| 1 |
|
| 2 | #include <linux/module.h>
|
| 3 | #include <linux/netfilter/nf_tables.h>
|
| 4 | #include <net/netfilter/nf_tables.h>
|
| 5 | #include <net/netfilter/nf_tables_core.h>
|
| 6 | #include <net/netfilter/nf_socket.h>
|
| 7 | #include <net/inet_sock.h>
|
| 8 | #include <net/tcp.h>
|
| 9 |
|
| 10 | struct nft_socket {
|
| 11 | enum nft_socket_keys key:8;
|
| 12 | u8 level;
|
| 13 | u8 level_user;
|
| 14 | u8 len;
|
| 15 | union {
|
| 16 | u8 dreg;
|
| 17 | };
|
| 18 | };
|
| 19 |
|
| 20 | static void nft_socket_wildcard(const struct nft_pktinfo *pkt,
|
| 21 | struct nft_regs *regs, struct sock *sk,
|
| 22 | u32 *dest)
|
| 23 | {
|
| 24 | switch (nft_pf(pkt)) {
|
| 25 | case NFPROTO_IPV4:
|
| 26 | nft_reg_store8(dest, inet_sk(sk)->inet_rcv_saddr == 0);
|
| 27 | break;
|
| 28 | #if IS_ENABLED(CONFIG_NF_TABLES_IPV6)
|
| 29 | case NFPROTO_IPV6:
|
| 30 | nft_reg_store8(dest, ipv6_addr_any(&sk->sk_v6_rcv_saddr));
|
| 31 | break;
|
| 32 | #endif
|
| 33 | default:
|
| 34 | regs->verdict.code = NFT_BREAK;
|
| 35 | return;
|
| 36 | }
|
| 37 | }
|
| 38 |
|
| 39 | #ifdef CONFIG_SOCK_CGROUP_DATA
|
| 40 | static noinline bool
|
| 41 | nft_sock_get_eval_cgroupv2(u32 *dest, struct sock *sk, const struct nft_pktinfo *pkt, u32 level)
|
| 42 | {
|
| 43 | struct cgroup *cgrp;
|
| 44 | u64 cgid;
|
| 45 |
|
| 46 | if (!sk_fullsock(sk))
|
| 47 | return false;
|
| 48 |
|
| 49 | cgrp = cgroup_ancestor(sock_cgroup_ptr(&sk->sk_cgrp_data), level);
|
| 50 | if (!cgrp)
|
| 51 | return false;
|
| 52 |
|
| 53 | cgid = cgroup_id(cgrp);
|
| 54 | memcpy(dest, &cgid, sizeof(u64));
|
| 55 | return true;
|
| 56 | }
|
| 57 |
|
| 58 |
|
| 59 | static noinline int nft_socket_cgroup_subtree_level(void)
|
| 60 | {
|
| 61 | struct cgroup *cgrp = cgroup_get_from_path("/");
|
| 62 | int level;
|
| 63 |
|
| 64 | if (IS_ERR(cgrp))
|
| 65 | return PTR_ERR(cgrp);
|
| 66 |
|
| 67 | level = cgrp->level;
|
| 68 |
|
| 69 | cgroup_put(cgrp);
|
| 70 |
|
| 71 | if (level > 255)
|
| 72 | return -ERANGE;
|
| 73 |
|
| 74 | if (unlikely(level < 0)) {
|
| 75 | DEBUG_NET_WARN_ON_ONCE(1);
|
| 76 | return -EINVAL;
|
| 77 | }
|
| 78 |
|
| 79 | return level;
|
| 80 | }
|
| 81 | #endif
|
| 82 |
|
| 83 | static struct sock *nft_socket_do_lookup(const struct nft_pktinfo *pkt)
|
| 84 | {
|
| 85 | const struct net_device *indev = nft_in(pkt);
|
| 86 | const struct sk_buff *skb = pkt->skb;
|
| 87 | struct sock *sk = NULL;
|
| 88 |
|
| 89 | if (!indev)
|
| 90 | return NULL;
|
| 91 |
|
| 92 | switch (nft_pf(pkt)) {
|
| 93 | case NFPROTO_IPV4:
|
| 94 | sk = nf_sk_lookup_slow_v4(nft_net(pkt), skb, indev);
|
| 95 | break;
|
| 96 | #if IS_ENABLED(CONFIG_NF_TABLES_IPV6)
|
| 97 | case NFPROTO_IPV6:
|
| 98 | sk = nf_sk_lookup_slow_v6(nft_net(pkt), skb, indev);
|
| 99 | break;
|
| 100 | #endif
|
| 101 | default:
|
| 102 | DEBUG_NET_WARN_ON_ONCE(1);
|
| 103 | break;
|
| 104 | }
|
| 105 |
|
| 106 | return sk;
|
| 107 | }
|
| 108 |
|
| 109 | static void nft_socket_eval(const struct nft_expr *expr,
|
| 110 | struct nft_regs *regs,
|
| 111 | const struct nft_pktinfo *pkt)
|
| 112 | {
|
| 113 | const struct nft_socket *priv = nft_expr_priv(expr);
|
| 114 | struct sk_buff *skb = pkt->skb;
|
| 115 | struct sock *sk = skb->sk;
|
| 116 | u32 *dest = ®s->data[priv->dreg];
|
| 117 |
|
| 118 | if (sk && !net_eq(nft_net(pkt), sock_net(sk)))
|
| 119 | sk = NULL;
|
| 120 |
|
| 121 | if (!sk)
|
| 122 | sk = nft_socket_do_lookup(pkt);
|
| 123 |
|
| 124 | if (!sk) {
|
| 125 | regs->verdict.code = NFT_BREAK;
|
| 126 | return;
|
| 127 | }
|
| 128 |
|
| 129 | switch(priv->key) {
|
| 130 | case NFT_SOCKET_TRANSPARENT:
|
| 131 | nft_reg_store8(dest, inet_sk_transparent(sk));
|
| 132 | break;
|
| 133 | case NFT_SOCKET_MARK:
|
| 134 | if (sk_fullsock(sk)) {
|
| 135 | *dest = READ_ONCE(sk->sk_mark);
|
| 136 | } else {
|
| 137 | regs->verdict.code = NFT_BREAK;
|
| 138 | goto out_put_sk;
|
| 139 | }
|
| 140 | break;
|
| 141 | case NFT_SOCKET_WILDCARD:
|
| 142 | if (!sk_fullsock(sk)) {
|
| 143 | regs->verdict.code = NFT_BREAK;
|
| 144 | goto out_put_sk;
|
| 145 | }
|
| 146 | nft_socket_wildcard(pkt, regs, sk, dest);
|
| 147 | break;
|
| 148 | #ifdef CONFIG_SOCK_CGROUP_DATA
|
| 149 | case NFT_SOCKET_CGROUPV2:
|
| 150 | if (!nft_sock_get_eval_cgroupv2(dest, sk, pkt, priv->level)) {
|
| 151 | regs->verdict.code = NFT_BREAK;
|
| 152 | goto out_put_sk;
|
| 153 | }
|
| 154 | break;
|
| 155 | #endif
|
| 156 | default:
|
| 157 | DEBUG_NET_WARN_ON_ONCE(1);
|
| 158 | regs->verdict.code = NFT_BREAK;
|
| 159 | }
|
| 160 |
|
| 161 | out_put_sk:
|
| 162 | if (sk != skb->sk)
|
| 163 | sock_gen_put(sk);
|
| 164 | }
|
| 165 |
|
| 166 | static const struct nla_policy nft_socket_policy[NFTA_SOCKET_MAX + 1] = {
|
| 167 | [NFTA_SOCKET_KEY] = NLA_POLICY_MAX(NLA_BE32, 255),
|
| 168 | [NFTA_SOCKET_DREG] = NLA_POLICY_MAX(NLA_BE32, NFT_REG32_MAX),
|
| 169 | [NFTA_SOCKET_LEVEL] = NLA_POLICY_MAX(NLA_BE32, 255),
|
| 170 | };
|
| 171 |
|
| 172 | static int nft_socket_init(const struct nft_ctx *ctx,
|
| 173 | const struct nft_expr *expr,
|
| 174 | const struct nlattr * const tb[])
|
| 175 | {
|
| 176 | struct nft_socket *priv = nft_expr_priv(expr);
|
| 177 | unsigned int len;
|
| 178 |
|
| 179 | if (!tb[NFTA_SOCKET_DREG] || !tb[NFTA_SOCKET_KEY])
|
| 180 | return -EINVAL;
|
| 181 |
|
| 182 | switch(ctx->family) {
|
| 183 | case NFPROTO_IPV4:
|
| 184 | #if IS_ENABLED(CONFIG_NF_TABLES_IPV6)
|
| 185 | case NFPROTO_IPV6:
|
| 186 | #endif
|
| 187 | case NFPROTO_INET:
|
| 188 | break;
|
| 189 | default:
|
| 190 | return -EOPNOTSUPP;
|
| 191 | }
|
| 192 |
|
| 193 | priv->key = ntohl(nla_get_be32(tb[NFTA_SOCKET_KEY]));
|
| 194 | switch(priv->key) {
|
| 195 | case NFT_SOCKET_TRANSPARENT:
|
| 196 | case NFT_SOCKET_WILDCARD:
|
| 197 | len = sizeof(u8);
|
| 198 | break;
|
| 199 | case NFT_SOCKET_MARK:
|
| 200 | len = sizeof(u32);
|
| 201 | break;
|
| 202 | #ifdef CONFIG_SOCK_CGROUP_DATA
|
| 203 | case NFT_SOCKET_CGROUPV2: {
|
| 204 | unsigned int level;
|
| 205 | int err;
|
| 206 |
|
| 207 | if (!tb[NFTA_SOCKET_LEVEL])
|
| 208 | return -EINVAL;
|
| 209 |
|
| 210 | level = ntohl(nla_get_be32(tb[NFTA_SOCKET_LEVEL]));
|
| 211 | if (level > 255)
|
| 212 | return -EOPNOTSUPP;
|
| 213 |
|
| 214 | err = nft_socket_cgroup_subtree_level();
|
| 215 | if (err < 0)
|
| 216 | return err;
|
| 217 |
|
| 218 | priv->level_user = level;
|
| 219 |
|
| 220 | level += err;
|
| 221 |
|
| 222 | if (level > 255)
|
| 223 | return -EOPNOTSUPP;
|
| 224 |
|
| 225 | priv->level = level;
|
| 226 | len = sizeof(u64);
|
| 227 | break;
|
| 228 | }
|
| 229 | #endif
|
| 230 | default:
|
| 231 | return -EOPNOTSUPP;
|
| 232 | }
|
| 233 |
|
| 234 | priv->len = len;
|
| 235 | return nft_parse_register_store(ctx, tb[NFTA_SOCKET_DREG], &priv->dreg,
|
| 236 | NULL, NFT_DATA_VALUE, len);
|
| 237 | }
|
| 238 |
|
| 239 | static int nft_socket_dump(struct sk_buff *skb,
|
| 240 | const struct nft_expr *expr, bool reset)
|
| 241 | {
|
| 242 | const struct nft_socket *priv = nft_expr_priv(expr);
|
| 243 |
|
| 244 | if (nla_put_be32(skb, NFTA_SOCKET_KEY, htonl(priv->key)))
|
| 245 | return -1;
|
| 246 | if (nft_dump_register(skb, NFTA_SOCKET_DREG, priv->dreg))
|
| 247 | return -1;
|
| 248 | if (priv->key == NFT_SOCKET_CGROUPV2 &&
|
| 249 | nla_put_be32(skb, NFTA_SOCKET_LEVEL, htonl(priv->level_user)))
|
| 250 | return -1;
|
| 251 | return 0;
|
| 252 | }
|
| 253 |
|
| 254 | static int nft_socket_validate(const struct nft_ctx *ctx,
|
| 255 | const struct nft_expr *expr)
|
| 256 | {
|
| 257 | if (ctx->family != NFPROTO_IPV4 &&
|
| 258 | ctx->family != NFPROTO_IPV6 &&
|
| 259 | ctx->family != NFPROTO_INET)
|
| 260 | return -EOPNOTSUPP;
|
| 261 |
|
| 262 | return nft_chain_validate_hooks(ctx->chain,
|
| 263 | (1 << NF_INET_PRE_ROUTING) |
|
| 264 | (1 << NF_INET_LOCAL_IN) |
|
| 265 | (1 << NF_INET_LOCAL_OUT));
|
| 266 | }
|
| 267 |
|
| 268 | static struct nft_expr_type nft_socket_type;
|
| 269 | static const struct nft_expr_ops nft_socket_ops = {
|
| 270 | .type = &nft_socket_type,
|
| 271 | .size = NFT_EXPR_SIZE(sizeof(struct nft_socket)),
|
| 272 | .eval = nft_socket_eval,
|
| 273 | .init = nft_socket_init,
|
| 274 | .dump = nft_socket_dump,
|
| 275 | .validate = nft_socket_validate,
|
| 276 | };
|
| 277 |
|
| 278 | static struct nft_expr_type nft_socket_type __read_mostly = {
|
| 279 | .name = "socket",
|
| 280 | .ops = &nft_socket_ops,
|
| 281 | .policy = nft_socket_policy,
|
| 282 | .maxattr = NFTA_SOCKET_MAX,
|
| 283 | .owner = THIS_MODULE,
|
| 284 | };
|
| 285 |
|
| 286 | static int __init nft_socket_module_init(void)
|
| 287 | {
|
| 288 | return nft_register_expr(&nft_socket_type);
|
| 289 | }
|
| 290 |
|
| 291 | static void __exit nft_socket_module_exit(void)
|
| 292 | {
|
| 293 | nft_unregister_expr(&nft_socket_type);
|
| 294 | }
|
| 295 |
|
| 296 | module_init(nft_socket_module_init);
|
| 297 | module_exit(nft_socket_module_exit);
|
| 298 |
|
| 299 | MODULE_LICENSE("GPL");
|
| 300 | MODULE_AUTHOR("Máté Eckl");
|
| 301 | MODULE_DESCRIPTION("nf_tables socket match module");
|
| 302 | MODULE_ALIAS_NFT_EXPR("socket");
|