->u.tun_info; 135f52f11ecSXin Long res = metadata_dst_alloc(src->options_len, METADATA_IP_TUNNEL, flags); 13663d008a4SJiri Benc if (!res) 13763d008a4SJiri Benc return NULL; 13863d008a4SJiri Benc 13963d008a4SJiri Benc dst = &res->u.tun_info; 14063d008a4SJiri Benc dst->key.tun_id = src->key.tun_id; 14163d008a4SJiri Benc if (src->mode & IP_TUNNEL_INFO_IPV6) 14263d008a4SJiri Benc memcpy(&dst->key.u.ipv6.dst, &src->key.u.ipv6.src, 14363d008a4SJiri Benc sizeof(struct in6_addr)); 14463d008a4SJiri Benc else 14563d008a4SJiri Benc dst->key.u.ipv4.dst = src->key.u.ipv4.src; 1467bdca378Swenxu dst->key.tun_flags = src->key.tun_flags; 14763d008a4SJiri Benc dst->mode = src->mode | IP_TUNNEL_INFO_TX; 148f52f11ecSXin Long ip_tunnel_info_opts_set(dst, ip_tunnel_info_opts(src), 149f52f11ecSXin Long src->options_len, 0); 15063d008a4SJiri Benc 15163d008a4SJiri Benc return res; 15263d008a4SJiri Benc } 15363d008a4SJiri Benc EXPORT_SYMBOL_GPL(iptunnel_metadata_reply); 15463d008a4SJiri Benc 155aed069dfSAlexander Duyck int iptunnel_handle_offloads(struct sk_buff *skb, 1562d26f0a3SEric Dumazet int gso_type_mask) 1572d26f0a3SEric Dumazet { 1582d26f0a3SEric Dumazet int err; 1592d26f0a3SEric Dumazet 1602d26f0a3SEric Dumazet if (likely(!skb->encapsulation)) { 1612d26f0a3SEric Dumazet skb_reset_inner_headers(skb); 1622d26f0a3SEric Dumazet skb->encapsulation = 1; 1632d26f0a3SEric Dumazet } 1642d26f0a3SEric Dumazet 1652d26f0a3SEric Dumazet if (skb_is_gso(skb)) { 1669580bf2eSEric Dumazet err = skb_header_unclone(skb, GFP_ATOMIC); 1672d26f0a3SEric Dumazet if (unlikely(err)) 168aed069dfSAlexander Duyck return err; 1692d26f0a3SEric Dumazet skb_shinfo(skb)->gso_type |= gso_type_mask; 170aed069dfSAlexander Duyck return 0; 1712d26f0a3SEric Dumazet } 1722d26f0a3SEric Dumazet 1736fa79666SEdward Cree if (skb->ip_summed != CHECKSUM_PARTIAL) { 1742d26f0a3SEric Dumazet skb->ip_summed = CHECKSUM_NONE; 1756fa79666SEdward Cree /* We clear encapsulation here to prevent badly-written 1766fa79666SEdward Cree * drivers potentially deciding to offload an inner checksum 1776fa79666SEdward Cree * if we set CHECKSUM_PARTIAL on the outer header. 1786fa79666SEdward Cree * This should go away when the drivers are all fixed. 1796fa79666SEdward Cree */ 180179bc67fSEdward Cree skb->encapsulation = 0; 181179bc67fSEdward Cree } 1822d26f0a3SEric Dumazet 183aed069dfSAlexander Duyck return 0; 1842d26f0a3SEric Dumazet } 1852d26f0a3SEric Dumazet EXPORT_SYMBOL_GPL(iptunnel_handle_offloads); 186ebe44f35SDavid S. Miller 187ebe44f35SDavid S. Miller /* Often modified stats are per cpu, other are shared (netdev->stats) */ 188bc1f4470Sstephen hemminger void ip_tunnel_get_stats64(struct net_device *dev, 189ebe44f35SDavid S. Miller struct rtnl_link_stats64 *tot) 190ebe44f35SDavid S. Miller { 191ebe44f35SDavid S. Miller int i; 192ebe44f35SDavid S. Miller 193c24a5964SAlexander Duyck netdev_stats_to_stats64(tot, &dev->stats); 194c24a5964SAlexander Duyck 195ebe44f35SDavid S. Miller for_each_possible_cpu(i) { 196ebe44f35SDavid S. Miller const struct pcpu_sw_netstats *tstats = 197ebe44f35SDavid S. Miller per_cpu_ptr(dev->tstats, i); 198ebe44f35SDavid S. Miller u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 199ebe44f35SDavid S. Miller unsigned int start; 200ebe44f35SDavid S. Miller 201ebe44f35SDavid S. Miller do { 20257a7744eSEric W. Biederman start = u64_stats_fetch_begin_irq(&tstats->syncp); 203ebe44f35SDavid S. Miller rx_packets = tstats->rx_packets; 204ebe44f35SDavid S. Miller tx_packets = tstats->tx_packets; 205ebe44f35SDavid S. Miller rx_bytes = tstats->rx_bytes; 206ebe44f35SDavid S. Miller tx_bytes = tstats->tx_bytes; 20757a7744eSEric W. Biederman } while (u64_stats_fetch_retry_irq(&tstats->syncp, start)); 208ebe44f35SDavid S. Miller 209ebe44f35SDavid S. Miller tot->rx_packets += rx_packets; 210ebe44f35SDavid S. Miller tot->tx_packets += tx_packets; 211ebe44f35SDavid S. Miller tot->rx_bytes += rx_bytes; 212ebe44f35SDavid S. Miller tot->tx_bytes += tx_bytes; 213ebe44f35SDavid S. Miller } 214ebe44f35SDavid S. Miller } 215ebe44f35SDavid S. Miller EXPORT_SYMBOL_GPL(ip_tunnel_get_stats64); 2163093fbe7SThomas Graf 217a1c234f9SJiri Benc static const struct nla_policy ip_tun_policy[LWTUNNEL_IP_MAX + 1] = { 2187b6a70f7SXin Long [LWTUNNEL_IP_UNSPEC] = { .strict_start_type = LWTUNNEL_IP_OPTS }, 219a1c234f9SJiri Benc [LWTUNNEL_IP_ID] = { .type = NLA_U64 }, 220a1c234f9SJiri Benc [LWTUNNEL_IP_DST] = { .type = NLA_U32 }, 221a1c234f9SJiri Benc [LWTUNNEL_IP_SRC] = { .type = NLA_U32 }, 222a1c234f9SJiri Benc [LWTUNNEL_IP_TTL] = { .type = NLA_U8 }, 223a1c234f9SJiri Benc [LWTUNNEL_IP_TOS] = { .type = NLA_U8 }, 224a1c234f9SJiri Benc [LWTUNNEL_IP_FLAGS] = { .type = NLA_U16 }, 2254ece4778SXin Long [LWTUNNEL_IP_OPTS] = { .type = NLA_NESTED }, 2263093fbe7SThomas Graf }; 2273093fbe7SThomas Graf 2284ece4778SXin Long static const struct nla_policy ip_opts_policy[LWTUNNEL_IP_OPTS_MAX + 1] = { 2294ece4778SXin Long [LWTUNNEL_IP_OPTS_GENEVE] = { .type = NLA_NESTED }, 230edf31cbbSXin Long [LWTUNNEL_IP_OPTS_VXLAN] = { .type = NLA_NESTED }, 231b0a21810SXin Long [LWTUNNEL_IP_OPTS_ERSPAN] = { .type = NLA_NESTED }, 2324ece4778SXin Long }; 2334ece4778SXin Long 2344ece4778SXin Long static const struct nla_policy 2354ece4778SXin Long geneve_opt_policy[LWTUNNEL_IP_OPT_GENEVE_MAX + 1] = { 2364ece4778SXin Long [LWTUNNEL_IP_OPT_GENEVE_CLASS] = { .type = NLA_U16 }, 2374ece4778SXin Long [LWTUNNEL_IP_OPT_GENEVE_TYPE] = { .type = NLA_U8 }, 2384ece4778SXin Long [LWTUNNEL_IP_OPT_GENEVE_DATA] = { .type = NLA_BINARY, .len = 128 }, 2394ece4778SXin Long }; 2404ece4778SXin Long 241edf31cbbSXin Long static const struct nla_policy 242edf31cbbSXin Long vxlan_opt_policy[LWTUNNEL_IP_OPT_VXLAN_MAX + 1] = { 243edf31cbbSXin Long [LWTUNNEL_IP_OPT_VXLAN_GBP] = { .type = NLA_U32 }, 244edf31cbbSXin Long }; 245edf31cbbSXin Long 246b0a21810SXin Long static const struct nla_policy 247b0a21810SXin Long erspan_opt_policy[LWTUNNEL_IP_OPT_ERSPAN_MAX + 1] = { 248b0a21810SXin Long [LWTUNNEL_IP_OPT_ERSPAN_VER] = { .type = NLA_U8 }, 249b0a21810SXin Long [LWTUNNEL_IP_OPT_ERSPAN_INDEX] = { .type = NLA_U32 }, 250b0a21810SXin Long [LWTUNNEL_IP_OPT_ERSPAN_DIR] = { .type = NLA_U8 }, 251b0a21810SXin Long [LWTUNNEL_IP_OPT_ERSPAN_HWID] = { .type = NLA_U8 }, 252b0a21810SXin Long }; 253b0a21810SXin Long 2544ece4778SXin Long static int ip_tun_parse_opts_geneve(struct nlattr *attr, 2552f1d370bSXin Long struct ip_tunnel_info *info, int opts_len, 2564ece4778SXin Long struct netlink_ext_ack *extack) 2574ece4778SXin Long { 2584ece4778SXin Long struct nlattr *tb[LWTUNNEL_IP_OPT_GENEVE_MAX + 1]; 2594ece4778SXin Long int data_len, err; 2604ece4778SXin Long 261ed02551fSXin Long err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_GENEVE_MAX, attr, 262ed02551fSXin Long geneve_opt_policy, extack); 2634ece4778SXin Long if (err) 2644ece4778SXin Long return err; 2654ece4778SXin Long 2664ece4778SXin Long if (!tb[LWTUNNEL_IP_OPT_GENEVE_CLASS] || 2674ece4778SXin Long !tb[LWTUNNEL_IP_OPT_GENEVE_TYPE] || 2684ece4778SXin Long !tb[LWTUNNEL_IP_OPT_GENEVE_DATA]) 2694ece4778SXin Long return -EINVAL; 2704ece4778SXin Long 2714ece4778SXin Long attr = tb[LWTUNNEL_IP_OPT_GENEVE_DATA]; 2724ece4778SXin Long data_len = nla_len(attr); 2734ece4778SXin Long if (data_len % 4) 2744ece4778SXin Long return -EINVAL; 2754ece4778SXin Long 2764ece4778SXin Long if (info) { 2772f1d370bSXin Long struct geneve_opt *opt = ip_tunnel_info_opts(info) + opts_len; 2784ece4778SXin Long 2794ece4778SXin Long memcpy(opt->opt_data, nla_data(attr), data_len); 2804ece4778SXin Long opt->length = data_len / 4; 2814ece4778SXin Long attr = tb[LWTUNNEL_IP_OPT_GENEVE_CLASS]; 2824ece4778SXin Long opt->opt_class = nla_get_be16(attr); 2834ece4778SXin Long attr = tb[LWTUNNEL_IP_OPT_GENEVE_TYPE]; 2844ece4778SXin Long opt->type = nla_get_u8(attr); 2854ece4778SXin Long info->key.tun_flags |= TUNNEL_GENEVE_OPT; 2864ece4778SXin Long } 2874ece4778SXin Long 2884ece4778SXin Long return sizeof(struct geneve_opt) + data_len; 2894ece4778SXin Long } 2904ece4778SXin Long 291edf31cbbSXin Long static int ip_tun_parse_opts_vxlan(struct nlattr *attr, 2922f1d370bSXin Long struct ip_tunnel_info *info, int opts_len, 293edf31cbbSXin Long struct netlink_ext_ack *extack) 294edf31cbbSXin Long { 295edf31cbbSXin Long struct nlattr *tb[LWTUNNEL_IP_OPT_VXLAN_MAX + 1]; 296edf31cbbSXin Long int err; 297edf31cbbSXin Long 298ed02551fSXin Long err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_VXLAN_MAX, attr, 299ed02551fSXin Long vxlan_opt_policy, extack); 300edf31cbbSXin Long if (err) 301edf31cbbSXin Long return err; 302edf31cbbSXin Long 303edf31cbbSXin Long if (!tb[LWTUNNEL_IP_OPT_VXLAN_GBP]) 304edf31cbbSXin Long return -EINVAL; 305edf31cbbSXin Long 306edf31cbbSXin Long if (info) { 3072f1d370bSXin Long struct vxlan_metadata *md = 3082f1d370bSXin Long ip_tunnel_info_opts(info) + opts_len; 309edf31cbbSXin Long 310edf31cbbSXin Long attr = tb[LWTUNNEL_IP_OPT_VXLAN_GBP]; 311edf31cbbSXin Long md->gbp = nla_get_u32(attr); 312edf31cbbSXin Long info->key.tun_flags |= TUNNEL_VXLAN_OPT; 313edf31cbbSXin Long } 314edf31cbbSXin Long 315edf31cbbSXin Long return sizeof(struct vxlan_metadata); 316edf31cbbSXin Long } 317edf31cbbSXin Long 318b0a21810SXin Long static int ip_tun_parse_opts_erspan(struct nlattr *attr, 3192f1d370bSXin Long struct ip_tunnel_info *info, int opts_len, 320b0a21810SXin Long struct netlink_ext_ack *extack) 321b0a21810SXin Long { 322b0a21810SXin Long struct nlattr *tb[LWTUNNEL_IP_OPT_ERSPAN_MAX + 1]; 323b0a21810SXin Long int err; 3241841b982SXin Long u8 ver; 325b0a21810SXin Long 326ed02551fSXin Long err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_ERSPAN_MAX, attr, 327ed02551fSXin Long erspan_opt_policy, extack); 328b0a21810SXin Long if (err) 329b0a21810SXin Long return err; 330b0a21810SXin Long 331b0a21810SXin Long if (!tb[LWTUNNEL_IP_OPT_ERSPAN_VER]) 332b0a21810SXin Long return -EINVAL; 333b0a21810SXin Long 3341841b982SXin Long ver = nla_get_u8(tb[LWTUNNEL_IP_OPT_ERSPAN_VER]); 3351841b982SXin Long if (ver == 1) { 3361841b982SXin Long if (!tb[LWTUNNEL_IP_OPT_ERSPAN_INDEX]) 3371841b982SXin Long return -EINVAL; 3381841b982SXin Long } else if (ver == 2) { 3391841b982SXin Long if (!tb[LWTUNNEL_IP_OPT_ERSPAN_DIR] || 3401841b982SXin Long !tb[LWTUNNEL_IP_OPT_ERSPAN_HWID]) 3411841b982SXin Long return -EINVAL; 3421841b982SXin Long } else { 3431841b982SXin Long return -EINVAL; 3441841b982SXin Long } 3451841b982SXin Long 346b0a21810SXin Long if (info) { 3472f1d370bSXin Long struct erspan_metadata *md = 3482f1d370bSXin Long ip_tunnel_info_opts(info) + opts_len; 349b0a21810SXin Long 3501841b982SXin Long md->version = ver; 3511841b982SXin Long if (ver == 1) { 352b0a21810SXin Long attr = tb[LWTUNNEL_IP_OPT_ERSPAN_INDEX]; 353b0a21810SXin Long md->u.index = nla_get_be32(attr); 3541841b982SXin Long } else { 355b0a21810SXin Long attr = tb[LWTUNNEL_IP_OPT_ERSPAN_DIR]; 356b0a21810SXin Long md->u.md2.dir = nla_get_u8(attr); 357b0a21810SXin Long attr = tb[LWTUNNEL_IP_OPT_ERSPAN_HWID]; 358b0a21810SXin Long set_hwid(&md->u.md2, nla_get_u8(attr)); 359b0a21810SXin Long } 360b0a21810SXin Long 361b0a21810SXin Long info->key.tun_flags |= TUNNEL_ERSPAN_OPT; 362b0a21810SXin Long } 363b0a21810SXin Long 364b0a21810SXin Long return sizeof(struct erspan_metadata); 365b0a21810SXin Long } 366b0a21810SXin Long 3674ece4778SXin Long static int ip_tun_parse_opts(struct nlattr *attr, struct ip_tunnel_info *info, 3684ece4778SXin Long struct netlink_ext_ack *extack) 3694ece4778SXin Long { 3702f1d370bSXin Long int err, rem, opt_len, opts_len = 0, type = 0; 3712f1d370bSXin Long struct nlattr *nla; 3724ece4778SXin Long 3734ece4778SXin Long if (!attr) 3744ece4778SXin Long return 0; 3754ece4778SXin Long 3762f1d370bSXin Long err = nla_validate(nla_data(attr), nla_len(attr), LWTUNNEL_IP_OPTS_MAX, 3774ece4778SXin Long ip_opts_policy, extack); 3784ece4778SXin Long if (err) 3794ece4778SXin Long return err; 3804ece4778SXin Long 3812f1d370bSXin Long nla_for_each_attr(nla, nla_data(attr), nla_len(attr), rem) { 3822f1d370bSXin Long switch (nla_type(nla)) { 3832f1d370bSXin Long case LWTUNNEL_IP_OPTS_GENEVE: 3842f1d370bSXin Long if (type && type != TUNNEL_GENEVE_OPT) 3852f1d370bSXin Long return -EINVAL; 3862f1d370bSXin Long opt_len = ip_tun_parse_opts_geneve(nla, info, opts_len, 3872f1d370bSXin Long extack); 3882f1d370bSXin Long if (opt_len < 0) 3892f1d370bSXin Long return opt_len; 3902f1d370bSXin Long opts_len += opt_len; 3912f1d370bSXin Long if (opts_len > IP_TUNNEL_OPTS_MAX) 3922f1d370bSXin Long return -EINVAL; 3932f1d370bSXin Long type = TUNNEL_GENEVE_OPT; 3942f1d370bSXin Long break; 3952f1d370bSXin Long case LWTUNNEL_IP_OPTS_VXLAN: 3962f1d370bSXin Long if (type) 3972f1d370bSXin Long return -EINVAL; 3982f1d370bSXin Long opt_len = ip_tun_parse_opts_vxlan(nla, info, opts_len, 3992f1d370bSXin Long extack); 4002f1d370bSXin Long if (opt_len < 0) 4012f1d370bSXin Long return opt_len; 4022f1d370bSXin Long opts_len += opt_len; 4032f1d370bSXin Long type = TUNNEL_VXLAN_OPT; 4042f1d370bSXin Long break; 4052f1d370bSXin Long case LWTUNNEL_IP_OPTS_ERSPAN: 4062f1d370bSXin Long if (type) 4072f1d370bSXin Long return -EINVAL; 4082f1d370bSXin Long opt_len = ip_tun_parse_opts_erspan(nla, info, opts_len, 4092f1d370bSXin Long extack); 4102f1d370bSXin Long if (opt_len < 0) 4112f1d370bSXin Long return opt_len; 4122f1d370bSXin Long opts_len += opt_len; 4132f1d370bSXin Long type = TUNNEL_ERSPAN_OPT; 4142f1d370bSXin Long break; 4152f1d370bSXin Long default: 4162f1d370bSXin Long return -EINVAL; 4172f1d370bSXin Long } 4182f1d370bSXin Long } 4194ece4778SXin Long 4202f1d370bSXin Long return opts_len; 4214ece4778SXin Long } 4224ece4778SXin Long 4234ece4778SXin Long static int ip_tun_get_optlen(struct nlattr *attr, 4244ece4778SXin Long struct netlink_ext_ack *extack) 4254ece4778SXin Long { 4264ece4778SXin Long return ip_tun_parse_opts(attr, NULL, extack); 4274ece4778SXin Long } 4284ece4778SXin Long 4294ece4778SXin Long static int ip_tun_set_opts(struct nlattr *attr, struct ip_tunnel_info *info, 4304ece4778SXin Long struct netlink_ext_ack *extack) 4314ece4778SXin Long { 4324ece4778SXin Long return ip_tun_parse_opts(attr, info, extack); 4334ece4778SXin Long } 4344ece4778SXin Long 435faee6769SAlexander Aring static int ip_tun_build_state(struct net *net, struct nlattr *attr, 436127eb7cdSTom Herbert unsigned int family, const void *cfg, 4379ae28727SDavid Ahern struct lwtunnel_state **ts, 4389ae28727SDavid Ahern struct netlink_ext_ack *extack) 4393093fbe7SThomas Graf { 440a1c234f9SJiri Benc struct nlattr *tb[LWTUNNEL_IP_MAX + 1]; 4414ece4778SXin Long struct lwtunnel_state *new_state; 4424ece4778SXin Long struct ip_tunnel_info *tun_info; 4434ece4778SXin Long int err, opt_len; 4443093fbe7SThomas Graf 4458cb08174SJohannes Berg err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP_MAX, attr, 4468cb08174SJohannes Berg ip_tun_policy, extack); 4473093fbe7SThomas Graf if (err < 0) 4483093fbe7SThomas Graf return err; 4493093fbe7SThomas Graf 4504ece4778SXin Long opt_len = ip_tun_get_optlen(tb[LWTUNNEL_IP_OPTS], extack); 4514ece4778SXin Long if (opt_len < 0) 4524ece4778SXin Long return opt_len; 4534ece4778SXin Long 4544ece4778SXin Long new_state = lwtunnel_state_alloc(sizeof(*tun_info) + opt_len); 4553093fbe7SThomas Graf if (!new_state) 4563093fbe7SThomas Graf return -ENOMEM; 4573093fbe7SThomas Graf 4583093fbe7SThomas Graf new_state->type = LWTUNNEL_ENCAP_IP; 4593093fbe7SThomas Graf 4603093fbe7SThomas Graf tun_info = lwt_tun_info(new_state); 4613093fbe7SThomas Graf 4624ece4778SXin Long err = ip_tun_set_opts(tb[LWTUNNEL_IP_OPTS], tun_info, extack); 4634ece4778SXin Long if (err < 0) { 4644ece4778SXin Long lwtstate_free(new_state); 4654ece4778SXin Long return err; 4664ece4778SXin Long } 4674ece4778SXin Long 4683d25eabbSwenxu #ifdef CONFIG_DST_CACHE 4693d25eabbSwenxu err = dst_cache_init(&tun_info->dst_cache, GFP_KERNEL); 4703d25eabbSwenxu if (err) { 4713d25eabbSwenxu lwtstate_free(new_state); 4723d25eabbSwenxu return err; 4733d25eabbSwenxu } 4743d25eabbSwenxu #endif 4753d25eabbSwenxu 476a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_ID]) 47730d3d83aSLance Richardson tun_info->key.tun_id = nla_get_be64(tb[LWTUNNEL_IP_ID]); 4783093fbe7SThomas Graf 479a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_DST]) 4807822ce73SHaishuang Yan tun_info->key.u.ipv4.dst = nla_get_in_addr(tb[LWTUNNEL_IP_DST]); 4813093fbe7SThomas Graf 482a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_SRC]) 4837822ce73SHaishuang Yan tun_info->key.u.ipv4.src = nla_get_in_addr(tb[LWTUNNEL_IP_SRC]); 4843093fbe7SThomas Graf 485a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_TTL]) 4867c383fb2SJiri Benc tun_info->key.ttl = nla_get_u8(tb[LWTUNNEL_IP_TTL]); 4873093fbe7SThomas Graf 488a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_TOS]) 4897c383fb2SJiri Benc tun_info->key.tos = nla_get_u8(tb[LWTUNNEL_IP_TOS]); 4903093fbe7SThomas Graf 491a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_FLAGS]) 4920c06d166SXin Long tun_info->key.tun_flags |= 4930c06d166SXin Long (nla_get_be16(tb[LWTUNNEL_IP_FLAGS]) & 4940c06d166SXin Long ~TUNNEL_OPTIONS_PRESENT); 4953093fbe7SThomas Graf 4963093fbe7SThomas Graf tun_info->mode = IP_TUNNEL_INFO_TX; 4974ece4778SXin Long tun_info->options_len = opt_len; 4983093fbe7SThomas Graf 4993093fbe7SThomas Graf *ts = new_state; 5003093fbe7SThomas Graf 5013093fbe7SThomas Graf return 0; 5023093fbe7SThomas Graf } 5033093fbe7SThomas Graf 5043d25eabbSwenxu static void ip_tun_destroy_state(struct lwtunnel_state *lwtstate) 5053d25eabbSwenxu { 5063d25eabbSwenxu #ifdef CONFIG_DST_CACHE 5073d25eabbSwenxu struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate); 5083d25eabbSwenxu 5093d25eabbSwenxu dst_cache_destroy(&tun_info->dst_cache); 5103d25eabbSwenxu #endif 5113d25eabbSwenxu } 5123d25eabbSwenxu 5134ece4778SXin Long static int ip_tun_fill_encap_opts_geneve(struct sk_buff *skb, 5144ece4778SXin Long struct ip_tunnel_info *tun_info) 5154ece4778SXin Long { 5164ece4778SXin Long struct geneve_opt *opt; 5174ece4778SXin Long struct nlattr *nest; 5182f1d370bSXin Long int offset = 0; 5194ece4778SXin Long 5204ece4778SXin Long nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_GENEVE); 5214ece4778SXin Long if (!nest) 5224ece4778SXin Long return -ENOMEM; 5234ece4778SXin Long 5242f1d370bSXin Long while (tun_info->options_len > offset) { 5252f1d370bSXin Long opt = ip_tunnel_info_opts(tun_info) + offset; 5262f1d370bSXin Long if (nla_put_be16(skb, LWTUNNEL_IP_OPT_GENEVE_CLASS, 5272f1d370bSXin Long opt->opt_class) || 5284ece4778SXin Long nla_put_u8(skb, LWTUNNEL_IP_OPT_GENEVE_TYPE, opt->type) || 5294ece4778SXin Long nla_put(skb, LWTUNNEL_IP_OPT_GENEVE_DATA, opt->length * 4, 5304ece4778SXin Long opt->opt_data)) { 5314ece4778SXin Long nla_nest_cancel(skb, nest); 5324ece4778SXin Long return -ENOMEM; 5334ece4778SXin Long } 5342f1d370bSXin Long offset += sizeof(*opt) + opt->length * 4; 5352f1d370bSXin Long } 5364ece4778SXin Long 5374ece4778SXin Long nla_nest_end(skb, nest); 5384ece4778SXin Long return 0; 5394ece4778SXin Long } 5404ece4778SXin Long 541edf31cbbSXin Long static int ip_tun_fill_encap_opts_vxlan(struct sk_buff *skb, 542edf31cbbSXin Long struct ip_tunnel_info *tun_info) 543edf31cbbSXin Long { 544edf31cbbSXin Long struct vxlan_metadata *md; 545edf31cbbSXin Long struct nlattr *nest; 546edf31cbbSXin Long 547edf31cbbSXin Long nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_VXLAN); 548edf31cbbSXin Long if (!nest) 549edf31cbbSXin Long return -ENOMEM; 550edf31cbbSXin Long 551edf31cbbSXin Long md = ip_tunnel_info_opts(tun_info); 552edf31cbbSXin Long if (nla_put_u32(skb, LWTUNNEL_IP_OPT_VXLAN_GBP, md->gbp)) { 553edf31cbbSXin Long nla_nest_cancel(skb, nest); 554edf31cbbSXin Long return -ENOMEM; 555edf31cbbSXin Long } 556edf31cbbSXin Long 557edf31cbbSXin Long nla_nest_end(skb, nest); 558edf31cbbSXin Long return 0; 559edf31cbbSXin Long } 560edf31cbbSXin Long 561b0a21810SXin Long static int ip_tun_fill_encap_opts_erspan(struct sk_buff *skb, 562b0a21810SXin Long struct ip_tunnel_info *tun_info) 563b0a21810SXin Long { 564b0a21810SXin Long struct erspan_metadata *md; 565b0a21810SXin Long struct nlattr *nest; 566b0a21810SXin Long 567b0a21810SXin Long nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_ERSPAN); 568b0a21810SXin Long if (!nest) 569b0a21810SXin Long return -ENOMEM; 570b0a21810SXin Long 571b0a21810SXin Long md = ip_tunnel_info_opts(tun_info); 5723132174bSXin Long if (nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_VER, md->version)) 573b0a21810SXin Long goto err; 574b0a21810SXin Long 575b0a21810SXin Long if (md->version == 1 && 576b0a21810SXin Long nla_put_be32(skb, LWTUNNEL_IP_OPT_ERSPAN_INDEX, md->u.index)) 577b0a21810SXin Long goto err; 578b0a21810SXin Long 579b0a21810SXin Long if (md->version == 2 && 580b0a21810SXin Long (nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_DIR, md->u.md2.dir) || 581b0a21810SXin Long nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_HWID, 582b0a21810SXin Long get_hwid(&md->u.md2)))) 583b0a21810SXin Long goto err; 584b0a21810SXin Long 585b0a21810SXin Long nla_nest_end(skb, nest); 586b0a21810SXin Long return 0; 587b0a21810SXin Long err: 588b0a21810SXin Long nla_nest_cancel(skb, nest); 589b0a21810SXin Long return -ENOMEM; 590b0a21810SXin Long } 591b0a21810SXin Long 5924ece4778SXin Long static int ip_tun_fill_encap_opts(struct sk_buff *skb, int type, 5934ece4778SXin Long struct ip_tunnel_info *tun_info) 5944ece4778SXin Long { 5954ece4778SXin Long struct nlattr *nest; 5964ece4778SXin Long int err = 0; 5974ece4778SXin Long 5980c06d166SXin Long if (!(tun_info->key.tun_flags & TUNNEL_OPTIONS_PRESENT)) 5994ece4778SXin Long return 0; 6004ece4778SXin Long 6014ece4778SXin Long nest = nla_nest_start_noflag(skb, type); 6024ece4778SXin Long if (!nest) 6034ece4778SXin Long return -ENOMEM; 6044ece4778SXin Long 6054ece4778SXin Long if (tun_info->key.tun_flags & TUNNEL_GENEVE_OPT) 6064ece4778SXin Long err = ip_tun_fill_encap_opts_geneve(skb, tun_info); 607edf31cbbSXin Long else if (tun_info->key.tun_flags & TUNNEL_VXLAN_OPT) 608edf31cbbSXin Long err = ip_tun_fill_encap_opts_vxlan(skb, tun_info); 609b0a21810SXin Long else if (tun_info->key.tun_flags & TUNNEL_ERSPAN_OPT) 610b0a21810SXin Long err = ip_tun_fill_encap_opts_erspan(skb, tun_info); 6114ece4778SXin Long 6124ece4778SXin Long if (err) { 6134ece4778SXin Long nla_nest_cancel(skb, nest); 6144ece4778SXin Long return err; 6154ece4778SXin Long } 6164ece4778SXin Long 6174ece4778SXin Long nla_nest_end(skb, nest); 6184ece4778SXin Long return 0; 6194ece4778SXin Long } 6204ece4778SXin Long 6213093fbe7SThomas Graf static int ip_tun_fill_encap_info(struct sk_buff *skb, 6223093fbe7SThomas Graf struct lwtunnel_state *lwtstate) 6233093fbe7SThomas Graf { 6243093fbe7SThomas Graf struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate); 6253093fbe7SThomas Graf 626b46f6dedSNicolas Dichtel if (nla_put_be64(skb, LWTUNNEL_IP_ID, tun_info->key.tun_id, 627b46f6dedSNicolas Dichtel LWTUNNEL_IP_PAD) || 6287822ce73SHaishuang Yan nla_put_in_addr(skb, LWTUNNEL_IP_DST, tun_info->key.u.ipv4.dst) || 6297822ce73SHaishuang Yan nla_put_in_addr(skb, LWTUNNEL_IP_SRC, tun_info->key.u.ipv4.src) || 6307c383fb2SJiri Benc nla_put_u8(skb, LWTUNNEL_IP_TOS, tun_info->key.tos) || 6317c383fb2SJiri Benc nla_put_u8(skb, LWTUNNEL_IP_TTL, tun_info->key.ttl) || 6324ece4778SXin Long nla_put_be16(skb, LWTUNNEL_IP_FLAGS, tun_info->key.tun_flags) || 6334ece4778SXin Long ip_tun_fill_encap_opts(skb, LWTUNNEL_IP_OPTS, tun_info)) 6343093fbe7SThomas Graf return -ENOMEM; 6353093fbe7SThomas Graf 6363093fbe7SThomas Graf return 0; 6373093fbe7SThomas Graf } 6383093fbe7SThomas Graf 6394ece4778SXin Long static int ip_tun_opts_nlsize(struct ip_tunnel_info *info) 6404ece4778SXin Long { 6414ece4778SXin Long int opt_len; 6424ece4778SXin Long 6430c06d166SXin Long if (!(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT)) 6444ece4778SXin Long return 0; 6454ece4778SXin Long 6464ece4778SXin Long opt_len = nla_total_size(0); /* LWTUNNEL_IP_OPTS */ 6474ece4778SXin Long if (info->key.tun_flags & TUNNEL_GENEVE_OPT) { 6482f1d370bSXin Long struct geneve_opt *opt; 6492f1d370bSXin Long int offset = 0; 6504ece4778SXin Long 6512f1d370bSXin Long opt_len += nla_total_size(0); /* LWTUNNEL_IP_OPTS_GENEVE */ 6522f1d370bSXin Long while (info->options_len > offset) { 6532f1d370bSXin Long opt = ip_tunnel_info_opts(info) + offset; 6542f1d370bSXin Long opt_len += nla_total_size(2) /* OPT_GENEVE_CLASS */ 6554ece4778SXin Long + nla_total_size(1) /* OPT_GENEVE_TYPE */ 6564ece4778SXin Long + nla_total_size(opt->length * 4); 6574ece4778SXin Long /* OPT_GENEVE_DATA */ 6582f1d370bSXin Long offset += sizeof(*opt) + opt->length * 4; 6592f1d370bSXin Long } 660edf31cbbSXin Long } else if (info->key.tun_flags & TUNNEL_VXLAN_OPT) { 661edf31cbbSXin Long opt_len += nla_total_size(0) /* LWTUNNEL_IP_OPTS_VXLAN */ 662edf31cbbSXin Long + nla_total_size(4); /* OPT_VXLAN_GBP */ 663b0a21810SXin Long } else if (info->key.tun_flags & TUNNEL_ERSPAN_OPT) { 66458e8494eSXin Long struct erspan_metadata *md = ip_tunnel_info_opts(info); 66558e8494eSXin Long 666b0a21810SXin Long opt_len += nla_total_size(0) /* LWTUNNEL_IP_OPTS_ERSPAN */ 667b0a21810SXin Long + nla_total_size(1) /* OPT_ERSPAN_VER */ 66858e8494eSXin Long + (md->version == 1 ? nla_total_size(4) 66958e8494eSXin Long /* OPT_ERSPAN_INDEX (v1) */ 67058e8494eSXin Long : nla_total_size(1) + 67158e8494eSXin Long nla_total_size(1)); 67258e8494eSXin Long /* OPT_ERSPAN_DIR + HWID (v2) */ 6734ece4778SXin Long } 6744ece4778SXin Long 6754ece4778SXin Long return opt_len; 6764ece4778SXin Long } 6774ece4778SXin Long 6783093fbe7SThomas Graf static int ip_tun_encap_nlsize(struct lwtunnel_state *lwtstate) 6793093fbe7SThomas Graf { 680b46f6dedSNicolas Dichtel return nla_total_size_64bit(8) /* LWTUNNEL_IP_ID */ 681a1c234f9SJiri Benc + nla_total_size(4) /* LWTUNNEL_IP_DST */ 682a1c234f9SJiri Benc + nla_total_size(4) /* LWTUNNEL_IP_SRC */ 683a1c234f9SJiri Benc + nla_total_size(1) /* LWTUNNEL_IP_TOS */ 684a1c234f9SJiri Benc + nla_total_size(1) /* LWTUNNEL_IP_TTL */ 6854ece4778SXin Long + nla_total_size(2) /* LWTUNNEL_IP_FLAGS */ 6864ece4778SXin Long + ip_tun_opts_nlsize(lwt_tun_info(lwtstate)); 6874ece4778SXin Long /* LWTUNNEL_IP_OPTS */ 6883093fbe7SThomas Graf } 6893093fbe7SThomas Graf 6902d798499SJiri Benc static int ip_tun_cmp_encap(struct lwtunnel_state *a, struct lwtunnel_state *b) 6912d798499SJiri Benc { 6920eb8eb2fSXin Long struct ip_tunnel_info *info_a = lwt_tun_info(a); 6930eb8eb2fSXin Long struct ip_tunnel_info *info_b = lwt_tun_info(b); 6940eb8eb2fSXin Long 6950eb8eb2fSXin Long return memcmp(info_a, info_b, sizeof(info_a->key)) || 6960eb8eb2fSXin Long info_a->mode != info_b->mode || 6970eb8eb2fSXin Long info_a->options_len != info_b->options_len || 6980eb8eb2fSXin Long memcmp(ip_tunnel_info_opts(info_a), 6990eb8eb2fSXin Long ip_tunnel_info_opts(info_b), info_a->options_len); 7002d798499SJiri Benc } 7012d798499SJiri Benc 7023093fbe7SThomas Graf static const struct lwtunnel_encap_ops ip_tun_lwt_ops = { 7033093fbe7SThomas Graf .build_state = ip_tun_build_state, 7043d25eabbSwenxu .destroy_state = ip_tun_destroy_state, 7053093fbe7SThomas Graf .fill_encap = ip_tun_fill_encap_info, 7063093fbe7SThomas Graf .get_encap_size = ip_tun_encap_nlsize, 7072d798499SJiri Benc .cmp_encap = ip_tun_cmp_encap, 70888ff7334SRobert Shearman .owner = THIS_MODULE, 7093093fbe7SThomas Graf }; 7103093fbe7SThomas Graf 71132a2b002SJiri Benc static const struct nla_policy ip6_tun_policy[LWTUNNEL_IP6_MAX + 1] = { 7127b6a70f7SXin Long [LWTUNNEL_IP6_UNSPEC] = { .strict_start_type = LWTUNNEL_IP6_OPTS }, 71332a2b002SJiri Benc [LWTUNNEL_IP6_ID] = { .type = NLA_U64 }, 71432a2b002SJiri Benc [LWTUNNEL_IP6_DST] = { .len = sizeof(struct in6_addr) }, 71532a2b002SJiri Benc [LWTUNNEL_IP6_SRC] = { .len = sizeof(struct in6_addr) }, 71632a2b002SJiri Benc [LWTUNNEL_IP6_HOPLIMIT] = { .type = NLA_U8 }, 71732a2b002SJiri Benc [LWTUNNEL_IP6_TC] = { .type = NLA_U8 }, 71832a2b002SJiri Benc [LWTUNNEL_IP6_FLAGS] = { .type = NLA_U16 }, 7197b6a70f7SXin Long [LWTUNNEL_IP6_OPTS] = { .type = NLA_NESTED }, 72032a2b002SJiri Benc }; 72132a2b002SJiri Benc 722faee6769SAlexander Aring static int ip6_tun_build_state(struct net *net, struct nlattr *attr, 723127eb7cdSTom Herbert unsigned int family, const void *cfg, 7249ae28727SDavid Ahern struct lwtunnel_state **ts, 7259ae28727SDavid Ahern struct netlink_ext_ack *extack) 72632a2b002SJiri Benc { 72732a2b002SJiri Benc struct nlattr *tb[LWTUNNEL_IP6_MAX + 1]; 7284ece4778SXin Long struct lwtunnel_state *new_state; 7294ece4778SXin Long struct ip_tunnel_info *tun_info; 7304ece4778SXin Long int err, opt_len; 73132a2b002SJiri Benc 7328cb08174SJohannes Berg err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP6_MAX, attr, 7338cb08174SJohannes Berg ip6_tun_policy, extack); 73432a2b002SJiri Benc if (err < 0) 73532a2b002SJiri Benc return err; 73632a2b002SJiri Benc 7374ece4778SXin Long opt_len = ip_tun_get_optlen(tb[LWTUNNEL_IP6_OPTS], extack); 7384ece4778SXin Long if (opt_len < 0) 7394ece4778SXin Long return opt_len; 7404ece4778SXin Long 7414ece4778SXin Long new_state = lwtunnel_state_alloc(sizeof(*tun_info) + opt_len); 74232a2b002SJiri Benc if (!new_state) 74332a2b002SJiri Benc return -ENOMEM; 74432a2b002SJiri Benc 74532a2b002SJiri Benc new_state->type = LWTUNNEL_ENCAP_IP6; 74632a2b002SJiri Benc 74732a2b002SJiri Benc tun_info = lwt_tun_info(new_state); 74832a2b002SJiri Benc 7494ece4778SXin Long err = ip_tun_set_opts(tb[LWTUNNEL_IP6_OPTS], tun_info, extack); 7504ece4778SXin Long if (err < 0) { 7514ece4778SXin Long lwtstate_free(new_state); 7524ece4778SXin Long return err; 7534ece4778SXin Long } 7544ece4778SXin Long 75532a2b002SJiri Benc if (tb[LWTUNNEL_IP6_ID]) 75630d3d83aSLance Richardson tun_info->key.tun_id = nla_get_be64(tb[LWTUNNEL_IP6_ID]); 75732a2b002SJiri Benc 75832a2b002SJiri Benc if (tb[LWTUNNEL_IP6_DST]) 75932a2b002SJiri Benc tun_info->key.u.ipv6.dst = nla_get_in6_addr(tb[LWTUNNEL_IP6_DST]); 76032a2b002SJiri Benc 76132a2b002SJiri Benc if (tb[LWTUNNEL_IP6_SRC]) 76232a2b002SJiri Benc tun_info->key.u.ipv6.src = nla_get_in6_addr(tb[LWTUNNEL_IP6_SRC]); 76332a2b002SJiri Benc 76432a2b002SJiri Benc if (tb[LWTUNNEL_IP6_HOPLIMIT]) 76532a2b002SJiri Benc tun_info->key.ttl = nla_get_u8(tb[LWTUNNEL_IP6_HOPLIMIT]); 76632a2b002SJiri Benc 76732a2b002SJiri Benc if (tb[LWTUNNEL_IP6_TC]) 76832a2b002SJiri Benc tun_info->key.tos = nla_get_u8(tb[LWTUNNEL_IP6_TC]); 76932a2b002SJiri Benc 77032a2b002SJiri Benc if (tb[LWTUNNEL_IP6_FLAGS]) 7710c06d166SXin Long tun_info->key.tun_flags |= 7720c06d166SXin Long (nla_get_be16(tb[LWTUNNEL_IP6_FLAGS]) & 7730c06d166SXin Long ~TUNNEL_OPTIONS_PRESENT); 77432a2b002SJiri Benc 7757f9562a1SJiri Benc tun_info->mode = IP_TUNNEL_INFO_TX | IP_TUNNEL_INFO_IPV6; 7764ece4778SXin Long tun_info->options_len = opt_len; 77732a2b002SJiri Benc 77832a2b002SJiri Benc *ts = new_state; 77932a2b002SJiri Benc 78032a2b002SJiri Benc return 0; 78132a2b002SJiri Benc } 78232a2b002SJiri Benc 78332a2b002SJiri Benc static int ip6_tun_fill_encap_info(struct sk_buff *skb, 78432a2b002SJiri Benc struct lwtunnel_state *lwtstate) 78532a2b002SJiri Benc { 78632a2b002SJiri Benc struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate); 78732a2b002SJiri Benc 788b46f6dedSNicolas Dichtel if (nla_put_be64(skb, LWTUNNEL_IP6_ID, tun_info->key.tun_id, 789b46f6dedSNicolas Dichtel LWTUNNEL_IP6_PAD) || 79032a2b002SJiri Benc nla_put_in6_addr(skb, LWTUNNEL_IP6_DST, &tun_info->key.u.ipv6.dst) || 79132a2b002SJiri Benc nla_put_in6_addr(skb, LWTUNNEL_IP6_SRC, &tun_info->key.u.ipv6.src) || 792995096a0SQuentin Armitage nla_put_u8(skb, LWTUNNEL_IP6_TC, tun_info->key.tos) || 793995096a0SQuentin Armitage nla_put_u8(skb, LWTUNNEL_IP6_HOPLIMIT, tun_info->key.ttl) || 7944ece4778SXin Long nla_put_be16(skb, LWTUNNEL_IP6_FLAGS, tun_info->key.tun_flags) || 7954ece4778SXin Long ip_tun_fill_encap_opts(skb, LWTUNNEL_IP6_OPTS, tun_info)) 79632a2b002SJiri Benc return -ENOMEM; 79732a2b002SJiri Benc 79832a2b002SJiri Benc return 0; 79932a2b002SJiri Benc } 80032a2b002SJiri Benc 80132a2b002SJiri Benc static int ip6_tun_encap_nlsize(struct lwtunnel_state *lwtstate) 80232a2b002SJiri Benc { 803b46f6dedSNicolas Dichtel return nla_total_size_64bit(8) /* LWTUNNEL_IP6_ID */ 80432a2b002SJiri Benc + nla_total_size(16) /* LWTUNNEL_IP6_DST */ 80532a2b002SJiri Benc + nla_total_size(16) /* LWTUNNEL_IP6_SRC */ 80632a2b002SJiri Benc + nla_total_size(1) /* LWTUNNEL_IP6_HOPLIMIT */ 80732a2b002SJiri Benc + nla_total_size(1) /* LWTUNNEL_IP6_TC */ 8084ece4778SXin Long + nla_total_size(2) /* LWTUNNEL_IP6_FLAGS */ 8094ece4778SXin Long + ip_tun_opts_nlsize(lwt_tun_info(lwtstate)); 8104ece4778SXin Long /* LWTUNNEL_IP6_OPTS */ 81132a2b002SJiri Benc } 81232a2b002SJiri Benc 81332a2b002SJiri Benc static const struct lwtunnel_encap_ops ip6_tun_lwt_ops = { 81432a2b002SJiri Benc .build_state = ip6_tun_build_state, 81532a2b002SJiri Benc .fill_encap = ip6_tun_fill_encap_info, 81632a2b002SJiri Benc .get_encap_size = ip6_tun_encap_nlsize, 81732a2b002SJiri Benc .cmp_encap = ip_tun_cmp_encap, 81888ff7334SRobert Shearman .owner = THIS_MODULE, 81932a2b002SJiri Benc }; 82032a2b002SJiri Benc 821045a0fa0SThomas Graf void __init ip_tunnel_core_init(void) 8223093fbe7SThomas Graf { 823fca5fdf6SDaniel Borkmann /* If you land here, make sure whether increasing ip_tunnel_info's 824fca5fdf6SDaniel Borkmann * options_len is a reasonable choice with its usage in front ends 825fca5fdf6SDaniel Borkmann * (f.e., it's part of flow keys, etc). 826fca5fdf6SDaniel Borkmann */ 827fca5fdf6SDaniel Borkmann BUILD_BUG_ON(IP_TUNNEL_OPTS_MAX != 255); 828fca5fdf6SDaniel Borkmann 8293093fbe7SThomas Graf lwtunnel_encap_add_ops(&ip_tun_lwt_ops, LWTUNNEL_ENCAP_IP); 83032a2b002SJiri Benc lwtunnel_encap_add_ops(&ip6_tun_lwt_ops, LWTUNNEL_ENCAP_IP6); 8313093fbe7SThomas Graf } 832e7030878SThomas Graf 8335263a98fSDavidlohr Bueso DEFINE_STATIC_KEY_FALSE(ip_tunnel_metadata_cnt); 834e7030878SThomas Graf EXPORT_SYMBOL(ip_tunnel_metadata_cnt); 835e7030878SThomas Graf 836e7030878SThomas Graf void ip_tunnel_need_metadata(void) 837e7030878SThomas Graf { 8385263a98fSDavidlohr Bueso static_branch_inc(&ip_tunnel_metadata_cnt); 839e7030878SThomas Graf } 840e7030878SThomas Graf EXPORT_SYMBOL_GPL(ip_tunnel_need_metadata); 841e7030878SThomas Graf 842e7030878SThomas Graf void ip_tunnel_unneed_metadata(void) 843e7030878SThomas Graf { 8445263a98fSDavidlohr Bueso static_branch_dec(&ip_tunnel_metadata_cnt); 845e7030878SThomas Graf } 846e7030878SThomas Graf EXPORT_SYMBOL_GPL(ip_tunnel_unneed_metadata); 847