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  2008-09-17 17:59:53.000000000 -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        2008-09-17 17:59:53.000000000 -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    2008-09-17 17:59:53.000000000 -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  2008-09-17 17:59:53.000000000 -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       2008-09-17 17:59:53.000000000 -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      2008-09-17 17:59:53.000000000 -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     2008-09-17 17:59:53.000000000 -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  2008-09-17 17:59:53.000000000 -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    2008-09-17 18:29:52.000000000 -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 +27,50 @@
 135  MODULE_ALIAS("ipt_MARK");
 136  MODULE_ALIAS("ip6t_MARK");
 137
 138 +#define PEERCRED_SET(x) ((x!=0) && (x!=(unsigned int)-1))
 139 +
 140 +static inline u_int16_t
 141 +get_dst_port(struct nf_conntrack_tuple *tuple)
 142 +{
 143 +       switch (tuple->dst.protonum) {
 144 +       case IPPROTO_GRE:
 145 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 146 +#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,11)
 147 +               return tuple->dst.u.gre.key;
 148 +#else
 149 +               return htons(ntohl(tuple->dst.u.gre.key));
 150 +#endif
 151 +       case IPPROTO_ICMP:
 152 +               /* Bind on ICMP echo ID */
 153 +               return tuple->src.u.icmp.id;
 154 +       case IPPROTO_TCP:
 155 +               return tuple->dst.u.tcp.port;
 156 +       case IPPROTO_UDP:
 157 +               return tuple->dst.u.udp.port;
 158 +       default:
 159 +               return tuple->dst.u.all;
 160 +       }
 161 +}
 162 +
 163 +static inline u_int16_t
 164 +get_src_port(struct nf_conntrack_tuple *tuple)
 165 +{
 166 +       switch (tuple->dst.protonum) {
 167 +       case IPPROTO_GRE:
 168 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 169 +               return htons(ntohl(tuple->src.u.gre.key));
 170 +       case IPPROTO_ICMP:
 171 +               /* Bind on ICMP echo ID */
 172 +               return tuple->src.u.icmp.id;
 173 +       case IPPROTO_TCP:
 174 +               return tuple->src.u.tcp.port;
 175 +       case IPPROTO_UDP:
 176 +               return tuple->src.u.udp.port;
 177 +       default:
 178 +               return tuple->src.u.all;
 179 +       }
 180 +}
 181 +
 182  static unsigned int
 183  target_v0(struct sk_buff **pskb,
 184           const struct net_device *in,
 185 @@ -35,6 +85,68 @@
 186         return XT_CONTINUE;
 187  }
 188
 189 +extern DEFINE_PER_CPU(int, sknid_elevator);
 190 +
 191 +static struct sock *__udp4_lib_lookup(__be32 saddr, __be16 sport,
 192 +                      __be32 daddr, __be16 dport,
 193 +                      int dif, struct hlist_head udptable[])
 194 +{
 195 +    struct sock *sk, *result = NULL;
 196 +    struct hlist_node *node;
 197 +    unsigned short hnum = ntohs(dport);
 198 +    int badness = -1;
 199 +
 200 +    read_lock(&udp_hash_lock);
 201 +
 202 +    sk_for_each(sk, node, &udptable[hnum & (UDP_HTABLE_SIZE - 1)]) {
 203 +        struct inet_sock *inet = inet_sk(sk);
 204 +
 205 +        if (sk->sk_hash == hnum && !ipv6_only_sock(sk)) {
 206 +            int score = (sk->sk_family == PF_INET ? 1 : 0);
 207 +
 208 +            if (inet->rcv_saddr) {
 209 +                if (inet->rcv_saddr != daddr)
 210 +                    continue;
 211 +                score+=2;
 212 +            } else {
 213 +                /* block non nx_info ips */
 214 +                if (!v4_addr_in_nx_info(sk->sk_nx_info,
 215 +                    daddr, NXA_MASK_BIND))
 216 +                    continue;
 217 +            }
 218 +            if (inet->daddr) {
 219 +                if (inet->daddr != saddr)
 220 +                    continue;
 221 +                score+=2;
 222 +            }
 223 +            if (inet->dport) {
 224 +                if (inet->dport != sport)
 225 +                    continue;
 226 +                score+=2;
 227 +            }
 228 +            if (sk->sk_bound_dev_if) {
 229 +                if (sk->sk_bound_dev_if != dif)
 230 +                    continue;
 231 +                score+=2;
 232 +            }
 233 +            if (score == 9) {
 234 +                result = sk;
 235 +                break;
 236 +            } else if (score > badness) {
 237 +                result = sk;
 238 +                badness = score;
 239 +            }
 240 +        }
 241 +    }
 242 +
 243 +    if (result)
 244 +        sock_hold(result);
 245 +    read_unlock(&udp_hash_lock);
 246 +    return result;
 247 +}
 248 +
 249 +#define related(ct) (ct==(IP_CT_IS_REPLY + IP_CT_RELATED))
 250 +
 251  static unsigned int
 252  target_v1(struct sk_buff **pskb,
 253           const struct net_device *in,
 254 @@ -44,7 +156,20 @@
 255           const void *targinfo)
 256  {
 257         const struct xt_mark_target_info_v1 *markinfo = targinfo;
 258 -       int mark = 0;
 259 +               enum ip_conntrack_info ctinfo;
 260 +               struct sock *connection_sk;
 261 +               int dif;
 262 +               struct nf_conn *ct;
 263 +               extern struct inet_hashinfo tcp_hashinfo;
 264 +               enum ip_conntrack_dir dir;
 265 +               int *curtag;
 266 +               u_int32_t src_ip;
 267 +               u_int32_t dst_ip;
 268 +               u_int16_t proto, src_port;
 269 +               u_int32_t ip;
 270 +               u_int16_t port;
 271 +
 272 +               int mark = -1;
 273
 274         switch (markinfo->mode) {
 275         case XT_MARK_SET:
 276 @@ -58,13 +183,126 @@
 277         case XT_MARK_OR:
 278                 mark = (*pskb)->mark | markinfo->mark;
 279                 break;
 280 +
 281 +                               case XT_MARK_COPYXID:
 282 +                                               dif = ((struct rtable *)(*pskb)->dst)->rt_iif;
 283 +
 284 +                                               ct = nf_ct_get((*pskb), &ctinfo);
 285 +                                               if (!ct)
 286 +                                                               break;
 287 +
 288 +                                               dir = CTINFO2DIR(ctinfo);
 289 +                                               src_ip = ct->tuplehash[dir].tuple.src.u3.ip;
 290 +                                               dst_ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 291 +                                               src_port = get_src_port(&ct->tuplehash[dir].tuple);
 292 +                                               proto = ct->tuplehash[dir].tuple.dst.protonum;
 293 +
 294 +                                               ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 295 +                                               port = get_dst_port(&ct->tuplehash[dir].tuple);
 296 +
 297 +                                               if (proto == 1) {
 298 +                                                               if ((*pskb)->mark>0) /* The packet is marked, it's going out */
 299 +                                                               {
 300 +                                                                               ct->xid[0]=(*pskb)->mark;
 301 +                                                               }
 302 +
 303 +                                                               if (ct->xid[0] > 0) {
 304 +                                                                               mark = ct->xid[0];
 305 +                                                               }
 306 +                                               }
 307 +                                               else if (proto == 17) {
 308 +                                                               struct sock *sk;
 309 +                                                               if (!(*pskb)->mark) {
 310 +                                                                               sk = __udp4_lib_lookup(src_ip, src_port, ip, port,
 311 +                                                                                                               dif, udp_hash);
 312 +
 313 +                                                                               if (sk && hooknum==NF_IP_LOCAL_IN) {
 314 +                                                                                               mark=sk->sk_nid;
 315 +                                                                               }
 316 +
 317 +                                                                               if (sk) {
 318 +                                                                                               sock_put(sk);
 319 +                                                                               }
 320 +                                                               }
 321 +                                                               else
 322 +                                                                               if ((*pskb)->mark>0) /* The packet is marked, it's going out */
 323 +                                                                               {
 324 +                                                                                               ct->xid[0]=(*pskb)->mark;
 325 +                                                                               }
 326 +                                               }
 327 +                                               else if (proto == 6) /* TCP */{
 328 +                                                               int sockettype=0; /* Established socket */
 329 +                                                               /* Looks for an established socket or a listening socket corresponding to the 4-tuple, in
 330 +                                                                * that order. The order is important for Codemux connections to be handled properly */
 331 +
 332 +                                                               connection_sk = inet_lookup_established(&tcp_hashinfo, src_ip, src_port, ip, port, dif);
 333 +
 334 +                                                               if (!connection_sk) {
 335 +                                                                               connection_sk = inet_lookup_listener(&tcp_hashinfo, ip, port, dif);
 336 +                                                                               sockettype=1; /* Listening socket */
 337 +                                                               }
 338 +
 339 +                                                               if (connection_sk) {
 340 +                                                                               /* The peercred is not set. We set it if the other side has an xid. */
 341 +                                                                               if (!PEERCRED_SET(connection_sk->sk_peercred.uid)
 342 +                                                                                                               && ct->xid[!dir]>0 && (sockettype==0)) {
 343 +                                                                                               connection_sk->sk_peercred.gid = connection_sk->sk_peercred.uid = ct->xid[!dir];
 344 +                                                                               }
 345 +
 346 +                                                                               /* The peercred is set, and is not equal to the XID of 'the other side' */
 347 +                                                                               else if (PEERCRED_SET(connection_sk->sk_peercred.uid) && (connection_sk->sk_peercred.uid != ct->xid[!dir]) && (sockettype==0)) {
 348 +                                                                                               mark = connection_sk->sk_peercred.uid;
 349 +                                                                               }
 350 +
 351 +                                                                               /* Has this connection already been tagged? */
 352 +                                                                               if (ct->xid[dir] < 1) {
 353 +                                                                                               /* No - let's tag it */
 354 +                                                                                               ct->xid[dir]=connection_sk->sk_nid;
 355 +
 356 +                                                                               }
 357 +
 358 +                                                                               if (mark==-1 && (ct->xid[dir]!= 0))
 359 +                                                                                               mark = ct->xid[dir];
 360 +
 361 +                                                                               if (connection_sk->sk_state == TCP_TIME_WAIT) {
 362 +                                                                                               inet_twsk_put(inet_twsk(connection_sk));
 363 +                                                                                               break;
 364 +                                                                               }
 365 +                                                                               else
 366 +                                                                                               sock_put(connection_sk);
 367         }
 368
 369 +                                                               /* All else failed. Is this a connection over raw sockets? That explains
 370 +                                                                * why we couldn't get anything out of skb->sk, or look up a "real" connection.*/
 371 +                                                               if (ct->xid[dir]<1) {
 372 +                                                                               if ((*pskb)->skb_tag) {
 373 +                                                                                               ct->xid[dir]=(*pskb)->skb_tag;
 374 +                                                                               }
 375 +                                                               }
 376 +
 377 +                                                               /* Covers CoDemux case */
 378 +                                                               if (mark < 1 && (ct->xid[dir]>0)) {
 379 +                                                                               mark = ct->xid[dir];
 380 +                                                               }
 381 +
 382 +                                                               if (mark < 1 && (ct->xid[!dir]>0)) {
 383 +                                                                               mark = ct->xid[!dir];
 384 +                                                               }
 385 +                                                               break;
 386 +                                               }
 387 +               }
 388 +               if (mark != -1) {
 389         (*pskb)->mark = mark;
 390 +               }
 391 +
 392 +               curtag=&__get_cpu_var(sknid_elevator);
 393 +               if (mark > 0 && *curtag==-2 && hooknum==NF_IP_LOCAL_IN)
 394 +               {
 395 +                               *curtag = mark;
 396 +               }
 397         return XT_CONTINUE;
 398  }
 399
 400 -
 401  static int
 402  checkentry_v0(const char *tablename,
 403               const void *entry,
 404 @@ -92,7 +330,8 @@
 405
 406         if (markinfo->mode != XT_MARK_SET
 407             && markinfo->mode != XT_MARK_AND
 408 -           && markinfo->mode != XT_MARK_OR) {
 409 +           && markinfo->mode != XT_MARK_OR
 410 +           && markinfo->mode != XT_MARK_COPYXID) {
 411                 printk(KERN_WARNING "MARK: unknown mode %u\n",
 412                        markinfo->mode);
 413                 return 0;
 414 diff -Nurb linux-2.6.22-521/net/netfilter/xt_SETXID.c linux-2.6.22-522/net/netfilter/xt_SETXID.c
 415 --- linux-2.6.22-521/net/netfilter/xt_SETXID.c  1969-12-31 19:00:00.000000000 -0500
 416 +++ linux-2.6.22-522/net/netfilter/xt_SETXID.c  2008-09-17 17:59:53.000000000 -0400
 417 @@ -0,0 +1,79 @@
 418 +#include <linux/module.h>
 419 +#include <linux/skbuff.h>
 420 +#include <linux/ip.h>
 421 +#include <net/checksum.h>
 422 +#include <linux/vs_network.h>
 423 +
 424 +#include <linux/netfilter/x_tables.h>
 425 +#include <linux/netfilter/xt_SETXID.h>
 426 +
 427 +MODULE_LICENSE("GPL");
 428 +MODULE_AUTHOR("");
 429 +MODULE_DESCRIPTION("");
 430 +MODULE_ALIAS("ipt_SETXID");
 431 +
 432 +static unsigned int
 433 +target_v1(struct sk_buff **pskb,
 434 +         const struct net_device *in,
 435 +         const struct net_device *out,
 436 +         unsigned int hooknum,
 437 +         const struct xt_target *target,
 438 +         const void *targinfo)
 439 +{
 440 +       const struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 441 +
 442 +       switch (setxidinfo->mode) {
 443 +       case XT_SET_PACKET_XID:
 444 +                (*pskb)->skb_tag = setxidinfo->mark;
 445 +               break;
 446 +       }
 447 +       return XT_CONTINUE;
 448 +}
 449 +
 450 +
 451 +static int
 452 +checkentry_v1(const char *tablename,
 453 +             const void *entry,
 454 +             const struct xt_target *target,
 455 +             void *targinfo,
 456 +             unsigned int hook_mask)
 457 +{
 458 +       struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 459 +
 460 +       if (setxidinfo->mode != XT_SET_PACKET_XID) {
 461 +               printk(KERN_WARNING "SETXID: unknown mode %u\n",
 462 +                      setxidinfo->mode);
 463 +               return 0;
 464 +       }
 465 +
 466 +       return 1;
 467 +}
 468 +
 469 +static struct xt_target xt_setxid_target[] = {
 470 +       {
 471 +               .name           = "SETXID",
 472 +               .family         = AF_INET,
 473 +               .revision       = 1,
 474 +               .checkentry     = checkentry_v1,
 475 +               .target         = target_v1,
 476 +               .targetsize     = sizeof(struct xt_setxid_target_info_v1),
 477 +               .table          = "mangle",
 478 +               .me             = THIS_MODULE,
 479 +       }
 480 +};
 481 +
 482 +static int __init init(void)
 483 +{
 484 +       int err;
 485 +
 486 +       err = xt_register_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 487 +       return err;
 488 +}
 489 +
 490 +static void __exit fini(void)
 491 +{
 492 +       xt_unregister_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 493 +}
 494 +
 495 +module_init(init);
 496 +module_exit(fini);