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_rcu *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->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    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
 162         iph->saddr    = saddr;
 163         iph->protocol = sk->sk_protocol;
 164         ip_select_ident(iph, &rt->dst, sk);
 165
 166         if (opt && opt->opt.optlen) {
 167                 iph->ihl += opt->opt.optlen>>2;
 168                 ip_options_build(skb, &opt->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         struct neighbour *neigh;
 186
 187         if (rt->rt_type == RTN_MULTICAST) {
 188                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
 189         } else if (rt->rt_type == RTN_BROADCAST)
 190                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
 191
 192         /* Be paranoid, rather than too clever. */
 193         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 194                 struct sk_buff *skb2;
 195
 196                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 197                 if (skb2 == NULL) {
 198                         kfree_skb(skb);
 199                         return -ENOMEM;
 200                 }
 201                 if (skb->sk)
 202                         skb_set_owner_w(skb2, skb->sk);
 203                 kfree_skb(skb);
 204                 skb = skb2;
 205         }
 206
 207         rcu_read_lock();
 208         neigh = dst_get_neighbour(dst);
 209         if (neigh) {
 210                 int res = neigh_output(neigh, skb);
 211
 212                 rcu_read_unlock();
 213                 return res;
 214         }
 215         rcu_read_unlock();
 216
 217         if (net_ratelimit())
 218                 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
 219         kfree_skb(skb);
 220         return -EINVAL;
 221 }
 222
 223 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
 224 {
 225         struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
 226
 227         return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
 228                skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
 229 }
 230
 231 static int ip_finish_output(struct sk_buff *skb)
 232 {
 233 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 234         /* Policy lookup after SNAT yielded a new policy */
 235         if (skb_dst(skb)->xfrm != NULL) {
 236                 IPCB(skb)->flags |= IPSKB_REROUTED;
 237                 return dst_output(skb);
 238         }
 239 #endif
 240         if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
 241                 return ip_fragment(skb, ip_finish_output2);
 242         else
 243                 return ip_finish_output2(skb);
 244 }
 245
 246 int ip_mc_output(struct sk_buff *skb)
 247 {
 248         struct sock *sk = skb->sk;
 249         struct rtable *rt = skb_rtable(skb);
 250         struct net_device *dev = rt->dst.dev;
 251
 252         /*
 253          *      If the indicated interface is up and running, send the packet.
 254          */
 255         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 256
 257         skb->dev = dev;
 258         skb->protocol = htons(ETH_P_IP);
 259
 260         /*
 261          *      Multicasts are looped back for other local users
 262          */
 263
 264         if (rt->rt_flags&RTCF_MULTICAST) {
 265                 if (sk_mc_loop(sk)
 266 #ifdef CONFIG_IP_MROUTE
 267                 /* Small optimization: do not loopback not local frames,
 268                    which returned after forwarding; they will be  dropped
 269                    by ip_mr_input in any case.
 270                    Note, that local frames are looped back to be delivered
 271                    to local recipients.
 272
 273                    This check is duplicated in ip_mr_input at the moment.
 274                  */
 275                     &&
 276                     ((rt->rt_flags & RTCF_LOCAL) ||
 277                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 278 #endif
 279                    ) {
 280                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 281                         if (newskb)
 282                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 283                                         newskb, NULL, newskb->dev,
 284                                         ip_dev_loopback_xmit);
 285                 }
 286
 287                 /* Multicasts with ttl 0 must not go beyond the host */
 288
 289                 if (ip_hdr(skb)->ttl == 0) {
 290                         kfree_skb(skb);
 291                         return 0;
 292                 }
 293         }
 294
 295         if (rt->rt_flags&RTCF_BROADCAST) {
 296                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 297                 if (newskb)
 298                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
 299                                 NULL, newskb->dev, ip_dev_loopback_xmit);
 300         }
 301
 302         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
 303                             skb->dev, ip_finish_output,
 304                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 305 }
 306
 307 int ip_output(struct sk_buff *skb)
 308 {
 309         struct net_device *dev = skb_dst(skb)->dev;
 310
 311         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 312
 313         skb->dev = dev;
 314         skb->protocol = htons(ETH_P_IP);
 315
 316         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
 317                             ip_finish_output,
 318                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 319 }
 320
 321 int ip_queue_xmit(struct sk_buff *skb, struct flowi *fl)
 322 {
 323         struct sock *sk = skb->sk;
 324         struct inet_sock *inet = inet_sk(sk);
 325         struct ip_options_rcu *inet_opt;
 326         struct flowi4 *fl4;
 327         struct rtable *rt;
 328         struct iphdr *iph;
 329         int res;
 330
 331         /* Skip all of this if the packet is already routed,
 332          * f.e. by something like SCTP.
 333          */
 334         rcu_read_lock();
 335         inet_opt = rcu_dereference(inet->inet_opt);
 336         fl4 = &fl->u.ip4;
 337         rt = skb_rtable(skb);
 338         if (rt != NULL)
 339                 goto packet_routed;
 340
 341         /* Make sure we can route this packet. */
 342         rt = (struct rtable *)__sk_dst_check(sk, 0);
 343         if (rt == NULL) {
 344                 __be32 daddr;
 345
 346                 /* Use correct destination address if we have options. */
 347                 daddr = inet->inet_daddr;
 348                 if (inet_opt && inet_opt->opt.srr)
 349                         daddr = inet_opt->opt.faddr;
 350
 351                 /* If this fails, retransmit mechanism of transport layer will
 352                  * keep trying until route appears or the connection times
 353                  * itself out.
 354                  */
 355                 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
 356                                            daddr, inet->inet_saddr,
 357                                            inet->inet_dport,
 358                                            inet->inet_sport,
 359                                            sk->sk_protocol,
 360                                            RT_CONN_FLAGS(sk),
 361                                            sk->sk_bound_dev_if);
 362                 if (IS_ERR(rt))
 363                         goto no_route;
 364                 sk_setup_caps(sk, &rt->dst);
 365         }
 366         skb_dst_set_noref(skb, &rt->dst);
 367
 368 packet_routed:
 369         if (inet_opt && inet_opt->opt.is_strictroute && fl4->daddr != rt->rt_gateway)
 370                 goto no_route;
 371
 372         /* OK, we know where to send it, allocate and build IP header. */
 373         skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
 374         skb_reset_network_header(skb);
 375         iph = ip_hdr(skb);
 376         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 377         if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
 378                 iph->frag_off = htons(IP_DF);
 379         else
 380                 iph->frag_off = 0;
 381         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 382         iph->protocol = sk->sk_protocol;
 383         iph->saddr    = fl4->saddr;
 384         iph->daddr    = fl4->daddr;
 385         /* Transport layer set skb->h.foo itself. */
 386
 387         if (inet_opt && inet_opt->opt.optlen) {
 388                 iph->ihl += inet_opt->opt.optlen >> 2;
 389                 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
 390         }
 391
 392         ip_select_ident_more(iph, &rt->dst, sk,
 393                              (skb_shinfo(skb)->gso_segs ?: 1) - 1);
 394
 395         skb->priority = sk->sk_priority;
 396         skb->mark = sk->sk_mark;
 397
 398         res = ip_local_out(skb);
 399         rcu_read_unlock();
 400         return res;
 401
 402 no_route:
 403         rcu_read_unlock();
 404         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 405         kfree_skb(skb);
 406         return -EHOSTUNREACH;
 407 }
 408 EXPORT_SYMBOL(ip_queue_xmit);
 409
 410
 411 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 412 {
 413         to->pkt_type = from->pkt_type;
 414         to->priority = from->priority;
 415         to->protocol = from->protocol;
 416         skb_dst_drop(to);
 417         skb_dst_copy(to, from);
 418         to->dev = from->dev;
 419         to->mark = from->mark;
 420
 421         /* Copy the flags to each fragment. */
 422         IPCB(to)->flags = IPCB(from)->flags;
 423
 424 #ifdef CONFIG_NET_SCHED
 425         to->tc_index = from->tc_index;
 426 #endif
 427         nf_copy(to, from);
 428 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
 429     defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
 430         to->nf_trace = from->nf_trace;
 431 #endif
 432 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 433         to->ipvs_property = from->ipvs_property;
 434 #endif
 435         skb_copy_secmark(to, from);
 436 }
 437
 438 /*
 439  *      This IP datagram is too large to be sent in one piece.  Break it up into
 440  *      smaller pieces (each of size equal to IP header plus
 441  *      a block of the data of the original IP data part) that will yet fit in a
 442  *      single device frame, and queue such a frame for sending.
 443  */
 444
 445 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 446 {
 447         struct iphdr *iph;
 448         int ptr;
 449         struct net_device *dev;
 450         struct sk_buff *skb2;
 451         unsigned int mtu, hlen, left, len, ll_rs;
 452         int offset;
 453         __be16 not_last_frag;
 454         struct rtable *rt = skb_rtable(skb);
 455         int err = 0;
 456
 457         dev = rt->dst.dev;
 458
 459         /*
 460          *      Point into the IP datagram header.
 461          */
 462
 463         iph = ip_hdr(skb);
 464
 465         if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
 466                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 467                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 468                           htonl(ip_skb_dst_mtu(skb)));
 469                 kfree_skb(skb);
 470                 return -EMSGSIZE;
 471         }
 472
 473         /*
 474          *      Setup starting values.
 475          */
 476
 477         hlen = iph->ihl * 4;
 478         mtu = dst_mtu(&rt->dst) - hlen; /* Size of data space */
 479 #ifdef CONFIG_BRIDGE_NETFILTER
 480         if (skb->nf_bridge)
 481                 mtu -= nf_bridge_mtu_reduction(skb);
 482 #endif
 483         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 484
 485         /* When frag_list is given, use it. First, check its validity:
 486          * some transformers could create wrong frag_list or break existing
 487          * one, it is not prohibited. In this case fall back to copying.
 488          *
 489          * LATER: this step can be merged to real generation of fragments,
 490          * we can switch to copy when see the first bad fragment.
 491          */
 492         if (skb_has_frag_list(skb)) {
 493                 struct sk_buff *frag, *frag2;
 494                 int first_len = skb_pagelen(skb);
 495
 496                 if (first_len - hlen > mtu ||
 497                     ((first_len - hlen) & 7) ||
 498                     ip_is_fragment(iph) ||
 499                     skb_cloned(skb))
 500                         goto slow_path;
 501
 502                 skb_walk_frags(skb, frag) {
 503                         /* Correct geometry. */
 504                         if (frag->len > mtu ||
 505                             ((frag->len & 7) && frag->next) ||
 506                             skb_headroom(frag) < hlen)
 507                                 goto slow_path_clean;
 508
 509                         /* Partially cloned skb? */
 510                         if (skb_shared(frag))
 511                                 goto slow_path_clean;
 512
 513                         BUG_ON(frag->sk);
 514                         if (skb->sk) {
 515                                 frag->sk = skb->sk;
 516                                 frag->destructor = sock_wfree;
 517                         }
 518                         skb->truesize -= frag->truesize;
 519                 }
 520
 521                 /* Everything is OK. Generate! */
 522
 523                 err = 0;
 524                 offset = 0;
 525                 frag = skb_shinfo(skb)->frag_list;
 526                 skb_frag_list_init(skb);
 527                 skb->data_len = first_len - skb_headlen(skb);
 528                 skb->len = first_len;
 529                 iph->tot_len = htons(first_len);
 530                 iph->frag_off = htons(IP_MF);
 531                 ip_send_check(iph);
 532
 533                 for (;;) {
 534                         /* Prepare header of the next frame,
 535                          * before previous one went down. */
 536                         if (frag) {
 537                                 frag->ip_summed = CHECKSUM_NONE;
 538                                 skb_reset_transport_header(frag);
 539                                 __skb_push(frag, hlen);
 540                                 skb_reset_network_header(frag);
 541                                 memcpy(skb_network_header(frag), iph, hlen);
 542                                 iph = ip_hdr(frag);
 543                                 iph->tot_len = htons(frag->len);
 544                                 ip_copy_metadata(frag, skb);
 545                                 if (offset == 0)
 546                                         ip_options_fragment(frag);
 547                                 offset += skb->len - hlen;
 548                                 iph->frag_off = htons(offset>>3);
 549                                 if (frag->next != NULL)
 550                                         iph->frag_off |= htons(IP_MF);
 551                                 /* Ready, complete checksum */
 552                                 ip_send_check(iph);
 553                         }
 554
 555                         err = output(skb);
 556
 557                         if (!err)
 558                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 559                         if (err || !frag)
 560                                 break;
 561
 562                         skb = frag;
 563                         frag = skb->next;
 564                         skb->next = NULL;
 565                 }
 566
 567                 if (err == 0) {
 568                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 569                         return 0;
 570                 }
 571
 572                 while (frag) {
 573                         skb = frag->next;
 574                         kfree_skb(frag);
 575                         frag = skb;
 576                 }
 577                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 578                 return err;
 579
 580 slow_path_clean:
 581                 skb_walk_frags(skb, frag2) {
 582                         if (frag2 == frag)
 583                                 break;
 584                         frag2->sk = NULL;
 585                         frag2->destructor = NULL;
 586                         skb->truesize += frag2->truesize;
 587                 }
 588         }
 589
 590 slow_path:
 591         left = skb->len - hlen;         /* Space per frame */
 592         ptr = hlen;             /* Where to start from */
 593
 594         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 595          * we need to make room for the encapsulating header
 596          */
 597         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
 598
 599         /*
 600          *      Fragment the datagram.
 601          */
 602
 603         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 604         not_last_frag = iph->frag_off & htons(IP_MF);
 605
 606         /*
 607          *      Keep copying data until we run out.
 608          */
 609
 610         while (left > 0) {
 611                 len = left;
 612                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 613                 if (len > mtu)
 614                         len = mtu;
 615                 /* IF: we are not sending up to and including the packet end
 616                    then align the next start on an eight byte boundary */
 617                 if (len < left) {
 618                         len &= ~7;
 619                 }
 620                 /*
 621                  *      Allocate buffer.
 622                  */
 623
 624                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 625                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 626                         err = -ENOMEM;
 627                         goto fail;
 628                 }
 629
 630                 /*
 631                  *      Set up data on packet
 632                  */
 633
 634                 ip_copy_metadata(skb2, skb);
 635                 skb_reserve(skb2, ll_rs);
 636                 skb_put(skb2, len + hlen);
 637                 skb_reset_network_header(skb2);
 638                 skb2->transport_header = skb2->network_header + hlen;
 639
 640                 /*
 641                  *      Charge the memory for the fragment to any owner
 642                  *      it might possess
 643                  */
 644
 645                 if (skb->sk)
 646                         skb_set_owner_w(skb2, skb->sk);
 647
 648                 /*
 649                  *      Copy the packet header into the new buffer.
 650                  */
 651
 652                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 653
 654                 /*
 655                  *      Copy a block of the IP datagram.
 656                  */
 657                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 658                         BUG();
 659                 left -= len;
 660
 661                 /*
 662                  *      Fill in the new header fields.
 663                  */
 664                 iph = ip_hdr(skb2);
 665                 iph->frag_off = htons((offset >> 3));
 666
 667                 /* ANK: dirty, but effective trick. Upgrade options only if
 668                  * the segment to be fragmented was THE FIRST (otherwise,
 669                  * options are already fixed) and make it ONCE
 670                  * on the initial skb, so that all the following fragments
 671                  * will inherit fixed options.
 672                  */
 673                 if (offset == 0)
 674                         ip_options_fragment(skb);
 675
 676                 /*
 677                  *      Added AC : If we are fragmenting a fragment that's not the
 678                  *                 last fragment then keep MF on each bit
 679                  */
 680                 if (left > 0 || not_last_frag)
 681                         iph->frag_off |= htons(IP_MF);
 682                 ptr += len;
 683                 offset += len;
 684
 685                 /*
 686                  *      Put this fragment into the sending queue.
 687                  */
 688                 iph->tot_len = htons(len + hlen);
 689
 690                 ip_send_check(iph);
 691
 692                 err = output(skb2);
 693                 if (err)
 694                         goto fail;
 695
 696                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 697         }
 698         kfree_skb(skb);
 699         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 700         return err;
 701
 702 fail:
 703         kfree_skb(skb);
 704         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 705         return err;
 706 }
 707 EXPORT_SYMBOL(ip_fragment);
 708
 709 int
 710 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 711 {
 712         struct iovec *iov = from;
 713
 714         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 715                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 716                         return -EFAULT;
 717         } else {
 718                 __wsum csum = 0;
 719                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 720                         return -EFAULT;
 721                 skb->csum = csum_block_add(skb->csum, csum, odd);
 722         }
 723         return 0;
 724 }
 725 EXPORT_SYMBOL(ip_generic_getfrag);
 726
 727 static inline __wsum
 728 csum_page(struct page *page, int offset, int copy)
 729 {
 730         char *kaddr;
 731         __wsum csum;
 732         kaddr = kmap(page);
 733         csum = csum_partial(kaddr + offset, copy, 0);
 734         kunmap(page);
 735         return csum;
 736 }
 737
 738 static inline int ip_ufo_append_data(struct sock *sk,
 739                         struct sk_buff_head *queue,
 740                         int getfrag(void *from, char *to, int offset, int len,
 741                                int odd, struct sk_buff *skb),
 742                         void *from, int length, int hh_len, int fragheaderlen,
 743                         int transhdrlen, int maxfraglen, unsigned int flags)
 744 {
 745         struct sk_buff *skb;
 746         int err;
 747
 748         /* There is support for UDP fragmentation offload by network
 749          * device, so create one single skb packet containing complete
 750          * udp datagram
 751          */
 752         if ((skb = skb_peek_tail(queue)) == NULL) {
 753                 skb = sock_alloc_send_skb(sk,
 754                         hh_len + fragheaderlen + transhdrlen + 20,
 755                         (flags & MSG_DONTWAIT), &err);
 756
 757                 if (skb == NULL)
 758                         return err;
 759
 760                 /* reserve space for Hardware header */
 761                 skb_reserve(skb, hh_len);
 762
 763                 /* create space for UDP/IP header */
 764                 skb_put(skb, fragheaderlen + transhdrlen);
 765
 766                 /* initialize network header pointer */
 767                 skb_reset_network_header(skb);
 768
 769                 /* initialize protocol header pointer */
 770                 skb->transport_header = skb->network_header + fragheaderlen;
 771
 772                 skb->ip_summed = CHECKSUM_PARTIAL;
 773                 skb->csum = 0;
 774
 775                 /* specify the length of each IP datagram fragment */
 776                 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
 777                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 778                 __skb_queue_tail(queue, skb);
 779         }
 780
 781         return skb_append_datato_frags(sk, skb, getfrag, from,
 782                                        (length - transhdrlen));
 783 }
 784
 785 static int __ip_append_data(struct sock *sk,
 786                             struct flowi4 *fl4,
 787                             struct sk_buff_head *queue,
 788                             struct inet_cork *cork,
 789                             int getfrag(void *from, char *to, int offset,
 790                                         int len, int odd, struct sk_buff *skb),
 791                             void *from, int length, int transhdrlen,
 792                             unsigned int flags)
 793 {
 794         struct inet_sock *inet = inet_sk(sk);
 795         struct sk_buff *skb;
 796
 797         struct ip_options *opt = cork->opt;
 798         int hh_len;
 799         int exthdrlen;
 800         int mtu;
 801         int copy;
 802         int err;
 803         int offset = 0;
 804         unsigned int maxfraglen, fragheaderlen;
 805         int csummode = CHECKSUM_NONE;
 806         struct rtable *rt = (struct rtable *)cork->dst;
 807
 808         skb = skb_peek_tail(queue);
 809
 810         exthdrlen = !skb ? rt->dst.header_len : 0;
 811         mtu = cork->fragsize;
 812
 813         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 814
 815         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 816         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 817
 818         if (cork->length + length > 0xFFFF - fragheaderlen) {
 819                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
 820                                mtu-exthdrlen);
 821                 return -EMSGSIZE;
 822         }
 823
 824         /*
 825          * transhdrlen > 0 means that this is the first fragment and we wish
 826          * it won't be fragmented in the future.
 827          */
 828         if (transhdrlen &&
 829             length + fragheaderlen <= mtu &&
 830             rt->dst.dev->features & NETIF_F_V4_CSUM &&
 831             !exthdrlen)
 832                 csummode = CHECKSUM_PARTIAL;
 833
 834         cork->length += length;
 835         if (((length > mtu) || (skb && skb_is_gso(skb))) &&
 836             (sk->sk_protocol == IPPROTO_UDP) &&
 837             (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len) {
 838                 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
 839                                          hh_len, fragheaderlen, transhdrlen,
 840                                          maxfraglen, flags);
 841                 if (err)
 842                         goto error;
 843                 return 0;
 844         }
 845
 846         /* So, what's going on in the loop below?
 847          *
 848          * We use calculated fragment length to generate chained skb,
 849          * each of segments is IP fragment ready for sending to network after
 850          * adding appropriate IP header.
 851          */
 852
 853         if (!skb)
 854                 goto alloc_new_skb;
 855
 856         while (length > 0) {
 857                 /* Check if the remaining data fits into current packet. */
 858                 copy = mtu - skb->len;
 859                 if (copy < length)
 860                         copy = maxfraglen - skb->len;
 861                 if (copy <= 0) {
 862                         char *data;
 863                         unsigned int datalen;
 864                         unsigned int fraglen;
 865                         unsigned int fraggap;
 866                         unsigned int alloclen;
 867                         struct sk_buff *skb_prev;
 868 alloc_new_skb:
 869                         skb_prev = skb;
 870                         if (skb_prev)
 871                                 fraggap = skb_prev->len - maxfraglen;
 872                         else
 873                                 fraggap = 0;
 874
 875                         /*
 876                          * If remaining data exceeds the mtu,
 877                          * we know we need more fragment(s).
 878                          */
 879                         datalen = length + fraggap;
 880                         if (datalen > mtu - fragheaderlen)
 881                                 datalen = maxfraglen - fragheaderlen;
 882                         fraglen = datalen + fragheaderlen;
 883
 884                         if ((flags & MSG_MORE) &&
 885                             !(rt->dst.dev->features&NETIF_F_SG))
 886                                 alloclen = mtu;
 887                         else
 888                                 alloclen = fraglen;
 889
 890                         alloclen += exthdrlen;
 891
 892                         /* The last fragment gets additional space at tail.
 893                          * Note, with MSG_MORE we overallocate on fragments,
 894                          * because we have no idea what fragment will be
 895                          * the last.
 896                          */
 897                         if (datalen == length + fraggap)
 898                                 alloclen += rt->dst.trailer_len;
 899
 900                         if (transhdrlen) {
 901                                 skb = sock_alloc_send_skb(sk,
 902                                                 alloclen + hh_len + 15,
 903                                                 (flags & MSG_DONTWAIT), &err);
 904                         } else {
 905                                 skb = NULL;
 906                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 907                                     2 * sk->sk_sndbuf)
 908                                         skb = sock_wmalloc(sk,
 909                                                            alloclen + hh_len + 15, 1,
 910                                                            sk->sk_allocation);
 911                                 if (unlikely(skb == NULL))
 912                                         err = -ENOBUFS;
 913                                 else
 914                                         /* only the initial fragment is
 915                                            time stamped */
 916                                         cork->tx_flags = 0;
 917                         }
 918                         if (skb == NULL)
 919                                 goto error;
 920
 921                         /*
 922                          *      Fill in the control structures
 923                          */
 924                         skb->ip_summed = csummode;
 925                         skb->csum = 0;
 926                         skb_reserve(skb, hh_len);
 927                         skb_shinfo(skb)->tx_flags = cork->tx_flags;
 928
 929                         /*
 930                          *      Find where to start putting bytes.
 931                          */
 932                         data = skb_put(skb, fraglen + exthdrlen);
 933                         skb_set_network_header(skb, exthdrlen);
 934                         skb->transport_header = (skb->network_header +
 935                                                  fragheaderlen);
 936                         data += fragheaderlen + exthdrlen;
 937
 938                         if (fraggap) {
 939                                 skb->csum = skb_copy_and_csum_bits(
 940                                         skb_prev, maxfraglen,
 941                                         data + transhdrlen, fraggap, 0);
 942                                 skb_prev->csum = csum_sub(skb_prev->csum,
 943                                                           skb->csum);
 944                                 data += fraggap;
 945                                 pskb_trim_unique(skb_prev, maxfraglen);
 946                         }
 947
 948                         copy = datalen - transhdrlen - fraggap;
 949                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
 950                                 err = -EFAULT;
 951                                 kfree_skb(skb);
 952                                 goto error;
 953                         }
 954
 955                         offset += copy;
 956                         length -= datalen - fraggap;
 957                         transhdrlen = 0;
 958                         exthdrlen = 0;
 959                         csummode = CHECKSUM_NONE;
 960
 961                         /*
 962                          * Put the packet on the pending queue.
 963                          */
 964                         __skb_queue_tail(queue, skb);
 965                         continue;
 966                 }
 967
 968                 if (copy > length)
 969                         copy = length;
 970
 971                 if (!(rt->dst.dev->features&NETIF_F_SG)) {
 972                         unsigned int off;
 973
 974                         off = skb->len;
 975                         if (getfrag(from, skb_put(skb, copy),
 976                                         offset, copy, off, skb) < 0) {
 977                                 __skb_trim(skb, off);
 978                                 err = -EFAULT;
 979                                 goto error;
 980                         }
 981                 } else {
 982                         int i = skb_shinfo(skb)->nr_frags;
 983                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
 984                         struct page *page = cork->page;
 985                         int off = cork->off;
 986                         unsigned int left;
 987
 988                         if (page && (left = PAGE_SIZE - off) > 0) {
 989                                 if (copy >= left)
 990                                         copy = left;
 991                                 if (page != frag->page) {
 992                                         if (i == MAX_SKB_FRAGS) {
 993                                                 err = -EMSGSIZE;
 994                                                 goto error;
 995                                         }
 996                                         get_page(page);
 997                                         skb_fill_page_desc(skb, i, page, off, 0);
 998                                         frag = &skb_shinfo(skb)->frags[i];
 999                                 }
1000                         } else if (i < MAX_SKB_FRAGS) {
1001                                 if (copy > PAGE_SIZE)
1002                                         copy = PAGE_SIZE;
1003                                 page = alloc_pages(sk->sk_allocation, 0);
1004                                 if (page == NULL)  {
1005                                         err = -ENOMEM;
1006                                         goto error;
1007                                 }
1008                                 cork->page = page;
1009                                 cork->off = 0;
1010
1011                                 skb_fill_page_desc(skb, i, page, 0, 0);
1012                                 frag = &skb_shinfo(skb)->frags[i];
1013                         } else {
1014                                 err = -EMSGSIZE;
1015                                 goto error;
1016                         }
1017                         if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1018                                 err = -EFAULT;
1019                                 goto error;
1020                         }
1021                         cork->off += copy;
1022                         frag->size += copy;
1023                         skb->len += copy;
1024                         skb->data_len += copy;
1025                         skb->truesize += copy;
1026                         atomic_add(copy, &sk->sk_wmem_alloc);
1027                 }
1028                 offset += copy;
1029                 length -= copy;
1030         }
1031
1032         return 0;
1033
1034 error:
1035         cork->length -= length;
1036         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1037         return err;
1038 }
1039
1040 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1041                          struct ipcm_cookie *ipc, struct rtable **rtp)
1042 {
1043         struct inet_sock *inet = inet_sk(sk);
1044         struct ip_options_rcu *opt;
1045         struct rtable *rt;
1046
1047         /*
1048          * setup for corking.
1049          */
1050         opt = ipc->opt;
1051         if (opt) {
1052                 if (cork->opt == NULL) {
1053                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1054                                             sk->sk_allocation);
1055                         if (unlikely(cork->opt == NULL))
1056                                 return -ENOBUFS;
1057                 }
1058                 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1059                 cork->flags |= IPCORK_OPT;
1060                 cork->addr = ipc->addr;
1061         }
1062         rt = *rtp;
1063         if (unlikely(!rt))
1064                 return -EFAULT;
1065         /*
1066          * We steal reference to this route, caller should not release it
1067          */
1068         *rtp = NULL;
1069         cork->fragsize = inet->pmtudisc == IP_PMTUDISC_PROBE ?
1070                          rt->dst.dev->mtu : dst_mtu(&rt->dst);
1071         cork->dst = &rt->dst;
1072         cork->length = 0;
1073         cork->tx_flags = ipc->tx_flags;
1074         cork->page = NULL;
1075         cork->off = 0;
1076
1077         return 0;
1078 }
1079
1080 /*
1081  *      ip_append_data() and ip_append_page() can make one large IP datagram
1082  *      from many pieces of data. Each pieces will be holded on the socket
1083  *      until ip_push_pending_frames() is called. Each piece can be a page
1084  *      or non-page data.
1085  *
1086  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1087  *      this interface potentially.
1088  *
1089  *      LATER: length must be adjusted by pad at tail, when it is required.
1090  */
1091 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1092                    int getfrag(void *from, char *to, int offset, int len,
1093                                int odd, struct sk_buff *skb),
1094                    void *from, int length, int transhdrlen,
1095                    struct ipcm_cookie *ipc, struct rtable **rtp,
1096                    unsigned int flags)
1097 {
1098         struct inet_sock *inet = inet_sk(sk);
1099         int err;
1100
1101         if (flags&MSG_PROBE)
1102                 return 0;
1103
1104         if (skb_queue_empty(&sk->sk_write_queue)) {
1105                 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1106                 if (err)
1107                         return err;
1108         } else {
1109                 transhdrlen = 0;
1110         }
1111
1112         return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base, getfrag,
1113                                 from, length, transhdrlen, flags);
1114 }
1115
1116 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1117                        int offset, size_t size, int flags)
1118 {
1119         struct inet_sock *inet = inet_sk(sk);
1120         struct sk_buff *skb;
1121         struct rtable *rt;
1122         struct ip_options *opt = NULL;
1123         struct inet_cork *cork;
1124         int hh_len;
1125         int mtu;
1126         int len;
1127         int err;
1128         unsigned int maxfraglen, fragheaderlen, fraggap;
1129
1130         if (inet->hdrincl)
1131                 return -EPERM;
1132
1133         if (flags&MSG_PROBE)
1134                 return 0;
1135
1136         if (skb_queue_empty(&sk->sk_write_queue))
1137                 return -EINVAL;
1138
1139         cork = &inet->cork.base;
1140         rt = (struct rtable *)cork->dst;
1141         if (cork->flags & IPCORK_OPT)
1142                 opt = cork->opt;
1143
1144         if (!(rt->dst.dev->features&NETIF_F_SG))
1145                 return -EOPNOTSUPP;
1146
1147         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1148         mtu = cork->fragsize;
1149
1150         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1151         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1152
1153         if (cork->length + size > 0xFFFF - fragheaderlen) {
1154                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport, mtu);
1155                 return -EMSGSIZE;
1156         }
1157
1158         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1159                 return -EINVAL;
1160
1161         cork->length += size;
1162         if ((size + skb->len > mtu) &&
1163             (sk->sk_protocol == IPPROTO_UDP) &&
1164             (rt->dst.dev->features & NETIF_F_UFO)) {
1165                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1166                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1167         }
1168
1169
1170         while (size > 0) {
1171                 int i;
1172
1173                 if (skb_is_gso(skb))
1174                         len = size;
1175                 else {
1176
1177                         /* Check if the remaining data fits into current packet. */
1178                         len = mtu - skb->len;
1179                         if (len < size)
1180                                 len = maxfraglen - skb->len;
1181                 }
1182                 if (len <= 0) {
1183                         struct sk_buff *skb_prev;
1184                         int alloclen;
1185
1186                         skb_prev = skb;
1187                         fraggap = skb_prev->len - maxfraglen;
1188
1189                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1190                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1191                         if (unlikely(!skb)) {
1192                                 err = -ENOBUFS;
1193                                 goto error;
1194                         }
1195
1196                         /*
1197                          *      Fill in the control structures
1198                          */
1199                         skb->ip_summed = CHECKSUM_NONE;
1200                         skb->csum = 0;
1201                         skb_reserve(skb, hh_len);
1202
1203                         /*
1204                          *      Find where to start putting bytes.
1205                          */
1206                         skb_put(skb, fragheaderlen + fraggap);
1207                         skb_reset_network_header(skb);
1208                         skb->transport_header = (skb->network_header +
1209                                                  fragheaderlen);
1210                         if (fraggap) {
1211                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1212                                                                    maxfraglen,
1213                                                     skb_transport_header(skb),
1214                                                                    fraggap, 0);
1215                                 skb_prev->csum = csum_sub(skb_prev->csum,
1216                                                           skb->csum);
1217                                 pskb_trim_unique(skb_prev, maxfraglen);
1218                         }
1219
1220                         /*
1221                          * Put the packet on the pending queue.
1222                          */
1223                         __skb_queue_tail(&sk->sk_write_queue, skb);
1224                         continue;
1225                 }
1226
1227                 i = skb_shinfo(skb)->nr_frags;
1228                 if (len > size)
1229                         len = size;
1230                 if (skb_can_coalesce(skb, i, page, offset)) {
1231                         skb_shinfo(skb)->frags[i-1].size += len;
1232                 } else if (i < MAX_SKB_FRAGS) {
1233                         get_page(page);
1234                         skb_fill_page_desc(skb, i, page, offset, len);
1235                 } else {
1236                         err = -EMSGSIZE;
1237                         goto error;
1238                 }
1239
1240                 if (skb->ip_summed == CHECKSUM_NONE) {
1241                         __wsum csum;
1242                         csum = csum_page(page, offset, len);
1243                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1244                 }
1245
1246                 skb->len += len;
1247                 skb->data_len += len;
1248                 skb->truesize += len;
1249                 atomic_add(len, &sk->sk_wmem_alloc);
1250                 offset += len;
1251                 size -= len;
1252         }
1253         return 0;
1254
1255 error:
1256         cork->length -= size;
1257         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1258         return err;
1259 }
1260
1261 static void ip_cork_release(struct inet_cork *cork)
1262 {
1263         cork->flags &= ~IPCORK_OPT;
1264         kfree(cork->opt);
1265         cork->opt = NULL;
1266         dst_release(cork->dst);
1267         cork->dst = NULL;
1268 }
1269
1270 /*
1271  *      Combined all pending IP fragments on the socket as one IP datagram
1272  *      and push them out.
1273  */
1274 struct sk_buff *__ip_make_skb(struct sock *sk,
1275                               struct flowi4 *fl4,
1276                               struct sk_buff_head *queue,
1277                               struct inet_cork *cork)
1278 {
1279         struct sk_buff *skb, *tmp_skb;
1280         struct sk_buff **tail_skb;
1281         struct inet_sock *inet = inet_sk(sk);
1282         struct net *net = sock_net(sk);
1283         struct ip_options *opt = NULL;
1284         struct rtable *rt = (struct rtable *)cork->dst;
1285         struct iphdr *iph;
1286         __be16 df = 0;
1287         __u8 ttl;
1288
1289         if ((skb = __skb_dequeue(queue)) == NULL)
1290                 goto out;
1291         tail_skb = &(skb_shinfo(skb)->frag_list);
1292
1293         /* move skb->data to ip header from ext header */
1294         if (skb->data < skb_network_header(skb))
1295                 __skb_pull(skb, skb_network_offset(skb));
1296         while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1297                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1298                 *tail_skb = tmp_skb;
1299                 tail_skb = &(tmp_skb->next);
1300                 skb->len += tmp_skb->len;
1301                 skb->data_len += tmp_skb->len;
1302                 skb->truesize += tmp_skb->truesize;
1303                 tmp_skb->destructor = NULL;
1304                 tmp_skb->sk = NULL;
1305         }
1306
1307         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1308          * to fragment the frame generated here. No matter, what transforms
1309          * how transforms change size of the packet, it will come out.
1310          */
1311         if (inet->pmtudisc < IP_PMTUDISC_DO)
1312                 skb->local_df = 1;
1313
1314         /* DF bit is set when we want to see DF on outgoing frames.
1315          * If local_df is set too, we still allow to fragment this frame
1316          * locally. */
1317         if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1318             (skb->len <= dst_mtu(&rt->dst) &&
1319              ip_dont_fragment(sk, &rt->dst)))
1320                 df = htons(IP_DF);
1321
1322         if (cork->flags & IPCORK_OPT)
1323                 opt = cork->opt;
1324
1325         if (rt->rt_type == RTN_MULTICAST)
1326                 ttl = inet->mc_ttl;
1327         else
1328                 ttl = ip_select_ttl(inet, &rt->dst);
1329
1330         iph = (struct iphdr *)skb->data;
1331         iph->version = 4;
1332         iph->ihl = 5;
1333         iph->tos = inet->tos;
1334         iph->frag_off = df;
1335         ip_select_ident(iph, &rt->dst, sk);
1336         iph->ttl = ttl;
1337         iph->protocol = sk->sk_protocol;
1338         iph->saddr = fl4->saddr;
1339         iph->daddr = fl4->daddr;
1340
1341         if (opt) {
1342                 iph->ihl += opt->optlen>>2;
1343                 ip_options_build(skb, opt, cork->addr, rt, 0);
1344         }
1345
1346         skb->priority = sk->sk_priority;
1347         skb->mark = sk->sk_mark;
1348         /*
1349          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1350          * on dst refcount
1351          */
1352         cork->dst = NULL;
1353         skb_dst_set(skb, &rt->dst);
1354
1355         if (iph->protocol == IPPROTO_ICMP)
1356                 icmp_out_count(net, ((struct icmphdr *)
1357                         skb_transport_header(skb))->type);
1358
1359         ip_cork_release(cork);
1360 out:
1361         return skb;
1362 }
1363
1364 int ip_send_skb(struct sk_buff *skb)
1365 {
1366         struct net *net = sock_net(skb->sk);
1367         int err;
1368
1369         err = ip_local_out(skb);
1370         if (err) {
1371                 if (err > 0)
1372                         err = net_xmit_errno(err);
1373                 if (err)
1374                         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1375         }
1376
1377         return err;
1378 }
1379
1380 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1381 {
1382         struct sk_buff *skb;
1383
1384         skb = ip_finish_skb(sk, fl4);
1385         if (!skb)
1386                 return 0;
1387
1388         /* Netfilter gets whole the not fragmented skb. */
1389         return ip_send_skb(skb);
1390 }
1391
1392 /*
1393  *      Throw away all pending data on the socket.
1394  */
1395 static void __ip_flush_pending_frames(struct sock *sk,
1396                                       struct sk_buff_head *queue,
1397                                       struct inet_cork *cork)
1398 {
1399         struct sk_buff *skb;
1400
1401         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1402                 kfree_skb(skb);
1403
1404         ip_cork_release(cork);
1405 }
1406
1407 void ip_flush_pending_frames(struct sock *sk)
1408 {
1409         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1410 }
1411
1412 struct sk_buff *ip_make_skb(struct sock *sk,
1413                             struct flowi4 *fl4,
1414                             int getfrag(void *from, char *to, int offset,
1415                                         int len, int odd, struct sk_buff *skb),
1416                             void *from, int length, int transhdrlen,
1417                             struct ipcm_cookie *ipc, struct rtable **rtp,
1418                             unsigned int flags)
1419 {
1420         struct inet_cork cork;
1421         struct sk_buff_head queue;
1422         int err;
1423
1424         if (flags & MSG_PROBE)
1425                 return NULL;
1426
1427         __skb_queue_head_init(&queue);
1428
1429         cork.flags = 0;
1430         cork.addr = 0;
1431         cork.opt = NULL;
1432         err = ip_setup_cork(sk, &cork, ipc, rtp);
1433         if (err)
1434                 return ERR_PTR(err);
1435
1436         err = __ip_append_data(sk, fl4, &queue, &cork, getfrag,
1437                                from, length, transhdrlen, flags);
1438         if (err) {
1439                 __ip_flush_pending_frames(sk, &queue, &cork);
1440                 return ERR_PTR(err);
1441         }
1442
1443         return __ip_make_skb(sk, fl4, &queue, &cork);
1444 }
1445
1446 /*
1447  *      Fetch data from kernel space and fill in checksum if needed.
1448  */
1449 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1450                               int len, int odd, struct sk_buff *skb)
1451 {
1452         __wsum csum;
1453
1454         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1455         skb->csum = csum_block_add(skb->csum, csum, odd);
1456         return 0;
1457 }
1458
1459 /*
1460  *      Generic function to send a packet as reply to another packet.
1461  *      Used to send TCP resets so far. ICMP should use this function too.
1462  *
1463  *      Should run single threaded per socket because it uses the sock
1464  *      structure to pass arguments.
1465  */
1466 void ip_send_reply(struct sock *sk, struct sk_buff *skb, __be32 daddr,
1467                    struct ip_reply_arg *arg, unsigned int len)
1468 {
1469         struct inet_sock *inet = inet_sk(sk);
1470         struct ip_options_data replyopts;
1471         struct ipcm_cookie ipc;
1472         struct flowi4 fl4;
1473         struct rtable *rt = skb_rtable(skb);
1474
1475         if (ip_options_echo(&replyopts.opt.opt, skb))
1476                 return;
1477
1478         ipc.addr = daddr;
1479         ipc.opt = NULL;
1480         ipc.tx_flags = 0;
1481
1482         if (replyopts.opt.opt.optlen) {
1483                 ipc.opt = &replyopts.opt;
1484
1485                 if (replyopts.opt.opt.srr)
1486                         daddr = replyopts.opt.opt.faddr;
1487         }
1488
1489         flowi4_init_output(&fl4, arg->bound_dev_if, 0,
1490                            RT_TOS(ip_hdr(skb)->tos),
1491                            RT_SCOPE_UNIVERSE, sk->sk_protocol,
1492                            ip_reply_arg_flowi_flags(arg),
1493                            daddr, rt->rt_spec_dst,
1494                            tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1495         security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1496         rt = ip_route_output_key(sock_net(sk), &fl4);
1497         if (IS_ERR(rt))
1498                 return;
1499
1500         /* And let IP do all the hard work.
1501
1502            This chunk is not reenterable, hence spinlock.
1503            Note that it uses the fact, that this function is called
1504            with locally disabled BH and that sk cannot be already spinlocked.
1505          */
1506         bh_lock_sock(sk);
1507         inet->tos = ip_hdr(skb)->tos;
1508         sk->sk_priority = skb->priority;
1509         sk->sk_protocol = ip_hdr(skb)->protocol;
1510         sk->sk_bound_dev_if = arg->bound_dev_if;
1511         ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1512                        &ipc, &rt, MSG_DONTWAIT);
1513         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1514                 if (arg->csumoffset >= 0)
1515                         *((__sum16 *)skb_transport_header(skb) +
1516                           arg->csumoffset) = csum_fold(csum_add(skb->csum,
1517                                                                 arg->csum));
1518                 skb->ip_summed = CHECKSUM_NONE;
1519                 ip_push_pending_frames(sk, &fl4);
1520         }
1521
1522         bh_unlock_sock(sk);
1523
1524         ip_rt_put(rt);
1525 }
1526
1527 void __init ip_init(void)
1528 {
1529         ip_rt_init();
1530         inet_initpeers();
1531
1532 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1533         igmp_mc_proc_init();
1534 #endif
1535 }