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

   1 diff -Nurb linux-2.6.22-521/include/linux/netfilter/xt_MARK.h linux-2.6.22-522/include/linux/netfilter/xt_MARK.h
   2 --- linux-2.6.22-521/include/linux/netfilter/xt_MARK.h  2007-07-08 19:32:17.000000000 -0400
   3 +++ linux-2.6.22-522/include/linux/netfilter/xt_MARK.h  2008-07-28 16:36:24.000000000 -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-2.6.22-521/include/linux/netfilter/xt_SETXID.h linux-2.6.22-522/include/linux/netfilter/xt_SETXID.h
  13 --- linux-2.6.22-521/include/linux/netfilter/xt_SETXID.h        1969-12-31 19:00:00.000000000 -0500
  14 +++ linux-2.6.22-522/include/linux/netfilter/xt_SETXID.h        2008-07-28 16:36:24.000000000 -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-2.6.22-521/include/linux/netfilter_ipv4/ipt_MARK.h linux-2.6.22-522/include/linux/netfilter_ipv4/ipt_MARK.h
  31 --- linux-2.6.22-521/include/linux/netfilter_ipv4/ipt_MARK.h    2007-07-08 19:32:17.000000000 -0400
  32 +++ linux-2.6.22-522/include/linux/netfilter_ipv4/ipt_MARK.h    2008-07-28 16:36:24.000000000 -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-2.6.22-521/include/linux/netfilter_ipv4/ipt_SETXID.h linux-2.6.22-522/include/linux/netfilter_ipv4/ipt_SETXID.h
  42 --- linux-2.6.22-521/include/linux/netfilter_ipv4/ipt_SETXID.h  1969-12-31 19:00:00.000000000 -0500
  43 +++ linux-2.6.22-522/include/linux/netfilter_ipv4/ipt_SETXID.h  2008-07-28 16:36:24.000000000 -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-2.6.22-521/include/net/netfilter/nf_conntrack.h linux-2.6.22-522/include/net/netfilter/nf_conntrack.h
  59 --- linux-2.6.22-521/include/net/netfilter/nf_conntrack.h       2007-07-08 19:32:17.000000000 -0400
  60 +++ linux-2.6.22-522/include/net/netfilter/nf_conntrack.h       2008-07-28 16:36:24.000000000 -0400
  61 @@ -131,6 +131,9 @@
  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         /* features dynamically at the end: helper, nat (both optional) */
  69         char data[0];
  70  };
  71 diff -Nurb linux-2.6.22-521/net/netfilter/Kconfig linux-2.6.22-522/net/netfilter/Kconfig
  72 --- linux-2.6.22-521/net/netfilter/Kconfig      2007-07-08 19:32:17.000000000 -0400
  73 +++ linux-2.6.22-522/net/netfilter/Kconfig      2008-07-28 16:36:24.000000000 -0400
  74 @@ -389,6 +389,13 @@
  75
  76           To compile it as a module, choose M here.  If unsure, say N.
  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-2.6.22-521/net/netfilter/Makefile linux-2.6.22-522/net/netfilter/Makefile
  89 --- linux-2.6.22-521/net/netfilter/Makefile     2007-07-08 19:32:17.000000000 -0400
  90 +++ linux-2.6.22-522/net/netfilter/Makefile     2008-07-28 16:36:24.000000000 -0400
  91 @@ -37,6 +37,7 @@
  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_DSCP) += xt_DSCP.o
  99 diff -Nurb linux-2.6.22-521/net/netfilter/nf_conntrack_core.c linux-2.6.22-522/net/netfilter/nf_conntrack_core.c
 100 --- linux-2.6.22-521/net/netfilter/nf_conntrack_core.c  2007-07-08 19:32:17.000000000 -0400
 101 +++ linux-2.6.22-522/net/netfilter/nf_conntrack_core.c  2008-07-28 16:36:24.000000000 -0400
 102 @@ -726,6 +726,8 @@
 103
 104         /* Overload tuple linked list to put us in unconfirmed list. */
 105         list_add(&conntrack->tuplehash[IP_CT_DIR_ORIGINAL].list, &unconfirmed);
 106 +       conntrack->xid[IP_CT_DIR_ORIGINAL] = -1;
 107 +       conntrack->xid[IP_CT_DIR_REPLY] = -1;
 108
 109         write_unlock_bh(&nf_conntrack_lock);
 110
 111 diff -Nurb linux-2.6.22-521/net/netfilter/xt_MARK.c linux-2.6.22-522/net/netfilter/xt_MARK.c
 112 --- linux-2.6.22-521/net/netfilter/xt_MARK.c    2007-07-08 19:32:17.000000000 -0400
 113 +++ linux-2.6.22-522/net/netfilter/xt_MARK.c    2008-08-04 04:59:30.000000000 -0400
 114 @@ -5,13 +5,19 @@
 115   * This program is free software; you can redistribute it and/or modify
 116   * it under the terms of the GNU General Public License version 2 as
 117   * published by the Free Software Foundation.
 118 + *
 119   */
 120
 121  #include <linux/module.h>
 122 +#include <linux/version.h>
 123  #include <linux/skbuff.h>
 124  #include <linux/ip.h>
 125 +#include <net/udp.h>
 126  #include <net/checksum.h>
 127 +#include <net/route.h>
 128 +#include <net/inet_hashtables.h>
 129
 130 +#include <net/netfilter/nf_conntrack.h>
 131  #include <linux/netfilter/x_tables.h>
 132  #include <linux/netfilter/xt_MARK.h>
 133
 134 @@ -21,6 +27,48 @@
 135  MODULE_ALIAS("ipt_MARK");
 136  MODULE_ALIAS("ip6t_MARK");
 137
 138 +static inline u_int16_t
 139 +get_dst_port(struct nf_conntrack_tuple *tuple)
 140 +{
 141 +       switch (tuple->dst.protonum) {
 142 +       case IPPROTO_GRE:
 143 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 144 +#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,11)
 145 +               return tuple->dst.u.gre.key;
 146 +#else
 147 +               return htons(ntohl(tuple->dst.u.gre.key));
 148 +#endif
 149 +       case IPPROTO_ICMP:
 150 +               /* Bind on ICMP echo ID */
 151 +               return tuple->src.u.icmp.id;
 152 +       case IPPROTO_TCP:
 153 +               return tuple->dst.u.tcp.port;
 154 +       case IPPROTO_UDP:
 155 +               return tuple->dst.u.udp.port;
 156 +       default:
 157 +               return tuple->dst.u.all;
 158 +       }
 159 +}
 160 +
 161 +static inline u_int16_t
 162 +get_src_port(struct nf_conntrack_tuple *tuple)
 163 +{
 164 +       switch (tuple->dst.protonum) {
 165 +       case IPPROTO_GRE:
 166 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 167 +               return htons(ntohl(tuple->src.u.gre.key));
 168 +       case IPPROTO_ICMP:
 169 +               /* Bind on ICMP echo ID */
 170 +               return tuple->src.u.icmp.id;
 171 +       case IPPROTO_TCP:
 172 +               return tuple->src.u.tcp.port;
 173 +       case IPPROTO_UDP:
 174 +               return tuple->src.u.udp.port;
 175 +       default:
 176 +               return tuple->src.u.all;
 177 +       }
 178 +}
 179 +
 180  static unsigned int
 181  target_v0(struct sk_buff **pskb,
 182           const struct net_device *in,
 183 @@ -35,6 +83,67 @@
 184         return XT_CONTINUE;
 185  }
 186
 187 +extern DEFINE_PER_CPU(int, sknid_elevator);
 188 +static struct sock *__udp4_lib_lookup(__be32 saddr, __be16 sport,
 189 +                      __be32 daddr, __be16 dport,
 190 +                      int dif, struct hlist_head udptable[])
 191 +{
 192 +    struct sock *sk, *result = NULL;
 193 +    struct hlist_node *node;
 194 +    unsigned short hnum = ntohs(dport);
 195 +    int badness = -1;
 196 +
 197 +    read_lock(&udp_hash_lock);
 198 +
 199 +    sk_for_each(sk, node, &udptable[hnum & (UDP_HTABLE_SIZE - 1)]) {
 200 +        struct inet_sock *inet = inet_sk(sk);
 201 +
 202 +        if (sk->sk_hash == hnum && !ipv6_only_sock(sk)) {
 203 +            int score = (sk->sk_family == PF_INET ? 1 : 0);
 204 +
 205 +            if (inet->rcv_saddr) {
 206 +                if (inet->rcv_saddr != daddr)
 207 +                    continue;
 208 +                score+=2;
 209 +            } else {
 210 +                /* block non nx_info ips */
 211 +                if (!v4_addr_in_nx_info(sk->sk_nx_info,
 212 +                    daddr, NXA_MASK_BIND))
 213 +                    continue;
 214 +            }
 215 +            if (inet->daddr) {
 216 +                if (inet->daddr != saddr)
 217 +                    continue;
 218 +                score+=2;
 219 +            }
 220 +            if (inet->dport) {
 221 +                if (inet->dport != sport)
 222 +                    continue;
 223 +                score+=2;
 224 +            }
 225 +            if (sk->sk_bound_dev_if) {
 226 +                if (sk->sk_bound_dev_if != dif)
 227 +                    continue;
 228 +                score+=2;
 229 +            }
 230 +            if (score == 9) {
 231 +                result = sk;
 232 +                break;
 233 +            } else if (score > badness) {
 234 +                result = sk;
 235 +                badness = score;
 236 +            }
 237 +        }
 238 +    }
 239 +
 240 +    if (result)
 241 +        sock_hold(result);
 242 +    read_unlock(&udp_hash_lock);
 243 +    return result;
 244 +}
 245 +
 246 +#define related(ct) (ct==(IP_CT_IS_REPLY + IP_CT_RELATED))
 247 +
 248  static unsigned int
 249  target_v1(struct sk_buff **pskb,
 250           const struct net_device *in,
 251 @@ -44,7 +153,20 @@
 252           const void *targinfo)
 253  {
 254         const struct xt_mark_target_info_v1 *markinfo = targinfo;
 255 -       int mark = 0;
 256 +       enum ip_conntrack_info ctinfo;
 257 +      struct sock *connection_sk;
 258 +      int dif;
 259 +      struct nf_conn *ct;
 260 +      extern struct inet_hashinfo tcp_hashinfo;
 261 +      enum ip_conntrack_dir dir;
 262 +      int *curtag;
 263 +      u_int32_t src_ip;
 264 +      u_int32_t dst_ip;
 265 +      u_int16_t proto, src_port;
 266 +      u_int32_t ip;
 267 +      u_int16_t port;
 268 +
 269 +       int mark = -1;
 270
 271         switch (markinfo->mode) {
 272         case XT_MARK_SET:
 273 @@ -58,13 +180,117 @@
 274         case XT_MARK_OR:
 275                 mark = (*pskb)->mark | markinfo->mark;
 276                 break;
 277 +
 278 +       case XT_MARK_COPYXID:
 279 +             dif = ((struct rtable *)(*pskb)->dst)->rt_iif;
 280 +
 281 +             ct = nf_ct_get((*pskb), &ctinfo);
 282 +             if (!ct)
 283 +                     break;
 284 +
 285 +             dir = CTINFO2DIR(ctinfo);
 286 +             src_ip = ct->tuplehash[dir].tuple.src.u3.ip;
 287 +             dst_ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 288 +             src_port = get_src_port(&ct->tuplehash[dir].tuple);
 289 +             proto = ct->tuplehash[dir].tuple.dst.protonum;
 290 +
 291 +             ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 292 +             port = get_dst_port(&ct->tuplehash[dir].tuple);
 293 +
 294 +             if (proto == 1) {
 295 +                                 if ((*pskb)->mark>0) /* The packet is marked, it's going out */
 296 +                                 {
 297 +                                                 ct->xid[0]=(*pskb)->mark;
 298 +                                 }
 299 +
 300 +                                 if (ct->xid[0] > 0) {
 301 +                                                 mark = ct->xid[0];
 302 +                                 }
 303 +             }
 304 +                 else if (proto == 17) {
 305 +                  struct sock *sk;
 306 +                  if (!(*pskb)->mark) {
 307 +                          sk = __udp4_lib_lookup(src_ip, src_port, ip, port,
 308 +                                          dif, udp_hash);
 309 +
 310 +                          if (sk && hooknum==NF_IP_LOCAL_IN) {
 311 +                                  mark=sk->sk_nid;
 312 +                          }
 313 +                  }
 314 +                  else
 315 +                  if ((*pskb)->mark>0) /* The packet is marked, it's going out */
 316 +                  {
 317 +                          ct->xid[0]=(*pskb)->mark;
 318 +                  }
 319 +          }
 320 +             else if (proto == 6) {
 321 +                     if ((*pskb)->sk) {
 322 +                             /* It's a listening socket */
 323 +                             connection_sk = (*pskb)->sk;
 324 +                             sock_hold(connection_sk);
 325 +                     }
 326 +                     else   /* It might be a connected socket */
 327 +                             connection_sk = inet_lookup_established(&tcp_hashinfo, src_ip, src_port, ip, port, dif);
 328 +
 329 +
 330 +                     if (connection_sk /* Well, some kind of TCP socket */) {
 331 +                             if (connection_sk->sk_peercred.uid == 0 || connection_sk->sk_peercred.uid == (__u32) -1) {
 332 +                                     /* Normal case - the peercred on the socket is not set */
 333 +                                     connection_sk->sk_peercred.gid = connection_sk->sk_peercred.uid = ct->xid[!dir];
 334 +                             }
 335 +                             else    /* Exceptional case - the peercred was set using SET_PEERCRED. Somebody wants us
 336 +                                        to mark packets with some arbitrary value.*/
 337 +                                     mark=connection_sk->sk_peercred.uid;
 338 +
 339 +                             /* Has this connection already been tagged? */
 340 +                             if (ct->xid[dir] < 1) {
 341 +                                     /* No - let's tag it */
 342 +                                     ct->xid[dir]=connection_sk->sk_nid;
 343 +                             }
 344 +
 345 +                             if (mark==-1 && (connection_sk->sk_nid != 0))
 346 +                                     mark = ct->xid[dir];
 347 +
 348 +
 349 +                             if (connection_sk->sk_state == TCP_TIME_WAIT) {
 350 +                                     inet_twsk_put(inet_twsk(connection_sk));
 351 +                                     break;
 352 +                             }
 353 +                             else
 354 +                                     sock_put(connection_sk);
 355 +                     }
 356 +
 357 +                     /* Covers CoDemux case */
 358 +                     if (mark < 1 && (ct->xid[dir]>0)) {
 359 +                             mark = ct->xid[dir];
 360         }
 361
 362 +                     if (mark < 1 && (ct->xid[!dir]>0)) {
 363 +                               mark = ct->xid[!dir];
 364 +                     }
 365 +
 366 +                     /* All else failed. Is this a connection over raw sockets? That explains
 367 +                      * why we couldn't get anything out of skb->sk, or look up a "real" connection.*/
 368 +                     if (ct->xid[dir]<1) {
 369 +                       if ((*pskb)->skb_tag) {
 370 +                               ct->xid[dir]=(*pskb)->skb_tag;
 371 +                       }
 372 +                     }
 373 +             }
 374 +             break;
 375 +       }
 376 +       if (mark != -1) {
 377         (*pskb)->mark = mark;
 378 +       }
 379 +
 380 +       curtag=&__get_cpu_var(sknid_elevator);
 381 +       if (mark > 0 && *curtag==-2 && hooknum==NF_IP_LOCAL_IN)
 382 +       {
 383 +               *curtag = mark;
 384 +       }
 385         return XT_CONTINUE;
 386  }
 387
 388 -
 389  static int
 390  checkentry_v0(const char *tablename,
 391               const void *entry,
 392 @@ -92,7 +318,8 @@
 393
 394         if (markinfo->mode != XT_MARK_SET
 395             && markinfo->mode != XT_MARK_AND
 396 -           && markinfo->mode != XT_MARK_OR) {
 397 +           && markinfo->mode != XT_MARK_OR
 398 +           && markinfo->mode != XT_MARK_COPYXID) {
 399                 printk(KERN_WARNING "MARK: unknown mode %u\n",
 400                        markinfo->mode);
 401                 return 0;
 402 diff -Nurb linux-2.6.22-521/net/netfilter/xt_SETXID.c linux-2.6.22-522/net/netfilter/xt_SETXID.c
 403 --- linux-2.6.22-521/net/netfilter/xt_SETXID.c  1969-12-31 19:00:00.000000000 -0500
 404 +++ linux-2.6.22-522/net/netfilter/xt_SETXID.c  2008-07-28 16:36:24.000000000 -0400
 405 @@ -0,0 +1,79 @@
 406 +#include <linux/module.h>
 407 +#include <linux/skbuff.h>
 408 +#include <linux/ip.h>
 409 +#include <net/checksum.h>
 410 +#include <linux/vs_network.h>
 411 +
 412 +#include <linux/netfilter/x_tables.h>
 413 +#include <linux/netfilter/xt_SETXID.h>
 414 +
 415 +MODULE_LICENSE("GPL");
 416 +MODULE_AUTHOR("");
 417 +MODULE_DESCRIPTION("");
 418 +MODULE_ALIAS("ipt_SETXID");
 419 +
 420 +static unsigned int
 421 +target_v1(struct sk_buff **pskb,
 422 +         const struct net_device *in,
 423 +         const struct net_device *out,
 424 +         unsigned int hooknum,
 425 +         const struct xt_target *target,
 426 +         const void *targinfo)
 427 +{
 428 +       const struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 429 +
 430 +       switch (setxidinfo->mode) {
 431 +       case XT_SET_PACKET_XID:
 432 +                (*pskb)->skb_tag = setxidinfo->mark;
 433 +               break;
 434 +       }
 435 +       return XT_CONTINUE;
 436 +}
 437 +
 438 +
 439 +static int
 440 +checkentry_v1(const char *tablename,
 441 +             const void *entry,
 442 +             const struct xt_target *target,
 443 +             void *targinfo,
 444 +             unsigned int hook_mask)
 445 +{
 446 +       struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 447 +
 448 +       if (setxidinfo->mode != XT_SET_PACKET_XID) {
 449 +               printk(KERN_WARNING "SETXID: unknown mode %u\n",
 450 +                      setxidinfo->mode);
 451 +               return 0;
 452 +       }
 453 +
 454 +       return 1;
 455 +}
 456 +
 457 +static struct xt_target xt_setxid_target[] = {
 458 +       {
 459 +               .name           = "SETXID",
 460 +               .family         = AF_INET,
 461 +               .revision       = 1,
 462 +               .checkentry     = checkentry_v1,
 463 +               .target         = target_v1,
 464 +               .targetsize     = sizeof(struct xt_setxid_target_info_v1),
 465 +               .table          = "mangle",
 466 +               .me             = THIS_MODULE,
 467 +       }
 468 +};
 469 +
 470 +static int __init init(void)
 471 +{
 472 +       int err;
 473 +
 474 +       err = xt_register_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 475 +       return err;
 476 +}
 477 +
 478 +static void __exit fini(void)
 479 +{
 480 +       xt_unregister_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 481 +}
 482 +
 483 +module_init(init);
 484 +module_exit(fini);