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 #include <linux/slab.h>
  55
  56 #include <linux/socket.h>
  57 #include <linux/sockios.h>
  58 #include <linux/in.h>
  59 #include <linux/inet.h>
  60 #include <linux/netdevice.h>
  61 #include <linux/etherdevice.h>
  62 #include <linux/proc_fs.h>
  63 #include <linux/stat.h>
  64 #include <linux/init.h>
  65
  66 #include <net/snmp.h>
  67 #include <net/ip.h>
  68 #include <net/protocol.h>
  69 #include <net/route.h>
  70 #include <net/xfrm.h>
  71 #include <linux/skbuff.h>
  72 #include <net/sock.h>
  73 #include <net/arp.h>
  74 #include <net/icmp.h>
  75 #include <net/checksum.h>
  76 #include <net/inetpeer.h>
  77 #include <linux/igmp.h>
  78 #include <linux/netfilter_ipv4.h>
  79 #include <linux/netfilter_bridge.h>
  80 #include <linux/mroute.h>
  81 #include <linux/netlink.h>
  82 #include <linux/tcp.h>
  83
  84 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
  85 EXPORT_SYMBOL(sysctl_ip_default_ttl);
  86
  87 /* Generate a checksum for an outgoing IP datagram. */
  88 __inline__ void ip_send_check(struct iphdr *iph)
  89 {
  90         iph->check = 0;
  91         iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
  92 }
  93 EXPORT_SYMBOL(ip_send_check);
  94
  95 int __ip_local_out(struct sk_buff *skb)
  96 {
  97         struct iphdr *iph = ip_hdr(skb);
  98
  99         iph->tot_len = htons(skb->len);
 100         ip_send_check(iph);
 101         return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
 102                        skb_dst(skb)->dev, dst_output);
 103 }
 104
 105 int ip_local_out(struct sk_buff *skb)
 106 {
 107         int err;
 108
 109         err = __ip_local_out(skb);
 110         if (likely(err == 1))
 111                 err = dst_output(skb);
 112
 113         return err;
 114 }
 115 EXPORT_SYMBOL_GPL(ip_local_out);
 116
 117 /* dev_loopback_xmit for use with netfilter. */
 118 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
 119 {
 120         skb_reset_mac_header(newskb);
 121         __skb_pull(newskb, skb_network_offset(newskb));
 122         newskb->pkt_type = PACKET_LOOPBACK;
 123         newskb->ip_summed = CHECKSUM_UNNECESSARY;
 124         WARN_ON(!skb_dst(newskb));
 125         netif_rx_ni(newskb);
 126         return 0;
 127 }
 128
 129 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
 130 {
 131         int ttl = inet->uc_ttl;
 132
 133         if (ttl < 0)
 134                 ttl = ip4_dst_hoplimit(dst);
 135         return ttl;
 136 }
 137
 138 /*
 139  *              Add an ip header to a skbuff and send it out.
 140  *
 141  */
 142 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
 143                           __be32 saddr, __be32 daddr, struct ip_options *opt)
 144 {
 145         struct inet_sock *inet = inet_sk(sk);
 146         struct rtable *rt = skb_rtable(skb);
 147         struct iphdr *iph;
 148
 149         /* Build the IP header. */
 150         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 151         skb_reset_network_header(skb);
 152         iph = ip_hdr(skb);
 153         iph->version  = 4;
 154         iph->ihl      = 5;
 155         iph->tos      = inet->tos;
 156         if (ip_dont_fragment(sk, &rt->dst))
 157                 iph->frag_off = htons(IP_DF);
 158         else
 159                 iph->frag_off = 0;
 160         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 161         iph->daddr    = rt->rt_dst;
 162         iph->saddr    = rt->rt_src;
 163         iph->protocol = sk->sk_protocol;
 164         ip_select_ident(iph, &rt->dst, sk);
 165
 166         if (opt && opt->optlen) {
 167                 iph->ihl += opt->optlen>>2;
 168                 ip_options_build(skb, opt, daddr, rt, 0);
 169         }
 170
 171         skb->priority = sk->sk_priority;
 172         skb->mark = sk->sk_mark;
 173
 174         /* Send it out. */
 175         return ip_local_out(skb);
 176 }
 177 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
 178
 179 static inline int ip_finish_output2(struct sk_buff *skb)
 180 {
 181         struct dst_entry *dst = skb_dst(skb);
 182         struct rtable *rt = (struct rtable *)dst;
 183         struct net_device *dev = dst->dev;
 184         unsigned int hh_len = LL_RESERVED_SPACE(dev);
 185
 186         if (rt->rt_type == RTN_MULTICAST) {
 187                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
 188         } else if (rt->rt_type == RTN_BROADCAST)
 189                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
 190
 191         /* Be paranoid, rather than too clever. */
 192         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 193                 struct sk_buff *skb2;
 194
 195                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 196                 if (skb2 == NULL) {
 197                         kfree_skb(skb);
 198                         return -ENOMEM;
 199                 }
 200                 if (skb->sk)
 201                         skb_set_owner_w(skb2, skb->sk);
 202                 kfree_skb(skb);
 203                 skb = skb2;
 204         }
 205
 206         if (dst->hh)
 207                 return neigh_hh_output(dst->hh, skb);
 208         else if (dst->neighbour)
 209                 return dst->neighbour->output(skb);
 210
 211         if (net_ratelimit())
 212                 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
 213         kfree_skb(skb);
 214         return -EINVAL;
 215 }
 216
 217 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
 218 {
 219         struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
 220
 221         return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
 222                skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
 223 }
 224
 225 static int ip_finish_output(struct sk_buff *skb)
 226 {
 227 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 228         /* Policy lookup after SNAT yielded a new policy */
 229         if (skb_dst(skb)->xfrm != NULL) {
 230                 IPCB(skb)->flags |= IPSKB_REROUTED;
 231                 return dst_output(skb);
 232         }
 233 #endif
 234         if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
 235                 return ip_fragment(skb, ip_finish_output2);
 236         else
 237                 return ip_finish_output2(skb);
 238 }
 239
 240 int ip_mc_output(struct sk_buff *skb)
 241 {
 242         struct sock *sk = skb->sk;
 243         struct rtable *rt = skb_rtable(skb);
 244         struct net_device *dev = rt->dst.dev;
 245
 246         /*
 247          *      If the indicated interface is up and running, send the packet.
 248          */
 249         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 250
 251         skb->dev = dev;
 252         skb->protocol = htons(ETH_P_IP);
 253
 254         /*
 255          *      Multicasts are looped back for other local users
 256          */
 257
 258         if (rt->rt_flags&RTCF_MULTICAST) {
 259                 if (sk_mc_loop(sk)
 260 #ifdef CONFIG_IP_MROUTE
 261                 /* Small optimization: do not loopback not local frames,
 262                    which returned after forwarding; they will be  dropped
 263                    by ip_mr_input in any case.
 264                    Note, that local frames are looped back to be delivered
 265                    to local recipients.
 266
 267                    This check is duplicated in ip_mr_input at the moment.
 268                  */
 269                     &&
 270                     ((rt->rt_flags & RTCF_LOCAL) ||
 271                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 272 #endif
 273                    ) {
 274                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 275                         if (newskb)
 276                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 277                                         newskb, NULL, newskb->dev,
 278                                         ip_dev_loopback_xmit);
 279                 }
 280
 281                 /* Multicasts with ttl 0 must not go beyond the host */
 282
 283                 if (ip_hdr(skb)->ttl == 0) {
 284                         kfree_skb(skb);
 285                         return 0;
 286                 }
 287         }
 288
 289         if (rt->rt_flags&RTCF_BROADCAST) {
 290                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 291                 if (newskb)
 292                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
 293                                 NULL, newskb->dev, ip_dev_loopback_xmit);
 294         }
 295
 296         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
 297                             skb->dev, ip_finish_output,
 298                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 299 }
 300
 301 int ip_output(struct sk_buff *skb)
 302 {
 303         struct net_device *dev = skb_dst(skb)->dev;
 304
 305         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 306
 307         skb->dev = dev;
 308         skb->protocol = htons(ETH_P_IP);
 309
 310         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
 311                             ip_finish_output,
 312                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 313 }
 314
 315 int ip_queue_xmit(struct sk_buff *skb)
 316 {
 317         struct sock *sk = skb->sk;
 318         struct inet_sock *inet = inet_sk(sk);
 319         struct ip_options *opt = inet->opt;
 320         struct rtable *rt;
 321         struct iphdr *iph;
 322         int res;
 323
 324         /* Skip all of this if the packet is already routed,
 325          * f.e. by something like SCTP.
 326          */
 327         rcu_read_lock();
 328         rt = skb_rtable(skb);
 329         if (rt != NULL)
 330                 goto packet_routed;
 331
 332         /* Make sure we can route this packet. */
 333         rt = (struct rtable *)__sk_dst_check(sk, 0);
 334         if (rt == NULL) {
 335                 __be32 daddr;
 336
 337                 /* Use correct destination address if we have options. */
 338                 daddr = inet->inet_daddr;
 339                 if(opt && opt->srr)
 340                         daddr = opt->faddr;
 341
 342                 {
 343                         struct flowi fl = { .oif = sk->sk_bound_dev_if,
 344                                             .mark = sk->sk_mark,
 345                                             .fl4_dst = daddr,
 346                                             .fl4_src = inet->inet_saddr,
 347                                             .fl4_tos = RT_CONN_FLAGS(sk),
 348                                             .proto = sk->sk_protocol,
 349                                             .flags = inet_sk_flowi_flags(sk),
 350                                             .fl_ip_sport = inet->inet_sport,
 351                                             .fl_ip_dport = inet->inet_dport };
 352
 353                         /* If this fails, retransmit mechanism of transport layer will
 354                          * keep trying until route appears or the connection times
 355                          * itself out.
 356                          */
 357                         security_sk_classify_flow(sk, &fl);
 358                         if (ip_route_output_flow(sock_net(sk), &rt, &fl, sk, 0))
 359                                 goto no_route;
 360                 }
 361                 sk_setup_caps(sk, &rt->dst);
 362         }
 363         skb_dst_set_noref(skb, &rt->dst);
 364
 365 packet_routed:
 366         if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
 367                 goto no_route;
 368
 369         /* OK, we know where to send it, allocate and build IP header. */
 370         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 371         skb_reset_network_header(skb);
 372         iph = ip_hdr(skb);
 373         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 374         if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
 375                 iph->frag_off = htons(IP_DF);
 376         else
 377                 iph->frag_off = 0;
 378         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 379         iph->protocol = sk->sk_protocol;
 380         iph->saddr    = rt->rt_src;
 381         iph->daddr    = rt->rt_dst;
 382         /* Transport layer set skb->h.foo itself. */
 383
 384         if (opt && opt->optlen) {
 385                 iph->ihl += opt->optlen >> 2;
 386                 ip_options_build(skb, opt, inet->inet_daddr, rt, 0);
 387         }
 388
 389         ip_select_ident_more(iph, &rt->dst, sk,
 390                              (skb_shinfo(skb)->gso_segs ?: 1) - 1);
 391
 392         skb->priority = sk->sk_priority;
 393         skb->mark = sk->sk_mark;
 394
 395         res = ip_local_out(skb);
 396         rcu_read_unlock();
 397         return res;
 398
 399 no_route:
 400         rcu_read_unlock();
 401         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 402         kfree_skb(skb);
 403         return -EHOSTUNREACH;
 404 }
 405 EXPORT_SYMBOL(ip_queue_xmit);
 406
 407
 408 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 409 {
 410         to->pkt_type = from->pkt_type;
 411         to->priority = from->priority;
 412         to->protocol = from->protocol;
 413         skb_dst_drop(to);
 414         skb_dst_copy(to, from);
 415         to->dev = from->dev;
 416         to->mark = from->mark;
 417
 418         /* Copy the flags to each fragment. */
 419         IPCB(to)->flags = IPCB(from)->flags;
 420
 421 #ifdef CONFIG_NET_SCHED
 422         to->tc_index = from->tc_index;
 423 #endif
 424         nf_copy(to, from);
 425 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
 426     defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
 427         to->nf_trace = from->nf_trace;
 428 #endif
 429 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 430         to->ipvs_property = from->ipvs_property;
 431 #endif
 432         skb_copy_secmark(to, from);
 433 }
 434
 435 /*
 436  *      This IP datagram is too large to be sent in one piece.  Break it up into
 437  *      smaller pieces (each of size equal to IP header plus
 438  *      a block of the data of the original IP data part) that will yet fit in a
 439  *      single device frame, and queue such a frame for sending.
 440  */
 441
 442 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 443 {
 444         struct iphdr *iph;
 445         int ptr;
 446         struct net_device *dev;
 447         struct sk_buff *skb2;
 448         unsigned int mtu, hlen, left, len, ll_rs;
 449         int offset;
 450         __be16 not_last_frag;
 451         struct rtable *rt = skb_rtable(skb);
 452         int err = 0;
 453
 454         dev = rt->dst.dev;
 455
 456         /*
 457          *      Point into the IP datagram header.
 458          */
 459
 460         iph = ip_hdr(skb);
 461
 462         if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
 463                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 464                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 465                           htonl(ip_skb_dst_mtu(skb)));
 466                 kfree_skb(skb);
 467                 return -EMSGSIZE;
 468         }
 469
 470         /*
 471          *      Setup starting values.
 472          */
 473
 474         hlen = iph->ihl * 4;
 475         mtu = dst_mtu(&rt->dst) - hlen; /* Size of data space */
 476 #ifdef CONFIG_BRIDGE_NETFILTER
 477         if (skb->nf_bridge)
 478                 mtu -= nf_bridge_mtu_reduction(skb);
 479 #endif
 480         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 481
 482         /* When frag_list is given, use it. First, check its validity:
 483          * some transformers could create wrong frag_list or break existing
 484          * one, it is not prohibited. In this case fall back to copying.
 485          *
 486          * LATER: this step can be merged to real generation of fragments,
 487          * we can switch to copy when see the first bad fragment.
 488          */
 489         if (skb_has_frag_list(skb)) {
 490                 struct sk_buff *frag, *frag2;
 491                 int first_len = skb_pagelen(skb);
 492
 493                 if (first_len - hlen > mtu ||
 494                     ((first_len - hlen) & 7) ||
 495                     (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
 496                     skb_cloned(skb))
 497                         goto slow_path;
 498
 499                 skb_walk_frags(skb, frag) {
 500                         /* Correct geometry. */
 501                         if (frag->len > mtu ||
 502                             ((frag->len & 7) && frag->next) ||
 503                             skb_headroom(frag) < hlen)
 504                                 goto slow_path_clean;
 505
 506                         /* Partially cloned skb? */
 507                         if (skb_shared(frag))
 508                                 goto slow_path_clean;
 509
 510                         BUG_ON(frag->sk);
 511                         if (skb->sk) {
 512                                 frag->sk = skb->sk;
 513                                 frag->destructor = sock_wfree;
 514                         }
 515                         skb->truesize -= frag->truesize;
 516                 }
 517
 518                 /* Everything is OK. Generate! */
 519
 520                 err = 0;
 521                 offset = 0;
 522                 frag = skb_shinfo(skb)->frag_list;
 523                 skb_frag_list_init(skb);
 524                 skb->data_len = first_len - skb_headlen(skb);
 525                 skb->len = first_len;
 526                 iph->tot_len = htons(first_len);
 527                 iph->frag_off = htons(IP_MF);
 528                 ip_send_check(iph);
 529
 530                 for (;;) {
 531                         /* Prepare header of the next frame,
 532                          * before previous one went down. */
 533                         if (frag) {
 534                                 frag->ip_summed = CHECKSUM_NONE;
 535                                 skb_reset_transport_header(frag);
 536                                 __skb_push(frag, hlen);
 537                                 skb_reset_network_header(frag);
 538                                 memcpy(skb_network_header(frag), iph, hlen);
 539                                 iph = ip_hdr(frag);
 540                                 iph->tot_len = htons(frag->len);
 541                                 ip_copy_metadata(frag, skb);
 542                                 if (offset == 0)
 543                                         ip_options_fragment(frag);
 544                                 offset += skb->len - hlen;
 545                                 iph->frag_off = htons(offset>>3);
 546                                 if (frag->next != NULL)
 547                                         iph->frag_off |= htons(IP_MF);
 548                                 /* Ready, complete checksum */
 549                                 ip_send_check(iph);
 550                         }
 551
 552                         err = output(skb);
 553
 554                         if (!err)
 555                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 556                         if (err || !frag)
 557                                 break;
 558
 559                         skb = frag;
 560                         frag = skb->next;
 561                         skb->next = NULL;
 562                 }
 563
 564                 if (err == 0) {
 565                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 566                         return 0;
 567                 }
 568
 569                 while (frag) {
 570                         skb = frag->next;
 571                         kfree_skb(frag);
 572                         frag = skb;
 573                 }
 574                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 575                 return err;
 576
 577 slow_path_clean:
 578                 skb_walk_frags(skb, frag2) {
 579                         if (frag2 == frag)
 580                                 break;
 581                         frag2->sk = NULL;
 582                         frag2->destructor = NULL;
 583                         skb->truesize += frag2->truesize;
 584                 }
 585         }
 586
 587 slow_path:
 588         left = skb->len - hlen;         /* Space per frame */
 589         ptr = hlen;             /* Where to start from */
 590
 591         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 592          * we need to make room for the encapsulating header
 593          */
 594         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
 595
 596         /*
 597          *      Fragment the datagram.
 598          */
 599
 600         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 601         not_last_frag = iph->frag_off & htons(IP_MF);
 602
 603         /*
 604          *      Keep copying data until we run out.
 605          */
 606
 607         while (left > 0) {
 608                 len = left;
 609                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 610                 if (len > mtu)
 611                         len = mtu;
 612                 /* IF: we are not sending upto and including the packet end
 613                    then align the next start on an eight byte boundary */
 614                 if (len < left) {
 615                         len &= ~7;
 616                 }
 617                 /*
 618                  *      Allocate buffer.
 619                  */
 620
 621                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 622                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 623                         err = -ENOMEM;
 624                         goto fail;
 625                 }
 626
 627                 /*
 628                  *      Set up data on packet
 629                  */
 630
 631                 ip_copy_metadata(skb2, skb);
 632                 skb_reserve(skb2, ll_rs);
 633                 skb_put(skb2, len + hlen);
 634                 skb_reset_network_header(skb2);
 635                 skb2->transport_header = skb2->network_header + hlen;
 636
 637                 /*
 638                  *      Charge the memory for the fragment to any owner
 639                  *      it might possess
 640                  */
 641
 642                 if (skb->sk)
 643                         skb_set_owner_w(skb2, skb->sk);
 644
 645                 /*
 646                  *      Copy the packet header into the new buffer.
 647                  */
 648
 649                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 650
 651                 /*
 652                  *      Copy a block of the IP datagram.
 653                  */
 654                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 655                         BUG();
 656                 left -= len;
 657
 658                 /*
 659                  *      Fill in the new header fields.
 660                  */
 661                 iph = ip_hdr(skb2);
 662                 iph->frag_off = htons((offset >> 3));
 663
 664                 /* ANK: dirty, but effective trick. Upgrade options only if
 665                  * the segment to be fragmented was THE FIRST (otherwise,
 666                  * options are already fixed) and make it ONCE
 667                  * on the initial skb, so that all the following fragments
 668                  * will inherit fixed options.
 669                  */
 670                 if (offset == 0)
 671                         ip_options_fragment(skb);
 672
 673                 /*
 674                  *      Added AC : If we are fragmenting a fragment that's not the
 675                  *                 last fragment then keep MF on each bit
 676                  */
 677                 if (left > 0 || not_last_frag)
 678                         iph->frag_off |= htons(IP_MF);
 679                 ptr += len;
 680                 offset += len;
 681
 682                 /*
 683                  *      Put this fragment into the sending queue.
 684                  */
 685                 iph->tot_len = htons(len + hlen);
 686
 687                 ip_send_check(iph);
 688
 689                 err = output(skb2);
 690                 if (err)
 691                         goto fail;
 692
 693                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 694         }
 695         kfree_skb(skb);
 696         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 697         return err;
 698
 699 fail:
 700         kfree_skb(skb);
 701         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 702         return err;
 703 }
 704 EXPORT_SYMBOL(ip_fragment);
 705
 706 int
 707 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 708 {
 709         struct iovec *iov = from;
 710
 711         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 712                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 713                         return -EFAULT;
 714         } else {
 715                 __wsum csum = 0;
 716                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 717                         return -EFAULT;
 718                 skb->csum = csum_block_add(skb->csum, csum, odd);
 719         }
 720         return 0;
 721 }
 722 EXPORT_SYMBOL(ip_generic_getfrag);
 723
 724 static inline __wsum
 725 csum_page(struct page *page, int offset, int copy)
 726 {
 727         char *kaddr;
 728         __wsum csum;
 729         kaddr = kmap(page);
 730         csum = csum_partial(kaddr + offset, copy, 0);
 731         kunmap(page);
 732         return csum;
 733 }
 734
 735 static inline int ip_ufo_append_data(struct sock *sk,
 736                         int getfrag(void *from, char *to, int offset, int len,
 737                                int odd, struct sk_buff *skb),
 738                         void *from, int length, int hh_len, int fragheaderlen,
 739                         int transhdrlen, int mtu, unsigned int flags)
 740 {
 741         struct sk_buff *skb;
 742         int err;
 743
 744         /* There is support for UDP fragmentation offload by network
 745          * device, so create one single skb packet containing complete
 746          * udp datagram
 747          */
 748         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
 749                 skb = sock_alloc_send_skb(sk,
 750                         hh_len + fragheaderlen + transhdrlen + 20,
 751                         (flags & MSG_DONTWAIT), &err);
 752
 753                 if (skb == NULL)
 754                         return err;
 755
 756                 /* reserve space for Hardware header */
 757                 skb_reserve(skb, hh_len);
 758
 759                 /* create space for UDP/IP header */
 760                 skb_put(skb, fragheaderlen + transhdrlen);
 761
 762                 /* initialize network header pointer */
 763                 skb_reset_network_header(skb);
 764
 765                 /* initialize protocol header pointer */
 766                 skb->transport_header = skb->network_header + fragheaderlen;
 767
 768                 skb->ip_summed = CHECKSUM_PARTIAL;
 769                 skb->csum = 0;
 770                 sk->sk_sndmsg_off = 0;
 771
 772                 /* specify the length of each IP datagram fragment */
 773                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
 774                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 775                 __skb_queue_tail(&sk->sk_write_queue, skb);
 776         }
 777
 778         return skb_append_datato_frags(sk, skb, getfrag, from,
 779                                        (length - transhdrlen));
 780 }
 781
 782 /*
 783  *      ip_append_data() and ip_append_page() can make one large IP datagram
 784  *      from many pieces of data. Each pieces will be holded on the socket
 785  *      until ip_push_pending_frames() is called. Each piece can be a page
 786  *      or non-page data.
 787  *
 788  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
 789  *      this interface potentially.
 790  *
 791  *      LATER: length must be adjusted by pad at tail, when it is required.
 792  */
 793 int ip_append_data(struct sock *sk,
 794                    int getfrag(void *from, char *to, int offset, int len,
 795                                int odd, struct sk_buff *skb),
 796                    void *from, int length, int transhdrlen,
 797                    struct ipcm_cookie *ipc, struct rtable **rtp,
 798                    unsigned int flags)
 799 {
 800         struct inet_sock *inet = inet_sk(sk);
 801         struct sk_buff *skb;
 802
 803         struct ip_options *opt = NULL;
 804         int hh_len;
 805         int exthdrlen;
 806         int mtu;
 807         int copy;
 808         int err;
 809         int offset = 0;
 810         unsigned int maxfraglen, fragheaderlen;
 811         int csummode = CHECKSUM_NONE;
 812         struct rtable *rt;
 813
 814         if (flags&MSG_PROBE)
 815                 return 0;
 816
 817         if (skb_queue_empty(&sk->sk_write_queue)) {
 818                 /*
 819                  * setup for corking.
 820                  */
 821                 opt = ipc->opt;
 822                 if (opt) {
 823                         if (inet->cork.opt == NULL) {
 824                                 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
 825                                 if (unlikely(inet->cork.opt == NULL))
 826                                         return -ENOBUFS;
 827                         }
 828                         memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
 829                         inet->cork.flags |= IPCORK_OPT;
 830                         inet->cork.addr = ipc->addr;
 831                 }
 832                 rt = *rtp;
 833                 if (unlikely(!rt))
 834                         return -EFAULT;
 835                 /*
 836                  * We steal reference to this route, caller should not release it
 837                  */
 838                 *rtp = NULL;
 839                 inet->cork.fragsize = mtu = inet->pmtudisc == IP_PMTUDISC_PROBE ?
 840                                             rt->dst.dev->mtu :
 841                                             dst_mtu(rt->dst.path);
 842                 inet->cork.dst = &rt->dst;
 843                 inet->cork.length = 0;
 844                 sk->sk_sndmsg_page = NULL;
 845                 sk->sk_sndmsg_off = 0;
 846                 exthdrlen = rt->dst.header_len;
 847                 length += exthdrlen;
 848                 transhdrlen += exthdrlen;
 849         } else {
 850                 rt = (struct rtable *)inet->cork.dst;
 851                 if (inet->cork.flags & IPCORK_OPT)
 852                         opt = inet->cork.opt;
 853
 854                 transhdrlen = 0;
 855                 exthdrlen = 0;
 856                 mtu = inet->cork.fragsize;
 857         }
 858         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 859
 860         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 861         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 862
 863         if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
 864                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport,
 865                                mtu-exthdrlen);
 866                 return -EMSGSIZE;
 867         }
 868
 869         /*
 870          * transhdrlen > 0 means that this is the first fragment and we wish
 871          * it won't be fragmented in the future.
 872          */
 873         if (transhdrlen &&
 874             length + fragheaderlen <= mtu &&
 875             rt->dst.dev->features & NETIF_F_V4_CSUM &&
 876             !exthdrlen)
 877                 csummode = CHECKSUM_PARTIAL;
 878
 879         skb = skb_peek_tail(&sk->sk_write_queue);
 880
 881         inet->cork.length += length;
 882         if (((length > mtu) || (skb && skb_is_gso(skb))) &&
 883             (sk->sk_protocol == IPPROTO_UDP) &&
 884             (rt->dst.dev->features & NETIF_F_UFO)) {
 885                 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
 886                                          fragheaderlen, transhdrlen, mtu,
 887                                          flags);
 888                 if (err)
 889                         goto error;
 890                 return 0;
 891         }
 892
 893         /* So, what's going on in the loop below?
 894          *
 895          * We use calculated fragment length to generate chained skb,
 896          * each of segments is IP fragment ready for sending to network after
 897          * adding appropriate IP header.
 898          */
 899
 900         if (!skb)
 901                 goto alloc_new_skb;
 902
 903         while (length > 0) {
 904                 /* Check if the remaining data fits into current packet. */
 905                 copy = mtu - skb->len;
 906                 if (copy < length)
 907                         copy = maxfraglen - skb->len;
 908                 if (copy <= 0) {
 909                         char *data;
 910                         unsigned int datalen;
 911                         unsigned int fraglen;
 912                         unsigned int fraggap;
 913                         unsigned int alloclen;
 914                         struct sk_buff *skb_prev;
 915 alloc_new_skb:
 916                         skb_prev = skb;
 917                         if (skb_prev)
 918                                 fraggap = skb_prev->len - maxfraglen;
 919                         else
 920                                 fraggap = 0;
 921
 922                         /*
 923                          * If remaining data exceeds the mtu,
 924                          * we know we need more fragment(s).
 925                          */
 926                         datalen = length + fraggap;
 927                         if (datalen > mtu - fragheaderlen)
 928                                 datalen = maxfraglen - fragheaderlen;
 929                         fraglen = datalen + fragheaderlen;
 930
 931                         if ((flags & MSG_MORE) &&
 932                             !(rt->dst.dev->features&NETIF_F_SG))
 933                                 alloclen = mtu;
 934                         else
 935                                 alloclen = fraglen;
 936
 937                         /* The last fragment gets additional space at tail.
 938                          * Note, with MSG_MORE we overallocate on fragments,
 939                          * because we have no idea what fragment will be
 940                          * the last.
 941                          */
 942                         if (datalen == length + fraggap) {
 943                                 alloclen += rt->dst.trailer_len;
 944                                 /* make sure mtu is not reached */
 945                                 if (datalen > mtu - fragheaderlen - rt->dst.trailer_len)
 946                                         datalen -= ALIGN(rt->dst.trailer_len, 8);
 947                         }
 948                         if (transhdrlen) {
 949                                 skb = sock_alloc_send_skb(sk,
 950                                                 alloclen + hh_len + 15,
 951                                                 (flags & MSG_DONTWAIT), &err);
 952                         } else {
 953                                 skb = NULL;
 954                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 955                                     2 * sk->sk_sndbuf)
 956                                         skb = sock_wmalloc(sk,
 957                                                            alloclen + hh_len + 15, 1,
 958                                                            sk->sk_allocation);
 959                                 if (unlikely(skb == NULL))
 960                                         err = -ENOBUFS;
 961                                 else
 962                                         /* only the initial fragment is
 963                                            time stamped */
 964                                         ipc->tx_flags = 0;
 965                         }
 966                         if (skb == NULL)
 967                                 goto error;
 968
 969                         /*
 970                          *      Fill in the control structures
 971                          */
 972                         skb->ip_summed = csummode;
 973                         skb->csum = 0;
 974                         skb_reserve(skb, hh_len);
 975                         skb_shinfo(skb)->tx_flags = ipc->tx_flags;
 976
 977                         /*
 978                          *      Find where to start putting bytes.
 979                          */
 980                         data = skb_put(skb, fraglen);
 981                         skb_set_network_header(skb, exthdrlen);
 982                         skb->transport_header = (skb->network_header +
 983                                                  fragheaderlen);
 984                         data += fragheaderlen;
 985
 986                         if (fraggap) {
 987                                 skb->csum = skb_copy_and_csum_bits(
 988                                         skb_prev, maxfraglen,
 989                                         data + transhdrlen, fraggap, 0);
 990                                 skb_prev->csum = csum_sub(skb_prev->csum,
 991                                                           skb->csum);
 992                                 data += fraggap;
 993                                 pskb_trim_unique(skb_prev, maxfraglen);
 994                         }
 995
 996                         copy = datalen - transhdrlen - fraggap;
 997                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
 998                                 err = -EFAULT;
 999                                 kfree_skb(skb);
1000                                 goto error;
1001                         }
1002
1003                         offset += copy;
1004                         length -= datalen - fraggap;
1005                         transhdrlen = 0;
1006                         exthdrlen = 0;
1007                         csummode = CHECKSUM_NONE;
1008
1009                         /*
1010                          * Put the packet on the pending queue.
1011                          */
1012                         __skb_queue_tail(&sk->sk_write_queue, skb);
1013                         continue;
1014                 }
1015
1016                 if (copy > length)
1017                         copy = length;
1018
1019                 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1020                         unsigned int off;
1021
1022                         off = skb->len;
1023                         if (getfrag(from, skb_put(skb, copy),
1024                                         offset, copy, off, skb) < 0) {
1025                                 __skb_trim(skb, off);
1026                                 err = -EFAULT;
1027                                 goto error;
1028                         }
1029                 } else {
1030                         int i = skb_shinfo(skb)->nr_frags;
1031                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
1032                         struct page *page = sk->sk_sndmsg_page;
1033                         int off = sk->sk_sndmsg_off;
1034                         unsigned int left;
1035
1036                         if (page && (left = PAGE_SIZE - off) > 0) {
1037                                 if (copy >= left)
1038                                         copy = left;
1039                                 if (page != frag->page) {
1040                                         if (i == MAX_SKB_FRAGS) {
1041                                                 err = -EMSGSIZE;
1042                                                 goto error;
1043                                         }
1044                                         get_page(page);
1045                                         skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1046                                         frag = &skb_shinfo(skb)->frags[i];
1047                                 }
1048                         } else if (i < MAX_SKB_FRAGS) {
1049                                 if (copy > PAGE_SIZE)
1050                                         copy = PAGE_SIZE;
1051                                 page = alloc_pages(sk->sk_allocation, 0);
1052                                 if (page == NULL)  {
1053                                         err = -ENOMEM;
1054                                         goto error;
1055                                 }
1056                                 sk->sk_sndmsg_page = page;
1057                                 sk->sk_sndmsg_off = 0;
1058
1059                                 skb_fill_page_desc(skb, i, page, 0, 0);
1060                                 frag = &skb_shinfo(skb)->frags[i];
1061                         } else {
1062                                 err = -EMSGSIZE;
1063                                 goto error;
1064                         }
1065                         if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1066                                 err = -EFAULT;
1067                                 goto error;
1068                         }
1069                         sk->sk_sndmsg_off += copy;
1070                         frag->size += copy;
1071                         skb->len += copy;
1072                         skb->data_len += copy;
1073                         skb->truesize += copy;
1074                         atomic_add(copy, &sk->sk_wmem_alloc);
1075                 }
1076                 offset += copy;
1077                 length -= copy;
1078         }
1079
1080         return 0;
1081
1082 error:
1083         inet->cork.length -= length;
1084         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1085         return err;
1086 }
1087
1088 ssize_t ip_append_page(struct sock *sk, struct page *page,
1089                        int offset, size_t size, int flags)
1090 {
1091         struct inet_sock *inet = inet_sk(sk);
1092         struct sk_buff *skb;
1093         struct rtable *rt;
1094         struct ip_options *opt = NULL;
1095         int hh_len;
1096         int mtu;
1097         int len;
1098         int err;
1099         unsigned int maxfraglen, fragheaderlen, fraggap;
1100
1101         if (inet->hdrincl)
1102                 return -EPERM;
1103
1104         if (flags&MSG_PROBE)
1105                 return 0;
1106
1107         if (skb_queue_empty(&sk->sk_write_queue))
1108                 return -EINVAL;
1109
1110         rt = (struct rtable *)inet->cork.dst;
1111         if (inet->cork.flags & IPCORK_OPT)
1112                 opt = inet->cork.opt;
1113
1114         if (!(rt->dst.dev->features&NETIF_F_SG))
1115                 return -EOPNOTSUPP;
1116
1117         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1118         mtu = inet->cork.fragsize;
1119
1120         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1121         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1122
1123         if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1124                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport, mtu);
1125                 return -EMSGSIZE;
1126         }
1127
1128         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1129                 return -EINVAL;
1130
1131         inet->cork.length += size;
1132         if ((size + skb->len > mtu) &&
1133             (sk->sk_protocol == IPPROTO_UDP) &&
1134             (rt->dst.dev->features & NETIF_F_UFO)) {
1135                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1136                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1137         }
1138
1139
1140         while (size > 0) {
1141                 int i;
1142
1143                 if (skb_is_gso(skb))
1144                         len = size;
1145                 else {
1146
1147                         /* Check if the remaining data fits into current packet. */
1148                         len = mtu - skb->len;
1149                         if (len < size)
1150                                 len = maxfraglen - skb->len;
1151                 }
1152                 if (len <= 0) {
1153                         struct sk_buff *skb_prev;
1154                         int alloclen;
1155
1156                         skb_prev = skb;
1157                         fraggap = skb_prev->len - maxfraglen;
1158
1159                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1160                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1161                         if (unlikely(!skb)) {
1162                                 err = -ENOBUFS;
1163                                 goto error;
1164                         }
1165
1166                         /*
1167                          *      Fill in the control structures
1168                          */
1169                         skb->ip_summed = CHECKSUM_NONE;
1170                         skb->csum = 0;
1171                         skb_reserve(skb, hh_len);
1172
1173                         /*
1174                          *      Find where to start putting bytes.
1175                          */
1176                         skb_put(skb, fragheaderlen + fraggap);
1177                         skb_reset_network_header(skb);
1178                         skb->transport_header = (skb->network_header +
1179                                                  fragheaderlen);
1180                         if (fraggap) {
1181                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1182                                                                    maxfraglen,
1183                                                     skb_transport_header(skb),
1184                                                                    fraggap, 0);
1185                                 skb_prev->csum = csum_sub(skb_prev->csum,
1186                                                           skb->csum);
1187                                 pskb_trim_unique(skb_prev, maxfraglen);
1188                         }
1189
1190                         /*
1191                          * Put the packet on the pending queue.
1192                          */
1193                         __skb_queue_tail(&sk->sk_write_queue, skb);
1194                         continue;
1195                 }
1196
1197                 i = skb_shinfo(skb)->nr_frags;
1198                 if (len > size)
1199                         len = size;
1200                 if (skb_can_coalesce(skb, i, page, offset)) {
1201                         skb_shinfo(skb)->frags[i-1].size += len;
1202                 } else if (i < MAX_SKB_FRAGS) {
1203                         get_page(page);
1204                         skb_fill_page_desc(skb, i, page, offset, len);
1205                 } else {
1206                         err = -EMSGSIZE;
1207                         goto error;
1208                 }
1209
1210                 if (skb->ip_summed == CHECKSUM_NONE) {
1211                         __wsum csum;
1212                         csum = csum_page(page, offset, len);
1213                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1214                 }
1215
1216                 skb->len += len;
1217                 skb->data_len += len;
1218                 skb->truesize += len;
1219                 atomic_add(len, &sk->sk_wmem_alloc);
1220                 offset += len;
1221                 size -= len;
1222         }
1223         return 0;
1224
1225 error:
1226         inet->cork.length -= size;
1227         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1228         return err;
1229 }
1230
1231 static void ip_cork_release(struct inet_sock *inet)
1232 {
1233         inet->cork.flags &= ~IPCORK_OPT;
1234         kfree(inet->cork.opt);
1235         inet->cork.opt = NULL;
1236         dst_release(inet->cork.dst);
1237         inet->cork.dst = NULL;
1238 }
1239
1240 /*
1241  *      Combined all pending IP fragments on the socket as one IP datagram
1242  *      and push them out.
1243  */
1244 int ip_push_pending_frames(struct sock *sk)
1245 {
1246         struct sk_buff *skb, *tmp_skb;
1247         struct sk_buff **tail_skb;
1248         struct inet_sock *inet = inet_sk(sk);
1249         struct net *net = sock_net(sk);
1250         struct ip_options *opt = NULL;
1251         struct rtable *rt = (struct rtable *)inet->cork.dst;
1252         struct iphdr *iph;
1253         __be16 df = 0;
1254         __u8 ttl;
1255         int err = 0;
1256
1257         if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1258                 goto out;
1259         tail_skb = &(skb_shinfo(skb)->frag_list);
1260
1261         /* move skb->data to ip header from ext header */
1262         if (skb->data < skb_network_header(skb))
1263                 __skb_pull(skb, skb_network_offset(skb));
1264         while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1265                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1266                 *tail_skb = tmp_skb;
1267                 tail_skb = &(tmp_skb->next);
1268                 skb->len += tmp_skb->len;
1269                 skb->data_len += tmp_skb->len;
1270                 skb->truesize += tmp_skb->truesize;
1271                 tmp_skb->destructor = NULL;
1272                 tmp_skb->sk = NULL;
1273         }
1274
1275         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1276          * to fragment the frame generated here. No matter, what transforms
1277          * how transforms change size of the packet, it will come out.
1278          */
1279         if (inet->pmtudisc < IP_PMTUDISC_DO)
1280                 skb->local_df = 1;
1281
1282         /* DF bit is set when we want to see DF on outgoing frames.
1283          * If local_df is set too, we still allow to fragment this frame
1284          * locally. */
1285         if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1286             (skb->len <= dst_mtu(&rt->dst) &&
1287              ip_dont_fragment(sk, &rt->dst)))
1288                 df = htons(IP_DF);
1289
1290         if (inet->cork.flags & IPCORK_OPT)
1291                 opt = inet->cork.opt;
1292
1293         if (rt->rt_type == RTN_MULTICAST)
1294                 ttl = inet->mc_ttl;
1295         else
1296                 ttl = ip_select_ttl(inet, &rt->dst);
1297
1298         iph = (struct iphdr *)skb->data;
1299         iph->version = 4;
1300         iph->ihl = 5;
1301         if (opt) {
1302                 iph->ihl += opt->optlen>>2;
1303                 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1304         }
1305         iph->tos = inet->tos;
1306         iph->frag_off = df;
1307         ip_select_ident(iph, &rt->dst, sk);
1308         iph->ttl = ttl;
1309         iph->protocol = sk->sk_protocol;
1310         iph->saddr = rt->rt_src;
1311         iph->daddr = rt->rt_dst;
1312
1313         skb->priority = sk->sk_priority;
1314         skb->mark = sk->sk_mark;
1315         /*
1316          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1317          * on dst refcount
1318          */
1319         inet->cork.dst = NULL;
1320         skb_dst_set(skb, &rt->dst);
1321
1322         if (iph->protocol == IPPROTO_ICMP)
1323                 icmp_out_count(net, ((struct icmphdr *)
1324                         skb_transport_header(skb))->type);
1325
1326         /* Netfilter gets whole the not fragmented skb. */
1327         err = ip_local_out(skb);
1328         if (err) {
1329                 if (err > 0)
1330                         err = net_xmit_errno(err);
1331                 if (err)
1332                         goto error;
1333         }
1334
1335 out:
1336         ip_cork_release(inet);
1337         return err;
1338
1339 error:
1340         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1341         goto out;
1342 }
1343
1344 /*
1345  *      Throw away all pending data on the socket.
1346  */
1347 void ip_flush_pending_frames(struct sock *sk)
1348 {
1349         struct sk_buff *skb;
1350
1351         while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1352                 kfree_skb(skb);
1353
1354         ip_cork_release(inet_sk(sk));
1355 }
1356
1357
1358 /*
1359  *      Fetch data from kernel space and fill in checksum if needed.
1360  */
1361 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1362                               int len, int odd, struct sk_buff *skb)
1363 {
1364         __wsum csum;
1365
1366         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1367         skb->csum = csum_block_add(skb->csum, csum, odd);
1368         return 0;
1369 }
1370
1371 /*
1372  *      Generic function to send a packet as reply to another packet.
1373  *      Used to send TCP resets so far. ICMP should use this function too.
1374  *
1375  *      Should run single threaded per socket because it uses the sock
1376  *      structure to pass arguments.
1377  */
1378 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1379                    unsigned int len)
1380 {
1381         struct inet_sock *inet = inet_sk(sk);
1382         struct {
1383                 struct ip_options       opt;
1384                 char                    data[40];
1385         } replyopts;
1386         struct ipcm_cookie ipc;
1387         __be32 daddr;
1388         struct rtable *rt = skb_rtable(skb);
1389
1390         if (ip_options_echo(&replyopts.opt, skb))
1391                 return;
1392
1393         daddr = ipc.addr = rt->rt_src;
1394         ipc.opt = NULL;
1395         ipc.tx_flags = 0;
1396
1397         if (replyopts.opt.optlen) {
1398                 ipc.opt = &replyopts.opt;
1399
1400                 if (ipc.opt->srr)
1401                         daddr = replyopts.opt.faddr;
1402         }
1403
1404         {
1405                 struct flowi fl = { .oif = arg->bound_dev_if,
1406                                     .fl4_dst = daddr,
1407                                     .fl4_src = rt->rt_spec_dst,
1408                                     .fl4_tos = RT_TOS(ip_hdr(skb)->tos),
1409                                     .fl_ip_sport = tcp_hdr(skb)->dest,
1410                                     .fl_ip_dport = tcp_hdr(skb)->source,
1411                                     .proto = sk->sk_protocol,
1412                                     .flags = ip_reply_arg_flowi_flags(arg) };
1413                 security_skb_classify_flow(skb, &fl);
1414                 if (ip_route_output_key(sock_net(sk), &rt, &fl))
1415                         return;
1416         }
1417
1418         /* And let IP do all the hard work.
1419
1420            This chunk is not reenterable, hence spinlock.
1421            Note that it uses the fact, that this function is called
1422            with locally disabled BH and that sk cannot be already spinlocked.
1423          */
1424         bh_lock_sock(sk);
1425         inet->tos = ip_hdr(skb)->tos;
1426         sk->sk_priority = skb->priority;
1427         sk->sk_protocol = ip_hdr(skb)->protocol;
1428         sk->sk_bound_dev_if = arg->bound_dev_if;
1429         ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1430                        &ipc, &rt, MSG_DONTWAIT);
1431         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1432                 if (arg->csumoffset >= 0)
1433                         *((__sum16 *)skb_transport_header(skb) +
1434                           arg->csumoffset) = csum_fold(csum_add(skb->csum,
1435                                                                 arg->csum));
1436                 skb->ip_summed = CHECKSUM_NONE;
1437                 ip_push_pending_frames(sk);
1438         }
1439
1440         bh_unlock_sock(sk);
1441
1442         ip_rt_put(rt);
1443 }
1444
1445 void __init ip_init(void)
1446 {
1447         ip_rt_init();
1448         inet_initpeers();
1449
1450 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1451         igmp_mc_proc_init();
1452 #endif
1453 }