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> 288ed54f16SStefano Brivio #include <net/ip6_checksum.h> 290e6fbc5bSPravin B Shelar #include <net/arp.h> 300e6fbc5bSPravin B Shelar #include <net/checksum.h> 310e6fbc5bSPravin B Shelar #include <net/dsfield.h> 320e6fbc5bSPravin B Shelar #include <net/inet_ecn.h> 330e6fbc5bSPravin B Shelar #include <net/xfrm.h> 340e6fbc5bSPravin B Shelar #include <net/net_namespace.h> 350e6fbc5bSPravin B Shelar #include <net/netns/generic.h> 360e6fbc5bSPravin B Shelar #include <net/rtnetlink.h> 3763d008a4SJiri Benc #include <net/dst_metadata.h> 384ece4778SXin Long #include <net/geneve.h> 39edf31cbbSXin Long #include <net/vxlan.h> 40b0a21810SXin Long #include <net/erspan.h> 410e6fbc5bSPravin B Shelar 4255c2bc14STom Herbert const struct ip_tunnel_encap_ops __rcu * 4355c2bc14STom Herbert iptun_encaps[MAX_IPTUN_ENCAP_OPS] __read_mostly; 4455c2bc14STom Herbert EXPORT_SYMBOL(iptun_encaps); 4555c2bc14STom Herbert 46058214a4STom Herbert const struct ip6_tnl_encap_ops __rcu * 47058214a4STom Herbert ip6tun_encaps[MAX_IPTUN_ENCAP_OPS] __read_mostly; 48058214a4STom Herbert EXPORT_SYMBOL(ip6tun_encaps); 49058214a4STom Herbert 50039f5062SPravin B Shelar void iptunnel_xmit(struct sock *sk, struct rtable *rt, struct sk_buff *skb, 510e6fbc5bSPravin B Shelar __be32 src, __be32 dst, __u8 proto, 52963a88b3SNicolas Dichtel __u8 tos, __u8 ttl, __be16 df, bool xnet) 530e6fbc5bSPravin B Shelar { 54bc22a0e2SNicolas Dichtel int pkt_len = skb->len - skb_inner_network_offset(skb); 55f859b0f6SEric W. Biederman struct net *net = dev_net(rt->dst.dev); 56039f5062SPravin B Shelar struct net_device *dev = skb->dev; 570e6fbc5bSPravin B Shelar struct iphdr *iph; 580e6fbc5bSPravin B Shelar int err; 590e6fbc5bSPravin B Shelar 60963a88b3SNicolas Dichtel skb_scrub_packet(skb, xnet); 61963a88b3SNicolas Dichtel 62bf8d85d4SEric Dumazet skb_clear_hash_if_not_l4(skb); 630e6fbc5bSPravin B Shelar skb_dst_set(skb, &rt->dst); 640e6fbc5bSPravin B Shelar memset(IPCB(skb), 0, sizeof(*IPCB(skb))); 650e6fbc5bSPravin B Shelar 660e6fbc5bSPravin B Shelar /* Push down and install the IP header. */ 6778a3694dSSteffen Klassert skb_push(skb, sizeof(struct iphdr)); 680e6fbc5bSPravin B Shelar skb_reset_network_header(skb); 690e6fbc5bSPravin B Shelar 700e6fbc5bSPravin B Shelar iph = ip_hdr(skb); 710e6fbc5bSPravin B Shelar 720e6fbc5bSPravin B Shelar iph->version = 4; 730e6fbc5bSPravin B Shelar iph->ihl = sizeof(struct iphdr) >> 2; 7416f7eb2bSSabrina Dubroca iph->frag_off = ip_mtu_locked(&rt->dst) ? 0 : df; 750e6fbc5bSPravin B Shelar iph->protocol = proto; 760e6fbc5bSPravin B Shelar iph->tos = tos; 770e6fbc5bSPravin B Shelar iph->daddr = dst; 780e6fbc5bSPravin B Shelar iph->saddr = src; 790e6fbc5bSPravin B Shelar iph->ttl = ttl; 80f859b0f6SEric W. Biederman __ip_select_ident(net, iph, skb_shinfo(skb)->gso_segs ?: 1); 810e6fbc5bSPravin B Shelar 8233224b16SEric W. Biederman err = ip_local_out(net, sk, skb); 835684abf7SXin Long 845684abf7SXin Long if (dev) { 850e6fbc5bSPravin B Shelar if (unlikely(net_xmit_eval(err))) 860e6fbc5bSPravin B Shelar pkt_len = 0; 87039f5062SPravin B Shelar iptunnel_xmit_stats(dev, pkt_len); 880e6fbc5bSPravin B Shelar } 895684abf7SXin Long } 900e6fbc5bSPravin B Shelar EXPORT_SYMBOL_GPL(iptunnel_xmit); 913d7b46cdSPravin B Shelar 92a6d5bbf3SJiri Benc int __iptunnel_pull_header(struct sk_buff *skb, int hdr_len, 93a6d5bbf3SJiri Benc __be16 inner_proto, bool raw_proto, bool xnet) 943d7b46cdSPravin B Shelar { 953d7b46cdSPravin B Shelar if (unlikely(!pskb_may_pull(skb, hdr_len))) 963d7b46cdSPravin B Shelar return -ENOMEM; 973d7b46cdSPravin B Shelar 983d7b46cdSPravin B Shelar skb_pull_rcsum(skb, hdr_len); 993d7b46cdSPravin B Shelar 100a6d5bbf3SJiri Benc if (!raw_proto && inner_proto == htons(ETH_P_TEB)) { 1011245dfc8SLi RongQing struct ethhdr *eh; 1023d7b46cdSPravin B Shelar 1033d7b46cdSPravin B Shelar if (unlikely(!pskb_may_pull(skb, ETH_HLEN))) 1043d7b46cdSPravin B Shelar return -ENOMEM; 1053d7b46cdSPravin B Shelar 1061245dfc8SLi RongQing eh = (struct ethhdr *)skb->data; 107d181ddcaSAlexander Duyck if (likely(eth_proto_is_802_3(eh->h_proto))) 1083d7b46cdSPravin B Shelar skb->protocol = eh->h_proto; 1093d7b46cdSPravin B Shelar else 1103d7b46cdSPravin B Shelar skb->protocol = htons(ETH_P_802_2); 1113d7b46cdSPravin B Shelar 1123d7b46cdSPravin B Shelar } else { 1133d7b46cdSPravin B Shelar skb->protocol = inner_proto; 1143d7b46cdSPravin B Shelar } 1153d7b46cdSPravin B Shelar 1167539fadcSTom Herbert skb_clear_hash_if_not_l4(skb); 1173e2ed0c2SMichał Mirosław __vlan_hwaccel_clear_tag(skb); 1183d7b46cdSPravin B Shelar skb_set_queue_mapping(skb, 0); 1197f290c94SJiri Benc skb_scrub_packet(skb, xnet); 120a09a4c8dSJesse Gross 121a09a4c8dSJesse Gross return iptunnel_pull_offloads(skb); 1223d7b46cdSPravin B Shelar } 123a6d5bbf3SJiri Benc EXPORT_SYMBOL_GPL(__iptunnel_pull_header); 1242d26f0a3SEric Dumazet 12563d008a4SJiri Benc struct metadata_dst *iptunnel_metadata_reply(struct metadata_dst *md, 12663d008a4SJiri Benc gfp_t flags) 12763d008a4SJiri Benc { 12863d008a4SJiri Benc struct metadata_dst *res; 12963d008a4SJiri Benc struct ip_tunnel_info *dst, *src; 13063d008a4SJiri Benc 1313fcece12SJakub Kicinski if (!md || md->type != METADATA_IP_TUNNEL || 1323fcece12SJakub Kicinski md->u.tun_info.mode & IP_TUNNEL_INFO_TX) 13363d008a4SJiri Benc return NULL; 13463d008a4SJiri Benc 135f52f11ecSXin Long src = &md->u.tun_info; 136f52f11ecSXin Long res = metadata_dst_alloc(src->options_len, METADATA_IP_TUNNEL, flags); 13763d008a4SJiri Benc if (!res) 13863d008a4SJiri Benc return NULL; 13963d008a4SJiri Benc 14063d008a4SJiri Benc dst = &res->u.tun_info; 14163d008a4SJiri Benc dst->key.tun_id = src->key.tun_id; 14263d008a4SJiri Benc if (src->mode & IP_TUNNEL_INFO_IPV6) 14363d008a4SJiri Benc memcpy(&dst->key.u.ipv6.dst, &src->key.u.ipv6.src, 14463d008a4SJiri Benc sizeof(struct in6_addr)); 14563d008a4SJiri Benc else 14663d008a4SJiri Benc dst->key.u.ipv4.dst = src->key.u.ipv4.src; 1477bdca378Swenxu dst->key.tun_flags = src->key.tun_flags; 14863d008a4SJiri Benc dst->mode = src->mode | IP_TUNNEL_INFO_TX; 149f52f11ecSXin Long ip_tunnel_info_opts_set(dst, ip_tunnel_info_opts(src), 150f52f11ecSXin Long src->options_len, 0); 15163d008a4SJiri Benc 15263d008a4SJiri Benc return res; 15363d008a4SJiri Benc } 15463d008a4SJiri Benc EXPORT_SYMBOL_GPL(iptunnel_metadata_reply); 15563d008a4SJiri Benc 156aed069dfSAlexander Duyck int iptunnel_handle_offloads(struct sk_buff *skb, 1572d26f0a3SEric Dumazet int gso_type_mask) 1582d26f0a3SEric Dumazet { 1592d26f0a3SEric Dumazet int err; 1602d26f0a3SEric Dumazet 1612d26f0a3SEric Dumazet if (likely(!skb->encapsulation)) { 1622d26f0a3SEric Dumazet skb_reset_inner_headers(skb); 1632d26f0a3SEric Dumazet skb->encapsulation = 1; 1642d26f0a3SEric Dumazet } 1652d26f0a3SEric Dumazet 1662d26f0a3SEric Dumazet if (skb_is_gso(skb)) { 1679580bf2eSEric Dumazet err = skb_header_unclone(skb, GFP_ATOMIC); 1682d26f0a3SEric Dumazet if (unlikely(err)) 169aed069dfSAlexander Duyck return err; 1702d26f0a3SEric Dumazet skb_shinfo(skb)->gso_type |= gso_type_mask; 171aed069dfSAlexander Duyck return 0; 1722d26f0a3SEric Dumazet } 1732d26f0a3SEric Dumazet 1746fa79666SEdward Cree if (skb->ip_summed != CHECKSUM_PARTIAL) { 1752d26f0a3SEric Dumazet skb->ip_summed = CHECKSUM_NONE; 1766fa79666SEdward Cree /* We clear encapsulation here to prevent badly-written 1776fa79666SEdward Cree * drivers potentially deciding to offload an inner checksum 1786fa79666SEdward Cree * if we set CHECKSUM_PARTIAL on the outer header. 1796fa79666SEdward Cree * This should go away when the drivers are all fixed. 1806fa79666SEdward Cree */ 181179bc67fSEdward Cree skb->encapsulation = 0; 182179bc67fSEdward Cree } 1832d26f0a3SEric Dumazet 184aed069dfSAlexander Duyck return 0; 1852d26f0a3SEric Dumazet } 1862d26f0a3SEric Dumazet EXPORT_SYMBOL_GPL(iptunnel_handle_offloads); 187ebe44f35SDavid S. Miller 1884cb47a86SStefano Brivio /** 1894cb47a86SStefano Brivio * iptunnel_pmtud_build_icmp() - Build ICMP error message for PMTUD 1904cb47a86SStefano Brivio * @skb: Original packet with L2 header 1914cb47a86SStefano Brivio * @mtu: MTU value for ICMP error 1924cb47a86SStefano Brivio * 1934cb47a86SStefano Brivio * Return: length on success, negative error code if message couldn't be built. 1944cb47a86SStefano Brivio */ 1954cb47a86SStefano Brivio static int iptunnel_pmtud_build_icmp(struct sk_buff *skb, int mtu) 1964cb47a86SStefano Brivio { 1974cb47a86SStefano Brivio const struct iphdr *iph = ip_hdr(skb); 1984cb47a86SStefano Brivio struct icmphdr *icmph; 1994cb47a86SStefano Brivio struct iphdr *niph; 2004cb47a86SStefano Brivio struct ethhdr eh; 2014cb47a86SStefano Brivio int len, err; 2024cb47a86SStefano Brivio 2034cb47a86SStefano Brivio if (!pskb_may_pull(skb, ETH_HLEN + sizeof(struct iphdr))) 2044cb47a86SStefano Brivio return -EINVAL; 2054cb47a86SStefano Brivio 2064cb47a86SStefano Brivio skb_copy_bits(skb, skb_mac_offset(skb), &eh, ETH_HLEN); 2074cb47a86SStefano Brivio pskb_pull(skb, ETH_HLEN); 2084cb47a86SStefano Brivio skb_reset_network_header(skb); 2094cb47a86SStefano Brivio 2104cb47a86SStefano Brivio err = pskb_trim(skb, 576 - sizeof(*niph) - sizeof(*icmph)); 2114cb47a86SStefano Brivio if (err) 2124cb47a86SStefano Brivio return err; 2134cb47a86SStefano Brivio 2144cb47a86SStefano Brivio len = skb->len + sizeof(*icmph); 2154cb47a86SStefano Brivio err = skb_cow(skb, sizeof(*niph) + sizeof(*icmph) + ETH_HLEN); 2164cb47a86SStefano Brivio if (err) 2174cb47a86SStefano Brivio return err; 2184cb47a86SStefano Brivio 2194cb47a86SStefano Brivio icmph = skb_push(skb, sizeof(*icmph)); 2204cb47a86SStefano Brivio *icmph = (struct icmphdr) { 2214cb47a86SStefano Brivio .type = ICMP_DEST_UNREACH, 2224cb47a86SStefano Brivio .code = ICMP_FRAG_NEEDED, 2234cb47a86SStefano Brivio .checksum = 0, 2244cb47a86SStefano Brivio .un.frag.__unused = 0, 225fda4fde2SJulian Wiedmann .un.frag.mtu = htons(mtu), 2264cb47a86SStefano Brivio }; 2274cb47a86SStefano Brivio icmph->checksum = ip_compute_csum(icmph, len); 2284cb47a86SStefano Brivio skb_reset_transport_header(skb); 2294cb47a86SStefano Brivio 2304cb47a86SStefano Brivio niph = skb_push(skb, sizeof(*niph)); 2314cb47a86SStefano Brivio *niph = (struct iphdr) { 2324cb47a86SStefano Brivio .ihl = sizeof(*niph) / 4u, 2334cb47a86SStefano Brivio .version = 4, 2344cb47a86SStefano Brivio .tos = 0, 2354cb47a86SStefano Brivio .tot_len = htons(len + sizeof(*niph)), 2364cb47a86SStefano Brivio .id = 0, 2374cb47a86SStefano Brivio .frag_off = htons(IP_DF), 2384cb47a86SStefano Brivio .ttl = iph->ttl, 2394cb47a86SStefano Brivio .protocol = IPPROTO_ICMP, 2404cb47a86SStefano Brivio .saddr = iph->daddr, 2414cb47a86SStefano Brivio .daddr = iph->saddr, 2424cb47a86SStefano Brivio }; 2434cb47a86SStefano Brivio ip_send_check(niph); 2444cb47a86SStefano Brivio skb_reset_network_header(skb); 2454cb47a86SStefano Brivio 2464cb47a86SStefano Brivio skb->ip_summed = CHECKSUM_NONE; 2474cb47a86SStefano Brivio 248fda4fde2SJulian Wiedmann eth_header(skb, skb->dev, ntohs(eh.h_proto), eh.h_source, eh.h_dest, 0); 2494cb47a86SStefano Brivio skb_reset_mac_header(skb); 2504cb47a86SStefano Brivio 2514cb47a86SStefano Brivio return skb->len; 2524cb47a86SStefano Brivio } 2534cb47a86SStefano Brivio 2544cb47a86SStefano Brivio /** 2554cb47a86SStefano Brivio * iptunnel_pmtud_check_icmp() - Trigger ICMP reply if needed and allowed 2564cb47a86SStefano Brivio * @skb: Buffer being sent by encapsulation, L2 headers expected 2574cb47a86SStefano Brivio * @mtu: Network MTU for path 2584cb47a86SStefano Brivio * 2594cb47a86SStefano Brivio * Return: 0 for no ICMP reply, length if built, negative value on error. 2604cb47a86SStefano Brivio */ 2614cb47a86SStefano Brivio static int iptunnel_pmtud_check_icmp(struct sk_buff *skb, int mtu) 2624cb47a86SStefano Brivio { 2634cb47a86SStefano Brivio const struct icmphdr *icmph = icmp_hdr(skb); 2644cb47a86SStefano Brivio const struct iphdr *iph = ip_hdr(skb); 2654cb47a86SStefano Brivio 26677a2d673SStefano Brivio if (mtu < 576 || iph->frag_off != htons(IP_DF)) 2674cb47a86SStefano Brivio return 0; 2684cb47a86SStefano Brivio 2694cb47a86SStefano Brivio if (ipv4_is_lbcast(iph->daddr) || ipv4_is_multicast(iph->daddr) || 2704cb47a86SStefano Brivio ipv4_is_zeronet(iph->saddr) || ipv4_is_loopback(iph->saddr) || 2714cb47a86SStefano Brivio ipv4_is_lbcast(iph->saddr) || ipv4_is_multicast(iph->saddr)) 2724cb47a86SStefano Brivio return 0; 2734cb47a86SStefano Brivio 2744cb47a86SStefano Brivio if (iph->protocol == IPPROTO_ICMP && icmp_is_err(icmph->type)) 2754cb47a86SStefano Brivio return 0; 2764cb47a86SStefano Brivio 2774cb47a86SStefano Brivio return iptunnel_pmtud_build_icmp(skb, mtu); 2784cb47a86SStefano Brivio } 2794cb47a86SStefano Brivio 2804cb47a86SStefano Brivio #if IS_ENABLED(CONFIG_IPV6) 2814cb47a86SStefano Brivio /** 2824cb47a86SStefano Brivio * iptunnel_pmtud_build_icmpv6() - Build ICMPv6 error message for PMTUD 2834cb47a86SStefano Brivio * @skb: Original packet with L2 header 2844cb47a86SStefano Brivio * @mtu: MTU value for ICMPv6 error 2854cb47a86SStefano Brivio * 2864cb47a86SStefano Brivio * Return: length on success, negative error code if message couldn't be built. 2874cb47a86SStefano Brivio */ 2884cb47a86SStefano Brivio static int iptunnel_pmtud_build_icmpv6(struct sk_buff *skb, int mtu) 2894cb47a86SStefano Brivio { 2904cb47a86SStefano Brivio const struct ipv6hdr *ip6h = ipv6_hdr(skb); 2914cb47a86SStefano Brivio struct icmp6hdr *icmp6h; 2924cb47a86SStefano Brivio struct ipv6hdr *nip6h; 2934cb47a86SStefano Brivio struct ethhdr eh; 2944cb47a86SStefano Brivio int len, err; 2954cb47a86SStefano Brivio __wsum csum; 2964cb47a86SStefano Brivio 2974cb47a86SStefano Brivio if (!pskb_may_pull(skb, ETH_HLEN + sizeof(struct ipv6hdr))) 2984cb47a86SStefano Brivio return -EINVAL; 2994cb47a86SStefano Brivio 3004cb47a86SStefano Brivio skb_copy_bits(skb, skb_mac_offset(skb), &eh, ETH_HLEN); 3014cb47a86SStefano Brivio pskb_pull(skb, ETH_HLEN); 3024cb47a86SStefano Brivio skb_reset_network_header(skb); 3034cb47a86SStefano Brivio 3044cb47a86SStefano Brivio err = pskb_trim(skb, IPV6_MIN_MTU - sizeof(*nip6h) - sizeof(*icmp6h)); 3054cb47a86SStefano Brivio if (err) 3064cb47a86SStefano Brivio return err; 3074cb47a86SStefano Brivio 3084cb47a86SStefano Brivio len = skb->len + sizeof(*icmp6h); 3094cb47a86SStefano Brivio err = skb_cow(skb, sizeof(*nip6h) + sizeof(*icmp6h) + ETH_HLEN); 3104cb47a86SStefano Brivio if (err) 3114cb47a86SStefano Brivio return err; 3124cb47a86SStefano Brivio 3134cb47a86SStefano Brivio icmp6h = skb_push(skb, sizeof(*icmp6h)); 3144cb47a86SStefano Brivio *icmp6h = (struct icmp6hdr) { 3154cb47a86SStefano Brivio .icmp6_type = ICMPV6_PKT_TOOBIG, 3164cb47a86SStefano Brivio .icmp6_code = 0, 3174cb47a86SStefano Brivio .icmp6_cksum = 0, 3184cb47a86SStefano Brivio .icmp6_mtu = htonl(mtu), 3194cb47a86SStefano Brivio }; 3204cb47a86SStefano Brivio skb_reset_transport_header(skb); 3214cb47a86SStefano Brivio 3224cb47a86SStefano Brivio nip6h = skb_push(skb, sizeof(*nip6h)); 3234cb47a86SStefano Brivio *nip6h = (struct ipv6hdr) { 3244cb47a86SStefano Brivio .priority = 0, 3254cb47a86SStefano Brivio .version = 6, 3264cb47a86SStefano Brivio .flow_lbl = { 0 }, 3274cb47a86SStefano Brivio .payload_len = htons(len), 3284cb47a86SStefano Brivio .nexthdr = IPPROTO_ICMPV6, 3294cb47a86SStefano Brivio .hop_limit = ip6h->hop_limit, 3304cb47a86SStefano Brivio .saddr = ip6h->daddr, 3314cb47a86SStefano Brivio .daddr = ip6h->saddr, 3324cb47a86SStefano Brivio }; 3334cb47a86SStefano Brivio skb_reset_network_header(skb); 3344cb47a86SStefano Brivio 3354cb47a86SStefano Brivio csum = csum_partial(icmp6h, len, 0); 3364cb47a86SStefano Brivio icmp6h->icmp6_cksum = csum_ipv6_magic(&nip6h->saddr, &nip6h->daddr, len, 3374cb47a86SStefano Brivio IPPROTO_ICMPV6, csum); 3384cb47a86SStefano Brivio 3394cb47a86SStefano Brivio skb->ip_summed = CHECKSUM_NONE; 3404cb47a86SStefano Brivio 341fda4fde2SJulian Wiedmann eth_header(skb, skb->dev, ntohs(eh.h_proto), eh.h_source, eh.h_dest, 0); 3424cb47a86SStefano Brivio skb_reset_mac_header(skb); 3434cb47a86SStefano Brivio 3444cb47a86SStefano Brivio return skb->len; 3454cb47a86SStefano Brivio } 3464cb47a86SStefano Brivio 3474cb47a86SStefano Brivio /** 3484cb47a86SStefano Brivio * iptunnel_pmtud_check_icmpv6() - Trigger ICMPv6 reply if needed and allowed 3494cb47a86SStefano Brivio * @skb: Buffer being sent by encapsulation, L2 headers expected 3504cb47a86SStefano Brivio * @mtu: Network MTU for path 3514cb47a86SStefano Brivio * 3524cb47a86SStefano Brivio * Return: 0 for no ICMPv6 reply, length if built, negative value on error. 3534cb47a86SStefano Brivio */ 3544cb47a86SStefano Brivio static int iptunnel_pmtud_check_icmpv6(struct sk_buff *skb, int mtu) 3554cb47a86SStefano Brivio { 3564cb47a86SStefano Brivio const struct ipv6hdr *ip6h = ipv6_hdr(skb); 3574cb47a86SStefano Brivio int stype = ipv6_addr_type(&ip6h->saddr); 3584cb47a86SStefano Brivio u8 proto = ip6h->nexthdr; 3594cb47a86SStefano Brivio __be16 frag_off; 3604cb47a86SStefano Brivio int offset; 3614cb47a86SStefano Brivio 36277a2d673SStefano Brivio if (mtu < IPV6_MIN_MTU) 3634cb47a86SStefano Brivio return 0; 3644cb47a86SStefano Brivio 3654cb47a86SStefano Brivio if (stype == IPV6_ADDR_ANY || stype == IPV6_ADDR_MULTICAST || 3664cb47a86SStefano Brivio stype == IPV6_ADDR_LOOPBACK) 3674cb47a86SStefano Brivio return 0; 3684cb47a86SStefano Brivio 3694cb47a86SStefano Brivio offset = ipv6_skip_exthdr(skb, sizeof(struct ipv6hdr), &proto, 3704cb47a86SStefano Brivio &frag_off); 3714cb47a86SStefano Brivio if (offset < 0 || (frag_off & htons(~0x7))) 3724cb47a86SStefano Brivio return 0; 3734cb47a86SStefano Brivio 3744cb47a86SStefano Brivio if (proto == IPPROTO_ICMPV6) { 3754cb47a86SStefano Brivio struct icmp6hdr *icmp6h; 3764cb47a86SStefano Brivio 3774cb47a86SStefano Brivio if (!pskb_may_pull(skb, skb_network_header(skb) + 3784cb47a86SStefano Brivio offset + 1 - skb->data)) 3794cb47a86SStefano Brivio return 0; 3804cb47a86SStefano Brivio 3814cb47a86SStefano Brivio icmp6h = (struct icmp6hdr *)(skb_network_header(skb) + offset); 3824cb47a86SStefano Brivio if (icmpv6_is_err(icmp6h->icmp6_type) || 3834cb47a86SStefano Brivio icmp6h->icmp6_type == NDISC_REDIRECT) 3844cb47a86SStefano Brivio return 0; 3854cb47a86SStefano Brivio } 3864cb47a86SStefano Brivio 3874cb47a86SStefano Brivio return iptunnel_pmtud_build_icmpv6(skb, mtu); 3884cb47a86SStefano Brivio } 3894cb47a86SStefano Brivio #endif /* IS_ENABLED(CONFIG_IPV6) */ 3904cb47a86SStefano Brivio 3914cb47a86SStefano Brivio /** 3924cb47a86SStefano Brivio * skb_tunnel_check_pmtu() - Check, update PMTU and trigger ICMP reply as needed 3934cb47a86SStefano Brivio * @skb: Buffer being sent by encapsulation, L2 headers expected 3944cb47a86SStefano Brivio * @encap_dst: Destination for tunnel encapsulation (outer IP) 3954cb47a86SStefano Brivio * @headroom: Encapsulation header size, bytes 3964cb47a86SStefano Brivio * @reply: Build matching ICMP or ICMPv6 message as a result 3974cb47a86SStefano Brivio * 3984cb47a86SStefano Brivio * L2 tunnel implementations that can carry IP and can be directly bridged 3994cb47a86SStefano Brivio * (currently UDP tunnels) can't always rely on IP forwarding paths to handle 4004cb47a86SStefano Brivio * PMTU discovery. In the bridged case, ICMP or ICMPv6 messages need to be built 4014cb47a86SStefano Brivio * based on payload and sent back by the encapsulation itself. 4024cb47a86SStefano Brivio * 4034cb47a86SStefano Brivio * For routable interfaces, we just need to update the PMTU for the destination. 4044cb47a86SStefano Brivio * 4054cb47a86SStefano Brivio * Return: 0 if ICMP error not needed, length if built, negative value on error 4064cb47a86SStefano Brivio */ 4074cb47a86SStefano Brivio int skb_tunnel_check_pmtu(struct sk_buff *skb, struct dst_entry *encap_dst, 4084cb47a86SStefano Brivio int headroom, bool reply) 4094cb47a86SStefano Brivio { 4104cb47a86SStefano Brivio u32 mtu = dst_mtu(encap_dst) - headroom; 4114cb47a86SStefano Brivio 4124cb47a86SStefano Brivio if ((skb_is_gso(skb) && skb_gso_validate_network_len(skb, mtu)) || 413853a7614SEric Dumazet (!skb_is_gso(skb) && (skb->len - skb_network_offset(skb)) <= mtu)) 4144cb47a86SStefano Brivio return 0; 4154cb47a86SStefano Brivio 4164cb47a86SStefano Brivio skb_dst_update_pmtu_no_confirm(skb, mtu); 4174cb47a86SStefano Brivio 4184cb47a86SStefano Brivio if (!reply || skb->pkt_type == PACKET_HOST) 4194cb47a86SStefano Brivio return 0; 4204cb47a86SStefano Brivio 4214cb47a86SStefano Brivio if (skb->protocol == htons(ETH_P_IP)) 4224cb47a86SStefano Brivio return iptunnel_pmtud_check_icmp(skb, mtu); 4234cb47a86SStefano Brivio 4244cb47a86SStefano Brivio #if IS_ENABLED(CONFIG_IPV6) 4254cb47a86SStefano Brivio if (skb->protocol == htons(ETH_P_IPV6)) 4264cb47a86SStefano Brivio return iptunnel_pmtud_check_icmpv6(skb, mtu); 4274cb47a86SStefano Brivio #endif 4284cb47a86SStefano Brivio return 0; 4294cb47a86SStefano Brivio } 4304cb47a86SStefano Brivio EXPORT_SYMBOL(skb_tunnel_check_pmtu); 4314cb47a86SStefano Brivio 432a1c234f9SJiri Benc static const struct nla_policy ip_tun_policy[LWTUNNEL_IP_MAX + 1] = { 4337b6a70f7SXin Long [LWTUNNEL_IP_UNSPEC] = { .strict_start_type = LWTUNNEL_IP_OPTS }, 434a1c234f9SJiri Benc [LWTUNNEL_IP_ID] = { .type = NLA_U64 }, 435a1c234f9SJiri Benc [LWTUNNEL_IP_DST] = { .type = NLA_U32 }, 436a1c234f9SJiri Benc [LWTUNNEL_IP_SRC] = { .type = NLA_U32 }, 437a1c234f9SJiri Benc [LWTUNNEL_IP_TTL] = { .type = NLA_U8 }, 438a1c234f9SJiri Benc [LWTUNNEL_IP_TOS] = { .type = NLA_U8 }, 439a1c234f9SJiri Benc [LWTUNNEL_IP_FLAGS] = { .type = NLA_U16 }, 4404ece4778SXin Long [LWTUNNEL_IP_OPTS] = { .type = NLA_NESTED }, 4413093fbe7SThomas Graf }; 4423093fbe7SThomas Graf 4434ece4778SXin Long static const struct nla_policy ip_opts_policy[LWTUNNEL_IP_OPTS_MAX + 1] = { 4444ece4778SXin Long [LWTUNNEL_IP_OPTS_GENEVE] = { .type = NLA_NESTED }, 445edf31cbbSXin Long [LWTUNNEL_IP_OPTS_VXLAN] = { .type = NLA_NESTED }, 446b0a21810SXin Long [LWTUNNEL_IP_OPTS_ERSPAN] = { .type = NLA_NESTED }, 4474ece4778SXin Long }; 4484ece4778SXin Long 4494ece4778SXin Long static const struct nla_policy 4504ece4778SXin Long geneve_opt_policy[LWTUNNEL_IP_OPT_GENEVE_MAX + 1] = { 4514ece4778SXin Long [LWTUNNEL_IP_OPT_GENEVE_CLASS] = { .type = NLA_U16 }, 4524ece4778SXin Long [LWTUNNEL_IP_OPT_GENEVE_TYPE] = { .type = NLA_U8 }, 4534ece4778SXin Long [LWTUNNEL_IP_OPT_GENEVE_DATA] = { .type = NLA_BINARY, .len = 128 }, 4544ece4778SXin Long }; 4554ece4778SXin Long 456edf31cbbSXin Long static const struct nla_policy 457edf31cbbSXin Long vxlan_opt_policy[LWTUNNEL_IP_OPT_VXLAN_MAX + 1] = { 458edf31cbbSXin Long [LWTUNNEL_IP_OPT_VXLAN_GBP] = { .type = NLA_U32 }, 459edf31cbbSXin Long }; 460edf31cbbSXin Long 461b0a21810SXin Long static const struct nla_policy 462b0a21810SXin Long erspan_opt_policy[LWTUNNEL_IP_OPT_ERSPAN_MAX + 1] = { 463b0a21810SXin Long [LWTUNNEL_IP_OPT_ERSPAN_VER] = { .type = NLA_U8 }, 464b0a21810SXin Long [LWTUNNEL_IP_OPT_ERSPAN_INDEX] = { .type = NLA_U32 }, 465b0a21810SXin Long [LWTUNNEL_IP_OPT_ERSPAN_DIR] = { .type = NLA_U8 }, 466b0a21810SXin Long [LWTUNNEL_IP_OPT_ERSPAN_HWID] = { .type = NLA_U8 }, 467b0a21810SXin Long }; 468b0a21810SXin Long 4694ece4778SXin Long static int ip_tun_parse_opts_geneve(struct nlattr *attr, 4702f1d370bSXin Long struct ip_tunnel_info *info, int opts_len, 4714ece4778SXin Long struct netlink_ext_ack *extack) 4724ece4778SXin Long { 4734ece4778SXin Long struct nlattr *tb[LWTUNNEL_IP_OPT_GENEVE_MAX + 1]; 4744ece4778SXin Long int data_len, err; 4754ece4778SXin Long 476ed02551fSXin Long err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_GENEVE_MAX, attr, 477ed02551fSXin Long geneve_opt_policy, extack); 4784ece4778SXin Long if (err) 4794ece4778SXin Long return err; 4804ece4778SXin Long 4814ece4778SXin Long if (!tb[LWTUNNEL_IP_OPT_GENEVE_CLASS] || 4824ece4778SXin Long !tb[LWTUNNEL_IP_OPT_GENEVE_TYPE] || 4834ece4778SXin Long !tb[LWTUNNEL_IP_OPT_GENEVE_DATA]) 4844ece4778SXin Long return -EINVAL; 4854ece4778SXin Long 4864ece4778SXin Long attr = tb[LWTUNNEL_IP_OPT_GENEVE_DATA]; 4874ece4778SXin Long data_len = nla_len(attr); 4884ece4778SXin Long if (data_len % 4) 4894ece4778SXin Long return -EINVAL; 4904ece4778SXin Long 4914ece4778SXin Long if (info) { 4922f1d370bSXin Long struct geneve_opt *opt = ip_tunnel_info_opts(info) + opts_len; 4934ece4778SXin Long 4944ece4778SXin Long memcpy(opt->opt_data, nla_data(attr), data_len); 4954ece4778SXin Long opt->length = data_len / 4; 4964ece4778SXin Long attr = tb[LWTUNNEL_IP_OPT_GENEVE_CLASS]; 4974ece4778SXin Long opt->opt_class = nla_get_be16(attr); 4984ece4778SXin Long attr = tb[LWTUNNEL_IP_OPT_GENEVE_TYPE]; 4994ece4778SXin Long opt->type = nla_get_u8(attr); 5004ece4778SXin Long info->key.tun_flags |= TUNNEL_GENEVE_OPT; 5014ece4778SXin Long } 5024ece4778SXin Long 5034ece4778SXin Long return sizeof(struct geneve_opt) + data_len; 5044ece4778SXin Long } 5054ece4778SXin Long 506edf31cbbSXin Long static int ip_tun_parse_opts_vxlan(struct nlattr *attr, 5072f1d370bSXin Long struct ip_tunnel_info *info, int opts_len, 508edf31cbbSXin Long struct netlink_ext_ack *extack) 509edf31cbbSXin Long { 510edf31cbbSXin Long struct nlattr *tb[LWTUNNEL_IP_OPT_VXLAN_MAX + 1]; 511edf31cbbSXin Long int err; 512edf31cbbSXin Long 513ed02551fSXin Long err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_VXLAN_MAX, attr, 514ed02551fSXin Long vxlan_opt_policy, extack); 515edf31cbbSXin Long if (err) 516edf31cbbSXin Long return err; 517edf31cbbSXin Long 518edf31cbbSXin Long if (!tb[LWTUNNEL_IP_OPT_VXLAN_GBP]) 519edf31cbbSXin Long return -EINVAL; 520edf31cbbSXin Long 521edf31cbbSXin Long if (info) { 5222f1d370bSXin Long struct vxlan_metadata *md = 5232f1d370bSXin Long ip_tunnel_info_opts(info) + opts_len; 524edf31cbbSXin Long 525edf31cbbSXin Long attr = tb[LWTUNNEL_IP_OPT_VXLAN_GBP]; 526edf31cbbSXin Long md->gbp = nla_get_u32(attr); 527681d2cfbSXin Long md->gbp &= VXLAN_GBP_MASK; 528edf31cbbSXin Long info->key.tun_flags |= TUNNEL_VXLAN_OPT; 529edf31cbbSXin Long } 530edf31cbbSXin Long 531edf31cbbSXin Long return sizeof(struct vxlan_metadata); 532edf31cbbSXin Long } 533edf31cbbSXin Long 534b0a21810SXin Long static int ip_tun_parse_opts_erspan(struct nlattr *attr, 5352f1d370bSXin Long struct ip_tunnel_info *info, int opts_len, 536b0a21810SXin Long struct netlink_ext_ack *extack) 537b0a21810SXin Long { 538b0a21810SXin Long struct nlattr *tb[LWTUNNEL_IP_OPT_ERSPAN_MAX + 1]; 539b0a21810SXin Long int err; 5401841b982SXin Long u8 ver; 541b0a21810SXin Long 542ed02551fSXin Long err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_ERSPAN_MAX, attr, 543ed02551fSXin Long erspan_opt_policy, extack); 544b0a21810SXin Long if (err) 545b0a21810SXin Long return err; 546b0a21810SXin Long 547b0a21810SXin Long if (!tb[LWTUNNEL_IP_OPT_ERSPAN_VER]) 548b0a21810SXin Long return -EINVAL; 549b0a21810SXin Long 5501841b982SXin Long ver = nla_get_u8(tb[LWTUNNEL_IP_OPT_ERSPAN_VER]); 5511841b982SXin Long if (ver == 1) { 5521841b982SXin Long if (!tb[LWTUNNEL_IP_OPT_ERSPAN_INDEX]) 5531841b982SXin Long return -EINVAL; 5541841b982SXin Long } else if (ver == 2) { 5551841b982SXin Long if (!tb[LWTUNNEL_IP_OPT_ERSPAN_DIR] || 5561841b982SXin Long !tb[LWTUNNEL_IP_OPT_ERSPAN_HWID]) 5571841b982SXin Long return -EINVAL; 5581841b982SXin Long } else { 5591841b982SXin Long return -EINVAL; 5601841b982SXin Long } 5611841b982SXin Long 562b0a21810SXin Long if (info) { 5632f1d370bSXin Long struct erspan_metadata *md = 5642f1d370bSXin Long ip_tunnel_info_opts(info) + opts_len; 565b0a21810SXin Long 5661841b982SXin Long md->version = ver; 5671841b982SXin Long if (ver == 1) { 568b0a21810SXin Long attr = tb[LWTUNNEL_IP_OPT_ERSPAN_INDEX]; 569b0a21810SXin Long md->u.index = nla_get_be32(attr); 5701841b982SXin Long } else { 571b0a21810SXin Long attr = tb[LWTUNNEL_IP_OPT_ERSPAN_DIR]; 572b0a21810SXin Long md->u.md2.dir = nla_get_u8(attr); 573b0a21810SXin Long attr = tb[LWTUNNEL_IP_OPT_ERSPAN_HWID]; 574b0a21810SXin Long set_hwid(&md->u.md2, nla_get_u8(attr)); 575b0a21810SXin Long } 576b0a21810SXin Long 577b0a21810SXin Long info->key.tun_flags |= TUNNEL_ERSPAN_OPT; 578b0a21810SXin Long } 579b0a21810SXin Long 580b0a21810SXin Long return sizeof(struct erspan_metadata); 581b0a21810SXin Long } 582b0a21810SXin Long 5834ece4778SXin Long static int ip_tun_parse_opts(struct nlattr *attr, struct ip_tunnel_info *info, 5844ece4778SXin Long struct netlink_ext_ack *extack) 5854ece4778SXin Long { 586fda4fde2SJulian Wiedmann int err, rem, opt_len, opts_len = 0; 5872f1d370bSXin Long struct nlattr *nla; 588fda4fde2SJulian Wiedmann __be16 type = 0; 5894ece4778SXin Long 5904ece4778SXin Long if (!attr) 5914ece4778SXin Long return 0; 5924ece4778SXin Long 5932f1d370bSXin Long err = nla_validate(nla_data(attr), nla_len(attr), LWTUNNEL_IP_OPTS_MAX, 5944ece4778SXin Long ip_opts_policy, extack); 5954ece4778SXin Long if (err) 5964ece4778SXin Long return err; 5974ece4778SXin Long 5982f1d370bSXin Long nla_for_each_attr(nla, nla_data(attr), nla_len(attr), rem) { 5992f1d370bSXin Long switch (nla_type(nla)) { 6002f1d370bSXin Long case LWTUNNEL_IP_OPTS_GENEVE: 6012f1d370bSXin Long if (type && type != TUNNEL_GENEVE_OPT) 6022f1d370bSXin Long return -EINVAL; 6032f1d370bSXin Long opt_len = ip_tun_parse_opts_geneve(nla, info, opts_len, 6042f1d370bSXin Long extack); 6052f1d370bSXin Long if (opt_len < 0) 6062f1d370bSXin Long return opt_len; 6072f1d370bSXin Long opts_len += opt_len; 6082f1d370bSXin Long if (opts_len > IP_TUNNEL_OPTS_MAX) 6092f1d370bSXin Long return -EINVAL; 6102f1d370bSXin Long type = TUNNEL_GENEVE_OPT; 6112f1d370bSXin Long break; 6122f1d370bSXin Long case LWTUNNEL_IP_OPTS_VXLAN: 6132f1d370bSXin Long if (type) 6142f1d370bSXin Long return -EINVAL; 6152f1d370bSXin Long opt_len = ip_tun_parse_opts_vxlan(nla, info, opts_len, 6162f1d370bSXin Long extack); 6172f1d370bSXin Long if (opt_len < 0) 6182f1d370bSXin Long return opt_len; 6192f1d370bSXin Long opts_len += opt_len; 6202f1d370bSXin Long type = TUNNEL_VXLAN_OPT; 6212f1d370bSXin Long break; 6222f1d370bSXin Long case LWTUNNEL_IP_OPTS_ERSPAN: 6232f1d370bSXin Long if (type) 6242f1d370bSXin Long return -EINVAL; 6252f1d370bSXin Long opt_len = ip_tun_parse_opts_erspan(nla, info, opts_len, 6262f1d370bSXin Long extack); 6272f1d370bSXin Long if (opt_len < 0) 6282f1d370bSXin Long return opt_len; 6292f1d370bSXin Long opts_len += opt_len; 6302f1d370bSXin Long type = TUNNEL_ERSPAN_OPT; 6312f1d370bSXin Long break; 6322f1d370bSXin Long default: 6332f1d370bSXin Long return -EINVAL; 6342f1d370bSXin Long } 6352f1d370bSXin Long } 6364ece4778SXin Long 6372f1d370bSXin Long return opts_len; 6384ece4778SXin Long } 6394ece4778SXin Long 6404ece4778SXin Long static int ip_tun_get_optlen(struct nlattr *attr, 6414ece4778SXin Long struct netlink_ext_ack *extack) 6424ece4778SXin Long { 6434ece4778SXin Long return ip_tun_parse_opts(attr, NULL, extack); 6444ece4778SXin Long } 6454ece4778SXin Long 6464ece4778SXin Long static int ip_tun_set_opts(struct nlattr *attr, struct ip_tunnel_info *info, 6474ece4778SXin Long struct netlink_ext_ack *extack) 6484ece4778SXin Long { 6494ece4778SXin Long return ip_tun_parse_opts(attr, info, extack); 6504ece4778SXin Long } 6514ece4778SXin Long 652faee6769SAlexander Aring static int ip_tun_build_state(struct net *net, struct nlattr *attr, 653127eb7cdSTom Herbert unsigned int family, const void *cfg, 6549ae28727SDavid Ahern struct lwtunnel_state **ts, 6559ae28727SDavid Ahern struct netlink_ext_ack *extack) 6563093fbe7SThomas Graf { 657a1c234f9SJiri Benc struct nlattr *tb[LWTUNNEL_IP_MAX + 1]; 6584ece4778SXin Long struct lwtunnel_state *new_state; 6594ece4778SXin Long struct ip_tunnel_info *tun_info; 6604ece4778SXin Long int err, opt_len; 6613093fbe7SThomas Graf 6628cb08174SJohannes Berg err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP_MAX, attr, 6638cb08174SJohannes Berg ip_tun_policy, extack); 6643093fbe7SThomas Graf if (err < 0) 6653093fbe7SThomas Graf return err; 6663093fbe7SThomas Graf 6674ece4778SXin Long opt_len = ip_tun_get_optlen(tb[LWTUNNEL_IP_OPTS], extack); 6684ece4778SXin Long if (opt_len < 0) 6694ece4778SXin Long return opt_len; 6704ece4778SXin Long 6714ece4778SXin Long new_state = lwtunnel_state_alloc(sizeof(*tun_info) + opt_len); 6723093fbe7SThomas Graf if (!new_state) 6733093fbe7SThomas Graf return -ENOMEM; 6743093fbe7SThomas Graf 6753093fbe7SThomas Graf new_state->type = LWTUNNEL_ENCAP_IP; 6763093fbe7SThomas Graf 6773093fbe7SThomas Graf tun_info = lwt_tun_info(new_state); 6783093fbe7SThomas Graf 6794ece4778SXin Long err = ip_tun_set_opts(tb[LWTUNNEL_IP_OPTS], tun_info, extack); 6804ece4778SXin Long if (err < 0) { 6814ece4778SXin Long lwtstate_free(new_state); 6824ece4778SXin Long return err; 6834ece4778SXin Long } 6844ece4778SXin Long 6853d25eabbSwenxu #ifdef CONFIG_DST_CACHE 6863d25eabbSwenxu err = dst_cache_init(&tun_info->dst_cache, GFP_KERNEL); 6873d25eabbSwenxu if (err) { 6883d25eabbSwenxu lwtstate_free(new_state); 6893d25eabbSwenxu return err; 6903d25eabbSwenxu } 6913d25eabbSwenxu #endif 6923d25eabbSwenxu 693a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_ID]) 69430d3d83aSLance Richardson tun_info->key.tun_id = nla_get_be64(tb[LWTUNNEL_IP_ID]); 6953093fbe7SThomas Graf 696a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_DST]) 6977822ce73SHaishuang Yan tun_info->key.u.ipv4.dst = nla_get_in_addr(tb[LWTUNNEL_IP_DST]); 6983093fbe7SThomas Graf 699a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_SRC]) 7007822ce73SHaishuang Yan tun_info->key.u.ipv4.src = nla_get_in_addr(tb[LWTUNNEL_IP_SRC]); 7013093fbe7SThomas Graf 702a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_TTL]) 7037c383fb2SJiri Benc tun_info->key.ttl = nla_get_u8(tb[LWTUNNEL_IP_TTL]); 7043093fbe7SThomas Graf 705a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_TOS]) 7067c383fb2SJiri Benc tun_info->key.tos = nla_get_u8(tb[LWTUNNEL_IP_TOS]); 7073093fbe7SThomas Graf 708a1c234f9SJiri Benc if (tb[LWTUNNEL_IP_FLAGS]) 7090c06d166SXin Long tun_info->key.tun_flags |= 7100c06d166SXin Long (nla_get_be16(tb[LWTUNNEL_IP_FLAGS]) & 7110c06d166SXin Long ~TUNNEL_OPTIONS_PRESENT); 7123093fbe7SThomas Graf 7133093fbe7SThomas Graf tun_info->mode = IP_TUNNEL_INFO_TX; 7144ece4778SXin Long tun_info->options_len = opt_len; 7153093fbe7SThomas Graf 7163093fbe7SThomas Graf *ts = new_state; 7173093fbe7SThomas Graf 7183093fbe7SThomas Graf return 0; 7193093fbe7SThomas Graf } 7203093fbe7SThomas Graf 7213d25eabbSwenxu static void ip_tun_destroy_state(struct lwtunnel_state *lwtstate) 7223d25eabbSwenxu { 7233d25eabbSwenxu #ifdef CONFIG_DST_CACHE 7243d25eabbSwenxu struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate); 7253d25eabbSwenxu 7263d25eabbSwenxu dst_cache_destroy(&tun_info->dst_cache); 7273d25eabbSwenxu #endif 7283d25eabbSwenxu } 7293d25eabbSwenxu 7304ece4778SXin Long static int ip_tun_fill_encap_opts_geneve(struct sk_buff *skb, 7314ece4778SXin Long struct ip_tunnel_info *tun_info) 7324ece4778SXin Long { 7334ece4778SXin Long struct geneve_opt *opt; 7344ece4778SXin Long struct nlattr *nest; 7352f1d370bSXin Long int offset = 0; 7364ece4778SXin Long 7374ece4778SXin Long nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_GENEVE); 7384ece4778SXin Long if (!nest) 7394ece4778SXin Long return -ENOMEM; 7404ece4778SXin Long 7412f1d370bSXin Long while (tun_info->options_len > offset) { 7422f1d370bSXin Long opt = ip_tunnel_info_opts(tun_info) + offset; 7432f1d370bSXin Long if (nla_put_be16(skb, LWTUNNEL_IP_OPT_GENEVE_CLASS, 7442f1d370bSXin Long opt->opt_class) || 7454ece4778SXin Long nla_put_u8(skb, LWTUNNEL_IP_OPT_GENEVE_TYPE, opt->type) || 7464ece4778SXin Long nla_put(skb, LWTUNNEL_IP_OPT_GENEVE_DATA, opt->length * 4, 7474ece4778SXin Long opt->opt_data)) { 7484ece4778SXin Long nla_nest_cancel(skb, nest); 7494ece4778SXin Long return -ENOMEM; 7504ece4778SXin Long } 7512f1d370bSXin Long offset += sizeof(*opt) + opt->length * 4; 7522f1d370bSXin Long } 7534ece4778SXin Long 7544ece4778SXin Long nla_nest_end(skb, nest); 7554ece4778SXin Long return 0; 7564ece4778SXin Long } 7574ece4778SXin Long 758edf31cbbSXin Long static int ip_tun_fill_encap_opts_vxlan(struct sk_buff *skb, 759edf31cbbSXin Long struct ip_tunnel_info *tun_info) 760edf31cbbSXin Long { 761edf31cbbSXin Long struct vxlan_metadata *md; 762edf31cbbSXin Long struct nlattr *nest; 763edf31cbbSXin Long 764edf31cbbSXin Long nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_VXLAN); 765edf31cbbSXin Long if (!nest) 766edf31cbbSXin Long return -ENOMEM; 767edf31cbbSXin Long 768edf31cbbSXin Long md = ip_tunnel_info_opts(tun_info); 769edf31cbbSXin Long if (nla_put_u32(skb, LWTUNNEL_IP_OPT_VXLAN_GBP, md->gbp)) { 770edf31cbbSXin Long nla_nest_cancel(skb, nest); 771edf31cbbSXin Long return -ENOMEM; 772edf31cbbSXin Long } 773edf31cbbSXin Long 774edf31cbbSXin Long nla_nest_end(skb, nest); 775edf31cbbSXin Long return 0; 776edf31cbbSXin Long } 777edf31cbbSXin Long 778b0a21810SXin Long static int ip_tun_fill_encap_opts_erspan(struct sk_buff *skb, 779b0a21810SXin Long struct ip_tunnel_info *tun_info) 780b0a21810SXin Long { 781b0a21810SXin Long struct erspan_metadata *md; 782b0a21810SXin Long struct nlattr *nest; 783b0a21810SXin Long 784b0a21810SXin Long nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_ERSPAN); 785b0a21810SXin Long if (!nest) 786b0a21810SXin Long return -ENOMEM; 787b0a21810SXin Long 788b0a21810SXin Long md = ip_tunnel_info_opts(tun_info); 7893132174bSXin Long if (nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_VER, md->version)) 790b0a21810SXin Long goto err; 791b0a21810SXin Long 792b0a21810SXin Long if (md->version == 1 && 793b0a21810SXin Long nla_put_be32(skb, LWTUNNEL_IP_OPT_ERSPAN_INDEX, md->u.index)) 794b0a21810SXin Long goto err; 795b0a21810SXin Long 796b0a21810SXin Long if (md->version == 2 && 797b0a21810SXin Long (nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_DIR, md->u.md2.dir) || 798b0a21810SXin Long nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_HWID, 799b0a21810SXin Long get_hwid(&md->u.md2)))) 800b0a21810SXin Long goto err; 801b0a21810SXin Long 802b0a21810SXin Long nla_nest_end(skb, nest); 803b0a21810SXin Long return 0; 804b0a21810SXin Long err: 805b0a21810SXin Long nla_nest_cancel(skb, nest); 806b0a21810SXin Long return -ENOMEM; 807b0a21810SXin Long } 808b0a21810SXin Long 8094ece4778SXin Long static int ip_tun_fill_encap_opts(struct sk_buff *skb, int type, 8104ece4778SXin Long struct ip_tunnel_info *tun_info) 8114ece4778SXin Long { 8124ece4778SXin Long struct nlattr *nest; 8134ece4778SXin Long int err = 0; 8144ece4778SXin Long 8150c06d166SXin Long if (!(tun_info->key.tun_flags & TUNNEL_OPTIONS_PRESENT)) 8164ece4778SXin Long return 0; 8174ece4778SXin Long 8184ece4778SXin Long nest = nla_nest_start_noflag(skb, type); 8194ece4778SXin Long if (!nest) 8204ece4778SXin Long return -ENOMEM; 8214ece4778SXin Long 8224ece4778SXin Long if (tun_info->key.tun_flags & TUNNEL_GENEVE_OPT) 8234ece4778SXin Long err = ip_tun_fill_encap_opts_geneve(skb, tun_info); 824edf31cbbSXin Long else if (tun_info->key.tun_flags & TUNNEL_VXLAN_OPT) 825edf31cbbSXin Long err = ip_tun_fill_encap_opts_vxlan(skb, tun_info); 826b0a21810SXin Long else if (tun_info->key.tun_flags & TUNNEL_ERSPAN_OPT) 827b0a21810SXin Long err = ip_tun_fill_encap_opts_erspan(skb, tun_info); 8284ece4778SXin Long 8294ece4778SXin Long if (err) { 8304ece4778SXin Long nla_nest_cancel(skb, nest); 8314ece4778SXin Long return err; 8324ece4778SXin Long } 8334ece4778SXin Long 8344ece4778SXin Long nla_nest_end(skb, nest); 8354ece4778SXin Long return 0; 8364ece4778SXin Long } 8374ece4778SXin Long 8383093fbe7SThomas Graf static int ip_tun_fill_encap_info(struct sk_buff *skb, 8393093fbe7SThomas Graf struct lwtunnel_state *lwtstate) 8403093fbe7SThomas Graf { 8413093fbe7SThomas Graf struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate); 8423093fbe7SThomas Graf 843b46f6dedSNicolas Dichtel if (nla_put_be64(skb, LWTUNNEL_IP_ID, tun_info->key.tun_id, 844b46f6dedSNicolas Dichtel LWTUNNEL_IP_PAD) || 8457822ce73SHaishuang Yan nla_put_in_addr(skb, LWTUNNEL_IP_DST, tun_info->key.u.ipv4.dst) || 8467822ce73SHaishuang Yan nla_put_in_addr(skb, LWTUNNEL_IP_SRC, tun_info->key.u.ipv4.src) || 8477c383fb2SJiri Benc nla_put_u8(skb, LWTUNNEL_IP_TOS, tun_info->key.tos) || 8487c383fb2SJiri Benc nla_put_u8(skb, LWTUNNEL_IP_TTL, tun_info->key.ttl) || 8494ece4778SXin Long nla_put_be16(skb, LWTUNNEL_IP_FLAGS, tun_info->key.tun_flags) || 8504ece4778SXin Long ip_tun_fill_encap_opts(skb, LWTUNNEL_IP_OPTS, tun_info)) 8513093fbe7SThomas Graf return -ENOMEM; 8523093fbe7SThomas Graf 8533093fbe7SThomas Graf return 0; 8543093fbe7SThomas Graf } 8553093fbe7SThomas Graf 8564ece4778SXin Long static int ip_tun_opts_nlsize(struct ip_tunnel_info *info) 8574ece4778SXin Long { 8584ece4778SXin Long int opt_len; 8594ece4778SXin Long 8600c06d166SXin Long if (!(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT)) 8614ece4778SXin Long return 0; 8624ece4778SXin Long 8634ece4778SXin Long opt_len = nla_total_size(0); /* LWTUNNEL_IP_OPTS */ 8644ece4778SXin Long if (info->key.tun_flags & TUNNEL_GENEVE_OPT) { 8652f1d370bSXin Long struct geneve_opt *opt; 8662f1d370bSXin Long int offset = 0; 8674ece4778SXin Long 8682f1d370bSXin Long opt_len += nla_total_size(0); /* LWTUNNEL_IP_OPTS_GENEVE */ 8692f1d370bSXin Long while (info->options_len > offset) { 8702f1d370bSXin Long opt = ip_tunnel_info_opts(info) + offset; 8712f1d370bSXin Long opt_len += nla_total_size(2) /* OPT_GENEVE_CLASS */ 8724ece4778SXin Long + nla_total_size(1) /* OPT_GENEVE_TYPE */ 8734ece4778SXin Long + nla_total_size(opt->length * 4); 8744ece4778SXin Long /* OPT_GENEVE_DATA */ 8752f1d370bSXin Long offset += sizeof(*opt) + opt->length * 4; 8762f1d370bSXin Long } 877edf31cbbSXin Long } else if (info->key.tun_flags & TUNNEL_VXLAN_OPT) { 878edf31cbbSXin Long opt_len += nla_total_size(0) /* LWTUNNEL_IP_OPTS_VXLAN */ 879edf31cbbSXin Long + nla_total_size(4); /* OPT_VXLAN_GBP */ 880b0a21810SXin Long } else if (info->key.tun_flags & TUNNEL_ERSPAN_OPT) { 88158e8494eSXin Long struct erspan_metadata *md = ip_tunnel_info_opts(info); 88258e8494eSXin Long 883b0a21810SXin Long opt_len += nla_total_size(0) /* LWTUNNEL_IP_OPTS_ERSPAN */ 884b0a21810SXin Long + nla_total_size(1) /* OPT_ERSPAN_VER */ 88558e8494eSXin Long + (md->version == 1 ? nla_total_size(4) 88658e8494eSXin Long /* OPT_ERSPAN_INDEX (v1) */ 88758e8494eSXin Long : nla_total_size(1) + 88858e8494eSXin Long nla_total_size(1)); 88958e8494eSXin Long /* OPT_ERSPAN_DIR + HWID (v2) */ 8904ece4778SXin Long } 8914ece4778SXin Long 8924ece4778SXin Long return opt_len; 8934ece4778SXin Long } 8944ece4778SXin Long 8953093fbe7SThomas Graf static int ip_tun_encap_nlsize(struct lwtunnel_state *lwtstate) 8963093fbe7SThomas Graf { 897b46f6dedSNicolas Dichtel return nla_total_size_64bit(8) /* LWTUNNEL_IP_ID */ 898a1c234f9SJiri Benc + nla_total_size(4) /* LWTUNNEL_IP_DST */ 899a1c234f9SJiri Benc + nla_total_size(4) /* LWTUNNEL_IP_SRC */ 900a1c234f9SJiri Benc + nla_total_size(1) /* LWTUNNEL_IP_TOS */ 901a1c234f9SJiri Benc + nla_total_size(1) /* LWTUNNEL_IP_TTL */ 9024ece4778SXin Long + nla_total_size(2) /* LWTUNNEL_IP_FLAGS */ 9034ece4778SXin Long + ip_tun_opts_nlsize(lwt_tun_info(lwtstate)); 9044ece4778SXin Long /* LWTUNNEL_IP_OPTS */ 9053093fbe7SThomas Graf } 9063093fbe7SThomas Graf 9072d798499SJiri Benc static int ip_tun_cmp_encap(struct lwtunnel_state *a, struct lwtunnel_state *b) 9082d798499SJiri Benc { 9090eb8eb2fSXin Long struct ip_tunnel_info *info_a = lwt_tun_info(a); 9100eb8eb2fSXin Long struct ip_tunnel_info *info_b = lwt_tun_info(b); 9110eb8eb2fSXin Long 9120eb8eb2fSXin Long return memcmp(info_a, info_b, sizeof(info_a->key)) || 9130eb8eb2fSXin Long info_a->mode != info_b->mode || 9140eb8eb2fSXin Long info_a->options_len != info_b->options_len || 9150eb8eb2fSXin Long memcmp(ip_tunnel_info_opts(info_a), 9160eb8eb2fSXin Long ip_tunnel_info_opts(info_b), info_a->options_len); 9172d798499SJiri Benc } 9182d798499SJiri Benc 9193093fbe7SThomas Graf static const struct lwtunnel_encap_ops ip_tun_lwt_ops = { 9203093fbe7SThomas Graf .build_state = ip_tun_build_state, 9213d25eabbSwenxu .destroy_state = ip_tun_destroy_state, 9223093fbe7SThomas Graf .fill_encap = ip_tun_fill_encap_info, 9233093fbe7SThomas Graf .get_encap_size = ip_tun_encap_nlsize, 9242d798499SJiri Benc .cmp_encap = ip_tun_cmp_encap, 92588ff7334SRobert Shearman .owner = THIS_MODULE, 9263093fbe7SThomas Graf }; 9273093fbe7SThomas Graf 92832a2b002SJiri Benc static const struct nla_policy ip6_tun_policy[LWTUNNEL_IP6_MAX + 1] = { 9297b6a70f7SXin Long [LWTUNNEL_IP6_UNSPEC] = { .strict_start_type = LWTUNNEL_IP6_OPTS }, 93032a2b002SJiri Benc [LWTUNNEL_IP6_ID] = { .type = NLA_U64 }, 93132a2b002SJiri Benc [LWTUNNEL_IP6_DST] = { .len = sizeof(struct in6_addr) }, 93232a2b002SJiri Benc [LWTUNNEL_IP6_SRC] = { .len = sizeof(struct in6_addr) }, 93332a2b002SJiri Benc [LWTUNNEL_IP6_HOPLIMIT] = { .type = NLA_U8 }, 93432a2b002SJiri Benc [LWTUNNEL_IP6_TC] = { .type = NLA_U8 }, 93532a2b002SJiri Benc [LWTUNNEL_IP6_FLAGS] = { .type = NLA_U16 }, 9367b6a70f7SXin Long [LWTUNNEL_IP6_OPTS] = { .type = NLA_NESTED }, 93732a2b002SJiri Benc }; 93832a2b002SJiri Benc 939faee6769SAlexander Aring static int ip6_tun_build_state(struct net *net, struct nlattr *attr, 940127eb7cdSTom Herbert unsigned int family, const void *cfg, 9419ae28727SDavid Ahern struct lwtunnel_state **ts, 9429ae28727SDavid Ahern struct netlink_ext_ack *extack) 94332a2b002SJiri Benc { 94432a2b002SJiri Benc struct nlattr *tb[LWTUNNEL_IP6_MAX + 1]; 9454ece4778SXin Long struct lwtunnel_state *new_state; 9464ece4778SXin Long struct ip_tunnel_info *tun_info; 9474ece4778SXin Long int err, opt_len; 94832a2b002SJiri Benc 9498cb08174SJohannes Berg err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP6_MAX, attr, 9508cb08174SJohannes Berg ip6_tun_policy, extack); 95132a2b002SJiri Benc if (err < 0) 95232a2b002SJiri Benc return err; 95332a2b002SJiri Benc 9544ece4778SXin Long opt_len = ip_tun_get_optlen(tb[LWTUNNEL_IP6_OPTS], extack); 9554ece4778SXin Long if (opt_len < 0) 9564ece4778SXin Long return opt_len; 9574ece4778SXin Long 9584ece4778SXin Long new_state = lwtunnel_state_alloc(sizeof(*tun_info) + opt_len); 95932a2b002SJiri Benc if (!new_state) 96032a2b002SJiri Benc return -ENOMEM; 96132a2b002SJiri Benc 96232a2b002SJiri Benc new_state->type = LWTUNNEL_ENCAP_IP6; 96332a2b002SJiri Benc 96432a2b002SJiri Benc tun_info = lwt_tun_info(new_state); 96532a2b002SJiri Benc 9664ece4778SXin Long err = ip_tun_set_opts(tb[LWTUNNEL_IP6_OPTS], tun_info, extack); 9674ece4778SXin Long if (err < 0) { 9684ece4778SXin Long lwtstate_free(new_state); 9694ece4778SXin Long return err; 9704ece4778SXin Long } 9714ece4778SXin Long 97232a2b002SJiri Benc if (tb[LWTUNNEL_IP6_ID]) 97330d3d83aSLance Richardson tun_info->key.tun_id = nla_get_be64(tb[LWTUNNEL_IP6_ID]); 97432a2b002SJiri Benc 97532a2b002SJiri Benc if (tb[LWTUNNEL_IP6_DST]) 97632a2b002SJiri Benc tun_info->key.u.ipv6.dst = nla_get_in6_addr(tb[LWTUNNEL_IP6_DST]); 97732a2b002SJiri Benc 97832a2b002SJiri Benc if (tb[LWTUNNEL_IP6_SRC]) 97932a2b002SJiri Benc tun_info->key.u.ipv6.src = nla_get_in6_addr(tb[LWTUNNEL_IP6_SRC]); 98032a2b002SJiri Benc 98132a2b002SJiri Benc if (tb[LWTUNNEL_IP6_HOPLIMIT]) 98232a2b002SJiri Benc tun_info->key.ttl = nla_get_u8(tb[LWTUNNEL_IP6_HOPLIMIT]); 98332a2b002SJiri Benc 98432a2b002SJiri Benc if (tb[LWTUNNEL_IP6_TC]) 98532a2b002SJiri Benc tun_info->key.tos = nla_get_u8(tb[LWTUNNEL_IP6_TC]); 98632a2b002SJiri Benc 98732a2b002SJiri Benc if (tb[LWTUNNEL_IP6_FLAGS]) 9880c06d166SXin Long tun_info->key.tun_flags |= 9890c06d166SXin Long (nla_get_be16(tb[LWTUNNEL_IP6_FLAGS]) & 9900c06d166SXin Long ~TUNNEL_OPTIONS_PRESENT); 99132a2b002SJiri Benc 9927f9562a1SJiri Benc tun_info->mode = IP_TUNNEL_INFO_TX | IP_TUNNEL_INFO_IPV6; 9934ece4778SXin Long tun_info->options_len = opt_len; 99432a2b002SJiri Benc 99532a2b002SJiri Benc *ts = new_state; 99632a2b002SJiri Benc 99732a2b002SJiri Benc return 0; 99832a2b002SJiri Benc } 99932a2b002SJiri Benc 100032a2b002SJiri Benc static int ip6_tun_fill_encap_info(struct sk_buff *skb, 100132a2b002SJiri Benc struct lwtunnel_state *lwtstate) 100232a2b002SJiri Benc { 100332a2b002SJiri Benc struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate); 100432a2b002SJiri Benc 1005b46f6dedSNicolas Dichtel if (nla_put_be64(skb, LWTUNNEL_IP6_ID, tun_info->key.tun_id, 1006b46f6dedSNicolas Dichtel LWTUNNEL_IP6_PAD) || 100732a2b002SJiri Benc nla_put_in6_addr(skb, LWTUNNEL_IP6_DST, &tun_info->key.u.ipv6.dst) || 100832a2b002SJiri Benc nla_put_in6_addr(skb, LWTUNNEL_IP6_SRC, &tun_info->key.u.ipv6.src) || 1009995096a0SQuentin Armitage nla_put_u8(skb, LWTUNNEL_IP6_TC, tun_info->key.tos) || 1010995096a0SQuentin Armitage nla_put_u8(skb, LWTUNNEL_IP6_HOPLIMIT, tun_info->key.ttl) || 10114ece4778SXin Long nla_put_be16(skb, LWTUNNEL_IP6_FLAGS, tun_info->key.tun_flags) || 10124ece4778SXin Long ip_tun_fill_encap_opts(skb, LWTUNNEL_IP6_OPTS, tun_info)) 101332a2b002SJiri Benc return -ENOMEM; 101432a2b002SJiri Benc 101532a2b002SJiri Benc return 0; 101632a2b002SJiri Benc } 101732a2b002SJiri Benc 101832a2b002SJiri Benc static int ip6_tun_encap_nlsize(struct lwtunnel_state *lwtstate) 101932a2b002SJiri Benc { 1020b46f6dedSNicolas Dichtel return nla_total_size_64bit(8) /* LWTUNNEL_IP6_ID */ 102132a2b002SJiri Benc + nla_total_size(16) /* LWTUNNEL_IP6_DST */ 102232a2b002SJiri Benc + nla_total_size(16) /* LWTUNNEL_IP6_SRC */ 102332a2b002SJiri Benc + nla_total_size(1) /* LWTUNNEL_IP6_HOPLIMIT */ 102432a2b002SJiri Benc + nla_total_size(1) /* LWTUNNEL_IP6_TC */ 10254ece4778SXin Long + nla_total_size(2) /* LWTUNNEL_IP6_FLAGS */ 10264ece4778SXin Long + ip_tun_opts_nlsize(lwt_tun_info(lwtstate)); 10274ece4778SXin Long /* LWTUNNEL_IP6_OPTS */ 102832a2b002SJiri Benc } 102932a2b002SJiri Benc 103032a2b002SJiri Benc static const struct lwtunnel_encap_ops ip6_tun_lwt_ops = { 103132a2b002SJiri Benc .build_state = ip6_tun_build_state, 103232a2b002SJiri Benc .fill_encap = ip6_tun_fill_encap_info, 103332a2b002SJiri Benc .get_encap_size = ip6_tun_encap_nlsize, 103432a2b002SJiri Benc .cmp_encap = ip_tun_cmp_encap, 103588ff7334SRobert Shearman .owner = THIS_MODULE, 103632a2b002SJiri Benc }; 103732a2b002SJiri Benc 1038045a0fa0SThomas Graf void __init ip_tunnel_core_init(void) 10393093fbe7SThomas Graf { 1040fca5fdf6SDaniel Borkmann /* If you land here, make sure whether increasing ip_tunnel_info's 1041fca5fdf6SDaniel Borkmann * options_len is a reasonable choice with its usage in front ends 1042fca5fdf6SDaniel Borkmann * (f.e., it's part of flow keys, etc). 1043fca5fdf6SDaniel Borkmann */ 1044fca5fdf6SDaniel Borkmann BUILD_BUG_ON(IP_TUNNEL_OPTS_MAX != 255); 1045fca5fdf6SDaniel Borkmann 10463093fbe7SThomas Graf lwtunnel_encap_add_ops(&ip_tun_lwt_ops, LWTUNNEL_ENCAP_IP); 104732a2b002SJiri Benc lwtunnel_encap_add_ops(&ip6_tun_lwt_ops, LWTUNNEL_ENCAP_IP6); 10483093fbe7SThomas Graf } 1049e7030878SThomas Graf 10505263a98fSDavidlohr Bueso DEFINE_STATIC_KEY_FALSE(ip_tunnel_metadata_cnt); 1051e7030878SThomas Graf EXPORT_SYMBOL(ip_tunnel_metadata_cnt); 1052e7030878SThomas Graf 1053e7030878SThomas Graf void ip_tunnel_need_metadata(void) 1054e7030878SThomas Graf { 10555263a98fSDavidlohr Bueso static_branch_inc(&ip_tunnel_metadata_cnt); 1056e7030878SThomas Graf } 1057e7030878SThomas Graf EXPORT_SYMBOL_GPL(ip_tunnel_need_metadata); 1058e7030878SThomas Graf 1059e7030878SThomas Graf void ip_tunnel_unneed_metadata(void) 1060e7030878SThomas Graf { 10615263a98fSDavidlohr Bueso static_branch_dec(&ip_tunnel_metadata_cnt); 1062e7030878SThomas Graf } 1063e7030878SThomas Graf EXPORT_SYMBOL_GPL(ip_tunnel_unneed_metadata); 10642606aff9SJason A. Donenfeld 10652606aff9SJason A. Donenfeld /* Returns either the correct skb->protocol value, or 0 if invalid. */ 10662606aff9SJason A. Donenfeld __be16 ip_tunnel_parse_protocol(const struct sk_buff *skb) 10672606aff9SJason A. Donenfeld { 10682606aff9SJason A. Donenfeld if (skb_network_header(skb) >= skb->head && 10692606aff9SJason A. Donenfeld (skb_network_header(skb) + sizeof(struct iphdr)) <= skb_tail_pointer(skb) && 10702606aff9SJason A. Donenfeld ip_hdr(skb)->version == 4) 10712606aff9SJason A. Donenfeld return htons(ETH_P_IP); 10722606aff9SJason A. Donenfeld if (skb_network_header(skb) >= skb->head && 10732606aff9SJason A. Donenfeld (skb_network_header(skb) + sizeof(struct ipv6hdr)) <= skb_tail_pointer(skb) && 10742606aff9SJason A. Donenfeld ipv6_hdr(skb)->version == 6) 10752606aff9SJason A. Donenfeld return htons(ETH_P_IPV6); 10762606aff9SJason A. Donenfeld return 0; 10772606aff9SJason A. Donenfeld } 10782606aff9SJason A. Donenfeld EXPORT_SYMBOL(ip_tunnel_parse_protocol); 10792606aff9SJason A. Donenfeld 10802606aff9SJason A. Donenfeld const struct header_ops ip_tunnel_header_ops = { .parse_protocol = ip_tunnel_parse_protocol }; 10812606aff9SJason A. Donenfeld EXPORT_SYMBOL(ip_tunnel_header_ops); 1082537dd2d9SLiu Jian 1083537dd2d9SLiu Jian /* This function returns true when ENCAP attributes are present in the nl msg */ 1084537dd2d9SLiu Jian bool ip_tunnel_netlink_encap_parms(struct nlattr *data[], 1085537dd2d9SLiu Jian struct ip_tunnel_encap *encap) 1086537dd2d9SLiu Jian { 1087537dd2d9SLiu Jian bool ret = false; 1088537dd2d9SLiu Jian 1089537dd2d9SLiu Jian memset(encap, 0, sizeof(*encap)); 1090537dd2d9SLiu Jian 1091537dd2d9SLiu Jian if (!data) 1092537dd2d9SLiu Jian return ret; 1093537dd2d9SLiu Jian 1094537dd2d9SLiu Jian if (data[IFLA_IPTUN_ENCAP_TYPE]) { 1095537dd2d9SLiu Jian ret = true; 1096537dd2d9SLiu Jian encap->type = nla_get_u16(data[IFLA_IPTUN_ENCAP_TYPE]); 1097537dd2d9SLiu Jian } 1098537dd2d9SLiu Jian 1099537dd2d9SLiu Jian if (data[IFLA_IPTUN_ENCAP_FLAGS]) { 1100537dd2d9SLiu Jian ret = true; 1101537dd2d9SLiu Jian encap->flags = nla_get_u16(data[IFLA_IPTUN_ENCAP_FLAGS]); 1102537dd2d9SLiu Jian } 1103537dd2d9SLiu Jian 1104537dd2d9SLiu Jian if (data[IFLA_IPTUN_ENCAP_SPORT]) { 1105537dd2d9SLiu Jian ret = true; 1106537dd2d9SLiu Jian encap->sport = nla_get_be16(data[IFLA_IPTUN_ENCAP_SPORT]); 1107537dd2d9SLiu Jian } 1108537dd2d9SLiu Jian 1109537dd2d9SLiu Jian if (data[IFLA_IPTUN_ENCAP_DPORT]) { 1110537dd2d9SLiu Jian ret = true; 1111537dd2d9SLiu Jian encap->dport = nla_get_be16(data[IFLA_IPTUN_ENCAP_DPORT]); 1112537dd2d9SLiu Jian } 1113537dd2d9SLiu Jian 1114537dd2d9SLiu Jian return ret; 1115537dd2d9SLiu Jian } 1116537dd2d9SLiu Jian EXPORT_SYMBOL_GPL(ip_tunnel_netlink_encap_parms); 1117*b86fca80SLiu Jian 1118*b86fca80SLiu Jian void ip_tunnel_netlink_parms(struct nlattr *data[], 1119*b86fca80SLiu Jian struct ip_tunnel_parm *parms) 1120*b86fca80SLiu Jian { 1121*b86fca80SLiu Jian if (data[IFLA_IPTUN_LINK]) 1122*b86fca80SLiu Jian parms->link = nla_get_u32(data[IFLA_IPTUN_LINK]); 1123*b86fca80SLiu Jian 1124*b86fca80SLiu Jian if (data[IFLA_IPTUN_LOCAL]) 1125*b86fca80SLiu Jian parms->iph.saddr = nla_get_be32(data[IFLA_IPTUN_LOCAL]); 1126*b86fca80SLiu Jian 1127*b86fca80SLiu Jian if (data[IFLA_IPTUN_REMOTE]) 1128*b86fca80SLiu Jian parms->iph.daddr = nla_get_be32(data[IFLA_IPTUN_REMOTE]); 1129*b86fca80SLiu Jian 1130*b86fca80SLiu Jian if (data[IFLA_IPTUN_TTL]) { 1131*b86fca80SLiu Jian parms->iph.ttl = nla_get_u8(data[IFLA_IPTUN_TTL]); 1132*b86fca80SLiu Jian if (parms->iph.ttl) 1133*b86fca80SLiu Jian parms->iph.frag_off = htons(IP_DF); 1134*b86fca80SLiu Jian } 1135*b86fca80SLiu Jian 1136*b86fca80SLiu Jian if (data[IFLA_IPTUN_TOS]) 1137*b86fca80SLiu Jian parms->iph.tos = nla_get_u8(data[IFLA_IPTUN_TOS]); 1138*b86fca80SLiu Jian 1139*b86fca80SLiu Jian if (!data[IFLA_IPTUN_PMTUDISC] || nla_get_u8(data[IFLA_IPTUN_PMTUDISC])) 1140*b86fca80SLiu Jian parms->iph.frag_off = htons(IP_DF); 1141*b86fca80SLiu Jian 1142*b86fca80SLiu Jian if (data[IFLA_IPTUN_FLAGS]) 1143*b86fca80SLiu Jian parms->i_flags = nla_get_be16(data[IFLA_IPTUN_FLAGS]); 1144*b86fca80SLiu Jian 1145*b86fca80SLiu Jian if (data[IFLA_IPTUN_PROTO]) 1146*b86fca80SLiu Jian parms->iph.protocol = nla_get_u8(data[IFLA_IPTUN_PROTO]); 1147*b86fca80SLiu Jian } 1148*b86fca80SLiu Jian EXPORT_SYMBOL_GPL(ip_tunnel_netlink_parms); 1149