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    2008-09-14 16:50:22.000000000 -0400
 114 @@ -5,13 +5,18 @@
 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/checksum.h>
 126 +#include <net/route.h>
 127 +#include <net/inet_hashtables.h>
 128
 129 +#include <net/netfilter/nf_conntrack.h>
 130  #include <linux/netfilter/x_tables.h>
 131  #include <linux/netfilter/xt_MARK.h>
 132
 133 @@ -21,6 +26,50 @@
 134  MODULE_ALIAS("ipt_MARK");
 135  MODULE_ALIAS("ip6t_MARK");
 136
 137 +#define PEERCRED_SET(x) ((x!=0) && (x!=(unsigned int)-1))
 138 +
 139 +static inline u_int16_t
 140 +get_dst_port(struct nf_conntrack_tuple *tuple)
 141 +{
 142 +       switch (tuple->dst.protonum) {
 143 +       case IPPROTO_GRE:
 144 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 145 +#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,11)
 146 +               return tuple->dst.u.gre.key;
 147 +#else
 148 +               return htons(ntohl(tuple->dst.u.gre.key));
 149 +#endif
 150 +       case IPPROTO_ICMP:
 151 +               /* Bind on ICMP echo ID */
 152 +               return tuple->src.u.icmp.id;
 153 +       case IPPROTO_TCP:
 154 +               return tuple->dst.u.tcp.port;
 155 +       case IPPROTO_UDP:
 156 +               return tuple->dst.u.udp.port;
 157 +       default:
 158 +               return tuple->dst.u.all;
 159 +       }
 160 +}
 161 +
 162 +static inline u_int16_t
 163 +get_src_port(struct nf_conntrack_tuple *tuple)
 164 +{
 165 +       switch (tuple->dst.protonum) {
 166 +       case IPPROTO_GRE:
 167 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 168 +               return htons(ntohl(tuple->src.u.gre.key));
 169 +       case IPPROTO_ICMP:
 170 +               /* Bind on ICMP echo ID */
 171 +               return tuple->src.u.icmp.id;
 172 +       case IPPROTO_TCP:
 173 +               return tuple->src.u.tcp.port;
 174 +       case IPPROTO_UDP:
 175 +               return tuple->src.u.udp.port;
 176 +       default:
 177 +               return tuple->src.u.all;
 178 +       }
 179 +}
 180 +
 181  static unsigned int
 182  target_v0(struct sk_buff **pskb,
 183           const struct net_device *in,
 184 @@ -35,6 +84,10 @@
 185         return XT_CONTINUE;
 186  }
 187
 188 +extern DEFINE_PER_CPU(int, sknid_elevator);
 189 +
 190 +#define related(ct) (ct==(IP_CT_IS_REPLY + IP_CT_RELATED))
 191 +
 192  static unsigned int
 193  target_v1(struct sk_buff **pskb,
 194           const struct net_device *in,
 195 @@ -44,7 +97,20 @@
 196           const void *targinfo)
 197  {
 198         const struct xt_mark_target_info_v1 *markinfo = targinfo;
 199 -       int mark = 0;
 200 +       enum ip_conntrack_info ctinfo;
 201 +      struct sock *connection_sk;
 202 +      int dif;
 203 +      struct nf_conn *ct;
 204 +      extern struct inet_hashinfo tcp_hashinfo;
 205 +      enum ip_conntrack_dir dir;
 206 +      int *curtag;
 207 +      u_int32_t src_ip;
 208 +      u_int32_t dst_ip;
 209 +      u_int16_t proto, src_port;
 210 +      u_int32_t ip;
 211 +      u_int16_t port;
 212 +
 213 +       int mark = -1;
 214
 215         switch (markinfo->mode) {
 216         case XT_MARK_SET:
 217 @@ -58,13 +124,107 @@
 218         case XT_MARK_OR:
 219                 mark = (*pskb)->mark | markinfo->mark;
 220                 break;
 221 +
 222 +                       case XT_MARK_COPYXID:
 223 +                                       dif = ((struct rtable *)(*pskb)->dst)->rt_iif;
 224 +
 225 +                                       ct = nf_ct_get((*pskb), &ctinfo);
 226 +                                       if (!ct)
 227 +                                                       break;
 228 +
 229 +                                       dir = CTINFO2DIR(ctinfo);
 230 +                                       src_ip = ct->tuplehash[dir].tuple.src.u3.ip;
 231 +                                       dst_ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 232 +                                       src_port = get_src_port(&ct->tuplehash[dir].tuple);
 233 +                                       proto = ct->tuplehash[dir].tuple.dst.protonum;
 234 +
 235 +                                       ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 236 +                                       port = get_dst_port(&ct->tuplehash[dir].tuple);
 237 +
 238 +                                       if (proto == 1 || proto == 17) {
 239 +                                                       if ((*pskb)->mark>0) /* The packet is marked, it's going out */
 240 +                                                       {
 241 +                                                                       ct->xid[0]=(*pskb)->mark;
 242         }
 243
 244 +                                                       if (ct->xid[0] > 0) {
 245 +                                                                       mark = ct->xid[0];
 246 +                                                       }
 247 +
 248 +                                       }
 249 +                                       else if (proto == 6) /* TCP */{
 250 +                                                       int sockettype=0; /* Established socket */
 251 +                                                       /* Looks for an established socket or a listening socket corresponding to the 4-tuple, in
 252 +                                                        * that order. The order is important for Codemux connections to be handled properly */
 253 +
 254 +                                                       connection_sk = inet_lookup_established(&tcp_hashinfo, src_ip, src_port, ip, port, dif);
 255 +
 256 +                                                       if (!connection_sk) {
 257 +                                                               connection_sk = inet_lookup_listener(&tcp_hashinfo, ip, port, dif);
 258 +                                                               sockettype=1; /* Listening socket */
 259 +                                                       }
 260 +
 261 +                                                       if (connection_sk) {
 262 +                                                                       /* The peercred is not set. We set it if the other side has an xid. */
 263 +                                                                       if (!PEERCRED_SET(connection_sk->sk_peercred.uid)
 264 +                                                                                                       && ct->xid[!dir]>0 && (sockettype==0)) {
 265 +                                                                                       connection_sk->sk_peercred.gid = connection_sk->sk_peercred.uid = ct->xid[!dir];
 266 +                                                                       }
 267 +
 268 +                                                                       /* The peercred is set, and is not equal to the XID of 'the other side' */
 269 +                                                                       else if (PEERCRED_SET(connection_sk->sk_peercred.uid) && (connection_sk->sk_peercred.uid != ct->xid[!dir]) && (sockettype==0)) {
 270 +                                                                                       mark = connection_sk->sk_peercred.uid;
 271 +                                                                       }
 272 +
 273 +                                                                       /* Has this connection already been tagged? */
 274 +                                                                       if (ct->xid[dir] < 1) {
 275 +                                                                                       /* No - let's tag it */
 276 +                                                                                       ct->xid[dir]=connection_sk->sk_nid;
 277 +
 278 +                                                                       }
 279 +
 280 +                                                                       if (mark==-1 && (ct->xid[dir]!= 0))
 281 +                                                                                       mark = ct->xid[dir];
 282 +
 283 +                                                                       if (connection_sk->sk_state == TCP_TIME_WAIT) {
 284 +                                                                                       inet_twsk_put(inet_twsk(connection_sk));
 285 +                                                                                       break;
 286 +                                                                       }
 287 +                                                                       else
 288 +                                                                                       sock_put(connection_sk);
 289 +                                                       }
 290 +
 291 +                                                       /* All else failed. Is this a connection over raw sockets? That explains
 292 +                                                        * why we couldn't get anything out of skb->sk, or look up a "real" connection.*/
 293 +                                                       if (ct->xid[dir]<1) {
 294 +                                                                       if ((*pskb)->skb_tag) {
 295 +                                                                                       ct->xid[dir]=(*pskb)->skb_tag;
 296 +                                                                       }
 297 +                                                       }
 298 +
 299 +                                                       /* Covers CoDemux case */
 300 +                                                       if (mark < 1 && (ct->xid[dir]>0)) {
 301 +                                                                       mark = ct->xid[dir];
 302 +                                                       }
 303 +
 304 +                                                       if (mark < 1 && (ct->xid[!dir]>0)) {
 305 +                                                                       mark = ct->xid[!dir];
 306 +                                                       }
 307 +                                                       break;
 308 +                                       }
 309 +       }
 310 +       if (mark != -1) {
 311         (*pskb)->mark = mark;
 312 +       }
 313 +
 314 +       curtag=&__get_cpu_var(sknid_elevator);
 315 +       if (mark > 0 && *curtag==-2 && hooknum==NF_IP_LOCAL_IN)
 316 +       {
 317 +               *curtag = mark;
 318 +       }
 319         return XT_CONTINUE;
 320  }
 321
 322 -
 323  static int
 324  checkentry_v0(const char *tablename,
 325               const void *entry,
 326 @@ -92,7 +252,8 @@
 327
 328         if (markinfo->mode != XT_MARK_SET
 329             && markinfo->mode != XT_MARK_AND
 330 -           && markinfo->mode != XT_MARK_OR) {
 331 +           && markinfo->mode != XT_MARK_OR
 332 +           && markinfo->mode != XT_MARK_COPYXID) {
 333                 printk(KERN_WARNING "MARK: unknown mode %u\n",
 334                        markinfo->mode);
 335                 return 0;
 336 diff -Nurb linux-2.6.22-521/net/netfilter/xt_SETXID.c linux-2.6.22-522/net/netfilter/xt_SETXID.c
 337 --- linux-2.6.22-521/net/netfilter/xt_SETXID.c  1969-12-31 19:00:00.000000000 -0500
 338 +++ linux-2.6.22-522/net/netfilter/xt_SETXID.c  2008-07-28 16:36:24.000000000 -0400
 339 @@ -0,0 +1,79 @@
 340 +#include <linux/module.h>
 341 +#include <linux/skbuff.h>
 342 +#include <linux/ip.h>
 343 +#include <net/checksum.h>
 344 +#include <linux/vs_network.h>
 345 +
 346 +#include <linux/netfilter/x_tables.h>
 347 +#include <linux/netfilter/xt_SETXID.h>
 348 +
 349 +MODULE_LICENSE("GPL");
 350 +MODULE_AUTHOR("");
 351 +MODULE_DESCRIPTION("");
 352 +MODULE_ALIAS("ipt_SETXID");
 353 +
 354 +static unsigned int
 355 +target_v1(struct sk_buff **pskb,
 356 +         const struct net_device *in,
 357 +         const struct net_device *out,
 358 +         unsigned int hooknum,
 359 +         const struct xt_target *target,
 360 +         const void *targinfo)
 361 +{
 362 +       const struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 363 +
 364 +       switch (setxidinfo->mode) {
 365 +       case XT_SET_PACKET_XID:
 366 +                (*pskb)->skb_tag = setxidinfo->mark;
 367 +               break;
 368 +       }
 369 +       return XT_CONTINUE;
 370 +}
 371 +
 372 +
 373 +static int
 374 +checkentry_v1(const char *tablename,
 375 +             const void *entry,
 376 +             const struct xt_target *target,
 377 +             void *targinfo,
 378 +             unsigned int hook_mask)
 379 +{
 380 +       struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 381 +
 382 +       if (setxidinfo->mode != XT_SET_PACKET_XID) {
 383 +               printk(KERN_WARNING "SETXID: unknown mode %u\n",
 384 +                      setxidinfo->mode);
 385 +               return 0;
 386 +       }
 387 +
 388 +       return 1;
 389 +}
 390 +
 391 +static struct xt_target xt_setxid_target[] = {
 392 +       {
 393 +               .name           = "SETXID",
 394 +               .family         = AF_INET,
 395 +               .revision       = 1,
 396 +               .checkentry     = checkentry_v1,
 397 +               .target         = target_v1,
 398 +               .targetsize     = sizeof(struct xt_setxid_target_info_v1),
 399 +               .table          = "mangle",
 400 +               .me             = THIS_MODULE,
 401 +       }
 402 +};
 403 +
 404 +static int __init init(void)
 405 +{
 406 +       int err;
 407 +
 408 +       err = xt_register_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 409 +       return err;
 410 +}
 411 +
 412 +static void __exit fini(void)
 413 +{
 414 +       xt_unregister_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 415 +}
 416 +
 417 +module_init(init);
 418 +module_exit(fini);