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 <linux/module.h>
  47 #include <linux/types.h>
  48 #include <linux/kernel.h>
  49 #include <linux/mm.h>
  50 #include <linux/string.h>
  51 #include <linux/errno.h>
  52 #include <linux/highmem.h>
  53 #include <linux/slab.h>
  54
  55 #include <linux/socket.h>
  56 #include <linux/sockios.h>
  57 #include <linux/in.h>
  58 #include <linux/inet.h>
  59 #include <linux/netdevice.h>
  60 #include <linux/etherdevice.h>
  61 #include <linux/proc_fs.h>
  62 #include <linux/stat.h>
  63 #include <linux/init.h>
  64
  65 #include <net/snmp.h>
  66 #include <net/ip.h>
  67 #include <net/protocol.h>
  68 #include <net/route.h>
  69 #include <net/xfrm.h>
  70 #include <linux/skbuff.h>
  71 #include <net/sock.h>
  72 #include <net/arp.h>
  73 #include <net/icmp.h>
  74 #include <net/checksum.h>
  75 #include <net/inetpeer.h>
  76 #include <linux/igmp.h>
  77 #include <linux/netfilter_ipv4.h>
  78 #include <linux/netfilter_bridge.h>
  79 #include <linux/mroute.h>
  80 #include <linux/netlink.h>
  81 #include <linux/tcp.h>
  82
  83 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
  84 EXPORT_SYMBOL(sysctl_ip_default_ttl);
  85
  86 /* Generate a checksum for an outgoing IP datagram. */
  87 void ip_send_check(struct iphdr *iph)
  88 {
  89         iph->check = 0;
  90         iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
  91 }
  92 EXPORT_SYMBOL(ip_send_check);
  93
  94 int __ip_local_out(struct sk_buff *skb)
  95 {
  96         struct iphdr *iph = ip_hdr(skb);
  97
  98         iph->tot_len = htons(skb->len);
  99         ip_send_check(iph);
 100         return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
 101                        skb_dst(skb)->dev, dst_output);
 102 }
 103
 104 int ip_local_out(struct sk_buff *skb)
 105 {
 106         int err;
 107
 108         err = __ip_local_out(skb);
 109         if (likely(err == 1))
 110                 err = dst_output(skb);
 111
 112         return err;
 113 }
 114 EXPORT_SYMBOL_GPL(ip_local_out);
 115
 116 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
 117 {
 118         int ttl = inet->uc_ttl;
 119
 120         if (ttl < 0)
 121                 ttl = ip4_dst_hoplimit(dst);
 122         return ttl;
 123 }
 124
 125 /*
 126  *              Add an ip header to a skbuff and send it out.
 127  *
 128  */
 129 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
 130                           __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
 131 {
 132         struct inet_sock *inet = inet_sk(sk);
 133         struct rtable *rt = skb_rtable(skb);
 134         struct iphdr *iph;
 135
 136         /* Build the IP header. */
 137         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
 138         skb_reset_network_header(skb);
 139         iph = ip_hdr(skb);
 140         iph->version  = 4;
 141         iph->ihl      = 5;
 142         iph->tos      = inet->tos;
 143         if (ip_dont_fragment(sk, &rt->dst))
 144                 iph->frag_off = htons(IP_DF);
 145         else
 146                 iph->frag_off = 0;
 147         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 148         iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
 149         iph->saddr    = saddr;
 150         iph->protocol = sk->sk_protocol;
 151         ip_select_ident(skb, sk);
 152
 153         if (opt && opt->opt.optlen) {
 154                 iph->ihl += opt->opt.optlen>>2;
 155                 ip_options_build(skb, &opt->opt, daddr, rt, 0);
 156         }
 157
 158         skb->priority = sk->sk_priority;
 159         skb->mark = sk->sk_mark;
 160
 161         /* Send it out. */
 162         return ip_local_out(skb);
 163 }
 164 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
 165
 166 static inline int ip_finish_output2(struct sk_buff *skb)
 167 {
 168         struct dst_entry *dst = skb_dst(skb);
 169         struct rtable *rt = (struct rtable *)dst;
 170         struct net_device *dev = dst->dev;
 171         unsigned int hh_len = LL_RESERVED_SPACE(dev);
 172         struct neighbour *neigh;
 173         u32 nexthop;
 174
 175         if (rt->rt_type == RTN_MULTICAST) {
 176                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
 177         } else if (rt->rt_type == RTN_BROADCAST)
 178                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
 179
 180         /* Be paranoid, rather than too clever. */
 181         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 182                 struct sk_buff *skb2;
 183
 184                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 185                 if (skb2 == NULL) {
 186                         kfree_skb(skb);
 187                         return -ENOMEM;
 188                 }
 189                 if (skb->sk)
 190                         skb_set_owner_w(skb2, skb->sk);
 191                 consume_skb(skb);
 192                 skb = skb2;
 193         }
 194
 195         rcu_read_lock_bh();
 196         nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
 197         neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
 198         if (unlikely(!neigh))
 199                 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
 200         if (!IS_ERR(neigh)) {
 201                 int res = dst_neigh_output(dst, neigh, skb);
 202
 203                 rcu_read_unlock_bh();
 204                 return res;
 205         }
 206         rcu_read_unlock_bh();
 207
 208         net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
 209                             __func__);
 210         kfree_skb(skb);
 211         return -EINVAL;
 212 }
 213
 214 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
 215 {
 216         struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
 217
 218         return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
 219                skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
 220 }
 221
 222 static int ip_finish_output(struct sk_buff *skb)
 223 {
 224 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 225         /* Policy lookup after SNAT yielded a new policy */
 226         if (skb_dst(skb)->xfrm != NULL) {
 227                 IPCB(skb)->flags |= IPSKB_REROUTED;
 228                 return dst_output(skb);
 229         }
 230 #endif
 231         if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
 232                 return ip_fragment(skb, ip_finish_output2);
 233         else
 234                 return ip_finish_output2(skb);
 235 }
 236
 237 int ip_mc_output(struct sk_buff *skb)
 238 {
 239         struct sock *sk = skb->sk;
 240         struct rtable *rt = skb_rtable(skb);
 241         struct net_device *dev = rt->dst.dev;
 242
 243         /*
 244          *      If the indicated interface is up and running, send the packet.
 245          */
 246         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 247
 248         skb->dev = dev;
 249         skb->protocol = htons(ETH_P_IP);
 250
 251         /*
 252          *      Multicasts are looped back for other local users
 253          */
 254
 255         if (rt->rt_flags&RTCF_MULTICAST) {
 256                 if (sk_mc_loop(sk)
 257 #ifdef CONFIG_IP_MROUTE
 258                 /* Small optimization: do not loopback not local frames,
 259                    which returned after forwarding; they will be  dropped
 260                    by ip_mr_input in any case.
 261                    Note, that local frames are looped back to be delivered
 262                    to local recipients.
 263
 264                    This check is duplicated in ip_mr_input at the moment.
 265                  */
 266                     &&
 267                     ((rt->rt_flags & RTCF_LOCAL) ||
 268                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 269 #endif
 270                    ) {
 271                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 272                         if (newskb)
 273                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 274                                         newskb, NULL, newskb->dev,
 275                                         dev_loopback_xmit);
 276                 }
 277
 278                 /* Multicasts with ttl 0 must not go beyond the host */
 279
 280                 if (ip_hdr(skb)->ttl == 0) {
 281                         kfree_skb(skb);
 282                         return 0;
 283                 }
 284         }
 285
 286         if (rt->rt_flags&RTCF_BROADCAST) {
 287                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 288                 if (newskb)
 289                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
 290                                 NULL, newskb->dev, dev_loopback_xmit);
 291         }
 292
 293         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
 294                             skb->dev, ip_finish_output,
 295                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 296 }
 297
 298 int ip_output(struct sk_buff *skb)
 299 {
 300         struct net_device *dev = skb_dst(skb)->dev;
 301
 302         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 303
 304         skb->dev = dev;
 305         skb->protocol = htons(ETH_P_IP);
 306
 307         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
 308                             ip_finish_output,
 309                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 310 }
 311
 312 /*
 313  * copy saddr and daddr, possibly using 64bit load/stores
 314  * Equivalent to :
 315  *   iph->saddr = fl4->saddr;
 316  *   iph->daddr = fl4->daddr;
 317  */
 318 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
 319 {
 320         BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
 321                      offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
 322         memcpy(&iph->saddr, &fl4->saddr,
 323                sizeof(fl4->saddr) + sizeof(fl4->daddr));
 324 }
 325
 326 int ip_queue_xmit(struct sk_buff *skb, struct flowi *fl)
 327 {
 328         struct sock *sk = skb->sk;
 329         struct inet_sock *inet = inet_sk(sk);
 330         struct ip_options_rcu *inet_opt;
 331         struct flowi4 *fl4;
 332         struct rtable *rt;
 333         struct iphdr *iph;
 334         int res;
 335
 336         /* Skip all of this if the packet is already routed,
 337          * f.e. by something like SCTP.
 338          */
 339         rcu_read_lock();
 340         inet_opt = rcu_dereference(inet->inet_opt);
 341         fl4 = &fl->u.ip4;
 342         rt = skb_rtable(skb);
 343         if (rt != NULL)
 344                 goto packet_routed;
 345
 346         /* Make sure we can route this packet. */
 347         rt = (struct rtable *)__sk_dst_check(sk, 0);
 348         if (rt == NULL) {
 349                 __be32 daddr;
 350
 351                 /* Use correct destination address if we have options. */
 352                 daddr = inet->inet_daddr;
 353                 if (inet_opt && inet_opt->opt.srr)
 354                         daddr = inet_opt->opt.faddr;
 355
 356                 /* If this fails, retransmit mechanism of transport layer will
 357                  * keep trying until route appears or the connection times
 358                  * itself out.
 359                  */
 360                 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
 361                                            daddr, inet->inet_saddr,
 362                                            inet->inet_dport,
 363                                            inet->inet_sport,
 364                                            sk->sk_protocol,
 365                                            RT_CONN_FLAGS(sk),
 366                                            sk->sk_bound_dev_if);
 367                 if (IS_ERR(rt))
 368                         goto no_route;
 369                 sk_setup_caps(sk, &rt->dst);
 370         }
 371         skb_dst_set_noref(skb, &rt->dst);
 372
 373 packet_routed:
 374         if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
 375                 goto no_route;
 376
 377         /* OK, we know where to send it, allocate and build IP header. */
 378         skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
 379         skb_reset_network_header(skb);
 380         iph = ip_hdr(skb);
 381         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 382         if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
 383                 iph->frag_off = htons(IP_DF);
 384         else
 385                 iph->frag_off = 0;
 386         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 387         iph->protocol = sk->sk_protocol;
 388         ip_copy_addrs(iph, fl4);
 389
 390         /* Transport layer set skb->h.foo itself. */
 391
 392         if (inet_opt && inet_opt->opt.optlen) {
 393                 iph->ihl += inet_opt->opt.optlen >> 2;
 394                 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
 395         }
 396
 397         ip_select_ident_segs(skb, sk, skb_shinfo(skb)->gso_segs ?: 1);
 398
 399         skb->priority = sk->sk_priority;
 400         skb->mark = sk->sk_mark;
 401
 402         res = ip_local_out(skb);
 403         rcu_read_unlock();
 404         return res;
 405
 406 no_route:
 407         rcu_read_unlock();
 408         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 409         kfree_skb(skb);
 410         return -EHOSTUNREACH;
 411 }
 412 EXPORT_SYMBOL(ip_queue_xmit);
 413
 414
 415 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 416 {
 417         to->pkt_type = from->pkt_type;
 418         to->priority = from->priority;
 419         to->protocol = from->protocol;
 420         skb_dst_drop(to);
 421         skb_dst_copy(to, from);
 422         to->dev = from->dev;
 423         to->mark = from->mark;
 424
 425         /* Copy the flags to each fragment. */
 426         IPCB(to)->flags = IPCB(from)->flags;
 427
 428 #ifdef CONFIG_NET_SCHED
 429         to->tc_index = from->tc_index;
 430 #endif
 431         nf_copy(to, from);
 432 #if IS_ENABLED(CONFIG_NETFILTER_XT_TARGET_TRACE)
 433         to->nf_trace = from->nf_trace;
 434 #endif
 435 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 436         to->ipvs_property = from->ipvs_property;
 437 #endif
 438         skb_copy_secmark(to, from);
 439 }
 440
 441 /*
 442  *      This IP datagram is too large to be sent in one piece.  Break it up into
 443  *      smaller pieces (each of size equal to IP header plus
 444  *      a block of the data of the original IP data part) that will yet fit in a
 445  *      single device frame, and queue such a frame for sending.
 446  */
 447
 448 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 449 {
 450         struct iphdr *iph;
 451         int ptr;
 452         struct net_device *dev;
 453         struct sk_buff *skb2;
 454         unsigned int mtu, hlen, left, len, ll_rs;
 455         int offset;
 456         __be16 not_last_frag;
 457         struct rtable *rt = skb_rtable(skb);
 458         int err = 0;
 459
 460         dev = rt->dst.dev;
 461
 462         /*
 463          *      Point into the IP datagram header.
 464          */
 465
 466         iph = ip_hdr(skb);
 467
 468         if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->local_df) ||
 469                      (IPCB(skb)->frag_max_size &&
 470                       IPCB(skb)->frag_max_size > dst_mtu(&rt->dst)))) {
 471                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 472                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 473                           htonl(ip_skb_dst_mtu(skb)));
 474                 kfree_skb(skb);
 475                 return -EMSGSIZE;
 476         }
 477
 478         /*
 479          *      Setup starting values.
 480          */
 481
 482         hlen = iph->ihl * 4;
 483         mtu = dst_mtu(&rt->dst) - hlen; /* Size of data space */
 484 #ifdef CONFIG_BRIDGE_NETFILTER
 485         if (skb->nf_bridge)
 486                 mtu -= nf_bridge_mtu_reduction(skb);
 487 #endif
 488         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 489
 490         /* When frag_list is given, use it. First, check its validity:
 491          * some transformers could create wrong frag_list or break existing
 492          * one, it is not prohibited. In this case fall back to copying.
 493          *
 494          * LATER: this step can be merged to real generation of fragments,
 495          * we can switch to copy when see the first bad fragment.
 496          */
 497         if (skb_has_frag_list(skb)) {
 498                 struct sk_buff *frag, *frag2;
 499                 int first_len = skb_pagelen(skb);
 500
 501                 if (first_len - hlen > mtu ||
 502                     ((first_len - hlen) & 7) ||
 503                     ip_is_fragment(iph) ||
 504                     skb_cloned(skb))
 505                         goto slow_path;
 506
 507                 skb_walk_frags(skb, frag) {
 508                         /* Correct geometry. */
 509                         if (frag->len > mtu ||
 510                             ((frag->len & 7) && frag->next) ||
 511                             skb_headroom(frag) < hlen)
 512                                 goto slow_path_clean;
 513
 514                         /* Partially cloned skb? */
 515                         if (skb_shared(frag))
 516                                 goto slow_path_clean;
 517
 518                         BUG_ON(frag->sk);
 519                         if (skb->sk) {
 520                                 frag->sk = skb->sk;
 521                                 frag->destructor = sock_wfree;
 522                         }
 523                         skb->truesize -= frag->truesize;
 524                 }
 525
 526                 /* Everything is OK. Generate! */
 527
 528                 err = 0;
 529                 offset = 0;
 530                 frag = skb_shinfo(skb)->frag_list;
 531                 skb_frag_list_init(skb);
 532                 skb->data_len = first_len - skb_headlen(skb);
 533                 skb->len = first_len;
 534                 iph->tot_len = htons(first_len);
 535                 iph->frag_off = htons(IP_MF);
 536                 ip_send_check(iph);
 537
 538                 for (;;) {
 539                         /* Prepare header of the next frame,
 540                          * before previous one went down. */
 541                         if (frag) {
 542                                 frag->ip_summed = CHECKSUM_NONE;
 543                                 skb_reset_transport_header(frag);
 544                                 __skb_push(frag, hlen);
 545                                 skb_reset_network_header(frag);
 546                                 memcpy(skb_network_header(frag), iph, hlen);
 547                                 iph = ip_hdr(frag);
 548                                 iph->tot_len = htons(frag->len);
 549                                 ip_copy_metadata(frag, skb);
 550                                 if (offset == 0)
 551                                         ip_options_fragment(frag);
 552                                 offset += skb->len - hlen;
 553                                 iph->frag_off = htons(offset>>3);
 554                                 if (frag->next != NULL)
 555                                         iph->frag_off |= htons(IP_MF);
 556                                 /* Ready, complete checksum */
 557                                 ip_send_check(iph);
 558                         }
 559
 560                         err = output(skb);
 561
 562                         if (!err)
 563                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 564                         if (err || !frag)
 565                                 break;
 566
 567                         skb = frag;
 568                         frag = skb->next;
 569                         skb->next = NULL;
 570                 }
 571
 572                 if (err == 0) {
 573                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 574                         return 0;
 575                 }
 576
 577                 while (frag) {
 578                         skb = frag->next;
 579                         kfree_skb(frag);
 580                         frag = skb;
 581                 }
 582                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 583                 return err;
 584
 585 slow_path_clean:
 586                 skb_walk_frags(skb, frag2) {
 587                         if (frag2 == frag)
 588                                 break;
 589                         frag2->sk = NULL;
 590                         frag2->destructor = NULL;
 591                         skb->truesize += frag2->truesize;
 592                 }
 593         }
 594
 595 slow_path:
 596         /* for offloaded checksums cleanup checksum before fragmentation */
 597         if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
 598                 goto fail;
 599         iph = ip_hdr(skb);
 600
 601         left = skb->len - hlen;         /* Space per frame */
 602         ptr = hlen;             /* Where to start from */
 603
 604         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 605          * we need to make room for the encapsulating header
 606          */
 607         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
 608
 609         /*
 610          *      Fragment the datagram.
 611          */
 612
 613         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 614         not_last_frag = iph->frag_off & htons(IP_MF);
 615
 616         /*
 617          *      Keep copying data until we run out.
 618          */
 619
 620         while (left > 0) {
 621                 len = left;
 622                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 623                 if (len > mtu)
 624                         len = mtu;
 625                 /* IF: we are not sending up to and including the packet end
 626                    then align the next start on an eight byte boundary */
 627                 if (len < left) {
 628                         len &= ~7;
 629                 }
 630                 /*
 631                  *      Allocate buffer.
 632                  */
 633
 634                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 635                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 636                         err = -ENOMEM;
 637                         goto fail;
 638                 }
 639
 640                 /*
 641                  *      Set up data on packet
 642                  */
 643
 644                 ip_copy_metadata(skb2, skb);
 645                 skb_reserve(skb2, ll_rs);
 646                 skb_put(skb2, len + hlen);
 647                 skb_reset_network_header(skb2);
 648                 skb2->transport_header = skb2->network_header + hlen;
 649
 650                 /*
 651                  *      Charge the memory for the fragment to any owner
 652                  *      it might possess
 653                  */
 654
 655                 if (skb->sk)
 656                         skb_set_owner_w(skb2, skb->sk);
 657
 658                 /*
 659                  *      Copy the packet header into the new buffer.
 660                  */
 661
 662                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 663
 664                 /*
 665                  *      Copy a block of the IP datagram.
 666                  */
 667                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 668                         BUG();
 669                 left -= len;
 670
 671                 /*
 672                  *      Fill in the new header fields.
 673                  */
 674                 iph = ip_hdr(skb2);
 675                 iph->frag_off = htons((offset >> 3));
 676
 677                 /* ANK: dirty, but effective trick. Upgrade options only if
 678                  * the segment to be fragmented was THE FIRST (otherwise,
 679                  * options are already fixed) and make it ONCE
 680                  * on the initial skb, so that all the following fragments
 681                  * will inherit fixed options.
 682                  */
 683                 if (offset == 0)
 684                         ip_options_fragment(skb);
 685
 686                 /*
 687                  *      Added AC : If we are fragmenting a fragment that's not the
 688                  *                 last fragment then keep MF on each bit
 689                  */
 690                 if (left > 0 || not_last_frag)
 691                         iph->frag_off |= htons(IP_MF);
 692                 ptr += len;
 693                 offset += len;
 694
 695                 /*
 696                  *      Put this fragment into the sending queue.
 697                  */
 698                 iph->tot_len = htons(len + hlen);
 699
 700                 ip_send_check(iph);
 701
 702                 err = output(skb2);
 703                 if (err)
 704                         goto fail;
 705
 706                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 707         }
 708         consume_skb(skb);
 709         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 710         return err;
 711
 712 fail:
 713         kfree_skb(skb);
 714         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 715         return err;
 716 }
 717 EXPORT_SYMBOL(ip_fragment);
 718
 719 int
 720 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 721 {
 722         struct iovec *iov = from;
 723
 724         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 725                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 726                         return -EFAULT;
 727         } else {
 728                 __wsum csum = 0;
 729                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 730                         return -EFAULT;
 731                 skb->csum = csum_block_add(skb->csum, csum, odd);
 732         }
 733         return 0;
 734 }
 735 EXPORT_SYMBOL(ip_generic_getfrag);
 736
 737 static inline __wsum
 738 csum_page(struct page *page, int offset, int copy)
 739 {
 740         char *kaddr;
 741         __wsum csum;
 742         kaddr = kmap(page);
 743         csum = csum_partial(kaddr + offset, copy, 0);
 744         kunmap(page);
 745         return csum;
 746 }
 747
 748 static inline int ip_ufo_append_data(struct sock *sk,
 749                         struct sk_buff_head *queue,
 750                         int getfrag(void *from, char *to, int offset, int len,
 751                                int odd, struct sk_buff *skb),
 752                         void *from, int length, int hh_len, int fragheaderlen,
 753                         int transhdrlen, int maxfraglen, unsigned int flags)
 754 {
 755         struct sk_buff *skb;
 756         int err;
 757
 758         /* There is support for UDP fragmentation offload by network
 759          * device, so create one single skb packet containing complete
 760          * udp datagram
 761          */
 762         if ((skb = skb_peek_tail(queue)) == NULL) {
 763                 skb = sock_alloc_send_skb(sk,
 764                         hh_len + fragheaderlen + transhdrlen + 20,
 765                         (flags & MSG_DONTWAIT), &err);
 766
 767                 if (skb == NULL)
 768                         return err;
 769
 770                 /* reserve space for Hardware header */
 771                 skb_reserve(skb, hh_len);
 772
 773                 /* create space for UDP/IP header */
 774                 skb_put(skb, fragheaderlen + transhdrlen);
 775
 776                 /* initialize network header pointer */
 777                 skb_reset_network_header(skb);
 778
 779                 /* initialize protocol header pointer */
 780                 skb->transport_header = skb->network_header + fragheaderlen;
 781
 782                 skb->ip_summed = CHECKSUM_PARTIAL;
 783                 skb->csum = 0;
 784
 785                 /* specify the length of each IP datagram fragment */
 786                 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
 787                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 788                 __skb_queue_tail(queue, skb);
 789         }
 790
 791         return skb_append_datato_frags(sk, skb, getfrag, from,
 792                                        (length - transhdrlen));
 793 }
 794
 795 static int __ip_append_data(struct sock *sk,
 796                             struct flowi4 *fl4,
 797                             struct sk_buff_head *queue,
 798                             struct inet_cork *cork,
 799                             struct page_frag *pfrag,
 800                             int getfrag(void *from, char *to, int offset,
 801                                         int len, int odd, struct sk_buff *skb),
 802                             void *from, int length, int transhdrlen,
 803                             unsigned int flags)
 804 {
 805         struct inet_sock *inet = inet_sk(sk);
 806         struct sk_buff *skb;
 807
 808         struct ip_options *opt = cork->opt;
 809         int hh_len;
 810         int exthdrlen;
 811         int mtu;
 812         int copy;
 813         int err;
 814         int offset = 0;
 815         unsigned int maxfraglen, fragheaderlen;
 816         int csummode = CHECKSUM_NONE;
 817         struct rtable *rt = (struct rtable *)cork->dst;
 818
 819         skb = skb_peek_tail(queue);
 820
 821         exthdrlen = !skb ? rt->dst.header_len : 0;
 822         mtu = cork->fragsize;
 823
 824         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 825
 826         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 827         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 828
 829         if (cork->length + length > 0xFFFF - fragheaderlen) {
 830                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
 831                                mtu-exthdrlen);
 832                 return -EMSGSIZE;
 833         }
 834
 835         /*
 836          * transhdrlen > 0 means that this is the first fragment and we wish
 837          * it won't be fragmented in the future.
 838          */
 839         if (transhdrlen &&
 840             length + fragheaderlen <= mtu &&
 841             rt->dst.dev->features & NETIF_F_V4_CSUM &&
 842             !exthdrlen)
 843                 csummode = CHECKSUM_PARTIAL;
 844
 845         cork->length += length;
 846         if (((length > mtu) || (skb && skb_has_frags(skb))) &&
 847             (sk->sk_protocol == IPPROTO_UDP) &&
 848             (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len) {
 849                 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
 850                                          hh_len, fragheaderlen, transhdrlen,
 851                                          maxfraglen, flags);
 852                 if (err)
 853                         goto error;
 854                 return 0;
 855         }
 856
 857         /* So, what's going on in the loop below?
 858          *
 859          * We use calculated fragment length to generate chained skb,
 860          * each of segments is IP fragment ready for sending to network after
 861          * adding appropriate IP header.
 862          */
 863
 864         if (!skb)
 865                 goto alloc_new_skb;
 866
 867         while (length > 0) {
 868                 /* Check if the remaining data fits into current packet. */
 869                 copy = mtu - skb->len;
 870                 if (copy < length)
 871                         copy = maxfraglen - skb->len;
 872                 if (copy <= 0) {
 873                         char *data;
 874                         unsigned int datalen;
 875                         unsigned int fraglen;
 876                         unsigned int fraggap;
 877                         unsigned int alloclen;
 878                         struct sk_buff *skb_prev;
 879 alloc_new_skb:
 880                         skb_prev = skb;
 881                         if (skb_prev)
 882                                 fraggap = skb_prev->len - maxfraglen;
 883                         else
 884                                 fraggap = 0;
 885
 886                         /*
 887                          * If remaining data exceeds the mtu,
 888                          * we know we need more fragment(s).
 889                          */
 890                         datalen = length + fraggap;
 891                         if (datalen > mtu - fragheaderlen)
 892                                 datalen = maxfraglen - fragheaderlen;
 893                         fraglen = datalen + fragheaderlen;
 894
 895                         if ((flags & MSG_MORE) &&
 896                             !(rt->dst.dev->features&NETIF_F_SG))
 897                                 alloclen = mtu;
 898                         else
 899                                 alloclen = fraglen;
 900
 901                         alloclen += exthdrlen;
 902
 903                         /* The last fragment gets additional space at tail.
 904                          * Note, with MSG_MORE we overallocate on fragments,
 905                          * because we have no idea what fragment will be
 906                          * the last.
 907                          */
 908                         if (datalen == length + fraggap)
 909                                 alloclen += rt->dst.trailer_len;
 910
 911                         if (transhdrlen) {
 912                                 skb = sock_alloc_send_skb(sk,
 913                                                 alloclen + hh_len + 15,
 914                                                 (flags & MSG_DONTWAIT), &err);
 915                         } else {
 916                                 skb = NULL;
 917                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 918                                     2 * sk->sk_sndbuf)
 919                                         skb = sock_wmalloc(sk,
 920                                                            alloclen + hh_len + 15, 1,
 921                                                            sk->sk_allocation);
 922                                 if (unlikely(skb == NULL))
 923                                         err = -ENOBUFS;
 924                                 else
 925                                         /* only the initial fragment is
 926                                            time stamped */
 927                                         cork->tx_flags = 0;
 928                         }
 929                         if (skb == NULL)
 930                                 goto error;
 931
 932                         /*
 933                          *      Fill in the control structures
 934                          */
 935                         skb->ip_summed = csummode;
 936                         skb->csum = 0;
 937                         skb_reserve(skb, hh_len);
 938                         skb_shinfo(skb)->tx_flags = cork->tx_flags;
 939
 940                         /*
 941                          *      Find where to start putting bytes.
 942                          */
 943                         data = skb_put(skb, fraglen + exthdrlen);
 944                         skb_set_network_header(skb, exthdrlen);
 945                         skb->transport_header = (skb->network_header +
 946                                                  fragheaderlen);
 947                         data += fragheaderlen + exthdrlen;
 948
 949                         if (fraggap) {
 950                                 skb->csum = skb_copy_and_csum_bits(
 951                                         skb_prev, maxfraglen,
 952                                         data + transhdrlen, fraggap, 0);
 953                                 skb_prev->csum = csum_sub(skb_prev->csum,
 954                                                           skb->csum);
 955                                 data += fraggap;
 956                                 pskb_trim_unique(skb_prev, maxfraglen);
 957                         }
 958
 959                         copy = datalen - transhdrlen - fraggap;
 960                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
 961                                 err = -EFAULT;
 962                                 kfree_skb(skb);
 963                                 goto error;
 964                         }
 965
 966                         offset += copy;
 967                         length -= datalen - fraggap;
 968                         transhdrlen = 0;
 969                         exthdrlen = 0;
 970                         csummode = CHECKSUM_NONE;
 971
 972                         /*
 973                          * Put the packet on the pending queue.
 974                          */
 975                         __skb_queue_tail(queue, skb);
 976                         continue;
 977                 }
 978
 979                 if (copy > length)
 980                         copy = length;
 981
 982                 if (!(rt->dst.dev->features&NETIF_F_SG)) {
 983                         unsigned int off;
 984
 985                         off = skb->len;
 986                         if (getfrag(from, skb_put(skb, copy),
 987                                         offset, copy, off, skb) < 0) {
 988                                 __skb_trim(skb, off);
 989                                 err = -EFAULT;
 990                                 goto error;
 991                         }
 992                 } else {
 993                         int i = skb_shinfo(skb)->nr_frags;
 994
 995                         err = -ENOMEM;
 996                         if (!sk_page_frag_refill(sk, pfrag))
 997                                 goto error;
 998
 999                         if (!skb_can_coalesce(skb, i, pfrag->page,
1000                                               pfrag->offset)) {
1001                                 err = -EMSGSIZE;
1002                                 if (i == MAX_SKB_FRAGS)
1003                                         goto error;
1004
1005                                 __skb_fill_page_desc(skb, i, pfrag->page,
1006                                                      pfrag->offset, 0);
1007                                 skb_shinfo(skb)->nr_frags = ++i;
1008                                 get_page(pfrag->page);
1009                         }
1010                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1011                         if (getfrag(from,
1012                                     page_address(pfrag->page) + pfrag->offset,
1013                                     offset, copy, skb->len, skb) < 0)
1014                                 goto error_efault;
1015
1016                         pfrag->offset += copy;
1017                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1018                         skb->len += copy;
1019                         skb->data_len += copy;
1020                         skb->truesize += copy;
1021                         atomic_add(copy, &sk->sk_wmem_alloc);
1022                 }
1023                 offset += copy;
1024                 length -= copy;
1025         }
1026
1027         return 0;
1028
1029 error_efault:
1030         err = -EFAULT;
1031 error:
1032         cork->length -= length;
1033         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1034         return err;
1035 }
1036
1037 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1038                          struct ipcm_cookie *ipc, struct rtable **rtp)
1039 {
1040         struct inet_sock *inet = inet_sk(sk);
1041         struct ip_options_rcu *opt;
1042         struct rtable *rt;
1043
1044         /*
1045          * setup for corking.
1046          */
1047         opt = ipc->opt;
1048         if (opt) {
1049                 if (cork->opt == NULL) {
1050                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1051                                             sk->sk_allocation);
1052                         if (unlikely(cork->opt == NULL))
1053                                 return -ENOBUFS;
1054                 }
1055                 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1056                 cork->flags |= IPCORK_OPT;
1057                 cork->addr = ipc->addr;
1058         }
1059         rt = *rtp;
1060         if (unlikely(!rt))
1061                 return -EFAULT;
1062         /*
1063          * We steal reference to this route, caller should not release it
1064          */
1065         *rtp = NULL;
1066         cork->fragsize = inet->pmtudisc == IP_PMTUDISC_PROBE ?
1067                          rt->dst.dev->mtu : dst_mtu(&rt->dst);
1068         cork->dst = &rt->dst;
1069         cork->length = 0;
1070         cork->tx_flags = ipc->tx_flags;
1071
1072         return 0;
1073 }
1074
1075 /*
1076  *      ip_append_data() and ip_append_page() can make one large IP datagram
1077  *      from many pieces of data. Each pieces will be holded on the socket
1078  *      until ip_push_pending_frames() is called. Each piece can be a page
1079  *      or non-page data.
1080  *
1081  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1082  *      this interface potentially.
1083  *
1084  *      LATER: length must be adjusted by pad at tail, when it is required.
1085  */
1086 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1087                    int getfrag(void *from, char *to, int offset, int len,
1088                                int odd, struct sk_buff *skb),
1089                    void *from, int length, int transhdrlen,
1090                    struct ipcm_cookie *ipc, struct rtable **rtp,
1091                    unsigned int flags)
1092 {
1093         struct inet_sock *inet = inet_sk(sk);
1094         int err;
1095
1096         if (flags&MSG_PROBE)
1097                 return 0;
1098
1099         if (skb_queue_empty(&sk->sk_write_queue)) {
1100                 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1101                 if (err)
1102                         return err;
1103         } else {
1104                 transhdrlen = 0;
1105         }
1106
1107         return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1108                                 sk_page_frag(sk), getfrag,
1109                                 from, length, transhdrlen, flags);
1110 }
1111
1112 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1113                        int offset, size_t size, int flags)
1114 {
1115         struct inet_sock *inet = inet_sk(sk);
1116         struct sk_buff *skb;
1117         struct rtable *rt;
1118         struct ip_options *opt = NULL;
1119         struct inet_cork *cork;
1120         int hh_len;
1121         int mtu;
1122         int len;
1123         int err;
1124         unsigned int maxfraglen, fragheaderlen, fraggap;
1125
1126         if (inet->hdrincl)
1127                 return -EPERM;
1128
1129         if (flags&MSG_PROBE)
1130                 return 0;
1131
1132         if (skb_queue_empty(&sk->sk_write_queue))
1133                 return -EINVAL;
1134
1135         cork = &inet->cork.base;
1136         rt = (struct rtable *)cork->dst;
1137         if (cork->flags & IPCORK_OPT)
1138                 opt = 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 = cork->fragsize;
1145
1146         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1147         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1148
1149         if (cork->length + size > 0xFFFF - fragheaderlen) {
1150                 ip_local_error(sk, EMSGSIZE, fl4->daddr, 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         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_frag_size_add(&skb_shinfo(skb)->frags[i-1], 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         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 struct sk_buff *__ip_make_skb(struct sock *sk,
1271                               struct flowi4 *fl4,
1272                               struct sk_buff_head *queue,
1273                               struct inet_cork *cork)
1274 {
1275         struct sk_buff *skb, *tmp_skb;
1276         struct sk_buff **tail_skb;
1277         struct inet_sock *inet = inet_sk(sk);
1278         struct net *net = sock_net(sk);
1279         struct ip_options *opt = NULL;
1280         struct rtable *rt = (struct rtable *)cork->dst;
1281         struct iphdr *iph;
1282         __be16 df = 0;
1283         __u8 ttl;
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 = ip_hdr(skb);
1327         iph->version = 4;
1328         iph->ihl = 5;
1329         iph->tos = inet->tos;
1330         iph->frag_off = df;
1331         iph->ttl = ttl;
1332         iph->protocol = sk->sk_protocol;
1333         ip_copy_addrs(iph, fl4);
1334         ip_select_ident(skb, sk);
1335
1336         if (opt) {
1337                 iph->ihl += opt->optlen>>2;
1338                 ip_options_build(skb, opt, cork->addr, rt, 0);
1339         }
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         ip_cork_release(cork);
1355 out:
1356         return skb;
1357 }
1358
1359 int ip_send_skb(struct net *net, struct sk_buff *skb)
1360 {
1361         int err;
1362
1363         err = ip_local_out(skb);
1364         if (err) {
1365                 if (err > 0)
1366                         err = net_xmit_errno(err);
1367                 if (err)
1368                         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1369         }
1370
1371         return err;
1372 }
1373
1374 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1375 {
1376         struct sk_buff *skb;
1377
1378         skb = ip_finish_skb(sk, fl4);
1379         if (!skb)
1380                 return 0;
1381
1382         /* Netfilter gets whole the not fragmented skb. */
1383         return ip_send_skb(sock_net(sk), skb);
1384 }
1385
1386 /*
1387  *      Throw away all pending data on the socket.
1388  */
1389 static void __ip_flush_pending_frames(struct sock *sk,
1390                                       struct sk_buff_head *queue,
1391                                       struct inet_cork *cork)
1392 {
1393         struct sk_buff *skb;
1394
1395         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1396                 kfree_skb(skb);
1397
1398         ip_cork_release(cork);
1399 }
1400
1401 void ip_flush_pending_frames(struct sock *sk)
1402 {
1403         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1404 }
1405
1406 struct sk_buff *ip_make_skb(struct sock *sk,
1407                             struct flowi4 *fl4,
1408                             int getfrag(void *from, char *to, int offset,
1409                                         int len, int odd, struct sk_buff *skb),
1410                             void *from, int length, int transhdrlen,
1411                             struct ipcm_cookie *ipc, struct rtable **rtp,
1412                             unsigned int flags)
1413 {
1414         struct inet_cork cork;
1415         struct sk_buff_head queue;
1416         int err;
1417
1418         if (flags & MSG_PROBE)
1419                 return NULL;
1420
1421         __skb_queue_head_init(&queue);
1422
1423         cork.flags = 0;
1424         cork.addr = 0;
1425         cork.opt = NULL;
1426         err = ip_setup_cork(sk, &cork, ipc, rtp);
1427         if (err)
1428                 return ERR_PTR(err);
1429
1430         err = __ip_append_data(sk, fl4, &queue, &cork,
1431                                &current->task_frag, getfrag,
1432                                from, length, transhdrlen, flags);
1433         if (err) {
1434                 __ip_flush_pending_frames(sk, &queue, &cork);
1435                 return ERR_PTR(err);
1436         }
1437
1438         return __ip_make_skb(sk, fl4, &queue, &cork);
1439 }
1440
1441 /*
1442  *      Fetch data from kernel space and fill in checksum if needed.
1443  */
1444 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1445                               int len, int odd, struct sk_buff *skb)
1446 {
1447         __wsum csum;
1448
1449         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1450         skb->csum = csum_block_add(skb->csum, csum, odd);
1451         return 0;
1452 }
1453
1454 /*
1455  *      Generic function to send a packet as reply to another packet.
1456  *      Used to send some TCP resets/acks so far.
1457  *
1458  *      Use a fake percpu inet socket to avoid false sharing and contention.
1459  */
1460 static DEFINE_PER_CPU(struct inet_sock, unicast_sock) = {
1461         .sk = {
1462                 .__sk_common = {
1463                         .skc_refcnt = ATOMIC_INIT(1),
1464                 },
1465                 .sk_wmem_alloc  = ATOMIC_INIT(1),
1466                 .sk_allocation  = GFP_ATOMIC,
1467                 .sk_flags       = (1UL << SOCK_USE_WRITE_QUEUE),
1468         },
1469         .pmtudisc       = IP_PMTUDISC_WANT,
1470         .uc_ttl         = -1,
1471 };
1472
1473 void ip_send_unicast_reply(struct net *net, struct sk_buff *skb, __be32 daddr,
1474                            __be32 saddr, const struct ip_reply_arg *arg,
1475                            unsigned int len)
1476 {
1477         struct ip_options_data replyopts;
1478         struct ipcm_cookie ipc;
1479         struct flowi4 fl4;
1480         struct rtable *rt = skb_rtable(skb);
1481         struct sk_buff *nskb;
1482         struct sock *sk;
1483         struct inet_sock *inet;
1484         int err;
1485
1486         if (ip_options_echo(&replyopts.opt.opt, skb))
1487                 return;
1488
1489         ipc.addr = daddr;
1490         ipc.opt = NULL;
1491         ipc.tx_flags = 0;
1492
1493         if (replyopts.opt.opt.optlen) {
1494                 ipc.opt = &replyopts.opt;
1495
1496                 if (replyopts.opt.opt.srr)
1497                         daddr = replyopts.opt.opt.faddr;
1498         }
1499
1500         flowi4_init_output(&fl4, arg->bound_dev_if,
1501                            IP4_REPLY_MARK(net, skb->mark),
1502                            RT_TOS(arg->tos),
1503                            RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1504                            ip_reply_arg_flowi_flags(arg),
1505                            daddr, saddr,
1506                            tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1507                            arg->uid);
1508         security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1509         rt = ip_route_output_key(net, &fl4);
1510         if (IS_ERR(rt))
1511                 return;
1512
1513         inet = &get_cpu_var(unicast_sock);
1514
1515         inet->tos = arg->tos;
1516         sk = &inet->sk;
1517         sk->sk_priority = skb->priority;
1518         sk->sk_protocol = ip_hdr(skb)->protocol;
1519         sk->sk_bound_dev_if = arg->bound_dev_if;
1520         sock_net_set(sk, net);
1521         __skb_queue_head_init(&sk->sk_write_queue);
1522         sk->sk_sndbuf = sysctl_wmem_default;
1523         err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1524                              len, 0, &ipc, &rt, MSG_DONTWAIT);
1525         if (unlikely(err)) {
1526                 ip_flush_pending_frames(sk);
1527                 goto out;
1528         }
1529
1530         nskb = skb_peek(&sk->sk_write_queue);
1531         if (nskb) {
1532                 if (arg->csumoffset >= 0)
1533                         *((__sum16 *)skb_transport_header(nskb) +
1534                           arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1535                                                                 arg->csum));
1536                 nskb->ip_summed = CHECKSUM_NONE;
1537                 skb_orphan(nskb);
1538                 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1539                 ip_push_pending_frames(sk, &fl4);
1540         }
1541 out:
1542         put_cpu_var(unicast_sock);
1543
1544         ip_rt_put(rt);
1545 }
1546
1547 void __init ip_init(void)
1548 {
1549         ip_rt_init();
1550         inet_initpeers();
1551
1552 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1553         igmp_mc_proc_init();
1554 #endif
1555 }