linux-2.6-520-vnet+.patch

   1 diff -Nurb linux-2.6.22-510/include/linux/netfilter/xt_MARK.h linux-2.6.22-520/include/linux/netfilter/xt_MARK.h
   2 --- linux-2.6.22-510/include/linux/netfilter/xt_MARK.h  2007-07-08 19:32:17.000000000 -0400
   3 +++ linux-2.6.22-520/include/linux/netfilter/xt_MARK.h  2008-05-05 00:06:39.000000000 -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-2.6.22-510/include/linux/netfilter/xt_SETXID.h linux-2.6.22-520/include/linux/netfilter/xt_SETXID.h
  13 --- linux-2.6.22-510/include/linux/netfilter/xt_SETXID.h        1969-12-31 19:00:00.000000000 -0500
  14 +++ linux-2.6.22-520/include/linux/netfilter/xt_SETXID.h        2008-05-05 00:06:39.000000000 -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-2.6.22-510/include/linux/netfilter_ipv4/ipt_MARK.h linux-2.6.22-520/include/linux/netfilter_ipv4/ipt_MARK.h
  31 --- linux-2.6.22-510/include/linux/netfilter_ipv4/ipt_MARK.h    2007-07-08 19:32:17.000000000 -0400
  32 +++ linux-2.6.22-520/include/linux/netfilter_ipv4/ipt_MARK.h    2008-05-05 00:06:39.000000000 -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-2.6.22-510/include/linux/netfilter_ipv4/ipt_SETXID.h linux-2.6.22-520/include/linux/netfilter_ipv4/ipt_SETXID.h
  42 --- linux-2.6.22-510/include/linux/netfilter_ipv4/ipt_SETXID.h  1969-12-31 19:00:00.000000000 -0500
  43 +++ linux-2.6.22-520/include/linux/netfilter_ipv4/ipt_SETXID.h  2008-05-05 00:06:39.000000000 -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-2.6.22-510/include/linux/skbuff.h linux-2.6.22-520/include/linux/skbuff.h
  59 --- linux-2.6.22-510/include/linux/skbuff.h     2007-07-08 19:32:17.000000000 -0400
  60 +++ linux-2.6.22-520/include/linux/skbuff.h     2008-05-05 00:06:39.000000000 -0400
  61 @@ -302,6 +302,7 @@
  62  #endif
  63
  64         __u32                   mark;
  65 +#define skb_tag                        mark
  66
  67         sk_buff_data_t          transport_header;
  68         sk_buff_data_t          network_header;
  69 diff -Nurb linux-2.6.22-510/include/linux/socket.h linux-2.6.22-520/include/linux/socket.h
  70 --- linux-2.6.22-510/include/linux/socket.h     2007-07-08 19:32:17.000000000 -0400
  71 +++ linux-2.6.22-520/include/linux/socket.h     2008-05-05 00:06:39.000000000 -0400
  72 @@ -288,6 +288,8 @@
  73  #define SOL_TIPC       271
  74  #define SOL_RXRPC      272
  75
  76 +#define SO_SETXID      SO_PEERCRED
  77 +
  78  /* IPX options */
  79  #define IPX_TYPE       1
  80
  81 diff -Nurb linux-2.6.22-510/include/linux/vserver/network.h linux-2.6.22-520/include/linux/vserver/network.h
  82 --- linux-2.6.22-510/include/linux/vserver/network.h    2008-05-05 00:06:36.000000000 -0400
  83 +++ linux-2.6.22-520/include/linux/vserver/network.h    2008-05-05 00:06:39.000000000 -0400
  84 @@ -47,6 +47,8 @@
  85  #define NXC_TUN_CREATE         0x00000001
  86
  87  #define NXC_RAW_ICMP           0x00000100
  88 +#define NXC_RAW_SOCKET         0x00000200
  89 +#define NXC_RAW_SEND           0x00000400
  90
  91
  92  /* address types */
  93 diff -Nurb linux-2.6.22-510/include/net/netfilter/nf_conntrack.h linux-2.6.22-520/include/net/netfilter/nf_conntrack.h
  94 --- linux-2.6.22-510/include/net/netfilter/nf_conntrack.h       2007-07-08 19:32:17.000000000 -0400
  95 +++ linux-2.6.22-520/include/net/netfilter/nf_conntrack.h       2008-05-05 00:06:39.000000000 -0400
  96 @@ -131,6 +131,9 @@
  97         /* Storage reserved for other modules: */
  98         union nf_conntrack_proto proto;
  99
 100 +       /* PLANETLAB. VNET-specific */
 101 +       xid_t xid[IP_CT_DIR_MAX];
 102 +
 103         /* features dynamically at the end: helper, nat (both optional) */
 104         char data[0];
 105  };
 106 diff -Nurb linux-2.6.22-510/include/net/raw.h linux-2.6.22-520/include/net/raw.h
 107 --- linux-2.6.22-510/include/net/raw.h  2007-07-08 19:32:17.000000000 -0400
 108 +++ linux-2.6.22-520/include/net/raw.h  2008-05-05 00:06:39.000000000 -0400
 109 @@ -36,7 +36,7 @@
 110
 111  extern struct sock *__raw_v4_lookup(struct sock *sk, unsigned short num,
 112                                     __be32 raddr, __be32 laddr,
 113 -                                   int dif);
 114 +                                   int dif, int tag);
 115
 116  extern int raw_v4_input(struct sk_buff *skb, struct iphdr *iph, int hash);
 117
 118 diff -Nurb linux-2.6.22-510/net/core/skbuff.c linux-2.6.22-520/net/core/skbuff.c
 119 --- linux-2.6.22-510/net/core/skbuff.c  2007-07-08 19:32:17.000000000 -0400
 120 +++ linux-2.6.22-520/net/core/skbuff.c  2008-05-05 00:06:39.000000000 -0400
 121 @@ -56,6 +56,7 @@
 122  #include <linux/rtnetlink.h>
 123  #include <linux/init.h>
 124  #include <linux/scatterlist.h>
 125 +#include <linux/vs_network.h>
 126
 127  #include <net/protocol.h>
 128  #include <net/dst.h>
 129 @@ -174,6 +175,7 @@
 130         skb->data = data;
 131         skb_reset_tail_pointer(skb);
 132         skb->end = skb->tail + size;
 133 +       skb->skb_tag = nx_current_nid();
 134         /* make sure we initialize shinfo sequentially */
 135         shinfo = skb_shinfo(skb);
 136         atomic_set(&shinfo->dataref, 1);
 137 @@ -443,6 +445,8 @@
 138         C(tail);
 139         C(end);
 140
 141 +       /* Sapan: Cloned skbs aren't owned by anyone. Let the cloner decide who it belongs to. */
 142 +
 143         atomic_inc(&(skb_shinfo(skb)->dataref));
 144         skb->cloned = 1;
 145
 146 @@ -492,6 +496,7 @@
 147         new->tc_index   = old->tc_index;
 148  #endif
 149         skb_copy_secmark(new, old);
 150 +       new->skb_tag = old->skb_tag;
 151         atomic_set(&new->users, 1);
 152         skb_shinfo(new)->gso_size = skb_shinfo(old)->gso_size;
 153         skb_shinfo(new)->gso_segs = skb_shinfo(old)->gso_segs;
 154 diff -Nurb linux-2.6.22-510/net/core/sock.c linux-2.6.22-520/net/core/sock.c
 155 --- linux-2.6.22-510/net/core/sock.c    2008-05-05 00:06:36.000000000 -0400
 156 +++ linux-2.6.22-520/net/core/sock.c    2008-05-05 00:06:39.000000000 -0400
 157 @@ -444,6 +444,19 @@
 158                 }
 159                 goto set_sndbuf;
 160
 161 +       case SO_SETXID:
 162 +               if (current_vx_info()) {
 163 +                       ret = -EPERM;
 164 +                       break;
 165 +               }
 166 +               if (val < 0 || val > MAX_S_CONTEXT) {
 167 +                       ret = -EINVAL;
 168 +                       break;
 169 +               }
 170 +               sk->sk_xid = val;
 171 +               sk->sk_nid = val;
 172 +               break;
 173 +
 174         case SO_RCVBUF:
 175                 /* Don't error on this BSD doesn't and if you think
 176                    about it this is right. Otherwise apps have to
 177 @@ -573,7 +586,7 @@
 178                 char devname[IFNAMSIZ];
 179
 180                 /* Sorry... */
 181 -               if (!capable(CAP_NET_RAW)) {
 182 +               if (!nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET)) {
 183                         ret = -EPERM;
 184                         break;
 185                 }
 186 diff -Nurb linux-2.6.22-510/net/ipv4/af_inet.c linux-2.6.22-520/net/ipv4/af_inet.c
 187 --- linux-2.6.22-510/net/ipv4/af_inet.c 2008-05-05 00:06:36.000000000 -0400
 188 +++ linux-2.6.22-520/net/ipv4/af_inet.c 2008-05-05 00:06:39.000000000 -0400
 189 @@ -178,6 +178,8 @@
 190                         return -EAGAIN;
 191                 }
 192                 inet->sport = htons(inet->num);
 193 +               sk->sk_xid = vx_current_xid();
 194 +               sk->sk_nid = nx_current_nid();
 195         }
 196         release_sock(sk);
 197         return 0;
 198 @@ -312,6 +314,9 @@
 199         if ((protocol == IPPROTO_ICMP) &&
 200                 nx_capable(answer->capability, NXC_RAW_ICMP))
 201                 goto override;
 202 +       if (sock->type == SOCK_RAW &&
 203 +               nx_capable(answer->capability, NXC_RAW_SOCKET))
 204 +               goto override;
 205         if (answer->capability > 0 && !capable(answer->capability))
 206                 goto out_rcu_unlock;
 207  override:
 208 diff -Nurb linux-2.6.22-510/net/ipv4/icmp.c linux-2.6.22-520/net/ipv4/icmp.c
 209 --- linux-2.6.22-510/net/ipv4/icmp.c    2008-05-05 00:06:39.000000000 -0400
 210 +++ linux-2.6.22-520/net/ipv4/icmp.c    2008-05-05 00:06:39.000000000 -0400
 211 @@ -709,7 +709,7 @@
 212         if ((raw_sk = sk_head(&raw_v4_htable[hash])) != NULL) {
 213                 while ((raw_sk = __raw_v4_lookup(raw_sk, protocol, iph->daddr,
 214                                                  iph->saddr,
 215 -                                                skb->dev->ifindex)) != NULL) {
 216 +                                                skb->dev->ifindex, skb->skb_tag)) != NULL) {
 217                         raw_err(raw_sk, skb, info);
 218                         raw_sk = sk_next(raw_sk);
 219                         iph = (struct iphdr *)skb->data;
 220 diff -Nurb linux-2.6.22-510/net/ipv4/ip_options.c linux-2.6.22-520/net/ipv4/ip_options.c
 221 --- linux-2.6.22-510/net/ipv4/ip_options.c      2007-07-08 19:32:17.000000000 -0400
 222 +++ linux-2.6.22-520/net/ipv4/ip_options.c      2008-05-05 00:06:39.000000000 -0400
 223 @@ -409,7 +409,7 @@
 224                                         optptr[2] += 8;
 225                                         break;
 226                                       default:
 227 -                                       if (!skb && !capable(CAP_NET_RAW)) {
 228 +                                       if (!skb && !nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET)) {
 229                                                 pp_ptr = optptr + 3;
 230                                                 goto error;
 231                                         }
 232 @@ -445,7 +445,7 @@
 233                                 opt->router_alert = optptr - iph;
 234                         break;
 235                       case IPOPT_CIPSO:
 236 -                       if ((!skb && !capable(CAP_NET_RAW)) || opt->cipso) {
 237 +                       if ((!skb && !nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET)) || opt->cipso) {
 238                                 pp_ptr = optptr;
 239                                 goto error;
 240                         }
 241 @@ -458,7 +458,7 @@
 242                       case IPOPT_SEC:
 243                       case IPOPT_SID:
 244                       default:
 245 -                       if (!skb && !capable(CAP_NET_RAW)) {
 246 +                       if (!skb && !nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET)) {
 247                                 pp_ptr = optptr;
 248                                 goto error;
 249                         }
 250 diff -Nurb linux-2.6.22-510/net/ipv4/netfilter/ipt_LOG.c linux-2.6.22-520/net/ipv4/netfilter/ipt_LOG.c
 251 --- linux-2.6.22-510/net/ipv4/netfilter/ipt_LOG.c       2008-05-05 00:06:26.000000000 -0400
 252 +++ linux-2.6.22-520/net/ipv4/netfilter/ipt_LOG.c       2008-05-05 00:06:39.000000000 -0400
 253 @@ -49,6 +49,8 @@
 254         else
 255                 logflags = NF_LOG_MASK;
 256
 257 +       printk("TAG=%d ", skb->skb_tag);
 258 +
 259         ih = skb_header_pointer(skb, iphoff, sizeof(_iph), &_iph);
 260         if (ih == NULL) {
 261                 printk("TRUNCATED");
 262 diff -Nurb linux-2.6.22-510/net/ipv4/raw.c linux-2.6.22-520/net/ipv4/raw.c
 263 --- linux-2.6.22-510/net/ipv4/raw.c     2008-05-05 00:06:36.000000000 -0400
 264 +++ linux-2.6.22-520/net/ipv4/raw.c     2008-05-05 00:06:39.000000000 -0400
 265 @@ -103,7 +103,7 @@
 266
 267  struct sock *__raw_v4_lookup(struct sock *sk, unsigned short num,
 268                              __be32 raddr, __be32 laddr,
 269 -                            int dif)
 270 +                            int dif, int tag)
 271  {
 272         struct hlist_node *node;
 273
 274 @@ -112,6 +112,7 @@
 275
 276                 if (inet->num == num                                    &&
 277                     !(inet->daddr && inet->daddr != raddr)              &&
 278 +                   (!sk->sk_nx_info || tag == 1 || sk->sk_nid == tag)  &&
 279                     v4_sock_addr_match(sk->sk_nx_info, inet, laddr)     &&
 280                     !(sk->sk_bound_dev_if && sk->sk_bound_dev_if != dif))
 281                         goto found; /* gotcha */
 282 @@ -161,7 +162,7 @@
 283                 goto out;
 284         sk = __raw_v4_lookup(__sk_head(head), iph->protocol,
 285                              iph->saddr, iph->daddr,
 286 -                            skb->dev->ifindex);
 287 +                            skb->dev->ifindex, skb->skb_tag);
 288
 289         while (sk) {
 290                 delivered = 1;
 291 @@ -174,7 +175,7 @@
 292                 }
 293                 sk = __raw_v4_lookup(sk_next(sk), iph->protocol,
 294                                      iph->saddr, iph->daddr,
 295 -                                    skb->dev->ifindex);
 296 +                                    skb->dev->ifindex, skb->skb_tag);
 297         }
 298  out:
 299         read_unlock(&raw_v4_lock);
 300 @@ -315,7 +316,7 @@
 301         }
 302
 303         err = -EPERM;
 304 -       if (!nx_check(0, VS_ADMIN) && !capable(CAP_NET_RAW) &&
 305 +       if (!nx_check(0, VS_ADMIN) && !nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET) &&
 306                 sk->sk_nx_info &&
 307                 !v4_addr_in_nx_info(sk->sk_nx_info, iph->saddr, NXA_MASK_BIND))
 308                 goto error_free;
 309 diff -Nurb linux-2.6.22-510/net/netfilter/Kconfig linux-2.6.22-520/net/netfilter/Kconfig
 310 --- linux-2.6.22-510/net/netfilter/Kconfig      2007-07-08 19:32:17.000000000 -0400
 311 +++ linux-2.6.22-520/net/netfilter/Kconfig      2008-05-05 00:06:39.000000000 -0400
 312 @@ -389,6 +389,13 @@
 313
 314           To compile it as a module, choose M here.  If unsure, say N.
 315
 316 +config NETFILTER_XT_TARGET_SETXID
 317 +       tristate '"SETXID" target support'
 318 +       depends on NETFILTER_XTABLES
 319 +       help
 320 +         This option adds a `SETXID' target, which allows you to alter the
 321 +         xid of a socket.
 322 +
 323  config NETFILTER_XT_MATCH_COMMENT
 324         tristate  '"comment" match support'
 325         depends on NETFILTER_XTABLES
 326 diff -Nurb linux-2.6.22-510/net/netfilter/Makefile linux-2.6.22-520/net/netfilter/Makefile
 327 --- linux-2.6.22-510/net/netfilter/Makefile     2007-07-08 19:32:17.000000000 -0400
 328 +++ linux-2.6.22-520/net/netfilter/Makefile     2008-05-05 00:06:39.000000000 -0400
 329 @@ -37,6 +37,7 @@
 330  obj-$(CONFIG_NETFILTER_XTABLES) += x_tables.o xt_tcpudp.o
 331
 332  # targets
 333 +obj-$(CONFIG_NETFILTER_XT_TARGET_SETXID) += xt_SETXID.o
 334  obj-$(CONFIG_NETFILTER_XT_TARGET_CLASSIFY) += xt_CLASSIFY.o
 335  obj-$(CONFIG_NETFILTER_XT_TARGET_CONNMARK) += xt_CONNMARK.o
 336  obj-$(CONFIG_NETFILTER_XT_TARGET_DSCP) += xt_DSCP.o
 337 diff -Nurb linux-2.6.22-510/net/netfilter/nf_conntrack_core.c linux-2.6.22-520/net/netfilter/nf_conntrack_core.c
 338 --- linux-2.6.22-510/net/netfilter/nf_conntrack_core.c  2007-07-08 19:32:17.000000000 -0400
 339 +++ linux-2.6.22-520/net/netfilter/nf_conntrack_core.c  2008-05-05 00:06:39.000000000 -0400
 340 @@ -726,6 +726,8 @@
 341
 342         /* Overload tuple linked list to put us in unconfirmed list. */
 343         list_add(&conntrack->tuplehash[IP_CT_DIR_ORIGINAL].list, &unconfirmed);
 344 +       conntrack->xid[IP_CT_DIR_ORIGINAL] = -1;
 345 +       conntrack->xid[IP_CT_DIR_REPLY] = -1;
 346
 347         write_unlock_bh(&nf_conntrack_lock);
 348
 349 diff -Nurb linux-2.6.22-510/net/netfilter/xt_MARK.c linux-2.6.22-520/net/netfilter/xt_MARK.c
 350 --- linux-2.6.22-510/net/netfilter/xt_MARK.c    2007-07-08 19:32:17.000000000 -0400
 351 +++ linux-2.6.22-520/net/netfilter/xt_MARK.c    2008-05-09 14:09:29.000000000 -0400
 352 @@ -5,13 +5,18 @@
 353   * This program is free software; you can redistribute it and/or modify
 354   * it under the terms of the GNU General Public License version 2 as
 355   * published by the Free Software Foundation.
 356 + *
 357   */
 358
 359  #include <linux/module.h>
 360 +#include <linux/version.h>
 361  #include <linux/skbuff.h>
 362  #include <linux/ip.h>
 363  #include <net/checksum.h>
 364 +#include <net/route.h>
 365 +#include <net/inet_hashtables.h>
 366
 367 +#include <net/netfilter/nf_conntrack.h>
 368  #include <linux/netfilter/x_tables.h>
 369  #include <linux/netfilter/xt_MARK.h>
 370
 371 @@ -21,6 +26,48 @@
 372  MODULE_ALIAS("ipt_MARK");
 373  MODULE_ALIAS("ip6t_MARK");
 374
 375 +static inline u_int16_t
 376 +get_dst_port(struct nf_conntrack_tuple *tuple)
 377 +{
 378 +       switch (tuple->dst.protonum) {
 379 +       case IPPROTO_GRE:
 380 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 381 +#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,11)
 382 +               return tuple->dst.u.gre.key;
 383 +#else
 384 +               return htons(ntohl(tuple->dst.u.gre.key));
 385 +#endif
 386 +       case IPPROTO_ICMP:
 387 +               /* Bind on ICMP echo ID */
 388 +               return tuple->src.u.icmp.id;
 389 +       case IPPROTO_TCP:
 390 +               return tuple->dst.u.tcp.port;
 391 +       case IPPROTO_UDP:
 392 +               return tuple->dst.u.udp.port;
 393 +       default:
 394 +               return tuple->dst.u.all;
 395 +       }
 396 +}
 397 +
 398 +static inline u_int16_t
 399 +get_src_port(struct nf_conntrack_tuple *tuple)
 400 +{
 401 +       switch (tuple->dst.protonum) {
 402 +       case IPPROTO_GRE:
 403 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 404 +               return htons(ntohl(tuple->src.u.gre.key));
 405 +       case IPPROTO_ICMP:
 406 +               /* Bind on ICMP echo ID */
 407 +               return tuple->src.u.icmp.id;
 408 +       case IPPROTO_TCP:
 409 +               return tuple->src.u.tcp.port;
 410 +       case IPPROTO_UDP:
 411 +               return tuple->src.u.udp.port;
 412 +       default:
 413 +               return tuple->src.u.all;
 414 +       }
 415 +}
 416 +
 417  static unsigned int
 418  target_v0(struct sk_buff **pskb,
 419           const struct net_device *in,
 420 @@ -44,7 +91,16 @@
 421           const void *targinfo)
 422  {
 423         const struct xt_mark_target_info_v1 *markinfo = targinfo;
 424 -       int mark = 0;
 425 +       struct nf_conn *ct;
 426 +       int mark = -1;
 427 +       extern struct inet_hashinfo tcp_hashinfo;
 428 +       enum ip_conntrack_dir dir;
 429 +       u_int32_t src_ip;
 430 +       u_int16_t src_port;
 431 +       u_int16_t proto;
 432 +
 433 +        u_int32_t ip;
 434 +        u_int16_t port;
 435
 436         switch (markinfo->mode) {
 437         case XT_MARK_SET:
 438 @@ -58,8 +114,58 @@
 439         case XT_MARK_OR:
 440                 mark = (*pskb)->mark | markinfo->mark;
 441                 break;
 442 +
 443 +               case XT_MARK_COPYXID: {
 444 +                                             enum ip_conntrack_info ctinfo;
 445 +                                             struct sock *connection_sk;
 446 +                                             int dif;
 447 +
 448 +                                             ct = nf_ct_get((*pskb), &ctinfo);
 449 +                                             if (!ct)
 450 +                                                     break;
 451 +
 452 +                                             dir = CTINFO2DIR(ctinfo);
 453 +                                             src_ip = ct->tuplehash[dir].tuple.src.u3.ip;
 454 +                                             src_port = get_src_port(&ct->tuplehash[dir].tuple);
 455 +                                             proto = ct->tuplehash[dir].tuple.dst.protonum;
 456 +
 457 +                                             dif = ((struct rtable *)(*pskb)->dst)->rt_iif;
 458 +                                             ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 459 +                                             port = get_dst_port(&ct->tuplehash[dir].tuple);
 460 +
 461 +                                             if (proto == 1 || proto == 17) {
 462 +                                                     if (((*pskb)->mark!=-1) && (*pskb)->mark)
 463 +                                                             ct->xid[0]=(*pskb)->mark;
 464 +                                                     if (ct->xid[0])
 465 +                                                             mark = ct->xid[0];
 466 +
 467 +                                             }
 468 +                                             else if (proto == 6) {
 469 +                                                     if ((*pskb)->sk) {
 470 +                                                             connection_sk = (*pskb)->sk;
 471 +                                                             sock_hold(connection_sk);
 472 +                                                     }
 473 +                                                     else
 474 +                                                             connection_sk = inet_lookup_established(&tcp_hashinfo, src_ip, src_port, ip, port, dif);
 475 +
 476 +
 477 +                                                     if (connection_sk) {
 478 +                                                             if (connection_sk->sk_state == TCP_TIME_WAIT) {
 479 +                                                                     inet_twsk_put(inet_twsk(connection_sk));
 480 +                                                                     break;
 481 +                                                             }
 482 +                                                             connection_sk->sk_peercred.gid = connection_sk->sk_peercred.uid = ct->xid[dir];
 483 +                                                             ct->xid[!dir]=connection_sk->sk_xid;
 484 +                                                             if (connection_sk->sk_xid != 0)
 485 +                                                                     mark = connection_sk->sk_xid;
 486 +                                                             sock_put(connection_sk);
 487 +                                                     }
 488 +                                             }
 489 +                                             break;
 490 +                                     }
 491         }
 492
 493 +       if (mark != -1)
 494         (*pskb)->mark = mark;
 495         return XT_CONTINUE;
 496  }
 497 @@ -92,7 +198,8 @@
 498
 499         if (markinfo->mode != XT_MARK_SET
 500             && markinfo->mode != XT_MARK_AND
 501 -           && markinfo->mode != XT_MARK_OR) {
 502 +           && markinfo->mode != XT_MARK_OR
 503 +           && markinfo->mode != XT_MARK_COPYXID) {
 504                 printk(KERN_WARNING "MARK: unknown mode %u\n",
 505                        markinfo->mode);
 506                 return 0;
 507 diff -Nurb linux-2.6.22-510/net/netfilter/xt_SETXID.c linux-2.6.22-520/net/netfilter/xt_SETXID.c
 508 --- linux-2.6.22-510/net/netfilter/xt_SETXID.c  1969-12-31 19:00:00.000000000 -0500
 509 +++ linux-2.6.22-520/net/netfilter/xt_SETXID.c  2008-05-05 00:06:39.000000000 -0400
 510 @@ -0,0 +1,79 @@
 511 +#include <linux/module.h>
 512 +#include <linux/skbuff.h>
 513 +#include <linux/ip.h>
 514 +#include <net/checksum.h>
 515 +#include <linux/vs_network.h>
 516 +
 517 +#include <linux/netfilter/x_tables.h>
 518 +#include <linux/netfilter/xt_SETXID.h>
 519 +
 520 +MODULE_LICENSE("GPL");
 521 +MODULE_AUTHOR("");
 522 +MODULE_DESCRIPTION("");
 523 +MODULE_ALIAS("ipt_SETXID");
 524 +
 525 +static unsigned int
 526 +target_v1(struct sk_buff **pskb,
 527 +         const struct net_device *in,
 528 +         const struct net_device *out,
 529 +         unsigned int hooknum,
 530 +         const struct xt_target *target,
 531 +         const void *targinfo)
 532 +{
 533 +       const struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 534 +
 535 +       switch (setxidinfo->mode) {
 536 +       case XT_SET_PACKET_XID:
 537 +                (*pskb)->skb_tag = setxidinfo->mark;
 538 +               break;
 539 +       }
 540 +       return XT_CONTINUE;
 541 +}
 542 +
 543 +
 544 +static int
 545 +checkentry_v1(const char *tablename,
 546 +             const void *entry,
 547 +             const struct xt_target *target,
 548 +             void *targinfo,
 549 +             unsigned int hook_mask)
 550 +{
 551 +       struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 552 +
 553 +       if (setxidinfo->mode != XT_SET_PACKET_XID) {
 554 +               printk(KERN_WARNING "SETXID: unknown mode %u\n",
 555 +                      setxidinfo->mode);
 556 +               return 0;
 557 +       }
 558 +
 559 +       return 1;
 560 +}
 561 +
 562 +static struct xt_target xt_setxid_target[] = {
 563 +       {
 564 +               .name           = "SETXID",
 565 +               .family         = AF_INET,
 566 +               .revision       = 1,
 567 +               .checkentry     = checkentry_v1,
 568 +               .target         = target_v1,
 569 +               .targetsize     = sizeof(struct xt_setxid_target_info_v1),
 570 +               .table          = "mangle",
 571 +               .me             = THIS_MODULE,
 572 +       }
 573 +};
 574 +
 575 +static int __init init(void)
 576 +{
 577 +       int err;
 578 +
 579 +       err = xt_register_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 580 +       return err;
 581 +}
 582 +
 583 +static void __exit fini(void)
 584 +{
 585 +       xt_unregister_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 586 +}
 587 +
 588 +module_init(init);
 589 +module_exit(fini);
 590 diff -Nurb linux-2.6.22-510/net/packet/af_packet.c linux-2.6.22-520/net/packet/af_packet.c
 591 --- linux-2.6.22-510/net/packet/af_packet.c     2007-07-08 19:32:17.000000000 -0400
 592 +++ linux-2.6.22-520/net/packet/af_packet.c     2008-05-05 00:06:39.000000000 -0400
 593 @@ -78,6 +78,7 @@
 594  #include <linux/poll.h>
 595  #include <linux/module.h>
 596  #include <linux/init.h>
 597 +#include <linux/vs_network.h>
 598
 599  #ifdef CONFIG_INET
 600  #include <net/inet_common.h>
 601 @@ -324,6 +325,9 @@
 602         __be16 proto=0;
 603         int err;
 604
 605 +       if (!nx_capable(CAP_NET_RAW, NXC_RAW_SEND))
 606 +               return -EPERM;
 607 +
 608         /*
 609          *      Get and verify the address.
 610          */
 611 @@ -420,6 +424,10 @@
 612                                       unsigned int res)
 613  {
 614         struct sk_filter *filter;
 615 +       int tag = skb->skb_tag;
 616 +
 617 +       if (sk->sk_nx_info && !(tag == 1 || sk->sk_nid == tag))
 618 +               return 0;
 619
 620         rcu_read_lock_bh();
 621         filter = rcu_dereference(sk->sk_filter);
 622 @@ -711,6 +719,9 @@
 623         unsigned char *addr;
 624         int ifindex, err, reserve = 0;
 625
 626 +       if (!nx_capable(CAP_NET_RAW, NXC_RAW_SEND))
 627 +               return -EPERM;
 628 +
 629         /*
 630          *      Get and verify the address.
 631          */
 632 @@ -984,7 +995,7 @@
 633         __be16 proto = (__force __be16)protocol; /* weird, but documented */
 634         int err;
 635
 636 -       if (!capable(CAP_NET_RAW))
 637 +       if (!nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET))
 638                 return -EPERM;
 639         if (sock->type != SOCK_DGRAM && sock->type != SOCK_RAW &&
 640             sock->type != SOCK_PACKET)
 641 diff -Nurb linux-2.6.22-510/xt_MARK.c linux-2.6.22-520/xt_MARK.c
 642 --- linux-2.6.22-510/xt_MARK.c  1969-12-31 19:00:00.000000000 -0500
 643 +++ linux-2.6.22-520/xt_MARK.c  2008-05-05 00:06:39.000000000 -0400
 644 @@ -0,0 +1,287 @@
 645 +/* This is a module which is used for setting the NFMARK field of an skb. */
 646 +
 647 +/* (C) 1999-2001 Marc Boucher <marc@mbsi.ca>
 648 + *
 649 + * This program is free software; you can redistribute it and/or modify
 650 + * it under the terms of the GNU General Public License version 2 as
 651 + * published by the Free Software Foundation.
 652 + *
 653 + */
 654 +
 655 +#include <linux/module.h>
 656 +#include <linux/version.h>
 657 +#include <linux/skbuff.h>
 658 +#include <linux/ip.h>
 659 +#include <net/checksum.h>
 660 +#include <net/route.h>
 661 +#include <net/inet_hashtables.h>
 662 +
 663 +#include <net/netfilter/nf_conntrack.h>
 664 +#include <linux/netfilter/x_tables.h>
 665 +#include <linux/netfilter/xt_MARK.h>
 666 +
 667 +MODULE_LICENSE("GPL");
 668 +MODULE_AUTHOR("Marc Boucher <marc@mbsi.ca>");
 669 +MODULE_DESCRIPTION("ip[6]tables MARK modification module");
 670 +MODULE_ALIAS("ipt_MARK");
 671 +MODULE_ALIAS("ip6t_MARK");
 672 +
 673 +static inline u_int16_t
 674 +get_dst_port(struct nf_conntrack_tuple *tuple)
 675 +{
 676 +       switch (tuple->dst.protonum) {
 677 +       case IPPROTO_GRE:
 678 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 679 +#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,11)
 680 +               return tuple->dst.u.gre.key;
 681 +#else
 682 +               return htons(ntohl(tuple->dst.u.gre.key));
 683 +#endif
 684 +       case IPPROTO_ICMP:
 685 +               /* Bind on ICMP echo ID */
 686 +               return tuple->src.u.icmp.id;
 687 +       case IPPROTO_TCP:
 688 +               return tuple->dst.u.tcp.port;
 689 +       case IPPROTO_UDP:
 690 +               return tuple->dst.u.udp.port;
 691 +       default:
 692 +               return tuple->dst.u.all;
 693 +       }
 694 +}
 695 +
 696 +static inline u_int16_t
 697 +get_src_port(struct nf_conntrack_tuple *tuple)
 698 +{
 699 +       switch (tuple->dst.protonum) {
 700 +       case IPPROTO_GRE:
 701 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 702 +               return htons(ntohl(tuple->src.u.gre.key));
 703 +       case IPPROTO_ICMP:
 704 +               /* Bind on ICMP echo ID */
 705 +               return tuple->src.u.icmp.id;
 706 +       case IPPROTO_TCP:
 707 +               return tuple->src.u.tcp.port;
 708 +       case IPPROTO_UDP:
 709 +               return tuple->src.u.udp.port;
 710 +       default:
 711 +               return tuple->src.u.all;
 712 +       }
 713 +}
 714 +
 715 +static unsigned int
 716 +target_v0(struct sk_buff **pskb,
 717 +         const struct net_device *in,
 718 +         const struct net_device *out,
 719 +         unsigned int hooknum,
 720 +         const struct xt_target *target,
 721 +         const void *targinfo)
 722 +{
 723 +       const struct xt_mark_target_info *markinfo = targinfo;
 724 +
 725 +       (*pskb)->mark = markinfo->mark;
 726 +       return XT_CONTINUE;
 727 +}
 728 +
 729 +static unsigned int
 730 +target_v1(struct sk_buff **pskb,
 731 +         const struct net_device *in,
 732 +         const struct net_device *out,
 733 +         unsigned int hooknum,
 734 +         const struct xt_target *target,
 735 +         const void *targinfo)
 736 +{
 737 +       const struct xt_mark_target_info_v1 *markinfo = targinfo;
 738 +       int mark = -1;
 739 +
 740 +       switch (markinfo->mode) {
 741 +               case XT_MARK_SET:
 742 +                       mark = markinfo->mark;
 743 +                       break;
 744 +
 745 +               case XT_MARK_AND:
 746 +                       mark = (*pskb)->mark & markinfo->mark;
 747 +                       break;
 748 +
 749 +               case XT_MARK_OR:
 750 +                       mark = (*pskb)->mark | markinfo->mark;
 751 +                       break;
 752 +
 753 +               case XT_MARK_COPYXID: {
 754 +                                             enum ip_conntrack_info ctinfo;
 755 +                                             struct sock *connection_sk;
 756 +                                             int dif;
 757 +
 758 +                                             struct nf_conn *ct = nf_ct_get((*pskb), &ctinfo);
 759 +                                             extern struct inet_hashinfo tcp_hashinfo;
 760 +                                             enum ip_conntrack_dir dir;
 761 +                                             if (!ct)
 762 +                                                     break;
 763 +
 764 +                                             dir = CTINFO2DIR(ctinfo);
 765 +                                             u_int32_t src_ip = ct->tuplehash[dir].tuple.src.u3.ip;
 766 +                                             u_int16_t src_port = get_src_port(&ct->tuplehash[dir].tuple);
 767 +                                             u_int16_t proto = ct->tuplehash[dir].tuple.dst.protonum;
 768 +
 769 +                                             u_int32_t ip;
 770 +                                             u_int16_t port;
 771 +
 772 +                                             dif = ((struct rtable *)(*pskb)->dst)->rt_iif;
 773 +                                             ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 774 +                                             port = get_dst_port(&ct->tuplehash[dir].tuple);
 775 +
 776 +                                             if (proto == 1 || proto == 17) {
 777 +                                                     if (((*pskb)->mark!=-1) && (*pskb)->mark)
 778 +                                                             ct->xid[0]=(*pskb)->mark;
 779 +                                                     if (ct->xid[0])
 780 +                                                             mark = ct->xid[0];
 781 +
 782 +                                             }
 783 +                                             else if (proto == 6) {
 784 +                                                     if ((*pskb)->sk) {
 785 +                                                             connection_sk = (*pskb)->sk;
 786 +                                                             sock_hold(connection_sk);
 787 +                                                     }
 788 +                                                     else {
 789 +                                                             connection_sk = inet_lookup(&tcp_hashinfo, src_ip, src_port, ip, port, dif);
 790 +                                                     }
 791 +
 792 +                                                     if (connection_sk) {
 793 +                                                             connection_sk->sk_peercred.gid = connection_sk->sk_peercred.uid = ct->xid[dir];
 794 +                                                             ct->xid[!dir]=connection_sk->sk_xid;
 795 +                                                             if (connection_sk->sk_xid != 0)
 796 +                                                                     mark = connection_sk->sk_xid;
 797 +                                                             if (connection_sk->sk_state == TCP_TIME_WAIT)
 798 +                                                                     inet_twsk_put(inet_twsk(connection_sk));
 799 +                                                             else
 800 +                                                                     sock_put(connection_sk);
 801 +                                                     }
 802 +                                             }
 803 +                                             break;
 804 +                                     }
 805 +       }
 806 +
 807 +       if (mark != -1)
 808 +       (*pskb)->mark = mark;
 809 +       return XT_CONTINUE;
 810 +}
 811 +
 812 +
 813 +static int
 814 +checkentry_v0(const char *tablename,
 815 +             const void *entry,
 816 +             const struct xt_target *target,
 817 +             void *targinfo,
 818 +             unsigned int hook_mask)
 819 +{
 820 +       struct xt_mark_target_info *markinfo = targinfo;
 821 +
 822 +       if (markinfo->mark > 0xffffffff) {
 823 +               printk(KERN_WARNING "MARK: Only supports 32bit wide mark\n");
 824 +               return 0;
 825 +       }
 826 +       return 1;
 827 +}
 828 +
 829 +static int
 830 +checkentry_v1(const char *tablename,
 831 +             const void *entry,
 832 +             const struct xt_target *target,
 833 +             void *targinfo,
 834 +             unsigned int hook_mask)
 835 +{
 836 +       struct xt_mark_target_info_v1 *markinfo = targinfo;
 837 +
 838 +       if (markinfo->mode != XT_MARK_SET
 839 +           && markinfo->mode != XT_MARK_AND
 840 +           && markinfo->mode != XT_MARK_OR
 841 +           && markinfo->mode != XT_MARK_COPYXID) {
 842 +               printk(KERN_WARNING "MARK: unknown mode %u\n",
 843 +                      markinfo->mode);
 844 +               return 0;
 845 +       }
 846 +       if (markinfo->mark > 0xffffffff) {
 847 +               printk(KERN_WARNING "MARK: Only supports 32bit wide mark\n");
 848 +               return 0;
 849 +       }
 850 +       return 1;
 851 +}
 852 +
 853 +#ifdef CONFIG_COMPAT
 854 +struct compat_xt_mark_target_info_v1 {
 855 +       compat_ulong_t  mark;
 856 +       u_int8_t        mode;
 857 +       u_int8_t        __pad1;
 858 +       u_int16_t       __pad2;
 859 +};
 860 +
 861 +static void compat_from_user_v1(void *dst, void *src)
 862 +{
 863 +       struct compat_xt_mark_target_info_v1 *cm = src;
 864 +       struct xt_mark_target_info_v1 m = {
 865 +               .mark   = cm->mark,
 866 +               .mode   = cm->mode,
 867 +       };
 868 +       memcpy(dst, &m, sizeof(m));
 869 +}
 870 +
 871 +static int compat_to_user_v1(void __user *dst, void *src)
 872 +{
 873 +       struct xt_mark_target_info_v1 *m = src;
 874 +       struct compat_xt_mark_target_info_v1 cm = {
 875 +               .mark   = m->mark,
 876 +               .mode   = m->mode,
 877 +       };
 878 +       return copy_to_user(dst, &cm, sizeof(cm)) ? -EFAULT : 0;
 879 +}
 880 +#endif /* CONFIG_COMPAT */
 881 +
 882 +static struct xt_target xt_mark_target[] = {
 883 +       {
 884 +               .name           = "MARK",
 885 +               .family         = AF_INET,
 886 +               .revision       = 0,
 887 +               .checkentry     = checkentry_v0,
 888 +               .target         = target_v0,
 889 +               .targetsize     = sizeof(struct xt_mark_target_info),
 890 +               .table          = "mangle",
 891 +               .me             = THIS_MODULE,
 892 +       },
 893 +       {
 894 +               .name           = "MARK",
 895 +               .family         = AF_INET,
 896 +               .revision       = 1,
 897 +               .checkentry     = checkentry_v1,
 898 +               .target         = target_v1,
 899 +               .targetsize     = sizeof(struct xt_mark_target_info_v1),
 900 +#ifdef CONFIG_COMPAT
 901 +               .compatsize     = sizeof(struct compat_xt_mark_target_info_v1),
 902 +               .compat_from_user = compat_from_user_v1,
 903 +               .compat_to_user = compat_to_user_v1,
 904 +#endif
 905 +               .table          = "mangle",
 906 +               .me             = THIS_MODULE,
 907 +       },
 908 +       {
 909 +               .name           = "MARK",
 910 +               .family         = AF_INET6,
 911 +               .revision       = 0,
 912 +               .checkentry     = checkentry_v0,
 913 +               .target         = target_v0,
 914 +               .targetsize     = sizeof(struct xt_mark_target_info),
 915 +               .table          = "mangle",
 916 +               .me             = THIS_MODULE,
 917 +       },
 918 +};
 919 +
 920 +static int __init xt_mark_init(void)
 921 +{
 922 +       return xt_register_targets(xt_mark_target, ARRAY_SIZE(xt_mark_target));
 923 +}
 924 +
 925 +static void __exit xt_mark_fini(void)
 926 +{
 927 +       xt_unregister_targets(xt_mark_target, ARRAY_SIZE(xt_mark_target));
 928 +}
 929 +
 930 +module_init(xt_mark_init);
 931 +module_exit(xt_mark_fini);