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, &rt->dst, 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 int ip_finish_output(struct sk_buff *skb)
 215 {
 216 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 217         /* Policy lookup after SNAT yielded a new policy */
 218         if (skb_dst(skb)->xfrm != NULL) {
 219                 IPCB(skb)->flags |= IPSKB_REROUTED;
 220                 return dst_output(skb);
 221         }
 222 #endif
 223         if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
 224                 return ip_fragment(skb, ip_finish_output2);
 225         else
 226                 return ip_finish_output2(skb);
 227 }
 228
 229 int ip_mc_output(struct sk_buff *skb)
 230 {
 231         struct sock *sk = skb->sk;
 232         struct rtable *rt = skb_rtable(skb);
 233         struct net_device *dev = rt->dst.dev;
 234
 235         /*
 236          *      If the indicated interface is up and running, send the packet.
 237          */
 238         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 239
 240         skb->dev = dev;
 241         skb->protocol = htons(ETH_P_IP);
 242
 243         /*
 244          *      Multicasts are looped back for other local users
 245          */
 246
 247         if (rt->rt_flags&RTCF_MULTICAST) {
 248                 if (sk_mc_loop(sk)
 249 #ifdef CONFIG_IP_MROUTE
 250                 /* Small optimization: do not loopback not local frames,
 251                    which returned after forwarding; they will be  dropped
 252                    by ip_mr_input in any case.
 253                    Note, that local frames are looped back to be delivered
 254                    to local recipients.
 255
 256                    This check is duplicated in ip_mr_input at the moment.
 257                  */
 258                     &&
 259                     ((rt->rt_flags & RTCF_LOCAL) ||
 260                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 261 #endif
 262                    ) {
 263                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 264                         if (newskb)
 265                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 266                                         newskb, NULL, newskb->dev,
 267                                         dev_loopback_xmit);
 268                 }
 269
 270                 /* Multicasts with ttl 0 must not go beyond the host */
 271
 272                 if (ip_hdr(skb)->ttl == 0) {
 273                         kfree_skb(skb);
 274                         return 0;
 275                 }
 276         }
 277
 278         if (rt->rt_flags&RTCF_BROADCAST) {
 279                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 280                 if (newskb)
 281                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
 282                                 NULL, newskb->dev, dev_loopback_xmit);
 283         }
 284
 285         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
 286                             skb->dev, ip_finish_output,
 287                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 288 }
 289
 290 int ip_output(struct sk_buff *skb)
 291 {
 292         struct net_device *dev = skb_dst(skb)->dev;
 293
 294         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 295
 296         skb->dev = dev;
 297         skb->protocol = htons(ETH_P_IP);
 298
 299         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
 300                             ip_finish_output,
 301                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 302 }
 303
 304 /*
 305  * copy saddr and daddr, possibly using 64bit load/stores
 306  * Equivalent to :
 307  *   iph->saddr = fl4->saddr;
 308  *   iph->daddr = fl4->daddr;
 309  */
 310 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
 311 {
 312         BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
 313                      offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
 314         memcpy(&iph->saddr, &fl4->saddr,
 315                sizeof(fl4->saddr) + sizeof(fl4->daddr));
 316 }
 317
 318 int ip_queue_xmit(struct sk_buff *skb, struct flowi *fl)
 319 {
 320         struct sock *sk = skb->sk;
 321         struct inet_sock *inet = inet_sk(sk);
 322         struct ip_options_rcu *inet_opt;
 323         struct flowi4 *fl4;
 324         struct rtable *rt;
 325         struct iphdr *iph;
 326         int res;
 327
 328         /* Skip all of this if the packet is already routed,
 329          * f.e. by something like SCTP.
 330          */
 331         rcu_read_lock();
 332         inet_opt = rcu_dereference(inet->inet_opt);
 333         fl4 = &fl->u.ip4;
 334         rt = skb_rtable(skb);
 335         if (rt != NULL)
 336                 goto packet_routed;
 337
 338         /* Make sure we can route this packet. */
 339         rt = (struct rtable *)__sk_dst_check(sk, 0);
 340         if (rt == NULL) {
 341                 __be32 daddr;
 342
 343                 /* Use correct destination address if we have options. */
 344                 daddr = inet->inet_daddr;
 345                 if (inet_opt && inet_opt->opt.srr)
 346                         daddr = inet_opt->opt.faddr;
 347
 348                 /* If this fails, retransmit mechanism of transport layer will
 349                  * keep trying until route appears or the connection times
 350                  * itself out.
 351                  */
 352                 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
 353                                            daddr, inet->inet_saddr,
 354                                            inet->inet_dport,
 355                                            inet->inet_sport,
 356                                            sk->sk_protocol,
 357                                            RT_CONN_FLAGS(sk),
 358                                            sk->sk_bound_dev_if);
 359                 if (IS_ERR(rt))
 360                         goto no_route;
 361                 sk_setup_caps(sk, &rt->dst);
 362         }
 363         skb_dst_set_noref(skb, &rt->dst);
 364
 365 packet_routed:
 366         if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_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) + (inet_opt ? inet_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         ip_copy_addrs(iph, fl4);
 381
 382         /* Transport layer set skb->h.foo itself. */
 383
 384         if (inet_opt && inet_opt->opt.optlen) {
 385                 iph->ihl += inet_opt->opt.optlen >> 2;
 386                 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
 387         }
 388
 389         ip_select_ident_more(skb, &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 IS_ENABLED(CONFIG_NETFILTER_XT_TARGET_TRACE)
 426         to->nf_trace = from->nf_trace;
 427 #endif
 428 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 429         to->ipvs_property = from->ipvs_property;
 430 #endif
 431         skb_copy_secmark(to, from);
 432 }
 433
 434 /*
 435  *      This IP datagram is too large to be sent in one piece.  Break it up into
 436  *      smaller pieces (each of size equal to IP header plus
 437  *      a block of the data of the original IP data part) that will yet fit in a
 438  *      single device frame, and queue such a frame for sending.
 439  */
 440
 441 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 442 {
 443         struct iphdr *iph;
 444         int ptr;
 445         struct net_device *dev;
 446         struct sk_buff *skb2;
 447         unsigned int mtu, hlen, left, len, ll_rs;
 448         int offset;
 449         __be16 not_last_frag;
 450         struct rtable *rt = skb_rtable(skb);
 451         int err = 0;
 452
 453         dev = rt->dst.dev;
 454
 455         /*
 456          *      Point into the IP datagram header.
 457          */
 458
 459         iph = ip_hdr(skb);
 460
 461         if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->local_df) ||
 462                      (IPCB(skb)->frag_max_size &&
 463                       IPCB(skb)->frag_max_size > dst_mtu(&rt->dst)))) {
 464                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 465                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 466                           htonl(ip_skb_dst_mtu(skb)));
 467                 kfree_skb(skb);
 468                 return -EMSGSIZE;
 469         }
 470
 471         /*
 472          *      Setup starting values.
 473          */
 474
 475         hlen = iph->ihl * 4;
 476         mtu = dst_mtu(&rt->dst) - hlen; /* Size of data space */
 477 #ifdef CONFIG_BRIDGE_NETFILTER
 478         if (skb->nf_bridge)
 479                 mtu -= nf_bridge_mtu_reduction(skb);
 480 #endif
 481         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 482
 483         /* When frag_list is given, use it. First, check its validity:
 484          * some transformers could create wrong frag_list or break existing
 485          * one, it is not prohibited. In this case fall back to copying.
 486          *
 487          * LATER: this step can be merged to real generation of fragments,
 488          * we can switch to copy when see the first bad fragment.
 489          */
 490         if (skb_has_frag_list(skb)) {
 491                 struct sk_buff *frag, *frag2;
 492                 int first_len = skb_pagelen(skb);
 493
 494                 if (first_len - hlen > mtu ||
 495                     ((first_len - hlen) & 7) ||
 496                     ip_is_fragment(iph) ||
 497                     skb_cloned(skb))
 498                         goto slow_path;
 499
 500                 skb_walk_frags(skb, frag) {
 501                         /* Correct geometry. */
 502                         if (frag->len > mtu ||
 503                             ((frag->len & 7) && frag->next) ||
 504                             skb_headroom(frag) < hlen)
 505                                 goto slow_path_clean;
 506
 507                         /* Partially cloned skb? */
 508                         if (skb_shared(frag))
 509                                 goto slow_path_clean;
 510
 511                         BUG_ON(frag->sk);
 512                         if (skb->sk) {
 513                                 frag->sk = skb->sk;
 514                                 frag->destructor = sock_wfree;
 515                         }
 516                         skb->truesize -= frag->truesize;
 517                 }
 518
 519                 /* Everything is OK. Generate! */
 520
 521                 err = 0;
 522                 offset = 0;
 523                 frag = skb_shinfo(skb)->frag_list;
 524                 skb_frag_list_init(skb);
 525                 skb->data_len = first_len - skb_headlen(skb);
 526                 skb->len = first_len;
 527                 iph->tot_len = htons(first_len);
 528                 iph->frag_off = htons(IP_MF);
 529                 ip_send_check(iph);
 530
 531                 for (;;) {
 532                         /* Prepare header of the next frame,
 533                          * before previous one went down. */
 534                         if (frag) {
 535                                 frag->ip_summed = CHECKSUM_NONE;
 536                                 skb_reset_transport_header(frag);
 537                                 __skb_push(frag, hlen);
 538                                 skb_reset_network_header(frag);
 539                                 memcpy(skb_network_header(frag), iph, hlen);
 540                                 iph = ip_hdr(frag);
 541                                 iph->tot_len = htons(frag->len);
 542                                 ip_copy_metadata(frag, skb);
 543                                 if (offset == 0)
 544                                         ip_options_fragment(frag);
 545                                 offset += skb->len - hlen;
 546                                 iph->frag_off = htons(offset>>3);
 547                                 if (frag->next != NULL)
 548                                         iph->frag_off |= htons(IP_MF);
 549                                 /* Ready, complete checksum */
 550                                 ip_send_check(iph);
 551                         }
 552
 553                         err = output(skb);
 554
 555                         if (!err)
 556                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 557                         if (err || !frag)
 558                                 break;
 559
 560                         skb = frag;
 561                         frag = skb->next;
 562                         skb->next = NULL;
 563                 }
 564
 565                 if (err == 0) {
 566                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 567                         return 0;
 568                 }
 569
 570                 while (frag) {
 571                         skb = frag->next;
 572                         kfree_skb(frag);
 573                         frag = skb;
 574                 }
 575                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 576                 return err;
 577
 578 slow_path_clean:
 579                 skb_walk_frags(skb, frag2) {
 580                         if (frag2 == frag)
 581                                 break;
 582                         frag2->sk = NULL;
 583                         frag2->destructor = NULL;
 584                         skb->truesize += frag2->truesize;
 585                 }
 586         }
 587
 588 slow_path:
 589         /* for offloaded checksums cleanup checksum before fragmentation */
 590         if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
 591                 goto fail;
 592         iph = ip_hdr(skb);
 593
 594         left = skb->len - hlen;         /* Space per frame */
 595         ptr = hlen;             /* Where to start from */
 596
 597         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 598          * we need to make room for the encapsulating header
 599          */
 600         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
 601
 602         /*
 603          *      Fragment the datagram.
 604          */
 605
 606         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 607         not_last_frag = iph->frag_off & htons(IP_MF);
 608
 609         /*
 610          *      Keep copying data until we run out.
 611          */
 612
 613         while (left > 0) {
 614                 len = left;
 615                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 616                 if (len > mtu)
 617                         len = mtu;
 618                 /* IF: we are not sending up to and including the packet end
 619                    then align the next start on an eight byte boundary */
 620                 if (len < left) {
 621                         len &= ~7;
 622                 }
 623                 /*
 624                  *      Allocate buffer.
 625                  */
 626
 627                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 628                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 629                         err = -ENOMEM;
 630                         goto fail;
 631                 }
 632
 633                 /*
 634                  *      Set up data on packet
 635                  */
 636
 637                 ip_copy_metadata(skb2, skb);
 638                 skb_reserve(skb2, ll_rs);
 639                 skb_put(skb2, len + hlen);
 640                 skb_reset_network_header(skb2);
 641                 skb2->transport_header = skb2->network_header + hlen;
 642
 643                 /*
 644                  *      Charge the memory for the fragment to any owner
 645                  *      it might possess
 646                  */
 647
 648                 if (skb->sk)
 649                         skb_set_owner_w(skb2, skb->sk);
 650
 651                 /*
 652                  *      Copy the packet header into the new buffer.
 653                  */
 654
 655                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 656
 657                 /*
 658                  *      Copy a block of the IP datagram.
 659                  */
 660                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 661                         BUG();
 662                 left -= len;
 663
 664                 /*
 665                  *      Fill in the new header fields.
 666                  */
 667                 iph = ip_hdr(skb2);
 668                 iph->frag_off = htons((offset >> 3));
 669
 670                 /* ANK: dirty, but effective trick. Upgrade options only if
 671                  * the segment to be fragmented was THE FIRST (otherwise,
 672                  * options are already fixed) and make it ONCE
 673                  * on the initial skb, so that all the following fragments
 674                  * will inherit fixed options.
 675                  */
 676                 if (offset == 0)
 677                         ip_options_fragment(skb);
 678
 679                 /*
 680                  *      Added AC : If we are fragmenting a fragment that's not the
 681                  *                 last fragment then keep MF on each bit
 682                  */
 683                 if (left > 0 || not_last_frag)
 684                         iph->frag_off |= htons(IP_MF);
 685                 ptr += len;
 686                 offset += len;
 687
 688                 /*
 689                  *      Put this fragment into the sending queue.
 690                  */
 691                 iph->tot_len = htons(len + hlen);
 692
 693                 ip_send_check(iph);
 694
 695                 err = output(skb2);
 696                 if (err)
 697                         goto fail;
 698
 699                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 700         }
 701         consume_skb(skb);
 702         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 703         return err;
 704
 705 fail:
 706         kfree_skb(skb);
 707         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 708         return err;
 709 }
 710 EXPORT_SYMBOL(ip_fragment);
 711
 712 int
 713 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 714 {
 715         struct iovec *iov = from;
 716
 717         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 718                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 719                         return -EFAULT;
 720         } else {
 721                 __wsum csum = 0;
 722                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 723                         return -EFAULT;
 724                 skb->csum = csum_block_add(skb->csum, csum, odd);
 725         }
 726         return 0;
 727 }
 728 EXPORT_SYMBOL(ip_generic_getfrag);
 729
 730 static inline __wsum
 731 csum_page(struct page *page, int offset, int copy)
 732 {
 733         char *kaddr;
 734         __wsum csum;
 735         kaddr = kmap(page);
 736         csum = csum_partial(kaddr + offset, copy, 0);
 737         kunmap(page);
 738         return csum;
 739 }
 740
 741 static inline int ip_ufo_append_data(struct sock *sk,
 742                         struct sk_buff_head *queue,
 743                         int getfrag(void *from, char *to, int offset, int len,
 744                                int odd, struct sk_buff *skb),
 745                         void *from, int length, int hh_len, int fragheaderlen,
 746                         int transhdrlen, int maxfraglen, unsigned int flags)
 747 {
 748         struct sk_buff *skb;
 749         int err;
 750
 751         /* There is support for UDP fragmentation offload by network
 752          * device, so create one single skb packet containing complete
 753          * udp datagram
 754          */
 755         if ((skb = skb_peek_tail(queue)) == NULL) {
 756                 skb = sock_alloc_send_skb(sk,
 757                         hh_len + fragheaderlen + transhdrlen + 20,
 758                         (flags & MSG_DONTWAIT), &err);
 759
 760                 if (skb == NULL)
 761                         return err;
 762
 763                 /* reserve space for Hardware header */
 764                 skb_reserve(skb, hh_len);
 765
 766                 /* create space for UDP/IP header */
 767                 skb_put(skb, fragheaderlen + transhdrlen);
 768
 769                 /* initialize network header pointer */
 770                 skb_reset_network_header(skb);
 771
 772                 /* initialize protocol header pointer */
 773                 skb->transport_header = skb->network_header + fragheaderlen;
 774
 775                 skb->csum = 0;
 776
 777
 778                 __skb_queue_tail(queue, skb);
 779         } else if (skb_is_gso(skb)) {
 780                 goto append;
 781         }
 782
 783         skb->ip_summed = CHECKSUM_PARTIAL;
 784         /* specify the length of each IP datagram fragment */
 785         skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
 786         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 787
 788 append:
 789         return skb_append_datato_frags(sk, skb, getfrag, from,
 790                                        (length - transhdrlen));
 791 }
 792
 793 static int __ip_append_data(struct sock *sk,
 794                             struct flowi4 *fl4,
 795                             struct sk_buff_head *queue,
 796                             struct inet_cork *cork,
 797                             struct page_frag *pfrag,
 798                             int getfrag(void *from, char *to, int offset,
 799                                         int len, int odd, struct sk_buff *skb),
 800                             void *from, int length, int transhdrlen,
 801                             unsigned int flags)
 802 {
 803         struct inet_sock *inet = inet_sk(sk);
 804         struct sk_buff *skb;
 805
 806         struct ip_options *opt = cork->opt;
 807         int hh_len;
 808         int exthdrlen;
 809         int mtu;
 810         int copy;
 811         int err;
 812         int offset = 0;
 813         unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
 814         int csummode = CHECKSUM_NONE;
 815         struct rtable *rt = (struct rtable *)cork->dst;
 816
 817         skb = skb_peek_tail(queue);
 818
 819         exthdrlen = !skb ? rt->dst.header_len : 0;
 820         mtu = cork->fragsize;
 821
 822         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 823
 824         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 825         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 826         maxnonfragsize = (inet->pmtudisc >= IP_PMTUDISC_DO) ?
 827                          mtu : 0xFFFF;
 828
 829         if (cork->length + length > maxnonfragsize - 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_is_gso(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 ip_options_rcu *opt;
1041         struct rtable *rt;
1042
1043         /*
1044          * setup for corking.
1045          */
1046         opt = ipc->opt;
1047         if (opt) {
1048                 if (cork->opt == NULL) {
1049                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1050                                             sk->sk_allocation);
1051                         if (unlikely(cork->opt == NULL))
1052                                 return -ENOBUFS;
1053                 }
1054                 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1055                 cork->flags |= IPCORK_OPT;
1056                 cork->addr = ipc->addr;
1057         }
1058         rt = *rtp;
1059         if (unlikely(!rt))
1060                 return -EFAULT;
1061         /*
1062          * We steal reference to this route, caller should not release it
1063          */
1064         *rtp = NULL;
1065         cork->fragsize = ip_sk_use_pmtu(sk) ?
1066                          dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1067         cork->dst = &rt->dst;
1068         cork->length = 0;
1069         cork->ttl = ipc->ttl;
1070         cork->tos = ipc->tos;
1071         cork->priority = ipc->priority;
1072         cork->tx_flags = ipc->tx_flags;
1073
1074         return 0;
1075 }
1076
1077 /*
1078  *      ip_append_data() and ip_append_page() can make one large IP datagram
1079  *      from many pieces of data. Each pieces will be holded on the socket
1080  *      until ip_push_pending_frames() is called. Each piece can be a page
1081  *      or non-page data.
1082  *
1083  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1084  *      this interface potentially.
1085  *
1086  *      LATER: length must be adjusted by pad at tail, when it is required.
1087  */
1088 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1089                    int getfrag(void *from, char *to, int offset, int len,
1090                                int odd, struct sk_buff *skb),
1091                    void *from, int length, int transhdrlen,
1092                    struct ipcm_cookie *ipc, struct rtable **rtp,
1093                    unsigned int flags)
1094 {
1095         struct inet_sock *inet = inet_sk(sk);
1096         int err;
1097
1098         if (flags&MSG_PROBE)
1099                 return 0;
1100
1101         if (skb_queue_empty(&sk->sk_write_queue)) {
1102                 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1103                 if (err)
1104                         return err;
1105         } else {
1106                 transhdrlen = 0;
1107         }
1108
1109         return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1110                                 sk_page_frag(sk), getfrag,
1111                                 from, length, transhdrlen, flags);
1112 }
1113
1114 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, 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         struct inet_cork *cork;
1122         int hh_len;
1123         int mtu;
1124         int len;
1125         int err;
1126         unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1127
1128         if (inet->hdrincl)
1129                 return -EPERM;
1130
1131         if (flags&MSG_PROBE)
1132                 return 0;
1133
1134         if (skb_queue_empty(&sk->sk_write_queue))
1135                 return -EINVAL;
1136
1137         cork = &inet->cork.base;
1138         rt = (struct rtable *)cork->dst;
1139         if (cork->flags & IPCORK_OPT)
1140                 opt = cork->opt;
1141
1142         if (!(rt->dst.dev->features&NETIF_F_SG))
1143                 return -EOPNOTSUPP;
1144
1145         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1146         mtu = cork->fragsize;
1147
1148         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1149         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1150         maxnonfragsize = (inet->pmtudisc >= IP_PMTUDISC_DO) ?
1151                          mtu : 0xFFFF;
1152
1153         if (cork->length + size > maxnonfragsize - fragheaderlen) {
1154                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport, mtu);
1155                 return -EMSGSIZE;
1156         }
1157
1158         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1159                 return -EINVAL;
1160
1161         cork->length += size;
1162         if ((size + skb->len > mtu) &&
1163             (sk->sk_protocol == IPPROTO_UDP) &&
1164             (rt->dst.dev->features & NETIF_F_UFO)) {
1165                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1166                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1167         }
1168
1169
1170         while (size > 0) {
1171                 int i;
1172
1173                 if (skb_is_gso(skb))
1174                         len = size;
1175                 else {
1176
1177                         /* Check if the remaining data fits into current packet. */
1178                         len = mtu - skb->len;
1179                         if (len < size)
1180                                 len = maxfraglen - skb->len;
1181                 }
1182                 if (len <= 0) {
1183                         struct sk_buff *skb_prev;
1184                         int alloclen;
1185
1186                         skb_prev = skb;
1187                         fraggap = skb_prev->len - maxfraglen;
1188
1189                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1190                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1191                         if (unlikely(!skb)) {
1192                                 err = -ENOBUFS;
1193                                 goto error;
1194                         }
1195
1196                         /*
1197                          *      Fill in the control structures
1198                          */
1199                         skb->ip_summed = CHECKSUM_NONE;
1200                         skb->csum = 0;
1201                         skb_reserve(skb, hh_len);
1202
1203                         /*
1204                          *      Find where to start putting bytes.
1205                          */
1206                         skb_put(skb, fragheaderlen + fraggap);
1207                         skb_reset_network_header(skb);
1208                         skb->transport_header = (skb->network_header +
1209                                                  fragheaderlen);
1210                         if (fraggap) {
1211                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1212                                                                    maxfraglen,
1213                                                     skb_transport_header(skb),
1214                                                                    fraggap, 0);
1215                                 skb_prev->csum = csum_sub(skb_prev->csum,
1216                                                           skb->csum);
1217                                 pskb_trim_unique(skb_prev, maxfraglen);
1218                         }
1219
1220                         /*
1221                          * Put the packet on the pending queue.
1222                          */
1223                         __skb_queue_tail(&sk->sk_write_queue, skb);
1224                         continue;
1225                 }
1226
1227                 i = skb_shinfo(skb)->nr_frags;
1228                 if (len > size)
1229                         len = size;
1230                 if (skb_can_coalesce(skb, i, page, offset)) {
1231                         skb_frag_size_add(&skb_shinfo(skb)->frags[i-1], len);
1232                 } else if (i < MAX_SKB_FRAGS) {
1233                         get_page(page);
1234                         skb_fill_page_desc(skb, i, page, offset, len);
1235                 } else {
1236                         err = -EMSGSIZE;
1237                         goto error;
1238                 }
1239
1240                 if (skb->ip_summed == CHECKSUM_NONE) {
1241                         __wsum csum;
1242                         csum = csum_page(page, offset, len);
1243                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1244                 }
1245
1246                 skb->len += len;
1247                 skb->data_len += len;
1248                 skb->truesize += len;
1249                 atomic_add(len, &sk->sk_wmem_alloc);
1250                 offset += len;
1251                 size -= len;
1252         }
1253         return 0;
1254
1255 error:
1256         cork->length -= size;
1257         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1258         return err;
1259 }
1260
1261 static void ip_cork_release(struct inet_cork *cork)
1262 {
1263         cork->flags &= ~IPCORK_OPT;
1264         kfree(cork->opt);
1265         cork->opt = NULL;
1266         dst_release(cork->dst);
1267         cork->dst = NULL;
1268 }
1269
1270 /*
1271  *      Combined all pending IP fragments on the socket as one IP datagram
1272  *      and push them out.
1273  */
1274 struct sk_buff *__ip_make_skb(struct sock *sk,
1275                               struct flowi4 *fl4,
1276                               struct sk_buff_head *queue,
1277                               struct inet_cork *cork)
1278 {
1279         struct sk_buff *skb, *tmp_skb;
1280         struct sk_buff **tail_skb;
1281         struct inet_sock *inet = inet_sk(sk);
1282         struct net *net = sock_net(sk);
1283         struct ip_options *opt = NULL;
1284         struct rtable *rt = (struct rtable *)cork->dst;
1285         struct iphdr *iph;
1286         __be16 df = 0;
1287         __u8 ttl;
1288
1289         if ((skb = __skb_dequeue(queue)) == NULL)
1290                 goto out;
1291         tail_skb = &(skb_shinfo(skb)->frag_list);
1292
1293         /* move skb->data to ip header from ext header */
1294         if (skb->data < skb_network_header(skb))
1295                 __skb_pull(skb, skb_network_offset(skb));
1296         while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1297                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1298                 *tail_skb = tmp_skb;
1299                 tail_skb = &(tmp_skb->next);
1300                 skb->len += tmp_skb->len;
1301                 skb->data_len += tmp_skb->len;
1302                 skb->truesize += tmp_skb->truesize;
1303                 tmp_skb->destructor = NULL;
1304                 tmp_skb->sk = NULL;
1305         }
1306
1307         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1308          * to fragment the frame generated here. No matter, what transforms
1309          * how transforms change size of the packet, it will come out.
1310          */
1311         if (inet->pmtudisc < IP_PMTUDISC_DO)
1312                 skb->local_df = 1;
1313
1314         /* DF bit is set when we want to see DF on outgoing frames.
1315          * If local_df is set too, we still allow to fragment this frame
1316          * locally. */
1317         if (inet->pmtudisc == IP_PMTUDISC_DO ||
1318             inet->pmtudisc == IP_PMTUDISC_PROBE ||
1319             (skb->len <= dst_mtu(&rt->dst) &&
1320              ip_dont_fragment(sk, &rt->dst)))
1321                 df = htons(IP_DF);
1322
1323         if (cork->flags & IPCORK_OPT)
1324                 opt = cork->opt;
1325
1326         if (cork->ttl != 0)
1327                 ttl = cork->ttl;
1328         else if (rt->rt_type == RTN_MULTICAST)
1329                 ttl = inet->mc_ttl;
1330         else
1331                 ttl = ip_select_ttl(inet, &rt->dst);
1332
1333         iph = ip_hdr(skb);
1334         iph->version = 4;
1335         iph->ihl = 5;
1336         iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1337         iph->frag_off = df;
1338         iph->ttl = ttl;
1339         iph->protocol = sk->sk_protocol;
1340         ip_copy_addrs(iph, fl4);
1341         ip_select_ident(skb, &rt->dst, sk);
1342
1343         if (opt) {
1344                 iph->ihl += opt->optlen>>2;
1345                 ip_options_build(skb, opt, cork->addr, rt, 0);
1346         }
1347
1348         skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1349         skb->mark = sk->sk_mark;
1350         /*
1351          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1352          * on dst refcount
1353          */
1354         cork->dst = NULL;
1355         skb_dst_set(skb, &rt->dst);
1356
1357         if (iph->protocol == IPPROTO_ICMP)
1358                 icmp_out_count(net, ((struct icmphdr *)
1359                         skb_transport_header(skb))->type);
1360
1361         ip_cork_release(cork);
1362 out:
1363         return skb;
1364 }
1365
1366 int ip_send_skb(struct net *net, struct sk_buff *skb)
1367 {
1368         int err;
1369
1370         err = ip_local_out(skb);
1371         if (err) {
1372                 if (err > 0)
1373                         err = net_xmit_errno(err);
1374                 if (err)
1375                         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1376         }
1377
1378         return err;
1379 }
1380
1381 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1382 {
1383         struct sk_buff *skb;
1384
1385         skb = ip_finish_skb(sk, fl4);
1386         if (!skb)
1387                 return 0;
1388
1389         /* Netfilter gets whole the not fragmented skb. */
1390         return ip_send_skb(sock_net(sk), skb);
1391 }
1392
1393 /*
1394  *      Throw away all pending data on the socket.
1395  */
1396 static void __ip_flush_pending_frames(struct sock *sk,
1397                                       struct sk_buff_head *queue,
1398                                       struct inet_cork *cork)
1399 {
1400         struct sk_buff *skb;
1401
1402         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1403                 kfree_skb(skb);
1404
1405         ip_cork_release(cork);
1406 }
1407
1408 void ip_flush_pending_frames(struct sock *sk)
1409 {
1410         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1411 }
1412
1413 struct sk_buff *ip_make_skb(struct sock *sk,
1414                             struct flowi4 *fl4,
1415                             int getfrag(void *from, char *to, int offset,
1416                                         int len, int odd, struct sk_buff *skb),
1417                             void *from, int length, int transhdrlen,
1418                             struct ipcm_cookie *ipc, struct rtable **rtp,
1419                             unsigned int flags)
1420 {
1421         struct inet_cork cork;
1422         struct sk_buff_head queue;
1423         int err;
1424
1425         if (flags & MSG_PROBE)
1426                 return NULL;
1427
1428         __skb_queue_head_init(&queue);
1429
1430         cork.flags = 0;
1431         cork.addr = 0;
1432         cork.opt = NULL;
1433         err = ip_setup_cork(sk, &cork, ipc, rtp);
1434         if (err)
1435                 return ERR_PTR(err);
1436
1437         err = __ip_append_data(sk, fl4, &queue, &cork,
1438                                &current->task_frag, getfrag,
1439                                from, length, transhdrlen, flags);
1440         if (err) {
1441                 __ip_flush_pending_frames(sk, &queue, &cork);
1442                 return ERR_PTR(err);
1443         }
1444
1445         return __ip_make_skb(sk, fl4, &queue, &cork);
1446 }
1447
1448 /*
1449  *      Fetch data from kernel space and fill in checksum if needed.
1450  */
1451 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1452                               int len, int odd, struct sk_buff *skb)
1453 {
1454         __wsum csum;
1455
1456         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1457         skb->csum = csum_block_add(skb->csum, csum, odd);
1458         return 0;
1459 }
1460
1461 /*
1462  *      Generic function to send a packet as reply to another packet.
1463  *      Used to send some TCP resets/acks so far.
1464  *
1465  *      Use a fake percpu inet socket to avoid false sharing and contention.
1466  */
1467 static DEFINE_PER_CPU(struct inet_sock, unicast_sock) = {
1468         .sk = {
1469                 .__sk_common = {
1470                         .skc_refcnt = ATOMIC_INIT(1),
1471                 },
1472                 .sk_wmem_alloc  = ATOMIC_INIT(1),
1473                 .sk_allocation  = GFP_ATOMIC,
1474                 .sk_flags       = (1UL << SOCK_USE_WRITE_QUEUE),
1475         },
1476         .pmtudisc       = IP_PMTUDISC_WANT,
1477         .uc_ttl         = -1,
1478 };
1479
1480 void ip_send_unicast_reply(struct net *net, struct sk_buff *skb, __be32 daddr,
1481                            __be32 saddr, const struct ip_reply_arg *arg,
1482                            unsigned int len)
1483 {
1484         struct ip_options_data replyopts;
1485         struct ipcm_cookie ipc;
1486         struct flowi4 fl4;
1487         struct rtable *rt = skb_rtable(skb);
1488         struct sk_buff *nskb;
1489         struct sock *sk;
1490         struct inet_sock *inet;
1491
1492         if (ip_options_echo(&replyopts.opt.opt, skb))
1493                 return;
1494
1495         ipc.addr = daddr;
1496         ipc.opt = NULL;
1497         ipc.tx_flags = 0;
1498         ipc.ttl = 0;
1499         ipc.tos = -1;
1500
1501         if (replyopts.opt.opt.optlen) {
1502                 ipc.opt = &replyopts.opt;
1503
1504                 if (replyopts.opt.opt.srr)
1505                         daddr = replyopts.opt.opt.faddr;
1506         }
1507
1508         flowi4_init_output(&fl4, arg->bound_dev_if, 0,
1509                            RT_TOS(arg->tos),
1510                            RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1511                            ip_reply_arg_flowi_flags(arg),
1512                            daddr, saddr,
1513                            tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1514         security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1515         rt = ip_route_output_key(net, &fl4);
1516         if (IS_ERR(rt))
1517                 return;
1518
1519         inet = &get_cpu_var(unicast_sock);
1520
1521         inet->tos = arg->tos;
1522         sk = &inet->sk;
1523         sk->sk_priority = skb->priority;
1524         sk->sk_protocol = ip_hdr(skb)->protocol;
1525         sk->sk_bound_dev_if = arg->bound_dev_if;
1526         sock_net_set(sk, net);
1527         __skb_queue_head_init(&sk->sk_write_queue);
1528         sk->sk_sndbuf = sysctl_wmem_default;
1529         ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1530                        &ipc, &rt, MSG_DONTWAIT);
1531         nskb = skb_peek(&sk->sk_write_queue);
1532         if (nskb) {
1533                 if (arg->csumoffset >= 0)
1534                         *((__sum16 *)skb_transport_header(nskb) +
1535                           arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1536                                                                 arg->csum));
1537                 nskb->ip_summed = CHECKSUM_NONE;
1538                 skb_orphan(nskb);
1539                 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1540                 ip_push_pending_frames(sk, &fl4);
1541         }
1542
1543         put_cpu_var(unicast_sock);
1544
1545         ip_rt_put(rt);
1546 }
1547
1548 void __init ip_init(void)
1549 {
1550         ip_rt_init();
1551         inet_initpeers();
1552
1553 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1554         igmp_mc_proc_init();
1555 #endif
1556 }