linux-2.6-522-iptables-connection-tagging.patch

   1 diff -Nurb linux-2.6.27-521/include/linux/netfilter/xt_MARK.h linux-2.6.27-522-ol/include/linux/netfilter/xt_MARK.h
   2 --- linux-2.6.27-521/include/linux/netfilter/xt_MARK.h  2008-10-09 18:13:53.000000000 -0400
   3 +++ linux-2.6.27-522-ol/include/linux/netfilter/xt_MARK.h       2009-06-02 10:02:16.000000000 -0400
   4 @@ -11,6 +11,7 @@
   5         XT_MARK_SET=0,
   6         XT_MARK_AND,
   7         XT_MARK_OR,
   8 +       XT_MARK_COPYXID,
   9  };
  10
  11  struct xt_mark_target_info_v1 {
  12 diff -Nurb linux-2.6.27-521/include/linux/netfilter/xt_SETXID.h linux-2.6.27-522-ol/include/linux/netfilter/xt_SETXID.h
  13 --- linux-2.6.27-521/include/linux/netfilter/xt_SETXID.h        1969-12-31 19:00:00.000000000 -0500
  14 +++ linux-2.6.27-522-ol/include/linux/netfilter/xt_SETXID.h     2009-06-02 10:02:16.000000000 -0400
  15 @@ -0,0 +1,14 @@
  16 +#ifndef _XT_SETXID_H_target
  17 +#define _XT_SETXID_H_target
  18 +
  19 +/* Version 1 */
  20 +enum {
  21 +       XT_SET_PACKET_XID=0
  22 +};
  23 +
  24 +struct xt_setxid_target_info_v1 {
  25 +       unsigned long mark;
  26 +       u_int8_t mode;
  27 +};
  28 +
  29 +#endif /*_XT_SETXID_H_target*/
  30 diff -Nurb linux-2.6.27-521/include/linux/netfilter_ipv4/ipt_MARK.h linux-2.6.27-522-ol/include/linux/netfilter_ipv4/ipt_MARK.h
  31 --- linux-2.6.27-521/include/linux/netfilter_ipv4/ipt_MARK.h    2008-10-09 18:13:53.000000000 -0400
  32 +++ linux-2.6.27-522-ol/include/linux/netfilter_ipv4/ipt_MARK.h 2009-06-02 10:02:16.000000000 -0400
  33 @@ -12,6 +12,7 @@
  34  #define IPT_MARK_SET   XT_MARK_SET
  35  #define IPT_MARK_AND   XT_MARK_AND
  36  #define        IPT_MARK_OR     XT_MARK_OR
  37 +#define IPT_MARK_COPYXID       XT_MARK_COPYXID
  38
  39  #define ipt_mark_target_info_v1 xt_mark_target_info_v1
  40
  41 diff -Nurb linux-2.6.27-521/include/linux/netfilter_ipv4/ipt_SETXID.h linux-2.6.27-522-ol/include/linux/netfilter_ipv4/ipt_SETXID.h
  42 --- linux-2.6.27-521/include/linux/netfilter_ipv4/ipt_SETXID.h  1969-12-31 19:00:00.000000000 -0500
  43 +++ linux-2.6.27-522-ol/include/linux/netfilter_ipv4/ipt_SETXID.h       2009-06-02 10:02:16.000000000 -0400
  44 @@ -0,0 +1,13 @@
  45 +#ifndef _IPT_SETXID_H_target
  46 +#define _IPT_SETXID_H_target
  47 +
  48 +/* Backwards compatibility for old userspace */
  49 +
  50 +#include <linux/netfilter/xt_SETXID.h>
  51 +
  52 +/* Version 1 */
  53 +#define IPT_SET_PACKET_XID     XT_SET_PACKET_XID
  54 +
  55 +#define ipt_setxid_target_info_v1 xt_setxid_target_info_v1
  56 +
  57 +#endif /*_IPT_SETXID_H_target*/
  58 diff -Nurb linux-2.6.27-521/include/net/netfilter/nf_conntrack.h linux-2.6.27-522-ol/include/net/netfilter/nf_conntrack.h
  59 --- linux-2.6.27-521/include/net/netfilter/nf_conntrack.h       2008-10-09 18:13:53.000000000 -0400
  60 +++ linux-2.6.27-522-ol/include/net/netfilter/nf_conntrack.h    2009-06-02 10:02:16.000000000 -0400
  61 @@ -121,6 +121,9 @@
  62         /* Storage reserved for other modules: */
  63         union nf_conntrack_proto proto;
  64
  65 +       /* PLANETLAB. VNET-specific */
  66 +       int xid[IP_CT_DIR_MAX];
  67 +
  68         /* Extensions */
  69         struct nf_ct_ext *ext;
  70
  71 diff -Nurb linux-2.6.27-521/net/netfilter/Kconfig linux-2.6.27-522-ol/net/netfilter/Kconfig
  72 --- linux-2.6.27-521/net/netfilter/Kconfig      2008-10-09 18:13:53.000000000 -0400
  73 +++ linux-2.6.27-522-ol/net/netfilter/Kconfig   2009-06-02 10:02:16.000000000 -0400
  74 @@ -477,6 +477,13 @@
  75           This option adds a "TCPOPTSTRIP" target, which allows you to strip
  76           TCP options from TCP packets.
  77
  78 +config NETFILTER_XT_TARGET_SETXID
  79 +       tristate '"SETXID" target support'
  80 +       depends on NETFILTER_XTABLES
  81 +       help
  82 +         This option adds a `SETXID' target, which allows you to alter the
  83 +         xid of a socket.
  84 +
  85  config NETFILTER_XT_MATCH_COMMENT
  86         tristate  '"comment" match support'
  87         depends on NETFILTER_XTABLES
  88 diff -Nurb linux-2.6.27-521/net/netfilter/Makefile linux-2.6.27-522-ol/net/netfilter/Makefile
  89 --- linux-2.6.27-521/net/netfilter/Makefile     2008-10-09 18:13:53.000000000 -0400
  90 +++ linux-2.6.27-522-ol/net/netfilter/Makefile  2009-06-02 10:02:16.000000000 -0400
  91 @@ -38,6 +38,7 @@
  92  obj-$(CONFIG_NETFILTER_XTABLES) += x_tables.o xt_tcpudp.o
  93
  94  # targets
  95 +obj-$(CONFIG_NETFILTER_XT_TARGET_SETXID) += xt_SETXID.o
  96  obj-$(CONFIG_NETFILTER_XT_TARGET_CLASSIFY) += xt_CLASSIFY.o
  97  obj-$(CONFIG_NETFILTER_XT_TARGET_CONNMARK) += xt_CONNMARK.o
  98  obj-$(CONFIG_NETFILTER_XT_TARGET_CONNSECMARK) += xt_CONNSECMARK.o
  99 diff -Nurb linux-2.6.27-521/net/netfilter/nf_conntrack_core.c linux-2.6.27-522-ol/net/netfilter/nf_conntrack_core.c
 100 --- linux-2.6.27-521/net/netfilter/nf_conntrack_core.c  2008-10-09 18:13:53.000000000 -0400
 101 +++ linux-2.6.27-522-ol/net/netfilter/nf_conntrack_core.c       2009-06-02 10:02:16.000000000 -0400
 102 @@ -595,6 +595,9 @@
 103         /* Overload tuple linked list to put us in unconfirmed list. */
 104         hlist_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnode, &unconfirmed);
 105
 106 +       ct->xid[IP_CT_DIR_ORIGINAL] = -1;
 107 +       ct->xid[IP_CT_DIR_REPLY] = -1;
 108 +
 109         spin_unlock_bh(&nf_conntrack_lock);
 110
 111         if (exp) {
 112 diff -Nurb linux-2.6.27-521/net/netfilter/xt_MARK.c linux-2.6.27-522/net/netfilter/xt_MARK.c
 113 --- linux-2.6.27-521/net/netfilter/xt_MARK.c    2008-10-09 18:13:53.000000000 -0400
 114 +++ linux-2.6.27-522/net/netfilter/xt_MARK.c    2009-06-02 11:12:59.000000000 -0400
 115 @@ -13,7 +13,13 @@
 116  #include <linux/module.h>
 117  #include <linux/skbuff.h>
 118  #include <linux/ip.h>
 119 +#include <net/udp.h>
 120  #include <net/checksum.h>
 121 +#include <net/route.h>
 122 +#include <net/inet_hashtables.h>
 123 +#include <net/net_namespace.h>
 124 +
 125 +#include <net/netfilter/nf_conntrack.h>
 126
 127  #include <linux/netfilter/x_tables.h>
 128  #include <linux/netfilter/xt_MARK.h>
 129 @@ -24,6 +30,8 @@
 130  MODULE_ALIAS("ipt_MARK");
 131  MODULE_ALIAS("ip6t_MARK");
 132
 133 +extern DEFINE_PER_CPU(int, sknid_elevator);
 134 +
 135  static unsigned int
 136  mark_tg_v0(struct sk_buff *skb, const struct net_device *in,
 137             const struct net_device *out, unsigned int hooknum,
 138 @@ -61,14 +69,242 @@
 139         return XT_CONTINUE;
 140  }
 141
 142 +#define PEERCRED_SET(x) ((x!=0) && (x!=(unsigned int)-1))
 143 +
 144 +
 145 +static inline u_int16_t
 146 +get_dst_port(struct nf_conntrack_tuple *tuple)
 147 +{
 148 +       switch (tuple->dst.protonum) {
 149 +       case IPPROTO_GRE:
 150 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 151 +               return tuple->dst.u.gre.key;
 152 +       case IPPROTO_ICMP:
 153 +               /* Bind on ICMP echo ID */
 154 +               return tuple->src.u.icmp.id;
 155 +       case IPPROTO_TCP:
 156 +               return tuple->dst.u.tcp.port;
 157 +       case IPPROTO_UDP:
 158 +               return tuple->dst.u.udp.port;
 159 +       default:
 160 +               return tuple->dst.u.all;
 161 +       }
 162 +}
 163 +
 164 +static inline u_int16_t
 165 +get_src_port(struct nf_conntrack_tuple *tuple)
 166 +{
 167 +       switch (tuple->dst.protonum) {
 168 +       case IPPROTO_GRE:
 169 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 170 +               return htons(ntohl(tuple->src.u.gre.key));
 171 +       case IPPROTO_ICMP:
 172 +               /* Bind on ICMP echo ID */
 173 +               return tuple->src.u.icmp.id;
 174 +       case IPPROTO_TCP:
 175 +               return tuple->src.u.tcp.port;
 176 +       case IPPROTO_UDP:
 177 +               return tuple->src.u.udp.port;
 178 +       default:
 179 +               return tuple->src.u.all;
 180 +       }
 181 +}
 182 +
 183 +static struct sock *__udp4_lib_lookup(__be32 saddr, __be16 sport,
 184 +                      __be32 daddr, __be16 dport,
 185 +                      int dif, struct hlist_head udptable[])
 186 +{
 187 +    struct sock *sk, *result = NULL;
 188 +    struct hlist_node *node;
 189 +    unsigned short hnum = ntohs(dport);
 190 +    int badness = -1;
 191 +
 192 +    read_lock(&udp_hash_lock);
 193 +
 194 +    sk_for_each(sk, node, &udptable[hnum & (UDP_HTABLE_SIZE - 1)]) {
 195 +        struct inet_sock *inet = inet_sk(sk);
 196 +
 197 +        if (sk->sk_hash == hnum && !ipv6_only_sock(sk)) {
 198 +            int score = (sk->sk_family == PF_INET ? 1 : 0);
 199 +
 200 +            if (inet->rcv_saddr) {
 201 +                if (inet->rcv_saddr != daddr)
 202 +                    continue;
 203 +                score+=2;
 204 +            } else {
 205 +                /* block non nx_info ips */
 206 +                if (!v4_addr_in_nx_info(sk->sk_nx_info,
 207 +                    daddr, NXA_MASK_BIND))
 208 +                    continue;
 209 +            }
 210 +            if (inet->daddr) {
 211 +                if (inet->daddr != saddr)
 212 +                    continue;
 213 +                score+=2;
 214 +            }
 215 +            if (inet->dport) {
 216 +                if (inet->dport != sport)
 217 +                    continue;
 218 +                score+=2;
 219 +            }
 220 +            if (sk->sk_bound_dev_if) {
 221 +                if (sk->sk_bound_dev_if != dif)
 222 +                    continue;
 223 +                score+=2;
 224 +            }
 225 +            if (score == 9) {
 226 +                result = sk;
 227 +                break;
 228 +            } else if (score > badness) {
 229 +                result = sk;
 230 +                badness = score;
 231 +            }
 232 +        }
 233 +    }
 234 +
 235 +    if (result)
 236 +        sock_hold(result);
 237 +    read_unlock(&udp_hash_lock);
 238 +    return result;
 239 +}
 240  static unsigned int
 241  mark_tg(struct sk_buff *skb, const struct net_device *in,
 242          const struct net_device *out, unsigned int hooknum,
 243          const struct xt_target *target, const void *targinfo)
 244  {
 245         const struct xt_mark_tginfo2 *info = targinfo;
 246 +    long mark = -1;
 247 +    enum ip_conntrack_info ctinfo;
 248 +    struct sock *connection_sk;
 249 +    int dif;
 250 +    struct nf_conn *ct;
 251 +    extern struct inet_hashinfo tcp_hashinfo;
 252 +    enum ip_conntrack_dir dir;
 253 +    int *curtag;
 254 +    u_int32_t src_ip;
 255 +    u_int32_t dst_ip;
 256 +    u_int16_t proto, src_port;
 257 +    u_int32_t ip;
 258 +    u_int16_t port;
 259 +
 260 +    if (info->mark == ~0U) {
 261 +        /* copy-xid */
 262 +        dif = ((struct rtable *)(skb->dst))->rt_iif;
 263 +
 264 +        ct = nf_ct_get(skb, &ctinfo);
 265 +        if (!ct)
 266 +            goto out_mark_finish;
 267 +
 268 +        dir = CTINFO2DIR(ctinfo);
 269 +        src_ip = ct->tuplehash[dir].tuple.src.u3.ip;
 270 +        dst_ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 271 +        src_port = get_src_port(&ct->tuplehash[dir].tuple);
 272 +        proto = ct->tuplehash[dir].tuple.dst.protonum;
 273 +
 274 +        ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 275 +        port = get_dst_port(&ct->tuplehash[dir].tuple);
 276 +
 277 +        if (proto == 1) {
 278 +            if (skb->mark > 0)
 279 +                /* The packet is marked, it's going out */
 280 +                ct->xid[0] = skb->mark;
 281 +
 282 +            if (ct->xid[0] > 0)
 283 +                mark = ct->xid[0];
 284 +        }
 285 +        else if (proto == 17) {
 286 +            struct sock *sk;
 287 +            if (!skb->mark) {
 288 +                sk = __udp4_lib_lookup(src_ip, src_port,
 289 +                        ip, port, dif, udp_hash);
 290 +
 291 +                if (sk && hooknum == NF_INET_LOCAL_IN)
 292 +                    mark = sk->sk_nid;
 293 +
 294 +                if (sk)
 295 +                    sock_put(sk);
 296 +            }
 297 +            else if (skb->mark > 0)
 298 +                /* The packet is marked, it's going out */
 299 +                ct->xid[0] = skb->mark;
 300 +        }
 301 +        else if (proto == 6) /* TCP */{
 302 +            int sockettype = 0; /* Established socket */
 303 +            struct net *net = &init_net;
 304 +
 305 +            /* Looks for an established socket or a listening
 306 +               socket corresponding to the 4-tuple, in that order.
 307 +               The order is important for Codemux connections
 308 +               to be handled properly */
 309 +
 310 +            connection_sk = inet_lookup_established(net,
 311 +                    &tcp_hashinfo, src_ip, src_port, ip, port, dif);
 312 +
 313 +            if (!connection_sk) {
 314 +                connection_sk = inet_lookup_listener(net,
 315 +                        &tcp_hashinfo, ip, port, dif);
 316 +                sockettype = 1; /* Listening socket */
 317 +            }
 318 +
 319 +            if (connection_sk) {
 320 +                /* The peercred is not set. We set it if the other side has an xid. */
 321 +                if (!PEERCRED_SET(connection_sk->sk_peercred.uid)
 322 +                        && ct->xid[!dir] > 0 && (sockettype == 0)) {
 323 +                    connection_sk->sk_peercred.gid =
 324 +                        connection_sk->sk_peercred.uid = ct->xid[!dir];
 325 +                }
 326 +
 327 +                /* The peercred is set, and is not equal to the XID of 'the other side' */
 328 +                else if (PEERCRED_SET(connection_sk->sk_peercred.uid) &&
 329 +                        (connection_sk->sk_peercred.uid != ct->xid[!dir]) &&
 330 +                        (sockettype == 0)) {
 331 +                    mark = connection_sk->sk_peercred.uid;
 332 +                }
 333 +
 334 +                /* Has this connection already been tagged? */
 335 +                if (ct->xid[dir] < 1) {
 336 +                    /* No - let's tag it */
 337 +                    ct->xid[dir]=connection_sk->sk_nid;
 338 +                }
 339 +
 340 +                if (mark == -1 && (ct->xid[dir] != 0))
 341 +                    mark = ct->xid[dir];
 342 +
 343 +                if (connection_sk->sk_state == TCP_TIME_WAIT) {
 344 +                    inet_twsk_put(inet_twsk(connection_sk));
 345 +                    goto out_mark_finish;
 346 +                } else
 347 +                    sock_put(connection_sk);
 348 +            }
 349 +
 350 +            /* All else failed. Is this a connection over raw sockets?
 351 +               That explains why we couldn't get anything out of skb->sk,
 352 +               or look up a "real" connection. */
 353 +            if (ct->xid[dir] < 1) {
 354 +                if (skb->skb_tag)
 355 +                    ct->xid[dir] = skb->skb_tag;
 356 +            }
 357 +
 358 +            /* Covers CoDemux case */
 359 +            if (mark < 1 && (ct->xid[dir] > 0))
 360 +                mark = ct->xid[dir];
 361 +
 362 +            if (mark < 1 && (ct->xid[!dir] > 0))
 363 +                mark = ct->xid[!dir];
 364 +            goto out_mark_finish;
 365 +        }
 366 +    }
 367 +    else
 368 +        mark = (skb->mark & ~info->mask) ^ info->mark;
 369 +
 370 +out_mark_finish:
 371 +    if (mark != -1)
 372 +        skb->mark = mark;
 373 +
 374 +    curtag = &__get_cpu_var(sknid_elevator);
 375 +    if (mark > 0 && *curtag == -2 && hooknum == NF_INET_LOCAL_IN)
 376 +        *curtag = mark;
 377
 378 -       skb->mark = (skb->mark & ~info->mask) ^ info->mark;
 379         return XT_CONTINUE;
 380  }
 381
 382 diff -Nurb linux-2.6.27-521/net/netfilter/xt_SETXID.c linux-2.6.27-522-ol/net/netfilter/xt_SETXID.c
 383 --- linux-2.6.27-521/net/netfilter/xt_SETXID.c  1969-12-31 19:00:00.000000000 -0500
 384 +++ linux-2.6.27-522-ol/net/netfilter/xt_SETXID.c       2009-06-02 10:02:16.000000000 -0400
 385 @@ -0,0 +1,79 @@
 386 +#include <linux/module.h>
 387 +#include <linux/skbuff.h>
 388 +#include <linux/ip.h>
 389 +#include <net/checksum.h>
 390 +#include <linux/vs_network.h>
 391 +
 392 +#include <linux/netfilter/x_tables.h>
 393 +#include <linux/netfilter/xt_SETXID.h>
 394 +
 395 +MODULE_LICENSE("GPL");
 396 +MODULE_AUTHOR("");
 397 +MODULE_DESCRIPTION("");
 398 +MODULE_ALIAS("ipt_SETXID");
 399 +
 400 +static unsigned int
 401 +target_v1(struct sk_buff **pskb,
 402 +         const struct net_device *in,
 403 +         const struct net_device *out,
 404 +         unsigned int hooknum,
 405 +         const struct xt_target *target,
 406 +         const void *targinfo)
 407 +{
 408 +       const struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 409 +
 410 +       switch (setxidinfo->mode) {
 411 +       case XT_SET_PACKET_XID:
 412 +                (*pskb)->skb_tag = setxidinfo->mark;
 413 +               break;
 414 +       }
 415 +       return XT_CONTINUE;
 416 +}
 417 +
 418 +
 419 +static int
 420 +checkentry_v1(const char *tablename,
 421 +             const void *entry,
 422 +             const struct xt_target *target,
 423 +             void *targinfo,
 424 +             unsigned int hook_mask)
 425 +{
 426 +       struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 427 +
 428 +       if (setxidinfo->mode != XT_SET_PACKET_XID) {
 429 +               printk(KERN_WARNING "SETXID: unknown mode %u\n",
 430 +                      setxidinfo->mode);
 431 +               return 0;
 432 +       }
 433 +
 434 +       return 1;
 435 +}
 436 +
 437 +static struct xt_target xt_setxid_target[] = {
 438 +       {
 439 +               .name           = "SETXID",
 440 +               .family         = AF_INET,
 441 +               .revision       = 1,
 442 +               .checkentry     = checkentry_v1,
 443 +               .target         = target_v1,
 444 +               .targetsize     = sizeof(struct xt_setxid_target_info_v1),
 445 +               .table          = "mangle",
 446 +               .me             = THIS_MODULE,
 447 +       }
 448 +};
 449 +
 450 +static int __init init(void)
 451 +{
 452 +       int err;
 453 +
 454 +       err = xt_register_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 455 +       return err;
 456 +}
 457 +
 458 +static void __exit fini(void)
 459 +{
 460 +       xt_unregister_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 461 +}
 462 +
 463 +module_init(init);
 464 +module_exit(fini);