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  2008-06-06 17:07:56.000000000 -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        2008-06-06 17:07:56.000000000 -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    2008-06-06 17:07:56.000000000 -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  2008-06-06 17:07:56.000000000 -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     2008-06-06 17:07:56.000000000 -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     2008-06-06 17:07:56.000000000 -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    2008-06-06 17:07:48.000000000 -0400
  83 +++ linux-2.6.22-520/include/linux/vserver/network.h    2008-06-06 17:07:56.000000000 -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       2008-06-06 17:07:56.000000000 -0400
  96 @@ -131,6 +131,9 @@
  97         /* Storage reserved for other modules: */
  98         union nf_conntrack_proto proto;
  99
 100 +       /* PLANETLAB. VNET-specific */
 101 +       int 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  2008-06-06 17:07:56.000000000 -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/dev.c linux-2.6.22-520/net/core/dev.c
 119 --- linux-2.6.22-510/net/core/dev.c     2008-06-06 17:07:48.000000000 -0400
 120 +++ linux-2.6.22-520/net/core/dev.c     2008-06-06 17:07:56.000000000 -0400
 121 @@ -1803,6 +1803,7 @@
 122   * the ingress scheduler, you just cant add policies on ingress.
 123   *
 124   */
 125 +
 126  static int ing_filter(struct sk_buff *skb)
 127  {
 128         struct Qdisc *q;
 129 @@ -1832,13 +1833,20 @@
 130  }
 131  #endif
 132
 133 +/* The code already makes the assumption that packet handlers run
 134 + * sequentially on the same CPU. -Sapan */
 135 +DEFINE_PER_CPU(int, sknid_elevator);
 136 +
 137  int netif_receive_skb(struct sk_buff *skb)
 138  {
 139         struct packet_type *ptype, *pt_prev;
 140         struct net_device *orig_dev;
 141         int ret = NET_RX_DROP;
 142 +       int *cur_elevator=&__get_cpu_var(sknid_elevator);
 143         __be16 type;
 144
 145 +       *cur_elevator = 0;
 146 +
 147         /* if we've gotten here through NAPI, check netpoll */
 148         if (skb->dev->poll && netpoll_rx(skb))
 149                 return NET_RX_DROP;
 150 @@ -1873,8 +1881,9 @@
 151
 152         list_for_each_entry_rcu(ptype, &ptype_all, list) {
 153                 if (!ptype->dev || ptype->dev == skb->dev) {
 154 -                       if (pt_prev)
 155 +                       if (pt_prev) {
 156                                 ret = deliver_skb(skb, pt_prev, orig_dev);
 157 +                       }
 158                         pt_prev = ptype;
 159                 }
 160         }
 161 @@ -1912,8 +1921,22 @@
 162                 }
 163         }
 164
 165 +       /* We don't want the packet handlers to throw the packet away
 166 +        * if we want the taps to treat it again - Sapan */
 167 +       if (*cur_elevator) {
 168 +               atomic_inc(&skb->users);
 169 +       }
 170 +
 171         if (pt_prev) {
 172                 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
 173 +                       if (*cur_elevator > 0) {
 174 +                               skb->skb_tag = *cur_elevator;
 175 +                               list_for_each_entry_rcu(ptype, &ptype_all, list) {
 176 +                                       if (!ptype->dev || ptype->dev == skb->dev) {
 177 +                                                       ret = deliver_skb(skb, ptype, orig_dev);
 178 +                                       }
 179 +                               }
 180 +                       }
 181         } else {
 182                 kfree_skb(skb);
 183                 /* Jamal, now you will not able to escape explaining
 184 @@ -1922,6 +1945,13 @@
 185                 ret = NET_RX_DROP;
 186         }
 187
 188 +       if (*cur_elevator) {
 189 +               /* We have a packet */
 190 +               kfree_skb(skb);
 191 +       }
 192 +
 193 +       *cur_elevator=0;
 194 +
 195  out:
 196         rcu_read_unlock();
 197         return ret;
 198 @@ -3780,6 +3810,7 @@
 199  EXPORT_SYMBOL(net_enable_timestamp);
 200  EXPORT_SYMBOL(net_disable_timestamp);
 201  EXPORT_SYMBOL(dev_get_flags);
 202 +EXPORT_PER_CPU_SYMBOL(sknid_elevator);
 203
 204  #if defined(CONFIG_BRIDGE) || defined(CONFIG_BRIDGE_MODULE)
 205  EXPORT_SYMBOL(br_handle_frame_hook);
 206 diff -Nurb linux-2.6.22-510/net/core/skbuff.c linux-2.6.22-520/net/core/skbuff.c
 207 --- linux-2.6.22-510/net/core/skbuff.c  2007-07-08 19:32:17.000000000 -0400
 208 +++ linux-2.6.22-520/net/core/skbuff.c  2008-06-06 17:07:56.000000000 -0400
 209 @@ -56,6 +56,7 @@
 210  #include <linux/rtnetlink.h>
 211  #include <linux/init.h>
 212  #include <linux/scatterlist.h>
 213 +#include <linux/vs_network.h>
 214
 215  #include <net/protocol.h>
 216  #include <net/dst.h>
 217 @@ -174,6 +175,7 @@
 218         skb->data = data;
 219         skb_reset_tail_pointer(skb);
 220         skb->end = skb->tail + size;
 221 +       if (!in_interrupt()) skb->skb_tag = nx_current_nid(); else skb->skb_tag = 0;
 222         /* make sure we initialize shinfo sequentially */
 223         shinfo = skb_shinfo(skb);
 224         atomic_set(&shinfo->dataref, 1);
 225 @@ -443,6 +445,8 @@
 226         C(tail);
 227         C(end);
 228
 229 +       /* Sapan: Cloned skbs aren't owned by anyone. Let the cloner decide who it belongs to. */
 230 +
 231         atomic_inc(&(skb_shinfo(skb)->dataref));
 232         skb->cloned = 1;
 233
 234 @@ -492,6 +496,7 @@
 235         new->tc_index   = old->tc_index;
 236  #endif
 237         skb_copy_secmark(new, old);
 238 +       new->skb_tag = old->skb_tag;
 239         atomic_set(&new->users, 1);
 240         skb_shinfo(new)->gso_size = skb_shinfo(old)->gso_size;
 241         skb_shinfo(new)->gso_segs = skb_shinfo(old)->gso_segs;
 242 diff -Nurb linux-2.6.22-510/net/core/sock.c linux-2.6.22-520/net/core/sock.c
 243 --- linux-2.6.22-510/net/core/sock.c    2008-06-06 17:07:48.000000000 -0400
 244 +++ linux-2.6.22-520/net/core/sock.c    2008-06-06 17:07:56.000000000 -0400
 245 @@ -444,6 +444,19 @@
 246                 }
 247                 goto set_sndbuf;
 248
 249 +       case SO_SETXID:
 250 +               if (current_vx_info()) {
 251 +                       ret = -EPERM;
 252 +                       break;
 253 +               }
 254 +               if (val < 0 || val > MAX_S_CONTEXT) {
 255 +                       ret = -EINVAL;
 256 +                       break;
 257 +               }
 258 +               sk->sk_xid = val;
 259 +               sk->sk_nid = val;
 260 +               break;
 261 +
 262         case SO_RCVBUF:
 263                 /* Don't error on this BSD doesn't and if you think
 264                    about it this is right. Otherwise apps have to
 265 @@ -573,7 +586,7 @@
 266                 char devname[IFNAMSIZ];
 267
 268                 /* Sorry... */
 269 -               if (!capable(CAP_NET_RAW)) {
 270 +               if (!nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET)) {
 271                         ret = -EPERM;
 272                         break;
 273                 }
 274 diff -Nurb linux-2.6.22-510/net/ipv4/af_inet.c linux-2.6.22-520/net/ipv4/af_inet.c
 275 --- linux-2.6.22-510/net/ipv4/af_inet.c 2008-06-06 17:07:48.000000000 -0400
 276 +++ linux-2.6.22-520/net/ipv4/af_inet.c 2008-06-06 17:07:56.000000000 -0400
 277 @@ -178,6 +178,8 @@
 278                         return -EAGAIN;
 279                 }
 280                 inet->sport = htons(inet->num);
 281 +               sk->sk_xid = vx_current_xid();
 282 +               if (!in_interrupt()) sk->sk_nid = nx_current_nid(); else sk->sk_nid=0;
 283         }
 284         release_sock(sk);
 285         return 0;
 286 @@ -312,6 +314,9 @@
 287         if ((protocol == IPPROTO_ICMP) &&
 288                 nx_capable(answer->capability, NXC_RAW_ICMP))
 289                 goto override;
 290 +       if (sock->type == SOCK_RAW &&
 291 +               nx_capable(answer->capability, NXC_RAW_SOCKET))
 292 +               goto override;
 293         if (answer->capability > 0 && !capable(answer->capability))
 294                 goto out_rcu_unlock;
 295  override:
 296 diff -Nurb linux-2.6.22-510/net/ipv4/icmp.c linux-2.6.22-520/net/ipv4/icmp.c
 297 --- linux-2.6.22-510/net/ipv4/icmp.c    2008-06-06 17:07:55.000000000 -0400
 298 +++ linux-2.6.22-520/net/ipv4/icmp.c    2008-06-06 17:07:56.000000000 -0400
 299 @@ -709,7 +709,7 @@
 300         if ((raw_sk = sk_head(&raw_v4_htable[hash])) != NULL) {
 301                 while ((raw_sk = __raw_v4_lookup(raw_sk, protocol, iph->daddr,
 302                                                  iph->saddr,
 303 -                                                skb->dev->ifindex)) != NULL) {
 304 +                                                skb->dev->ifindex, skb->skb_tag)) != NULL) {
 305                         raw_err(raw_sk, skb, info);
 306                         raw_sk = sk_next(raw_sk);
 307                         iph = (struct iphdr *)skb->data;
 308 diff -Nurb linux-2.6.22-510/net/ipv4/ip_options.c linux-2.6.22-520/net/ipv4/ip_options.c
 309 --- linux-2.6.22-510/net/ipv4/ip_options.c      2007-07-08 19:32:17.000000000 -0400
 310 +++ linux-2.6.22-520/net/ipv4/ip_options.c      2008-06-06 17:07:56.000000000 -0400
 311 @@ -409,7 +409,7 @@
 312                                         optptr[2] += 8;
 313                                         break;
 314                                       default:
 315 -                                       if (!skb && !capable(CAP_NET_RAW)) {
 316 +                                       if (!skb && !nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET)) {
 317                                                 pp_ptr = optptr + 3;
 318                                                 goto error;
 319                                         }
 320 @@ -445,7 +445,7 @@
 321                                 opt->router_alert = optptr - iph;
 322                         break;
 323                       case IPOPT_CIPSO:
 324 -                       if ((!skb && !capable(CAP_NET_RAW)) || opt->cipso) {
 325 +                       if ((!skb && !nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET)) || opt->cipso) {
 326                                 pp_ptr = optptr;
 327                                 goto error;
 328                         }
 329 @@ -458,7 +458,7 @@
 330                       case IPOPT_SEC:
 331                       case IPOPT_SID:
 332                       default:
 333 -                       if (!skb && !capable(CAP_NET_RAW)) {
 334 +                       if (!skb && !nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET)) {
 335                                 pp_ptr = optptr;
 336                                 goto error;
 337                         }
 338 diff -Nurb linux-2.6.22-510/net/ipv4/netfilter/ipt_LOG.c linux-2.6.22-520/net/ipv4/netfilter/ipt_LOG.c
 339 --- linux-2.6.22-510/net/ipv4/netfilter/ipt_LOG.c       2008-06-06 17:07:43.000000000 -0400
 340 +++ linux-2.6.22-520/net/ipv4/netfilter/ipt_LOG.c       2008-06-06 17:07:56.000000000 -0400
 341 @@ -49,6 +49,8 @@
 342         else
 343                 logflags = NF_LOG_MASK;
 344
 345 +       printk("TAG=%d ", skb->skb_tag);
 346 +
 347         ih = skb_header_pointer(skb, iphoff, sizeof(_iph), &_iph);
 348         if (ih == NULL) {
 349                 printk("TRUNCATED");
 350 diff -Nurb linux-2.6.22-510/net/ipv4/raw.c linux-2.6.22-520/net/ipv4/raw.c
 351 --- linux-2.6.22-510/net/ipv4/raw.c     2008-06-06 17:07:48.000000000 -0400
 352 +++ linux-2.6.22-520/net/ipv4/raw.c     2008-06-06 17:07:56.000000000 -0400
 353 @@ -103,7 +103,7 @@
 354
 355  struct sock *__raw_v4_lookup(struct sock *sk, unsigned short num,
 356                              __be32 raddr, __be32 laddr,
 357 -                            int dif)
 358 +                            int dif, int tag)
 359  {
 360         struct hlist_node *node;
 361
 362 @@ -112,6 +112,7 @@
 363
 364                 if (inet->num == num                                    &&
 365                     !(inet->daddr && inet->daddr != raddr)              &&
 366 +                   (!sk->sk_nx_info || tag == 1 || sk->sk_nid == tag)  &&
 367                     v4_sock_addr_match(sk->sk_nx_info, inet, laddr)     &&
 368                     !(sk->sk_bound_dev_if && sk->sk_bound_dev_if != dif))
 369                         goto found; /* gotcha */
 370 @@ -161,7 +162,7 @@
 371                 goto out;
 372         sk = __raw_v4_lookup(__sk_head(head), iph->protocol,
 373                              iph->saddr, iph->daddr,
 374 -                            skb->dev->ifindex);
 375 +                            skb->dev->ifindex, skb->skb_tag);
 376
 377         while (sk) {
 378                 delivered = 1;
 379 @@ -174,7 +175,7 @@
 380                 }
 381                 sk = __raw_v4_lookup(sk_next(sk), iph->protocol,
 382                                      iph->saddr, iph->daddr,
 383 -                                    skb->dev->ifindex);
 384 +                                    skb->dev->ifindex, skb->skb_tag);
 385         }
 386  out:
 387         read_unlock(&raw_v4_lock);
 388 @@ -315,7 +316,7 @@
 389         }
 390
 391         err = -EPERM;
 392 -       if (!nx_check(0, VS_ADMIN) && !capable(CAP_NET_RAW) &&
 393 +       if (!nx_check(0, VS_ADMIN) && !nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET) &&
 394                 sk->sk_nx_info &&
 395                 !v4_addr_in_nx_info(sk->sk_nx_info, iph->saddr, NXA_MASK_BIND))
 396                 goto error_free;
 397 diff -Nurb linux-2.6.22-510/net/netfilter/Kconfig linux-2.6.22-520/net/netfilter/Kconfig
 398 --- linux-2.6.22-510/net/netfilter/Kconfig      2007-07-08 19:32:17.000000000 -0400
 399 +++ linux-2.6.22-520/net/netfilter/Kconfig      2008-06-06 17:07:56.000000000 -0400
 400 @@ -389,6 +389,13 @@
 401
 402           To compile it as a module, choose M here.  If unsure, say N.
 403
 404 +config NETFILTER_XT_TARGET_SETXID
 405 +       tristate '"SETXID" target support'
 406 +       depends on NETFILTER_XTABLES
 407 +       help
 408 +         This option adds a `SETXID' target, which allows you to alter the
 409 +         xid of a socket.
 410 +
 411  config NETFILTER_XT_MATCH_COMMENT
 412         tristate  '"comment" match support'
 413         depends on NETFILTER_XTABLES
 414 diff -Nurb linux-2.6.22-510/net/netfilter/Makefile linux-2.6.22-520/net/netfilter/Makefile
 415 --- linux-2.6.22-510/net/netfilter/Makefile     2007-07-08 19:32:17.000000000 -0400
 416 +++ linux-2.6.22-520/net/netfilter/Makefile     2008-06-06 17:07:56.000000000 -0400
 417 @@ -37,6 +37,7 @@
 418  obj-$(CONFIG_NETFILTER_XTABLES) += x_tables.o xt_tcpudp.o
 419
 420  # targets
 421 +obj-$(CONFIG_NETFILTER_XT_TARGET_SETXID) += xt_SETXID.o
 422  obj-$(CONFIG_NETFILTER_XT_TARGET_CLASSIFY) += xt_CLASSIFY.o
 423  obj-$(CONFIG_NETFILTER_XT_TARGET_CONNMARK) += xt_CONNMARK.o
 424  obj-$(CONFIG_NETFILTER_XT_TARGET_DSCP) += xt_DSCP.o
 425 diff -Nurb linux-2.6.22-510/net/netfilter/nf_conntrack_core.c linux-2.6.22-520/net/netfilter/nf_conntrack_core.c
 426 --- linux-2.6.22-510/net/netfilter/nf_conntrack_core.c  2007-07-08 19:32:17.000000000 -0400
 427 +++ linux-2.6.22-520/net/netfilter/nf_conntrack_core.c  2008-06-06 17:07:56.000000000 -0400
 428 @@ -726,6 +726,8 @@
 429
 430         /* Overload tuple linked list to put us in unconfirmed list. */
 431         list_add(&conntrack->tuplehash[IP_CT_DIR_ORIGINAL].list, &unconfirmed);
 432 +       conntrack->xid[IP_CT_DIR_ORIGINAL] = -1;
 433 +       conntrack->xid[IP_CT_DIR_REPLY] = -1;
 434
 435         write_unlock_bh(&nf_conntrack_lock);
 436
 437 diff -Nurb linux-2.6.22-510/net/netfilter/xt_MARK.c linux-2.6.22-520/net/netfilter/xt_MARK.c
 438 --- linux-2.6.22-510/net/netfilter/xt_MARK.c    2007-07-08 19:32:17.000000000 -0400
 439 +++ linux-2.6.22-520/net/netfilter/xt_MARK.c    2008-06-07 17:55:26.000000000 -0400
 440 @@ -5,13 +5,18 @@
 441   * This program is free software; you can redistribute it and/or modify
 442   * it under the terms of the GNU General Public License version 2 as
 443   * published by the Free Software Foundation.
 444 + *
 445   */
 446
 447  #include <linux/module.h>
 448 +#include <linux/version.h>
 449  #include <linux/skbuff.h>
 450  #include <linux/ip.h>
 451  #include <net/checksum.h>
 452 +#include <net/route.h>
 453 +#include <net/inet_hashtables.h>
 454
 455 +#include <net/netfilter/nf_conntrack.h>
 456  #include <linux/netfilter/x_tables.h>
 457  #include <linux/netfilter/xt_MARK.h>
 458
 459 @@ -21,6 +26,48 @@
 460  MODULE_ALIAS("ipt_MARK");
 461  MODULE_ALIAS("ip6t_MARK");
 462
 463 +static inline u_int16_t
 464 +get_dst_port(struct nf_conntrack_tuple *tuple)
 465 +{
 466 +       switch (tuple->dst.protonum) {
 467 +       case IPPROTO_GRE:
 468 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 469 +#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,11)
 470 +               return tuple->dst.u.gre.key;
 471 +#else
 472 +               return htons(ntohl(tuple->dst.u.gre.key));
 473 +#endif
 474 +       case IPPROTO_ICMP:
 475 +               /* Bind on ICMP echo ID */
 476 +               return tuple->src.u.icmp.id;
 477 +       case IPPROTO_TCP:
 478 +               return tuple->dst.u.tcp.port;
 479 +       case IPPROTO_UDP:
 480 +               return tuple->dst.u.udp.port;
 481 +       default:
 482 +               return tuple->dst.u.all;
 483 +       }
 484 +}
 485 +
 486 +static inline u_int16_t
 487 +get_src_port(struct nf_conntrack_tuple *tuple)
 488 +{
 489 +       switch (tuple->dst.protonum) {
 490 +       case IPPROTO_GRE:
 491 +               /* XXX Truncate 32-bit GRE key to 16 bits */
 492 +               return htons(ntohl(tuple->src.u.gre.key));
 493 +       case IPPROTO_ICMP:
 494 +               /* Bind on ICMP echo ID */
 495 +               return tuple->src.u.icmp.id;
 496 +       case IPPROTO_TCP:
 497 +               return tuple->src.u.tcp.port;
 498 +       case IPPROTO_UDP:
 499 +               return tuple->src.u.udp.port;
 500 +       default:
 501 +               return tuple->src.u.all;
 502 +       }
 503 +}
 504 +
 505  static unsigned int
 506  target_v0(struct sk_buff **pskb,
 507           const struct net_device *in,
 508 @@ -35,6 +82,8 @@
 509         return XT_CONTINUE;
 510  }
 511
 512 +extern DEFINE_PER_CPU(int, sknid_elevator);
 513 +
 514  static unsigned int
 515  target_v1(struct sk_buff **pskb,
 516           const struct net_device *in,
 517 @@ -44,7 +93,20 @@
 518           const void *targinfo)
 519  {
 520         const struct xt_mark_target_info_v1 *markinfo = targinfo;
 521 -       int mark = 0;
 522 +       enum ip_conntrack_info ctinfo;
 523 +      struct sock *connection_sk;
 524 +      int dif;
 525 +      struct nf_conn *ct;
 526 +      extern struct inet_hashinfo tcp_hashinfo;
 527 +      enum ip_conntrack_dir dir;
 528 +      int *curtag;
 529 +      u_int32_t src_ip;
 530 +      u_int32_t dst_ip;
 531 +      u_int16_t proto, src_port;
 532 +      u_int32_t ip;
 533 +      u_int16_t port;
 534 +
 535 +       int mark = -1;
 536
 537         switch (markinfo->mode) {
 538         case XT_MARK_SET:
 539 @@ -58,13 +120,74 @@
 540         case XT_MARK_OR:
 541                 mark = (*pskb)->mark | markinfo->mark;
 542                 break;
 543 +
 544 +               case XT_MARK_COPYXID:
 545 +
 546 +                                             ct = nf_ct_get((*pskb), &ctinfo);
 547 +                                             if (!ct)
 548 +                                                     break;
 549 +
 550 +                                             dir = CTINFO2DIR(ctinfo);
 551 +                                             src_ip = ct->tuplehash[dir].tuple.src.u3.ip;
 552 +                                             dst_ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 553 +                                             src_port = get_src_port(&ct->tuplehash[dir].tuple);
 554 +                                             proto = ct->tuplehash[dir].tuple.dst.protonum;
 555 +
 556 +                                             dif = ((struct rtable *)(*pskb)->dst)->rt_iif;
 557 +                                             ip = ct->tuplehash[dir].tuple.dst.u3.ip;
 558 +                                             port = get_dst_port(&ct->tuplehash[dir].tuple);
 559 +
 560 +                                             if (proto == 1 || proto == 17) {
 561 +                                                     if ((*pskb)->mark>0) /* The packet is marked, it's going out */
 562 +                                                     {
 563 +                                                               //if (ct->xid[0]>0 && ct->xid[0]!=(*pskb)->mark)
 564 +                                                                       /*printk(KERN_CRIT "xt_MARK log: %d/%d/%d/%d\n",ct->xid[0],(*pskb)->mark,hooknum==NF_IP_LOCAL_IN,proto);*/
 565 +
 566 +                                                               ct->xid[0]=(*pskb)->mark;
 567         }
 568
 569 +                                                     if (ct->xid[0] > 0) {
 570 +                                                             mark = ct->xid[0];
 571 +                                                     }
 572 +
 573 +                                             }
 574 +                                             else if (proto == 6) {
 575 +                                                     if ((*pskb)->sk) {
 576 +                                                             connection_sk = (*pskb)->sk;
 577 +                                                             sock_hold(connection_sk);
 578 +                                                     }
 579 +                                                     else
 580 +                                                             connection_sk = inet_lookup_established(&tcp_hashinfo, src_ip, src_port, ip, port, dif);
 581 +
 582 +
 583 +                                                     if (connection_sk) {
 584 +                                                             if (connection_sk->sk_state == TCP_TIME_WAIT) {
 585 +                                                                     inet_twsk_put(inet_twsk(connection_sk));
 586 +                                                                     break;
 587 +                                                             }
 588 +                                                             connection_sk->sk_peercred.gid = connection_sk->sk_peercred.uid = ct->xid[dir];
 589 +                                                             ct->xid[!dir]=connection_sk->sk_nid;
 590 +                                                             if (connection_sk->sk_nid != 0)
 591 +                                                                     mark = connection_sk->sk_nid;
 592 +                                                             sock_put(connection_sk);
 593 +                                                     }
 594 +                                                     else
 595 +                                                             mark = -1 ;
 596 +                                             }
 597 +                                             break;
 598 +       }
 599 +       if (mark != -1) {
 600         (*pskb)->mark = mark;
 601 +       }
 602 +
 603 +       curtag=&__get_cpu_var(sknid_elevator);
 604 +       if (mark > 0 && *curtag==-2)
 605 +       {
 606 +               *curtag = mark;
 607 +       }
 608         return XT_CONTINUE;
 609  }
 610
 611 -
 612  static int
 613  checkentry_v0(const char *tablename,
 614               const void *entry,
 615 @@ -92,7 +215,8 @@
 616
 617         if (markinfo->mode != XT_MARK_SET
 618             && markinfo->mode != XT_MARK_AND
 619 -           && markinfo->mode != XT_MARK_OR) {
 620 +           && markinfo->mode != XT_MARK_OR
 621 +           && markinfo->mode != XT_MARK_COPYXID) {
 622                 printk(KERN_WARNING "MARK: unknown mode %u\n",
 623                        markinfo->mode);
 624                 return 0;
 625 diff -Nurb linux-2.6.22-510/net/netfilter/xt_SETXID.c linux-2.6.22-520/net/netfilter/xt_SETXID.c
 626 --- linux-2.6.22-510/net/netfilter/xt_SETXID.c  1969-12-31 19:00:00.000000000 -0500
 627 +++ linux-2.6.22-520/net/netfilter/xt_SETXID.c  2008-06-06 17:07:56.000000000 -0400
 628 @@ -0,0 +1,79 @@
 629 +#include <linux/module.h>
 630 +#include <linux/skbuff.h>
 631 +#include <linux/ip.h>
 632 +#include <net/checksum.h>
 633 +#include <linux/vs_network.h>
 634 +
 635 +#include <linux/netfilter/x_tables.h>
 636 +#include <linux/netfilter/xt_SETXID.h>
 637 +
 638 +MODULE_LICENSE("GPL");
 639 +MODULE_AUTHOR("");
 640 +MODULE_DESCRIPTION("");
 641 +MODULE_ALIAS("ipt_SETXID");
 642 +
 643 +static unsigned int
 644 +target_v1(struct sk_buff **pskb,
 645 +         const struct net_device *in,
 646 +         const struct net_device *out,
 647 +         unsigned int hooknum,
 648 +         const struct xt_target *target,
 649 +         const void *targinfo)
 650 +{
 651 +       const struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 652 +
 653 +       switch (setxidinfo->mode) {
 654 +       case XT_SET_PACKET_XID:
 655 +                (*pskb)->skb_tag = setxidinfo->mark;
 656 +               break;
 657 +       }
 658 +       return XT_CONTINUE;
 659 +}
 660 +
 661 +
 662 +static int
 663 +checkentry_v1(const char *tablename,
 664 +             const void *entry,
 665 +             const struct xt_target *target,
 666 +             void *targinfo,
 667 +             unsigned int hook_mask)
 668 +{
 669 +       struct xt_setxid_target_info_v1 *setxidinfo = targinfo;
 670 +
 671 +       if (setxidinfo->mode != XT_SET_PACKET_XID) {
 672 +               printk(KERN_WARNING "SETXID: unknown mode %u\n",
 673 +                      setxidinfo->mode);
 674 +               return 0;
 675 +       }
 676 +
 677 +       return 1;
 678 +}
 679 +
 680 +static struct xt_target xt_setxid_target[] = {
 681 +       {
 682 +               .name           = "SETXID",
 683 +               .family         = AF_INET,
 684 +               .revision       = 1,
 685 +               .checkentry     = checkentry_v1,
 686 +               .target         = target_v1,
 687 +               .targetsize     = sizeof(struct xt_setxid_target_info_v1),
 688 +               .table          = "mangle",
 689 +               .me             = THIS_MODULE,
 690 +       }
 691 +};
 692 +
 693 +static int __init init(void)
 694 +{
 695 +       int err;
 696 +
 697 +       err = xt_register_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 698 +       return err;
 699 +}
 700 +
 701 +static void __exit fini(void)
 702 +{
 703 +       xt_unregister_targets(xt_setxid_target, ARRAY_SIZE(xt_setxid_target));
 704 +}
 705 +
 706 +module_init(init);
 707 +module_exit(fini);
 708 diff -Nurb linux-2.6.22-510/net/packet/af_packet.c linux-2.6.22-520/net/packet/af_packet.c
 709 --- linux-2.6.22-510/net/packet/af_packet.c     2007-07-08 19:32:17.000000000 -0400
 710 +++ linux-2.6.22-520/net/packet/af_packet.c     2008-06-07 18:30:41.000000000 -0400
 711 @@ -78,6 +78,7 @@
 712  #include <linux/poll.h>
 713  #include <linux/module.h>
 714  #include <linux/init.h>
 715 +#include <linux/vs_network.h>
 716
 717  #ifdef CONFIG_INET
 718  #include <net/inet_common.h>
 719 @@ -246,10 +247,13 @@
 720
 721  static const struct proto_ops packet_ops_spkt;
 722
 723 +extern DEFINE_PER_CPU(int, sknid_elevator);
 724  static int packet_rcv_spkt(struct sk_buff *skb, struct net_device *dev,  struct packet_type *pt, struct net_device *orig_dev)
 725  {
 726         struct sock *sk;
 727         struct sockaddr_pkt *spkt;
 728 +       int tag = skb->skb_tag;
 729 +       int *elevator=&__get_cpu_var(sknid_elevator);
 730
 731         /*
 732          *      When we registered the protocol we saved the socket in the data
 733 @@ -269,6 +273,22 @@
 734          *      so that this procedure is noop.
 735          */
 736
 737 +       /*
 738 +        * (18:05:41) daniel_hozac: where?
 739 +        * (18:05:58) daniel_hozac: we already have filters on PF_PACKET, don't we?
 740 +        * (18:05:58) er: in packet_rcv_skpt
 741 +        * (18:07:33) daniel_hozac: oh, that's evil.
 742 +        */
 743 +
 744 +       if (sk->sk_nx_info && !(tag == 1 || sk->sk_nid == tag)) {
 745 +               *elevator=-2;
 746 +               goto out;
 747 +       }
 748 +       else if (!sk->sk_nx_info && *elevator) {
 749 +               /* Root has already seen this packet */
 750 +               goto out;
 751 +       }
 752 +
 753         if (skb->pkt_type == PACKET_LOOPBACK)
 754                 goto out;
 755
 756 @@ -324,6 +344,9 @@
 757         __be16 proto=0;
 758         int err;
 759
 760 +       if (!nx_capable(CAP_NET_RAW, NXC_RAW_SEND))
 761 +               return -EPERM;
 762 +
 763         /*
 764          *      Get and verify the address.
 765          */
 766 @@ -420,6 +443,17 @@
 767                                       unsigned int res)
 768  {
 769         struct sk_filter *filter;
 770 +       int tag = skb->skb_tag;
 771 +       int *elevator=&__get_cpu_var(sknid_elevator);
 772 +
 773 +       if (sk->sk_nx_info && !(tag == 1 || sk->sk_nid == tag)) {
 774 +               *elevator=-2;
 775 +               return 0;
 776 +       }
 777 +       else if (!sk->sk_nx_info && *elevator) {
 778 +               /* Root has already seen this packet */
 779 +               return 0;
 780 +       }
 781
 782         rcu_read_lock_bh();
 783         filter = rcu_dereference(sk->sk_filter);
 784 @@ -711,6 +745,9 @@
 785         unsigned char *addr;
 786         int ifindex, err, reserve = 0;
 787
 788 +       if (!nx_capable(CAP_NET_RAW, NXC_RAW_SEND))
 789 +               return -EPERM;
 790 +
 791         /*
 792          *      Get and verify the address.
 793          */
 794 @@ -984,8 +1021,9 @@
 795         __be16 proto = (__force __be16)protocol; /* weird, but documented */
 796         int err;
 797
 798 -       if (!capable(CAP_NET_RAW))
 799 +       if (!nx_capable(CAP_NET_RAW, NXC_RAW_SOCKET))
 800                 return -EPERM;
 801 +
 802         if (sock->type != SOCK_DGRAM && sock->type != SOCK_RAW &&
 803             sock->type != SOCK_PACKET)
 804                 return -ESOCKTNOSUPPORT;