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  2008-07-27 22:06:25.000000000 -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        2008-07-27 22:06:25.000000000 -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    2008-07-27 22:06:25.000000000 -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  2008-07-27 22:06:25.000000000 -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       2008-07-27 22:06:25.000000000 -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      2008-07-27 22:06:25.000000000 -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     2008-07-27 22:06:25.000000000 -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  2008-07-27 22:06:25.000000000 -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    2008-07-28 16:21:20.000000000 -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 +26,48 @@
 134  MODULE_ALIAS("ipt_MARK");
 135  MODULE_ALIAS("ip6t_MARK");
 136
 137 +static inline u_int16_t
 138 +get_dst_port(struct nf_conntrack_tuple *tuple)
 139 +{
 140 +       switch (tuple->dst.protonum) {
 141 +       case IPPROTO_GRE:
 142 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 143 +#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,11)
 144 +               return tuple->dst.u.gre.key;
 145 +#else
 146 +               return htons(ntohl(tuple->dst.u.gre.key));
 147 +#endif
 148 +       case IPPROTO_ICMP:
 149 +               /* Bind on ICMP echo ID */
 150 +               return tuple->src.u.icmp.id;
 151 +       case IPPROTO_TCP:
 152 +               return tuple->dst.u.tcp.port;
 153 +       case IPPROTO_UDP:
 154 +               return tuple->dst.u.udp.port;
 155 +       default:
 156 +               return tuple->dst.u.all;
 157 +       }
 158 +}
 159 +
 160 +static inline u_int16_t
 161 +get_src_port(struct nf_conntrack_tuple *tuple)
 162 +{
 163 +       switch (tuple->dst.protonum) {
 164 +       case IPPROTO_GRE:
 165 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 166 +               return htons(ntohl(tuple->src.u.gre.key));
 167 +       case IPPROTO_ICMP:
 168 +               /* Bind on ICMP echo ID */
 169 +               return tuple->src.u.icmp.id;
 170 +       case IPPROTO_TCP:
 171 +               return tuple->src.u.tcp.port;
 172 +       case IPPROTO_UDP:
 173 +               return tuple->src.u.udp.port;
 174 +       default:
 175 +               return tuple->src.u.all;
 176 +       }
 177 +}
 178 +
 179  static unsigned int
 180  target_v0(struct sk_buff **pskb,
 181           const struct net_device *in,
 182 @@ -35,6 +82,10 @@
 183         return XT_CONTINUE;
 184  }
 185
 186 +extern DEFINE_PER_CPU(int, sknid_elevator);
 187 +
 188 +#define related(ct) (ct==(IP_CT_IS_REPLY + IP_CT_RELATED))
 189 +
 190  static unsigned int
 191  target_v1(struct sk_buff **pskb,
 192           const struct net_device *in,
 193 @@ -44,7 +95,20 @@
 194           const void *targinfo)
 195  {
 196         const struct xt_mark_target_info_v1 *markinfo = targinfo;
 197 -       int mark = 0;
 198 +       enum ip_conntrack_info ctinfo;
 199 +      struct sock *connection_sk;
 200 +      int dif;
 201 +      struct nf_conn *ct;
 202 +      extern struct inet_hashinfo tcp_hashinfo;
 203 +      enum ip_conntrack_dir dir;
 204 +      int *curtag;
 205 +      u_int32_t src_ip;
 206 +      u_int32_t dst_ip;
 207 +      u_int16_t proto, src_port;
 208 +      u_int32_t ip;
 209 +      u_int16_t port;
 210 +
 211 +       int mark = -1;
 212
 213         switch (markinfo->mode) {
 214         case XT_MARK_SET:
 215 @@ -58,13 +122,102 @@
 216         case XT_MARK_OR:
 217                 mark = (*pskb)->mark | markinfo->mark;
 218                 break;
 219 +
 220 +       case XT_MARK_COPYXID:
 221 +             dif = ((struct rtable *)(*pskb)->dst)->rt_iif;
 222 +
 223 +             ct = nf_ct_get((*pskb), &ctinfo);
 224 +             if (!ct)
 225 +                     break;
 226 +
 227 +             dir = CTINFO2DIR(ctinfo);
 228 +             src_ip = ct->tuplehash[dir].tuple.src.u3.ip;
 229 +             dst_ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 230 +             src_port = get_src_port(&ct->tuplehash[dir].tuple);
 231 +             proto = ct->tuplehash[dir].tuple.dst.protonum;
 232 +
 233 +             ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 234 +             port = get_dst_port(&ct->tuplehash[dir].tuple);
 235 +
 236 +             if (proto == 1 || proto == 17) {
 237 +                     if ((*pskb)->mark>0) /* The packet is marked, it's going out */
 238 +                     {
 239 +                             ct->xid[0]=(*pskb)->mark;
 240 +                     }
 241 +
 242 +                     if (ct->xid[0] > 0) {
 243 +                             mark = ct->xid[0];
 244 +                     }
 245 +
 246 +             }
 247 +             else if (proto == 6) {
 248 +                     if ((*pskb)->sk) {
 249 +                             /* It's a listening socket */
 250 +                             connection_sk = (*pskb)->sk;
 251 +                             sock_hold(connection_sk);
 252 +                     }
 253 +                     else   /* It might be a connected socket */
 254 +                             connection_sk = inet_lookup_established(&tcp_hashinfo, src_ip, src_port, ip, port, dif);
 255 +
 256 +
 257 +                     if (connection_sk /* Well, some kind of TCP socket */) {
 258 +                             if (connection_sk->sk_peercred.uid < 1) {
 259 +                                     /* Normal case - the peercred on the socket is not set */
 260 +                                     connection_sk->sk_peercred.gid = connection_sk->sk_peercred.uid = ct->xid[!dir];
 261 +                             }
 262 +                             else    /* Exceptional case - the peercred was set using SET_PEERCRED. Somebody wants us
 263 +                                        to mark packets with some arbitrary value.*/
 264 +                                     mark=connection_sk->sk_peercred.uid;
 265 +
 266 +                             /* Has this connection already been tagged? */
 267 +                             if (ct->xid[dir] < 1) {
 268 +                                     /* No - let's tag it */
 269 +                                     ct->xid[dir]=connection_sk->sk_nid;
 270 +                             }
 271 +
 272 +                             if (mark==-1 && (connection_sk->sk_nid != 0))
 273 +                                     mark = ct->xid[dir];
 274 +
 275 +
 276 +                             if (connection_sk->sk_state == TCP_TIME_WAIT) {
 277 +                                     inet_twsk_put(inet_twsk(connection_sk));
 278 +                                     break;
 279 +                             }
 280 +                             else
 281 +                                     sock_put(connection_sk);
 282 +                     }
 283 +
 284 +                     /* Covers CoDemux case */
 285 +                     if (mark < 1 && (ct->xid[dir]>0)) {
 286 +                               mark = ct->xid[dir];
 287         }
 288
 289 +                     if (mark < 1 && (ct->xid[!dir]>0)) {
 290 +                               mark = ct->xid[!dir];
 291 +                     }
 292 +
 293 +                     /* All else failed. Is this a connection over raw sockets? That explains
 294 +                      * why we couldn't get anything out of skb->sk, or look up a "real" connection.*/
 295 +                     if (ct->xid[dir]<1) {
 296 +                       if ((*pskb)->skb_tag) {
 297 +                               ct->xid[dir]=(*pskb)->skb_tag;
 298 +                       }
 299 +                     }
 300 +             }
 301 +             break;
 302 +       }
 303 +       if (mark != -1) {
 304         (*pskb)->mark = mark;
 305 +       }
 306 +
 307 +       curtag=&__get_cpu_var(sknid_elevator);
 308 +       if (mark > 0 && *curtag==-2 && hooknum==NF_IP_LOCAL_IN)
 309 +       {
 310 +               *curtag = mark;
 311 +       }
 312         return XT_CONTINUE;
 313  }
 314
 315 -
 316  static int
 317  checkentry_v0(const char *tablename,
 318               const void *entry,
 319 @@ -92,7 +245,8 @@
 320
 321         if (markinfo->mode != XT_MARK_SET
 322             && markinfo->mode != XT_MARK_AND
 323 -           && markinfo->mode != XT_MARK_OR) {
 324 +           && markinfo->mode != XT_MARK_OR
 325 +           && markinfo->mode != XT_MARK_COPYXID) {
 326                 printk(KERN_WARNING "MARK: unknown mode %u\n",
 327                        markinfo->mode);
 328                 return 0;
 329 diff -Nurb linux-2.6.22-521/net/netfilter/xt_SETXID.c linux-2.6.22-522/net/netfilter/xt_SETXID.c
 330 --- linux-2.6.22-521/net/netfilter/xt_SETXID.c  1969-12-31 19:00:00.000000000 -0500
 331 +++ linux-2.6.22-522/net/netfilter/xt_SETXID.c  2008-07-27 22:06:25.000000000 -0400
 332 @@ -0,0 +1,79 @@
 333 +#include <linux/module.h>
 334 +#include <linux/skbuff.h>
 335 +#include <linux/ip.h>
 336 +#include <net/checksum.h>
 337 +#include <linux/vs_network.h>
 338 +
 339 +#include <linux/netfilter/x_tables.h>
 340 +#include <linux/netfilter/xt_SETXID.h>
 341 +
 342 +MODULE_LICENSE("GPL");
 343 +MODULE_AUTHOR("");
 344 +MODULE_DESCRIPTION("");
 345 +MODULE_ALIAS("ipt_SETXID");
 346 +
 347 +static unsigned int
 348 +target_v1(struct sk_buff **pskb,
 349 +         const struct net_device *in,
 350 +         const struct net_device *out,
 351 +         unsigned int hooknum,
 352 +         const struct xt_target *target,
 353 +         const void *targinfo)
 354 +{
 355 +       const struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 356 +
 357 +       switch (setxidinfo->mode) {
 358 +       case XT_SET_PACKET_XID:
 359 +                (*pskb)->skb_tag = setxidinfo->mark;
 360 +               break;
 361 +       }
 362 +       return XT_CONTINUE;
 363 +}
 364 +
 365 +
 366 +static int
 367 +checkentry_v1(const char *tablename,
 368 +             const void *entry,
 369 +             const struct xt_target *target,
 370 +             void *targinfo,
 371 +             unsigned int hook_mask)
 372 +{
 373 +       struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 374 +
 375 +       if (setxidinfo->mode != XT_SET_PACKET_XID) {
 376 +               printk(KERN_WARNING "SETXID: unknown mode %u\n",
 377 +                      setxidinfo->mode);
 378 +               return 0;
 379 +       }
 380 +
 381 +       return 1;
 382 +}
 383 +
 384 +static struct xt_target xt_setxid_target[] = {
 385 +       {
 386 +               .name           = "SETXID",
 387 +               .family         = AF_INET,
 388 +               .revision       = 1,
 389 +               .checkentry     = checkentry_v1,
 390 +               .target         = target_v1,
 391 +               .targetsize     = sizeof(struct xt_setxid_target_info_v1),
 392 +               .table          = "mangle",
 393 +               .me             = THIS_MODULE,
 394 +       }
 395 +};
 396 +
 397 +static int __init init(void)
 398 +{
 399 +       int err;
 400 +
 401 +       err = xt_register_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 402 +       return err;
 403 +}
 404 +
 405 +static void __exit fini(void)
 406 +{
 407 +       xt_unregister_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 408 +}
 409 +
 410 +module_init(init);
 411 +module_exit(fini);