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, <[email protected]>
  10  *              Donald Becker, <[email protected]>
  11  *              Alan Cox, <[email protected]>
  12  *              Richard Underwood
  13  *              Stefan Becker, <[email protected]>
  14  *              Jorge Cwik, <[email protected]>
  15  *              Arnt Gulbrandsen, <[email protected]>
  16  *              Hirokazu Takahashi, <[email protected]>
  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 static int
  87 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
  88             unsigned int mtu,
  89             int (*output)(struct net *, struct sock *, struct sk_buff *));
  90
  91 /* Generate a checksum for an outgoing IP datagram. */
  92 void ip_send_check(struct iphdr *iph)
  93 {
  94         iph->check = 0;
  95         iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
  96 }
  97 EXPORT_SYMBOL(ip_send_check);
  98
  99 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
 100 {
 101         struct iphdr *iph = ip_hdr(skb);
 102
 103         iph->tot_len = htons(skb->len);
 104         ip_send_check(iph);
 105         return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
 106                        net, sk, skb, NULL, skb_dst(skb)->dev,
 107                        dst_output);
 108 }
 109
 110 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
 111 {
 112         int err;
 113
 114         err = __ip_local_out(net, sk, skb);
 115         if (likely(err == 1))
 116                 err = dst_output(net, sk, skb);
 117
 118         return err;
 119 }
 120 EXPORT_SYMBOL_GPL(ip_local_out);
 121
 122 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
 123 {
 124         int ttl = inet->uc_ttl;
 125
 126         if (ttl < 0)
 127                 ttl = ip4_dst_hoplimit(dst);
 128         return ttl;
 129 }
 130
 131 /*
 132  *              Add an ip header to a skbuff and send it out.
 133  *
 134  */
 135 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
 136                           __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
 137 {
 138         struct inet_sock *inet = inet_sk(sk);
 139         struct rtable *rt = skb_rtable(skb);
 140         struct net *net = sock_net(sk);
 141         struct iphdr *iph;
 142
 143         /* Build the IP header. */
 144         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
 145         skb_reset_network_header(skb);
 146         iph = ip_hdr(skb);
 147         iph->version  = 4;
 148         iph->ihl      = 5;
 149         iph->tos      = inet->tos;
 150         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 151         iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
 152         iph->saddr    = saddr;
 153         iph->protocol = sk->sk_protocol;
 154         if (ip_dont_fragment(sk, &rt->dst)) {
 155                 iph->frag_off = htons(IP_DF);
 156                 iph->id = 0;
 157         } else {
 158                 iph->frag_off = 0;
 159                 __ip_select_ident(net, iph, 1);
 160         }
 161
 162         if (opt && opt->opt.optlen) {
 163                 iph->ihl += opt->opt.optlen>>2;
 164                 ip_options_build(skb, &opt->opt, daddr, rt, 0);
 165         }
 166
 167         skb->priority = sk->sk_priority;
 168         skb->mark = sk->sk_mark;
 169
 170         /* Send it out. */
 171         return ip_local_out(net, skb->sk, skb);
 172 }
 173 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
 174
 175 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
 176 {
 177         struct dst_entry *dst = skb_dst(skb);
 178         struct rtable *rt = (struct rtable *)dst;
 179         struct net_device *dev = dst->dev;
 180         unsigned int hh_len = LL_RESERVED_SPACE(dev);
 181         struct neighbour *neigh;
 182         u32 nexthop;
 183
 184         if (rt->rt_type == RTN_MULTICAST) {
 185                 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
 186         } else if (rt->rt_type == RTN_BROADCAST)
 187                 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
 188
 189         /* Be paranoid, rather than too clever. */
 190         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 191                 struct sk_buff *skb2;
 192
 193                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 194                 if (!skb2) {
 195                         kfree_skb(skb);
 196                         return -ENOMEM;
 197                 }
 198                 if (skb->sk)
 199                         skb_set_owner_w(skb2, skb->sk);
 200                 consume_skb(skb);
 201                 skb = skb2;
 202         }
 203
 204         rcu_read_lock_bh();
 205         nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
 206         neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
 207         if (unlikely(!neigh))
 208                 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
 209         if (!IS_ERR(neigh)) {
 210                 int res = dst_neigh_output(dst, neigh, skb);
 211
 212                 rcu_read_unlock_bh();
 213                 return res;
 214         }
 215         rcu_read_unlock_bh();
 216
 217         net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
 218                             __func__);
 219         kfree_skb(skb);
 220         return -EINVAL;
 221 }
 222
 223 static int ip_finish_output_gso(struct net *net, struct sock *sk,
 224                                 struct sk_buff *skb, unsigned int mtu)
 225 {
 226         netdev_features_t features;
 227         struct sk_buff *segs;
 228         int ret = 0;
 229
 230         /* common case: locally created skb or seglen is <= mtu */
 231         if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
 232               skb_gso_network_seglen(skb) <= mtu)
 233                 return ip_finish_output2(net, sk, skb);
 234
 235         /* Slowpath -  GSO segment length is exceeding the dst MTU.
 236          *
 237          * This can happen in two cases:
 238          * 1) TCP GRO packet, DF bit not set
 239          * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
 240          * from host network stack.
 241          */
 242         features = netif_skb_features(skb);
 243         segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
 244         if (IS_ERR_OR_NULL(segs)) {
 245                 kfree_skb(skb);
 246                 return -ENOMEM;
 247         }
 248
 249         consume_skb(skb);
 250
 251         do {
 252                 struct sk_buff *nskb = segs->next;
 253                 int err;
 254
 255                 segs->next = NULL;
 256                 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
 257
 258                 if (err && ret == 0)
 259                         ret = err;
 260                 segs = nskb;
 261         } while (segs);
 262
 263         return ret;
 264 }
 265
 266 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
 267 {
 268         unsigned int mtu;
 269
 270 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 271         /* Policy lookup after SNAT yielded a new policy */
 272         if (skb_dst(skb)->xfrm) {
 273                 IPCB(skb)->flags |= IPSKB_REROUTED;
 274                 return dst_output(net, sk, skb);
 275         }
 276 #endif
 277         mtu = ip_skb_dst_mtu(skb);
 278         if (skb_is_gso(skb))
 279                 return ip_finish_output_gso(net, sk, skb, mtu);
 280
 281         if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
 282                 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
 283
 284         return ip_finish_output2(net, sk, skb);
 285 }
 286
 287 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
 288 {
 289         struct rtable *rt = skb_rtable(skb);
 290         struct net_device *dev = rt->dst.dev;
 291
 292         /*
 293          *      If the indicated interface is up and running, send the packet.
 294          */
 295         IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
 296
 297         skb->dev = dev;
 298         skb->protocol = htons(ETH_P_IP);
 299
 300         /*
 301          *      Multicasts are looped back for other local users
 302          */
 303
 304         if (rt->rt_flags&RTCF_MULTICAST) {
 305                 if (sk_mc_loop(sk)
 306 #ifdef CONFIG_IP_MROUTE
 307                 /* Small optimization: do not loopback not local frames,
 308                    which returned after forwarding; they will be  dropped
 309                    by ip_mr_input in any case.
 310                    Note, that local frames are looped back to be delivered
 311                    to local recipients.
 312
 313                    This check is duplicated in ip_mr_input at the moment.
 314                  */
 315                     &&
 316                     ((rt->rt_flags & RTCF_LOCAL) ||
 317                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 318 #endif
 319                    ) {
 320                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 321                         if (newskb)
 322                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 323                                         net, sk, newskb, NULL, newskb->dev,
 324                                         dev_loopback_xmit);
 325                 }
 326
 327                 /* Multicasts with ttl 0 must not go beyond the host */
 328
 329                 if (ip_hdr(skb)->ttl == 0) {
 330                         kfree_skb(skb);
 331                         return 0;
 332                 }
 333         }
 334
 335         if (rt->rt_flags&RTCF_BROADCAST) {
 336                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 337                 if (newskb)
 338                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 339                                 net, sk, newskb, NULL, newskb->dev,
 340                                 dev_loopback_xmit);
 341         }
 342
 343         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 344                             net, sk, skb, NULL, skb->dev,
 345                             ip_finish_output,
 346                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 347 }
 348
 349 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
 350 {
 351         struct net_device *dev = skb_dst(skb)->dev;
 352
 353         IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
 354
 355         skb->dev = dev;
 356         skb->protocol = htons(ETH_P_IP);
 357
 358         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 359                             net, sk, skb, NULL, dev,
 360                             ip_finish_output,
 361                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 362 }
 363
 364 /*
 365  * copy saddr and daddr, possibly using 64bit load/stores
 366  * Equivalent to :
 367  *   iph->saddr = fl4->saddr;
 368  *   iph->daddr = fl4->daddr;
 369  */
 370 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
 371 {
 372         BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
 373                      offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
 374         memcpy(&iph->saddr, &fl4->saddr,
 375                sizeof(fl4->saddr) + sizeof(fl4->daddr));
 376 }
 377
 378 /* Note: skb->sk can be different from sk, in case of tunnels */
 379 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
 380 {
 381         struct inet_sock *inet = inet_sk(sk);
 382         struct net *net = sock_net(sk);
 383         struct ip_options_rcu *inet_opt;
 384         struct flowi4 *fl4;
 385         struct rtable *rt;
 386         struct iphdr *iph;
 387         int res;
 388
 389         /* Skip all of this if the packet is already routed,
 390          * f.e. by something like SCTP.
 391          */
 392         rcu_read_lock();
 393         inet_opt = rcu_dereference(inet->inet_opt);
 394         fl4 = &fl->u.ip4;
 395         rt = skb_rtable(skb);
 396         if (rt)
 397                 goto packet_routed;
 398
 399         /* Make sure we can route this packet. */
 400         rt = (struct rtable *)__sk_dst_check(sk, 0);
 401         if (!rt) {
 402                 __be32 daddr;
 403
 404                 /* Use correct destination address if we have options. */
 405                 daddr = inet->inet_daddr;
 406                 if (inet_opt && inet_opt->opt.srr)
 407                         daddr = inet_opt->opt.faddr;
 408
 409                 /* If this fails, retransmit mechanism of transport layer will
 410                  * keep trying until route appears or the connection times
 411                  * itself out.
 412                  */
 413                 rt = ip_route_output_ports(net, fl4, sk,
 414                                            daddr, inet->inet_saddr,
 415                                            inet->inet_dport,
 416                                            inet->inet_sport,
 417                                            sk->sk_protocol,
 418                                            RT_CONN_FLAGS(sk),
 419                                            sk->sk_bound_dev_if);
 420                 if (IS_ERR(rt))
 421                         goto no_route;
 422                 sk_setup_caps(sk, &rt->dst);
 423         }
 424         skb_dst_set_noref(skb, &rt->dst);
 425
 426 packet_routed:
 427         if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
 428                 goto no_route;
 429
 430         /* OK, we know where to send it, allocate and build IP header. */
 431         skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
 432         skb_reset_network_header(skb);
 433         iph = ip_hdr(skb);
 434         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 435         if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
 436                 iph->frag_off = htons(IP_DF);
 437         else
 438                 iph->frag_off = 0;
 439         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 440         iph->protocol = sk->sk_protocol;
 441         ip_copy_addrs(iph, fl4);
 442
 443         /* Transport layer set skb->h.foo itself. */
 444
 445         if (inet_opt && inet_opt->opt.optlen) {
 446                 iph->ihl += inet_opt->opt.optlen >> 2;
 447                 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
 448         }
 449
 450         ip_select_ident_segs(net, skb, sk,
 451                              skb_shinfo(skb)->gso_segs ?: 1);
 452
 453         /* TODO : should we use skb->sk here instead of sk ? */
 454         skb->priority = sk->sk_priority;
 455         skb->mark = sk->sk_mark;
 456
 457         res = ip_local_out(net, sk, skb);
 458         rcu_read_unlock();
 459         return res;
 460
 461 no_route:
 462         rcu_read_unlock();
 463         IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 464         kfree_skb(skb);
 465         return -EHOSTUNREACH;
 466 }
 467 EXPORT_SYMBOL(ip_queue_xmit);
 468
 469 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 470 {
 471         to->pkt_type = from->pkt_type;
 472         to->priority = from->priority;
 473         to->protocol = from->protocol;
 474         skb_dst_drop(to);
 475         skb_dst_copy(to, from);
 476         to->dev = from->dev;
 477         to->mark = from->mark;
 478
 479         /* Copy the flags to each fragment. */
 480         IPCB(to)->flags = IPCB(from)->flags;
 481
 482 #ifdef CONFIG_NET_SCHED
 483         to->tc_index = from->tc_index;
 484 #endif
 485         nf_copy(to, from);
 486 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 487         to->ipvs_property = from->ipvs_property;
 488 #endif
 489         skb_copy_secmark(to, from);
 490 }
 491
 492 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
 493                        unsigned int mtu,
 494                        int (*output)(struct net *, struct sock *, struct sk_buff *))
 495 {
 496         struct iphdr *iph = ip_hdr(skb);
 497
 498         if ((iph->frag_off & htons(IP_DF)) == 0)
 499                 return ip_do_fragment(net, sk, skb, output);
 500
 501         if (unlikely(!skb->ignore_df ||
 502                      (IPCB(skb)->frag_max_size &&
 503                       IPCB(skb)->frag_max_size > mtu))) {
 504                 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
 505                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 506                           htonl(mtu));
 507                 kfree_skb(skb);
 508                 return -EMSGSIZE;
 509         }
 510
 511         return ip_do_fragment(net, sk, skb, output);
 512 }
 513
 514 /*
 515  *      This IP datagram is too large to be sent in one piece.  Break it up into
 516  *      smaller pieces (each of size equal to IP header plus
 517  *      a block of the data of the original IP data part) that will yet fit in a
 518  *      single device frame, and queue such a frame for sending.
 519  */
 520
 521 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
 522                    int (*output)(struct net *, struct sock *, struct sk_buff *))
 523 {
 524         struct iphdr *iph;
 525         int ptr;
 526         struct net_device *dev;
 527         struct sk_buff *skb2;
 528         unsigned int mtu, hlen, left, len, ll_rs;
 529         int offset;
 530         __be16 not_last_frag;
 531         struct rtable *rt = skb_rtable(skb);
 532         int err = 0;
 533
 534         dev = rt->dst.dev;
 535
 536         /* for offloaded checksums cleanup checksum before fragmentation */
 537         if (skb->ip_summed == CHECKSUM_PARTIAL &&
 538             (err = skb_checksum_help(skb)))
 539                 goto fail;
 540
 541         /*
 542          *      Point into the IP datagram header.
 543          */
 544
 545         iph = ip_hdr(skb);
 546
 547         mtu = ip_skb_dst_mtu(skb);
 548         if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
 549                 mtu = IPCB(skb)->frag_max_size;
 550
 551         /*
 552          *      Setup starting values.
 553          */
 554
 555         hlen = iph->ihl * 4;
 556         mtu = mtu - hlen;       /* Size of data space */
 557         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 558
 559         /* When frag_list is given, use it. First, check its validity:
 560          * some transformers could create wrong frag_list or break existing
 561          * one, it is not prohibited. In this case fall back to copying.
 562          *
 563          * LATER: this step can be merged to real generation of fragments,
 564          * we can switch to copy when see the first bad fragment.
 565          */
 566         if (skb_has_frag_list(skb)) {
 567                 struct sk_buff *frag, *frag2;
 568                 int first_len = skb_pagelen(skb);
 569
 570                 if (first_len - hlen > mtu ||
 571                     ((first_len - hlen) & 7) ||
 572                     ip_is_fragment(iph) ||
 573                     skb_cloned(skb))
 574                         goto slow_path;
 575
 576                 skb_walk_frags(skb, frag) {
 577                         /* Correct geometry. */
 578                         if (frag->len > mtu ||
 579                             ((frag->len & 7) && frag->next) ||
 580                             skb_headroom(frag) < hlen)
 581                                 goto slow_path_clean;
 582
 583                         /* Partially cloned skb? */
 584                         if (skb_shared(frag))
 585                                 goto slow_path_clean;
 586
 587                         BUG_ON(frag->sk);
 588                         if (skb->sk) {
 589                                 frag->sk = skb->sk;
 590                                 frag->destructor = sock_wfree;
 591                         }
 592                         skb->truesize -= frag->truesize;
 593                 }
 594
 595                 /* Everything is OK. Generate! */
 596
 597                 err = 0;
 598                 offset = 0;
 599                 frag = skb_shinfo(skb)->frag_list;
 600                 skb_frag_list_init(skb);
 601                 skb->data_len = first_len - skb_headlen(skb);
 602                 skb->len = first_len;
 603                 iph->tot_len = htons(first_len);
 604                 iph->frag_off = htons(IP_MF);
 605                 ip_send_check(iph);
 606
 607                 for (;;) {
 608                         /* Prepare header of the next frame,
 609                          * before previous one went down. */
 610                         if (frag) {
 611                                 frag->ip_summed = CHECKSUM_NONE;
 612                                 skb_reset_transport_header(frag);
 613                                 __skb_push(frag, hlen);
 614                                 skb_reset_network_header(frag);
 615                                 memcpy(skb_network_header(frag), iph, hlen);
 616                                 iph = ip_hdr(frag);
 617                                 iph->tot_len = htons(frag->len);
 618                                 ip_copy_metadata(frag, skb);
 619                                 if (offset == 0)
 620                                         ip_options_fragment(frag);
 621                                 offset += skb->len - hlen;
 622                                 iph->frag_off = htons(offset>>3);
 623                                 if (frag->next)
 624                                         iph->frag_off |= htons(IP_MF);
 625                                 /* Ready, complete checksum */
 626                                 ip_send_check(iph);
 627                         }
 628
 629                         err = output(net, sk, skb);
 630
 631                         if (!err)
 632                                 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
 633                         if (err || !frag)
 634                                 break;
 635
 636                         skb = frag;
 637                         frag = skb->next;
 638                         skb->next = NULL;
 639                 }
 640
 641                 if (err == 0) {
 642                         IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
 643                         return 0;
 644                 }
 645
 646                 while (frag) {
 647                         skb = frag->next;
 648                         kfree_skb(frag);
 649                         frag = skb;
 650                 }
 651                 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
 652                 return err;
 653
 654 slow_path_clean:
 655                 skb_walk_frags(skb, frag2) {
 656                         if (frag2 == frag)
 657                                 break;
 658                         frag2->sk = NULL;
 659                         frag2->destructor = NULL;
 660                         skb->truesize += frag2->truesize;
 661                 }
 662         }
 663
 664 slow_path:
 665         iph = ip_hdr(skb);
 666
 667         left = skb->len - hlen;         /* Space per frame */
 668         ptr = hlen;             /* Where to start from */
 669
 670         ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
 671
 672         /*
 673          *      Fragment the datagram.
 674          */
 675
 676         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 677         not_last_frag = iph->frag_off & htons(IP_MF);
 678
 679         /*
 680          *      Keep copying data until we run out.
 681          */
 682
 683         while (left > 0) {
 684                 len = left;
 685                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 686                 if (len > mtu)
 687                         len = mtu;
 688                 /* IF: we are not sending up to and including the packet end
 689                    then align the next start on an eight byte boundary */
 690                 if (len < left) {
 691                         len &= ~7;
 692                 }
 693
 694                 /* Allocate buffer */
 695                 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
 696                 if (!skb2) {
 697                         err = -ENOMEM;
 698                         goto fail;
 699                 }
 700
 701                 /*
 702                  *      Set up data on packet
 703                  */
 704
 705                 ip_copy_metadata(skb2, skb);
 706                 skb_reserve(skb2, ll_rs);
 707                 skb_put(skb2, len + hlen);
 708                 skb_reset_network_header(skb2);
 709                 skb2->transport_header = skb2->network_header + hlen;
 710
 711                 /*
 712                  *      Charge the memory for the fragment to any owner
 713                  *      it might possess
 714                  */
 715
 716                 if (skb->sk)
 717                         skb_set_owner_w(skb2, skb->sk);
 718
 719                 /*
 720                  *      Copy the packet header into the new buffer.
 721                  */
 722
 723                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 724
 725                 /*
 726                  *      Copy a block of the IP datagram.
 727                  */
 728                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 729                         BUG();
 730                 left -= len;
 731
 732                 /*
 733                  *      Fill in the new header fields.
 734                  */
 735                 iph = ip_hdr(skb2);
 736                 iph->frag_off = htons((offset >> 3));
 737
 738                 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
 739                         iph->frag_off |= htons(IP_DF);
 740
 741                 /* ANK: dirty, but effective trick. Upgrade options only if
 742                  * the segment to be fragmented was THE FIRST (otherwise,
 743                  * options are already fixed) and make it ONCE
 744                  * on the initial skb, so that all the following fragments
 745                  * will inherit fixed options.
 746                  */
 747                 if (offset == 0)
 748                         ip_options_fragment(skb);
 749
 750                 /*
 751                  *      Added AC : If we are fragmenting a fragment that's not the
 752                  *                 last fragment then keep MF on each bit
 753                  */
 754                 if (left > 0 || not_last_frag)
 755                         iph->frag_off |= htons(IP_MF);
 756                 ptr += len;
 757                 offset += len;
 758
 759                 /*
 760                  *      Put this fragment into the sending queue.
 761                  */
 762                 iph->tot_len = htons(len + hlen);
 763
 764                 ip_send_check(iph);
 765
 766                 err = output(net, sk, skb2);
 767                 if (err)
 768                         goto fail;
 769
 770                 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
 771         }
 772         consume_skb(skb);
 773         IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
 774         return err;
 775
 776 fail:
 777         kfree_skb(skb);
 778         IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
 779         return err;
 780 }
 781 EXPORT_SYMBOL(ip_do_fragment);
 782
 783 int
 784 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 785 {
 786         struct msghdr *msg = from;
 787
 788         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 789                 if (copy_from_iter(to, len, &msg->msg_iter) != len)
 790                         return -EFAULT;
 791         } else {
 792                 __wsum csum = 0;
 793                 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
 794                         return -EFAULT;
 795                 skb->csum = csum_block_add(skb->csum, csum, odd);
 796         }
 797         return 0;
 798 }
 799 EXPORT_SYMBOL(ip_generic_getfrag);
 800
 801 static inline __wsum
 802 csum_page(struct page *page, int offset, int copy)
 803 {
 804         char *kaddr;
 805         __wsum csum;
 806         kaddr = kmap(page);
 807         csum = csum_partial(kaddr + offset, copy, 0);
 808         kunmap(page);
 809         return csum;
 810 }
 811
 812 static inline int ip_ufo_append_data(struct sock *sk,
 813                         struct sk_buff_head *queue,
 814                         int getfrag(void *from, char *to, int offset, int len,
 815                                int odd, struct sk_buff *skb),
 816                         void *from, int length, int hh_len, int fragheaderlen,
 817                         int transhdrlen, int maxfraglen, unsigned int flags)
 818 {
 819         struct sk_buff *skb;
 820         int err;
 821
 822         /* There is support for UDP fragmentation offload by network
 823          * device, so create one single skb packet containing complete
 824          * udp datagram
 825          */
 826         skb = skb_peek_tail(queue);
 827         if (!skb) {
 828                 skb = sock_alloc_send_skb(sk,
 829                         hh_len + fragheaderlen + transhdrlen + 20,
 830                         (flags & MSG_DONTWAIT), &err);
 831
 832                 if (!skb)
 833                         return err;
 834
 835                 /* reserve space for Hardware header */
 836                 skb_reserve(skb, hh_len);
 837
 838                 /* create space for UDP/IP header */
 839                 skb_put(skb, fragheaderlen + transhdrlen);
 840
 841                 /* initialize network header pointer */
 842                 skb_reset_network_header(skb);
 843
 844                 /* initialize protocol header pointer */
 845                 skb->transport_header = skb->network_header + fragheaderlen;
 846
 847                 skb->csum = 0;
 848
 849                 __skb_queue_tail(queue, skb);
 850         } else if (skb_is_gso(skb)) {
 851                 goto append;
 852         }
 853
 854         skb->ip_summed = CHECKSUM_PARTIAL;
 855         /* specify the length of each IP datagram fragment */
 856         skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
 857         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 858
 859 append:
 860         return skb_append_datato_frags(sk, skb, getfrag, from,
 861                                        (length - transhdrlen));
 862 }
 863
 864 static int __ip_append_data(struct sock *sk,
 865                             struct flowi4 *fl4,
 866                             struct sk_buff_head *queue,
 867                             struct inet_cork *cork,
 868                             struct page_frag *pfrag,
 869                             int getfrag(void *from, char *to, int offset,
 870                                         int len, int odd, struct sk_buff *skb),
 871                             void *from, int length, int transhdrlen,
 872                             unsigned int flags)
 873 {
 874         struct inet_sock *inet = inet_sk(sk);
 875         struct sk_buff *skb;
 876
 877         struct ip_options *opt = cork->opt;
 878         int hh_len;
 879         int exthdrlen;
 880         int mtu;
 881         int copy;
 882         int err;
 883         int offset = 0;
 884         unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
 885         int csummode = CHECKSUM_NONE;
 886         struct rtable *rt = (struct rtable *)cork->dst;
 887         u32 tskey = 0;
 888
 889         skb = skb_peek_tail(queue);
 890
 891         exthdrlen = !skb ? rt->dst.header_len : 0;
 892         mtu = cork->fragsize;
 893         if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
 894             sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
 895                 tskey = sk->sk_tskey++;
 896
 897         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 898
 899         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 900         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 901         maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
 902
 903         if (cork->length + length > maxnonfragsize - fragheaderlen) {
 904                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
 905                                mtu - (opt ? opt->optlen : 0));
 906                 return -EMSGSIZE;
 907         }
 908
 909         /*
 910          * transhdrlen > 0 means that this is the first fragment and we wish
 911          * it won't be fragmented in the future.
 912          */
 913         if (transhdrlen &&
 914             length + fragheaderlen <= mtu &&
 915             rt->dst.dev->features & NETIF_F_V4_CSUM &&
 916             !(flags & MSG_MORE) &&
 917             !exthdrlen)
 918                 csummode = CHECKSUM_PARTIAL;
 919
 920         cork->length += length;
 921         if (((length > mtu) || (skb && skb_is_gso(skb))) &&
 922             (sk->sk_protocol == IPPROTO_UDP) &&
 923             (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
 924             (sk->sk_type == SOCK_DGRAM)) {
 925                 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
 926                                          hh_len, fragheaderlen, transhdrlen,
 927                                          maxfraglen, flags);
 928                 if (err)
 929                         goto error;
 930                 return 0;
 931         }
 932
 933         /* So, what's going on in the loop below?
 934          *
 935          * We use calculated fragment length to generate chained skb,
 936          * each of segments is IP fragment ready for sending to network after
 937          * adding appropriate IP header.
 938          */
 939
 940         if (!skb)
 941                 goto alloc_new_skb;
 942
 943         while (length > 0) {
 944                 /* Check if the remaining data fits into current packet. */
 945                 copy = mtu - skb->len;
 946                 if (copy < length)
 947                         copy = maxfraglen - skb->len;
 948                 if (copy <= 0) {
 949                         char *data;
 950                         unsigned int datalen;
 951                         unsigned int fraglen;
 952                         unsigned int fraggap;
 953                         unsigned int alloclen;
 954                         struct sk_buff *skb_prev;
 955 alloc_new_skb:
 956                         skb_prev = skb;
 957                         if (skb_prev)
 958                                 fraggap = skb_prev->len - maxfraglen;
 959                         else
 960                                 fraggap = 0;
 961
 962                         /*
 963                          * If remaining data exceeds the mtu,
 964                          * we know we need more fragment(s).
 965                          */
 966                         datalen = length + fraggap;
 967                         if (datalen > mtu - fragheaderlen)
 968                                 datalen = maxfraglen - fragheaderlen;
 969                         fraglen = datalen + fragheaderlen;
 970
 971                         if ((flags & MSG_MORE) &&
 972                             !(rt->dst.dev->features&NETIF_F_SG))
 973                                 alloclen = mtu;
 974                         else
 975                                 alloclen = fraglen;
 976
 977                         alloclen += exthdrlen;
 978
 979                         /* The last fragment gets additional space at tail.
 980                          * Note, with MSG_MORE we overallocate on fragments,
 981                          * because we have no idea what fragment will be
 982                          * the last.
 983                          */
 984                         if (datalen == length + fraggap)
 985                                 alloclen += rt->dst.trailer_len;
 986
 987                         if (transhdrlen) {
 988                                 skb = sock_alloc_send_skb(sk,
 989                                                 alloclen + hh_len + 15,
 990                                                 (flags & MSG_DONTWAIT), &err);
 991                         } else {
 992                                 skb = NULL;
 993                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 994                                     2 * sk->sk_sndbuf)
 995                                         skb = sock_wmalloc(sk,
 996                                                            alloclen + hh_len + 15, 1,
 997                                                            sk->sk_allocation);
 998                                 if (unlikely(!skb))
 999                                         err = -ENOBUFS;
1000                         }
1001                         if (!skb)
1002                                 goto error;
1003
1004                         /*
1005                          *      Fill in the control structures
1006                          */
1007                         skb->ip_summed = csummode;
1008                         skb->csum = 0;
1009                         skb_reserve(skb, hh_len);
1010
1011                         /* only the initial fragment is time stamped */
1012                         skb_shinfo(skb)->tx_flags = cork->tx_flags;
1013                         cork->tx_flags = 0;
1014                         skb_shinfo(skb)->tskey = tskey;
1015                         tskey = 0;
1016
1017                         /*
1018                          *      Find where to start putting bytes.
1019                          */
1020                         data = skb_put(skb, fraglen + exthdrlen);
1021                         skb_set_network_header(skb, exthdrlen);
1022                         skb->transport_header = (skb->network_header +
1023                                                  fragheaderlen);
1024                         data += fragheaderlen + exthdrlen;
1025
1026                         if (fraggap) {
1027                                 skb->csum = skb_copy_and_csum_bits(
1028                                         skb_prev, maxfraglen,
1029                                         data + transhdrlen, fraggap, 0);
1030                                 skb_prev->csum = csum_sub(skb_prev->csum,
1031                                                           skb->csum);
1032                                 data += fraggap;
1033                                 pskb_trim_unique(skb_prev, maxfraglen);
1034                         }
1035
1036                         copy = datalen - transhdrlen - fraggap;
1037                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1038                                 err = -EFAULT;
1039                                 kfree_skb(skb);
1040                                 goto error;
1041                         }
1042
1043                         offset += copy;
1044                         length -= datalen - fraggap;
1045                         transhdrlen = 0;
1046                         exthdrlen = 0;
1047                         csummode = CHECKSUM_NONE;
1048
1049                         /*
1050                          * Put the packet on the pending queue.
1051                          */
1052                         __skb_queue_tail(queue, skb);
1053                         continue;
1054                 }
1055
1056                 if (copy > length)
1057                         copy = length;
1058
1059                 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1060                         unsigned int off;
1061
1062                         off = skb->len;
1063                         if (getfrag(from, skb_put(skb, copy),
1064                                         offset, copy, off, skb) < 0) {
1065                                 __skb_trim(skb, off);
1066                                 err = -EFAULT;
1067                                 goto error;
1068                         }
1069                 } else {
1070                         int i = skb_shinfo(skb)->nr_frags;
1071
1072                         err = -ENOMEM;
1073                         if (!sk_page_frag_refill(sk, pfrag))
1074                                 goto error;
1075
1076                         if (!skb_can_coalesce(skb, i, pfrag->page,
1077                                               pfrag->offset)) {
1078                                 err = -EMSGSIZE;
1079                                 if (i == MAX_SKB_FRAGS)
1080                                         goto error;
1081
1082                                 __skb_fill_page_desc(skb, i, pfrag->page,
1083                                                      pfrag->offset, 0);
1084                                 skb_shinfo(skb)->nr_frags = ++i;
1085                                 get_page(pfrag->page);
1086                         }
1087                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1088                         if (getfrag(from,
1089                                     page_address(pfrag->page) + pfrag->offset,
1090                                     offset, copy, skb->len, skb) < 0)
1091                                 goto error_efault;
1092
1093                         pfrag->offset += copy;
1094                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1095                         skb->len += copy;
1096                         skb->data_len += copy;
1097                         skb->truesize += copy;
1098                         atomic_add(copy, &sk->sk_wmem_alloc);
1099                 }
1100                 offset += copy;
1101                 length -= copy;
1102         }
1103
1104         return 0;
1105
1106 error_efault:
1107         err = -EFAULT;
1108 error:
1109         cork->length -= length;
1110         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1111         return err;
1112 }
1113
1114 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1115                          struct ipcm_cookie *ipc, struct rtable **rtp)
1116 {
1117         struct ip_options_rcu *opt;
1118         struct rtable *rt;
1119
1120         /*
1121          * setup for corking.
1122          */
1123         opt = ipc->opt;
1124         if (opt) {
1125                 if (!cork->opt) {
1126                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1127                                             sk->sk_allocation);
1128                         if (unlikely(!cork->opt))
1129                                 return -ENOBUFS;
1130                 }
1131                 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1132                 cork->flags |= IPCORK_OPT;
1133                 cork->addr = ipc->addr;
1134         }
1135         rt = *rtp;
1136         if (unlikely(!rt))
1137                 return -EFAULT;
1138         /*
1139          * We steal reference to this route, caller should not release it
1140          */
1141         *rtp = NULL;
1142         cork->fragsize = ip_sk_use_pmtu(sk) ?
1143                          dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1144         cork->dst = &rt->dst;
1145         cork->length = 0;
1146         cork->ttl = ipc->ttl;
1147         cork->tos = ipc->tos;
1148         cork->priority = ipc->priority;
1149         cork->tx_flags = ipc->tx_flags;
1150
1151         return 0;
1152 }
1153
1154 /*
1155  *      ip_append_data() and ip_append_page() can make one large IP datagram
1156  *      from many pieces of data. Each pieces will be holded on the socket
1157  *      until ip_push_pending_frames() is called. Each piece can be a page
1158  *      or non-page data.
1159  *
1160  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1161  *      this interface potentially.
1162  *
1163  *      LATER: length must be adjusted by pad at tail, when it is required.
1164  */
1165 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1166                    int getfrag(void *from, char *to, int offset, int len,
1167                                int odd, struct sk_buff *skb),
1168                    void *from, int length, int transhdrlen,
1169                    struct ipcm_cookie *ipc, struct rtable **rtp,
1170                    unsigned int flags)
1171 {
1172         struct inet_sock *inet = inet_sk(sk);
1173         int err;
1174
1175         if (flags&MSG_PROBE)
1176                 return 0;
1177
1178         if (skb_queue_empty(&sk->sk_write_queue)) {
1179                 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1180                 if (err)
1181                         return err;
1182         } else {
1183                 transhdrlen = 0;
1184         }
1185
1186         return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1187                                 sk_page_frag(sk), getfrag,
1188                                 from, length, transhdrlen, flags);
1189 }
1190
1191 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1192                        int offset, size_t size, int flags)
1193 {
1194         struct inet_sock *inet = inet_sk(sk);
1195         struct sk_buff *skb;
1196         struct rtable *rt;
1197         struct ip_options *opt = NULL;
1198         struct inet_cork *cork;
1199         int hh_len;
1200         int mtu;
1201         int len;
1202         int err;
1203         unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1204
1205         if (inet->hdrincl)
1206                 return -EPERM;
1207
1208         if (flags&MSG_PROBE)
1209                 return 0;
1210
1211         if (skb_queue_empty(&sk->sk_write_queue))
1212                 return -EINVAL;
1213
1214         cork = &inet->cork.base;
1215         rt = (struct rtable *)cork->dst;
1216         if (cork->flags & IPCORK_OPT)
1217                 opt = cork->opt;
1218
1219         if (!(rt->dst.dev->features&NETIF_F_SG))
1220                 return -EOPNOTSUPP;
1221
1222         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1223         mtu = cork->fragsize;
1224
1225         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1226         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1227         maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1228
1229         if (cork->length + size > maxnonfragsize - fragheaderlen) {
1230                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1231                                mtu - (opt ? opt->optlen : 0));
1232                 return -EMSGSIZE;
1233         }
1234
1235         skb = skb_peek_tail(&sk->sk_write_queue);
1236         if (!skb)
1237                 return -EINVAL;
1238
1239         cork->length += size;
1240         if ((size + skb->len > mtu) &&
1241             (sk->sk_protocol == IPPROTO_UDP) &&
1242             (rt->dst.dev->features & NETIF_F_UFO)) {
1243                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1244                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1245         }
1246
1247         while (size > 0) {
1248                 if (skb_is_gso(skb)) {
1249                         len = size;
1250                 } else {
1251
1252                         /* Check if the remaining data fits into current packet. */
1253                         len = mtu - skb->len;
1254                         if (len < size)
1255                                 len = maxfraglen - skb->len;
1256                 }
1257                 if (len <= 0) {
1258                         struct sk_buff *skb_prev;
1259                         int alloclen;
1260
1261                         skb_prev = skb;
1262                         fraggap = skb_prev->len - maxfraglen;
1263
1264                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1265                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1266                         if (unlikely(!skb)) {
1267                                 err = -ENOBUFS;
1268                                 goto error;
1269                         }
1270
1271                         /*
1272                          *      Fill in the control structures
1273                          */
1274                         skb->ip_summed = CHECKSUM_NONE;
1275                         skb->csum = 0;
1276                         skb_reserve(skb, hh_len);
1277
1278                         /*
1279                          *      Find where to start putting bytes.
1280                          */
1281                         skb_put(skb, fragheaderlen + fraggap);
1282                         skb_reset_network_header(skb);
1283                         skb->transport_header = (skb->network_header +
1284                                                  fragheaderlen);
1285                         if (fraggap) {
1286                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1287                                                                    maxfraglen,
1288                                                     skb_transport_header(skb),
1289                                                                    fraggap, 0);
1290                                 skb_prev->csum = csum_sub(skb_prev->csum,
1291                                                           skb->csum);
1292                                 pskb_trim_unique(skb_prev, maxfraglen);
1293                         }
1294
1295                         /*
1296                          * Put the packet on the pending queue.
1297                          */
1298                         __skb_queue_tail(&sk->sk_write_queue, skb);
1299                         continue;
1300                 }
1301
1302                 if (len > size)
1303                         len = size;
1304
1305                 if (skb_append_pagefrags(skb, page, offset, len)) {
1306                         err = -EMSGSIZE;
1307                         goto error;
1308                 }
1309
1310                 if (skb->ip_summed == CHECKSUM_NONE) {
1311                         __wsum csum;
1312                         csum = csum_page(page, offset, len);
1313                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1314                 }
1315
1316                 skb->len += len;
1317                 skb->data_len += len;
1318                 skb->truesize += len;
1319                 atomic_add(len, &sk->sk_wmem_alloc);
1320                 offset += len;
1321                 size -= len;
1322         }
1323         return 0;
1324
1325 error:
1326         cork->length -= size;
1327         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1328         return err;
1329 }
1330
1331 static void ip_cork_release(struct inet_cork *cork)
1332 {
1333         cork->flags &= ~IPCORK_OPT;
1334         kfree(cork->opt);
1335         cork->opt = NULL;
1336         dst_release(cork->dst);
1337         cork->dst = NULL;
1338 }
1339
1340 /*
1341  *      Combined all pending IP fragments on the socket as one IP datagram
1342  *      and push them out.
1343  */
1344 struct sk_buff *__ip_make_skb(struct sock *sk,
1345                               struct flowi4 *fl4,
1346                               struct sk_buff_head *queue,
1347                               struct inet_cork *cork)
1348 {
1349         struct sk_buff *skb, *tmp_skb;
1350         struct sk_buff **tail_skb;
1351         struct inet_sock *inet = inet_sk(sk);
1352         struct net *net = sock_net(sk);
1353         struct ip_options *opt = NULL;
1354         struct rtable *rt = (struct rtable *)cork->dst;
1355         struct iphdr *iph;
1356         __be16 df = 0;
1357         __u8 ttl;
1358
1359         skb = __skb_dequeue(queue);
1360         if (!skb)
1361                 goto out;
1362         tail_skb = &(skb_shinfo(skb)->frag_list);
1363
1364         /* move skb->data to ip header from ext header */
1365         if (skb->data < skb_network_header(skb))
1366                 __skb_pull(skb, skb_network_offset(skb));
1367         while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1368                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1369                 *tail_skb = tmp_skb;
1370                 tail_skb = &(tmp_skb->next);
1371                 skb->len += tmp_skb->len;
1372                 skb->data_len += tmp_skb->len;
1373                 skb->truesize += tmp_skb->truesize;
1374                 tmp_skb->destructor = NULL;
1375                 tmp_skb->sk = NULL;
1376         }
1377
1378         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1379          * to fragment the frame generated here. No matter, what transforms
1380          * how transforms change size of the packet, it will come out.
1381          */
1382         skb->ignore_df = ip_sk_ignore_df(sk);
1383
1384         /* DF bit is set when we want to see DF on outgoing frames.
1385          * If ignore_df is set too, we still allow to fragment this frame
1386          * locally. */
1387         if (inet->pmtudisc == IP_PMTUDISC_DO ||
1388             inet->pmtudisc == IP_PMTUDISC_PROBE ||
1389             (skb->len <= dst_mtu(&rt->dst) &&
1390              ip_dont_fragment(sk, &rt->dst)))
1391                 df = htons(IP_DF);
1392
1393         if (cork->flags & IPCORK_OPT)
1394                 opt = cork->opt;
1395
1396         if (cork->ttl != 0)
1397                 ttl = cork->ttl;
1398         else if (rt->rt_type == RTN_MULTICAST)
1399                 ttl = inet->mc_ttl;
1400         else
1401                 ttl = ip_select_ttl(inet, &rt->dst);
1402
1403         iph = ip_hdr(skb);
1404         iph->version = 4;
1405         iph->ihl = 5;
1406         iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1407         iph->frag_off = df;
1408         iph->ttl = ttl;
1409         iph->protocol = sk->sk_protocol;
1410         ip_copy_addrs(iph, fl4);
1411         ip_select_ident(net, skb, sk);
1412
1413         if (opt) {
1414                 iph->ihl += opt->optlen>>2;
1415                 ip_options_build(skb, opt, cork->addr, rt, 0);
1416         }
1417
1418         skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1419         skb->mark = sk->sk_mark;
1420         /*
1421          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1422          * on dst refcount
1423          */
1424         cork->dst = NULL;
1425         skb_dst_set(skb, &rt->dst);
1426
1427         if (iph->protocol == IPPROTO_ICMP)
1428                 icmp_out_count(net, ((struct icmphdr *)
1429                         skb_transport_header(skb))->type);
1430
1431         ip_cork_release(cork);
1432 out:
1433         return skb;
1434 }
1435
1436 int ip_send_skb(struct net *net, struct sk_buff *skb)
1437 {
1438         int err;
1439
1440         err = ip_local_out(net, skb->sk, skb);
1441         if (err) {
1442                 if (err > 0)
1443                         err = net_xmit_errno(err);
1444                 if (err)
1445                         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1446         }
1447
1448         return err;
1449 }
1450
1451 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1452 {
1453         struct sk_buff *skb;
1454
1455         skb = ip_finish_skb(sk, fl4);
1456         if (!skb)
1457                 return 0;
1458
1459         /* Netfilter gets whole the not fragmented skb. */
1460         return ip_send_skb(sock_net(sk), skb);
1461 }
1462
1463 /*
1464  *      Throw away all pending data on the socket.
1465  */
1466 static void __ip_flush_pending_frames(struct sock *sk,
1467                                       struct sk_buff_head *queue,
1468                                       struct inet_cork *cork)
1469 {
1470         struct sk_buff *skb;
1471
1472         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1473                 kfree_skb(skb);
1474
1475         ip_cork_release(cork);
1476 }
1477
1478 void ip_flush_pending_frames(struct sock *sk)
1479 {
1480         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1481 }
1482
1483 struct sk_buff *ip_make_skb(struct sock *sk,
1484                             struct flowi4 *fl4,
1485                             int getfrag(void *from, char *to, int offset,
1486                                         int len, int odd, struct sk_buff *skb),
1487                             void *from, int length, int transhdrlen,
1488                             struct ipcm_cookie *ipc, struct rtable **rtp,
1489                             unsigned int flags)
1490 {
1491         struct inet_cork cork;
1492         struct sk_buff_head queue;
1493         int err;
1494
1495         if (flags & MSG_PROBE)
1496                 return NULL;
1497
1498         __skb_queue_head_init(&queue);
1499
1500         cork.flags = 0;
1501         cork.addr = 0;
1502         cork.opt = NULL;
1503         err = ip_setup_cork(sk, &cork, ipc, rtp);
1504         if (err)
1505                 return ERR_PTR(err);
1506
1507         err = __ip_append_data(sk, fl4, &queue, &cork,
1508                                &current->task_frag, getfrag,
1509                                from, length, transhdrlen, flags);
1510         if (err) {
1511                 __ip_flush_pending_frames(sk, &queue, &cork);
1512                 return ERR_PTR(err);
1513         }
1514
1515         return __ip_make_skb(sk, fl4, &queue, &cork);
1516 }
1517
1518 /*
1519  *      Fetch data from kernel space and fill in checksum if needed.
1520  */
1521 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1522                               int len, int odd, struct sk_buff *skb)
1523 {
1524         __wsum csum;
1525
1526         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1527         skb->csum = csum_block_add(skb->csum, csum, odd);
1528         return 0;
1529 }
1530
1531 /*
1532  *      Generic function to send a packet as reply to another packet.
1533  *      Used to send some TCP resets/acks so far.
1534  */
1535 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1536                            const struct ip_options *sopt,
1537                            __be32 daddr, __be32 saddr,
1538                            const struct ip_reply_arg *arg,
1539                            unsigned int len)
1540 {
1541         struct ip_options_data replyopts;
1542         struct ipcm_cookie ipc;
1543         struct flowi4 fl4;
1544         struct rtable *rt = skb_rtable(skb);
1545         struct net *net = sock_net(sk);
1546         struct sk_buff *nskb;
1547         int err;
1548         int oif;
1549
1550         if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1551                 return;
1552
1553         ipc.addr = daddr;
1554         ipc.opt = NULL;
1555         ipc.tx_flags = 0;
1556         ipc.ttl = 0;
1557         ipc.tos = -1;
1558
1559         if (replyopts.opt.opt.optlen) {
1560                 ipc.opt = &replyopts.opt;
1561
1562                 if (replyopts.opt.opt.srr)
1563                         daddr = replyopts.opt.opt.faddr;
1564         }
1565
1566         oif = arg->bound_dev_if;
1567         if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1568                 oif = skb->skb_iif;
1569
1570         flowi4_init_output(&fl4, oif,
1571                            IP4_REPLY_MARK(net, skb->mark),
1572                            RT_TOS(arg->tos),
1573                            RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1574                            ip_reply_arg_flowi_flags(arg),
1575                            daddr, saddr,
1576                            tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1577         security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1578         rt = ip_route_output_key(net, &fl4);
1579         if (IS_ERR(rt))
1580                 return;
1581
1582         inet_sk(sk)->tos = arg->tos;
1583
1584         sk->sk_priority = skb->priority;
1585         sk->sk_protocol = ip_hdr(skb)->protocol;
1586         sk->sk_bound_dev_if = arg->bound_dev_if;
1587         sk->sk_sndbuf = sysctl_wmem_default;
1588         err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1589                              len, 0, &ipc, &rt, MSG_DONTWAIT);
1590         if (unlikely(err)) {
1591                 ip_flush_pending_frames(sk);
1592                 goto out;
1593         }
1594
1595         nskb = skb_peek(&sk->sk_write_queue);
1596         if (nskb) {
1597                 if (arg->csumoffset >= 0)
1598                         *((__sum16 *)skb_transport_header(nskb) +
1599                           arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1600                                                                 arg->csum));
1601                 nskb->ip_summed = CHECKSUM_NONE;
1602                 ip_push_pending_frames(sk, &fl4);
1603         }
1604 out:
1605         ip_rt_put(rt);
1606 }
1607
1608 void __init ip_init(void)
1609 {
1610         ip_rt_init();
1611         inet_initpeers();
1612
1613 #if defined(CONFIG_IP_MULTICAST)
1614         igmp_mc_init();
1615 #endif
1616 }