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                         struct sk_buff_head *queue,
 737                         int getfrag(void *from, char *to, int offset, int len,
 738                                int odd, struct sk_buff *skb),
 739                         void *from, int length, int hh_len, int fragheaderlen,
 740                         int transhdrlen, int mtu, unsigned int flags)
 741 {
 742         struct sk_buff *skb;
 743         int err;
 744
 745         /* There is support for UDP fragmentation offload by network
 746          * device, so create one single skb packet containing complete
 747          * udp datagram
 748          */
 749         if ((skb = skb_peek_tail(queue)) == NULL) {
 750                 skb = sock_alloc_send_skb(sk,
 751                         hh_len + fragheaderlen + transhdrlen + 20,
 752                         (flags & MSG_DONTWAIT), &err);
 753
 754                 if (skb == NULL)
 755                         return err;
 756
 757                 /* reserve space for Hardware header */
 758                 skb_reserve(skb, hh_len);
 759
 760                 /* create space for UDP/IP header */
 761                 skb_put(skb, fragheaderlen + transhdrlen);
 762
 763                 /* initialize network header pointer */
 764                 skb_reset_network_header(skb);
 765
 766                 /* initialize protocol header pointer */
 767                 skb->transport_header = skb->network_header + fragheaderlen;
 768
 769                 skb->ip_summed = CHECKSUM_PARTIAL;
 770                 skb->csum = 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(queue, skb);
 776         }
 777
 778         return skb_append_datato_frags(sk, skb, getfrag, from,
 779                                        (length - transhdrlen));
 780 }
 781
 782 static int __ip_append_data(struct sock *sk, struct sk_buff_head *queue,
 783                             struct inet_cork *cork,
 784                             int getfrag(void *from, char *to, int offset,
 785                                         int len, int odd, struct sk_buff *skb),
 786                             void *from, int length, int transhdrlen,
 787                             unsigned int flags)
 788 {
 789         struct inet_sock *inet = inet_sk(sk);
 790         struct sk_buff *skb;
 791
 792         struct ip_options *opt = inet->cork.opt;
 793         int hh_len;
 794         int exthdrlen;
 795         int mtu;
 796         int copy;
 797         int err;
 798         int offset = 0;
 799         unsigned int maxfraglen, fragheaderlen;
 800         int csummode = CHECKSUM_NONE;
 801         struct rtable *rt = (struct rtable *)cork->dst;
 802
 803         exthdrlen = transhdrlen ? rt->dst.header_len : 0;
 804         length += exthdrlen;
 805         transhdrlen += exthdrlen;
 806         mtu = inet->cork.fragsize;
 807
 808         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 809
 810         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 811         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 812
 813         if (cork->length + length > 0xFFFF - fragheaderlen) {
 814                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport,
 815                                mtu-exthdrlen);
 816                 return -EMSGSIZE;
 817         }
 818
 819         /*
 820          * transhdrlen > 0 means that this is the first fragment and we wish
 821          * it won't be fragmented in the future.
 822          */
 823         if (transhdrlen &&
 824             length + fragheaderlen <= mtu &&
 825             rt->dst.dev->features & NETIF_F_V4_CSUM &&
 826             !exthdrlen)
 827                 csummode = CHECKSUM_PARTIAL;
 828
 829         skb = skb_peek_tail(queue);
 830
 831         cork->length += length;
 832         if (((length > mtu) || (skb && skb_is_gso(skb))) &&
 833             (sk->sk_protocol == IPPROTO_UDP) &&
 834             (rt->dst.dev->features & NETIF_F_UFO)) {
 835                 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
 836                                          hh_len, fragheaderlen, transhdrlen,
 837                                          mtu, flags);
 838                 if (err)
 839                         goto error;
 840                 return 0;
 841         }
 842
 843         /* So, what's going on in the loop below?
 844          *
 845          * We use calculated fragment length to generate chained skb,
 846          * each of segments is IP fragment ready for sending to network after
 847          * adding appropriate IP header.
 848          */
 849
 850         if (!skb)
 851                 goto alloc_new_skb;
 852
 853         while (length > 0) {
 854                 /* Check if the remaining data fits into current packet. */
 855                 copy = mtu - skb->len;
 856                 if (copy < length)
 857                         copy = maxfraglen - skb->len;
 858                 if (copy <= 0) {
 859                         char *data;
 860                         unsigned int datalen;
 861                         unsigned int fraglen;
 862                         unsigned int fraggap;
 863                         unsigned int alloclen;
 864                         struct sk_buff *skb_prev;
 865 alloc_new_skb:
 866                         skb_prev = skb;
 867                         if (skb_prev)
 868                                 fraggap = skb_prev->len - maxfraglen;
 869                         else
 870                                 fraggap = 0;
 871
 872                         /*
 873                          * If remaining data exceeds the mtu,
 874                          * we know we need more fragment(s).
 875                          */
 876                         datalen = length + fraggap;
 877                         if (datalen > mtu - fragheaderlen)
 878                                 datalen = maxfraglen - fragheaderlen;
 879                         fraglen = datalen + fragheaderlen;
 880
 881                         if ((flags & MSG_MORE) &&
 882                             !(rt->dst.dev->features&NETIF_F_SG))
 883                                 alloclen = mtu;
 884                         else
 885                                 alloclen = fraglen;
 886
 887                         /* The last fragment gets additional space at tail.
 888                          * Note, with MSG_MORE we overallocate on fragments,
 889                          * because we have no idea what fragment will be
 890                          * the last.
 891                          */
 892                         if (datalen == length + fraggap) {
 893                                 alloclen += rt->dst.trailer_len;
 894                                 /* make sure mtu is not reached */
 895                                 if (datalen > mtu - fragheaderlen - rt->dst.trailer_len)
 896                                         datalen -= ALIGN(rt->dst.trailer_len, 8);
 897                         }
 898                         if (transhdrlen) {
 899                                 skb = sock_alloc_send_skb(sk,
 900                                                 alloclen + hh_len + 15,
 901                                                 (flags & MSG_DONTWAIT), &err);
 902                         } else {
 903                                 skb = NULL;
 904                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 905                                     2 * sk->sk_sndbuf)
 906                                         skb = sock_wmalloc(sk,
 907                                                            alloclen + hh_len + 15, 1,
 908                                                            sk->sk_allocation);
 909                                 if (unlikely(skb == NULL))
 910                                         err = -ENOBUFS;
 911                                 else
 912                                         /* only the initial fragment is
 913                                            time stamped */
 914                                         cork->tx_flags = 0;
 915                         }
 916                         if (skb == NULL)
 917                                 goto error;
 918
 919                         /*
 920                          *      Fill in the control structures
 921                          */
 922                         skb->ip_summed = csummode;
 923                         skb->csum = 0;
 924                         skb_reserve(skb, hh_len);
 925                         skb_shinfo(skb)->tx_flags = cork->tx_flags;
 926
 927                         /*
 928                          *      Find where to start putting bytes.
 929                          */
 930                         data = skb_put(skb, fraglen);
 931                         skb_set_network_header(skb, exthdrlen);
 932                         skb->transport_header = (skb->network_header +
 933                                                  fragheaderlen);
 934                         data += fragheaderlen;
 935
 936                         if (fraggap) {
 937                                 skb->csum = skb_copy_and_csum_bits(
 938                                         skb_prev, maxfraglen,
 939                                         data + transhdrlen, fraggap, 0);
 940                                 skb_prev->csum = csum_sub(skb_prev->csum,
 941                                                           skb->csum);
 942                                 data += fraggap;
 943                                 pskb_trim_unique(skb_prev, maxfraglen);
 944                         }
 945
 946                         copy = datalen - transhdrlen - fraggap;
 947                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
 948                                 err = -EFAULT;
 949                                 kfree_skb(skb);
 950                                 goto error;
 951                         }
 952
 953                         offset += copy;
 954                         length -= datalen - fraggap;
 955                         transhdrlen = 0;
 956                         exthdrlen = 0;
 957                         csummode = CHECKSUM_NONE;
 958
 959                         /*
 960                          * Put the packet on the pending queue.
 961                          */
 962                         __skb_queue_tail(queue, skb);
 963                         continue;
 964                 }
 965
 966                 if (copy > length)
 967                         copy = length;
 968
 969                 if (!(rt->dst.dev->features&NETIF_F_SG)) {
 970                         unsigned int off;
 971
 972                         off = skb->len;
 973                         if (getfrag(from, skb_put(skb, copy),
 974                                         offset, copy, off, skb) < 0) {
 975                                 __skb_trim(skb, off);
 976                                 err = -EFAULT;
 977                                 goto error;
 978                         }
 979                 } else {
 980                         int i = skb_shinfo(skb)->nr_frags;
 981                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
 982                         struct page *page = cork->page;
 983                         int off = cork->off;
 984                         unsigned int left;
 985
 986                         if (page && (left = PAGE_SIZE - off) > 0) {
 987                                 if (copy >= left)
 988                                         copy = left;
 989                                 if (page != frag->page) {
 990                                         if (i == MAX_SKB_FRAGS) {
 991                                                 err = -EMSGSIZE;
 992                                                 goto error;
 993                                         }
 994                                         get_page(page);
 995                                         skb_fill_page_desc(skb, i, page, off, 0);
 996                                         frag = &skb_shinfo(skb)->frags[i];
 997                                 }
 998                         } else if (i < MAX_SKB_FRAGS) {
 999                                 if (copy > PAGE_SIZE)
1000                                         copy = PAGE_SIZE;
1001                                 page = alloc_pages(sk->sk_allocation, 0);
1002                                 if (page == NULL)  {
1003                                         err = -ENOMEM;
1004                                         goto error;
1005                                 }
1006                                 cork->page = page;
1007                                 cork->off = 0;
1008
1009                                 skb_fill_page_desc(skb, i, page, 0, 0);
1010                                 frag = &skb_shinfo(skb)->frags[i];
1011                         } else {
1012                                 err = -EMSGSIZE;
1013                                 goto error;
1014                         }
1015                         if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1016                                 err = -EFAULT;
1017                                 goto error;
1018                         }
1019                         cork->off += copy;
1020                         frag->size += copy;
1021                         skb->len += copy;
1022                         skb->data_len += copy;
1023                         skb->truesize += copy;
1024                         atomic_add(copy, &sk->sk_wmem_alloc);
1025                 }
1026                 offset += copy;
1027                 length -= copy;
1028         }
1029
1030         return 0;
1031
1032 error:
1033         cork->length -= length;
1034         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1035         return err;
1036 }
1037
1038 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1039                          struct ipcm_cookie *ipc, struct rtable **rtp)
1040 {
1041         struct inet_sock *inet = inet_sk(sk);
1042         struct ip_options *opt;
1043         struct rtable *rt;
1044
1045         /*
1046          * setup for corking.
1047          */
1048         opt = ipc->opt;
1049         if (opt) {
1050                 if (cork->opt == NULL) {
1051                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1052                                             sk->sk_allocation);
1053                         if (unlikely(cork->opt == NULL))
1054                                 return -ENOBUFS;
1055                 }
1056                 memcpy(cork->opt, opt, sizeof(struct ip_options) + opt->optlen);
1057                 cork->flags |= IPCORK_OPT;
1058                 cork->addr = ipc->addr;
1059         }
1060         rt = *rtp;
1061         if (unlikely(!rt))
1062                 return -EFAULT;
1063         /*
1064          * We steal reference to this route, caller should not release it
1065          */
1066         *rtp = NULL;
1067         cork->fragsize = inet->pmtudisc == IP_PMTUDISC_PROBE ?
1068                          rt->dst.dev->mtu : dst_mtu(rt->dst.path);
1069         cork->dst = &rt->dst;
1070         cork->length = 0;
1071         cork->tx_flags = ipc->tx_flags;
1072         cork->page = NULL;
1073         cork->off = 0;
1074
1075         return 0;
1076 }
1077
1078 /*
1079  *      ip_append_data() and ip_append_page() can make one large IP datagram
1080  *      from many pieces of data. Each pieces will be holded on the socket
1081  *      until ip_push_pending_frames() is called. Each piece can be a page
1082  *      or non-page data.
1083  *
1084  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1085  *      this interface potentially.
1086  *
1087  *      LATER: length must be adjusted by pad at tail, when it is required.
1088  */
1089 int ip_append_data(struct sock *sk,
1090                    int getfrag(void *from, char *to, int offset, int len,
1091                                int odd, struct sk_buff *skb),
1092                    void *from, int length, int transhdrlen,
1093                    struct ipcm_cookie *ipc, struct rtable **rtp,
1094                    unsigned int flags)
1095 {
1096         struct inet_sock *inet = inet_sk(sk);
1097         int err;
1098
1099         if (flags&MSG_PROBE)
1100                 return 0;
1101
1102         if (skb_queue_empty(&sk->sk_write_queue)) {
1103                 err = ip_setup_cork(sk, &inet->cork, ipc, rtp);
1104                 if (err)
1105                         return err;
1106         } else {
1107                 transhdrlen = 0;
1108         }
1109
1110         return __ip_append_data(sk, &sk->sk_write_queue, &inet->cork, getfrag,
1111                                 from, length, transhdrlen, flags);
1112 }
1113
1114 ssize_t ip_append_page(struct sock *sk, struct page *page,
1115                        int offset, size_t size, int flags)
1116 {
1117         struct inet_sock *inet = inet_sk(sk);
1118         struct sk_buff *skb;
1119         struct rtable *rt;
1120         struct ip_options *opt = NULL;
1121         int hh_len;
1122         int mtu;
1123         int len;
1124         int err;
1125         unsigned int maxfraglen, fragheaderlen, fraggap;
1126
1127         if (inet->hdrincl)
1128                 return -EPERM;
1129
1130         if (flags&MSG_PROBE)
1131                 return 0;
1132
1133         if (skb_queue_empty(&sk->sk_write_queue))
1134                 return -EINVAL;
1135
1136         rt = (struct rtable *)inet->cork.dst;
1137         if (inet->cork.flags & IPCORK_OPT)
1138                 opt = inet->cork.opt;
1139
1140         if (!(rt->dst.dev->features&NETIF_F_SG))
1141                 return -EOPNOTSUPP;
1142
1143         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1144         mtu = inet->cork.fragsize;
1145
1146         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1147         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1148
1149         if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1150                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport, mtu);
1151                 return -EMSGSIZE;
1152         }
1153
1154         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1155                 return -EINVAL;
1156
1157         inet->cork.length += size;
1158         if ((size + skb->len > mtu) &&
1159             (sk->sk_protocol == IPPROTO_UDP) &&
1160             (rt->dst.dev->features & NETIF_F_UFO)) {
1161                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1162                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1163         }
1164
1165
1166         while (size > 0) {
1167                 int i;
1168
1169                 if (skb_is_gso(skb))
1170                         len = size;
1171                 else {
1172
1173                         /* Check if the remaining data fits into current packet. */
1174                         len = mtu - skb->len;
1175                         if (len < size)
1176                                 len = maxfraglen - skb->len;
1177                 }
1178                 if (len <= 0) {
1179                         struct sk_buff *skb_prev;
1180                         int alloclen;
1181
1182                         skb_prev = skb;
1183                         fraggap = skb_prev->len - maxfraglen;
1184
1185                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1186                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1187                         if (unlikely(!skb)) {
1188                                 err = -ENOBUFS;
1189                                 goto error;
1190                         }
1191
1192                         /*
1193                          *      Fill in the control structures
1194                          */
1195                         skb->ip_summed = CHECKSUM_NONE;
1196                         skb->csum = 0;
1197                         skb_reserve(skb, hh_len);
1198
1199                         /*
1200                          *      Find where to start putting bytes.
1201                          */
1202                         skb_put(skb, fragheaderlen + fraggap);
1203                         skb_reset_network_header(skb);
1204                         skb->transport_header = (skb->network_header +
1205                                                  fragheaderlen);
1206                         if (fraggap) {
1207                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1208                                                                    maxfraglen,
1209                                                     skb_transport_header(skb),
1210                                                                    fraggap, 0);
1211                                 skb_prev->csum = csum_sub(skb_prev->csum,
1212                                                           skb->csum);
1213                                 pskb_trim_unique(skb_prev, maxfraglen);
1214                         }
1215
1216                         /*
1217                          * Put the packet on the pending queue.
1218                          */
1219                         __skb_queue_tail(&sk->sk_write_queue, skb);
1220                         continue;
1221                 }
1222
1223                 i = skb_shinfo(skb)->nr_frags;
1224                 if (len > size)
1225                         len = size;
1226                 if (skb_can_coalesce(skb, i, page, offset)) {
1227                         skb_shinfo(skb)->frags[i-1].size += len;
1228                 } else if (i < MAX_SKB_FRAGS) {
1229                         get_page(page);
1230                         skb_fill_page_desc(skb, i, page, offset, len);
1231                 } else {
1232                         err = -EMSGSIZE;
1233                         goto error;
1234                 }
1235
1236                 if (skb->ip_summed == CHECKSUM_NONE) {
1237                         __wsum csum;
1238                         csum = csum_page(page, offset, len);
1239                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1240                 }
1241
1242                 skb->len += len;
1243                 skb->data_len += len;
1244                 skb->truesize += len;
1245                 atomic_add(len, &sk->sk_wmem_alloc);
1246                 offset += len;
1247                 size -= len;
1248         }
1249         return 0;
1250
1251 error:
1252         inet->cork.length -= size;
1253         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1254         return err;
1255 }
1256
1257 static void ip_cork_release(struct inet_cork *cork)
1258 {
1259         cork->flags &= ~IPCORK_OPT;
1260         kfree(cork->opt);
1261         cork->opt = NULL;
1262         dst_release(cork->dst);
1263         cork->dst = NULL;
1264 }
1265
1266 /*
1267  *      Combined all pending IP fragments on the socket as one IP datagram
1268  *      and push them out.
1269  */
1270 static int __ip_push_pending_frames(struct sock *sk,
1271                                     struct sk_buff_head *queue,
1272                                     struct inet_cork *cork)
1273 {
1274         struct sk_buff *skb, *tmp_skb;
1275         struct sk_buff **tail_skb;
1276         struct inet_sock *inet = inet_sk(sk);
1277         struct net *net = sock_net(sk);
1278         struct ip_options *opt = NULL;
1279         struct rtable *rt = (struct rtable *)cork->dst;
1280         struct iphdr *iph;
1281         __be16 df = 0;
1282         __u8 ttl;
1283         int err = 0;
1284
1285         if ((skb = __skb_dequeue(queue)) == NULL)
1286                 goto out;
1287         tail_skb = &(skb_shinfo(skb)->frag_list);
1288
1289         /* move skb->data to ip header from ext header */
1290         if (skb->data < skb_network_header(skb))
1291                 __skb_pull(skb, skb_network_offset(skb));
1292         while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1293                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1294                 *tail_skb = tmp_skb;
1295                 tail_skb = &(tmp_skb->next);
1296                 skb->len += tmp_skb->len;
1297                 skb->data_len += tmp_skb->len;
1298                 skb->truesize += tmp_skb->truesize;
1299                 tmp_skb->destructor = NULL;
1300                 tmp_skb->sk = NULL;
1301         }
1302
1303         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1304          * to fragment the frame generated here. No matter, what transforms
1305          * how transforms change size of the packet, it will come out.
1306          */
1307         if (inet->pmtudisc < IP_PMTUDISC_DO)
1308                 skb->local_df = 1;
1309
1310         /* DF bit is set when we want to see DF on outgoing frames.
1311          * If local_df is set too, we still allow to fragment this frame
1312          * locally. */
1313         if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1314             (skb->len <= dst_mtu(&rt->dst) &&
1315              ip_dont_fragment(sk, &rt->dst)))
1316                 df = htons(IP_DF);
1317
1318         if (cork->flags & IPCORK_OPT)
1319                 opt = cork->opt;
1320
1321         if (rt->rt_type == RTN_MULTICAST)
1322                 ttl = inet->mc_ttl;
1323         else
1324                 ttl = ip_select_ttl(inet, &rt->dst);
1325
1326         iph = (struct iphdr *)skb->data;
1327         iph->version = 4;
1328         iph->ihl = 5;
1329         if (opt) {
1330                 iph->ihl += opt->optlen>>2;
1331                 ip_options_build(skb, opt, cork->addr, rt, 0);
1332         }
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 = rt->rt_src;
1339         iph->daddr = rt->rt_dst;
1340
1341         skb->priority = sk->sk_priority;
1342         skb->mark = sk->sk_mark;
1343         /*
1344          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1345          * on dst refcount
1346          */
1347         cork->dst = NULL;
1348         skb_dst_set(skb, &rt->dst);
1349
1350         if (iph->protocol == IPPROTO_ICMP)
1351                 icmp_out_count(net, ((struct icmphdr *)
1352                         skb_transport_header(skb))->type);
1353
1354         /* Netfilter gets whole the not fragmented skb. */
1355         err = ip_local_out(skb);
1356         if (err) {
1357                 if (err > 0)
1358                         err = net_xmit_errno(err);
1359                 if (err)
1360                         goto error;
1361         }
1362
1363 out:
1364         ip_cork_release(cork);
1365         return err;
1366
1367 error:
1368         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1369         goto out;
1370 }
1371
1372 int ip_push_pending_frames(struct sock *sk)
1373 {
1374         return __ip_push_pending_frames(sk, &sk->sk_write_queue,
1375                                         &inet_sk(sk)->cork);
1376 }
1377
1378 /*
1379  *      Throw away all pending data on the socket.
1380  */
1381 static void __ip_flush_pending_frames(struct sock *sk,
1382                                       struct sk_buff_head *queue,
1383                                       struct inet_cork *cork)
1384 {
1385         struct sk_buff *skb;
1386
1387         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1388                 kfree_skb(skb);
1389
1390         ip_cork_release(cork);
1391 }
1392
1393 void ip_flush_pending_frames(struct sock *sk)
1394 {
1395         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork);
1396 }
1397
1398
1399 /*
1400  *      Fetch data from kernel space and fill in checksum if needed.
1401  */
1402 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1403                               int len, int odd, struct sk_buff *skb)
1404 {
1405         __wsum csum;
1406
1407         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1408         skb->csum = csum_block_add(skb->csum, csum, odd);
1409         return 0;
1410 }
1411
1412 /*
1413  *      Generic function to send a packet as reply to another packet.
1414  *      Used to send TCP resets so far. ICMP should use this function too.
1415  *
1416  *      Should run single threaded per socket because it uses the sock
1417  *      structure to pass arguments.
1418  */
1419 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1420                    unsigned int len)
1421 {
1422         struct inet_sock *inet = inet_sk(sk);
1423         struct {
1424                 struct ip_options       opt;
1425                 char                    data[40];
1426         } replyopts;
1427         struct ipcm_cookie ipc;
1428         __be32 daddr;
1429         struct rtable *rt = skb_rtable(skb);
1430
1431         if (ip_options_echo(&replyopts.opt, skb))
1432                 return;
1433
1434         daddr = ipc.addr = rt->rt_src;
1435         ipc.opt = NULL;
1436         ipc.tx_flags = 0;
1437
1438         if (replyopts.opt.optlen) {
1439                 ipc.opt = &replyopts.opt;
1440
1441                 if (ipc.opt->srr)
1442                         daddr = replyopts.opt.faddr;
1443         }
1444
1445         {
1446                 struct flowi fl = { .oif = arg->bound_dev_if,
1447                                     .fl4_dst = daddr,
1448                                     .fl4_src = rt->rt_spec_dst,
1449                                     .fl4_tos = RT_TOS(ip_hdr(skb)->tos),
1450                                     .fl_ip_sport = tcp_hdr(skb)->dest,
1451                                     .fl_ip_dport = tcp_hdr(skb)->source,
1452                                     .proto = sk->sk_protocol,
1453                                     .flags = ip_reply_arg_flowi_flags(arg) };
1454                 security_skb_classify_flow(skb, &fl);
1455                 if (ip_route_output_key(sock_net(sk), &rt, &fl))
1456                         return;
1457         }
1458
1459         /* And let IP do all the hard work.
1460
1461            This chunk is not reenterable, hence spinlock.
1462            Note that it uses the fact, that this function is called
1463            with locally disabled BH and that sk cannot be already spinlocked.
1464          */
1465         bh_lock_sock(sk);
1466         inet->tos = ip_hdr(skb)->tos;
1467         sk->sk_priority = skb->priority;
1468         sk->sk_protocol = ip_hdr(skb)->protocol;
1469         sk->sk_bound_dev_if = arg->bound_dev_if;
1470         ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1471                        &ipc, &rt, MSG_DONTWAIT);
1472         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1473                 if (arg->csumoffset >= 0)
1474                         *((__sum16 *)skb_transport_header(skb) +
1475                           arg->csumoffset) = csum_fold(csum_add(skb->csum,
1476                                                                 arg->csum));
1477                 skb->ip_summed = CHECKSUM_NONE;
1478                 ip_push_pending_frames(sk);
1479         }
1480
1481         bh_unlock_sock(sk);
1482
1483         ip_rt_put(rt);
1484 }
1485
1486 void __init ip_init(void)
1487 {
1488         ip_rt_init();
1489         inet_initpeers();
1490
1491 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1492         igmp_mc_proc_init();
1493 #endif
1494 }