net/ipv4/ip_output.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              The Internet Protocol (IP) output module.
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Donald Becker, <becker@super.org>
  11  *              Alan Cox, <Alan.Cox@linux.org>
  12  *              Richard Underwood
  13  *              Stefan Becker, <stefanb@yello.ping.de>
  14  *              Jorge Cwik, <jorge@laser.satlink.net>
  15  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  16  *              Hirokazu Takahashi, <taka@valinux.co.jp>
  17  *
  18  *      See ip_input.c for original log
  19  *
  20  *      Fixes:
  21  *              Alan Cox        :       Missing nonblock feature in ip_build_xmit.
  22  *              Mike Kilburn    :       htons() missing in ip_build_xmit.
  23  *              Bradford Johnson:       Fix faulty handling of some frames when
  24  *                                      no route is found.
  25  *              Alexander Demenshin:    Missing sk/skb free in ip_queue_xmit
  26  *                                      (in case if packet not accepted by
  27  *                                      output firewall rules)
  28  *              Mike McLagan    :       Routing by source
  29  *              Alexey Kuznetsov:       use new route cache
  30  *              Andi Kleen:             Fix broken PMTU recovery and remove
  31  *                                      some redundant tests.
  32  *      Vitaly E. Lavrov        :       Transparent proxy revived after year coma.
  33  *              Andi Kleen      :       Replace ip_reply with ip_send_reply.
  34  *              Andi Kleen      :       Split fast and slow ip_build_xmit path
  35  *                                      for decreased register pressure on x86
  36  *                                      and more readibility.
  37  *              Marc Boucher    :       When call_out_firewall returns FW_QUEUE,
  38  *                                      silently drop skb instead of failing with -EPERM.
  39  *              Detlev Wengorz  :       Copy protocol for fragments.
  40  *              Hirokazu Takahashi:     HW checksumming for outgoing UDP
  41  *                                      datagrams.
  42  *              Hirokazu Takahashi:     sendfile() on UDP works now.
  43  */
  44
  45 #include <asm/uaccess.h>
  46 #include <asm/system.h>
  47 #include <linux/module.h>
  48 #include <linux/types.h>
  49 #include <linux/kernel.h>
  50 #include <linux/mm.h>
  51 #include <linux/string.h>
  52 #include <linux/errno.h>
  53 #include <linux/highmem.h>
  54
  55 #include <linux/socket.h>
  56 #include <linux/sockios.h>
  57 #include <linux/in.h>
  58 #include <linux/inet.h>
  59 #include <linux/netdevice.h>
  60 #include <linux/etherdevice.h>
  61 #include <linux/proc_fs.h>
  62 #include <linux/stat.h>
  63 #include <linux/init.h>
  64
  65 #include <net/snmp.h>
  66 #include <net/ip.h>
  67 #include <net/protocol.h>
  68 #include <net/route.h>
  69 #include <net/xfrm.h>
  70 #include <linux/skbuff.h>
  71 #include <net/sock.h>
  72 #include <net/arp.h>
  73 #include <net/icmp.h>
  74 #include <net/checksum.h>
  75 #include <net/inetpeer.h>
  76 #include <linux/igmp.h>
  77 #include <linux/netfilter_ipv4.h>
  78 #include <linux/netfilter_bridge.h>
  79 #include <linux/mroute.h>
  80 #include <linux/netlink.h>
  81 #include <linux/tcp.h>
  82
  83 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
  84
  85 /* Generate a checksum for an outgoing IP datagram. */
  86 __inline__ void ip_send_check(struct iphdr *iph)
  87 {
  88         iph->check = 0;
  89         iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
  90 }
  91
  92 int __ip_local_out(struct sk_buff *skb)
  93 {
  94         struct iphdr *iph = ip_hdr(skb);
  95
  96         iph->tot_len = htons(skb->len);
  97         ip_send_check(iph);
  98         return nf_hook(PF_INET, NF_INET_LOCAL_OUT, skb, NULL, skb_dst(skb)->dev,
  99                        dst_output);
 100 }
 101
 102 int ip_local_out(struct sk_buff *skb)
 103 {
 104         int err;
 105
 106         err = __ip_local_out(skb);
 107         if (likely(err == 1))
 108                 err = dst_output(skb);
 109
 110         return err;
 111 }
 112 EXPORT_SYMBOL_GPL(ip_local_out);
 113
 114 /* dev_loopback_xmit for use with netfilter. */
 115 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
 116 {
 117         skb_reset_mac_header(newskb);
 118         __skb_pull(newskb, skb_network_offset(newskb));
 119         newskb->pkt_type = PACKET_LOOPBACK;
 120         newskb->ip_summed = CHECKSUM_UNNECESSARY;
 121         WARN_ON(!skb_dst(newskb));
 122         netif_rx(newskb);
 123         return 0;
 124 }
 125
 126 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
 127 {
 128         int ttl = inet->uc_ttl;
 129
 130         if (ttl < 0)
 131                 ttl = dst_metric(dst, RTAX_HOPLIMIT);
 132         return ttl;
 133 }
 134
 135 /*
 136  *              Add an ip header to a skbuff and send it out.
 137  *
 138  */
 139 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
 140                           __be32 saddr, __be32 daddr, struct ip_options *opt)
 141 {
 142         struct inet_sock *inet = inet_sk(sk);
 143         struct rtable *rt = skb_rtable(skb);
 144         struct iphdr *iph;
 145
 146         /* Build the IP header. */
 147         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 148         skb_reset_network_header(skb);
 149         iph = ip_hdr(skb);
 150         iph->version  = 4;
 151         iph->ihl      = 5;
 152         iph->tos      = inet->tos;
 153         if (ip_dont_fragment(sk, &rt->u.dst))
 154                 iph->frag_off = htons(IP_DF);
 155         else
 156                 iph->frag_off = 0;
 157         iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
 158         iph->daddr    = rt->rt_dst;
 159         iph->saddr    = rt->rt_src;
 160         iph->protocol = sk->sk_protocol;
 161         ip_select_ident(iph, &rt->u.dst, sk);
 162
 163         if (opt && opt->optlen) {
 164                 iph->ihl += opt->optlen>>2;
 165                 ip_options_build(skb, opt, daddr, rt, 0);
 166         }
 167
 168         skb->priority = sk->sk_priority;
 169         skb->mark = sk->sk_mark;
 170
 171         /* Send it out. */
 172         return ip_local_out(skb);
 173 }
 174
 175 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
 176
 177 static inline int ip_finish_output2(struct sk_buff *skb)
 178 {
 179         struct dst_entry *dst = skb_dst(skb);
 180         struct rtable *rt = (struct rtable *)dst;
 181         struct net_device *dev = dst->dev;
 182         unsigned int hh_len = LL_RESERVED_SPACE(dev);
 183
 184         if (rt->rt_type == RTN_MULTICAST) {
 185                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
 186         } else if (rt->rt_type == RTN_BROADCAST)
 187                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
 188
 189         /* Be paranoid, rather than too clever. */
 190         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 191                 struct sk_buff *skb2;
 192
 193                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 194                 if (skb2 == NULL) {
 195                         kfree_skb(skb);
 196                         return -ENOMEM;
 197                 }
 198                 if (skb->sk)
 199                         skb_set_owner_w(skb2, skb->sk);
 200                 kfree_skb(skb);
 201                 skb = skb2;
 202         }
 203
 204         if (dst->hh)
 205                 return neigh_hh_output(dst->hh, skb);
 206         else if (dst->neighbour)
 207                 return dst->neighbour->output(skb);
 208
 209         if (net_ratelimit())
 210                 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
 211         kfree_skb(skb);
 212         return -EINVAL;
 213 }
 214
 215 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
 216 {
 217         struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
 218
 219         return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
 220                skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
 221 }
 222
 223 static int ip_finish_output(struct sk_buff *skb)
 224 {
 225 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 226         /* Policy lookup after SNAT yielded a new policy */
 227         if (skb_dst(skb)->xfrm != NULL) {
 228                 IPCB(skb)->flags |= IPSKB_REROUTED;
 229                 return dst_output(skb);
 230         }
 231 #endif
 232         if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
 233                 return ip_fragment(skb, ip_finish_output2);
 234         else
 235                 return ip_finish_output2(skb);
 236 }
 237
 238 int ip_mc_output(struct sk_buff *skb)
 239 {
 240         struct sock *sk = skb->sk;
 241         struct rtable *rt = skb_rtable(skb);
 242         struct net_device *dev = rt->u.dst.dev;
 243
 244         /*
 245          *      If the indicated interface is up and running, send the packet.
 246          */
 247         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 248
 249         skb->dev = dev;
 250         skb->protocol = htons(ETH_P_IP);
 251
 252         /*
 253          *      Multicasts are looped back for other local users
 254          */
 255
 256         if (rt->rt_flags&RTCF_MULTICAST) {
 257                 if ((!sk || inet_sk(sk)->mc_loop)
 258 #ifdef CONFIG_IP_MROUTE
 259                 /* Small optimization: do not loopback not local frames,
 260                    which returned after forwarding; they will be  dropped
 261                    by ip_mr_input in any case.
 262                    Note, that local frames are looped back to be delivered
 263                    to local recipients.
 264
 265                    This check is duplicated in ip_mr_input at the moment.
 266                  */
 267                     && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
 268 #endif
 269                 ) {
 270                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 271                         if (newskb)
 272                                 NF_HOOK(PF_INET, NF_INET_POST_ROUTING, newskb,
 273                                         NULL, newskb->dev,
 274                                         ip_dev_loopback_xmit);
 275                 }
 276
 277                 /* Multicasts with ttl 0 must not go beyond the host */
 278
 279                 if (ip_hdr(skb)->ttl == 0) {
 280                         kfree_skb(skb);
 281                         return 0;
 282                 }
 283         }
 284
 285         if (rt->rt_flags&RTCF_BROADCAST) {
 286                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 287                 if (newskb)
 288                         NF_HOOK(PF_INET, NF_INET_POST_ROUTING, newskb, NULL,
 289                                 newskb->dev, ip_dev_loopback_xmit);
 290         }
 291
 292         return NF_HOOK_COND(PF_INET, NF_INET_POST_ROUTING, skb, NULL, skb->dev,
 293                             ip_finish_output,
 294                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 295 }
 296
 297 int ip_output(struct sk_buff *skb)
 298 {
 299         struct net_device *dev = skb_dst(skb)->dev;
 300
 301         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 302
 303         skb->dev = dev;
 304         skb->protocol = htons(ETH_P_IP);
 305
 306         return NF_HOOK_COND(PF_INET, NF_INET_POST_ROUTING, skb, NULL, dev,
 307                             ip_finish_output,
 308                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 309 }
 310
 311 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
 312 {
 313         struct sock *sk = skb->sk;
 314         struct inet_sock *inet = inet_sk(sk);
 315         struct ip_options *opt = inet->opt;
 316         struct rtable *rt;
 317         struct iphdr *iph;
 318
 319         /* Skip all of this if the packet is already routed,
 320          * f.e. by something like SCTP.
 321          */
 322         rt = skb_rtable(skb);
 323         if (rt != NULL)
 324                 goto packet_routed;
 325
 326         /* Make sure we can route this packet. */
 327         rt = (struct rtable *)__sk_dst_check(sk, 0);
 328         if (rt == NULL) {
 329                 __be32 daddr;
 330
 331                 /* Use correct destination address if we have options. */
 332                 daddr = inet->daddr;
 333                 if(opt && opt->srr)
 334                         daddr = opt->faddr;
 335
 336                 {
 337                         struct flowi fl = { .oif = sk->sk_bound_dev_if,
 338                                             .mark = sk->sk_mark,
 339                                             .nl_u = { .ip4_u =
 340                                                       { .daddr = daddr,
 341                                                         .saddr = inet->saddr,
 342                                                         .tos = RT_CONN_FLAGS(sk) } },
 343                                             .proto = sk->sk_protocol,
 344                                             .flags = inet_sk_flowi_flags(sk),
 345                                             .uli_u = { .ports =
 346                                                        { .sport = inet->sport,
 347                                                          .dport = inet->dport } } };
 348
 349                         /* If this fails, retransmit mechanism of transport layer will
 350                          * keep trying until route appears or the connection times
 351                          * itself out.
 352                          */
 353                         security_sk_classify_flow(sk, &fl);
 354                         if (ip_route_output_flow(sock_net(sk), &rt, &fl, sk, 0))
 355                                 goto no_route;
 356                 }
 357                 sk_setup_caps(sk, &rt->u.dst);
 358         }
 359         skb_dst_set(skb, dst_clone(&rt->u.dst));
 360
 361 packet_routed:
 362         if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
 363                 goto no_route;
 364
 365         /* OK, we know where to send it, allocate and build IP header. */
 366         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 367         skb_reset_network_header(skb);
 368         iph = ip_hdr(skb);
 369         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 370         if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
 371                 iph->frag_off = htons(IP_DF);
 372         else
 373                 iph->frag_off = 0;
 374         iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
 375         iph->protocol = sk->sk_protocol;
 376         iph->saddr    = rt->rt_src;
 377         iph->daddr    = rt->rt_dst;
 378         /* Transport layer set skb->h.foo itself. */
 379
 380         if (opt && opt->optlen) {
 381                 iph->ihl += opt->optlen >> 2;
 382                 ip_options_build(skb, opt, inet->daddr, rt, 0);
 383         }
 384
 385         ip_select_ident_more(iph, &rt->u.dst, sk,
 386                              (skb_shinfo(skb)->gso_segs ?: 1) - 1);
 387
 388         skb->priority = sk->sk_priority;
 389         skb->mark = sk->sk_mark;
 390
 391         return ip_local_out(skb);
 392
 393 no_route:
 394         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 395         kfree_skb(skb);
 396         return -EHOSTUNREACH;
 397 }
 398
 399
 400 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 401 {
 402         to->pkt_type = from->pkt_type;
 403         to->priority = from->priority;
 404         to->protocol = from->protocol;
 405         skb_dst_drop(to);
 406         skb_dst_set(to, dst_clone(skb_dst(from)));
 407         to->dev = from->dev;
 408         to->mark = from->mark;
 409
 410         /* Copy the flags to each fragment. */
 411         IPCB(to)->flags = IPCB(from)->flags;
 412
 413 #ifdef CONFIG_NET_SCHED
 414         to->tc_index = from->tc_index;
 415 #endif
 416         nf_copy(to, from);
 417 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
 418     defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
 419         to->nf_trace = from->nf_trace;
 420 #endif
 421 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 422         to->ipvs_property = from->ipvs_property;
 423 #endif
 424         skb_copy_secmark(to, from);
 425 }
 426
 427 /*
 428  *      This IP datagram is too large to be sent in one piece.  Break it up into
 429  *      smaller pieces (each of size equal to IP header plus
 430  *      a block of the data of the original IP data part) that will yet fit in a
 431  *      single device frame, and queue such a frame for sending.
 432  */
 433
 434 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 435 {
 436         struct iphdr *iph;
 437         int raw = 0;
 438         int ptr;
 439         struct net_device *dev;
 440         struct sk_buff *skb2;
 441         unsigned int mtu, hlen, left, len, ll_rs, pad;
 442         int offset;
 443         __be16 not_last_frag;
 444         struct rtable *rt = skb_rtable(skb);
 445         int err = 0;
 446
 447         dev = rt->u.dst.dev;
 448
 449         /*
 450          *      Point into the IP datagram header.
 451          */
 452
 453         iph = ip_hdr(skb);
 454
 455         if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
 456                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 457                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 458                           htonl(ip_skb_dst_mtu(skb)));
 459                 kfree_skb(skb);
 460                 return -EMSGSIZE;
 461         }
 462
 463         /*
 464          *      Setup starting values.
 465          */
 466
 467         hlen = iph->ihl * 4;
 468         mtu = dst_mtu(&rt->u.dst) - hlen;       /* Size of data space */
 469         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 470
 471         /* When frag_list is given, use it. First, check its validity:
 472          * some transformers could create wrong frag_list or break existing
 473          * one, it is not prohibited. In this case fall back to copying.
 474          *
 475          * LATER: this step can be merged to real generation of fragments,
 476          * we can switch to copy when see the first bad fragment.
 477          */
 478         if (skb_has_frags(skb)) {
 479                 struct sk_buff *frag;
 480                 int first_len = skb_pagelen(skb);
 481                 int truesizes = 0;
 482
 483                 if (first_len - hlen > mtu ||
 484                     ((first_len - hlen) & 7) ||
 485                     (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
 486                     skb_cloned(skb))
 487                         goto slow_path;
 488
 489                 skb_walk_frags(skb, frag) {
 490                         /* Correct geometry. */
 491                         if (frag->len > mtu ||
 492                             ((frag->len & 7) && frag->next) ||
 493                             skb_headroom(frag) < hlen)
 494                             goto slow_path;
 495
 496                         /* Partially cloned skb? */
 497                         if (skb_shared(frag))
 498                                 goto slow_path;
 499
 500                         BUG_ON(frag->sk);
 501                         if (skb->sk) {
 502                                 frag->sk = skb->sk;
 503                                 frag->destructor = sock_wfree;
 504                                 truesizes += frag->truesize;
 505                         }
 506                 }
 507
 508                 /* Everything is OK. Generate! */
 509
 510                 err = 0;
 511                 offset = 0;
 512                 frag = skb_shinfo(skb)->frag_list;
 513                 skb_frag_list_init(skb);
 514                 skb->data_len = first_len - skb_headlen(skb);
 515                 skb->truesize -= truesizes;
 516                 skb->len = first_len;
 517                 iph->tot_len = htons(first_len);
 518                 iph->frag_off = htons(IP_MF);
 519                 ip_send_check(iph);
 520
 521                 for (;;) {
 522                         /* Prepare header of the next frame,
 523                          * before previous one went down. */
 524                         if (frag) {
 525                                 frag->ip_summed = CHECKSUM_NONE;
 526                                 skb_reset_transport_header(frag);
 527                                 __skb_push(frag, hlen);
 528                                 skb_reset_network_header(frag);
 529                                 memcpy(skb_network_header(frag), iph, hlen);
 530                                 iph = ip_hdr(frag);
 531                                 iph->tot_len = htons(frag->len);
 532                                 ip_copy_metadata(frag, skb);
 533                                 if (offset == 0)
 534                                         ip_options_fragment(frag);
 535                                 offset += skb->len - hlen;
 536                                 iph->frag_off = htons(offset>>3);
 537                                 if (frag->next != NULL)
 538                                         iph->frag_off |= htons(IP_MF);
 539                                 /* Ready, complete checksum */
 540                                 ip_send_check(iph);
 541                         }
 542
 543                         err = output(skb);
 544
 545                         if (!err)
 546                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 547                         if (err || !frag)
 548                                 break;
 549
 550                         skb = frag;
 551                         frag = skb->next;
 552                         skb->next = NULL;
 553                 }
 554
 555                 if (err == 0) {
 556                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 557                         return 0;
 558                 }
 559
 560                 while (frag) {
 561                         skb = frag->next;
 562                         kfree_skb(frag);
 563                         frag = skb;
 564                 }
 565                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 566                 return err;
 567         }
 568
 569 slow_path:
 570         left = skb->len - hlen;         /* Space per frame */
 571         ptr = raw + hlen;               /* Where to start from */
 572
 573         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 574          * we need to make room for the encapsulating header
 575          */
 576         pad = nf_bridge_pad(skb);
 577         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, pad);
 578         mtu -= pad;
 579
 580         /*
 581          *      Fragment the datagram.
 582          */
 583
 584         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 585         not_last_frag = iph->frag_off & htons(IP_MF);
 586
 587         /*
 588          *      Keep copying data until we run out.
 589          */
 590
 591         while (left > 0) {
 592                 len = left;
 593                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 594                 if (len > mtu)
 595                         len = mtu;
 596                 /* IF: we are not sending upto and including the packet end
 597                    then align the next start on an eight byte boundary */
 598                 if (len < left) {
 599                         len &= ~7;
 600                 }
 601                 /*
 602                  *      Allocate buffer.
 603                  */
 604
 605                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 606                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 607                         err = -ENOMEM;
 608                         goto fail;
 609                 }
 610
 611                 /*
 612                  *      Set up data on packet
 613                  */
 614
 615                 ip_copy_metadata(skb2, skb);
 616                 skb_reserve(skb2, ll_rs);
 617                 skb_put(skb2, len + hlen);
 618                 skb_reset_network_header(skb2);
 619                 skb2->transport_header = skb2->network_header + hlen;
 620
 621                 /*
 622                  *      Charge the memory for the fragment to any owner
 623                  *      it might possess
 624                  */
 625
 626                 if (skb->sk)
 627                         skb_set_owner_w(skb2, skb->sk);
 628
 629                 /*
 630                  *      Copy the packet header into the new buffer.
 631                  */
 632
 633                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 634
 635                 /*
 636                  *      Copy a block of the IP datagram.
 637                  */
 638                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 639                         BUG();
 640                 left -= len;
 641
 642                 /*
 643                  *      Fill in the new header fields.
 644                  */
 645                 iph = ip_hdr(skb2);
 646                 iph->frag_off = htons((offset >> 3));
 647
 648                 /* ANK: dirty, but effective trick. Upgrade options only if
 649                  * the segment to be fragmented was THE FIRST (otherwise,
 650                  * options are already fixed) and make it ONCE
 651                  * on the initial skb, so that all the following fragments
 652                  * will inherit fixed options.
 653                  */
 654                 if (offset == 0)
 655                         ip_options_fragment(skb);
 656
 657                 /*
 658                  *      Added AC : If we are fragmenting a fragment that's not the
 659                  *                 last fragment then keep MF on each bit
 660                  */
 661                 if (left > 0 || not_last_frag)
 662                         iph->frag_off |= htons(IP_MF);
 663                 ptr += len;
 664                 offset += len;
 665
 666                 /*
 667                  *      Put this fragment into the sending queue.
 668                  */
 669                 iph->tot_len = htons(len + hlen);
 670
 671                 ip_send_check(iph);
 672
 673                 err = output(skb2);
 674                 if (err)
 675                         goto fail;
 676
 677                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 678         }
 679         kfree_skb(skb);
 680         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 681         return err;
 682
 683 fail:
 684         kfree_skb(skb);
 685         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 686         return err;
 687 }
 688
 689 EXPORT_SYMBOL(ip_fragment);
 690
 691 int
 692 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 693 {
 694         struct iovec *iov = from;
 695
 696         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 697                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 698                         return -EFAULT;
 699         } else {
 700                 __wsum csum = 0;
 701                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 702                         return -EFAULT;
 703                 skb->csum = csum_block_add(skb->csum, csum, odd);
 704         }
 705         return 0;
 706 }
 707
 708 static inline __wsum
 709 csum_page(struct page *page, int offset, int copy)
 710 {
 711         char *kaddr;
 712         __wsum csum;
 713         kaddr = kmap(page);
 714         csum = csum_partial(kaddr + offset, copy, 0);
 715         kunmap(page);
 716         return csum;
 717 }
 718
 719 static inline int ip_ufo_append_data(struct sock *sk,
 720                         int getfrag(void *from, char *to, int offset, int len,
 721                                int odd, struct sk_buff *skb),
 722                         void *from, int length, int hh_len, int fragheaderlen,
 723                         int transhdrlen, int mtu, unsigned int flags)
 724 {
 725         struct sk_buff *skb;
 726         int err;
 727
 728         /* There is support for UDP fragmentation offload by network
 729          * device, so create one single skb packet containing complete
 730          * udp datagram
 731          */
 732         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
 733                 skb = sock_alloc_send_skb(sk,
 734                         hh_len + fragheaderlen + transhdrlen + 20,
 735                         (flags & MSG_DONTWAIT), &err);
 736
 737                 if (skb == NULL)
 738                         return err;
 739
 740                 /* reserve space for Hardware header */
 741                 skb_reserve(skb, hh_len);
 742
 743                 /* create space for UDP/IP header */
 744                 skb_put(skb, fragheaderlen + transhdrlen);
 745
 746                 /* initialize network header pointer */
 747                 skb_reset_network_header(skb);
 748
 749                 /* initialize protocol header pointer */
 750                 skb->transport_header = skb->network_header + fragheaderlen;
 751
 752                 skb->ip_summed = CHECKSUM_PARTIAL;
 753                 skb->csum = 0;
 754                 sk->sk_sndmsg_off = 0;
 755
 756                 /* specify the length of each IP datagram fragment */
 757                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
 758                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 759                 __skb_queue_tail(&sk->sk_write_queue, skb);
 760         }
 761
 762         return skb_append_datato_frags(sk, skb, getfrag, from,
 763                                        (length - transhdrlen));
 764 }
 765
 766 /*
 767  *      ip_append_data() and ip_append_page() can make one large IP datagram
 768  *      from many pieces of data. Each pieces will be holded on the socket
 769  *      until ip_push_pending_frames() is called. Each piece can be a page
 770  *      or non-page data.
 771  *
 772  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
 773  *      this interface potentially.
 774  *
 775  *      LATER: length must be adjusted by pad at tail, when it is required.
 776  */
 777 int ip_append_data(struct sock *sk,
 778                    int getfrag(void *from, char *to, int offset, int len,
 779                                int odd, struct sk_buff *skb),
 780                    void *from, int length, int transhdrlen,
 781                    struct ipcm_cookie *ipc, struct rtable **rtp,
 782                    unsigned int flags)
 783 {
 784         struct inet_sock *inet = inet_sk(sk);
 785         struct sk_buff *skb;
 786
 787         struct ip_options *opt = NULL;
 788         int hh_len;
 789         int exthdrlen;
 790         int mtu;
 791         int copy;
 792         int err;
 793         int offset = 0;
 794         unsigned int maxfraglen, fragheaderlen;
 795         int csummode = CHECKSUM_NONE;
 796         struct rtable *rt;
 797
 798         if (flags&MSG_PROBE)
 799                 return 0;
 800
 801         if (skb_queue_empty(&sk->sk_write_queue)) {
 802                 /*
 803                  * setup for corking.
 804                  */
 805                 opt = ipc->opt;
 806                 if (opt) {
 807                         if (inet->cork.opt == NULL) {
 808                                 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
 809                                 if (unlikely(inet->cork.opt == NULL))
 810                                         return -ENOBUFS;
 811                         }
 812                         memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
 813                         inet->cork.flags |= IPCORK_OPT;
 814                         inet->cork.addr = ipc->addr;
 815                 }
 816                 rt = *rtp;
 817                 if (unlikely(!rt))
 818                         return -EFAULT;
 819                 /*
 820                  * We steal reference to this route, caller should not release it
 821                  */
 822                 *rtp = NULL;
 823                 inet->cork.fragsize = mtu = inet->pmtudisc == IP_PMTUDISC_PROBE ?
 824                                             rt->u.dst.dev->mtu :
 825                                             dst_mtu(rt->u.dst.path);
 826                 inet->cork.dst = &rt->u.dst;
 827                 inet->cork.length = 0;
 828                 sk->sk_sndmsg_page = NULL;
 829                 sk->sk_sndmsg_off = 0;
 830                 if ((exthdrlen = rt->u.dst.header_len) != 0) {
 831                         length += exthdrlen;
 832                         transhdrlen += exthdrlen;
 833                 }
 834         } else {
 835                 rt = (struct rtable *)inet->cork.dst;
 836                 if (inet->cork.flags & IPCORK_OPT)
 837                         opt = inet->cork.opt;
 838
 839                 transhdrlen = 0;
 840                 exthdrlen = 0;
 841                 mtu = inet->cork.fragsize;
 842         }
 843         hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
 844
 845         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 846         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 847
 848         if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
 849                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
 850                 return -EMSGSIZE;
 851         }
 852
 853         /*
 854          * transhdrlen > 0 means that this is the first fragment and we wish
 855          * it won't be fragmented in the future.
 856          */
 857         if (transhdrlen &&
 858             length + fragheaderlen <= mtu &&
 859             rt->u.dst.dev->features & NETIF_F_V4_CSUM &&
 860             !exthdrlen)
 861                 csummode = CHECKSUM_PARTIAL;
 862
 863         inet->cork.length += length;
 864         if (((length> mtu) || !skb_queue_empty(&sk->sk_write_queue)) &&
 865             (sk->sk_protocol == IPPROTO_UDP) &&
 866             (rt->u.dst.dev->features & NETIF_F_UFO)) {
 867                 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
 868                                          fragheaderlen, transhdrlen, mtu,
 869                                          flags);
 870                 if (err)
 871                         goto error;
 872                 return 0;
 873         }
 874
 875         /* So, what's going on in the loop below?
 876          *
 877          * We use calculated fragment length to generate chained skb,
 878          * each of segments is IP fragment ready for sending to network after
 879          * adding appropriate IP header.
 880          */
 881
 882         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
 883                 goto alloc_new_skb;
 884
 885         while (length > 0) {
 886                 /* Check if the remaining data fits into current packet. */
 887                 copy = mtu - skb->len;
 888                 if (copy < length)
 889                         copy = maxfraglen - skb->len;
 890                 if (copy <= 0) {
 891                         char *data;
 892                         unsigned int datalen;
 893                         unsigned int fraglen;
 894                         unsigned int fraggap;
 895                         unsigned int alloclen;
 896                         struct sk_buff *skb_prev;
 897 alloc_new_skb:
 898                         skb_prev = skb;
 899                         if (skb_prev)
 900                                 fraggap = skb_prev->len - maxfraglen;
 901                         else
 902                                 fraggap = 0;
 903
 904                         /*
 905                          * If remaining data exceeds the mtu,
 906                          * we know we need more fragment(s).
 907                          */
 908                         datalen = length + fraggap;
 909                         if (datalen > mtu - fragheaderlen)
 910                                 datalen = maxfraglen - fragheaderlen;
 911                         fraglen = datalen + fragheaderlen;
 912
 913                         if ((flags & MSG_MORE) &&
 914                             !(rt->u.dst.dev->features&NETIF_F_SG))
 915                                 alloclen = mtu;
 916                         else
 917                                 alloclen = datalen + fragheaderlen;
 918
 919                         /* The last fragment gets additional space at tail.
 920                          * Note, with MSG_MORE we overallocate on fragments,
 921                          * because we have no idea what fragment will be
 922                          * the last.
 923                          */
 924                         if (datalen == length + fraggap)
 925                                 alloclen += rt->u.dst.trailer_len;
 926
 927                         if (transhdrlen) {
 928                                 skb = sock_alloc_send_skb(sk,
 929                                                 alloclen + hh_len + 15,
 930                                                 (flags & MSG_DONTWAIT), &err);
 931                         } else {
 932                                 skb = NULL;
 933                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 934                                     2 * sk->sk_sndbuf)
 935                                         skb = sock_wmalloc(sk,
 936                                                            alloclen + hh_len + 15, 1,
 937                                                            sk->sk_allocation);
 938                                 if (unlikely(skb == NULL))
 939                                         err = -ENOBUFS;
 940                                 else
 941                                         /* only the initial fragment is
 942                                            time stamped */
 943                                         ipc->shtx.flags = 0;
 944                         }
 945                         if (skb == NULL)
 946                                 goto error;
 947
 948                         /*
 949                          *      Fill in the control structures
 950                          */
 951                         skb->ip_summed = csummode;
 952                         skb->csum = 0;
 953                         skb_reserve(skb, hh_len);
 954                         *skb_tx(skb) = ipc->shtx;
 955
 956                         /*
 957                          *      Find where to start putting bytes.
 958                          */
 959                         data = skb_put(skb, fraglen);
 960                         skb_set_network_header(skb, exthdrlen);
 961                         skb->transport_header = (skb->network_header +
 962                                                  fragheaderlen);
 963                         data += fragheaderlen;
 964
 965                         if (fraggap) {
 966                                 skb->csum = skb_copy_and_csum_bits(
 967                                         skb_prev, maxfraglen,
 968                                         data + transhdrlen, fraggap, 0);
 969                                 skb_prev->csum = csum_sub(skb_prev->csum,
 970                                                           skb->csum);
 971                                 data += fraggap;
 972                                 pskb_trim_unique(skb_prev, maxfraglen);
 973                         }
 974
 975                         copy = datalen - transhdrlen - fraggap;
 976                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
 977                                 err = -EFAULT;
 978                                 kfree_skb(skb);
 979                                 goto error;
 980                         }
 981
 982                         offset += copy;
 983                         length -= datalen - fraggap;
 984                         transhdrlen = 0;
 985                         exthdrlen = 0;
 986                         csummode = CHECKSUM_NONE;
 987
 988                         /*
 989                          * Put the packet on the pending queue.
 990                          */
 991                         __skb_queue_tail(&sk->sk_write_queue, skb);
 992                         continue;
 993                 }
 994
 995                 if (copy > length)
 996                         copy = length;
 997
 998                 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
 999                         unsigned int off;
1000
1001                         off = skb->len;
1002                         if (getfrag(from, skb_put(skb, copy),
1003                                         offset, copy, off, skb) < 0) {
1004                                 __skb_trim(skb, off);
1005                                 err = -EFAULT;
1006                                 goto error;
1007                         }
1008                 } else {
1009                         int i = skb_shinfo(skb)->nr_frags;
1010                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
1011                         struct page *page = sk->sk_sndmsg_page;
1012                         int off = sk->sk_sndmsg_off;
1013                         unsigned int left;
1014
1015                         if (page && (left = PAGE_SIZE - off) > 0) {
1016                                 if (copy >= left)
1017                                         copy = left;
1018                                 if (page != frag->page) {
1019                                         if (i == MAX_SKB_FRAGS) {
1020                                                 err = -EMSGSIZE;
1021                                                 goto error;
1022                                         }
1023                                         get_page(page);
1024                                         skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1025                                         frag = &skb_shinfo(skb)->frags[i];
1026                                 }
1027                         } else if (i < MAX_SKB_FRAGS) {
1028                                 if (copy > PAGE_SIZE)
1029                                         copy = PAGE_SIZE;
1030                                 page = alloc_pages(sk->sk_allocation, 0);
1031                                 if (page == NULL)  {
1032                                         err = -ENOMEM;
1033                                         goto error;
1034                                 }
1035                                 sk->sk_sndmsg_page = page;
1036                                 sk->sk_sndmsg_off = 0;
1037
1038                                 skb_fill_page_desc(skb, i, page, 0, 0);
1039                                 frag = &skb_shinfo(skb)->frags[i];
1040                         } else {
1041                                 err = -EMSGSIZE;
1042                                 goto error;
1043                         }
1044                         if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1045                                 err = -EFAULT;
1046                                 goto error;
1047                         }
1048                         sk->sk_sndmsg_off += copy;
1049                         frag->size += copy;
1050                         skb->len += copy;
1051                         skb->data_len += copy;
1052                         skb->truesize += copy;
1053                         atomic_add(copy, &sk->sk_wmem_alloc);
1054                 }
1055                 offset += copy;
1056                 length -= copy;
1057         }
1058
1059         return 0;
1060
1061 error:
1062         inet->cork.length -= length;
1063         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1064         return err;
1065 }
1066
1067 ssize_t ip_append_page(struct sock *sk, struct page *page,
1068                        int offset, size_t size, int flags)
1069 {
1070         struct inet_sock *inet = inet_sk(sk);
1071         struct sk_buff *skb;
1072         struct rtable *rt;
1073         struct ip_options *opt = NULL;
1074         int hh_len;
1075         int mtu;
1076         int len;
1077         int err;
1078         unsigned int maxfraglen, fragheaderlen, fraggap;
1079
1080         if (inet->hdrincl)
1081                 return -EPERM;
1082
1083         if (flags&MSG_PROBE)
1084                 return 0;
1085
1086         if (skb_queue_empty(&sk->sk_write_queue))
1087                 return -EINVAL;
1088
1089         rt = (struct rtable *)inet->cork.dst;
1090         if (inet->cork.flags & IPCORK_OPT)
1091                 opt = inet->cork.opt;
1092
1093         if (!(rt->u.dst.dev->features&NETIF_F_SG))
1094                 return -EOPNOTSUPP;
1095
1096         hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1097         mtu = inet->cork.fragsize;
1098
1099         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1100         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1101
1102         if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1103                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1104                 return -EMSGSIZE;
1105         }
1106
1107         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1108                 return -EINVAL;
1109
1110         inet->cork.length += size;
1111         if ((sk->sk_protocol == IPPROTO_UDP) &&
1112             (rt->u.dst.dev->features & NETIF_F_UFO)) {
1113                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1114                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1115         }
1116
1117
1118         while (size > 0) {
1119                 int i;
1120
1121                 if (skb_is_gso(skb))
1122                         len = size;
1123                 else {
1124
1125                         /* Check if the remaining data fits into current packet. */
1126                         len = mtu - skb->len;
1127                         if (len < size)
1128                                 len = maxfraglen - skb->len;
1129                 }
1130                 if (len <= 0) {
1131                         struct sk_buff *skb_prev;
1132                         int alloclen;
1133
1134                         skb_prev = skb;
1135                         fraggap = skb_prev->len - maxfraglen;
1136
1137                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1138                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1139                         if (unlikely(!skb)) {
1140                                 err = -ENOBUFS;
1141                                 goto error;
1142                         }
1143
1144                         /*
1145                          *      Fill in the control structures
1146                          */
1147                         skb->ip_summed = CHECKSUM_NONE;
1148                         skb->csum = 0;
1149                         skb_reserve(skb, hh_len);
1150
1151                         /*
1152                          *      Find where to start putting bytes.
1153                          */
1154                         skb_put(skb, fragheaderlen + fraggap);
1155                         skb_reset_network_header(skb);
1156                         skb->transport_header = (skb->network_header +
1157                                                  fragheaderlen);
1158                         if (fraggap) {
1159                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1160                                                                    maxfraglen,
1161                                                     skb_transport_header(skb),
1162                                                                    fraggap, 0);
1163                                 skb_prev->csum = csum_sub(skb_prev->csum,
1164                                                           skb->csum);
1165                                 pskb_trim_unique(skb_prev, maxfraglen);
1166                         }
1167
1168                         /*
1169                          * Put the packet on the pending queue.
1170                          */
1171                         __skb_queue_tail(&sk->sk_write_queue, skb);
1172                         continue;
1173                 }
1174
1175                 i = skb_shinfo(skb)->nr_frags;
1176                 if (len > size)
1177                         len = size;
1178                 if (skb_can_coalesce(skb, i, page, offset)) {
1179                         skb_shinfo(skb)->frags[i-1].size += len;
1180                 } else if (i < MAX_SKB_FRAGS) {
1181                         get_page(page);
1182                         skb_fill_page_desc(skb, i, page, offset, len);
1183                 } else {
1184                         err = -EMSGSIZE;
1185                         goto error;
1186                 }
1187
1188                 if (skb->ip_summed == CHECKSUM_NONE) {
1189                         __wsum csum;
1190                         csum = csum_page(page, offset, len);
1191                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1192                 }
1193
1194                 skb->len += len;
1195                 skb->data_len += len;
1196                 skb->truesize += len;
1197                 atomic_add(len, &sk->sk_wmem_alloc);
1198                 offset += len;
1199                 size -= len;
1200         }
1201         return 0;
1202
1203 error:
1204         inet->cork.length -= size;
1205         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1206         return err;
1207 }
1208
1209 static void ip_cork_release(struct inet_sock *inet)
1210 {
1211         inet->cork.flags &= ~IPCORK_OPT;
1212         kfree(inet->cork.opt);
1213         inet->cork.opt = NULL;
1214         dst_release(inet->cork.dst);
1215         inet->cork.dst = NULL;
1216 }
1217
1218 /*
1219  *      Combined all pending IP fragments on the socket as one IP datagram
1220  *      and push them out.
1221  */
1222 int ip_push_pending_frames(struct sock *sk)
1223 {
1224         struct sk_buff *skb, *tmp_skb;
1225         struct sk_buff **tail_skb;
1226         struct inet_sock *inet = inet_sk(sk);
1227         struct net *net = sock_net(sk);
1228         struct ip_options *opt = NULL;
1229         struct rtable *rt = (struct rtable *)inet->cork.dst;
1230         struct iphdr *iph;
1231         __be16 df = 0;
1232         __u8 ttl;
1233         int err = 0;
1234
1235         if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1236                 goto out;
1237         tail_skb = &(skb_shinfo(skb)->frag_list);
1238
1239         /* move skb->data to ip header from ext header */
1240         if (skb->data < skb_network_header(skb))
1241                 __skb_pull(skb, skb_network_offset(skb));
1242         while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1243                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1244                 *tail_skb = tmp_skb;
1245                 tail_skb = &(tmp_skb->next);
1246                 skb->len += tmp_skb->len;
1247                 skb->data_len += tmp_skb->len;
1248                 skb->truesize += tmp_skb->truesize;
1249                 tmp_skb->destructor = NULL;
1250                 tmp_skb->sk = NULL;
1251         }
1252
1253         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1254          * to fragment the frame generated here. No matter, what transforms
1255          * how transforms change size of the packet, it will come out.
1256          */
1257         if (inet->pmtudisc < IP_PMTUDISC_DO)
1258                 skb->local_df = 1;
1259
1260         /* DF bit is set when we want to see DF on outgoing frames.
1261          * If local_df is set too, we still allow to fragment this frame
1262          * locally. */
1263         if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1264             (skb->len <= dst_mtu(&rt->u.dst) &&
1265              ip_dont_fragment(sk, &rt->u.dst)))
1266                 df = htons(IP_DF);
1267
1268         if (inet->cork.flags & IPCORK_OPT)
1269                 opt = inet->cork.opt;
1270
1271         if (rt->rt_type == RTN_MULTICAST)
1272                 ttl = inet->mc_ttl;
1273         else
1274                 ttl = ip_select_ttl(inet, &rt->u.dst);
1275
1276         iph = (struct iphdr *)skb->data;
1277         iph->version = 4;
1278         iph->ihl = 5;
1279         if (opt) {
1280                 iph->ihl += opt->optlen>>2;
1281                 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1282         }
1283         iph->tos = inet->tos;
1284         iph->frag_off = df;
1285         ip_select_ident(iph, &rt->u.dst, sk);
1286         iph->ttl = ttl;
1287         iph->protocol = sk->sk_protocol;
1288         iph->saddr = rt->rt_src;
1289         iph->daddr = rt->rt_dst;
1290
1291         skb->priority = sk->sk_priority;
1292         skb->mark = sk->sk_mark;
1293         /*
1294          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1295          * on dst refcount
1296          */
1297         inet->cork.dst = NULL;
1298         skb_dst_set(skb, &rt->u.dst);
1299
1300         if (iph->protocol == IPPROTO_ICMP)
1301                 icmp_out_count(net, ((struct icmphdr *)
1302                         skb_transport_header(skb))->type);
1303
1304         /* Netfilter gets whole the not fragmented skb. */
1305         err = ip_local_out(skb);
1306         if (err) {
1307                 if (err > 0)
1308                         err = net_xmit_errno(err);
1309                 if (err)
1310                         goto error;
1311         }
1312
1313 out:
1314         ip_cork_release(inet);
1315         return err;
1316
1317 error:
1318         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1319         goto out;
1320 }
1321
1322 /*
1323  *      Throw away all pending data on the socket.
1324  */
1325 void ip_flush_pending_frames(struct sock *sk)
1326 {
1327         struct sk_buff *skb;
1328
1329         while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1330                 kfree_skb(skb);
1331
1332         ip_cork_release(inet_sk(sk));
1333 }
1334
1335
1336 /*
1337  *      Fetch data from kernel space and fill in checksum if needed.
1338  */
1339 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1340                               int len, int odd, struct sk_buff *skb)
1341 {
1342         __wsum csum;
1343
1344         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1345         skb->csum = csum_block_add(skb->csum, csum, odd);
1346         return 0;
1347 }
1348
1349 /*
1350  *      Generic function to send a packet as reply to another packet.
1351  *      Used to send TCP resets so far. ICMP should use this function too.
1352  *
1353  *      Should run single threaded per socket because it uses the sock
1354  *      structure to pass arguments.
1355  */
1356 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1357                    unsigned int len)
1358 {
1359         struct inet_sock *inet = inet_sk(sk);
1360         struct {
1361                 struct ip_options       opt;
1362                 char                    data[40];
1363         } replyopts;
1364         struct ipcm_cookie ipc;
1365         __be32 daddr;
1366         struct rtable *rt = skb_rtable(skb);
1367
1368         if (ip_options_echo(&replyopts.opt, skb))
1369                 return;
1370
1371         daddr = ipc.addr = rt->rt_src;
1372         ipc.opt = NULL;
1373         ipc.shtx.flags = 0;
1374
1375         if (replyopts.opt.optlen) {
1376                 ipc.opt = &replyopts.opt;
1377
1378                 if (ipc.opt->srr)
1379                         daddr = replyopts.opt.faddr;
1380         }
1381
1382         {
1383                 struct flowi fl = { .oif = arg->bound_dev_if,
1384                                     .nl_u = { .ip4_u =
1385                                               { .daddr = daddr,
1386                                                 .saddr = rt->rt_spec_dst,
1387                                                 .tos = RT_TOS(ip_hdr(skb)->tos) } },
1388                                     /* Not quite clean, but right. */
1389                                     .uli_u = { .ports =
1390                                                { .sport = tcp_hdr(skb)->dest,
1391                                                  .dport = tcp_hdr(skb)->source } },
1392                                     .proto = sk->sk_protocol,
1393                                     .flags = ip_reply_arg_flowi_flags(arg) };
1394                 security_skb_classify_flow(skb, &fl);
1395                 if (ip_route_output_key(sock_net(sk), &rt, &fl))
1396                         return;
1397         }
1398
1399         /* And let IP do all the hard work.
1400
1401            This chunk is not reenterable, hence spinlock.
1402            Note that it uses the fact, that this function is called
1403            with locally disabled BH and that sk cannot be already spinlocked.
1404          */
1405         bh_lock_sock(sk);
1406         inet->tos = ip_hdr(skb)->tos;
1407         sk->sk_priority = skb->priority;
1408         sk->sk_protocol = ip_hdr(skb)->protocol;
1409         sk->sk_bound_dev_if = arg->bound_dev_if;
1410         ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1411                        &ipc, &rt, MSG_DONTWAIT);
1412         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1413                 if (arg->csumoffset >= 0)
1414                         *((__sum16 *)skb_transport_header(skb) +
1415                           arg->csumoffset) = csum_fold(csum_add(skb->csum,
1416                                                                 arg->csum));
1417                 skb->ip_summed = CHECKSUM_NONE;
1418                 ip_push_pending_frames(sk);
1419         }
1420
1421         bh_unlock_sock(sk);
1422
1423         ip_rt_put(rt);
1424 }
1425
1426 void __init ip_init(void)
1427 {
1428         ip_rt_init();
1429         inet_initpeers();
1430
1431 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1432         igmp_mc_proc_init();
1433 #endif
1434 }
1435
1436 EXPORT_SYMBOL(ip_generic_getfrag);
1437 EXPORT_SYMBOL(ip_queue_xmit);
1438 EXPORT_SYMBOL(ip_send_check);